id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
1804.07461-en-3
|
1804.07461-en-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2019}}$
<table>
<tbody>
<tr>
<th>Coarse-Grained Categories</th>
<th>Fine-Grained Categories</th>
</tr>
<tr>
<td>Lexical Semantics</td>
<td>Lexical Entailment, Morphological Negation, Factivity, Symmetry/Collectivity, Redundancy, Named Entities, Quantifiers</td>
</tr>
<tr>
<td>Predicate-Argument Structure</td>
<td>Core Arguments, Prepositional Phrases, Ellipsis/Implicits Anaphora/Coreference Active/Passive, Nominalization, Genitives/Partitives, Datives, Relative Clauses, Coordination Scope, Intersectivity, Restrictivity</td>
</tr>
<tr>
<td>Logic</td>
<td>Negation, Double Negation, Intervals/Numbers, Conjunction, Disjunction Conditionals, Universal, Existential, Temporal, Upward Monotone, Downward Monotone, Non-Monotone</td>
</tr>
<tr>
<td>Knowledge</td>
<td>Common World Knowledge sense</td>
</tr>
</tbody>
</table>
Table 2: The types of linguistic phenomena annotated in the diagnostic dataset, organized under four
major categories. For a description of each phenomenon, see Appendix E.
Each pair is human-annotated with a similarity score from 1 to 5; the task is to predict these scores
Follow common practice, we evaluate using Pearson and Spearman correlation coefficients.
# 3.3 INFERENCE TASKS
MNLI The Multi-Genre Natural Language Inference Corpus (Williams et al., 2018) is a crowdsourced collection of sentence pairs with textual entailment annotations. Given a premise sentence and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis (entail- $ment)$, contradicts the hypothesis $(contradiction)$, or neither $(neutral).$ The premise sentences are gathered from ten different sources, including transcribed speech, fiction, and government reports. $\tilde{\text{We use the standard test set, for which we obtained private labels from the authors, and evaluate}}$ on both the matched (in-domain) and mismatched (cross-domain) sections. We also use and recommend the SNLI corpus (Bowman et al., 2015) as 550k examples of auxiliary training data
$\mathbf{QNLI}$ The Stanford Question Answering Dataset (Rajpurkar et al.2016) is a question-answering dataset consisting of question-paragraph pairs, where one of the sentences in the paragraph (drawn from Wikipedia) contains the answer to the corresponding question (written by an annotator). We convert the task into sentence pair classification by forming a pair between each question and each sentence in the corresponding context, and filtering out pairs with low lexical overlap between the question and the context sentence. The task is to determine whether the context sentence contains the answer to the question. This modified version of the original task removes the requirement that the model select the exact answer, but also removes the simplifying assumptions that the answer is always present in the input and that lexical overlap is a reliable cue. This process of recasting existing datasets into NLI is similar to methods introduced in White et al.(2017) and expanded upon in Demszky et al. (2018). We call the converted dataset QNLI (Question-answering NÎLI).$^3$
RTE The Recognizing Textual Entailment (RTE) datasets come from a series of annual textual entailment challenges. We combine the data from RTE1 (Dagan et al., 2006), RTE2 (Bar Haim et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009).$^{4}$ Examples are constructed based on news and Wikipedia text. We convert all datasets to a two-class split, where for three-class datasets we collapse neutral and contradiction into not entailment, for consistency. $\mathbf{WNLI}$ The Winograd Schema Challenge (Levesque et al., 2011) is a reading comprehension task in which a system must read a sentence with a pronoun and select the referent of that pronoun from
a list of choices. The examples are manually constructed to foil simple statistical methods: Each
one is contingent on contextual information provided by a single word or phrase in the sentence.
To convert the problem into sentence pair classification, we construct sentence pairs by replacing
the ambiguous pronoun with each possible referent. The task is to predict if the sentence with the
$^{3}$An earlier release of QNLI had an artifact where the task could be modeled and solved as an easier task
than we describe here. We have since released an updated version of QNLI that removes this possibility.
$^{4}$RTE4 is not publicly available, while RTE6 and RTE7 do not fit the standard NLI task.
4
|
[
[
0.17236328125,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.17236328125,
0.04736328125
],
[
0.1829833984375,
0.1060791015625,
0.361083984375,
0.1060791015625,
0.361083984375,
0.11932373046875,
0.1829833984375,
0.11932373046875
],
[
0.3798828125,
0.1060791015625,
0.53759765625,
0.1060791015625,
0.53759765625,
0.11932373046875,
0.3798828125,
0.11932373046875
],
[
0.1829833984375,
0.1307373046875,
0.295654296875,
0.1307373046875,
0.295654296875,
0.1458740234375,
0.1829833984375,
0.1458740234375
],
[
0.37744140625,
0.125,
0.705078125,
0.125,
0.705078125,
0.1383056640625,
0.37744140625,
0.1383056640625
],
[
0.378173828125,
0.1383056640625,
0.76318359375,
0.1383056640625,
0.76318359375,
0.1514892578125,
0.378173828125,
0.1514892578125
],
[
0.378173828125,
0.1558837890625,
0.72119140625,
0.1558837890625,
0.72119140625,
0.169189453125,
0.378173828125,
0.169189453125
],
[
0.1829833984375,
0.1685791015625,
0.366943359375,
0.1685791015625,
0.366943359375,
0.1817626953125,
0.1829833984375,
0.1817626953125
],
[
0.378173828125,
0.17041015625,
0.705078125,
0.17041015625,
0.705078125,
0.18115234375,
0.378173828125,
0.18115234375
],
[
0.3798828125,
0.1817626953125,
0.650390625,
0.1817626953125,
0.650390625,
0.1956787109375,
0.3798828125,
0.1956787109375
],
[
0.3798828125,
0.1932373046875,
0.6611328125,
0.1932373046875,
0.6611328125,
0.20703125,
0.3798828125,
0.20703125
],
[
0.3798828125,
0.212158203125,
0.81884765625,
0.212158203125,
0.81884765625,
0.2266845703125,
0.3798828125,
0.2266845703125
],
[
0.181396484375,
0.2183837890625,
0.223876953125,
0.2183837890625,
0.223876953125,
0.2342529296875,
0.181396484375,
0.2342529296875
],
[
0.381591796875,
0.2259521484375,
0.775390625,
0.2259521484375,
0.775390625,
0.2392578125,
0.381591796875,
0.2392578125
],
[
0.378173828125,
0.2374267578125,
0.60888671875,
0.2392578125,
0.60888671875,
0.25244140625,
0.378173828125,
0.250732421875
],
[
0.1829833984375,
0.2568359375,
0.2548828125,
0.2568359375,
0.2548828125,
0.271484375,
0.1829833984375,
0.271484375
],
[
0.3798828125,
0.2568359375,
0.5888671875,
0.2568359375,
0.5888671875,
0.270751953125,
0.3798828125,
0.270751953125
],
[
0.17236328125,
0.285400390625,
0.82275390625,
0.285888671875,
0.82275390625,
0.300537109375,
0.17236328125,
0.2998046875
],
[
0.1732177734375,
0.300537109375,
0.65869140625,
0.300537109375,
0.65869140625,
0.314453125,
0.1732177734375,
0.314453125
],
[
0.1732177734375,
0.339111328125,
0.8203125,
0.339111328125,
0.8203125,
0.352783203125,
0.1732177734375,
0.352783203125
],
[
0.1732177734375,
0.352783203125,
0.77294921875,
0.354248046875,
0.77294921875,
0.36865234375,
0.1732177734375,
0.367431640625
],
[
0.1748046875,
0.386474609375,
0.2091064453125,
0.386474609375,
0.2091064453125,
0.397705078125,
0.1748046875,
0.397705078125
],
[
0.2017822265625,
0.385009765625,
0.346435546875,
0.385009765625,
0.346435546875,
0.3984375,
0.2017822265625,
0.3984375
],
[
0.1748046875,
0.410888671875,
0.82177734375,
0.410888671875,
0.82177734375,
0.4248046875,
0.1748046875,
0.4248046875
],
[
0.1748046875,
0.425537109375,
0.82275390625,
0.425537109375,
0.82275390625,
0.439453125,
0.1748046875,
0.439453125
],
[
0.1748046875,
0.439453125,
0.82275390625,
0.439453125,
0.82275390625,
0.45263671875,
0.1748046875,
0.45263671875
],
[
0.1748046875,
0.453857421875,
0.82421875,
0.453857421875,
0.82421875,
0.46728515625,
0.1748046875,
0.46728515625
],
[
0.1748046875,
0.46728515625,
0.82275390625,
0.46728515625,
0.82275390625,
0.48095703125,
0.1748046875,
0.48095703125
],
[
0.1748046875,
0.48095703125,
0.82421875,
0.48095703125,
0.82421875,
0.494384765625,
0.1748046875,
0.494384765625
],
[
0.1732177734375,
0.494384765625,
0.82177734375,
0.494384765625,
0.82177734375,
0.50830078125,
0.1732177734375,
0.50830078125
],
[
0.1748046875,
0.5087890625,
0.7646484375,
0.5087890625,
0.7646484375,
0.52294921875,
0.1748046875,
0.52294921875
],
[
0.1748046875,
0.53857421875,
0.82275390625,
0.53857421875,
0.82275390625,
0.5517578125,
0.1748046875,
0.5517578125
],
[
0.1748046875,
0.5517578125,
0.82421875,
0.5517578125,
0.82421875,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.1732177734375,
0.56494140625,
0.82177734375,
0.56494140625,
0.82177734375,
0.5791015625,
0.1732177734375,
0.5791015625
],
[
0.1748046875,
0.580078125,
0.82421875,
0.580078125,
0.82421875,
0.59326171875,
0.1748046875,
0.59326171875
],
[
0.1748046875,
0.59326171875,
0.82421875,
0.59326171875,
0.82421875,
0.60693359375,
0.1748046875,
0.60693359375
],
[
0.1748046875,
0.60791015625,
0.82275390625,
0.60791015625,
0.82275390625,
0.6220703125,
0.1748046875,
0.6220703125
],
[
0.1748046875,
0.6220703125,
0.82421875,
0.6220703125,
0.82421875,
0.63525390625,
0.1748046875,
0.63525390625
],
[
0.1748046875,
0.63525390625,
0.82275390625,
0.63525390625,
0.82275390625,
0.6484375,
0.1748046875,
0.6484375
],
[
0.17236328125,
0.6484375,
0.8203125,
0.6484375,
0.8203125,
0.662109375,
0.17236328125,
0.662109375
],
[
0.1748046875,
0.66357421875,
0.82275390625,
0.66357421875,
0.82275390625,
0.6767578125,
0.1748046875,
0.6767578125
],
[
0.1748046875,
0.67529296875,
0.7998046875,
0.67529296875,
0.7998046875,
0.689453125,
0.1748046875,
0.689453125
],
[
0.1748046875,
0.70556640625,
0.82421875,
0.70556640625,
0.82421875,
0.72021484375,
0.1748046875,
0.72021484375
],
[
0.1748046875,
0.72021484375,
0.82421875,
0.72021484375,
0.82421875,
0.734375,
0.1748046875,
0.734375
],
[
0.1748046875,
0.734375,
0.82275390625,
0.734375,
0.82275390625,
0.74755859375,
0.1748046875,
0.74755859375
],
[
0.1748046875,
0.74853515625,
0.82275390625,
0.74853515625,
0.82275390625,
0.76220703125,
0.1748046875,
0.76220703125
],
[
0.1748046875,
0.76220703125,
0.814453125,
0.76220703125,
0.814453125,
0.77587890625,
0.1748046875,
0.77587890625
],
[
0.1748046875,
0.79150390625,
0.82275390625,
0.79150390625,
0.82275390625,
0.80615234375,
0.1748046875,
0.80615234375
],
[
0.1732177734375,
0.80615234375,
0.82275390625,
0.80615234375,
0.82275390625,
0.8193359375,
0.1732177734375,
0.8193359375
],
[
0.1748046875,
0.8193359375,
0.82421875,
0.8193359375,
0.82421875,
0.83349609375,
0.1748046875,
0.83349609375
],
[
0.176513671875,
0.83544921875,
0.8203125,
0.83544921875,
0.8203125,
0.84521484375,
0.176513671875,
0.84521484375
],
[
0.1748046875,
0.8466796875,
0.82275390625,
0.8466796875,
0.82275390625,
0.85986328125,
0.1748046875,
0.85986328125
],
[
0.1748046875,
0.861328125,
0.82421875,
0.861328125,
0.82421875,
0.875,
0.1748046875,
0.875
],
[
0.19775390625,
0.88525390625,
0.82421875,
0.88525390625,
0.82421875,
0.8984375,
0.19775390625,
0.8984375
],
[
0.17236328125,
0.89794921875,
0.78515625,
0.89794921875,
0.78515625,
0.9111328125,
0.17236328125,
0.9111328125
],
[
0.1968994140625,
0.9111328125,
0.71337890625,
0.9111328125,
0.71337890625,
0.9248046875,
0.1968994140625,
0.9248046875
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2019",
"Coarse-Grained Categories",
"Fine-Grained Categories",
"Lexical Semantics",
"Lexical Entailment, Morphological Negation, Factivity",
"SymmetryCollectivity,Redundancy,Named Entities,Quantifier",
"Core Arguments, Prepositional Phrases, Ellipsis/lmplicits.",
"Predicate-Argument Structure",
"Anaphora/CoreferenceActive/Passive,Nominalization",
"Genitives/Partitives, Datives, Relative Clauses",
"Coordination Scope, Intersectivity,Restrictivity",
"Negation, Double Negation, Intervals/Numbers, Conjunction, Disjunction,",
"Logic",
"Conditionals, Universal, Existential, Temporal, Upward Monotone,",
"Downward Monotone,Non-Monotone",
"Knowledge",
"Common Sense, World Knowledge",
"Table 2: The types of linguistic phenomena annotated in the diagnostic dataset, organized under four",
"major categories. For a description of each phenomenon, see Appendix E.",
"Each pair is human-annotated with a similarity score from 1 to 5; the task is to predict these scores",
"Follow common practice, we evaluate using Pearson and Spearman correlation coefficients.",
"3.3",
"INFERENCETASKS",
"MNLI The Multi-Genre Natural Language Inference Corpus (Williams et al., 2018) is a crowd-",
"sourced collection of sentence pairs with textual entailment annotations. Given a premise sentence",
"and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis (entail-",
"ment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are",
"gathered from ten different sources, including transcribed speech, fiction, and government reports.",
"We use the standard test set, for which we obtained private labels from the authors, and evaluate",
"on both the matched (in-domain) and mismatched (cross-domain) sections. We also use and recom-",
"mend the SNLI corpus (Bowman et al., 2015) as 550k examples of auxiliary training data.",
"QNLI The Stanford Question Answering Dataset (Rajpurkar et al. 2016) is a question-answering",
"dataset consisting of question-paragraph pairs, where one of the sentences in the paragraph (drawn",
"from Wikipedia) contains the answer to the corresponding question (written by an annotator). We",
"convert the task into sentence pair classification by forming a pair between each question and each",
"sentence in the corresponding context, and filtering out pairs with low lexical overlap between the",
"question and the context sentence. The task is to determine whether the context sentence contains",
"the answer to the question. This modified version of the original task removes the requirement that",
"the model select the exact answer, but also removes the simplifying assumptions that the answer",
"is always present in the input and that lexical overlap is a reliable cue. This process of recasting",
"existing datasets into NLI is similar to methods introduced in White et al. (2017) and expanded",
"upon in Demszky et al. (2018). We call the converted dataset QNLI (Question-answering NLI).",
"RTE The Recognizing Textual Entailment (RTE) datasets come from a series of annual textual",
"entailment challenges. We combine the data from RTE1 (Dagan et al., 2006), RTE2 (Bar Haim",
"et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009).4 Examples are",
"constructed based on news and Wikipedia text. We convert all datasets to a two-class split, where",
"for three-class datasets we collapse neutral and contradiction into not_entailment, for consistency.",
"WNLI The Winograd Schema Challenge (Levesque et al., 2011) is a reading comprehension task",
"in which a system must read a sentence with a pronoun and select the referent of that pronoun from",
"a list of choices. The examples are manually constructed to foil simple statistical methods: Each",
"oneiscontingentoncontextualinformationprovidedbyasinglewordorphraseinthesentence",
"To convert the problem into sentence pair classification, we construct sentence pairs by replacing",
"the ambiguous pronoun with each possible referent. The task is to predict if the sentence with the",
"3 An earlier release of QNLI had an artifact where the task could be modeled and solved as an easier task",
"than we describe here. We have since released an updated version of QNLI that removes this possibility",
"ARTE4 is not publicly available, while RTE6 and RTE7 do not fit the standard NLI task.",
"4"
] |
{
"arxiv_id": "1804.07461",
"language": "en",
"page_idx": 3
}
|
||
1804.07461-en-4
|
1804.07461-en-4.jpg
|
Published as a conference paper at ICLR 2019
<table>
<tbody>
<tr>
<th>$\textbf{Tags}$</th>
<th>Sentence 1</th>
<th>Sentence 2</th>
<th>Fwd</th>
<th>Bwd</th>
</tr>
<tr>
<td>Lexical Entailment $t~(Lexi-$ $cal~Semantics),~Downward$ ${Monotone}\left(Logic\right)$</td>
<td>The timing of the meeting has not been set, according to a Starbucks s spokesper- $son.$</td>
<td>The timing of the meet- ing has $\frac{1}{2}=\frac{1}{2}$ not $\frac{1}{2}=\frac{1}{2}=\frac{1}{2}$ been consid- $ered$, according to a Star- bucks spokesperson.</td>
<td>$N$</td>
<td>$E$</td>
</tr>
<tr>
<td>Universal Quantifers $Logic)$</td>
<td>Our deepest sympathies are with all those affected $1by$ this accident.</td>
<td>Our deepest sympathies are with a victim who was af- fected by this accident.</td>
<td></td>
<td>$N$</td>
</tr>
<tr>
<td>Quantifers $(Lexical$ $Se-$ $mantics)$, , Double Negation $(Logic)$</td>
<td>I have never seen 1 a hum- mingbird not flying.</td>
<td>I have never seen a hum- ${\mathrm{mingbird}}.$</td>
<td>$N$</td>
<td>$E$</td>
</tr>
</tbody>
</table>
Table 3: Examples from the diagnostic set. Fwd (resp. $Bwd)$ denotes the label when sentence 1 (resp. sentence 2)is the premise. Labels are entailment (E), neutral (N), or contradiction (C). Examples are tagged with the phenomena they demonstrate, and each phenomenon belongs to one of four broad categories (in parentheses).
pronoun substituted is entailed by the original sentence. We use a small evaluation set consisting of new examples derived from fiction books$^{5}$ that was shared privately by the authors of the original corpus. While the included training set is balanced between two classes, the test set is imbalanced $\hat{\text{between them }}(65\%$ not entailment). Also, due to a data quirk, the development set is adversarial: hypotheses are sometimes shared between training and development examples, so if a model memorizes the training examples, they will predict the wrong label on corresponding development set example. As with QNLI, each example is evaluated separately, so there is not a systematic correspondence between a model's score on this task and its score on the unconverted original task. We call converted dataset WNLI (Winograd NLI).
## 3.4 EVALUATION
The GLUE benchmark follows the same evaluation model as SemEval and Kaggle. To evaluate a system on the benchmark, one must run the system on the provided test data for the tasks, then upload the results to the website gluebenchmark.com for scoring. The benchmark site shows per-task scores and a macro-average of those scores to determine a system's position on the leader- $\hat{\text{board. For tasks with multiple metrics (e.g., accuracy and F1), we use an unweighted average of}}$ the metrics as the score for the task when computing the overall macro-average. The website also provides fine- and coarse-grained results on the diagnostic dataset. See Appendix D for details.
## 4 DIAGNOSTIC DATASET
Drawing inspiration from the FraCaS suite (Cooper et al., 1996) and the recent Build-It-Break-It competition $\hat{\text{(Ettinger et al., 2017), we include a small, manually-curated test set for the analysis of}}$ system performance. While the main benchmark mostly reflects an application-driven distribution $\hat{\text{of examples, our diagnostic dataset highlights a pre-defined set of phenomena that we believe are}}$ $\hat{\text{interesting and important for models to capture. }}\hat{\text{We show the full set of phenomena in Table 2.}}$
Each diagnostic example is an NLI sentence pair with tags for the phenomena demonstrated. The NLI task is well-suited to this kind of analysis, as it can easily evaluate the full set of skills involved in (ungrounded) sentence understanding, from resolution of syntactic ambiguity to pragmatic reasoning with world knowledge. We ensure the data is reasonably diverse by producing examples for a variety of linguistic phenomena and basing our examples on naturally-occurring sentences from several domains (news, Reddit, Wikipedia, academic papers). This approaches differs from that of FraCaS, which was designed to test linguistic theories with a minimal and uniform set of examples. A sample from our dataset is shown in Table 3.
See similar examples at cs.nyu.edu/faculty/davise/papers/Winogradschemas/
$WS.html$
5
|
[
[
0.1732177734375,
0.0328369140625,
0.4794921875,
0.032196044921875,
0.4794921875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.734375,
0.10479736328125,
0.7685546875,
0.10479736328125,
0.7685546875,
0.121826171875,
0.734375,
0.121826171875
],
[
0.78125,
0.10479736328125,
0.81787109375,
0.10479736328125,
0.81787109375,
0.12060546875,
0.78125,
0.12060546875
],
[
0.365234375,
0.1060791015625,
0.43701171875,
0.1060791015625,
0.43701171875,
0.12060546875,
0.365234375,
0.12060546875
],
[
0.181396484375,
0.1060791015625,
0.2181396484375,
0.1060791015625,
0.2181396484375,
0.121826171875,
0.181396484375,
0.121826171875
],
[
0.55126953125,
0.1060791015625,
0.62353515625,
0.1060791015625,
0.62353515625,
0.12060546875,
0.55126953125,
0.12060546875
],
[
0.1829833984375,
0.125,
0.351318359375,
0.125,
0.351318359375,
0.1395263671875,
0.1829833984375,
0.1395263671875
],
[
0.366943359375,
0.125,
0.53759765625,
0.125,
0.53759765625,
0.1395263671875,
0.366943359375,
0.1395263671875
],
[
0.5498046875,
0.125,
0.72119140625,
0.125,
0.72119140625,
0.1395263671875,
0.5498046875,
0.1395263671875
],
[
0.734375,
0.126953125,
0.7490234375,
0.126953125,
0.7490234375,
0.136962890625,
0.734375,
0.136962890625
],
[
0.78125,
0.125,
0.79833984375,
0.125,
0.79833984375,
0.1395263671875,
0.78125,
0.1395263671875
],
[
0.1829833984375,
0.1383056640625,
0.3544921875,
0.1383056640625,
0.3544921875,
0.1514892578125,
0.1829833984375,
0.1514892578125
],
[
0.365234375,
0.136962890625,
0.53759765625,
0.1383056640625,
0.53759765625,
0.15283203125,
0.365234375,
0.1514892578125
],
[
0.54833984375,
0.1383056640625,
0.7197265625,
0.136962890625,
0.7197265625,
0.1514892578125,
0.54833984375,
0.15283203125
],
[
0.181396484375,
0.149658203125,
0.294189453125,
0.15087890625,
0.294189453125,
0.1654052734375,
0.181396484375,
0.1641845703125
],
[
0.36767578125,
0.15283203125,
0.53515625,
0.15283203125,
0.53515625,
0.1634521484375,
0.36767578125,
0.1634521484375
],
[
0.55126953125,
0.1514892578125,
0.71826171875,
0.1514892578125,
0.71826171875,
0.1622314453125,
0.55126953125,
0.1622314453125
],
[
0.36767578125,
0.1654052734375,
0.396240234375,
0.1654052734375,
0.396240234375,
0.177978515625,
0.36767578125,
0.177978515625
],
[
0.55126953125,
0.1641845703125,
0.67333984375,
0.1641845703125,
0.67333984375,
0.177978515625,
0.55126953125,
0.177978515625
],
[
0.365234375,
0.179931640625,
0.5361328125,
0.18115234375,
0.5361328125,
0.1956787109375,
0.365234375,
0.1944580078125
],
[
0.5498046875,
0.179931640625,
0.72119140625,
0.18115234375,
0.72119140625,
0.1956787109375,
0.5498046875,
0.1944580078125
],
[
0.78125,
0.18115234375,
0.79833984375,
0.18115234375,
0.79833984375,
0.1956787109375,
0.78125,
0.1956787109375
],
[
0.283447265625,
0.18310546875,
0.35302734375,
0.18310546875,
0.35302734375,
0.1944580078125,
0.283447265625,
0.1944580078125
],
[
0.1829833984375,
0.18115234375,
0.2467041015625,
0.18115234375,
0.2467041015625,
0.1956787109375,
0.1829833984375,
0.1956787109375
],
[
0.734375,
0.18310546875,
0.74853515625,
0.18310546875,
0.74853515625,
0.1932373046875,
0.734375,
0.1932373046875
],
[
0.181396484375,
0.1944580078125,
0.2320556640625,
0.1944580078125,
0.2320556640625,
0.208984375,
0.181396484375,
0.208984375
],
[
0.366943359375,
0.1932373046875,
0.53515625,
0.1932373046875,
0.53515625,
0.20703125,
0.366943359375,
0.20703125
],
[
0.55126953125,
0.1944580078125,
0.7197265625,
0.1944580078125,
0.7197265625,
0.2083740234375,
0.55126953125,
0.2083740234375
],
[
0.365234375,
0.2052001953125,
0.44775390625,
0.20703125,
0.44775390625,
0.2215576171875,
0.365234375,
0.2197265625
],
[
0.5498046875,
0.205810546875,
0.68896484375,
0.20703125,
0.68896484375,
0.2215576171875,
0.5498046875,
0.2203369140625
],
[
0.1846923828125,
0.2259521484375,
0.2548828125,
0.2259521484375,
0.2548828125,
0.240478515625,
0.1846923828125,
0.240478515625
],
[
0.365234375,
0.2259521484375,
0.5390625,
0.2259521484375,
0.5390625,
0.2392578125,
0.365234375,
0.2392578125
],
[
0.54833984375,
0.2259521484375,
0.72119140625,
0.2259521484375,
0.72119140625,
0.2392578125,
0.54833984375,
0.2392578125
],
[
0.78125,
0.2247314453125,
0.79833984375,
0.2247314453125,
0.79833984375,
0.240478515625,
0.78125,
0.240478515625
],
[
0.32275390625,
0.2266845703125,
0.35302734375,
0.2266845703125,
0.35302734375,
0.238037109375,
0.32275390625,
0.238037109375
],
[
0.265625,
0.2266845703125,
0.32177734375,
0.2266845703125,
0.32177734375,
0.238037109375,
0.265625,
0.238037109375
],
[
0.736328125,
0.2266845703125,
0.7490234375,
0.2266845703125,
0.7490234375,
0.238037109375,
0.736328125,
0.238037109375
],
[
0.181396484375,
0.2374267578125,
0.3544921875,
0.238037109375,
0.3544921875,
0.25244140625,
0.181396484375,
0.251953125
],
[
0.365234375,
0.2392578125,
0.488525390625,
0.2392578125,
0.488525390625,
0.25244140625,
0.365234375,
0.25244140625
],
[
0.5498046875,
0.238037109375,
0.61279296875,
0.238037109375,
0.61279296875,
0.25244140625,
0.5498046875,
0.25244140625
],
[
0.1829833984375,
0.250732421875,
0.2320556640625,
0.250732421875,
0.2320556640625,
0.26513671875,
0.1829833984375,
0.26513671875
],
[
0.1748046875,
0.27978515625,
0.82275390625,
0.27978515625,
0.82275390625,
0.294189453125,
0.1748046875,
0.294189453125
],
[
0.1748046875,
0.294189453125,
0.82275390625,
0.294189453125,
0.82275390625,
0.30810546875,
0.1748046875,
0.30810546875
],
[
0.1732177734375,
0.30810546875,
0.82421875,
0.306884765625,
0.82421875,
0.3212890625,
0.1732177734375,
0.322509765625
],
[
0.1748046875,
0.3212890625,
0.4453125,
0.3212890625,
0.4453125,
0.3359375,
0.1748046875,
0.3359375
],
[
0.1732177734375,
0.365478515625,
0.826171875,
0.3642578125,
0.826171875,
0.37890625,
0.1732177734375,
0.380126953125
],
[
0.1748046875,
0.380126953125,
0.82421875,
0.380126953125,
0.82421875,
0.393310546875,
0.1748046875,
0.393310546875
],
[
0.1748046875,
0.393310546875,
0.82421875,
0.393310546875,
0.82421875,
0.4072265625,
0.1748046875,
0.4072265625
],
[
0.1748046875,
0.4072265625,
0.82421875,
0.4072265625,
0.82421875,
0.42041015625,
0.1748046875,
0.42041015625
],
[
0.1748046875,
0.421630859375,
0.82177734375,
0.421630859375,
0.82177734375,
0.43505859375,
0.1748046875,
0.43505859375
],
[
0.1732177734375,
0.43505859375,
0.82275390625,
0.43505859375,
0.82275390625,
0.448974609375,
0.1732177734375,
0.448974609375
],
[
0.1748046875,
0.448974609375,
0.82177734375,
0.448974609375,
0.82177734375,
0.462158203125,
0.1748046875,
0.462158203125
],
[
0.1732177734375,
0.462158203125,
0.82421875,
0.4609375,
0.82421875,
0.47607421875,
0.1732177734375,
0.4765625
],
[
0.176513671875,
0.477783203125,
0.4755859375,
0.477783203125,
0.4755859375,
0.488037109375,
0.176513671875,
0.488037109375
],
[
0.176513671875,
0.51220703125,
0.21728515625,
0.51220703125,
0.21728515625,
0.5234375,
0.176513671875,
0.5234375
],
[
0.2091064453125,
0.51220703125,
0.306396484375,
0.51220703125,
0.306396484375,
0.5234375,
0.2091064453125,
0.5234375
],
[
0.1748046875,
0.53857421875,
0.82421875,
0.53857421875,
0.82421875,
0.5517578125,
0.1748046875,
0.5517578125
],
[
0.1748046875,
0.55322265625,
0.82421875,
0.55322265625,
0.82421875,
0.56640625,
0.1748046875,
0.56640625
],
[
0.1732177734375,
0.5673828125,
0.82421875,
0.5673828125,
0.82421875,
0.58056640625,
0.1732177734375,
0.58056640625
],
[
0.1748046875,
0.58056640625,
0.82421875,
0.58056640625,
0.82421875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.1748046875,
0.5947265625,
0.826171875,
0.5947265625,
0.826171875,
0.60791015625,
0.1748046875,
0.60791015625
],
[
0.1748046875,
0.609375,
0.82421875,
0.609375,
0.82421875,
0.62255859375,
0.1748046875,
0.62255859375
],
[
0.1732177734375,
0.62255859375,
0.7958984375,
0.62255859375,
0.7958984375,
0.63623046875,
0.1732177734375,
0.63623046875
],
[
0.176513671875,
0.66357421875,
0.1895751953125,
0.66357421875,
0.1895751953125,
0.67236328125,
0.176513671875,
0.67236328125
],
[
0.2034912109375,
0.6611328125,
0.396240234375,
0.6611328125,
0.396240234375,
0.6748046875,
0.2034912109375,
0.6748046875
],
[
0.1748046875,
0.693359375,
0.82421875,
0.693359375,
0.82421875,
0.70703125,
0.1748046875,
0.70703125
],
[
0.1748046875,
0.70849609375,
0.82421875,
0.70849609375,
0.82421875,
0.7216796875,
0.1748046875,
0.7216796875
],
[
0.1732177734375,
0.7216796875,
0.82421875,
0.7216796875,
0.82421875,
0.73486328125,
0.1732177734375,
0.73486328125
],
[
0.1732177734375,
0.736328125,
0.82421875,
0.736328125,
0.82421875,
0.75,
0.1732177734375,
0.75
],
[
0.17236328125,
0.75,
0.79345703125,
0.75,
0.79345703125,
0.76318359375,
0.17236328125,
0.76318359375
],
[
0.1748046875,
0.76953125,
0.82275390625,
0.76953125,
0.82275390625,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1732177734375,
0.7841796875,
0.82275390625,
0.7841796875,
0.82275390625,
0.79736328125,
0.1732177734375,
0.79736328125
],
[
0.1732177734375,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.1732177734375,
0.81298828125
],
[
0.1732177734375,
0.81298828125,
0.82421875,
0.81298828125,
0.82421875,
0.8271484375,
0.1732177734375,
0.8271484375
],
[
0.1748046875,
0.8271484375,
0.82421875,
0.8271484375,
0.82421875,
0.84033203125,
0.1748046875,
0.84033203125
],
[
0.176513671875,
0.84228515625,
0.82421875,
0.84228515625,
0.82421875,
0.85302734375,
0.176513671875,
0.85302734375
],
[
0.1732177734375,
0.85302734375,
0.82421875,
0.85400390625,
0.82421875,
0.86865234375,
0.1732177734375,
0.8671875
],
[
0.1732177734375,
0.8671875,
0.483642578125,
0.8671875,
0.483642578125,
0.880859375,
0.1732177734375,
0.880859375
],
[
0.193603515625,
0.896484375,
0.2296142578125,
0.896484375,
0.2296142578125,
0.912109375,
0.193603515625,
0.912109375
],
[
0.2344970703125,
0.8994140625,
0.281005859375,
0.8994140625,
0.281005859375,
0.9111328125,
0.2344970703125,
0.9111328125
],
[
0.293212890625,
0.90087890625,
0.35302734375,
0.90087890625,
0.35302734375,
0.9111328125,
0.293212890625,
0.9111328125
],
[
0.36279296875,
0.9013671875,
0.378173828125,
0.9013671875,
0.378173828125,
0.91015625,
0.36279296875,
0.91015625
],
[
0.3857421875,
0.8994140625,
0.82275390625,
0.8984375,
0.82275390625,
0.912109375,
0.3857421875,
0.91357421875
],
[
0.1748046875,
0.91357421875,
0.23779296875,
0.91357421875,
0.23779296875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.494384765625,
0.94775390625,
0.50634765625,
0.94775390625,
0.50634765625,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2019",
"Fwd",
"Bwd",
"Sentence 1",
"Tags",
"Sentence 2",
"Lexical Entailment (Lexi-",
"The timing of the meeting",
"The timing of the meet-",
"N",
"E",
"cal Semantics),Downward",
"has not been set, according",
"ing has not been consid-",
"Monotone (Logic)",
"toa Starbucks spokesper-",
"ered, according to a Star-",
"son.",
"bucks spokesperson.",
" Our deepest sympathies are",
"Our deepest sympathies are",
"N",
"Quantifiers",
"Universal",
"E",
"(Logic)",
"with all those affected by",
"with a victim who was af-",
"this accident.",
"fected by this accident.",
"Quantifiers",
"I have never seen a hum-",
"I have never seen a hum-",
"E",
"Se-",
"(Lexical",
"N",
"mantics),DoubleNegation",
"mingbird not flying.",
"mingbird.",
"(Logic)",
"Table 3: Examples from the diagnostic set. Fwd (resp. Bwd) denotes the label when sentence 1",
"(resp. sentence 2) is the premise. Labels are entailment (E), neutral (N), or contradiction (C).",
"Examples are tagged with the phenomena they demonstrate, and each phenomenon belongs to one",
"of four broad categories (in parentheses).",
"pronoun substituted is entailed by the original sentence. We use a small evaluation set consisting of",
"new examples derived from fiction books that was shared privately by the authors of the original",
"corpus. While the included training set is balanced between two classes, the test set is imbalanced",
"between them (65% not entailment). Also, due to a data quirk, the development set is adversarial:",
"hypotheses are sometimes shared between training and development examples, so if a model mem-",
"orizes the training examples, they will predict the wrong label on corresponding development set",
"example. As with QNLI, each example is evaluated separately, so there is not a systematic corre-",
"spondence between a model's score on this task and its score on the unconverted original task. We",
"callconverted dataset WNLI (Winograd NLI)",
"3.4",
"EVALUATION",
"The GLUE benchmark follows the same evaluation model as SemEval and Kaggle. To evaluate",
"a system on the benchmark, one must run the system on the provided test data for the tasks, then",
"upload the results to the website gluebenchmark.com for scoring. The benchmark site shows",
"per-task scores and a macro-average of those scores to determine a system's position on the leader-",
"board. For tasks with multiple metrics (e.g., accuracy and Fl), we use an unweighted average of",
"the metrics as the score for the task when computing the overall macro-average. The website also",
"provides fine- and coarse-grained results on the diagnostic dataset. See Appendix D for details.",
"4",
"DIAGNOSTIC DATASET",
"Drawing inspiration from the FraCaS suite (Cooper et al., 1996) and the recent Build-It-Break-It",
"competition (Ettinger et al., 2017), we include a small, manually-curated test set for the analysis of",
"system performance. While the main benchmark mostly reflects an application-driven distribution",
"of examples, our diagnostic dataset highlights a pre-defined set of phenomena that we believe are",
"interesting and important for models to capture. We show the full set of phenomena in Table 2.",
"Each diagnostic example is an NLI sentence pair with tags for the phenomena demonstrated. The",
"NLI task is well-suited to this kind of analysis, as it can easily evaluate the full set of skills involved",
"in (ungrounded) sentence understanding,from resolution of syntactic ambiguity to pragmatic rea-",
"soning with world knowledge. We ensure the data is reasonably diverse by producing examples for",
"a variety of linguistic phenomena and basing our examples on naturally-occurring sentences from",
"several domains(news,Reddit,Wikipedia,academicpapers).This approaches differsfromthatof",
"FraCaS, which was designed to test linguistic theories with a minimal and uniform set of examples.",
"A sample from our dataset is shown in Table 3.",
"'See",
"similar",
"examples",
"at",
"cs.nyu.edu/faculty/davise/papers/winogradschemas/",
"wS.html",
"5"
] |
{
"arxiv_id": "1804.07461",
"language": "en",
"page_idx": 4
}
|
||
2209.14389-en-0
|
2209.14389-en-0.jpg
|
$\textbf{Downstream Datasets Make Surprisingly Good Pretraining Corpora}$
Kundan Krishna, Saurabh Garg, Jeffrey P. Bigham, Zachary C. Lipton
Carnegie Mellon University
$\{$kundank,sgarg2,jbigham,zlipton}eandrew.cmu.edu
### Abstract
For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gains are attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning. In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around $10\times-500\times$ less data), outperforming the latter on 7 and 5 datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks, including the GLUE benchmark. Self-pretraining also provides benefits on structured output prediction tasks such as question answering and commonsense inference, often providing more than 50% improvements compared to standard pretraining. Our results hint that often performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.$^{\mathrm{l}}$
## 1 Introduction
For training predictive models operating on natural language data, the current best practice is to pretrain models on large unlabeled upstream corpora to optimize self-supervised objectives, for example, masked language modeling (MLM); the resulting weights are then used to initialize models that
$^{1}$Pretrained models can be downloaded from https://
github.com/acmi-lab/self-pretrain
are subsequently trained (fnetuned) on the labeled downstream data available for the task at hand. Large-scale pretrained models typically provide significant performance boosts when compared to models trained directly on the downstream task (with random initializations) (Peters et al., 2018; Devlin et al., 2019; Chiang and Lee, 2020; Krishna et al., 2021). Upstream corpora tend to be significantly larger than the downstream corpora and the success of this approach is often attributed to its ability to leverage these massive upstream corpora (Liu et al., 2019; Yang et al., 2019). For example, the seminal BERT model (Devlin et al., 2019) was pretrained using the BookWiki corpus which is a combination of English Wikipedia and BooksCorpus (Zhu et al., 2015), totaling 13GB of plain text. Subsequent models have moved on to web-scale data. For example, XLNet (Yang et al., 2019), RoBERTa (Liu et al., 2019), and T5 (Raffel et al., 2020)), were trained on 158GB, 160GB and 750GB of data, respectively.
As upstream corpus size and downstream per-
formance have gone up, popular attempts at explaining these gains have focused on themes of “knowledge transfer" from the upstream corpus, attributing them to shared linguistic structure, semantics (Lina et al., 2019; Tenney et al., 2019), and facts about the world (Petroni et al., 2019). However, since the introduction of large-scale pretraining corpora occurred together with the invention of self-supervised pretraining objectives (e.g. masked language modeling (Devlin et al., 2019) and replaced token detection (Clark et al., 2019)), it remains unclear to what extent large-scale corpora are integral to these leaps in performance. For several tasks, especially summarization, recent works achieved surprising performance gains in settings where the upstream corpus is created synthetically with arbitrary symbols, but the pretraining objective is designed to capture some of the structure of the task (Krishna et al., 2021; Wu et al., 2022).
|
[
[
0.14697265625,
0.0914306640625,
0.85205078125,
0.0926513671875,
0.85205078125,
0.10986328125,
0.14697265625,
0.1092529296875
],
[
0.197265625,
0.144287109375,
0.8076171875,
0.144287109375,
0.8076171875,
0.157958984375,
0.197265625,
0.157958984375
],
[
0.38623046875,
0.1602783203125,
0.61376953125,
0.160888671875,
0.61376953125,
0.17578125,
0.38623046875,
0.174560546875
],
[
0.267822265625,
0.1776123046875,
0.73876953125,
0.1776123046875,
0.73876953125,
0.1923828125,
0.267822265625,
0.1923828125
],
[
0.2626953125,
0.2529296875,
0.343505859375,
0.2529296875,
0.343505859375,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.509765625,
0.254150390625,
0.8818359375,
0.2529296875,
0.8818359375,
0.267333984375,
0.509765625,
0.268310546875
],
[
0.509765625,
0.26953125,
0.880859375,
0.26953125,
0.880859375,
0.283203125,
0.509765625,
0.283203125
],
[
0.145263671875,
0.281982421875,
0.45751953125,
0.281982421875,
0.45751953125,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.51123046875,
0.28515625,
0.8818359375,
0.2861328125,
0.8818359375,
0.300537109375,
0.51123046875,
0.2998046875
],
[
0.0235137939453125,
0.300537109375,
0.06463623046875,
0.300537109375,
0.06463623046875,
0.439453125,
0.0235137939453125,
0.439453125
],
[
0.145263671875,
0.295654296875,
0.45751953125,
0.295654296875,
0.45751953125,
0.308837890625,
0.145263671875,
0.308837890625
],
[
0.51123046875,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51123046875,
0.31640625
],
[
0.145263671875,
0.31005859375,
0.459228515625,
0.31005859375,
0.459228515625,
0.32373046875,
0.145263671875,
0.32373046875
],
[
0.1435546875,
0.322509765625,
0.459228515625,
0.32373046875,
0.459228515625,
0.33740234375,
0.1435546875,
0.336181640625
],
[
0.51318359375,
0.320556640625,
0.880859375,
0.320556640625,
0.880859375,
0.330810546875,
0.51318359375,
0.330810546875
],
[
0.145263671875,
0.33837890625,
0.459228515625,
0.33837890625,
0.459228515625,
0.3515625,
0.145263671875,
0.3515625
],
[
0.509765625,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.347412109375,
0.509765625,
0.347412109375
],
[
0.145263671875,
0.3515625,
0.459228515625,
0.3515625,
0.459228515625,
0.365234375,
0.145263671875,
0.365234375
],
[
0.51318359375,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.1485595703125,
0.3681640625,
0.45751953125,
0.3681640625,
0.45751953125,
0.37890625,
0.1485595703125,
0.37890625
],
[
0.51123046875,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.14697265625,
0.38134765625,
0.45751953125,
0.38134765625,
0.45751953125,
0.394287109375,
0.14697265625,
0.394287109375
],
[
0.509765625,
0.383056640625,
0.8818359375,
0.3818359375,
0.8818359375,
0.396728515625,
0.509765625,
0.39794921875
],
[
0.145263671875,
0.394287109375,
0.459228515625,
0.394287109375,
0.459228515625,
0.407958984375,
0.145263671875,
0.407958984375
],
[
0.51123046875,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.51123046875,
0.413330078125
],
[
0.145263671875,
0.4091796875,
0.459228515625,
0.4091796875,
0.459228515625,
0.4228515625,
0.145263671875,
0.4228515625
],
[
0.51123046875,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.145263671875,
0.42333984375,
0.45751953125,
0.42333984375,
0.45751953125,
0.43701171875,
0.145263671875,
0.43701171875
],
[
0.0201568603515625,
0.435302734375,
0.0621337890625,
0.43408203125,
0.06463623046875,
0.51708984375,
0.0235137939453125,
0.5185546875
],
[
0.51318359375,
0.431640625,
0.88330078125,
0.431640625,
0.88330078125,
0.4453125,
0.51318359375,
0.4453125
],
[
0.1485595703125,
0.439453125,
0.45751953125,
0.439453125,
0.45751953125,
0.449462890625,
0.1485595703125,
0.449462890625
],
[
0.14697265625,
0.451904296875,
0.459228515625,
0.451904296875,
0.459228515625,
0.465576171875,
0.14697265625,
0.465576171875
],
[
0.509765625,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.14697265625,
0.466064453125,
0.459228515625,
0.466064453125,
0.459228515625,
0.479736328125,
0.14697265625,
0.479736328125
],
[
0.51123046875,
0.463134765625,
0.880859375,
0.463134765625,
0.880859375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.14697265625,
0.479736328125,
0.459228515625,
0.479736328125,
0.459228515625,
0.493408203125,
0.14697265625,
0.493408203125
],
[
0.509765625,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.509765625,
0.4921875
],
[
0.1435546875,
0.4921875,
0.45751953125,
0.493408203125,
0.45751953125,
0.50830078125,
0.1435546875,
0.50732421875
],
[
0.509765625,
0.49462890625,
0.88330078125,
0.495361328125,
0.88330078125,
0.51025390625,
0.509765625,
0.5087890625
],
[
0.1435546875,
0.50732421875,
0.45751953125,
0.50830078125,
0.45751953125,
0.5224609375,
0.1435546875,
0.521484375
],
[
0.51123046875,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51123046875,
0.525390625
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.145263671875,
0.5224609375,
0.459228515625,
0.5224609375,
0.459228515625,
0.5361328125,
0.145263671875,
0.5361328125
],
[
0.51318359375,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.51318359375,
0.54150390625
],
[
0.145263671875,
0.53759765625,
0.456787109375,
0.5361328125,
0.456787109375,
0.5498046875,
0.145263671875,
0.55029296875
],
[
0.51123046875,
0.5419921875,
0.880859375,
0.5419921875,
0.880859375,
0.5556640625,
0.51123046875,
0.5556640625
],
[
0.145263671875,
0.55029296875,
0.45751953125,
0.55029296875,
0.45751953125,
0.56396484375,
0.145263671875,
0.56396484375
],
[
0.145263671875,
0.56396484375,
0.45751953125,
0.5654296875,
0.45751953125,
0.5791015625,
0.145263671875,
0.57763671875
],
[
0.51123046875,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.51123046875,
0.572265625
],
[
0.145263671875,
0.580078125,
0.459228515625,
0.580078125,
0.459228515625,
0.59326171875,
0.145263671875,
0.59326171875
],
[
0.51123046875,
0.57470703125,
0.72314453125,
0.576171875,
0.72314453125,
0.59033203125,
0.51123046875,
0.5888671875
],
[
0.14697265625,
0.595703125,
0.45751953125,
0.595703125,
0.45751953125,
0.60546875,
0.14697265625,
0.60546875
],
[
0.53076171875,
0.5986328125,
0.88330078125,
0.5986328125,
0.88330078125,
0.61328125,
0.53076171875,
0.61328125
],
[
0.145263671875,
0.60791015625,
0.459228515625,
0.60693359375,
0.459228515625,
0.62060546875,
0.145263671875,
0.62158203125
],
[
0.51123046875,
0.615234375,
0.88330078125,
0.615234375,
0.88330078125,
0.6298828125,
0.51123046875,
0.6298828125
],
[
0.145263671875,
0.62255859375,
0.459228515625,
0.62255859375,
0.459228515625,
0.6357421875,
0.145263671875,
0.6357421875
],
[
0.145263671875,
0.6357421875,
0.459228515625,
0.6357421875,
0.459228515625,
0.6494140625,
0.145263671875,
0.6494140625
],
[
0.51318359375,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.51318359375,
0.64453125
],
[
0.14697265625,
0.65087890625,
0.459228515625,
0.65087890625,
0.459228515625,
0.66455078125,
0.14697265625,
0.66455078125
],
[
0.509765625,
0.6455078125,
0.88330078125,
0.64794921875,
0.88330078125,
0.662109375,
0.509765625,
0.66015625
],
[
0.1485595703125,
0.66650390625,
0.456787109375,
0.66650390625,
0.456787109375,
0.6767578125,
0.1485595703125,
0.6767578125
],
[
0.51123046875,
0.66455078125,
0.88330078125,
0.66455078125,
0.88330078125,
0.677734375,
0.51123046875,
0.677734375
],
[
0.14697265625,
0.6787109375,
0.459228515625,
0.6787109375,
0.459228515625,
0.6923828125,
0.14697265625,
0.6923828125
],
[
0.51123046875,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.51123046875,
0.693359375
],
[
0.145263671875,
0.693359375,
0.459228515625,
0.693359375,
0.459228515625,
0.70703125,
0.145263671875,
0.70703125
],
[
0.51123046875,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.145263671875,
0.70703125,
0.459228515625,
0.70703125,
0.459228515625,
0.72021484375,
0.145263671875,
0.72021484375
],
[
0.51318359375,
0.7119140625,
0.8818359375,
0.7119140625,
0.8818359375,
0.7255859375,
0.51318359375,
0.7255859375
],
[
0.1435546875,
0.7216796875,
0.459228515625,
0.7216796875,
0.459228515625,
0.7353515625,
0.1435546875,
0.7353515625
],
[
0.51123046875,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1435546875,
0.736328125,
0.403076171875,
0.736328125,
0.403076171875,
0.75,
0.1435546875,
0.75
],
[
0.51318359375,
0.74462890625,
0.8818359375,
0.74462890625,
0.8818359375,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.1175537109375,
0.76318359375,
0.26025390625,
0.76318359375,
0.26025390625,
0.77685546875,
0.1175537109375,
0.77685546875
],
[
0.51318359375,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.51123046875,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51123046875,
0.7880859375
],
[
0.1175537109375,
0.7890625,
0.48779296875,
0.7890625,
0.48779296875,
0.80419921875,
0.1175537109375,
0.80419921875
],
[
0.51123046875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.1175537109375,
0.8056640625,
0.491943359375,
0.8056640625,
0.491943359375,
0.8193359375,
0.1175537109375,
0.8193359375
],
[
0.509765625,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1158447265625,
0.82177734375,
0.489501953125,
0.82177734375,
0.489501953125,
0.8349609375,
0.1158447265625,
0.8349609375
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1158447265625,
0.83837890625,
0.49365234375,
0.83837890625,
0.49365234375,
0.8515625,
0.1158447265625,
0.8515625
],
[
0.5087890625,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.5087890625,
0.85498046875
],
[
0.1175537109375,
0.85400390625,
0.490234375,
0.85400390625,
0.490234375,
0.86767578125,
0.1175537109375,
0.86767578125
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1175537109375,
0.869140625,
0.489501953125,
0.869140625,
0.489501953125,
0.8828125,
0.1175537109375,
0.8828125
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.139404296875,
0.892578125,
0.490234375,
0.892578125,
0.490234375,
0.90673828125,
0.139404296875,
0.90673828125
],
[
0.509765625,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.372802734375,
0.90478515625,
0.372802734375,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.85546875,
0.90478515625,
0.85546875,
0.91845703125,
0.509765625,
0.91845703125
],
[
0.49365234375,
0.9404296875,
0.50537109375,
0.9404296875,
0.50537109375,
0.95166015625,
0.49365234375,
0.95166015625
]
] |
[
"Downstream Datasets Make Surprisingly Good Pretraining Corpora",
"Kundan Krishna, Saurabh Garg, Jeffrey P. Bigham, Zachary C. Lipton",
"Carnegie Mellon University",
"[kundank,sgarg2,jbigham,zliptonj@andrew.cmu.edu",
"Abstract",
"are subsequently trained (finetuned) on the labeled",
"downstream data available for the task at hand",
"For most natural language processing tasks, the",
"Large-scale pretrained models typically provide",
"",
"dominant practice is to finetune large pretrained",
"significant performance boosts when compared to",
"transformer models (e.g., BERT) using smaller",
"downstream datasets. Despite the success of",
"models trained directly on the downstream task",
"this approach, it remains unclear to what ex-",
"(with random initializations) (Peters et al., 2018:",
"tent these gains are attributable to the massive",
"Devlin et al., 2019; Chiang and Lee, 2020; Kr-",
"backgroundcorporaemployedforpretraining",
"ishna et al., 2021). Upstream corpora tend to be",
"versus to the pretraining objectives themselves",
" significantly larger than the downstream corpora",
"This paper introduces a large-scale study of",
"and the success of this approach is often attributed",
"self-pretraining, where the same (downstream)",
"to its ability to leverage these massive upstream",
"training data is used for both pretraining and",
"['SS]",
"corpora (Liu et al., 2019; Yang et al., 2019). For",
"finetuning.Inexperimentsaddressingboth",
"ELECTRA and RoBERTa models and 10 dis-",
"example, the seminal BERT model (Devlin et al.,",
"tinct downstream classification datasets, we ob-",
"2019) was pretrained using the BookWiki corpus",
"serve that self-pretraining rivals standard pre-",
"which is a combination of English Wikipedia and",
"training on the BookWiki corpus (despite using",
"BooksCorpus (Zhu et al., 2015), totaling 13GB of",
"around 10×-500× less data), outperforming",
"plain text. Subsequent models have moved on to",
"81602:10",
"the latter on 7 and 5 datasets, respectively. Sur-",
"web-scale data. For example, XLNet (Yang et al.",
"prisingly, these task-specific pretrained models",
"2019), RoBERTa (Liu et al., 2019), and T5 (Raffel",
"often perform well on other tasks, including the",
"GLUE benchmark. Self-pretraining also pro",
"et al., 2020), were trained on 158GB, 160GB and",
"vides benefits on structured output prediction",
"750GB of data, respectively.",
"taskssuchasquestionansweringandcommon-",
"As upstream corpus size and downstream per-",
"sense inference,often providingmore than 50%",
"formance have gone up, popular attempts at ex-",
"improvements compared to standard pretrain-",
"ing. Our results hint that often performance",
"plaining these gains have focused on themes of",
"gains attributable to pretraining are driven pri-",
"\"knowledge transfer\"from the upstream corpus, at-",
"marilybythepretrainingobjectiveitself and",
"tributing them to shared linguistic structure, seman-",
"are not always attributable to the use of exter-",
"tics (Lina et al., 2019; Tenney et al., 2019), and",
"nal pretraining data in massive amounts. These",
"facts about the world (Petroni et al., 2019). How-",
"findings are especially relevant in light of con-",
"ever, since the introduction of large-scale pretrain",
"cerns about intellectual property and offensive",
"ing corpora occurred together with the invention of",
"content in web-scale pretraining data.",
"self-supervised pretraining objectives (e.g. masked",
"1 Introduction",
"language modeling (Devlin et al., 2019) and re",
"placed token detection (Clark et al., 2019), it re-",
"For training predictive models operating on natural",
"mains unclear to what extent large-scale corpora",
"language data, the current best practice is to pre-",
" are integral to these leaps in performance. For sev-",
"train models on large unlabeled upstream corpora",
"eral tasks, especially summarization, recent works",
"to optimize self-supervised objectives, for exam-",
"achieved surprising performance gains in settings",
"ple, masked language modeling (MLM); the result-",
"where the upstream corpus is created synthetically",
"ing weights are then used to initialize models that",
"with arbitrary symbols, but the pretraining objec-",
"'Pretrained models can be downloaded from https://",
"tive is designed to capture some of the structure of",
"github.com/acmi-lab/self-pretrain",
"the task (Krishna et al., 2021; Wu et al., 2022)",
"1"
] |
{
"arxiv_id": "2209.14389",
"language": "en",
"page_idx": 0
}
|
||
2209.14389-en-2
|
2209.14389-en-2.jpg
|
## 2 $\textbf{Related Work}$
Self-pretraining in Computer Vision Most relevant to our work, recent/concurrent works in computer vision explore self-pretraining (He et al., 2022; El-Nouby et al., 2021). In a contemporary work, He et al. (2022) showed that pretraining with a Masked AutoEncoder (MAE) objective (analogue of MLM objective for images) boosts the performance of ViT models on the Imagenet-1K dataset. El-Nouby et al. (2021) showed that pretraining solely on downstream datasets for object detection and segmentation tasks reaches the performance of Imagenet-pretrained models. Our work establishes that a similar phenomenon is observed for NLP tasks too across a wide range of datasets.
Pretraining on Downstream Data in NLP Task- $Adaptive\textit{ PreTraining ( TAPT ( Gururangan et al. , }$ 2020)) consists of taking off-the-shelf pretrained models like BERT and RoBERTa and engaging in further pretraining on the downstream datasets before finetuning them to the task at hand. TAPT has been shown to improve performance of offthe-shelf models in a variety of works (Logeswaran et al., 2019; Han and Eisenstein, 2019; Chakrabarty et al., 2019). Another way in which downstream data has been used is for retrieval to create a small pretraining corpus for efficient pretraining (Yao et al., 2022). By contrast, our work pretrains models only on the downstream dataset, enabling a head-to-head comparison between the performance of off-the-shelf and self-pretrained models, and (in some situations) challenging the necessity of upstream corpora altogether.
Claims about Knowledge transfer Many works claim that pretraining extracts generally useful knowledge from the upstream corpus such as linguistic patterns (Lina et al., 2019; Tenney et al., 2019; Manning et al., 2020) and facts (Petroni et al., 2019), and that this accounts for the performance gains that they enjoy on downstream tasks. Several works, e.g., in the probing literature (Tenney et al., 2019; Manning et al., 2020; Petroni et al., 2019), demonstrate that from the internal representations of a model, it is easy (e.g., via linear models) to predict certain linguistic features or real-world
facts. However, these studies do not clarify the mechanism by which these observations relate to performance gains on downstream tasks. Tenney et al. (2019) recognizes this limitation, stating “the observation of a (linguistic) pattern does not tell us how it is used". Our work suggests that to the
3
extent that such knowledge extraction plays a role in pretraining's benefits, sufficient knowledge is often present in the downstream dataset and need not be transferred from huge upstream corpora.
Challenges to the Knowledge Transfer Narrative Multiple previous works have questioned whether knowledge transfer can fully account for the efficacy of pretraining. Improvements in performance on downstream NLP tasks have resulted from pretraining on other modalities like music and code (Papadimitriou and Jurafsky, 2020), sequences of meaningless symbols (Chiang and Lee, 2020; Krishna et al., 2021; Wu et al., 2022), and language denatured via shuffling of words (Sinha et al., 2021). On the other hand, models pretrained on language have shown improved performance on tasks dealing with other modalities such as image classification (Lu et al., 2021) and reinforcement learning for games (Reid et al., 2022). By contrast, we show that without surplus upstream data of any modality, self-pretraining alone can often perform comparably or even better than standard pretraining with a large upstream corpus. In a similar vein with these papers, our work suggests that a large portion of pretraining's success may come from alternative, unexplored mechanisms which have more to do with the pretraining objective than knowledge transfer from upstream corpora.
# 3 Experimental Setup
Our experiments center around the ELECTRA model (Clark et al., 2019) and the RoBERTa-base model (Liu et al., 2019). On the broadest set of experiments, for which we can only afford to train one model, we employ ELECTRA because it performs better than RoBERTa given comparable compute budgets (Clark et al., 2019). In particular, we use the small variant of ELECTRA (14 million parameters), which performs similarly to BERTbase on GLUE (difference of $\approx2$ points) while training much faster (Clark et al., 2019). However, we replicate many of these results on the larger RoBERTa-base model revealing similar results and thus establishing the generality of our findings.
During pretraining, a text sequence is fed into the model with some tokens masked out. While MLMonly models like RoBERTa only have a generator network that predicts the content of the masked tokens, ELECTRA has an additional discriminator module that predicts if those predictions were correct. Both the generator and the discriminator
|
[
[
0.1158447265625,
0.08489990234375,
0.268798828125,
0.08489990234375,
0.268798828125,
0.09857177734375,
0.1158447265625,
0.09857177734375
],
[
0.51318359375,
0.08612060546875,
0.8818359375,
0.08612060546875,
0.8818359375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.51318359375,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.51318359375,
0.114013671875
],
[
0.1175537109375,
0.1134033203125,
0.490234375,
0.1134033203125,
0.490234375,
0.1270751953125,
0.1175537109375,
0.1270751953125
],
[
0.509765625,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1158447265625,
0.1300048828125,
0.490234375,
0.1300048828125,
0.490234375,
0.1436767578125,
0.1158447265625,
0.1436767578125
],
[
0.509765625,
0.1353759765625,
0.861328125,
0.1353759765625,
0.861328125,
0.1484375,
0.509765625,
0.1484375
],
[
0.1175537109375,
0.146728515625,
0.490234375,
0.146728515625,
0.490234375,
0.1602783203125,
0.1175537109375,
0.1602783203125
],
[
0.51123046875,
0.149658203125,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.51123046875,
0.163330078125
],
[
0.1158447265625,
0.160888671875,
0.48779296875,
0.162109375,
0.48779296875,
0.177001953125,
0.1158447265625,
0.17578125
],
[
0.509765625,
0.16748046875,
0.8818359375,
0.16748046875,
0.8818359375,
0.18115234375,
0.509765625,
0.18115234375
],
[
0.1175537109375,
0.177001953125,
0.486083984375,
0.177001953125,
0.486083984375,
0.190673828125,
0.1175537109375,
0.190673828125
],
[
0.51123046875,
0.185302734375,
0.880859375,
0.185302734375,
0.880859375,
0.1953125,
0.51123046875,
0.1953125
],
[
0.1158447265625,
0.19482421875,
0.489501953125,
0.19482421875,
0.489501953125,
0.2078857421875,
0.1158447265625,
0.2078857421875
],
[
0.51123046875,
0.1995849609375,
0.8818359375,
0.1995849609375,
0.8818359375,
0.213134765625,
0.51123046875,
0.213134765625
],
[
0.1175537109375,
0.210205078125,
0.489501953125,
0.210205078125,
0.489501953125,
0.223876953125,
0.1175537109375,
0.223876953125
],
[
0.509765625,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.1175537109375,
0.2257080078125,
0.489501953125,
0.2257080078125,
0.489501953125,
0.2392578125,
0.1175537109375,
0.2392578125
],
[
0.509765625,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.509765625,
0.2452392578125
],
[
0.1158447265625,
0.2410888671875,
0.48779296875,
0.2423095703125,
0.48779296875,
0.257080078125,
0.1158447265625,
0.255859375
],
[
0.509765625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.262451171875,
0.509765625,
0.262451171875
],
[
0.118408203125,
0.260009765625,
0.486083984375,
0.260009765625,
0.486083984375,
0.270751953125,
0.118408203125,
0.270751953125
],
[
0.509765625,
0.26416015625,
0.8818359375,
0.26318359375,
0.8818359375,
0.27783203125,
0.509765625,
0.279052734375
],
[
0.1175537109375,
0.27490234375,
0.48779296875,
0.27490234375,
0.48779296875,
0.2880859375,
0.1175537109375,
0.2880859375
],
[
0.51318359375,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.1175537109375,
0.290283203125,
0.489501953125,
0.290283203125,
0.489501953125,
0.303955078125,
0.1175537109375,
0.303955078125
],
[
0.509765625,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.31005859375,
0.509765625,
0.31005859375
],
[
0.1158447265625,
0.30712890625,
0.486083984375,
0.30712890625,
0.486083984375,
0.320556640625,
0.1158447265625,
0.320556640625
],
[
0.51318359375,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.325927734375,
0.51318359375,
0.325927734375
],
[
0.1175537109375,
0.32470703125,
0.4189453125,
0.32470703125,
0.4189453125,
0.3349609375,
0.1175537109375,
0.3349609375
],
[
0.51318359375,
0.328857421875,
0.8818359375,
0.328857421875,
0.8818359375,
0.342529296875,
0.51318359375,
0.342529296875
],
[
0.1175537109375,
0.34033203125,
0.489501953125,
0.34033203125,
0.489501953125,
0.35400390625,
0.1175537109375,
0.35400390625
],
[
0.509765625,
0.34326171875,
0.8818359375,
0.344482421875,
0.8818359375,
0.359375,
0.509765625,
0.358154296875
],
[
0.1158447265625,
0.355712890625,
0.489501953125,
0.355712890625,
0.489501953125,
0.37060546875,
0.1158447265625,
0.37060546875
],
[
0.51318359375,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.51318359375,
0.37353515625
],
[
0.1175537109375,
0.372314453125,
0.48779296875,
0.372314453125,
0.48779296875,
0.385986328125,
0.1175537109375,
0.385986328125
],
[
0.51123046875,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.51123046875,
0.3896484375
],
[
0.11419677734375,
0.38720703125,
0.48779296875,
0.388427734375,
0.48779296875,
0.402587890625,
0.11419677734375,
0.402099609375
],
[
0.5087890625,
0.391357421875,
0.880859375,
0.392578125,
0.880859375,
0.40673828125,
0.5087890625,
0.40625
],
[
0.11419677734375,
0.40380859375,
0.48779296875,
0.402587890625,
0.48779296875,
0.41748046875,
0.11419677734375,
0.418701171875
],
[
0.509765625,
0.4091796875,
0.880859375,
0.4091796875,
0.880859375,
0.4228515625,
0.509765625,
0.4228515625
],
[
0.11419677734375,
0.419189453125,
0.486083984375,
0.418701171875,
0.486083984375,
0.432861328125,
0.11419677734375,
0.43408203125
],
[
0.509765625,
0.42578125,
0.880859375,
0.42578125,
0.880859375,
0.439453125,
0.509765625,
0.439453125
],
[
0.1158447265625,
0.435791015625,
0.48779296875,
0.435791015625,
0.48779296875,
0.449462890625,
0.1158447265625,
0.449462890625
],
[
0.509765625,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.509765625,
0.454833984375
],
[
0.1175537109375,
0.453125,
0.48779296875,
0.453125,
0.48779296875,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.51123046875,
0.457763671875,
0.8818359375,
0.457763671875,
0.8818359375,
0.471435546875,
0.51123046875,
0.471435546875
],
[
0.1158447265625,
0.46728515625,
0.48779296875,
0.468505859375,
0.48779296875,
0.482666015625,
0.1158447265625,
0.482177734375
],
[
0.509765625,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.509765625,
0.48681640625
],
[
0.1158447265625,
0.485107421875,
0.486083984375,
0.485107421875,
0.486083984375,
0.498779296875,
0.1158447265625,
0.498779296875
],
[
0.509765625,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.5029296875,
0.509765625,
0.5029296875
],
[
0.1175537109375,
0.5029296875,
0.484375,
0.5029296875,
0.484375,
0.51318359375,
0.1175537109375,
0.51318359375
],
[
0.509765625,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.509765625,
0.51953125
],
[
0.1158447265625,
0.51708984375,
0.48779296875,
0.51708984375,
0.48779296875,
0.53076171875,
0.1158447265625,
0.53076171875
],
[
0.5087890625,
0.52099609375,
0.748046875,
0.521484375,
0.748046875,
0.5361328125,
0.5087890625,
0.53515625
],
[
0.1175537109375,
0.533203125,
0.48779296875,
0.533203125,
0.48779296875,
0.54638671875,
0.1175537109375,
0.54638671875
],
[
0.1158447265625,
0.548828125,
0.489501953125,
0.5498046875,
0.489501953125,
0.5634765625,
0.1158447265625,
0.5625
],
[
0.515625,
0.5556640625,
0.53662109375,
0.5556640625,
0.53662109375,
0.5634765625,
0.515625,
0.5634765625
],
[
0.53662109375,
0.55712890625,
0.54150390625,
0.55712890625,
0.54150390625,
0.5625,
0.53662109375,
0.5625
],
[
0.5400390625,
0.55419921875,
0.71533203125,
0.55419921875,
0.71533203125,
0.56787109375,
0.5400390625,
0.56787109375
],
[
0.1158447265625,
0.5654296875,
0.486083984375,
0.5654296875,
0.486083984375,
0.5791015625,
0.1158447265625,
0.5791015625
],
[
0.1175537109375,
0.5830078125,
0.48779296875,
0.5830078125,
0.48779296875,
0.59326171875,
0.1175537109375,
0.59326171875
],
[
0.51123046875,
0.58056640625,
0.8818359375,
0.58056640625,
0.8818359375,
0.595703125,
0.51123046875,
0.595703125
],
[
0.11419677734375,
0.5966796875,
0.490234375,
0.59716796875,
0.490234375,
0.6123046875,
0.11419677734375,
0.61083984375
],
[
0.509765625,
0.59716796875,
0.8818359375,
0.59716796875,
0.8818359375,
0.61083984375,
0.509765625,
0.61083984375
],
[
0.11419677734375,
0.61328125,
0.30908203125,
0.61328125,
0.30908203125,
0.626953125,
0.11419677734375,
0.626953125
],
[
0.509765625,
0.61328125,
0.88330078125,
0.61328125,
0.88330078125,
0.626953125,
0.509765625,
0.626953125
],
[
0.118408203125,
0.63134765625,
0.489501953125,
0.63134765625,
0.489501953125,
0.64453125,
0.118408203125,
0.64453125
],
[
0.509765625,
0.62890625,
0.8818359375,
0.62890625,
0.8818359375,
0.6435546875,
0.509765625,
0.6435546875
],
[
0.1175537109375,
0.6455078125,
0.48779296875,
0.6455078125,
0.48779296875,
0.66015625,
0.1175537109375,
0.66015625
],
[
0.5087890625,
0.64453125,
0.880859375,
0.64453125,
0.880859375,
0.6591796875,
0.5087890625,
0.6591796875
],
[
0.1175537109375,
0.6630859375,
0.49365234375,
0.6630859375,
0.49365234375,
0.6767578125,
0.1175537109375,
0.6767578125
],
[
0.509765625,
0.662109375,
0.88330078125,
0.662109375,
0.88330078125,
0.67578125,
0.509765625,
0.67578125
],
[
0.118408203125,
0.6796875,
0.49365234375,
0.6796875,
0.49365234375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.5087890625,
0.677734375,
0.8818359375,
0.6767578125,
0.8818359375,
0.69140625,
0.5087890625,
0.6923828125
],
[
0.1175537109375,
0.693359375,
0.49365234375,
0.69482421875,
0.49365234375,
0.708984375,
0.1175537109375,
0.7080078125
],
[
0.509765625,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.509765625,
0.7080078125
],
[
0.118408203125,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.118408203125,
0.724609375
],
[
0.5087890625,
0.71044921875,
0.8818359375,
0.708984375,
0.8818359375,
0.7236328125,
0.5087890625,
0.724609375
],
[
0.1175537109375,
0.72802734375,
0.491943359375,
0.72802734375,
0.491943359375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.7255859375,
0.8818359375,
0.724609375,
0.8818359375,
0.7392578125,
0.509765625,
0.74072265625
],
[
0.1175537109375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.509765625,
0.7421875,
0.88330078125,
0.7421875,
0.88330078125,
0.755859375,
0.509765625,
0.755859375
],
[
0.1175537109375,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.509765625,
0.75830078125,
0.8818359375,
0.75830078125,
0.8818359375,
0.771484375,
0.509765625,
0.771484375
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.77490234375,
0.8818359375,
0.77490234375,
0.8818359375,
0.7880859375,
0.51123046875,
0.7880859375
],
[
0.1158447265625,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.51123046875,
0.79052734375,
0.85791015625,
0.79052734375,
0.85791015625,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1158447265625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.52880859375,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.52880859375,
0.82177734375
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.509765625,
0.82373046875,
0.88330078125,
0.82373046875,
0.88330078125,
0.83740234375,
0.509765625,
0.83740234375
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.5087890625,
0.8388671875,
0.88330078125,
0.84033203125,
0.88330078125,
0.85498046875,
0.5087890625,
0.85400390625
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.509765625,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.8681640625,
0.509765625,
0.8681640625
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.5087890625,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.5087890625,
0.8857421875
],
[
0.1158447265625,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.5029296875,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.5029296875,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
],
[
0.484375,
0.9453125,
0.50146484375,
0.93603515625,
0.51318359375,
0.9453125,
0.496337890625,
0.955078125
]
] |
[
"2Related Work",
"extent that such knowledge extraction plays a role",
"inpretraining's benefits.sufficientknowledgeis",
"Self-pretraining in Computer Vision Most rele-",
"often present in the downstream dataset and need",
"vant to our work, recent/concurrent works in com-",
"not be transferred from huge upstream corpora",
"puter vision explore self-pretraining (He et al.,",
"Challenges to the Knowledge Transfer Narra-",
"2022; El-Nouby et al., 2021). In a contemporary",
"tiveMultiple previous works have questioned",
"work, He et al. (2022) showed that pretraining with",
"whetherknowledgetransfercanfullyaccountfor",
"a Masked AutoEncoder (MAE) objective (analogue",
"the efficacy of pretraining. Improvements in per-",
"of MLM objective for images) boosts the perfor-",
"formance on downstream NLP tasks have resulted",
"mance of ViT models on the Imagenet-1K dataset.",
"from pretraining on other modalities like music",
"El-Nouby et al. (2021) showed that pretraining",
"and code (Papadimitriou and Jurafsky, 2020), se",
"solely ondownstreamdatasetsforobjectdetection",
"quences of meaningless symbols (Chiang and Lee",
"and segmentation tasks reaches the performance of",
"2020; Krishna et al., 2021; Wu et al., 2022), and",
"Imagenet-pretrained models. Our work establishes",
"language denatured via shuffling of words (Sinha",
"that a similar phenomenon is observed for NLP",
"et al., 2021). On the other hand, models pretrained",
"taskstooacrossa widerangeof datasets.",
"on language have shown improved performance on",
"Pretraining on Downstream Data in NLP Task-",
"tasks dealing with other modalities such as image",
"Adaptive PreTraining (TAPT (Gururangan et al.,",
"classification (Lu et al., 2021) and reinforcement",
"2020) consists of taking off-the-shelf pretrained",
"learning for games (Reid et al., 2022). By contrast.",
"models like BERT and RoBERTa and engaging",
"we show that without surplus upstream data of any",
"in further pretraining on the downstream datasets",
"modality, self-pretraining alone can often perform",
"before finetuning them to the task at hand. TAPT",
"comparably or even better than standard pretraining",
"has been shown to improve performance of off-",
"with a large upstream corpus. In a similar vein with",
"the-shelf models in a variety of works (Logeswaran",
"these papers, our work suggests that a large portion",
"et al., 2019; Han and Eisenstein, 2019; Chakrabarty",
"of pretraining's success may come from alterna-",
"et al., 2019). Another way in which downstream",
"tive, unexplored mechanisms which have more to",
"data has been used is for retrieval to create a small",
" do with the pretraining objective than knowledge",
"pretraining corpus for efficient pretraining (Yao",
"transfer from upstream corpora.",
"et al., 2022). By contrast, our work pretrains mod-",
"els only on the downstream dataset, enabling a",
"3",
"",
"Experimental Setup",
"head-to-head comparison between the performance",
"ofoff-the-shelfand self-pretrained models.and(in",
"Our experiments center around the ELECTRA",
"some situations) challenging the necessity of up-",
"model (Clark et al., 2019) and the RoBERTa-base",
" stream corpora altogether",
"model (Liu et al., 2019). On the broadest set of",
"Claims about Knowledge transfer Many works",
" experiments, for which we can only afford to train",
"claim that pretraining extracts generally useful",
" one model, we employ ELECTRA because it per-",
"knowledge from the upstream corpus such as lin-",
"forms better than RoBERTa given comparable com-",
"guistic patterns (Lina et al., 2019; Tenney et al.,",
"pute budgets (Clark et al., 2019). In particular, we",
"2019; Manning et al., 2020) and facts (Petroni et al.,",
"use the small variant of ELECTRA (14 million",
"2019), and that this accounts for the performance",
"parameters), which performs similarly to BERT",
"gains that they enjoy on downstream tasks. Sev-",
"base on GLUE (difference of ~2 points) while",
"eral works, e.g., in the probing literature (Tenney",
"training much faster (Clark et al., 2019). However,",
"et al., 2019; Manning et al., 2020; Petroni et al.,",
"we replicate many of these results on the larger",
"2019), demonstrate that from the internal represen-",
"RoBERTa-base model revealing similar results and",
"tations of a model, it is easy (e.g., via linear models)",
"thus establishing the generality of our findings.",
"to predict certain linguistic features or real-world",
"During pretraining, a text sequence is fed into the",
"facts. However, these studies do not clarify the",
"model with some tokens masked out. While MLM-",
"mechanism by which these observations relate to",
" only models like RoBERTa only have a generator",
"performance gains on downstream tasks. Tenney",
"network that predictsthe content of the masked",
"et al. (2019) recognizes this limitation, stating “the",
" tokens, ELECTRA has an additional discrimina-",
"observation of a (linguistic) pattern does not tell",
"tor module that predicts if those predictions were",
"us how it is used\". Our work suggests that to the",
" correct. Both the generator and the discriminator",
""
] |
{
"arxiv_id": "2209.14389",
"language": "en",
"page_idx": 2
}
|
||
2209.14389-en-3
|
2209.14389-en-3.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Size $(\mathbf{MB})$</th>
<th>Classes</th>
<th>$\mathbf{D}_{0}$main</th>
<th>Task</th>
</tr>
<tr>
<td>AGNews s (Zhang et al., 2015)</td>
<td>27</td>
<td>4</td>
<td>News</td>
<td>topic classification</td>
</tr>
<tr>
<td>QQP(Wang et al., 2018)</td>
<td>43</td>
<td>2</td>
<td>Online forum questions</td>
<td>paraphrase detection</td>
</tr>
<tr>
<td>Jigsaw Toxicity ${\mathrm{v}}($Kaggle.com, 2018</td>
<td>59</td>
<td>6</td>
<td>Wikipedia comments</td>
<td>toxicity detection</td>
</tr>
<tr>
<td>MNLI (Williams et al.,2018)</td>
<td>65</td>
<td>3</td>
<td>Diverse</td>
<td>natural language inference</td>
</tr>
<tr>
<td>Sentiment140 (Go et al..2009)</td>
<td>114</td>
<td>5</td>
<td>Tweets</td>
<td>sentiment classification</td>
</tr>
<tr>
<td>PAWS (Zhang et al., 2019)</td>
<td>139</td>
<td>2</td>
<td>Wikipedia</td>
<td>paraphrase detection</td>
</tr>
<tr>
<td>DBPedia14 ( Zhang et al. , 2015) </td>
<td>151</td>
<td>14</td>
<td>Wikipedia</td>
<td>topic classification</td>
</tr>
<tr>
<td>Discovery $\nu($Sileo et al., 2019)</td>
<td>293</td>
<td>174</td>
<td>Web crawl</td>
<td>discourse marker predictior</td>
</tr>
<tr>
<td>Yahoo Answertopics (Zhang et al., 2015) 1</td>
<td>461</td>
<td>10</td>
<td>Online forum answers</td>
<td>topic classification</td>
</tr>
<tr>
<td>(Zhang et al., 2015) Amazon Polaritv</td>
<td>1427</td>
<td>2</td>
<td>Product reviews</td>
<td>sentiment classification</td>
</tr>
</tbody>
</table>
Table 1: The suite of downstream datasets used in this work along with their training set sizes
networks' parameters are updated simultaneously during pretraining. After pretraining, the generator is discarded and the discriminator is used as an encoder for finetuning on downstream tasks.
We experimented with 10 different downstream datasets (Table 1). We chose these datasets in our testbed to span different dataset sizes ranging from 27 megabytes to about 1.4 gigabytes of text in the training split. These datasets are for different tasks such as topic classification, sentiment classification, natural language inference etc., and are created using data sourced from diverse domains. Most of them are multi-class classification tasks except Jigsaw Toxicity which is a multi-label classification task, and Sentiment140 which is modeled as a regression task. For finetuning a pretrained model on any dataset, we passed the input through the model, took the vector representation of the CLS token in the final layer, and passed it through a classification head with one hidden layer to get the output.
multi-class classification problems).
Notably, all self-pretrained models deliver significant performance boosts on their respective datasets (Table 2), and over half of them perform even better than the off-the-shelf model. We measured a model's benefit as the increase in performance metric that it achieves over a randomly initialized model, divided by the increase in performance metric achieved by the off-the-shelf ELECTRA model against the same baseline. The average benefit of self-pretraining across all datasets is 103.70%. We do not see a clear correlation between the size of the dataset and the performance of self-pretraining. For example, the highest beneft of 131.33% is achieved for the smallest dataset (AGNews), which is merely 27MB in size, while the minimum benefit is achieved on the Discovery dataset, which is the third largest dataset measuring 293MB. For each downstream dataset, we also pretrain a model on a randomly sampled subset of Wikipedia of the same size as the dataset's training corpus, and finetune it on the downstream task. This approach (called WikiSub) provides a sizeadjusted comparision between using separate upstream data vs the downstream data for pretraining. We see that self-pretraining performs better than WikiSub in the majority of cases (Table 2).
We also evaluated the alternate pretraining technique $TAPT$ as described in Gururangan et al. (2020). In this technique, we take the off-the-shelf ELECTRA model, which has already been pretrained on the upstream BookWiki corpus, and further pretrain it on the downstream dataset for 100 epochs. Self-pretraining outperforms TAPT on 6 datasets, notably including the two datasets where it outperformed the off-the-shelf models by the greatest benefit margin- AGNews and Yahoo Answertopics. Interestingly, TAPT performs worse
### 4 Self-pretraining Performance
In our first set of experiments, we compare selfpretraining's performance with other pretraining techniques. For each dataset, we pretrain an ELECTRA model on text from its training split and then finetune it on the same training data using the associated labels. To create a pretraining corpus from a downstream dataset, we concatenate the input text from each of the examples, assembling them in random order. We evaluate the performance of each finetuned model on the corresponding dataset's test split. For QQP and MNLI we just use the validation split because test set labels are private. For all datasets, we evaluate performance by accuracy, except for Sentiment140 and Jigsaw Toxicity, for which we use Pearson correlation and microaveraged AUC scores, respectively (these are not
4
|
[
[
0.124267578125,
0.08612060546875,
0.181396484375,
0.0885009765625,
0.1796875,
0.10272216796875,
0.1234130859375,
0.1009521484375
],
[
0.393798828125,
0.08905029296875,
0.462646484375,
0.08905029296875,
0.462646484375,
0.1009521484375,
0.393798828125,
0.1009521484375
],
[
0.478515625,
0.08905029296875,
0.53076171875,
0.08905029296875,
0.53076171875,
0.1009521484375,
0.478515625,
0.1009521484375
],
[
0.5439453125,
0.08905029296875,
0.60107421875,
0.08905029296875,
0.60107421875,
0.1009521484375,
0.5439453125,
0.1009521484375
],
[
0.7080078125,
0.0885009765625,
0.744140625,
0.0885009765625,
0.744140625,
0.1009521484375,
0.7080078125,
0.1009521484375
],
[
0.1259765625,
0.1092529296875,
0.30908203125,
0.1092529296875,
0.30908203125,
0.122314453125,
0.1259765625,
0.122314453125
],
[
0.441650390625,
0.10809326171875,
0.46337890625,
0.10809326171875,
0.46337890625,
0.122314453125,
0.441650390625,
0.122314453125
],
[
0.5439453125,
0.1092529296875,
0.5859375,
0.1092529296875,
0.5859375,
0.122314453125,
0.5439453125,
0.122314453125
],
[
0.7060546875,
0.1092529296875,
0.8251953125,
0.1092529296875,
0.8251953125,
0.122314453125,
0.7060546875,
0.122314453125
],
[
0.443359375,
0.1217041015625,
0.46337890625,
0.1217041015625,
0.46337890625,
0.135986328125,
0.443359375,
0.135986328125
],
[
0.51708984375,
0.10986328125,
0.53076171875,
0.10986328125,
0.53076171875,
0.12054443359375,
0.51708984375,
0.12054443359375
],
[
0.124267578125,
0.122314453125,
0.279541015625,
0.122314453125,
0.279541015625,
0.135986328125,
0.124267578125,
0.135986328125
],
[
0.51708984375,
0.12353515625,
0.53076171875,
0.12353515625,
0.53076171875,
0.1341552734375,
0.51708984375,
0.1341552734375
],
[
0.5439453125,
0.122314453125,
0.69287109375,
0.122314453125,
0.69287109375,
0.135986328125,
0.5439453125,
0.135986328125
],
[
0.7060546875,
0.12353515625,
0.83642578125,
0.122314453125,
0.83642578125,
0.135986328125,
0.7060546875,
0.13720703125
],
[
0.1259765625,
0.13720703125,
0.349365234375,
0.13720703125,
0.349365234375,
0.15087890625,
0.1259765625,
0.15087890625
],
[
0.441650390625,
0.135986328125,
0.46337890625,
0.135986328125,
0.46337890625,
0.149658203125,
0.441650390625,
0.149658203125
],
[
0.51904296875,
0.1395263671875,
0.53076171875,
0.1395263671875,
0.53076171875,
0.1484375,
0.51904296875,
0.1484375
],
[
0.5458984375,
0.13720703125,
0.67919921875,
0.13720703125,
0.67919921875,
0.15087890625,
0.5458984375,
0.15087890625
],
[
0.7080078125,
0.1395263671875,
0.81787109375,
0.1395263671875,
0.81787109375,
0.149658203125,
0.7080078125,
0.149658203125
],
[
0.1259765625,
0.1519775390625,
0.30908203125,
0.1519775390625,
0.30908203125,
0.1650390625,
0.1259765625,
0.1650390625
],
[
0.441650390625,
0.1519775390625,
0.46337890625,
0.1519775390625,
0.46337890625,
0.1644287109375,
0.441650390625,
0.1644287109375
],
[
0.51904296875,
0.152587890625,
0.53076171875,
0.152587890625,
0.53076171875,
0.163330078125,
0.51904296875,
0.163330078125
],
[
0.5458984375,
0.152587890625,
0.59619140625,
0.152587890625,
0.59619140625,
0.1644287109375,
0.5458984375,
0.1644287109375
],
[
0.7060546875,
0.1519775390625,
0.873046875,
0.1519775390625,
0.873046875,
0.1650390625,
0.7060546875,
0.1650390625
],
[
0.435791015625,
0.1650390625,
0.46337890625,
0.1650390625,
0.46337890625,
0.1787109375,
0.435791015625,
0.1787109375
],
[
0.5439453125,
0.1644287109375,
0.591796875,
0.166259765625,
0.59033203125,
0.179931640625,
0.54248046875,
0.1776123046875
],
[
0.7060546875,
0.166259765625,
0.85400390625,
0.166259765625,
0.85400390625,
0.179931640625,
0.7060546875,
0.179931640625
],
[
0.51708984375,
0.16748046875,
0.53076171875,
0.16748046875,
0.53076171875,
0.177001953125,
0.51708984375,
0.177001953125
],
[
0.1275634765625,
0.16748046875,
0.31494140625,
0.16748046875,
0.31494140625,
0.1776123046875,
0.1275634765625,
0.1776123046875
],
[
0.1259765625,
0.179931640625,
0.29296875,
0.179931640625,
0.29296875,
0.193603515625,
0.1259765625,
0.193603515625
],
[
0.435791015625,
0.179931640625,
0.46337890625,
0.179931640625,
0.46337890625,
0.1923828125,
0.435791015625,
0.1923828125
],
[
0.5439453125,
0.18115234375,
0.61279296875,
0.18115234375,
0.61279296875,
0.1923828125,
0.5439453125,
0.1923828125
],
[
0.705078125,
0.179931640625,
0.8369140625,
0.1787109375,
0.8369140625,
0.193603515625,
0.705078125,
0.19482421875
],
[
0.1259765625,
0.19482421875,
0.322509765625,
0.19482421875,
0.322509765625,
0.2078857421875,
0.1259765625,
0.2078857421875
],
[
0.51904296875,
0.1817626953125,
0.53076171875,
0.1817626953125,
0.53076171875,
0.191162109375,
0.51904296875,
0.191162109375
],
[
0.435791015625,
0.19482421875,
0.462646484375,
0.19482421875,
0.462646484375,
0.207275390625,
0.435791015625,
0.207275390625
],
[
0.51123046875,
0.19482421875,
0.5322265625,
0.19482421875,
0.5322265625,
0.207275390625,
0.51123046875,
0.207275390625
],
[
0.5458984375,
0.1953125,
0.61279296875,
0.1953125,
0.61279296875,
0.207275390625,
0.5458984375,
0.207275390625
],
[
0.7080078125,
0.196533203125,
0.82373046875,
0.196533203125,
0.82373046875,
0.207275390625,
0.7080078125,
0.207275390625
],
[
0.124267578125,
0.2078857421875,
0.307373046875,
0.207275390625,
0.307373046875,
0.221435546875,
0.124267578125,
0.22265625
],
[
0.435791015625,
0.2078857421875,
0.46337890625,
0.2078857421875,
0.46337890625,
0.221435546875,
0.435791015625,
0.221435546875
],
[
0.50537109375,
0.208984375,
0.5322265625,
0.208984375,
0.5322265625,
0.2203369140625,
0.50537109375,
0.2203369140625
],
[
0.54248046875,
0.207275390625,
0.615234375,
0.208984375,
0.615234375,
0.223876953125,
0.54248046875,
0.221435546875
],
[
0.7060546875,
0.208984375,
0.8818359375,
0.208984375,
0.8818359375,
0.22265625,
0.7060546875,
0.22265625
],
[
0.1259765625,
0.221435546875,
0.3779296875,
0.2203369140625,
0.3779296875,
0.235107421875,
0.1259765625,
0.236328125
],
[
0.435791015625,
0.221435546875,
0.462646484375,
0.221435546875,
0.462646484375,
0.235107421875,
0.435791015625,
0.235107421875
],
[
0.509765625,
0.22265625,
0.5341796875,
0.22265625,
0.5341796875,
0.235107421875,
0.509765625,
0.235107421875
],
[
0.54248046875,
0.22265625,
0.68603515625,
0.22265625,
0.68603515625,
0.236328125,
0.54248046875,
0.236328125
],
[
0.7060546875,
0.22265625,
0.8251953125,
0.22265625,
0.8251953125,
0.236328125,
0.7060546875,
0.236328125
],
[
0.1259765625,
0.236328125,
0.355224609375,
0.236328125,
0.355224609375,
0.2493896484375,
0.1259765625,
0.2493896484375
],
[
0.42822265625,
0.236328125,
0.46337890625,
0.236328125,
0.46337890625,
0.248779296875,
0.42822265625,
0.248779296875
],
[
0.51708984375,
0.2369384765625,
0.5341796875,
0.2369384765625,
0.5341796875,
0.2476806640625,
0.51708984375,
0.2476806640625
],
[
0.5439453125,
0.2369384765625,
0.646484375,
0.2369384765625,
0.646484375,
0.248779296875,
0.5439453125,
0.248779296875
],
[
0.7080078125,
0.2381591796875,
0.85205078125,
0.2381591796875,
0.85205078125,
0.248779296875,
0.7080078125,
0.248779296875
],
[
0.1822509765625,
0.2666015625,
0.8154296875,
0.2666015625,
0.8154296875,
0.27978515625,
0.1822509765625,
0.27978515625
],
[
0.118408203125,
0.30810546875,
0.486083984375,
0.30810546875,
0.486083984375,
0.318359375,
0.118408203125,
0.318359375
],
[
0.5087890625,
0.3046875,
0.77734375,
0.305908203125,
0.77734375,
0.320556640625,
0.5087890625,
0.319580078125
],
[
0.1175537109375,
0.322509765625,
0.490234375,
0.322509765625,
0.490234375,
0.33740234375,
0.1175537109375,
0.33740234375
],
[
0.5283203125,
0.322509765625,
0.8818359375,
0.32373046875,
0.8818359375,
0.33837890625,
0.5283203125,
0.33740234375
],
[
0.1158447265625,
0.339111328125,
0.489501953125,
0.339111328125,
0.489501953125,
0.352783203125,
0.1158447265625,
0.352783203125
],
[
0.509765625,
0.34033203125,
0.8818359375,
0.34033203125,
0.8818359375,
0.35400390625,
0.509765625,
0.35400390625
],
[
0.1158447265625,
0.355224609375,
0.447509765625,
0.355224609375,
0.447509765625,
0.3681640625,
0.1158447265625,
0.3681640625
],
[
0.51123046875,
0.355712890625,
0.8818359375,
0.355712890625,
0.8818359375,
0.369384765625,
0.51123046875,
0.369384765625
],
[
0.1368408203125,
0.369384765625,
0.48779296875,
0.369384765625,
0.48779296875,
0.383056640625,
0.1368408203125,
0.383056640625
],
[
0.5087890625,
0.37060546875,
0.88330078125,
0.371826171875,
0.88330078125,
0.385986328125,
0.5087890625,
0.385498046875
],
[
0.118408203125,
0.388427734375,
0.48779296875,
0.388427734375,
0.48779296875,
0.3984375,
0.118408203125,
0.3984375
],
[
0.509765625,
0.388427734375,
0.88330078125,
0.388427734375,
0.88330078125,
0.402099609375,
0.509765625,
0.402099609375
],
[
0.1175537109375,
0.402587890625,
0.48779296875,
0.402587890625,
0.48779296875,
0.416259765625,
0.1175537109375,
0.416259765625
],
[
0.51123046875,
0.40625,
0.8818359375,
0.40625,
0.8818359375,
0.416259765625,
0.51123046875,
0.416259765625
],
[
0.1158447265625,
0.418701171875,
0.486083984375,
0.418701171875,
0.486083984375,
0.431640625,
0.1158447265625,
0.431640625
],
[
0.509765625,
0.42041015625,
0.88330078125,
0.42041015625,
0.88330078125,
0.43408203125,
0.509765625,
0.43408203125
],
[
0.1158447265625,
0.435302734375,
0.48779296875,
0.435302734375,
0.48779296875,
0.448974609375,
0.1158447265625,
0.448974609375
],
[
0.509765625,
0.435791015625,
0.88330078125,
0.435791015625,
0.88330078125,
0.449462890625,
0.509765625,
0.449462890625
],
[
0.1175537109375,
0.451904296875,
0.489501953125,
0.451904296875,
0.489501953125,
0.465576171875,
0.1175537109375,
0.465576171875
],
[
0.509765625,
0.453125,
0.88330078125,
0.453125,
0.88330078125,
0.466064453125,
0.509765625,
0.466064453125
],
[
0.118408203125,
0.4697265625,
0.489501953125,
0.4697265625,
0.489501953125,
0.479736328125,
0.118408203125,
0.479736328125
],
[
0.51123046875,
0.468505859375,
0.8818359375,
0.468505859375,
0.8818359375,
0.482177734375,
0.51123046875,
0.482177734375
],
[
0.1175537109375,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.50732421875,
0.482666015625,
0.8818359375,
0.48388671875,
0.8818359375,
0.498779296875,
0.50732421875,
0.49755859375
],
[
0.1175537109375,
0.498779296875,
0.490234375,
0.49951171875,
0.490234375,
0.51318359375,
0.1175537109375,
0.5126953125
],
[
0.5087890625,
0.50048828125,
0.8818359375,
0.50048828125,
0.8818359375,
0.51416015625,
0.5087890625,
0.51416015625
],
[
0.118408203125,
0.51708984375,
0.486083984375,
0.51708984375,
0.486083984375,
0.52783203125,
0.118408203125,
0.52783203125
],
[
0.51123046875,
0.5185546875,
0.87890625,
0.5185546875,
0.87890625,
0.529296875,
0.51123046875,
0.529296875
],
[
0.1158447265625,
0.5322265625,
0.490234375,
0.5322265625,
0.490234375,
0.5458984375,
0.1158447265625,
0.5458984375
],
[
0.51318359375,
0.53369140625,
0.880859375,
0.53369140625,
0.880859375,
0.54443359375,
0.51318359375,
0.54443359375
],
[
0.1175537109375,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5625,
0.1175537109375,
0.5625
],
[
0.51123046875,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5625,
0.51123046875,
0.5625
],
[
0.1158447265625,
0.56396484375,
0.489501953125,
0.56396484375,
0.489501953125,
0.57763671875,
0.1158447265625,
0.57763671875
],
[
0.509765625,
0.56396484375,
0.8818359375,
0.5654296875,
0.8818359375,
0.580078125,
0.509765625,
0.5791015625
],
[
0.1158447265625,
0.580078125,
0.489501953125,
0.580078125,
0.489501953125,
0.59326171875,
0.1158447265625,
0.59326171875
],
[
0.51318359375,
0.5830078125,
0.8818359375,
0.5830078125,
0.8818359375,
0.59326171875,
0.51318359375,
0.59326171875
],
[
0.1175537109375,
0.595703125,
0.489501953125,
0.595703125,
0.489501953125,
0.60986328125,
0.1175537109375,
0.60986328125
],
[
0.51123046875,
0.59716796875,
0.88330078125,
0.59716796875,
0.88330078125,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.1158447265625,
0.61328125,
0.447509765625,
0.61328125,
0.447509765625,
0.626953125,
0.1158447265625,
0.626953125
],
[
0.51123046875,
0.6142578125,
0.88330078125,
0.6142578125,
0.88330078125,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.51318359375,
0.6298828125,
0.88330078125,
0.6298828125,
0.88330078125,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.12005615234375,
0.64013671875,
0.1435546875,
0.64013671875,
0.1435546875,
0.6484375,
0.12005615234375,
0.6484375
],
[
0.1435546875,
0.63818359375,
0.403076171875,
0.63818359375,
0.403076171875,
0.65283203125,
0.1435546875,
0.65283203125
],
[
0.509765625,
0.6455078125,
0.88330078125,
0.64453125,
0.88330078125,
0.6591796875,
0.509765625,
0.66015625
],
[
0.11419677734375,
0.662109375,
0.489501953125,
0.662109375,
0.489501953125,
0.6767578125,
0.11419677734375,
0.6767578125
],
[
0.5087890625,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.6748046875,
0.5087890625,
0.6748046875
],
[
0.1158447265625,
0.6796875,
0.489501953125,
0.6796875,
0.489501953125,
0.693359375,
0.1158447265625,
0.693359375
],
[
0.51123046875,
0.677734375,
0.88330078125,
0.677734375,
0.88330078125,
0.69140625,
0.51123046875,
0.69140625
],
[
0.1175537109375,
0.6953125,
0.490234375,
0.6953125,
0.490234375,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.509765625,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.509765625,
0.7080078125
],
[
0.1158447265625,
0.71044921875,
0.489501953125,
0.71142578125,
0.489501953125,
0.7255859375,
0.1158447265625,
0.724609375
],
[
0.509765625,
0.71044921875,
0.8818359375,
0.71044921875,
0.8818359375,
0.7236328125,
0.509765625,
0.7236328125
],
[
0.1175537109375,
0.72802734375,
0.490234375,
0.72802734375,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.7255859375,
0.82861328125,
0.7255859375,
0.82861328125,
0.74072265625,
0.509765625,
0.74072265625
],
[
0.1158447265625,
0.7421875,
0.48779296875,
0.74365234375,
0.48779296875,
0.75830078125,
0.1158447265625,
0.75732421875
],
[
0.5283203125,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.5283203125,
0.755859375
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.509765625,
0.7587890625,
0.88330078125,
0.7587890625,
0.88330078125,
0.77392578125,
0.509765625,
0.77392578125
],
[
0.1175537109375,
0.775390625,
0.490234375,
0.775390625,
0.490234375,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.509765625,
0.7890625
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.79052734375,
0.88330078125,
0.79150390625,
0.88330078125,
0.8056640625,
0.51123046875,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.5087890625,
0.82373046875,
0.8818359375,
0.822265625,
0.8818359375,
0.83740234375,
0.5087890625,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.490234375,
0.84130859375,
0.490234375,
0.85498046875,
0.1158447265625,
0.85400390625
],
[
0.5087890625,
0.84033203125,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.5087890625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.490234375,
0.85693359375,
0.490234375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.88720703125,
0.1175537109375,
0.88720703125
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.49365234375,
0.8896484375,
0.49365234375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.8896484375,
0.88330078125,
0.8896484375,
0.88330078125,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.509765625,
0.91943359375
],
[
0.491943359375,
0.9404296875,
0.50732421875,
0.9404296875,
0.50732421875,
0.95166015625,
0.491943359375,
0.95166015625
]
] |
[
"Dataset",
"Size (MB)",
"Classes",
"Domain",
"Task",
"AGNews (Zhang et al., 2015)",
"27",
"News",
"topic classification",
"43",
"4",
"QQP (Wang et al., 2018)",
"2",
"Online forum questions",
"paraphrase detection",
"Jigsaw Toxicity (Kaggle.com, 2018)",
"59",
"6",
"Wikipedia comments",
"toxicity detection",
"MNLI (Williams et al., 2018)",
"65",
"3",
"Diverse",
" natural language inference",
"114",
"Tweets",
" sentiment classification",
"5",
"Sentiment140(Goet al.,2009)",
"PAWS (Zhang et al., 2019)",
"139",
"Wikipedia",
" paraphrase detection",
"DBPedia14 (Zhang et al., 2015)",
"2",
"151",
"14 ",
"Wikipedia",
"topic classification",
"Discovery (Sileo et al., 2019)",
"293",
"174",
"Webcrawl",
"discourse marker prediction",
"Yahoo Answertopics (Zhang et al., 2015)",
"461",
"10",
" Online forum answers",
"topic classification",
"Amazon Polarity (Zhang et al., 2015)",
"1427",
"2 ",
"Product reviews",
"sentiment classification",
"Table I: The suite of downstream datasets used in this work along with their training set sizes",
"networks'parameters areupdated simultaneously",
"multi-class classification problems).",
"during pretraining. After pretraining, the genera-",
"Notably, all self-pretrained models deliver sig",
"tor is discarded and the discriminator is used as an",
"nificant performance boosts on their respective",
"encoder for finetuning on downstream tasks.",
"datasets (Table 2), and over half of them perform",
"We experimented with 10 different downstream",
"even better than the off-the-shelf model. We mea-",
"datasets(Table1). Wechosethese datasetsin our",
"sured a model's benefit as the increase in perfor-",
"testbed to span different dataset sizes ranging from",
"mance metric that it achieves over a randomly ini-",
"27 megabytes to about 1.4 gigabytes of text in the",
"tialized model, divided by the increase in perfor-",
"training split. These datasets are for different tasks",
"mance metric achieved by the off-the-shelf ELEC-",
"such as topic classification, sentiment classification,",
"TRA model against the same baseline. The aver-",
"natural language inference etc.,and are created us-",
"age benefit of self-pretraining across all datasets",
"ing data sourcedfrom diverse domains.Most of",
"is 103.70%. We do not see a clear correlation be-",
"them are multi-class classification tasks except Jig-",
" tween the size of the dataset and the performance",
"saw Toxicity whichis a multi-label classification",
"of self-pretraining.Forexample,the highestbene",
"task, and Sentiment140 which is modeled as a re-",
"fitof131.33%isachievedforthesmallestdataset",
"gression task. For finetuning a pretrained model on",
"(AGNews), which is merely 27MB in size, while",
" any dataset, we passed the input through the model,",
"the minimum benefit is achieved on the Discovery",
"took the vector representation of the CLS token in",
"dataset,which isthe third largest dataset measur-",
"the final layer, and passed it through a classification",
"ing 293MB. For each downstream dataset, we also",
"head with one hidden layer to get the output.",
"pretrain a model on a randomly sampled subset of",
"Wikipedia of the same size as the dataset's train-",
"4",
"Self-pretraining Performance",
"ing corpus, and finetune it on the downstream task.",
"In our first set of experiments, we compare self-",
"This approach (called WikiSub) provides a size-",
"pretraining's performance with other pretraining",
"adjusted comparision between using separate up-",
"techniques. For each dataset, we pretrain an ELEC-",
" stream data vs the downstream data for pretraining",
"TRA model on text from its training split and then",
"We see that self-pretraining performs better than",
"finetune it on the same training data using the asso-",
"WikiSub in the majority of cases (Table 2).",
"ciated labels. To create a pretraining corpus from a",
"We also evaluated the alternate pretraining tech-",
"downstream dataset, we concatenate the input text",
"nique TAPT as described in Gururangan et al.",
"from each of the examples, assembling them in ran-",
"(2020). In this technique, we take the off-the-shelf",
"dom order. We evaluate the performance of each",
"ELECTRA model, which has already been pre-",
"finetuned model on the corresponding dataset's",
"trained on the upstream Book Wiki corpus, and fur-",
"test split. For QQP and MNLI we just use the",
"ther pretrain it on the downstream dataset for 100",
"validation split because test set labels are private.",
"epochs. Self-pretraining outperforms TAPT on 6",
"For all datasets, we evaluate performance by accu-",
"datasets, notably including the two datasets where",
"racy, except for Sentiment140 and Jigsaw Toxicity,",
"it outperformed the off-the-shelf models by the",
"for which we use Pearson correlation and micro-",
"greatest benefit margin - AGNews and Yahoo An-",
"averaged AUC scores, respectively (these are not",
"swertopics. Interestingly, TAPT performs worse",
"4"
] |
{
"arxiv_id": "2209.14389",
"language": "en",
"page_idx": 3
}
|
||
2209.14389-en-4
|
2209.14389-en-4.jpg
|
than off-the-shelf model on the same 3 datasets where self-pretraining performs worse than off-theshelf model (except Sentiment140). None of the three pretraining approaches seem to be uniformly better than any other.
Finally, we also evaluate the self-pretrained models on the GLUE benchmark and report results on the dev set $^2.$ The performance of the models on their pretraining dataset does not correlate strongly with its GLUE score. The GLUE score also does not monotonically go up with increasing dataset size, indicating that the data domain makes some difference. For example, the Amazon Polarity corpus scores just 66.14 on GLUE despite being about 1.4GB in size, while AGNews which is 27MB in size, scores 74.30. The highest GLUE score is achieved by pretraining on Yahoo Answertopics.
## 5 Cross Dataset Finetuning
In this set of experiments, we investigated if the models pretrained on a dataset are only useful for that specific task, or are they useful across the whole spectrum of tasks that we consider. We took each model pretrained on a dataset in our testbed and finetuned and evaluated it on all other datasets in the testbed. The performance benefits provided in all cases are shown as a heatmap in Figure 2.
We found that for almost all downstream datasets, pretraining on any other dataset provides significant advantage (Figure 2). In most cases, pretraining on the downstream dataset itself performs the best. Among datasets where self-pretraining performs better than off-the-shelf model (i.e. the diagonal entry is greater than 1), pretraining on datasets of larger size does not help further. How- this question, we investigate the difference in pre-
ever, for the datasets where self-pretraining's benefit is much less than 100% (i.e. MNLI and Discovery), pretraining on a larger dataset (e.g., Yahoo Answertopics) performs better than self-pretraining.
Among all the pretrained models, a few models perform consistently good or bad across different downstream datasets (Figure 2). For example, the model pretrained on Yahoo Answertopics gets the highest average score of 0.90 across all datasets, while the PAWS-pretrained model gives the lowest aggregate score of 0.64. Similarly, there are downstream datasets that are benefited consistently by either a large or a small margin by pretraining on different datasets. For example, performance on
$^{2}$Following Clark et al.(2019) we exclude the WNLI task
from the results.
5
QQP and PAWS receives huge boosts by pretraining on most datasets. In contrast, performance on sentiment140 is mostly low , even dropping below 20% for 3 pretrained models
We perform an ablation to investigate that given a fixed dataset to finetune on, is it better to pretrain on the exact same data (i.e., using the same set of inputs), or is it better to pretrain on different data with an identical distribution. To test this hypothesis, we divided the training splits of the downstream datasets randomly into two equal subsets (denoted as A and B). We pretrained one model on each subset and then finetuned them on both subsets separately. The validation and test sets used for finetuning are the same as in the original dataset.
We do not see any consistent benefits with pretraining and finetuning on the same dataset (Table 3). Instead, we found consistent patterns where models pretrained on one split (either A or B) outperformed models pretrained on the other, irrespective of the split used for finetuning. This suggests that the pretraining data has greater influence on the final performance than the finetuning data. Additionally, we observe that finetuning the superior pretrained model, using the downstream split other than the one used for pretraining, performs the best, suggesting overall exposure to more data helps.
## 6 Difference in Outputs of Self-pretrained and Off-the-shelf Models
Since self-pretrained models and off-the-shelf models perform similarly in terms of classification accuracy, a natural question to ask is: do these models make errors on the same set of inputs? To answer dictions made by models pretrained with different strategies across all multi-class classification tasks. In particular, given model $f_A$ and $f_B$, we compute error inconsistency, defined as follows:
$$\begin{aligned}\sum_{i=1}^{n}\frac{\mathbb{1}\left[f_{A}(x_{i})\neq y_{i}\wedge f_{B}(x_{i})=y_{i}\right]}{n}\\&+\frac{\mathbb{1}\left[f_{A}(x_{i})=y_{i}\wedge f_{B}(x_{i})\neq y_{i}\right]}{n}\:,\end{aligned}$$
where $\{x_i,y_i\}_{i=1}^n$ is the test set. Intuitively, error inconsistency captures the fraction of examples where exactly one model is correct. This definition has been commonly used to estimate diversity in model prediction (Gontijo-Lopes et al., 2022; Geirhos et al., 2020). Across all the multi-class
|
[
[
0.1175537109375,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51318359375,
0.08489990234375,
0.880859375,
0.08489990234375,
0.880859375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.491943359375,
0.1015625,
0.491943359375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.10394287109375,
0.87744140625,
0.10394287109375,
0.87744140625,
0.114013671875,
0.509765625,
0.114013671875
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1158447265625,
0.1341552734375,
0.486083984375,
0.133056640625,
0.486083984375,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.509765625,
0.1341552734375,
0.72900390625,
0.1341552734375,
0.72900390625,
0.1484375,
0.509765625,
0.1484375
],
[
0.1158447265625,
0.1484375,
0.273681640625,
0.1484375,
0.273681640625,
0.162109375,
0.1158447265625,
0.162109375
],
[
0.53076171875,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.163330078125,
0.53076171875,
0.163330078125
],
[
0.1351318359375,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1351318359375,
0.179931640625
],
[
0.51123046875,
0.168701171875,
0.880859375,
0.168701171875,
0.880859375,
0.1787109375,
0.51123046875,
0.1787109375
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.509765625,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.50732421875,
0.1995849609375,
0.8818359375,
0.1995849609375,
0.8818359375,
0.213134765625,
0.50732421875,
0.213134765625
],
[
0.1175537109375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.509765625,
0.215576171875,
0.8818359375,
0.215576171875,
0.8818359375,
0.2286376953125,
0.509765625,
0.2286376953125
],
[
0.1175537109375,
0.23095703125,
0.48779296875,
0.23095703125,
0.48779296875,
0.24462890625,
0.1175537109375,
0.24462890625
],
[
0.509765625,
0.232177734375,
0.880859375,
0.232177734375,
0.880859375,
0.2452392578125,
0.509765625,
0.2452392578125
],
[
0.1158447265625,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.1158447265625,
0.262451171875
],
[
0.51123046875,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.1175537109375,
0.26416015625,
0.489501953125,
0.26416015625,
0.489501953125,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.509765625,
0.26416015625,
0.88330078125,
0.26416015625,
0.88330078125,
0.27783203125,
0.509765625,
0.27783203125
],
[
0.1175537109375,
0.27978515625,
0.491943359375,
0.27978515625,
0.491943359375,
0.29345703125,
0.1175537109375,
0.29345703125
],
[
0.5087890625,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.5087890625,
0.292236328125
],
[
0.118408203125,
0.297607421875,
0.48779296875,
0.297607421875,
0.48779296875,
0.30810546875,
0.118408203125,
0.30810546875
],
[
0.5087890625,
0.295654296875,
0.8818359375,
0.29443359375,
0.8818359375,
0.308837890625,
0.5087890625,
0.31005859375
],
[
0.1175537109375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51123046875,
0.312255859375,
0.873046875,
0.312255859375,
0.873046875,
0.325927734375,
0.51123046875,
0.325927734375
],
[
0.1175537109375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.53076171875,
0.328857421875,
0.88330078125,
0.328857421875,
0.88330078125,
0.342529296875,
0.53076171875,
0.342529296875
],
[
0.11419677734375,
0.342529296875,
0.476806640625,
0.34326171875,
0.476806640625,
0.358154296875,
0.11419677734375,
0.35693359375
],
[
0.509765625,
0.344482421875,
0.8818359375,
0.344482421875,
0.8818359375,
0.358154296875,
0.509765625,
0.358154296875
],
[
0.51318359375,
0.361083984375,
0.8818359375,
0.361083984375,
0.8818359375,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.118408203125,
0.3759765625,
0.1435546875,
0.3759765625,
0.1435546875,
0.38427734375,
0.118408203125,
0.38427734375
],
[
0.145263671875,
0.37353515625,
0.366943359375,
0.37353515625,
0.366943359375,
0.388427734375,
0.145263671875,
0.388427734375
],
[
0.51318359375,
0.377197265625,
0.8818359375,
0.377197265625,
0.8818359375,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.51318359375,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.40673828125,
0.51318359375,
0.40673828125
],
[
0.1158447265625,
0.399658203125,
0.48779296875,
0.3984375,
0.48779296875,
0.413330078125,
0.1158447265625,
0.41455078125
],
[
0.51123046875,
0.407958984375,
0.8818359375,
0.4091796875,
0.8818359375,
0.42333984375,
0.51123046875,
0.4228515625
],
[
0.1158447265625,
0.416259765625,
0.486083984375,
0.416259765625,
0.486083984375,
0.429931640625,
0.1158447265625,
0.429931640625
],
[
0.5146484375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.43701171875,
0.5146484375,
0.43701171875
],
[
0.1175537109375,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.446533203125,
0.1175537109375,
0.446533203125
],
[
0.51318359375,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.1175537109375,
0.448974609375,
0.48779296875,
0.448974609375,
0.48779296875,
0.4619140625,
0.1175537109375,
0.4619140625
],
[
0.51318359375,
0.457763671875,
0.88330078125,
0.457763671875,
0.88330078125,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.1158447265625,
0.46435546875,
0.486083984375,
0.46435546875,
0.486083984375,
0.47802734375,
0.1158447265625,
0.47802734375
],
[
0.51318359375,
0.473876953125,
0.88330078125,
0.473876953125,
0.88330078125,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.1175537109375,
0.48095703125,
0.48779296875,
0.48095703125,
0.48779296875,
0.49462890625,
0.1175537109375,
0.49462890625
],
[
0.509765625,
0.490478515625,
0.88330078125,
0.490478515625,
0.88330078125,
0.50439453125,
0.509765625,
0.50439453125
],
[
0.1175537109375,
0.496337890625,
0.486083984375,
0.496337890625,
0.486083984375,
0.51025390625,
0.1175537109375,
0.51025390625
],
[
0.11419677734375,
0.51123046875,
0.466796875,
0.5126953125,
0.466796875,
0.52685546875,
0.11419677734375,
0.525390625
],
[
0.509765625,
0.50732421875,
0.8623046875,
0.505859375,
0.8623046875,
0.52099609375,
0.509765625,
0.521484375
],
[
0.1351318359375,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.5419921875,
0.1351318359375,
0.5419921875
],
[
0.5380859375,
0.53369140625,
0.880859375,
0.53515625,
0.880859375,
0.5498046875,
0.5380859375,
0.548828125
],
[
0.1158447265625,
0.54443359375,
0.486083984375,
0.54443359375,
0.486083984375,
0.55810546875,
0.1158447265625,
0.55810546875
],
[
0.51708984375,
0.54052734375,
0.52294921875,
0.54052734375,
0.52294921875,
0.5458984375,
0.51708984375,
0.5458984375
],
[
0.54248046875,
0.5517578125,
0.759765625,
0.5517578125,
0.759765625,
0.5654296875,
0.54248046875,
0.5654296875
],
[
0.1175537109375,
0.5625,
0.489501953125,
0.5625,
0.489501953125,
0.576171875,
0.1175537109375,
0.576171875
],
[
0.1175537109375,
0.57763671875,
0.490234375,
0.57763671875,
0.490234375,
0.59130859375,
0.1175537109375,
0.59130859375
],
[
0.51123046875,
0.5791015625,
0.88330078125,
0.5791015625,
0.88330078125,
0.5927734375,
0.51123046875,
0.5927734375
],
[
0.118408203125,
0.595703125,
0.486083984375,
0.595703125,
0.486083984375,
0.60546875,
0.118408203125,
0.60546875
],
[
0.5087890625,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.60791015625,
0.5087890625,
0.60791015625
],
[
0.1175537109375,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.1175537109375,
0.62353515625
],
[
0.51123046875,
0.6123046875,
0.8818359375,
0.6123046875,
0.8818359375,
0.6259765625,
0.51123046875,
0.6259765625
],
[
0.1158447265625,
0.62451171875,
0.48779296875,
0.6259765625,
0.48779296875,
0.64013671875,
0.1158447265625,
0.6396484375
],
[
0.51123046875,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.118408203125,
0.6435546875,
0.490234375,
0.6435546875,
0.490234375,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.509765625,
0.642578125,
0.8818359375,
0.6435546875,
0.8818359375,
0.65771484375,
0.509765625,
0.65673828125
],
[
0.1175537109375,
0.65771484375,
0.490234375,
0.65771484375,
0.490234375,
0.67138671875,
0.1175537109375,
0.67138671875
],
[
0.51318359375,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51318359375,
0.67333984375
],
[
0.1175537109375,
0.67333984375,
0.490234375,
0.67333984375,
0.490234375,
0.68701171875,
0.1175537109375,
0.68701171875
],
[
0.5146484375,
0.677734375,
0.8818359375,
0.677734375,
0.8818359375,
0.6884765625,
0.5146484375,
0.6884765625
],
[
0.1158447265625,
0.6904296875,
0.490234375,
0.689453125,
0.490234375,
0.70361328125,
0.1158447265625,
0.705078125
],
[
0.51318359375,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.70703125,
0.51318359375,
0.70703125
],
[
0.1158447265625,
0.70703125,
0.478515625,
0.70703125,
0.478515625,
0.72021484375,
0.1158447265625,
0.72021484375
],
[
0.51318359375,
0.708984375,
0.8037109375,
0.708984375,
0.8037109375,
0.72265625,
0.51318359375,
0.72265625
],
[
0.1368408203125,
0.72265625,
0.48779296875,
0.72265625,
0.48779296875,
0.736328125,
0.1368408203125,
0.736328125
],
[
0.1175537109375,
0.7392578125,
0.48779296875,
0.7392578125,
0.48779296875,
0.7529296875,
0.1175537109375,
0.7529296875
],
[
0.52490234375,
0.73681640625,
0.771484375,
0.7392578125,
0.771484375,
0.75732421875,
0.52490234375,
0.7548828125
],
[
0.1175537109375,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.76953125,
0.1175537109375,
0.76953125
],
[
0.6484375,
0.7587890625,
0.66748046875,
0.7587890625,
0.66748046875,
0.7724609375,
0.6484375,
0.7724609375
],
[
0.1158447265625,
0.771484375,
0.48779296875,
0.771484375,
0.48779296875,
0.78515625,
0.1158447265625,
0.78515625
],
[
0.5146484375,
0.7685546875,
0.54150390625,
0.76318359375,
0.5458984375,
0.775390625,
0.51904296875,
0.78076171875
],
[
0.615234375,
0.77685546875,
0.87158203125,
0.771484375,
0.873046875,
0.79736328125,
0.6171875,
0.802734375
],
[
0.1158447265625,
0.7861328125,
0.489501953125,
0.78759765625,
0.489501953125,
0.8017578125,
0.1158447265625,
0.80029296875
],
[
0.1175537109375,
0.802734375,
0.48779296875,
0.802734375,
0.48779296875,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.74462890625,
0.7998046875,
0.76416015625,
0.7998046875,
0.76416015625,
0.8134765625,
0.74462890625,
0.8134765625
],
[
0.1158447265625,
0.8193359375,
0.489501953125,
0.818359375,
0.489501953125,
0.8330078125,
0.1158447265625,
0.83447265625
],
[
0.509765625,
0.82373046875,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.118408203125,
0.83740234375,
0.484375,
0.83740234375,
0.484375,
0.84716796875,
0.118408203125,
0.84716796875
],
[
0.51123046875,
0.8427734375,
0.880859375,
0.8427734375,
0.880859375,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.1175537109375,
0.8515625,
0.486083984375,
0.8515625,
0.486083984375,
0.865234375,
0.1175537109375,
0.865234375
],
[
0.509765625,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.869140625,
0.509765625,
0.869140625
],
[
0.1158447265625,
0.8662109375,
0.48779296875,
0.86767578125,
0.48779296875,
0.8818359375,
0.1158447265625,
0.88134765625
],
[
0.509765625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1368408203125,
0.892578125,
0.489501953125,
0.892578125,
0.489501953125,
0.90625,
0.1368408203125,
0.90625
],
[
0.509765625,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90625,
0.21826171875,
0.90625,
0.21826171875,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51123046875,
0.90625,
0.87890625,
0.90625,
0.87890625,
0.91650390625,
0.51123046875,
0.91650390625
],
[
0.490234375,
0.9384765625,
0.5087890625,
0.9384765625,
0.5087890625,
0.953125,
0.490234375,
0.953125
]
] |
[
"than off-the-shelf model on the same 3 datasets",
"QQP and PAWS receives huge boosts by pretrain.",
"where self-pretraining performs worse than off-the-",
"ing on most datasets.In contrast.performance or",
"shelf model (except Sentiment140). None of the",
"sentiment140 is mostly low,even dropping below",
"three pretraining approaches seem to be uniformly",
"20% for 3 pretrained models.",
"better than any other.",
"We perform an ablation to investigate that given",
"Finally, we also evaluate the self-pretrained mod-",
"afixed datasettofinetune on,is itbettertopretrain",
"els on the GLUE benchmark and report results on",
" on the exact same data (i.e., using the same set of",
"the dev set 2. The performance of the models on",
"inputs), or is it better to pretrain on different data",
"their pretraining dataset does not correlate strongly",
"with an identical distribution. To test this hypothe-",
"with its GLUE score. The GLUE score also does",
" sis, we divided the training splits of the downstream",
"not monotonically go up with increasing dataset",
"datasets randomly into two equal subsets (denoted",
"size, indicating that the data domain makes some",
"as A and B). We pretrained one model on each",
"difference. For example, the Amazon Polarity cor-",
" subset and then finetuned them on both subsets",
"pusscores just 66.14onGLUEdespitebeingabout",
"separately. The validation and test sets used for",
"1.4GB in size, while AGNews which is 27MB in",
"finetuning are the same as in the original dataset.",
"size, scores 74.30. The highest GLUE score is",
"We do not see any consistent benefits with pre-",
"achieved by pretraining on Yahoo Answertopics.",
"training and finetuning on the same dataset (Ta-",
"ble 3). Instead, we found consistent patterns where",
"5",
"Cross Dataset Finetuning",
"models pretrained on one split (either A or B) out-",
"performed models pretrained on the other, irrespec-",
"In this set of experiments, we investigated if the",
"tive of the split used for finetuning. This suggests",
"models pretrained on a dataset are only useful for",
"that the pretraining datahas greater influence on",
"that specific task, or are they useful across the",
"the final performance than the finetuning data. Ad",
"whole spectrum of tasks that we consider. We took",
"ditionally, we observe that finetuning the superior",
"each model pretrained on a dataset in our testbed",
"pretrained model, using the downstream split other",
"and finetuned and evaluated it on all other datasets",
"than the one used for pretraining, performs the best,",
"in the testbed. The performance benefits provided",
"in all cases are shown as a heatmap in Figure 2.",
"suggesting overall exposure to more data helps.",
"We found that for almost all downstream",
"Difference in Outputs of Self-pretrained",
"datasets, pretraining on any other dataset provides",
"D",
"and Off-the-shelf Models",
"significant advantage (Figure 2). In most cases, pre-",
"training on the downstream dataset itself performs",
"Since self-pretrained models and off-the-shelf mod-",
"thebest.Among datasets where self-pretraining",
"els perform similarly in terms of classification ac-",
"performs better than off-the-shelf model (i.e. the",
"curacy, a natural question to ask is: do these models",
"diagonal entry is greater than 1), pretraining on",
"make errors on the same set of inputs? To answer",
"datasets of larger size does not helpfurther.How-",
"this question, we investigate the difference in pre-",
"ever, for the datasets where self-pretraining's bene-",
"dictions made by models pretrained with different",
"fit is much less than 100% (i.e. MNLI and Discov-",
"strategiesacross allmulti-class classificationtasks.",
"ery), pretraining on a larger dataset (e.g., Yahoo An-",
"In particular, given model fA and fB, we compute",
"swertopics) performs better than self-pretraining.",
"error inconsistency, defined as follows:",
"Among all the pretrained models, a few models",
"perform consistently good or bad across different",
"nI[fA(ci) ≠ yi^ fB(ci) =yi]",
"downstream datasets (Figure 2). For example, the",
"n",
"model pretrained on Yahoo Answertopics gets the",
"=1",
"+ 1 fA(e) = yi ^ fB(a) + yl)",
"highest average score of 0.90 across all datasets,",
"while the PAWS-pretrained model gives the lowest",
"n",
"aggregate score of 0.64. Similarly, there are down-",
"where [ci, yiln-, is the test set. Intuitively, er-",
"streamdatasetsthatarebenefited consistentlyby",
"rorinconsistencycapturesthefractionofexamples",
"either a large or a small margin by pretraining on",
"where exactly one model is correct.This defini",
"different datasets. For example, performance on",
"tion has been commonly used to estimate diversity",
"2Following Clark et al. (2019) we exclude the WNLI task",
"in model prediction (Gontijo-Lopes et al., 2022;",
"from the results.",
"Geirhos et al..2020).Across all the multi-class",
"5"
] |
{
"arxiv_id": "2209.14389",
"language": "en",
"page_idx": 4
}
|
||
2302.05330-en-0
|
2302.05330-en-0.jpg
|
### Action Dynamics Task Graphs for Learning Plannable Representations of Procedural Tasks
Weichao Mao,$^{1,2*}\textbf{Ruta Desai,}^2\textbf{Michael Louis Iuzzolino,}^2\textbf{Nitin Kamra}^2$
$^{1}$ University of Illinois Urbana-Champaign
$^{2}$ Reality Labs Research, Meta
weichao2@ illinois.edu, $\{$rutadesai, mliuzzolino, nitinkamra$\}@$meta.com
## Abstract
Given video demonstrations and paired narrations of an athome procedural task such as changing a tire, we present an approach to extract the underlying task structure-relevant actions and their temporal dependencies - via actioncentric task graphs. Learnt structured representations from our method, Action Dynamics Task Graphs (ADTG), can then be used for understanding such tasks in unseen videos of humans performing them. Furthermore, ADTG can enable providing user-centric guidance to humans in these tasks, either for performing them better or for learning new tasks. Specifically, we show how ADTG can be used for: (1) tracking an ongoing task, (2) recommending next actions, and (3) planning a sequence of actions to accomplish a procedural task. We compare against state-of-the-art Neural Task Graph method and demonstrate substantial gains on 18 procedural tasks from the CrossTask dataset, including 30.1% improvement in task tracking accuracy and 20.3% accuracy gain in next action prediction.
## 1 Introduction
With the advent of augmented reality and advanced visionpowered AI systems, we envision a future of next generation AI assistants that will be able to deeply understand the athome tasks that users are doing from visual data and assist them to accomplish these tasks. These AI assistants with reasoning capabilities would be able to track the user's actions in an ongoing complex task, detect mistakes, and provide actionable guidance to the users such as next steps to take. Such user-centric guidance can either help the user better perform a task or help them learn a new task more efficiently.
To make progress toward such assistants, we focus on at-home procedural tasks, where humans routinely require guidance. Examples of such tasks include assembling furniture, making lemonade, preparing fish curry, changing a tire, and more. Procedural tasks typically involve executing specific durative actions in certain temporal order. We refer to the actions needed to accomplish a given procedural task and their temporal dependencies as task structure. Our goal is to learn representations that capture such underlying task
$^*$Work done while interning at Reality Labs Research, Meta.
Copyright $\odot2023$, Association for the Ädvancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
structure for downstream guidance generation from visual demonstrations such as videos and annotated action labels.
Learning such task structure from videos is challenging for multiple reasons. First, procedural tasks require a representation to track the state of the task and identify durative actions grounded in visual observations. Both these challenges require dealing with immense variations in visual observations even for demonstrations from a single task. Also, procedural tasks often have multiple acceptable action sequences; i.e., the ordering of some actions may be interchanged without affecting the final outcome while certain actions are temporally dependent on others. Such interchangeability and temporal dependence between actions must also be learnt directly from data.
To mitigate these challenges, we propose a structured representation - Action Dynamics Task Graph (ADTG), a graph data-structure centered around actions that inherently captures the temporal dependence in procedural tasks. ADTG focuses solely on actions and avoids representing states in the graph, thereby making the size of the graph much smaller than typical task graph representations. It uses robust visual representations of actions learnt by treating actions as“transformations between states”. We also present an approach to learn:(i) task tracking and (ii) next action prediction models based on ADTG using video demonstrations and paired action annotations of a procedural task.
Our approach allows us to observe users while they perform procedural tasks and generate actionable plans for them from direct visual observations. Specifically, ADTG can enable: (1) tracking of an ongoing task using the learnt graph structure, (2) recommendation of a next step to the user, and (3) planning a sequence of potential next steps to complete the task. We compare our method against the Neural Task Graph (NTG) method (Huang et al. 2019) and demonstrate substantial performance gains on the CrossTask dataset (Zhukov et al. 2019). Specifically, our method achieves 30.1% improvement in task tracking accuracy and $20.3\%$ improvement in next action prediction accuracy. We also present an analysis of its plan generation capability, which is not possible with NTG, and show further ablation studies to understand its strengths and weaknesses.
|
[
[
0.1298828125,
0.121826171875,
0.87109375,
0.12249755859375,
0.87109375,
0.14013671875,
0.1298828125,
0.1395263671875
],
[
0.41259765625,
0.14453125,
0.58837890625,
0.14453125,
0.58837890625,
0.1590576171875,
0.41259765625,
0.1590576171875
],
[
0.20751953125,
0.175537109375,
0.7958984375,
0.175537109375,
0.7958984375,
0.1932373046875,
0.20751953125,
0.1932373046875
],
[
0.35693359375,
0.1976318359375,
0.640625,
0.199462890625,
0.640625,
0.2139892578125,
0.35693359375,
0.212158203125
],
[
0.39794921875,
0.2108154296875,
0.60107421875,
0.212158203125,
0.60107421875,
0.2266845703125,
0.39794921875,
0.2259521484375
],
[
0.259033203125,
0.2279052734375,
0.73876953125,
0.2279052734375,
0.73876953125,
0.2412109375,
0.259033203125,
0.2412109375
],
[
0.0236968994140625,
0.267578125,
0.05963134765625,
0.267578125,
0.05963134765625,
0.330810546875,
0.0236968994140625,
0.330810546875
],
[
0.25,
0.2734375,
0.316162109375,
0.2734375,
0.316162109375,
0.28857421875,
0.25,
0.28857421875
],
[
0.521484375,
0.274658203125,
0.91259765625,
0.274658203125,
0.91259765625,
0.28857421875,
0.521484375,
0.28857421875
],
[
0.5185546875,
0.28857421875,
0.90625,
0.287353515625,
0.90625,
0.3017578125,
0.5185546875,
0.302978515625
],
[
0.101318359375,
0.298583984375,
0.463134765625,
0.2998046875,
0.463134765625,
0.313232421875,
0.101318359375,
0.3125
],
[
0.10296630859375,
0.314453125,
0.461669921875,
0.314453125,
0.461669921875,
0.324462890625,
0.10296630859375,
0.324462890625
],
[
0.53515625,
0.318115234375,
0.91259765625,
0.318115234375,
0.91259765625,
0.33203125,
0.53515625,
0.33203125
],
[
0.101318359375,
0.32568359375,
0.463134765625,
0.324462890625,
0.463134765625,
0.33837890625,
0.101318359375,
0.339111328125
],
[
0.0236968994140625,
0.3193359375,
0.05963134765625,
0.3193359375,
0.05963134765625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.51953125,
0.33203125,
0.9111328125,
0.33203125,
0.9111328125,
0.34521484375,
0.51953125,
0.34521484375
],
[
0.10455322265625,
0.33837890625,
0.4599609375,
0.33837890625,
0.4599609375,
0.348388671875,
0.10455322265625,
0.348388671875
],
[
0.10455322265625,
0.3515625,
0.461669921875,
0.3515625,
0.461669921875,
0.3623046875,
0.10455322265625,
0.3623046875
],
[
0.51953125,
0.347900390625,
0.9111328125,
0.347900390625,
0.9111328125,
0.35791015625,
0.51953125,
0.35791015625
],
[
0.10455322265625,
0.363037109375,
0.4599609375,
0.363037109375,
0.4599609375,
0.373779296875,
0.10455322265625,
0.373779296875
],
[
0.51953125,
0.35986328125,
0.91259765625,
0.35986328125,
0.91259765625,
0.373779296875,
0.51953125,
0.373779296875
],
[
0.0269622802734375,
0.3662109375,
0.0555419921875,
0.3662109375,
0.0555419921875,
0.39453125,
0.0269622802734375,
0.39453125
],
[
0.10296630859375,
0.376953125,
0.461669921875,
0.376953125,
0.461669921875,
0.386962890625,
0.10296630859375,
0.386962890625
],
[
0.521484375,
0.375732421875,
0.91015625,
0.375732421875,
0.91015625,
0.386474609375,
0.521484375,
0.386474609375
],
[
0.10455322265625,
0.389404296875,
0.461669921875,
0.389404296875,
0.461669921875,
0.399658203125,
0.10455322265625,
0.399658203125
],
[
0.51953125,
0.386962890625,
0.91259765625,
0.386962890625,
0.91259765625,
0.40087890625,
0.51953125,
0.40087890625
],
[
0.0187835693359375,
0.399658203125,
0.056365966796875,
0.397705078125,
0.063720703125,
0.48486328125,
0.02532958984375,
0.4873046875
],
[
0.10455322265625,
0.401611328125,
0.461669921875,
0.401611328125,
0.461669921875,
0.412353515625,
0.10455322265625,
0.412353515625
],
[
0.521484375,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.521484375,
0.415283203125
],
[
0.10455322265625,
0.4140625,
0.4599609375,
0.4140625,
0.4599609375,
0.4248046875,
0.10455322265625,
0.4248046875
],
[
0.51953125,
0.415283203125,
0.9140625,
0.415283203125,
0.9140625,
0.4287109375,
0.51953125,
0.4287109375
],
[
0.10455322265625,
0.4267578125,
0.4599609375,
0.4267578125,
0.4599609375,
0.4375,
0.10455322265625,
0.4375
],
[
0.51953125,
0.4287109375,
0.9111328125,
0.4287109375,
0.9111328125,
0.442626953125,
0.51953125,
0.442626953125
],
[
0.10455322265625,
0.438232421875,
0.463134765625,
0.438232421875,
0.463134765625,
0.451904296875,
0.10455322265625,
0.451904296875
],
[
0.51953125,
0.442626953125,
0.91259765625,
0.442626953125,
0.91259765625,
0.455810546875,
0.51953125,
0.455810546875
],
[
0.10296630859375,
0.45068359375,
0.463134765625,
0.45068359375,
0.463134765625,
0.464111328125,
0.10296630859375,
0.464111328125
],
[
0.51953125,
0.45703125,
0.9140625,
0.45703125,
0.9140625,
0.47021484375,
0.51953125,
0.47021484375
],
[
0.10296630859375,
0.46337890625,
0.463134765625,
0.46337890625,
0.463134765625,
0.4765625,
0.10296630859375,
0.4765625
],
[
0.10296630859375,
0.47607421875,
0.463134765625,
0.47607421875,
0.463134765625,
0.4892578125,
0.10296630859375,
0.4892578125
],
[
0.51953125,
0.47021484375,
0.76953125,
0.47021484375,
0.76953125,
0.484130859375,
0.51953125,
0.484130859375
],
[
0.10296630859375,
0.488037109375,
0.461669921875,
0.488037109375,
0.461669921875,
0.501953125,
0.10296630859375,
0.501953125
],
[
0.0228729248046875,
0.49560546875,
0.060455322265625,
0.49560546875,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.101318359375,
0.50048828125,
0.464111328125,
0.50048828125,
0.464111328125,
0.51513671875,
0.101318359375,
0.51513671875
],
[
0.53515625,
0.498779296875,
0.91259765625,
0.5,
0.91259765625,
0.5146484375,
0.53515625,
0.51318359375
],
[
0.10296630859375,
0.51513671875,
0.233642578125,
0.51513671875,
0.233642578125,
0.52587890625,
0.10296630859375,
0.52587890625
],
[
0.51953125,
0.5146484375,
0.91259765625,
0.5146484375,
0.91259765625,
0.52783203125,
0.51953125,
0.52783203125
],
[
0.521484375,
0.52978515625,
0.91015625,
0.52978515625,
0.91015625,
0.54052734375,
0.521484375,
0.54052734375
],
[
0.51953125,
0.54150390625,
0.91259765625,
0.54150390625,
0.91259765625,
0.5546875,
0.51953125,
0.5546875
],
[
0.2132568359375,
0.5517578125,
0.351318359375,
0.5517578125,
0.351318359375,
0.56640625,
0.2132568359375,
0.56640625
],
[
0.521484375,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.5693359375,
0.521484375,
0.5693359375
],
[
0.08660888671875,
0.57275390625,
0.4755859375,
0.57373046875,
0.4755859375,
0.58837890625,
0.08660888671875,
0.5869140625
],
[
0.521484375,
0.5693359375,
0.91259765625,
0.5693359375,
0.91259765625,
0.58349609375,
0.521484375,
0.58349609375
],
[
0.08660888671875,
0.58837890625,
0.4794921875,
0.58837890625,
0.4794921875,
0.6015625,
0.08660888671875,
0.6015625
],
[
0.51953125,
0.583984375,
0.91259765625,
0.583984375,
0.91259765625,
0.59765625,
0.51953125,
0.59765625
],
[
0.5185546875,
0.59521484375,
0.9111328125,
0.5966796875,
0.9111328125,
0.611328125,
0.5185546875,
0.60986328125
],
[
0.0882568359375,
0.60400390625,
0.4755859375,
0.60400390625,
0.4755859375,
0.6142578125,
0.0882568359375,
0.6142578125
],
[
0.51953125,
0.60986328125,
0.9140625,
0.60986328125,
0.9140625,
0.62353515625,
0.51953125,
0.62353515625
],
[
0.0849609375,
0.6162109375,
0.4794921875,
0.6162109375,
0.4794921875,
0.6298828125,
0.0849609375,
0.6298828125
],
[
0.0849609375,
0.6298828125,
0.47802734375,
0.6298828125,
0.47802734375,
0.64306640625,
0.0849609375,
0.64306640625
],
[
0.51953125,
0.625,
0.9140625,
0.625,
0.9140625,
0.63818359375,
0.51953125,
0.63818359375
],
[
0.521484375,
0.63818359375,
0.91259765625,
0.63818359375,
0.91259765625,
0.65283203125,
0.521484375,
0.65283203125
],
[
0.0849609375,
0.64453125,
0.47802734375,
0.64453125,
0.47802734375,
0.65771484375,
0.0849609375,
0.65771484375
],
[
0.51953125,
0.65283203125,
0.8857421875,
0.65283203125,
0.8857421875,
0.66650390625,
0.51953125,
0.66650390625
],
[
0.08660888671875,
0.6591796875,
0.47802734375,
0.6591796875,
0.47802734375,
0.66943359375,
0.08660888671875,
0.66943359375
],
[
0.0849609375,
0.671875,
0.476318359375,
0.671875,
0.476318359375,
0.68505859375,
0.0849609375,
0.68505859375
],
[
0.08660888671875,
0.68505859375,
0.4794921875,
0.68505859375,
0.4794921875,
0.69873046875,
0.08660888671875,
0.69873046875
],
[
0.5361328125,
0.681640625,
0.9111328125,
0.681640625,
0.9111328125,
0.69580078125,
0.5361328125,
0.69580078125
],
[
0.0849609375,
0.69970703125,
0.4755859375,
0.69970703125,
0.4755859375,
0.71337890625,
0.0849609375,
0.71337890625
],
[
0.51953125,
0.69580078125,
0.91259765625,
0.69580078125,
0.91259765625,
0.708984375,
0.51953125,
0.708984375
],
[
0.10296630859375,
0.7138671875,
0.4794921875,
0.7138671875,
0.4794921875,
0.72802734375,
0.10296630859375,
0.72802734375
],
[
0.51953125,
0.708984375,
0.91259765625,
0.708984375,
0.91259765625,
0.72265625,
0.51953125,
0.72265625
],
[
0.0849609375,
0.72802734375,
0.4794921875,
0.72802734375,
0.4794921875,
0.7412109375,
0.0849609375,
0.7412109375
],
[
0.51953125,
0.7236328125,
0.9140625,
0.7236328125,
0.9140625,
0.7373046875,
0.51953125,
0.7373046875
],
[
0.08660888671875,
0.7421875,
0.476318359375,
0.7421875,
0.476318359375,
0.755859375,
0.08660888671875,
0.755859375
],
[
0.51953125,
0.73779296875,
0.9140625,
0.736328125,
0.9140625,
0.75,
0.51953125,
0.751953125
],
[
0.0849609375,
0.755859375,
0.4794921875,
0.755859375,
0.4794921875,
0.76953125,
0.0849609375,
0.76953125
],
[
0.521484375,
0.751953125,
0.91259765625,
0.751953125,
0.91259765625,
0.76513671875,
0.521484375,
0.76513671875
],
[
0.51953125,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.51953125,
0.77880859375
],
[
0.08660888671875,
0.771484375,
0.47802734375,
0.771484375,
0.47802734375,
0.7822265625,
0.08660888671875,
0.7822265625
],
[
0.08660888671875,
0.78271484375,
0.47802734375,
0.78271484375,
0.47802734375,
0.796875,
0.08660888671875,
0.796875
],
[
0.51953125,
0.77880859375,
0.9140625,
0.77880859375,
0.9140625,
0.79248046875,
0.51953125,
0.79248046875
],
[
0.08660888671875,
0.796875,
0.47802734375,
0.796875,
0.47802734375,
0.81103515625,
0.08660888671875,
0.81103515625
],
[
0.51953125,
0.79248046875,
0.9140625,
0.79248046875,
0.9140625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.81103515625,
0.47802734375,
0.81103515625,
0.47802734375,
0.82470703125,
0.0849609375,
0.82470703125
],
[
0.51953125,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.51953125,
0.82080078125,
0.91259765625,
0.82080078125,
0.91259765625,
0.833984375,
0.51953125,
0.833984375
],
[
0.08660888671875,
0.8271484375,
0.47802734375,
0.8271484375,
0.47802734375,
0.8369140625,
0.08660888671875,
0.8369140625
],
[
0.51953125,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.51953125,
0.84765625
],
[
0.1070556640625,
0.8515625,
0.4599609375,
0.8515625,
0.4599609375,
0.86572265625,
0.1070556640625,
0.86572265625
],
[
0.51953125,
0.8486328125,
0.9111328125,
0.8486328125,
0.9111328125,
0.8623046875,
0.51953125,
0.8623046875
],
[
0.0849609375,
0.86279296875,
0.4794921875,
0.86279296875,
0.4794921875,
0.876953125,
0.0849609375,
0.876953125
],
[
0.51953125,
0.8623046875,
0.9140625,
0.8623046875,
0.9140625,
0.87548828125,
0.51953125,
0.87548828125
],
[
0.0849609375,
0.876953125,
0.373291015625,
0.876953125,
0.373291015625,
0.89013671875,
0.0849609375,
0.89013671875
],
[
0.51953125,
0.87548828125,
0.8544921875,
0.87548828125,
0.8544921875,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"Action Dynamics Task Graphs for Learning Plannable Representations of",
"ProceduralTasks",
"Weichao Mao,1,2* Ruta Desai,? Michael Louis Iuzzolino, Nitin Kamra?",
"1 University of Illinois Urbana-Champaign",
"2 Reality Labs Research, Meta",
"weichao2@illinois.edu,(rutadesai,mliuzzolino,nitinkamra)@meta.com",
"",
"Abstract",
"structure for downstream guidance generation from visual",
"demonstrations such as videos and annotated action labels.",
"Given video demonstrations and paired narrations of an at-",
"home procedural task such as changing a tire,we present",
"Learning such task structure from videos is challenging",
"an approach to extract the underlying task structure-rele-",
"er",
"for multiple reasons. First, procedural tasks require a rep-",
"vant actions and their temporal dependencies-via action-",
"centric task graphs.Learnt structured representations from",
"resentation to track the state of the task and identify du-",
"our method,ActionDynamicsTask Graphs (ADTG),can then",
"rative actions grounded in visual observations. Both these",
"=",
"beused for understanding such tasks in unseen videos of hu-",
"challenges require dealing withimmensevariations in vi-",
"mansperformingthem.Furthermore,ADTGcan enablepro-",
"sual observations even for demonstrations from a single task.",
"[']",
"viding user-centric guidance to humans in these tasks, ei-",
"Also, procedural tasks often have multiple acceptable ac-",
"therfor performing them better or for learning new tasks",
"tion sequences; i.e., the ordering of some actions may be",
"Specifically,weshowhowADTG canbeusedfor:(1)track",
"interchanged without affecting the final outcome while cer-",
"ing an ongoing task, (2) recommending next actions, and (3)",
"tain actions are temporally dependent on others. Such in-",
"planning a sequence of actions to accomplish a procedural",
"terchangeability and temporal dependence between actions",
"task. We compare against state-of-the-art Neural Task Graph",
"method and demonstrate substantial gains on 18 procedural",
"must also be learnt directly from data.",
"tasks from the CrossTask dataset, including 30.1% improve-",
"E",
"ment in task tracking accuracy and 20.3% accuracy gain in",
"To mitigate these challenges, we propose a structured",
"nextactionprediction",
"representation - Action Dynamics Task Graph (ADTG),",
"a graph data-structure centered around actions that inher-",
"ently captures the temporal dependence in procedural tasks.",
"1 Introduction",
"ADTG focuses solely on actions and avoids representing",
"With the advent of augmented reality and advanced vision-",
"states in the graph, thereby making the size of the graph",
"powered AI systems, we envision a future of next generation",
"much smaller than typical task graph representations. It uses",
"robust visual representations of actions learnt by treating ac-",
"A assistants that will be ableto deeplyunderstandthe at-",
"tions as “transformations between states\". We also present",
"home tasks that users are doing from visual data and assist",
"them to accomplish these tasks. These AI assistants with rea-",
"an approach to learn: (i) task tracking and (i) next action",
"prediction models based on ADTG using video demonstra-",
" soning capabilities would be able to track the user's actions",
"tions and paired action annotations of a procedural task.",
"inan ongoingcomplextask,detectmistakes,and provide",
"actionable guidance to the users such as next steps to take.",
"Such user-centric guidance can either help the user better",
"Our approach allows us to observe users while they per-",
"perform a task or help them learn a new task more efficiently",
"form procedural tasks and generate actionable plans for",
"To make progress toward such assistants, we focus on",
"them from direct visual observations. Specifically, ADTG",
"at-home procedural tasks, where humans routinely require",
"can enable: (1) tracking of an ongoing task using the learnt",
"guidance. Examples of such tasks include assembling fur-",
"graph structure, (2) recommendation of a next step to the",
"niture, making lemonade, preparing fish curry, changing a",
"user, and (3) planning a sequence of potential next steps",
"to complete the task. We compare our method against the",
"tire,andmore.Procedural tasks typically involveexecuting",
"specific durative actions in certain temporal order. We refer",
"Neural Task Graph (NTG) method (Huang et al. 2019) and",
"to the actions needed to accomplish a given procedural task",
"demonstrate substantial performance gains on the CrossTask",
"and their temporal dependencies as task structure. Our goal",
"dataset (Zhukov et al. 2019). Specifically, our method",
"achieves 30.1% improvement in task tracking accuracy and",
"istolearnrepresentationsthatcapturesuchunderlyingtask",
"20.3% improvement in next action prediction accuracy. We",
"*Work done while interning at Reality Labs Research, Meta",
"also present an analysis of its plan generation capability,",
"Copyright 2023, Association for the Advancement of Artificial",
"which is not possible with NTG, and show further ablation",
"Intelligence (www.aaai.org).All rights reserved.",
"studies to understand its strengths and weaknesses."
] |
{
"arxiv_id": "2302.05330",
"language": "en",
"page_idx": 0
}
|
||
2302.05330-en-1
|
2302.05330-en-1.jpg
|
# 2 Related Work
Existing work in learning to plan from video demonstrations can be broadly categorized into implicit and explicit approaches based on whether it explicitly maintains structural representations of the tasks.
Implicit Representations Many existing approaches learn to plan directly from visual observations without explicitly characterizing the underlying structure of the task (Srinivas et al. 2018; Zhukov et al. 2019; Sun et al. 2022; Zhao et al. 2022). In particular, under the assumption of a differentiable action space, Universal Planning Networks (UPN) (Srinivas et al. $\hat{2018}$) use gradient descent to directly learn the planner and its representations in an end-to-end fashion, by optimizing a supervised imitation learning objective. Kurutach et al.(2018) combines representation learning and planning using an InfoGAN to learn a generative model of sequential observations and a low-dimensional planning model. More recently, a Transformer-based planning network named PlaTe (Sun et al. 2022) has been proposed for procedure planning in instructional videos, which simultaneously learns the planning model and the latent semantic representations. In comparison, our approach tries to explicitly learn structural representations of tasks, leading to a modular pipeline that allows us to easily achieve various downstream learning objectives such as task tracking, action recommendation, and planning.
Explicit State-Centric Representations Another line of work uses more explicit representations for planning. These are often the state transition model and the policy from a Markov Decision Process (MDP), and are learnt using a combination of model-based deep reinforcement learning and imitation learning approaches directly from user demonstrations (Fang et al. 2020; Chang et al. 2020; Bi, Luo, and Xu 2021). In case of visual observations, one often resorts to learning a partially observable MDP structure (Hafner et al. 2019). Other approaches try to represent state transitions via transition graphs (Liu et al. 2016) and learn them from annotated data (Pan et al. 2020; Xu et al. 2020). However, directly working on the high-dimensional visual observations often leads to graphs with a prohibitively large number of states, which are computationally intractable for planning.
Explicit Action-Centric Representations To avoid the intractability of state-centric representations, some recent works have focused on action-centeric representations because the action space of a task is typically much smaller than its observation/state space. The key idea is to leverage the Conjugate Task Graph (CTG) initially proposed by (Hayes and Scassellati 2016), which reverses the roles of states and actions in the task representations. More recent approaches have proposed other easier to learn variants of CTGs (Huang et al. 2019; Chang et al. 2020), but often do not support flexible multi-step planning. Our approach is centered around building a variant of the CTG that abstracts out states and learns the inter-dependence amongst actions, with a key focus on flexible multi-step plan generation at runtime.
3 Problem Formulation
In this section, we introduce procedural tasks and graphbased representations used in the recent literature to capture their structure. We also briefly describe the CrossTask dataset used in our experiments.
## 3.1 Procedural Tasks
Definition Let the assistive agent be located in a world with an underlying state $s$, which also includes information about the user being assisted. Let $g$ be a set of desired goal states. The agent can suggest durative actions from the set $A=\{a_{1},\ldots,a_{N}\}$ to the user. Each action $a_i$ has preconditions which must be met by the current state $s$,before the action can be enacted. Further, each action has effects (a.k.a. post-conditions) which change the current state $s$, thereby transitioning the world into a new state $s^\prime.$ Accomplishing a procedural task requires finding a sequence of actions from the current world state $s$ to achieve a desired goal $g.$ Many at-home tasks like: make lemonade, change a tire, etc. can be captured by this formulation.
Assumptions In our work, we assume the set of durative actions $A$ for the procedural task is known a priori. The trained agent should observe a user perform a procedural task and recommend a relevant next action towards a desired goal state for the task. At test time, we only have access to input video frames$^1$ of the user performing the task. At training time, we additionally also have action labels annotated on the input videos along with their temporal start and end boundaries. We do not have direct access to the current state $s$ at any time, neither the pre-conditions which enable actions, nor the post-conditions of actions. Hence, our agent must represent the current task state, learn to track the state, detect the feasible next actions, plan to achieve a goal state and recommend the next action in the plan to the user
Challenges This is challenging for three reasons. The first challenge this introduces is to define and learn a state representation in order to track the state of a procedural task. The second challenge is to ground the action set $A$ in visual observations. Both these challenges require dealing with immense variations in visual input even for demonstrations from a single task. Finally, there are often multiple acceptable action sequences for procedural tasks. The ordering of some actions may be interchanged without affecting the final outcome. For instance, in the task "Make lemonade", the actions“pour lemon juice” and“pour water” may be interchanged. However, not all sequences are valid since there exist temporal dependencies between certain actions, e.g., the action“squeeze lemon”must occur before“pour lemon juice" because the former's effects are pre-conditions for the latter. Hence, the third challenge is to learn the interchangeability and temporal dependence between actions from demonstrations of the task.
3.2 Graph-based Task Representations
Many approaches in existing literature (Chang et al. 2020; Bi, Luo, and Xu 2021) have traditionally used $Task\textit{Graphs}$
$^{1}$optionally, also audio
|
[
[
0.20751953125,
0.06756591796875,
0.35693359375,
0.06756591796875,
0.35693359375,
0.08209228515625,
0.20751953125,
0.08209228515625
],
[
0.60693359375,
0.0662841796875,
0.82177734375,
0.06756591796875,
0.82177734375,
0.08331298828125,
0.60693359375,
0.08209228515625
],
[
0.08660888671875,
0.08837890625,
0.476318359375,
0.08837890625,
0.476318359375,
0.1029052734375,
0.08660888671875,
0.1029052734375
],
[
0.51953125,
0.08837890625,
0.91015625,
0.08837890625,
0.91015625,
0.1029052734375,
0.51953125,
0.1029052734375
],
[
0.08660888671875,
0.10418701171875,
0.47802734375,
0.10418701171875,
0.47802734375,
0.1142578125,
0.08660888671875,
0.1142578125
],
[
0.51953125,
0.1029052734375,
0.9111328125,
0.1029052734375,
0.9111328125,
0.11614990234375,
0.51953125,
0.11614990234375
],
[
0.0849609375,
0.11614990234375,
0.47802734375,
0.11614990234375,
0.47802734375,
0.1300048828125,
0.0849609375,
0.1300048828125
],
[
0.51953125,
0.117431640625,
0.9140625,
0.117431640625,
0.9140625,
0.1307373046875,
0.51953125,
0.1307373046875
],
[
0.0849609375,
0.1300048828125,
0.306396484375,
0.1300048828125,
0.306396484375,
0.143310546875,
0.0849609375,
0.143310546875
],
[
0.51953125,
0.1307373046875,
0.73046875,
0.1307373046875,
0.73046875,
0.14453125,
0.51953125,
0.14453125
],
[
0.0849609375,
0.15087890625,
0.47802734375,
0.1514892578125,
0.47802734375,
0.1666259765625,
0.0849609375,
0.1654052734375
],
[
0.5185546875,
0.1572265625,
0.69677734375,
0.1578369140625,
0.69677734375,
0.17236328125,
0.5185546875,
0.1717529296875
],
[
0.0849609375,
0.1666259765625,
0.4794921875,
0.1666259765625,
0.4794921875,
0.179931640625,
0.0849609375,
0.179931640625
],
[
0.08660888671875,
0.179931640625,
0.476318359375,
0.179931640625,
0.476318359375,
0.1932373046875,
0.08660888671875,
0.1932373046875
],
[
0.5185546875,
0.175537109375,
0.91259765625,
0.1767578125,
0.91259765625,
0.1912841796875,
0.5185546875,
0.1900634765625
],
[
0.0849609375,
0.1925048828125,
0.476318359375,
0.1932373046875,
0.476318359375,
0.2083740234375,
0.0849609375,
0.20703125
],
[
0.5185546875,
0.1912841796875,
0.9111328125,
0.1925048828125,
0.9111328125,
0.205810546875,
0.5185546875,
0.2052001953125
],
[
0.08416748046875,
0.20703125,
0.47802734375,
0.2083740234375,
0.47802734375,
0.222900390625,
0.08416748046875,
0.2215576171875
],
[
0.51953125,
0.205810546875,
0.91259765625,
0.205810546875,
0.91259765625,
0.2197265625,
0.51953125,
0.2197265625
],
[
0.0849609375,
0.2203369140625,
0.4794921875,
0.2215576171875,
0.4794921875,
0.23486328125,
0.0849609375,
0.2342529296875
],
[
0.521484375,
0.2197265625,
0.91259765625,
0.2197265625,
0.91259765625,
0.23291015625,
0.521484375,
0.23291015625
],
[
0.08660888671875,
0.23486328125,
0.4794921875,
0.23486328125,
0.4794921875,
0.248779296875,
0.08660888671875,
0.248779296875
],
[
0.521484375,
0.2342529296875,
0.9111328125,
0.2342529296875,
0.9111328125,
0.2474365234375,
0.521484375,
0.2474365234375
],
[
0.0849609375,
0.25,
0.4794921875,
0.25,
0.4794921875,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.521484375,
0.2474365234375,
0.91015625,
0.2474365234375,
0.91015625,
0.261474609375,
0.521484375,
0.261474609375
],
[
0.08660888671875,
0.26318359375,
0.47802734375,
0.26318359375,
0.47802734375,
0.276611328125,
0.08660888671875,
0.276611328125
],
[
0.521484375,
0.261474609375,
0.9111328125,
0.261474609375,
0.9111328125,
0.274658203125,
0.521484375,
0.274658203125
],
[
0.0849609375,
0.276611328125,
0.47802734375,
0.276611328125,
0.47802734375,
0.290283203125,
0.0849609375,
0.290283203125
],
[
0.521484375,
0.276611328125,
0.91259765625,
0.276611328125,
0.91259765625,
0.287353515625,
0.521484375,
0.287353515625
],
[
0.0849609375,
0.291748046875,
0.47802734375,
0.291748046875,
0.47802734375,
0.304931640625,
0.0849609375,
0.304931640625
],
[
0.521484375,
0.2890625,
0.9111328125,
0.2890625,
0.9111328125,
0.302978515625,
0.521484375,
0.302978515625
],
[
0.0849609375,
0.304931640625,
0.47802734375,
0.304931640625,
0.47802734375,
0.318115234375,
0.0849609375,
0.318115234375
],
[
0.52294921875,
0.304931640625,
0.9111328125,
0.304931640625,
0.9111328125,
0.31494140625,
0.52294921875,
0.31494140625
],
[
0.0849609375,
0.318115234375,
0.47802734375,
0.318115234375,
0.47802734375,
0.33203125,
0.0849609375,
0.33203125
],
[
0.52294921875,
0.318115234375,
0.9111328125,
0.318115234375,
0.9111328125,
0.328857421875,
0.52294921875,
0.328857421875
],
[
0.08660888671875,
0.33203125,
0.47802734375,
0.33203125,
0.47802734375,
0.34521484375,
0.08660888671875,
0.34521484375
],
[
0.521484375,
0.330078125,
0.9140625,
0.330078125,
0.9140625,
0.3447265625,
0.521484375,
0.3447265625
],
[
0.08660888671875,
0.3466796875,
0.47802734375,
0.3466796875,
0.47802734375,
0.35986328125,
0.08660888671875,
0.35986328125
],
[
0.51953125,
0.3447265625,
0.814453125,
0.3447265625,
0.814453125,
0.35791015625,
0.51953125,
0.35791015625
],
[
0.08660888671875,
0.35986328125,
0.47802734375,
0.35986328125,
0.47802734375,
0.373779296875,
0.08660888671875,
0.373779296875
],
[
0.52294921875,
0.35986328125,
0.90869140625,
0.35986328125,
0.90869140625,
0.37060546875,
0.52294921875,
0.37060546875
],
[
0.08660888671875,
0.373779296875,
0.47802734375,
0.373779296875,
0.47802734375,
0.386962890625,
0.08660888671875,
0.386962890625
],
[
0.51953125,
0.37255859375,
0.91259765625,
0.37255859375,
0.91259765625,
0.386474609375,
0.51953125,
0.386474609375
],
[
0.0849609375,
0.38818359375,
0.47802734375,
0.38818359375,
0.47802734375,
0.401611328125,
0.0849609375,
0.401611328125
],
[
0.521484375,
0.38818359375,
0.9111328125,
0.38818359375,
0.9111328125,
0.3984375,
0.521484375,
0.3984375
],
[
0.0849609375,
0.401611328125,
0.4794921875,
0.401611328125,
0.4794921875,
0.415283203125,
0.0849609375,
0.415283203125
],
[
0.521484375,
0.40087890625,
0.91259765625,
0.40087890625,
0.91259765625,
0.4140625,
0.521484375,
0.4140625
],
[
0.0849609375,
0.415283203125,
0.47802734375,
0.415283203125,
0.47802734375,
0.4287109375,
0.0849609375,
0.4287109375
],
[
0.521484375,
0.4140625,
0.91015625,
0.4140625,
0.91015625,
0.427978515625,
0.521484375,
0.427978515625
],
[
0.0849609375,
0.429931640625,
0.47802734375,
0.429931640625,
0.47802734375,
0.443115234375,
0.0849609375,
0.443115234375
],
[
0.521484375,
0.427978515625,
0.9111328125,
0.427978515625,
0.9111328125,
0.44140625,
0.521484375,
0.44140625
],
[
0.0849609375,
0.44140625,
0.17724609375,
0.443115234375,
0.17724609375,
0.458251953125,
0.0849609375,
0.455810546875
],
[
0.521484375,
0.443115234375,
0.9111328125,
0.443115234375,
0.9111328125,
0.453857421875,
0.521484375,
0.453857421875
],
[
0.521484375,
0.455810546875,
0.91259765625,
0.455810546875,
0.91259765625,
0.4697265625,
0.521484375,
0.4697265625
],
[
0.08660888671875,
0.46533203125,
0.4794921875,
0.46533203125,
0.4794921875,
0.479248046875,
0.08660888671875,
0.479248046875
],
[
0.521484375,
0.47021484375,
0.9111328125,
0.47021484375,
0.9111328125,
0.48095703125,
0.521484375,
0.48095703125
],
[
0.08660888671875,
0.479736328125,
0.4794921875,
0.479736328125,
0.4794921875,
0.49365234375,
0.08660888671875,
0.49365234375
],
[
0.51953125,
0.48291015625,
0.9111328125,
0.48291015625,
0.9111328125,
0.496826171875,
0.51953125,
0.496826171875
],
[
0.0849609375,
0.49365234375,
0.4794921875,
0.49365234375,
0.4794921875,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.51953125,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.51025390625,
0.51953125,
0.51025390625
],
[
0.0849609375,
0.5068359375,
0.4794921875,
0.5068359375,
0.4794921875,
0.52099609375,
0.0849609375,
0.52099609375
],
[
0.5185546875,
0.51025390625,
0.91259765625,
0.51123046875,
0.91259765625,
0.5244140625,
0.5185546875,
0.5234375
],
[
0.0849609375,
0.521484375,
0.4794921875,
0.521484375,
0.4794921875,
0.53515625,
0.0849609375,
0.53515625
],
[
0.51953125,
0.5244140625,
0.9140625,
0.5244140625,
0.9140625,
0.53857421875,
0.51953125,
0.53857421875
],
[
0.08660888671875,
0.53515625,
0.47802734375,
0.53515625,
0.47802734375,
0.548828125,
0.08660888671875,
0.548828125
],
[
0.5185546875,
0.537109375,
0.87353515625,
0.537109375,
0.87353515625,
0.55029296875,
0.5185546875,
0.55029296875
],
[
0.0849609375,
0.548828125,
0.4794921875,
0.548828125,
0.4794921875,
0.5625,
0.0849609375,
0.5625
],
[
0.521484375,
0.5517578125,
0.91259765625,
0.5517578125,
0.91259765625,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.08660888671875,
0.564453125,
0.47802734375,
0.564453125,
0.47802734375,
0.57470703125,
0.08660888671875,
0.57470703125
],
[
0.51953125,
0.5673828125,
0.9111328125,
0.5673828125,
0.9111328125,
0.58056640625,
0.51953125,
0.58056640625
],
[
0.08660888671875,
0.5771484375,
0.47802734375,
0.5771484375,
0.47802734375,
0.59033203125,
0.08660888671875,
0.59033203125
],
[
0.51953125,
0.58056640625,
0.91015625,
0.58056640625,
0.91015625,
0.5947265625,
0.51953125,
0.5947265625
],
[
0.08660888671875,
0.5888671875,
0.47802734375,
0.5888671875,
0.47802734375,
0.60302734375,
0.08660888671875,
0.60302734375
],
[
0.521484375,
0.5947265625,
0.91259765625,
0.5947265625,
0.91259765625,
0.60791015625,
0.521484375,
0.60791015625
],
[
0.08660888671875,
0.60498046875,
0.476318359375,
0.60498046875,
0.476318359375,
0.61865234375,
0.08660888671875,
0.61865234375
],
[
0.51953125,
0.609375,
0.9111328125,
0.609375,
0.9111328125,
0.62255859375,
0.51953125,
0.62255859375
],
[
0.08660888671875,
0.61865234375,
0.47802734375,
0.61865234375,
0.47802734375,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.521484375,
0.6220703125,
0.91259765625,
0.6220703125,
0.91259765625,
0.63525390625,
0.521484375,
0.63525390625
],
[
0.08660888671875,
0.6318359375,
0.47802734375,
0.6318359375,
0.47802734375,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.521484375,
0.63720703125,
0.90869140625,
0.63720703125,
0.90869140625,
0.64794921875,
0.521484375,
0.64794921875
],
[
0.08660888671875,
0.646484375,
0.47802734375,
0.646484375,
0.47802734375,
0.66015625,
0.08660888671875,
0.66015625
],
[
0.51953125,
0.6494140625,
0.9140625,
0.6494140625,
0.9140625,
0.66357421875,
0.51953125,
0.66357421875
],
[
0.0849609375,
0.65771484375,
0.42236328125,
0.66015625,
0.42236328125,
0.6748046875,
0.0849609375,
0.67236328125
],
[
0.51953125,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.6767578125,
0.51953125,
0.6767578125
],
[
0.08660888671875,
0.68115234375,
0.38232421875,
0.68115234375,
0.38232421875,
0.69580078125,
0.08660888671875,
0.69580078125
],
[
0.375732421875,
0.681640625,
0.47802734375,
0.681640625,
0.47802734375,
0.693359375,
0.375732421875,
0.693359375
],
[
0.521484375,
0.6787109375,
0.9111328125,
0.6787109375,
0.9111328125,
0.689453125,
0.521484375,
0.689453125
],
[
0.51953125,
0.68994140625,
0.9140625,
0.68994140625,
0.9140625,
0.7041015625,
0.51953125,
0.7041015625
],
[
0.08660888671875,
0.6962890625,
0.47802734375,
0.6962890625,
0.47802734375,
0.70703125,
0.08660888671875,
0.70703125
],
[
0.0849609375,
0.71044921875,
0.47802734375,
0.71044921875,
0.47802734375,
0.7236328125,
0.0849609375,
0.7236328125
],
[
0.51953125,
0.705078125,
0.91259765625,
0.705078125,
0.91259765625,
0.71826171875,
0.51953125,
0.71826171875
],
[
0.0849609375,
0.7236328125,
0.4794921875,
0.7236328125,
0.4794921875,
0.7373046875,
0.0849609375,
0.7373046875
],
[
0.5185546875,
0.71826171875,
0.91259765625,
0.7197265625,
0.91259765625,
0.73291015625,
0.5185546875,
0.73193359375
],
[
0.0849609375,
0.7373046875,
0.476318359375,
0.7373046875,
0.476318359375,
0.75048828125,
0.0849609375,
0.75048828125
],
[
0.51953125,
0.73193359375,
0.91259765625,
0.73193359375,
0.91259765625,
0.74560546875,
0.51953125,
0.74560546875
],
[
0.0849609375,
0.75048828125,
0.4794921875,
0.75048828125,
0.4794921875,
0.76513671875,
0.0849609375,
0.76513671875
],
[
0.521484375,
0.7470703125,
0.91259765625,
0.7470703125,
0.91259765625,
0.76025390625,
0.521484375,
0.76025390625
],
[
0.0849609375,
0.76513671875,
0.4794921875,
0.76513671875,
0.4794921875,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.521484375,
0.76220703125,
0.91015625,
0.76220703125,
0.91015625,
0.77294921875,
0.521484375,
0.77294921875
],
[
0.08660888671875,
0.77978515625,
0.476318359375,
0.77978515625,
0.476318359375,
0.79052734375,
0.08660888671875,
0.79052734375
],
[
0.51953125,
0.7744140625,
0.9140625,
0.7744140625,
0.9140625,
0.78857421875,
0.51953125,
0.78857421875
],
[
0.0849609375,
0.79248046875,
0.47802734375,
0.79248046875,
0.47802734375,
0.80615234375,
0.0849609375,
0.80615234375
],
[
0.51953125,
0.78857421875,
0.73779296875,
0.78857421875,
0.73779296875,
0.8017578125,
0.51953125,
0.8017578125
],
[
0.08660888671875,
0.80615234375,
0.47802734375,
0.80615234375,
0.47802734375,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.08660888671875,
0.82080078125,
0.47802734375,
0.82080078125,
0.47802734375,
0.833984375,
0.08660888671875,
0.833984375
],
[
0.51953125,
0.81494140625,
0.8310546875,
0.81494140625,
0.8310546875,
0.830078125,
0.51953125,
0.830078125
],
[
0.0849609375,
0.833984375,
0.47802734375,
0.833984375,
0.47802734375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.51953125,
0.83544921875,
0.91259765625,
0.83544921875,
0.91259765625,
0.849609375,
0.51953125,
0.849609375
],
[
0.0882568359375,
0.849609375,
0.476318359375,
0.849609375,
0.476318359375,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.51953125,
0.849609375,
0.9140625,
0.849609375,
0.9140625,
0.86279296875,
0.51953125,
0.86279296875
],
[
0.0849609375,
0.8623046875,
0.48046875,
0.8623046875,
0.48046875,
0.87548828125,
0.0849609375,
0.87548828125
],
[
0.0849609375,
0.876953125,
0.143798828125,
0.876953125,
0.143798828125,
0.88818359375,
0.0849609375,
0.88818359375
],
[
0.54345703125,
0.87548828125,
0.67626953125,
0.87548828125,
0.67626953125,
0.8896484375,
0.54345703125,
0.8896484375
]
] |
[
"2 Related Work",
"3ProblemFormulation",
"Existing work in learning to plan from video demonstra-",
"In this section, we introduce procedural tasks and graph-",
"tions can be broadlycategorized into implicit and explicit",
"based representations used in the recent literature to cap-",
"approaches based on whether it explicitly maintains struc-",
"ture their structure. We also briefly describe the CrossTask",
"tural representations of the tasks.",
"dataset used in our experiments.",
"Implicit Representations Many existing approaches",
"3.1 Procedural Tasks",
"learn to plan directly from visual observations without",
"explicitly characterizing the underlying structure of the",
"Definition Let the assistive agent be located in a world",
"task (Srinivas et al. 2018; Zhukov et al. 2019; Sun et al.",
"with an underlying state s, which also includes informa-",
"2022; Zhao et al. 2022). In particular, under the assump-",
"tion about the user being assisted. Let g be a set of desired",
"tion of a differentiable action space, Universal Planning",
"goal states. The agent can suggest durative actions from the",
"Networks (UPN) (Srinivas et al. 2018) use gradient descent",
"set A = {ai,: .., an} to the user. Each action a, has pre-",
"to directly learn the planner and its representations in",
"conditions which must be met by the current state s, be-",
"an end-to-end fashion, by optimizing a supervised imita-",
"fore the action can be enacted. Further, each action has ef-",
"tion learning objective. Kurutach et al. (2018) combines",
"fects (a.k.a.post-conditions)which change the current state",
"representation learning and planning using an InfoGAN",
"s, thereby transitioning the world into a new state s. Ac-",
"to learn a generative model of sequential observations",
"complishingaproceduraltaskrequiresfindingasequence",
"and a low-dimensional planning model. More recently, a",
"of actionsfrom the current worldstate sto achieve a desired",
"Transformer-based planning network named PlaTe (Sun",
"goal g. Many at-home tasks like: make lemonade, change a",
"et al. 2022) has been proposed for procedure planning",
"tire, etc. can be captured by this formulation.",
"in instructional videos, which simultaneously learns the",
"Assumptions Inourwork,weassumethesetofdurativeac.",
"planning model and the latent semantic representations. In",
"tions A for the procedural task is known a priori. The trained",
"comparison, our approach tries to explicitly learn structural",
"agentshould observe a userperform aprocedural task and",
"representations of tasks, leading to a modular pipeline that",
"recommend a relevant next action towards a desired goal",
"allows us to easily achieve various downstream learning",
"state for the task. At test time, we only have access to in-",
"objectives such as task tracking, action recommendation,",
"put video frames' of the user performing the task. At train-",
"and planning.",
"ing time,we additionally alsohave action labels annotated",
"on the input videos along with their temporal start and end",
"Explicit State-Centric Representations Another line of",
"boundaries.We do not have direct access to the current state",
"work uses more explicit representations for planning. These",
"s at any time, neither the pre-conditions which enable ac-",
"are often the state transition model and the policy from a",
"tions, nor the post-conditions of actions. Hence, our agent",
"Markov Decision Process (MDP), and are learnt using a",
"must represent the current task state, learn to track the state,",
"combination of model-based deep reinforcement learning",
"detect the feasible next actions, plan to achieve a goal state",
"and imitation learning approaches directly from user demon-",
"and recommend the next action in the plan to the user.",
" strations (Fang et al. 2020; Chang et al. 2020; Bi, Luo, and",
"Challenges This is challenging for three reasons. The first",
"Xu2021).Incaseofvisual observations,oneoftenresortsto",
"challenge this introduces is to define and learn a state rep-",
"learning a partially observable MDP structure (Hafner et al.",
"resentation in order to track the state of a procedural task.",
"2019). Other approaches try to represent state transitions via",
"The second challenge is to ground the action set A in visual",
"transition graphs (Liu et al. 2016) and learn them from anno-",
"observations. Both these challenges require dealing with im-",
"tated data (Pan et al. 2020; Xu et al. 2020). However, directly",
"mense variations in visual input even for demonstrations",
"working on the high-dimensional visual observations often",
"from a single task.Finally,there are often multiple accept",
"leads to graphs with a prohibitively large number of states,",
"able action sequences for procedural tasks. The ordering of",
"which are computationally intractable for planning.",
"some actions may be interchanged without affecting the fi-",
"Explicit Action-Centric Representations",
"Toavoidthe",
"nal outcome.For instance,in the task\"Makelemonade",
"the actions “pour lemon juice\" and “pour water\" may be",
"intractability of state-centric representations, some recent",
"works have focused on action-centeric representations be-",
"interchanged. However, not all sequences are valid since",
"cause the action space of a task is typically much smaller",
"there exist temporal dependencies between certain actions,",
"than its observation/state space. The key idea is to lever-",
"e.g. the action“\"squeeze lemon\"must occur before “pour",
"age the Conjugate Task Graph (CTG) initially proposed",
"lemon juice\" because the former's effects are pre-conditions",
"by (Hayes and Scassellati 2016), which reverses the roles",
"for the latter. Hence,the third challenge is to learn the in-",
"of states and actions in the task representations.More re-",
"terchangeability and temporal dependence between actions",
"cent approaches have proposed other easier to learn variants",
"from demonstrations of the task.",
"of CTGs (Huang et al. 2019; Chang et al. 2020), but often",
"do not support flexible multi-step planning. Our approach is",
"3.2Graph-based Task Representations",
"centered around building a variant of the CTG that abstracts",
"Many approaches in existing literature (Chang et al. 2020;",
"outstates andlearnstheinter-dependenceamongstactions",
"Bi, Luo, and Xu 2021) have traditionally used Task Graphs",
"with a key focus on flexible multi-step plan generation at",
"runtime.",
"'optionally,also audio"
] |
{
"arxiv_id": "2302.05330",
"language": "en",
"page_idx": 1
}
|
||
2302.05330-en-4
|
2302.05330-en-4.jpg
|
<FigureHere>
Figure 3: Action embedding network
form a procedural task. We consider multiple scenarios of guidance, including task tracking, next action recommendation, and plan generation. A working pipeline of using ADTG is illustrated in Figure 4.
come from either the very beginning of the task (i.e., complete plan generation) or from a random midpoint in the execution of the task (i.e., planning after a prefix observation sequence). In our approach, plan generation is done by applying one step of task tracking, followed by multiple steps of next action recommendation in an autoregressive way. Specifically, given an initial visual observation and initial action history, we pass them through the task tracking module to recognize the current action, and localize to an action node in the ADTG. Using the embedding of the localized action and the action history information as input, we obtain the next action to perform from the next action recommendation module. Finally, we update the action history information by passing the newest action through the history RNN, and recursively invoke the next action recommendation module to generate a sequence of actions to perform. Repeating this process will lead to a plan to accomplish the task. In this sense, our plan generation is built entirely upon existing modules.
In the recursive process of plan generation, instead of following the greedy action choice outputted by the next action recommendation module at each step, we use beam search to select the output sequence with higher overall likelihood. Specifically, for a beam search width of $k$, we maintain $k$ candidate action trajectories, perform rollouts based on these trajectories, and only keep the rollout trajectories with the top $\tilde{k}$ highest likelihoods at each step. In this way, our method strikes a good balance between computation complexity and the optimality of the selected action sequence.
Task Tracking and Next Action Recommendation In task tracking, our model tries to track the human's progress when performing a certain task, by localizing the current visual observation in the corresponding ADTG graph. For any time step $t\in\{1,\ldots,T\}$ of a given video, the inputs to the task tracking module include the current visual observation $x_t$, the action history information $h_t-1$ up to time step $t-1$, and the embedding of a candidate action $e_a$ being evaluated. The task tracking module then outputs a confidence score indicating whether the candidate action $a$ captures the given visual observation $x_t.$ We then enumerate all the candidate actions and localize to the action node $a_t$ with the highest confidence score in the ADTG. In particular, to get a succinct representation of the action history, we pass the sequence of history action embeddings $(e_{a_1},\ldots,e_{a_{t-1}})$ up to time $t-1$ through a recurrent neural network (RNN), and use $\begin{array}{l}{\text{the RNN hidden state as the action history information }h_{t-1}.}\\{\text{We train the task tracking module and the history RNN over}}\end{array}$ all the time steps of the training videos, and use the cross entropy loss to update the network parameters for both of them using back-propagation.
In next action recommendation, our model recommends to the user an action to perform in the next time step. For any time step $t$,the next action recommendation module takes as input the currently localized action node embedding $e_{a_t}$,the action history information $h_t$, and the embedding of a candidate next action $e_a$ being evaluated. It maps these inputs to a confidence score as the output, indicating where it believes that the evaluated candidate action is a suitable one for the next time step. We then enumerate all the candidate next actions from the outgoing edges of the localized action node in the ADTG, and select the action with the highest confidence score as the recommendation $a_t+1$ to the user. The action history information comes from the hidden state of the same history RNN as in the task tracking module. We again use the cross entropy loss to train the next action recommendation module over any consecutive action pairs $(a_t,a_{t+1})$ in the training videos.
Plan Generation Given a visual observation, plan generation is used to generate a sequence of actions to accomplish
## 4.3 Implementation Details
In our implementation, the condition generator, transformation predictor, task tracking module, and the next action recommendation module are all instantiated as two-layer feedforward neural networks. The pre- and post-conditions are 128-dimensional semantic vectors, and the action embedding vectors are 96-dimensional. The hidden size of the history RNN is 128. The margin threshold for contrastive loss in Equation (2) is set to be $M=0.5.$ In the plan generation module, the beam search width is $k=5.$ We use the ADAM optimizer (Kingma and Ba 2014) with a learning rate of $1e-5$ for training the action embeddings, and learning rate $5e-5$ for training the task tracking and next action recommendation modules. The action embeddings and task tracking module are trained for 50 epochs, while the next action recommendation modules takes 100 training epochs.
the task from this point onward. The visual observation can
<FigureHere>
Figure 4: Using ADTG for guidance
|
[
[
0.191162109375,
0.07135009765625,
0.2303466796875,
0.07135009765625,
0.2303466796875,
0.07891845703125,
0.191162109375,
0.07891845703125
],
[
0.2607421875,
0.08587646484375,
0.306396484375,
0.08587646484375,
0.306396484375,
0.09661865234375,
0.2607421875,
0.09661865234375
],
[
0.41015625,
0.08209228515625,
0.42236328125,
0.08209228515625,
0.42236328125,
0.085205078125,
0.41015625,
0.085205078125
],
[
0.461669921875,
0.08587646484375,
0.53125,
0.08587646484375,
0.53125,
0.09783935546875,
0.461669921875,
0.09783935546875
],
[
0.2607421875,
0.09661865234375,
0.306396484375,
0.09661865234375,
0.306396484375,
0.106689453125,
0.2607421875,
0.106689453125
],
[
0.473876953125,
0.09661865234375,
0.5185546875,
0.09661865234375,
0.5185546875,
0.10797119140625,
0.473876953125,
0.10797119140625
],
[
0.68896484375,
0.0985107421875,
0.7197265625,
0.0985107421875,
0.7197265625,
0.10986328125,
0.68896484375,
0.10986328125
],
[
0.69677734375,
0.0947265625,
0.71142578125,
0.0947265625,
0.71142578125,
0.0985107421875,
0.69677734375,
0.0985107421875
],
[
0.488525390625,
0.11297607421875,
0.5048828125,
0.11297607421875,
0.5048828125,
0.1300048828125,
0.488525390625,
0.1300048828125
],
[
0.491943359375,
0.11871337890625,
0.496826171875,
0.11871337890625,
0.496826171875,
0.1256103515625,
0.491943359375,
0.1256103515625
],
[
0.60107421875,
0.117431640625,
0.6259765625,
0.117431640625,
0.6259765625,
0.125,
0.60107421875,
0.125
],
[
0.70166015625,
0.11932373046875,
0.70751953125,
0.11932373046875,
0.70751953125,
0.125,
0.70166015625,
0.125
],
[
0.3505859375,
0.1300048828125,
0.369384765625,
0.1300048828125,
0.369384765625,
0.136962890625,
0.3505859375,
0.136962890625
],
[
0.386474609375,
0.1319580078125,
0.4697265625,
0.1319580078125,
0.4697265625,
0.1427001953125,
0.386474609375,
0.1427001953125
],
[
0.48193359375,
0.1275634765625,
0.50732421875,
0.1275634765625,
0.50732421875,
0.1395263671875,
0.48193359375,
0.1395263671875
],
[
0.259033203125,
0.1363525390625,
0.306396484375,
0.1363525390625,
0.306396484375,
0.146484375,
0.259033203125,
0.146484375
],
[
0.35693359375,
0.14013671875,
0.365234375,
0.14013671875,
0.365234375,
0.1458740234375,
0.35693359375,
0.1458740234375
],
[
0.69384765625,
0.1363525390625,
0.71826171875,
0.1363525390625,
0.71826171875,
0.1483154296875,
0.69384765625,
0.1483154296875
],
[
0.39208984375,
0.143310546875,
0.418212890625,
0.143310546875,
0.418212890625,
0.146484375,
0.39208984375,
0.146484375
],
[
0.76123046875,
0.1427001953125,
0.814453125,
0.1427001953125,
0.814453125,
0.154052734375,
0.76123046875,
0.154052734375
],
[
0.75634765625,
0.15283203125,
0.82861328125,
0.15283203125,
0.82861328125,
0.1634521484375,
0.75634765625,
0.1634521484375
],
[
0.2607421875,
0.146484375,
0.306396484375,
0.146484375,
0.306396484375,
0.1572265625,
0.2607421875,
0.1572265625
],
[
0.69775390625,
0.1717529296875,
0.734375,
0.1717529296875,
0.734375,
0.1817626953125,
0.69775390625,
0.1817626953125
],
[
0.771484375,
0.17041015625,
0.81201171875,
0.17041015625,
0.81201171875,
0.1817626953125,
0.771484375,
0.1817626953125
],
[
0.84716796875,
0.17236328125,
0.865234375,
0.17236328125,
0.865234375,
0.18115234375,
0.84716796875,
0.18115234375
],
[
0.68310546875,
0.173583984375,
0.6953125,
0.173583984375,
0.6953125,
0.177978515625,
0.68310546875,
0.177978515625
],
[
0.1732177734375,
0.1749267578125,
0.1871337890625,
0.1787109375,
0.1829833984375,
0.186279296875,
0.1690673828125,
0.18310546875
],
[
0.330078125,
0.1749267578125,
0.37744140625,
0.1749267578125,
0.37744140625,
0.186279296875,
0.330078125,
0.186279296875
],
[
0.40283203125,
0.1749267578125,
0.43701171875,
0.1749267578125,
0.43701171875,
0.186279296875,
0.40283203125,
0.186279296875
],
[
0.90185546875,
0.1749267578125,
0.90869140625,
0.1749267578125,
0.90869140625,
0.1787109375,
0.90185546875,
0.1787109375
],
[
0.471435546875,
0.175537109375,
0.5185546875,
0.175537109375,
0.5185546875,
0.186279296875,
0.471435546875,
0.186279296875
],
[
0.08905029296875,
0.1787109375,
0.146240234375,
0.18310546875,
0.1446533203125,
0.1912841796875,
0.0882568359375,
0.1875
],
[
0.1968994140625,
0.175537109375,
0.240234375,
0.175537109375,
0.240234375,
0.186279296875,
0.1968994140625,
0.186279296875
],
[
0.58837890625,
0.1787109375,
0.59716796875,
0.1787109375,
0.59716796875,
0.184326171875,
0.58837890625,
0.184326171875
],
[
0.765625,
0.18115234375,
0.81787109375,
0.18115234375,
0.81787109375,
0.1912841796875,
0.765625,
0.1912841796875
],
[
0.68115234375,
0.1817626953125,
0.73388671875,
0.1817626953125,
0.73388671875,
0.189453125,
0.68115234375,
0.189453125
],
[
0.143798828125,
0.1817626953125,
0.158447265625,
0.1817626953125,
0.158447265625,
0.189453125,
0.143798828125,
0.189453125
],
[
0.1854248046875,
0.186279296875,
0.2303466796875,
0.186279296875,
0.2303466796875,
0.1962890625,
0.1854248046875,
0.1962890625
],
[
0.328369140625,
0.186279296875,
0.44775390625,
0.186279296875,
0.44775390625,
0.1962890625,
0.328369140625,
0.1962890625
],
[
0.45751953125,
0.186279296875,
0.53173828125,
0.186279296875,
0.53173828125,
0.1962890625,
0.45751953125,
0.1962890625
],
[
0.85546875,
0.1817626953125,
0.90283203125,
0.1817626953125,
0.90283203125,
0.189453125,
0.85546875,
0.189453125
],
[
0.62744140625,
0.2020263671875,
0.86669921875,
0.2020263671875,
0.86669921875,
0.21533203125,
0.62744140625,
0.21533203125
],
[
0.19287109375,
0.2083740234375,
0.435546875,
0.2083740234375,
0.435546875,
0.2215576171875,
0.19287109375,
0.2215576171875
],
[
0.0849609375,
0.25,
0.4794921875,
0.25,
0.4794921875,
0.26318359375,
0.0849609375,
0.26318359375
],
[
0.51953125,
0.25,
0.9111328125,
0.25,
0.9111328125,
0.26318359375,
0.51953125,
0.26318359375
],
[
0.08660888671875,
0.264404296875,
0.47802734375,
0.264404296875,
0.47802734375,
0.27783203125,
0.08660888671875,
0.27783203125
],
[
0.51953125,
0.264404296875,
0.9111328125,
0.264404296875,
0.9111328125,
0.27783203125,
0.51953125,
0.27783203125
],
[
0.0849609375,
0.27783203125,
0.4794921875,
0.27783203125,
0.4794921875,
0.292236328125,
0.0849609375,
0.292236328125
],
[
0.51953125,
0.27783203125,
0.91259765625,
0.27783203125,
0.91259765625,
0.291748046875,
0.51953125,
0.291748046875
],
[
0.08660888671875,
0.292236328125,
0.298095703125,
0.292236328125,
0.298095703125,
0.30615234375,
0.08660888671875,
0.30615234375
],
[
0.521484375,
0.292236328125,
0.9111328125,
0.292236328125,
0.9111328125,
0.30615234375,
0.521484375,
0.30615234375
],
[
0.08660888671875,
0.311279296875,
0.455810546875,
0.311279296875,
0.455810546875,
0.32568359375,
0.08660888671875,
0.32568359375
],
[
0.51953125,
0.30615234375,
0.9111328125,
0.30615234375,
0.9111328125,
0.3193359375,
0.51953125,
0.3193359375
],
[
0.453369140625,
0.31640625,
0.476318359375,
0.31640625,
0.476318359375,
0.324462890625,
0.453369140625,
0.324462890625
],
[
0.51953125,
0.3193359375,
0.91015625,
0.32080078125,
0.91015625,
0.33203125,
0.51953125,
0.330810546875
],
[
0.0849609375,
0.32763671875,
0.47802734375,
0.32763671875,
0.47802734375,
0.341552734375,
0.0849609375,
0.341552734375
],
[
0.51953125,
0.333984375,
0.91259765625,
0.333984375,
0.91259765625,
0.347900390625,
0.51953125,
0.347900390625
],
[
0.0849609375,
0.341552734375,
0.47802734375,
0.341552734375,
0.47802734375,
0.354736328125,
0.0849609375,
0.354736328125
],
[
0.521484375,
0.347900390625,
0.9111328125,
0.347900390625,
0.9111328125,
0.361083984375,
0.521484375,
0.361083984375
],
[
0.08660888671875,
0.354248046875,
0.476318359375,
0.354248046875,
0.476318359375,
0.36865234375,
0.08660888671875,
0.36865234375
],
[
0.521484375,
0.3623046875,
0.9111328125,
0.3623046875,
0.9111328125,
0.37255859375,
0.521484375,
0.37255859375
],
[
0.0849609375,
0.369384765625,
0.47802734375,
0.369384765625,
0.47802734375,
0.38330078125,
0.0849609375,
0.38330078125
],
[
0.51953125,
0.375,
0.9140625,
0.375,
0.9140625,
0.38818359375,
0.51953125,
0.38818359375
],
[
0.0849609375,
0.38330078125,
0.4794921875,
0.38330078125,
0.4794921875,
0.396484375,
0.0849609375,
0.396484375
],
[
0.51953125,
0.38818359375,
0.91259765625,
0.38818359375,
0.91259765625,
0.401611328125,
0.51953125,
0.401611328125
],
[
0.08660888671875,
0.396484375,
0.476318359375,
0.396484375,
0.476318359375,
0.410400390625,
0.08660888671875,
0.410400390625
],
[
0.521484375,
0.404052734375,
0.90869140625,
0.404052734375,
0.90869140625,
0.4140625,
0.521484375,
0.4140625
],
[
0.08660888671875,
0.410400390625,
0.476318359375,
0.410400390625,
0.476318359375,
0.423583984375,
0.08660888671875,
0.423583984375
],
[
0.51953125,
0.416748046875,
0.9111328125,
0.416748046875,
0.9111328125,
0.429931640625,
0.51953125,
0.429931640625
],
[
0.08660888671875,
0.4248046875,
0.4794921875,
0.4248046875,
0.4794921875,
0.438232421875,
0.08660888671875,
0.438232421875
],
[
0.51953125,
0.429931640625,
0.9111328125,
0.429931640625,
0.9111328125,
0.4443359375,
0.51953125,
0.4443359375
],
[
0.0849609375,
0.438232421875,
0.4794921875,
0.438232421875,
0.4794921875,
0.451904296875,
0.0849609375,
0.451904296875
],
[
0.521484375,
0.443115234375,
0.91015625,
0.443115234375,
0.91015625,
0.45703125,
0.521484375,
0.45703125
],
[
0.08660888671875,
0.451904296875,
0.4794921875,
0.451904296875,
0.4794921875,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.51953125,
0.458251953125,
0.91259765625,
0.458251953125,
0.91259765625,
0.4716796875,
0.51953125,
0.4716796875
],
[
0.0849609375,
0.466552734375,
0.4794921875,
0.466552734375,
0.4794921875,
0.479736328125,
0.0849609375,
0.479736328125
],
[
0.521484375,
0.4716796875,
0.9140625,
0.4716796875,
0.9140625,
0.48486328125,
0.521484375,
0.48486328125
],
[
0.0849609375,
0.479736328125,
0.476318359375,
0.479736328125,
0.476318359375,
0.49365234375,
0.0849609375,
0.49365234375
],
[
0.51953125,
0.486083984375,
0.91259765625,
0.486083984375,
0.91259765625,
0.5,
0.51953125,
0.5
],
[
0.08660888671875,
0.49365234375,
0.476318359375,
0.49365234375,
0.476318359375,
0.50830078125,
0.08660888671875,
0.50830078125
],
[
0.51953125,
0.5,
0.63818359375,
0.5,
0.63818359375,
0.51318359375,
0.51953125,
0.51318359375
],
[
0.08660888671875,
0.50830078125,
0.47802734375,
0.50830078125,
0.47802734375,
0.52294921875,
0.08660888671875,
0.52294921875
],
[
0.53515625,
0.51318359375,
0.9111328125,
0.51318359375,
0.9111328125,
0.5263671875,
0.53515625,
0.5263671875
],
[
0.08660888671875,
0.521484375,
0.4794921875,
0.521484375,
0.4794921875,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.521484375,
0.52880859375,
0.9111328125,
0.52880859375,
0.9111328125,
0.5390625,
0.521484375,
0.5390625
],
[
0.08660888671875,
0.53515625,
0.459228515625,
0.53515625,
0.459228515625,
0.548828125,
0.08660888671875,
0.548828125
],
[
0.4599609375,
0.54150390625,
0.47216796875,
0.54150390625,
0.47216796875,
0.54541015625,
0.4599609375,
0.54541015625
],
[
0.521484375,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55322265625,
0.521484375,
0.55322265625
],
[
0.08660888671875,
0.548828125,
0.4794921875,
0.548828125,
0.4794921875,
0.5625,
0.08660888671875,
0.5625
],
[
0.51953125,
0.5546875,
0.91259765625,
0.5546875,
0.91259765625,
0.568359375,
0.51953125,
0.568359375
],
[
0.0882568359375,
0.564453125,
0.476318359375,
0.564453125,
0.476318359375,
0.57470703125,
0.0882568359375,
0.57470703125
],
[
0.521484375,
0.568359375,
0.91259765625,
0.568359375,
0.91259765625,
0.58203125,
0.521484375,
0.58203125
],
[
0.0849609375,
0.5771484375,
0.4794921875,
0.5771484375,
0.4794921875,
0.59033203125,
0.0849609375,
0.59033203125
],
[
0.521484375,
0.58203125,
0.91259765625,
0.58203125,
0.91259765625,
0.59521484375,
0.521484375,
0.59521484375
],
[
0.0849609375,
0.59033203125,
0.2484130859375,
0.59033203125,
0.2484130859375,
0.60400390625,
0.0849609375,
0.60400390625
],
[
0.521484375,
0.59765625,
0.9111328125,
0.59765625,
0.9111328125,
0.60791015625,
0.521484375,
0.60791015625
],
[
0.10455322265625,
0.60595703125,
0.476318359375,
0.60595703125,
0.476318359375,
0.6162109375,
0.10455322265625,
0.6162109375
],
[
0.51953125,
0.60986328125,
0.91259765625,
0.60986328125,
0.91259765625,
0.62353515625,
0.51953125,
0.62353515625
],
[
0.0849609375,
0.61865234375,
0.476318359375,
0.61865234375,
0.476318359375,
0.6318359375,
0.0849609375,
0.6318359375
],
[
0.51953125,
0.62353515625,
0.91015625,
0.62353515625,
0.91015625,
0.63720703125,
0.51953125,
0.63720703125
],
[
0.0849609375,
0.6318359375,
0.47802734375,
0.6318359375,
0.47802734375,
0.64599609375,
0.0849609375,
0.64599609375
],
[
0.51953125,
0.63818359375,
0.900390625,
0.63818359375,
0.900390625,
0.6513671875,
0.51953125,
0.6513671875
],
[
0.0849609375,
0.646484375,
0.47802734375,
0.646484375,
0.47802734375,
0.66015625,
0.0849609375,
0.66015625
],
[
0.0849609375,
0.6591796875,
0.4755859375,
0.6591796875,
0.4755859375,
0.67236328125,
0.0849609375,
0.67236328125
],
[
0.521484375,
0.66357421875,
0.74169921875,
0.66357421875,
0.74169921875,
0.67822265625,
0.521484375,
0.67822265625
],
[
0.0849609375,
0.673828125,
0.4794921875,
0.673828125,
0.4794921875,
0.6875,
0.0849609375,
0.6875
],
[
0.51953125,
0.681640625,
0.9111328125,
0.681640625,
0.9111328125,
0.69580078125,
0.51953125,
0.69580078125
],
[
0.08660888671875,
0.689453125,
0.476318359375,
0.689453125,
0.476318359375,
0.69970703125,
0.08660888671875,
0.69970703125
],
[
0.51953125,
0.69580078125,
0.9111328125,
0.69580078125,
0.9111328125,
0.708984375,
0.51953125,
0.708984375
],
[
0.08660888671875,
0.7021484375,
0.47802734375,
0.7021484375,
0.47802734375,
0.71533203125,
0.08660888671875,
0.71533203125
],
[
0.0849609375,
0.71533203125,
0.476318359375,
0.71533203125,
0.476318359375,
0.72900390625,
0.0849609375,
0.72900390625
],
[
0.521484375,
0.71044921875,
0.9111328125,
0.71044921875,
0.9111328125,
0.7236328125,
0.521484375,
0.7236328125
],
[
0.51953125,
0.7236328125,
0.91259765625,
0.7236328125,
0.91259765625,
0.7373046875,
0.51953125,
0.7373046875
],
[
0.08660888671875,
0.73095703125,
0.47802734375,
0.73095703125,
0.47802734375,
0.7412109375,
0.08660888671875,
0.7412109375
],
[
0.0849609375,
0.7421875,
0.476318359375,
0.7421875,
0.476318359375,
0.755859375,
0.0849609375,
0.755859375
],
[
0.521484375,
0.7373046875,
0.9111328125,
0.7373046875,
0.9111328125,
0.75048828125,
0.521484375,
0.75048828125
],
[
0.08416748046875,
0.7568359375,
0.4794921875,
0.755859375,
0.4794921875,
0.77099609375,
0.08416748046875,
0.771484375
],
[
0.521484375,
0.751953125,
0.91015625,
0.751953125,
0.91015625,
0.76513671875,
0.521484375,
0.76513671875
],
[
0.51953125,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.51953125,
0.77880859375
],
[
0.08660888671875,
0.77294921875,
0.476318359375,
0.77294921875,
0.476318359375,
0.78271484375,
0.08660888671875,
0.78271484375
],
[
0.0849609375,
0.7841796875,
0.4794921875,
0.7841796875,
0.4794921875,
0.79736328125,
0.0849609375,
0.79736328125
],
[
0.51953125,
0.77880859375,
0.9111328125,
0.77880859375,
0.9111328125,
0.79248046875,
0.51953125,
0.79248046875
],
[
0.51953125,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.798828125,
0.476318359375,
0.798828125,
0.476318359375,
0.8125,
0.0849609375,
0.8125
],
[
0.08416748046875,
0.81103515625,
0.4794921875,
0.81298828125,
0.4794921875,
0.8271484375,
0.08416748046875,
0.82470703125
],
[
0.521484375,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.0849609375,
0.82568359375,
0.2156982421875,
0.82568359375,
0.2156982421875,
0.8388671875,
0.0849609375,
0.8388671875
],
[
0.521484375,
0.8212890625,
0.91015625,
0.8212890625,
0.91015625,
0.83203125,
0.521484375,
0.83203125
],
[
0.51953125,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.51953125,
0.84765625
],
[
0.0849609375,
0.8466796875,
0.476318359375,
0.84765625,
0.476318359375,
0.8623046875,
0.0849609375,
0.861328125
],
[
0.521484375,
0.849609375,
0.9111328125,
0.849609375,
0.9111328125,
0.85986328125,
0.521484375,
0.85986328125
],
[
0.08660888671875,
0.86279296875,
0.47802734375,
0.86279296875,
0.47802734375,
0.87353515625,
0.08660888671875,
0.87353515625
],
[
0.51708984375,
0.8623046875,
0.9140625,
0.861328125,
0.9140625,
0.875,
0.51708984375,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.5185546875,
0.87548828125,
0.91015625,
0.87548828125,
0.91015625,
0.8896484375,
0.5185546875,
0.8896484375
]
] |
[
"3200Dx2",
"Condition",
"6",
"Transformatior",
"generator",
"predictor",
"tracking",
"Las",
"+",
"?",
"ADTO",
"4",
"28D",
"(cosine similarity)",
"128D",
"Condition",
"",
"128D",
"",
"Next action",
"ecommendation",
"generator",
"oryRNN",
"Localized",
"Reco",
"Hie",
"Pre",
"Pre/Post-",
"Action",
"pr",
"Predicted",
"Videosegm",
"omputed",
"fis",
"action node",
"hiddenstate",
"ent",
"features",
"conditionsEmbedding",
"post-condition",
"nextaction",
"Figure 4: Using ADTG for guidance",
"Figure3:Action embedding network",
"form a procedural task. We consider multiple scenarios of",
"come from either the very beginning of the task (i.e., com-",
"guidance, including task tracking, next action recommen-",
"plete plan generation) or from a random midpoint in the ex-",
"dation, and plan generation. A working pipeline of using",
"ecution of the task (i.e., planning after a prefix observation",
"ADTG is illustrated in Figure 4.",
"sequence). In our approach, plan generation is done by ap-",
"Task Tracking and Next Action Recommendation",
"plying one step of task tracking, followed by multiple steps",
"In",
"ofnextactionrecommendation in an autoregressiveway",
"task tracking, our model tries to track the human's progress",
"Specifically, given an initial visual observation and initial",
"when performing a certain task, by localizing the current vi-",
"action history, we pass them through the task tracking mod-",
"sual observation in the corresponding ADTG graph. For any",
"uletorecognize the currentaction,and localizetoanaction",
"time step t e {1, :.., T] of a given video, the inputs to the",
"node in the ADTG. Using the embedding of the localized",
"task tracking module include the current visual observation",
"action and the action history information as input, we ob-",
"Ct, the action history information ht-1 up to time step t- 1,",
"tainthenextactionto performfromthenextactionrecom",
"and the embedding of a candidate action ea being evaluated.",
"mendation module. Finally, we update the action history in-",
"The task tracking module then outputs a confidence score",
"formation by passing the newest action through the history",
"indicating whether the candidate action a captures the given",
"RNN, and recursively invoke the next action recommenda-",
"visual observation t. We then enumerate all the candidate",
"tion module to generate a sequence of actions to perform.",
"actions and localize to the action node at with the highest",
"Repeating this process will lead to a plan to accomplish the",
"confidence score in the ADTG. In particular, to get a suc-",
"task. In this sense, our plan generation is built entirely upon",
"cinct representation of the action history, we pass the se-",
"existing modules.",
"quence of history action embeddings (ea1, : .:, Eat-1) up to",
"In the recursive process of plan generation, instead of fol-",
"time t-1 through a recurrent neural network (RNN), and use",
"lowingthegreedyactionchoiceoutputtedbythenextaction",
"the RNN hidden state as the action history information ht",
"1",
"recommendationmoduleateachstep,weusebeam search",
"We train the task tracking module and the history RNN over",
"to select the output sequence with higher overall likelihood.",
"allthetimesteps of thetrainingvideos,andusethecross en-",
"Specifically, for a beam search width of k, we maintain",
"tropy loss to update the network parameters for both of them",
"k candidate action trajectories,perform rollouts based on",
"using back-propagation.",
"thesetrajectories,and onlykeeptherollouttrajectorieswith",
"In nextaction recommendation,our modelrecommends",
"the top k highest likelihoods at each step. In this way, our",
"to the user an action to perform in the next time step. For any",
"method strikes a good balance between computation com-",
"time step t, the next action recommendation module takes as",
"plexity and the optimality of the selected action sequence.",
"input the currently localized action node embedding eat, the",
"action history information ht, and the embedding of a candi-",
"4.3Implementation Details",
"date next action ea being evaluated. It maps these inputs to a",
"In our implementation, the condition generator, transforma-",
"confidencescoreastheoutput,indicating whereitbelieves",
"tion predictor, task tracking module, and the next action rec-",
"that the evaluated candidate action is a suitable one for the",
"next time step. We then enumerate all the candidate next ac-",
"ommendation module are all instantiated as two-layer feed-",
"forward neural networks. The pre- and post-conditions are",
"tionsfromtheoutgoingedgesof thelocalizedactionnodein",
"the ADTG, and select the action with the highest confidence",
"128-dimensional semantic vectors, and the action embed-",
"score as the recommendation at+1 to the user. The action",
"ding vectors are 96-dimensional. The hidden size of the his-",
"tory RNN is 128. The margin threshold for contrastive loss",
"historyinformationcomesfromthehidden stateofthesame",
"history RNN as in the task tracking module. We again use",
"in Equation (2) is set to be M = 0.5. In the plan genera-",
"tion module, the beam search width is k = 5. We use the",
"the cross entropy loss to train the next action recommenda-",
"tion module over any consecutive action pairs (at,at+i)in",
"ADAM optimizer (Kingma and Ba 2014) with a learning",
"the training videos.",
"rateofle-5fortrainingtheactionembeddings,andlearn-",
"ing rate 5e -5 for training the task tracking and next action",
"Plan Generation Given a visual observation, plan gener.",
"recommendationmodules.Theactionembeddingsandtask",
"ationis used to generatea sequence of actions to accomplish",
"tracking modulearetrained for50 epochs,whilethe next",
"the task from this point onward. The visual observation can",
"action recommendation modules takes 100 training epochs."
] |
{
"arxiv_id": "2302.05330",
"language": "en",
"page_idx": 4
}
|
||
2306.09192-en-0
|
2306.09192-en-0.jpg
|
# Training Diffusion Classifiers with $\textbf{Denoising Assistance}$
Sri Harsha Dumpala
Dalhousie University &Vector Institute
sriharsha.d.ece@gmail.com
Chandramouli Sastry
Dalhousie University & Vector Institute chandramouli.sastry@gmail.com
Sageev Oore
Dalhousie University &Vector Institute
osageev@gmail.com
### $\mathbf{Abstract}$
Score-matching and diffusion models have emerged as state-of-the-art generative models for both conditional and unconditional generation. Classifier-guided diffusion models are created by training a classifier on samples obtained from the forward-diffusion process(i.e., from data to noise). In this paper, we propose denoising-assisted $\hat{(}$DA) classifiers wherein the diffusion classifier is trained using both noisy and denoised examples as simultaneous inputs to the model. We differentiate between denoising-assisted (DA) classifiers and noisy classifiers, which are diffusion classifiers that are only trained on noisy examples. Our experiments on Cifar10 and Imagenet show that DA-classifiers improve over noisy classifiers both quantitatively in terms of generalization to test data and qualitatively in terms of perceptually-aligned classifier-gradients and generative modeling metrics. Finally, we describe a semi-supervised framework for training diffusion classifiers and our experiments, that also include positive-unlabeled settings, demonstrate improved generalization of DA-classifiers over noisy classifiers.
## 1 Introduction
Score models are unnormalized probabilistic models that model the probability density in terms of its score function- that is, the gradient of the log-likelihood. Score models are advantageous as compared to contemporary generative models such as GANs, VAEs, Autoregressive models and normalizing flows as they require neither adversarial optimization nor restricted architecture families, while achieving state-of-the-art performance across multiple modalities. Score models can be trained using any score-matching objective such as the implicit score-matching [Hyvärinen, 2005],the sliced score-matching [Song et al., 2020b] or the denoising score-matching [Vincent, 2011] methods. Song et al.[2021] use denoising score-matching to unify Multi-scale Score Matching [Song and Ermon, 2019, 2020] and Denoising Diffusion Probabilistic Models [Ho et al., 2020] under a stochastic diffusion framework. We briefly review the stochastic diffusion score-matching framework introduced by Song et al.[2021] in Section 2.1.
In this work, we focus on classifier-guided diffusion models which utilize the gradients of a classifier, with respect to its input, that maximize the log-probability of a target class to guide the sampling. To this end, a classifier is trained on noisy images sampled from the forward diffusion as described in Section 2.2; the forward diffusion is usually described such that the distribution of examples at a given diffusion time can be specified as a closed-form Gaussian implying that the examples sampled from the diffusion contain Gaussian noise. We use the term diffusion classifiers to refer to classifrers trained on diffused examples. We introduce a simple modification to the training of the
Preprint. Under review.
|
[
[
0.291748046875,
0.1256103515625,
0.7060546875,
0.1256103515625,
0.7060546875,
0.146484375,
0.291748046875,
0.146484375
],
[
0.371826171875,
0.15087890625,
0.6259765625,
0.1514892578125,
0.6259765625,
0.17236328125,
0.371826171875,
0.1717529296875
],
[
0.26953125,
0.2279052734375,
0.420654296875,
0.2279052734375,
0.420654296875,
0.2412109375,
0.26953125,
0.2412109375
],
[
0.57861328125,
0.2259521484375,
0.72802734375,
0.2266845703125,
0.72802734375,
0.2412109375,
0.57861328125,
0.240478515625
],
[
0.2132568359375,
0.240478515625,
0.4755859375,
0.2412109375,
0.4755859375,
0.255615234375,
0.2132568359375,
0.255126953125
],
[
0.52294921875,
0.242431640625,
0.78369140625,
0.242431640625,
0.78369140625,
0.255615234375,
0.52294921875,
0.255615234375
],
[
0.221435546875,
0.255615234375,
0.471435546875,
0.255615234375,
0.471435546875,
0.26953125,
0.221435546875,
0.26953125
],
[
0.5458984375,
0.255615234375,
0.76123046875,
0.255615234375,
0.76123046875,
0.26953125,
0.5458984375,
0.26953125
],
[
0.0228729248046875,
0.266357421875,
0.060455322265625,
0.266357421875,
0.060455322265625,
0.40283203125,
0.0228729248046875,
0.40283203125
],
[
0.455078125,
0.290283203125,
0.5439453125,
0.290283203125,
0.5439453125,
0.304931640625,
0.455078125,
0.304931640625
],
[
0.369384765625,
0.3037109375,
0.6298828125,
0.3037109375,
0.6298828125,
0.318115234375,
0.369384765625,
0.318115234375
],
[
0.4248046875,
0.3193359375,
0.57421875,
0.3193359375,
0.57421875,
0.333251953125,
0.4248046875,
0.333251953125
],
[
0.461669921875,
0.3662109375,
0.53759765625,
0.36865234375,
0.53759765625,
0.3837890625,
0.461669921875,
0.3818359375
],
[
0.021240234375,
0.396484375,
0.06451416015625,
0.396484375,
0.06451416015625,
0.488037109375,
0.021240234375,
0.488037109375
],
[
0.233642578125,
0.399658203125,
0.765625,
0.399658203125,
0.765625,
0.41357421875,
0.233642578125,
0.41357421875
],
[
0.2320556640625,
0.41357421875,
0.765625,
0.41357421875,
0.765625,
0.4267578125,
0.2320556640625,
0.4267578125
],
[
0.2344970703125,
0.4287109375,
0.7646484375,
0.4287109375,
0.7646484375,
0.439453125,
0.2344970703125,
0.439453125
],
[
0.2320556640625,
0.43994140625,
0.765625,
0.44140625,
0.765625,
0.455810546875,
0.2320556640625,
0.455078125
],
[
0.2320556640625,
0.453857421875,
0.765625,
0.455078125,
0.765625,
0.4697265625,
0.2320556640625,
0.468505859375
],
[
0.233642578125,
0.468505859375,
0.76708984375,
0.468505859375,
0.76708984375,
0.481689453125,
0.233642578125,
0.481689453125
],
[
0.047393798828125,
0.481689453125,
0.052276611328125,
0.481689453125,
0.052276611328125,
0.4873046875,
0.047393798828125,
0.4873046875
],
[
0.233642578125,
0.48291015625,
0.765625,
0.48291015625,
0.765625,
0.496826171875,
0.233642578125,
0.496826171875
],
[
0.0228729248046875,
0.4912109375,
0.05963134765625,
0.4912109375,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.041656494140625,
0.496826171875,
0.0482177734375,
0.496826171875,
0.0482177734375,
0.5,
0.041656494140625,
0.5
],
[
0.233642578125,
0.496826171875,
0.765625,
0.496826171875,
0.765625,
0.51025390625,
0.233642578125,
0.51025390625
],
[
0.233642578125,
0.51025390625,
0.76708984375,
0.51025390625,
0.76708984375,
0.5234375,
0.233642578125,
0.5234375
],
[
0.233642578125,
0.5244140625,
0.76708984375,
0.5244140625,
0.76708984375,
0.53857421875,
0.233642578125,
0.53857421875
],
[
0.233642578125,
0.53857421875,
0.76708984375,
0.53857421875,
0.76708984375,
0.5517578125,
0.233642578125,
0.5517578125
],
[
0.233642578125,
0.5517578125,
0.765625,
0.5517578125,
0.765625,
0.56494140625,
0.233642578125,
0.56494140625
],
[
0.233642578125,
0.56494140625,
0.765625,
0.56494140625,
0.765625,
0.5791015625,
0.233642578125,
0.5791015625
],
[
0.2320556640625,
0.5791015625,
0.58642578125,
0.57763671875,
0.58642578125,
0.59228515625,
0.2320556640625,
0.59326171875
],
[
0.17724609375,
0.619140625,
0.188720703125,
0.619140625,
0.188720703125,
0.62890625,
0.17724609375,
0.62890625
],
[
0.188720703125,
0.62255859375,
0.1953125,
0.62255859375,
0.1953125,
0.62548828125,
0.188720703125,
0.62548828125
],
[
0.1968994140625,
0.61865234375,
0.31201171875,
0.61865234375,
0.31201171875,
0.6318359375,
0.1968994140625,
0.6318359375
],
[
0.1732177734375,
0.6484375,
0.826171875,
0.6484375,
0.826171875,
0.662109375,
0.1732177734375,
0.662109375
],
[
0.1732177734375,
0.662109375,
0.826171875,
0.66357421875,
0.826171875,
0.6767578125,
0.1732177734375,
0.67529296875
],
[
0.17236328125,
0.67529296875,
0.82421875,
0.67529296875,
0.82421875,
0.689453125,
0.17236328125,
0.689453125
],
[
0.1748046875,
0.68994140625,
0.82421875,
0.68994140625,
0.82421875,
0.7041015625,
0.1748046875,
0.7041015625
],
[
0.176513671875,
0.70556640625,
0.82421875,
0.70556640625,
0.82421875,
0.71630859375,
0.176513671875,
0.71630859375
],
[
0.1748046875,
0.71826171875,
0.826171875,
0.71826171875,
0.826171875,
0.73193359375,
0.1748046875,
0.73193359375
],
[
0.1748046875,
0.73193359375,
0.826171875,
0.73193359375,
0.826171875,
0.74560546875,
0.1748046875,
0.74560546875
],
[
0.1732177734375,
0.744140625,
0.82421875,
0.74560546875,
0.82421875,
0.7587890625,
0.1732177734375,
0.75830078125
],
[
0.1732177734375,
0.7587890625,
0.826171875,
0.7587890625,
0.826171875,
0.77294921875,
0.1732177734375,
0.77294921875
],
[
0.1732177734375,
0.771484375,
0.826171875,
0.77294921875,
0.826171875,
0.787109375,
0.1732177734375,
0.7861328125
],
[
0.176513671875,
0.78857421875,
0.483642578125,
0.78857421875,
0.483642578125,
0.798828125,
0.176513671875,
0.798828125
],
[
0.1732177734375,
0.806640625,
0.82666015625,
0.806640625,
0.82666015625,
0.82080078125,
0.1732177734375,
0.82080078125
],
[
0.1732177734375,
0.8212890625,
0.826171875,
0.8212890625,
0.826171875,
0.83544921875,
0.1732177734375,
0.83544921875
],
[
0.1748046875,
0.83544921875,
0.82421875,
0.83544921875,
0.82421875,
0.8486328125,
0.1748046875,
0.8486328125
],
[
0.17236328125,
0.84765625,
0.82421875,
0.8486328125,
0.82421875,
0.8623046875,
0.17236328125,
0.861328125
],
[
0.176513671875,
0.8642578125,
0.82275390625,
0.8642578125,
0.82275390625,
0.875,
0.176513671875,
0.875
],
[
0.1748046875,
0.876953125,
0.82421875,
0.876953125,
0.82421875,
0.89013671875,
0.1748046875,
0.89013671875
],
[
0.1748046875,
0.89013671875,
0.82421875,
0.89013671875,
0.82421875,
0.90380859375,
0.1748046875,
0.90380859375
],
[
0.1748046875,
0.92529296875,
0.310546875,
0.92529296875,
0.310546875,
0.93603515625,
0.1748046875,
0.93603515625
]
] |
[
"Training Diffusion Classifiers with",
"Denoising Assistance",
"Chandramouli Sastry",
"Sri Harsha Dumpala",
"DalhousieUniversity&Vector Institute",
"DalhousieUniversity&VectorInstitute",
"chandramouli.sastry@gmail.com",
"sriharsha.d.ece@gmail.com",
" ",
"Sageev Oore",
"Dalhousie University & Vector Institute",
"osageev@gmail.com",
"Abstract",
"[A's]",
"Score-matching and diffusion models have emerged as state-of-the-art generative",
"models for both conditional and unconditional generation. Classifier-guided dif-",
"fusionmodelsarecreatedbytraininga classifieron samples obtainedfromthe",
"forward-diffusion process (i.e., from data to noise). In this paper, we propose",
"denoising-assisted (DA) classifiers wherein the diffusion classifier is trained using",
"both noisy and denoised examples as simultaneous inputs to the model. We differ-",
"",
"entiate between denoising-assisted (DA) classifiers and noisy classifiers, which are",
"261609051",
"",
"diffusion classifiers that are only trained on noisy examples. Our experiments on",
"Cifar1O and Imagenet show that DA-classifiers improve over noisy classifiers both",
"quantitatively in terms of generalization to test data and qualitatively in terms of",
"perceptually-aligned classifier-gradients and generative modeling metrics. Finally,",
"we describe a semi-supervised framework for training diffusion classifiers and our",
"experiments, that also include positive-unlabeled settings, demonstrate improved",
"generalization of DA-classifiers over noisy classifiers.",
"1",
"",
"Introduction",
"Score models are unnormalized probabilistic models that model the probability density in terms",
"of its scorefunction-that is, thegradient of the log-likelihood.Scoremodels are advantageous",
"as compared to contemporary generative models such as GANs, VAEs, Autoregressive models",
"and normalizing flows as they require neither adversarial optimization nor restricted architecture",
"families,whileachieving state-of-the-artperformance acrossmultiplemodalities.Scoremodels",
"can be trained using any score-matching objective such as the implicit score-matching [Hyvarinen,",
"2005], the sliced score-matching [Song et al., 2020b] or the denoising score-matching [Vincent, 2011]",
"methods. Song et al. [2021] use denoising score-matching to unify Multi-scale Score Matching [Song",
"and Ermon, 2019, 2020] and Denoising Diffusion Probabilistic Models [Ho et al., 2020] under a",
"stochastic diffusion framework. We briefly review the stochastic diffusion score-matching framework",
"introducedbySongetal.[2021]inSection2.1",
"In this work, we focus on classifier-guided diffusion models which utilize the gradients of a classifier,",
"with respect to its input, that maximize the log-probability of a target class to guide the sampling.",
"To this end, a classifier is trained on noisy images sampled from the forward diffusion as described",
"in Section 2.2; theforward diffusion is usuallydescribed such that the distribution of examples",
"at agiven diffusiontime can be specifiedasa closed-form Gaussian implying thatthe examples",
"sampled from the diffusion contain Gaussian noise. We use the term diffusion classifiers to refer to",
"classifiers trained on diffused examples. We introduce a simple modification to the training of the",
"Preprint.Underreview"
] |
{
"arxiv_id": "2306.09192",
"language": "en",
"page_idx": 0
}
|
||
2306.09192-en-1
|
2306.09192-en-1.jpg
|
classifer: instead of providing only the noisy image to the classifier, we give both the noisy image and also its corresponding denoised image as simultaneous inputs to the classifier; we refer to this as a denoising-assisted (DA) classifier. To train the DA-classifier, we reuse the same denoising models used to train the diffusion model. Our experiments show that this simple modification results in improved generalization to unseen examples. We also qualitatively analyse the gradients and observe that the gradients of DA-classifiers are improved in terms of perceptual alignment.
Finally, we propose a framework for semi-supervised training of classifier-guided conditional scoremodels and evaluate the benefits of DA-classifiers in label-limited settings: more specifically, the training data consists of a largely unlabelled dataset, along with a relatively small number of labelled examples. We consider the following semi-supervised learning settings:(a) classical multi-class semi-supervised learning wherein the labeled dataset consists of a few examples from each class, and (b) binary positive-unlabeled learning in which a few examples from the positive class are labeled and the unlabeled examples are a mix of positive and negative examples
Contributions. In summary:
(1) We introduce denoising-assisted (DA) classifiers.
(2) We analyze DA-classifiers empirically with Imagenet and CIFAR10 datasets, showing improvements over noisy diffusion classifiers including: (a) improved generalization to unseen examples; (b) qualitatively improved classifier-gradients; and (c) improved image generation.
(3) We introduce a semi-supervised framework for training diffusion classiffers, inspired by state-ofthe-art semi-supervised methods, wherein we train classifiers with (a) a novel temporal-consistency loss and (b) show that denoising-assisted classifiers generalize better.
# 2 Background
2.1 $\textbf{Unconditional Score- based SDE models}$
Score models are probabilistic models of the data that enable sampling and exact inference of loglikelihoods. Song et al.[2021] propose a framework generalizing Multi-scale score matching Song and Ermon [2019, 2020] and Denoising Diffusion Probabilistic Models Ho et al.[2020]. Concretely, the framework consists of two components: 1) the forward-diffusion (i.e., data to noise) stochastic process, and 2) a learnable score-function that can then be used for the reverse-diffusion (i.e., noise to data) stochastic process.
The forward diffusion stochastic process $\{\mathbf{x}(t)\}_{t\in[0,T]}$ starts at data, $\mathbf{x}(0)$, and ends at noise, $\mathbf{x}(T).$ We let $p_t(\mathbf{x})$ denote the probability density of x at time $t$,so, e.g.$,p_0(\mathbf{x})$ is the distribution of the data, and $p_T(\mathbf{x})$ is the distribution of the noise. The diffusion is structured so that $p_T(\mathbf{x})$ is independent of the starting point at $t=0.$ This process is defined with a stochastic-differential-equation (SDE):
$$d\mathbf{x}=\mathbf{f}(\mathbf{x},t)\:dt+g(t)\:d\mathbf{w},$$
(1)
where w denotes a standard Wiener process, $\mathbf{f}(\mathbf{x}(t),t)$ is a drift coefficient, and $g(t)$ is a diffusion coefficient. The drift and diffusion coefficients are usually manually specified without learnable parameters; this lets us obtain closed-form solutions to the forward-diffusion SDE. For example, if f $\hat{\text{is linear in x, the solution to the SDE is a time-varying gaussian distribution whose mean }\hat{\mu}(t)\mathrm{~and}}$ standard deviation $\sigma(t)$ can be exactly computed. We use $p_t(\mathbf{x}|\mathbf{x}_0)$ to denote the probability density function of x$(t)$ when the diffusion is seeded at $\mathbf{x}_0.$
To sample from $p_0(\mathbf{x})$ starting with samples from $p_T(\mathbf{x})$,we have to solve the reverse diffusion SDE
[Anderson, 1982]:
(2)
$$d\mathbf{x}=\left[\mathbf{f}(\mathbf{x},t)-g(t)^2\nabla_\mathbf{x}\log p_t(\mathbf{x})\right]dt+g(t)\:d\mathbf{\bar{w}},$$
where $d\bar{\mathbf{w}}$ is a standard Wiener process when time flows from T to 0, and $dt$ is an infinitesimal negative timestep. In practice, the score function $\nabla_\mathbf{x}\log p_{t}(\mathbf{x})$ is estimated by a neural network $s_{\theta }( \mathbf{x} , t) , \bar {\text{parameterized by }\theta }$,trained to optimize the following score-matching loss:
(3)
$$\int_0^T\mathbb{E}_{\mathbf{x}\sim p_t(\mathbf{x})}[\lambda(t)||\nabla_{\mathbf{x}}\log p_t(\mathbf{x})-\mathbf{s}_\theta(\mathbf{x},t)||_2^2]dt$$
where $\lambda(t)$ is a positive real number introduced to balance out the score-matching objective across various time steps. Using samples from the training dataset, we can define an empirical density ${\mathrm{function~for~}}t=0{\mathrm{~as~}}p_{0}({\mathbf{x}})={\frac{1}{N}}\sum_{{\mathbf{x}}_{i}\in p_{data}}\delta({\mathbf{x}}-{\mathbf{x}}_{i}){\mathrm{~and~then~obtain~samples~from~}}p_{t}({\mathbf{x}}){\mathrm{~by~first}}$
2
|
[
[
0.1732177734375,
0.091552734375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.106689453125,
0.1732177734375,
0.1060791015625
],
[
0.1732177734375,
0.1060791015625,
0.82421875,
0.106689453125,
0.82421875,
0.12060546875,
0.1732177734375,
0.11932373046875
],
[
0.176513671875,
0.121826171875,
0.82275390625,
0.121826171875,
0.82275390625,
0.1319580078125,
0.176513671875,
0.1319580078125
],
[
0.1732177734375,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.1732177734375,
0.146484375
],
[
0.1748046875,
0.147705078125,
0.82275390625,
0.147705078125,
0.82275390625,
0.1610107421875,
0.1748046875,
0.1610107421875
],
[
0.1748046875,
0.1622314453125,
0.7158203125,
0.1622314453125,
0.7158203125,
0.175537109375,
0.1748046875,
0.175537109375
],
[
0.1748046875,
0.1817626953125,
0.826171875,
0.1817626953125,
0.826171875,
0.1962890625,
0.1748046875,
0.1962890625
],
[
0.176513671875,
0.1976318359375,
0.82177734375,
0.1976318359375,
0.82177734375,
0.2083740234375,
0.176513671875,
0.2083740234375
],
[
0.1748046875,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1748046875,
0.2235107421875,
0.82177734375,
0.2235107421875,
0.82177734375,
0.2374267578125,
0.1748046875,
0.2374267578125
],
[
0.1748046875,
0.2374267578125,
0.82275390625,
0.2374267578125,
0.82275390625,
0.250732421875,
0.1748046875,
0.250732421875
],
[
0.1748046875,
0.251953125,
0.82421875,
0.251953125,
0.82421875,
0.26513671875,
0.1748046875,
0.26513671875
],
[
0.1748046875,
0.26513671875,
0.646484375,
0.26513671875,
0.646484375,
0.279052734375,
0.1748046875,
0.279052734375
],
[
0.1748046875,
0.285400390625,
0.363525390625,
0.285400390625,
0.363525390625,
0.298583984375,
0.1748046875,
0.298583984375
],
[
0.1748046875,
0.2998046875,
0.51708984375,
0.2998046875,
0.51708984375,
0.313232421875,
0.1748046875,
0.313232421875
],
[
0.1748046875,
0.3125,
0.82275390625,
0.3125,
0.82275390625,
0.326904296875,
0.1748046875,
0.326904296875
],
[
0.1748046875,
0.326904296875,
0.826171875,
0.326904296875,
0.826171875,
0.34033203125,
0.1748046875,
0.34033203125
],
[
0.1748046875,
0.341552734375,
0.69140625,
0.341552734375,
0.69140625,
0.354736328125,
0.1748046875,
0.354736328125
],
[
0.1748046875,
0.354736328125,
0.82421875,
0.354736328125,
0.82421875,
0.36865234375,
0.1748046875,
0.36865234375
],
[
0.1748046875,
0.36865234375,
0.82421875,
0.36865234375,
0.82421875,
0.3818359375,
0.1748046875,
0.3818359375
],
[
0.17236328125,
0.3818359375,
0.62744140625,
0.3818359375,
0.62744140625,
0.395751953125,
0.17236328125,
0.395751953125
],
[
0.1732177734375,
0.417236328125,
0.19775390625,
0.417236328125,
0.19775390625,
0.429931640625,
0.1732177734375,
0.429931640625
],
[
0.1968994140625,
0.416748046875,
0.3095703125,
0.416748046875,
0.3095703125,
0.43115234375,
0.1968994140625,
0.43115234375
],
[
0.1732177734375,
0.447509765625,
0.488525390625,
0.447509765625,
0.488525390625,
0.4609375,
0.1732177734375,
0.4609375
],
[
0.1732177734375,
0.472900390625,
0.82275390625,
0.473388671875,
0.82275390625,
0.488037109375,
0.1732177734375,
0.4873046875
],
[
0.1748046875,
0.488037109375,
0.82275390625,
0.488037109375,
0.82275390625,
0.501953125,
0.1748046875,
0.501953125
],
[
0.1732177734375,
0.501953125,
0.826171875,
0.501953125,
0.826171875,
0.51513671875,
0.1732177734375,
0.51513671875
],
[
0.17236328125,
0.5146484375,
0.82275390625,
0.51513671875,
0.82275390625,
0.52880859375,
0.17236328125,
0.52783203125
],
[
0.1748046875,
0.52978515625,
0.8203125,
0.52978515625,
0.8203125,
0.54052734375,
0.1748046875,
0.54052734375
],
[
0.1732177734375,
0.54248046875,
0.3505859375,
0.54248046875,
0.3505859375,
0.55615234375,
0.1732177734375,
0.55615234375
],
[
0.1748046875,
0.56298828125,
0.82275390625,
0.56298828125,
0.82275390625,
0.57763671875,
0.1748046875,
0.57763671875
],
[
0.1748046875,
0.5791015625,
0.82421875,
0.5791015625,
0.82421875,
0.59326171875,
0.1748046875,
0.59326171875
],
[
0.1732177734375,
0.59228515625,
0.826171875,
0.59130859375,
0.826171875,
0.60595703125,
0.1732177734375,
0.60693359375
],
[
0.1748046875,
0.60693359375,
0.7998046875,
0.60693359375,
0.7998046875,
0.62060546875,
0.1748046875,
0.62060546875
],
[
0.4150390625,
0.62548828125,
0.5810546875,
0.626953125,
0.5810546875,
0.6416015625,
0.4150390625,
0.64013671875
],
[
0.8056640625,
0.62744140625,
0.826171875,
0.62744140625,
0.826171875,
0.6416015625,
0.8056640625,
0.6416015625
],
[
0.1732177734375,
0.6484375,
0.82421875,
0.6484375,
0.82421875,
0.662109375,
0.1732177734375,
0.662109375
],
[
0.176513671875,
0.6640625,
0.82177734375,
0.6640625,
0.82177734375,
0.6748046875,
0.176513671875,
0.6748046875
],
[
0.1732177734375,
0.67529296875,
0.82177734375,
0.67529296875,
0.82177734375,
0.689453125,
0.1732177734375,
0.689453125
],
[
0.1748046875,
0.68994140625,
0.82421875,
0.68994140625,
0.82421875,
0.7041015625,
0.1748046875,
0.7041015625
],
[
0.1748046875,
0.705078125,
0.82275390625,
0.705078125,
0.82275390625,
0.71826171875,
0.1748046875,
0.71826171875
],
[
0.1748046875,
0.71826171875,
0.5107421875,
0.71826171875,
0.5107421875,
0.73193359375,
0.1748046875,
0.73193359375
],
[
0.1748046875,
0.73779296875,
0.82421875,
0.73779296875,
0.82421875,
0.75244140625,
0.1748046875,
0.75244140625
],
[
0.1732177734375,
0.75244140625,
0.29736328125,
0.75244140625,
0.29736328125,
0.7666015625,
0.1732177734375,
0.7666015625
],
[
0.348876953125,
0.76513671875,
0.65185546875,
0.76513671875,
0.65185546875,
0.77978515625,
0.348876953125,
0.77978515625
],
[
0.8056640625,
0.7666015625,
0.82421875,
0.7666015625,
0.82421875,
0.77978515625,
0.8056640625,
0.77978515625
],
[
0.1732177734375,
0.78271484375,
0.82275390625,
0.78271484375,
0.82275390625,
0.79736328125,
0.1732177734375,
0.79736328125
],
[
0.1748046875,
0.798828125,
0.82421875,
0.798828125,
0.82421875,
0.8125,
0.1748046875,
0.8125
],
[
0.1732177734375,
0.8125,
0.72216796875,
0.81103515625,
0.72216796875,
0.82568359375,
0.1732177734375,
0.8271484375
],
[
0.384033203125,
0.84130859375,
0.650390625,
0.8388671875,
0.650390625,
0.85693359375,
0.384033203125,
0.859375
],
[
0.8037109375,
0.84130859375,
0.826171875,
0.84130859375,
0.826171875,
0.85595703125,
0.8037109375,
0.85595703125
],
[
0.1748046875,
0.87060546875,
0.82275390625,
0.87060546875,
0.82275390625,
0.8837890625,
0.1748046875,
0.8837890625
],
[
0.1748046875,
0.88525390625,
0.82275390625,
0.88525390625,
0.82275390625,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1707763671875,
0.896484375,
0.82666015625,
0.896484375,
0.82666015625,
0.91748046875,
0.1707763671875,
0.91748046875
],
[
0.490234375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"classifier: instead of providing only the noisy image to the classifier, we give both the noisy image",
"andalsoitscorrespondingdenoised imageas simultaneous inputs totheclassifier; werefertothis as",
"adenoising-assisted(DA)classifier.TotraintheDA-classifier,wereusethe samedenoisingmodels",
"used to train the diffusion model. Our experiments show that this simple modification results in",
"improved generalization to unseen examples. We also qualitatively analyse the gradients and observe",
"that the gradients of DA-classifiers are improved in terms of perceptual alignment.",
"Finally, we propose a framework for semi-supervised training of classifier-guided conditional score-",
"models and evaluatethebenefits ofDA-classifiers inlabel-limited settings:morespecifically,the",
"training data consists of a largely unlabelled dataset, along with a relatively small number of labelled",
"examples. We consider the following semi-supervised learning settings: (a) classical multi-class",
"semi-supervised learning wherein the labeled dataset consists of a few examples from each class, and",
"(b) binary positive-unlabeled learning in which a few examples from the positive class are labeled",
"and the unlabeled examples are a mix of positive and negative examples",
"Contributions.In summary",
"(1)We introduce denoising-assisted (DA)classifiers",
"(2) We analyze DA-classifiers empirically with Imagenet and CIFAR10 datasets, showing improve",
"ments over noisy diffusion classifiers including: (a) improved generalization to unseen examples; (b)",
"qualitatively improved classifier-gradients; and (c) improved image generation.",
"(3) We introduce a semi-supervised framework for training diffusion classifiers, inspired by state-of-",
"the-art semi-supervised methods, wherein we train classifiers with (a) a novel temporal-consistency",
"loss and (b) show that denoising-assisted classifiers generalize better.",
"2",
" Background",
"2.1 Unconditional Score-based SDE models",
"Score models are probabilistic models of the data that enable sampling and exact inference of log",
"likelihoods. Song et al. [2021] propose a framework generalizing Multi-scale score matching Song",
"and Ermon [2019, 2020] and Denoising Diffusion Probabilistic Models Ho et al. [2020]. Concretely,",
"the framework consists oftwo components:1)theforward-diffusion (i.e.,data to noise)stochastic",
"process,and 2)alearnable score-functionthat can thenbe usedforthe reverse-diffusion(i.e..noise",
"to data) stochastic process.",
"The forward diffusion stochastic process (x(t))tejo,Ti starts at data, x(O), and ends at noise, x(T)",
"We let pt(x) denote the probability density of x at time t, so, e.g., po(x) is the distribution of the data",
"and p(x) is the distribution of the noise. The diffusion is structured so that pT(x) is independent of",
"the starting point at t = O. This process is defined with a stochastic-differential-equation (SDE):",
"dx = f(x,t)dt +g(t)dw;",
"(1)",
"where w denotes a standard Wiener process, f(x(t),t) is a drift coefficient, and g(t) is a diffusion",
"coefficient.The drift and diffusion coefficients areusuallymanually specified withoutlearnable",
"parameters; this lets us obtain closed-form solutions to the forward-diffusion SDE. For example, if f",
"is linear in x, the solution to the SDE is a time-varying gaussian distribution whose mean μ(t) and",
"standard deviation o(t) can be exactly computed. We use pt(x xo) to denote the probability density",
"function of x(t) when the diffusion is seeded at Xo.",
"To sample from po(x) starting with samples from pr(x), we have to solve the reverse diffusion SDE",
"[Anderson, 1982]:",
"dx = [f(x, t) - g(t)\"Vx log pt(x) dt + g(t) dw,",
"(2)",
"where dw is a standard Wiener process when time flows from T to O, and dt is an infinitesimal",
"negative timestep. In practice, the score function Vx log pt(x) is estimated by a neural network",
"Se(x, t), parameterized by 0, trained to optimize the following score-matching loss:",
"Ex~p:(x)[△(t)/x log pt(x) - se(x, )/]dt",
"(3)",
"where ^(t) is a positive real number introduced to balance out the score-matching objective across",
"various time steps. Using samples from the training dataset, we can define an empirical density",
"function for t = 0 as po(x) = Zx;epdata d(x -xi) and then obtain samples from pt(x) by first",
"2"
] |
{
"arxiv_id": "2306.09192",
"language": "en",
"page_idx": 1
}
|
||
2306.09192-en-2
|
2306.09192-en-2.jpg
|
sampling $\mathbf{x}(0)\sim p_0$ and then solving the forward-diffusion SDE (Eq. 1). If the solution to the SDE is a Gaussian distribution whose mean $\mu(t)$ and covariance matrix $\Sigma(t)$ can be determined in a closed-form, we can empirically define $p_t(\mathbf{x})$ as a mixture of $N$ gaussians; for such SDE's, we can also estimate the score-function $\nabla_\mathbf{x}\log p_{t}(\mathbf{x})$ in the closed-form for evaluating the score-matching loss: this is usually referred to as denoising score matching as the score-function points in the denoising direction-i.e., towards $\mu(t).$
2.2 Class-Conditional Score-based SDE models
Given a data distribution whose samples can be classified into $C$ classes, class-conditional scoremodels are trained to estimate $\nabla_\mathbf{x}\log p(\mathbf{x},t|y)$ where $y\in[1,C]$ is the class label. Classifier-free conditional models directly learn $s_\theta(\mathbf{x},t|y)$ by taking $y$ as an additional input. On the other hand, $classifter-guided$ conditional models learn the probability distribution $p(y|\mathbf{x},t)$ using a classifier and then combine this with the learnt unconditional score (i.e., $s_\theta(\mathbf{x},t))$ using Bayes rule: $p(\mathbf{x},t|y)=$ $\frac{p(y|\mathbf{x},t)p(\mathbf{x},t)}{p(y)}\colon$ applying log on both sides and taking the derivative with respect to x, we get
(4)
$$s_{\Theta}(\mathbf{x},t|y)=\nabla_{\mathbf{x}}\log p_{\phi}(y|\mathbf{x},t)+s_{\theta}(\mathbf{x},t)$$
where $\phi$ denotes the parameters of the classifier, so the full model is parameterized by $\Theta=\{\theta,\phi\}.$ In this work, we use classifier-guided score models as this allows us to reuse pre-trained larger score-models while training smaller classifier-models: for example, Song et al.[2021] generate classconditional CIFAR10 examples using a score-model having 107M parameters and a classifier-model having 1.5M parameters. Classifier-guidance also allows us to flexibly incorporate new definitions of classes without requiring to retrain the larger score-model.
Song et al.[2021] suggest a simple sum of the cross-entropy losses over examples sampled from the
diffusion for training the classifier $p_\phi{:}$
(5)
$$\mathcal{L}_{\mathrm{CE}}=\mathbb{E}_{\begin{array}{c}t\sim\mathcal{U}(0,T)\\(\mathbf{x}_{0},y)\sim p_{0}(\mathbf{x})\\\mathbf{x}\sim p_{t}(\mathbf{x}|\mathbf{x}_{0})\end{array}}[-\log p_{\phi}(y|\mathbf{x},t)]$$
Dhariwal and Nichol [2021a] introduce a classifier gradient scale $\lambda_s-$ in which, the classifier gradient in Eq. 4 is scaled with $\lambda_\mathrm{s}-$as a control for trading off fidelity vs diversity. Higher scale corresponds to higher fidelity and lower diversity.
## 3 Denoising-Assisted (DA) Classifier
Given samples $(\mathbf{x}_0,y)\sim p_0(\mathbf{x})$ and a forward diffusion SDE, our goal is to train a classifier on examples sampled from the forward diffusion. Suppose that the solution to the SDE at time $t-$when seeded with $\mathbf{x}(0)$ at time 0-can be specified as a Gaussian distribution with mean $\mu(\mathbf{x}(0),t)$ and diagonal covariance $\sigma^2(t)$I or in other words, $p_t( \mathbf{x} | \mathbf{x} _0) = \mathcal{N} ( \mathbf{x} \mid \mu ( \mathbf{x} ( 0) , t) $, $\sigma ^2( t)$I). For training the DA-classifier, rather than use the typical (noisy) input, we propose to use as input both $\mathbf{x}\sim p_{t}(\tilde{\mathbf{x}}|\mathbf{x}_{0})$ and the denoised image $\hat{\mathbf{x}}:$
(6)
$$\mathcal{L}_{\mathrm{CE}}=\mathbb{E}_{\begin{array}{c}t\sim\mathcal{U}(0,T)\\(\mathbf{x}_{0},y)\sim p_{0}(\mathbf{x})\\\mathbf{x}\sim p_{t}(\mathbf{x}|\mathbf{x}_{0})\end{array}}[-\log p_{\phi}(y|\mathbf{x},\hat{\mathbf{x}},t)]$$
We compute $\hat{\mathbf{x}}$ using a pretrained score network $s_\theta$ as $\hat{\mathbf{x} } = \mathbf{x} + \sigma ^2( t) s_\theta ( \mathbf{x} , t) . \textbf{More specifically, the}$ denoised image refers to the estimated mean of the Gaussian that x was sampled from. For example, the mean does not change with diffusion time $t$ in Variance-exploding SDEs while the mean decays to zero with diffusion time for DDPMs. In the following, we evaluate—both quantitatively and qualitatively—the improvements obtained by using $\hat{\mathbf{x}}$ as an additional input to the classifier.
Experiment setup. We conduct our experiments on CIFAR10 and 256×256 Imagenet in the VEDiffusion[Song et al., 2021] and Improved-DDPM [Nichol and Dhariwal, 2021, Dhariwal and Nichol, 2021b] settings respectively. We use pretrained score-networks for these experiments: specifically, we use the deep NCSN++ (continuous) model released by Song et al.[2021] as the score-network for CIFAR10 and the unconditional model open-sourced by Dhariwal and Nichol [2021b] as the score-network for Imagenet. We use the pretrained noisy Imagenet classifier released by Dhariwal and Nichol [2021b] while we trained the noisy CIFAR10 classifier ourselves; the Imagenet classifier is the downsampling half of the UNET with attention pooling classifier-head while we use WideResNet- 28-2 as the architecture for CIFAR10. For the DA-classifier, we simply add an extra convolution that
3
|
[
[
0.1732177734375,
0.0921630859375,
0.82275390625,
0.091552734375,
0.82275390625,
0.1060791015625,
0.1732177734375,
0.106689453125
],
[
0.1732177734375,
0.1060791015625,
0.82421875,
0.106689453125,
0.82421875,
0.12060546875,
0.1732177734375,
0.11932373046875
],
[
0.176513671875,
0.12249755859375,
0.82177734375,
0.12249755859375,
0.82177734375,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.1748046875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.1483154296875,
0.1748046875,
0.1483154296875
],
[
0.1732177734375,
0.146484375,
0.82421875,
0.147705078125,
0.82421875,
0.1622314453125,
0.1732177734375,
0.1610107421875
],
[
0.1732177734375,
0.1610107421875,
0.435546875,
0.1622314453125,
0.435546875,
0.1767578125,
0.1732177734375,
0.175537109375
],
[
0.1748046875,
0.1925048828125,
0.5146484375,
0.1925048828125,
0.5146484375,
0.205810546875,
0.1748046875,
0.205810546875
],
[
0.1748046875,
0.2171630859375,
0.826171875,
0.2171630859375,
0.826171875,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.1732177734375,
0.2310791015625,
0.82275390625,
0.2298583984375,
0.82275390625,
0.2442626953125,
0.1732177734375,
0.24560546875
],
[
0.1748046875,
0.24560546875,
0.826171875,
0.24560546875,
0.826171875,
0.2587890625,
0.1748046875,
0.2587890625
],
[
0.1748046875,
0.2587890625,
0.82421875,
0.2587890625,
0.82421875,
0.2734375,
0.1748046875,
0.2734375
],
[
0.1732177734375,
0.271484375,
0.8203125,
0.2734375,
0.8203125,
0.287353515625,
0.1732177734375,
0.285400390625
],
[
0.17236328125,
0.2822265625,
0.77392578125,
0.287353515625,
0.77392578125,
0.3125,
0.17236328125,
0.306884765625
],
[
0.200927734375,
0.29736328125,
0.2303466796875,
0.29736328125,
0.2303466796875,
0.309326171875,
0.200927734375,
0.309326171875
],
[
0.371826171875,
0.313232421875,
0.6259765625,
0.313232421875,
0.6259765625,
0.32763671875,
0.371826171875,
0.32763671875
],
[
0.8037109375,
0.3125,
0.826171875,
0.3125,
0.826171875,
0.326904296875,
0.8037109375,
0.326904296875
],
[
0.1748046875,
0.333251953125,
0.826171875,
0.333251953125,
0.826171875,
0.3466796875,
0.1748046875,
0.3466796875
],
[
0.1748046875,
0.3466796875,
0.82275390625,
0.3466796875,
0.82275390625,
0.35986328125,
0.1748046875,
0.35986328125
],
[
0.1732177734375,
0.35986328125,
0.82275390625,
0.35986328125,
0.82275390625,
0.373779296875,
0.1732177734375,
0.373779296875
],
[
0.1748046875,
0.375,
0.82421875,
0.375,
0.82421875,
0.38818359375,
0.1748046875,
0.38818359375
],
[
0.1748046875,
0.38818359375,
0.82421875,
0.38818359375,
0.82421875,
0.401611328125,
0.1748046875,
0.401611328125
],
[
0.1748046875,
0.401611328125,
0.55810546875,
0.401611328125,
0.55810546875,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.1748046875,
0.42236328125,
0.82421875,
0.42236328125,
0.82421875,
0.436279296875,
0.1748046875,
0.436279296875
],
[
0.1732177734375,
0.43505859375,
0.42724609375,
0.4375,
0.42724609375,
0.451904296875,
0.1732177734375,
0.449462890625
],
[
0.3798828125,
0.453857421875,
0.619140625,
0.453857421875,
0.619140625,
0.4716796875,
0.3798828125,
0.4716796875
],
[
0.80224609375,
0.453857421875,
0.82666015625,
0.453857421875,
0.82666015625,
0.4716796875,
0.80224609375,
0.4716796875
],
[
0.438720703125,
0.4697265625,
0.50732421875,
0.4697265625,
0.50732421875,
0.48095703125,
0.438720703125,
0.48095703125
],
[
0.44287109375,
0.479248046875,
0.50634765625,
0.479248046875,
0.50634765625,
0.4892578125,
0.44287109375,
0.4892578125
],
[
0.1748046875,
0.492431640625,
0.82177734375,
0.492431640625,
0.82177734375,
0.5068359375,
0.1748046875,
0.5068359375
],
[
0.1748046875,
0.50830078125,
0.406982421875,
0.50830078125,
0.406982421875,
0.521484375,
0.1748046875,
0.521484375
],
[
0.4150390625,
0.50830078125,
0.82275390625,
0.50830078125,
0.82275390625,
0.521484375,
0.4150390625,
0.521484375
],
[
0.1748046875,
0.521484375,
0.50244140625,
0.521484375,
0.50244140625,
0.53515625,
0.1748046875,
0.53515625
],
[
0.1732177734375,
0.5546875,
0.490234375,
0.5546875,
0.490234375,
0.5693359375,
0.1732177734375,
0.5693359375
],
[
0.1748046875,
0.58544921875,
0.82275390625,
0.58544921875,
0.82275390625,
0.599609375,
0.1748046875,
0.599609375
],
[
0.1748046875,
0.599609375,
0.76953125,
0.599609375,
0.76953125,
0.61279296875,
0.1748046875,
0.61279296875
],
[
0.78369140625,
0.60107421875,
0.8203125,
0.60107421875,
0.8203125,
0.6123046875,
0.78369140625,
0.6123046875
],
[
0.1748046875,
0.61279296875,
0.373291015625,
0.61279296875,
0.373291015625,
0.626953125,
0.1748046875,
0.626953125
],
[
0.365234375,
0.61279296875,
0.82421875,
0.61279296875,
0.82421875,
0.626953125,
0.365234375,
0.626953125
],
[
0.1748046875,
0.626953125,
0.82275390625,
0.626953125,
0.82275390625,
0.64013671875,
0.1748046875,
0.64013671875
],
[
0.1732177734375,
0.6396484375,
0.82421875,
0.64013671875,
0.82421875,
0.65478515625,
0.1732177734375,
0.65380859375
],
[
0.1748046875,
0.65478515625,
0.35302734375,
0.65478515625,
0.35302734375,
0.66845703125,
0.1748046875,
0.66845703125
],
[
0.371826171875,
0.67236328125,
0.6259765625,
0.671875,
0.6259765625,
0.689453125,
0.371826171875,
0.68994140625
],
[
0.8037109375,
0.673828125,
0.826171875,
0.673828125,
0.826171875,
0.68798828125,
0.8037109375,
0.68798828125
],
[
0.431396484375,
0.6875,
0.50244140625,
0.6875,
0.50244140625,
0.69775390625,
0.431396484375,
0.69775390625
],
[
0.439453125,
0.69580078125,
0.498291015625,
0.69580078125,
0.498291015625,
0.70703125,
0.439453125,
0.70703125
],
[
0.1748046875,
0.7119140625,
0.82275390625,
0.7119140625,
0.82275390625,
0.72607421875,
0.1748046875,
0.72607421875
],
[
0.1748046875,
0.7265625,
0.82421875,
0.7265625,
0.82421875,
0.74072265625,
0.1748046875,
0.74072265625
],
[
0.1748046875,
0.74072265625,
0.82275390625,
0.74072265625,
0.82275390625,
0.75390625,
0.1748046875,
0.75390625
],
[
0.1748046875,
0.75390625,
0.65625,
0.75390625,
0.65625,
0.767578125,
0.1748046875,
0.767578125
],
[
0.65625,
0.7587890625,
0.66259765625,
0.7587890625,
0.66259765625,
0.76220703125,
0.65625,
0.76220703125
],
[
0.66259765625,
0.7548828125,
0.82275390625,
0.7548828125,
0.82275390625,
0.76513671875,
0.66259765625,
0.76513671875
],
[
0.1748046875,
0.76806640625,
0.77392578125,
0.76806640625,
0.77392578125,
0.7822265625,
0.1748046875,
0.7822265625
],
[
0.176513671875,
0.78857421875,
0.826171875,
0.78857421875,
0.826171875,
0.8017578125,
0.176513671875,
0.8017578125
],
[
0.1748046875,
0.8017578125,
0.826171875,
0.8017578125,
0.826171875,
0.81494140625,
0.1748046875,
0.81494140625
],
[
0.1748046875,
0.81640625,
0.826171875,
0.81640625,
0.826171875,
0.830078125,
0.1748046875,
0.830078125
],
[
0.1748046875,
0.830078125,
0.82421875,
0.830078125,
0.82421875,
0.84326171875,
0.1748046875,
0.84326171875
],
[
0.1748046875,
0.84326171875,
0.82275390625,
0.84326171875,
0.82275390625,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.1748046875,
0.85791015625,
0.82421875,
0.85791015625,
0.82421875,
0.8720703125,
0.1748046875,
0.8720703125
],
[
0.1732177734375,
0.869140625,
0.82421875,
0.87060546875,
0.82421875,
0.8837890625,
0.1732177734375,
0.88330078125
],
[
0.1732177734375,
0.8837890625,
0.826171875,
0.88330078125,
0.826171875,
0.89794921875,
0.1732177734375,
0.8984375
],
[
0.1732177734375,
0.89794921875,
0.826171875,
0.8984375,
0.826171875,
0.912109375,
0.1732177734375,
0.9111328125
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"sampling x(O) ~ po and then solving the forward-diffusion SDE (Eg 1). If the solution to the",
"SDEis a Gaussian distribution whose mean μ(t)and covariance matrix (t) can bedetermined in a",
"closed-form,wecan empiricallydefinept(x)as amixture of N gaussians;for suchSDE's,wecan",
"also estimate the score-function Vx log pt(x) in the closed-form for evaluating the score-matching",
"loss: this is usually referred to as denoising score matching as the score-function points in the",
"denoising direction -i.e., towards μ(t).",
"2.2Class-Conditional Score-based SDE models",
"Given a data distribution whose samples can be classified into C classes, class-conditional score-",
"models are trained to estimate Vx log p(x, tly) where y e [1, C] is the class label. Classifier-free",
"conditional models directly learn se(x, tly) by taking y as an additional input. On the other hand,",
"classifier-guided conditional models learn the probability distribution p(yx, t) using a classifier and",
"then combine this with the learnt unconditional score (i.e., se(x,t)) using Bayes rule: p(x, tly) =",
"pplynggonbothsesandtakinghdrivatvwithrepttw",
"p(y)",
"se(x,tly) = Vxlogpe(yx,t) + se(x,t)",
"(4)",
"where @ denotes the parameters of the classifier, so the full model is parameterized by = [0, ].",
"In this work, we use classifier-guided score models as this allows us to reuse pre-trained larger",
"score-models while training smaller classifier-models: for example, Song et al. [2021] generate class",
"conditional CIFAR10 examples using a score-model having 107M parameters and a classifier-model",
"having 1.5M parameters. Classifier-guidance also allows us to flexibly incorporate new definitions of",
"classes without requiring to retrain the larger score-model.",
"Song et al. [2021] suggest a simple sum of the cross-entropy losses over examples sampled from the",
"diffusion for training the classifier P:",
"CcE = E t~u(o,T) [-log Pa(ylx,t)]",
"(5)",
"(xo,y)~po(x",
"x~pt(x|xo)",
"Dhariwal and Nichol [2021a] introduce a classifier gradient scale As — in which, the classifie",
"gradient in Eq.4 is scaled with ^s",
"as acontrol fortradingoff fidelityvsdiversity.Higherscale",
"corresponds to higher fidelity and lower diversity.",
"3Denoising-Assisted (DA) Classifier",
"Given samples (xo, y) ~ po(x) and a forward diffusion SDE, our goal is to train a classifier on",
"examples sampled from the forward diffusion. Suppose that the solution to the SDE at time t",
"wher",
"seeded with x(0) at time 0 -",
"- can be specified as a Gaussian distribution with mean μ(x(O), t) and",
"diagonal covariance 2(t)I or in other words, Pt(xxo) = N(x μ(x(O), t), α2(t)1). For training the",
"DA-classifier, rather than use the typical (noisy) input, we propose to use as input both x ~ pt(x|xo)",
"and the denoised image x:",
"LcE = E,t~u(o,T)[-logPs(y|x,x,t)]",
"(6)",
"(xo,y)~po(x)",
"<~pt(x|x0)",
"We compute x using a pretrained score network sg as x = x + g2(t)se(x, t). More specifically, the",
"denoised image refers to the estimated mean of the Gaussian that x was sampled from. For example,",
"the mean does not change with diffusion time t in Variance-exploding SDEs while the mean decays",
"to zero with diffusion time for DDPMs.In the following, we evaluate-",
"",
"-bothquantitativelyand",
"qualitatively-the improvements obtained by using x as an additional input to the classifier.",
"Experiment setup. We conduct our experiments on CIFAR10 and 256×256 Imagenet in the VE-",
"Diffusion [Song et al., 2021] and Improved-DDPM [Nichol and Dhariwal, 2021, Dhariwal and Nichol,",
"2021b] settings respectively. We use pretrained score-networks for these experiments: specifically,",
"we use the deep NCSN++ (continuous) model released by Song et al. [2021] as the score-network",
"for CIFAR10 and the unconditional model open-sourced by Dhariwal and Nichol [2021b] as the",
"score-network for Imagenet. We use the pretrained noisy Imagenet classifier released by Dhariwal and",
"Nichol [2021b] while we trained the noisy CIFAR10classifier ourselves; the Imagenet classifier is",
"the downsampling half of the UNET with attention pooling classifier-head while we use WideResNet-",
"28-2as the architectureforCIFAR10.For theDA-classifier,we simplyadd an extraconvolution that",
"3"
] |
{
"arxiv_id": "2306.09192",
"language": "en",
"page_idx": 2
}
|
||
2306.09192-en-3
|
2306.09192-en-3.jpg
|
Table 1: Summary of Test Accuracies for CIFAR10 and Imagenet: each test example is diffused to a random uniformly sampled diffusion time. Both classifiers are shown the same diffused example.
<table>
<tbody>
<tr>
<th rowspan="2">Dataset Method</th>
<th rowspan="2">CIFAR10</th>
<th colspan="2">Imagenet</th>
</tr>
<tr>
<th>$Top-1$</th>
<th>$Top-5$</th>
</tr>
<tr>
<td>Noisy Classifier</td>
<td>54.79</td>
<td>33.78</td>
<td>49.86</td>
</tr>
<tr>
<td>Denoising Assisted Classifier ( Ours) 1</td>
<td>57.16</td>
<td>36.11</td>
<td>52.34</td>
</tr>
</tbody>
</table>
<FigureHere>
<FigureHere>
Figure 2: Min-max normalized gradients on samples ${\mathrm{diffused~to~}t=300~(T=999):~\mathrm{~The~top~panel~shows}}$ the gradients obtained with noisy classifier while the bottom panel shows the gradients obtained with the DAclassifier. The middle panel shows the clean Imagenet samples considered. We recommend zooming in to
Figure 1: CIFAR10: Test Accuracy vs. Noise Scale. see clearer differences between gradients. This is also
shown in Figure 4.
can process the denoised input: for Imagenet, we finetune the pretrained noisy classifier by adding an additional input-convolution module while we train the denoising-assisted CIFAR 10 classiffer from scratch. The details of the optimization are as follows: (1) for Imagenet, we fine-tune the entire network along with the new convolution-module (initialized with very small weights) using AdamW optimizer with a learning-rate of le-5 and a weight-decay of 0.05 for 50k steps with a batch size of 128. (2) For CIFAR10, we train both noisy and DA-classifiers for 150k steps with a batch size of 512 using AdamW optimizer with a learning-rate of 3e-4 and weight decay of 0.05. For CIFAR10 classifiers, we use the Exponential Moving Average of the parameters with decay-rate equal to 0.999. Classification Accuracy. We first compare diffusion classifiers in terms of their test accuracies as a measure of their generalization. We compute accuracy aggregated over all diffusion times by sampling examples from the forward diffusion- seeded with test examples- ensuring that classiffers are evaluated on the same set of diffused examples. We summarize accuracies for CIFAR10 and Imagenet in Table l and plot accuracy as a function of noise-scale for CIFAR10 in Figure 1. We find that DA-classifiers generalize better to the test data; this is particularly interesting as randomperturbation training is often used to improve generalization — for example, Chapelle et al.[2000] $\hat{\text{discuss Vicinal Risk Minimization for improved generalization by consiđering Gaussian Vicinities.}}$ We hypothesize that denoised images can be viewed as augmentations of the original image that are relatively easier to classify as compared to their noisy counterparts while being diverse enough to improve generalization. Alternatively, introducing the denoising step before classifying the noisy example can be seen as an inductive bias for classifying noisy examples-on its own, cross-entropy loss does not encourage the classifier to denoise before assigning a label. To determine the relative importance of noisy and denoised examples, we zeroed out one of the input images and measured the classification accuracies: while removing either one of the inputs causes the accuracy to drop below the noisy classifier, the drop is higher when the denoised image is masked out.
Classifier Gradients. The diffusion classifiers participate in the sampling through their gradients and we qualitatively compare the classifiers in terms of their class-likelihood gradients $\nabla_\mathbf{x}\log p_\phi(y|\mathbf{x},t)$ -for the DA-classifier, we use the total derivative of $\log p_\phi(y|\mathbf{x},\hat{\mathbf{x}},t)$ with respect to x. We show the
4
|
[
[
0.1748046875,
0.12060546875,
0.5009765625,
0.12060546875,
0.5009765625,
0.1337890625,
0.1748046875,
0.1337890625
],
[
0.1748046875,
0.1337890625,
0.5009765625,
0.1337890625,
0.5009765625,
0.147705078125,
0.1748046875,
0.147705078125
],
[
0.1732177734375,
0.1458740234375,
0.5009765625,
0.1458740234375,
0.5009765625,
0.1590576171875,
0.1732177734375,
0.1590576171875
],
[
0.1748046875,
0.1610107421875,
0.375732421875,
0.1610107421875,
0.375732421875,
0.1717529296875,
0.1748046875,
0.1717529296875
],
[
0.253173828125,
0.186279296875,
0.301513671875,
0.186279296875,
0.301513671875,
0.1976318359375,
0.253173828125,
0.1976318359375
],
[
0.316162109375,
0.1900634765625,
0.378173828125,
0.1900634765625,
0.378173828125,
0.2052001953125,
0.316162109375,
0.2052001953125
],
[
0.41015625,
0.1849365234375,
0.468017578125,
0.1849365234375,
0.468017578125,
0.199462890625,
0.41015625,
0.199462890625
],
[
0.1854248046875,
0.1962890625,
0.23779296875,
0.1962890625,
0.23779296875,
0.2108154296875,
0.1854248046875,
0.2108154296875
],
[
0.39208984375,
0.203857421875,
0.43310546875,
0.203857421875,
0.43310546875,
0.2183837890625,
0.39208984375,
0.2183837890625
],
[
0.44677734375,
0.203857421875,
0.48779296875,
0.203857421875,
0.48779296875,
0.2183837890625,
0.44677734375,
0.2183837890625
],
[
0.328369140625,
0.222900390625,
0.366943359375,
0.222900390625,
0.366943359375,
0.238037109375,
0.328369140625,
0.238037109375
],
[
0.39208984375,
0.222900390625,
0.431396484375,
0.222900390625,
0.431396484375,
0.238037109375,
0.39208984375,
0.238037109375
],
[
0.451904296875,
0.222900390625,
0.49267578125,
0.222900390625,
0.49267578125,
0.238037109375,
0.451904296875,
0.238037109375
],
[
0.178955078125,
0.236083984375,
0.298095703125,
0.236083984375,
0.298095703125,
0.25,
0.178955078125,
0.25
],
[
0.191162109375,
0.2247314453125,
0.285888671875,
0.2247314453125,
0.285888671875,
0.236083984375,
0.191162109375,
0.236083984375
],
[
0.326904296875,
0.242431640625,
0.36767578125,
0.242431640625,
0.36767578125,
0.2568359375,
0.326904296875,
0.2568359375
],
[
0.393798828125,
0.24365234375,
0.431396484375,
0.24365234375,
0.431396484375,
0.2568359375,
0.393798828125,
0.2568359375
],
[
0.449462890625,
0.242431640625,
0.490234375,
0.240478515625,
0.491943359375,
0.255615234375,
0.450927734375,
0.25830078125
],
[
0.188720703125,
0.250732421875,
0.28759765625,
0.250732421875,
0.28759765625,
0.261474609375,
0.188720703125,
0.261474609375
],
[
0.24267578125,
0.27978515625,
0.43701171875,
0.27978515625,
0.43701171875,
0.287353515625,
0.24267578125,
0.287353515625
],
[
0.200927734375,
0.29736328125,
0.2132568359375,
0.29736328125,
0.2132568359375,
0.306884765625,
0.200927734375,
0.306884765625
],
[
0.358642578125,
0.291748046875,
0.459228515625,
0.291748046875,
0.459228515625,
0.3017578125,
0.358642578125,
0.3017578125
],
[
0.361083984375,
0.300537109375,
0.414306640625,
0.300537109375,
0.414306640625,
0.30810546875,
0.361083984375,
0.30810546875
],
[
0.19775390625,
0.31005859375,
0.2156982421875,
0.31005859375,
0.2156982421875,
0.324462890625,
0.19775390625,
0.324462890625
],
[
0.1993408203125,
0.32763671875,
0.2132568359375,
0.32763671875,
0.2132568359375,
0.339111328125,
0.1993408203125,
0.339111328125
],
[
0.1854248046875,
0.34033203125,
0.2181396484375,
0.34033203125,
0.2181396484375,
0.37744140625,
0.1854248046875,
0.37744140625
],
[
0.2017822265625,
0.3818359375,
0.2115478515625,
0.3818359375,
0.2115478515625,
0.385009765625,
0.2017822265625,
0.385009765625
],
[
0.50341796875,
0.3818359375,
0.82421875,
0.38330078125,
0.82421875,
0.396484375,
0.50341796875,
0.395751953125
],
[
0.1993408203125,
0.392578125,
0.2132568359375,
0.392578125,
0.2132568359375,
0.404052734375,
0.1993408203125,
0.404052734375
],
[
0.5048828125,
0.397705078125,
0.82421875,
0.397705078125,
0.82421875,
0.40771484375,
0.5048828125,
0.40771484375
],
[
0.19775390625,
0.4072265625,
0.214111328125,
0.4072265625,
0.214111328125,
0.421630859375,
0.19775390625,
0.421630859375
],
[
0.50341796875,
0.4091796875,
0.826171875,
0.4091796875,
0.826171875,
0.42236328125,
0.50341796875,
0.42236328125
],
[
0.1993408203125,
0.4248046875,
0.2132568359375,
0.4248046875,
0.2132568359375,
0.434326171875,
0.1993408203125,
0.434326171875
],
[
0.50341796875,
0.421630859375,
0.826171875,
0.421630859375,
0.826171875,
0.43505859375,
0.50341796875,
0.43505859375
],
[
0.50244140625,
0.434326171875,
0.82275390625,
0.434326171875,
0.82275390625,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.50341796875,
0.4462890625,
0.826171875,
0.4462890625,
0.826171875,
0.460205078125,
0.50341796875,
0.460205078125
],
[
0.1805419921875,
0.462158203125,
0.5,
0.462158203125,
0.5,
0.4765625,
0.1805419921875,
0.4765625
],
[
0.50244140625,
0.458984375,
0.826171875,
0.458984375,
0.826171875,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.50244140625,
0.4716796875,
0.61767578125,
0.4716796875,
0.61767578125,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.1748046875,
0.52587890625,
0.82421875,
0.52587890625,
0.82421875,
0.54052734375,
0.1748046875,
0.54052734375
],
[
0.1748046875,
0.54052734375,
0.82421875,
0.54052734375,
0.82421875,
0.5537109375,
0.1748046875,
0.5537109375
],
[
0.1732177734375,
0.55322265625,
0.82421875,
0.5537109375,
0.82421875,
0.5673828125,
0.1732177734375,
0.56640625
],
[
0.1732177734375,
0.56640625,
0.82177734375,
0.56640625,
0.82177734375,
0.580078125,
0.1732177734375,
0.580078125
],
[
0.176513671875,
0.58203125,
0.826171875,
0.58203125,
0.826171875,
0.59521484375,
0.176513671875,
0.59521484375
],
[
0.17724609375,
0.5966796875,
0.82421875,
0.5966796875,
0.82421875,
0.60693359375,
0.17724609375,
0.60693359375
],
[
0.1748046875,
0.609375,
0.82421875,
0.609375,
0.82421875,
0.62255859375,
0.1748046875,
0.62255859375
],
[
0.1748046875,
0.62353515625,
0.826171875,
0.62353515625,
0.826171875,
0.63720703125,
0.1748046875,
0.63720703125
],
[
0.1748046875,
0.642578125,
0.82421875,
0.642578125,
0.82421875,
0.6572265625,
0.1748046875,
0.6572265625
],
[
0.1732177734375,
0.65576171875,
0.82421875,
0.6572265625,
0.82421875,
0.671875,
0.1732177734375,
0.67041015625
],
[
0.1732177734375,
0.67041015625,
0.82275390625,
0.67041015625,
0.82275390625,
0.6845703125,
0.1732177734375,
0.6845703125
],
[
0.1748046875,
0.68505859375,
0.82421875,
0.68505859375,
0.82421875,
0.69873046875,
0.1748046875,
0.69873046875
],
[
0.1748046875,
0.69873046875,
0.82421875,
0.69873046875,
0.82421875,
0.7119140625,
0.1748046875,
0.7119140625
],
[
0.1748046875,
0.7119140625,
0.826171875,
0.7119140625,
0.826171875,
0.72607421875,
0.1748046875,
0.72607421875
],
[
0.17724609375,
0.72802734375,
0.57861328125,
0.72802734375,
0.57861328125,
0.73779296875,
0.17724609375,
0.73779296875
],
[
0.58837890625,
0.72607421875,
0.826171875,
0.72607421875,
0.826171875,
0.7392578125,
0.58837890625,
0.7392578125
],
[
0.1732177734375,
0.7392578125,
0.82421875,
0.7392578125,
0.82421875,
0.75244140625,
0.1732177734375,
0.75244140625
],
[
0.1748046875,
0.75390625,
0.82421875,
0.75390625,
0.82421875,
0.767578125,
0.1748046875,
0.767578125
],
[
0.1748046875,
0.767578125,
0.826171875,
0.767578125,
0.826171875,
0.7822265625,
0.1748046875,
0.7822265625
],
[
0.1732177734375,
0.78076171875,
0.82275390625,
0.78076171875,
0.82275390625,
0.79443359375,
0.1732177734375,
0.79443359375
],
[
0.1732177734375,
0.79443359375,
0.82421875,
0.79541015625,
0.82421875,
0.8095703125,
0.1732177734375,
0.80810546875
],
[
0.1748046875,
0.8095703125,
0.82421875,
0.8095703125,
0.82421875,
0.82275390625,
0.1748046875,
0.82275390625
],
[
0.1748046875,
0.82275390625,
0.82421875,
0.82275390625,
0.82421875,
0.8359375,
0.1748046875,
0.8359375
],
[
0.1748046875,
0.8369140625,
0.826171875,
0.8369140625,
0.826171875,
0.85107421875,
0.1748046875,
0.85107421875
],
[
0.1748046875,
0.85107421875,
0.68896484375,
0.85107421875,
0.68896484375,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.17724609375,
0.8720703125,
0.82275390625,
0.8720703125,
0.82275390625,
0.8818359375,
0.17724609375,
0.8818359375
],
[
0.1732177734375,
0.88330078125,
0.826171875,
0.8837890625,
0.826171875,
0.8984375,
0.1732177734375,
0.89794921875
],
[
0.188720703125,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.91357421875,
0.188720703125,
0.91357421875
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"Table 1: Summary of Test Accuracies for CIFAR10",
"and Imagenet: each test example is diffused to a random",
"uniformly sampled diffusion time.Both classifiers are",
"shownthe samediffusedexample",
"Dataset",
"CIFAR10",
"Imagenet",
"Method",
"Top-1",
"Top-5",
"54.79",
"33.78",
"49.86",
"Denoising Assisted",
"Noisy Classifier",
"57.16",
"36.11",
"52.34",
"Classifier(Ours)",
"NoisyClassifiervsDenoisingAssistedClassifie",
"0.9",
"Denoising Assisted Classifie",
"Noisy Classifier",
"0.8",
"0.7",
"",
".4",
"Figure 2: Min-max normalized gradients on samples",
"0.3",
"diffusedtot=300(T=999):Thetoppanel shows",
"0.2",
"the gradients obtained with noisy classifier while the",
"0.1",
"bottom panel shows the gradients obtained with the DA-",
"classifier. The middle panel shows the clean Imagenet",
"samples considered. We recommend zooming in to",
"Figure 1: CIFAR10: Test Accuracy vs. Noise Scale.",
"see clearer differences between gradients. This is also",
"shown in Figure 4.",
"can process the denoised input: for Imagenet, we finetune the pretrained noisy classifier by adding",
"an additional input-convolution module while we train the denoising-assisted CIFAR1O classifier",
"from scratch.The details of the optimization are as follows: (1)for Imagenet, we fine-tune the entire",
"network along with the new convolution-module (initialized with very small weights) using Adamw",
"optimizer with a learning-rate of le-5 and a weight-decay of 0.05 for 50k steps with a batch size of",
"128.(2)ForCIFAR10.wetrainbothnoisyandDA-classifiersfor150kstepswithabatchsizeof",
"512 using AdamW optimizer with a learning-rate of 3e-4 and weight decay of 0.05. For CIFAR10",
"classifiers, we use the Exponential Moving Average of the parameters with decay-rate equal to 0.999.",
"Classification Accuracy. We first compare diffusion classifiers in terms of their test accuracies",
"as a measure of their generalization. We compute accuracy aggregated over all diffusion times by",
"sampling examples from the forward diffusion - seeded with test examples - ensuring that classifiers",
"are evaluated on the same set of diffused examples. We summarize accuracies for CIFAR10 and",
"Imagenet in Table 1 and plot accuracy as a function of noise-scale for CIFAR10 in Figure 1. We",
"find that DA-classifiers generalize better to the test data; this is particularly interesting as random-",
"perturbationtrainingisoftenusedtoimprovegeneralization",
" for example, Chapelle et al. [2000]",
"discuss Vicinal Risk Minimization for improved generalization by considering Gaussian Vicinities.",
"We hypothesize that denoised images can be viewed as augmentations of the original image that are",
"relatively easier to classify as compared to their noisy counterparts while being diverse enough to",
"improve generalization. Alternatively, introducing the denoising step before classifying the noisy",
"example can be seen as an inductive bias for classifying noisy examples - on its own, cross-entropy",
"loss does not encourage the classifier to denoise before assigning a label. To determine the relative",
"importance of noisy and denoised examples, we zeroed out one of the input images and measured the",
"classification accuracies: while removing either one of the inputs causes the accuracy to drop below",
"the noisy classifier, the drop is higher when the denoised image is masked out.",
"ClassifierGradients.Thediffusionclassifiersparticipateinthe samplingthroughtheirgradientsand",
"we qualitatively compare the classifiers in terms of their class-likelihood gradients Vx log Po(y x,t)",
"for the DA-classifier, we use the total derivative of log Ps(yx, x, t) with respect to x. We show the",
"4"
] |
{
"arxiv_id": "2306.09192",
"language": "en",
"page_idx": 3
}
|
||
1910.03876-en-1
|
1910.03876-en-1.jpg
|
sists of two sub-networks: a denoising network and a rectification network. Motivated by the success of U-Net [31] in recovering the object details, we employ U-Net structure as an image recovery backbone network, attempting to extract visual content at structural-level details. In the denoising sub-network (DSN), we try to transform a low-quality image to a high-quality image pixel by pixel directly. The DSN can penalize the loss between noisy and noise-free image pairs and thus acquire the output image with the fine textures of the clean component, learning an independent realization of the noise. However, even with such sophisticated DSN, denoising images are unsatisfactory because they still have arbitrary geometric variations. Therefore, the rectification sub-network (RSN) is proposed to correct geometric distortions of denoising license plates and generate more accurate correction image distortion. Furthermore, we propose to leverage the new auxiliary tasks to further optimize the image recovery sub-networks (DSN, RSN) of SNIDER. There are two auxiliary tasks: a text counting module and a segment prediction module. Specifically, we solve each auxiliary module using CNN as a decoder. The counting module is used to predict the number of text in the image as a classification problem. In this module, despite the ambiguous boundary of consecutive text, text counting can distinguish single text, which makes the image quality suitable for text detection. For the segment prediction module, we propose a binary segmentation to emphasize the foreground over the background. The generated segmentation result makes the license plate clean for text recognition. Finally, learning the auxiliary tasks will lead the intermediate features of the recovery main task networks to enhance the difficulties such as geometric variations and low-quality information. More importantly, we introduce a new loss function that trains the SNIDER with auxiliary tasks, which provide significantly higher license plate quality for robust $LPR.$
To sum up, we highlight the main contributions of this
paper as follows:
$\bullet$ We propose a novel end-to-end license plate recovery
network, where denoising and rectification network are used to generate a clear recovery image for robust LPR performance.
$\bullet$ We present the auxiliary tasks to leverage the quality of
the license plate recovery from low-quality. Mainly a new loss is introduced to provide regularization effects to the backbone SNIDER for robust representation and license plate recovery.
$\bullet$ Finally, we demonstrate the effectiveness of the pro-
posed method in recovering a high-quality license plate from a low-quality license plate in the real-world and show that the LPR performance outperforms the state-of-the-art methods on two challenging datasets,
AOLP-RP[13] and VTLPs dataset newly collected on
the most challenging real-world environments.
## $2. \textbf{Related Work}$
In this section, we briefly review on low-quality image recovery methods and license plate recognition methods that is most related to this work.
# 2.1. Low-Quality Image Recovery
To obtain the high-quality image, most of the existing methods depend on the assumption that both signal and noise arise from particular statistical regularities by using hand-crafted algorithms, such as anisotropic diffusion [28] and total variation [32]. Besides, non-parametric models [8,26] were developed to model image noise, but they were also not robust to the unconstrained environment in the wild due to priors estimated from limited observations. Recently, due to the advances in deep learning, most denoising algorithms are designed with deep neural network architectures and data-driven approach rather than relying on the priors. Burger et al.[3] employ multi-layer perceptrons with a data-driven technique based on an extensive image database. Zhang et al. [38] train the deep CNN by utilizing batch normalization (BN) [14] and residual learning [12].
Though useful for estimating a clean image, text classifers are still hard to recognize due to the irregular text geometry. It motivates research for image recovery to extend image rectification. Shi et al.[34] develop a spatial transformer network (STN) for rectifying text distortion. Cheng et al.[6] adopt more in-depth representations of images by a residual network. Different from the existing methods, in this paper, we extract deep representations of images using the U-Net-based CNN for denoising as well as rectification. To the best of our knowledge, our research may be first work to apply the two modules mentioned above for LPR at the same time.
## 2.2. License Plate Recognition
Before the advent of deep learning, most of the traditional LPR methods [16,1,13,40] employ two-stage process flow, involving text detection and following text recognition. After the advancement of deep learning, many approaches employ a one-stage process flow without text detection. Li $etal.[20]$ extract deep feature representations by using RNN with LSTM for acquiring sequential features of the license plate. Bulan $etal.[2]$ estimate domain shifts between target and multiple source domains for selecting a domain that yields the best recognition performance based on fully convolutional network [23]. However, these methods only consider high-quality license plate image except for low-quality image, which is easily led to low performance in real-world scenes. Moreover, their methods lack little or
|
[
[
0.08087158203125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.53515625,
0.09027099609375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.53515625,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.5361328125,
0.10797119140625,
0.84228515625,
0.10797119140625,
0.84228515625,
0.121826171875,
0.5361328125,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.146484375,
0.640625,
0.146484375,
0.640625,
0.1610107421875,
0.50341796875,
0.1610107421875
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.08087158203125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.08087158203125,
0.18115234375
],
[
0.51953125,
0.17041015625,
0.890625,
0.1717529296875,
0.890625,
0.186279296875,
0.51953125,
0.1849365234375
],
[
0.07843017578125,
0.1817626953125,
0.468017578125,
0.18115234375,
0.468017578125,
0.1956787109375,
0.07843017578125,
0.1962890625
],
[
0.50341796875,
0.1875,
0.89208984375,
0.1875,
0.89208984375,
0.2020263671875,
0.50341796875,
0.2020263671875
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.2020263671875,
0.7138671875,
0.2020263671875,
0.7138671875,
0.21533203125,
0.50244140625,
0.21533203125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.50341796875,
0.2266845703125,
0.7607421875,
0.2266845703125,
0.7607421875,
0.2412109375,
0.50341796875,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.521484375,
0.248779296875,
0.890625,
0.25,
0.890625,
0.264404296875,
0.521484375,
0.26318359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.266357421875,
0.8896484375,
0.266357421875,
0.8896484375,
0.276611328125,
0.50341796875,
0.276611328125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.5009765625,
0.279052734375,
0.89208984375,
0.27978515625,
0.89208984375,
0.294189453125,
0.5009765625,
0.29345703125
],
[
0.08087158203125,
0.28857421875,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.08087158203125,
0.3017578125
],
[
0.50341796875,
0.296630859375,
0.8896484375,
0.296630859375,
0.8896484375,
0.306884765625,
0.50341796875,
0.306884765625
],
[
0.07843017578125,
0.302978515625,
0.4697265625,
0.3037109375,
0.4697265625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50341796875,
0.31005859375,
0.89208984375,
0.31005859375,
0.89208984375,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.5048828125,
0.326904296875,
0.890625,
0.326904296875,
0.890625,
0.337158203125,
0.5048828125,
0.337158203125
],
[
0.08087158203125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.50341796875,
0.34033203125,
0.89208984375,
0.34033203125,
0.89208984375,
0.354248046875,
0.50341796875,
0.354248046875
],
[
0.08087158203125,
0.348388671875,
0.46728515625,
0.348388671875,
0.46728515625,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.50244140625,
0.35595703125,
0.890625,
0.35595703125,
0.890625,
0.369384765625,
0.50244140625,
0.369384765625
],
[
0.08087158203125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.50244140625,
0.37060546875,
0.89208984375,
0.37060546875,
0.89208984375,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.07843017578125,
0.37890625,
0.4697265625,
0.37744140625,
0.4697265625,
0.392578125,
0.07843017578125,
0.393310546875
],
[
0.50341796875,
0.386474609375,
0.890625,
0.386474609375,
0.890625,
0.399658203125,
0.50341796875,
0.399658203125
],
[
0.07843017578125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.07843017578125,
0.40771484375
],
[
0.50244140625,
0.401611328125,
0.890625,
0.401611328125,
0.890625,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.07843017578125,
0.4091796875,
0.468017578125,
0.40771484375,
0.468017578125,
0.42236328125,
0.07843017578125,
0.423583984375
],
[
0.50244140625,
0.416748046875,
0.89208984375,
0.416748046875,
0.89208984375,
0.429931640625,
0.50244140625,
0.429931640625
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50341796875,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50341796875,
0.4443359375
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.44580078125,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.50244140625,
0.458984375
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.4609375,
0.8798828125,
0.4609375,
0.8798828125,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.52294921875,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.490478515625,
0.52294921875,
0.490478515625
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.490478515625,
0.890625,
0.4912109375,
0.890625,
0.505859375,
0.50244140625,
0.5048828125
],
[
0.08087158203125,
0.50048828125,
0.46728515625,
0.50048828125,
0.46728515625,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50244140625,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.50244140625,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.08087158203125,
0.52978515625,
0.468017578125,
0.52978515625,
0.468017578125,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.50341796875,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.080078125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.5048828125,
0.5537109375,
0.8896484375,
0.5537109375,
0.8896484375,
0.564453125,
0.5048828125,
0.564453125
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.58203125,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.59521484375
],
[
0.080078125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50341796875,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.611328125,
0.50341796875,
0.611328125
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50341796875,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50341796875,
0.62548828125
],
[
0.08087158203125,
0.61865234375,
0.1168212890625,
0.62060546875,
0.11517333984375,
0.63330078125,
0.080078125,
0.630859375
],
[
0.5009765625,
0.626953125,
0.89208984375,
0.62548828125,
0.89208984375,
0.64013671875,
0.5009765625,
0.6416015625
],
[
0.098876953125,
0.63525390625,
0.468017578125,
0.6337890625,
0.468017578125,
0.6484375,
0.098876953125,
0.6494140625
],
[
0.50341796875,
0.64306640625,
0.57275390625,
0.64306640625,
0.57275390625,
0.65478515625,
0.50341796875,
0.65478515625
],
[
0.07843017578125,
0.6513671875,
0.193603515625,
0.65087890625,
0.193603515625,
0.66552734375,
0.07843017578125,
0.66650390625
],
[
0.09722900390625,
0.67236328125,
0.46728515625,
0.67236328125,
0.46728515625,
0.68603515625,
0.09722900390625,
0.68603515625
],
[
0.5048828125,
0.66650390625,
0.734375,
0.66650390625,
0.734375,
0.68115234375,
0.5048828125,
0.68115234375
],
[
0.1094970703125,
0.68798828125,
0.4697265625,
0.6875,
0.4697265625,
0.7021484375,
0.1094970703125,
0.70263671875
],
[
0.51953125,
0.68798828125,
0.890625,
0.689453125,
0.890625,
0.7041015625,
0.51953125,
0.70263671875
],
[
0.111083984375,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.111083984375,
0.71728515625
],
[
0.5009765625,
0.7041015625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.5009765625,
0.71728515625
],
[
0.11273193359375,
0.72021484375,
0.2034912109375,
0.72021484375,
0.2034912109375,
0.73193359375,
0.11273193359375,
0.73193359375
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.50244140625,
0.734375
],
[
0.50341796875,
0.73486328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50341796875,
0.74853515625
],
[
0.098876953125,
0.7412109375,
0.468017578125,
0.7412109375,
0.468017578125,
0.7548828125,
0.098876953125,
0.7548828125
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75,
0.890625,
0.7646484375,
0.50244140625,
0.76513671875
],
[
0.11358642578125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.76953125,
0.11358642578125,
0.76953125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.11273193359375,
0.77294921875,
0.468017578125,
0.77294921875,
0.468017578125,
0.7861328125,
0.11273193359375,
0.7861328125
],
[
0.111083984375,
0.7861328125,
0.468017578125,
0.787109375,
0.468017578125,
0.8017578125,
0.111083984375,
0.80029296875
],
[
0.50341796875,
0.78076171875,
0.8935546875,
0.78076171875,
0.8935546875,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.111083984375,
0.8017578125,
0.259033203125,
0.80322265625,
0.259033203125,
0.8173828125,
0.111083984375,
0.81640625
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.096435546875,
0.82470703125,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.096435546875,
0.8388671875
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.1094970703125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.1094970703125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.11273193359375,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.11273193359375,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.11273193359375,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.11273193359375,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.8935546875,
0.87060546875,
0.8935546875,
0.88525390625,
0.5009765625,
0.88623046875
],
[
0.11273193359375,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.11273193359375,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.8935546875,
0.88623046875,
0.8935546875,
0.8994140625,
0.50244140625,
0.8994140625
]
] |
[
"sists of two sub-networks: a denoising network and a recti-",
"AOLP-RP [13] and VTLPs dataset newly collected on",
"ficationnetwork.MotivatedbythesuccessofU-Net[31lin",
"the most challenging real-world environments.",
"recovering the object details, we employ U-Net structure as",
"an image recovery backbone network, attempting to extract",
"2. Related Work",
"visual content at structural-level details. In the denoising",
"sub-network (DSN), we try to transform a low-quality im-",
"In this section, we briefly review on low-quality image",
"age to a high-quality image pixel by pixel directly.The DSN",
"recovery methods and license plate recognition methods",
"can penalize the loss between noisy and noise-free image",
"that is most related to this work.",
"pairs and thus acquire the output image with the fine tex-",
"tures of the clean component, learning an independent real-",
"2.1.Low-Quality Image Recovery",
"ization of the noise. However, even with such sophisticated",
"To obtain the high-quality image, most of the existing",
"DSN, denoising images are unsatisfactory because they still",
"methods depend on the assumption that both signal and",
"have arbitrary geometric variations. Therefore, the rectifi-",
"noise arise from particular statistical regularities by using",
"cation sub-network (RSN) is proposed to correct geometric",
"hand-crafted algorithms.suchasanisotropicdiffusion[28]",
"distortions of denoising license plates and generate more",
"and total variation [32]. Besides, non-parametric models",
"accuratecorrection imagedistortion.Furthermore.wepro-",
"[8,26lweredevelopedtomodelimagenoise.buttheywere",
"pose to leverage the new auxiliary tasks to further optimize",
"also not robust to the unconstrained environment in the wild",
"the image recovery sub-networks (DSN, RSN) of SNIDER.",
"due to priors estimated from limited observations. Recently,",
"There are two auxiliary tasks: a text counting module and",
"due to the advances in deep learning, most denoising al-",
"a segment prediction module.Specifically,we solve each",
"gorithms are designed with deep neural network architec-",
"auxiliary module using CNN as a decoder. The counting",
"tures and data-driven approach rather than relying on the",
"module is used to predict the number of text in the im-",
"priors. Burger et al.[3] employ multi-layer perceptrons",
"age as a classification problem. In this module, despite the",
"with a data-driven technique based on an extensive image",
"ambiguous boundary of consecutive text, text counting can",
"database.Zhang et al. [38] train the deep CNN by utilizing",
"distinguish single text, which makes the image quality suit-",
"batch normalization (BN) [14] and residual learning [12].",
"able for text detection. For the segment prediction module,",
"Though useful for estimating a clean image, text classi-",
"we propose a binary segmentation to emphasize the fore-",
"fiers are still hard to recognize due to the irregular text ge-",
"ground over the background. The generated segmentation",
"ometry. It motivates research for image recovery to extend",
"result makes the license plate clean for text recognition. Fi-",
"image rectification. Shi et al. [34] develop a spatial trans-",
"nally, learning the auxiliary tasks will lead the intermedi-",
"former network (STN) for rectifying text distortion. Cheng",
"ate features of the recovery main task networks to enhance",
"etal.[6ladoptmorein-depthrepresentationsofimagesby",
"the difficulties such as geometric variations and low-quality",
"a residual network. Different from the existing methods, in",
"information.More importantly,we introduce a new loss",
"this paper, we extract deep representations of images using",
"function that trains the SNIDER with auxiliary tasks, which",
"the U-Net-based CNN for denoising as well as rectification.",
"provide significantly higher license plate quality for robust",
"To the best of our knowledge, our research may be first work",
"LPR.",
"to apply the two modules mentioned above for LPR at the",
"To sum up, we highlight the main contributions of this",
"sametime",
"paper as follows:",
".We propose a novel end-to-end license plate recovery",
"2.2.License Plate Recognition",
"network, where denoising and rectification network are",
"Before the advent of deep learning, most of the tradi-",
"used to generate a clear recovery image for robust LPR",
"tional LPR methods[16, 1,13,40] employ two-stage pro-",
"performance.",
"cess flow, involving text detection and following text recog-",
"nition. After the advancement of deep learning, many ap-",
". We present the auxiliary tasks to leverage the quality of",
"proaches employ a one-stage process flow without text de-",
"thelicenseplaterecoveryfromlow-qualityMainlya",
"tection. Li et al. [20] extract deep feature representations by",
"new loss is introduced to provide regularization effects",
"to the backbone SNIDER for robust representation and",
"using RNN with LSTM for acquiring sequential features of",
"license plate recovery.",
"the license plate. Bulan et al. [2] estimate domain shifts be-",
"tween target and multiple source domains for selecting a do-",
". Finally, we demonstrate the effectiveness of the pro-",
"main that yields the best recognition performance based on",
"posed method in recovering a high-quality license",
"fully convolutional network [23]. However, these methods",
"plate from a low-quality license plate in the real-world",
"only consider high-quality license plate image except for",
"and show that the LPR performance outperforms the",
"low-quality image, which is easily led to low performance",
"state-of-the-art methods on two challenging datasets,",
"in real-world scenes. Moreover, their methods lack little or"
] |
{
"arxiv_id": "1910.03876",
"language": "en",
"page_idx": 1
}
|
||
1910.03876-en-2
|
1910.03876-en-2.jpg
|
<FigureHere>
Figure 2. The training and testing process of the proposed approach with the learning of two auxiliary tasks:(a) The input images are fed into SNIDER for the image recovery; (b, c) SNIDER consists of main tasks (i.e. DSN, RSN) and auxiliary tasks, they transform low-quality data into high-quality data for training the DSN, RSN and auxiliary tasks networks; (d, e) LPR network is for testing and outputs a LPR result. The DSN is trained to generate a denoising image from the low-quality input image. Also, the RSN is trained to generate a rectified image from the result of DSN. The auxiliary tasks include text counting and binary segmentation, which are formulated as classification and regression simultaneously.
no effort to improve image quality, while requiring a lot of computing power. In this work, unlike existing methods, we adopt image recovery for high LPR performance under the low-quality image in real-world scenes. To the best of our knowledge, this is the first time we apply sophisticated image recovery to handle a challenging real-world environment. Besides, our methods are computationally efficient and capable of real-time recognition despite additional recovery modules.
<FigureHere>
# 3. Proposed Method
Figure 3. Label generation for the training of the proposed method. From a high-quality image as ground truth, the rotated images can be obtained using a simple linear transformation, and the lowquality image is processed through downsampling $\times1/4$ of rotation image. The segmented image is inferred through binarization [4] of the low-quality image.
The proposed approach consists of three parts: 1) main tasks prediction networks $G_D$ and $G_R$ for denoising and rectification; 2) auxiliary tasks prediction networks $D_c$ and $D_s$ for count classification and segment prediction; 3) LPR network for text detection and classification. The proposed architecture is illustrated in Figure 2. For training, dataset for main tasks and auxiliary tasks can be inferred from the intentionally transform operation by simple rotation (for rectification) and down-resizing (for denoising), as shown in Figure 3. In particular, only one sample of original image $I^{HQ}$ simply can generate four training samples that have been transformed by different angles. Given the training samples $I_i^{HQ}$ for $G_D,I_i^{LQ}$ for $G_R,I_i^{seg}$ for $D_s$ and $c$ for $D_{c},i\in\{-30^{\circ},-15^{\circ},+15^{\circ},+30^{\circ}\}$, the main tasks
$G_{D}$ and $G_R$ extract recovery result from input image $I_i^{LQ}$
and corresponding samples. LPR network $LPR$ then takes
$G_{R}(G_{D}(\tilde{I}_{i}^{LQ}))$ to recognize a recovery image.
In the following subsections, we introduce the method to predict the main tasks in Section 3.1. Then, we also address
|
[
[
0.3095703125,
0.09344482421875,
0.646484375,
0.0947265625,
0.646484375,
0.10797119140625,
0.3095703125,
0.106689453125
],
[
0.27294921875,
0.11932373046875,
0.38232421875,
0.11932373046875,
0.38232421875,
0.1300048828125,
0.27294921875,
0.1300048828125
],
[
0.56884765625,
0.11932373046875,
0.69140625,
0.11932373046875,
0.69140625,
0.1300048828125,
0.56884765625,
0.1300048828125
],
[
0.31201171875,
0.1300048828125,
0.3447265625,
0.1300048828125,
0.3447265625,
0.141357421875,
0.31201171875,
0.141357421875
],
[
0.61669921875,
0.1307373046875,
0.6435546875,
0.1307373046875,
0.6435546875,
0.1395263671875,
0.61669921875,
0.1395263671875
],
[
0.77294921875,
0.147705078125,
0.85546875,
0.147705078125,
0.85546875,
0.18310546875,
0.77294921875,
0.18310546875
],
[
0.12335205078125,
0.1685791015625,
0.176513671875,
0.1622314453125,
0.178955078125,
0.1717529296875,
0.1258544921875,
0.177978515625
],
[
0.44287109375,
0.1654052734375,
0.5009765625,
0.1610107421875,
0.50244140625,
0.17236328125,
0.443603515625,
0.1767578125
],
[
0.8056640625,
0.184326171875,
0.814453125,
0.184326171875,
0.814453125,
0.1875,
0.8056640625,
0.1875
],
[
0.318603515625,
0.1956787109375,
0.3857421875,
0.1956787109375,
0.3857421875,
0.205810546875,
0.318603515625,
0.205810546875
],
[
0.435546875,
0.1956787109375,
0.5087890625,
0.1956787109375,
0.5087890625,
0.205810546875,
0.435546875,
0.205810546875
],
[
0.60693359375,
0.1956787109375,
0.68896484375,
0.1956787109375,
0.68896484375,
0.205810546875,
0.60693359375,
0.205810546875
],
[
0.77294921875,
0.1956787109375,
0.85888671875,
0.1956787109375,
0.85888671875,
0.205810546875,
0.77294921875,
0.205810546875
],
[
0.40185546875,
0.2171630859375,
0.5439453125,
0.2171630859375,
0.5439453125,
0.2310791015625,
0.40185546875,
0.2310791015625
],
[
0.0882568359375,
0.2310791015625,
0.21728515625,
0.2310791015625,
0.21728515625,
0.2412109375,
0.0882568359375,
0.2412109375
],
[
0.78759765625,
0.2291259765625,
0.8447265625,
0.2291259765625,
0.8447265625,
0.240478515625,
0.78759765625,
0.240478515625
],
[
0.1192626953125,
0.24365234375,
0.1846923828125,
0.24365234375,
0.1846923828125,
0.255126953125,
0.1192626953125,
0.255126953125
],
[
0.765625,
0.266357421875,
0.86279296875,
0.266357421875,
0.86279296875,
0.281005859375,
0.765625,
0.281005859375
],
[
0.78173828125,
0.281005859375,
0.84619140625,
0.281005859375,
0.84619140625,
0.291748046875,
0.78173828125,
0.291748046875
],
[
0.5146484375,
0.29736328125,
0.5458984375,
0.29736328125,
0.5458984375,
0.304931640625,
0.5146484375,
0.304931640625
],
[
0.32275390625,
0.304931640625,
0.358642578125,
0.304931640625,
0.358642578125,
0.3125,
0.32275390625,
0.3125
],
[
0.410888671875,
0.311279296875,
0.441162109375,
0.311279296875,
0.441162109375,
0.318115234375,
0.410888671875,
0.318115234375
],
[
0.50732421875,
0.30615234375,
0.54345703125,
0.30615234375,
0.54345703125,
0.313232421875,
0.50732421875,
0.313232421875
],
[
0.373291015625,
0.318115234375,
0.48193359375,
0.318115234375,
0.48193359375,
0.328857421875,
0.373291015625,
0.328857421875
],
[
0.283447265625,
0.330810546875,
0.378173828125,
0.33203125,
0.378173828125,
0.343505859375,
0.283447265625,
0.34228515625
],
[
0.471435546875,
0.33203125,
0.57177734375,
0.33203125,
0.57177734375,
0.34228515625,
0.471435546875,
0.34228515625
],
[
0.39794921875,
0.348388671875,
0.56201171875,
0.348388671875,
0.56201171875,
0.3623046875,
0.39794921875,
0.3623046875
],
[
0.78173828125,
0.34521484375,
0.84619140625,
0.34521484375,
0.84619140625,
0.35986328125,
0.78173828125,
0.35986328125
],
[
0.76953125,
0.359130859375,
0.85888671875,
0.356689453125,
0.85888671875,
0.371826171875,
0.76953125,
0.373779296875
],
[
0.080078125,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.080078125,
0.40771484375,
0.890625,
0.40771484375,
0.890625,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.07843017578125,
0.421630859375,
0.89208984375,
0.421630859375,
0.89208984375,
0.43505859375,
0.07843017578125,
0.43505859375
],
[
0.07843017578125,
0.436279296875,
0.89208984375,
0.436279296875,
0.89208984375,
0.449462890625,
0.07843017578125,
0.449462890625
],
[
0.080078125,
0.449462890625,
0.89208984375,
0.449462890625,
0.89208984375,
0.46337890625,
0.080078125,
0.46337890625
],
[
0.07843017578125,
0.464111328125,
0.26318359375,
0.464111328125,
0.26318359375,
0.477783203125,
0.07843017578125,
0.477783203125
],
[
0.080078125,
0.50244140625,
0.4697265625,
0.50244140625,
0.4697265625,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.64892578125,
0.505859375,
0.71142578125,
0.5166015625,
0.70751953125,
0.52783203125,
0.64453125,
0.517578125
],
[
0.74609375,
0.50830078125,
0.79736328125,
0.5166015625,
0.7939453125,
0.52783203125,
0.74267578125,
0.51953125
],
[
0.080078125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.61669921875,
0.51025390625,
0.6396484375,
0.51025390625,
0.6396484375,
0.521484375,
0.61669921875,
0.521484375
],
[
0.66845703125,
0.5322265625,
0.69140625,
0.5322265625,
0.69140625,
0.54443359375,
0.66845703125,
0.54443359375
],
[
0.7490234375,
0.5322265625,
0.77294921875,
0.5322265625,
0.77294921875,
0.54345703125,
0.7490234375,
0.54345703125
],
[
0.8408203125,
0.53076171875,
0.865234375,
0.53076171875,
0.865234375,
0.54345703125,
0.8408203125,
0.54345703125
],
[
0.61669921875,
0.546875,
0.63818359375,
0.546875,
0.63818359375,
0.5546875,
0.61669921875,
0.5546875
],
[
0.080078125,
0.5341796875,
0.4697265625,
0.5341796875,
0.4697265625,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.080078125,
0.548828125,
0.4697265625,
0.548828125,
0.4697265625,
0.5625,
0.080078125,
0.5625
],
[
0.85107421875,
0.546875,
0.86376953125,
0.546875,
0.86376953125,
0.5498046875,
0.85107421875,
0.5498046875
],
[
0.65283203125,
0.5537109375,
0.7099609375,
0.5595703125,
0.708984375,
0.57080078125,
0.65185546875,
0.56494140625
],
[
0.74169921875,
0.5537109375,
0.79833984375,
0.56005859375,
0.7958984375,
0.57080078125,
0.740234375,
0.56494140625
],
[
0.080078125,
0.564453125,
0.4697265625,
0.564453125,
0.4697265625,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.5400390625,
0.5595703125,
0.57861328125,
0.5595703125,
0.57861328125,
0.57470703125,
0.5400390625,
0.57470703125
],
[
0.66845703125,
0.57470703125,
0.69287109375,
0.57470703125,
0.69287109375,
0.58837890625,
0.66845703125,
0.58837890625
],
[
0.7509765625,
0.57568359375,
0.77392578125,
0.57568359375,
0.77392578125,
0.5869140625,
0.7509765625,
0.5869140625
],
[
0.8408203125,
0.57568359375,
0.865234375,
0.57568359375,
0.865234375,
0.5869140625,
0.8408203125,
0.5869140625
],
[
0.5107421875,
0.5791015625,
0.611328125,
0.5791015625,
0.611328125,
0.5966796875,
0.5107421875,
0.5966796875
],
[
0.080078125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.080078125,
0.5947265625,
0.471435546875,
0.5947265625,
0.471435546875,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.65283203125,
0.5986328125,
0.705078125,
0.5947265625,
0.7060546875,
0.60595703125,
0.654296875,
0.60986328125
],
[
0.740234375,
0.59765625,
0.7900390625,
0.59521484375,
0.7900390625,
0.60693359375,
0.740234375,
0.609375
],
[
0.5498046875,
0.60498046875,
0.56884765625,
0.60302734375,
0.57177734375,
0.6123046875,
0.55224609375,
0.6142578125
],
[
0.615234375,
0.60693359375,
0.63818359375,
0.60693359375,
0.63818359375,
0.61865234375,
0.615234375,
0.61865234375
],
[
0.7509765625,
0.619140625,
0.771484375,
0.619140625,
0.771484375,
0.6318359375,
0.7509765625,
0.6318359375
],
[
0.07843017578125,
0.609375,
0.4697265625,
0.609375,
0.4697265625,
0.62255859375,
0.07843017578125,
0.62255859375
],
[
0.6650390625,
0.62060546875,
0.68896484375,
0.6162109375,
0.69287109375,
0.630859375,
0.6689453125,
0.63525390625
],
[
0.84326171875,
0.619140625,
0.86376953125,
0.619140625,
0.86376953125,
0.6318359375,
0.84326171875,
0.6318359375
],
[
0.53173828125,
0.62255859375,
0.5947265625,
0.62255859375,
0.5947265625,
0.63330078125,
0.53173828125,
0.63330078125
],
[
0.080078125,
0.625,
0.1895751953125,
0.625,
0.1895751953125,
0.63818359375,
0.080078125,
0.63818359375
],
[
0.50634765625,
0.63330078125,
0.61767578125,
0.63330078125,
0.61767578125,
0.64306640625,
0.50634765625,
0.64306640625
],
[
0.6689453125,
0.642578125,
0.69384765625,
0.64013671875,
0.6953125,
0.6484375,
0.6708984375,
0.65087890625
],
[
0.611328125,
0.65283203125,
0.6357421875,
0.6494140625,
0.63818359375,
0.6611328125,
0.61376953125,
0.66357421875
],
[
0.07843017578125,
0.6591796875,
0.250732421875,
0.66015625,
0.250732421875,
0.67822265625,
0.07843017578125,
0.6767578125
],
[
0.66845703125,
0.66357421875,
0.69287109375,
0.66357421875,
0.69287109375,
0.6767578125,
0.66845703125,
0.6767578125
],
[
0.7490234375,
0.66357421875,
0.77392578125,
0.66357421875,
0.77392578125,
0.6767578125,
0.7490234375,
0.6767578125
],
[
0.84326171875,
0.662109375,
0.86669921875,
0.662109375,
0.86669921875,
0.6767578125,
0.84326171875,
0.6767578125
],
[
0.65625,
0.67529296875,
0.69677734375,
0.67529296875,
0.69677734375,
0.68603515625,
0.65625,
0.68603515625
],
[
0.74609375,
0.673828125,
0.7861328125,
0.67529296875,
0.78515625,
0.6875,
0.744140625,
0.68505859375
],
[
0.83251953125,
0.67529296875,
0.87353515625,
0.67529296875,
0.87353515625,
0.6875,
0.83251953125,
0.6875
],
[
0.098876953125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.705078125,
0.098876953125,
0.705078125
],
[
0.07843017578125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.5009765625,
0.70556640625,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.5009765625,
0.72021484375
],
[
0.080078125,
0.7216796875,
0.4697265625,
0.7216796875,
0.4697265625,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.50244140625,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50244140625,
0.734375
],
[
0.08251953125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08251953125,
0.74755859375
],
[
0.50341796875,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.7470703125,
0.50341796875,
0.7470703125
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.07843017578125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.07843017578125,
0.77978515625
],
[
0.50244140625,
0.76220703125,
0.89208984375,
0.76220703125,
0.89208984375,
0.77587890625,
0.50244140625,
0.77587890625
],
[
0.50244140625,
0.7763671875,
0.67236328125,
0.7763671875,
0.67236328125,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.07843017578125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5,
0.82080078125,
0.8935546875,
0.81494140625,
0.8935546875,
0.8369140625,
0.5,
0.84228515625
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.0767822265625,
0.83837890625,
0.4697265625,
0.83837890625,
0.4697265625,
0.85595703125,
0.0767822265625,
0.85595703125
],
[
0.50244140625,
0.83837890625,
0.89208984375,
0.83837890625,
0.89208984375,
0.8515625,
0.50244140625,
0.8515625
],
[
0.5009765625,
0.8486328125,
0.81640625,
0.8486328125,
0.81640625,
0.869140625,
0.5009765625,
0.869140625
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.869140625
],
[
0.07598876953125,
0.86572265625,
0.473876953125,
0.8671875,
0.473876953125,
0.8896484375,
0.07598876953125,
0.88720703125
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"SNIDER: Single Noisy Image DEnoising and Rectification",
"Denoising Sub-Network",
"Rectification Sub-Network",
"(DSN)",
"(RSN)",
"322910",
"",
"3292910",
"-",
"Denoising Loss",
"Denoising Result",
"Rectification Loss",
"Rectification Result",
"(b) Main Tasks Prediction",
"(a)Low-qualitylicense",
"LPRNetwork",
"plate image",
"(d)License plate",
"recognition",
"egment",
"Decode",
"(Sumo)",
"Decoder",
"each last encoder feature)",
"Text Counting Loss",
"Semantic Pixel:wise Loss",
"(c)Auxiliary Tasks Prediction",
"'322910'",
"(e) LPR Result",
"Figure 2. The training and testing process of the proposed approach with the learning of two auxiliary tasks: (a) The input images are fed",
"into SNIDER for the image recovery; (b, c) SNIDER consists of main tasks (i.e. DSN, RSN) and auxiliary tasks, they transform low-quality",
"data into high-quality data for training the DSN, RSN and auxiliary tasks networks; (d, e) LPR network is for testing and outputs a LPR",
"result. The DSN is trained to generate a denoising image from the low-quality input image. Also, the RSN is trained to generate a rectified",
"image from the result of DSN. The auxiliary tasks include text counting and binary segmentation, which are formulated as classification",
" and regression simultaneously.",
"no effort to improve image quality, while requiring a lot of",
"36N8746",
"36W0743",
"computing power. In this work, unlike existing methods,",
"30°",
"1H30",
"1-30",
"1520",
"-15°",
"we adopt image recovery for high LPR performance under",
"the low-quality image in real-world scenes. To the best of",
"",
"36#8746",
"36W0746",
"our knowledge, this is the first time we apply sophisticated",
"D",
"1415.",
"1205",
"1-15",
"368746",
"image recovery to handle a challenging real-world environ-",
"ment. Besides, our methods are computationally efficient",
"3608746",
"360874",
"THO",
"+15",
"1415",
"and capable of real-time recognition despite additional re-",
"THO",
"1415",
"Text Num:6",
"covery modules.",
"Recognition :\"36 8746\"",
"87",
"+30",
"3. Proposed Method",
"IH3O",
"1430",
"1530",
"Rotation",
"Resizing",
"Segment",
"The proposed approach consists of three parts: 1) main",
"tasks prediction networks Gp and GR for denoising and",
"Figure 3. Label generation for the training of the proposed method.",
"rectification; 2) auxiliary tasks prediction networks De and",
"From a high-quality image as ground truth, the rotated images",
"Dforcountclassificationandsegmentprediction:3)LPR",
"can be obtained using a simple linear transformation, and the low-",
"network for text detection and classification. The proposed",
"quality image is processed through downsampling × 1/4 of rota-",
"architecture is illustrated in Figure 2. For training, dataset",
"tion image. The segmented image is inferred through binarization",
"[4] of the low-quality image.",
"for main tasks and auxiliarytasks can be inferred from the",
"intentionally transform operation by simple rotation (for",
"rectification) and down-resizing (for denoising), as shown",
"GD and GR extract recovery result from input image ILQ",
"in Figure 3. In particular, only one sample of original im-",
" age jHQ simply can generate four training samples that",
"and corresponding samples. LPR network LPR then takes",
"Gr(Gp(I/) to recognize a recovery image.",
"have been transformed by different angles. Given the train-",
"ing samples IHQ for GD, IJQ for GR, Ige9 for D, and c",
"In the following subsections, we introduce the method to",
"for De, i {-30°,-15°,+15°,+30°], the main tasks",
"predict the main tasks in Section 3.1. Then, we also address"
] |
{
"arxiv_id": "1910.03876",
"language": "en",
"page_idx": 2
}
|
||
1910.03876-en-3
|
1910.03876-en-3.jpg
|
the auxiliary tasks for prediction in Section 3.2. Then, we describe the network training of the proposed architecture in Section 3.3. Finally, we illustrate the testing process in Section 3.4.
# 3.1. Denoising and Rectification Network
Our main task networks include two sub-networks (i.e. denoising sub-network and rectification sub-network), and the first sub-network takes the low-quality image as the input, and the output is the recovered image. In this paper, we design the rectification network to rectify the denoising results from the denoising network.
The image recovery results[15] have shown the effectiveness of the U-Net since it can provide high-quality overall details of an image object, without a negative impact on the image generation. Therefore, we adopt a U-Net-based architecture adding skip connections that shuffle low-level information shared between input and output across the network. In contrast to their network, our recovery network includes two sub-networks, which are also the U-Net architecture. As shown in Table 1 and Figure 2.(b,c), our denoising network $G_D$ and rectification network $G_R$ consist of the encoder and decoder module.
To achieve the main tasks, we first feed $I_i^{LQ}$ into $G_D$ to generate denoising results. Given a pair of input image and non-rectified ground-truth denoising image $\left\{I_{i,j}^{LQ},I_{i,j}^{HQ}\right\}_{(i,j)}^{N}$,loss function for the $G_D$ is the pixel-wise MSE loss, and it is calculated as Eq. (1):
$$\mathcal{L}_{G_D}(w)=\frac{1}{N}\sum\limits_{(i,j)\in N}^{N}\parallel G_{D_w}(I_{(i,j)}^{LQ})-I_{(i,j)}^{HQ}\parallel^2,$$
(1)
where $w$ is the parameters of denoising network. Such loss function encourages the $G_D$ to not only extract the content information of input image but also generate a high-quality natural image in pixel level.
Then, the rectification sub-network $G_R$ processes the output from $G_D$, and outputs a rectified high-quality image, which is easier for the LPR network to recognize the identification text. With the training pairs of $\{G_D(I_{i,j}^{LQ}),I_{i=0,j}^{HQ}\}_{(i,j)}^N$, the $G_R$ can be trained using a L1 loss for the predicted result $G_R(G_D(I_{i,j}^{LQ})):$
$$\mathcal{L}_{G_{R}}(w)=\frac{1}{N}\sum_{(i,j)\in N}\parallel G_{R_{w}}(G_{D_{w}}(I_{(i,j)}^{LQ}))-I_{(i=0,j)}^{HQ}\parallel^{1},$$
(2)
where $w$ is the parameters of the rectification network. Unlike L2 loss, using L1 loss in the pixel level helps to preserve the appearance of an object, such as image color, intensity, and illumination, and leads to denoising result capable of only geometric transformation. Therefore, we can
Table 1. Details of different
network architectures
<table>
<tbody>
<tr>
<th> </th>
<th>SNIDER</th>
<th>SNIDER-Tiny</th>
</tr>
<tr>
<td>Encoder (for$G_D$, $G_{R})$</td>
<td>$\operatorname{convl}(1,2):(3\times3\times32$ conv$)\times2$, stride 2 pooll$:(2\times2)$max pooling, stride 2 conv$2(1,2):(3\times3\times64$ conv$)\times2$, stride 2 pool$2:(2\times2)$ max pooling, stride 2 conv$3(1,2):(3\times3\times128$ conv$)\times2$, stride 2 $pool3:(2\times2) max pooling, stride 2$ conv$4(1,2):(3\times3\times256$ conv$)\times2$, stride 2 pool$4:(2\times2)$ max pooling, stride 2</td>
<td>$7\times7\times32$ conv stride 2 $7\times7\times64$ conv stride 2 $5\times5\times128$ conv</td>
</tr>
<tr>
<td>Decoder (for$G_D$, $G_{R}$,</td>
<td>$\mathbf{r}^{-1}=\mathbf{L}^{\prime})$ conv6( 1, 2) : ( 3$\times$3$\times$256) , stride 2 x2 upsample: concat(conv6,2, conv3.2) $conv7(1,2):(3\times3\times128),stride 2$ x2 upsample: concat(conv7-2, conv2-2)</td>
<td>$5\times5\times64$ conv $7\times7\times32$ deconv dilate 2</td>
</tr>
<tr>
<td>Decoder (for $D_c)$</td>
<td>$\mathbf{N\times N\times512}$conv $1\times1\times256$ conv $1\times1\times128$ conv $1\times1\times64$ conv $1\times1\times1$ conv</td>
<td>$\mathbf{N\times N\times128}$conv $1\times1\times64$ conv $1\times1\times1$ conv</td>
</tr>
</tbody>
</table>
only perform geometric transformations without the appearance damage of the image during the rectification process, which forces the recognizer to be helpful.
## 3.2. Auxiliary Tasks Prediction
Due to the complex real-world environments such as the extremely irregular geometric shape of text as well as the complicated image background, the binary information of the license plate is often noisy. Although we intend $G_D$ and $G_R$ to capture robust features for image recovery, the results by this structure do not always guarantee a well-enhanced output. Therefore, our work involves an additional learning branch where a richer feature representation is obtained from the backbone network. Motivated by multi-task learning [5], we employ the auxiliary tasks, $i.e.$, binary segmentation and count estimation, which will contribute our main task networks produce more discriminative feature representations. Towards this problem, we sum the weights of the last layer of encoders in order to guide auxiliary task networks to help main task networks effectively extract critical information from the low-quality image.
For the binary segmentation task, we introduce the segment decoder $D_s$ based on U-Net architecture. Detailed architectures of the $D_s$ are shown in Table 1. The $D_s$ accepts feature set $F$ summed from the last features of each main task's encoder and outputs a license plate segment with val-
|
[
[
0.50244140625,
0.08587646484375,
0.85693359375,
0.08709716796875,
0.85693359375,
0.10162353515625,
0.50244140625,
0.10101318359375
],
[
0.080078125,
0.0921630859375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.654296875,
0.10479736328125,
0.70166015625,
0.10479736328125,
0.70166015625,
0.11553955078125,
0.654296875,
0.11553955078125
],
[
0.7890625,
0.10479736328125,
0.86279296875,
0.10479736328125,
0.86279296875,
0.11614990234375,
0.7890625,
0.11614990234375
],
[
0.576171875,
0.11932373046875,
0.77392578125,
0.11871337890625,
0.77392578125,
0.1319580078125,
0.576171875,
0.1331787109375
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.08087158203125,
0.1395263671875,
0.16015625,
0.1395263671875,
0.16015625,
0.15087890625,
0.08087158203125,
0.15087890625
],
[
0.57666015625,
0.1337890625,
0.74853515625,
0.1337890625,
0.74853515625,
0.147705078125,
0.57666015625,
0.147705078125
],
[
0.580078125,
0.15087890625,
0.771484375,
0.15087890625,
0.771484375,
0.1610107421875,
0.580078125,
0.1610107421875
],
[
0.79345703125,
0.149658203125,
0.86669921875,
0.149658203125,
0.86669921875,
0.1610107421875,
0.79345703125,
0.1610107421875
],
[
0.080078125,
0.160400390625,
0.39453125,
0.160400390625,
0.39453125,
0.1749267578125,
0.080078125,
0.1749267578125
],
[
0.57666015625,
0.1634521484375,
0.74853515625,
0.1634521484375,
0.74853515625,
0.1767578125,
0.57666015625,
0.1767578125
],
[
0.7900390625,
0.1634521484375,
0.8349609375,
0.1634521484375,
0.8349609375,
0.177978515625,
0.7900390625,
0.177978515625
],
[
0.7900390625,
0.179931640625,
0.86669921875,
0.179931640625,
0.86669921875,
0.1900634765625,
0.7900390625,
0.1900634765625
],
[
0.515625,
0.1641845703125,
0.5595703125,
0.1641845703125,
0.5595703125,
0.175537109375,
0.515625,
0.175537109375
],
[
0.515625,
0.18115234375,
0.56396484375,
0.18115234375,
0.56396484375,
0.1925048828125,
0.515625,
0.1925048828125
],
[
0.57861328125,
0.18115234375,
0.779296875,
0.18115234375,
0.779296875,
0.1912841796875,
0.57861328125,
0.1912841796875
],
[
0.515625,
0.1956787109375,
0.5439453125,
0.1956787109375,
0.5439453125,
0.2083740234375,
0.515625,
0.2083740234375
],
[
0.098876953125,
0.18310546875,
0.465576171875,
0.18310546875,
0.465576171875,
0.1962890625,
0.098876953125,
0.1962890625
],
[
0.580078125,
0.1962890625,
0.74658203125,
0.1962890625,
0.74658203125,
0.20703125,
0.580078125,
0.20703125
],
[
0.79150390625,
0.1956787109375,
0.83251953125,
0.1956787109375,
0.83251953125,
0.20703125,
0.79150390625,
0.20703125
],
[
0.580078125,
0.2108154296875,
0.779296875,
0.2108154296875,
0.779296875,
0.2215576171875,
0.580078125,
0.2215576171875
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.080078125,
0.2139892578125,
0.46728515625,
0.2139892578125,
0.46728515625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.79150390625,
0.2108154296875,
0.87158203125,
0.2108154296875,
0.87158203125,
0.2215576171875,
0.79150390625,
0.2215576171875
],
[
0.07843017578125,
0.2279052734375,
0.468017578125,
0.2291259765625,
0.468017578125,
0.24365234375,
0.07843017578125,
0.242431640625
],
[
0.57861328125,
0.2247314453125,
0.74853515625,
0.2247314453125,
0.74853515625,
0.238037109375,
0.57861328125,
0.238037109375
],
[
0.080078125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.080078125,
0.25830078125
],
[
0.57666015625,
0.2392578125,
0.779296875,
0.2392578125,
0.779296875,
0.25244140625,
0.57666015625,
0.25244140625
],
[
0.080078125,
0.2587890625,
0.31201171875,
0.2587890625,
0.31201171875,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.57666015625,
0.255615234375,
0.7685546875,
0.255615234375,
0.7685546875,
0.26953125,
0.57666015625,
0.26953125
],
[
0.098876953125,
0.274658203125,
0.468017578125,
0.274658203125,
0.468017578125,
0.28857421875,
0.098876953125,
0.28857421875
],
[
0.580078125,
0.270751953125,
0.73779296875,
0.270751953125,
0.73779296875,
0.281005859375,
0.580078125,
0.281005859375
],
[
0.580078125,
0.287353515625,
0.76708984375,
0.287353515625,
0.76708984375,
0.29736328125,
0.580078125,
0.29736328125
],
[
0.79345703125,
0.285888671875,
0.865234375,
0.285888671875,
0.865234375,
0.29736328125,
0.79345703125,
0.29736328125
],
[
0.515625,
0.29345703125,
0.560546875,
0.29345703125,
0.560546875,
0.304931640625,
0.515625,
0.304931640625
],
[
0.580078125,
0.3017578125,
0.736328125,
0.3017578125,
0.736328125,
0.3125,
0.580078125,
0.3125
],
[
0.07843017578125,
0.28857421875,
0.468017578125,
0.2890625,
0.468017578125,
0.3037109375,
0.07843017578125,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.5146484375,
0.30615234375,
0.56396484375,
0.30810546875,
0.56201171875,
0.322509765625,
0.51318359375,
0.32080078125
],
[
0.79150390625,
0.3017578125,
0.87744140625,
0.3017578125,
0.87744140625,
0.313232421875,
0.79150390625,
0.313232421875
],
[
0.5810546875,
0.317626953125,
0.765625,
0.317626953125,
0.765625,
0.32763671875,
0.5810546875,
0.32763671875
],
[
0.79150390625,
0.31640625,
0.83251953125,
0.31640625,
0.83251953125,
0.32763671875,
0.79150390625,
0.32763671875
],
[
0.5146484375,
0.322509765625,
0.54345703125,
0.324462890625,
0.54150390625,
0.337158203125,
0.51318359375,
0.335205078125
],
[
0.7900390625,
0.328857421875,
0.87158203125,
0.330810546875,
0.87158203125,
0.34521484375,
0.7900390625,
0.343505859375
],
[
0.07843017578125,
0.3193359375,
0.468017578125,
0.318115234375,
0.468017578125,
0.333251953125,
0.07843017578125,
0.333984375
],
[
0.080078125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.57666015625,
0.330810546875,
0.734375,
0.330810546875,
0.734375,
0.3447265625,
0.57666015625,
0.3447265625
],
[
0.51708984375,
0.339111328125,
0.54345703125,
0.339111328125,
0.54345703125,
0.3515625,
0.51708984375,
0.3515625
],
[
0.57861328125,
0.3466796875,
0.76708984375,
0.3466796875,
0.76708984375,
0.35986328125,
0.57861328125,
0.35986328125
],
[
0.79150390625,
0.34521484375,
0.833984375,
0.34521484375,
0.833984375,
0.35986328125,
0.79150390625,
0.35986328125
],
[
0.08087158203125,
0.3515625,
0.46728515625,
0.3515625,
0.46728515625,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.08087158203125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.57861328125,
0.361083984375,
0.73388671875,
0.361083984375,
0.73388671875,
0.375,
0.57861328125,
0.375
],
[
0.080078125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.57861328125,
0.376953125,
0.6708984375,
0.376953125,
0.6708984375,
0.38818359375,
0.57861328125,
0.38818359375
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.63818359375,
0.392578125,
0.71826171875,
0.392578125,
0.71826171875,
0.404052734375,
0.63818359375,
0.404052734375
],
[
0.63818359375,
0.40771484375,
0.71337890625,
0.40771484375,
0.71337890625,
0.41845703125,
0.63818359375,
0.41845703125
],
[
0.79345703125,
0.4072265625,
0.875,
0.4072265625,
0.875,
0.41845703125,
0.79345703125,
0.41845703125
],
[
0.5146484375,
0.41357421875,
0.56201171875,
0.41357421875,
0.56201171875,
0.427978515625,
0.5146484375,
0.427978515625
],
[
0.63623046875,
0.42236328125,
0.7138671875,
0.42236328125,
0.7138671875,
0.434326171875,
0.63623046875,
0.434326171875
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.080078125,
0.4267578125,
0.271240234375,
0.4267578125,
0.271240234375,
0.4375,
0.080078125,
0.4375
],
[
0.79345703125,
0.42236328125,
0.86669921875,
0.42236328125,
0.86669921875,
0.434326171875,
0.79345703125,
0.434326171875
],
[
0.09478759765625,
0.43310546875,
0.47216796875,
0.43310546875,
0.47216796875,
0.45703125,
0.09478759765625,
0.45703125
],
[
0.5146484375,
0.4287109375,
0.56396484375,
0.4287109375,
0.56396484375,
0.443115234375,
0.5146484375,
0.443115234375
],
[
0.63818359375,
0.438232421875,
0.708984375,
0.438232421875,
0.708984375,
0.448974609375,
0.63818359375,
0.448974609375
],
[
0.7939453125,
0.438232421875,
0.859375,
0.438232421875,
0.859375,
0.449462890625,
0.7939453125,
0.449462890625
],
[
0.08087158203125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.63818359375,
0.45263671875,
0.70361328125,
0.45263671875,
0.70361328125,
0.464111328125,
0.63818359375,
0.464111328125
],
[
0.080078125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.0743408203125,
0.4765625,
0.4755859375,
0.481689453125,
0.4755859375,
0.51025390625,
0.0743408203125,
0.5048828125
],
[
0.07843017578125,
0.5,
0.34716796875,
0.50048828125,
0.34716796875,
0.51513671875,
0.07843017578125,
0.5146484375
],
[
0.50244140625,
0.5068359375,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.50244140625,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.2181396484375,
0.5341796875,
0.23779296875,
0.5341796875,
0.23779296875,
0.546875,
0.2181396484375,
0.546875
],
[
0.50244140625,
0.537109375,
0.77685546875,
0.537109375,
0.77685546875,
0.55029296875,
0.50244140625,
0.55029296875
],
[
0.238525390625,
0.5390625,
0.35302734375,
0.54443359375,
0.3505859375,
0.5693359375,
0.236083984375,
0.564453125
],
[
0.193603515625,
0.5390625,
0.2017822265625,
0.54541015625,
0.193603515625,
0.5517578125,
0.1854248046875,
0.54541015625
],
[
0.2156982421875,
0.546875,
0.2484130859375,
0.546875,
0.2484130859375,
0.5625,
0.2156982421875,
0.5625
],
[
0.3544921875,
0.54248046875,
0.4345703125,
0.54248046875,
0.4345703125,
0.568359375,
0.3544921875,
0.568359375
],
[
0.10540771484375,
0.548828125,
0.1854248046875,
0.548828125,
0.1854248046875,
0.56298828125,
0.10540771484375,
0.56298828125
],
[
0.3203125,
0.548828125,
0.33251953125,
0.548828125,
0.33251953125,
0.5546875,
0.3203125,
0.5546875
],
[
0.181396484375,
0.55615234375,
0.2034912109375,
0.55615234375,
0.2034912109375,
0.568359375,
0.181396484375,
0.568359375
],
[
0.4306640625,
0.55029296875,
0.46728515625,
0.55029296875,
0.46728515625,
0.5625,
0.4306640625,
0.5625
],
[
0.5009765625,
0.55810546875,
0.74169921875,
0.56005859375,
0.74169921875,
0.57763671875,
0.5009765625,
0.57568359375
],
[
0.2059326171875,
0.56640625,
0.25244140625,
0.56640625,
0.25244140625,
0.57763671875,
0.2059326171875,
0.57763671875
],
[
0.080078125,
0.58544921875,
0.4697265625,
0.58544921875,
0.4697265625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.521484375,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.080078125,
0.60107421875,
0.4697265625,
0.60107421875,
0.4697265625,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.50244140625,
0.60107421875,
0.89208984375,
0.60107421875,
0.89208984375,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.07843017578125,
0.630859375,
0.2646484375,
0.630859375,
0.2646484375,
0.64453125,
0.07843017578125,
0.64453125
],
[
0.5009765625,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.5009765625,
0.642578125
],
[
0.098876953125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.66015625,
0.098876953125,
0.66015625
],
[
0.50341796875,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.080078125,
0.6611328125,
0.468017578125,
0.6611328125,
0.468017578125,
0.6748046875,
0.080078125,
0.6748046875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.07843017578125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.68994140625,
0.07843017578125,
0.68994140625
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.07843017578125,
0.689453125,
0.471435546875,
0.68994140625,
0.471435546875,
0.705078125,
0.07843017578125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07598876953125,
0.69873046875,
0.473876953125,
0.69873046875,
0.473876953125,
0.7265625,
0.07598876953125,
0.7265625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.0767822265625,
0.7216796875,
0.371826171875,
0.7216796875,
0.371826171875,
0.74365234375,
0.0767822265625,
0.74365234375
],
[
0.5009765625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.734375,
0.5009765625,
0.734375
],
[
0.50244140625,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.50244140625,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.1854248046875,
0.7646484375,
0.431396484375,
0.7646484375,
0.431396484375,
0.796875,
0.1854248046875,
0.796875
],
[
0.1993408203125,
0.76025390625,
0.214111328125,
0.76025390625,
0.214111328125,
0.76953125,
0.1993408203125,
0.76953125
],
[
0.1641845703125,
0.77099609375,
0.1732177734375,
0.7646484375,
0.1805419921875,
0.76953125,
0.17236328125,
0.7763671875
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08416748046875,
0.77294921875,
0.16259765625,
0.77099609375,
0.16259765625,
0.7861328125,
0.08416748046875,
0.78857421875
],
[
0.158447265625,
0.78271484375,
0.178955078125,
0.77783203125,
0.1846923828125,
0.79052734375,
0.1641845703125,
0.79541015625
],
[
0.386474609375,
0.77880859375,
0.461669921875,
0.76806640625,
0.465576171875,
0.787109375,
0.390625,
0.79736328125
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.181396484375,
0.79052734375,
0.2303466796875,
0.78857421875,
0.2320556640625,
0.80322265625,
0.1829833984375,
0.8046875
],
[
0.50341796875,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.44677734375,
0.8046875,
0.4697265625,
0.8046875,
0.4697265625,
0.818359375,
0.44677734375,
0.818359375
],
[
0.50244140625,
0.81103515625,
0.7939453125,
0.81103515625,
0.7939453125,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.521484375,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.521484375,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85595703125,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5009765625,
0.869140625
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Table 1. Details of different proposed network architectures.",
"the auxiliary tasks for prediction in Section 3.2. Then, we",
"SNIDER",
"SNIDER-Tiny",
"convl(1,2): (3×3×32 conv)×2, stride 2",
"describe the network training of the proposed architecture",
"in Section 3.3. Finally, we illustrate the testing process in",
"Section 3.4.",
"pooll : (2×2) max pooling, stride 2",
"conv2(1.2):(3×3×64conv)×2,stride2",
"7×7×32conv,",
"3.1. Denoising and Rectification Network",
"pool2 : (2×2) max pooling. stride 2",
"stride2",
"7×7×64conv.",
"Encoder",
"(for GD,",
"conv3(1.2):(3×3×128conv)×2,stride2",
"GR)",
"Our main task networks include two sub-networks (i.e.",
"pool3 : (2x2)max pooling, stride 2",
"stride 2",
"conv4(1.2):(3×3×256 conv)×2, stride2",
"denoising sub-network and rectification sub-network), and",
"the first sub-network takes the low-quality image as the in-",
"5×5×128conv",
"put, and the output is the recovered image. In this paper,",
"pool4 : (2x2) max pooling, stride 2",
"we design the rectification network to rectify the denoising",
"conv5(1,2) : (3×3×512 conv)×2, stride 2",
"results from the denoising network.",
"x2upsample:concat(conv5.2.conv4.2)",
"The image recovery results [15] have shown the effec-",
"conv6(1.2):(3×3×256),stride2",
"x2upsample:concat(conv6_2,conv3_2)",
"5×5×64conv",
"Decoder",
"conv7(1,2) : (3×3×128), stride 2",
"tiveness of the U-Net since it can provide high-quality over-",
"all details of an image object, without a negative impact on",
"(for GD.",
"7×7×32deconv,",
"x2upsample:concat(conv7_2,conv2.2)",
"dilate2",
"GR.",
"7×7×3deconv.",
"the image generation. Therefore, we adopt a U-Net-based",
"architecture adding skip connections that shuffle low-level",
"conv8(1,2) : (3×3×64), stride 2",
"Ds)",
"x2upsample:concat(conv8.2.conv2.2)",
"dilate2",
"information shared between input and output acrossthenet-",
"work.Incontrasttotheirnetwork,ourrecoverynetwork",
"conv9(1,2) : (3×3×32), stride 2",
"includes two sub-networks, which are also the U-Net archi-",
"conv10:(1×1x3)",
"tecture. As shown in Table I and Figure 2.(b,c), our denois-",
"N×N×512conv",
"1×1×256conv",
"N×Nx128conv",
"Decoder",
"1×1×128 conv",
"ing network GD and rectification network GR consist of the",
"encoderanddecodermodule",
"1×1×64 conv",
"To achieve the main tasks, we first feed IIQ into",
"(for De)",
"1×1×64conv",
"1×1×1 conv",
"GD to generate denoising results. Given a pair of in-",
"1x1x1 conv",
"put image and non-rectified ground-truth denoising image",
"",
"MSE loss, and'it is calculated as Eq. (1):",
"only perform geometric transformations without the appear-",
"ance damage of the image during the rectification process,",
"N",
"which forces the recognizer to be helpful.",
"IIGD(g)",
"1",
"7",
" 12",
"LGp(w) =",
"LC",
"N",
"()",
"3.2. Auxiliary Tasks Prediction",
"(i,j)EN",
"where w is the parameters of denoising network. Such loss",
"Due to the complex real-world environments such as the",
"function encourages the GD to not only extract the content",
"extremely irregular geometric shape of text as well as the",
"information of input image but also generate a high-quality",
"complicated image background, the binary information of",
" natural image in pixel level.",
"the license plate is often noisy. Although we intend Gp and",
"Then, the rectification sub-network GR processes the",
"G r to capture robust features for image recovery, the results",
"output from GD, and outputs a rectified high-quality",
"by this structure do not always guarantee a well-enhanced",
"image, which is easier for the LPR network to recog-",
"output. Therefore, our work involves an additional learn-",
"nize the identification text.With the training pairs of",
"ing branch where a richer feature representation is obtained",
"(G(9),8),theGncan be trained using aI",
"from the backbone network. Motivated by multi-task learn-",
"lossfor the predicted result GR(Gp(I):",
"ing [5], we employ the auxiliary tasks, i.e., binary segmen-",
"tation and count estimation, which will contribute our main",
"task networks produce more discriminative feature repre-",
" I GRu (GD.() - 10",
"N",
"",
"sentations. Towards this problem, we sum the weights of",
"CGR(W) =",
"N",
"(i=0,) I1,",
"the last layer of encoders in order to guide auxiliary task",
"(i.j)EN",
"networkstohelpmaintasknetworks effectivelyextractcrit-",
"(2)",
"ical information from the low-quality image.",
"where w is the parameters of the rectification network. Un-",
"For the binary segmentation task, we introduce the seg-",
"like L2 loss, using Ll loss in the pixel level helps to pre-",
"ment decoder D, based on U-Net architecture. Detailed ar-",
"serve the appearance of an object, such as image color, in-",
"chitecturesof theD,areshowninTable1.TheD,accepts",
"tensity, and illumination, and leads to denoising result ca-",
"feature set F summed from the last features of each main",
"pable of only geometric transformation. Therefore, we can",
"task's encoder and outputs a license plate segment with val-"
] |
{
"arxiv_id": "1910.03876",
"language": "en",
"page_idx": 3
}
|
||
1910.03876-en-4
|
1910.03876-en-4.jpg
|
ues indicating the probability of pixels belonging to the license plate. Also, ground-truth labels for segmentation can be inferred from the dotted annotations by [4]'s method as Otsu Thresholding, as shown in Figure 3. Although our segmentation annotations by [4] do not fully reflect the actual detail appearance of an image, we have shown in the experiments that this auxiliary and straightforward learning strategy leads to effective advances in image recovery. Given a pair of $F$ and the ground-truth segmentation result in $I^{seg}$, loss function for the $D_{s}$ is the binary cross-entropy loss:
(3)
$$\begin{aligned}\mathcal{L}_{D_{s}}(w)=&\frac{1}{N}\sum_{(x,y)\in N}^{N}I_{(x,y)}^{seg}log(D_{s_{w}}(F)_{(x,y)})\\&+(1-I_{(x,y)}^{seg})log(1-D_{s_{w}}(F)_{(x,y)}),\end{aligned}$$
where $I_(x,y)^{seg}\in\{0,1\}$ is the real classes of pixels in $I^{seg}$ with 1 for the license plate area and 0 for the background, $D_{s}(F)_{(x,y)}$denotes the pixel-wise probability by $D_{s}.$
Also, we find that the generated recovery samples cannot usually distinguish successive texts due to close to each other. Motivated by the observations, we add a counting decoder $D_{c}$, which predicts the number of characters in the image. As a result, our $D_c$ plays two roles, where the first is to cause separation between adjacent texts more clearly. The other role is to promote the encoders of each main task to generate a higher quality image while backpropagating the penalty. The loss function for the $D_c$ is the L2 loss:
(4)
$$\mathcal{L}_{D_c}=\parallel C_{pred}-C_{G.T}\parallel^2,$$
where $C_{pred}$ and $C_{G.T}$ are the predicted value and the
ground-truth, respectively.
# 3.3. Network Training
The full objective function is a weighted sum of all the
losses from Eq. (1) to (4):
: (5)
$$\mathcal{L}=\lambda_{G_{D}}\mathcal{L}_{G_{D}}+\lambda_{G_{R}}\mathcal{L}_{G_{R}}+\lambda_{D_{s}}\mathcal{L}_{D_{s}}+\lambda_{D_{c}}\mathcal{L}_{D_{C}}$$
We employ a stage-wise training strategy to optimize main tasks with auxiliary tasks and empirically set the weights of each loss as detailed in Section 5.3.
# 3.4. Testing
At the testing phase, the auxiliary tasks are removed. Given a low-quality test image $I_{test},G_{D}$ and $G_R$ output the recovered image via denoising and rectification. Then LPR network $LPR$ based on a YOLO v3 detector [29] by pre-trained on ImageNet [9] takes the recovered image and generates the recognition result $LPR_{result}$ of $I_{test}$, and it is denoted as Eq. (6):
$$LPR_{result}=LPR(G_{R}(G_{D}(I_{test}))).$$
(6)
# 4. Experimental Setting
In this section, we describe a list of datasets, metric, and
implementation details for the proposed method.
## 4.1. Datasets
We use LPR datasets AOLP [13] and newly collected
dataset, named VTLP.
$\mathbf{AOLP-RP:AOLP-RP[13]consists~of~611~images~col-}$ lected in Taiwan, including ten numbers and 25 letters (except "O"). This dataset has a challenging factor that the angle of the LP contains oblique samples in terms of distortion. On the other hand, in terms of resolution, all images are relatively easy because they consist of high-resolution samples rather than other datasets.
$\hat{\mathbf{VTLP}}:$ We introduce a new challenging large-scale dataset collected in South Korea. The dataset contains 10,650 LP images, which are divided into 6,400/4,250 images for training and testing, respectively. All Korean letters are hidden for privacy protection. Images in VTLP consist of text(only 10 digits, not Korean). Compared with the public LPR datasets, our dataset has challenging factors: 1) We apply the manual annotation of large-scale images selected from unconstrained real-world, covering a variety of challenging situations using bounding box coordination; 2) Distance from vehicles to the camera is far from other dataset; 3) Various scene-texts interfere with the detection, low-resolution appearance, and very oblique LP.
## 4.2. Evaluation Metric
We follow the evaluation metric that has been widely used in LPR research [13,41]. Therefore, if only one of the consecutive characters is misclassified or not detected, it is treated as a failure case. We denote this metric as a recognition accuracy. Also, we address the 36 characters, including 26 letters and 10 digits for text recognition.
## 4.3. Implementation Details
All the reported implementations are based on the TensorFlow framework, and our method has done on one NVIDIA TITAN X GPU and one Intel Core i7-6700K CPU In all the experiments, we resize all images to $320\times320.$ For stable training, we use a gradient clipping trick and the Adam optimizer [17] with high momentum. The proposed network is trained in 1 million iterations with a batch size of 16. The weights in all SNIDER layers are initialized from a zero-mean Gaussian distribution with a standard deviation of 0.01, and the constant 0 as the biases in all layers. All models are trained for the first 100 epochs with a learning rate of 10$^{-4}$ despite higher values, and then for the remaining epochs at the learning rate of $10^{-5}.$ Batch normalization [14] and LeakyReLU [24] are used in all layers of our networks. Also, for $LPR$ network as baseline, we use the YOLO v3 detector [29] model pre-trained on ImageNet[9].
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0889892578125,
0.70166015625,
0.091552734375,
0.70166015625,
0.1060791015625,
0.50244140625,
0.10418701171875
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.521484375,
0.11614990234375,
0.89208984375,
0.11614990234375,
0.89208984375,
0.1300048828125,
0.521484375,
0.1300048828125
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50244140625,
0.1319580078125,
0.82177734375,
0.1319580078125,
0.82177734375,
0.1458740234375,
0.50244140625,
0.1458740234375
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.60302734375,
0.15283203125,
0.60302734375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.07843017578125,
0.1673583984375,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.521484375,
0.1767578125,
0.89208984375,
0.1767578125,
0.89208984375,
0.1900634765625,
0.521484375,
0.1900634765625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.1912841796875,
0.64892578125,
0.1900634765625,
0.64892578125,
0.2052001953125,
0.5009765625,
0.205810546875
],
[
0.07843017578125,
0.1988525390625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.07843017578125,
0.212158203125,
0.46728515625,
0.2108154296875,
0.46728515625,
0.2259521484375,
0.07843017578125,
0.2266845703125
],
[
0.521484375,
0.20703125,
0.890625,
0.20703125,
0.890625,
0.2203369140625,
0.521484375,
0.2203369140625
],
[
0.5009765625,
0.2203369140625,
0.89208984375,
0.2215576171875,
0.89208984375,
0.236083984375,
0.5009765625,
0.23486328125
],
[
0.07843017578125,
0.2266845703125,
0.449462890625,
0.2279052734375,
0.449462890625,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.50244140625,
0.2374267578125,
0.89208984375,
0.2374267578125,
0.89208984375,
0.250732421875,
0.50244140625,
0.250732421875
],
[
0.214111328125,
0.251953125,
0.2296142578125,
0.251953125,
0.2296142578125,
0.260009765625,
0.214111328125,
0.260009765625
],
[
0.50341796875,
0.251953125,
0.8896484375,
0.251953125,
0.8896484375,
0.26513671875,
0.50341796875,
0.26513671875
],
[
0.281982421875,
0.26318359375,
0.40283203125,
0.26318359375,
0.40283203125,
0.27783203125,
0.281982421875,
0.27783203125
],
[
0.1829833984375,
0.2568359375,
0.1895751953125,
0.2568359375,
0.1895751953125,
0.26318359375,
0.1829833984375,
0.26318359375
],
[
0.1070556640625,
0.26513671875,
0.1641845703125,
0.26513671875,
0.1641845703125,
0.276611328125,
0.1070556640625,
0.276611328125
],
[
0.2115478515625,
0.26513671875,
0.236083984375,
0.26513671875,
0.236083984375,
0.276611328125,
0.2115478515625,
0.276611328125
],
[
0.178955078125,
0.2734375,
0.191162109375,
0.2734375,
0.191162109375,
0.2822265625,
0.178955078125,
0.2822265625
],
[
0.2607421875,
0.2734375,
0.281982421875,
0.2734375,
0.281982421875,
0.27783203125,
0.2607421875,
0.27783203125
],
[
0.50244140625,
0.267578125,
0.890625,
0.267578125,
0.890625,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.19775390625,
0.281005859375,
0.2467041015625,
0.281005859375,
0.2467041015625,
0.292236328125,
0.19775390625,
0.292236328125
],
[
0.44677734375,
0.27587890625,
0.4697265625,
0.27587890625,
0.4697265625,
0.28857421875,
0.44677734375,
0.28857421875
],
[
0.5048828125,
0.2841796875,
0.890625,
0.2841796875,
0.890625,
0.294189453125,
0.5048828125,
0.294189453125
],
[
0.176513671875,
0.29541015625,
0.42724609375,
0.29541015625,
0.42724609375,
0.317626953125,
0.176513671875,
0.317626953125
],
[
0.50244140625,
0.29736328125,
0.7294921875,
0.29736328125,
0.7294921875,
0.311279296875,
0.50244140625,
0.311279296875
],
[
0.521484375,
0.31005859375,
0.890625,
0.3125,
0.890625,
0.326904296875,
0.521484375,
0.324462890625
],
[
0.08251953125,
0.324462890625,
0.1478271484375,
0.324462890625,
0.1478271484375,
0.3359375,
0.08251953125,
0.3359375
],
[
0.1396484375,
0.322509765625,
0.468017578125,
0.3212890625,
0.468017578125,
0.3359375,
0.1396484375,
0.337158203125
],
[
0.50244140625,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.341552734375,
0.50244140625,
0.341552734375
],
[
0.080078125,
0.339111328125,
0.468017578125,
0.339111328125,
0.468017578125,
0.352783203125,
0.080078125,
0.352783203125
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.08087158203125,
0.352783203125,
0.4345703125,
0.352783203125,
0.4345703125,
0.37060546875,
0.08087158203125,
0.37060546875
],
[
0.50244140625,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.50244140625,
0.371826171875
],
[
0.098876953125,
0.36865234375,
0.46728515625,
0.36865234375,
0.46728515625,
0.3818359375,
0.098876953125,
0.3818359375
],
[
0.50244140625,
0.373779296875,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.50244140625,
0.386962890625
],
[
0.080078125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.397705078125,
0.080078125,
0.397705078125
],
[
0.50244140625,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.401611328125,
0.50244140625,
0.401611328125
],
[
0.07843017578125,
0.3984375,
0.4697265625,
0.399658203125,
0.4697265625,
0.41357421875,
0.07843017578125,
0.412353515625
],
[
0.50341796875,
0.40283203125,
0.89208984375,
0.40283203125,
0.89208984375,
0.416748046875,
0.50341796875,
0.416748046875
],
[
0.080078125,
0.4140625,
0.4697265625,
0.4140625,
0.4697265625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.50244140625,
0.416748046875,
0.89208984375,
0.417236328125,
0.89208984375,
0.431884765625,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50341796875,
0.43505859375,
0.8896484375,
0.43505859375,
0.8896484375,
0.44580078125,
0.50341796875,
0.44580078125
],
[
0.08087158203125,
0.4462890625,
0.465576171875,
0.4462890625,
0.465576171875,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.08087158203125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.473388671875,
0.08087158203125,
0.473388671875
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.07843017578125,
0.473388671875,
0.4697265625,
0.474853515625,
0.4697265625,
0.4892578125,
0.07843017578125,
0.488037109375
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.080078125,
0.490478515625,
0.44287109375,
0.490478515625,
0.44287109375,
0.50390625,
0.080078125,
0.50390625
],
[
0.50244140625,
0.494384765625,
0.82177734375,
0.494384765625,
0.82177734375,
0.50830078125,
0.50244140625,
0.50830078125
],
[
0.1829833984375,
0.51318359375,
0.366943359375,
0.51318359375,
0.366943359375,
0.53076171875,
0.1829833984375,
0.53076171875
],
[
0.44677734375,
0.51513671875,
0.4697265625,
0.51513671875,
0.4697265625,
0.52783203125,
0.44677734375,
0.52783203125
],
[
0.50341796875,
0.51513671875,
0.67626953125,
0.51513671875,
0.67626953125,
0.52880859375,
0.50341796875,
0.52880859375
],
[
0.07843017578125,
0.53857421875,
0.471435546875,
0.53857421875,
0.471435546875,
0.55615234375,
0.07843017578125,
0.55615234375
],
[
0.51953125,
0.537109375,
0.890625,
0.53857421875,
0.890625,
0.55322265625,
0.51953125,
0.5517578125
],
[
0.080078125,
0.55615234375,
0.2548828125,
0.55615234375,
0.2548828125,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.5048828125,
0.55615234375,
0.890625,
0.55615234375,
0.890625,
0.56640625,
0.5048828125,
0.56640625
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.080078125,
0.57763671875,
0.25244140625,
0.57763671875,
0.25244140625,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.098876953125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.098876953125,
0.6142578125
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.07843017578125,
0.61572265625,
0.253173828125,
0.6162109375,
0.253173828125,
0.630859375,
0.07843017578125,
0.6298828125
],
[
0.50341796875,
0.6142578125,
0.85302734375,
0.6142578125,
0.85302734375,
0.62744140625,
0.50341796875,
0.62744140625
],
[
0.098876953125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65576171875,
0.098876953125,
0.65576171875
],
[
0.50341796875,
0.63623046875,
0.71728515625,
0.63623046875,
0.71728515625,
0.65087890625,
0.50341796875,
0.65087890625
],
[
0.521484375,
0.65771484375,
0.890625,
0.6591796875,
0.890625,
0.673828125,
0.521484375,
0.67236328125
],
[
0.080078125,
0.66552734375,
0.468017578125,
0.66650390625,
0.468017578125,
0.68115234375,
0.080078125,
0.6796875
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.08087158203125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.50341796875,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.080078125,
0.69775390625,
0.313720703125,
0.69775390625,
0.313720703125,
0.7109375,
0.080078125,
0.7109375
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.07843017578125,
0.71826171875,
0.1707763671875,
0.72021484375,
0.1707763671875,
0.73486328125,
0.07843017578125,
0.73291015625
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.50341796875,
0.73486328125
],
[
0.50341796875,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.098876953125,
0.7421875,
0.46728515625,
0.7412109375,
0.46728515625,
0.755859375,
0.098876953125,
0.7568359375
],
[
0.07843017578125,
0.755859375,
0.471435546875,
0.7568359375,
0.471435546875,
0.7744140625,
0.07843017578125,
0.7734375
],
[
0.50341796875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.080078125,
0.77294921875,
0.468017578125,
0.77294921875,
0.468017578125,
0.7861328125,
0.080078125,
0.7861328125
],
[
0.5048828125,
0.767578125,
0.890625,
0.767578125,
0.890625,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.787109375,
0.46728515625,
0.787109375,
0.46728515625,
0.80029296875,
0.080078125,
0.80029296875
],
[
0.50341796875,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.080078125,
0.8037109375,
0.468017578125,
0.8037109375,
0.468017578125,
0.8173828125,
0.080078125,
0.8173828125
],
[
0.5009765625,
0.81005859375,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.5009765625,
0.82373046875
],
[
0.080078125,
0.818359375,
0.4697265625,
0.818359375,
0.4697265625,
0.83349609375,
0.080078125,
0.83349609375
],
[
0.50244140625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.07843017578125,
0.833984375,
0.221435546875,
0.833984375,
0.221435546875,
0.84765625,
0.07843017578125,
0.84765625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.1446533203125,
0.87060546875,
0.400390625,
0.87255859375,
0.400390625,
0.89013671875,
0.1446533203125,
0.88818359375
],
[
0.44677734375,
0.87353515625,
0.4697265625,
0.87353515625,
0.4697265625,
0.88720703125,
0.44677734375,
0.88720703125
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.50244140625,
0.88525390625,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.50244140625,
0.8994140625
]
] |
[
"ues indicating the probability of pixels belonging to the li-",
"4.Experimental Setting",
"censeplate.Also.ground-truthlabelsfor segmentationcan",
"In this section, we describe a list of datasets, metric, and",
"be inferred from the dotted annotations by [4]'s method as",
"implementation details for the proposed method.",
"Otsu Thresholding, as shown in Figure 3. Although our seg-",
"mentation annotations by [4] do not fully reflect the actual",
"4.1.Datasets",
"detail appearance of an image, we have shown in the exper-",
"We use LPR datasets AOLP [13] and newly collected",
"iments that this auxiliary and straightforward learning strat-",
"dataset, named VTLP.",
"egy leads to effective advances in image recovery.Givena",
"pair of F and the ground-truth segmentation result in Iseg,",
"AOLP-RP : AOLP-RP [13] consists of 611 images col-",
"lected in Taiwan, including ten numbers and 25 letters (ex-",
"loss function for the Ds is the binary cross-entropy loss:",
"cept \"O\").This dataset has a challenging factor that the",
"N",
"angle of the LP contains oblique samples in terms of distor-",
"log(Ds(F)(ty))",
"1",
"CD.(w)",
">",
"N",
"TH",
"tion. On the other hand, in terms of resolution, all images",
"(T.y)EN",
"(3)",
"arerelatively easybecausetheyconsistofhigh-resolution",
"+(1- I(eu)log(1-Dsu(F)(a,u),",
"samples rather than other datasets.",
"VTLP : We introduce a new challenging large-scale",
"whereIse",
"seg, E f0, 1} is the real classes of pixels in Iseg",
"dataset collected in South Korea.The dataset contains",
"with 1 for the license plate area and O for the background",
"10.650 LP images, which are divided into 6,400/4,250 im-",
"D.(F)(r,y) denotes the pixel-wise probability by Ds.",
"ages for training and testing, respectively. All Korean letters",
"Also, we find that the generated recovery samples can-",
"are hidden for privacy protection. Images in VTLP con-",
"not usually distinguish successive texts due to close to each",
"sist of text(only 10 digits, not Korean). Compared with the",
"other.Motivated by the observations, we add a counting",
"public LPR datasets, our dataset has challenging factors:",
"decoder De, which predicts the number of characters in the",
"1) We apply the manual annotation of large-scale images",
"image. As a result, our Deplays two roles, where the first",
"selected from unconstrained real-world,covering a variety",
"istocause separation between adjacenttextsmoreclearly",
"of challenging situations using bounding box coordination;",
"The other role is to promote the encoders of each main task",
"2) Distance from vehicles to the camera is far from other",
"to generate a higher quality image while backpropagating",
"dataset; 3) Various scene-texts interfere with the detection,",
"the penalty. The loss function for the D. is the L2 loss:",
"low-resolution appearance, and very oblique LP.",
"CD。= II Cpred - CG.T II,",
"(4)",
"4.2.Evaluation Metric",
"where Cpred and CG.T are the predicted value and the",
"We follow the evaluation metric that has been widely",
"ground-truth, respectively.",
"usedinLPRresearch[13.41l.Therefore,ifonlyone of",
"the consecutive characters is misclassified or not detected,",
"3.3. Network Training",
"it is treated as a failure case. We denote this metric as a",
"The full objective function is a weighted sum of all the",
"recognition accuracy. Also, we address the 36 characters,",
"losses from Eq. (1) to (4):",
"including 26 letters and 10 digits for text recognition.",
"C=AGPCG+AGCGR+ADCD+DCDC(5)",
"4.3.Implementation Details",
"All the reported implementations are based on the Ten-",
"We employ a stage-wise training strategy to optimize main",
"sorFlow framework, and our method has done on one",
"tasks with auxiliarytasks andempirically settheweights of",
"NVIDIA TITAN X GPU and one Intel Core i7-6700K CPU.",
"each loss as detailed in Section 5.3.",
"In all the experiments, we resize all images to 320 × 320",
"3.4. Testing",
"For stable training, we use a gradient clipping trick and the",
"Adam optimizer [17] with high momentum. The proposed",
"At the testing phase, the auxiliary tasks are removed.",
"Given a low-quality test image Itest, Gp and GR output",
"network is trained in 1 million iterations with a batch size of",
"the recovered image via denoising and rectification. Then",
"16.TheweightsinallSNIDERlayersareinitialized froma",
"zero-mean Gaussian distribution with a standard deviation",
"LPR network LPR based on a YOLO v3 detector [29] by",
"of 0.0l, and the constant 0 as the biases in all layers. All",
"pre-trained on ImageNet [9] takes the recovered image and",
"models are trained for the first 100 epochs with a learning",
"generates the recognition result LPRresult of Itest, and it",
"rate of 10-4 despite higher values, and then for the remain-",
"is denoted as Eq. (6):",
"ing epochs at the learning rate of 10-5. Batch normaliza-",
"tion[14]andLeakyReLU[24areused in all layers of our",
"LPRresult = LPR(GR(GD(Itest).",
"(6)",
"networks. Also, for LPR network as baseline, we use the",
"YOLO v3 detector [29] model pre-trained on ImageNet[9]."
] |
{
"arxiv_id": "1910.03876",
"language": "en",
"page_idx": 4
}
|
||
2307.08504-en-0
|
2307.08504-en-0.jpg
|
## $\mathbf{BUS:Efficient~and~Effective~Vision-language~Pre-training~with~Bottom-Up}$ Patch Summarization.
Chaoya Jiang$^{1}$, Haiyang Xu$^{2*}$; Wei Ye$^{1*}$,Qinghao Ye$^2$, Chenliang Li$^2$, Ming Yan$^2$, Bin Bi$^2$
Shikun Zhang$^1$,Fei Huang$^2$,Songfang Huang$^2$
$^{1}$National Engineering Research Center for Software Engineering, Peking University
$^{2}$DAMO Academy, Alibaba Group
shuofeng.xhy@alibaba-inc.com, wye@pku.edu.cn
### $\mathbf{Abstract}$
Vision Transformer (ViT) based Vision-Language Pretraining (VLP) models have demonstrated impressive performance in various tasks. However, the lengthy visual to-
$\sum_{\begin{array}{lll}&\text{ken sequences fed into ViT can lead to training inefficiency}\\&\text{and ineffectiveness. Existing efforts address the challenge}\\&by\text{ either bottom-level patch extraction in the ViT backbone}\end{array}}$ $\begin{array}{ccc}\dot{o}&or\:top-level\:patch\:abstraction\:outside,\:not\:balancing\:train-\\\dot{O}&ing\:efficiency\:and\:effectiveness\:well.\:Inspired\:by\:text\:sum-\end{array}$
marization in natural language processing, we propose a Bottom-Up Patch Summarization approach named BUS , coordinating bottom-level extraction and top-level abstraction to learn a concise summary of lengthy visual token sequences efficiently. Specifically, We incorporate a TextSemantics-Aware Patch Selector (TSPS) into the ViT backbone to perform a coarse-grained visual token extraction and then attach a flexible Transformer- based Patch Abstraction Decoder (PAD) upon the backbone for top-level visual abstraction. This bottom- up collaboration enables our BUS to yield high training efficiency while maintaining or even improving effectiveness. We evaluate our approach on various visual-language understanding and generation tasks and show competitive downstream task performance while boosting the training efficiency by 50%. Additionally, our model achieves state-of-the-art performance on many downstream tasks by increasing input image resolu- $tionwithoutincreasingcomputationalcostsoverbaselines.$
<FigureHere>
Figure 1: Subfigure (a) shows the examples of selected textrelevant patch in the VQA scenario. Subfigure(b) shows the overview of our proposed bottom-up patch summarization.
large number of image-text pairs by aligning the visual and
linguistic modalities.
Most recent works [14,52,29,23,11] adopt ViT as the visual encoder or cross-modal fusion encoder, due to its excellent ability to model the fne-grained long visual sequences from the image grids or patches.
Despite the impressive progress of ViT-based VLP models, they still face challenges of training inefficiency and ineffectiveness caused by lengthy visual token sequences. Firstly, long visual sequences will bring heavy self-attention calculation for visual representation modeling and crossmodal fusion, leading to time-consuming training. Secondly, long visual sequences contain many redundant patches irrelevant to the text semantics. For instance, as illustrated in Figure 1 (a), during the VQA task, when answering the question “What is the color of the buildings in the far left of the photo?", about 80% of the image patches may be irrelevant with the question. On the one hand, those text-irrelevant patches (e.g., the yellow building in the im-
## $1. \textbf{Introduction}$
Large-scale pre-training of vision-language models has recently received tremendous success on a wide range of cross-modal tasks [48,7,14,31,59,29,52]. Such visionlanguage models learn cross-modal representations from a
$^*$ Corresponding Author.
1
|
[
[
0.1094970703125,
0.1331787109375,
0.859375,
0.1331787109375,
0.859375,
0.15087890625,
0.1094970703125,
0.15087890625
],
[
0.374267578125,
0.1572265625,
0.5986328125,
0.1572265625,
0.5986328125,
0.1717529296875,
0.374267578125,
0.1717529296875
],
[
0.1396484375,
0.2052001953125,
0.83251953125,
0.2052001953125,
0.83251953125,
0.2197265625,
0.1396484375,
0.2197265625
],
[
0.301513671875,
0.222900390625,
0.6708984375,
0.222900390625,
0.6708984375,
0.2374267578125,
0.301513671875,
0.2374267578125
],
[
0.1544189453125,
0.2392578125,
0.81640625,
0.2412109375,
0.81640625,
0.255615234375,
0.1544189453125,
0.25390625
],
[
0.346435546875,
0.255615234375,
0.625,
0.2568359375,
0.625,
0.274658203125,
0.346435546875,
0.2734375
],
[
0.0236968994140625,
0.26318359375,
0.056365966796875,
0.26318359375,
0.056365966796875,
0.397705078125,
0.0236968994140625,
0.397705078125
],
[
0.28759765625,
0.276611328125,
0.6845703125,
0.276611328125,
0.6845703125,
0.291748046875,
0.28759765625,
0.291748046875
],
[
0.2344970703125,
0.324462890625,
0.313720703125,
0.324462890625,
0.313720703125,
0.34228515625,
0.2344970703125,
0.34228515625
],
[
0.101318359375,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.375,
0.101318359375,
0.375
],
[
0.85107421875,
0.356689453125,
0.861328125,
0.356689453125,
0.861328125,
0.35986328125,
0.85107421875,
0.35986328125
],
[
0.08087158203125,
0.375732421875,
0.46728515625,
0.375732421875,
0.46728515625,
0.389404296875,
0.08087158203125,
0.389404296875
],
[
0.08087158203125,
0.39013671875,
0.468017578125,
0.39013671875,
0.468017578125,
0.404052734375,
0.08087158203125,
0.404052734375
],
[
0.0228729248046875,
0.404052734375,
0.053924560546875,
0.40283203125,
0.056365966796875,
0.449462890625,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.406005859375,
0.468017578125,
0.406005859375,
0.468017578125,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.080078125,
0.419189453125,
0.468017578125,
0.42041015625,
0.468017578125,
0.43505859375,
0.080078125,
0.434326171875
],
[
0.54736328125,
0.427978515625,
0.58837890625,
0.427978515625,
0.58837890625,
0.43505859375,
0.54736328125,
0.43505859375
],
[
0.7099609375,
0.4287109375,
0.73388671875,
0.4287109375,
0.73388671875,
0.43310546875,
0.7099609375,
0.43310546875
],
[
0.7939453125,
0.4287109375,
0.826171875,
0.4287109375,
0.826171875,
0.43310546875,
0.7939453125,
0.43310546875
],
[
0.0318603515625,
0.443115234375,
0.056365966796875,
0.443115234375,
0.056365966796875,
0.479736328125,
0.0318603515625,
0.479736328125
],
[
0.08251953125,
0.4375,
0.46728515625,
0.4375,
0.46728515625,
0.447509765625,
0.08251953125,
0.447509765625
],
[
0.07843017578125,
0.451904296875,
0.46728515625,
0.45068359375,
0.46728515625,
0.464111328125,
0.07843017578125,
0.46533203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.650390625,
0.47021484375,
0.65673828125,
0.47021484375,
0.65673828125,
0.473388671875,
0.650390625,
0.473388671875
],
[
0.7861328125,
0.4697265625,
0.8310546875,
0.4697265625,
0.8310546875,
0.473388671875,
0.7861328125,
0.473388671875
],
[
0.080078125,
0.481689453125,
0.468017578125,
0.481689453125,
0.468017578125,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.78125,
0.4765625,
0.833984375,
0.4765625,
0.833984375,
0.48095703125,
0.78125,
0.48095703125
],
[
0.02532958984375,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.5888671875,
0.02532958984375,
0.5888671875
],
[
0.080078125,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51123046875,
0.080078125,
0.51123046875
],
[
0.8310546875,
0.498779296875,
0.85888671875,
0.498779296875,
0.85888671875,
0.50244140625,
0.8310546875,
0.50244140625
],
[
0.59326171875,
0.505859375,
0.61767578125,
0.505859375,
0.61767578125,
0.5234375,
0.59326171875,
0.5234375
],
[
0.833984375,
0.50244140625,
0.84228515625,
0.50244140625,
0.84228515625,
0.5068359375,
0.833984375,
0.5068359375
],
[
0.08087158203125,
0.51220703125,
0.468017578125,
0.51220703125,
0.468017578125,
0.52587890625,
0.08087158203125,
0.52587890625
],
[
0.080078125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.080078125,
0.54052734375
],
[
0.689453125,
0.5322265625,
0.69775390625,
0.5322265625,
0.69775390625,
0.53515625,
0.689453125,
0.53515625
],
[
0.08087158203125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.5546875,
0.08087158203125,
0.5546875
],
[
0.5009765625,
0.53857421875,
0.8896484375,
0.5390625,
0.8896484375,
0.5537109375,
0.5009765625,
0.55322265625
],
[
0.037567138671875,
0.55322265625,
0.047393798828125,
0.5625,
0.037567138671875,
0.568359375,
0.02777099609375,
0.5595703125
],
[
0.08087158203125,
0.55810546875,
0.46728515625,
0.55810546875,
0.46728515625,
0.568359375,
0.08087158203125,
0.568359375
],
[
0.5009765625,
0.5546875,
0.89208984375,
0.5537109375,
0.89208984375,
0.568359375,
0.5009765625,
0.5693359375
],
[
0.080078125,
0.5712890625,
0.4697265625,
0.5712890625,
0.4697265625,
0.5859375,
0.080078125,
0.5859375
],
[
0.50244140625,
0.57080078125,
0.88818359375,
0.57080078125,
0.88818359375,
0.583984375,
0.50244140625,
0.583984375
],
[
0.0318603515625,
0.59326171875,
0.04412841796875,
0.59326171875,
0.04412841796875,
0.60498046875,
0.0318603515625,
0.60498046875
],
[
0.08251953125,
0.58837890625,
0.46728515625,
0.58837890625,
0.46728515625,
0.5986328125,
0.08251953125,
0.5986328125
],
[
0.080078125,
0.6015625,
0.46728515625,
0.6015625,
0.46728515625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.0294189453125,
0.6123046875,
0.052276611328125,
0.6123046875,
0.052276611328125,
0.6318359375,
0.0294189453125,
0.6318359375
],
[
0.50244140625,
0.609375,
0.89208984375,
0.60986328125,
0.89208984375,
0.625,
0.50244140625,
0.62353515625
],
[
0.08087158203125,
0.6171875,
0.468017578125,
0.6171875,
0.468017578125,
0.630859375,
0.08087158203125,
0.630859375
],
[
0.0294189453125,
0.62890625,
0.052276611328125,
0.62890625,
0.052276611328125,
0.64306640625,
0.0294189453125,
0.64306640625
],
[
0.080078125,
0.630859375,
0.4697265625,
0.6318359375,
0.4697265625,
0.646484375,
0.080078125,
0.64599609375
],
[
0.50341796875,
0.62744140625,
0.64208984375,
0.62744140625,
0.64208984375,
0.63818359375,
0.50341796875,
0.63818359375
],
[
0.035125732421875,
0.642578125,
0.051483154296875,
0.642578125,
0.051483154296875,
0.6640625,
0.035125732421875,
0.6640625
],
[
0.080078125,
0.64794921875,
0.471435546875,
0.64794921875,
0.471435546875,
0.6611328125,
0.080078125,
0.6611328125
],
[
0.521484375,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.521484375,
0.65576171875
],
[
0.080078125,
0.66357421875,
0.471435546875,
0.66357421875,
0.471435546875,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.080078125,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.69140625,
0.080078125,
0.69140625
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.035125732421875,
0.68310546875,
0.0555419921875,
0.68310546875,
0.0555419921875,
0.70703125,
0.035125732421875,
0.70703125
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50244140625,
0.689453125,
0.775390625,
0.689453125,
0.775390625,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.08087158203125,
0.70849609375,
0.4697265625,
0.70849609375,
0.4697265625,
0.7216796875,
0.08087158203125,
0.7216796875
],
[
0.51953125,
0.705078125,
0.890625,
0.7041015625,
0.890625,
0.71826171875,
0.51953125,
0.7197265625
],
[
0.080078125,
0.72265625,
0.468017578125,
0.72265625,
0.468017578125,
0.736328125,
0.080078125,
0.736328125
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.080078125,
0.73779296875,
0.468017578125,
0.73779296875,
0.468017578125,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.50341796875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.5048828125,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.08087158203125,
0.78857421875,
0.20751953125,
0.78857421875,
0.20751953125,
0.80322265625,
0.08087158203125,
0.80322265625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.098876953125,
0.814453125,
0.4697265625,
0.814453125,
0.4697265625,
0.82763671875,
0.098876953125,
0.82763671875
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.830078125,
0.471435546875,
0.830078125,
0.471435546875,
0.84326171875,
0.08087158203125,
0.84326171875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08251953125,
0.8466796875,
0.468017578125,
0.8466796875,
0.468017578125,
0.85693359375,
0.08251953125,
0.85693359375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85986328125,
0.4697265625,
0.859375,
0.4697265625,
0.87353515625,
0.07843017578125,
0.875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.098876953125,
0.88623046875,
0.233642578125,
0.88720703125,
0.233642578125,
0.90087890625,
0.098876953125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"BUS : Efficient and Effective Vision-language Pre-training with Bottom-Up",
"Patch Summarization.",
"Chaoya Jiangl, Haiyang Xu?* Wei Yel* Qinghao Ye?, Chenliang Li?, Ming Yan?, Bin Bi?",
"Shikun Zhangl, Fei Huang?, Songfang Huang?",
"1National Engineering Research Center for Software Engineering, Peking University",
"2DAMO Academy, Alibaba Group",
"",
"shuofeng.xhy@alibaba-inc.com, wye@pku.edu.cn",
"Abstract",
"Vision Transformer (ViT) based Vision-Language Pre-",
"",
"training (VLP) models have demonstrated impressive per-",
"formance in various tasks. However, the lengthy visual to-",
"LA",
"ken sequences fed into ViT can lead to training inefficiency",
"and ineffectiveness.Existing efforts address the challenge",
"Origin Image",
"afPatel",
"Text-reles",
"S",
"byeitherbottom-levelpatchextractionintheViTbackbone",
"or top-level patch abstraction outside, not balancing train-",
"ing efficiency and effectiveness well. Inspired by text sum-",
"K",
"etided",
"marization in natural language processing, we propose a",
"PatehNhsirictid",
"",
"Bottom-Up Patch Summarization approach named BUS,",
"lisualPat",
"",
"",
"coordinating bottom-level extraction and top-level abstrac-",
"tion to learn a concise summary of lengthy visual token",
"",
"sequences efficiently Specifically We incorporate a Text-",
"Figure l: Subfigure (a) shows the examples of selected text-",
"",
"Semantics-AwarePatchSelector(TSPS)intotheViTback-",
"relevant patch in the VQA scenario. Subfigure (b) shows the",
"bone to perform a coarse-grained visual token extraction",
"overview of our proposed bottom-up patch summarization.",
"C",
"andthenattachaflexibleTransformer-basedPatchAbstrac-",
"tion Decoder (PAD) upon the backbone for top-level vi-",
"3",
"large number of image-text pairs by aligning the visual and",
"sual abstraction. This bottom-up collaboration enables our",
"2",
"BUS to yield high training efficiency while maintaining or",
"linguisticmodalities.",
".A",
"even improving effectiveness.We evaluate ourapproach",
"Most recent works [14, 52, 29, 23, 11] adopt ViT as the",
"on various visual-language understanding and generation",
"visual encoder or cross-modal fusion encoder, due to its",
"tasks and show competitive downstream task performance",
"excellent ability to model the fine-grained long visual se-",
"",
"while boosting the training eficiency by 50%. Addition-",
"quences from the image grids or patches.",
"ally, our model achieves state-of-the-art performance on",
"Despite the impressive progress of ViT-based VLP mod-",
"many downstream tasks by increasing input image resolu-",
"els, they still face challenges of training inefficiency and",
"tion without increasing computational costs over baselines.",
"ineffectiveness causedbylengthyvisualtoken sequences.",
"Firstly, long visual sequences will bring heavy self-attention",
"calculation for visualrepresentation modeling and cross-",
"modal fusion, leading to time-consuming training. Sec-",
"1.Introduction",
"ondly, long visual sequences contain many redundant",
"Large-scale pre-training of vision-language models has",
"patches irrelevant to the text semantics. For instance, as",
"recently received tremendous success on a wide range of",
"illustrated in Figure 1 (a), during the VQA task, when an-",
"cross-modaltasks[48,7.14.31.59,29,521.Suchvision-",
"swering the question *What is the color of the buildings in",
"language models learn cross-modal representations from a",
"thefarleftofthephoto?about80%of theimagepatches",
"may be irrelevant with the question. On the one hand, those",
"$ Corresponding Author.",
"text-irrelevant patches (e.g., the yellow building in the im-",
"-"
] |
{
"arxiv_id": "2307.08504",
"language": "en",
"page_idx": 0
}
|
||
2307.08504-en-1
|
2307.08504-en-1.jpg
|
tion and fine-grained patch-text alignment. Next, we introduce a lightweight Transformer-based Patch Abstraction Decoder (PAD) for top-level abstraction. It takes the top-K text-relevant patch tokens from the output sequence of ViTbackbone as the input and the overall visual sequence as the encoder hidden states to generate the final visual patch summary.
We evaluate BUS on various representative VL understanding and generation tasks, including visual question answering, cross-modal retrieval, and image captioning. We find that by reducing the length of the patch sequence to $20\%$ of its original length, we can not only get competitive or better downstream task performance but also enjoy a significant increase in efficiency over previous similar VLP models. For instance, BUS reduces about 51% of the inference time (see Table 7) and even improves by about 0.3 on the VQA test-dev with the same experimental settings. Furthermore, by increasing the input image resolution, BUS achieves state-of-the-art downstream task performance (e.g., 78.28 on VQA test-dev) which benefits from processing more image tokens without increasing computational costs.
age) will hinder the fine-grained alignment between the textual and visual modalities. On the other hand, they will lead to the overshadowing of brief linguistic signals (e.g., of short image captions) by complex visual ones during the cross-modal fusion, namely the“vanishing information” problem of textual information [26].
The limitations above underscore the importance of reducing visual token sequences. Recent related efforts can be categorized into two lines.
$\cdot \textbf{Top- level Abstraction. }$ The frst line tackles the issue outside the ViT-based visual backbones from a top-level perspective [27, 3]. Specifically, these works use a fixed number of learnable latent query vectors to query the long visual sequences output, obtaining the final fixed-length visual sequence representations, in an abstractive way. An obvious bottleneck is that they can not optimize the costly and potentially unnecessary self-attention calculation in the visual backbones. Meanwhile, this visual representation abstraction process only considers the semantics of the visual modality, ignoring the textual guidance and consequently leading to representation deficiency.
$\cdot \textbf{Bottom- level Extraction. The second line focuses on}$ reducing patch tokens in the ViT-backbone from the bottom-level perspective, usually in an extractive manner [42,32,17]. The problem here lies in that overly reducing the visual sequence by extracting critical tokens in the backbone, while accelerating the attention calculation, may deconstruct images' structural information. Therefore, balancing efficiency and effectiveness remains a bottleneck.
To achieve a better trade-off between the efficiency and effectiveness of VLP, we propose integrating the merits of top-level abstraction and bottom-level extraction. Inspired by bottom-up text summarization[4,20],which first select key phrases and then abstractively generate the final text summaries, we design a bottom-up summarization process for visual tokens. We fırst exploit coarse-grained key patch extraction in the ViT backbone, with regulation from text modality, to identify text-relevant tokens and remove potentially redundant ones, reducing the computational cost in the ViT backbone. Then fine-grained text-guided patch abstraction is performed upon the output sequence of the ViT backbone to obtain a further condensed visual representation sequence.
Specifically, we incorporate a Text Semantic-aware Patch Selector (TSPS) module into the ViT-based backbone for bottom-level extraction. We transform object/region annotations to patch-level annotations to train an effective extractor with a novel auxiliary pre-training task named Patch-Text Matching (PTM), which facilitates patch extrac-
### $2. \textbf{Related Work}$
# 2.1. Vision- Language Pre-training
Previous research on vision language pre-training has primarily fallen into two categories: Detector-based VLP models and CNN/ViT-based VLP models. Detector-based methods, such as [35, 30, 48, 31, 7, 59, 55, 58], use a twostep training pipeline that first extracts visual features using a pre-trained object detector and then aligns text and visual features using a cross-modal pre-training model. While some region-based methods, such as [50], use lightweight model architectures to reduce computation costs, they still face expensive computational and time-consuming object detection. Recently, Vision Transformer (ViT) based methods, such as [29, 23, 41, 51, 28, 26, 52, 23, 18, 19, 17] have emerged as a promising alternative to detector-based approaches. ViT-based models eliminate the need for object detectors in feature extraction, enabling end-to-end vision language learning. However, they struggle with lengthy visual token sequences and lack fine-grained cross-modal alignment information. These long visual sequences also increase computation costs and introduce noise visual information for cross-modal fusion. To address these challenges, we propose a Bottom-Up Patch Summarization approach that coordinates bottom-level extraction and top-level abstraction to efficiently learn a concise summary of lengthy visual token sequences.
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.080078125,
0.11932373046875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.5048828125,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.080078125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.46728515625,
0.15283203125,
0.46728515625,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08087158203125,
0.1685791015625,
0.313720703125,
0.1685791015625,
0.313720703125,
0.1787109375,
0.08087158203125,
0.1787109375
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.098876953125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.098876953125,
0.1962890625
],
[
0.5,
0.18115234375,
0.56884765625,
0.18310546875,
0.56884765625,
0.1976318359375,
0.5,
0.1956787109375
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.521484375,
0.205810546875,
0.89208984375,
0.205810546875,
0.89208984375,
0.2203369140625,
0.521484375,
0.2203369140625
],
[
0.07843017578125,
0.2139892578125,
0.275390625,
0.2139892578125,
0.275390625,
0.2279052734375,
0.07843017578125,
0.2279052734375
],
[
0.50244140625,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.1004638671875,
0.2412109375,
0.468017578125,
0.2412109375,
0.468017578125,
0.255126953125,
0.1004638671875,
0.255126953125
],
[
0.50244140625,
0.2374267578125,
0.89208984375,
0.2374267578125,
0.89208984375,
0.250732421875,
0.50244140625,
0.250732421875
],
[
0.111083984375,
0.2568359375,
0.471435546875,
0.2568359375,
0.471435546875,
0.270751953125,
0.111083984375,
0.270751953125
],
[
0.50244140625,
0.251953125,
0.89208984375,
0.251953125,
0.89208984375,
0.26513671875,
0.50244140625,
0.26513671875
],
[
0.50244140625,
0.266357421875,
0.89208984375,
0.266357421875,
0.89208984375,
0.27978515625,
0.50244140625,
0.27978515625
],
[
0.111083984375,
0.272705078125,
0.4697265625,
0.272705078125,
0.4697265625,
0.285888671875,
0.111083984375,
0.285888671875
],
[
0.1094970703125,
0.285888671875,
0.471435546875,
0.287353515625,
0.471435546875,
0.3017578125,
0.1094970703125,
0.300537109375
],
[
0.50244140625,
0.28271484375,
0.89208984375,
0.28271484375,
0.89208984375,
0.296630859375,
0.50244140625,
0.296630859375
],
[
0.111083984375,
0.3017578125,
0.468017578125,
0.3017578125,
0.468017578125,
0.31494140625,
0.111083984375,
0.31494140625
],
[
0.5009765625,
0.29736328125,
0.89208984375,
0.296630859375,
0.89208984375,
0.311279296875,
0.5009765625,
0.3125
],
[
0.11273193359375,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.11273193359375,
0.33203125
],
[
0.50341796875,
0.3125,
0.89208984375,
0.3125,
0.89208984375,
0.32568359375,
0.50341796875,
0.32568359375
],
[
0.11273193359375,
0.333251953125,
0.4697265625,
0.333251953125,
0.4697265625,
0.3466796875,
0.11273193359375,
0.3466796875
],
[
0.50244140625,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.341552734375,
0.50244140625,
0.341552734375
],
[
0.111083984375,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.111083984375,
0.3623046875
],
[
0.50244140625,
0.343505859375,
0.890625,
0.343505859375,
0.890625,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.111083984375,
0.3623046875,
0.46728515625,
0.3623046875,
0.46728515625,
0.375732421875,
0.111083984375,
0.375732421875
],
[
0.50244140625,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.50244140625,
0.371826171875
],
[
0.50244140625,
0.37255859375,
0.89208984375,
0.37255859375,
0.89208984375,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.111083984375,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.111083984375,
0.392578125
],
[
0.50244140625,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.111083984375,
0.39453125,
0.465576171875,
0.39453125,
0.465576171875,
0.40478515625,
0.111083984375,
0.40478515625
],
[
0.111083984375,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.111083984375,
0.42236328125
],
[
0.50244140625,
0.404052734375,
0.89208984375,
0.404052734375,
0.89208984375,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.111083984375,
0.423583984375,
0.4150390625,
0.423583984375,
0.4150390625,
0.4375,
0.111083984375,
0.4375
],
[
0.50244140625,
0.419189453125,
0.583984375,
0.419189453125,
0.583984375,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.098876953125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.46337890625,
0.098876953125,
0.46337890625
],
[
0.111083984375,
0.46533203125,
0.4697265625,
0.46533203125,
0.4697265625,
0.479248046875,
0.111083984375,
0.479248046875
],
[
0.50244140625,
0.473388671875,
0.640625,
0.473388671875,
0.640625,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.111083984375,
0.48095703125,
0.468017578125,
0.48095703125,
0.468017578125,
0.494384765625,
0.111083984375,
0.494384765625
],
[
0.111083984375,
0.49560546875,
0.468017578125,
0.49560546875,
0.468017578125,
0.5087890625,
0.111083984375,
0.5087890625
],
[
0.50244140625,
0.505859375,
0.7646484375,
0.5068359375,
0.7646484375,
0.521484375,
0.50244140625,
0.52099609375
],
[
0.11358642578125,
0.51220703125,
0.465576171875,
0.51220703125,
0.465576171875,
0.52294921875,
0.11358642578125,
0.52294921875
],
[
0.1094970703125,
0.52587890625,
0.468017578125,
0.52587890625,
0.468017578125,
0.5390625,
0.1094970703125,
0.5390625
],
[
0.11358642578125,
0.54248046875,
0.468017578125,
0.54248046875,
0.468017578125,
0.55322265625,
0.11358642578125,
0.55322265625
],
[
0.521484375,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.521484375,
0.55322265625
],
[
0.111083984375,
0.55615234375,
0.468017578125,
0.55615234375,
0.468017578125,
0.5693359375,
0.111083984375,
0.5693359375
],
[
0.5009765625,
0.5537109375,
0.89208984375,
0.55322265625,
0.89208984375,
0.5673828125,
0.5009765625,
0.568359375
],
[
0.111083984375,
0.57080078125,
0.285888671875,
0.57080078125,
0.285888671875,
0.583984375,
0.111083984375,
0.583984375
],
[
0.50244140625,
0.5693359375,
0.890625,
0.5693359375,
0.890625,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.50244140625,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.098876953125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.098876953125,
0.6123046875
],
[
0.5009765625,
0.599609375,
0.89208984375,
0.5986328125,
0.89208984375,
0.61279296875,
0.5009765625,
0.6142578125
],
[
0.07843017578125,
0.61279296875,
0.471435546875,
0.6142578125,
0.471435546875,
0.62890625,
0.07843017578125,
0.62744140625
],
[
0.5009765625,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.5009765625,
0.62890625
],
[
0.080078125,
0.6298828125,
0.471435546875,
0.6298828125,
0.471435546875,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.5009765625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.5009765625,
0.64453125
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.5009765625,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.5009765625,
0.6591796875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.70556640625,
0.471435546875,
0.70556640625,
0.471435546875,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.72265625,
0.468017578125,
0.72265625,
0.468017578125,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.471435546875,
0.73486328125,
0.471435546875,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.07843017578125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.07843017578125,
0.7646484375
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.5009765625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.5009765625,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.79736328125,
0.2091064453125,
0.79736328125,
0.2091064453125,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.50244140625,
0.796875,
0.8935546875,
0.796875,
0.8935546875,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.09722900390625,
0.81005859375,
0.4697265625,
0.8095703125,
0.4697265625,
0.82373046875,
0.09722900390625,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.5,
0.82568359375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.5,
0.8388671875
],
[
0.07843017578125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.5,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.66015625,
0.88720703125,
0.66015625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"age) will hinder the fine-grained alignment between the tex-",
"tion and fine-grained patch-text alignment. Next, we in-",
"tual and visual modalities.On the other hand. they will",
"troducealightweightTransformer-basedPatchAbstraction",
"lead to the overshadowing of brief linguistic signals (e.g.,",
"Decoder(PAD)fortop-levelabstraction.Ittakesthetop-K",
"of short image captions) by complex visual ones during",
"text-relevant patch tokens from the output sequence of ViT-",
"the cross-modal fusion, namely the \"vanishing information\"",
"backbone as the input and the overall visual sequence as",
"problemoftextualinformation[26]",
"the encoder hidden states to generate the final visual patch",
"The limitations above underscore the importance of re-",
"summary.",
"ducing visual token sequences. Recent related efforts can",
"We evaluate BUS on various representative VL under-",
"becategorized into two lines.",
"standing and generation tasks, including visual question an-",
". Top-level Abstraction. The first line tackles the is-",
"swering, cross-modal retrieval, and image captioning. We",
"sue outside the ViT-based visual backbones from a",
"find that by reducing the length of the patch sequence to",
"2o% of its original length, we can not only get compet-",
"top-level perspective [27, 3]. Specifically, these works",
"use a fixed number of learnable latent query vectors",
"itive or better downstream task performance but also en-",
"to query the long visual sequences output, obtaining",
"joy a significant increase in efficiency over previous similar",
"the final fixed-length visual sequence representations,",
"VLP models. For instance, BUS reduces about 51% of the",
"in an abstractive way. An obvious bottleneck is that",
"inference time (see Table 7)and even improves by about",
"they can not optimize the costly and potentially un-",
"0.3 on the VQA test-dev with the same experimental set-",
"necessary self-attention calculation in the visual back-",
"tings. Furthermore, by increasing the input image resolu-",
"tion, BUS achieves state-of-the-art downstream task perfor-",
"bones. Meanwhile, this visual representation abstrac-",
"mance (e.g., 78.28 on VQA test-dev) which benefits from",
"tion process onlyconsidersthesemantics of thevi",
" sual modality, ignoring the textual guidance and con-",
"processing more image tokens without increasing computa-",
" sequently leading to representation deficiency.",
"tional costs.",
".Bottom-level Extraction. The second line focuses on",
"reducing patch tokens in the ViT-backbone from the",
"2. Related Work",
"bottom-level perspective, usually in an extractive man-",
"ner [42, 32, 17] . The problem here lies in that overly",
"2.1.Vision-LanguagePre-training",
"reducingthe visual sequence by extracting critical to-",
"kens in the backbone, while accelerating the attention",
"calculation,may deconstructimages'structuralinfor-",
"Previous research on vision language pre-training has",
"mation. Therefore, balancing efficiency and effective-",
"primarily fallen into two categories: Detector-based VLP",
"ness remains a bottleneck.",
"models and CNN/ViT-based VLP models. Detector-based",
"methods, such as [35, 30, 48, 31, 7, 59, 55, 58], use a two-",
"To achieve a better trade-off between the efficiency and",
"step training pipeline that first extracts visual features us-",
"effectiveness of VLP, we propose integrating the merits of",
"ing a pre-trained object detector and then aligns text and vi-",
"top-level abstraction and bottom-level extraction. Inspired",
"sual features using a cross-modal pre-training model. While",
"by bottom-up text summarization[4, 20], which first select",
"some region-based methods, such as [50], use lightweight",
"key phrases and then abstractively generate the final text",
"model architectures to reduce computation costs, they still",
"summaries, we design a bottom-up summarization process",
"face expensive computational and time-consuming object",
"for visual tokens. We first exploit coarse-grained key patch",
"detection. Recently, Vision Transformer (ViT) based meth-",
"extraction in the ViT backbone, with regulation from text",
"ods, such as [29,23, 41, 51, 28, 26, 52, 23, 18, 19, 17] have",
"modality,to identify text-relevant tokens and remove po-",
"emerged as a promising alternative to detector-based ap-",
"tentially redundant ones, reducing the computational cost",
"proaches. ViT-based models eliminate the need for object",
"in the ViT backbone. Then fine-grained text-guided patch",
"detectors infeature extraction,enabling end-to-end vision",
"abstractionis performed upon the output sequence of the",
"language learning. However, they struggle with lengthy",
"ViT backbone to obtain a further condensed visual repre-",
"visual token sequences and lack fine-grained cross-modal",
"sentation sequence.",
"alignment information. These long visual sequences also",
"Specifically, we incorporate a Text Semantic-aware",
"increase computation costs and introduce noise visual infor-",
"Patch Selector (TSPS) module into the ViT-based back-",
"mation for cross-modal fusion.To address these challenges,",
"bone for bottom-level extraction. We transform object/re-",
"we propose a Bottom-Up Patch Summarization approach",
"gion annotations topatch-levelannotationstotrain aneffec-",
"that coordinates bottom-level extraction and top-level ab-",
"tive extractor with a novel auxiliary pre-training task named",
"straction to efficiently learn a concise summary of lengthy",
"Patch-Text Matching (PTM), which facilitates patch extrac-",
"visual token sequences."
] |
{
"arxiv_id": "2307.08504",
"language": "en",
"page_idx": 1
}
|
||
2307.08504-en-3
|
2307.08504-en-3.jpg
|
final visual summary of the long patch sequence, which can be denoted as $\hat{V}=\{\hat{v}_{cls},\hat{v}_{1},\tilde{v}_{2},\cdots,\hat{v}_{s}\}$,where $s\ll u.$ The image and text representations are concatenated and fed into the cross-modal encoder. We obtain the cross-modal representations $\{c_{cls},c_{1},c_{2},\ldots,c_{l}\}$,where $l=s+m.$ The cross-modal representations can be used to fine-tune downstream multi-modal understanding tasks. Additionally, the output cross-modal representations $\{c_{cls},c_{1},c_{2},\ldots,c_{l}\}$ of the multi-modal encoder are fed into a Transformer decoder for sequence-to-sequence learning
## 3.2. Bottom-Up Patch Summarization
The proposed Bottom-Up Patch Summarization consists of two steps: Key Patch Extraction (KPE) and Text-Guided Patch Abstraction (TPA). KPE is conducted within the visual backbone to select a coarse-grained subset of textrelevant patch tokens. Then, outside the backbone, TextGuided Patch Abstraction is conducted to abstract the selected tokens and obtain a more prominent visual summary.
## 3.2.1 Key Patch Extraction
Figure 2(b) shows how we perform Key Patch Extraction (KPE) in the ViT-based visual backbone based on the Text Semantic-aware Patch Selector (TSPS). Similar to the text extractive summarization methods[53,64],we view the KPE as a patch classification task that aims to classify whether a patch is aligned with the text semantic and should be selected. Specially, suppose the TSPS is plugged between the $k_{th}(1\leq k<N)$ Transformer layer and $(k+1)_{th}$ Transformer layer and the output patch sequence features of $k_{th}$ Transformer layer is $v^k=\{\hat{v}_{cls}^{k},v_{1}^{k},\cdots,v_{n}^{k}\}.$ The text [CLS] feature is output by the text encoder and represents global information of the input text $T.$ We will frst concatenate the text [CLS] feature with each image patch token as follows:
$$\dot{v}_{i}^{k}=concat(v_{i}^{k},t_{cls})$$
where
$$^{k}\:\in\:R^{d},t_{cls}\:\in\:R^{d},\dot{v}_{i}^{k}\:\in\:R^{2d},i\:\in\:\{1,2,\ldots,n$$
Then the concatenated patch features $\dot{v}_i^k$ are fed to TSPS, which is a Multi-layer Perceptron (MLP) that contains three linear layers and is used to predict the alignment score between patches and the input text $T.$ The first two linear layers will linearly project the concatenated patch features $\{\dot{v}_{i}^{k}\}$ to the hidden representations $\{h_i^{k}\}$ and then the hidden representations $\{h_{i}^{k}\}$ is fed to the last linear layer denoted as $\mathbf{F}_\theta$ which can be seen as a classifier to predict whether the patches are relevant to the input text. The output of the last linear layer has only one dimension and will be fed to a Sigmoid activation function. Formally, the alignment score $a_i$ between the $i_th$ image patch and input text T can be calculated as follow:
$$a_{i}=Sigmoid(\mathbf{F}_{\theta}(h_{i}^{k})),i\in\{1,2,\ldots,n\}$$
To learn an effective patch selector, we generate patch-level training labels by transforming the bounding box annotations (as described in Appendix A.2). However, this approach introduces bias, as illustrated in Figure 2(b), where TSPS predicts all patches within the bounding box with a high alignment score, even if some of them are irrelevant. To address this issue, we incorporate the attention map of the image [CLS] token into the selection process. Previous work has shown that in ViTs, the image [CLS] token pays more attention to class-specific tokens than to tokens in non-object regions [6]. By weighting the alignment score $a$ and the attention map $p$ of the image [CLS] token to other tokens, we can highlight the patches corresponding to objects in the detection box and suppress other patches. This approach reduces the bias introduced by the bounding box labels and can be formulated as follows:
$$\dot{a}_{i}=\beta*\mathbf{F}_{N}\left(a_{i}\right)+\left(1-\beta\right)*\mathbf{F}_{N}\left(p_{i}\right)$$
Here, $p_{i}$ is the attention value of the image[CLS] token to the $i$-th patch token, calculated in the $k_th$ Transformer layer. We define $\dot{a}_i$ as the text saliency score of the $i$-th patch in the sequence, $\beta$ as a hyper-parameter, and $\mathbf{F}_N$ as a normalization function for $p_i$ and $a_i.$ Then, we select the top-$u$ image patch tokens from the patch sequence $\{v_{cls}^k,v_1^k,\cdots,v_n^k\}$ $\hat{\text{based on their text saliency scores}}\{\dot{a}_{1},\cdots,\dot{a}_{n}\}$,where $u$ = $n\times \alpha$ and $\alpha$ is the selection ratio that controls the proportion of selected patches to total patches. The selected top-$u$ image patch tokens are kept and we reconstruct the $k_{th}$ visual sequence as $v^k=\{v_{cls}^{\hat{k}},v_{1}^{k},\cdots,v_{u}^{k},v_{u}^{k}\}.$ Then the reconstructed visual sequence is fed to the next $(k+1)_{th}$ Transformer layer.
## 3.2.2 Text-Guided Patch Abstraction
While the coarse-grained key patch extraction removes some redundant visual tokens, many text-irrelevant tokens remain in the ViT backbone. However, we need to be cautious not to remove too many visual tokens in the ViT backbone, as this can lead to the loss of important structural information and affect the distribution of hidden representations in the backbone. To address this issue, we propose the Text-guided Patch Abstraction (TPA) outside the ViT backbone, using a lightweight Patch Abstraction Decoder (PAD). As shown in Figure 2(b), the PAD consists of two transformer modules, each of which includes a self-attention layer and a cross-attention layer. Similar to [3] and [27], the model structure of PAD employs a similar way, but with a key difference: we select top-s image patch tokens $\{\overline{v}_{cls},\overline{v}_{1},\cdots,\overline{v}_{s}\}$ from the output patch sequence $\{v_{cls},v_{1},\cdots,v_{s}\}$ of ViT backbone based on the text saliency scores $\{\dot{a}_1,\cdots,\dot{a}_n\}$, and take them as the input to guide the PAD to condense visual information while they used the static learnable embeddings without taking any prior of text information. Noted that $s=\gamma*u$ and $\gamma$
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.5048828125,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.5048828125,
0.11932373046875
],
[
0.08087158203125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.08087158203125,
0.1395263671875,
0.468017578125,
0.1395263671875,
0.468017578125,
0.149658203125,
0.08087158203125,
0.149658203125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.136962890625,
0.89208984375,
0.15087890625,
0.50244140625,
0.1514892578125
],
[
0.07843017578125,
0.15283203125,
0.4697265625,
0.1514892578125,
0.4697265625,
0.1666259765625,
0.07843017578125,
0.1673583984375
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.50341796875,
0.1685791015625,
0.88818359375,
0.1685791015625,
0.88818359375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50341796875,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.08087158203125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.5009765625,
0.1962890625,
0.890625,
0.1976318359375,
0.890625,
0.212158203125,
0.5009765625,
0.2108154296875
],
[
0.08087158203125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.07843017578125,
0.2266845703125,
0.30810546875,
0.2291259765625,
0.30810546875,
0.24365234375,
0.07843017578125,
0.2412109375
],
[
0.50244140625,
0.2291259765625,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.50244140625,
0.242431640625
],
[
0.50341796875,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.080078125,
0.25390625,
0.370849609375,
0.25390625,
0.370849609375,
0.267578125,
0.080078125,
0.267578125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.098876953125,
0.27783203125,
0.4697265625,
0.27783203125,
0.4697265625,
0.292236328125,
0.098876953125,
0.292236328125
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.08087158203125,
0.29345703125,
0.4697265625,
0.29345703125,
0.4697265625,
0.306884765625,
0.08087158203125,
0.306884765625
],
[
0.5009765625,
0.28857421875,
0.890625,
0.28857421875,
0.890625,
0.3017578125,
0.5009765625,
0.3017578125
],
[
0.080078125,
0.30810546875,
0.468017578125,
0.30810546875,
0.468017578125,
0.3212890625,
0.080078125,
0.3212890625
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.5009765625,
0.317626953125,
0.7685546875,
0.317626953125,
0.7685546875,
0.330810546875,
0.5009765625,
0.330810546875
],
[
0.080078125,
0.323974609375,
0.468017578125,
0.323974609375,
0.468017578125,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.080078125,
0.33837890625,
0.46728515625,
0.33837890625,
0.46728515625,
0.3515625,
0.080078125,
0.3515625
],
[
0.56787109375,
0.3466796875,
0.826171875,
0.3466796875,
0.826171875,
0.361083984375,
0.56787109375,
0.361083984375
],
[
0.08087158203125,
0.354248046875,
0.46728515625,
0.354248046875,
0.46728515625,
0.367431640625,
0.08087158203125,
0.367431640625
],
[
0.07843017578125,
0.367431640625,
0.46728515625,
0.36865234375,
0.46728515625,
0.38330078125,
0.07843017578125,
0.3818359375
],
[
0.50341796875,
0.375,
0.89208984375,
0.375,
0.89208984375,
0.38818359375,
0.50341796875,
0.38818359375
],
[
0.50244140625,
0.389404296875,
0.890625,
0.389404296875,
0.890625,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.07843017578125,
0.40087890625,
0.281005859375,
0.401611328125,
0.281005859375,
0.416748046875,
0.07843017578125,
0.415283203125
],
[
0.50341796875,
0.40478515625,
0.89208984375,
0.40478515625,
0.89208984375,
0.41845703125,
0.50341796875,
0.41845703125
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.50341796875,
0.42041015625,
0.890625,
0.42041015625,
0.890625,
0.434326171875,
0.50341796875,
0.434326171875
],
[
0.080078125,
0.439453125,
0.4697265625,
0.439453125,
0.4697265625,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50341796875,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.50341796875,
0.448974609375
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50341796875,
0.45068359375,
0.88818359375,
0.45068359375,
0.88818359375,
0.46533203125,
0.50341796875,
0.46533203125
],
[
0.08087158203125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50244140625,
0.466552734375,
0.890625,
0.466552734375,
0.890625,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.48095703125,
0.890625,
0.479736328125,
0.890625,
0.49365234375,
0.50244140625,
0.494384765625
],
[
0.50341796875,
0.49560546875,
0.890625,
0.49560546875,
0.890625,
0.5087890625,
0.50341796875,
0.5087890625
],
[
0.08251953125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51220703125,
0.08251953125,
0.51220703125
],
[
0.50341796875,
0.51123046875,
0.89208984375,
0.51123046875,
0.89208984375,
0.5244140625,
0.50341796875,
0.5244140625
],
[
0.08087158203125,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.07843017578125,
0.52880859375,
0.468017578125,
0.53076171875,
0.468017578125,
0.54541015625,
0.07843017578125,
0.54345703125
],
[
0.50341796875,
0.52587890625,
0.89208984375,
0.52587890625,
0.89208984375,
0.54052734375,
0.50341796875,
0.54052734375
],
[
0.08251953125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.08251953125,
0.5595703125
],
[
0.50244140625,
0.54150390625,
0.89208984375,
0.54150390625,
0.89208984375,
0.5546875,
0.50244140625,
0.5546875
],
[
0.080078125,
0.5595703125,
0.468017578125,
0.56005859375,
0.468017578125,
0.57470703125,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.55615234375,
0.6259765625,
0.55615234375,
0.6259765625,
0.5693359375,
0.50244140625,
0.5693359375
],
[
0.08087158203125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.08087158203125,
0.5888671875
],
[
0.50244140625,
0.5771484375,
0.771484375,
0.5771484375,
0.771484375,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.08087158203125,
0.59033203125,
0.46728515625,
0.59033203125,
0.46728515625,
0.60400390625,
0.08087158203125,
0.60400390625
],
[
0.080078125,
0.60498046875,
0.46728515625,
0.60498046875,
0.46728515625,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.50341796875,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50341796875,
0.62890625
],
[
0.080078125,
0.62255859375,
0.1544189453125,
0.62255859375,
0.1544189453125,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.200927734375,
0.63525390625,
0.3447265625,
0.63525390625,
0.3447265625,
0.6494140625,
0.200927734375,
0.6494140625
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.50341796875,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.08087158203125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.6748046875,
0.08087158203125,
0.6748046875
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.08251953125,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.68994140625,
0.08251953125,
0.68994140625
],
[
0.5009765625,
0.673828125,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.5009765625,
0.68798828125
],
[
0.08087158203125,
0.6923828125,
0.468017578125,
0.6923828125,
0.468017578125,
0.70556640625,
0.08087158203125,
0.70556640625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70703125,
0.468017578125,
0.70703125,
0.468017578125,
0.72021484375,
0.080078125,
0.72021484375
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.07843017578125,
0.7216796875,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.07843017578125,
0.736328125
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.4697265625,
0.7373046875,
0.4697265625,
0.751953125,
0.08087158203125,
0.751953125
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.08251953125,
0.751953125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.08251953125,
0.7666015625
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.08251953125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.7841796875,
0.468017578125,
0.7841796875,
0.468017578125,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.50341796875,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.08087158203125,
0.79736328125,
0.4697265625,
0.79736328125,
0.4697265625,
0.81103515625,
0.08087158203125,
0.81103515625
],
[
0.5009765625,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.08087158203125,
0.81298828125,
0.471435546875,
0.81298828125,
0.471435546875,
0.8271484375,
0.08087158203125,
0.8271484375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82568359375,
0.50244140625,
0.82568359375
],
[
0.08087158203125,
0.82763671875,
0.468017578125,
0.82763671875,
0.468017578125,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82470703125,
0.89208984375,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.08087158203125,
0.84326171875,
0.468017578125,
0.84326171875,
0.468017578125,
0.85693359375,
0.08087158203125,
0.85693359375
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85791015625,
0.1993408203125,
0.85791015625,
0.1993408203125,
0.8720703125,
0.080078125,
0.8720703125
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85498046875,
0.89208984375,
0.869140625,
0.50244140625,
0.87060546875
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.1356201171875,
0.88623046875,
0.41259765625,
0.88623046875,
0.41259765625,
0.90087890625,
0.1356201171875,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"final visual summary of the long patch sequence, which can",
"To learn an effective patch selector, we generate patch-level",
"be denoted as V = [ocls, oi, U2, .., os], where s < u.",
"training labels bytransformingthe boundingbox annota-",
"The image and text representations are concatenated and fed",
"tions (as described in Appendix A.2). However, this ap-",
"into the cross-modal encoder.We obtain the cross-modal",
"proach introduces bias,as illustrated in Figure 2 (b),where",
"representations[cls,Ci,C2,..,ci],wherel =s+m.The",
"TSPS predicts all patches within the bounding box with a",
"cross-modal representations can be used to fine-tune down-",
"high alignment score, even if some of them are irrelevant",
"stream multi-modal understanding tasks.Additionally, the",
"To address this issue, we incorporate the attention map of",
"output cross-modal representations [cels, Ci, C2, .., ct] of",
"the image [CLS] token into the selection process. Previ-",
"the multi-modal encoder are fed into a Transformer decoder",
"ous work has shown that in ViTs, the image [CLS] token",
"for sequence-to-sequence learning",
"pays more attention to class-specific tokens than to tokens",
"in non-object regions [6]. By weighting the alignment score",
"3.2.Bottom-Up Patch Summarization",
"a and the attention map p of the image [CLS] token to other",
"The proposed Bottom-Up Patch Summarization consists",
"tokens, we can highlight the patches corresponding to ob-",
"of two steps: Key Patch Extraction (KPE) and Text-Guided",
"jects in the detection box and suppress other patches. This",
"Patch Abstraction (TPA). KPE is conducted within the vi-",
"approach reduces the bias introduced by the bounding box",
"labels and can be formulated as follows:",
"sual backbone to select a coarse-grained subset of text-",
"relevant patch tokens. Then, outside the backbone, Text-",
"ai=β*F (ai)+ (1-β)*F (pi)",
"Guided Patch Abstraction is conducted to abstract the se-",
"lected tokens and obtain a more prominent visual summary.",
"Here, P; is the attention value of the image [CLS] token to",
"the i-th patch token, calculated in the kth Transformer layer.",
"3.2.1 KeyPatchExtraction",
"We define a as the text saliency score of the i-th patch in the",
"Figure 2 (b) shows how we perform Key Patch Extrac-",
"sequence, β as a hyper-parameter, and F as a normaliza-",
"tion (KPE) in the ViT-based visual backbone based on the",
"tion function for p; and az. Then, we select the top-u image",
"Text Semantic-aware Patch Selector (TSPS). Similar to the",
"patch tokens from the patch sequence {vas,u,...,uh]",
"text extractive summarization methods53.64lwe view",
"based on their text saliency scores [ai,..,anJ, where",
"the KPE as a patch classification task that aims to classify",
"u= nx Q and αis the selection ratio that controls the",
"proportion of selected patches to total patches. The selected",
"whetherapatchisalignedwiththetextsemanticandshould",
"top-u image patch tokens are kept and we reconstruct the",
"beselected.Specially,supposetheTSPSispluggedbe-",
"tween the kth (1 ≤ k<N)Transformer layer and (k+1)th",
"kth visual sequence as uk = (uas, uh, ... , uh, vh]. Then",
"Transformer layer and the output patch sequence features of",
"the reconstructed visual sequence is fed to the next (k +1)th",
"kth Transformer layer is h =fuhls,uf,...,uh].The text",
"Transformerlayer.",
"[CLS] feature is output by the text encoder and represents",
"3.2.2Text-Guided Patch Abstraction",
"global information of the input text T. We will first con-",
"catenate the text [CLS] feature with each image patch token",
"While the coarse-grained key patch extraction removes",
"some redundant visual tokens, many text-irrelevant tokens",
"asfollows:",
"uk = concat(uk,tcs)",
"remain in the ViT backbone. However, we need to be",
"cautious not to remove too many visual tokens in the ViT",
"where vh e Rd,tcls e Rd, uk e R2d, i e [1,2,...,n].",
"backbone, as this can lead to the loss of important struc-",
"Then the concatenated patch features uk are fed to TSPS,",
"tural information and affect the distribution of hidden rep-",
"which is a Multi-layer Perceptron (MLP) that contains three",
"resentations in the backbone. To address this issue, we",
"linear layers and is used to predict the alignment score be-",
"propose the Text-guided Patch Abstraction (TPA) outside",
"tween patches and the input text T. The first two linear",
"the ViT backbone, using a lightweight Patch Abstraction",
"layers will linearly project the concatenated patch features",
"Decoder (PAD). As shown in Figure 2 (b), the PAD con-",
"[u] to the hidden representations (hk] and then the hidden",
"sists of two transformer modules, each of which includes a",
"representations {h) is fed to the last linear layer denoted",
"self-attention layer and a cross-attention layer. Similar to",
"asFewhichcanbeseenasaclassifiertopredict whether",
"[3] and [27], the model structure of PAD employs a sim-",
"the patches are relevant to the input text. The output of the",
"ilar way, but with a key difference: we select top-s im-",
"last linear layer has only one dimension and will be fed to a",
"age patch tokens {ucls, Ui,**, Us] from the output patch",
"Sigmoid activation function. Formally, the alignment score",
"sequence {vcls, Ui, .. -, Us] of ViT backbone based on the",
"ai between the ith image patch and input text T can be cal-",
"text saliency scores {ai,... , anJ, and take them as the in-",
"culated as follow:",
"put to guide the PAD to condense visual information while",
"they used the static learnable embeddings without taking",
"ai = Sigmoid(Fe(hk)),i E ([1,2,...,n)",
"any prior of text information. Noted that s = * u and "
] |
{
"arxiv_id": "2307.08504",
"language": "en",
"page_idx": 3
}
|
||
2307.08504-en-4
|
2307.08504-en-4.jpg
|
is the selection ratio for TPA. These top-s patch tokens are highly relevant to the text semantics and provide a strong prior to help the PAD learn a condensed visual summary and highlight text-relevant visual information. After we select the top-s image patch tokens $\{\overline{v}_{cls},\overline{v}_{1},\cdots,\overline{v}_{s}\}$,we will feed them to PAD. In each transformer module of the PAD, $\{\overline{v}_{cls},\overline{v}_{1},\cdots,\overline{v}_{l}\}$ first undergoes the self-attention layer, followed by a cross-attention with the output sequence $\{v_{cls},v_{1},\cdots,v_{s}\}.$ Finally, the output of the PAD denoted as $\hat{V}$ = $\{ \hat{v} _{cls}, \dot{\hat{v} } _{1}, \hat{v} _{2}, \cdots , \hat{v} _{s}\}$ serves as the final visual summary.
## $4. \textbf{Experiments}$
## 4.1. Data & Setup
In our work, we follow the pre-training setup established in [29] and utilize the same pre-training dataset consisting of4 million images with associated texts. The dataset comprises two in-domain datasets, MS COCO [33] and Visual Genome[24], as well as three out-domain datasets, Conceptual Captions [45], and SBU Captions [38]. Further details on the dataset can be found in Appendix A.1.
We pre-train the model for 30 epochs with a total batch size of 1024 on 8 NVIDIA A100 GPUs. We use a 10-layer Transformer for the text encoder, a 2-layer transformer for the patch abstraction decoder, a 3-layer for the cross-modal encoder network, and a 12-layer Transformer for the crossmodal decoder. Specifically, we initialize the text encoder using the first 10 layers of the BERTbase [10] model, initialize the patch abstraction decoder using the last 2 layers of BERT$base$, initialize the cross-modal encoder network using the last 3 layers of BERT$base.$
## 4.2. Main Result
We evaluate the effectiveness of our proposed model, BUS , on four well-established vision-language downstream tasks: Visual Question Answering (VQA), Crossmodal Retrieval, Image Captioning, and Visual Grounding (VG). The Key Patch Extraction (KPE) is performed after the 6th Transformer layer in the ViT encoder, where about 70% text-relevant patch tokens are selected after the coarsegrained extraction. For the Text-Guided Patch Abstraction (TPA), we select the top 20% of text-relevant patch tokens as the input for the Patch Abstraction Decoder (PAD). Our experiments in subsection 4.5 and subsection 4.5 demonstrate that this setting achieves the desired trade-off between downstream task performance and model inference speed. Please refer to Appendix A.1 for more information on the datasets and fine-tuning hyper-parameters, and to Appendix C for details on the comparison methods used in the experiments.
## 4.2.1 Visual Question Answering
The VQA task[1] requires the model to answer natural language questions given an image. Following the approach proposed in [29], we treat VQA as an answer-generation problem. We evaluate the performance of our proposed model, BUS , by submitting our results to the evaluation server $^{\text{l in Table l. and report the test-dev and test-std}}$ scores in Table 1. Our results show that BUS achieves comparable performance with state-of-the-art models under the same image resolution $(384\times384)$ while being about 50% faster in terms of model inference time (as reported in subsection 4.6). Furthermore, when we increase the image resolution to 512 × 512, our model achieves state-of-the-art performance while maintaining a similar inference computation cost to other baselines. These results demonstrate the effectiveness and efficiency of BUS in VQA tasks.
## 4.2.2 Image Captioning
For the image captioning task, where there is no textual input, we set the hyper-parameter $\beta$ to 0 and select patches based on the attention weight of the image[CLS] token to other image tokens. Following [31],we fine-tuned BUS with cross-entropy loss and then with CIDEr optimization for an extra 5 epochs. Our experiments, as shown in Table 1, demonstrate that BUS achieves comparable results with SOTA models when the image resolution is set to $384\times384.$ Moreover, when we set the image resolution to $512\times512$, BUS performs even better on CIDEr evaluation, surpassing state-of-the-art models.
## 4.2.3 Image-Text Retrieval
We conduct experiments on MSCOCO [33] and Flickr30K [40] datasets for both image-to-text retrieval (TR) and textto-image retrieval (IR), and jointly optimize the ITC loss and the ITM loss during fine-tuning. The results are reported in Table 2. Our model demonstrates comparable performance with other VLP baselines, as shown in the experimental results.
## 4.2.4 Visual Grounding
Table 3 demonstrates the performance of BUS in the visual grounding task. When the image resolution is set to $384\times384$, our model achieves comparable results with competitive baseline methods. Thanks to our bottom-up summarization mechanism, we are able to improve the image resolution and obtain better results without incurring additional computational costs compared to other methods. When we increase the image resolution to $512\times512$, our model outperforms all state-of-the-art methods, demonstrating the effectiveness and efficiency of BUS .
$^{\mathrm{l}}$https://eval.ai/web/challenges/challenge-page/830/overview
|
[
[
0.080078125,
0.091552734375,
0.47216796875,
0.091552734375,
0.47216796875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.744140625,
0.091552734375,
0.744140625,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.07843017578125,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.07843017578125,
0.12249755859375,
0.468017578125,
0.121826171875,
0.468017578125,
0.1363525390625,
0.07843017578125,
0.136962890625
],
[
0.5009765625,
0.11871337890625,
0.8896484375,
0.11871337890625,
0.8896484375,
0.1319580078125,
0.5009765625,
0.1319580078125
],
[
0.080078125,
0.1383056640625,
0.471435546875,
0.1383056640625,
0.471435546875,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1337890625,
0.89208984375,
0.1337890625,
0.89208984375,
0.147705078125,
0.50244140625,
0.147705078125
],
[
0.080078125,
0.154052734375,
0.4697265625,
0.154052734375,
0.4697265625,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.5009765625,
0.1483154296875,
0.89208984375,
0.147705078125,
0.89208984375,
0.1622314453125,
0.5009765625,
0.1634521484375
],
[
0.080078125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.5009765625,
0.1622314453125,
0.89208984375,
0.1634521484375,
0.89208984375,
0.177978515625,
0.5009765625,
0.1767578125
],
[
0.080078125,
0.1817626953125,
0.468017578125,
0.1817626953125,
0.468017578125,
0.1956787109375,
0.080078125,
0.1956787109375
],
[
0.5009765625,
0.1767578125,
0.89208984375,
0.1787109375,
0.89208984375,
0.1932373046875,
0.5009765625,
0.1912841796875
],
[
0.50244140625,
0.1932373046875,
0.89208984375,
0.1932373046875,
0.89208984375,
0.20703125,
0.50244140625,
0.20703125
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.2083740234375,
0.89208984375,
0.2083740234375,
0.89208984375,
0.2215576171875,
0.50244140625,
0.2215576171875
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2247314453125,
0.89208984375,
0.2247314453125,
0.89208984375,
0.238037109375,
0.50244140625,
0.238037109375
],
[
0.07843017578125,
0.242431640625,
0.1895751953125,
0.24365234375,
0.1895751953125,
0.25830078125,
0.07843017578125,
0.2568359375
],
[
0.50244140625,
0.2392578125,
0.89208984375,
0.2392578125,
0.89208984375,
0.25244140625,
0.50244140625,
0.25244140625
],
[
0.50341796875,
0.255615234375,
0.890625,
0.255615234375,
0.890625,
0.266357421875,
0.50341796875,
0.266357421875
],
[
0.5048828125,
0.270751953125,
0.8896484375,
0.270751953125,
0.8896484375,
0.281005859375,
0.5048828125,
0.281005859375
],
[
0.08087158203125,
0.2822265625,
0.20751953125,
0.2822265625,
0.20751953125,
0.296630859375,
0.08087158203125,
0.296630859375
],
[
0.50244140625,
0.2841796875,
0.89208984375,
0.2841796875,
0.89208984375,
0.29736328125,
0.50244140625,
0.29736328125
],
[
0.50244140625,
0.2998046875,
0.89208984375,
0.2998046875,
0.89208984375,
0.313232421875,
0.50244140625,
0.313232421875
],
[
0.080078125,
0.31005859375,
0.2181396484375,
0.311279296875,
0.2181396484375,
0.326904296875,
0.080078125,
0.32568359375
],
[
0.50341796875,
0.31640625,
0.890625,
0.31640625,
0.890625,
0.326904296875,
0.50341796875,
0.326904296875
],
[
0.50244140625,
0.330078125,
0.8349609375,
0.330078125,
0.8349609375,
0.343505859375,
0.50244140625,
0.343505859375
],
[
0.09722900390625,
0.339111328125,
0.4697265625,
0.339111328125,
0.4697265625,
0.352783203125,
0.09722900390625,
0.352783203125
],
[
0.07843017578125,
0.352783203125,
0.4697265625,
0.354248046875,
0.4697265625,
0.36865234375,
0.07843017578125,
0.367431640625
],
[
0.5009765625,
0.35595703125,
0.6806640625,
0.35791015625,
0.6806640625,
0.37255859375,
0.5009765625,
0.37060546875
],
[
0.080078125,
0.369384765625,
0.468017578125,
0.369384765625,
0.468017578125,
0.38330078125,
0.080078125,
0.38330078125
],
[
0.5009765625,
0.376953125,
0.890625,
0.37744140625,
0.890625,
0.392578125,
0.5009765625,
0.391357421875
],
[
0.08087158203125,
0.385009765625,
0.468017578125,
0.385009765625,
0.468017578125,
0.3984375,
0.08087158203125,
0.3984375
],
[
0.50244140625,
0.393310546875,
0.89208984375,
0.392578125,
0.89208984375,
0.4072265625,
0.50244140625,
0.40771484375
],
[
0.08251953125,
0.401611328125,
0.46728515625,
0.401611328125,
0.46728515625,
0.412353515625,
0.08251953125,
0.412353515625
],
[
0.50244140625,
0.4091796875,
0.89208984375,
0.4091796875,
0.89208984375,
0.42236328125,
0.50244140625,
0.42236328125
],
[
0.080078125,
0.415283203125,
0.4697265625,
0.415283203125,
0.4697265625,
0.4287109375,
0.080078125,
0.4287109375
],
[
0.50244140625,
0.423583984375,
0.89208984375,
0.423583984375,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.43115234375,
0.37744140625,
0.43115234375,
0.37744140625,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.50341796875,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50341796875,
0.45263671875
],
[
0.1004638671875,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.462158203125,
0.1004638671875,
0.462158203125
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.07843017578125,
0.462158203125,
0.4697265625,
0.46337890625,
0.4697265625,
0.477783203125,
0.07843017578125,
0.4765625
],
[
0.50244140625,
0.468505859375,
0.89208984375,
0.468505859375,
0.89208984375,
0.481689453125,
0.50244140625,
0.481689453125
],
[
0.08087158203125,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.50244140625,
0.484130859375,
0.890625,
0.484130859375,
0.890625,
0.49755859375,
0.50244140625,
0.49755859375
],
[
0.08087158203125,
0.49560546875,
0.46728515625,
0.49560546875,
0.46728515625,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.50341796875,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.51123046875,
0.50341796875,
0.51123046875
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.5146484375,
0.89208984375,
0.5146484375,
0.89208984375,
0.52783203125,
0.50244140625,
0.52783203125
],
[
0.08087158203125,
0.52587890625,
0.468017578125,
0.52587890625,
0.468017578125,
0.5361328125,
0.08087158203125,
0.5361328125
],
[
0.50244140625,
0.52978515625,
0.7646484375,
0.52978515625,
0.7646484375,
0.54345703125,
0.50244140625,
0.54345703125
],
[
0.07843017578125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.07843017578125,
0.55322265625
],
[
0.07843017578125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.07843017578125,
0.568359375
],
[
0.50341796875,
0.55810546875,
0.70166015625,
0.55810546875,
0.70166015625,
0.5712890625,
0.50341796875,
0.5712890625
],
[
0.080078125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.57568359375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59130859375,
0.50244140625,
0.59033203125
],
[
0.07843017578125,
0.58544921875,
0.32275390625,
0.58544921875,
0.32275390625,
0.5986328125,
0.07843017578125,
0.5986328125
],
[
0.50341796875,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50341796875,
0.60693359375
],
[
0.50244140625,
0.60791015625,
0.8935546875,
0.60791015625,
0.8935546875,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.08087158203125,
0.6171875,
0.2091064453125,
0.6171875,
0.2091064453125,
0.630859375,
0.08087158203125,
0.630859375
],
[
0.50341796875,
0.62353515625,
0.890625,
0.62353515625,
0.890625,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.1004638671875,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.1004638671875,
0.65771484375
],
[
0.50244140625,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.07843017578125,
0.65771484375,
0.468017578125,
0.6591796875,
0.468017578125,
0.673828125,
0.07843017578125,
0.67236328125
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50341796875,
0.66943359375,
0.6044921875,
0.66943359375,
0.6044921875,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.07843017578125,
0.689453125,
0.468017578125,
0.68994140625,
0.468017578125,
0.705078125,
0.07843017578125,
0.7041015625
],
[
0.50244140625,
0.69580078125,
0.6787109375,
0.6962890625,
0.6787109375,
0.7109375,
0.50244140625,
0.71044921875
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.72265625,
0.890625,
0.72265625,
0.890625,
0.736328125,
0.50244140625,
0.736328125
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73779296875,
0.89208984375,
0.73779296875,
0.89208984375,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.080078125,
0.751953125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.76513671875
],
[
0.50341796875,
0.7548828125,
0.890625,
0.7548828125,
0.890625,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.08087158203125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.5009765625,
0.767578125,
0.890625,
0.76806640625,
0.890625,
0.78271484375,
0.5009765625,
0.7822265625
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.7841796875,
0.890625,
0.7841796875,
0.890625,
0.79736328125,
0.50244140625,
0.79736328125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.7998046875,
0.89208984375,
0.7998046875,
0.89208984375,
0.81298828125,
0.50244140625,
0.81298828125
],
[
0.08087158203125,
0.8125,
0.46728515625,
0.8125,
0.46728515625,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.50244140625,
0.814453125,
0.89208984375,
0.814453125,
0.89208984375,
0.82763671875,
0.50244140625,
0.82763671875
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50341796875,
0.83056640625,
0.890625,
0.83056640625,
0.890625,
0.84130859375,
0.50341796875,
0.84130859375
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.8447265625,
0.890625,
0.8447265625,
0.890625,
0.85791015625,
0.50244140625,
0.85791015625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.859375,
0.79345703125,
0.859375,
0.79345703125,
0.87255859375,
0.50244140625,
0.87255859375
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.07843017578125,
0.8837890625
],
[
0.080078125,
0.88818359375,
0.131591796875,
0.88818359375,
0.131591796875,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.521484375,
0.88720703125,
0.8408203125,
0.88720703125,
0.8408203125,
0.9013671875,
0.521484375,
0.9013671875
]
] |
[
"is the selection ratio for TPA, These top-s patch tokens are",
"4.2.1 Visual Question Answering",
"highly relevant to the text semantics and provide a strong",
"prior to help the PAD learn a condensed visual summary",
"The VQA task [l] requires the model to answer natural lan-",
"and highlight text-relevant visual information. After we",
"guage questions given an image. Following the approach",
"select the top-s image patch tokens [Ucls, Ui, ... , Ug], we",
"proposed in [29], we treat VQA as an answer-generation",
"will feed them to PAD. In each transformer module of the",
"problem.We evaluate the performance of our proposed",
"PAD, {ucls,Ui,...,uil first undergoes the self-attention",
"model, BUS ,by submitting our results to the evaluation",
"server 1 in Table 1. and report the test-dev and test-std",
"layer, followed by a cross-attention with the output se-",
"quence [ucls, Ui, .. , Us]. Finally, the output of the PAD",
"scores in Table 1. Our results show that BUS achieves com-",
"denoted as V = [ocls, i, u2, .-, 0,] serves as the final",
"parable performance with state-of-the-art models under the",
"visual summary.",
"same image resolution (384 × 384) while being about 50%",
"fasterin terms of model inferencetime(as reportedin sub-",
"section4.6).Furthermore,whenweincreasetheimageres-",
"4.Experiments",
"olution to 512 × 512, our model achieves state-of-the-art",
"performance while maintaining a similar inference compu-",
"4.1. Data & Setup",
"tation cost to otherbaselines.Theseresults demonstratethe",
"effectiveness and efficiency of BUS in VQA tasks.",
"In our work,we follow the pre-training setup established",
"in [29] and utilize the same pre-training dataset consisting",
"4.2.2Image Captioning",
"of 4 million images with associated texts. The dataset com-",
"For the image captioning task, where there is no textual in-",
"prises two in-domain datasets, MS COCO [33] and Visual",
"put, we set the hyper-parameter β to O and select patches",
"Genome[24l.aswellasthreeout-domaindatasets,Concep-",
"based on the attention weight of the image [CLS] token to",
"tual Captions [45], and SBU Captions [38]. Further details",
"other image tokens.Following [31], we fine-tuned BUS",
"on the dataset can be found in Appendix A.1.",
"with cross-entropy loss and then with CIDEr optimization",
"We pre-train the model for 30 epochs with a total batch",
"for an extra 5epochs.Ourexperiments, as shown in Ta-",
"size of 1024 on 8 NVIDIA A100 GPUs. We use a 10-layer",
"ble 1, demonstrate that BUS achieves comparable results",
"Transformer for the text encoder, a 2-layer transformer for",
"with SOTA models when the image resolution is set to",
"thepatchabstractiondecoder,a3-layerforthecross-modal",
"384 x 384.Moreover, when we set the image resolution",
"encoder network, and a 12-layer Transformer for the cross-",
"to 512 × 512, BUS performs even better on CIDEr evalua-",
"modal decoder.Specifically,we initializethe textencoder",
"tion, surpassing state-of-the-art models.",
"using the first 10 layers of the BERTbase [10] model, ini-",
"tialize the patch abstraction decoder using the last 2 layers",
"4.2.3Image-Text Retrieval",
"of BERTbase, initialize the cross-modal encoder network",
"We conduct experiments on MSCOCO [33] and Flickr30K",
"using the last 3 layers of BERTbase.",
"[40] datasets for both image-to-text retrieval (TR) and text-",
"to-image retrieval (IR), and jointly optimize the ITC loss",
"4.2. Main Result",
"and the ITM loss during fine-tuning. The results are re-",
"We evaluate the effectiveness of our proposed model,",
"ported in Table 2. Our model demonstrates comparable per-",
"BUS , on four well-established vision-language down-",
"formance with other VLP baselines, as shown in the exper-",
"stream tasks: Visual Question Answering (VQA), Cross-",
"imental results.",
"modal Retrieval, Image Captioning, and Visual Grounding",
"4.2.4Visual Grounding",
"(VG). The Key Patch Extraction (KPE) is performed after",
"the 6th Transformer layer in the ViT encoder, where about",
"Table 3 demonstrates the performance of BUS in the vi-",
"70%text-relevantpatchtokensare selectedafterthecoarse-",
"sual grounding task. When the image resolution is set to",
"grained extraction.For theText-Guided Patch Abstraction",
"384x384.ourmodelachievescomparableresultswith",
"(TPA), we select the top 20% of text-relevant patch tokens",
"competitive baseline methods.Thanks to our bottom-up",
"as theinputforthePatchAbstractionDecoder(PAD).Our",
"summarization mechanism, we are able to improve the im-",
"experiments in subsection 4.5 and subsection 4.5 demon-",
"age resolution and obtain better results without incurring",
"stratethatthis settingachievesthedesiredtrade-offbetween",
"additional computational costs compared to other methods.",
"downstream task performance and model inference speed.",
"Whenweincreasetheimageresolutionto512X512,our",
"Please refer to Appendix A.1 for more information on the",
"model outperforms all state-of-the-art methods, demonstrat-",
"datasets and fine-tuning hyper-parameters, and to Appendix",
"ing the effectiveness and efficiency of BUS.",
"Cfor details on the comparison methods used in the exper-",
"iments.",
"'https://eval.ai/web/challenges/challenge-page/830/overview"
] |
{
"arxiv_id": "2307.08504",
"language": "en",
"page_idx": 4
}
|
||
1705.01759-en-1
|
1705.01759-en-1.jpg
|
and VR can cause a viewer to feel discomfort. Just imagine how hard it is to follow an X-game skateboarder in a $360^{\circ} video. Hence, a mechanism to automatically navi-$ gate a 360° video in a way that captures most of the interest events for a viewer would be beneficial.
Conceptually, a 360°-video viewer is a human agent: at each frame, the agent observes a panoramic image (i.e., the observed state) and steers the viewing angle (i.e. the action) to cover the next preferred viewing angle (i.e., the goal). We refer to this process as 360 piloting. Based on this analogy and, more importantly, to relieve the viewer from constantly steering the viewing angle while watching 360° videos, we argue for an intelligent agent that can automatically piloting through 360° sports videos.
Using an automatic mechanism for displaying video contents is not a new idea. For example, video summarizationcondensing a long video into a short summary video[58]- has been used in reviewing hourly long surveillance videos. However, while a video summarization algorithm makes binary decisions on whether to select a frame or not, an agent for 360 piloting needs to operate on a spatial space to steer the viewing angle to consider events of interest in a 360° video. On the other hand, in virtual cinematography, most camera manipulation tasks are performed within relatively simpler virtual environments [8, 22, 12, 40] and there is no need to deal with viewers' perception difficulty because 3- D positions and poses of all entities are known. However, a practical agent for 360 piloting needs to directly work with raw 360° videos. For displaying 360° videos, Su et al.[53] proposed firstly detecting candidate events of interest in the entire video, and then applying dynamic programming to link detected events. However, as this method requires observing an entire video, it is non-suited for video streaming applications such as foveated rendering [45]. We argue that being able to make a selection based on the current and previous frames (like a human agent does) is critical for 360 piloting. Finally, both[53] and recent virtual cinematography works [7,6] aim for smooth viewing angle transition. Such transition should also be enforced for 360° piloting.
We propose“deep 360 pilot”—a deep learning-based agent that navigates a 360° sports video in a way that smoothly captures interesting moments in the video. Our “deep 360 pilot" agent not only follows foreground objects of interest but also steers the viewing angle smoothly to increase viewers' comfort. We propose the following online pipeline to learn an online policy from human agents to model how a human agent takes actions in watching sports videos. First, because in sports videos foreground objects are those of viewers' interest, we leverage a state-of-the-art object detector [50] to identify candidate objects of interest. Then, a Recurrent Neural Network (RNN) is used to select the main object among candidate objects. Given the main object and previously selected viewing angles, our method
predicts how to steer the viewing angle to the preferred one by learning a regressor. In addition, our pipeline is jointly trained with the following functions: (1) a regression loss measuring the distance between the selected and ground truth viewing angles, (2) a smoothness loss to encourage smooth transition in viewing angle, and (3) an expected reward of focusing on a foreground object. We used the policy gradient technique [62] to train the pipeline since it involves making an intermediate discrete decision corresponding to selecting the main object. To evaluate our method, we collected a new 360° sports video dataset consisting of five domains and trained an agent for each domain (referred to as 360-Sports). These domain-specific agents achieve the best performance in regression accuracy and transition smoothness in viewing angle.
Our main contributions are as follows: (1) We develop the first human-like online agent for automatically navigating 360° videos for viewers. The online processing nature suits the agent for streaming videos and predicting views for foveated VR rendering.(2) We propose a jointly trainable pipeline for learning the agent. Since the main object selection objective is non-differentiable, we employ a policy gradient technique to optimize the pipeline. (3) Our agent considers both viewing angle selection accuracy and transition smoothness. (4) We build the first 360° sports videos dataset to train and evaluate our “deep 360 pilot" agent.
$2. \textbf{Related Work}$
## We review related works in video summarization, saliency detection, and virtual cinematography. $2. 1. \textbf{ Video Summarization}$
We selectively review several most relevant video sum-
marization works from a large body of literature [58].
Important frame sampling. [33,19,43,27] proposed to sample a few important frames as the summary of a video. [47,54,63] focused on sampling domain-specific highlights. [67,54, 18] proposed weakly-supervised methods to select important frames. Recently, a few deep learningbased methods [66,66,65] have shown impressive performance. [48,49,55] focused on extracting highlights and generating synopses which showed several spatially nonoverlapping actions from different times of a video. Several methods [17, 25] involving user interaction have also been proposed in the graphics and the HCI communities.
Ego-centric video summarization. In ego-centric videos, cues from hands and objects become easier to extract compare to third-person videos. [30] proposed video summarization based on the interestingness and diverseness of objects and faces. [36] further proposed tracking objects and measuring the influence of individual frames. [28] proposed a novel approach to speed-up ego-centric videos while removing unpleasant camera movements.
In contrary to most video summarization methods which concern whether to select a frame or not, a method for 360
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50341796875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.4697265625,
0.10919189453125,
0.4697265625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.07843017578125,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.08087158203125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.34228515625,
0.15283203125,
0.34228515625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.098876953125,
0.17041015625,
0.468017578125,
0.17041015625,
0.468017578125,
0.184326171875,
0.098876953125,
0.184326171875
],
[
0.50341796875,
0.1673583984375,
0.890625,
0.1673583984375,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08087158203125,
0.1875,
0.46728515625,
0.1875,
0.46728515625,
0.1976318359375,
0.08087158203125,
0.1976318359375
],
[
0.50341796875,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.08087158203125,
0.20263671875,
0.46728515625,
0.20263671875,
0.46728515625,
0.21337890625,
0.08087158203125,
0.21337890625
],
[
0.50341796875,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.080078125,
0.216552734375,
0.468017578125,
0.216552734375,
0.468017578125,
0.2298583984375,
0.080078125,
0.2298583984375
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.231689453125,
0.468017578125,
0.231689453125,
0.468017578125,
0.24560546875,
0.080078125,
0.24560546875
],
[
0.5048828125,
0.2298583984375,
0.8896484375,
0.2298583984375,
0.8896484375,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.080078125,
0.246826171875,
0.46728515625,
0.246826171875,
0.46728515625,
0.260009765625,
0.080078125,
0.260009765625
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.07843017578125,
0.261474609375,
0.468017578125,
0.260009765625,
0.468017578125,
0.274658203125,
0.07843017578125,
0.27587890625
],
[
0.50244140625,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.271484375,
0.50244140625,
0.271484375
],
[
0.080078125,
0.276611328125,
0.468017578125,
0.276611328125,
0.468017578125,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.50341796875,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.080078125,
0.292236328125,
0.2646484375,
0.292236328125,
0.2646484375,
0.30615234375,
0.080078125,
0.30615234375
],
[
0.50341796875,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50341796875,
0.302978515625
],
[
0.09722900390625,
0.30810546875,
0.468017578125,
0.309326171875,
0.468017578125,
0.323974609375,
0.09722900390625,
0.322509765625
],
[
0.50341796875,
0.304931640625,
0.650390625,
0.304931640625,
0.650390625,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.521484375,
0.317626953125,
0.89208984375,
0.318115234375,
0.89208984375,
0.333251953125,
0.521484375,
0.33203125
],
[
0.080078125,
0.324462890625,
0.468017578125,
0.324462890625,
0.468017578125,
0.33837890625,
0.080078125,
0.33837890625
],
[
0.080078125,
0.339111328125,
0.465576171875,
0.339111328125,
0.465576171875,
0.352783203125,
0.080078125,
0.352783203125
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.37060546875,
0.468017578125,
0.37060546875,
0.468017578125,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.50341796875,
0.365478515625,
0.89208984375,
0.365478515625,
0.89208984375,
0.37890625,
0.50341796875,
0.37890625
],
[
0.50244140625,
0.37890625,
0.890625,
0.37890625,
0.890625,
0.392578125,
0.50244140625,
0.392578125
],
[
0.08087158203125,
0.386962890625,
0.468017578125,
0.386962890625,
0.468017578125,
0.397705078125,
0.08087158203125,
0.397705078125
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.4140625,
0.080078125,
0.4140625
],
[
0.50341796875,
0.395751953125,
0.890625,
0.395751953125,
0.890625,
0.4091796875,
0.50341796875,
0.4091796875
],
[
0.5009765625,
0.40771484375,
0.890625,
0.4091796875,
0.890625,
0.423583984375,
0.5009765625,
0.42236328125
],
[
0.080078125,
0.415283203125,
0.46728515625,
0.415283203125,
0.46728515625,
0.4287109375,
0.080078125,
0.4287109375
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.08087158203125,
0.43115234375,
0.4697265625,
0.43115234375,
0.4697265625,
0.4443359375,
0.08087158203125,
0.4443359375
],
[
0.50341796875,
0.44140625,
0.8896484375,
0.44140625,
0.8896484375,
0.451904296875,
0.50341796875,
0.451904296875
],
[
0.08087158203125,
0.447509765625,
0.46728515625,
0.447509765625,
0.46728515625,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.080078125,
0.4609375,
0.468017578125,
0.4609375,
0.468017578125,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.07843017578125,
0.47607421875,
0.468017578125,
0.474853515625,
0.468017578125,
0.4892578125,
0.07843017578125,
0.490478515625
],
[
0.50341796875,
0.47021484375,
0.86767578125,
0.47021484375,
0.86767578125,
0.484130859375,
0.50341796875,
0.484130859375
],
[
0.07843017578125,
0.490478515625,
0.4697265625,
0.4912109375,
0.4697265625,
0.5048828125,
0.07843017578125,
0.50390625
],
[
0.50341796875,
0.490478515625,
0.640625,
0.490478515625,
0.640625,
0.50390625,
0.50341796875,
0.50390625
],
[
0.07843017578125,
0.505859375,
0.468017578125,
0.5048828125,
0.468017578125,
0.51953125,
0.07843017578125,
0.52099609375
],
[
0.521484375,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.521484375,
0.52099609375
],
[
0.080078125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.080078125,
0.53515625
],
[
0.50244140625,
0.52099609375,
0.810546875,
0.52294921875,
0.810546875,
0.537109375,
0.50244140625,
0.53515625
],
[
0.07843017578125,
0.5361328125,
0.468017578125,
0.53515625,
0.468017578125,
0.5498046875,
0.07843017578125,
0.55029296875
],
[
0.50341796875,
0.537109375,
0.7099609375,
0.537109375,
0.7099609375,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.080078125,
0.5517578125,
0.4697265625,
0.5517578125,
0.4697265625,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.52294921875,
0.5537109375,
0.890625,
0.5537109375,
0.890625,
0.5673828125,
0.52294921875,
0.5673828125
],
[
0.080078125,
0.56640625,
0.46728515625,
0.56640625,
0.46728515625,
0.580078125,
0.080078125,
0.580078125
],
[
0.5048828125,
0.57080078125,
0.85107421875,
0.57080078125,
0.85107421875,
0.58056640625,
0.5048828125,
0.58056640625
],
[
0.080078125,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.59521484375,
0.080078125,
0.59521484375
],
[
0.5009765625,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.59765625,
0.5009765625,
0.59765625
],
[
0.080078125,
0.5966796875,
0.4697265625,
0.59521484375,
0.4697265625,
0.60986328125,
0.080078125,
0.611328125
],
[
0.50341796875,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50341796875,
0.61279296875
],
[
0.08087158203125,
0.61279296875,
0.46728515625,
0.61279296875,
0.46728515625,
0.62353515625,
0.08087158203125,
0.62353515625
],
[
0.5048828125,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.626953125,
0.5048828125,
0.626953125
],
[
0.08087158203125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.08087158203125,
0.64013671875
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.080078125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.65576171875,
0.080078125,
0.65576171875
],
[
0.50244140625,
0.64453125,
0.890625,
0.64599609375,
0.890625,
0.6591796875,
0.50244140625,
0.65771484375
],
[
0.08087158203125,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.50244140625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.07843017578125,
0.67041015625,
0.455810546875,
0.671875,
0.455810546875,
0.68603515625,
0.07843017578125,
0.68505859375
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.098876953125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.705078125,
0.471435546875,
0.705078125,
0.471435546875,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.5009765625,
0.705078125,
0.89208984375,
0.70263671875,
0.89208984375,
0.71728515625,
0.5009765625,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50341796875,
0.7216796875,
0.890625,
0.7216796875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.50341796875,
0.73486328125,
0.8388671875,
0.73486328125,
0.8388671875,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.50341796875,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5009765625,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79345703125,
0.5009765625,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.08087158203125,
0.869140625
],
[
0.5009765625,
0.85693359375,
0.7646484375,
0.85693359375,
0.7646484375,
0.87060546875,
0.5009765625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"and VR can cause a viewer to feel discomfort. Just imag-",
"predicts how to steer the viewing angle to the preferred one",
"ine how hard it is to follow an X-game skateboarder in a",
"by learning a regressor. In addition, our pipeline is jointly",
"360ovideo.Hence,amechanismtoautomaticallynavi-",
"trained with the following functions: (1) a regression loss",
"gate a 360° video in a way that captures most of the interest",
"measuring the distance between the selected and ground",
"events for a viewer would be beneficial.",
"truth viewing angles, (2) a smoothness loss to encourage",
"Conceptually, a 360°-video viewer is a human agent: at",
"smooth transition in viewing angle, and (3) an expected re-",
"eachframe,the agent observes apanoramicimage (i.e.,the",
"ward of focusing on a foreground object. We used the policy",
"observed state)and steers theviewing angle(i.e.theaction)",
"gradient technique [62] to train the pipeline since it involves",
"to cover the next preferred viewing angle (i.e., the goal). We",
"making an intermediate discrete decision corresponding to",
"refer to this process as 360 piloting. Based on this analogy",
"selecting the main object.To evaluate our method,we col-",
"and, more importantly, to relieve the viewer from constantly",
"lected a new 360° sports video dataset consisting of five do-",
"steering the viewing angle while watching 360 videos, we",
"mains and trained an agent for each domain (referred to as",
"argue for an intelligent agent that can automatically piloting",
"360-Sports). These domain-specific agents achieve the best",
"through 360° sports videos.",
"performance in regression accuracy and transition smooth-",
"Using an automatic mechanism for displaying video con-",
"ness in viewing angle.",
"Our main contributions are as follows: (1) We develop",
"tents is not a new idea. For example, video summarization-",
"condensing a long video into a short summary video [58]-",
"the first human-like online agent for automatically navigat-",
"has been used in reviewing hourly long surveillance videos.",
"ing 360videos for viewers.The online processing nature",
"However, while a video summarization algorithm makes bi-",
"suits the agent for streaming videos and predicting views for",
"foveated VR rendering. (2) We propose a jointly trainable",
"narydecisions on whetherto selecta frame or not, an agent",
"for 360 piloting needs to operate on a spatial space to steer",
"pipeline for learning the agent. Since the main object se-",
"lection objective is non-differentiable, we employ a policy",
"the viewing angle to consider events of interest in a 3600",
"gradient technique to optimize the pipeline. (3) Our agent",
"video. On the other hand, in virtual cinematography, most",
"considersbothviewing angle selection accuracyandtransi-",
"cameramanipulationtasks areperformedwithinrelatively",
"simpler virtual environments [8, 22, 12, 40] and there is no",
"tion smoothness. (4) We build the first 360 sports videos",
"need to deal with viewers'perception difficulty because 3-",
"dataset to train and evaluate our \"deep 360 pilot\" agent.",
"Dpositionsandposes of all entities areknown.However,a",
"2.Related Work",
"practical agent for 360 piloting needs to directly work with",
"We review related works in video summarization,",
"raw 360°videos. For displaying 360° videos, Su et al. [53]",
"saliency detection, and virtual cinematography",
"proposed firstly detecting candidate events of interest in the",
"2.1. Video Summarization",
"entire video, and then applying dynamic programming to",
"We selectively review several most relevant video sum-",
"link detected events. However, as this method requires ob-",
"marizationworksfromalargebodyofliterature[58]",
"serving an entire video, it is non-suited for video streaming",
"Important frame sampling. [33, 19, 43, 27] proposed to",
"applications such as foveated rendering [45]. We argue that",
"sample a few important frames as the summary of a video.",
"beingableto makeaselection based onthecurrent and pre-",
"[47,54,63]focused onsamplingdomain-specifichigh-",
"vious frames (like a human agent does) is critical for 360",
"lights. [67, 54, 18] proposed weakly-supervised methods",
"piloting. Finally, both [53] and recent virtual cinematogra-",
"to select important frames. Recently, a few deep learning-",
"phy works [7, 6] aim for smooth viewing angle transition.",
"based methods [66, 66, 65] have shown impressive perfor-",
"Such transition should also be enforced for 360 piloting.",
"mance.[48, 49, 55] focused on extracting highlights and",
"We propose “deep 360 pilot\"—a deep learning-based",
"generating synopses which showed several spatially non-",
"agent that navigates a 360o sports video in a way that",
"overlapping actions from different times of a video. Several",
"smoothly captures interesting moments in the video. Our",
"methods[17.251involving userinteractionhave also been",
"“deep 360 pilot\" agent not only follows foreground objects",
"proposed in the graphics and the HCI communities.",
"of interest but also steers the viewing angle smoothly to",
"Ego-centric video summarization. In ego-centric videos,",
"increase viewers' comfort. We propose the following on-",
"cuesfromhandsand objectsbecomeeasiertoextractcom-",
"line pipeline to learn an online policy from human agents to",
"pare to third-person videos. [30] proposed video summa-",
"model how a human agent takes actions in watching sports",
"rization based on the interestingness and diverseness of ob-",
"videos. First, because in sports videos foreground objects",
"jects and faces. [36] further proposed tracking objects and",
"are those of viewers' interest, we leverage a state-of-the-art",
"measuring the influence of individual frames. [28] proposed",
"object detector [50] to identify candidate objects of interest.",
"a novel approach to speed-up ego-centric videos while re-",
"Then, a Recurrent Neural Network (RNN) is used to select",
"moving unpleasant camera movements.",
"the main object among candidate objects. Given the main",
"In contrary to most video summarization methods which",
"object and previously selected viewing angles, our method",
"concern whether to select a frame or not, a method for 360"
] |
{
"arxiv_id": "1705.01759",
"language": "en",
"page_idx": 1
}
|
||
1705.01759-en-2
|
1705.01759-en-2.jpg
|
piloting concerns which viewing angle to select for each
panoramic frame in a 360° video.
## 2.2. Saliency Detection
Many methods have been proposed to detect salient regions typically measured by human gaze. [35,21,1,46, 59,64,46] focused on detecting salient regions on images. Recently,[34,24,9,44,5,61,60,57] leveraged deep learning and achieved significant performance gain. For videos, [10,20,37,52,41,29] relied on low-level appearance and motion cues as inputs. In addition,[26,16,51,39,13] included information about face, people, objects, or other contexts. Note that saliency detection methods do not select views directly, but output a saliency score map. Our method is also different to visual attention methods for object detection [38, 3, 42] in that it considers view transition smoothness as selecting views, which is crucial for video watching experience.
Ranking foreground objects of interest. Since regions of interest in sports videos are typically foreground objects, [55] proposed to use an object detector [4] to extract candidate objects of interest, then rank the saliency of these candidate objects. For 360 piloting, we propose a similar baseline which first detects objects using RCNN [50], then select the viewing angle focusing on the most salient object according to a saliency detector [64].
# $2. 3. \textbf{ Virtual Cinematography}$
Finally, existing virtual cinematography works focused on camera manipulation in simple virtual environments/video games [8,22,12,40] and did not deal with the perception difficulty problem. [14,56,7,6] relaxed the assumption and controlled virtual cameras within restricted static wide field-of-view video of a classroom, video conference, or basketball court, where objects of interest could be easily extracted. In contrast, our method handles raw 360° sports videos downloaded from YouTube$^{\mathrm{l}}$ in five domains (e.g., basketball, parkour, etc.). Recently, Su et al.[53] also proposed handling raw 360° videos download from YouTube. They referred to this problem as Pano2Vid-automatic cinematography in 360° videos-and proposed an offline method. In contrast, we propose an online humanlike agent acting based on both present and previous observations. We argue that for handling streaming videos and other human-in-the-loop applications (e.g., foveated rendering[45]) a human-like online agent is necessary in order to provide more effective video-watching support.
## 3.0ur Approach
We first define the 360 piloting problem in details (Sec. 3.1). Then, we introduce our deep 360 pilot approach (Sec. 3.2-Sec. 3.6). Finally, we describe the training procedure of our model (Sec. 3.7).
$^{1}$https://www.youtube.com/
## 3.1. Definitions We formulate the 360 piloting task as the following on-
line viewing angle selection task.
Observation. At time $t$,the agent observes a new frame $v_t$, which is the $t$-th frame of the 360° video. The sequence of frames that the agent has observed up to this time is referred ${\mathrm{to~as~}\mathbf{V}}_{t}=\{v_{1},...,v_{t}\}.$
Goal. The goal of the agent is to select a viewing angle $l_{t}$ at time $t$ so that the sequence of viewing angles $\mathbf{L}_t=$ $\{l_1,...,l_t\}$ smoothly capture events of interest in the $360^\circ$ video. Note that $l_t=(\dot{\theta}_t,\phi_t)$ is a point on the 360° viewing sphere, parameterized by the azimuth angle $\theta_t\in[0^\circ,360^\circ]$ and elevation angle $\phi_t\in[-90^\circ,90^\circ]$
Action. In order to achieve the goal, the agent takes the action of steering the viewing angle by $\Delta_t$ at time $t.$ Given the previous viewing angle $l_t-1$ and current action $\Delta_t$, the current viewing angle $l_t$ is computed as follows,
$$l_t=\Delta_t+l_{t-1}.$$
(1)
Online policy. We assume that the agent takes an action $\Delta_t$
at frame $t$ according to an online policy $\pi$ as follows,
(2)
$$\Delta_t=\pi(\mathbf{V}_t,\mathbf{L}_{t-1}),$$
where the online policy depends on both the current and previous observation $\mathbf{V}_t$ and previous viewing angles $\mathbf{L}_t-1.$ This implies that the previous viewing angles affect the current action similar to what a human viewer acts when viewing a 360° sports video. Hence, the main task of 360 piloting is about learning the online policy from data.
In the following, we discuss various design choices of our proposed deep 360 pilot where the online policy in Eq. 2 is modeled as a deep neural network. $3. 2. \textbf{ Observing in Object Level}$
Instead of extracting information from the whole 360° panoramic frame at each time instance, we propose to focus on foreground objects (Fig. 2(b)) for two reasons. Firstly, in sports videos, foreground objects are typically the targets to be followed. Moreover, the relative size of foreground objects is small compared to the whole panoramic image. If processing is done at the frame level, information of object fine details would be diluted. Working with object-based observations help our method extract subtle appearance and motion cues to take an action. We define object-level observation $\mathbf{V}_t^O$ as,
(3)
(4)
$$\begin{aligned}&\mathbf{V}_{t}^{O}=\{v_{1}^{O},...,v_{t}^{O}\}\\&\mathrm{where~}v_{t}^{O}\mathrm{~is~given~by~}v_{t}^{O}=\mathrm{con}_{V}(O_{t},P_{t},M_{t})\\&\mathrm{and~}O_{t}=\mathrm{con}_{H}(\{o_{t}^{i}\}),P_{t}=\mathrm{con}_{H}(\{p_{t}^{i}\}),\\&M_{t}=\mathrm{con}_{H}(\{m_{t}^{i}\}).\end{aligned}$$
(5)
(6)
Note that con$_H()$ and con$_V()$ denote horizontal and vertical concatenation of vectors, respectively. The vector
|
[
[
0.08087158203125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.62109375,
0.091552734375,
0.62109375,
0.10479736328125,
0.50244140625,
0.10479736328125
],
[
0.07843017578125,
0.106689453125,
0.303955078125,
0.1060791015625,
0.303955078125,
0.12060546875,
0.07843017578125,
0.121826171875
],
[
0.521484375,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.521484375,
0.12060546875
],
[
0.08087158203125,
0.1275634765625,
0.2548828125,
0.1275634765625,
0.2548828125,
0.1427001953125,
0.08087158203125,
0.1427001953125
],
[
0.50341796875,
0.12249755859375,
0.72216796875,
0.12249755859375,
0.72216796875,
0.1363525390625,
0.50341796875,
0.1363525390625
],
[
0.50244140625,
0.136962890625,
0.8896484375,
0.1383056640625,
0.8896484375,
0.1514892578125,
0.50244140625,
0.15087890625
],
[
0.098876953125,
0.146484375,
0.46728515625,
0.146484375,
0.46728515625,
0.1610107421875,
0.098876953125,
0.1610107421875
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.08087158203125,
0.1610107421875,
0.468017578125,
0.1610107421875,
0.468017578125,
0.1749267578125,
0.08087158203125,
0.1749267578125
],
[
0.50341796875,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.080078125,
0.175537109375,
0.465576171875,
0.175537109375,
0.465576171875,
0.189453125,
0.080078125,
0.189453125
],
[
0.50244140625,
0.184326171875,
0.65625,
0.184326171875,
0.65625,
0.1988525390625,
0.50244140625,
0.1988525390625
],
[
0.08087158203125,
0.1912841796875,
0.468017578125,
0.1912841796875,
0.468017578125,
0.2052001953125,
0.08087158203125,
0.2052001953125
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.20703125,
0.468017578125,
0.20703125,
0.468017578125,
0.2203369140625,
0.080078125,
0.2203369140625
],
[
0.5009765625,
0.212158203125,
0.8837890625,
0.21337890625,
0.8837890625,
0.2279052734375,
0.5009765625,
0.2266845703125
],
[
0.07843017578125,
0.2203369140625,
0.4697265625,
0.2215576171875,
0.4697265625,
0.236083984375,
0.07843017578125,
0.23486328125
],
[
0.50341796875,
0.2291259765625,
0.88818359375,
0.2291259765625,
0.88818359375,
0.242431640625,
0.50341796875,
0.242431640625
],
[
0.080078125,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.251953125,
0.4697265625,
0.251953125,
0.4697265625,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50244140625,
0.25830078125,
0.890625,
0.25830078125,
0.890625,
0.271484375,
0.50244140625,
0.271484375
],
[
0.080078125,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.5009765625,
0.272705078125,
0.74609375,
0.272705078125,
0.74609375,
0.285888671875,
0.5009765625,
0.285888671875
],
[
0.080078125,
0.2822265625,
0.468017578125,
0.2822265625,
0.468017578125,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.50341796875,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50341796875,
0.302978515625
],
[
0.080078125,
0.296630859375,
0.46728515625,
0.296630859375,
0.46728515625,
0.31005859375,
0.080078125,
0.31005859375
],
[
0.50244140625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.08087158203125,
0.314453125,
0.468017578125,
0.314453125,
0.468017578125,
0.324462890625,
0.08087158203125,
0.324462890625
],
[
0.5009765625,
0.318115234375,
0.8896484375,
0.318115234375,
0.8896484375,
0.33203125,
0.5009765625,
0.33203125
],
[
0.080078125,
0.32763671875,
0.468017578125,
0.32763671875,
0.468017578125,
0.341552734375,
0.080078125,
0.341552734375
],
[
0.5009765625,
0.333984375,
0.81884765625,
0.333251953125,
0.81884765625,
0.347900390625,
0.5009765625,
0.348388671875
],
[
0.080078125,
0.343505859375,
0.2197265625,
0.343505859375,
0.2197265625,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.08087158203125,
0.35986328125,
0.4697265625,
0.35986328125,
0.4697265625,
0.373779296875,
0.08087158203125,
0.373779296875
],
[
0.6259765625,
0.35986328125,
0.73388671875,
0.35986328125,
0.73388671875,
0.375,
0.6259765625,
0.375
],
[
0.869140625,
0.361083984375,
0.8935546875,
0.361083984375,
0.8935546875,
0.375,
0.869140625,
0.375
],
[
0.080078125,
0.375732421875,
0.468017578125,
0.375732421875,
0.468017578125,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.08087158203125,
0.39013671875,
0.468017578125,
0.39013671875,
0.468017578125,
0.404052734375,
0.08087158203125,
0.404052734375
],
[
0.50341796875,
0.386962890625,
0.890625,
0.386962890625,
0.890625,
0.40087890625,
0.50341796875,
0.40087890625
],
[
0.080078125,
0.40478515625,
0.471435546875,
0.40478515625,
0.471435546875,
0.41845703125,
0.080078125,
0.41845703125
],
[
0.50244140625,
0.40283203125,
0.8505859375,
0.40283203125,
0.8505859375,
0.416748046875,
0.50244140625,
0.416748046875
],
[
0.080078125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.434326171875,
0.080078125,
0.434326171875
],
[
0.61669921875,
0.4248046875,
0.74609375,
0.4267578125,
0.74609375,
0.4443359375,
0.61669921875,
0.442626953125
],
[
0.86767578125,
0.4267578125,
0.89453125,
0.4267578125,
0.89453125,
0.443115234375,
0.86767578125,
0.443115234375
],
[
0.080078125,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.50341796875,
0.453857421875,
0.8935546875,
0.453857421875,
0.8935546875,
0.468505859375,
0.50341796875,
0.468505859375
],
[
0.080078125,
0.466552734375,
0.32275390625,
0.466552734375,
0.32275390625,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50341796875,
0.47021484375,
0.89208984375,
0.47021484375,
0.89208984375,
0.484130859375,
0.50341796875,
0.484130859375
],
[
0.080078125,
0.486083984375,
0.301513671875,
0.486083984375,
0.301513671875,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.09722900390625,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.51806640625,
0.09722900390625,
0.51806640625
],
[
0.50244140625,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.51953125,
0.46728515625,
0.51953125,
0.46728515625,
0.53271484375,
0.080078125,
0.53271484375
],
[
0.5009765625,
0.51513671875,
0.890625,
0.5146484375,
0.890625,
0.52880859375,
0.5009765625,
0.52978515625
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.5009765625,
0.52978515625,
0.8203125,
0.52978515625,
0.8203125,
0.54345703125,
0.5009765625,
0.54345703125
],
[
0.51953125,
0.54541015625,
0.89208984375,
0.54541015625,
0.89208984375,
0.56005859375,
0.51953125,
0.56005859375
],
[
0.08087158203125,
0.5517578125,
0.46728515625,
0.5517578125,
0.46728515625,
0.5625,
0.08087158203125,
0.5625
],
[
0.08087158203125,
0.56640625,
0.468017578125,
0.56640625,
0.468017578125,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.50244140625,
0.56103515625,
0.89208984375,
0.56103515625,
0.89208984375,
0.57470703125,
0.50244140625,
0.57470703125
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50341796875,
0.57568359375,
0.74609375,
0.57568359375,
0.74609375,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.080078125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50244140625,
0.59033203125,
0.73876953125,
0.59033203125,
0.73876953125,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.080078125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.51953125,
0.60498046875,
0.8896484375,
0.60400390625,
0.8896484375,
0.61865234375,
0.51953125,
0.619140625
],
[
0.080078125,
0.625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63818359375,
0.080078125,
0.6396484375
],
[
0.5009765625,
0.619140625,
0.890625,
0.619140625,
0.890625,
0.6337890625,
0.5009765625,
0.6337890625
],
[
0.080078125,
0.64013671875,
0.46728515625,
0.64013671875,
0.46728515625,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.63623046875,
0.890625,
0.63623046875,
0.890625,
0.6494140625,
0.50244140625,
0.6494140625
],
[
0.07843017578125,
0.65478515625,
0.468017578125,
0.65380859375,
0.468017578125,
0.66845703125,
0.07843017578125,
0.66943359375
],
[
0.50244140625,
0.6513671875,
0.890625,
0.6513671875,
0.890625,
0.66552734375,
0.50244140625,
0.66552734375
],
[
0.08087158203125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.08087158203125,
0.6845703125
],
[
0.50341796875,
0.66650390625,
0.89208984375,
0.66650390625,
0.89208984375,
0.6796875,
0.50341796875,
0.6796875
],
[
0.080078125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50341796875,
0.68310546875,
0.890625,
0.68310546875,
0.890625,
0.693359375,
0.50341796875,
0.693359375
],
[
0.080078125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50341796875,
0.6962890625,
0.89208984375,
0.6962890625,
0.89208984375,
0.71044921875,
0.50341796875,
0.71044921875
],
[
0.080078125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.50341796875,
0.7109375,
0.890625,
0.7109375,
0.890625,
0.724609375,
0.50341796875,
0.724609375
],
[
0.080078125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.080078125,
0.744140625
],
[
0.50244140625,
0.7265625,
0.89208984375,
0.7265625,
0.89208984375,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.07843017578125,
0.74560546875,
0.4697265625,
0.74560546875,
0.4697265625,
0.76025390625,
0.07843017578125,
0.76025390625
],
[
0.50244140625,
0.7412109375,
0.890625,
0.7412109375,
0.890625,
0.7548828125,
0.50244140625,
0.7548828125
],
[
0.080078125,
0.76123046875,
0.46728515625,
0.76123046875,
0.46728515625,
0.7744140625,
0.080078125,
0.7744140625
],
[
0.50244140625,
0.755859375,
0.59619140625,
0.755859375,
0.59619140625,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.080078125,
0.7763671875,
0.431396484375,
0.7763671875,
0.431396484375,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.54150390625,
0.77880859375,
0.67333984375,
0.78076171875,
0.67333984375,
0.798828125,
0.54150390625,
0.796875
],
[
0.86767578125,
0.7822265625,
0.89453125,
0.7822265625,
0.89453125,
0.798828125,
0.86767578125,
0.798828125
],
[
0.080078125,
0.80322265625,
0.223876953125,
0.80322265625,
0.223876953125,
0.8173828125,
0.080078125,
0.8173828125
],
[
0.5439453125,
0.80322265625,
0.85302734375,
0.80322265625,
0.85302734375,
0.8173828125,
0.5439453125,
0.8173828125
],
[
0.86669921875,
0.8037109375,
0.89208984375,
0.8037109375,
0.89208984375,
0.8173828125,
0.86669921875,
0.8173828125
],
[
0.09722900390625,
0.8193359375,
0.468017578125,
0.818359375,
0.468017578125,
0.83349609375,
0.09722900390625,
0.833984375
],
[
0.5390625,
0.8212890625,
0.81884765625,
0.82275390625,
0.81884765625,
0.84033203125,
0.5390625,
0.8388671875
],
[
0.86669921875,
0.82373046875,
0.89453125,
0.82373046875,
0.89453125,
0.84033203125,
0.86669921875,
0.84033203125
],
[
0.080078125,
0.833984375,
0.468017578125,
0.83544921875,
0.468017578125,
0.849609375,
0.080078125,
0.8486328125
],
[
0.5400390625,
0.84130859375,
0.67724609375,
0.84326171875,
0.67724609375,
0.861328125,
0.5400390625,
0.859375
],
[
0.080078125,
0.8486328125,
0.468017578125,
0.85107421875,
0.468017578125,
0.86572265625,
0.080078125,
0.86279296875
],
[
0.86767578125,
0.84326171875,
0.89453125,
0.84326171875,
0.89453125,
0.85986328125,
0.86767578125,
0.85986328125
],
[
0.080078125,
0.8662109375,
0.26953125,
0.8662109375,
0.26953125,
0.8798828125,
0.080078125,
0.8798828125
],
[
0.50341796875,
0.87060546875,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.096435546875,
0.88525390625,
0.2459716796875,
0.88623046875,
0.2459716796875,
0.90087890625,
0.096435546875,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"piloting concerns which viewing angle to select for each",
"3.1. Definitions",
"panoramicframe in a 360video.",
"We formulate the 360 piloting task as the following on-",
"2.2. Saliency Detection",
"line viewing angle selection task.",
"Observation. At time t, the agent observes a new frame t",
"Many methods have been proposed to detect salient re-",
"which is the t-th frame of the 3600 video. The sequence of",
"gions typically measured by human gaze. [35, 21, 1, 46,",
"frames that the agent has observed up to this time is referred",
"59, 64, 46] focused on detecting salient regions on images.",
"to as Vt = [Vi, .., vt].",
"Recently, [34, 24, 9, 44, 5, 61, 60,57] leveraged deep learn-",
"Goal. The goal of the agent is to select a viewing angle",
"ing and achieved significant performance gain. For videos,",
"lt at time t so that the sequence of viewing angles Lt",
"[10,20, 37, 52,41,29] relied on low-level appearance and",
"{li, .., lt] smoothly capture events of interest in the 360",
"motion cues as inputs. In addition, [26, 16, 51, 39, 13]",
"video. Note that lt = (0t, Φt)is a point on the 360 viewing",
"included information about face,people, objects, or other",
"sphere,parameterized by the azimuth angle t J0,3600]",
"contexts. Note that saliency detection methods do not se-",
"and elevation angle Φt E[-900,90]",
"lect views directly, but output a saliency score map. Our",
"Action. In order to achieve the goal, the agent takes the",
"method is also different to visual attention methods for ob-",
"action of steering the viewing angle by At at time t. Given",
"jectdetection[38,3,421inthatitconsidersviewtransition",
"the previous viewing angle lt-i and current action At, the",
"smoothness as selecting views, which is crucial for video",
"current viewing angle lt is computed as follows,",
"watching experience.",
"Ranking foreground objects of interest. Since regions of",
"It = At + lt-1.",
"(1)",
"interest in sports videos are typically foreground objects,",
"[55] proposed to use an object detector [4] to extract can-",
"Online policy. We assume that the agent takes an action △t",
"didate objects of interest, then rank the saliency of these",
"at frame t according to an online policy as follows,",
"candidate objects.For 360 piloting,we propose a similar",
"At = π(Vt,Lt-1),",
"(2)",
"baseline which first detects objects using RCNN [50], then",
"select the viewing angle focusing on the most salient object",
"where the online policy depends on both the current and",
"according to a saliency detector [64].",
"previous observation Vt and previous viewing angles Lt-1.",
"2.3. Virtual Cinematography",
"This implies that the previous viewing angles affect the cur-",
"Finally, existing virtual cinematography works fo-",
"rent action similar to what a human viewer acts when view-",
"cused on camera manipulation in simple virtual environ-",
"ing a 360° sports video. Hence, the main task of 360 pilot-",
"ments/video games [8, 22, 12, 40] and did not deal with",
"ing is about learning the online policy from data.",
"In the following, we discuss various design choices of",
"theperceptiondifficultyproblem.[14,56,7,6]relaxedthe",
"assumption and controlled virtual cameras withinrestricted",
"our proposed deep360pilot where the onlinepolicy in Eq.2",
"static wide field-of-view video of a classroom, video confer-",
"is modeled as a deep neural network.",
"ence, or basketball court, where objects of interest could be",
"3.2. Observing in Object Level",
"easily extracted. In contrast, our method handles raw 3600",
"Instead of extracting information from the whole 3600",
"sports videos downloaded from YouTube! in five domains",
"panoramic frame at each time instance, we propose to focus",
"(e.g., basketball, parkour, etc.). Recently, Su et al. [53]",
"on foreground objects (Fig: 2(b)) for two reasons. Firstly,",
"also proposed handling raw 360 videos download from",
"in sports videos, foreground objects are typically the targets",
"YouTube. They referred to this problem as Pano2Vid - au-",
"to be followed.Moreover, the relative size of foreground",
"tomatic cinematography in 360 videos -- and proposed an",
"objectsissmallcomparedtothewholepanoramicimage.If",
"offline method. In contrast, we propose an online human-",
"processing is done at the frame level, information of object",
"like agent acting based on both present and previous ob-",
"fine details would be diluted. Working with object-based",
"servations. We argue that for handling streaming videos",
"observations help our method extract subtle appearance and",
"and other human-in-the-loop applications (e.g., foveated",
"motion cues to take an action. We define object-level obser-",
"rendering[45]) a human-like online agent is necessary in or-",
"vation V as,",
"der to provide more effective video-watching support.",
"VO - [vo... vO]",
"(3)",
"3.Our Approach",
"where vo is given by u = conv(Ot, Pt, Mt).",
"(4)",
"We first define the 360 piloting problem in details",
"and Ot = conH((ot), Pt = conH(p),",
"(5)",
"(Sec. 3.1). Then, we introduce our deep 360 pilot approach",
"Mt = conH(mi).",
"(Sec. 3.2-Sec. 3.6). Finally, we describe the training proce-",
"(6)",
"dure of our model (Sec. 3.7).",
"Note that con( and conv O denote horizontal and ver-",
"\"https://www.youtube.com/",
"tical concatenation of vectors, respectively. The vector"
] |
{
"arxiv_id": "1705.01759",
"language": "en",
"page_idx": 2
}
|
||
2209.01491-en-1
|
2209.01491-en-1.jpg
|
KDD'22, August 14-18, 2022, Washington, DC, USA
function of time instead of a multivariate function of other variables and time as in partial differential equation (PDE), it remains challenging to understand what variables dominate the dynamics and how do they depend on each other. In fact, if we can extract explicit forms of differential equations from data, such mathematical expressions can naturally provide us with knowledge about correlations between variables and how they evolve through time.
While the learning of explicit PDEs for interpretable time series modeling is tempting, two unresolved challenges need to be addressed. First, how do we represent the differential equations in explicit forms and how can we learn them from data? Although previous works [12, 36] have pointed out that trainable convolutional (Conv) kernels can approximate differential operators and form a PDE model for two-dimensional physical data, multivariate analyses could have much more variables (dimensions). It is computationally intractable for current methods to represent differential operations in time series. Second, many unobserved factors and unpredictable external conditions can change the dynamics. For example, many factors can switch the sequential pattern of stock prices, ranging from macroeconomic factors such as economic cycles to microscopic factors in investor behavior such as shifting investment preferences. It is unlikely that a single PDE model can adapt to the ever-changing dynamics in the real world.
In this paper, we address both questions with a new learning framework that can automatically obtain interpretable differential operators from multivariate time series data. First, we propose a novel P-block (Polynomial-block) to capture the multivariate dynamics. It frst passes the input variables through Conv layers and then fuses the output derivatives to other variables through element-wise product operation to calculate higher-dimensional differential operators. In this way, it is possible to learn the polynomial combination of target variables and any orders of derivatives to other variables and time. We prove that the learning of $P$-block can approximate any time-evolving continuous function in theory. Second, we propose a hybrid PDE model, which contains multiple PDEs that each is trained on the different time spans and sampling rates of the historical time series. A meta-learning controller is proposed to optimize the hyperparameters of this hybrid PDE model according to the current sequential pattern. This would allow the meta controller to detect and reflect the change of sequential dynamics in the time series.
To be noted, our proposed learning of differential operators can build a hybrid PDE model that fits the time series data and learn to capture the change of sequential dynamics. To summarize, our contributions are listed as follows:
$\bullet$ For the first time, we propose a novel $P$-block to learn interpretable partial differential equations from multivariate time series with unknown and ever-changing dynamics.
$\bullet$ We propose a meta-learning model that optimizes hyperparameters of the hybrid PDE model according to the current sequential pattern, which can effectively reflect the change of sequential dynamics for interpretability.
$\bullet$ Extensive experiments on a simulated dataset and three public datasets show that the proposed method has comparable accuracy to state-of-the-art models, with clear advantages in efficiency and interpretability.
Yingtao Luo et al.
$\bullet$ We give theoretical proof that the proposed model is a universal polynomial approximator with bounded estimation error on the smoothness of time series.
# 2 RELATED WORK
In this section, we review the existing works for sequential model-
ing and machine learning differential equations.
# 2.1 Time Series Forecasting
Time series forecasting has been studied for many decades [4,49]. Statistical models and deep learning models are two representative schools of methods in the literature. For example, autoregressive integrated moving average (ARIMA) [1] is a well-known statistical model for time series forecasting. ARIMA regresses on the distribution of a multivariate variable over functions, where the data values have been replaced with the difference between their values and the previous values to handle non-stationary. The State Space Models (SSM) [14] also models time series by using a state transition function to capture the dynamics and generate observations via an observation function. These models rarely consider multivariate time series forecasting, while Gaussian Process models [6] can be applied naturally to model multivariate time series data. Also, deep learning models [31, 38] are rising as the new state-of-the-arts for many time series forecasting tasks. Many of them use the sequenceto-sequence neural structures [13,48] based on recurrent neural networks [20,26,43], Convolutional neural networks [2,18],transformers [29,53,57],and graph neural networks [54]. Deep learning models are also efficient in modeling multivariate time series [59].
# 2.2 Neural Differential Equations
Neural Differential Equations represent a series of recent works that combine neural networks with differential equations to model complex dynamics. Neural Ordinary Differential Equations (Neural ODE) [9] as a family of continuous-time models is proposed to use neural networks to mimic an ODE for the dynamics and use ODE solvers to obtain the future prediction. The backpropagation can be efficiently performed by solving additional ODEs with adjoint methods. Further studies extend the capabilities of the basic neural ODE. For example, ODE-RNN [41] and GRU-ODE-Bayes [7] can model the continuous evolution of the hidden states of the RNN between two observations. Neural Jump ODE [19] learns the conditional expectation of a stochastic process that is continuous in time. Neural controlled differential equations (Neural CDEs) [23] are proposed to incorporate incoming information into the dynamics of ODEs. Neural stochastic differential equations [34,50] and neural rough differential equations [39] are also analyzed. Expert knowledge is also fused into neural ODE to guide the model optimization [40]. Although these models can capture time-dependent dynamics using univariate differentials with respect to time, the influences of other variables are not considered
# 2.3 Dynamical System Modeling
The modeling of dynamical systems by differential equations can date back to the early works of solving ODEs and PDEs by neural networks [25,28]. Recently, many new techniques to learn PDEs have been proposed [16,58]. For example, PDEs are reformualated
|
[
[
0.0849609375,
0.07574462890625,
0.33251953125,
0.07574462890625,
0.33251953125,
0.0889892578125,
0.0849609375,
0.0889892578125
],
[
0.82666015625,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08837890625,
0.82666015625,
0.08837890625
],
[
0.08660888671875,
0.10919189453125,
0.48193359375,
0.10919189453125,
0.48193359375,
0.11932373046875,
0.08660888671875,
0.11932373046875
],
[
0.5458984375,
0.106689453125,
0.9150390625,
0.106689453125,
0.9150390625,
0.12060546875,
0.5458984375,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.55615234375,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1363525390625,
0.55615234375,
0.1363525390625
],
[
0.08660888671875,
0.136962890625,
0.48046875,
0.136962890625,
0.48046875,
0.147705078125,
0.08660888671875,
0.147705078125
],
[
0.55615234375,
0.1351318359375,
0.79345703125,
0.1351318359375,
0.79345703125,
0.1483154296875,
0.55615234375,
0.1483154296875
],
[
0.0849609375,
0.1483154296875,
0.48046875,
0.1483154296875,
0.48046875,
0.1622314453125,
0.0849609375,
0.1622314453125
],
[
0.0849609375,
0.1634521484375,
0.48193359375,
0.1622314453125,
0.48193359375,
0.175537109375,
0.0849609375,
0.1767578125
],
[
0.51953125,
0.1641845703125,
0.53515625,
0.1641845703125,
0.53515625,
0.17236328125,
0.51953125,
0.17236328125
],
[
0.54150390625,
0.1622314453125,
0.69140625,
0.1622314453125,
0.69140625,
0.175537109375,
0.54150390625,
0.175537109375
],
[
0.08660888671875,
0.1787109375,
0.48046875,
0.1787109375,
0.48046875,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.51708984375,
0.179931640625,
0.9140625,
0.179931640625,
0.9140625,
0.1944580078125,
0.51708984375,
0.1944580078125
],
[
0.0849609375,
0.1912841796875,
0.483642578125,
0.1912841796875,
0.483642578125,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.5185546875,
0.1944580078125,
0.80615234375,
0.1944580078125,
0.80615234375,
0.2083740234375,
0.5185546875,
0.2083740234375
],
[
0.10296630859375,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.10296630859375,
0.2183837890625
],
[
0.0849609375,
0.2183837890625,
0.483642578125,
0.2183837890625,
0.483642578125,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.5185546875,
0.222900390625,
0.55615234375,
0.222900390625,
0.55615234375,
0.23291015625,
0.5185546875,
0.23291015625
],
[
0.54833984375,
0.2183837890625,
0.75244140625,
0.2215576171875,
0.75244140625,
0.236083984375,
0.54833984375,
0.23291015625
],
[
0.0849609375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.51708984375,
0.2392578125,
0.9140625,
0.2392578125,
0.9140625,
0.25244140625,
0.51708984375,
0.25244140625
],
[
0.08660888671875,
0.2474365234375,
0.48046875,
0.2474365234375,
0.48046875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.0849609375,
0.260009765625,
0.483642578125,
0.260009765625,
0.483642578125,
0.2734375,
0.0849609375,
0.2734375
],
[
0.5185546875,
0.255126953125,
0.91259765625,
0.255126953125,
0.91259765625,
0.26513671875,
0.5185546875,
0.26513671875
],
[
0.51708984375,
0.267578125,
0.91259765625,
0.267578125,
0.91259765625,
0.281005859375,
0.51708984375,
0.281005859375
],
[
0.08660888671875,
0.27587890625,
0.48193359375,
0.27587890625,
0.48193359375,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.51708984375,
0.27978515625,
0.9140625,
0.27978515625,
0.9140625,
0.29345703125,
0.51708984375,
0.29345703125
],
[
0.08660888671875,
0.2890625,
0.48046875,
0.2890625,
0.48046875,
0.2998046875,
0.08660888671875,
0.2998046875
],
[
0.51708984375,
0.294189453125,
0.9140625,
0.294189453125,
0.9140625,
0.30810546875,
0.51708984375,
0.30810546875
],
[
0.0849609375,
0.300537109375,
0.48193359375,
0.300537109375,
0.48193359375,
0.314453125,
0.0849609375,
0.314453125
],
[
0.5185546875,
0.31005859375,
0.91259765625,
0.31005859375,
0.91259765625,
0.32080078125,
0.5185546875,
0.32080078125
],
[
0.08660888671875,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.08660888671875,
0.32763671875
],
[
0.51708984375,
0.322509765625,
0.9140625,
0.322509765625,
0.9140625,
0.3359375,
0.51708984375,
0.3359375
],
[
0.08660888671875,
0.330810546875,
0.48046875,
0.330810546875,
0.48046875,
0.341552734375,
0.08660888671875,
0.341552734375
],
[
0.0849609375,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.51953125,
0.33837890625,
0.91259765625,
0.33837890625,
0.91259765625,
0.348388671875,
0.51953125,
0.348388671875
],
[
0.5185546875,
0.3515625,
0.91259765625,
0.3515625,
0.91259765625,
0.3623046875,
0.5185546875,
0.3623046875
],
[
0.08660888671875,
0.359130859375,
0.48046875,
0.359130859375,
0.48046875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.51708984375,
0.363037109375,
0.9140625,
0.363037109375,
0.9140625,
0.376953125,
0.51708984375,
0.376953125
],
[
0.0849609375,
0.371826171875,
0.48193359375,
0.371826171875,
0.48193359375,
0.385009765625,
0.0849609375,
0.385009765625
],
[
0.0849609375,
0.385009765625,
0.48193359375,
0.385009765625,
0.48193359375,
0.3984375,
0.0849609375,
0.3984375
],
[
0.51953125,
0.380126953125,
0.9111328125,
0.380126953125,
0.9111328125,
0.39013671875,
0.51953125,
0.39013671875
],
[
0.0849609375,
0.3984375,
0.48193359375,
0.3984375,
0.48193359375,
0.412353515625,
0.0849609375,
0.412353515625
],
[
0.5185546875,
0.393310546875,
0.91259765625,
0.393310546875,
0.91259765625,
0.404052734375,
0.5185546875,
0.404052734375
],
[
0.5185546875,
0.406005859375,
0.91259765625,
0.406005859375,
0.91259765625,
0.419189453125,
0.5185546875,
0.419189453125
],
[
0.08416748046875,
0.41357421875,
0.416748046875,
0.412353515625,
0.416748046875,
0.423583984375,
0.08416748046875,
0.4248046875
],
[
0.098876953125,
0.425537109375,
0.48193359375,
0.4248046875,
0.48193359375,
0.439453125,
0.098876953125,
0.43994140625
],
[
0.5185546875,
0.42041015625,
0.9111328125,
0.42041015625,
0.9111328125,
0.43115234375,
0.5185546875,
0.43115234375
],
[
0.51708984375,
0.43310546875,
0.9140625,
0.43310546875,
0.9140625,
0.4462890625,
0.51708984375,
0.4462890625
],
[
0.08660888671875,
0.44140625,
0.48193359375,
0.44140625,
0.48193359375,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.51708984375,
0.4462890625,
0.9140625,
0.4462890625,
0.9140625,
0.460205078125,
0.51708984375,
0.460205078125
],
[
0.0849609375,
0.453857421875,
0.48193359375,
0.453857421875,
0.48193359375,
0.46728515625,
0.0849609375,
0.46728515625
],
[
0.0849609375,
0.46728515625,
0.48193359375,
0.46728515625,
0.48193359375,
0.48095703125,
0.0849609375,
0.48095703125
],
[
0.5185546875,
0.462158203125,
0.91259765625,
0.462158203125,
0.91259765625,
0.472900390625,
0.5185546875,
0.472900390625
],
[
0.51708984375,
0.474853515625,
0.9140625,
0.474853515625,
0.9140625,
0.488037109375,
0.51708984375,
0.488037109375
],
[
0.0849609375,
0.481689453125,
0.48193359375,
0.481689453125,
0.48193359375,
0.49560546875,
0.0849609375,
0.49560546875
],
[
0.51708984375,
0.488037109375,
0.9140625,
0.488037109375,
0.9140625,
0.501953125,
0.51708984375,
0.501953125
],
[
0.08416748046875,
0.494384765625,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.08416748046875,
0.50830078125
],
[
0.08660888671875,
0.51123046875,
0.48046875,
0.51123046875,
0.48046875,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.51708984375,
0.51513671875,
0.80126953125,
0.51513671875,
0.80126953125,
0.52880859375,
0.51708984375,
0.52880859375
],
[
0.08660888671875,
0.5244140625,
0.48193359375,
0.5244140625,
0.48193359375,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.0849609375,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.51708984375,
0.53271484375,
0.91259765625,
0.53271484375,
0.91259765625,
0.546875,
0.51708984375,
0.546875
],
[
0.0849609375,
0.55029296875,
0.48193359375,
0.55029296875,
0.48193359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.51708984375,
0.54541015625,
0.91259765625,
0.54541015625,
0.91259765625,
0.5595703125,
0.51708984375,
0.5595703125
],
[
0.0849609375,
0.56494140625,
0.483642578125,
0.56494140625,
0.483642578125,
0.5791015625,
0.0849609375,
0.5791015625
],
[
0.51708984375,
0.56103515625,
0.91259765625,
0.56103515625,
0.91259765625,
0.57470703125,
0.51708984375,
0.57470703125
],
[
0.0849609375,
0.5791015625,
0.48193359375,
0.5791015625,
0.48193359375,
0.59228515625,
0.0849609375,
0.59228515625
],
[
0.51953125,
0.5771484375,
0.9111328125,
0.5771484375,
0.9111328125,
0.5869140625,
0.51953125,
0.5869140625
],
[
0.08660888671875,
0.59326171875,
0.483642578125,
0.59326171875,
0.483642578125,
0.60400390625,
0.08660888671875,
0.60400390625
],
[
0.5185546875,
0.59033203125,
0.9111328125,
0.59033203125,
0.9111328125,
0.60107421875,
0.5185546875,
0.60107421875
],
[
0.0849609375,
0.60595703125,
0.48193359375,
0.60595703125,
0.48193359375,
0.619140625,
0.0849609375,
0.619140625
],
[
0.51953125,
0.60302734375,
0.91259765625,
0.60302734375,
0.91259765625,
0.61279296875,
0.51953125,
0.61279296875
],
[
0.0849609375,
0.61865234375,
0.48046875,
0.61865234375,
0.48046875,
0.6318359375,
0.0849609375,
0.6318359375
],
[
0.51708984375,
0.6162109375,
0.9140625,
0.6162109375,
0.9140625,
0.6298828125,
0.51708984375,
0.6298828125
],
[
0.0849609375,
0.6337890625,
0.48193359375,
0.6337890625,
0.48193359375,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.51708984375,
0.6298828125,
0.9140625,
0.6298828125,
0.9140625,
0.64306640625,
0.51708984375,
0.64306640625
],
[
0.0849609375,
0.64794921875,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.51708984375,
0.64306640625,
0.9140625,
0.64453125,
0.9140625,
0.65771484375,
0.51708984375,
0.6572265625
],
[
0.08660888671875,
0.662109375,
0.25244140625,
0.662109375,
0.25244140625,
0.67236328125,
0.08660888671875,
0.67236328125
],
[
0.51708984375,
0.65771484375,
0.9140625,
0.65771484375,
0.9140625,
0.671875,
0.51708984375,
0.671875
],
[
0.101318359375,
0.67529296875,
0.48193359375,
0.67529296875,
0.48193359375,
0.689453125,
0.101318359375,
0.689453125
],
[
0.51708984375,
0.671875,
0.9140625,
0.671875,
0.9140625,
0.68505859375,
0.51708984375,
0.68505859375
],
[
0.0849609375,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.5185546875,
0.68505859375,
0.9140625,
0.68505859375,
0.9140625,
0.69873046875,
0.5185546875,
0.69873046875
],
[
0.08660888671875,
0.705078125,
0.4794921875,
0.705078125,
0.4794921875,
0.71533203125,
0.08660888671875,
0.71533203125
],
[
0.51708984375,
0.69970703125,
0.91259765625,
0.69970703125,
0.91259765625,
0.71337890625,
0.51708984375,
0.71337890625
],
[
0.0849609375,
0.71826171875,
0.295654296875,
0.71826171875,
0.295654296875,
0.72900390625,
0.0849609375,
0.72900390625
],
[
0.51708984375,
0.71337890625,
0.9140625,
0.71337890625,
0.9140625,
0.7265625,
0.51708984375,
0.7265625
],
[
0.51708984375,
0.7265625,
0.9140625,
0.7265625,
0.9140625,
0.74072265625,
0.51708984375,
0.74072265625
],
[
0.11273193359375,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.11273193359375,
0.75830078125
],
[
0.51708984375,
0.7412109375,
0.9140625,
0.7412109375,
0.9140625,
0.7548828125,
0.51708984375,
0.7548828125
],
[
0.125,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.771484375,
0.125,
0.771484375
],
[
0.515625,
0.7548828125,
0.9140625,
0.7548828125,
0.9140625,
0.76806640625,
0.515625,
0.76806640625
],
[
0.125,
0.77294921875,
0.435546875,
0.77294921875,
0.435546875,
0.7861328125,
0.125,
0.7861328125
],
[
0.5185546875,
0.76806640625,
0.9140625,
0.76806640625,
0.9140625,
0.7822265625,
0.5185546875,
0.7822265625
],
[
0.111083984375,
0.7841796875,
0.483642578125,
0.7861328125,
0.483642578125,
0.7998046875,
0.111083984375,
0.79736328125
],
[
0.5185546875,
0.78271484375,
0.9140625,
0.78271484375,
0.9140625,
0.796875,
0.5185546875,
0.796875
],
[
0.125,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.125,
0.81298828125
],
[
0.51708984375,
0.796875,
0.80810546875,
0.796875,
0.80810546875,
0.81005859375,
0.51708984375,
0.81005859375
],
[
0.125,
0.81298828125,
0.48046875,
0.814453125,
0.48046875,
0.82763671875,
0.125,
0.8271484375
],
[
0.1258544921875,
0.8291015625,
0.381591796875,
0.8291015625,
0.381591796875,
0.8388671875,
0.1258544921875,
0.8388671875
],
[
0.51708984375,
0.82373046875,
0.79150390625,
0.82373046875,
0.79150390625,
0.8369140625,
0.51708984375,
0.8369140625
],
[
0.11273193359375,
0.84033203125,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.11273193359375,
0.85400390625
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.125,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.125,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.125,
0.86865234375,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.125,
0.8818359375
],
[
0.5185546875,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.5185546875,
0.880859375
],
[
0.125,
0.88525390625,
0.324462890625,
0.88525390625,
0.324462890625,
0.89599609375,
0.125,
0.89599609375
],
[
0.5146484375,
0.88330078125,
0.9140625,
0.8818359375,
0.9140625,
0.896484375,
0.5146484375,
0.89794921875
]
] |
[
"KDD '22, August 14-18, 2022, Washington, DC, USA",
"Yingtao Luo et al.",
"function of time instead of a multivariatefunction of othervari-",
".We give theoretical proof that the proposed model is a uni-",
"ables and time as in partial differential equation (PDE), it remains",
"versal polynomial approximator with bounded estimation",
"challenging to understand what variables dominate the dynamics",
"error on the smoothness of time series.",
"and how do they depend on each other. In fact, if we can extract",
"explicitforms of differential equationsfrom data,such mathemat-",
"2",
"RELATEDWORK",
"ical expressionscannaturallyprovideuswithknowledgeabout",
"In this section, we review the existing works for sequential model-",
"correlations between variables and how they evolve through time.",
"ing and machine learning differential equations.",
"While the learning of explicit PDEs for interpretable time series",
"modeling is tempting, two unresolved challenges need to be ad-",
"2.1",
"Time Series Forecasting",
"dressed. First, how do we represent the differential equations in",
"Time series forecasting has been studied for many decades [4, 49].",
"explicitforms and how can welearn themfrom data?Although",
"previous works [12, 36] have pointed out that trainable convolu-",
"Statistical models and deep learning modelsaretwo representative",
"schools of methods in the literature. For example, autoregressive",
"tional (Conv)kernels can approximatedifferential operators and",
"integrated moving average (ARIMA) [1] is a well-known statistical",
"form aPDEmodel fortwo-dimensional physical data,multivariate",
"model for time series forecasting. ARIMA regresses on the distribu-",
"analyses could have much more variables (dimensions). It is compu",
"tion of a multivariatevariable overfunctions, where the data values",
"tationallyintractablefor current methods to represent differential",
"have been replaced with the difference between their values and",
"operations in time series.Second, many unobserved factors and",
"unpredictable external conditions can change the dynamics. For",
"thepreviousvaluestohandlenon-stationary.The StateSpaceMod-",
"els (SSM) [14] also models time series by using a state transition",
"example,manyfactorscanswitchthesequentialpatternofstock",
"function to capture the dynamics and generate observations via an",
"prices, ranging from macroeconomic factors such as economic cy-",
"cles to microscopic factors in investor behavior such as shifting",
"observationfunction.Thesemodelsrarely considermultivariate",
"investment preferences. It is unlikely that a single PDE model can",
"time series forecasting.while Gaussian Process models [6] can be",
"applied naturally to model multivariate time series data. Also, deep",
"adapt to the ever-changing dynamics in the real world.",
"In this paper, we address both questions with a new learning",
"learning models[31, 38] arerising as the new state-of-the-arts for",
"many time series forecasting tasks. Many of them use the sequence-",
"frameworkthat canautomaticallyobtaininterpretabledifferen-",
"to-sequence neural structures [13, 48] based on recurrent neural",
"tial operators from multivariate time series data. First, we propose",
"a novel P-block (Polynomial-block) to capture the multivariate",
"networks[20,26,43],Convolutional neural networks[2,18],trans-",
"formers [29, 53, 57], and graph neural networks [54]. Deep learning",
"dynamics. It first passes the input variables through Conv layers",
"models are also efficient in modeling multivariate time series [59]",
"and then fuses the output derivatives to other variables through",
"element-wiseproductoperationtocalculatehigher-dimensional",
"2.2 Neural Differential Equations",
"differential operators.Inthis way,it is possibleto learn the polyno-",
"mial combination of target variables and any orders of derivatives",
"Neural Differential Equations represent a series of recent works",
"to other variables and time. We prove that the learning of P-block",
"that combine neural networks with differential equations to model",
"can approximate any time-evolving continuous function in theory.",
"complex dynamics. Neural Ordinary Differential Equations (Neural",
"Second, we propose a hybrid PDE model, which contains multiple",
"ODE)[9]asafamilyofcontinuous-timemodelsisproposedtouse",
"PDEs that each is trained on the different time spans and sam-",
"neural networks to mimic an ODEfor the dynamics and use ODE",
"pling rates of the historical time series. A meta-learning controller",
"solverstoobtainthefutureprediction.Thebackpropagation can",
"is proposed to optimize the hyperparameters of this hybrid PDE",
"be efficiently performed by solving additional ODEs with adjoint",
"model according to the current sequential pattern.This would allow",
"methods. Further studies extend the capabilities of the basic neural",
"the meta controller to detect and reflect the change of sequential",
"ODE.For example, ODE-RNN [41] and GRU-ODE-Bayes [7] can",
"dynamics in the time series",
"model the continuous evolution of the hidden states of the RNN",
"To be noted, our proposed learning of differential operators can",
"between two observations. Neural Jump ODE [19] learns the con-",
"build a hybrid PDE model that fits the time series data and learn",
"ditional expectation of a stochastic process that is continuous in",
"to capture the change of sequential dynamics.To summarize, our",
"time. Neural controlled differential equations (Neural CDEs) [23]",
"contributions are listedas follows:",
"are proposed to incorporate incoming information into the dynam-",
"ics of ODEs. Neural stochastic differential equations [34, 50] and",
" For the first time, we propose a novel P-block to learn inter",
"neural rough differential equations [39] are also analyzed. Expert",
"pretable partial differential equations from multivariate time",
"knowledge is also fused into neural ODE to guide the model opti-",
"series with unknown and ever-changing dynamics.",
"mization [40]. Although these models can capture time-dependent",
".Wepropose a meta-learning model that optimizes hyperpa-",
"dynamics using univariate differentials with respect to time, the",
"rameters of the hybrid PDE model according to the current",
"influences of other variables are not considered.",
"sequential pattern,which can effectively reflect the change",
"of sequentialdynamicsforinterpretability",
"2.3Dynamical SystemModeling",
".Extensive experiments on a simulated dataset and three pub-",
"The modeling of dynamical systems by differential equations can",
"lic datasets show that the proposed method has comparable",
"date back to the early works of solving ODEs and PDEs by neural",
"accuracy to state-of-the-art models,with clear advantages",
"networks[25,28].Recently,manynewtechniquestolearnPDEs",
"inefficiencyand interpretability.",
"have been proposed [16, 58]. For example, PDEs are reformualated"
] |
{
"arxiv_id": "2209.01491",
"language": "en",
"page_idx": 1
}
|
||
2209.01491-en-2
|
2209.01491-en-2.jpg
|
Learning Differential Operators for Interpretable Time Series Modeling
as backward stochastic differential equations to apply deep neural networks to efficiently solve high-dimensional PDEs [17]. Moreover, physics-informed approaches [22,51,52] are extensively studied to leverage prior knowledge of PDEs as optimization constraints to solve PDEs with sparse data. In addition, black-box neural operators [24,30,37] are proposed to learn the infinite-dimensional mapping with neural networks for nonlinear PDEs. Many efforts are also made to learn PDEs with arbitrary geometry and domain [3,44]. Except for the solving of PDEs, the scientific discoveries of PDEs by symbolic learning [10,11,45] and sparse regression [8,42,56] have also been studied. To learn the nonlinear responses without fixing certain fnite difference approximations, convolutional kernels are widely used as learnable differential operators [35, 36]. However, most works discuss PDE systems on physical data without the variation of correlations between different variables through time that changes the basic PDE structure in unforeseen ways.
# 3 метноросоєт
In the following, we first give a a formal problem definition of time
series forecasting and then introduce our methods.
## 3.1 Problem Formulation
Let us assume multivariate time series $x_{1:m}=\begin{bmatrix}x_{1},...,x_{m}\end{bmatrix}$ and $y_{1:m}=[y_1,...,y_m]$, in which $x_m$ denotes the value of the correlated multivariate time series $x$ at $m$-th time stamp, $y_m$ denotes the value of target time series $y$ at $m$-th time stamp, and $t_1:m=[t_1,...,t_m]$ denotes the time stamps of the time series. We assume that each $x_{m}=[x_{m,1},...,x_{m,k}]$ comprises of $k$ variables and $x_i,j$ represents the $j$-th variable at the $i$-th time. Each $y_m$ has one variable and $y_i$ represents the variable at the $i$-th time. We aim to continuously predict the target variable $y_m+1:m+\tau$ at next $\tau$ time steps. We are interested in the following conditional distribution
(1)
$$p(y_{m+1:m+\tau}|x_{1:m},y_{1:m})_{single}=\prod\limits_{i=m+1}^{m+\tau}p(y_{i}|x_{1:i-1},y_{1:i-1}),$$
In other words, whenever we have new data, we concatenate it to the historical time series and use the time series to predict the target variable for the next time. In addition, we can also consider the multi-step prediction $p(y_{m+1:m+\tau}|x_{1:m},y_{1:m})_{multi}$ directly.
Let us assume that the target variable $y=y(x,t)$ is a multivariate function of $t$ and other variables $x.$ Let us further assume the response of a multivariate time series dynamical system as
$$\frac{\partial y}{\partial t}=F(x,t,y),$$
(2)
where $\frac{\partial y}{\partial t}$ denotes the time derivative of the target time series. with input linear, nonlinear, and differential terms to form PDE. Eq. 2 describes a dynamical system, where we assume that the time derivative is an indispensable part of the system. The equation aims at uncovering the closed-form $F$ as a partial differential equation (PDE), given time series of correlated variables and time.
KDD'22, August 14-18, 2022, Washington, DC, USA
## 3.2 Learnable Differential Block
Obtaining the numerical derivatives for constructing the derivative $\begin{array}{c}{\mathrm{terms~\frac{\partial y}{\partial t}~and~\frac{\partial y}{\partial x_{;j}}~is~crucial~to~the~equation~discovery.~Tradition-}}\\{\mathrm{u~r~i~n~a~r~i~n~a~}}\end{array}$ ally, Finite Difference (FD) can calculate the derivative by dividing
value differences by time differences. However, FD needs a complete ensemble of all possible differential terms according to prior knowledge. In multivariate time series, the various combinations of different variables lead to an exponentially growing number of differential terms, which makes it computationally intractable to include all possible terms. To handle this issue, we introduce a learnable differential block, namely, the P-block. The architecture of the $P$-block is shown in Figure 1. The input state variable to P-block would firstly be mapped to the hidden space through multiple parallel Conv layers, whose output is then fused via the element-wise product operation. A Conv layer with $1\times1$ kernel [32] is subsequently used to linearly combine multiple channels into the output to approximated $F.$ The use of Conv layers is to approximate the differential operators as previous works [12,36] point out and replace the FD with a learnable differential method. $\dot{\text{With learnable differential blocks, we can obtain the differential op-}}$ erators via training without assuming a term library to the system a priori. Here, a P-block seeks to approximate the function $F$ via polynomial combinations of target variable $y$ and its differential terms, given by
$$F(x,t,y)=\sum_{c=1}^{N_c}f_c\cdot[\prod_{l=1}^{N_l}\prod_{j=1}^k(K_{(c,l)}\otimes y)/(K_{(c,l)}\otimes x_j)T],$$
where N$_c$ and $N_l$ denote the number of channels and parallel Conv layers respectively. $\otimes$ denotes the Conv operation with $N\times1$ kernel. $(c,l)$ indicates the Conv filter of $K$ of $l$-th layer and $c$-th channel. $f_c$ denotes the weights in $1\times1$ Conv layers with bias terms omitted. $T=Dropout(t)$ can introduce time into the system when needed. The proposed block only calculates the division of $y$ differentiation by $x$ differentiation of subsequent time steps, thus it can work on different distributions of time series.
The diagram of the proposed learnable differential block is shown in Figure 1. First, the target variable $y$ and correlated variables $x$ are processed by the differential Conv operators $K_{(c,l)}$ as shown in Eq. 3. This allows the system to create differential terms such as $\partial y$ and $\partial x_{i,1}.$ Then, these terms are processed by two element-wise product operations together with $t$ to create multivariate polynomials of ${\mathrm{differential~terms~for~both~}y\mathrm{~and~}x,\mathrm{~such~as~}(\partial y^{2}),(x_{,2}\cdot\partial x_{;,1}),(t\cdot\partial^{2}y).}$ InEq. 3, the inner product operation $\prod_j$ can multiply differentials of different variables to create multivariate differentials, and the outer product operation $\prod_l$ can multiply the same differentials multiple times to create higher-order terms. Afterwards, by the division
Alternatively, we could also use other time differential operatorsoperation and weighting, we can obtain the additive combination such as $\frac {\partial ^2y}{\partial t^2}$,the second- order time derivative, as the left hand side of multiple nonlinear polynomial deri
uncover the most fitted differential terms into the PDE without constructing a large ensemble of differential terms a priori. Limited to the pair-correlation of the variables in the multivariate time series, one Conv layer can only approximate a derivative of one variable. However, the proposed element-wise product operations make it
|
[
[
0.08660888671875,
0.0770263671875,
0.419189453125,
0.0770263671875,
0.419189453125,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6650390625,
0.0738525390625,
0.9111328125,
0.07452392578125,
0.9111328125,
0.08837890625,
0.6650390625,
0.08709716796875
],
[
0.0849609375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.515625,
0.1060791015625,
0.7900390625,
0.10479736328125,
0.7900390625,
0.11932373046875,
0.515625,
0.12060546875
],
[
0.0849609375,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.125,
0.91259765625,
0.125,
0.91259765625,
0.1383056640625,
0.51708984375,
0.1383056640625
],
[
0.08660888671875,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.08660888671875,
0.1483154296875
],
[
0.51318359375,
0.1351318359375,
0.63818359375,
0.1351318359375,
0.63818359375,
0.160400390625,
0.51318359375,
0.160400390625
],
[
0.625,
0.14013671875,
0.9140625,
0.14013671875,
0.9140625,
0.154052734375,
0.625,
0.154052734375
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.5185546875,
0.1572265625,
0.9140625,
0.1572265625,
0.9140625,
0.17041015625,
0.5185546875,
0.17041015625
],
[
0.0849609375,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.0849609375,
0.1767578125
],
[
0.08660888671875,
0.1767578125,
0.48193359375,
0.1767578125,
0.48193359375,
0.1900634765625,
0.08660888671875,
0.1900634765625
],
[
0.5185546875,
0.1717529296875,
0.91259765625,
0.1717529296875,
0.91259765625,
0.1817626953125,
0.5185546875,
0.1817626953125
],
[
0.5185546875,
0.184326171875,
0.9140625,
0.184326171875,
0.9140625,
0.1976318359375,
0.5185546875,
0.1976318359375
],
[
0.08660888671875,
0.1925048828125,
0.48046875,
0.1925048828125,
0.48046875,
0.20263671875,
0.08660888671875,
0.20263671875
],
[
0.08660888671875,
0.2052001953125,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2197265625,
0.08660888671875,
0.2197265625
],
[
0.5185546875,
0.199462890625,
0.91259765625,
0.199462890625,
0.91259765625,
0.210205078125,
0.5185546875,
0.210205078125
],
[
0.08660888671875,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.08660888671875,
0.231689453125
],
[
0.51953125,
0.21337890625,
0.91259765625,
0.21337890625,
0.91259765625,
0.2235107421875,
0.51953125,
0.2235107421875
],
[
0.51708984375,
0.2259521484375,
0.9140625,
0.2259521484375,
0.9140625,
0.2392578125,
0.51708984375,
0.2392578125
],
[
0.0849609375,
0.23291015625,
0.48046875,
0.23291015625,
0.48046875,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.51708984375,
0.2392578125,
0.9140625,
0.2392578125,
0.9140625,
0.25244140625,
0.51708984375,
0.25244140625
],
[
0.08660888671875,
0.2474365234375,
0.48046875,
0.2474365234375,
0.48046875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.51953125,
0.255126953125,
0.91259765625,
0.255126953125,
0.91259765625,
0.26513671875,
0.51953125,
0.26513671875
],
[
0.0849609375,
0.2734375,
0.483642578125,
0.2734375,
0.483642578125,
0.287353515625,
0.0849609375,
0.287353515625
],
[
0.51708984375,
0.267578125,
0.91259765625,
0.267578125,
0.91259765625,
0.281005859375,
0.51708984375,
0.281005859375
],
[
0.08416748046875,
0.285888671875,
0.48046875,
0.287353515625,
0.48046875,
0.300537109375,
0.08416748046875,
0.2998046875
],
[
0.5185546875,
0.281005859375,
0.9140625,
0.281005859375,
0.9140625,
0.294189453125,
0.5185546875,
0.294189453125
],
[
0.0849609375,
0.3017578125,
0.48046875,
0.3017578125,
0.48046875,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.51953125,
0.296630859375,
0.9111328125,
0.296630859375,
0.9111328125,
0.306884765625,
0.51953125,
0.306884765625
],
[
0.08416748046875,
0.31494140625,
0.43310546875,
0.31494140625,
0.43310546875,
0.328857421875,
0.08416748046875,
0.328857421875
],
[
0.5185546875,
0.309326171875,
0.9140625,
0.309326171875,
0.9140625,
0.322509765625,
0.5185546875,
0.322509765625
],
[
0.5185546875,
0.322509765625,
0.91259765625,
0.322509765625,
0.91259765625,
0.3359375,
0.5185546875,
0.3359375
],
[
0.5185546875,
0.3359375,
0.9140625,
0.3359375,
0.9140625,
0.349853515625,
0.5185546875,
0.349853515625
],
[
0.0882568359375,
0.347900390625,
0.1094970703125,
0.347900390625,
0.1094970703125,
0.35595703125,
0.0882568359375,
0.35595703125
],
[
0.10455322265625,
0.3447265625,
0.253173828125,
0.3447265625,
0.253173828125,
0.35791015625,
0.10455322265625,
0.35791015625
],
[
0.515625,
0.349853515625,
0.91259765625,
0.348388671875,
0.91259765625,
0.363037109375,
0.515625,
0.3642578125
],
[
0.0849609375,
0.363037109375,
0.48046875,
0.363037109375,
0.48046875,
0.376953125,
0.0849609375,
0.376953125
],
[
0.5185546875,
0.365478515625,
0.91259765625,
0.365478515625,
0.91259765625,
0.375732421875,
0.5185546875,
0.375732421875
],
[
0.0849609375,
0.376953125,
0.390625,
0.376953125,
0.390625,
0.39013671875,
0.0849609375,
0.39013671875
],
[
0.51708984375,
0.37744140625,
0.9140625,
0.37744140625,
0.9140625,
0.391357421875,
0.51708984375,
0.391357421875
],
[
0.5185546875,
0.393310546875,
0.91259765625,
0.393310546875,
0.91259765625,
0.404052734375,
0.5185546875,
0.404052734375
],
[
0.08416748046875,
0.406005859375,
0.301513671875,
0.406005859375,
0.301513671875,
0.42041015625,
0.08416748046875,
0.42041015625
],
[
0.51708984375,
0.406005859375,
0.9140625,
0.406005859375,
0.9140625,
0.419189453125,
0.51708984375,
0.419189453125
],
[
0.08416748046875,
0.4248046875,
0.48193359375,
0.4248046875,
0.48193359375,
0.438232421875,
0.08416748046875,
0.438232421875
],
[
0.515625,
0.419189453125,
0.91259765625,
0.41845703125,
0.91259765625,
0.43310546875,
0.515625,
0.434326171875
],
[
0.08416748046875,
0.439453125,
0.48193359375,
0.4375,
0.48193359375,
0.451904296875,
0.08416748046875,
0.453857421875
],
[
0.5185546875,
0.43505859375,
0.61279296875,
0.43505859375,
0.61279296875,
0.4462890625,
0.5185546875,
0.4462890625
],
[
0.08660888671875,
0.455078125,
0.4794921875,
0.455078125,
0.4794921875,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.08416748046875,
0.46533203125,
0.48046875,
0.466552734375,
0.48046875,
0.48095703125,
0.08416748046875,
0.479736328125
],
[
0.62158203125,
0.46533203125,
0.63623046875,
0.46533203125,
0.63623046875,
0.47607421875,
0.62158203125,
0.47607421875
],
[
0.5498046875,
0.4765625,
0.625,
0.4765625,
0.625,
0.488037109375,
0.5498046875,
0.488037109375
],
[
0.70166015625,
0.47607421875,
0.91259765625,
0.47607421875,
0.91259765625,
0.490478515625,
0.70166015625,
0.490478515625
],
[
0.619140625,
0.48095703125,
0.66015625,
0.479248046875,
0.6611328125,
0.4892578125,
0.62109375,
0.4912109375
],
[
0.0849609375,
0.479736328125,
0.48193359375,
0.479736328125,
0.48193359375,
0.49365234375,
0.0849609375,
0.49365234375
],
[
0.08416748046875,
0.49365234375,
0.48046875,
0.492431640625,
0.48046875,
0.5068359375,
0.08416748046875,
0.50830078125
],
[
0.619140625,
0.4912109375,
0.63623046875,
0.4912109375,
0.63623046875,
0.5,
0.619140625,
0.5
],
[
0.67236328125,
0.494384765625,
0.70361328125,
0.494384765625,
0.70361328125,
0.498779296875,
0.67236328125,
0.498779296875
],
[
0.0849609375,
0.50830078125,
0.4794921875,
0.50830078125,
0.4794921875,
0.521484375,
0.0849609375,
0.521484375
],
[
0.0849609375,
0.521484375,
0.48193359375,
0.521484375,
0.48193359375,
0.53515625,
0.0849609375,
0.53515625
],
[
0.5185546875,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.5185546875,
0.53076171875
],
[
0.08416748046875,
0.5341796875,
0.48193359375,
0.53515625,
0.48193359375,
0.5498046875,
0.08416748046875,
0.548828125
],
[
0.5185546875,
0.5322265625,
0.91259765625,
0.5322265625,
0.91259765625,
0.54541015625,
0.5185546875,
0.54541015625
],
[
0.08416748046875,
0.5498046875,
0.393798828125,
0.5498046875,
0.393798828125,
0.56298828125,
0.08416748046875,
0.56298828125
],
[
0.5185546875,
0.54541015625,
0.9111328125,
0.546875,
0.9111328125,
0.55810546875,
0.5185546875,
0.556640625
],
[
0.5185546875,
0.5595703125,
0.9140625,
0.5595703125,
0.9140625,
0.57275390625,
0.5185546875,
0.57275390625
],
[
0.30224609375,
0.5771484375,
0.3203125,
0.5771484375,
0.3203125,
0.583984375,
0.30224609375,
0.583984375
],
[
0.5185546875,
0.57470703125,
0.91259765625,
0.57470703125,
0.91259765625,
0.58544921875,
0.5185546875,
0.58544921875
],
[
0.1094970703125,
0.5859375,
0.293212890625,
0.5859375,
0.293212890625,
0.60107421875,
0.1094970703125,
0.60107421875
],
[
0.3203125,
0.58544921875,
0.443603515625,
0.58544921875,
0.443603515625,
0.599609375,
0.3203125,
0.599609375
],
[
0.464111328125,
0.5869140625,
0.4794921875,
0.5869140625,
0.4794921875,
0.5966796875,
0.464111328125,
0.5966796875
],
[
0.51953125,
0.58837890625,
0.91259765625,
0.58837890625,
0.91259765625,
0.5986328125,
0.51953125,
0.5986328125
],
[
0.295654296875,
0.60302734375,
0.326904296875,
0.60302734375,
0.326904296875,
0.60986328125,
0.295654296875,
0.60986328125
],
[
0.51708984375,
0.60107421875,
0.9140625,
0.60107421875,
0.9140625,
0.6142578125,
0.51708984375,
0.6142578125
],
[
0.515625,
0.61279296875,
0.73779296875,
0.6142578125,
0.73779296875,
0.62744140625,
0.515625,
0.626953125
],
[
0.08660888671875,
0.6220703125,
0.47802734375,
0.6220703125,
0.47802734375,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.53369140625,
0.62744140625,
0.91259765625,
0.62744140625,
0.91259765625,
0.6416015625,
0.53369140625,
0.6416015625
],
[
0.0849609375,
0.63525390625,
0.48046875,
0.63525390625,
0.48046875,
0.6484375,
0.0849609375,
0.6484375
],
[
0.51708984375,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.51708984375,
0.65576171875
],
[
0.08660888671875,
0.65087890625,
0.4794921875,
0.65087890625,
0.4794921875,
0.6611328125,
0.08660888671875,
0.6611328125
],
[
0.51708984375,
0.65478515625,
0.9140625,
0.65576171875,
0.9140625,
0.67041015625,
0.51708984375,
0.66943359375
],
[
0.0849609375,
0.66357421875,
0.4599609375,
0.66357421875,
0.4599609375,
0.67822265625,
0.0849609375,
0.67822265625
],
[
0.101318359375,
0.6767578125,
0.48193359375,
0.6767578125,
0.48193359375,
0.68994140625,
0.101318359375,
0.68994140625
],
[
0.51953125,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.681640625,
0.51953125,
0.681640625
],
[
0.08416748046875,
0.68994140625,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.08416748046875,
0.7041015625
],
[
0.51708984375,
0.6845703125,
0.9140625,
0.6845703125,
0.9140625,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.51708984375,
0.69775390625,
0.9140625,
0.6962890625,
0.9140625,
0.71044921875,
0.51708984375,
0.7109375
],
[
0.0849609375,
0.705078125,
0.441162109375,
0.705078125,
0.441162109375,
0.71533203125,
0.0849609375,
0.71533203125
],
[
0.5185546875,
0.7119140625,
0.9140625,
0.7119140625,
0.9140625,
0.72607421875,
0.5185546875,
0.72607421875
],
[
0.240234375,
0.72802734375,
0.2607421875,
0.72802734375,
0.2607421875,
0.74365234375,
0.240234375,
0.74365234375
],
[
0.51708984375,
0.72607421875,
0.9140625,
0.72607421875,
0.9140625,
0.7392578125,
0.51708984375,
0.7392578125
],
[
0.253173828125,
0.73291015625,
0.330078125,
0.73486328125,
0.330078125,
0.75,
0.253173828125,
0.74755859375
],
[
0.463134765625,
0.73486328125,
0.48193359375,
0.73486328125,
0.48193359375,
0.74853515625,
0.463134765625,
0.74853515625
],
[
0.23779296875,
0.7421875,
0.25732421875,
0.7421875,
0.25732421875,
0.7548828125,
0.23779296875,
0.7548828125
],
[
0.51708984375,
0.7392578125,
0.9140625,
0.7392578125,
0.9140625,
0.75244140625,
0.51708984375,
0.75244140625
],
[
0.51708984375,
0.75390625,
0.9111328125,
0.75244140625,
0.9111328125,
0.7646484375,
0.51708984375,
0.76513671875
],
[
0.08416748046875,
0.7646484375,
0.483642578125,
0.7646484375,
0.483642578125,
0.7822265625,
0.08416748046875,
0.7822265625
],
[
0.5185546875,
0.76806640625,
0.9111328125,
0.76806640625,
0.9111328125,
0.77880859375,
0.5185546875,
0.77880859375
],
[
0.0849609375,
0.7822265625,
0.48193359375,
0.7822265625,
0.48193359375,
0.79541015625,
0.0849609375,
0.79541015625
],
[
0.51708984375,
0.78076171875,
0.91259765625,
0.78076171875,
0.91259765625,
0.79443359375,
0.51708984375,
0.79443359375
],
[
0.8408203125,
0.796875,
0.85302734375,
0.796875,
0.85302734375,
0.80029296875,
0.8408203125,
0.80029296875
],
[
0.890625,
0.796875,
0.904296875,
0.796875,
0.904296875,
0.80029296875,
0.890625,
0.80029296875
],
[
0.51708984375,
0.79736328125,
0.82275390625,
0.79736328125,
0.82275390625,
0.81103515625,
0.51708984375,
0.81103515625
],
[
0.08416748046875,
0.798828125,
0.48193359375,
0.798828125,
0.48193359375,
0.81640625,
0.08416748046875,
0.81640625
],
[
0.8408203125,
0.80810546875,
0.86376953125,
0.80810546875,
0.86376953125,
0.8125,
0.8408203125,
0.8125
],
[
0.87744140625,
0.80615234375,
0.90673828125,
0.80615234375,
0.90673828125,
0.81298828125,
0.87744140625,
0.81298828125
],
[
0.0849609375,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.53173828125,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.53173828125,
0.8271484375
],
[
0.08416748046875,
0.82568359375,
0.48193359375,
0.82763671875,
0.48193359375,
0.84130859375,
0.08416748046875,
0.8388671875
],
[
0.5185546875,
0.8291015625,
0.91259765625,
0.8291015625,
0.91259765625,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.51708984375,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.428955078125,
0.88330078125,
0.428955078125,
0.896484375,
0.0849609375,
0.896484375
],
[
0.51708984375,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"Learning Differential Operators for Interpretable Time Series Modeling",
"KDD22, August 1418, 2022, Washington, DC, USA",
"as backward stochastic differential equations to apply deep neural",
"3.2 LearnableDifferentialBlock",
"networks to efficiently solve high-dimensional PDEs [17].Moreover,",
"Obtaining the numerical derivatives for constructing the derivative",
"physics-informed approaches [22, 51, 52] are extensively studied",
"terms 9% and Sy",
" is crucial to the equation discovery. Tradition-",
"to leverage prior knowledge of PDEs as optimization constraints to",
"ally, Finite Difference (FD) can calculate the derivative by dividing",
"solve PDEs with sparse data. In addition, black-box neural operators",
"[24, 30, 37] are proposed to learn the infinite-dimensional mapping",
"value differences bytime differences.However,FD needs a com-",
"plete ensemble of all possible differential terms according to prior",
"withneural networksfornonlinearPDEs.Manyeffortsarealso",
"made to learn PDEs with arbitrary geometry and domain [3, 44]]",
"knowledge.In multivariatetime series,thevarious combinations",
"Exceptfor the solving of PDEs, the scientific discoveries of PDEs by",
"ofdifferentvariablesleadtoanexponentiallygrowingnumber",
"of differential terms, which makes it computationally intractable",
"symbolic learning [10, 11, 45] and sparse regression [8, 42, 56] have",
"to include all possible terms. To handle this issue, we introduce",
"also been studied.Tolearn the nonlinear responses without fixing",
"certain finite difference approximations, convolutional kernels are",
"alearnabledifferential block,namely,theP-block.Thearchitec",
"widely used as learnable differential operators [35, 36]. However,",
"ture of the P-block is shown in Figure 1. The input state variable",
"most works discuss PDE systems on physical data without the",
"to P-block would firstly be mapped to the hidden space through",
"variation of correlations between different variables through time",
"multipleparallelConvlayers,whoseoutputisthenfused viathe",
"that changes the basic PDE structure in unforeseen ways.",
"element-wise product operation. A Conv layer with 1 × 1 kernel",
"[32] is subsequently used to linearly combine multiple channels",
"into the output to approximated F. The use of Conv layers is to",
"3",
"METHODOLOGY",
"approximate the differential operators as previous works [12, 36]",
"In the following, we first give a a formal problem definition of time",
"pointoutand replacetheFDwithalearnabledifferential method",
"series forecasting and then introduce our methods.",
"With learnable differential blocks, we can obtain the differential op-",
"erators via training without assuming a term library to the system",
"3.1 ProblemFormulation",
"a priori. Here, a P-block seeks to approximate the function F via",
"Let us assume multivariate time series xi:m = [xi,...Xm] and",
"polynomial combinations of target variable y and its differential",
"yi:m = [y.,. ym], in which xm denotes the value of the correlated",
"terms,givenby",
"multivariatetimeseries x at m-thtime stamp,ymdenotes thevalue",
"of target time series y at m-th time stamp, and ti:m = [ti.., m]",
"N.",
"F(x,t,y) =",
"[(K(c,D) ③y)/(K(c,I) ?xj)T],(3)",
"2fe:",
"denotes the time stamps of the time series. We assume that each",
"xm = [xm,],.., xm.k] comprises of k variables and xi.j represents",
"c=1",
"=11=",
"the j-th variable at the i-th time. Each ym has one variable and yi",
"represents the variable at the i-th time. We aim to continuously",
"where Nc and N; denote the number of channels and parallel Conv",
"predict the target variable ym+1:m+r at next t time steps. We are",
"layers respectively.denotes the Conv operation with N×1 kernel",
"interested in the following conditional distribution",
"(c.I)indicatestheConvfilterof K of I-thlayer and c-th channel.fe",
"denotes the weights in 1 x 1 Conv layers with bias terms omitted.",
"m+",
"T =Dropout(t)can introduce time into the system when needed",
"p(ym+1:m+/x1:m, y1:m)single =",
"p(yi|x1:i-1, y1:i-1),",
"(1)",
"Theproposedblockonlycalculatesthedivisionofydifferentiation",
"i=m+1",
"by x differentiation of subsequent time steps, thus it can work on",
"different distributions of time series.",
"In otherwords,wheneverwehavenew data.we concatenateit",
"Thediagram ofthe proposed learnable differential block is shown",
"to the historical time series and use the time series to predict the",
"in Figure 1. First, the target variable y and correlated variables x are",
"target variable for the next time. In addition, we can also consider",
"processed by the differential Conv operators K(c.) as shown in Eq.",
"the multi-step prediction p(ym+1:m+/x1:m.y1:m)multi directly.",
"Let us assume that the target variable y = y(x, t) is a multivari-",
"3.This allows the system to create differential terms such as y and",
"ate function of t and other variables x. Let us further assume the",
"Ox:,i. Then, these terms are processed by two element-wise product",
"operations togetherwitht to createmultivariatepolynomials of",
"response of a multivariatetimeseriesdynamical systemas",
"differential terms for both y and x, such as (ay2), (x:2-ax:,1), (ty).",
"ay",
"In Eq. 3, the inner product operation I, can multiply differentials of",
"= F(x,t,y),",
"(2)",
"at",
"different variables to create multivariate differentials, and the outer",
"productoperationcanmultiplythesamedifferentialsmultiple",
"where u denotes the time derivative of the target time series.",
"times to create higher-order terms. Afterwards, by the division",
"Alternatively, we could also use other time differential operators",
"operation and weighting, we can obtain the additive combination",
"2",
"21",
"of multiple nonlinear polynomial derivatives like t .",
"such as y, the second-order time derivative, as the left hand side",
"1ox",
"ox.",
"of this equation.F is a function parameterizing the system dynamics",
"Due to the learnable property of the proposed P-block, we can",
"with input linear,nonlinear,and differential terms toformPDE.Eq",
"uncoverthemostfitted differentialtermsintothePDEwithoutcon-",
"2 describes a dynamical system, where we assume that the time",
"structing a large ensemble of differential terms a priori. Limited to",
"derivative is an indispensable part of the system.The equation aims",
"the pair-correlation of the variables in the multivariate time series,",
"at uncovering the closed-form F as a partial differential equation",
"one Conv layer can only approximate a derivative of one variable.",
"(PDE), given time series of correlated variables and time.",
"However, the proposed element-wise product operations make it"
] |
{
"arxiv_id": "2209.01491",
"language": "en",
"page_idx": 2
}
|
||
2209.01491-en-3
|
2209.01491-en-3.jpg
|
KDD'22, August 14-18, 2022, Washington, DC, USA
Yingtao Luo et al.
<FigureHere>
Figure 1: The learning framework of P-block (left) and the dynamic PDE model (right). The P-block can take the multivariate time series as input to generate the function $F\textbf{that parameterizes the time- evolving system dynamics, which can iteratively}$ forecast future points via time spacing. The meta-learning controller takes various combinations of time series and hyperparameters as inputs to learn to capture the current dynamics and predict the loss. It guides us to search for the optimal
hyperparameters.
3.4 Recurrent Differential Model
possible to multiply univariate derivative terms to form multivariate derivative terms. By $P$-block, we can adapt the modeling of the dynamical system to high-dimensional multivariate time series.
## 3.3 Sparse Regression
The proposed P-block is used to construct the nonlinear function F in Eq. 2. How to utilize the proposed block for time series forecasting requires a comprehensive neural architecture. In neural ODE, such forecasting function is undertaken by a black-box ODE solver. However, directly solving PDEs can be much harder than ODEs, which calls for a different approach. Herein, we propose a recurrent differential model to use P-block for inference, which mimics the forward Euler time stepping
$$y_{i+1}=y_{i}+\hat{F}(y;\omega)\delta t,$$
(6)
Based on the proposed P-block to calculate $F$ in Eqs.(2-3), we use the sparse regression algorithm to optimize the model and make sure the obtained $\hat{F}$ has a concise form. The motivation to have a sparse representation of $\hat{F}$ is to maintain an interpretable mathematical expression of the PDE. Without this constraint, the discovered PDE model will take up an enormously large equation space with many derivatives in the function. Moreover, when the orders of derivatives are too high, the differential accuracy also drops drastically, which adds extra errors to the model. Therefore, the sparse regression roots on the observation that only a few terms can reflect the major trend, and many other terms that are not influential enough are often considered trivial. Let us denote the number of terms in $F$ as $n$,we have $\frac{\partial y}{\partial t}\in R^{m\times1}.$ We can concatenate all the candidate terms in $F$ and form a candidate matrix $X\in R^{m\times n}.$We can denote the sparse coefficient vector as $W\in R^{n\times1}$ and reformulate Eq.2 to a sparse representation of the dynamical system as
where $\delta t=t_{i+1}-t_t$ denotes the time difference of the two adjacent target variables in the time series. $\omega$ denotes the trained parameters in the P-block to approximate $F.$ As shown in Figure 1, the time stepping together with the identity mapping of previous $y_i$ can obtain the estimated $y_{i+1}$ to realize Eq. 6. For the multi-step time series, we can stack one $P$-block for multiple times for $y_i\to y_{i+1}$, $y_{i+1}\to y_{i+2}$,etc., as shown in Figure 1. In this way, we can use $F$ as a nonlinear differential operator on the variables to mimic each time's variation throughout the entire time series and make future predictions. It should be noted that the accuracy of approximating F is at the core of reducing forecasting errors, which will accumulate when stacking predicted future time series to historical time series and further making multi-step predictions.
$$\frac{\partial y}{\partial t}=XW.$$
(4)
## 3.5 Hybrid and Dynamic PDE Model with Meta-learning Mechanism
To compute a proper $W$ so that it is both accurate in regression and sparse enough to infer interpretable mathematical insights, here we use the accelerated proximal gradient (FISTA) approximation of LASSO regression. The overall optimization of the coefficient vector accords to the following criteria:
$$W^*=\underset{W}{\mathrm{argmin}}\left\|\frac{\partial y}{\partial t}-XW\right\|_2^2+\lambda\left\|W\right\|_1,$$
(5)
Many observational studies have informed us that the sequential dynamics of most time series are not fixed, but have an ever-changing pattern. When some unobserved nontrivial variables or some external conditions change, the dynamics will change unpredictably. This phenomenon is sometimes referred to as concept drift, or outof-distribution shift, meaning that the test set does not necessarily share the same data distribution with training data. To tackle this issue, we first assume that the dynamics within a shorter period are generally smoother than the dynamics of the longer period. In
where $\|W\|_1$ measures the sparsity of coefficient vector, i.e. the sparsity of differential terms in the PDE. \lambda is another coefficient to balance the regression errors and sparsity, which can be adjusted flexibly due to the preference of sparsity or accuracy.
|
[
[
0.08416748046875,
0.0738525390625,
0.330810546875,
0.07452392578125,
0.330810546875,
0.08837890625,
0.08416748046875,
0.08709716796875
],
[
0.82666015625,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08837890625,
0.82666015625,
0.08837890625
],
[
0.43701171875,
0.11297607421875,
0.4755859375,
0.11553955078125,
0.473876953125,
0.1275634765625,
0.435546875,
0.1256103515625
],
[
0.69775390625,
0.12371826171875,
0.72216796875,
0.12371826171875,
0.72216796875,
0.126953125,
0.69775390625,
0.126953125
],
[
0.830078125,
0.121826171875,
0.859375,
0.121826171875,
0.859375,
0.1331787109375,
0.830078125,
0.1331787109375
],
[
0.82421875,
0.1319580078125,
0.86669921875,
0.1319580078125,
0.86669921875,
0.143310546875,
0.82421875,
0.143310546875
],
[
0.44677734375,
0.14453125,
0.473876953125,
0.146484375,
0.47216796875,
0.1546630859375,
0.4453125,
0.15283203125
],
[
0.8203125,
0.1427001953125,
0.869140625,
0.1427001953125,
0.869140625,
0.154052734375,
0.8203125,
0.154052734375
],
[
0.81396484375,
0.1666259765625,
0.84619140625,
0.1666259765625,
0.84619140625,
0.177978515625,
0.81396484375,
0.177978515625
],
[
0.176513671875,
0.1483154296875,
0.2017822265625,
0.15087890625,
0.200927734375,
0.1590576171875,
0.1748046875,
0.1572265625
],
[
0.30810546875,
0.17236328125,
0.318603515625,
0.1767578125,
0.31201171875,
0.1881103515625,
0.301513671875,
0.184326171875
],
[
0.35693359375,
0.1787109375,
0.390625,
0.1787109375,
0.390625,
0.1912841796875,
0.35693359375,
0.1912841796875
],
[
0.459228515625,
0.1817626953125,
0.48046875,
0.1817626953125,
0.48046875,
0.1900634765625,
0.459228515625,
0.1900634765625
],
[
0.52392578125,
0.18115234375,
0.5498046875,
0.18115234375,
0.5498046875,
0.1944580078125,
0.52392578125,
0.1944580078125
],
[
0.82275390625,
0.1817626953125,
0.86279296875,
0.184326171875,
0.861328125,
0.1962890625,
0.82177734375,
0.1944580078125
],
[
0.435546875,
0.18310546875,
0.4453125,
0.18310546875,
0.4453125,
0.1875,
0.435546875,
0.1875
],
[
0.3056640625,
0.1849365234375,
0.314453125,
0.1849365234375,
0.314453125,
0.1962890625,
0.3056640625,
0.1962890625
],
[
0.81640625,
0.1932373046875,
0.87353515625,
0.1912841796875,
0.87353515625,
0.203857421875,
0.81640625,
0.205810546875
],
[
0.1192626953125,
0.20263671875,
0.125,
0.20263671875,
0.125,
0.2083740234375,
0.1192626953125,
0.2083740234375
],
[
0.8203125,
0.203857421875,
0.869140625,
0.203857421875,
0.869140625,
0.2139892578125,
0.8203125,
0.2139892578125
],
[
0.1805419921875,
0.21533203125,
0.2017822265625,
0.21533203125,
0.2017822265625,
0.2235107421875,
0.1805419921875,
0.2235107421875
],
[
0.8310546875,
0.212158203125,
0.859375,
0.212158203125,
0.859375,
0.2203369140625,
0.8310546875,
0.2203369140625
],
[
0.42724609375,
0.222900390625,
0.450927734375,
0.222900390625,
0.450927734375,
0.2342529296875,
0.42724609375,
0.2342529296875
],
[
0.82861328125,
0.2197265625,
0.86669921875,
0.2197265625,
0.86669921875,
0.2298583984375,
0.82861328125,
0.2298583984375
],
[
0.6298828125,
0.2266845703125,
0.65673828125,
0.2266845703125,
0.65673828125,
0.2310791015625,
0.6298828125,
0.2310791015625
],
[
0.10540771484375,
0.23486328125,
0.146240234375,
0.23486328125,
0.146240234375,
0.246826171875,
0.10540771484375,
0.246826171875
],
[
0.188720703125,
0.238037109375,
0.2017822265625,
0.2392578125,
0.2017822265625,
0.24365234375,
0.188720703125,
0.242431640625
],
[
0.35693359375,
0.240478515625,
0.384033203125,
0.240478515625,
0.384033203125,
0.2474365234375,
0.35693359375,
0.2474365234375
],
[
0.51318359375,
0.240478515625,
0.56201171875,
0.2392578125,
0.56201171875,
0.246826171875,
0.51318359375,
0.2474365234375
],
[
0.72607421875,
0.2392578125,
0.7490234375,
0.2392578125,
0.7490234375,
0.242431640625,
0.72607421875,
0.242431640625
],
[
0.3544921875,
0.2474365234375,
0.396240234375,
0.2474365234375,
0.396240234375,
0.25830078125,
0.3544921875,
0.25830078125
],
[
0.4306640625,
0.2442626953125,
0.471435546875,
0.2442626953125,
0.471435546875,
0.251953125,
0.4306640625,
0.251953125
],
[
0.51318359375,
0.248779296875,
0.56396484375,
0.248779296875,
0.56396484375,
0.255615234375,
0.51318359375,
0.255615234375
],
[
0.82421875,
0.248779296875,
0.865234375,
0.248779296875,
0.865234375,
0.2587890625,
0.82421875,
0.2587890625
],
[
0.1094970703125,
0.255126953125,
0.11767578125,
0.255126953125,
0.11767578125,
0.25830078125,
0.1094970703125,
0.25830078125
],
[
0.6640625,
0.255615234375,
0.705078125,
0.255615234375,
0.705078125,
0.267578125,
0.6640625,
0.267578125
],
[
0.82666015625,
0.2568359375,
0.86376953125,
0.2568359375,
0.86376953125,
0.267578125,
0.82666015625,
0.267578125
],
[
0.0849609375,
0.29345703125,
0.91259765625,
0.29345703125,
0.91259765625,
0.306884765625,
0.0849609375,
0.306884765625
],
[
0.0849609375,
0.306884765625,
0.91259765625,
0.306884765625,
0.91259765625,
0.32080078125,
0.0849609375,
0.32080078125
],
[
0.0849609375,
0.3212890625,
0.9140625,
0.3212890625,
0.9140625,
0.335205078125,
0.0849609375,
0.335205078125
],
[
0.08660888671875,
0.335205078125,
0.9140625,
0.335205078125,
0.9140625,
0.348388671875,
0.08660888671875,
0.348388671875
],
[
0.0849609375,
0.349853515625,
0.20751953125,
0.349853515625,
0.20751953125,
0.363037109375,
0.0849609375,
0.363037109375
],
[
0.0849609375,
0.3837890625,
0.483642578125,
0.3837890625,
0.483642578125,
0.397705078125,
0.0849609375,
0.397705078125
],
[
0.51708984375,
0.3837890625,
0.7958984375,
0.3837890625,
0.7958984375,
0.397705078125,
0.51708984375,
0.397705078125
],
[
0.0849609375,
0.3984375,
0.48193359375,
0.3984375,
0.48193359375,
0.412353515625,
0.0849609375,
0.412353515625
],
[
0.51708984375,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.51708984375,
0.415283203125
],
[
0.0849609375,
0.412353515625,
0.4697265625,
0.412353515625,
0.4697265625,
0.425537109375,
0.0849609375,
0.425537109375
],
[
0.51708984375,
0.416748046875,
0.9140625,
0.416748046875,
0.9140625,
0.429931640625,
0.51708984375,
0.429931640625
],
[
0.51708984375,
0.429931640625,
0.9140625,
0.429931640625,
0.9140625,
0.443115234375,
0.51708984375,
0.443115234375
],
[
0.08416748046875,
0.439453125,
0.273681640625,
0.43994140625,
0.273681640625,
0.455078125,
0.08416748046875,
0.453857421875
],
[
0.51708984375,
0.443115234375,
0.9150390625,
0.443115234375,
0.9150390625,
0.45703125,
0.51708984375,
0.45703125
],
[
0.08416748046875,
0.458251953125,
0.48046875,
0.45703125,
0.48046875,
0.4716796875,
0.08416748046875,
0.472900390625
],
[
0.51708984375,
0.45703125,
0.9140625,
0.45703125,
0.9140625,
0.47021484375,
0.51708984375,
0.47021484375
],
[
0.08660888671875,
0.474853515625,
0.4794921875,
0.474853515625,
0.4794921875,
0.48486328125,
0.08660888671875,
0.48486328125
],
[
0.515625,
0.4716796875,
0.9140625,
0.4716796875,
0.9140625,
0.48486328125,
0.515625,
0.48486328125
],
[
0.08416748046875,
0.486083984375,
0.48193359375,
0.4873046875,
0.48193359375,
0.50048828125,
0.08416748046875,
0.5
],
[
0.51708984375,
0.48486328125,
0.9140625,
0.48486328125,
0.9140625,
0.498779296875,
0.51708984375,
0.498779296875
],
[
0.0849609375,
0.501953125,
0.48193359375,
0.501953125,
0.48193359375,
0.51513671875,
0.0849609375,
0.51513671875
],
[
0.515625,
0.49755859375,
0.69140625,
0.498779296875,
0.69140625,
0.51318359375,
0.515625,
0.51220703125
],
[
0.08416748046875,
0.51513671875,
0.48193359375,
0.51513671875,
0.48193359375,
0.52880859375,
0.08416748046875,
0.52880859375
],
[
0.64794921875,
0.521484375,
0.779296875,
0.52099609375,
0.779296875,
0.53515625,
0.64794921875,
0.5361328125
],
[
0.89453125,
0.52294921875,
0.9150390625,
0.52294921875,
0.9150390625,
0.53515625,
0.89453125,
0.53515625
],
[
0.08660888671875,
0.53076171875,
0.48046875,
0.53076171875,
0.48046875,
0.54150390625,
0.08660888671875,
0.54150390625
],
[
0.08416748046875,
0.54248046875,
0.48046875,
0.54248046875,
0.48046875,
0.55615234375,
0.08416748046875,
0.55615234375
],
[
0.51708984375,
0.54345703125,
0.9140625,
0.54345703125,
0.9140625,
0.556640625,
0.51708984375,
0.556640625
],
[
0.08660888671875,
0.55810546875,
0.48046875,
0.55810546875,
0.48046875,
0.568359375,
0.08660888671875,
0.568359375
],
[
0.515625,
0.55810546875,
0.9140625,
0.556640625,
0.9140625,
0.57080078125,
0.515625,
0.5712890625
],
[
0.08416748046875,
0.5693359375,
0.48193359375,
0.57080078125,
0.48193359375,
0.583984375,
0.08416748046875,
0.58349609375
],
[
0.51708984375,
0.5712890625,
0.9140625,
0.5712890625,
0.9140625,
0.58544921875,
0.51708984375,
0.58544921875
],
[
0.0849609375,
0.583984375,
0.48193359375,
0.583984375,
0.48193359375,
0.59765625,
0.0849609375,
0.59765625
],
[
0.51708984375,
0.58544921875,
0.91259765625,
0.58544921875,
0.91259765625,
0.5986328125,
0.51708984375,
0.5986328125
],
[
0.0849609375,
0.599609375,
0.48046875,
0.599609375,
0.48046875,
0.60986328125,
0.0849609375,
0.60986328125
],
[
0.51708984375,
0.599609375,
0.9140625,
0.599609375,
0.9140625,
0.61279296875,
0.51708984375,
0.61279296875
],
[
0.0849609375,
0.6123046875,
0.48193359375,
0.6123046875,
0.48193359375,
0.62548828125,
0.0849609375,
0.62548828125
],
[
0.515625,
0.6123046875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.515625,
0.62548828125
],
[
0.08251953125,
0.62353515625,
0.483642578125,
0.62255859375,
0.483642578125,
0.64013671875,
0.08251953125,
0.6416015625
],
[
0.515625,
0.626953125,
0.9140625,
0.62548828125,
0.9140625,
0.64013671875,
0.515625,
0.6416015625
],
[
0.0849609375,
0.6396484375,
0.483642578125,
0.6396484375,
0.483642578125,
0.65283203125,
0.0849609375,
0.65283203125
],
[
0.5185546875,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65283203125,
0.5185546875,
0.65283203125
],
[
0.0849609375,
0.65380859375,
0.483642578125,
0.65380859375,
0.483642578125,
0.66748046875,
0.0849609375,
0.66748046875
],
[
0.5185546875,
0.65478515625,
0.91259765625,
0.65478515625,
0.91259765625,
0.66552734375,
0.5185546875,
0.66552734375
],
[
0.0849609375,
0.66845703125,
0.384033203125,
0.66845703125,
0.384033203125,
0.681640625,
0.0849609375,
0.681640625
],
[
0.51708984375,
0.66845703125,
0.91259765625,
0.66845703125,
0.91259765625,
0.681640625,
0.51708984375,
0.681640625
],
[
0.515625,
0.68115234375,
0.9140625,
0.681640625,
0.9140625,
0.69580078125,
0.515625,
0.6943359375
],
[
0.25244140625,
0.689453125,
0.275390625,
0.689453125,
0.275390625,
0.7021484375,
0.25244140625,
0.7021484375
],
[
0.268798828125,
0.69580078125,
0.314453125,
0.69580078125,
0.314453125,
0.70703125,
0.268798828125,
0.70703125
],
[
0.463134765625,
0.6962890625,
0.483642578125,
0.6962890625,
0.483642578125,
0.708984375,
0.463134765625,
0.708984375
],
[
0.515625,
0.6962890625,
0.9140625,
0.6962890625,
0.9140625,
0.71044921875,
0.515625,
0.71044921875
],
[
0.25244140625,
0.7041015625,
0.27294921875,
0.7041015625,
0.27294921875,
0.71630859375,
0.25244140625,
0.71630859375
],
[
0.51708984375,
0.71044921875,
0.775390625,
0.71044921875,
0.775390625,
0.7236328125,
0.51708984375,
0.7236328125
],
[
0.0849609375,
0.7216796875,
0.483642578125,
0.7216796875,
0.483642578125,
0.736328125,
0.0849609375,
0.736328125
],
[
0.0849609375,
0.736328125,
0.48193359375,
0.736328125,
0.48193359375,
0.75,
0.0849609375,
0.75
],
[
0.5185546875,
0.73779296875,
0.55126953125,
0.73779296875,
0.55126953125,
0.75,
0.5185546875,
0.75
],
[
0.5556640625,
0.7373046875,
0.865234375,
0.7373046875,
0.865234375,
0.75048828125,
0.5556640625,
0.75048828125
],
[
0.0849609375,
0.75048828125,
0.48193359375,
0.75048828125,
0.48193359375,
0.7646484375,
0.0849609375,
0.7646484375
],
[
0.55615234375,
0.75390625,
0.77685546875,
0.75390625,
0.77685546875,
0.76806640625,
0.55615234375,
0.76806640625
],
[
0.0849609375,
0.76318359375,
0.48046875,
0.76318359375,
0.48046875,
0.7763671875,
0.0849609375,
0.7763671875
],
[
0.51708984375,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.0849609375,
0.77978515625,
0.32275390625,
0.77978515625,
0.32275390625,
0.79052734375,
0.0849609375,
0.79052734375
],
[
0.515625,
0.78515625,
0.9140625,
0.7861328125,
0.9140625,
0.7998046875,
0.515625,
0.798828125
],
[
0.25732421875,
0.8017578125,
0.27783203125,
0.80615234375,
0.275390625,
0.81640625,
0.253173828125,
0.8125
],
[
0.3203125,
0.8017578125,
0.33251953125,
0.8017578125,
0.33251953125,
0.81005859375,
0.3203125,
0.81005859375
],
[
0.51708984375,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.1690673828125,
0.80810546875,
0.2548828125,
0.81103515625,
0.2548828125,
0.82470703125,
0.1690673828125,
0.8212890625
],
[
0.330810546875,
0.8095703125,
0.396240234375,
0.8095703125,
0.396240234375,
0.82373046875,
0.330810546875,
0.82373046875
],
[
0.464111328125,
0.806640625,
0.486083984375,
0.8125,
0.48046875,
0.82568359375,
0.459228515625,
0.82080078125
],
[
0.30224609375,
0.81494140625,
0.31201171875,
0.81494140625,
0.31201171875,
0.818359375,
0.30224609375,
0.818359375
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.82763671875,
0.51708984375,
0.82763671875
],
[
0.22216796875,
0.82470703125,
0.238525390625,
0.82568359375,
0.238525390625,
0.830078125,
0.22216796875,
0.8291015625
],
[
0.256591796875,
0.8212890625,
0.27294921875,
0.8212890625,
0.27294921875,
0.8291015625,
0.256591796875,
0.8291015625
],
[
0.3203125,
0.82568359375,
0.330078125,
0.82568359375,
0.330078125,
0.83056640625,
0.3203125,
0.83056640625
],
[
0.515625,
0.82568359375,
0.9140625,
0.82763671875,
0.9140625,
0.84130859375,
0.515625,
0.8388671875
],
[
0.08660888671875,
0.84228515625,
0.4794921875,
0.84228515625,
0.4794921875,
0.85302734375,
0.08660888671875,
0.85302734375
],
[
0.51708984375,
0.84033203125,
0.9111328125,
0.84130859375,
0.9111328125,
0.85498046875,
0.51708984375,
0.85400390625
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0849609375,
0.86865234375,
0.48193359375,
0.86865234375,
0.48193359375,
0.8818359375,
0.0849609375,
0.8818359375
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.08416748046875,
0.8818359375,
0.404296875,
0.88330078125,
0.404296875,
0.896484375,
0.08416748046875,
0.89599609375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"KDD'22, August 14-18, 2022, Washington, DC, USA",
"Yingtao Luo et al.",
"P-Block",
"o",
"Meta-",
"Learning",
"Identity",
"Controller",
"Optimizc",
"Cont",
"0",
"Conv",
"++",
"yi+1",
"Hyper-",
"X",
"1O",
"parameters:",
"x",
"1.Sampling.2.",
"Conv",
"Span3.",
"8t",
"Weighting",
"3.Weioht",
"Miltivarate",
"erat",
"Weiehtin",
"TargetVariabd",
"nre",
"getF(xy)",
"TimeSpacing",
"atif1)-thtime",
"MSELoss",
"24",
"Yfuture",
"Function",
"Figure l: The learning framework of P-block (left) and the dynamic PDE model (right). The P-block can take the multivariate",
"time series as input to generate the function F that parameterizes the time-evolving system dynamics, which can iteratively",
"forecast future points via time spacing. The meta-learning controller takes various combinations of time series and hyper-",
"parameters as inputs to learn to capture the current dynamics and predict the loss. It guides us to search for the optimal",
"hyperparameters.",
"possible to multiply univariate derivative terms to form multivari-",
"3.4 RecurrentDifferential Model",
"ate derivative terms. By P-block, we can adapt the modeling of the",
"The proposed P-block is used to construct the nonlinear function F",
"dynamical system to high-dimensional multivariate time series.",
"in Eq. 2. How to utilize the proposed block for time series forecast-",
"ing requires a comprehensive neural architecture. In neural ODE,",
"3.3SparseRegression",
"such forecasting function is undertaken by a black-box ODE solver.",
"Based on the proposed P-block to calculate F in Eqs. (2-3), we use the",
"However, directly solving PDEs can be much harder than ODEs,",
"sparseregressionalgorithm to optimize the model and make sure",
"which calls for a different approach. Herein, we propose a recurrent",
"the obtained F has a concise form.The motivation to have a sparse",
"differential model to use P-block for inference, which mimics the",
"representation of F is to maintain an interpretable mathematical",
"forward Euler time stepping",
"expression of the PDE.Without this constraint, the discovered PDE",
"yi+1 = yi + F(yw)8t,",
"(6)",
"model willtakeup an enormouslylargeequation spacewithmany",
"derivatives in the function. Moreover, when the orders of derivatives",
"where St = ti+1 - tt denotes the time difference of the two adjacent",
"aretoo high,the differential accuracyalso drops drastically,which",
"target variables in the time series. denotes the trained parameters",
"adds extra errors to themodel.Therefore,the sparse regression",
"in the P-block to approximate F. As shown in Figure 1, the time",
"roots on the observation that only a few terms can reflect the major",
"stepping together with the identity mapping of previous yi can",
"trend,andmanyothertermsthatarenot influentialenoughare",
"obtain the estimated yi+i to realize Eq. 6. For the multi-step time",
"often considered trivial. Let us denote the number of terms in F as",
"series,we can stack oneP-block formultipletimes foryi→yi+1,",
"n, we have S e Rmx1. We can concatenate all the candidate terms",
"Yi+1 yi+2, etc., as shown in Figure 1. In this way, we can use F",
"in F and form a candidate matrix X e Rmxn. We can denote the",
"asanonlineardifferentialoperatoronthevariablestomimiceach",
"sparse coefficient vector as W e Rnxl and reformulate Eq. 2 to a",
"time's variation throughout the entire time series and make future",
"sparse representation of the dynamical system as",
"predictions. It should be noted that the accuracy of approximating F",
"is at the core of reducing forecasting errors, whichwill accumulate",
"ay",
"-XW",
"(4)",
"when stacking predicted future time series to historical time series",
"at",
"and further making multi-step predictions.",
"To compute a proper W so that it is both accurate in regression and",
"sparse enough to infer interpretable mathematical insights, here",
"3.5",
"Hybrid and Dynamic PDE Model with",
"we use the accelerated proximal gradient (FISTA) approximation",
"Meta-learning Mechanism",
"of LASSO regression.The overall optimization of the coefficient",
"Many observational studies have informed us that the sequential dy-",
"vector accordstothefollowing criteria:",
"namics of most time series are notfixed, but have an ever-changing",
"lay",
"112",
"pattern. When some unobserved nontrivial variables or some ex",
"W* = argmin",
"+a IIWill1,",
"(5)",
".7",
"ternal conditions change, the dynamics will change unpredictably.",
"14",
"a",
"12",
"This phenomenon is sometimes referred to as concept drift, or out-",
"where//wll measures the sparsity of coefficientvector,i.e.the",
"of-distribution shift,meaning that thetest set does not necessarily",
"sparsity of differential terms in the PDE. A is another coefficient to",
"share the same data distribution with training data. To tackle this",
"balance the regression errors and sparsity, which can be adjusted",
"issue,wefirstassumethatthedynamicswithin a shorter period",
"flexibly due to the preference of sparsity or accuracy.",
" are generally smoother than the dynamics of the longer period. In"
] |
{
"arxiv_id": "2209.01491",
"language": "en",
"page_idx": 3
}
|
||
2209.01491-en-4
|
2209.01491-en-4.jpg
|
Learning Differential Operators for Interpretable Time Series Modeling
fact, non-stationary processes are usually treated as stationary in local areas when only a slight time difference $dt$ is considered. On the other hand, the change of external conditions or unobserved variables could have inherent patterns. For example, the periodic patterns such as the last trading day of each quarter for the stock markets have an impact on the change of dynamics. Based on the two facts, it seems wiser to design a mechanism by which we can detect the dynamics change and use different strategies for time series forecasting. Unlike neural networks with long-short memory mechanisms, a PDE as a single equation is not designed to balance the global-local dynamics. Also, it cannot capture dynamic change and make robust predictions under this scenario.
In this paper, we propose a hybrid PDE model to improve the performance of PDE models. First, we create $h$ reorganized time series by hyperparameters time spans $s=[s_1,...,s_h]$ and sampling rates $r=[r_1,...,r_h].$ By time span $s_i$,we can extract a proportion of data $[x_{m-s_i:m}]$ and $[y_{m-s_i:m}]$ from the original time series; by sampling rate $r_i$,we can extract $[x_1,x_{1+r_i},x_{1+2r_i},...]$ and $[y_{1},y_{1+r_{i}},y_{1+2r_{i}},...].$Combining the two hyperparameters, we can $\operatorname{obtain}x^{(i)}=[x_{m-s_{i}},x_{m-s_{i}+r},...],y^{(i)}=[y_{m-s_{i}},y_{m-s_{i}+r_{i}},...]$ and $t^{(i)}=[t_{m-s_{i}},t_{m-s_{i}+r_{i}},\ldots].$We propose a hybrid PDE model that consists of multiple PDEs as an weighted ensemble model to predict the time derivative, which reformulate Eq. 2 to
$$\frac{\partial y}{\partial t}=H(x,t,y,s,r,\epsilon)=\sum_{i=1}^{h}\epsilon_{i}F_{i}(x^{(i)},t^{(i)},y^{(i)}),$$
(7)
where $\epsilon_i$ is the weight coefficient, a normalized hyperparameter scaling the contribution of each PDE in the hybrid model. $\sum _{i= 1}^{h}\epsilon _{i}=$ $1. H$ represents the prediction of time derivative of the hybrid model, in contrast to $F.$ In essence, each $F_i$ is independently trained with different time series such that we can find the optimal long-short balance for the data distribution.
Furthermore, we propose a meta-learning model to capture the dynamics change and update the hyperparameters in the hybrid model dynamically. We hope to efficiently optimize the hyperparameters in real time for every prediction, instead of tuning the hyperparameters once for the entire dataset. To this end, we propose to extract latent states from the time series that reflect the change of dynamics to guide the optimization of hyperparameters in the hybrid model. We can obtain a meta-controller $G$ by
(8)
$$\underset{\theta}{\operatorname*{argmin}}G(x,t,y,s,r,\epsilon;\theta)-H(x,t,y,s,r,\epsilon;\omega)+\frac{\partial y}{\partial t},$$
where $\omega$, i.e. the model parameters of the hybrid PDE, is fixed, and we optimize $\theta$, i.e. the model parameters of the meta-learning controller. G seeks to predict the estimation error of the hybrid PDE model $H$ by time series information and hyperparameters. It is trained by minimizing the regression error of $\frac{\partial y}{\partial t}-H$ under various combinations of time series and hyperparameters. The training process of G is shown in Algorithm 1 in Appendix. When the training completes, we can use $G$ for inference to quickly search for the optimal hyperparameters in real time whenever we have new data coming into the time series. In our paper, $G$ is set to be the combination of an LSTM model processing the latent states of the time series and an MLP processing the hyperparameters
KDD'22, August 14-18, 2022, Washington, DC, USA
## 4 THEORETICAL ANALYSIS
First, we prove that the proposed P-block is a universal polynomial approximator, meaning that the proposed block can approximate any polynomials of nonlinear differential operators. To support this claim, we need to recognize that P-block roots on the numerical differentiation (i.e., forward Euler time stepping), which can guarantee the convergence of the solution under two conditions: (i) the right-hand side of Eq. 1, i.e. $F$, can be estimated accurately; and (ii) the time spacing $\delta t$ is sufficiently small. The first condition can be satisfied by the optimization of sparse regression stated by Eq. 5. The second condition can be satisfied by selecting an appropriately small $\delta t$ to reflect the time derivative.
The design of P-block is essentially the flexible combination of additive representations and multiplicative representations of differential operators. With these two representations, we can express any polynomials of differential operators that we construct on the observable variables, such as $\frac{\partial^{2}y}{\partial x_{;2}^{2}},t\cdot\frac{\partial^{2}y}{\partial x_{;1}\partial x_{;2}}.$ This can be argued by proving that each monomial can be approximated by multiplying the output of a given number of Conv layers. To this end, we provide the following theoretical proof reproduced from [35].
For $N\times1$ Conv filter $q$, define the vector of $q$ as
(9)
$$V(q)=(v_i)_{N\times1},$$
where
$$v_i=\dfrac{1}{i!}\sum_{g=-\frac{N-1}{2}}^{\frac{N-1}{2}}g^iq[g],i=0,1,...,N-1$$
(10)
By Taylor's expansion, we can formulate the convolution on any
smooth function $f$,i.e. the differential operator on $f$,as
$$\begin{aligned}L(y)&=\sum_{g=-\frac{N-1}{2}}^{\frac{N-1}{2}}q[g]f(x+g\delta x)\\&=\sum_{g=-\frac{N-1}{2}}^{\frac{N-1}{2}}q[g]\sum_{i=0}^{N-1}\prod_{j=0}^{k}\left.\frac{\partial^{i}f}{\partial^{i}x_{j}}\right|_{x}\frac{g^{i}}{i!}\delta x^{i}+o(\sum_{j=0}^{N-1}|\delta x_{j}|^{N-1})\\&=\sum_{i=0}^{N-1}v_{i}\delta x_{i}\cdot\prod_{j=0}^{k}\left.\frac{\partial^{i}f}{\partial^{i}x}\right|_{x}+o(\sum_{j=0}^{N-1}|\delta x_{j}|^{N-1}).&\text{(1}\end{aligned}$$
From Eq. 11, we can conclude that filter $q$ can approximate any differential operator with the prescribed order of accuracy by imposing a constraint on $V.$ Without loss of generality, if we consider a sufficiently large N, the proposed P-block is a universal approximator to polynomial functions. With this relaxation, we show that the proposed P-block can approximate any time-evolving continuous function in theory by the generalized Stone-Weierstrass theorem [47]. The estimation error of P-block is bounded by the minimal terms related to the variable value difference $\delta x.$ Therefore, for time series that are generally smooth, P-block can accurately approximate the desired differential operators. A special case of P-block is that if the weights in the convolutional kernels are all fixed to one, the model only calculates the polynomials of linear and nonlinear terms without differential operators.
|
[
[
0.08660888671875,
0.0770263671875,
0.420654296875,
0.0770263671875,
0.420654296875,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6640625,
0.0738525390625,
0.9140625,
0.07452392578125,
0.9140625,
0.08837890625,
0.6640625,
0.08709716796875
],
[
0.525390625,
0.10797119140625,
0.53173828125,
0.11297607421875,
0.525390625,
0.11871337890625,
0.5185546875,
0.11297607421875
],
[
0.5400390625,
0.10479736328125,
0.7646484375,
0.1060791015625,
0.7646484375,
0.12060546875,
0.5400390625,
0.11932373046875
],
[
0.0849609375,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.0882568359375,
0.10919189453125,
0.48046875,
0.10919189453125,
0.48046875,
0.11932373046875,
0.0882568359375,
0.11932373046875
],
[
0.515625,
0.125,
0.9140625,
0.12371826171875,
0.9140625,
0.1383056640625,
0.515625,
0.1395263671875
],
[
0.0849609375,
0.1337890625,
0.48046875,
0.1337890625,
0.48046875,
0.147705078125,
0.0849609375,
0.147705078125
],
[
0.51708984375,
0.1395263671875,
0.91259765625,
0.1395263671875,
0.91259765625,
0.15283203125,
0.51708984375,
0.15283203125
],
[
0.08416748046875,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.08416748046875,
0.1634521484375
],
[
0.5185546875,
0.1546630859375,
0.9111328125,
0.1546630859375,
0.9111328125,
0.1654052734375,
0.5185546875,
0.1654052734375
],
[
0.0849609375,
0.1622314453125,
0.48046875,
0.1622314453125,
0.48046875,
0.175537109375,
0.0849609375,
0.175537109375
],
[
0.51708984375,
0.1666259765625,
0.9140625,
0.1666259765625,
0.9140625,
0.179931640625,
0.51708984375,
0.179931640625
],
[
0.0849609375,
0.1767578125,
0.48193359375,
0.1767578125,
0.48193359375,
0.1900634765625,
0.0849609375,
0.1900634765625
],
[
0.515625,
0.1787109375,
0.9140625,
0.18115234375,
0.9140625,
0.1944580078125,
0.515625,
0.1925048828125
],
[
0.08660888671875,
0.1912841796875,
0.4794921875,
0.1912841796875,
0.4794921875,
0.2020263671875,
0.08660888671875,
0.2020263671875
],
[
0.5185546875,
0.1956787109375,
0.9111328125,
0.1956787109375,
0.9111328125,
0.205810546875,
0.5185546875,
0.205810546875
],
[
0.08416748046875,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.08416748046875,
0.2183837890625
],
[
0.51708984375,
0.2083740234375,
0.9140625,
0.2083740234375,
0.9140625,
0.2215576171875,
0.51708984375,
0.2215576171875
],
[
0.0849609375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.51708984375,
0.2215576171875,
0.9140625,
0.2215576171875,
0.9140625,
0.23486328125,
0.51708984375,
0.23486328125
],
[
0.08660888671875,
0.2342529296875,
0.48046875,
0.2342529296875,
0.48046875,
0.2442626953125,
0.08660888671875,
0.2442626953125
],
[
0.5185546875,
0.2374267578125,
0.91259765625,
0.2374267578125,
0.91259765625,
0.2474365234375,
0.5185546875,
0.2474365234375
],
[
0.08660888671875,
0.2474365234375,
0.48046875,
0.2474365234375,
0.48046875,
0.25830078125,
0.08660888671875,
0.25830078125
],
[
0.51708984375,
0.25,
0.9140625,
0.25,
0.9140625,
0.26318359375,
0.51708984375,
0.26318359375
],
[
0.0849609375,
0.260009765625,
0.3798828125,
0.260009765625,
0.3798828125,
0.2734375,
0.0849609375,
0.2734375
],
[
0.51708984375,
0.26318359375,
0.74267578125,
0.26318359375,
0.74267578125,
0.276611328125,
0.51708984375,
0.276611328125
],
[
0.101318359375,
0.2734375,
0.48193359375,
0.2734375,
0.48193359375,
0.287353515625,
0.101318359375,
0.287353515625
],
[
0.53173828125,
0.276611328125,
0.9140625,
0.276611328125,
0.9140625,
0.290283203125,
0.53173828125,
0.290283203125
],
[
0.0849609375,
0.285888671875,
0.48046875,
0.285888671875,
0.48046875,
0.2998046875,
0.0849609375,
0.2998046875
],
[
0.51708984375,
0.291748046875,
0.9140625,
0.291748046875,
0.9140625,
0.304931640625,
0.51708984375,
0.304931640625
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.51708984375,
0.304931640625,
0.9140625,
0.304931640625,
0.9140625,
0.318115234375,
0.51708984375,
0.318115234375
],
[
0.0849609375,
0.31640625,
0.48193359375,
0.31640625,
0.48193359375,
0.330078125,
0.0849609375,
0.330078125
],
[
0.51123046875,
0.322509765625,
0.9208984375,
0.330810546875,
0.9189453125,
0.3623046875,
0.5107421875,
0.354248046875
],
[
0.51708984375,
0.3193359375,
0.91259765625,
0.318115234375,
0.91259765625,
0.33203125,
0.51708984375,
0.333251953125
],
[
0.08416748046875,
0.330078125,
0.483642578125,
0.330078125,
0.483642578125,
0.3447265625,
0.08416748046875,
0.3447265625
],
[
0.7587890625,
0.335205078125,
0.77392578125,
0.335205078125,
0.77392578125,
0.339111328125,
0.7587890625,
0.339111328125
],
[
0.0849609375,
0.34228515625,
0.48046875,
0.34228515625,
0.48046875,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.69287109375,
0.3447265625,
0.7158203125,
0.3447265625,
0.7158203125,
0.3515625,
0.69287109375,
0.3515625
],
[
0.0849609375,
0.356689453125,
0.48193359375,
0.35595703125,
0.48193359375,
0.37060546875,
0.0849609375,
0.371826171875
],
[
0.51708984375,
0.354736328125,
0.9140625,
0.354736328125,
0.9140625,
0.36865234375,
0.51708984375,
0.36865234375
],
[
0.08251953125,
0.36865234375,
0.483642578125,
0.37060546875,
0.483642578125,
0.38818359375,
0.08251953125,
0.386474609375
],
[
0.51708984375,
0.367431640625,
0.9140625,
0.367431640625,
0.9140625,
0.380615234375,
0.51708984375,
0.380615234375
],
[
0.08251953125,
0.38330078125,
0.483642578125,
0.385009765625,
0.483642578125,
0.40283203125,
0.08251953125,
0.40087890625
],
[
0.51708984375,
0.38330078125,
0.88818359375,
0.38330078125,
0.88818359375,
0.396484375,
0.51708984375,
0.396484375
],
[
0.08416748046875,
0.40087890625,
0.48046875,
0.40087890625,
0.48046875,
0.4140625,
0.08416748046875,
0.4140625
],
[
0.53173828125,
0.396484375,
0.82421875,
0.396484375,
0.82421875,
0.410400390625,
0.53173828125,
0.410400390625
],
[
0.0849609375,
0.4140625,
0.36767578125,
0.4140625,
0.36767578125,
0.427978515625,
0.0849609375,
0.427978515625
],
[
0.6650390625,
0.41845703125,
0.7646484375,
0.42041015625,
0.7646484375,
0.43505859375,
0.6650390625,
0.43310546875
],
[
0.8935546875,
0.419189453125,
0.9140625,
0.419189453125,
0.9140625,
0.434326171875,
0.8935546875,
0.434326171875
],
[
0.1502685546875,
0.43994140625,
0.435546875,
0.438232421875,
0.435546875,
0.486083984375,
0.1502685546875,
0.488037109375
],
[
0.28515625,
0.438232421875,
0.301513671875,
0.43505859375,
0.306396484375,
0.45068359375,
0.2900390625,
0.453857421875
],
[
0.515625,
0.442626953125,
0.560546875,
0.442626953125,
0.560546875,
0.45703125,
0.515625,
0.45703125
],
[
0.463134765625,
0.45068359375,
0.483642578125,
0.45068359375,
0.483642578125,
0.46533203125,
0.463134765625,
0.46533203125
],
[
0.1446533203125,
0.443115234375,
0.1641845703125,
0.443115234375,
0.1641845703125,
0.45703125,
0.1446533203125,
0.45703125
],
[
0.1446533203125,
0.458984375,
0.1641845703125,
0.458984375,
0.1641845703125,
0.4716796875,
0.1446533203125,
0.4716796875
],
[
0.654296875,
0.4609375,
0.67724609375,
0.4609375,
0.67724609375,
0.47607421875,
0.654296875,
0.47607421875
],
[
0.6298828125,
0.472900390625,
0.640625,
0.472900390625,
0.640625,
0.484130859375,
0.6298828125,
0.484130859375
],
[
0.64892578125,
0.4716796875,
0.67626953125,
0.46728515625,
0.68310546875,
0.49560546875,
0.65625,
0.5
],
[
0.685546875,
0.479736328125,
0.8310546875,
0.479736328125,
0.8310546875,
0.494384765625,
0.685546875,
0.494384765625
],
[
0.88623046875,
0.479736328125,
0.9150390625,
0.479736328125,
0.9150390625,
0.494384765625,
0.88623046875,
0.494384765625
],
[
0.08660888671875,
0.4873046875,
0.48046875,
0.4873046875,
0.48046875,
0.49755859375,
0.08660888671875,
0.49755859375
],
[
0.60302734375,
0.474853515625,
0.64892578125,
0.4873046875,
0.640625,
0.50390625,
0.5947265625,
0.4912109375
],
[
0.08660888671875,
0.5,
0.48046875,
0.5,
0.48046875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.640625,
0.494384765625,
0.68701171875,
0.494384765625,
0.68701171875,
0.51025390625,
0.640625,
0.51025390625
],
[
0.08660888671875,
0.5146484375,
0.48193359375,
0.5146484375,
0.48193359375,
0.52783203125,
0.08660888671875,
0.52783203125
],
[
0.51708984375,
0.51806640625,
0.91259765625,
0.51806640625,
0.91259765625,
0.5322265625,
0.51708984375,
0.5322265625
],
[
0.0849609375,
0.52880859375,
0.48046875,
0.52783203125,
0.48046875,
0.5390625,
0.0849609375,
0.54052734375
],
[
0.5146484375,
0.53076171875,
0.8544921875,
0.5322265625,
0.8544921875,
0.546875,
0.5146484375,
0.54541015625
],
[
0.08660888671875,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.5537109375,
0.08660888671875,
0.5537109375
],
[
0.0849609375,
0.5546875,
0.285888671875,
0.55615234375,
0.285888671875,
0.5693359375,
0.0849609375,
0.568359375
],
[
0.58642578125,
0.5498046875,
0.61669921875,
0.5498046875,
0.61669921875,
0.59228515625,
0.58642578125,
0.59228515625
],
[
0.101318359375,
0.5693359375,
0.48046875,
0.5693359375,
0.48046875,
0.58349609375,
0.101318359375,
0.58349609375
],
[
0.53369140625,
0.5712890625,
0.580078125,
0.5712890625,
0.580078125,
0.5859375,
0.53369140625,
0.5859375
],
[
0.62158203125,
0.5712890625,
0.71826171875,
0.5712890625,
0.71826171875,
0.5859375,
0.62158203125,
0.5859375
],
[
0.0849609375,
0.583984375,
0.48193359375,
0.583984375,
0.48193359375,
0.59765625,
0.0849609375,
0.59765625
],
[
0.576171875,
0.58203125,
0.625,
0.5859375,
0.62158203125,
0.60693359375,
0.57275390625,
0.60302734375
],
[
0.0849609375,
0.5966796875,
0.48046875,
0.5966796875,
0.48046875,
0.60986328125,
0.0849609375,
0.60986328125
],
[
0.58642578125,
0.60107421875,
0.61767578125,
0.60107421875,
0.61767578125,
0.642578125,
0.58642578125,
0.642578125
],
[
0.654296875,
0.60791015625,
0.68701171875,
0.60791015625,
0.68701171875,
0.62060546875,
0.654296875,
0.62060546875
],
[
0.740234375,
0.609375,
0.83642578125,
0.6162109375,
0.833984375,
0.6416015625,
0.73779296875,
0.6337890625
],
[
0.8095703125,
0.60791015625,
0.833984375,
0.60791015625,
0.833984375,
0.619140625,
0.8095703125,
0.619140625
],
[
0.66015625,
0.6142578125,
0.6787109375,
0.62255859375,
0.66650390625,
0.63818359375,
0.64794921875,
0.6298828125
],
[
0.7060546875,
0.6142578125,
0.736328125,
0.6142578125,
0.736328125,
0.626953125,
0.7060546875,
0.626953125
],
[
0.67724609375,
0.62060546875,
0.69677734375,
0.62060546875,
0.69677734375,
0.62890625,
0.67724609375,
0.62890625
],
[
0.08660888671875,
0.61279296875,
0.48046875,
0.61279296875,
0.48046875,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.62353515625,
0.6220703125,
0.65625,
0.6220703125,
0.65625,
0.63720703125,
0.62353515625,
0.63720703125
],
[
0.826171875,
0.62060546875,
0.86767578125,
0.62060546875,
0.86767578125,
0.63525390625,
0.826171875,
0.63525390625
],
[
0.86376953125,
0.6220703125,
0.8857421875,
0.6220703125,
0.8857421875,
0.62890625,
0.86376953125,
0.62890625
],
[
0.56884765625,
0.625,
0.576171875,
0.625,
0.576171875,
0.6318359375,
0.56884765625,
0.6318359375
],
[
0.0849609375,
0.62548828125,
0.484375,
0.62548828125,
0.484375,
0.6396484375,
0.0849609375,
0.6396484375
],
[
0.6845703125,
0.62890625,
0.6953125,
0.62890625,
0.6953125,
0.63330078125,
0.6845703125,
0.63330078125
],
[
0.6953125,
0.62890625,
0.76123046875,
0.62744140625,
0.76123046875,
0.64306640625,
0.6953125,
0.64453125
],
[
0.810546875,
0.6337890625,
0.833984375,
0.6337890625,
0.833984375,
0.64794921875,
0.810546875,
0.64794921875
],
[
0.654296875,
0.6337890625,
0.701171875,
0.6337890625,
0.701171875,
0.6484375,
0.654296875,
0.6484375
],
[
0.0849609375,
0.63818359375,
0.483642578125,
0.63818359375,
0.483642578125,
0.6513671875,
0.0849609375,
0.6513671875
],
[
0.57666015625,
0.63623046875,
0.62353515625,
0.63623046875,
0.62353515625,
0.6513671875,
0.57666015625,
0.6513671875
],
[
0.08660888671875,
0.65283203125,
0.4794921875,
0.65283203125,
0.4794921875,
0.66357421875,
0.08660888671875,
0.66357421875
],
[
0.57275390625,
0.6572265625,
0.5986328125,
0.646484375,
0.611328125,
0.66552734375,
0.5849609375,
0.67529296875
],
[
0.6708984375,
0.66015625,
0.69921875,
0.66015625,
0.69921875,
0.6748046875,
0.6708984375,
0.6748046875
],
[
0.73388671875,
0.65478515625,
0.75634765625,
0.65478515625,
0.75634765625,
0.66650390625,
0.73388671875,
0.66650390625
],
[
0.08251953125,
0.66552734375,
0.43701171875,
0.66552734375,
0.43701171875,
0.6787109375,
0.08251953125,
0.6787109375
],
[
0.580078125,
0.66650390625,
0.60546875,
0.66650390625,
0.60546875,
0.6962890625,
0.580078125,
0.6962890625
],
[
0.69921875,
0.6640625,
0.82421875,
0.662109375,
0.82421875,
0.6845703125,
0.69921875,
0.68603515625
],
[
0.8837890625,
0.66650390625,
0.9150390625,
0.66650390625,
0.9150390625,
0.681640625,
0.8837890625,
0.681640625
],
[
0.654296875,
0.67041015625,
0.6787109375,
0.6767578125,
0.67333984375,
0.68798828125,
0.64892578125,
0.681640625
],
[
0.59619140625,
0.66748046875,
0.646484375,
0.66552734375,
0.64794921875,
0.68115234375,
0.59716796875,
0.68310546875
],
[
0.6708984375,
0.671875,
0.7099609375,
0.6767578125,
0.7060546875,
0.6923828125,
0.66650390625,
0.6875
],
[
0.64794921875,
0.6845703125,
0.66845703125,
0.6845703125,
0.66845703125,
0.6923828125,
0.64794921875,
0.6923828125
],
[
0.734375,
0.68115234375,
0.75634765625,
0.68115234375,
0.75634765625,
0.6943359375,
0.734375,
0.6943359375
],
[
0.1298828125,
0.6962890625,
0.416748046875,
0.6962890625,
0.416748046875,
0.71044921875,
0.1298828125,
0.71044921875
],
[
0.408447265625,
0.6923828125,
0.428955078125,
0.6875,
0.4345703125,
0.69970703125,
0.414306640625,
0.705078125
],
[
0.461669921875,
0.6962890625,
0.48193359375,
0.6962890625,
0.48193359375,
0.708984375,
0.461669921875,
0.708984375
],
[
0.414306640625,
0.70556640625,
0.4345703125,
0.70556640625,
0.4345703125,
0.71630859375,
0.414306640625,
0.71630859375
],
[
0.53369140625,
0.7021484375,
0.91259765625,
0.7021484375,
0.91259765625,
0.71533203125,
0.53369140625,
0.71533203125
],
[
0.15283203125,
0.70849609375,
0.16015625,
0.7119140625,
0.15283203125,
0.7197265625,
0.146240234375,
0.71533203125
],
[
0.5185546875,
0.71533203125,
0.9140625,
0.71533203125,
0.9140625,
0.72900390625,
0.5185546875,
0.72900390625
],
[
0.08416748046875,
0.72802734375,
0.48193359375,
0.7265625,
0.48193359375,
0.7412109375,
0.08416748046875,
0.7421875
],
[
0.51708984375,
0.72998046875,
0.9140625,
0.72998046875,
0.9140625,
0.74365234375,
0.51708984375,
0.74365234375
],
[
0.08416748046875,
0.7412109375,
0.48193359375,
0.7421875,
0.48193359375,
0.7568359375,
0.08416748046875,
0.755859375
],
[
0.51708984375,
0.744140625,
0.9140625,
0.744140625,
0.9140625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.08660888671875,
0.75830078125,
0.48046875,
0.75830078125,
0.48046875,
0.76806640625,
0.08660888671875,
0.76806640625
],
[
0.51708984375,
0.75830078125,
0.9140625,
0.75830078125,
0.9140625,
0.771484375,
0.51708984375,
0.771484375
],
[
0.0849609375,
0.77099609375,
0.48193359375,
0.771484375,
0.48193359375,
0.78271484375,
0.0849609375,
0.7822265625
],
[
0.51708984375,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.08251953125,
0.78271484375,
0.419189453125,
0.78076171875,
0.419189453125,
0.798828125,
0.08251953125,
0.80029296875
],
[
0.41259765625,
0.7861328125,
0.48046875,
0.7861328125,
0.48046875,
0.79736328125,
0.41259765625,
0.79736328125
],
[
0.51708984375,
0.7841796875,
0.9140625,
0.7841796875,
0.9140625,
0.79736328125,
0.51708984375,
0.79736328125
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.5185546875,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.5185546875,
0.81298828125
],
[
0.08416748046875,
0.81298828125,
0.48193359375,
0.8125,
0.48193359375,
0.8271484375,
0.08416748046875,
0.82763671875
],
[
0.51708984375,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.51708984375,
0.8271484375,
0.9140625,
0.8271484375,
0.9140625,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.08660888671875,
0.85693359375,
0.48046875,
0.85693359375,
0.48046875,
0.8671875,
0.08660888671875,
0.8671875
],
[
0.515625,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.515625,
0.86865234375
],
[
0.08660888671875,
0.87060546875,
0.48193359375,
0.87060546875,
0.48193359375,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.51708984375,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.0849609375,
0.88330078125,
0.451904296875,
0.88330078125,
0.451904296875,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.740234375,
0.88330078125,
0.740234375,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Learning Differential Operators for Interpretable Time Series Modeling",
"KDD'22, August 1418, 2022,Washington,DC, UA",
"A",
"THEORETICALANALYSIS",
"local areas when only a slight time difference dt is considered. On",
"fact,non-stationaryprocesses areusuallytreated asstationaryin",
"First, we prove that the proposed P-block is a universal polynomial",
"the other hand, the change of external conditions or unobserved",
"approximator, meaning that the proposed block can approximate",
"variables could have inherent patterns. For example, the periodic",
"anypolynomialsof nonlinear differential operators.Tosupportthis",
"patterns such as the last trading day of each quarter for the stock",
"claim, we need to recognize that P-block roots on the numerical",
"markets have an impact on the change of dynamics. Based on the",
"differentiation (i.e.,forward Euler time stepping),which can guar-",
"two facts, it seems wiser to design a mechanism by which we can",
"anteetheconvergenceofthe solution undertwo conditions:(i)the",
"detect the dynamics change and use different strategies for time",
"right-hand side of Eq. 1, i.e. F, can be estimated accurately; and (ii)",
"series forecasting. Unlike neural networks with long-short memory",
"the time spacing St is sufficiently small. The first condition can be",
"mechanisms,aPDEasasingleequationisnotdesigned to balance",
"satisfied bythe optimizationof sparseregression stated byEq.5",
"the global-local dynamics.Also, it cannot capture dynamic change",
"The second condition can be satisfied by selecting an appropriately",
"and make robust predictions under this scenario.",
"small St to reflect the time derivative",
"In this paper, we propose a hybrid PDE model to improve the",
"The design of P-block is essentially the flexible combination of",
"performance of PDE models. First, we create h reorganized time",
"additive representations and multiplicative representations of dif-",
"series by hyperparameters time spans s = [si, ..Sh] and sampling",
"ferential operators. With these two representations, we can express",
"rates r = [ri,... rhl. By time span si, we can extract a propor-",
"eableableshhisca eaued",
"any polynomials of differential operators that we construct on the",
"tion of data [xm-si:m] and [ym-si:m] from the original time se-",
"",
"ries; by sampling rate ri, we can extract [xi, xi+r, xi+2r..]J and",
"ax2",
"[yi,yi+ri,y1+2ri..Combining the two hyperparameters,wecan",
"by proving that each monomial can be approximated by multiply-",
"obtain x(i) = [xm-si, m-s$+,...], y(i) = [ym-st. m-si+ri..J and",
"ing the output of a given number of Conv layers. To this end, we",
"t(i) = [tm-s, tm-s$+ri..]. We propose a hybrid PDE model that",
"provide the following theoretical proof reproduced from [35].",
"consists of multiple PDEs as an weighted ensemble model to predict",
"For N × 1 Conv filter q, define the vector of q as",
"the time derivative, which reformulate Eq.2 to",
"V(q) = (ui)Nx1,",
"(9)",
"- - 6",
"h",
"where",
"(7)",
"ay",
"at",
"N",
"1",
"2",
"g'q[g],i = o, ,., N - 1",
"(10)",
"where ei istheweight coefficient, a normalized hyperparameter",
"Ui=i",
"scaling the contribution of each PDE in the hybrid model. Zh-1 ej =",
"g=-N1",
"1. H represents the prediction of time derivative of the hybrid model,",
"By Taylor's expansion, we can formulate the convolution on any",
"incontrasttoF.In essence,each Fis independentlytrained with",
"smooth function f,ie. the differential operator on f,as",
"differenttime series suchthatwe canfind the optimal long-short",
"balance for the data distribution.",
"",
"Furthermore, we propose a meta-learning model to capture the",
"L(y) =",
"q[glf(x +gx)",
"dynamics change and update the hyperparameters in the hybrid",
"9A",
"model dynamically.We hope to efficiently optimize the hyperpa",
"",
"N-1",
"%8x+o(2",
"N-1",
"S",
"a'f",
"",
"rameters inreal timefor every prediction, instead of tuning the",
"q[g] ",
"18x;",
"N-",
"一",
"hyperparameters once for the entire dataset. To this end, we pro-",
"",
"aixii",
"=0",
"= j=0",
"pose to extract latent states from the time series that reflect the",
"g-N1",
"changeof dynamics to guidetheoptimizationof hvperparameters",
"!",
"aif",
"N-1",
"in the hybrid model. We can obtain a meta-controller G by",
"Ni",
" + o(Z 18x;/N-1).",
"(11)",
"",
"oioxi*",
"atxx",
"j=0",
"j=0",
"argminG(x,t, y, s, r, e; 0) - H(x, t, y, s, r, e;o) +",
"oy",
"(8)",
"at",
"From Eq.11, we can conclude that filter g can approximate any",
"0",
"differential operator with the prescribed order of accuracy by im-",
"where , i.e. the model parameters of the hybrid PDE, is fixed.",
"posing a constraint on V. Without loss of generality, if we consider a",
"and we optimize , i.e. the model parameters of the meta-learning",
"sufficiently large N, the proposed P-block is a universal approxima-",
"controller. G seeks to predict the estimation error of thehybrid",
"tor to polynomial functions. With this relaxation, we show that the",
"PDE model Hby time series information and hyperparameters.",
"proposed P-block can approximate any time-evolving continuous",
"It is trained by minimizing the regression error of y",
"-H under",
"function in theory by the generalized Stone-Weierstrass theorem",
"various combinations of time series and hyperparameters. The",
"[47]. The estimation error of P-block is bounded by the minimal",
"training process of G is shown in Algorithm 1 in Appendix. When",
"terms related to the variable value difference Sx.Therefore,for time",
"thetrainingcompletes,wecanuseGforinferencetoquicklysearch",
"series that are generally smooth, P-block can accurately approxi-",
"for the optimal hyperparameters in real time whenever we have",
"mate the desired differential operators. A special case of P-block is",
"new data coming into the time series. In our paper, G is set to be",
"that if the weights in the convolutional kernels are all fixed to one,",
"thecombinationof anLSTMmodel processingthelatentstatesof",
"the model only calculates the polynomials of linear and nonlinear",
"the time series and an MLP processing the hyperparameters.",
"terms without differential operators."
] |
{
"arxiv_id": "2209.01491",
"language": "en",
"page_idx": 4
}
|
||
2211.14666-en-0
|
2211.14666-en-0.jpg
|
# Synergies between Disentanglement and Sparsity: Generalization and Identifiability in Multi-Task Learning
Sébastien Lachapelle $^{*1}$ Tristan Deleu $^{*1}$ Divyat Mahajan$^{1}$ Ioannis Mitliagkas$^{12}$ Yoshua Bengio$^{12}$
Simon Lacoste-Julien $^{12}$ Quentin Bertrand
## $\mathbf{Abstract}$
image generation (Ramesh et al., 2022). However, preliminary assessments of the latter have highlighted shortcomings related to compositionality (Marcus et al., 2022), suggesting new algorithmic innovations are needed.
Although disentangled representations are often said to be beneficial for downstream tasks, current empirical and theoretical understanding is limited. In this work, we provide evidence that disentangled representations coupled with sparse task-specific predictors improve generalization. In the context of multi-task learning, we prove a new identifiability result that provides conditions under which maximally sparse predictors yield disentangled representations. Motivated by this theoretical result, we propose a practical approach to learn disentangled representations based on a sparsity-promoting bi-level optimization problem. Finally, we explore a meta-learning version of this algorithm based on group Lasso multiclass SVM predictors, for which we derive a tractable dual formulation. It obtains competitive results on standard few-shot classification benchmarks, while each task is using only a fraction of the learned representations.
Another line of work has argued for the integration of ideas from causality to make progress towards more robust and transferable machine learning systems (Pearl, 2019; Schölkopf, 2019; Goyal & Bengio, 2022). Causal representation learning has emerged recently as a field aiming to define and learn representations suited for causal reasoning (Schölkopf et al., 2021). This set of ideas is strongly related to learning disentangled representations (Bengio et al., 2013). Informally, a representation is considered disentangled when its components are in one-to-one correspondence with natural and interpretable factors of variations, such as object positions, colors or shapes. Although a plethora of works have investigated theoretically under which conditions disentanglement is possible through the lens of identifiability (Hyvärinen&Morioka,2016;2017; Hyvärinen et al., 2019; Khemakhem et al., 2020a; Locatello et al., 2020a; Klindt et al., 2021; Von Kügelgen et al., 2021; Gresele et al., 2021; Lachapelle et al., 2022; Lippe et al., 2022b; Ahuja et al., 2022c), fewer works have tackled how a disentangled representation could be beneficial for downstream tasks. Those who did mainly provide empirical rather than theoretical evidence for or against its usefulness (Locatello et al., 2019; van Steenkiste et al., 2019; Miladinović et al., 2019; Dittadi et al., 2021; Montero et al., 2021). We believe our work can bring some theoretical insights as to when and why disentanglement can help.
In this work, we explore synergies between disentanglement and sparse task-specific predictors in the context of multi-task learning. At the heart of our contributions is the assumption that only a small subset of all factors of variations are useful for each downstream task, and this subset might change from one task to another. We will refer to such tasks as sparse tasks, and their corresponding sets of useful factors as their supports. This assumption was initially suggested by Bengio et al. (2013, Section 3.5):“the feature set being trained may be destined to be used in multiple tasks that may have distinct [and unknown] subsets of relevant
### $1. \textbf{Introduction}$
The recent literature on self-supervised learning has provided evidence that learning a representation on large corpuses of data can yield strong performances on a wide variety of downstream tasks (Devlin et al., 2018; Chen et al., 2020), especially in few-shot learning scenarios where the training data for these tasks is limited (Brown et al., 2020b; Dosovitskiy et al., 2021; Radford et al., 2021). Beyond transferring across multiple tasks, these learned representations also lead to improved robustness against distribution shifts (Wortsman et al., 2022) as well as stunning text-conditioned
$^{*}$Equal contribution $^{\mathrm{l}}$Mila &DIRO, Université de Montréal $^{2}$Canada CIFAR AI Chair. Correspondence to: Sébastien Lachapelle <lachaseb@mila.quebec>, Tristan Deleu <deleutri@ mila.quebec>.
Proceedings of the 40$^{th}$ International Conference on Machine $Learning$,Honolulu,$^\prime$Hawaii, USA. PMLR 202, 2023. Copyright 2023 by the author(s).
1
|
[
[
0.23779296875,
0.11236572265625,
0.73876953125,
0.11297607421875,
0.73876953125,
0.1307373046875,
0.23779296875,
0.1300048828125
],
[
0.1993408203125,
0.1351318359375,
0.77392578125,
0.1363525390625,
0.77392578125,
0.154052734375,
0.1993408203125,
0.15283203125
],
[
0.1298828125,
0.1988525390625,
0.8408203125,
0.1988525390625,
0.8408203125,
0.216552734375,
0.1298828125,
0.216552734375
],
[
0.326904296875,
0.216552734375,
0.640625,
0.21533203125,
0.640625,
0.2298583984375,
0.326904296875,
0.2310791015625
],
[
0.24267578125,
0.2568359375,
0.32177734375,
0.2568359375,
0.32177734375,
0.274658203125,
0.24267578125,
0.274658203125
],
[
0.5009765625,
0.260009765625,
0.8857421875,
0.260009765625,
0.8857421875,
0.2734375,
0.5009765625,
0.2734375
],
[
0.0228729248046875,
0.274658203125,
0.060455322265625,
0.274658203125,
0.060455322265625,
0.39453125,
0.0228729248046875,
0.39453125
],
[
0.5009765625,
0.27587890625,
0.8857421875,
0.27587890625,
0.8857421875,
0.2890625,
0.5009765625,
0.2890625
],
[
0.1217041015625,
0.2841796875,
0.441162109375,
0.2841796875,
0.441162109375,
0.29736328125,
0.1217041015625,
0.29736328125
],
[
0.5,
0.290283203125,
0.8857421875,
0.290283203125,
0.8857421875,
0.3037109375,
0.5,
0.3037109375
],
[
0.1217041015625,
0.2998046875,
0.441162109375,
0.2998046875,
0.441162109375,
0.313232421875,
0.1217041015625,
0.313232421875
],
[
0.5009765625,
0.30615234375,
0.765625,
0.30615234375,
0.765625,
0.3193359375,
0.5009765625,
0.3193359375
],
[
0.12091064453125,
0.31494140625,
0.438720703125,
0.31494140625,
0.438720703125,
0.32568359375,
0.12091064453125,
0.32568359375
],
[
0.1217041015625,
0.330078125,
0.441162109375,
0.330078125,
0.441162109375,
0.343505859375,
0.1217041015625,
0.343505859375
],
[
0.5009765625,
0.32763671875,
0.88623046875,
0.32763671875,
0.88623046875,
0.341552734375,
0.5009765625,
0.341552734375
],
[
0.1217041015625,
0.3447265625,
0.441162109375,
0.3447265625,
0.441162109375,
0.35791015625,
0.1217041015625,
0.35791015625
],
[
0.5009765625,
0.343505859375,
0.8857421875,
0.343505859375,
0.8857421875,
0.356689453125,
0.5009765625,
0.356689453125
],
[
0.12091064453125,
0.359130859375,
0.439453125,
0.359130859375,
0.439453125,
0.37255859375,
0.12091064453125,
0.37255859375
],
[
0.5009765625,
0.35791015625,
0.88623046875,
0.35791015625,
0.88623046875,
0.371826171875,
0.5009765625,
0.371826171875
],
[
0.12091064453125,
0.373779296875,
0.441162109375,
0.375,
0.441162109375,
0.389404296875,
0.12091064453125,
0.38818359375
],
[
0.5009765625,
0.373779296875,
0.88623046875,
0.373779296875,
0.88623046875,
0.386962890625,
0.5009765625,
0.386962890625
],
[
0.12091064453125,
0.39013671875,
0.441162109375,
0.39013671875,
0.441162109375,
0.404052734375,
0.12091064453125,
0.404052734375
],
[
0.5,
0.389404296875,
0.8857421875,
0.389404296875,
0.8857421875,
0.40283203125,
0.5,
0.40283203125
],
[
0.021240234375,
0.396484375,
0.056365966796875,
0.395751953125,
0.060455322265625,
0.479736328125,
0.02532958984375,
0.48095703125
],
[
0.1217041015625,
0.40478515625,
0.441162109375,
0.40478515625,
0.441162109375,
0.419189453125,
0.1217041015625,
0.419189453125
],
[
0.5009765625,
0.404052734375,
0.88623046875,
0.404052734375,
0.88623046875,
0.417236328125,
0.5009765625,
0.417236328125
],
[
0.1217041015625,
0.42041015625,
0.441162109375,
0.42041015625,
0.441162109375,
0.434326171875,
0.1217041015625,
0.434326171875
],
[
0.5009765625,
0.41845703125,
0.8837890625,
0.41845703125,
0.8837890625,
0.431884765625,
0.5009765625,
0.431884765625
],
[
0.1217041015625,
0.43505859375,
0.441162109375,
0.43505859375,
0.441162109375,
0.448974609375,
0.1217041015625,
0.448974609375
],
[
0.5,
0.434326171875,
0.88623046875,
0.434326171875,
0.88623046875,
0.447509765625,
0.5,
0.447509765625
],
[
0.12091064453125,
0.45068359375,
0.441162109375,
0.45068359375,
0.441162109375,
0.464111328125,
0.12091064453125,
0.464111328125
],
[
0.5,
0.448974609375,
0.8837890625,
0.448974609375,
0.8837890625,
0.462158203125,
0.5,
0.462158203125
],
[
0.1217041015625,
0.466552734375,
0.441162109375,
0.466552734375,
0.441162109375,
0.479736328125,
0.1217041015625,
0.479736328125
],
[
0.5009765625,
0.464111328125,
0.88623046875,
0.464111328125,
0.88623046875,
0.477783203125,
0.5009765625,
0.477783203125
],
[
0.1217041015625,
0.48095703125,
0.44287109375,
0.48095703125,
0.44287109375,
0.494384765625,
0.1217041015625,
0.494384765625
],
[
0.498291015625,
0.479248046875,
0.88623046875,
0.477783203125,
0.88623046875,
0.492431640625,
0.498291015625,
0.49365234375
],
[
0.1217041015625,
0.49560546875,
0.44287109375,
0.49560546875,
0.44287109375,
0.5087890625,
0.1217041015625,
0.5087890625
],
[
0.0228729248046875,
0.484130859375,
0.060455322265625,
0.484130859375,
0.060455322265625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.50244140625,
0.49560546875,
0.8857421875,
0.49560546875,
0.8857421875,
0.505859375,
0.50244140625,
0.505859375
],
[
0.12091064453125,
0.51025390625,
0.439453125,
0.51025390625,
0.439453125,
0.5234375,
0.12091064453125,
0.5234375
],
[
0.5009765625,
0.5087890625,
0.88623046875,
0.5087890625,
0.88623046875,
0.52294921875,
0.5009765625,
0.52294921875
],
[
0.12091064453125,
0.52587890625,
0.441162109375,
0.52587890625,
0.441162109375,
0.5390625,
0.12091064453125,
0.5390625
],
[
0.50244140625,
0.52587890625,
0.8837890625,
0.52587890625,
0.8837890625,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.12091064453125,
0.54150390625,
0.439453125,
0.54150390625,
0.439453125,
0.5517578125,
0.12091064453125,
0.5517578125
],
[
0.5009765625,
0.5390625,
0.88623046875,
0.5390625,
0.88623046875,
0.55322265625,
0.5009765625,
0.55322265625
],
[
0.12091064453125,
0.55615234375,
0.44287109375,
0.55615234375,
0.44287109375,
0.5693359375,
0.12091064453125,
0.5693359375
],
[
0.50244140625,
0.55615234375,
0.8857421875,
0.55615234375,
0.8857421875,
0.56640625,
0.50244140625,
0.56640625
],
[
0.12091064453125,
0.5712890625,
0.279296875,
0.5712890625,
0.279296875,
0.58544921875,
0.12091064453125,
0.58544921875
],
[
0.50244140625,
0.5693359375,
0.88623046875,
0.5693359375,
0.88623046875,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.5009765625,
0.58544921875,
0.8857421875,
0.58544921875,
0.8857421875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.035125732421875,
0.5986328125,
0.043304443359375,
0.59521484375,
0.047393798828125,
0.60302734375,
0.039215087890625,
0.60595703125
],
[
0.5009765625,
0.599609375,
0.8857421875,
0.599609375,
0.8857421875,
0.61279296875,
0.5009765625,
0.61279296875
],
[
0.0906982421875,
0.61865234375,
0.2156982421875,
0.61865234375,
0.2156982421875,
0.6318359375,
0.0906982421875,
0.6318359375
],
[
0.50244140625,
0.6162109375,
0.8857421875,
0.6162109375,
0.8857421875,
0.626953125,
0.50244140625,
0.626953125
],
[
0.498291015625,
0.62890625,
0.8837890625,
0.62890625,
0.8837890625,
0.642578125,
0.498291015625,
0.642578125
],
[
0.08905029296875,
0.64306640625,
0.473876953125,
0.64306640625,
0.473876953125,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.5009765625,
0.64599609375,
0.88623046875,
0.64599609375,
0.88623046875,
0.6591796875,
0.5009765625,
0.6591796875
],
[
0.08905029296875,
0.65771484375,
0.47216796875,
0.65771484375,
0.47216796875,
0.671875,
0.08905029296875,
0.671875
],
[
0.5,
0.6591796875,
0.8857421875,
0.6591796875,
0.8857421875,
0.67236328125,
0.5,
0.67236328125
],
[
0.08905029296875,
0.6748046875,
0.473876953125,
0.6748046875,
0.473876953125,
0.68798828125,
0.08905029296875,
0.68798828125
],
[
0.5,
0.6748046875,
0.8857421875,
0.6748046875,
0.8857421875,
0.68798828125,
0.5,
0.68798828125
],
[
0.08905029296875,
0.689453125,
0.4755859375,
0.689453125,
0.4755859375,
0.70263671875,
0.08905029296875,
0.70263671875
],
[
0.5,
0.68994140625,
0.88623046875,
0.68994140625,
0.88623046875,
0.7041015625,
0.5,
0.7041015625
],
[
0.0906982421875,
0.70556640625,
0.473876953125,
0.70556640625,
0.473876953125,
0.71630859375,
0.0906982421875,
0.71630859375
],
[
0.5,
0.70556640625,
0.76953125,
0.70556640625,
0.76953125,
0.7197265625,
0.5,
0.7197265625
],
[
0.08905029296875,
0.7197265625,
0.473876953125,
0.7197265625,
0.473876953125,
0.73291015625,
0.08905029296875,
0.73291015625
],
[
0.5009765625,
0.72802734375,
0.8857421875,
0.72802734375,
0.8857421875,
0.7421875,
0.5009765625,
0.7421875
],
[
0.0906982421875,
0.736328125,
0.47216796875,
0.736328125,
0.47216796875,
0.7470703125,
0.0906982421875,
0.7470703125
],
[
0.5009765625,
0.74365234375,
0.8857421875,
0.74365234375,
0.8857421875,
0.7568359375,
0.5009765625,
0.7568359375
],
[
0.08905029296875,
0.75,
0.473876953125,
0.75,
0.473876953125,
0.76318359375,
0.08905029296875,
0.76318359375
],
[
0.5,
0.7587890625,
0.8857421875,
0.7587890625,
0.8857421875,
0.77294921875,
0.5,
0.77294921875
],
[
0.0906982421875,
0.7666015625,
0.47216796875,
0.7666015625,
0.47216796875,
0.7763671875,
0.0906982421875,
0.7763671875
],
[
0.08660888671875,
0.77880859375,
0.4755859375,
0.78076171875,
0.4755859375,
0.79443359375,
0.08660888671875,
0.79248046875
],
[
0.5,
0.7734375,
0.88623046875,
0.7734375,
0.88623046875,
0.787109375,
0.5,
0.787109375
],
[
0.5,
0.78857421875,
0.8857421875,
0.78857421875,
0.8857421875,
0.8017578125,
0.5,
0.8017578125
],
[
0.108642578125,
0.80322265625,
0.473876953125,
0.80322265625,
0.473876953125,
0.8173828125,
0.108642578125,
0.8173828125
],
[
0.498291015625,
0.80322265625,
0.8857421875,
0.8017578125,
0.8857421875,
0.81640625,
0.498291015625,
0.8173828125
],
[
0.0882568359375,
0.8173828125,
0.473876953125,
0.8173828125,
0.473876953125,
0.83056640625,
0.0882568359375,
0.83056640625
],
[
0.5009765625,
0.818359375,
0.8857421875,
0.818359375,
0.8857421875,
0.83203125,
0.5009765625,
0.83203125
],
[
0.0882568359375,
0.82763671875,
0.473876953125,
0.8291015625,
0.473876953125,
0.84326171875,
0.0882568359375,
0.84228515625
],
[
0.5,
0.833984375,
0.8857421875,
0.833984375,
0.8857421875,
0.84765625,
0.5,
0.84765625
],
[
0.08905029296875,
0.84326171875,
0.19775390625,
0.84326171875,
0.19775390625,
0.85400390625,
0.08905029296875,
0.85400390625
],
[
0.498291015625,
0.8486328125,
0.88623046875,
0.84765625,
0.88623046875,
0.8623046875,
0.498291015625,
0.86279296875
],
[
0.08660888671875,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.5009765625,
0.8642578125,
0.8857421875,
0.8642578125,
0.8857421875,
0.87744140625,
0.5009765625,
0.87744140625
],
[
0.08905029296875,
0.8798828125,
0.4755859375,
0.8798828125,
0.4755859375,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.498291015625,
0.8798828125,
0.88623046875,
0.8798828125,
0.88623046875,
0.89306640625,
0.498291015625,
0.89306640625
],
[
0.0882568359375,
0.892578125,
0.221435546875,
0.892578125,
0.221435546875,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.48193359375,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.9345703125,
0.48193359375,
0.9345703125
]
] |
[
"Synergies between Disentanglement and Sparsity:",
"Generalization and Identifiability in Multi-Task Learning",
"Sebastien Lachapelle *1 Tristan Deleu*1 Divyat Mahajan ! Ioannis Mitliagkas 12 Yoshua Bengio 12",
"Simon Lacoste-Julien 12 Quentin Bertrand !",
"Abstract",
"image generation (Ramesh et al., 2022). However, prelimi",
" 9",
"nary assessments of the latter have highlighted shortcomings",
"Although disentangled representations are often",
"related to compositionality (Marcus et al., 2022), suggesting",
"said to be beneficial for downstream tasks, cur-",
"new algorithmic innovations are needed.",
"rent empirical and theoretical understanding is",
"limited. In this work, we provide evidence that",
"Another line of work has argued for the integration of",
"disentangled representations coupled with sparse",
"ideas from causality to make progress towards more robust",
"task-specific predictors improve generalization",
"and transferable machine learning systems (Pearl, 2019;",
"In the context of multi-task learning, we prove a",
"Scholkopf, 2019; Goyal & Bengio, 2022). Causal repre-",
"new identifiability result that provides conditions",
"sentation learning has emerged recently as a field aiming",
"[DT']",
"under which maximally sparse predictors yield",
"to define and learn representations suited for causal reason-",
"disentangled representations. Motivated by this",
"ing (Scholkopf et al., 2021). This set of ideas is strongly",
"theoretical result,we propose a practical approach",
"related to learning disentangled representations (Bengio",
"to learn disentangled representations based on a",
"et al., 2013). Informally, a representation is considered",
"sparsity-promoting bi-level optimization problem.",
"disentangled when its components are in one-to-one corre-",
"Finally, we explore a meta-learning version of",
"spondence with natural and interpretable factors of varia-",
"this algorithm based on group Lasso multiclass",
"999117",
"tions.suchasobjectpositions,colorsorshapes.Although",
"SvM predictors,for which we derive a tractable",
"a plethora of works have investigated theoretically under",
"dual formulation. It obtains competitive results",
"whichconditionsdisentanglementispossiblethroughthe",
"onstandardfew-shotclassificationbenchmarks",
"lens of identifiability (Hyvarinen & Morioka, 2016; 2017",
"while each task is using only a fraction of the",
"Hyvarinenetal..2019:Khemakhemetal.,2020a:Locatello",
"learned representations.",
"et al., 2020a; Klindt et al., 2021; Von Kuigelgen et al., 2021;",
"Gresele et al., 2021; Lachapelle et al., 2022; Lippe et al.,",
"",
"2022b; Ahuja et al., 2022c), fewer works have tackled how",
"1.Introduction",
"a disentangledrepresentationcouldbebeneficial fordown",
"stream tasks. Those who did mainly provide empirical rather",
"The recent literature on self-supervised learning has pro-",
"than theoretical evidence for or against its usefulness (Lo-",
"vided evidence that learning a representation on large cor",
"catello et al., 2019; van Steenkiste et al., 2019; Miladinovic",
"puses of data can yield strong performances on a wide va",
"et al., 2019; Dittadi et al., 2021; Montero et al., 2021). We",
"riety of downstream tasks (Devlin et al., 2018; Chen et al.,",
"believe our work can bring some theoretical insights as to",
"2020),especiallyinfew-shotlearning scenarios wherethe",
"when and why disentanglement can help.",
"training data for these tasks is limited (Brown et al., 2020b;",
"In this work, we explore synergies between disentangle-",
"Dosovitskiyet al.2021;Radford et al.,2021).Beyond trans",
"ment and sparse task-specific predictors in the context of",
"ferring across multiple tasks, these learned representations",
"multi-task learning. At the heart of our contributions is the",
"alsoleadtoimprovedrobustnessagainstdistribution shifts",
"(Wortsman et al., 2022) as well as stunning text-conditioned",
"assumption that only a small subset of all factors of varia-",
"tions are useful for each downstream task, and this subset",
"Equal contribution 'Mila & DIRO, Université de Mon-",
"might change from one task to another. We will refer to such",
"treal °Canada CIFAR AI Chair. Correspondence to: Sébastien",
"tasks as sparse tasks, and their corresponding sets of useful",
"Lachapelle <lachaseb@mila.quebec>, Tristan Deleu <deleu-",
"factors as their supports. This assumption was initially sug-",
"tri@mila.quebec>",
"gested by Bengio et al. (2013, Section 3.5):\"the feature set",
"Proceedings of the 4oth International Conference on Machine",
"being trained may be destined to be used in multiple tasks",
"Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright",
"that may have distinct [and unknown] subsets of relevant",
"2023 by the author(s).",
"1"
] |
{
"arxiv_id": "2211.14666",
"language": "en",
"page_idx": 0
}
|
||
2211.14666-en-1
|
2211.14666-en-1.jpg
|
Synergies Between Disentanglement and Sparsity
features. Considerations such as these lead us to the conclusion that the most robust approach to feature learning is to disentangle as many factors as possible, discarding as little information about the data as is practical". This strategy is in line with the current self-supervised learning trend (Radford et al., 2021), except for its focus on disentanglement.
## 1.1. Contributions
1. We formalize this“sparse task assumption”and argue
theoretically and empirically how, when it holds, a disentangled representation coupled with a sparsityregularized task-specific predictor can generalize better than their entangled counterparts (Section 2).
2.We introduce a novel identifiability result (Theo-
rem 3.1) which shows how one can leverage multiple sparse supervised tasks to learn a shared disentangled representation by regularizing the task-specific predictors to be maximally sparse (Section 3.2). We note that the usage of supervision is in line with many recent results which leverages more or less weak forms of supervision to guarantee identifiability. Contrary to many existing identifiability results, ours allows for statistically dependent latent factors and a non-invertible map between observations and latents.
3. Motivated by this result, we propose a tractable bi-level
optimization (Problem (6)) to learn the shared representation while regularizing the task-specific predictors to be sparse (Section 3.4). We validate our theory by showing our approach can indeed disentangle latent factors on tasks constructed from the 3D Shapes dataset (Burgess & Kim, 2018).
4.Finally, we draw a connection between this bi- levelces. For a matrix $\boldsymbol{A}\in \mathbb{R} ^{k\times m}, \left \| \boldsymbol{A}\right \| _{2, 1}= \sum _{j= 1}^{m}\left \| \boldsymbol{A}_{j}\right \| $,and optimization problem and formulations from the meta- $\left \| \boldsymbol{A}\right \| _{2, 0}= \sum _{j= 1}^{m}1_{\left \| \boldsymbol{A}_{j, j}\right \| \neq 0}$,where l is
sult, we enhance an existing method (Lee et al., 2019), where the task-specific predictors are now groupsparse SVMs. We show that this new meta-learning algorithm achieves competitive performance on the miniImageNet benchmark (Vinyals et al., 2016), while only using a fraction of the representation.
We emphasize that, although related, the theoretical contributions of Sections 2 & 3 are distinct and stand of their
own. Indeed, Section 2 shows how disentangled representations combined with sparsity regularization can improve generalization, while Section 3 shows how regularizing taskspecific predictors to be sparse can induce disentanglement in a multi-task learning setting.
## $1. 2. \textbf{Background}$
We start by introducing formally the notion of entangled
and disentangled representations.
First, we assume the existence of some ground-truth en-
coder function $f_\theta:\mathbb{R}^d\to\mathbb{R}^m$ that maps observations $x$ $\in$ $\mathcal{X}$ $\subseteq$ $\mathbb{R} ^{d}$, $e. g. $, images, to its corresponding interpretable and usually lower dimensional representation $f_{\boldsymbol{\theta}}(x)\in\mathbb{R}^m,m\leq d.$ The exact form of this groundtruth encoder depends on the task at hand, but also on what the machine learning practitioner considers as interpretable The learned encoder function is denoted by $f_{\hat{\boldsymbol{\theta}}}:\mathbb{R}^{\hat{d}}\to\mathbb{R}^{m}$, and should not be conflated with the ground-truth representation $f_{\theta}.$ For example, $f_{\hat{\theta}}$ can be parametrized by a neural network. Throughout, we are going to use the following definition of disentanglement.
Definition 1.1 (Disentangled Representation, Khemakhem et al. 2020a; Lachapelle et al. 2022). A learned encoder function $f_{\hat{\boldsymbol{\theta}}}:\mathbb{R}^{d}\to\mathbb{R}^{m}$ is said to be disentangled w.r.t. the ground-truth representation $f_{\theta}$ when there exists an invertible diagonal matrix $D$ and a permutation matrix $P$ such that, for all $x\in\mathcal{X},f_{\hat{\boldsymbol{\theta}}}(x)=DPf_{\boldsymbol{\theta}}(x).$
Intuitively, a representation is disentangled when there is a one-to-one correspondence between its components and those of the ground-truth representation, up to rescaling. When an encoder $f_\hat{\theta}$ is not disentangled, we say it is entangled. Note that there exist less stringent notions of disentanglement which allow for component-wise nonlinear invertible transformations of the factors (Hyvärinen&Morioka, 2017; Hyvärinen et al., 2019).
Notation. Capital bold letters denote matrices and lowercase bold letters denote vectors. The set of integers from 1 to $n$ is denoted by $[n].$ We write $\|\cdot\|$ for the Euclidean norm on vectors and the Frobenius norm on matri- $\theta$, while that of the learned representation is $\hat{\theta}.$ We follow this convention for all the parameters throughout. Table 1 in Appendix A summarizes all the notation.
## $2. \textbf{ Disentanglement and Sparse Task- Specific}$ Predictors Improve Generalization
In this section, we show that for any linearly equivalent representation (entangled or disentangled), the maximum likelihood estimator defined in Problem (1) yields the same model (Proposition 2.2). However, we also show that disentangled representations have better generalization properties when the task-specific predictor is regularized to be sparse. (Proposition 2.4 and Fig. 1). Our analysis is centred around the following assumption.
Assumption 2.1 (Linear equivalence). The learned encoder $f_{\hat{\theta}}$ is linearly equivalent to the ground-truth encoder $f_\theta$, $i.e.$, there exists an invertible matrix $L$ such that, for all $x\in\mathcal{X},f_{\hat{\boldsymbol{\theta}}}(x)=\boldsymbol{L}f_{\boldsymbol{\theta}}(x).$
2
|
[
[
0.33251953125,
0.05682373046875,
0.64208984375,
0.05682373046875,
0.64208984375,
0.0706787109375,
0.33251953125,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.47802734375,
0.08587646484375,
0.47802734375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.498291015625,
0.085205078125,
0.8837890625,
0.085205078125,
0.8837890625,
0.0997314453125,
0.498291015625,
0.0997314453125
],
[
0.08905029296875,
0.10162353515625,
0.473876953125,
0.10162353515625,
0.473876953125,
0.11553955078125,
0.08905029296875,
0.11553955078125
],
[
0.5009765625,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.08905029296875,
0.11614990234375,
0.473876953125,
0.11614990234375,
0.473876953125,
0.1300048828125,
0.08905029296875,
0.1300048828125
],
[
0.5,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.5,
0.1300048828125
],
[
0.08905029296875,
0.1319580078125,
0.473876953125,
0.1319580078125,
0.473876953125,
0.1458740234375,
0.08905029296875,
0.1458740234375
],
[
0.5009765625,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.08905029296875,
0.146484375,
0.4755859375,
0.146484375,
0.4755859375,
0.1610107421875,
0.08905029296875,
0.1610107421875
],
[
0.5,
0.146484375,
0.88623046875,
0.146484375,
0.88623046875,
0.160400390625,
0.5,
0.160400390625
],
[
0.0882568359375,
0.1622314453125,
0.438720703125,
0.1622314453125,
0.438720703125,
0.175537109375,
0.0882568359375,
0.175537109375
],
[
0.5009765625,
0.1622314453125,
0.88623046875,
0.1622314453125,
0.88623046875,
0.175537109375,
0.5009765625,
0.175537109375
],
[
0.5009765625,
0.1767578125,
0.8857421875,
0.1767578125,
0.8857421875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.0906982421875,
0.1932373046875,
0.2181396484375,
0.1932373046875,
0.2181396484375,
0.20703125,
0.0906982421875,
0.20703125
],
[
0.5009765625,
0.1912841796875,
0.8857421875,
0.1912841796875,
0.8857421875,
0.2052001953125,
0.5009765625,
0.2052001953125
],
[
0.5009765625,
0.20703125,
0.8857421875,
0.20703125,
0.8857421875,
0.2215576171875,
0.5009765625,
0.2215576171875
],
[
0.098876953125,
0.2139892578125,
0.473876953125,
0.21533203125,
0.473876953125,
0.2298583984375,
0.098876953125,
0.2291259765625
],
[
0.5,
0.2215576171875,
0.8857421875,
0.2215576171875,
0.8857421875,
0.236083984375,
0.5,
0.236083984375
],
[
0.1217041015625,
0.231689453125,
0.473876953125,
0.231689453125,
0.473876953125,
0.24560546875,
0.1217041015625,
0.24560546875
],
[
0.5009765625,
0.2374267578125,
0.69921875,
0.2374267578125,
0.69921875,
0.250732421875,
0.5009765625,
0.250732421875
],
[
0.12091064453125,
0.246826171875,
0.473876953125,
0.246826171875,
0.473876953125,
0.260009765625,
0.12091064453125,
0.260009765625
],
[
0.1192626953125,
0.261474609375,
0.473876953125,
0.261474609375,
0.473876953125,
0.274658203125,
0.1192626953125,
0.274658203125
],
[
0.5009765625,
0.261962890625,
0.8857421875,
0.261962890625,
0.8857421875,
0.27587890625,
0.5009765625,
0.27587890625
],
[
0.12091064453125,
0.276611328125,
0.416748046875,
0.276611328125,
0.416748046875,
0.290283203125,
0.12091064453125,
0.290283203125
],
[
0.5009765625,
0.27783203125,
0.8857421875,
0.27783203125,
0.8857421875,
0.291748046875,
0.5009765625,
0.291748046875
],
[
0.1004638671875,
0.291748046875,
0.4755859375,
0.291748046875,
0.4755859375,
0.304931640625,
0.1004638671875,
0.304931640625
],
[
0.498291015625,
0.290283203125,
0.88818359375,
0.2890625,
0.88818359375,
0.306884765625,
0.498291015625,
0.30810546875
],
[
0.12091064453125,
0.306884765625,
0.473876953125,
0.306884765625,
0.473876953125,
0.32080078125,
0.12091064453125,
0.32080078125
],
[
0.5009765625,
0.30810546875,
0.8857421875,
0.30810546875,
0.8857421875,
0.3212890625,
0.5009765625,
0.3212890625
],
[
0.1192626953125,
0.3212890625,
0.473876953125,
0.32080078125,
0.473876953125,
0.335205078125,
0.1192626953125,
0.3359375
],
[
0.5,
0.322509765625,
0.8857421875,
0.322509765625,
0.8857421875,
0.3359375,
0.5,
0.3359375
],
[
0.12091064453125,
0.337158203125,
0.473876953125,
0.337158203125,
0.473876953125,
0.35107421875,
0.12091064453125,
0.35107421875
],
[
0.5,
0.33837890625,
0.8037109375,
0.33837890625,
0.8037109375,
0.352783203125,
0.5,
0.352783203125
],
[
0.12091064453125,
0.35107421875,
0.47216796875,
0.35107421875,
0.47216796875,
0.3642578125,
0.12091064453125,
0.3642578125
],
[
0.12091064453125,
0.367431640625,
0.473876953125,
0.367431640625,
0.473876953125,
0.380615234375,
0.12091064453125,
0.380615234375
],
[
0.5009765625,
0.37060546875,
0.8857421875,
0.37060546875,
0.8857421875,
0.385009765625,
0.5009765625,
0.385009765625
],
[
0.1217041015625,
0.3837890625,
0.47216796875,
0.3837890625,
0.47216796875,
0.39453125,
0.1217041015625,
0.39453125
],
[
0.5009765625,
0.386474609375,
0.8857421875,
0.386474609375,
0.8857421875,
0.399658203125,
0.5009765625,
0.399658203125
],
[
0.12091064453125,
0.397705078125,
0.473876953125,
0.397705078125,
0.473876953125,
0.410888671875,
0.12091064453125,
0.410888671875
],
[
0.5009765625,
0.401611328125,
0.8857421875,
0.401611328125,
0.8857421875,
0.415283203125,
0.5009765625,
0.415283203125
],
[
0.12091064453125,
0.412353515625,
0.4755859375,
0.412353515625,
0.4755859375,
0.425537109375,
0.12091064453125,
0.425537109375
],
[
0.5009765625,
0.416748046875,
0.88623046875,
0.416748046875,
0.88623046875,
0.43115234375,
0.5009765625,
0.43115234375
],
[
0.12091064453125,
0.427978515625,
0.473876953125,
0.427978515625,
0.473876953125,
0.44140625,
0.12091064453125,
0.44140625
],
[
0.5009765625,
0.431884765625,
0.88623046875,
0.431884765625,
0.88623046875,
0.44580078125,
0.5009765625,
0.44580078125
],
[
0.12091064453125,
0.442626953125,
0.34228515625,
0.442626953125,
0.34228515625,
0.455810546875,
0.12091064453125,
0.455810546875
],
[
0.5009765625,
0.4462890625,
0.88623046875,
0.4462890625,
0.88623046875,
0.460205078125,
0.5009765625,
0.460205078125
],
[
0.1004638671875,
0.45703125,
0.473876953125,
0.45703125,
0.473876953125,
0.4716796875,
0.1004638671875,
0.4716796875
],
[
0.5009765625,
0.462158203125,
0.88623046875,
0.462158203125,
0.88623046875,
0.47607421875,
0.5009765625,
0.47607421875
],
[
0.12091064453125,
0.472900390625,
0.473876953125,
0.472900390625,
0.473876953125,
0.486083984375,
0.12091064453125,
0.486083984375
],
[
0.5,
0.4765625,
0.69921875,
0.4765625,
0.69921875,
0.490478515625,
0.5,
0.490478515625
],
[
0.12091064453125,
0.488037109375,
0.473876953125,
0.488037109375,
0.473876953125,
0.501953125,
0.12091064453125,
0.501953125
],
[
0.1217041015625,
0.50244140625,
0.473876953125,
0.50244140625,
0.473876953125,
0.5166015625,
0.1217041015625,
0.5166015625
],
[
0.498291015625,
0.5,
0.8857421875,
0.5,
0.8857421875,
0.51318359375,
0.498291015625,
0.51318359375
],
[
0.5009765625,
0.5166015625,
0.744140625,
0.5166015625,
0.744140625,
0.5263671875,
0.5009765625,
0.5263671875
],
[
0.740234375,
0.5166015625,
0.8837890625,
0.5166015625,
0.8837890625,
0.5263671875,
0.740234375,
0.5263671875
],
[
0.5,
0.52978515625,
0.88623046875,
0.52978515625,
0.88623046875,
0.54443359375,
0.5,
0.54443359375
],
[
0.12091064453125,
0.51806640625,
0.473876953125,
0.51806640625,
0.473876953125,
0.5322265625,
0.12091064453125,
0.5322265625
],
[
0.12091064453125,
0.5322265625,
0.47216796875,
0.5322265625,
0.47216796875,
0.54541015625,
0.12091064453125,
0.54541015625
],
[
0.1217041015625,
0.5498046875,
0.328369140625,
0.5498046875,
0.328369140625,
0.56005859375,
0.1217041015625,
0.56005859375
],
[
0.5009765625,
0.54443359375,
0.8857421875,
0.54443359375,
0.8857421875,
0.55810546875,
0.5009765625,
0.55810546875
],
[
0.496826171875,
0.5546875,
0.88818359375,
0.55615234375,
0.88818359375,
0.5771484375,
0.496826171875,
0.57568359375
],
[
0.1004638671875,
0.56298828125,
0.4755859375,
0.56298828125,
0.4755859375,
0.5771484375,
0.1004638671875,
0.5771484375
],
[
0.12091064453125,
0.5791015625,
0.4755859375,
0.5791015625,
0.4755859375,
0.59228515625,
0.12091064453125,
0.59228515625
],
[
0.5009765625,
0.57568359375,
0.88623046875,
0.57568359375,
0.88623046875,
0.59326171875,
0.5009765625,
0.59326171875
],
[
0.1217041015625,
0.59326171875,
0.473876953125,
0.59326171875,
0.473876953125,
0.60693359375,
0.1217041015625,
0.60693359375
],
[
0.5,
0.59130859375,
0.88623046875,
0.59033203125,
0.88623046875,
0.60498046875,
0.5,
0.60595703125
],
[
0.12091064453125,
0.60791015625,
0.473876953125,
0.60791015625,
0.473876953125,
0.6220703125,
0.12091064453125,
0.6220703125
],
[
0.5009765625,
0.60595703125,
0.8837890625,
0.60595703125,
0.8837890625,
0.619140625,
0.5009765625,
0.619140625
],
[
0.1192626953125,
0.62255859375,
0.473876953125,
0.62353515625,
0.473876953125,
0.63720703125,
0.1192626953125,
0.63623046875
],
[
0.5009765625,
0.62255859375,
0.88623046875,
0.62255859375,
0.88623046875,
0.63623046875,
0.5009765625,
0.63623046875
],
[
0.1192626953125,
0.63720703125,
0.4755859375,
0.63818359375,
0.4755859375,
0.65283203125,
0.1192626953125,
0.6513671875
],
[
0.5,
0.63818359375,
0.78759765625,
0.63818359375,
0.78759765625,
0.6513671875,
0.5,
0.6513671875
],
[
0.1192626953125,
0.65380859375,
0.473876953125,
0.65283203125,
0.473876953125,
0.66650390625,
0.1192626953125,
0.66748046875
],
[
0.12091064453125,
0.66845703125,
0.473876953125,
0.66845703125,
0.473876953125,
0.681640625,
0.12091064453125,
0.681640625
],
[
0.50244140625,
0.671875,
0.87158203125,
0.671875,
0.87158203125,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.1192626953125,
0.68310546875,
0.398681640625,
0.68310546875,
0.398681640625,
0.6962890625,
0.1192626953125,
0.6962890625
],
[
0.521484375,
0.689453125,
0.81396484375,
0.689453125,
0.81396484375,
0.7041015625,
0.521484375,
0.7041015625
],
[
0.0882568359375,
0.70556640625,
0.473876953125,
0.70703125,
0.473876953125,
0.7216796875,
0.0882568359375,
0.72021484375
],
[
0.5,
0.71533203125,
0.8857421875,
0.71533203125,
0.8857421875,
0.72900390625,
0.5,
0.72900390625
],
[
0.08905029296875,
0.724609375,
0.47216796875,
0.724609375,
0.47216796875,
0.73486328125,
0.08905029296875,
0.73486328125
],
[
0.5009765625,
0.73193359375,
0.8837890625,
0.73193359375,
0.8837890625,
0.7421875,
0.5009765625,
0.7421875
],
[
0.0882568359375,
0.73779296875,
0.4755859375,
0.73779296875,
0.4755859375,
0.751953125,
0.0882568359375,
0.751953125
],
[
0.5009765625,
0.74560546875,
0.8857421875,
0.74560546875,
0.8857421875,
0.7587890625,
0.5009765625,
0.7587890625
],
[
0.08905029296875,
0.75390625,
0.473876953125,
0.75390625,
0.473876953125,
0.767578125,
0.08905029296875,
0.767578125
],
[
0.5009765625,
0.76025390625,
0.8857421875,
0.76025390625,
0.8857421875,
0.7734375,
0.5009765625,
0.7734375
],
[
0.08905029296875,
0.76806640625,
0.47216796875,
0.76806640625,
0.47216796875,
0.7822265625,
0.08905029296875,
0.7822265625
],
[
0.5,
0.77587890625,
0.8857421875,
0.77587890625,
0.8857421875,
0.7890625,
0.5,
0.7890625
],
[
0.08905029296875,
0.7841796875,
0.473876953125,
0.7841796875,
0.473876953125,
0.79736328125,
0.08905029296875,
0.79736328125
],
[
0.50244140625,
0.79248046875,
0.8837890625,
0.79248046875,
0.8837890625,
0.80322265625,
0.50244140625,
0.80322265625
],
[
0.0882568359375,
0.7998046875,
0.294189453125,
0.7998046875,
0.294189453125,
0.81298828125,
0.0882568359375,
0.81298828125
],
[
0.5,
0.80615234375,
0.88623046875,
0.80615234375,
0.88623046875,
0.8193359375,
0.5,
0.8193359375
],
[
0.5009765625,
0.8212890625,
0.67333984375,
0.8212890625,
0.67333984375,
0.83544921875,
0.5009765625,
0.83544921875
],
[
0.0906982421875,
0.830078125,
0.2059326171875,
0.830078125,
0.2059326171875,
0.84326171875,
0.0906982421875,
0.84326171875
],
[
0.5009765625,
0.84521484375,
0.8857421875,
0.84521484375,
0.8857421875,
0.859375,
0.5009765625,
0.859375
],
[
0.08660888671875,
0.8515625,
0.473876953125,
0.85302734375,
0.473876953125,
0.8671875,
0.08660888671875,
0.8662109375
],
[
0.08660888671875,
0.8671875,
0.3095703125,
0.86865234375,
0.3095703125,
0.88330078125,
0.08660888671875,
0.8818359375
],
[
0.5009765625,
0.8623046875,
0.8857421875,
0.85986328125,
0.8857421875,
0.875,
0.5009765625,
0.876953125
],
[
0.5009765625,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.0882568359375,
0.8916015625,
0.476318359375,
0.8916015625,
0.476318359375,
0.90478515625,
0.0882568359375,
0.90478515625
],
[
0.5,
0.892578125,
0.67626953125,
0.892578125,
0.67626953125,
0.90576171875,
0.5,
0.90576171875
],
[
0.476318359375,
0.92529296875,
0.491943359375,
0.921875,
0.496826171875,
0.931640625,
0.48193359375,
0.93603515625
]
] |
[
"Synergies Between Disentanglement and Sparsity",
"features. Considerations such as these lead us to the conclu-",
"coder function fe : Rd → Rm that maps observations",
"sion that the most robust approach to feature learning is to",
" e X C Rd, e.g., images, to its corresponding in",
"disentangle as many factors as possible, discarding as little",
"terpretable and usually lower dimensional representation",
"information about the data as is practical\". This strategy is in",
"fe(aα) E Rm, m ≤ d. The exact form of this ground",
"line with the current self-supervised learning trend (Radford",
"truth encoder depends on the task at hand, but also on what",
"et al., 2021), except for its focus on disentanglement.",
"the machine learning practitioner considers as interpretable.",
"The learned encoder function is denoted by fa : Rd→ Rm,",
"1.1.Contributions",
"and should not be conflated with the ground-truth represen",
"tation fe. For example, f can be parametrized by a neural",
"1. We formalize this “sparse task assumption\" and argue",
"network. Throughout, we are going to use the following",
"theoretically and empirically how, when it holds, a",
"definition of disentanglement.",
"disentangled representation coupled with a sparsity",
"regularized task-specific predictor can generalize better",
"Definition 1.1 (Disentangled Representation, Khemakhem",
"than their entangled counterparts (Section 2).",
"et al. 2020a; Lachapelle et al. 2022). A learned encoder",
"2. We introduce a novel identifiability result (Theo-",
"function fo : IRd = Rm is said to be disentangled wrt.",
"rem 3.1) which shows how one can leverage multiple",
"the ground-truth representation fe when there exists an",
"sparse supervised tasks to learn a shared disentangled",
"invertible diagonal matrix D and a permutation matrix P",
"representation by regularizing the task-specific predic-",
"such that, for all a E X, fa(α) = DPfe(α).",
"tors to be maximally sparse (Section 3.2). We note that",
"the usage of supervision is in line with many recent",
"Intuitively, a representation is disentangled when there is",
"resultswhichleveragesmoreorlessweakformsof su",
"a one-to-one correspondence between its components and",
"pervision to guarantee identifiability. Contrary to many",
"those of the ground-truth representation, up to rescaling",
"existing identifiability results, ours allows for statisti-",
"When an encoder f is not disentangled, we say it is entan-",
"cally dependent latent factors and a non-invertible map",
"gled. Note that there exist less stringent notions of disentan-",
"between observations and latents.",
"glement which allow for component-wise nonlinear invert-",
"3. Motivated by this result, we propose a tractable bi-level",
"ible transformations of the factors (Hyvarinen & Morioka,",
"optimization (Problem (6)) to learn the shared represen-",
"2017; Hyvarinen et al., 2019).",
"tation while regularizing the task-specific predictors",
"to be sparse (Section 3.4). We validate our theory",
"Notation. Capital bold letters denote matrices and low-",
"ercase bold lettersdenotevectors.",
"Thesetof integers",
"from 1 to n is denoted by [n]. We write Il-l for the Eu-",
"by showing our approach can indeed disentangle la-",
"tent factors on tasks constructed from the 3D Shapes",
"dataset (Burgess &Kim.2018)",
"clidean norm on vectors and the Frobenius norm on matri",
"ces. For a matrix A Rkxm, /Al2,1 - ,-1 /A:;ll, and",
"4. Finally, we draw a connection between this bi-level",
"optimization problem and formulations from the meta-",
"IAll2,0== A;ll0,where 1 is the indicator func-",
"learning literature. Inspired by our identifiability re-",
"tion. The ground-truth parameter of the encoder function is",
"sult, we enhance an existing method (Lee et al., 2019),",
", while that of the learned representation is . We follow",
"where the task-specific predictors are now group",
"this convention for all the parameters throughout. Table 1",
"sparse SvMs. We show that this new meta-learning",
"in Appendix A summarizes all the notation.",
"algorithm achieves competitive performance on the",
"minilmageNet benchmark (Vinyals et al.,2016), while",
"2. Disentanglement and Sparse Task-Specific",
"only using a fraction of the representation.",
"Predictors Improve Generalization",
"We emphasize that, although related, the theoretical con",
"In this section, we show that for any linearly equivalent",
"tributions ofSections2&3aredistinctandstandof their",
"representation(entangled ordisentangled),themaximum",
"own. Indeed, Section 2 shows how disentangled represen-",
"likelihood estimator defined in Problem (1) yields the same",
"tations combined with sparsity regularization can improve",
"model (Proposition 2.2). However, we also show that disen",
"generalization, while Section 3 shows how regularizing task",
"tangled representations have better generalization properties",
"specific predictors to be sparse can induce disentanglement",
"whenthetask-specificpredictorisregularizedtobesparse",
"in a multi-task learning setting.",
"(Proposition 2.4 and Fig. 1). Our analysis is centred around",
"the following assumption.",
"1.2.Background",
"Assumption 2.1 (Linear equivalence). The learned encoder",
"We start by introducing formally the notion of entangled",
"and disentangled representations.",
"f is linearly equivalent to the ground-truth encoder fe,",
"i.e., there exists an invertible matrix L such that, for all",
"First, we assume the existence of some ground-truth en-",
" EX, fa(α) = Lfe(α).",
"2"
] |
{
"arxiv_id": "2211.14666",
"language": "en",
"page_idx": 1
}
|
||
2211.14666-en-3
|
2211.14666-en-3.jpg
|
Synergies Between Disentanglement and Sparsity
error (bias). In contrast, an entangled representation might not benefit from sparsity regularization if the increase in approximation error is more important than the reduction in estimation error.
Empirical validation (Fig. 1). We now present a simple simulated experiment that illustrates the above claim that disentangled representations coupled with sparsity regularization can yield better generalization. Fig.1 compares the generalization performances of $L_1$ and $L_2$-penalized linear regressions (Tibshirani, 1996; Hoerl & Kennard, 1970), computed on the top of both disentangled and linearly entangled representations, which are frozen during training. $L_1$-penalized linear regression coupled with the disentangled representation yields better generalization than other alternatives when $\ell/m=5\%$ and when the number of samples is very small. One can also see that disentanglement, sparsity regularization, and sufficient sparsity in the ground-truth data generating process are necessary for significant improvements, in line with our discussion. Lastly, all methods yield similar performance when the number of samples grows. More details and discussions can be found in Appendix D.1.
## 3. Sparse Multi-Task Learning for Disentanglement
In Section 2, we argued that disentangled representations can improve generalization when combined with sparse taskspecific predictors, but we did not mention how to obtain a disentangled representation in the first place. In this section, we first provide a new identification result (Theorem 3.1, Section $\hat{3}.2)$, which states that in the multi-task learning setting, regularizing the task-specific predictors to be sparse can yield disentangled representations. Then, in Section 3.4, we provide a practical way to learn disentangled representations motivated by our identifiability result.
## 3.1. Task & data generating process
Throughout this section, we assume the learner is given a set of $T$ datasets $\{\mathcal{D}_1,\ldots,\mathcal{D}_T\}$ where each dataset $\mathcal{D}_t:=$ $\{(x^{(t,i)},y^{(t,i)})\}_{i=1}^n$ consists of $n$ couples of input $x\in\mathbb{R}^d$ and label $y\in\mathcal{Y}.$ The set of labels $\mathcal{Y}$ might contain either class indices or real values, depending on whether we are concerned with classification or regression tasks.
Our theory relies on the assumption that, for each task $t$,the dataset $\mathcal{D}_t$ is made of i.i.d. samples from the distribution
$$p(\boldsymbol{x},y\mid\boldsymbol{W}^{(t)}):=p(y;\boldsymbol{W}^{(t)}\boldsymbol{f_{\theta}}(\boldsymbol{x}))p(\boldsymbol{x}\mid\boldsymbol{W}^{(t)})\:,$$
(3)
where $W^{(t)}\in\mathbb{R}^{k\times m}$ is the task-specific ground-truth coefficient matrix. We emphasize that the representation $f_{\theta}$ is shared across all the tasks while the coefficient matrices $W^{(t)}$ are task-specific. Also note that the distribution over
<FigureHere>
Figure 1. Test performance for the entangled and disentangled rep-
resentation using Lasso and Ridge regression. All the results are averaged over 10 seeds, with standard error shown in error bars.
$x$ is allowed to change from one task to another. However,
we assume that its support, $\mathcal{X}$, is fixed across tasks.
We further assume that the task-specific matrices $W^{(t)}$ are i.i.d. samples from some probability measure $\mathbb{P}_W$ with support $\mathcal{W}.$ We will see in Section 3.3 that the most critical assumptions of our theory concern $\mathbb{P}_W.$
## 3.2. Main identifiability result
We are now ready to show the main theoretical result of this work, which provides a bi-level optimization problem for which the optimal representations are guaranteed to be disentangled. It assumes infinitely many tasks are observed, with task-specific ground-truth matrices $W$ sampled from $\mathbb{P}_{W}.$ We denote by $\hat{\boldsymbol{W}}^{(\boldsymbol{W})}$ the task-specific estimator of $W.$ We delay the presentation of its technical assumptions to Section 3.3. See Appendix B.2 for a proof.
Theorem 3.1 (Sparse multi-task learning for disentangle-
ment). Let $\hat{\boldsymbol{\theta }} \textit{be a minimizer of}$
$$\begin{aligned}&\operatorname*{min}_{\hat{\boldsymbol{\theta}}}\mathbb{E}_{\mathbb{P}_{\boldsymbol{W}}}\mathbb{E}_{p(\boldsymbol{x},y|\boldsymbol{W})}-\log p(y;\hat{\boldsymbol{W}}^{(\boldsymbol{W})}\boldsymbol{f}_{\hat{\boldsymbol{\theta}}}(\boldsymbol{x}))&\text{(4)}\\&\mathrm{s.t.}\quad\hat{\boldsymbol{W}}^{(\boldsymbol{W})}\in\arg\min\mathbb{E}_{p(\boldsymbol{x},y|\boldsymbol{W})}-\log p(y;\tilde{\boldsymbol{W}}\boldsymbol{f}_{\hat{\boldsymbol{\theta}}}(\boldsymbol{x}))\:,\end{aligned}$$
where the constraint holds for all W EW and where $\mathbb{P}_W$ and p(x,y|W) are described in Section 3.1. Under Assumptions 3.2 to 3.6 and if $f_{\tilde{\boldsymbol{\theta}}}$ is continuous for all $\tilde{\boldsymbol{\theta}},f_{\hat{\boldsymbol{\theta}}}$ is disentangled w.r.t. f f 0 (Definition 1.1).
Intuitively, this optimization problem effectively selects a representation $f_{\hat{\theta}}$ that (i) allows a perfect fit of the data dis- $\hat{\text{tribution, and (ii) allows the task-specific estimators }}\hat{\boldsymbol{W}}^{(\boldsymbol{W})}$ to be as sparse as the ground-truth $W.$ The theorem guarantees that such a representation must be disentangled
Under the same assumptions and with the same disentanglement guarantees, Theorem B.6 in Appendix B presents a variation of Problem (4) which enforces the weaker constraint $\mathbb{E}_{\mathbb{P}_W}\|\hat{\boldsymbol{W}}^{(\boldsymbol{W})}\|_{2,0}\leq\mathbb{E}_{\mathbb{P}_W}\|W\|_{2,0}$, instead of $\|\hat{W}^{(W)}\|_{2,0}\leq\|\boldsymbol{W}\|_{2,0}$ for each task $W$ individually.
Characteristic features of our theory.(i) Contrary to most identifiability results for disentanglement (Section 4), we do not assume the observations $x$ are generated by trans-
4
|
[
[
0.330810546875,
0.05682373046875,
0.640625,
0.05682373046875,
0.640625,
0.0706787109375,
0.330810546875,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.473876953125,
0.08587646484375,
0.473876953125,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.6044921875,
0.0889892578125,
0.67236328125,
0.0889892578125,
0.67236328125,
0.09661865234375,
0.6044921875,
0.09661865234375
],
[
0.72607421875,
0.0889892578125,
0.7890625,
0.0889892578125,
0.7890625,
0.09661865234375,
0.72607421875,
0.09661865234375
],
[
0.609375,
0.09783935546875,
0.68115234375,
0.09783935546875,
0.68115234375,
0.10797119140625,
0.609375,
0.10797119140625
],
[
0.72802734375,
0.09783935546875,
0.8056640625,
0.09783935546875,
0.8056640625,
0.10797119140625,
0.72802734375,
0.10797119140625
],
[
0.56884765625,
0.111083984375,
0.57861328125,
0.111083984375,
0.57861328125,
0.1142578125,
0.56884765625,
0.1142578125
],
[
0.0882568359375,
0.10162353515625,
0.47216796875,
0.10162353515625,
0.47216796875,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.68896484375,
0.111083984375,
0.6953125,
0.111083984375,
0.6953125,
0.1142578125,
0.68896484375,
0.1142578125
],
[
0.83642578125,
0.111083984375,
0.84619140625,
0.111083984375,
0.84619140625,
0.1142578125,
0.83642578125,
0.1142578125
],
[
0.08660888671875,
0.1331787109375,
0.19775390625,
0.1319580078125,
0.19775390625,
0.143310546875,
0.08660888671875,
0.14453125
],
[
0.0882568359375,
0.11614990234375,
0.473876953125,
0.11614990234375,
0.473876953125,
0.1300048828125,
0.0882568359375,
0.1300048828125
],
[
0.08905029296875,
0.154052734375,
0.47216796875,
0.154052734375,
0.47216796875,
0.1673583984375,
0.08905029296875,
0.1673583984375
],
[
0.58837890625,
0.1572265625,
0.60107421875,
0.1572265625,
0.60107421875,
0.1610107421875,
0.58837890625,
0.1610107421875
],
[
0.62353515625,
0.1572265625,
0.63623046875,
0.1572265625,
0.63623046875,
0.1610107421875,
0.62353515625,
0.1610107421875
],
[
0.7099609375,
0.1572265625,
0.7197265625,
0.1572265625,
0.7197265625,
0.160400390625,
0.7099609375,
0.160400390625
],
[
0.74853515625,
0.1572265625,
0.75634765625,
0.1572265625,
0.75634765625,
0.1610107421875,
0.74853515625,
0.1610107421875
],
[
0.56787109375,
0.1641845703125,
0.6044921875,
0.1641845703125,
0.6044921875,
0.1717529296875,
0.56787109375,
0.1717529296875
],
[
0.8095703125,
0.1641845703125,
0.8447265625,
0.1641845703125,
0.8447265625,
0.1717529296875,
0.8095703125,
0.1717529296875
],
[
0.5009765625,
0.177978515625,
0.8857421875,
0.177978515625,
0.8857421875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.08905029296875,
0.169189453125,
0.473876953125,
0.169189453125,
0.473876953125,
0.18310546875,
0.08905029296875,
0.18310546875
],
[
0.0882568359375,
0.184326171875,
0.4755859375,
0.184326171875,
0.4755859375,
0.1976318359375,
0.0882568359375,
0.1976318359375
],
[
0.498291015625,
0.1912841796875,
0.8837890625,
0.1900634765625,
0.8837890625,
0.203857421875,
0.498291015625,
0.2052001953125
],
[
0.08905029296875,
0.2008056640625,
0.47216796875,
0.2008056640625,
0.47216796875,
0.2108154296875,
0.08905029296875,
0.2108154296875
],
[
0.498291015625,
0.2052001953125,
0.87841796875,
0.203857421875,
0.87841796875,
0.2183837890625,
0.498291015625,
0.2197265625
],
[
0.0882568359375,
0.2139892578125,
0.473876953125,
0.21337890625,
0.473876953125,
0.2279052734375,
0.0882568359375,
0.2291259765625
],
[
0.0882568359375,
0.2298583984375,
0.473876953125,
0.2298583984375,
0.473876953125,
0.24365234375,
0.0882568359375,
0.24365234375
],
[
0.0882568359375,
0.2442626953125,
0.473876953125,
0.2442626953125,
0.473876953125,
0.25830078125,
0.0882568359375,
0.25830078125
],
[
0.5009765625,
0.240478515625,
0.88623046875,
0.240478515625,
0.88623046875,
0.25390625,
0.5009765625,
0.25390625
],
[
0.08905029296875,
0.260009765625,
0.473876953125,
0.260009765625,
0.473876953125,
0.2734375,
0.08905029296875,
0.2734375
],
[
0.498291015625,
0.255615234375,
0.8349609375,
0.255615234375,
0.8349609375,
0.26953125,
0.498291015625,
0.26953125
],
[
0.08905029296875,
0.274658203125,
0.4755859375,
0.274658203125,
0.4755859375,
0.28857421875,
0.08905029296875,
0.28857421875
],
[
0.498291015625,
0.27587890625,
0.8857421875,
0.2734375,
0.8857421875,
0.291748046875,
0.498291015625,
0.29345703125
],
[
0.08905029296875,
0.2890625,
0.47216796875,
0.2890625,
0.47216796875,
0.302978515625,
0.08905029296875,
0.302978515625
],
[
0.5,
0.292236328125,
0.8857421875,
0.292236328125,
0.8857421875,
0.306884765625,
0.5,
0.306884765625
],
[
0.08660888671875,
0.304931640625,
0.4755859375,
0.3037109375,
0.4755859375,
0.318115234375,
0.08660888671875,
0.3193359375
],
[
0.496826171875,
0.306884765625,
0.88623046875,
0.30615234375,
0.88623046875,
0.32080078125,
0.496826171875,
0.3212890625
],
[
0.0882568359375,
0.32080078125,
0.473876953125,
0.32080078125,
0.473876953125,
0.333984375,
0.0882568359375,
0.333984375
],
[
0.5,
0.323974609375,
0.76318359375,
0.323974609375,
0.76318359375,
0.337158203125,
0.5,
0.337158203125
],
[
0.0906982421875,
0.337158203125,
0.47216796875,
0.337158203125,
0.47216796875,
0.347900390625,
0.0906982421875,
0.347900390625
],
[
0.08905029296875,
0.35107421875,
0.473876953125,
0.35107421875,
0.473876953125,
0.3642578125,
0.08905029296875,
0.3642578125
],
[
0.5009765625,
0.354736328125,
0.708984375,
0.354736328125,
0.708984375,
0.36865234375,
0.5009765625,
0.36865234375
],
[
0.08905029296875,
0.3642578125,
0.47216796875,
0.3642578125,
0.47216796875,
0.37744140625,
0.08905029296875,
0.37744140625
],
[
0.0882568359375,
0.380615234375,
0.4755859375,
0.380615234375,
0.4755859375,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.5009765625,
0.37744140625,
0.88623046875,
0.37744140625,
0.88623046875,
0.391357421875,
0.5009765625,
0.391357421875
],
[
0.0882568359375,
0.395751953125,
0.4755859375,
0.395751953125,
0.4755859375,
0.4091796875,
0.0882568359375,
0.4091796875
],
[
0.5009765625,
0.393310546875,
0.8857421875,
0.393310546875,
0.8857421875,
0.4072265625,
0.5009765625,
0.4072265625
],
[
0.0882568359375,
0.410888671875,
0.2034912109375,
0.410888671875,
0.2034912109375,
0.4248046875,
0.0882568359375,
0.4248046875
],
[
0.5009765625,
0.4091796875,
0.8857421875,
0.4091796875,
0.8857421875,
0.42236328125,
0.5009765625,
0.42236328125
],
[
0.5009765625,
0.423583984375,
0.88623046875,
0.423583984375,
0.88623046875,
0.4375,
0.5009765625,
0.4375
],
[
0.08660888671875,
0.442626953125,
0.375732421875,
0.443115234375,
0.375732421875,
0.4609375,
0.08660888671875,
0.460205078125
],
[
0.5009765625,
0.439453125,
0.8857421875,
0.439453125,
0.8857421875,
0.45263671875,
0.5009765625,
0.45263671875
],
[
0.5009765625,
0.45068359375,
0.88818359375,
0.45068359375,
0.88818359375,
0.468505859375,
0.5009765625,
0.468505859375
],
[
0.108642578125,
0.46337890625,
0.250732421875,
0.46337890625,
0.250732421875,
0.477783203125,
0.108642578125,
0.477783203125
],
[
0.5009765625,
0.46728515625,
0.88623046875,
0.468505859375,
0.88623046875,
0.48291015625,
0.5009765625,
0.481689453125
],
[
0.08905029296875,
0.488037109375,
0.473876953125,
0.488037109375,
0.473876953125,
0.501953125,
0.08905029296875,
0.501953125
],
[
0.498291015625,
0.484130859375,
0.79833984375,
0.484130859375,
0.79833984375,
0.49755859375,
0.498291015625,
0.49755859375
],
[
0.0882568359375,
0.50390625,
0.4755859375,
0.50390625,
0.4755859375,
0.517578125,
0.0882568359375,
0.517578125
],
[
0.5,
0.501953125,
0.8857421875,
0.50244140625,
0.8857421875,
0.517578125,
0.5,
0.5166015625
],
[
0.0906982421875,
0.52099609375,
0.47216796875,
0.52099609375,
0.47216796875,
0.53076171875,
0.0906982421875,
0.53076171875
],
[
0.5,
0.517578125,
0.708984375,
0.51806640625,
0.708984375,
0.53271484375,
0.5,
0.5322265625
],
[
0.0882568359375,
0.53515625,
0.473876953125,
0.53515625,
0.473876953125,
0.548828125,
0.0882568359375,
0.548828125
],
[
0.495849609375,
0.54150390625,
0.8056640625,
0.5390625,
0.8056640625,
0.56005859375,
0.495849609375,
0.5625
],
[
0.08660888671875,
0.5498046875,
0.473876953125,
0.548828125,
0.473876953125,
0.5625,
0.08660888671875,
0.56298828125
],
[
0.86376953125,
0.54443359375,
0.88623046875,
0.54443359375,
0.88623046875,
0.55810546875,
0.86376953125,
0.55810546875
],
[
0.08905029296875,
0.56298828125,
0.47216796875,
0.56298828125,
0.47216796875,
0.5771484375,
0.08905029296875,
0.5771484375
],
[
0.5,
0.564453125,
0.8935546875,
0.5693359375,
0.8935546875,
0.5947265625,
0.5,
0.5888671875
],
[
0.0882568359375,
0.580078125,
0.473876953125,
0.580078125,
0.473876953125,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.62158203125,
0.58837890625,
0.64208984375,
0.58837890625,
0.64208984375,
0.59521484375,
0.62158203125,
0.59521484375
],
[
0.0882568359375,
0.5947265625,
0.473876953125,
0.5947265625,
0.473876953125,
0.60791015625,
0.0882568359375,
0.60791015625
],
[
0.58251953125,
0.59521484375,
0.69140625,
0.59521484375,
0.69140625,
0.60986328125,
0.58251953125,
0.60986328125
],
[
0.0882568359375,
0.60986328125,
0.473876953125,
0.60986328125,
0.473876953125,
0.62353515625,
0.0882568359375,
0.62353515625
],
[
0.08660888671875,
0.62353515625,
0.371826171875,
0.62353515625,
0.371826171875,
0.63720703125,
0.08660888671875,
0.63720703125
],
[
0.5009765625,
0.61865234375,
0.88232421875,
0.61865234375,
0.88232421875,
0.63330078125,
0.5009765625,
0.63330078125
],
[
0.5009765625,
0.63525390625,
0.88623046875,
0.63525390625,
0.88623046875,
0.6484375,
0.5009765625,
0.6484375
],
[
0.5009765625,
0.6494140625,
0.88623046875,
0.6494140625,
0.88623046875,
0.6640625,
0.5009765625,
0.6640625
],
[
0.0882568359375,
0.65576171875,
0.339111328125,
0.65576171875,
0.339111328125,
0.66943359375,
0.0882568359375,
0.66943359375
],
[
0.5,
0.6640625,
0.75244140625,
0.6640625,
0.75244140625,
0.6787109375,
0.5,
0.6787109375
],
[
0.08660888671875,
0.6787109375,
0.4755859375,
0.6796875,
0.4755859375,
0.6943359375,
0.08660888671875,
0.693359375
],
[
0.08905029296875,
0.69580078125,
0.464111328125,
0.69580078125,
0.464111328125,
0.708984375,
0.08905029296875,
0.708984375
],
[
0.5,
0.69140625,
0.8857421875,
0.69140625,
0.8857421875,
0.705078125,
0.5,
0.705078125
],
[
0.08905029296875,
0.708984375,
0.473876953125,
0.708984375,
0.473876953125,
0.7236328125,
0.08905029296875,
0.7236328125
],
[
0.498291015625,
0.70703125,
0.88623046875,
0.70703125,
0.88623046875,
0.7216796875,
0.498291015625,
0.7216796875
],
[
0.0882568359375,
0.724609375,
0.473876953125,
0.724609375,
0.473876953125,
0.73779296875,
0.0882568359375,
0.73779296875
],
[
0.496826171875,
0.7216796875,
0.88623046875,
0.7197265625,
0.88623046875,
0.7373046875,
0.496826171875,
0.7392578125
],
[
0.0882568359375,
0.74072265625,
0.473876953125,
0.74072265625,
0.473876953125,
0.75390625,
0.0882568359375,
0.75390625
],
[
0.5,
0.7392578125,
0.88623046875,
0.7392578125,
0.88623046875,
0.75244140625,
0.5,
0.75244140625
],
[
0.0882568359375,
0.7548828125,
0.41015625,
0.7548828125,
0.41015625,
0.76806640625,
0.0882568359375,
0.76806640625
],
[
0.5,
0.7548828125,
0.84326171875,
0.7548828125,
0.84326171875,
0.76806640625,
0.5,
0.76806640625
],
[
0.08905029296875,
0.7763671875,
0.47216796875,
0.7763671875,
0.47216796875,
0.79052734375,
0.08905029296875,
0.79052734375
],
[
0.5,
0.77587890625,
0.8837890625,
0.77587890625,
0.8837890625,
0.7890625,
0.5,
0.7890625
],
[
0.0882568359375,
0.79345703125,
0.4599609375,
0.79345703125,
0.4599609375,
0.806640625,
0.0882568359375,
0.806640625
],
[
0.5009765625,
0.79248046875,
0.88623046875,
0.79248046875,
0.88623046875,
0.80615234375,
0.5009765625,
0.80615234375
],
[
0.5009765625,
0.806640625,
0.88623046875,
0.806640625,
0.88623046875,
0.82080078125,
0.5009765625,
0.82080078125
],
[
0.10540771484375,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.833984375,
0.10540771484375,
0.833984375
],
[
0.498291015625,
0.818359375,
0.8896484375,
0.818359375,
0.8896484375,
0.8388671875,
0.498291015625,
0.8388671875
],
[
0.5,
0.83544921875,
0.85693359375,
0.8369140625,
0.85693359375,
0.85498046875,
0.5,
0.85302734375
],
[
0.0849609375,
0.84228515625,
0.4755859375,
0.84326171875,
0.4755859375,
0.861328125,
0.0849609375,
0.85986328125
],
[
0.08660888671875,
0.85986328125,
0.473876953125,
0.861328125,
0.473876953125,
0.87548828125,
0.08660888671875,
0.875
],
[
0.5009765625,
0.861328125,
0.8857421875,
0.861328125,
0.8857421875,
0.875,
0.5009765625,
0.875
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.5,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5,
0.89013671875
],
[
0.0882568359375,
0.8896484375,
0.473876953125,
0.8916015625,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90380859375
],
[
0.5,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.5,
0.90576171875
],
[
0.483642578125,
0.9267578125,
0.49267578125,
0.9267578125,
0.49267578125,
0.9345703125,
0.483642578125,
0.9345703125
]
] |
[
"Synergies Between Disentanglement and Sparsity",
"error (bias). In contrast, an entangled representation might",
"Entangled-Lasso",
"Entangled-Ridg",
"Disentangled-Lassd",
"Disentangled-Ridge",
"t",
"not benefit from sparsity regularization if the increase in",
"",
"00",
"estimationerror.",
"approximation error is more important than the reduction in",
"Empirical validation (Fig. 1). We now present a simple",
"100",
"15",
"",
"5",
"samples",
"#sample",
"Figure 1. Test performance for the entangled and disentangled rep-",
"simulated experiment that illustrates the above claim that",
"disentangled representations coupled with sparsity regular-",
"resentation using Lasso and Ridge regression. All the results are",
"izationcan yield bettergeneralization.Fig.l comparesthe",
"averaged over 10 seeds, with standard error shown in error bars.",
"generalization performances of L and L2-penalized lin-",
"ear regressions (Tibshirani, 1996; Hoerl & Kennard, 1970),",
"computed on the top of both disentangled and linearly en-",
" is allowed to change from one task to another. However,",
"tangled representations, which are frozen during training",
"we assume that its support, X, is fixed across tasks.",
"Li-penalized linear regression coupled with the disentan-",
"We further assume that the task-specific matrices W(t) are",
"gled representation yields better generalization than other",
"i.i.d. samples from some probability measure Pw with",
"alternatives when e/m = 5% and when the number of",
"support W. We will see in Section 3.3 that the most critical",
"samples is very small. One can also see that disentangle-",
"assumptions of our theory concern Pw.",
"ment,sparsityregularization,and sufficient sparsityin the",
"ground-truth data generating process are necessary for sig-",
"3.2. Main identifiability result",
"nificant improvements, in line with our discussion. Lastly",
"all methods yield similar performance when the number of",
"We are now ready to show the main theoretical result of",
"samples grows. More details and discussions can be found",
"this work, which provides a bi-level optimization problem",
"in Appendix D.1.",
"for which the optimal representations are guaranteed to be",
"disentangled. It assumes infinitely many tasks are observed,",
"3. Sparse Multi-Task Learning for",
"with task-specific ground-truth matrices W sampled from",
"Pw. We denote by W(w) the task-specific estimator of",
"Disentanglement",
"W. We delay the presentation of its technical assumptions",
"In Section 2, we argued that disentangled representations",
"to Section 3.3. See Appendix B.2 for a proof.",
"can improve generalization when combined with sparse task-",
"Theorem 3.1 (Sparse multi-task learning for disentangle-",
"specificpredictors.butwedidnotmentionhowtoobtaina",
"ment). Let be a minimizer of",
"disentangled representation in the first place. In this section,",
"min EpwEp(a,9lw) -log p(y; W(W) fe(a)",
"we first provide a new identification result (Theorem 3.1",
"(4)",
"Section 3.2), which states that in the multi-task learning",
"s.t. W(W) e argmin Ep(a,y/w) -logp(y; Wfe(a) ,",
"setting, regularizing the task-specific predictors to be sparse",
"W",
"can yield disentangled representations. Then, in Section 3.4.",
"IwIl2,0≤l/wil2,0",
"we provide a practical way to learn disentangled representa-",
"tions motivated by our identifiability result.",
"where the constraint holds for all W e W and where Pw",
"and p(, y | W) are described in Section 3.1. Under As-",
"sumptions 3.2 to 3.6 and if f is continuous for all 0, fe is",
"3.1. Task & data generating process",
"disentangled w.r.t. fe (Definition I.1).",
"Throughout this section, we assume the learner is given a",
"set of T datasets [Di, ... , DT] where each dataset Dt :",
"Intuitively, this optimization problem effectively selects a",
"((α(t,i), y(t,i))in-, consists of n couples of input e Rd",
"representation f that (i) allows a perfect fit of the data dis-",
"and label y E y. The set of labels Y might contain either",
"tribution, and (i) allows the task-specific estimators W(w)",
"class indices or real values, depending on whether we are",
"to be as sparse as the ground-truth W. The theorem guaran-",
"concerned with classification or regression tasks.",
"tees that such a representation must be disentangled.",
"Our theory relies on the assumption that, for each task t, the",
"Under the same assumptions and with the same disentan",
"dataset Dt is made of i.i.d. samples from the distribution",
"glement guarantees, Theorem B.6 in Appendix B presents",
"a variation of Problem (4) which enforces the weaker",
"p(c, y / w(t) := p(y; w(t) fe(a)p(α / w(t), (3)",
"constraint EPw lW(W)l2,0 ≤ Epw lWil2.0, instead of",
"IW(W) /2,0 ≤ W/2.0 for each task W individually.",
"where W(t) e IRkxm is the task-specific ground-truth co-",
"efficient matrix. We emphasize that the representation fe",
"Characteristic features of our theory. (i) Contrary to",
"is shared across all the tasks while the coefficient matrices",
"most identifiability results for disentanglement (Section 4),",
"W(t) are task-specific. Also note that the distribution over",
"we do not assume the observations a are generated by trans-",
"4"
] |
{
"arxiv_id": "2211.14666",
"language": "en",
"page_idx": 3
}
|
||
2211.14666-en-4
|
2211.14666-en-4.jpg
|
Synergies Between Disentanglement and Sparsity
<FigureHere>
Figure 2. Illustration of Assumption 3.5 showing three examples of distribution $\mathbb{P}_{W|S}.$ The red distribution satisfies the assumption, but the blue and orange distributions do not. The red lines are level sets of a Gaussian distribution with full rank covariance. The blue line represents the support of a Gaussian distribution with a low-rank covariance. The orange dots represent a distribution with finite support. The green vector $a$ shows that the condition is violated for both the blue and the orange distribution, since, in both cases, $\boldsymbol W_{1,S}$ and $\boldsymbol a$ are orthogonal $(\boldsymbol W_1,S\boldsymbol{a}=0)$ with probability greater than zero.
forming a latent random vector $z$ through a bijective decoder $g.$ Instead, we assume the existence of a not necessarily invertible ground-truth feature extractor $f_{\theta}(x)$ from which the labels can be predicted using only a subset of its components in every task. (ii) Most previous works make assumptions about the distribution of latent factors, e.g., (conditional) independence, exponential family or other parametric assumptions. In contrast, we make no such assumption except a rather weak assumption on the support of the ground-truth features (Assumption 3.3). Crucially, this allows for statistically dependent latent factors, which we explore empirically in Section 5.1.
## 3.3. Assumptions of Theorem 3.1
We now present the technical assumptions of Theorem 3.1. Perhaps unsurprisingly, the parameters $\eta$ have to be identifiable from $p(y;\boldsymbol{\eta})$ in order for $f_{\theta}$ to be identifiable.
Assumption 3.2 (Identifiability of $\eta$ from $p(y;\boldsymbol{\eta})).$ KL$( p( y; \boldsymbol{\eta }) \parallel p( y; \tilde{\boldsymbol{\eta }} ) )$ = 0 $\Longrightarrow$ $\eta$ = $\tilde{\boldsymbol{\eta }} $,where KL denotes the Kullback-Leibler divergence.
This property holds, $e.g.$, when $p(y;\boldsymbol{\eta})$ is a Gaussian in the usual $\mu,\sigma^2$ parameterization. Generally, it also holds for minimal parameterizations of exponential families (Wainwright & Jordan, 2008).
The following assumption requires the ground-truth representation $f_\theta(x)$ to vary enough such that its image cannot be trapped inside a proper subspace.
Assumption 3.3 (Sufficient representation variability). There exists $x^{(1)},\ldots,x^{(m)}\in\mathcal{X}$ such that the matrix $F:=[f_{\boldsymbol{\theta}}(x^{(1)}),\ldots,f_{\boldsymbol{\theta}}(x^{(m)})]$ is invertible.
The following assumption requires that the support of the
<FigureHere>
Figure 3. The leftmost figure represents $S$, the set of task supports observed under the ground-truth distribution $p(S).$ The other figures form a verification that Assumption 3.6 holds for $\mathcal{S}.$
distribution $\mathbb{P}_W$ is sufficiently rich.
Assumption 3.4 (Sufficient task variability). There exists $\boldsymbol{W}^{(1)},\ldots,\boldsymbol{W}^{(m)}\in\mathcal{W}$ and indices $i_1,\ldots,i_m\in[k]$ such that the rows $W_i_1,:^{(1)},\ldots,W_{i_m,:}^{(m)}$ are linearly independent
Under Assumptions 3.2 to 3.4, the representation $f_\mathrm{\theta}$ is identifiable up to linear equivalence (see Theorem B.4 in Appendix B). Similar results were shown by Roeder et al. (2021); Ahuja et al. (2022c). The next assumptions will guarantee disentanglement.
In order to formalize the intuitive idea that most tasks do not require all features, we will denote by $S^{(t)}$ the support of the matrix $W^{(t)},i.e.$,
$$S^{(t)}:=\{j\in[m]\mid\boldsymbol{W}_{:j}^{(t)}\neq\boldsymbol{0}\}\:.$$
In other words, $S^{(t)}$ is the set of features which are useful to predict $y$ in the $t$-th task; note that it is unknown to the learner. For our analysis, we decompose $\mathbb{P}_W$ as
(5)
$$\mathbb{P}_{\boldsymbol{W}}=\sum_{S\in\mathcal{P}([m])}p(S)\mathbb{P}_{\boldsymbol{W}|S}\:,$$
where $\mathcal{P}([m])$ is the collection of all subsets of $[m],p(S)$ is the probability that the support of $W$ is $S$ and $\mathbb{P}_{W|S}$ is the conditional distribution of $W$ given that its support is $S.$ Let $\mathcal{S}$ be the support of the distribution $p(S),i.e.,\mathcal{S}:=\{S\in$ $\mathcal{P}([m])\mid p(S)>0\}.$ The set $\mathcal{S}$ will have an important role in Assumption 3.6.
The following assumption requires that $\mathbb{P}_{W|S}$ does not con-
centrate mass on certain proper subspaces.
Assumption 3.5 (Intra-support sufficient task variability).
$\operatorname{For}$ all $\bar{S}\in\mathcal{S}$ and all $\boldsymbol a\in\mathbb{R}^|S|\backslash\{0\}$,
$$\mathbb{P}_{\boldsymbol{W}|S}\{\boldsymbol{W}\in\mathbb{R}^{k\times m}\mid\boldsymbol{W}_{:S}\boldsymbol{a}=\boldsymbol{0}\}=0\:.$$
We illustrate the above assumption in the simpler case where $k=1.$ For instance, Assumption 3.5 holds when the distribution of $W_{1,S}\mid S$ has a density w.r.t. the Lebesgue measure on $\mathbb{R}^{|S|}$, which is true for example when $W_{1,S}\mid S\sim\mathcal{N}(\mathbf{0},\boldsymbol{\Sigma})$ and the covariance matrix $\Sigma$ is full rank (red distribution in Fig. 2). However, if $\Sigma$ is not full rank, the probability distribution of $W_{1,S}\mid S$ concentrates
5
|
[
[
0.330810546875,
0.05682373046875,
0.64208984375,
0.05682373046875,
0.64208984375,
0.0706787109375,
0.330810546875,
0.0706787109375
],
[
0.2059326171875,
0.08587646484375,
0.2296142578125,
0.08587646484375,
0.2296142578125,
0.0947265625,
0.2059326171875,
0.0947265625
],
[
0.53515625,
0.09661865234375,
0.56201171875,
0.09661865234375,
0.56201171875,
0.10479736328125,
0.53515625,
0.10479736328125
],
[
0.62353515625,
0.0985107421875,
0.6337890625,
0.0985107421875,
0.6337890625,
0.1029052734375,
0.62353515625,
0.1029052734375
],
[
0.51123046875,
0.11871337890625,
0.55810546875,
0.11871337890625,
0.55810546875,
0.1300048828125,
0.51123046875,
0.1300048828125
],
[
0.5908203125,
0.11932373046875,
0.6044921875,
0.11932373046875,
0.6044921875,
0.1275634765625,
0.5908203125,
0.1275634765625
],
[
0.66650390625,
0.121826171875,
0.68115234375,
0.11614990234375,
0.68701171875,
0.125,
0.6708984375,
0.1300048828125
],
[
0.82421875,
0.11932373046875,
0.84228515625,
0.11932373046875,
0.84228515625,
0.126953125,
0.82421875,
0.126953125
],
[
0.53125,
0.136962890625,
0.54345703125,
0.136962890625,
0.54345703125,
0.147705078125,
0.53125,
0.147705078125
],
[
0.60546875,
0.1363525390625,
0.6298828125,
0.1363525390625,
0.6298828125,
0.143310546875,
0.60546875,
0.143310546875
],
[
0.68115234375,
0.136962890625,
0.70751953125,
0.136962890625,
0.70751953125,
0.1458740234375,
0.68115234375,
0.1458740234375
],
[
0.7587890625,
0.1383056640625,
0.78515625,
0.1383056640625,
0.78515625,
0.1458740234375,
0.7587890625,
0.1458740234375
],
[
0.340576171875,
0.1483154296875,
0.3505859375,
0.1483154296875,
0.3505859375,
0.15283203125,
0.340576171875,
0.15283203125
],
[
0.498291015625,
0.1641845703125,
0.8857421875,
0.1654052734375,
0.8857421875,
0.179931640625,
0.498291015625,
0.1787109375
],
[
0.498291015625,
0.1787109375,
0.8857421875,
0.1787109375,
0.8857421875,
0.1925048828125,
0.498291015625,
0.1925048828125
],
[
0.313720703125,
0.186279296875,
0.35693359375,
0.186279296875,
0.35693359375,
0.1976318359375,
0.313720703125,
0.1976318359375
],
[
0.31787109375,
0.1976318359375,
0.34228515625,
0.1976318359375,
0.34228515625,
0.205810546875,
0.31787109375,
0.205810546875
],
[
0.5,
0.1932373046875,
0.8505859375,
0.1932373046875,
0.8505859375,
0.20703125,
0.5,
0.20703125
],
[
0.0882568359375,
0.2203369140625,
0.473876953125,
0.2203369140625,
0.473876953125,
0.2342529296875,
0.0882568359375,
0.2342529296875
],
[
0.0882568359375,
0.23486328125,
0.473876953125,
0.23486328125,
0.473876953125,
0.248779296875,
0.0882568359375,
0.248779296875
],
[
0.498291015625,
0.2342529296875,
0.73193359375,
0.2342529296875,
0.73193359375,
0.2474365234375,
0.498291015625,
0.2474365234375
],
[
0.0882568359375,
0.248779296875,
0.47216796875,
0.25,
0.47216796875,
0.261474609375,
0.0882568359375,
0.260009765625
],
[
0.5009765625,
0.25390625,
0.8857421875,
0.25390625,
0.8857421875,
0.267578125,
0.5009765625,
0.267578125
],
[
0.08905029296875,
0.264404296875,
0.47216796875,
0.264404296875,
0.47216796875,
0.274658203125,
0.08905029296875,
0.274658203125
],
[
0.5009765625,
0.268310546875,
0.5400390625,
0.266357421875,
0.54150390625,
0.279052734375,
0.50244140625,
0.281005859375
],
[
0.53515625,
0.26513671875,
0.88623046875,
0.267578125,
0.88623046875,
0.285400390625,
0.53515625,
0.28271484375
],
[
0.0882568359375,
0.276611328125,
0.473876953125,
0.276611328125,
0.473876953125,
0.290283203125,
0.0882568359375,
0.290283203125
],
[
0.498291015625,
0.2841796875,
0.63134765625,
0.27978515625,
0.63232421875,
0.298583984375,
0.498291015625,
0.302978515625
],
[
0.6259765625,
0.281005859375,
0.875,
0.27978515625,
0.875,
0.3017578125,
0.6259765625,
0.302978515625
],
[
0.08660888671875,
0.290283203125,
0.473876953125,
0.2890625,
0.473876953125,
0.302978515625,
0.08660888671875,
0.3037109375
],
[
0.0882568359375,
0.304931640625,
0.473876953125,
0.304931640625,
0.473876953125,
0.318115234375,
0.0882568359375,
0.318115234375
],
[
0.0882568359375,
0.318115234375,
0.473876953125,
0.318115234375,
0.473876953125,
0.33203125,
0.0882568359375,
0.33203125
],
[
0.498291015625,
0.313232421875,
0.88623046875,
0.314453125,
0.88623046875,
0.328857421875,
0.498291015625,
0.32763671875
],
[
0.0882568359375,
0.330810546875,
0.47216796875,
0.330810546875,
0.47216796875,
0.3447265625,
0.0882568359375,
0.3447265625
],
[
0.5,
0.330078125,
0.88623046875,
0.330078125,
0.88623046875,
0.343505859375,
0.5,
0.343505859375
],
[
0.0906982421875,
0.347900390625,
0.19287109375,
0.347900390625,
0.19287109375,
0.35791015625,
0.0906982421875,
0.35791015625
],
[
0.5009765625,
0.3447265625,
0.88623046875,
0.3447265625,
0.88623046875,
0.35791015625,
0.5009765625,
0.35791015625
],
[
0.5,
0.35986328125,
0.88623046875,
0.35986328125,
0.88623046875,
0.373779296875,
0.5,
0.373779296875
],
[
0.5,
0.375732421875,
0.68115234375,
0.375732421875,
0.68115234375,
0.389404296875,
0.5,
0.389404296875
],
[
0.08905029296875,
0.38330078125,
0.4755859375,
0.38330078125,
0.4755859375,
0.396484375,
0.08905029296875,
0.396484375
],
[
0.0882568359375,
0.3984375,
0.4755859375,
0.3984375,
0.4755859375,
0.412353515625,
0.0882568359375,
0.412353515625
],
[
0.5,
0.396484375,
0.88623046875,
0.396484375,
0.88623046875,
0.410400390625,
0.5,
0.410400390625
],
[
0.0882568359375,
0.41357421875,
0.4755859375,
0.41357421875,
0.4755859375,
0.4267578125,
0.0882568359375,
0.4267578125
],
[
0.496826171875,
0.412353515625,
0.8857421875,
0.412353515625,
0.8857421875,
0.4267578125,
0.496826171875,
0.4267578125
],
[
0.08905029296875,
0.4287109375,
0.473876953125,
0.4287109375,
0.473876953125,
0.442626953125,
0.08905029296875,
0.442626953125
],
[
0.498291015625,
0.425537109375,
0.66015625,
0.4267578125,
0.66015625,
0.44140625,
0.498291015625,
0.43994140625
],
[
0.0882568359375,
0.443115234375,
0.473876953125,
0.443115234375,
0.473876953125,
0.45703125,
0.0882568359375,
0.45703125
],
[
0.5849609375,
0.45068359375,
0.80126953125,
0.45068359375,
0.80126953125,
0.4716796875,
0.5849609375,
0.4716796875
],
[
0.0882568359375,
0.458984375,
0.473876953125,
0.458984375,
0.473876953125,
0.472900390625,
0.0882568359375,
0.472900390625
],
[
0.08660888671875,
0.474853515625,
0.473876953125,
0.472900390625,
0.473876953125,
0.486083984375,
0.08660888671875,
0.488037109375
],
[
0.0882568359375,
0.4892578125,
0.4755859375,
0.4892578125,
0.4755859375,
0.50244140625,
0.0882568359375,
0.50244140625
],
[
0.5009765625,
0.48486328125,
0.8857421875,
0.48486328125,
0.8857421875,
0.5,
0.5009765625,
0.5
],
[
0.08905029296875,
0.5048828125,
0.4755859375,
0.5048828125,
0.4755859375,
0.51806640625,
0.08905029296875,
0.51806640625
],
[
0.5,
0.501953125,
0.8857421875,
0.501953125,
0.8857421875,
0.51513671875,
0.5,
0.51513671875
],
[
0.08905029296875,
0.52099609375,
0.47216796875,
0.52099609375,
0.47216796875,
0.53076171875,
0.08905029296875,
0.53076171875
],
[
0.5,
0.5166015625,
0.81640625,
0.5166015625,
0.81640625,
0.52978515625,
0.5,
0.52978515625
],
[
0.08905029296875,
0.5341796875,
0.473876953125,
0.5341796875,
0.473876953125,
0.54736328125,
0.08905029296875,
0.54736328125
],
[
0.5947265625,
0.54052734375,
0.7900390625,
0.54345703125,
0.7900390625,
0.564453125,
0.5947265625,
0.56103515625
],
[
0.08905029296875,
0.55029296875,
0.338134765625,
0.55029296875,
0.338134765625,
0.56103515625,
0.08905029296875,
0.56103515625
],
[
0.86279296875,
0.546875,
0.88623046875,
0.546875,
0.88623046875,
0.56005859375,
0.86279296875,
0.56005859375
],
[
0.64453125,
0.564453125,
0.70751953125,
0.564453125,
0.70751953125,
0.57568359375,
0.64453125,
0.57568359375
],
[
0.0882568359375,
0.58056640625,
0.31787109375,
0.58056640625,
0.31787109375,
0.5947265625,
0.0882568359375,
0.5947265625
],
[
0.5009765625,
0.5888671875,
0.8857421875,
0.5888671875,
0.8857421875,
0.60302734375,
0.5009765625,
0.60302734375
],
[
0.0882568359375,
0.60400390625,
0.473876953125,
0.60400390625,
0.473876953125,
0.61865234375,
0.0882568359375,
0.61865234375
],
[
0.5,
0.60400390625,
0.8857421875,
0.60400390625,
0.8857421875,
0.61865234375,
0.5,
0.61865234375
],
[
0.5009765625,
0.61865234375,
0.8857421875,
0.61865234375,
0.8857421875,
0.6318359375,
0.5009765625,
0.6318359375
],
[
0.0882568359375,
0.626953125,
0.473876953125,
0.626953125,
0.473876953125,
0.64013671875,
0.0882568359375,
0.64013671875
],
[
0.5009765625,
0.6337890625,
0.8857421875,
0.6337890625,
0.8857421875,
0.6484375,
0.5009765625,
0.6484375
],
[
0.0882568359375,
0.642578125,
0.42236328125,
0.642578125,
0.42236328125,
0.6572265625,
0.0882568359375,
0.6572265625
],
[
0.5,
0.64794921875,
0.8857421875,
0.6484375,
0.8857421875,
0.66357421875,
0.5,
0.662109375
],
[
0.0882568359375,
0.66015625,
0.4755859375,
0.6611328125,
0.4755859375,
0.67529296875,
0.0882568359375,
0.6748046875
],
[
0.5,
0.6640625,
0.62744140625,
0.6640625,
0.62744140625,
0.67822265625,
0.5,
0.67822265625
],
[
0.08905029296875,
0.6767578125,
0.283447265625,
0.6767578125,
0.283447265625,
0.68994140625,
0.08905029296875,
0.68994140625
],
[
0.28759765625,
0.6767578125,
0.473876953125,
0.6767578125,
0.473876953125,
0.68994140625,
0.28759765625,
0.68994140625
],
[
0.0882568359375,
0.68994140625,
0.361083984375,
0.69140625,
0.361083984375,
0.70556640625,
0.0882568359375,
0.705078125
],
[
0.5009765625,
0.6875,
0.88623046875,
0.6875,
0.88623046875,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.5,
0.70263671875,
0.779296875,
0.70263671875,
0.779296875,
0.71630859375,
0.5,
0.71630859375
],
[
0.0882568359375,
0.71826171875,
0.473876953125,
0.71826171875,
0.473876953125,
0.73193359375,
0.0882568359375,
0.73193359375
],
[
0.5009765625,
0.7216796875,
0.88818359375,
0.7216796875,
0.88818359375,
0.736328125,
0.5009765625,
0.736328125
],
[
0.0882568359375,
0.73291015625,
0.473876953125,
0.73291015625,
0.473876953125,
0.7470703125,
0.0882568359375,
0.7470703125
],
[
0.498291015625,
0.736328125,
0.740234375,
0.7373046875,
0.740234375,
0.751953125,
0.498291015625,
0.75048828125
],
[
0.0882568359375,
0.74853515625,
0.47216796875,
0.74853515625,
0.47216796875,
0.76220703125,
0.0882568359375,
0.76220703125
],
[
0.0882568359375,
0.76318359375,
0.2484130859375,
0.76318359375,
0.2484130859375,
0.7763671875,
0.0882568359375,
0.7763671875
],
[
0.5595703125,
0.76123046875,
0.82275390625,
0.76025390625,
0.82275390625,
0.78076171875,
0.5595703125,
0.7822265625
],
[
0.0882568359375,
0.78515625,
0.47216796875,
0.78515625,
0.47216796875,
0.798828125,
0.0882568359375,
0.798828125
],
[
0.0882568359375,
0.80029296875,
0.4755859375,
0.80029296875,
0.4755859375,
0.814453125,
0.0882568359375,
0.814453125
],
[
0.5009765625,
0.80029296875,
0.8857421875,
0.80029296875,
0.8857421875,
0.814453125,
0.5009765625,
0.814453125
],
[
0.08905029296875,
0.81640625,
0.326904296875,
0.81640625,
0.326904296875,
0.830078125,
0.08905029296875,
0.830078125
],
[
0.5,
0.81640625,
0.8857421875,
0.81640625,
0.8857421875,
0.830078125,
0.5,
0.830078125
],
[
0.08905029296875,
0.833984375,
0.47216796875,
0.833984375,
0.47216796875,
0.8486328125,
0.08905029296875,
0.8486328125
],
[
0.5,
0.83056640625,
0.8857421875,
0.83056640625,
0.8857421875,
0.84521484375,
0.5,
0.84521484375
],
[
0.08660888671875,
0.8466796875,
0.4755859375,
0.84765625,
0.4755859375,
0.86572265625,
0.08660888671875,
0.8642578125
],
[
0.5009765625,
0.84521484375,
0.8857421875,
0.84521484375,
0.8857421875,
0.859375,
0.5009765625,
0.859375
],
[
0.0882568359375,
0.86279296875,
0.384033203125,
0.86279296875,
0.384033203125,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.5009765625,
0.861328125,
0.8857421875,
0.861328125,
0.8857421875,
0.87548828125,
0.5009765625,
0.87548828125
],
[
0.5009765625,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.0882568359375,
0.89013671875,
0.473876953125,
0.8916015625,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90478515625
],
[
0.5,
0.892578125,
0.88623046875,
0.892578125,
0.88623046875,
0.90576171875,
0.5,
0.90576171875
],
[
0.476318359375,
0.92529296875,
0.488525390625,
0.92041015625,
0.496826171875,
0.931640625,
0.484375,
0.9375
]
] |
[
"Synergies Between Disentanglement and Sparsity",
"PwIs",
"2",
"",
"3X4",
"3",
"3",
"30",
"s",
"=1",
"=2",
"7=3",
"",
"Figure 3. The leftmost figure represents S, the set of task supports",
"observed under the ground-truth distribution p(S). The other",
"S := (1.2)",
"k=1",
"figures form a verification that Assumption 3.6 holds for S.",
"Figure 2. Ilustration of Assumption 3.5 showing three examples",
"of distribution Pwis. The red distribution satisfies the assumption,",
"distribution Pw is sufficiently rich.",
"but the blue and orange distributions do not.The red lines are",
"Assumption 3.4 (Sufficient task variability). There exists",
"level sets of a Gaussian distribution withfull rank covariance.The",
"W(1)",
"...., W(m) e W and indices ii,..,im e [k] such",
"blue line represents the support of a Gaussian distribution with",
"that the rows w(1),",
",.., Wm) are linearly independent.",
"a low-rank covariance.The orange dots represent a distribution",
"with finite support. The green vector a shows that the condition is",
"violated for both the blue and the orange distribution, since, in both",
"Under Assumptions 3.2 to 3.4, the representation fe is iden-",
"cases, Wi,s and a are orthogonal (Wi,sa = O) with probability",
"tifiable up to linear equivalence (see Theorem B.4 in Ap-",
"greater than zero.",
"pendix B). Similar results were shown by Roeder et al.",
"(2021); Ahuja et al. (2022c). The next assumptions will",
"guarantee disentanglement.",
"forming a latent random vector z through a bijective de-",
"coder g. Instead, we assume the existence of a not nec-",
"In order to formalize the intuitive idea that most tasks do",
"essarily invertible ground-truth feature extractor fe()",
" not require all features, we will denote by s(t) the support",
"from which the labels can be predicted using only a subset",
"of the matrix W(t), i.e.",
"of its components in every task. (ii) Most previous works",
"s(t) := (je [ml I w(t) +0] .",
"make assumptions about the distribution of latent factors,",
"e.g.,(conditional)independence,exponential family or other",
"parametric assumptions. In contrast, we make no such as-",
"In other words, s(t) is the set of features which are useful",
"sumption except a rather weak assumption on the support of",
"to predict y in the t-th task; note that it is unknown to the",
"theground-truthfeatures(Assumption3.3).Crucially,this",
"learner. For our analysis, we decompose Pw as",
"allows for statistically dependent latent factors, which",
"Pw= p(S)PwiS,",
"we exploreempiricallyin Section 5.1.",
"(5)",
"SEP([m])",
"3.3. Assumptions of Theorem 3.1",
"where P([ml) is the collection of all subsets of [ml, p(S) is",
"We now present the technical assumptions of Theorem 3.1.",
"the probability that the support of W is S and Pwis is the",
"conditional distribution of W given that its support is S. Let",
"Perhaps unsurprisingly, the parameters n have to be identifi-",
"S be the support of the distribution p(S), i.e., S := {S E",
"able from p(y; n) in order for fe to be identifiable.",
"P([ml) I p(S) > 0). The set S will have an important role",
"Assumption 3.2 (Identifiability of n from p(y;n)).",
"in Assumption 3.6.",
"KL(p(y;) ll p(y;) = 0",
"→n=,whereKL",
"denotes the Kullback-Leibler divergence.",
"The following assumption requires that Pwis does not con-",
"centrate mass on certain proper subspaces.",
"This property holds, e.g., when p(y; ) is a Gaussian in the",
"Assumption 3.5 (Intra-support sufficient task variability).",
"usual μ, o? parameterization. Generally, it also holds for",
"For all S S and all a E Ris//(0],",
"minimal parameterizations of exponential families (Wain",
"wright & Jordan,2008).",
"Pwis(W e Rkxm [ W:sa = 0) = 0.",
"The following assumption requires the ground-truth repre",
"sentation fe() to vary enough such that its image cannot",
"We illustrate the above assumption in the simpler case",
"be trapped inside a proper subspace.",
"where k = 1. For instance, Assumption 3.5 holds when",
"Assumption 3.3 (Sufficient representation variability)",
"the distribution of Wi,s I S has a density w.r.t. the",
"There exists e(1),.,e(m) e X such that the matrix",
"Lebesgue measure on RIsl, which is true for example when",
"F :- [fe(ae(1),..., fe(a(m)] is invertible.",
"Wi,s I S ~ N(o, ) and the covariance matrix is full",
"rank (red distribution in Fig. 2). However, if is not full",
"The following assumption requires that the support of the",
"rank, the probability distribution of Wi,s I S concentrates",
"5"
] |
{
"arxiv_id": "2211.14666",
"language": "en",
"page_idx": 4
}
|
||
2005.00123-en-0
|
2005.00123-en-0.jpg
|
$\textbf{Unsupervised Learning of KB Queries in Task- Oriented Dialogs}$
Dinesh Raghu$^{*12},\textbf{Nikhil Gupta}^{\dagger3}$ and Mausam$^{1}$
$^{1}$IIT Delhi, New Delhi, India
$^{2}$ IBM Research, New Delhi, India
$^3$ LimeChat, Gurgaon, India
$diraghul@in.ibm.com,nikhil@limechat.ai,mausam@cse.iitd.ac.in$
### $\mathbf{Abstract}$
Task-oriented dialog (TOD) systems often need to formulate knowledge base (KB) queries corresponding to the user intent and use the query results to generate system responses. Existing approaches require dialog datasets to explicitly annotate these KB queries- these annotations can be time consuming, and expensive. In response, we define the novel problems of predicting the KB query and training the dialog agent, without explicit KB query annotation. For query prediction, we propose a reinforcement learning (RL) baseline, which rewards the generation of those queries whose KB results cover the entities mentioned in subsequent dialog. Further analysis reveals that correlation among query attributes in KB can significantly confuse memory augmented policy optimization (MAPO), an existing state of the art RL agent. To address this, we improve the MAPO baseline with simple but important modifications suited to our task.
To train the full TOD system for our setting, we propose a pipelined approach: it independently predicts when to make a KB query (query position predictor), then predicts a KB query at the predicted position (query predictor), and uses the results of predicted query in subsequent dialog (next response predictor). Overall, our work proposes first solutions to our novel problem, and our analysis highlights the research challenges in training TOD systems without query annotation.
## 1 Introduction
Task-oriented dialog (TOD) systems converse with users to accomplish specific tasks such as restaurant reservation (Henderson et al., 2014b), movie
$^*$D.Raghu is an employee at IBM Research. This work
was carried out as part of PhD research at IIT Delhi.
$^{\dagger}$This work was done while Nikhil Gupta was a graduate
student at IIT Delhi.
ticket booking (Li et al., 2017) or bus enquiry (Raux et al., 2005). In addition to the ability to converse, it is crucial for TOD systems to learn to formulate knowledge base (KB) queries based on the user needs, and generate responses using the query results. An example task oriented dialog is shown in Figure 1, where during the conversation (at turn 2), the agent queries the KB based on the user needs, and then suggests the Peking Restaurant based on the retrieved results. Existing end-to-end approaches (Bordes and Weston, 2017; Madotto et al., 2018; Reddy et al., 2019) learn to formulate KB queries using manually annotated queries.
In real-world scenarios, human agents chat with users on messaging platforms. When the need to query the KB arises, the agent fires the query on a back-end KB application and uses the retrieved results to compose a response back on the messaging platform. The dialogs retrieved from these platforms contain just the user and the agent utterances, but the KB queries typically go undocumented. As existing approaches require KB annotations, they have to be manually annotated, which is expensive and hinders scalability. To eliminate the need for such annotations, we define the novel problem of training a TOD system without explicit KB query annotation. A key subtask of such a system is the unsupervised prediction of KB queries.
While there is no explicit query annotation, we observe that dialog data still offers weak supervision to induce KB queries- all the entities used by the agent in subsequent dialog should be returned by the correct query. For example, in Figure 1, the correct query should retrieve Peking restaurant and its phone number. This suggests a reinforcement learning (RL) formulation for query prediction, not unlike similar tasks in Question Answering (QA) (Artzi and Zettlemoyer, 2013) and conversational QA (Yu et al., 2019). However, there is one key difference between TOD and QA settings- in TOD only a few entities from the query results are used
|
[
[
0.17041015625,
0.08074951171875,
0.8310546875,
0.08197021484375,
0.8310546875,
0.09979248046875,
0.17041015625,
0.09857177734375
],
[
0.300537109375,
0.12054443359375,
0.70947265625,
0.12054443359375,
0.70947265625,
0.1383056640625,
0.300537109375,
0.1383056640625
],
[
0.382080078125,
0.1383056640625,
0.623046875,
0.14013671875,
0.623046875,
0.155029296875,
0.382080078125,
0.152587890625
],
[
0.359375,
0.152587890625,
0.6455078125,
0.155029296875,
0.6455078125,
0.1728515625,
0.359375,
0.17041015625
],
[
0.3896484375,
0.17333984375,
0.6171875,
0.17333984375,
0.6171875,
0.188232421875,
0.3896484375,
0.188232421875
],
[
0.2333984375,
0.190673828125,
0.77685546875,
0.190673828125,
0.77685546875,
0.204833984375,
0.2333984375,
0.204833984375
],
[
0.264404296875,
0.243408203125,
0.343505859375,
0.243408203125,
0.343505859375,
0.25830078125,
0.264404296875,
0.25830078125
],
[
0.51318359375,
0.243408203125,
0.88330078125,
0.24462890625,
0.88330078125,
0.2587890625,
0.51318359375,
0.25830078125
],
[
0.51318359375,
0.2587890625,
0.88330078125,
0.260009765625,
0.88330078125,
0.27490234375,
0.51318359375,
0.273681640625
],
[
0.1561279296875,
0.270751953125,
0.45166015625,
0.270751953125,
0.45166015625,
0.283935546875,
0.1561279296875,
0.283935546875
],
[
0.5146484375,
0.276611328125,
0.88330078125,
0.276611328125,
0.88330078125,
0.290283203125,
0.5146484375,
0.290283203125
],
[
0.1561279296875,
0.28515625,
0.453369140625,
0.28515625,
0.453369140625,
0.298583984375,
0.1561279296875,
0.298583984375
],
[
0.5146484375,
0.292236328125,
0.88330078125,
0.292236328125,
0.88330078125,
0.30712890625,
0.5146484375,
0.30712890625
],
[
0.1561279296875,
0.2998046875,
0.453369140625,
0.2998046875,
0.453369140625,
0.3134765625,
0.1561279296875,
0.3134765625
],
[
0.51123046875,
0.30810546875,
0.8818359375,
0.30810546875,
0.8818359375,
0.32177734375,
0.51123046875,
0.32177734375
],
[
0.154541015625,
0.314208984375,
0.453369140625,
0.314208984375,
0.453369140625,
0.327880859375,
0.154541015625,
0.327880859375
],
[
0.0235137939453125,
0.32177734375,
0.060455322265625,
0.32177734375,
0.060455322265625,
0.4150390625,
0.0235137939453125,
0.4150390625
],
[
0.154541015625,
0.328857421875,
0.453369140625,
0.32666015625,
0.453369140625,
0.34033203125,
0.154541015625,
0.342529296875
],
[
0.51318359375,
0.32373046875,
0.8818359375,
0.32373046875,
0.8818359375,
0.33740234375,
0.51318359375,
0.33740234375
],
[
0.1561279296875,
0.342529296875,
0.45166015625,
0.342529296875,
0.45166015625,
0.355712890625,
0.1561279296875,
0.355712890625
],
[
0.51318359375,
0.34033203125,
0.8818359375,
0.34033203125,
0.8818359375,
0.35400390625,
0.51318359375,
0.35400390625
],
[
0.1561279296875,
0.355712890625,
0.453369140625,
0.355712890625,
0.453369140625,
0.369384765625,
0.1561279296875,
0.369384765625
],
[
0.51318359375,
0.35693359375,
0.88330078125,
0.355712890625,
0.88330078125,
0.37060546875,
0.51318359375,
0.371826171875
],
[
0.1561279296875,
0.37060546875,
0.453369140625,
0.37060546875,
0.453369140625,
0.38427734375,
0.1561279296875,
0.38427734375
],
[
0.51318359375,
0.37353515625,
0.88330078125,
0.37353515625,
0.88330078125,
0.38720703125,
0.51318359375,
0.38720703125
],
[
0.1561279296875,
0.38427734375,
0.45166015625,
0.38427734375,
0.45166015625,
0.39794921875,
0.1561279296875,
0.39794921875
],
[
0.5146484375,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.402099609375,
0.5146484375,
0.402099609375
],
[
0.1561279296875,
0.3984375,
0.45166015625,
0.3984375,
0.45166015625,
0.412109375,
0.1561279296875,
0.412109375
],
[
0.0277099609375,
0.40625,
0.056243896484375,
0.40625,
0.056243896484375,
0.427001953125,
0.0277099609375,
0.427001953125
],
[
0.5146484375,
0.40625,
0.884765625,
0.40625,
0.884765625,
0.419189453125,
0.5146484375,
0.419189453125
],
[
0.1561279296875,
0.413330078125,
0.453369140625,
0.413330078125,
0.453369140625,
0.427001953125,
0.1561279296875,
0.427001953125
],
[
0.1561279296875,
0.427001953125,
0.45166015625,
0.427001953125,
0.45166015625,
0.440673828125,
0.1561279296875,
0.440673828125
],
[
0.51318359375,
0.421630859375,
0.884765625,
0.421630859375,
0.884765625,
0.435302734375,
0.51318359375,
0.435302734375
],
[
0.0201568603515625,
0.43408203125,
0.05877685546875,
0.432861328125,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51123046875
],
[
0.154541015625,
0.440673828125,
0.45166015625,
0.441162109375,
0.45166015625,
0.454833984375,
0.154541015625,
0.45361328125
],
[
0.5146484375,
0.438232421875,
0.85400390625,
0.438232421875,
0.85400390625,
0.451904296875,
0.5146484375,
0.451904296875
],
[
0.154541015625,
0.454833984375,
0.45166015625,
0.45361328125,
0.45166015625,
0.468505859375,
0.154541015625,
0.4697265625
],
[
0.53076171875,
0.457275390625,
0.88330078125,
0.457275390625,
0.88330078125,
0.47021484375,
0.53076171875,
0.47021484375
],
[
0.1561279296875,
0.4697265625,
0.453369140625,
0.4697265625,
0.453369140625,
0.482666015625,
0.1561279296875,
0.482666015625
],
[
0.51318359375,
0.47265625,
0.884765625,
0.471435546875,
0.884765625,
0.486328125,
0.51318359375,
0.48681640625
],
[
0.154541015625,
0.48388671875,
0.45166015625,
0.48388671875,
0.45166015625,
0.49755859375,
0.154541015625,
0.49755859375
],
[
0.51318359375,
0.4892578125,
0.884765625,
0.4892578125,
0.884765625,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.154541015625,
0.498779296875,
0.45166015625,
0.49755859375,
0.45166015625,
0.51123046875,
0.154541015625,
0.5126953125
],
[
0.51318359375,
0.505859375,
0.88671875,
0.505859375,
0.88671875,
0.51953125,
0.51318359375,
0.51953125
],
[
0.0218353271484375,
0.51416015625,
0.0621337890625,
0.51416015625,
0.0621337890625,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.1561279296875,
0.5126953125,
0.453369140625,
0.5126953125,
0.453369140625,
0.525390625,
0.1561279296875,
0.525390625
],
[
0.1561279296875,
0.525390625,
0.453369140625,
0.525390625,
0.453369140625,
0.5390625,
0.1561279296875,
0.5390625
],
[
0.51123046875,
0.52099609375,
0.884765625,
0.521484375,
0.884765625,
0.5361328125,
0.51123046875,
0.53515625
],
[
0.1561279296875,
0.54052734375,
0.453369140625,
0.54052734375,
0.453369140625,
0.55419921875,
0.1561279296875,
0.55419921875
],
[
0.51318359375,
0.5380859375,
0.88671875,
0.5380859375,
0.88671875,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.154541015625,
0.5546875,
0.453369140625,
0.5546875,
0.453369140625,
0.568359375,
0.154541015625,
0.568359375
],
[
0.51318359375,
0.55419921875,
0.88671875,
0.55419921875,
0.88671875,
0.56787109375,
0.51318359375,
0.56787109375
],
[
0.156982421875,
0.57080078125,
0.276123046875,
0.57080078125,
0.276123046875,
0.58056640625,
0.156982421875,
0.58056640625
],
[
0.51318359375,
0.5693359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.58447265625,
0.51318359375,
0.58447265625
],
[
0.15283203125,
0.5859375,
0.453369140625,
0.58740234375,
0.453369140625,
0.6015625,
0.15283203125,
0.60107421875
],
[
0.51318359375,
0.5859375,
0.88330078125,
0.5849609375,
0.88330078125,
0.599609375,
0.51318359375,
0.60107421875
],
[
0.1561279296875,
0.6015625,
0.453369140625,
0.6015625,
0.453369140625,
0.615234375,
0.1561279296875,
0.615234375
],
[
0.5146484375,
0.60400390625,
0.8818359375,
0.60400390625,
0.8818359375,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.1561279296875,
0.6162109375,
0.45166015625,
0.6162109375,
0.45166015625,
0.6298828125,
0.1561279296875,
0.6298828125
],
[
0.5146484375,
0.6181640625,
0.88330078125,
0.6181640625,
0.88330078125,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.1561279296875,
0.63134765625,
0.453369140625,
0.63134765625,
0.453369140625,
0.64453125,
0.1561279296875,
0.64453125
],
[
0.51318359375,
0.6328125,
0.88330078125,
0.6328125,
0.88330078125,
0.646484375,
0.51318359375,
0.646484375
],
[
0.15869140625,
0.646484375,
0.450927734375,
0.646484375,
0.450927734375,
0.65673828125,
0.15869140625,
0.65673828125
],
[
0.51318359375,
0.65087890625,
0.8818359375,
0.65087890625,
0.8818359375,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.1561279296875,
0.6591796875,
0.45166015625,
0.6591796875,
0.45166015625,
0.6728515625,
0.1561279296875,
0.6728515625
],
[
0.51318359375,
0.66650390625,
0.88330078125,
0.66650390625,
0.88330078125,
0.6796875,
0.51318359375,
0.6796875
],
[
0.1561279296875,
0.6728515625,
0.45166015625,
0.6728515625,
0.45166015625,
0.68603515625,
0.1561279296875,
0.68603515625
],
[
0.1561279296875,
0.68603515625,
0.45166015625,
0.68603515625,
0.45166015625,
0.69970703125,
0.1561279296875,
0.69970703125
],
[
0.51318359375,
0.68310546875,
0.80419921875,
0.68310546875,
0.80419921875,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.154541015625,
0.70166015625,
0.453369140625,
0.70166015625,
0.453369140625,
0.71533203125,
0.154541015625,
0.71533203125
],
[
0.5322265625,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.5322265625,
0.71533203125
],
[
0.154541015625,
0.71533203125,
0.453369140625,
0.71533203125,
0.453369140625,
0.728515625,
0.154541015625,
0.728515625
],
[
0.51318359375,
0.71875,
0.884765625,
0.71875,
0.884765625,
0.732421875,
0.51318359375,
0.732421875
],
[
0.154541015625,
0.72998046875,
0.455078125,
0.72998046875,
0.455078125,
0.74365234375,
0.154541015625,
0.74365234375
],
[
0.5146484375,
0.73388671875,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.154541015625,
0.74462890625,
0.18896484375,
0.74462890625,
0.18896484375,
0.75732421875,
0.154541015625,
0.75732421875
],
[
0.51318359375,
0.75,
0.88330078125,
0.7490234375,
0.88330078125,
0.76318359375,
0.51318359375,
0.76416015625
],
[
0.5146484375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.124267578125,
0.7822265625,
0.130126953125,
0.7822265625,
0.130126953125,
0.7880859375,
0.124267578125,
0.7880859375
],
[
0.1351318359375,
0.783203125,
0.1427001953125,
0.783203125,
0.1427001953125,
0.78759765625,
0.1351318359375,
0.78759765625
],
[
0.1427001953125,
0.783203125,
0.1485595703125,
0.783203125,
0.1485595703125,
0.7880859375,
0.1427001953125,
0.7880859375
],
[
0.1435546875,
0.779296875,
0.26123046875,
0.779296875,
0.26123046875,
0.79345703125,
0.1435546875,
0.79345703125
],
[
0.5146484375,
0.78369140625,
0.8818359375,
0.78369140625,
0.8818359375,
0.79443359375,
0.5146484375,
0.79443359375
],
[
0.1175537109375,
0.8017578125,
0.489501953125,
0.802734375,
0.489501953125,
0.81787109375,
0.1175537109375,
0.81640625
],
[
0.51123046875,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.51123046875,
0.8125
],
[
0.51318359375,
0.81396484375,
0.884765625,
0.81396484375,
0.884765625,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.1175537109375,
0.82080078125,
0.491943359375,
0.82080078125,
0.491943359375,
0.83447265625,
0.1175537109375,
0.83447265625
],
[
0.118408203125,
0.8359375,
0.489501953125,
0.8359375,
0.489501953125,
0.849609375,
0.118408203125,
0.849609375
],
[
0.51318359375,
0.83056640625,
0.884765625,
0.83056640625,
0.884765625,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.51318359375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.145263671875,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.145263671875,
0.869140625
],
[
0.5146484375,
0.8623046875,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.5146484375,
0.876953125
],
[
0.118408203125,
0.87158203125,
0.438232421875,
0.87158203125,
0.438232421875,
0.8818359375,
0.118408203125,
0.8818359375
],
[
0.145263671875,
0.88134765625,
0.495361328125,
0.8818359375,
0.495361328125,
0.896484375,
0.145263671875,
0.8955078125
],
[
0.5146484375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.12176513671875,
0.89794921875,
0.2418212890625,
0.89794921875,
0.2418212890625,
0.90771484375,
0.12176513671875,
0.90771484375
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Unsupervised Learning of KB Queries in Task-Oriented Dialogs",
"Dinesh Raghu*1 2, Nikhil Guptat3 and Mausam'",
"1 IIT Delhi, New Delhi, India",
"2 IBM Research, New Delhi, India",
"3 LimeChat, Gurgaon, India",
"diraghul@in.ibm.com, nikhil@ limechat.ai, mausam@cse.itd.ac.in",
"Abstract",
"ticket booking (Li et al., 2017) or bus enquiry (Raux",
"et al., 2005). In addition to the ability to converse,",
"Task-oriented dialog (TOD) systems often",
"it is crucial for TOD systems to learn to formulate",
"need to formulate knowledge base (KB)",
"knowledge base (KB) queries based on the user",
"queries corresponding to the user intent and",
"needs, and generate responses using the query re",
"use the query results to generate system re-",
"12z",
"sponses. Existing approaches require dia-",
"sults. An example task oriented dialog is shown",
"log datasets to explicitly annotate these KB",
"in Figure 1, where during the conversation (at turn",
"queries - these annotations can be time con-",
"2), the agent queries the KB based on the user",
"suming, and expensive. In response, we de-",
"needs, and then suggests the Peking Restaurant",
"fine the novel problems of predicting the KB",
"basedontheretrievedresults.Existingend-to-end",
"query and training the dialog agent, without",
"3",
"approaches (Bordes and Weston, 2017; Madotto",
"explicit KB query annotation. For query pre-",
"diction, we propose a reinforcement learning",
"et al., 2018; Reddy et al., 2019) learn to formulate",
"[DT's]",
"(RL)baseline,which rewards the generation",
"KB queries using manually annotated queries.",
"of those queries whose KB results cover the",
"In real-world scenarios, human agents chat with",
"entities mentioned in subsequent dialog. Fur-",
"users on messaging platforms. When the need to",
"ther analysis reveals that correlation among",
"query the KB arises, the agent fires the query on a",
"query attributes in KB can significantly con",
"back-end KB application and uses the retrieved re-",
"ZAEZ100'S02:AIX",
"fuse memory augmented policy optimization",
"(MAPO), an existing state of the art RL agent.",
"sults to compose a response back on the messaging",
"To address this, we improve the MAPO base-",
"platform. The dialogs retrieved from these plat-",
"line with simple but important modifications",
"forms contain just the user and the agent utterances,",
"suited to our task.",
"but the KB queries typically go undocumented. As",
"To train the full TOD system for our setting,",
"existing approaches require KB annotations, they",
"we propose a pipelined approach: it indepen-",
"havetobemanuallyannotated.whichis expensive",
"dently predicts when to make a KB query",
"and hinders scalability. To eliminate the need for",
"(query position predictor), then predicts a",
"such annotations, we define the novel problem of",
"KBquery atthepredictedposition(query",
"training a TOD system without explicit KB query",
"predictor), and uses the results of predicted",
"annotation. A key subtask of such a system is the",
"query in subsequent dialog (next response",
"predictor). Overall, our work proposes first",
"unsupervised prediction of KB queries.",
"solutions to our novel problem, and our anal-",
"While there is no explicit query annotation, we",
"ysis highlights the research challenges in",
"observe that dialog data still offers weak supervi-",
"training TOD systems without query annota-",
"sion to induce KB queries - all the entities used by",
"tion.",
"the agent in subsequent dialog should be returned",
"by the correct query. For example, in Figure l, the",
"",
"",
"",
"Introduction",
"correct query should retrieve Peking restaurant and",
"Task-oriented dialog (TOD) systems converse with",
"its phone number. This suggests a reinforcement",
"learning (RL) formulation for query prediction, not",
"users to accomplish specific tasks such as restau-",
"rant reservation (Henderson et al., 2014b), movie",
"unlike similar tasks in Question Answering (QA)",
"(Artzi and Zettlemoyer, 2013) and conversational",
"*D. Raghu is an employee at IBM Research. This work",
"QA (Yu et al., 2019). However, there is one key",
"was carried out as part of PhDresearch at IIT Delhi.",
"t This work was done while Nikhil Gupta was a graduate",
"difference between TOD and QA settings - in TOD",
"student atIIT Delhi",
"only a few entities from the query results are used"
] |
{
"arxiv_id": "2005.00123",
"language": "en",
"page_idx": 0
}
|
||
2005.00123-en-3
|
2005.00123-en-3.jpg
|
hence negligible contribution to the gradient estimation. To ensure queries in the buffer contribute significantly to the gradient estimates, MAPO clips $\pi_{\mathcal{B}}$ to $\alpha.$ The modified gradient estimate is,
$$\nabla_{\theta}\mathcal{O}_{ER}^{c}=\pi_{\mathcal{B}}^{c}\mathbb{E}_{\mathbf{a}\sim\pi_{\theta}^{+}(\mathbf{a})}\nabla_{\theta}\mathrm{log}\pi_{\theta}(\mathbf{a})R(\mathbf{a})\\+(1-\pi_{\mathcal{B}}^{c})\mathbb{E}_{\mathbf{a}\sim\pi_{\theta}^{-}(\mathbf{a})}\nabla_{\theta}\mathrm{log}\pi_{\theta}(\mathbf{a})R(\mathbf{a})$$
(3)
where $\pi_{\mathcal{B}}^{c}=\max(\pi_{\mathcal{B}},\alpha).$ When the training begins, $\alpha>\pi_{\mathcal{B}}$ resulting in gradient estimates biased towards the queries in the buffer. Once the policy stabilizes, $\pi_{\mathcal{B}}$ gets larger than $\alpha$ and the estimates becomes unbiased.
Experience Replay: Experience replay (Lin, 1992) stores past experience in a buffer and reuses that to stabilize training and improve sample effciency. Prioritized experience replay (Schaul et al., 2016) assigns priorities to the experiences and samples them based on the priorities to efficiently learn using them. MAPO and our approach use replay buffers which store non-zero reward queries. These queries include both past experiences and queries identified using systematic search. The sampling strategy is similar to prioritized experience replay, where the priorities are defined by the probability of the query in the current policy.
# 3 Task Definition & Baseline System
We first define our novel task of learning TOD system using unannotated dialogs-we name it uTOD. We then describe the KB query predictor in detail. Finally, we describe our proposed baseline uTOD system, which uses the KB query predictor.
## 3.1 Problem Definition
We represent a dialog $d$ between a user $u$ and an agent $s$ as $\{c_1^u,c_1^s,c_2^u,c_2^s,\ldots,c_m^u,c_m^s\}$ where $m$ denotes the number of turns in the dialog. Our goal is to train a uTOD system, which, for all turns $i$, takes the partial dialog-so-far $\{c_{1}^{u},c_{1}^{s},\ldots,c_{i-1}^{u},c_{i-1}^{s},c_{i}^{u}\}$ as input and predicts the next system response $\hat{c}_i^\mathrm{s}$. Such a uTOD system is trained by the training data $\mathcal{D}$ comprising complete dialogs $\{d_{j}\}_{j=1}^{|\mathcal{D}|}$, and an associated knowledge base $KB.$
System responses in a dialog often use certain entities from $KB.$ To accomplish this, a uTOD system can fre an appropriate query a to $KB$ and fetch the query results $E^a$, and use those to generate its response. No explicit supervision is provided on the gold queries, at either training or test time.
I.e., a uTOD system neither knows which query was fired to $KB$, nor knows when it was fired. In this work we assume that the system can make only one KB query per dialog. Our overall uTOD baseline uses a pipeline of multiple components - a query position predictor, a query predictor and a next response predictor.
## 3.2 KB Query Predictor
We first assume that we have access to the specific turn number $q,1\leq q\leq m$, where a KB query gets fired by the system (an assumption we relax in the next section). Thus, formally, we are given the dialog context $c=\{c_1^u,c_1^s,c_2^u,c_2^s,\ldots,c_q^u\}$,and $KB$, and the goal of query predictor is to output a that appropriately captures the user intent. We use the term subsequent dialog to refer to all utterances that follow the KB query$-\{c_q^s,c_{q+1}^u,\ldots,c_m^s\}.$In Figure 1, the subsequent dialog starts from the second agent utterance. Let $E^s$ be the set of KB entities present in the subsequent dialog. In Figure 1, $E^{s}=\{PekingRestaurant,2343-4040\}.$ We train the query predictor using RL, by providing feedback based on the predicted query's ability to retrieve the entities in $E^{s}.$
Our proposed baseline follows the literature on semantic parsing with weak supervision using RL, and treats the query prediction as equivalent to learning policy $\pi_\theta(\mathbf{a}|c)$ that takes the dialog context $c$ as the input and generates a KB query $\mathbf{a}=a_1a_2\cdots a_T.$ The KB query is a sequence of $T$ actions, where each action is a word predicted by the query predictor. For a given context, the set of all possible actions is a union of a set of keywords$^{2}$in the SQL query language, set of field names in $KB$, the set of all words in the context $c$ and an <eoq> token to indicate the end of query. The query is generated auto-regressively as:
$$\pi_\theta(\mathbf{a}|c)=\prod\limits_{t=1}^T\pi_\theta(a_t|a_{1:t-1},c)$$
(4)
As the environment is deterministic, at each time $t$, the RL state can be fully defined by the dialog context $c$ and the actions predicted so far $a_1:t-1.$ RL gradients can be computed using REINFORCE or the MAPO algorithm.
The policy network is implemented with a standard encoder-decoder architecture for TOD systems. The context is encoded using a multi-hop
$^{2}\{$SELECT$,^*,=$,FROM,AND,WHERE$\}$
|
[
[
0.12005615234375,
0.07659912109375,
0.490234375,
0.07659912109375,
0.490234375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.12005615234375,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51318359375,
0.0926513671875,
0.88330078125,
0.0926513671875,
0.88330078125,
0.105712890625,
0.51318359375,
0.105712890625
],
[
0.12005615234375,
0.10809326171875,
0.491943359375,
0.10809326171875,
0.491943359375,
0.122314453125,
0.12005615234375,
0.122314453125
],
[
0.51318359375,
0.10986328125,
0.880859375,
0.10986328125,
0.880859375,
0.12054443359375,
0.51318359375,
0.12054443359375
],
[
0.12005615234375,
0.12469482421875,
0.438232421875,
0.12469482421875,
0.438232421875,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.51123046875,
0.12469482421875,
0.884765625,
0.12353515625,
0.884765625,
0.1383056640625,
0.51123046875,
0.1395263671875
],
[
0.5146484375,
0.141357421875,
0.884765625,
0.141357421875,
0.884765625,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.1427001953125,
0.152587890625,
0.44580078125,
0.152587890625,
0.44580078125,
0.17333984375,
0.1427001953125,
0.17333984375
],
[
0.5146484375,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.171630859375,
0.5146484375,
0.171630859375
],
[
0.16455078125,
0.177001953125,
0.491943359375,
0.177001953125,
0.491943359375,
0.19775390625,
0.16455078125,
0.19775390625
],
[
0.51318359375,
0.17333984375,
0.69287109375,
0.171630859375,
0.69287109375,
0.1859130859375,
0.51318359375,
0.188232421875
],
[
0.51318359375,
0.20068359375,
0.71875,
0.20068359375,
0.71875,
0.21435546875,
0.51318359375,
0.21435546875
],
[
0.1376953125,
0.210205078125,
0.486083984375,
0.210205078125,
0.486083984375,
0.2244873046875,
0.1376953125,
0.2244873046875
],
[
0.118408203125,
0.226806640625,
0.48779296875,
0.226806640625,
0.48779296875,
0.240478515625,
0.118408203125,
0.240478515625
],
[
0.51123046875,
0.22265625,
0.880859375,
0.22265625,
0.880859375,
0.236328125,
0.51123046875,
0.236328125
],
[
0.12005615234375,
0.2423095703125,
0.490234375,
0.2423095703125,
0.490234375,
0.257080078125,
0.12005615234375,
0.257080078125
],
[
0.51318359375,
0.2392578125,
0.88330078125,
0.2392578125,
0.88330078125,
0.2529296875,
0.51318359375,
0.2529296875
],
[
0.118408203125,
0.2587890625,
0.491943359375,
0.2587890625,
0.491943359375,
0.273681640625,
0.118408203125,
0.273681640625
],
[
0.51318359375,
0.254638671875,
0.88330078125,
0.254638671875,
0.88330078125,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.118408203125,
0.27490234375,
0.33154296875,
0.27490234375,
0.33154296875,
0.2880859375,
0.118408203125,
0.2880859375
],
[
0.51318359375,
0.26953125,
0.884765625,
0.270751953125,
0.884765625,
0.28515625,
0.51318359375,
0.283935546875
],
[
0.1376953125,
0.29150390625,
0.491943359375,
0.29150390625,
0.491943359375,
0.3046875,
0.1376953125,
0.3046875
],
[
0.51318359375,
0.2861328125,
0.88671875,
0.2861328125,
0.88671875,
0.303955078125,
0.51318359375,
0.303955078125
],
[
0.118408203125,
0.30712890625,
0.489501953125,
0.30712890625,
0.489501953125,
0.32177734375,
0.118408203125,
0.32177734375
],
[
0.5146484375,
0.3017578125,
0.88330078125,
0.302734375,
0.88330078125,
0.317626953125,
0.5146484375,
0.31640625
],
[
0.12005615234375,
0.32373046875,
0.490234375,
0.32373046875,
0.490234375,
0.33740234375,
0.12005615234375,
0.33740234375
],
[
0.5146484375,
0.319580078125,
0.88330078125,
0.319580078125,
0.88330078125,
0.334228515625,
0.5146484375,
0.334228515625
],
[
0.118408203125,
0.34033203125,
0.491943359375,
0.34033203125,
0.491943359375,
0.35400390625,
0.118408203125,
0.35400390625
],
[
0.5146484375,
0.3349609375,
0.88330078125,
0.3349609375,
0.88330078125,
0.3486328125,
0.5146484375,
0.3486328125
],
[
0.118408203125,
0.355224609375,
0.490234375,
0.355224609375,
0.490234375,
0.3681640625,
0.118408203125,
0.3681640625
],
[
0.51123046875,
0.349853515625,
0.88330078125,
0.3515625,
0.88330078125,
0.3681640625,
0.51123046875,
0.366455078125
],
[
0.12005615234375,
0.371826171875,
0.491943359375,
0.371826171875,
0.491943359375,
0.385498046875,
0.12005615234375,
0.385498046875
],
[
0.51318359375,
0.366455078125,
0.88330078125,
0.36767578125,
0.88330078125,
0.3818359375,
0.51318359375,
0.38134765625
],
[
0.118408203125,
0.38720703125,
0.489501953125,
0.38720703125,
0.489501953125,
0.40087890625,
0.118408203125,
0.40087890625
],
[
0.51318359375,
0.38427734375,
0.88330078125,
0.38427734375,
0.88330078125,
0.39794921875,
0.51318359375,
0.39794921875
],
[
0.118408203125,
0.40380859375,
0.490234375,
0.40380859375,
0.490234375,
0.41748046875,
0.118408203125,
0.41748046875
],
[
0.51318359375,
0.399658203125,
0.88330078125,
0.399658203125,
0.88330078125,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.12005615234375,
0.42041015625,
0.489501953125,
0.42041015625,
0.489501953125,
0.43408203125,
0.12005615234375,
0.43408203125
],
[
0.5146484375,
0.4150390625,
0.88330078125,
0.4150390625,
0.88330078125,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.12005615234375,
0.435791015625,
0.489501953125,
0.435791015625,
0.489501953125,
0.449462890625,
0.12005615234375,
0.449462890625
],
[
0.5146484375,
0.431640625,
0.88330078125,
0.431640625,
0.88330078125,
0.4453125,
0.5146484375,
0.4453125
],
[
0.12005615234375,
0.451904296875,
0.490234375,
0.451904296875,
0.490234375,
0.465576171875,
0.12005615234375,
0.465576171875
],
[
0.51318359375,
0.44775390625,
0.88330078125,
0.44775390625,
0.88330078125,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.118408203125,
0.46728515625,
0.489501953125,
0.46728515625,
0.489501953125,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51318359375,
0.46435546875,
0.68994140625,
0.46435546875,
0.68994140625,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.1175537109375,
0.482666015625,
0.365234375,
0.48388671875,
0.365234375,
0.498779296875,
0.1175537109375,
0.49755859375
],
[
0.52880859375,
0.479736328125,
0.88330078125,
0.479736328125,
0.88330078125,
0.49462890625,
0.52880859375,
0.49462890625
],
[
0.51123046875,
0.49755859375,
0.8818359375,
0.49755859375,
0.8818359375,
0.51123046875,
0.51123046875,
0.51123046875
],
[
0.12005615234375,
0.5126953125,
0.447509765625,
0.5126953125,
0.447509765625,
0.525390625,
0.12005615234375,
0.525390625
],
[
0.5146484375,
0.51318359375,
0.884765625,
0.51318359375,
0.884765625,
0.52685546875,
0.5146484375,
0.52685546875
],
[
0.51318359375,
0.529296875,
0.88330078125,
0.529296875,
0.88330078125,
0.54345703125,
0.51318359375,
0.54345703125
],
[
0.12005615234375,
0.5380859375,
0.48779296875,
0.5380859375,
0.48779296875,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.51318359375,
0.5458984375,
0.88330078125,
0.5458984375,
0.88330078125,
0.55859375,
0.51318359375,
0.55859375
],
[
0.1175537109375,
0.55419921875,
0.48779296875,
0.55419921875,
0.48779296875,
0.56787109375,
0.1175537109375,
0.56787109375
],
[
0.51318359375,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.118408203125,
0.57080078125,
0.491943359375,
0.57080078125,
0.491943359375,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.51318359375,
0.576171875,
0.88330078125,
0.57666015625,
0.88330078125,
0.59033203125,
0.51318359375,
0.5888671875
],
[
0.12005615234375,
0.58740234375,
0.490234375,
0.58740234375,
0.490234375,
0.60107421875,
0.12005615234375,
0.60107421875
],
[
0.51318359375,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.118408203125,
0.6025390625,
0.490234375,
0.6025390625,
0.490234375,
0.61767578125,
0.118408203125,
0.61767578125
],
[
0.51318359375,
0.60791015625,
0.884765625,
0.609375,
0.884765625,
0.62353515625,
0.51318359375,
0.62255859375
],
[
0.12005615234375,
0.6298828125,
0.313232421875,
0.6298828125,
0.313232421875,
0.6435546875,
0.12005615234375,
0.6435546875
],
[
0.5146484375,
0.62451171875,
0.88330078125,
0.62451171875,
0.88330078125,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.5146484375,
0.64111328125,
0.88330078125,
0.64111328125,
0.88330078125,
0.65478515625,
0.5146484375,
0.65478515625
],
[
0.118408203125,
0.65185546875,
0.490234375,
0.65185546875,
0.490234375,
0.6650390625,
0.118408203125,
0.6650390625
],
[
0.5146484375,
0.65771484375,
0.884765625,
0.65771484375,
0.884765625,
0.67138671875,
0.5146484375,
0.67138671875
],
[
0.12005615234375,
0.66845703125,
0.490234375,
0.66845703125,
0.490234375,
0.681640625,
0.12005615234375,
0.681640625
],
[
0.51318359375,
0.6728515625,
0.8369140625,
0.6748046875,
0.8369140625,
0.689453125,
0.51318359375,
0.68701171875
],
[
0.118408203125,
0.681640625,
0.489501953125,
0.68310546875,
0.489501953125,
0.69775390625,
0.118408203125,
0.69677734375
],
[
0.12005615234375,
0.69970703125,
0.489501953125,
0.69970703125,
0.489501953125,
0.71337890625,
0.12005615234375,
0.71337890625
],
[
0.12005615234375,
0.71630859375,
0.484375,
0.71630859375,
0.484375,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.58837890625,
0.71630859375,
0.6572265625,
0.71630859375,
0.6572265625,
0.73095703125,
0.58837890625,
0.73095703125
],
[
0.67333984375,
0.71435546875,
0.81005859375,
0.71435546875,
0.81005859375,
0.732421875,
0.67333984375,
0.732421875
],
[
0.85791015625,
0.71630859375,
0.884765625,
0.71630859375,
0.884765625,
0.72998046875,
0.85791015625,
0.72998046875
],
[
0.12176513671875,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74755859375,
0.12176513671875,
0.74755859375
],
[
0.6689453125,
0.73388671875,
0.6884765625,
0.73388671875,
0.6884765625,
0.7421875,
0.6689453125,
0.7421875
],
[
0.118408203125,
0.74755859375,
0.48779296875,
0.74755859375,
0.48779296875,
0.7626953125,
0.118408203125,
0.7626953125
],
[
0.5146484375,
0.75732421875,
0.8818359375,
0.75732421875,
0.8818359375,
0.77099609375,
0.5146484375,
0.77099609375
],
[
0.1175537109375,
0.76318359375,
0.490234375,
0.76416015625,
0.490234375,
0.779296875,
0.1175537109375,
0.77783203125
],
[
0.51318359375,
0.77392578125,
0.88330078125,
0.77392578125,
0.88330078125,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.118408203125,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.798828125,
0.118408203125,
0.798828125
],
[
0.51318359375,
0.7890625,
0.884765625,
0.7890625,
0.884765625,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.1175537109375,
0.79736328125,
0.235107421875,
0.79638671875,
0.235107421875,
0.81103515625,
0.1175537109375,
0.8125
],
[
0.51123046875,
0.8046875,
0.884765625,
0.80419921875,
0.884765625,
0.818359375,
0.51123046875,
0.8193359375
],
[
0.1368408203125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.1368408203125,
0.82763671875
],
[
0.51318359375,
0.82177734375,
0.7001953125,
0.82177734375,
0.7001953125,
0.8349609375,
0.51318359375,
0.8349609375
],
[
0.118408203125,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.52880859375,
0.83740234375,
0.88671875,
0.8359375,
0.88671875,
0.85107421875,
0.52880859375,
0.8515625
],
[
0.118408203125,
0.8466796875,
0.490234375,
0.845703125,
0.490234375,
0.85986328125,
0.118408203125,
0.86083984375
],
[
0.51318359375,
0.85400390625,
0.88671875,
0.85400390625,
0.88671875,
0.86767578125,
0.51318359375,
0.86767578125
],
[
0.118408203125,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.87060546875,
0.88330078125,
0.87060546875,
0.88330078125,
0.88427734375,
0.51318359375,
0.88427734375
],
[
0.1175537109375,
0.87890625,
0.490234375,
0.87744140625,
0.490234375,
0.892578125,
0.1175537109375,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.53564453125,
0.8955078125,
0.79443359375,
0.8955078125,
0.79443359375,
0.9091796875,
0.53564453125,
0.9091796875
]
] |
[
"hence negligible contribution to the gradient esti-",
"I.e., a uTOD system neither knows which query",
"mation. To ensure queries in the buffer contribute",
"was fired to K B, nor knows when it was fired. In",
"significantly to the gradient estimates, MAPO clips",
"thisworkweassumethatthesystemcanmakeonly",
"TB to Q. The modified gradient estimate is,",
" one KB query per dialog. Our overall uTOD base-",
"line uses a pipeline of multiple components - a",
"VeOER = Ea~(a) Vologe(a)R(a)",
"query position predictor, a query predictor and a",
"+(1 -)Ea~ (a) Vloge(a)R(a) (3)",
"next response predictor.",
"3.2KB QueryPredictor",
"where g = max(B,Q). When the training",
"begins, α > T resulting in gradient estimates",
"We first assume that we have access to the specific",
"biased towards the queries in the buffer. Once the",
"turn number q,1 ≤ q ≤ m, where a KB query",
"policy stabilizes, π gets larger than Q and the",
"gets fired by the system (an assumption we relax",
"estimates becomes unbiased.",
"in the next section). Thus, formally, we are given",
"Experience Replay: Experience replay (Lin,",
"the dialog context c = [ci, ci, c, c,..., cu], and",
"1992) stores past experience in a buffer and reuses",
"K B, and the goal of query predictor is to output a",
"that to stabilize training and improve sample effi-",
"that appropriately captures the user intent. We use",
"ciency. Prioritized experience replay (Schaul et al.,",
"the term subsequent dialog to refer to all utterances",
"2016) assigns priorities to the experiences and sam-",
"that follow the KB query - {cg, cu+1, .., cm]. In",
"ples them based on the priorities to efficiently learn",
"Figure 1, the subsequent dialog starts from the sec-",
"using them. MAPO and our approach use replay",
"ond agent utterance. Let Es be the set of KB enti-",
"buffers which store non-zero reward queries. These",
"ties present in the subsequent dialog. In Figure 1.",
"queries include both past experiences and queries",
"Es = {Peking Restaurant, 2343-4040]. We train",
"identified using systematic search. The sampling",
"the query predictor using RL, by providing feed-",
"strategy is similar to prioritized experience replay,",
"back based on the predicted query's ability to re-",
"where the priorities are defined by the probability",
"trieve the entities in Es",
"of the query in the current policy.",
"Our proposed baseline follows the literature",
" on semantic parsing with weak supervision using",
"3TaskDefinition &Baseline System",
"RL, and treats the query prediction as equivalent",
"to learning policy e(ac) that takes the dialog",
"We first define our novel task of learning TOD",
"context c as the input and generates a KB query",
"system using unannotated dialogs -- we name it",
"a = aia2 .. . aT. The KB query is a sequence of",
"uTOD. We then describe the KB query predictor in",
"T actions. where each action is a word predicted",
"detail. Finally, we describe our proposed baseline",
"by the query predictor. For a given context, the",
"uTOD system, which uses the KB query predictor.",
"set of all possible actions is a union of a set of",
"3.1Problem Definition",
"keywords? in the SQL query language, set of field",
"names in K B, the set of all words in the context c",
"We represent a dialog d between a user u and",
"and an <eoq> token to indicate the end of query.",
"an agent s as [c',ci,c,c,...,cm,cm] where",
"The query is generated auto-regressively as:",
"m denotes the number of turns in the dialog.",
"Our goal is to train a uTOD system, which,",
"for all turns i. takes the partial dialog-so-fau",
"e(ac) =",
"I 元e(at a1:t-1,c)",
"(4)",
"[c, ci,..., cu-i, ci-i, c\" as input and predicts",
"t=",
"the next system response cs. Such a uTOD system",
"As the environment is deterministic, at each time",
"is trained by the training data D comprising com-",
"t, the RL state can be fully defined by the dialog",
"plete dialogs [d;]j-1, and an associated knowl-",
"context c and the actions predicted so far ai:t-1.",
"edge base K B.",
"RL gradients can be computed using REINFORCE",
"System responses in a dialog often use certain",
"or the MAPO algorithm.",
"entities from KB. To accomplish this, a uTOD",
"The policy network is implemented with a stan-",
"system can fire an appropriate query a to K B and",
"dard encoder-decoder architecture for TOD sys-",
"fetch the query results E, and use those to gener-",
"tems. The context is encoded using a multi-hop",
"ate its response. No explicit supervision is provided",
"on the gold queries, at either training or test time.",
"2 {SELECT, *, =, FROM, AND, WHERE)"
] |
{
"arxiv_id": "2005.00123",
"language": "en",
"page_idx": 3
}
|
||
2005.00123-en-4
|
2005.00123-en-4.jpg
|
memory encoder (Sukhbaatar et al., 2015) with a bag of sequences memory (Raghu et al., 2019). In the bag of sequences memory, the context is represented as a set of utterances and each utterance as a sequence of words. Each utterance is encoded using a bi-directional GRU (Cho et al., 2014).
The KB query is generated one word at a time by a copy-augmented sequence decoder (Gu et al., 2016). The search space is the output space of this variable length, copy-augmented sequence decoder. At each time step, the decoder computes a copy distribution over words in the dialog context and a generate distribution over words KB feld names and SQL keywords. Finally, a soft gate (See et al., 2017) is used to combine the two distributions and a word is sampled from the combined distribution. By allowing only copying words from the dialog context to the values in the SQL WHERE clauses, we reduce the decoder's ability to predict spurious queries. Following Ghazvininejad et al.(2016), we use beam search guided by the SQL grammar to generate only syntactically correct queries.
## 3.2.1 Reward Functions
We now describe a vanilla reward function considered for training the RL agent. For a given $(c,E^{s})$ pair, the query a is predicted using $c$ and the reward is computed using $E^\mathrm{s}.$ Since partial query isn't meaningful, the reward is computed only at time $t=T$, when the complete query is generated.
Let $E^a$ be the set of entities retrieved by the query a, then the reward at time $T$ must ensure that all entities present in the subsequent dialog are retrieved by the query. Moreover, the query should be penalized for retrieving more than required entities. This can be operationalized as follows:
(5)
$$R(\mathbf{a}|c,E^s)=\mathbb{1}_{\text{recall}(E^s,E^a)=1}.\text{prec}(E^s,E^a)$$
The recall based indicator ensures that all entities are retrieved, and precision penalizes retrieval of a large number of entities. While a reasonable first solution, our initial experiments showed that MAPO with this reward function does not achieve satisfactory results. We now discuss the challenging aspects of our task, and our improved baseline, multi buffer MAPO, to resolve those issues.
3.2.2 Multi Buffer MAPO (mB-MAPO)
Our problem can become particularly challenging if $KB$ has correlated attributes. To understand this, consider two types of KB queries: (1) partial intent query and (2) complete intent query. A partial
query is one that is partially correct, i.e., captures a part of the user's intent, and does not include any incorrect clauses. A complete query contains all (and only) correct clauses expressed in the intent. For example (a) in Table 1, the first two queries are complete ones, whereas rest are partial. Query 4 in column (b) is a spurious query that can retrieve the same result as the complete intent query. This spurious query is not a partial query, as it does not capture any of the user needs. Our query decoder allows only words to be copied from the context when predicting the query, thus reducing the policy network's ability to learn spurious queries.
Because of the nature of weak supervision in uTOD (we get to see only a subset of entities in subsequent dialog), it is possible that multiple queries can achieve non-zero rewards for a dialog context. It is further possible that a specific partial query can achieve non-zero rewards in many dialog contexts. Example, the query “price=moderate" fetches nonzero rewards for both intents in Table 1. Such phenomena can confuse an RL agent, to the extent that it may end up incorrectly learning a single partial query as the best query for many contexts.
This problem is further exacerbated if some query attributes in $KB$ are correlated. For example (a) in Table 1, let us assume that $KB$ has 8 chinese restaurants out of which 7 have moderate price range, i.e., cuisine and price range are highly correlated. Here, results of the partial query “cuisine=chinese" would contain only one additional restaurant compared to the complete query. As a result, they will receive almost the same reward during training. This could potentially get extreme in certain cases, where presence or absence of an attribute makes no difference, giving little signal to an RL agent. As the number of attributes in an intent increases, the problem can get harder and harder. In our preliminary experiments using MAPO with reward from Equation 5, the RL agent often produced partial queries leading to further errors in subsequent dialog.
In our datasets intents can be described by SELECT queries with multiple WHERE clauses; thus, every partial query is a prefix of some complete query (e.g,query 3 in Table 1). Upon further analysis of model behavior, we observed that MAPO often maintained both partial prefix queries and complete queries, but still learned a model to output the partial ones.
To understand this surprising observation, we
|
[
[
0.12005615234375,
0.07598876953125,
0.491943359375,
0.07598876953125,
0.491943359375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.5146484375,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.118408203125,
0.0926513671875,
0.490234375,
0.0914306640625,
0.490234375,
0.105712890625,
0.118408203125,
0.10687255859375
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.118408203125,
0.1092529296875,
0.491943359375,
0.1092529296875,
0.491943359375,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.51318359375,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.51318359375,
0.122314453125
],
[
0.118408203125,
0.12469482421875,
0.491943359375,
0.12469482421875,
0.491943359375,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.51318359375,
0.12469482421875,
0.88671875,
0.12469482421875,
0.88671875,
0.1383056640625,
0.51318359375,
0.1383056640625
],
[
0.1175537109375,
0.14013671875,
0.490234375,
0.1395263671875,
0.490234375,
0.15380859375,
0.1175537109375,
0.155029296875
],
[
0.51318359375,
0.14013671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.1175537109375,
0.1561279296875,
0.456787109375,
0.1561279296875,
0.456787109375,
0.169189453125,
0.1175537109375,
0.169189453125
],
[
0.51318359375,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.1351318359375,
0.171630859375,
0.49365234375,
0.1728515625,
0.49365234375,
0.18701171875,
0.1351318359375,
0.1859130859375
],
[
0.51318359375,
0.17333984375,
0.884765625,
0.17333984375,
0.884765625,
0.18701171875,
0.51318359375,
0.18701171875
],
[
0.118408203125,
0.189453125,
0.495361328125,
0.189453125,
0.495361328125,
0.203125,
0.118408203125,
0.203125
],
[
0.51318359375,
0.189453125,
0.88671875,
0.189453125,
0.88671875,
0.203125,
0.51318359375,
0.203125
],
[
0.118408203125,
0.204833984375,
0.49365234375,
0.204833984375,
0.49365234375,
0.218505859375,
0.118408203125,
0.218505859375
],
[
0.51318359375,
0.2060546875,
0.88671875,
0.2060546875,
0.88671875,
0.2197265625,
0.51318359375,
0.2197265625
],
[
0.118408203125,
0.221435546875,
0.49365234375,
0.221435546875,
0.49365234375,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.51318359375,
0.221435546875,
0.88671875,
0.221435546875,
0.88671875,
0.235107421875,
0.51318359375,
0.235107421875
],
[
0.118408203125,
0.2369384765625,
0.490234375,
0.2369384765625,
0.490234375,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.51123046875,
0.2369384765625,
0.88671875,
0.236328125,
0.88671875,
0.25048828125,
0.51123046875,
0.251708984375
],
[
0.118408203125,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.51318359375,
0.254150390625,
0.88330078125,
0.254150390625,
0.88330078125,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.12005615234375,
0.26953125,
0.489501953125,
0.26953125,
0.489501953125,
0.283203125,
0.12005615234375,
0.283203125
],
[
0.51123046875,
0.268310546875,
0.8310546875,
0.26953125,
0.8310546875,
0.283935546875,
0.51123046875,
0.283203125
],
[
0.12005615234375,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.52880859375,
0.287353515625,
0.88330078125,
0.2880859375,
0.88330078125,
0.302734375,
0.52880859375,
0.3017578125
],
[
0.118408203125,
0.300537109375,
0.48779296875,
0.300537109375,
0.48779296875,
0.314208984375,
0.118408203125,
0.314208984375
],
[
0.5146484375,
0.3046875,
0.88671875,
0.3046875,
0.88671875,
0.319580078125,
0.5146484375,
0.319580078125
],
[
0.12005615234375,
0.317626953125,
0.490234375,
0.317626953125,
0.490234375,
0.330810546875,
0.12005615234375,
0.330810546875
],
[
0.51318359375,
0.32177734375,
0.88671875,
0.32177734375,
0.88671875,
0.3349609375,
0.51318359375,
0.3349609375
],
[
0.12005615234375,
0.334228515625,
0.490234375,
0.334228515625,
0.490234375,
0.347412109375,
0.12005615234375,
0.347412109375
],
[
0.51318359375,
0.33740234375,
0.88671875,
0.33740234375,
0.88671875,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.118408203125,
0.349853515625,
0.491943359375,
0.349853515625,
0.491943359375,
0.363525390625,
0.118408203125,
0.363525390625
],
[
0.51123046875,
0.352783203125,
0.88330078125,
0.352783203125,
0.88330078125,
0.366455078125,
0.51123046875,
0.366455078125
],
[
0.12005615234375,
0.366455078125,
0.490234375,
0.366455078125,
0.490234375,
0.380126953125,
0.12005615234375,
0.380126953125
],
[
0.5146484375,
0.369384765625,
0.88671875,
0.369384765625,
0.88671875,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.1175537109375,
0.3818359375,
0.491943359375,
0.3818359375,
0.491943359375,
0.3955078125,
0.1175537109375,
0.3955078125
],
[
0.51318359375,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.51318359375,
0.39794921875
],
[
0.118408203125,
0.3984375,
0.490234375,
0.3984375,
0.490234375,
0.412109375,
0.118408203125,
0.412109375
],
[
0.5146484375,
0.402099609375,
0.88330078125,
0.402099609375,
0.88330078125,
0.4150390625,
0.5146484375,
0.4150390625
],
[
0.1175537109375,
0.4150390625,
0.455078125,
0.4150390625,
0.455078125,
0.4287109375,
0.1175537109375,
0.4287109375
],
[
0.5146484375,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.5146484375,
0.43115234375
],
[
0.118408203125,
0.43701171875,
0.32080078125,
0.438232421875,
0.32080078125,
0.453125,
0.118408203125,
0.451904296875
],
[
0.5146484375,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.51318359375,
0.45068359375,
0.8251953125,
0.45068359375,
0.8251953125,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.1175537109375,
0.457275390625,
0.490234375,
0.4560546875,
0.490234375,
0.47021484375,
0.1175537109375,
0.471435546875
],
[
0.1175537109375,
0.473876953125,
0.489501953125,
0.47265625,
0.489501953125,
0.48681640625,
0.1175537109375,
0.488037109375
],
[
0.53076171875,
0.4697265625,
0.880859375,
0.4697265625,
0.880859375,
0.482666015625,
0.53076171875,
0.482666015625
],
[
0.12005615234375,
0.490478515625,
0.490234375,
0.490478515625,
0.490234375,
0.50439453125,
0.12005615234375,
0.50439453125
],
[
0.51318359375,
0.485107421875,
0.880859375,
0.48388671875,
0.880859375,
0.498779296875,
0.51318359375,
0.49951171875
],
[
0.118408203125,
0.505859375,
0.489501953125,
0.505859375,
0.489501953125,
0.51953125,
0.118408203125,
0.51953125
],
[
0.5146484375,
0.501953125,
0.8818359375,
0.501953125,
0.8818359375,
0.515625,
0.5146484375,
0.515625
],
[
0.118408203125,
0.5224609375,
0.490234375,
0.5224609375,
0.490234375,
0.5361328125,
0.118408203125,
0.5361328125
],
[
0.51318359375,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.1175537109375,
0.53759765625,
0.49365234375,
0.5380859375,
0.49365234375,
0.552734375,
0.1175537109375,
0.5517578125
],
[
0.5146484375,
0.53369140625,
0.8818359375,
0.53369140625,
0.8818359375,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.1351318359375,
0.5546875,
0.491943359375,
0.5546875,
0.491943359375,
0.568359375,
0.1351318359375,
0.568359375
],
[
0.5146484375,
0.5517578125,
0.880859375,
0.5517578125,
0.880859375,
0.5625,
0.5146484375,
0.5625
],
[
0.1175537109375,
0.57080078125,
0.490234375,
0.5693359375,
0.490234375,
0.58447265625,
0.1175537109375,
0.5849609375
],
[
0.51318359375,
0.56640625,
0.88330078125,
0.56640625,
0.88330078125,
0.580078125,
0.51318359375,
0.580078125
],
[
0.12005615234375,
0.58740234375,
0.48779296875,
0.58740234375,
0.48779296875,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.51318359375,
0.5830078125,
0.88330078125,
0.5830078125,
0.88330078125,
0.5966796875,
0.51318359375,
0.5966796875
],
[
0.118408203125,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.118408203125,
0.615234375
],
[
0.51123046875,
0.59716796875,
0.8818359375,
0.59716796875,
0.8818359375,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.51123046875,
0.61328125,
0.88330078125,
0.6142578125,
0.88330078125,
0.62890625,
0.51123046875,
0.62744140625
],
[
0.12176513671875,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63134765625,
0.12176513671875,
0.63134765625
],
[
0.1175537109375,
0.6328125,
0.41064453125,
0.63427734375,
0.41064453125,
0.6484375,
0.1175537109375,
0.64794921875
],
[
0.51318359375,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.51318359375,
0.64453125
],
[
0.51123046875,
0.6455078125,
0.88330078125,
0.64794921875,
0.88330078125,
0.6611328125,
0.51123046875,
0.6591796875
],
[
0.1292724609375,
0.66015625,
0.491943359375,
0.66015625,
0.491943359375,
0.68115234375,
0.1292724609375,
0.68115234375
],
[
0.51318359375,
0.6630859375,
0.884765625,
0.6630859375,
0.884765625,
0.6767578125,
0.51318359375,
0.6767578125
],
[
0.51318359375,
0.6787109375,
0.884765625,
0.6787109375,
0.884765625,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.1368408203125,
0.689453125,
0.490234375,
0.689453125,
0.490234375,
0.703125,
0.1368408203125,
0.703125
],
[
0.5146484375,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.708984375,
0.5146484375,
0.708984375
],
[
0.118408203125,
0.70703125,
0.490234375,
0.70703125,
0.490234375,
0.72021484375,
0.118408203125,
0.72021484375
],
[
0.51318359375,
0.71142578125,
0.88330078125,
0.71142578125,
0.88330078125,
0.724609375,
0.51318359375,
0.724609375
],
[
0.118408203125,
0.72265625,
0.490234375,
0.72265625,
0.490234375,
0.736328125,
0.118408203125,
0.736328125
],
[
0.51318359375,
0.72705078125,
0.88330078125,
0.72705078125,
0.88330078125,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.12005615234375,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.7529296875,
0.12005615234375,
0.7529296875
],
[
0.51123046875,
0.7412109375,
0.71875,
0.74365234375,
0.71875,
0.75830078125,
0.51123046875,
0.755859375
],
[
0.1175537109375,
0.75439453125,
0.489501953125,
0.75439453125,
0.489501953125,
0.76708984375,
0.1175537109375,
0.76708984375
],
[
0.5283203125,
0.76123046875,
0.884765625,
0.76123046875,
0.884765625,
0.77490234375,
0.5283203125,
0.77490234375
],
[
0.118408203125,
0.77099609375,
0.490234375,
0.771484375,
0.490234375,
0.78515625,
0.118408203125,
0.78369140625
],
[
0.51318359375,
0.77783203125,
0.88671875,
0.77783203125,
0.88671875,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.118408203125,
0.78759765625,
0.491943359375,
0.78759765625,
0.491943359375,
0.80029296875,
0.118408203125,
0.80029296875
],
[
0.51318359375,
0.7958984375,
0.884765625,
0.7958984375,
0.884765625,
0.80859375,
0.51318359375,
0.80859375
],
[
0.118408203125,
0.802734375,
0.443359375,
0.802734375,
0.443359375,
0.81640625,
0.118408203125,
0.81640625
],
[
0.51123046875,
0.81103515625,
0.88330078125,
0.81005859375,
0.88330078125,
0.82470703125,
0.51123046875,
0.826171875
],
[
0.118408203125,
0.82666015625,
0.441650390625,
0.82666015625,
0.441650390625,
0.84033203125,
0.118408203125,
0.84033203125
],
[
0.51123046875,
0.82666015625,
0.88330078125,
0.826171875,
0.88330078125,
0.84033203125,
0.51123046875,
0.84130859375
],
[
0.12005615234375,
0.8466796875,
0.490234375,
0.8466796875,
0.490234375,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.51123046875,
0.84326171875,
0.8818359375,
0.84326171875,
0.8818359375,
0.85693359375,
0.51123046875,
0.85693359375
],
[
0.12005615234375,
0.8642578125,
0.491943359375,
0.8642578125,
0.491943359375,
0.87451171875,
0.12005615234375,
0.87451171875
],
[
0.51318359375,
0.85986328125,
0.884765625,
0.85986328125,
0.884765625,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.118408203125,
0.87890625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.51318359375,
0.8759765625,
0.66015625,
0.8759765625,
0.66015625,
0.8896484375,
0.51318359375,
0.8896484375
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.53076171875,
0.8935546875,
0.884765625,
0.89453125,
0.884765625,
0.9091796875,
0.53076171875,
0.90771484375
]
] |
[
"memory encoder (Sukhbaatar et al., 2015) with a",
"query is one that is partially correct, i.e., captures a",
"bag of sequences memory (Raghu et al., 2019). In",
"part of the user's intent, and does not include any",
"the bag of sequences memory, the context is rep-",
"incorrect clauses. A complete query contains all",
"resented as a set of utterances and each utterance",
"(and only) correct clauses expressed in the intent.",
"as a sequence of words. Each utterance is encoded",
"For example (a) in Table 1, the first two queries are",
"using a bi-directional GRU (Cho et al., 2014).",
"complete ones, whereas rest are partial. Query 4",
"The KB query is generated one word at a time",
"in column (b) is a spurious query that can retrieve",
"by a copy-augmented sequence decoder (Gu et al.,",
"the same result as the complete intent query. This",
"2016). The search space is the output space of this",
"spurious query is not a partial query, as it does not",
"variable length, copy-augmented sequence decoder.",
"capture any of the user needs. Our query decoder",
"At each time step, the decoder computes a copy",
"allows only words to be copied from the context",
"distribution over words in the dialog context and",
"when predicting the query, thus reducing the policy",
"a generate distribution over words KB field names",
"network's ability to learn spurious queries.",
"and SOL keywords. Finally, a soft gate (See et al..",
"Because of the nature of weak supervision in",
"2017) is used to combine the two distributions and",
"uTOD (we get to see only a subset of entities in sub-",
"a word is sampled from the combined distribution.",
"sequent dialog), it is possible that multiple queries",
"By allowing only copying words from the dialog",
"can achieve non-zero rewards for a dialog context.",
"context to the values in the SQL WHERE clauses,",
"It is further possible that a specific partial query can",
"we reduce the decoder's ability to predict spurious",
"achieve non-zero rewards in many dialog contexts.",
"queries. Following Ghazvininejad et al. (2016),",
"Example,the query“price=moderate\"fetches non",
"we use beam search guided by the SQL grammar",
"zero rewards for both intents in Table 1. Such phe-",
"to generate only syntactically correct queries.",
"nomena can confuse an RL agent, to the extent that",
"3.2.1 RewardFunctions",
"it may end up incorrectly learning a single partial",
"query as the best query for many contexts.",
"We now describe a vanilla reward function consid-",
"ered for training the RL agent. For a given (c, Es)",
"This problem is further exacerbated if some",
"pair, the query a is predicted using c and the re-",
"query attributes in KB are correlated.For exam",
"ward is computed using Es. Since partial query",
"ple (a) in Table 1, let us assume that KB has 8",
"isn't meaningful, the reward is computed only at",
"chinese restaurants out of which 7 have moderate",
"time t = T, when the complete query is generated.",
"price range, i.e., cuisine and price range are highly",
"Let Ea be the set of entities retrieved by the",
"correlated.Here,results of thepartial query\"cui",
"query a, then the reward at time T must ensure that",
"sine=chinese\" would contain only one additional",
"all entities present in the subseguent dialog are re",
"restaurant compared to the complete query. As a",
"trieved by the query. Moreover, the query should be",
"result, they will receive almost the same reward",
"during training. This could potentially get extreme",
"penalizedforretrievingmorethanrequired entities.",
"This can be operationalized as follows:",
"in certain cases, where presence or absence of an",
"attribute makes no difference,giving little signal",
"R(alc, E°) = Irecll(Es,Ea)=1-prec(Es, E) (5)",
"to an RL agent. As the number of attributes in",
"an intent increases, the problem can get harder",
"The recall based indicator ensures that all enti-",
"and harder. In our preliminary experiments using",
"ties are retrieved, and precision penalizes retrieval",
"MAPO with reward from Equation 5, the RL agent",
"of a large number of entities. While a reasonable",
"often produced partial queries leading to further",
"first solution, our initial experiments showed that",
"errors in subsequent dialog.",
"MAPO with this reward function does not achieve",
"In our datasets intents can be described by SE-",
"satisfactory results.We now discuss the challeng-",
"LECT queries with multiple WHERE clauses; thus,",
"ing aspects of our task, and our improved baseline,",
"every partial query is a prefix of some complete",
"multi buffer MAPO, to resolve those issues.",
"query (e.g, query 3 in Table 1). Upon further anal",
"3.2.2 Multi Buffer MAPO (mB-MAPO)",
"ysis of model behavior, we observed that MAPO",
"Our problem can become particularly challenging",
" often maintained both partial prefix queries and",
"if KB has correlated attributes.To understand this,",
"complete queries, but still learned a model to out-",
"consider two types of KB queries: (1) partial intent",
"put the partial ones.",
"query and (2) complete intent query. A partial",
"To understand this surprising observation, we"
] |
{
"arxiv_id": "2005.00123",
"language": "en",
"page_idx": 4
}
|
||
2101.06351-en-2
|
2101.06351-en-2.jpg
|
$\boxed{\text{Connection with Prior Work}}$
$\overline{\text{Strategy}\:\textbf{Definition and Examples}}$
<table>
<tbody>
<tr>
<th>Commitment</th>
<th>The persuaders indicating their intentions to take acts or justify their earlier decisions to convince others that they have made the correct choice. e.g., I just lent to Auntie Fine's Donut Shop. (Kiva)</th>
<th>$[ LBOS] Commitment$ (Yang et al. 2019)$, $Self-modeling~($Wang et al. 2019), Commitment (Vargheese et al. 2020b</th>
</tr>
<tr>
<td>Emotion</td>
<td>Making request full of emotional valence and arousal affect to influence others. $I've~been~in~the~lowest~depressive~state~of~my~life.~(RAOP)$ e.g., Guys $I’m$ desperate. $(Borrow)$</td>
<td>$Ethos\left(\mathrm{Carlıle~et~al.~2018}\right)$, $Emotion~appeal~($Carlile et al. 2018), Sentiment (Durmus et al. 2018), $Emotion~words~($Luu et al, 2019a), $Emotion\left(\mathrm{Asai~et~al.~2020}\right)$</td>
</tr>
<tr>
<td>Politeness</td>
<td>The usage of polite language in requests. e.g., Your help is deeply appreciated! (Borrow)</td>
<td>$$Politeness($Durmus et al. 2018)$, $$Politeness($Althoff et al. 2014)$, $Politeness($Nashruddin et al. 2020)</td>
</tr>
<tr>
<td>Reciprocity</td>
<td>Responding to a positive action with another positive action. People are more likely to help if they have received help themselves. e.g., I will pay 5% interest no later than May I, 2016. (Borrow) $I'll~pay~it~forward~with~my~frst~check.~(RAOP)$</td>
<td>$[ LBOS] Reciprocity$ (Althoff et al. 2014)$, Reciprocity (Roethke et al. 2020), $Reciprocity($ (Vargheese et al.2020b)</td>
</tr>
<tr>
<td>Scarcity</td>
<td>People emphasizing on the urgency, rare of their needs. e.g., Need this loan urgently. (Borrow) $I\textit{haven' t ate a meal in two days. }( RAOP)$ Loan expiring todav and still needs 8650. (Kiva)</td>
<td>$[ LBOS] Scarcity$ (Vargheese et al. 2020b)$, Scarcity (Yang et al. 2019), $Scarcity\left(\mathrm{Lawson~et~al.~2020}\right)$</td>
</tr>
<tr>
<td>Credibility</td>
<td>The uses of credentials impacts to establish credibility and earn others' trust. e.g., Can provide any documentation needed. (Borrow) She has already repaid 2 previous $loans.$ $( Kiva)$</td>
<td>$Credibility~appeal~($Wang et al. 2019) $Social~Proof~($Roethke et al. 2020), Social Proof ( Vargheese et al. 2020a)</td>
</tr>
<tr>
<td>$\operatorname{Evidence}$</td>
<td>Providing concrete facts or evidence for the narrative or request. e.g. My insurance was canceled today. (Borrow) There is a Pizza Hut and a Dominos near me. (RAOP) $\$225$ to go and I A+ member on the loan. (Kiva)</td>
<td>$Evidentiality~({\mathrm{Althoff~et~al.~2014}})$, $[ LBOS] Evidence$ (Carlile et al. 2018)$, Evidence (Stab and Gurevych 2014), $Concreteness($Yang et al. 2019) $Evidence($Durmus et al. 2018)</td>
</tr>
<tr>
<td>Impact</td>
<td>Emphasizing the importance or impact of the request. e.g., I will use this loan to pay my rent. ( Borrow)</td>
<td>Logos (Carlile et al. 2018), $Logic~appeal~($Wang et al. 2019) lmnact (Yang et al 2019)</td>
</tr>
</tbody>
</table>
Table 1: The generic taxonomy of persuasive strategies, their definitions, example sentences, and connections with prior work.
than 8 sentences, and 85% posts in Borrow has at most 7 sentences. We recruited four research assistants to label persuasion strategies for each sentence in sampled documents. Definitions and examples of different persuasion strategies were provided, together with a training session where we asked annotators to annotate a number of example sentences and walked them through any disagreed annotations. To assess the reliability of the annotated labels, we then asked them to annotate the same 100 documents with 400 sentences and computed Cohen's Kappa coefficient to measure inter-rater reliability. We obtained an average score of 0.538 on Kiva, 0.613 on RAOP, and 0.623 on Borrow, which indicates moderate agreement (McHugh 2012). Annotators then annotated the rest 1200 documents by themselves independently.
The dataset statistics are shown in Table 2, and the sentence-level label distribution in each dataset is shown in Figure 1. We merge rare strategies into the Other category. Specifically, we merge Commitment, Scarcity, and Emotion in Borrow, Credibility and Commitment in RAOP, Reciprocity and Emotion in Kiva, as Other. We utilized whether the requester received pizzas or loans from the subreddits as the document-level labels for RAOP and Borrow. 30.1% of people successfully got pizzas on RAOP and 48.5% of people received loans on Borrow. In Kiva, we utilized the number of people who lent loans as the document-level labels. The numbers are further labeled based on buckets: [1,2),
$[2,3),[3,4),[4,\infty)$, accounting for $44.1\%,20.3\%,12.4\%$
and 33.2% of all documents.
# $\mathbf{Method}$
To alleviate the dependencies on labeled data, we propose a hierarchical weakly-supervised latent variable model to leverage partially labeled data to predict sentence-level persuasive strategies. Specifically, we introduce a sentencelevel latent variable model to reconstruct the input sentence and predict the sentence-level persuasion labels spontaneously, supervised by the global or document-level labels (e.g., overall persuasiveness of the documents). The overall architecture of our method is shown in Figure 2.
## Weakly Supervised Latent Model
Given a corpus of $N$ documents $\mathbf{D}=\{\mathbf{d}_{i}\}_{i=1}^{N}$,where each documents $\mathbf{d}_{i}=\{\mathbf{s}_{i}^{j}\}_{j=1}^{M}.$ For each document $\mathbf{d}_i\in\mathbf{D}$, its document level label is denoted as t$_i$, representing the overall persuasiveness of the documents. We divide the corpus into two parts: D= D $\cup\mathbf{D}_U$, where $\mathbf{D}_L\left(\mathbf{D}_U\right)$denotes the set of documents with (with- $\begin{array}{l}{\mathrm{out)~sentence~labels.~For~each~document~d}_{i}\in\mathbf{D}_{L},\mathrm{~the~cor-}}\\{\mathrm{responding~sentence~labels~are~\{y_{i}^{j}\}_{j=1}^{M},~where~y_{i}^{j}\in~C=}}\\{\mathrm{responding~sentence~labels~are~\{y_{i}^{j}\}_{j=1}^{M},~where~y_{i}^{j}\in~C=}}\end{array}$ $\{c_{k}\}_{k=1}^{K}$ and represents the persuasive strategy of a given sentence. In many practical scenarios, getting documentlevel labels $\{\mathbf{t}_i\}$ is much easier and cheaper than the fine-
|
[
[
0.1192626953125,
0.06756591796875,
0.3505859375,
0.06756591796875,
0.3505859375,
0.080810546875,
0.1192626953125,
0.080810546875
],
[
0.67626953125,
0.06817626953125,
0.85693359375,
0.06817626953125,
0.85693359375,
0.07891845703125,
0.67626953125,
0.07891845703125
],
[
0.191162109375,
0.080810546875,
0.63818359375,
0.08209228515625,
0.63818359375,
0.0953369140625,
0.191162109375,
0.0947265625
],
[
0.67626953125,
0.08209228515625,
0.865234375,
0.080810546875,
0.865234375,
0.0947265625,
0.67626953125,
0.0953369140625
],
[
0.096435546875,
0.0947265625,
0.17724609375,
0.0953369140625,
0.17724609375,
0.106689453125,
0.096435546875,
0.1060791015625
],
[
0.193603515625,
0.09661865234375,
0.59228515625,
0.09661865234375,
0.59228515625,
0.106689453125,
0.193603515625,
0.106689453125
],
[
0.67724609375,
0.09661865234375,
0.875,
0.09661865234375,
0.875,
0.106689453125,
0.67724609375,
0.106689453125
],
[
0.19287109375,
0.10919189453125,
0.494384765625,
0.10919189453125,
0.494384765625,
0.11932373046875,
0.19287109375,
0.11932373046875
],
[
0.67724609375,
0.10919189453125,
0.90185546875,
0.10919189453125,
0.90185546875,
0.11932373046875,
0.67724609375,
0.11932373046875
],
[
0.6748046875,
0.12060546875,
0.83642578125,
0.12060546875,
0.83642578125,
0.1337890625,
0.6748046875,
0.1337890625
],
[
0.19287109375,
0.1331787109375,
0.65869140625,
0.1331787109375,
0.65869140625,
0.146484375,
0.19287109375,
0.146484375
],
[
0.67626953125,
0.1331787109375,
0.89453125,
0.1331787109375,
0.89453125,
0.146484375,
0.67626953125,
0.146484375
],
[
0.12335205078125,
0.146484375,
0.17724609375,
0.146484375,
0.17724609375,
0.1578369140625,
0.12335205078125,
0.1578369140625
],
[
0.193603515625,
0.146484375,
0.398681640625,
0.146484375,
0.398681640625,
0.1572265625,
0.193603515625,
0.1572265625
],
[
0.67724609375,
0.1458740234375,
0.865234375,
0.1458740234375,
0.865234375,
0.1590576171875,
0.67724609375,
0.1590576171875
],
[
0.19287109375,
0.1590576171875,
0.53515625,
0.1590576171875,
0.53515625,
0.17236328125,
0.19287109375,
0.17236328125
],
[
0.67626953125,
0.1572265625,
0.8798828125,
0.1578369140625,
0.8798828125,
0.17236328125,
0.67626953125,
0.1717529296875
],
[
0.67626953125,
0.17041015625,
0.833984375,
0.169189453125,
0.833984375,
0.18310546875,
0.67626953125,
0.184326171875
],
[
0.67626953125,
0.1817626953125,
0.86767578125,
0.1817626953125,
0.86767578125,
0.1962890625,
0.67626953125,
0.1962890625
],
[
0.191162109375,
0.1881103515625,
0.435546875,
0.189453125,
0.435546875,
0.203857421875,
0.191162109375,
0.20263671875
],
[
0.11358642578125,
0.1956787109375,
0.176513671875,
0.1976318359375,
0.176513671875,
0.208984375,
0.11358642578125,
0.20703125
],
[
0.67626953125,
0.1962890625,
0.86279296875,
0.1962890625,
0.86279296875,
0.210205078125,
0.67626953125,
0.210205078125
],
[
0.193603515625,
0.2052001953125,
0.4697265625,
0.2052001953125,
0.4697265625,
0.21533203125,
0.193603515625,
0.21533203125
],
[
0.67626953125,
0.208984375,
0.88623046875,
0.208984375,
0.88623046875,
0.222900390625,
0.67626953125,
0.222900390625
],
[
0.191162109375,
0.222900390625,
0.650390625,
0.222900390625,
0.650390625,
0.236083984375,
0.191162109375,
0.236083984375
],
[
0.191162109375,
0.23486328125,
0.498291015625,
0.23486328125,
0.498291015625,
0.248779296875,
0.191162109375,
0.248779296875
],
[
0.67724609375,
0.2291259765625,
0.87109375,
0.2291259765625,
0.87109375,
0.242431640625,
0.67724609375,
0.242431640625
],
[
0.1070556640625,
0.242431640625,
0.176513671875,
0.242431640625,
0.176513671875,
0.25390625,
0.1070556640625,
0.25390625
],
[
0.67724609375,
0.2412109375,
0.875,
0.2412109375,
0.875,
0.255126953125,
0.67724609375,
0.255126953125
],
[
0.191162109375,
0.2474365234375,
0.56884765625,
0.2474365234375,
0.56884765625,
0.261474609375,
0.191162109375,
0.261474609375
],
[
0.67626953125,
0.25390625,
0.89208984375,
0.25390625,
0.89208984375,
0.267578125,
0.67626953125,
0.267578125
],
[
0.19287109375,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.19287109375,
0.2734375
],
[
0.191162109375,
0.2734375,
0.51708984375,
0.2734375,
0.51708984375,
0.287353515625,
0.191162109375,
0.287353515625
],
[
0.67724609375,
0.279052734375,
0.87744140625,
0.279052734375,
0.87744140625,
0.292236328125,
0.67724609375,
0.292236328125
],
[
0.1258544921875,
0.290283203125,
0.176513671875,
0.292236328125,
0.1748046875,
0.304931640625,
0.125,
0.302978515625
],
[
0.1895751953125,
0.285888671875,
0.42236328125,
0.285400390625,
0.42236328125,
0.298583984375,
0.1895751953125,
0.2998046875
],
[
0.67626953125,
0.290283203125,
0.8388671875,
0.290283203125,
0.8388671875,
0.3037109375,
0.67626953125,
0.3037109375
],
[
0.1895751953125,
0.296630859375,
0.435546875,
0.296630859375,
0.435546875,
0.31005859375,
0.1895751953125,
0.31005859375
],
[
0.191162109375,
0.311279296875,
0.48046875,
0.311279296875,
0.48046875,
0.324462890625,
0.191162109375,
0.324462890625
],
[
0.67724609375,
0.30615234375,
0.85107421875,
0.30615234375,
0.85107421875,
0.31640625,
0.67724609375,
0.31640625
],
[
0.193603515625,
0.326904296875,
0.64453125,
0.326904296875,
0.64453125,
0.34033203125,
0.193603515625,
0.34033203125
],
[
0.67626953125,
0.326904296875,
0.90625,
0.326904296875,
0.90625,
0.34033203125,
0.67626953125,
0.34033203125
],
[
0.1070556640625,
0.3359375,
0.17724609375,
0.33837890625,
0.17724609375,
0.352783203125,
0.1070556640625,
0.35107421875
],
[
0.191162109375,
0.33837890625,
0.5185546875,
0.33837890625,
0.5185546875,
0.3515625,
0.191162109375,
0.3515625
],
[
0.67626953125,
0.339111328125,
0.8837890625,
0.339111328125,
0.8837890625,
0.352783203125,
0.67626953125,
0.352783203125
],
[
0.19287109375,
0.3515625,
0.4755859375,
0.3515625,
0.4755859375,
0.365478515625,
0.19287109375,
0.365478515625
],
[
0.67626953125,
0.3515625,
0.900390625,
0.3515625,
0.900390625,
0.365478515625,
0.67626953125,
0.365478515625
],
[
0.191162109375,
0.369384765625,
0.57177734375,
0.371826171875,
0.57177734375,
0.385009765625,
0.191162109375,
0.38330078125
],
[
0.67626953125,
0.365478515625,
0.88232421875,
0.365478515625,
0.88232421875,
0.37890625,
0.67626953125,
0.37890625
],
[
0.67626953125,
0.37744140625,
0.85302734375,
0.37744140625,
0.85302734375,
0.38818359375,
0.67626953125,
0.38818359375
],
[
0.11767578125,
0.38818359375,
0.178955078125,
0.38818359375,
0.178955078125,
0.40283203125,
0.11767578125,
0.40283203125
],
[
0.193603515625,
0.385009765625,
0.4794921875,
0.385009765625,
0.4794921875,
0.395751953125,
0.193603515625,
0.395751953125
],
[
0.67626953125,
0.389404296875,
0.89453125,
0.389404296875,
0.89453125,
0.40283203125,
0.67626953125,
0.40283203125
],
[
0.19287109375,
0.395751953125,
0.51318359375,
0.395751953125,
0.51318359375,
0.4091796875,
0.19287109375,
0.4091796875
],
[
0.67626953125,
0.401611328125,
0.869140625,
0.401611328125,
0.869140625,
0.415283203125,
0.67626953125,
0.415283203125
],
[
0.19287109375,
0.4091796875,
0.483642578125,
0.4091796875,
0.483642578125,
0.42236328125,
0.19287109375,
0.42236328125
],
[
0.67626953125,
0.4140625,
0.861328125,
0.4140625,
0.861328125,
0.427978515625,
0.67626953125,
0.427978515625
],
[
0.193603515625,
0.429931640625,
0.5087890625,
0.429931640625,
0.5087890625,
0.43994140625,
0.193603515625,
0.43994140625
],
[
0.67626953125,
0.429931640625,
0.83642578125,
0.4287109375,
0.83642578125,
0.43994140625,
0.67626953125,
0.44140625
],
[
0.131591796875,
0.439453125,
0.178955078125,
0.44140625,
0.17724609375,
0.455810546875,
0.1298828125,
0.453857421875
],
[
0.193603515625,
0.442626953125,
0.48046875,
0.442626953125,
0.48046875,
0.45263671875,
0.193603515625,
0.45263671875
],
[
0.6748046875,
0.43994140625,
0.869140625,
0.439453125,
0.869140625,
0.453857421875,
0.6748046875,
0.455078125
],
[
0.193603515625,
0.455078125,
0.47802734375,
0.455078125,
0.47802734375,
0.46533203125,
0.193603515625,
0.46533203125
],
[
0.6748046875,
0.453857421875,
0.830078125,
0.453857421875,
0.830078125,
0.46728515625,
0.6748046875,
0.46728515625
],
[
0.08660888671875,
0.477783203125,
0.90869140625,
0.479248046875,
0.90869140625,
0.49365234375,
0.08660888671875,
0.492431640625
],
[
0.08660888671875,
0.52099609375,
0.4794921875,
0.52099609375,
0.4794921875,
0.5341796875,
0.08660888671875,
0.5341796875
],
[
0.521484375,
0.52099609375,
0.9111328125,
0.51953125,
0.9111328125,
0.5341796875,
0.521484375,
0.53515625
],
[
0.0849609375,
0.53515625,
0.476318359375,
0.53515625,
0.476318359375,
0.548828125,
0.0849609375,
0.548828125
],
[
0.51953125,
0.53515625,
0.708984375,
0.53515625,
0.708984375,
0.548828125,
0.51953125,
0.548828125
],
[
0.08660888671875,
0.55029296875,
0.4755859375,
0.55029296875,
0.4755859375,
0.56103515625,
0.08660888671875,
0.56103515625
],
[
0.0849609375,
0.56298828125,
0.4794921875,
0.56298828125,
0.4794921875,
0.5771484375,
0.0849609375,
0.5771484375
],
[
0.68310546875,
0.55810546875,
0.75244140625,
0.55810546875,
0.75244140625,
0.57470703125,
0.68310546875,
0.57470703125
],
[
0.08660888671875,
0.5771484375,
0.47802734375,
0.5771484375,
0.47802734375,
0.59033203125,
0.08660888671875,
0.59033203125
],
[
0.51953125,
0.57763671875,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.51953125,
0.59130859375
],
[
0.08660888671875,
0.59228515625,
0.47802734375,
0.59228515625,
0.47802734375,
0.60302734375,
0.08660888671875,
0.60302734375
],
[
0.51953125,
0.59228515625,
0.9140625,
0.59228515625,
0.9140625,
0.60595703125,
0.51953125,
0.60595703125
],
[
0.0849609375,
0.60400390625,
0.47802734375,
0.60400390625,
0.47802734375,
0.6171875,
0.0849609375,
0.6171875
],
[
0.5185546875,
0.60595703125,
0.91015625,
0.60693359375,
0.91015625,
0.62060546875,
0.5185546875,
0.619140625
],
[
0.08660888671875,
0.61865234375,
0.47802734375,
0.61865234375,
0.47802734375,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.521484375,
0.6220703125,
0.90869140625,
0.6220703125,
0.90869140625,
0.6318359375,
0.521484375,
0.6318359375
],
[
0.0849609375,
0.6318359375,
0.4794921875,
0.6318359375,
0.4794921875,
0.64599609375,
0.0849609375,
0.64599609375
],
[
0.521484375,
0.63525390625,
0.90869140625,
0.63525390625,
0.90869140625,
0.64599609375,
0.521484375,
0.64599609375
],
[
0.0849609375,
0.646484375,
0.4794921875,
0.646484375,
0.4794921875,
0.66015625,
0.0849609375,
0.66015625
],
[
0.51953125,
0.64794921875,
0.91259765625,
0.64794921875,
0.91259765625,
0.6611328125,
0.51953125,
0.6611328125
],
[
0.0849609375,
0.6591796875,
0.47802734375,
0.6591796875,
0.47802734375,
0.673828125,
0.0849609375,
0.673828125
],
[
0.51953125,
0.662109375,
0.91259765625,
0.662109375,
0.91259765625,
0.67529296875,
0.51953125,
0.67529296875
],
[
0.0849609375,
0.673828125,
0.47802734375,
0.673828125,
0.47802734375,
0.6875,
0.0849609375,
0.6875
],
[
0.521484375,
0.67529296875,
0.91259765625,
0.67529296875,
0.91259765625,
0.689453125,
0.521484375,
0.689453125
],
[
0.08416748046875,
0.6875,
0.4794921875,
0.68603515625,
0.4794921875,
0.70068359375,
0.08416748046875,
0.7021484375
],
[
0.51953125,
0.689453125,
0.8349609375,
0.689453125,
0.8349609375,
0.70263671875,
0.51953125,
0.70263671875
],
[
0.0849609375,
0.70068359375,
0.44287109375,
0.70068359375,
0.44287109375,
0.71533203125,
0.0849609375,
0.71533203125
],
[
0.52294921875,
0.7109375,
0.77685546875,
0.7109375,
0.77685546875,
0.72607421875,
0.52294921875,
0.72607421875
],
[
0.10296630859375,
0.72265625,
0.4794921875,
0.72265625,
0.4794921875,
0.736328125,
0.10296630859375,
0.736328125
],
[
0.51953125,
0.72802734375,
0.9150390625,
0.72802734375,
0.9150390625,
0.74560546875,
0.51953125,
0.74560546875
],
[
0.0849609375,
0.7373046875,
0.4794921875,
0.7373046875,
0.4794921875,
0.75048828125,
0.0849609375,
0.75048828125
],
[
0.51708984375,
0.7421875,
0.9150390625,
0.7421875,
0.9150390625,
0.76318359375,
0.51708984375,
0.76318359375
],
[
0.0849609375,
0.75,
0.476318359375,
0.751953125,
0.476318359375,
0.7666015625,
0.0849609375,
0.7646484375
],
[
0.08660888671875,
0.76513671875,
0.4794921875,
0.76513671875,
0.4794921875,
0.77880859375,
0.08660888671875,
0.77880859375
],
[
0.51953125,
0.76025390625,
0.91259765625,
0.76025390625,
0.91259765625,
0.7734375,
0.51953125,
0.7734375
],
[
0.0849609375,
0.77880859375,
0.47802734375,
0.77880859375,
0.47802734375,
0.79248046875,
0.0849609375,
0.79248046875
],
[
0.521484375,
0.77587890625,
0.91015625,
0.77587890625,
0.91015625,
0.7861328125,
0.521484375,
0.7861328125
],
[
0.5185546875,
0.7861328125,
0.91259765625,
0.787109375,
0.91259765625,
0.8017578125,
0.5185546875,
0.80029296875
],
[
0.08660888671875,
0.79248046875,
0.4794921875,
0.79248046875,
0.4794921875,
0.80615234375,
0.08660888671875,
0.80615234375
],
[
0.0849609375,
0.80615234375,
0.47802734375,
0.80615234375,
0.47802734375,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.51953125,
0.8017578125,
0.9111328125,
0.8017578125,
0.9111328125,
0.81494140625,
0.51953125,
0.81494140625
],
[
0.08660888671875,
0.8193359375,
0.4794921875,
0.8193359375,
0.4794921875,
0.83349609375,
0.08660888671875,
0.83349609375
],
[
0.521484375,
0.8173828125,
0.91015625,
0.8173828125,
0.91015625,
0.82763671875,
0.521484375,
0.82763671875
],
[
0.08660888671875,
0.833984375,
0.476318359375,
0.833984375,
0.476318359375,
0.84765625,
0.08660888671875,
0.84765625
],
[
0.51953125,
0.83203125,
0.9111328125,
0.83203125,
0.9111328125,
0.8466796875,
0.51953125,
0.8466796875
],
[
0.08660888671875,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.51953125,
0.84765625,
0.91259765625,
0.8466796875,
0.91259765625,
0.861328125,
0.51953125,
0.8623046875
],
[
0.08660888671875,
0.8623046875,
0.47802734375,
0.8623046875,
0.47802734375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.521484375,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.521484375,
0.87548828125
],
[
0.0849609375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51953125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"StrategyDefinition and Examples",
"Connection withPrior Work",
"The persuaders indicating their intentions to take acts or justify their earlier",
"Commitment(Yanget al.2019)",
"Commitment",
"decisions to convince others that they have made the correct choice.",
"Self-modeling (Wang et al.2019)",
"e.g.,I just lent to Auntie Fine's Donut Shop. (Kiva)",
"Commitment (Vargheeseetal.2020b)",
"Ethos (Carlile et al. 2018),",
"Making request full of emotional valence and arousal affect to influence others.",
"Emotion appeal (Carlile et al. 2018),",
"Emotion",
"e.g.,GuysI'm desperate.(Borrow)",
"Sentiment (Durmus et al. 2018)",
"I've been in the lowest depressive state of my life. (RAOP)",
"Emotion words (Luu et al, 2019a),",
"Emotion (Asai et al. 2020)",
"Politeness (Durmus et al. 2018),",
"The usage of polite language in requests.",
"Politeness",
"Politeness (Althoff et al. 2014)",
"e.g.,Yourhelp isdeeplyappreciated!(Borrow)",
"Politeness (Nashruddin et al. 2020)",
"Responding to a positive action with another positive action. People are more",
"likely to help if they have received help themselves.",
"Reciprocity (Althoff et al. 2014),",
"Reciprocity",
"Reciprocity (Roethke et al. 2020)",
"e.g., I will pay 5% interest no later than May I, 2016. (Borrow)",
"Reciprocity (Vargheese et al. 2020b)",
"I'll pay it forward with my first check. (RAOP)",
"People emphasizing on the urgency, rare of their needs",
"Scarcity (Vargheese et al. 2020b),",
"Scarcity",
"e.g.,Need this loan urgently. (Borrow)",
"Scarcity (Yang et al. 2019),",
"I haven't ate a meal in two days. (RAOP)",
"Loan expiring today and still needs $650. (Kiva)",
"Scarcity (Lawson etal.2020)",
"The uses of credentials impacts to establish credibility and earn others trust.",
"Credibility appeal (Wang et al.2019),",
"Credibility",
"e.g., Can provide any documentation needed. (Borrow)",
"Social Proof (Roethke et al. 2020)",
"She has already repaid 2 previous loans. (Kiva)",
"Social Proof (Vargheese et al. 2020a)",
"Providing concrete facts or evidence for the narrative or request.",
"Evidentiality (Althoff et al.2014),",
"Evidence (Carlile et al 2018)",
"Evidence",
"e.g.Myinsurancewascanceledtoday(Borrow)",
"Evidence (Stab and Gurevych 2014),",
"There is a Pizza Hut and a Dominos near me. (RAOP)",
"Concreteness (Yang et al. 2019)",
"$225 to go and 1 A+ member on the loan. (Kiva)",
"Evidence (Durmus et al. 2018)",
"Emphasizingtheimportance orimpactoftherequest.",
"Logos (Carlile et al.2018)",
"Impact",
"e.g.,I will usethis loan to pay myrent.(Borrow)",
"Logic appeal (Wang et al. 2019)",
"This loan will helphim withhis business. (Kiva)",
"Impact (Yang et al. 2019)",
"Table : The generic taxonomy of persuasive strategies, their definitions, example sentences, and connections with prior work",
"than 8 sentences, and 85% posts in Borrow has at most 7 sen-",
"[2,3],[3,4),[4,oo), accounting for 44.1%, 20.3%, 12.4%",
"tences. We recruited four research assistants to label persua-",
"and 33.2% of all documents.",
"sion strategiesforeachsentencein sampleddocuments.Def-",
"initions and examples of different persuasion strategies were",
"Method",
"provided, together with a training session where we asked",
"To alleviate the dependencies on labeled data, we propose",
"annotatorstoannotateanumberof examplesentencesand",
"a hierarchical weakly-supervised latent variable model to",
"walked them through any disagreed annotations. To assess",
"leveragepartiallylabeleddatatopredict sentence-level per-",
"the reliability of the annotated labels, we then asked them",
"suasive strategies. Specifically, we introduce a sentence",
"to annotate the same 100 documents with 400 sentences and",
"level latent variable model to reconstruct the input sen",
"computed Cohen's Kappa coefficient to measure inter-rater",
"tence and predict the sentence-level persuasion labels spon-",
"reliability. We obtained an average score of 0.538 on Kiva,",
"taneously, supervised by the global or document-level labels",
"0.613 on RAOP, and 0.623 on Borrow, which indicates mod-",
"(e.g., overall persuasiveness of the documents). The overall",
"erate agreement (McHugh 2012).Annotators then annotated",
"architecture of our method is shown in Figure 2.",
"the rest 1200 documents by themselves independently.",
"Weakly Supervised Latent Model",
"The dataset statistics are shown in Table 2, and the",
"Given a corpus of N documents D = {d,}1, where each",
" sentence-level label distribution in each dataset is shown in",
"document d consists of M sentences d, - ($)1. For",
"Figure 1. We merge rare strategies into the Other category",
"Specifically, we merge Commitment, Scarcity, and Emotion",
"each document d; E D, its document level label is denoted",
"in Borrow, Credibility and Commitment in RAOP, Reci-",
"ast,representingtheoverall persuasivenessofthedocu-",
"ments. We divide the corpus into two parts: D = DL UDu,",
"procity and Emotion in Kiva, as Other. We utilized whether",
"the requester received pizzas or loans from the subreddits as",
"where DL (Du) denotes the set of documents with (with-",
"the document-level labels for RAOP and Borrow. 30.1% of",
"out)sentencelabels.Foreach document d, E DL,the cor-",
"people successfully got pizzas on RAOP and 48.5% of peo-",
"responding sentence labels are (y-i, where y? E C=",
"ple received loans on Borrow. In Kiva, we utilized the num-",
"[ck]k-} and represents the persuasive strategy of a given",
"ber of people who lent loans as the document-level labels.",
"sentence. In many practical scenarios, getting document-",
"The numbers are further labeled based on buckets: [1, 2),",
"level labels [ti] is much easier and cheaper than the fine-"
] |
{
"arxiv_id": "2101.06351",
"language": "en",
"page_idx": 2
}
|
||
1410.6264-en-0
|
1410.6264-en-0.jpg
|
# Capturing spatial interdependence in image features: the counting grid, an epitomic representation for bags of features
Alessandro Perina, Nebojsa Jojic Microsoft Research Redmond, USA
Abstract
In recent scene recognition research images or large image regions are often represented as disorganized "bags" of features which can then be analyzed using models originally developed to capture co-variation of word counts in text. However, image feature counts are likely to be constrained in different ways than word counts in text. For example, as a camera pans upwards from a building entrance over its first few floors and then further up into the sky Fig. 1, some feature counts in the image drop while others rise- only to drop again giving way to features found more often at higher elevations. The space of all possible feature count combinations is constrained both by the properties of the larger scene and the size and the location of the window into it. To capture such variation, in this paper we propose the use of the counting grid model. This generative model is based on a grid of feature counts, considerably larger than any of the modeled images, and considerably smaller than the real estate needed to tile the images next to each other tightly. Each modeled image is assumed to have a representative window in the grid in which the feature counts mimic the feature distribution in the image. We provide a learning procedure that jointly maps all images in the training set to the counting grid and estimates the appropriate local counts in it. Experimentally, we demonstrate that the resulting representation captures the space of feature count combinations more accurately than the traditional models, not only when the input images come from a panning camera, but even when modeling images of different scenes from the same category.
Index Terms
Bag of Features Spatial Layout Scene AnalysisBag of Features Spatial Layout Scene AnalysisB
$\spadesuit$
$\bullet$ $A.$ Perina and $N. Jojic$ are with Microsoft $Research$, $Redmond, WA.$
E-mail: alperina@microsoft.com
|
[
[
0.0931396484375,
0.06817626953125,
0.90673828125,
0.06817626953125,
0.90673828125,
0.0889892578125,
0.0931396484375,
0.0889892578125
],
[
0.125,
0.09783935546875,
0.8759765625,
0.09661865234375,
0.8759765625,
0.11871337890625,
0.125,
0.11932373046875
],
[
0.439453125,
0.1287841796875,
0.55615234375,
0.1287841796875,
0.55615234375,
0.147705078125,
0.439453125,
0.147705078125
],
[
0.3798828125,
0.1610107421875,
0.619140625,
0.1610107421875,
0.619140625,
0.175537109375,
0.3798828125,
0.175537109375
],
[
0.369384765625,
0.1767578125,
0.62841796875,
0.1767578125,
0.62841796875,
0.1900634765625,
0.369384765625,
0.1900634765625
],
[
0.471435546875,
0.205810546875,
0.52783203125,
0.205810546875,
0.52783203125,
0.2203369140625,
0.471435546875,
0.2203369140625
],
[
0.1683349609375,
0.2266845703125,
0.85546875,
0.2266845703125,
0.85546875,
0.240478515625,
0.1683349609375,
0.240478515625
],
[
0.1446533203125,
0.2412109375,
0.85302734375,
0.2412109375,
0.85302734375,
0.251953125,
0.1446533203125,
0.251953125
],
[
0.143798828125,
0.255126953125,
0.8544921875,
0.255126953125,
0.8544921875,
0.26513671875,
0.143798828125,
0.26513671875
],
[
0.0228729248046875,
0.26953125,
0.05963134765625,
0.26953125,
0.05963134765625,
0.4072265625,
0.0228729248046875,
0.4072265625
],
[
0.143798828125,
0.267578125,
0.8544921875,
0.267578125,
0.8544921875,
0.27783203125,
0.143798828125,
0.27783203125
],
[
0.1446533203125,
0.281005859375,
0.85302734375,
0.281005859375,
0.85302734375,
0.291748046875,
0.1446533203125,
0.291748046875
],
[
0.1446533203125,
0.29345703125,
0.44775390625,
0.29345703125,
0.44775390625,
0.3037109375,
0.1446533203125,
0.3037109375
],
[
0.455078125,
0.29736328125,
0.465576171875,
0.29736328125,
0.465576171875,
0.300537109375,
0.455078125,
0.300537109375
],
[
0.476318359375,
0.29541015625,
0.85302734375,
0.29541015625,
0.85302734375,
0.302978515625,
0.476318359375,
0.302978515625
],
[
0.146240234375,
0.306884765625,
0.85546875,
0.306884765625,
0.85546875,
0.317626953125,
0.146240234375,
0.317626953125
],
[
0.143798828125,
0.318115234375,
0.8544921875,
0.318115234375,
0.8544921875,
0.33203125,
0.143798828125,
0.33203125
],
[
0.146240234375,
0.33203125,
0.85302734375,
0.33203125,
0.85302734375,
0.34228515625,
0.146240234375,
0.34228515625
],
[
0.1446533203125,
0.34521484375,
0.8544921875,
0.34521484375,
0.8544921875,
0.35595703125,
0.1446533203125,
0.35595703125
],
[
0.1446533203125,
0.359130859375,
0.8544921875,
0.359130859375,
0.8544921875,
0.369384765625,
0.1446533203125,
0.369384765625
],
[
0.1422119140625,
0.37060546875,
0.85546875,
0.369384765625,
0.85546875,
0.38330078125,
0.1422119140625,
0.3837890625
],
[
0.1422119140625,
0.38330078125,
0.85546875,
0.3837890625,
0.85546875,
0.397705078125,
0.1422119140625,
0.396484375
],
[
0.143798828125,
0.396484375,
0.63134765625,
0.397705078125,
0.63134765625,
0.4091796875,
0.143798828125,
0.40771484375
],
[
0.0228729248046875,
0.406005859375,
0.062103271484375,
0.406005859375,
0.062103271484375,
0.50244140625,
0.0228729248046875,
0.50244140625
],
[
0.4599609375,
0.423583984375,
0.53759765625,
0.423583984375,
0.53759765625,
0.43505859375,
0.4599609375,
0.43505859375
],
[
0.1683349609375,
0.442626953125,
0.71826171875,
0.442626953125,
0.71826171875,
0.455810546875,
0.1683349609375,
0.455810546875
],
[
0.0228729248046875,
0.4912109375,
0.05963134765625,
0.4912109375,
0.05963134765625,
0.7041015625,
0.0228729248046875,
0.7041015625
],
[
0.07843017578125,
0.90380859375,
0.455078125,
0.90380859375,
0.455078125,
0.91748046875,
0.07843017578125,
0.91748046875
],
[
0.0931396484375,
0.91845703125,
0.26318359375,
0.91845703125,
0.26318359375,
0.9287109375,
0.0931396484375,
0.9287109375
]
] |
[
"Capturing spatial interdependence in image features:the",
"counting grid, an epitomic representation for bags of",
"features",
"Alessandro Perina,Nebojsa Jojic",
"MicrosoftResearchRedmond,USA",
"Abstract",
"In recent scene recognition research images or large image regions are often represented as disorganized \"bags\" of",
"features which can then be analyzed using models originally developedto capture co-variation of word counts intext",
"However,image feature counts are likelyto be constrained in different ways than word counts in text.For example,asa",
"1",
"camera pansupwards froma building entrance overitsfirstfewfloors andthenfurtherup intothe skyFig.1,somefeature",
"counts in the imagedropwhileothers rise-onlytodropagaingivingwaytofeaturesfound moreoftenat higherelevations",
"The spaceofall possiblefeaturecountcombinations",
"",
"onstrainedbothbythepropertiesofthelargersceneandthesize",
"and the location of the window into it.To capture such variation, in this paper we propose the use of the counting grid",
"model. This generative model is based on a grid of feature counts, considerably larger than any of the modeled images,",
"and considerablysmallerthanthe real estateneededtotilethe images nexttoeachothertightly.Eachmodeled image",
"is assumed to have a representative window in the grid in which the feature counts mimic the feature distribution in the",
"image.Weprovidealearningprocedurethat jointlymapsall images inthetrainingsettothecountinggrid andestimatesthe",
"appropriate local counts in it.Experimentally,wedemonstrate that theresulting representation captures the space of feature",
"count combinations more accurately than the traditional models, not only when the input images come from a panning",
"camera,but evenwhenmodeling images ofdifferent scenes fromthe samecategory",
"[']",
"IndexTerms",
"Bag of Features Spatial Layout Scene AnalysisBag of Features Spatial Layout SceneAnalysisB",
"",
"· A. Perina and N.Jojic are with Microsoft Research, Redmond, WA.",
"E-mail:alperina@microsoft.com"
] |
{
"arxiv_id": "1410.6264",
"language": "en",
"page_idx": 0
}
|
||
1410.6264-en-2
|
1410.6264-en-2.jpg
|
2
$i)$
<FigureHere>
$\square$Sky $\blacksquare$Mount $\square$Grass $\square$Bldg $\blacksquare$Roof $\square$Train $\square$Tracks ◼Fence
Fig.2. Counting grid illustration. i) Images and their Bag of feature representation. ii) Images of a train station, taken as windows into the larger scene. iii) Hand labeled features. iv) Scene reconstructed (e,g, counting grid) starting from bags taken from 50 windows (see the Text for details).
<FigureHere>
소수하다.가다.기다.지다.기다.기다.기다.기다.기다.
$\square$ Sky ◼Mount □Felds DSand ECars DBldgs ▣Street ▣Sea\_
$i)$
$\square$ Sky ◼Mount ◼Grass □Bldg ▣Roof $\square$Train DTracks ◼Fence
Fig. 3. Illustration of counting grids using different data at different levels of abstraction. At the top of each CG we show few examples of the training images from which we extracted the bags. The bottom row of each panel, illustrates the features. i) A counting grid learned using patches extracted from the train station toy example of Fig.2.In this case all the bags come from windows into the same image, which is reconstructed on the grid. ii) A counting grid estimated from images taken with a wearable camera[18]. In this case we learned a dictionary of features (illustrated by the textons on the bottom) clustering image patches, as in [19]. To illustrate the office scene (e.g., the computer screens- see images on the top), we overlap these textons by as much as the patches were overlapping during feature extraction process, and then average to create a clearer visual representation. iii) A counting grid estimated starting from LabelMe annotations (see the top row) [20].
original window location information, the counting grid was images of different scenes. For example in Fig. 3-iii) we used as computed so that for each training image, a window into the features the human-supplied labels for LabelMe[20] dataset, counting grid can be found so that the appropriate sections and in- $ii)$ the outputs of hundreds of simple computational have matching histograms. The resolution of the reconstructed
feature layout of the large scene goes well beyond what would the SenseCam dataset [18]. As opposed to the train station be expected from a crude $2\times2$ tessellation of the input images example, input images are not subimages of a larger single (the height of each section is roughly 20% of the large scene scene, but rather images of the same types of scenes. Each and only the feature counts in each section were used, not window into a counting grid represents a possible feature their spatial layout within the section). Although none of the training examples was taken from the area close to D where all five of D's features can be seen in a single image, that part patterns very coarsely, but depending on feature count coof the scene is reconstructed as well, and D's histogram can variation for most of its reasoning power. For example, despite be matched well.
feature detectors applied to images from various scenes in combination$^1$ present in the dataset (3- $ii))$ and the model is able to reconstruct the feature layout only exploiting the spatial only relying on a 2×2 tessellation of each input image, full resolution panoramas of office and corridors are visible in the ${\mathrm{CG~in~Fig.~3-}ii}).$
This paper presents and extends the counting grid model [21,22]. The basic model is extended to include priors which help with overfitting issues. We also formally introduce the tessellated counting grid model and analyzed the the extreme
In this simple example, the training images are different views of a single scene. However, at the feature level, images of other train stations are likely have a similar layout, and so they could be used to learn a counting grid. In practice, we rarely have access to highly discriminative and reliable features, and so instead of the 8 fake features in our example, in our experiments we had to use hundreds of simpler automatically derived features, and infer the counting grids from related
1.The CG model also learns a prior over the grid window usage
which may prevent some combinations.
|
[
[
0.72119140625,
0.07135009765625,
0.73046875,
0.0706787109375,
0.73193359375,
0.07574462890625,
0.72216796875,
0.0770263671875
],
[
0.82177734375,
0.06817626953125,
0.82666015625,
0.06817626953125,
0.82666015625,
0.07452392578125,
0.82177734375,
0.07452392578125
],
[
0.53173828125,
0.18310546875,
0.90673828125,
0.18310546875,
0.90673828125,
0.1932373046875,
0.53173828125,
0.1932373046875
],
[
0.080078125,
0.212158203125,
0.9189453125,
0.212158203125,
0.9189453125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.0767822265625,
0.2259521484375,
0.9208984375,
0.2266845703125,
0.9208984375,
0.2412109375,
0.0767822265625,
0.240478515625
],
[
0.07843017578125,
0.2412109375,
0.298095703125,
0.2412109375,
0.298095703125,
0.255126953125,
0.07843017578125,
0.255126953125
],
[
0.77392578125,
0.27978515625,
0.7998046875,
0.27978515625,
0.7998046875,
0.287353515625,
0.77392578125,
0.287353515625
],
[
0.7255859375,
0.300537109375,
0.7509765625,
0.300537109375,
0.7509765625,
0.309326171875,
0.7255859375,
0.309326171875
],
[
0.73046875,
0.32763671875,
0.7685546875,
0.32763671875,
0.7685546875,
0.33837890625,
0.73046875,
0.33837890625
],
[
0.81396484375,
0.322509765625,
0.861328125,
0.322509765625,
0.861328125,
0.333251953125,
0.81396484375,
0.333251953125
],
[
0.69287109375,
0.328857421875,
0.73046875,
0.328857421875,
0.73046875,
0.339111328125,
0.69287109375,
0.339111328125
],
[
0.82275390625,
0.3359375,
0.8544921875,
0.3359375,
0.8544921875,
0.343505859375,
0.82275390625,
0.343505859375
],
[
0.5009765625,
0.3466796875,
0.57861328125,
0.3466796875,
0.57861328125,
0.35791015625,
0.5009765625,
0.35791015625
],
[
0.82861328125,
0.347900390625,
0.85888671875,
0.347900390625,
0.85888671875,
0.35595703125,
0.82861328125,
0.35595703125
],
[
0.69921875,
0.359130859375,
0.7294921875,
0.359130859375,
0.7294921875,
0.37060546875,
0.69921875,
0.37060546875
],
[
0.420654296875,
0.386962890625,
0.471435546875,
0.386962890625,
0.471435546875,
0.397705078125,
0.420654296875,
0.397705078125
],
[
0.72802734375,
0.3837890625,
0.76953125,
0.3837890625,
0.76953125,
0.39453125,
0.72802734375,
0.39453125
],
[
0.69140625,
0.401611328125,
0.72119140625,
0.401611328125,
0.72119140625,
0.4091796875,
0.69140625,
0.4091796875
],
[
0.75732421875,
0.3984375,
0.78515625,
0.3984375,
0.78515625,
0.406005859375,
0.75732421875,
0.406005859375
],
[
0.830078125,
0.406005859375,
0.84228515625,
0.406005859375,
0.84228515625,
0.4091796875,
0.830078125,
0.4091796875
],
[
0.7294921875,
0.41357421875,
0.7587890625,
0.41357421875,
0.7587890625,
0.421630859375,
0.7294921875,
0.421630859375
],
[
0.85546875,
0.417236328125,
0.8759765625,
0.417236328125,
0.8759765625,
0.4248046875,
0.85546875,
0.4248046875
],
[
0.6708984375,
0.443115234375,
0.701171875,
0.443115234375,
0.701171875,
0.4462890625,
0.6708984375,
0.4462890625
],
[
0.7255859375,
0.4443359375,
0.7490234375,
0.4443359375,
0.7490234375,
0.448974609375,
0.7255859375,
0.448974609375
],
[
0.82861328125,
0.45068359375,
0.84326171875,
0.45068359375,
0.84326171875,
0.453857421875,
0.82861328125,
0.453857421875
],
[
0.101318359375,
0.466552734375,
0.351318359375,
0.46533203125,
0.351318359375,
0.479248046875,
0.101318359375,
0.479736328125
],
[
0.38232421875,
0.46533203125,
0.6259765625,
0.46533203125,
0.6259765625,
0.479736328125,
0.38232421875,
0.479736328125
],
[
0.66650390625,
0.466552734375,
0.90625,
0.466552734375,
0.90625,
0.479736328125,
0.66650390625,
0.479736328125
],
[
0.07843017578125,
0.498779296875,
0.9208984375,
0.498779296875,
0.9208984375,
0.51220703125,
0.07843017578125,
0.51220703125
],
[
0.07843017578125,
0.51318359375,
0.9189453125,
0.51318359375,
0.9189453125,
0.5263671875,
0.07843017578125,
0.5263671875
],
[
0.07843017578125,
0.5263671875,
0.9208984375,
0.5263671875,
0.9208984375,
0.54052734375,
0.07843017578125,
0.54052734375
],
[
0.080078125,
0.54248046875,
0.9189453125,
0.54248046875,
0.9189453125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.07843017578125,
0.5546875,
0.9208984375,
0.5546875,
0.9208984375,
0.568359375,
0.07843017578125,
0.568359375
],
[
0.07843017578125,
0.568359375,
0.9208984375,
0.568359375,
0.9208984375,
0.58203125,
0.07843017578125,
0.58203125
],
[
0.07843017578125,
0.58349609375,
0.9208984375,
0.58349609375,
0.9208984375,
0.5966796875,
0.07843017578125,
0.5966796875
],
[
0.07843017578125,
0.5966796875,
0.6044921875,
0.5966796875,
0.6044921875,
0.60986328125,
0.07843017578125,
0.60986328125
],
[
0.07843017578125,
0.63623046875,
0.5087890625,
0.63623046875,
0.5087890625,
0.6494140625,
0.07843017578125,
0.6494140625
],
[
0.50634765625,
0.63623046875,
0.9189453125,
0.63623046875,
0.9189453125,
0.6494140625,
0.50634765625,
0.6494140625
],
[
0.07843017578125,
0.65087890625,
0.5,
0.65087890625,
0.5,
0.6640625,
0.07843017578125,
0.6640625
],
[
0.5009765625,
0.65478515625,
0.5107421875,
0.65478515625,
0.5107421875,
0.6591796875,
0.5009765625,
0.6591796875
],
[
0.50634765625,
0.65087890625,
0.9189453125,
0.65087890625,
0.9189453125,
0.6640625,
0.50634765625,
0.6640625
],
[
0.07843017578125,
0.6640625,
0.496826171875,
0.6640625,
0.496826171875,
0.67822265625,
0.07843017578125,
0.67822265625
],
[
0.50341796875,
0.6640625,
0.9208984375,
0.6640625,
0.9208984375,
0.67822265625,
0.50341796875,
0.67822265625
],
[
0.080078125,
0.6787109375,
0.491943359375,
0.6787109375,
0.491943359375,
0.689453125,
0.080078125,
0.689453125
],
[
0.5087890625,
0.6796875,
0.9189453125,
0.6796875,
0.9189453125,
0.68994140625,
0.5087890625,
0.68994140625
],
[
0.08087158203125,
0.693359375,
0.49267578125,
0.693359375,
0.49267578125,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.50732421875,
0.6923828125,
0.9208984375,
0.6923828125,
0.9208984375,
0.70556640625,
0.50732421875,
0.70556640625
],
[
0.0767822265625,
0.705078125,
0.494384765625,
0.70556640625,
0.494384765625,
0.7197265625,
0.0767822265625,
0.71826171875
],
[
0.50634765625,
0.70556640625,
0.9208984375,
0.70556640625,
0.9208984375,
0.7197265625,
0.50634765625,
0.7197265625
],
[
0.080078125,
0.7216796875,
0.49267578125,
0.7216796875,
0.49267578125,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50732421875,
0.72021484375,
0.9189453125,
0.72021484375,
0.9189453125,
0.73095703125,
0.50732421875,
0.73095703125
],
[
0.080078125,
0.73486328125,
0.49267578125,
0.73486328125,
0.49267578125,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50634765625,
0.734375,
0.9189453125,
0.734375,
0.9189453125,
0.74755859375,
0.50634765625,
0.74755859375
],
[
0.07843017578125,
0.74755859375,
0.494384765625,
0.74755859375,
0.494384765625,
0.76123046875,
0.07843017578125,
0.76123046875
],
[
0.5048828125,
0.74755859375,
0.9208984375,
0.74755859375,
0.9208984375,
0.76123046875,
0.5048828125,
0.76123046875
],
[
0.07843017578125,
0.76123046875,
0.494384765625,
0.76123046875,
0.494384765625,
0.7744140625,
0.07843017578125,
0.7744140625
],
[
0.5048828125,
0.76025390625,
0.9208984375,
0.76123046875,
0.9208984375,
0.7744140625,
0.5048828125,
0.7734375
],
[
0.07843017578125,
0.7744140625,
0.494384765625,
0.7744140625,
0.494384765625,
0.78857421875,
0.07843017578125,
0.78857421875
],
[
0.50634765625,
0.77587890625,
0.9189453125,
0.77587890625,
0.9189453125,
0.7890625,
0.50634765625,
0.7890625
],
[
0.0767822265625,
0.78857421875,
0.49267578125,
0.78857421875,
0.49267578125,
0.8017578125,
0.0767822265625,
0.8017578125
],
[
0.50634765625,
0.78857421875,
0.9208984375,
0.78857421875,
0.9208984375,
0.8017578125,
0.50634765625,
0.8017578125
],
[
0.0767822265625,
0.80322265625,
0.193603515625,
0.80322265625,
0.193603515625,
0.81640625,
0.0767822265625,
0.81640625
],
[
0.5087890625,
0.8037109375,
0.9189453125,
0.8037109375,
0.9189453125,
0.8173828125,
0.5087890625,
0.8173828125
],
[
0.50732421875,
0.8173828125,
0.9208984375,
0.8173828125,
0.9208984375,
0.83056640625,
0.50732421875,
0.83056640625
],
[
0.0906982421875,
0.83349609375,
0.49267578125,
0.833984375,
0.49267578125,
0.8486328125,
0.0906982421875,
0.84765625
],
[
0.5087890625,
0.83203125,
0.61376953125,
0.83203125,
0.61376953125,
0.8447265625,
0.5087890625,
0.8447265625
],
[
0.0767822265625,
0.8486328125,
0.491943359375,
0.8486328125,
0.491943359375,
0.8623046875,
0.0767822265625,
0.8623046875
],
[
0.51953125,
0.84326171875,
0.92236328125,
0.8447265625,
0.92236328125,
0.8623046875,
0.51953125,
0.861328125
],
[
0.07843017578125,
0.8623046875,
0.491943359375,
0.86279296875,
0.491943359375,
0.875,
0.07843017578125,
0.87353515625
],
[
0.5087890625,
0.85986328125,
0.9189453125,
0.85986328125,
0.9189453125,
0.87353515625,
0.5087890625,
0.87353515625
],
[
0.07843017578125,
0.876953125,
0.491943359375,
0.876953125,
0.491943359375,
0.89013671875,
0.07843017578125,
0.89013671875
],
[
0.50732421875,
0.875,
0.9189453125,
0.875,
0.9189453125,
0.88818359375,
0.50732421875,
0.88818359375
],
[
0.07843017578125,
0.8916015625,
0.49267578125,
0.8916015625,
0.49267578125,
0.9013671875,
0.07843017578125,
0.9013671875
],
[
0.50634765625,
0.88818359375,
0.9208984375,
0.88818359375,
0.9208984375,
0.90283203125,
0.50634765625,
0.90283203125
],
[
0.0767822265625,
0.90283203125,
0.491943359375,
0.90380859375,
0.491943359375,
0.91748046875,
0.0767822265625,
0.91650390625
],
[
0.07843017578125,
0.91845703125,
0.490234375,
0.91845703125,
0.490234375,
0.931640625,
0.07843017578125,
0.931640625
],
[
0.51953125,
0.91943359375,
0.9208984375,
0.92041015625,
0.9208984375,
0.93505859375,
0.51953125,
0.9345703125
],
[
0.0767822265625,
0.931640625,
0.49267578125,
0.931640625,
0.49267578125,
0.9462890625,
0.0767822265625,
0.9462890625
],
[
0.50732421875,
0.9345703125,
0.744140625,
0.9345703125,
0.744140625,
0.9443359375,
0.50732421875,
0.9443359375
]
] |
[
"V",
"LL",
"SkyMountGrassBldg Roof Train TracksFence",
"Fig. 2. Counting grid illustration. i) Images and their Bag of feature representation. i) Images of a train station, taken as windows",
"into the larger scene. ili) Hand labeled features. iv) Scene reconstructed (e.g., counting grid) starting from bags taken from 50",
"windows (see the Text for details).",
"",
"",
"ROCKS",
"BUILDINGS",
"SEA",
"vehicles",
"CORRIDOR",
"STREET",
"SNOW",
"OFFICE",
"MOUNTAIN",
"FIELDS",
"ROCKS",
"SFA",
"WATER",
"SAND",
"BDING",
"MATER",
"",
"SkyMount Grass Bldg Roof Train Tracks Fence",
"",
"SkyMount Fields Sand CarsBldgs Street Sea.",
"Fig. 3. Illustration of counting grids using different data at different levels of abstraction. At the top of each CG we show few",
"examples of the training images from which we extracted the bags. The bottom row of each panel, illustrates the features. i) A",
"counting grid learned using patches extracted from the train station toy example of Fig. 2. In this case all the bags come from",
"windowsintothesameimage,whichisreconstructedonthegrid.ii)Acountinggridestimatedfromimagestakenwithawearable",
"camera [18]. In this case we learned a dictionary of features (illustrated by the textons on the bottom) clustering image patches, as",
"in [19]. To illustrate the office scene (e.g., the computer screens - see images on the top), we overlap these textons by as much as",
"the patches were overlapping during feature extraction process, and then average to create a clearer visual representation. ili) A",
"countinggridestimated startingfrom LabelMeannotations (seethetoprow) [2o].",
"original window location information, the counting grid was",
"images of different scenes. For example in Fig. 3-ii) we used as",
"computed so that for each training image, a window into the",
"",
"features the human-supplied labels for LabelMe [20] dataset,",
"counting grid can be found so that the appropriate sections",
"and in - ii) the outputs of hundreds of simple computational",
"havematchinghistograms.Theresolutionofthereconstructed",
"featuredetectors appliedto images from variousscenes in",
"featurelayoutofthelargescenegoeswellbeyondwhatwould",
"the SenseCam dataset [18]. As opposed to the train station",
"be expected from a crude 2 × 2 tessellation of the input images",
"example, input images are not subimages of a larger single",
"(theheightof eachsectionisroughly20%ofthelargescene",
"scene, but rather images of the same tvpes of scenes. Each",
"and only the feature counts in each section were used, not",
"window into a counting grid represents a possible feature",
"their spatial layout within the section). Although none of the",
"combination' present in the dataset (3- i) and the model is",
"training examples was taken from the area close to D where",
"able to reconstruct the feature layout only exploiting the spatial",
"all five of D's features can be seen in a single image, that part",
"patterns very coarsely, but depending on feature count co-",
"of the scene is reconstructed as well, and D's histogram can",
"variation for most of its reasoning power. For example, despite",
"be matched well.",
"only relying on a 2x2 tessellation of each input image, full",
"resolution panoramas of office and corridors are visible in the",
"In this simple example, the training images are different",
"CG in Fig. 3-ii).",
"views of a single scene. However, at the feature level, images of",
"This paper presents and extends the counting grid model",
"othertrain stations arelikelyhavea similarlayout,and sothey",
"[21,22]. The basic model is extended to include priors which",
"could be used to learn a counting grid. In practice, we rarely",
"help with overfitting issues. We also formally introduce the",
"haveaccessto highlydiscriminative and reliable features,",
"tessellated counting grid model and analyzed the the extreme",
"and so instead of the 8 fake features in our example, in our",
"experiments we had to use hundreds of simpler automatically",
"1. The CG model also learns a prior over the grid window usage",
"derived features, and infer the counting grids from related",
"which may prevent some combinations."
] |
{
"arxiv_id": "1410.6264",
"language": "en",
"page_idx": 2
}
|
||
2310.04444-en-0
|
2310.04444-en-0.jpg
|
$What^{\prime }s\textit{ the Magic Word? A CoNTROL THEORY OF LLM}$
# $\mathsf{PROMPTING}$
$\textbf{Aman Bhargava}^1,\textbf{Cameron Witkowski}^2,\textbf{Manav Shah}^2,\textbf{Matt Thomson}^1{}^*$
$^{1}$California Institute of Technology, 2University of Toronto
$\{$abhargav,mthomson$\}$ecaltech.edu
$\{$cameron.witkowski,manav.shah$\}$@mail.utoronto.ca
### ABSTRACT
Prompt engineering is crucial for deploying LLMs but is poorly understood mathematically. We formalize LLM systems as a class of discrete stochastic dynamical systems to explore prompt engineering through the lens of control theory. We investigate the reachable set of output token sequences $\mathcal{R}_y(\mathbf{x}_0)$ for which there exists a control input sequence u for each $\mathbf{y}\in\mathcal{R}_y(\mathbf{x}_0)$ that steers the LLM to output y from initial state sequence $\mathbf{x}_0.$ We offer analytic analysis on the limitations on the controllability of self-attention in terms of reachable set, where we prove an upper bound on the reachable set of outputs $\mathcal{R}_y(\mathbf{x}_0)$ as a function of the singular values of the parameter matrices. We present complementary empirical analysis on the controllability of a panel of LLMs, including Falcon-7b, Llama-7b, and Falcon-40b. Our results demonstrate a lower bound on the reachable set of outputs $\mathcal{R}_{y}(\mathbf{x}_{0})$ w.r.t. initial state sequences $\mathbf{x}_0$ sampled from the Wikitext dataset. We find that the correct next Wikitext token following sequence $\mathbf{x}_0$ is reachable over 97% of the time with prompts of $k\leq10$ tokens. We also establish that the top 75 most likely next tokens, as estimated by the LLM itself, are reachable at least 85% of the time with prompts of $k\leq10$ tokens. Intriguingly, short prompt sequences can dramatically alter the likelihood of specific outputs, even making the least likely tokens become the most likely ones. This control-centric analysis of LLMs demonstrates the significant and poorly understood role of input sequences in steering output probabilities, offering a foundational perspective for enhancing language model system capabilities.
1 INTRODUCTION
LLMs pre-trained on unsupervised next token prediction objectives exhibit unprecedented dynamic reprogrammability achieved through“prompting”, often referred to as zero-shot learning (Brown et al., 2020; Wei et al., 2022; Hagendorff, $\hat{2}023;$ Noever &McKee, 2023; OpenAI, 2023; 2022). These capabilities appear to emerge as the model's size, training data, and training time are scaled. The dynamic reprogrammability of LLMs is akin to the adaptable computational capacities observed in biological systems. This feature finds applications across domains such as machine translation (Wang et al., $\dot{2}023$a), code generation (Rôzière et al., 2023), and chatbots Bai et al.(2022). A rigorous understanding of the prompt's influence over LLM generation would be of great utility for understanding LLMs and building more robust and capable systems leveraging LLMs.
Strategies for controlling pre-trained LLM generation today fall into three broad categories (Zhang
et al., 2022):
1. Input Optimization (Prompting): Adjusting the input tokens (e.g., rewording the
prompt) to improve subsequent text generation.
2. Model Optimization: Adjusting the weights of the network (e.g., fine-tuning, RLHF) to
improve model behavior during inference.
3. Post-processing: Adjusting or re-ranking generated text (e.g., surrogate ranking algo-
$rithm).$
$^{*}$Code will be made available at https://github.com/amanb2000/Magic-Words.
1
|
[
[
0.178955078125,
0.10162353515625,
0.82177734375,
0.10162353515625,
0.82177734375,
0.11932373046875,
0.178955078125,
0.11932373046875
],
[
0.1748046875,
0.126953125,
0.32177734375,
0.1287841796875,
0.32177734375,
0.146484375,
0.1748046875,
0.14453125
],
[
0.1854248046875,
0.17041015625,
0.69677734375,
0.17041015625,
0.69677734375,
0.1849365234375,
0.1854248046875,
0.1849365234375
],
[
0.1854248046875,
0.186279296875,
0.57177734375,
0.186279296875,
0.57177734375,
0.199462890625,
0.1854248046875,
0.199462890625
],
[
0.188720703125,
0.2008056640625,
0.484375,
0.2008056640625,
0.484375,
0.2139892578125,
0.188720703125,
0.2139892578125
],
[
0.1871337890625,
0.21533203125,
0.64208984375,
0.21533203125,
0.64208984375,
0.2291259765625,
0.1871337890625,
0.2291259765625
],
[
0.453369140625,
0.264404296875,
0.5458984375,
0.264404296875,
0.5458984375,
0.279052734375,
0.453369140625,
0.279052734375
],
[
0.0228729248046875,
0.27587890625,
0.060455322265625,
0.27587890625,
0.060455322265625,
0.397705078125,
0.0228729248046875,
0.397705078125
],
[
0.2344970703125,
0.294189453125,
0.76318359375,
0.294189453125,
0.76318359375,
0.309326171875,
0.2344970703125,
0.309326171875
],
[
0.2344970703125,
0.309326171875,
0.765625,
0.309326171875,
0.765625,
0.322509765625,
0.2344970703125,
0.322509765625
],
[
0.2344970703125,
0.322509765625,
0.73779296875,
0.322509765625,
0.73779296875,
0.3359375,
0.2344970703125,
0.3359375
],
[
0.736328125,
0.323974609375,
0.7646484375,
0.323974609375,
0.7646484375,
0.335205078125,
0.736328125,
0.335205078125
],
[
0.2344970703125,
0.337158203125,
0.765625,
0.337158203125,
0.765625,
0.35107421875,
0.2344970703125,
0.35107421875
],
[
0.233642578125,
0.35107421875,
0.765625,
0.349853515625,
0.765625,
0.3642578125,
0.233642578125,
0.365478515625
],
[
0.2344970703125,
0.365478515625,
0.7646484375,
0.365478515625,
0.7646484375,
0.37890625,
0.2344970703125,
0.37890625
],
[
0.233642578125,
0.37890625,
0.765625,
0.37744140625,
0.765625,
0.391357421875,
0.233642578125,
0.392578125
],
[
0.0228729248046875,
0.399658203125,
0.062103271484375,
0.399658203125,
0.062103271484375,
0.48486328125,
0.0228729248046875,
0.48486328125
],
[
0.2344970703125,
0.39453125,
0.7646484375,
0.39453125,
0.7646484375,
0.40478515625,
0.2344970703125,
0.40478515625
],
[
0.2344970703125,
0.406005859375,
0.461669921875,
0.406005859375,
0.461669921875,
0.419189453125,
0.2344970703125,
0.419189453125
],
[
0.453369140625,
0.406005859375,
0.7646484375,
0.406005859375,
0.7646484375,
0.419189453125,
0.453369140625,
0.419189453125
],
[
0.2344970703125,
0.419189453125,
0.765625,
0.419189453125,
0.765625,
0.43310546875,
0.2344970703125,
0.43310546875
],
[
0.233642578125,
0.43310546875,
0.765625,
0.434326171875,
0.765625,
0.447509765625,
0.233642578125,
0.4462890625
],
[
0.2344970703125,
0.447509765625,
0.765625,
0.447509765625,
0.765625,
0.4609375,
0.2344970703125,
0.4609375
],
[
0.236083984375,
0.46337890625,
0.7646484375,
0.46337890625,
0.7646484375,
0.473388671875,
0.236083984375,
0.473388671875
],
[
0.233642578125,
0.473388671875,
0.765625,
0.47607421875,
0.765625,
0.4892578125,
0.233642578125,
0.4873046875
],
[
0.0228729248046875,
0.48486328125,
0.05963134765625,
0.48486328125,
0.05963134765625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.2344970703125,
0.4892578125,
0.765625,
0.4892578125,
0.765625,
0.50244140625,
0.2344970703125,
0.50244140625
],
[
0.2344970703125,
0.50244140625,
0.76318359375,
0.50244140625,
0.76318359375,
0.5166015625,
0.2344970703125,
0.5166015625
],
[
0.2344970703125,
0.517578125,
0.765625,
0.517578125,
0.765625,
0.53076171875,
0.2344970703125,
0.53076171875
],
[
0.2344970703125,
0.53076171875,
0.765625,
0.53076171875,
0.765625,
0.54443359375,
0.2344970703125,
0.54443359375
],
[
0.233642578125,
0.54443359375,
0.765625,
0.54443359375,
0.765625,
0.55810546875,
0.233642578125,
0.55810546875
],
[
0.2344970703125,
0.5595703125,
0.765625,
0.5595703125,
0.765625,
0.57275390625,
0.2344970703125,
0.57275390625
],
[
0.2320556640625,
0.5712890625,
0.468017578125,
0.57080078125,
0.468017578125,
0.58544921875,
0.2320556640625,
0.5859375
],
[
0.178955078125,
0.61572265625,
0.1846923828125,
0.61572265625,
0.1846923828125,
0.62060546875,
0.178955078125,
0.62060546875
],
[
0.1968994140625,
0.611328125,
0.338134765625,
0.611328125,
0.338134765625,
0.625,
0.1968994140625,
0.625
],
[
0.1748046875,
0.64013671875,
0.82421875,
0.64013671875,
0.82421875,
0.65478515625,
0.1748046875,
0.65478515625
],
[
0.1748046875,
0.65576171875,
0.82275390625,
0.65576171875,
0.82275390625,
0.66943359375,
0.1748046875,
0.66943359375
],
[
0.1748046875,
0.66943359375,
0.82177734375,
0.66943359375,
0.82177734375,
0.68310546875,
0.1748046875,
0.68310546875
],
[
0.1748046875,
0.6845703125,
0.82177734375,
0.6845703125,
0.82177734375,
0.69775390625,
0.1748046875,
0.69775390625
],
[
0.1732177734375,
0.6962890625,
0.82275390625,
0.6962890625,
0.82275390625,
0.71044921875,
0.1732177734375,
0.71044921875
],
[
0.1732177734375,
0.7109375,
0.82275390625,
0.7109375,
0.82275390625,
0.724609375,
0.1732177734375,
0.724609375
],
[
0.1748046875,
0.72607421875,
0.82421875,
0.72607421875,
0.82421875,
0.7392578125,
0.1748046875,
0.7392578125
],
[
0.1748046875,
0.73779296875,
0.82421875,
0.73779296875,
0.82421875,
0.75244140625,
0.1748046875,
0.75244140625
],
[
0.1748046875,
0.75244140625,
0.73779296875,
0.75244140625,
0.73779296875,
0.7666015625,
0.1748046875,
0.7666015625
],
[
0.1748046875,
0.7734375,
0.82275390625,
0.7734375,
0.82275390625,
0.78857421875,
0.1748046875,
0.78857421875
],
[
0.176513671875,
0.78857421875,
0.25732421875,
0.78857421875,
0.25732421875,
0.7998046875,
0.176513671875,
0.7998046875
],
[
0.2132568359375,
0.8125,
0.82275390625,
0.8125,
0.82275390625,
0.8271484375,
0.2132568359375,
0.8271484375
],
[
0.233642578125,
0.82763671875,
0.5400390625,
0.82763671875,
0.5400390625,
0.84130859375,
0.233642578125,
0.84130859375
],
[
0.2115478515625,
0.8447265625,
0.82275390625,
0.8447265625,
0.82275390625,
0.85791015625,
0.2115478515625,
0.85791015625
],
[
0.233642578125,
0.859375,
0.5087890625,
0.859375,
0.5087890625,
0.87255859375,
0.233642578125,
0.87255859375
],
[
0.2115478515625,
0.875,
0.82177734375,
0.87744140625,
0.82177734375,
0.892578125,
0.2115478515625,
0.8896484375
],
[
0.233642578125,
0.89013671875,
0.281005859375,
0.89013671875,
0.281005859375,
0.90478515625,
0.233642578125,
0.90478515625
],
[
0.19287109375,
0.91015625,
0.73876953125,
0.9111328125,
0.73876953125,
0.92529296875,
0.19287109375,
0.9248046875
],
[
0.494384765625,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"What's the Magic Word? A CoNTROL THEORY OF LLM",
"PROMPTING",
"Aman Bhargava', Cameron Witkowski?, Manav Shah?, Matt Thomson' *",
"1California Institute of Technology, 2University of Toronto",
"[abhargav,mthomson]@caltech.edu",
"(cameron.witkowski,manav.shah)@mail.utoronto.ca",
"ABSTRACT",
"e ",
"Prompt engineering is crucial for deploying LLMs but is poorly understood math-",
"ematically. We formalize LLM systems as a class of discrete stochastic dynamical",
"systems to explore prompt engineering through the lens of control theory.",
"We",
"investigate the reachable set of output token sequences R(xo) for which there",
"exists a control input sequence u for each y E Ry (xo) that steers the LLM to out-",
"put y from initial state sequence Xo. We offer analytic analysis on the limitations",
"on the controllability of self-attention in terms of reachable set, where we prove an",
"['S]",
"upperboundonthereachablesetofoutputsRu(xo)asafunctionofthesingular",
"values of the parameter matrices.",
":Wepresent complementary empirical analy-",
"sis on the controllability of a panel of LLMs, including Falcon-7b, Llama-7b, and",
"Falcon-4Ob. Our results demonstrate a lower bound on the reachable set of outputs",
"Ry (xo) w.r.t. initial state sequences xo sampled from the Wikitext dataset. We",
"findthatthecorrectnextWikitexttokenfollowingsequenceXoisreachableover",
"97%of thetimewithprompts of k≤10tokens.Wealsoestablish that thetop",
"",
"75 most likely next tokens, as estimated by the LLM itself, are reachable at least",
"85% of the time with prompts of k ≤ 10 tokens. Intriguingly, short prompt se-",
"quences can dramatically alter the likelihood of specific outputs, even making the",
"least likely tokens become the most likely ones. This control-centric analysis of",
"LLMs demonstrates the significant and poorly understood role of input sequences",
"in steering output probabilities, offering a foundational perspective for enhancing",
"language model system capabilities",
"1",
"INTRODUCTION",
"LLMs pre-trained on unsupervised next token prediction objectives exhibit unprecedented dynamic",
"reprogrammability achieved through prompting\", often referred to as zero-shot learning (Brown",
"et al., 2020; Wei et al., 2022; Hagendorff, 2023; Noever & McKee, 2023; OpenAI, 2023; 2022)",
"These capabilities appear to emerge as the model's size, training data, and training time are scaled.",
"The dynamic reprogrammability of LLMs is akin to the adaptable computational capacities observed",
"in biological systems. This feature finds applications across domains such as machine translation",
"(Wang et al., 2023a), code generation (Roziere et al., 2023), and chatbots Bai et al. (2022).A",
"rigorous understanding of the prompt's influence over LLM generation would be of great utility for",
"understanding LLMs and building more robust and capable systems leveraging LLMs.",
"Strategies for controlling pre-trained LLM generation today fall into three broad categories (Zhang",
"etal.,2022)",
"1. Input Optimization (Prompting):Adjusting the input tokens (e.g., rewording the",
"prompt) to improve subsequent text generation",
"2. Model Optimization: Adjusting the weights of the network (e.g., fine-tuning, RLHF) to",
"improve model behavior during inference.",
"3. Post-processing:Adjusting or re-ranking generated text (e.g., surrogate ranking algo-",
"rithm).",
"*Codewill bemadeavailableat https://github.com/amanb2000/Magic_Words.",
"1"
] |
{
"arxiv_id": "2310.04444",
"language": "en",
"page_idx": 0
}
|
||
2310.04444-en-1
|
2310.04444-en-1.jpg
|
Of all these approaches, input optimization (i.e., prompting) is the least invasive and lowest-cost method-and the least understood. Prompt optimization is also deeply connected to the zero-shot capabilities of LLMs-the mysterious emergent capabilities of LLMs such as problem-solving, knowledge retrieval, reasoning, and apparent general intelligence (Bubeck et al., 2023). With such a view, we seek to characterize the controllability of LLMs via prompting.
# 1.1 CONTRIBUTION
We formalize LLM systems in the mathematical framework of control theory in Section 3. Our analysis focuses on the reachable set of outputs $\mathcal{R}_y(\mathbf{x}_0)$ for an LLM system. The reachable set is a fundamental concept in control theory that underlies notions of controllability, stability, and observability (cf. Appendix A). The reachable output set $R_y(\mathbf{x}_0)$ is the set of output sequences y for which there exists a control input sequence u$^*$ that steers the LLM from initial state $\mathbf{x}_0$ to output $\mathbf{y}($cf. Definitions 3, 11).
Our analytic results in Section 4 prove an upper bound on the contents of the reachable output set for a self-attention head as a function of the singular values of its parameter matrices. Since selfattention is the only component in a transformer block where significant information is exchanged between token representations, this bound provides a foothold for analysis of LLM controllability from the perspective of mechanistic interpretability (e.g, Bricken et al. (2023); Chefer et al. (2021); Conmy et al. (2023)). Moreover, this bound represents a necessary condition for an output to be in the reachable set.
Our empirical results apply state-of-the-art prompt optimization techniques (Section 5.1) to demonstrate a lower bound on the contents of the reachable output set for a panel of LLMs, including Llama-7b (Touvron et al., 2023), Falcon-7b, and Falcon-40b (Almazrouei et al., 2023). Specifically, we sample initial states $\mathbf{x}_0$ from the Wikitext dataset (Merity et al., 2016) and probe the reachable output tokens $y$ under length-constrained control input sequences $\mathbf{u}:|\mathbf{u}|\leq k.$ The length constraint $k$ is highly relevant for optimal control of LLMs, as prompts with fewer tokens require fewer computation and memory resources. We find that the reachable output set contains the“correct”next Wikitext token following $\mathbf{x}_0$ over 97% of the time with prompts of $k\leq10$ tokens. We expand our analysis of the contents of $R_y(\mathbf{x}_0)$ by sampling target output tokens $y$ based on the LLMs initial estimate of output likelihood $P_{LM}(y|\mathbf{x}_{0}).$ We find that the top 75 most likely output tokens $y$ are reachable at least 85% of the time with prompts of $k\leq10$ tokens. Intriguingly, some tokens drawn from the set of least likely outputs are controllable to the most likely output with $k\leq4$ control input tokens. Our results suggest that prior likelihood-based metrics, such as cross-entropy loss, cannot guarantee exclusion from the reachable set, emphasizing the gap in our current understanding of LLM systems and control. Implications of our results and open questions in LLM control theory are further discussed in Section 6.
## 2 RELATED WORK
Much of the work on prompt optimization is concerned with finding prompts that induce higher LLM performance on "fill-in-the-blank" or“cloze” tasks (Taylor, 1953). Ône can frame a range of tasks including knowledge retrieval (Petroni et al., 2019), reasoning (Weston et al., 2016), and sentiment analysis (Wang et al., 2023b) as fill-in-the-blank tasks:
$\cdot$ Knowledge Retrieval:“The Titanic sank in the year[MASK]”(Answer:“1912”)
$•$ Reasoning: “A is taller than B. B is taller than C. Is A taller than C? Answer: [MASK]”
(Answer:“Yes”)
$\cdot$ Sentiment Analysis:“I am sad today. The sentiment of the previous sentence was
$[MASK]”($Answer:“Negative")
Notably, there is some freedom in the bolded“prompt text”that surrounds the question to convert it into a“fill-in-the-blank”task. As it turns out, the prompt tokens have a large effect on LLM performance (Brown et al., 2020; Zhang et al., 2022; Jiang et al., 2020).
Modern prompt optimization algorithms generally consist of two iterated steps: a sampling step where new prompts are generated and a testing step where the utility of the new prompts is evaluated,
2
|
[
[
0.1748046875,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11932373046875,
0.1748046875,
0.11932373046875
],
[
0.176513671875,
0.12060546875,
0.82177734375,
0.12060546875,
0.82177734375,
0.1307373046875,
0.176513671875,
0.1307373046875
],
[
0.1748046875,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.1748046875,
0.146484375
],
[
0.17236328125,
0.1458740234375,
0.82177734375,
0.1458740234375,
0.82177734375,
0.1590576171875,
0.17236328125,
0.1590576171875
],
[
0.17236328125,
0.1590576171875,
0.66015625,
0.160400390625,
0.66015625,
0.1749267578125,
0.17236328125,
0.173583984375
],
[
0.17724609375,
0.1925048828125,
0.2059326171875,
0.1925048828125,
0.2059326171875,
0.203857421875,
0.17724609375,
0.203857421875
],
[
0.2059326171875,
0.1932373046875,
0.324462890625,
0.1932373046875,
0.324462890625,
0.203857421875,
0.2059326171875,
0.203857421875
],
[
0.1732177734375,
0.216552734375,
0.82275390625,
0.2171630859375,
0.82275390625,
0.231689453125,
0.1732177734375,
0.2310791015625
],
[
0.1748046875,
0.23291015625,
0.82275390625,
0.23291015625,
0.82275390625,
0.246826171875,
0.1748046875,
0.246826171875
],
[
0.1732177734375,
0.246826171875,
0.82421875,
0.246826171875,
0.82421875,
0.260009765625,
0.1732177734375,
0.260009765625
],
[
0.1748046875,
0.260009765625,
0.82177734375,
0.260009765625,
0.82177734375,
0.2734375,
0.1748046875,
0.2734375
],
[
0.1732177734375,
0.272705078125,
0.82177734375,
0.272705078125,
0.82177734375,
0.287353515625,
0.1732177734375,
0.287353515625
],
[
0.17236328125,
0.28857421875,
0.340576171875,
0.285888671875,
0.340576171875,
0.2998046875,
0.17236328125,
0.3017578125
],
[
0.1748046875,
0.30810546875,
0.82275390625,
0.30810546875,
0.82275390625,
0.322509765625,
0.1748046875,
0.322509765625
],
[
0.1748046875,
0.322509765625,
0.82275390625,
0.322509765625,
0.82275390625,
0.3359375,
0.1748046875,
0.3359375
],
[
0.176513671875,
0.33837890625,
0.82177734375,
0.33837890625,
0.82177734375,
0.348388671875,
0.176513671875,
0.348388671875
],
[
0.1748046875,
0.35107421875,
0.82275390625,
0.35107421875,
0.82275390625,
0.3642578125,
0.1748046875,
0.3642578125
],
[
0.1748046875,
0.3642578125,
0.82275390625,
0.3642578125,
0.82275390625,
0.37744140625,
0.1748046875,
0.37744140625
],
[
0.1748046875,
0.37744140625,
0.82275390625,
0.37744140625,
0.82275390625,
0.391357421875,
0.1748046875,
0.391357421875
],
[
0.1732177734375,
0.392578125,
0.291748046875,
0.392578125,
0.291748046875,
0.406005859375,
0.1732177734375,
0.406005859375
],
[
0.1732177734375,
0.410888671875,
0.82275390625,
0.410888671875,
0.82275390625,
0.4287109375,
0.1732177734375,
0.4287109375
],
[
0.176513671875,
0.4287109375,
0.82177734375,
0.4287109375,
0.82177734375,
0.439453125,
0.176513671875,
0.439453125
],
[
0.1748046875,
0.43994140625,
0.82177734375,
0.43994140625,
0.82177734375,
0.455078125,
0.1748046875,
0.455078125
],
[
0.176513671875,
0.455810546875,
0.82177734375,
0.455810546875,
0.82177734375,
0.466552734375,
0.176513671875,
0.466552734375
],
[
0.1748046875,
0.468505859375,
0.82275390625,
0.468505859375,
0.82275390625,
0.481689453125,
0.1748046875,
0.481689453125
],
[
0.1732177734375,
0.48291015625,
0.82177734375,
0.48291015625,
0.82177734375,
0.496826171875,
0.1732177734375,
0.496826171875
],
[
0.176513671875,
0.49755859375,
0.82275390625,
0.49755859375,
0.82275390625,
0.50830078125,
0.176513671875,
0.50830078125
],
[
0.1748046875,
0.51025390625,
0.82275390625,
0.51025390625,
0.82275390625,
0.5234375,
0.1748046875,
0.5234375
],
[
0.1748046875,
0.5244140625,
0.82421875,
0.5244140625,
0.82421875,
0.53857421875,
0.1748046875,
0.53857421875
],
[
0.1748046875,
0.537109375,
0.82275390625,
0.537109375,
0.82275390625,
0.55029296875,
0.1748046875,
0.55029296875
],
[
0.1748046875,
0.55029296875,
0.82275390625,
0.55029296875,
0.82275390625,
0.564453125,
0.1748046875,
0.564453125
],
[
0.1732177734375,
0.56640625,
0.82275390625,
0.56640625,
0.82275390625,
0.580078125,
0.1732177734375,
0.580078125
],
[
0.1732177734375,
0.580078125,
0.82421875,
0.580078125,
0.82421875,
0.59326171875,
0.1732177734375,
0.59326171875
],
[
0.176513671875,
0.59326171875,
0.82421875,
0.59326171875,
0.82421875,
0.60693359375,
0.176513671875,
0.60693359375
],
[
0.1732177734375,
0.60595703125,
0.82421875,
0.60693359375,
0.82421875,
0.6220703125,
0.1732177734375,
0.62060546875
],
[
0.1732177734375,
0.62060546875,
0.375732421875,
0.62060546875,
0.375732421875,
0.6337890625,
0.1732177734375,
0.6337890625
],
[
0.176513671875,
0.65771484375,
0.191162109375,
0.65771484375,
0.191162109375,
0.66845703125,
0.176513671875,
0.66845703125
],
[
0.200927734375,
0.65478515625,
0.346435546875,
0.65576171875,
0.346435546875,
0.67041015625,
0.200927734375,
0.66943359375
],
[
0.1748046875,
0.6875,
0.82275390625,
0.6875,
0.82275390625,
0.70068359375,
0.1748046875,
0.70068359375
],
[
0.176513671875,
0.70068359375,
0.82275390625,
0.70068359375,
0.82275390625,
0.71533203125,
0.176513671875,
0.71533203125
],
[
0.1748046875,
0.71533203125,
0.82421875,
0.71533203125,
0.82421875,
0.72900390625,
0.1748046875,
0.72900390625
],
[
0.1732177734375,
0.72998046875,
0.60302734375,
0.72998046875,
0.60302734375,
0.74365234375,
0.1732177734375,
0.74365234375
],
[
0.221435546875,
0.7548828125,
0.779296875,
0.7548828125,
0.779296875,
0.76953125,
0.221435546875,
0.76953125
],
[
0.2181396484375,
0.7734375,
0.8203125,
0.7734375,
0.8203125,
0.78857421875,
0.2181396484375,
0.78857421875
],
[
0.233642578125,
0.78857421875,
0.346435546875,
0.78857421875,
0.346435546875,
0.8017578125,
0.233642578125,
0.8017578125
],
[
0.2181396484375,
0.806640625,
0.82177734375,
0.806640625,
0.82177734375,
0.82080078125,
0.2181396484375,
0.82080078125
],
[
0.2303466796875,
0.82080078125,
0.450927734375,
0.8193359375,
0.450927734375,
0.833984375,
0.2303466796875,
0.83544921875
],
[
0.1732177734375,
0.8466796875,
0.82275390625,
0.8466796875,
0.82275390625,
0.861328125,
0.1732177734375,
0.861328125
],
[
0.1732177734375,
0.8623046875,
0.82421875,
0.8623046875,
0.82421875,
0.87548828125,
0.1732177734375,
0.87548828125
],
[
0.1748046875,
0.87548828125,
0.64453125,
0.87548828125,
0.64453125,
0.8896484375,
0.1748046875,
0.8896484375
],
[
0.1748046875,
0.896484375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1732177734375,
0.9111328125,
0.82177734375,
0.9111328125,
0.82177734375,
0.9248046875,
0.1732177734375,
0.9248046875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Of all these approaches, input optimization (i.e., prompting) is the least invasive and lowest-cost",
"method-andtheleastunderstood.Prompt optimization isalsodeeply connected tothezero-shot",
"capabilities of LLMs - the mysterious emergent capabilities of LLMs such as problem-solving",
"knowledge retrieval, reasoning, and apparent general intelligence (Bubeck et al.,2023). With such",
"a view, we seek to characterize the controllability of LLMs via prompting.",
"1.1",
"CONTRIBUTION",
"We formalize LLM systems in themathematical framework of control theory in Section 3. Our",
"analysis focuses on the reachable set of outputs Ry(xo) for an LLM system. The reachable set",
"is a fundamental concept in control theory that underlies notions of controllability, stability, and",
"observability (cf. Appendix A). The reachable output set Ru (xo) is the set of output sequences y",
"for which there exists a control input sequence u* that steers the LLM from initial state xo to outpu",
"y (cf.Definitions 3, 11).",
"Our analytic results in Section 4 prove an upper bound on the contents of the reachable output set",
"for a self-attention head as a function of the singular values of its parameter matrices. Since self-",
"attentionistheonlycomponentinatransformerblockwhere significantinformationisexchanged",
"between token representations, this bound provides a foothold for analysis of LLM controllability",
"from the perspective of mechanistic interpretability (e.g., Bricken et al. (2023); Chefer et al. (2021)",
"Conmy et al. (2023). Moreover, this bound represents a necessary condition for an output to be in",
"the reachable set.",
"Our empirical results apply state-of-the-art prompt optimization techniques (Section 5.1) to demon-",
"strate alower bound on the contents of the reachable output set for a panel of LLMs,including",
"Llama-7b (Touvron et al., 2023), Falcon-7b, and Falcon-40b (Almazrouei et al., 2023). Specifically",
"wesampleinitial statesXofromtheWikitextdataset(Merityetal.,20l6)andprobethereachable",
"output tokens y under length-constrained control input sequences u : ul ≤ k. The length constraint",
"k is highly relevant for optimal control of LLMs, as prompts with fewer tokens require fewer com-",
"putation and memoryresources.Wefind thatthereachable output set contains thecorrect\"next",
"Wikitext token following Xo over 97% of the time with prompts of k ≤ 10 tokens. We expand our",
"analysis of the contents of Ry(xo) by sampling target output tokens y based on the LLMs initial",
"estimate of output likelihood PLM(y|xo). We find that the top 75 most likely output tokens y are",
"reachable at least 85% of the time with prompts of k < 10 tokens. Intriguingly, some tokens drawn",
"from the set of least likely outputs are controllable to the most likely output with k ≤ 4 control input",
"tokens. Our results suggest that prior likelihood-based metrics, such as cross-entropy loss, cannot",
"guarantee exclusion from the reachable set, emphasizing the gap in our current understanding of",
"LLM systems and control. Implications of our results and open questions in LLM control theory are",
"further discussed in Section 6.",
"2",
"RELATED WORK",
"Much of the work on prompt optimization is concerned with finding prompts that induce higher",
"LLM performance on “fill-in-the-blank\" or “cloze\" tasks (Taylor, 1953). One can frame a range",
"of tasks including knowledge retrieval (Petroni et al., 2019), reasoning (Weston et al., 2016), and",
"sentiment analysis (Wang et al., 2023b) as fill-in-the-blank tasks:",
": Knowledge Retrieval:\"The Titanic sank in the year [MASK]\" (Answer: \"1912\")",
". Reasoning:\"A is taller than B. B is taller than C. Is A taller than C? Answer: [MASK]'",
"(Answer: \"Yes\")",
". Sentiment Analysis:\"I am sad today. The sentiment of the previous sentence was",
"[MASK]\"(Answer:\"Negative\")",
"Notably, there is some freedom in the bolded “prompt text\" that surrounds the question to convert",
"it into a \"fill-in-the-blank\" task. As it turns out, the prompt tokens have a large effect on LLM",
"performance (Brown et al., 2020; Zhang et al., 2022; Jiang et al., 2020).",
"Modern prompt optimization algorithms generally consist of two iterated steps: a sampling step",
"where new prompts are generated and a testing step where the utility of the new prompts is evaluated",
"2"
] |
{
"arxiv_id": "2310.04444",
"language": "en",
"page_idx": 1
}
|
||
2310.04444-en-2
|
2310.04444-en-2.jpg
|
and the best are selected for the next iteration. Algorithms primarily differ in the sampling procedure, where various heuristics may be used to pick high-value swaps (Wen et al., 2023; Zhou et al., 2023; Reynolds& McDonell, 2021). Overall, AutoPrompt and its derivative algorithms have been the most numerically successful prompt optimization methods, with the greedy coordinate gradient (GCG) algorithm having state-of-the-art performance (Zou et al., 2023).
The AutoPrompt Family: AutoPrompt (Shin et al., 2020) pioneered the current wave of prompt optimization. Shin et al propose a prompt optimization technique and demonstrate its effectiveness for engineering prompts to improve LLM performance on knowledge and sentiment analysis tasks. At its core, the AutoPrompt algorithm leverages gradient information at the token embedding layer to inform iterative token exchanges within the prompt. This method was extended in Zou et al. (2023) as the greedy coordinate gradient (GCG) algorithm. Taking inspiration from adversarial examples (Goodfellow et al., 2015), Zou et al applied this AutoPrompt variant to generate “jailbreak” prompts that cause aligned LLMs to generate objectionable content.
Other Prompt Optimization Methods: Other investigations on LLMs as prompt optimizers (Zhou et al., 2023) and further analysis of manual prompt optimization (Reynolds&McDonell, 2021) are informative but do not exceed the AutoPrompt family's performance. Some other methods include GBDA (Guo et al., 2021), an approach based on the Gumbel-Softmax reparametrization, the PEZ algorithm (Wen et al., 2023), which directly optimizes embeddings via gradient information, and FluentPrompt (Shi et al., 2022), which differs from AutoPrompt by incorporating Langevin dynamics. Despite the variety of alternatives, GCG retains state-of-the-art performance.
Control Theory for LLMs: To our knowledge, the only other work to date on the controllability of LLMs is Soatto et al.(2023). Soatto et al analyze the controllability of LLMs in terms of “meaningful sentences", defined as the sigma-algebra generated by snippets of text written on the Internet. Their empirical analysis revolves around demonstrating that LLMs are capable of attributing meaning. The theoretical analysis of LLM controllability is limited to“meaningful sentences”, eliminating the possibility of out-of-distribution inputs and outputs. These restrictions render their results challenging to leverage toward a practical understanding of LLM controllability. We situate our work as a practically oriented exploration of LLM controllability. Motivated by challenges in developing LLM systems, we do not eliminate“meaningless sentences”from the state space or input space. We aim to establish a rigorous, general framework for understanding LLM systems and controllability that is amenable to the development of theory and practical engineering insights on systems design.
3 CONTROL THEORY FOR LLMS
Control theory originates from the study of automatic control systems in engineering. It seeks to understand how a“plant”system can be influenced toward a desired state using a“control signal”- often in the presence of disturbances and uncertainty.
Control theory is central to a variety of engineering problems, from electrical engineering to autopilot to telecommunications to manufacturing. Surprisingly, control theory has also been highly applicable to a diverse range of scientific disciplines. Analyzing systems through the lens of controllability has proven fruitful for generating insight into biological systems such as cell signaling pathways and neural networks (Yi et al., 2000), the economics of central banking (Anița et al., 2011), and controlling the spread of infectious diseases (Roy et al., 2009). One of the central benefits of studying systems via controllability is that a range of questions and problems naturally emerge from the framing: when is control possible? What is the cost of control? How computationally intensive $iscontrol?$ These questions are both practically useful and often lead to fundamental insights about the nature of the system in question.
To develop a control theory of LLMs, we provide fundamental abstract definitions of systems and control (Appendix A). We apply them to define LLM systems and outline specific canonical control concepts and problems such as controllability and reachability that arise naturally for LLM systems. Language Model Notation: We denote a causal language model using $P_{LM}.P_{LM}$ maps from an ordered list of tokens from a vocabulary set $\mathcal{V}($e.g., $\mathbf{x}\in\mathcal{V}^n)$ to the probability distribution over
3
|
[
[
0.1732177734375,
0.10418701171875,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11932373046875,
0.1732177734375,
0.11871337890625
],
[
0.1748046875,
0.11871337890625,
0.82275390625,
0.11871337890625,
0.82275390625,
0.1319580078125,
0.1748046875,
0.1319580078125
],
[
0.1748046875,
0.1331787109375,
0.82275390625,
0.1331787109375,
0.82275390625,
0.146484375,
0.1748046875,
0.146484375
],
[
0.1748046875,
0.146484375,
0.82275390625,
0.146484375,
0.82275390625,
0.160400390625,
0.1748046875,
0.160400390625
],
[
0.1732177734375,
0.160400390625,
0.59619140625,
0.160400390625,
0.59619140625,
0.173583984375,
0.1732177734375,
0.173583984375
],
[
0.1748046875,
0.189453125,
0.35693359375,
0.189453125,
0.35693359375,
0.20263671875,
0.1748046875,
0.20263671875
],
[
0.35693359375,
0.1881103515625,
0.82275390625,
0.189453125,
0.82275390625,
0.203857421875,
0.35693359375,
0.20263671875
],
[
0.176513671875,
0.2052001953125,
0.82275390625,
0.2052001953125,
0.82275390625,
0.21533203125,
0.176513671875,
0.21533203125
],
[
0.1748046875,
0.2171630859375,
0.82177734375,
0.2171630859375,
0.82177734375,
0.2310791015625,
0.1748046875,
0.2310791015625
],
[
0.1748046875,
0.2310791015625,
0.82177734375,
0.2310791015625,
0.82177734375,
0.2442626953125,
0.1748046875,
0.2442626953125
],
[
0.176513671875,
0.246826171875,
0.8203125,
0.246826171875,
0.8203125,
0.2568359375,
0.176513671875,
0.2568359375
],
[
0.1748046875,
0.2587890625,
0.82177734375,
0.2587890625,
0.82177734375,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1748046875,
0.272705078125,
0.82177734375,
0.272705078125,
0.82177734375,
0.285888671875,
0.1748046875,
0.285888671875
],
[
0.1748046875,
0.285888671875,
0.62109375,
0.285888671875,
0.62109375,
0.300537109375,
0.1748046875,
0.300537109375
],
[
0.1748046875,
0.31494140625,
0.455810546875,
0.31494140625,
0.455810546875,
0.328857421875,
0.1748046875,
0.328857421875
],
[
0.4599609375,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.323974609375,
0.4599609375,
0.323974609375
],
[
0.468017578125,
0.31494140625,
0.82177734375,
0.31494140625,
0.82177734375,
0.328857421875,
0.468017578125,
0.328857421875
],
[
0.1748046875,
0.330078125,
0.82275390625,
0.330078125,
0.82275390625,
0.343505859375,
0.1748046875,
0.343505859375
],
[
0.1732177734375,
0.34228515625,
0.82275390625,
0.34228515625,
0.82275390625,
0.35595703125,
0.1732177734375,
0.35595703125
],
[
0.176513671875,
0.359130859375,
0.82275390625,
0.359130859375,
0.82275390625,
0.369384765625,
0.176513671875,
0.369384765625
],
[
0.1748046875,
0.371826171875,
0.82275390625,
0.371826171875,
0.82275390625,
0.385009765625,
0.1748046875,
0.385009765625
],
[
0.1732177734375,
0.38330078125,
0.82275390625,
0.38330078125,
0.82275390625,
0.397705078125,
0.1732177734375,
0.397705078125
],
[
0.1748046875,
0.40087890625,
0.74853515625,
0.40087890625,
0.74853515625,
0.410888671875,
0.1748046875,
0.410888671875
],
[
0.1748046875,
0.4267578125,
0.369384765625,
0.4267578125,
0.369384765625,
0.43994140625,
0.1748046875,
0.43994140625
],
[
0.369384765625,
0.4267578125,
0.82177734375,
0.4267578125,
0.82177734375,
0.43994140625,
0.369384765625,
0.43994140625
],
[
0.1748046875,
0.442626953125,
0.82421875,
0.442626953125,
0.82421875,
0.455810546875,
0.1748046875,
0.455810546875
],
[
0.1748046875,
0.455810546875,
0.82275390625,
0.455810546875,
0.82275390625,
0.4697265625,
0.1748046875,
0.4697265625
],
[
0.1748046875,
0.4697265625,
0.82177734375,
0.4697265625,
0.82177734375,
0.48291015625,
0.1748046875,
0.48291015625
],
[
0.176513671875,
0.48486328125,
0.81884765625,
0.48486328125,
0.81884765625,
0.49560546875,
0.176513671875,
0.49560546875
],
[
0.1732177734375,
0.496826171875,
0.82177734375,
0.496826171875,
0.82177734375,
0.51025390625,
0.1732177734375,
0.51025390625
],
[
0.17236328125,
0.51025390625,
0.82275390625,
0.5087890625,
0.82275390625,
0.5234375,
0.17236328125,
0.5244140625
],
[
0.1732177734375,
0.5244140625,
0.82177734375,
0.5244140625,
0.82177734375,
0.53857421875,
0.1732177734375,
0.53857421875
],
[
0.1732177734375,
0.53857421875,
0.82177734375,
0.53857421875,
0.82177734375,
0.5517578125,
0.1732177734375,
0.5517578125
],
[
0.1748046875,
0.5517578125,
0.82275390625,
0.5517578125,
0.82275390625,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.1748046875,
0.56640625,
0.82421875,
0.56640625,
0.82421875,
0.580078125,
0.1748046875,
0.580078125
],
[
0.1748046875,
0.58203125,
0.279296875,
0.58203125,
0.279296875,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.17236328125,
0.61279296875,
0.461669921875,
0.6142578125,
0.461669921875,
0.62890625,
0.17236328125,
0.62744140625
],
[
0.1748046875,
0.64453125,
0.82275390625,
0.64453125,
0.82275390625,
0.65771484375,
0.1748046875,
0.65771484375
],
[
0.1748046875,
0.6591796875,
0.8203125,
0.6591796875,
0.8203125,
0.67236328125,
0.1748046875,
0.67236328125
],
[
0.1732177734375,
0.673828125,
0.52392578125,
0.673828125,
0.52392578125,
0.6875,
0.1732177734375,
0.6875
],
[
0.1748046875,
0.693359375,
0.82177734375,
0.693359375,
0.82177734375,
0.70703125,
0.1748046875,
0.70703125
],
[
0.1732177734375,
0.708984375,
0.82177734375,
0.708984375,
0.82177734375,
0.72265625,
0.1732177734375,
0.72265625
],
[
0.1732177734375,
0.7216796875,
0.82177734375,
0.7216796875,
0.82177734375,
0.73486328125,
0.1732177734375,
0.73486328125
],
[
0.1748046875,
0.73486328125,
0.82275390625,
0.73486328125,
0.82275390625,
0.74853515625,
0.1748046875,
0.74853515625
],
[
0.1732177734375,
0.74853515625,
0.82177734375,
0.74755859375,
0.82177734375,
0.76220703125,
0.1732177734375,
0.76318359375
],
[
0.1732177734375,
0.76220703125,
0.82421875,
0.76220703125,
0.82421875,
0.77587890625,
0.1732177734375,
0.77587890625
],
[
0.1748046875,
0.7763671875,
0.82275390625,
0.7763671875,
0.82275390625,
0.79052734375,
0.1748046875,
0.79052734375
],
[
0.1748046875,
0.79150390625,
0.82275390625,
0.79150390625,
0.82275390625,
0.8046875,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.8046875,
0.82421875,
0.8046875,
0.82421875,
0.818359375,
0.1748046875,
0.818359375
],
[
0.1732177734375,
0.818359375,
0.41259765625,
0.818359375,
0.41259765625,
0.83203125,
0.1732177734375,
0.83203125
],
[
0.1748046875,
0.8388671875,
0.82275390625,
0.8388671875,
0.82275390625,
0.85302734375,
0.1748046875,
0.85302734375
],
[
0.1748046875,
0.85400390625,
0.82275390625,
0.85400390625,
0.82275390625,
0.86865234375,
0.1748046875,
0.86865234375
],
[
0.1732177734375,
0.86865234375,
0.8203125,
0.86865234375,
0.8203125,
0.8818359375,
0.1732177734375,
0.8818359375
],
[
0.1732177734375,
0.896484375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.1732177734375,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.9609375,
0.491943359375,
0.9609375
]
] |
[
"and the best are selected for the next iteration. Algorithms primarily differ in the sampling procedure",
"where various heuristics may be used to pick high-value swaps (Wen et al., 2023; Zhou et al., 2023",
"Reynolds & McDonell, 2021). Overall, AutoPrompt and its derivative algorithms have been the most",
"numerically successful prompt optimization methods, with the greedy coordinate gradient (GCG)",
"algorithm having state-of-the-art performance (Zou et al., 2023).",
"The AutoPrompt Family:",
"AutoPrompt (Shin et al.,2020)pioneered the currentwave of prompt",
"optimization.Shinetalproposeapromptoptimizationtechniqueanddemonstrateitseffectiveness",
"for engineering prompts to improve LLM performance on knowledge and sentiment analysis tasks",
"At its core, the AutoPrompt algorithm leverages gradient information at the token embedding layer",
"to inform iterative token exchanges within the prompt.This method was extended in Zou et al",
"(2023) as the greedy coordinate gradient (GCG) algorithm. Taking inspiration from adversarial ex-",
"amples (Goodfellow et al., 2015), Zou et al applied this AutoPrompt variant to generate \"jailbreak\"",
"prompts that cause aligned LLMs to generate objectionable content.",
"Other Prompt Optimization Methods:",
"",
"Other investigations on LLMs as prompt optimizers",
"(Zhou et al., 2023) and further analysis of manual prompt optimization (Reynolds & McDonell,",
"2021) are informative but do not exceed the AutoPrompt family's performance. Some other methods",
"includeGBDA(Guoetal.,2021),anapproachbasedontheGumbel-Softmaxreparametrization,the",
"PEZ algorithm (Wen et al., 2023), which directly optimizes embeddings via gradient information,",
"and FluentPrompt (Shi et al., 2022), which differs from AutoPrompt by incorporating Langevin",
"dynamics.Despitethevarietyofalternatives,GCGretainsstate-of-the-artperformance.",
"Control Theory for LLMs:",
". To our knowledge, the only other work to date on the controllabil-",
"ity of LLMs is Soatto et al. (2023). Soatto et al analyze the controllability of LLMs in terms of",
"\"meaningful sentences\", defined as the sigma-algebra generated by snippets of text written on the",
"Internet. Their empirical analysis revolves around demonstrating that LLMs are capable of attribut",
"ingmeaning.Thetheoretical analysis of LLM controllabilityis limited tomeaningful sentences",
"eliminating the possibility of out-of-distribution inputs and outputs. These restrictions render their",
"results challenging to leverage toward a practical understanding of LLM controllability. We situate",
"our work as a practically oriented exploration of LLM controllability. Motivated by challenges in",
"developing LLM systems, we do not eliminate \"meaningless sentences\" from the state space or in-",
"put space. We aim to establish a rigorous, general framework for understanding LLM systems and",
"controllability that is amenable to the development of theory and practical engineering insights on",
"systems design.",
"3CONTROLTHEORYFORLLMS",
"Control theory originates from the study of automatic control systems in engineering. It seeks to",
"understand how a plant\" system can be influenced toward a desired state using a “control signal\"",
"often in the presence of disturbances and uncertainty.",
"Control theory is central to a variety of engineering problems, from electrical engineering to au-",
"topilot to telecommunications to manufacturing. Surprisingly, control theory has also been highly",
"applicable to a diverse range of scientific disciplines. Analyzing systems through the lens of con-",
"trollability has proven fruitful for generating insight into biological systems such as cell signaling",
"pathways and neural networks (Yi et al.,2000), the economics of central banking (Anita et al.,2011)",
"and controlling the spread of infectious diseases (Roy et al., 2009). One of the central benefits of",
"studying systems via controllability is that a range of questions and problems naturally emerge from",
"the framing: when is control possible? What is the cost of control? How computationally intensive",
"is control?These questions are both practically useful and often lead to fundamental insights about",
"the nature of the system in question.",
"To develop a control theory of LLMs, we provide fundamental abstract definitions of systems and",
"control (Appendix A). We apply them to define LLM systems and outline specific canonical control",
"concepts and problems such as controllability and reachability that arise naturally for LLM systems",
"Language Model Notation:We denote a causal language model using PLM. PLM maps from",
"an ordered list of tokens from a vocabulary set V (e.g., x E Vn) to the probability distribution over",
"3"
] |
{
"arxiv_id": "2310.04444",
"language": "en",
"page_idx": 2
}
|
||
2310.04444-en-3
|
2310.04444-en-3.jpg
|
the next token $P_LM(x_{n+1}|\mathbf{x})\in[0,1]^{|\mathcal{V}|}.$ We use $\mathcal{V}^*$ to denote the set of all possible sequences of any length composed of tokens from $\mathcal{V}.$ The addition operator indicates the concatenation of token sequences. Bolded lowercase variables (e.g., $\mathbf{x}=[x^1,\ldots,x^n])$ denote token sequences while unbolded lowercase variables refer to individual tokens (e.g., $x\in\mathcal{V}).$ The length of a token sequence is denoted $|\mathbf{x}|.$
While LLMs are at times leveraged in a manner that masks the iterative aspects of generation, the reality is that token generation and externally imposed “control input" sequences are generated and processed sequentially, leading to non-trivial system dynamics. Several key differences remain between LLM-based systems and systems typically modeled through ordinary differential equations (ODEs), which have long been a cornerstone in the study of continuous-time dynamical systems:
1. Discrete state and time: LLM systems operate on sequences of discrete tokens over a
discrete time set, in contrast to the continuous state spaces and time sets studied in classical
control theory.
2. Shift-and-Grow State Dynamics: Whereas the system state in an ODE-based system has
a fixed size over time, the system state $\mathbf{x}(t)$ for LLM systems grows as tokens are added to
the state sequence.
3. Mutual exclusion on control input token vs. generated token: The LLM system state
$\mathbf{x}(t)$ is written to one token at a time. The newest token is either drawn from the control input $u(t)$ or is generated by the LLM by sampling $x^\prime\sim P_{LM}(x^{\prime}|\mathbf{x}(t)).$ This differs from traditional discrete stochastic systems, where the control sequence and internal dynamics generally affect the state synchronously.
We begin by rigorously defining LLM systems with user input, drawing from the abstract mathe-
matical defnition of a system (Definition 7).
Definition 1 (LLM System with Control Input). An autoregressive $LLM$ system with control input
$\Sigma = \left ( \mathcal{V} , P_{LM}\right ) \textit{consists of: }$
$\bullet\mathcal{T}=\mathbb{N}:The$ time set is the natural numbers.
$\bullet X=\mathcal{V}^*:The$ state space consists of all possible token sequences of any length drawn
from ${\mathcal{V} }.$ $We$ denote the state $at$ time $t$ $as$ $\mathbf{x} ( t) = [ x^{0}( t) , \ldots , x^{t}( t) ] .$
$\bullet U=\mathcal{V}\cup\varnothing:The$ input takes values from the vocabulary set V or null
$\bullet$ $\phi : \mathcal{X} \times U\times \mathcal{T} ^2\to \mathcal{X} : The$ transition map $is$
$$\phi(\mathbf{x}(t),u(t),t,t+1)=\left\{\begin{matrix}\mathbf{x}(t)+u(t)&if\:u(t)\neq\varnothing\\\mathbf{x}(t)+x'|x'\sim P_{LM}(x'|\mathbf{x}(t))&else\end{matrix}\right.$$
(1)
Note that the general multi-step transition map $\phi(\mathbf{x}(t),u,t,t+N)$ can be achieved by
iterating equation 1 for control sequences u defined over the interval [t,t+N].
$h( \mathbf{x} ( t) ; r) = [ x^{t- r}( t) , \ldots , x^{t}( t) ] : The$ readout map returns the most recent r tokens from $state\textbf{ x}( t) .$
We note that this LLM system definition is generalizable to a variety of LLM augmentation, including chain-of-thought (Wei et al., 2023), retrieval-augmented generation (Lewis et al., 2020), and chatbot interaction. For example, chain-of-thought is equivalent to sampling the readout map $h(x(t),r)$ at time $T>|\mathbf{u}|+|\mathbf{x}_0|+r$ for prompt u and initial state $\mathbf{x}_0.$ A similar formulation may ba papliad ta U Mavata andouad uith aroramatic toola $(a,a,b,b_{\mathrm{otil~at~ol~}}(20))$ be applied to LLM systems endowed with programmatic tools (e.g., Patil et al. (2023)).
In Definition 1,we assume that the control input gets to“decide” whether to yield token generation to the LLM $(u(t)=\varnothing)$ or override the LLM and add some token $u(t)\neq\varnothing$ to the state x$(t).$ This assumption generally holds when building LLM systems, though it may not hold when using the family of $k$-long input sequences $\mathbf{u}\in\mathcal{V}^k-$the value of $\tilde u(\ell):\ell>\tilde{k}$ is implicitly$\varnothing$ unless otherwise stated.
While next token generation $x^\prime\sim P_{LM}(x^{\prime}|\mathbf{x}(t))$ in equation 1 is probabilistic, we may render the system deterministic by sampling with zero temperature (i.e., greedy decoding). The greedy
4
|
[
[
0.17236328125,
0.1029052734375,
0.82421875,
0.10162353515625,
0.82421875,
0.11932373046875,
0.17236328125,
0.12060546875
],
[
0.176513671875,
0.12060546875,
0.82275390625,
0.12060546875,
0.82275390625,
0.1307373046875,
0.176513671875,
0.1307373046875
],
[
0.176513671875,
0.1337890625,
0.82177734375,
0.1337890625,
0.82177734375,
0.14453125,
0.176513671875,
0.14453125
],
[
0.1748046875,
0.146484375,
0.82275390625,
0.146484375,
0.82275390625,
0.1610107421875,
0.1748046875,
0.1610107421875
],
[
0.17236328125,
0.1578369140625,
0.26953125,
0.160400390625,
0.26953125,
0.1749267578125,
0.17236328125,
0.17236328125
],
[
0.1732177734375,
0.179931640625,
0.82177734375,
0.18115234375,
0.82177734375,
0.1956787109375,
0.1732177734375,
0.1944580078125
],
[
0.1748046875,
0.1956787109375,
0.82275390625,
0.1956787109375,
0.82275390625,
0.208984375,
0.1748046875,
0.208984375
],
[
0.1748046875,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.2235107421875,
0.1748046875,
0.2235107421875
],
[
0.1732177734375,
0.2235107421875,
0.82275390625,
0.2235107421875,
0.82275390625,
0.2374267578125,
0.1732177734375,
0.2374267578125
],
[
0.1748046875,
0.2374267578125,
0.80810546875,
0.2374267578125,
0.80810546875,
0.250732421875,
0.1748046875,
0.250732421875
],
[
0.214111328125,
0.261474609375,
0.82177734375,
0.261474609375,
0.82177734375,
0.27587890625,
0.214111328125,
0.27587890625
],
[
0.233642578125,
0.276611328125,
0.82421875,
0.276611328125,
0.82421875,
0.290283203125,
0.233642578125,
0.290283203125
],
[
0.2344970703125,
0.291748046875,
0.330078125,
0.291748046875,
0.330078125,
0.302978515625,
0.2344970703125,
0.302978515625
],
[
0.2132568359375,
0.30810546875,
0.82275390625,
0.30810546875,
0.82275390625,
0.3212890625,
0.2132568359375,
0.3212890625
],
[
0.233642578125,
0.322509765625,
0.826171875,
0.322509765625,
0.826171875,
0.3359375,
0.233642578125,
0.3359375
],
[
0.236083984375,
0.33837890625,
0.359375,
0.33837890625,
0.359375,
0.348388671875,
0.236083984375,
0.348388671875
],
[
0.2132568359375,
0.354248046875,
0.82275390625,
0.354248046875,
0.82275390625,
0.36865234375,
0.2132568359375,
0.36865234375
],
[
0.2344970703125,
0.36865234375,
0.82275390625,
0.36865234375,
0.82275390625,
0.3818359375,
0.2344970703125,
0.3818359375
],
[
0.233642578125,
0.38330078125,
0.82275390625,
0.38330078125,
0.82275390625,
0.396484375,
0.233642578125,
0.396484375
],
[
0.233642578125,
0.396484375,
0.82421875,
0.396484375,
0.82421875,
0.410400390625,
0.233642578125,
0.410400390625
],
[
0.233642578125,
0.410400390625,
0.496826171875,
0.410400390625,
0.496826171875,
0.423583984375,
0.233642578125,
0.423583984375
],
[
0.1748046875,
0.43505859375,
0.82177734375,
0.43505859375,
0.82177734375,
0.448974609375,
0.1748046875,
0.448974609375
],
[
0.1732177734375,
0.449462890625,
0.464111328125,
0.449462890625,
0.464111328125,
0.46337890625,
0.1732177734375,
0.46337890625
],
[
0.1748046875,
0.46728515625,
0.82421875,
0.46728515625,
0.82421875,
0.48095703125,
0.1748046875,
0.48095703125
],
[
0.1748046875,
0.481689453125,
0.35302734375,
0.481689453125,
0.35302734375,
0.49560546875,
0.1748046875,
0.49560546875
],
[
0.2181396484375,
0.5048828125,
0.52783203125,
0.505859375,
0.52783203125,
0.52099609375,
0.2181396484375,
0.51953125
],
[
0.2197265625,
0.52978515625,
0.826171875,
0.52978515625,
0.826171875,
0.54443359375,
0.2197265625,
0.54443359375
],
[
0.2279052734375,
0.54443359375,
0.6650390625,
0.54443359375,
0.6650390625,
0.5595703125,
0.2279052734375,
0.5595703125
],
[
0.2181396484375,
0.5673828125,
0.69384765625,
0.5673828125,
0.69384765625,
0.58056640625,
0.2181396484375,
0.58056640625
],
[
0.2197265625,
0.59130859375,
0.5400390625,
0.59130859375,
0.5400390625,
0.60595703125,
0.2197265625,
0.60595703125
],
[
0.4599609375,
0.61279296875,
0.54736328125,
0.61279296875,
0.54736328125,
0.62744140625,
0.4599609375,
0.62744140625
],
[
0.6845703125,
0.61279296875,
0.76318359375,
0.61279296875,
0.76318359375,
0.62744140625,
0.6845703125,
0.62744140625
],
[
0.293212890625,
0.6220703125,
0.459228515625,
0.6220703125,
0.459228515625,
0.63623046875,
0.293212890625,
0.63623046875
],
[
0.80224609375,
0.6220703125,
0.826171875,
0.6220703125,
0.826171875,
0.63525390625,
0.80224609375,
0.63525390625
],
[
0.459228515625,
0.62890625,
0.71337890625,
0.62744140625,
0.71337890625,
0.64599609375,
0.459228515625,
0.646484375
],
[
0.233642578125,
0.6513671875,
0.82275390625,
0.6513671875,
0.82275390625,
0.66650390625,
0.233642578125,
0.66650390625
],
[
0.2303466796875,
0.66650390625,
0.7548828125,
0.66552734375,
0.7548828125,
0.6796875,
0.2303466796875,
0.68115234375
],
[
0.2197265625,
0.689453125,
0.82421875,
0.689453125,
0.82421875,
0.7041015625,
0.2197265625,
0.7041015625
],
[
0.2296142578125,
0.7021484375,
0.30224609375,
0.7041015625,
0.30224609375,
0.71826171875,
0.2296142578125,
0.71630859375
],
[
0.176513671875,
0.72900390625,
0.8203125,
0.72900390625,
0.8203125,
0.7421875,
0.176513671875,
0.7421875
],
[
0.1748046875,
0.74365234375,
0.82177734375,
0.74365234375,
0.82177734375,
0.7568359375,
0.1748046875,
0.7568359375
],
[
0.1748046875,
0.7568359375,
0.82275390625,
0.75830078125,
0.82275390625,
0.771484375,
0.1748046875,
0.77099609375
],
[
0.1732177734375,
0.77099609375,
0.82177734375,
0.77099609375,
0.82177734375,
0.7841796875,
0.1732177734375,
0.7841796875
],
[
0.1748046875,
0.78515625,
0.744140625,
0.78515625,
0.744140625,
0.798828125,
0.1748046875,
0.798828125
],
[
0.1748046875,
0.80615234375,
0.82275390625,
0.80615234375,
0.82275390625,
0.8193359375,
0.1748046875,
0.8193359375
],
[
0.1732177734375,
0.8193359375,
0.82275390625,
0.818359375,
0.82275390625,
0.83349609375,
0.1732177734375,
0.833984375
],
[
0.1748046875,
0.833984375,
0.82421875,
0.833984375,
0.82421875,
0.84765625,
0.1748046875,
0.84765625
],
[
0.1732177734375,
0.84765625,
0.82177734375,
0.84765625,
0.82177734375,
0.8623046875,
0.1732177734375,
0.8623046875
],
[
0.1748046875,
0.861328125,
0.826171875,
0.861328125,
0.826171875,
0.87548828125,
0.1748046875,
0.87548828125
],
[
0.1732177734375,
0.87548828125,
0.285888671875,
0.87548828125,
0.285888671875,
0.8896484375,
0.1732177734375,
0.8896484375
],
[
0.176513671875,
0.896484375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"the next token PLM(rn+1|x) e [0, 1]lvl. We use V* to denote the set of all possible sequences",
"of anylength composed oftokensfrom V.The addition operator indicates the concatenation of",
"tokensequences.Boldedlowercasevariables(e.g.,x=[,...,rn)denotetokensequenceswhile",
"unbolded lowercase variables refer to individual tokens (e.g., a e V). The length of a token sequence",
"is denoted x|.",
"While LLMs are at times leveraged in a manner that masks the iterative aspects of generation",
"the reality is that token generation and externally imposed \"control input\" sequences are generated",
"and processed sequentially, leading to non-trivial system dynamics. Several key differences remain",
"between LLM-based systems and systems typically modeled through ordinary differential equations",
"(ODEs), which have long been a cornerstone in the study of continuous-time dynamical systems:",
"1. Discrete state and time: LLM systems operate on sequences of discrete tokens over a",
"discrete time set, in contrast to the continuous state spaces and time sets studied in classical",
"controltheory.",
"2. Shift-and-Grow State Dynamics: Whereas the system state in an ODE-based system has",
"a fixed size over time, the system state x(t) for LLM systems grows as tokens are added to",
"thestatesequence.",
"3. Mutual exclusion on control input token vs. generated token: The LLM system state",
"x(t) is written to one token at a time. The newest token is either drawn from the control",
"input u(t) or is generated by the LLM by sampling r' ~ PLM ('|x(t)). This differs from",
"traditional discrete stochastic systems, where the control sequence and internal dynamics",
"generally affectthe state synchronously.",
"We begin by rigorously defining LLM systems with user input, drawing from the abstract mathe-",
"matical definition of a system (Definition 7)",
"Definition 1 (LLM System with Control Input). An autoregressive LLM system with control input",
" = (V, PLM) consists of:",
".T=N:Thetime set is thenatural numbers.",
". X = V* : The state space consists of all possible token sequences of any length drawn",
"from V. We denote the state at time t as x(t) = [c(t), ..., &t(t)].",
"•U=VU : The input takes values from the vocabulary set V or null",
". Φ : X × U × T2→X : The transition map is",
"[x(t) +u(t)",
"ifu(t)0",
"Φ(x(t), u(t),t,t + 1) =",
"(1)",
"(x(t) +a'|' ~ PLm(r'x(t) else",
"Note that the general multi-step transition map o(x(t), u,t,t + N) can be achieved by",
"iterating equation I for control sequences u defined over the interval [t,t + N].",
": h(x(t);r) =[rt-r(t), ..., rt(t)] : The readout map returns the most recent r tokens from",
"state x(t).",
"We note that this LLM system definition is generalizable to a variety of LLM augmentation, in.",
"cluding chain-of-thought (Wei et al., 2023), retrieval-augmented generation (Lewis et al., 2020)",
"and chatbot interaction.For example, chain-of-thought is equivalent to sampling the readout map",
"h(r(t), r) at time T > |ul + |xol + r for prompt u and initial state xo. A similar formulation may",
"be applied to LLM systems endowed with programmatic tools (e.g., Patil et al. (2023).",
"In Definition l, we assume that the control input gets to decide\" whether to yield token generation",
"to the LLM (u(t) = @) or override the LLM and add some token u(t) ± to the state x(t)",
"This assumption generally holds when building LLM systems, though it may not hold when using",
"existing systems (e.g., via non-streaming API). When discussing finite-length control inputs - e.g.",
"the family of k-long input sequences u e Vk - the value of u(e) : l > k is implicitly unless",
"otherwise stated.",
"While next token generation α' ~ PLM(r'x(t) in equation 1 is probabilistic, we may render",
"the system deterministic by sampling with zero temperature (i.e., greedy decoding). The greedy",
"4"
] |
{
"arxiv_id": "2310.04444",
"language": "en",
"page_idx": 3
}
|
||
2310.04444-en-4
|
2310.04444-en-4.jpg
|
decoding assumption provides a foothold to analyze the reachable sets and controllability of LLM systems without invoking notions of stochastic control as in Sivaramakrishnan et al. (2023); Soatto et al. (2023). Moreover, it remains connected to temperature-based stochastic decoding strategies as a limiting case of temperature-based sampling as zero-temperature sampling.
We now extend Definition 10 to define output controllability for LLM systems:
Definition 2 (LLM Output Reachability). Output token sequence $\mathbf{y}\in\mathcal{V}^r$ is reachable from initial $state\mathbf{x} _{0}\in \mathcal{V} ^{* }for$ $LLM$ system $\Sigma ( \mathcal{V} , P_{LM})$ iff there exists some time $T$ and input $\mathbf{u} ^{* }\in U^{T}$ that steers the LLM from initial state x$_0\textit{to output y}= h( \mathbf{x} ( T) , r) \textit{at time T}.$
We disregard the trivial solution wherein the control input $\mathbf{u}^*(t)$ overrides the LLM to force the state
sequence to take on the desired output value y.
The reachable output set definition for LLM systems follows from Definition 11:
Definition 3 (LLM Reachable Output Set). The reachable output set from initial state $\mathbf{x}_0\in\mathcal{V}^*for$ $LLM$ system $\Sigma=(\mathcal{V},P_{LM})$ is denoted $R_y(\mathbf{x}_0)$ and consists of all reachable outputs $\mathbf{y}\in\mathcal{V}^*$ from initial state x$_0.$
Output controllability for LLMs follows from Definition 13.
$\textbf{Definition 4 ( LLM Output Controllability) . An LLM system }\Sigma = ( \mathcal{V} , P_{LM})$ is output controllable
$iff,for$ every initial state $\mathbf{x}_0\in\mathcal{V}^*,the$ reachable output set $\mathcal{R}_y(\mathbf{x}_0)=\mathcal{V}^*.$
The turn-based nature of writing to the LLM state sequence $\mathbf{x}(t)$ invites the question of whether the prompt u should preempt the imposed state $\mathbf{x}_0$ or come after the state $^{\mathrm{i}}$ We focus our efforts on cases where u comes before imposed state sequence $\mathbf{x}_0$ due to its importance for developing system prompts and controlling text completion-based generation where the desired output is $\mathbf{x}_0+\mathbf{y}^*$ for some desired continuation $\mathbf{y}^*$ of partial string $\mathbf{x}_0.$ Due to the costly nature of long prompts, we are especially interested in the existence of prompts u* with minimal length $|\mathbf{u}^*|.$
Definitions 3 and 4 form the basis for our control theory of LLMs. While amenable to analytic analysis as in Section 4 and Soatto et al. (2023), empirical analysis of the reachable set and controllability is challenging due to the intractable size of $\hat{\mathcal{V}}^*.$ We propose the following statistical measure of controllability for practically assessing the controllability of an LLM system w.r.t. a dataset $\mathcal{D}$ under prompt length constraint $|\mathbf{u}|\leq k{:}$
$\textbf{Definition 5}( k- \epsilon$ Controllability). Consider a dataset of state-output pairs $\mathcal{D}=\{(\mathbf{x}_0^i,\mathbf{y}^i)\}_{i\in[N]}.$ $An$ $LLM$ $\Sigma =$ $( \mathcal{V} , P_{LM})$ $is$ $k- \epsilon$ controllable $w. r. t.$ $\mathcal{D}$ $if, for$ $at$ least $a$ proportion $of$ $( 1- \epsilon )$ $of$ $(\mathbf{x}_0^i,\mathbf{y}^i)\in\mathcal{D},thefollowing$ condition holds:
$$\mathbf{y}^i\in\mathcal{R}_y^k(\mathbf{x}_0^i)$$
(2)
Where $\mathcal{R} _y^k( \mathbf{x} _0^i)$ is the reachable set of outputs as in Definition 3 under the constraint that prompts u
must have length $|\mathbf{u}|\leq k.$
Our empirical work in Section 5.2 explores $k-\epsilon$ controllability w.r.t. initial states $\mathbf{x}_0$ sampled from the Wikitext dataset. While empirical analysis of LLM controllability is challenging due to the lack of apparent structure in LLM dynamics and the combinatorially large state space, we may still experimentally establish the existence of optimal prompts $\mathbf{u}^*$ that elicit a given output, and thus establish a lower bound on the content of the reachable set. Meanwhile, our theoretical work in Section 4 establishes upper bounds on the content of the reachable set for self-attention. We hope these complementary approaches aid in unifying our understanding of LLM systems.
4 МАТНЕМАТIСАL АNALYSIS ON THE СОNTROLLABILITY OF
## SELF- ATTENTION
Self-attention is a central component in modern transformer-based language models (Brown et al., 2020; Touvron et al., 2023; Radford et al., 2019; Min et al., 2023). Introduced in Vaswani et al.
$^{1}$Both situations are reasonable in developing LLM systems: u preceding $\mathbf{x}_0$ may arise when prompting an LLM to complete a partial string $\mathbf{x}_0.\mathbf{u}$ proceeding $\mathbf{x}_0$ may arise when prompting an LLM in the presence of an imposed system prompt $\mathbf{x}_0.$ Therefore, how an initial state $\mathbf{x}_0$ is interleaved with control input u is largely a design decision.
5
|
[
[
0.1732177734375,
0.10479736328125,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11871337890625,
0.1732177734375,
0.11871337890625
],
[
0.176513671875,
0.12060546875,
0.82177734375,
0.12060546875,
0.82177734375,
0.1307373046875,
0.176513671875,
0.1307373046875
],
[
0.1748046875,
0.1331787109375,
0.82421875,
0.1331787109375,
0.82421875,
0.146484375,
0.1748046875,
0.146484375
],
[
0.17236328125,
0.1458740234375,
0.6787109375,
0.146484375,
0.6787109375,
0.1610107421875,
0.17236328125,
0.160400390625
],
[
0.1748046875,
0.1673583984375,
0.69384765625,
0.1673583984375,
0.69384765625,
0.18115234375,
0.1748046875,
0.18115234375
],
[
0.1748046875,
0.184326171875,
0.82421875,
0.184326171875,
0.82421875,
0.1988525390625,
0.1748046875,
0.1988525390625
],
[
0.1748046875,
0.1988525390625,
0.82421875,
0.1988525390625,
0.82421875,
0.21337890625,
0.1748046875,
0.21337890625
],
[
0.1748046875,
0.21337890625,
0.64892578125,
0.21337890625,
0.64892578125,
0.2266845703125,
0.1748046875,
0.2266845703125
],
[
0.1732177734375,
0.23486328125,
0.82275390625,
0.236083984375,
0.82275390625,
0.250732421875,
0.1732177734375,
0.25
],
[
0.17236328125,
0.251953125,
0.48193359375,
0.251953125,
0.48193359375,
0.26513671875,
0.17236328125,
0.26513671875
],
[
0.1748046875,
0.272705078125,
0.70361328125,
0.272705078125,
0.70361328125,
0.285888671875,
0.1748046875,
0.285888671875
],
[
0.1748046875,
0.2890625,
0.82421875,
0.2890625,
0.82421875,
0.302978515625,
0.1748046875,
0.302978515625
],
[
0.1748046875,
0.3037109375,
0.82421875,
0.3037109375,
0.82421875,
0.318115234375,
0.1748046875,
0.318115234375
],
[
0.17236328125,
0.31640625,
0.275390625,
0.317626953125,
0.275390625,
0.33203125,
0.17236328125,
0.330810546875
],
[
0.1732177734375,
0.341552734375,
0.56787109375,
0.34033203125,
0.56787109375,
0.354736328125,
0.1732177734375,
0.35595703125
],
[
0.1748046875,
0.359130859375,
0.82421875,
0.359130859375,
0.82421875,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.1748046875,
0.373779296875,
0.654296875,
0.373779296875,
0.654296875,
0.386962890625,
0.1748046875,
0.386962890625
],
[
0.1732177734375,
0.397705078125,
0.82421875,
0.397705078125,
0.82421875,
0.410888671875,
0.1732177734375,
0.410888671875
],
[
0.1732177734375,
0.412353515625,
0.82275390625,
0.410888671875,
0.82275390625,
0.4248046875,
0.1732177734375,
0.425537109375
],
[
0.1748046875,
0.425537109375,
0.82275390625,
0.425537109375,
0.82275390625,
0.439453125,
0.1748046875,
0.439453125
],
[
0.1732177734375,
0.438232421875,
0.82421875,
0.439453125,
0.82421875,
0.453857421875,
0.1732177734375,
0.45263671875
],
[
0.1732177734375,
0.451904296875,
0.82421875,
0.453857421875,
0.82421875,
0.468505859375,
0.1732177734375,
0.466552734375
],
[
0.1748046875,
0.46728515625,
0.68310546875,
0.46728515625,
0.68310546875,
0.48095703125,
0.1748046875,
0.48095703125
],
[
0.1732177734375,
0.488037109375,
0.82177734375,
0.488037109375,
0.82177734375,
0.501953125,
0.1732177734375,
0.501953125
],
[
0.176513671875,
0.50390625,
0.8203125,
0.50390625,
0.8203125,
0.5146484375,
0.176513671875,
0.5146484375
],
[
0.1748046875,
0.5166015625,
0.82275390625,
0.5166015625,
0.82275390625,
0.52978515625,
0.1748046875,
0.52978515625
],
[
0.17236328125,
0.52978515625,
0.82421875,
0.52880859375,
0.82421875,
0.54345703125,
0.17236328125,
0.54443359375
],
[
0.1748046875,
0.54443359375,
0.438720703125,
0.54443359375,
0.438720703125,
0.55810546875,
0.1748046875,
0.55810546875
],
[
0.1707763671875,
0.556640625,
0.82177734375,
0.56005859375,
0.82177734375,
0.57763671875,
0.1707763671875,
0.57470703125
],
[
0.17236328125,
0.57470703125,
0.826171875,
0.57568359375,
0.826171875,
0.59033203125,
0.17236328125,
0.5888671875
],
[
0.1732177734375,
0.59033203125,
0.46728515625,
0.5888671875,
0.46728515625,
0.60400390625,
0.1732177734375,
0.60498046875
],
[
0.453369140625,
0.60986328125,
0.54150390625,
0.60986328125,
0.54150390625,
0.625,
0.453369140625,
0.625
],
[
0.80224609375,
0.60986328125,
0.82421875,
0.60986328125,
0.82421875,
0.625,
0.80224609375,
0.625
],
[
0.176513671875,
0.630859375,
0.82421875,
0.630859375,
0.82421875,
0.64599609375,
0.176513671875,
0.64599609375
],
[
0.1732177734375,
0.64599609375,
0.346435546875,
0.646484375,
0.346435546875,
0.6611328125,
0.1732177734375,
0.66015625
],
[
0.1748046875,
0.67041015625,
0.498291015625,
0.67041015625,
0.498291015625,
0.6845703125,
0.1748046875,
0.6845703125
],
[
0.5048828125,
0.671875,
0.82421875,
0.671875,
0.82421875,
0.68505859375,
0.5048828125,
0.68505859375
],
[
0.1748046875,
0.68505859375,
0.82421875,
0.68505859375,
0.82421875,
0.69873046875,
0.1748046875,
0.69873046875
],
[
0.1748046875,
0.69970703125,
0.82275390625,
0.69970703125,
0.82275390625,
0.71337890625,
0.1748046875,
0.71337890625
],
[
0.1748046875,
0.71337890625,
0.82421875,
0.71337890625,
0.82421875,
0.7265625,
0.1748046875,
0.7265625
],
[
0.1748046875,
0.7265625,
0.82421875,
0.7265625,
0.82421875,
0.74072265625,
0.1748046875,
0.74072265625
],
[
0.1748046875,
0.7412109375,
0.82275390625,
0.7412109375,
0.82275390625,
0.7548828125,
0.1748046875,
0.7548828125
],
[
0.1748046875,
0.7548828125,
0.73046875,
0.7548828125,
0.73046875,
0.76953125,
0.1748046875,
0.76953125
],
[
0.17724609375,
0.79052734375,
0.19287109375,
0.79248046875,
0.1895751953125,
0.8017578125,
0.1748046875,
0.7998046875
],
[
0.2017822265625,
0.7890625,
0.69677734375,
0.7890625,
0.69677734375,
0.80322265625,
0.2017822265625,
0.80322265625
],
[
0.2059326171875,
0.806640625,
0.35546875,
0.806640625,
0.35546875,
0.82080078125,
0.2059326171875,
0.82080078125
],
[
0.1732177734375,
0.83544921875,
0.8203125,
0.8359375,
0.8203125,
0.849609375,
0.1732177734375,
0.8486328125
],
[
0.1748046875,
0.85107421875,
0.82275390625,
0.85107421875,
0.82275390625,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.193603515625,
0.8720703125,
0.82177734375,
0.87255859375,
0.82177734375,
0.88720703125,
0.193603515625,
0.88623046875
],
[
0.17236328125,
0.8837890625,
0.826171875,
0.88525390625,
0.826171875,
0.8994140625,
0.17236328125,
0.8984375
],
[
0.1732177734375,
0.8994140625,
0.82275390625,
0.8994140625,
0.82275390625,
0.91357421875,
0.1732177734375,
0.91357421875
],
[
0.1748046875,
0.91357421875,
0.28515625,
0.91357421875,
0.28515625,
0.923828125,
0.1748046875,
0.923828125
],
[
0.494384765625,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.494384765625,
0.96044921875
]
] |
[
"decoding assumption provides a foothold to analyze the reachable sets and controllability of LLM",
"systems withoutinvoking notions ofstochasticcontrolas inSivaramakrishnanetal.(2023):Soatto",
"et al. (2023).Moreover, it remains connected to temperature-based stochastic decoding strategies as",
"a limiting case of temperature-based sampling as zero-temperature sampling.",
"We now extend Definition 1O to define output controllability for LLM systems:",
"Definition 2 (LLM Output Reachability). Output token sequence y E Vr is reachable from initial",
"state Xo E V* for LLM system E(V, PLM) iff there exists some time T and input u* E uT that",
"steers the LLM from initial state Xo to output y = h(x(T), r) at time T.",
"We disregard the trivial solution wherein the control input u* (t) overrides the LLM to force the state",
"sequence to take on the desired output value y.",
"The reachable output set definition for LLM systems follows from Definition 11:",
"Definition 3 (LLM Reachable Output Set). The reachable output set from initial state Xo E V* for",
"LLM system = (V, PLM) is denoted Ry(xo) and consists of all reachable outputs y E V* from",
"initial state Xo.",
"Output controllability for LLMs follows from Definition 13.",
"Definition 4 (LLM Output Controllability). An LLM system = (V, PLM) is output controllable",
"iff for every initial state Xo E V*, the reachable output set Ry (xo) = V*",
"The turn-based nature of writing to the LLM state sequence x(t) invites the question of whether the",
"promptushouldpreemptthe imposed statexoorcomeafterthe state.Wefocus ourefforts on",
"cases where u comes before imposed state sequence xo due to its importance for developing system",
"prompts and controlling text completion-based generation where the desired output is xo + y* for",
"some desired continuation y* of partial string xo. Due to the costly nature of long prompts, we are",
"especially interested in the existence of prompts u* with minimal length u*|.",
"Definitions 3 and 4 form the basis for our control theory of LLMs. While amenable to analytic",
"analysisasinSection4and Soattoetal.(2023),empiricalanalysisofthereachablesetand control",
"lability is challenging due to the intractable size of V*.We propose the following statistical measure",
"of controllability for practically assessing the controllability of an LLM system w.r.t. a dataset D",
"under prompt length constraint u ≤ k:",
"Definition 5 (k - e Controllability). Consider a dataset of state-output pairs D = [(x, y\"))ie[N)",
"An LLM =(V, PLM)is k-E controllablew.rt.D if. for at least a proportion of (1-e)of",
"(xo,y') E D, the following condition holds:",
"yiERh(xo)",
"(2)",
"Where R(x) is the reachable set of outputs as in Definition 3 under the constraint that prompts u",
"musthavelength|u|≤k.",
"Our empirical work in Section 5.2 explores k -",
"E controllability w.r.t. initial states Xo sampled",
"from the Wikitext dataset. While empirical analysis of LLM controllability is challenging due to",
"the lack of apparent structure in LLM dynamics and the combinatorially large state space, we may",
"still experimentally establish the existence of optimal prompts u* that elicit a given output, and thus",
"establish a lower bound on the content of the reachable set. Meanwhile, our theoretical work in",
"Section 4 establishes upper bounds on the content of the reachable set for self-attention. We hope",
"these complementary approaches aid in unifying our understanding of LLM systems.",
"4.",
"MATHEMATICAL ANALYSIS ONTHE CONTROLLABILITY OF",
"SELF-ATTENTION",
"Self-attention is a central component in modern transformer-based language models (Brown et al.",
"2020; Touvron et al., 2023; Radford et al., 2019; Min et al., 2023). Introduced in Vaswani et al.",
"'Both situations are reasonable in developing LLM systems: u preceding xo may arise when prompting an",
"LLM to complete a partial string xo.u proceeding xo may arise when prompting an LLM in the presence of",
"an imposed system prompt xo. Therefore, how an initial state xo is interleaved with control input u is largely",
"adesigndecision.",
"5"
] |
{
"arxiv_id": "2310.04444",
"language": "en",
"page_idx": 4
}
|
||
2312.04893-en-0
|
2312.04893-en-0.jpg
|
Annotation-Free Group Robustness via Loss-Based Resampling
Hesam Asadollahzadeh
Mahdi Ghaznavi
mahdi.ghaznavißce.sharif.edu
hesam.asadzadeh26Qresearcher.sharif.edu
Fahimeh HosseiniNoohdani
HamidReza Yaghoubi Araghi
hamidreza.yaghoubiaraghi@sharif.edu
fhosseini@ce.sharif.edu
Mohammad Hossein Rohban
Mahdieh Soleymani Baghshah
rohban@sharif.edu
soleymani@sharif.edu
Sharif University of Technology
## Tehran, Iran
## $\mathbf{Abstract}$
It is well-known that training neural networks for image classificationwithempirical risk minimization (ERM) makes them vulnerable to relying on spurious attributes instead of causal ones for prediction. Previously, deep feature re-weighting (DFR) has proposed retraining the last layer of a pre-trained network on balanced data concerning spurious attributes, making it robust to spurious correlation. However, spurious attribute annotations are not always available. In order to provide group robustness without such annotations, we propose a new method, called $loss- based\textit{feature re- weighting ( LFR) , in which we infer}$ a grouping of the data by evaluating an ERM-pre-trained model on a small left-out split of the training data. Then, a balanced number of samples is chosen by selecting high-loss samples from misclassifred data points and low-loss samples from correctly-classifted ones. Finally, we retrain the last layer on the selected balanced groups to make the model robust to spurious correlation. For a complete assessment, we evaluate LFR on various versions $of$ Waterbirds and CelebA datasets with different spurious correlations, which is a novel technique for observing the model's performance in a wide range of spuriosity rates. While LFR is extremely fast and straightforward, it outperforms the previous methods that do not assume group label availability, as well as the DFR with group annotations provided, in cases of high spurious correlation in the training data.
## $1. \textbf{Introduction}$
Neural networks have the tendency to rely on patterns in input data that are highly correlated with the target class in image classification. If the correlations that a model relies on are spurious, i.e. non-causal, the model's performance will drop under spurious correlation shift in the test time [21, 10, 18]. In such situations, the model performs well on majority groups of training data which have the spurious pattern, while the performance is poor on $mi-$ $noritygroups[1,2].$ As an example, in Waterbirds[16] dataset, there is a spurious correlation between the waterbird/landbird label and the water/land background, which implies waterbirds on water and landbirds on land as the majority groups and the other two combinations as minority groups. The final goal is to achieve the best worst group accuracy $(WGA)$ among all groups for the sake of robustness to correlation shift.
Some earlier methods consider accessing group annotations for the whole data[16,7] or a small portion of it[20,9]. [16] proposes group distributionally robust optimization (GDRO), which softly optimizes the worst-case loss among groups during training. [7] suggests that retraining the last layer of a classification model on a balanced dataset concerning group annotation and label is sufficient for making the model robust to spurious correlation shift. Their deep feature re-weighting (DFR) method has achieved good performance while being simple and fast. Nonetheless, group annotations are not always available. Some works do not assume such annotation requirement [11,14,19,3,12], but they need to train or fine-tune the whole model to achieve robustness.
As stated in [7], models trained with empirical risk minimization (ERM) learn core (causal) features besides the
1
|
[
[
0.1683349609375,
0.1319580078125,
0.8037109375,
0.1331787109375,
0.8037109375,
0.15087890625,
0.1683349609375,
0.149658203125
],
[
0.22216796875,
0.18310546875,
0.35546875,
0.18310546875,
0.35546875,
0.1976318359375,
0.22216796875,
0.1976318359375
],
[
0.54150390625,
0.18310546875,
0.72802734375,
0.18310546875,
0.72802734375,
0.1976318359375,
0.54150390625,
0.1976318359375
],
[
0.1641845703125,
0.2020263671875,
0.414306640625,
0.2020263671875,
0.414306640625,
0.21533203125,
0.1641845703125,
0.21533203125
],
[
0.4599609375,
0.2020263671875,
0.8056640625,
0.2020263671875,
0.8056640625,
0.21533203125,
0.4599609375,
0.21533203125
],
[
0.2320556640625,
0.2259521484375,
0.464111328125,
0.2259521484375,
0.464111328125,
0.240478515625,
0.2320556640625,
0.240478515625
],
[
0.55810546875,
0.2259521484375,
0.77783203125,
0.2259521484375,
0.77783203125,
0.240478515625,
0.55810546875,
0.240478515625
],
[
0.19287109375,
0.24560546875,
0.50341796875,
0.24560546875,
0.50341796875,
0.2587890625,
0.19287109375,
0.2587890625
],
[
0.56787109375,
0.246826171875,
0.7685546875,
0.246826171875,
0.7685546875,
0.2568359375,
0.56787109375,
0.2568359375
],
[
0.0196075439453125,
0.266357421875,
0.058013916015625,
0.266357421875,
0.060455322265625,
0.477783203125,
0.0228729248046875,
0.477783203125
],
[
0.221435546875,
0.268310546875,
0.450927734375,
0.268310546875,
0.450927734375,
0.28271484375,
0.221435546875,
0.28271484375
],
[
0.5087890625,
0.26953125,
0.7490234375,
0.26953125,
0.7490234375,
0.28271484375,
0.5087890625,
0.28271484375
],
[
0.259033203125,
0.287353515625,
0.410888671875,
0.287353515625,
0.410888671875,
0.300537109375,
0.259033203125,
0.300537109375
],
[
0.54345703125,
0.28857421875,
0.72119140625,
0.28857421875,
0.72119140625,
0.3017578125,
0.54345703125,
0.3017578125
],
[
0.358642578125,
0.309326171875,
0.61279296875,
0.31005859375,
0.61279296875,
0.32763671875,
0.358642578125,
0.326904296875
],
[
0.43701171875,
0.330078125,
0.5361328125,
0.330078125,
0.5361328125,
0.3447265625,
0.43701171875,
0.3447265625
],
[
0.23779296875,
0.380615234375,
0.313720703125,
0.380615234375,
0.313720703125,
0.395751953125,
0.23779296875,
0.395751953125
],
[
0.50244140625,
0.380615234375,
0.6298828125,
0.380615234375,
0.6298828125,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.51953125,
0.40771484375,
0.89208984375,
0.4091796875,
0.89208984375,
0.423583984375,
0.51953125,
0.42236328125
],
[
0.50244140625,
0.425537109375,
0.89208984375,
0.425537109375,
0.89208984375,
0.439453125,
0.50244140625,
0.439453125
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.50244140625,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.1004638671875,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.1004638671875,
0.477783203125
],
[
0.50244140625,
0.47021484375,
0.89208984375,
0.47021484375,
0.89208984375,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.08087158203125,
0.479248046875,
0.46728515625,
0.479248046875,
0.46728515625,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.5009765625,
0.48486328125,
0.890625,
0.486083984375,
0.890625,
0.5,
0.5009765625,
0.498779296875
],
[
0.02532958984375,
0.496826171875,
0.058013916015625,
0.496826171875,
0.058013916015625,
0.58544921875,
0.02532958984375,
0.58544921875
],
[
0.08087158203125,
0.494384765625,
0.4697265625,
0.494384765625,
0.4697265625,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.5009765625,
0.5,
0.89208984375,
0.50048828125,
0.89208984375,
0.51513671875,
0.5009765625,
0.5146484375
],
[
0.07843017578125,
0.50830078125,
0.468017578125,
0.5087890625,
0.468017578125,
0.5234375,
0.07843017578125,
0.52294921875
],
[
0.50341796875,
0.5166015625,
0.890625,
0.5166015625,
0.890625,
0.52978515625,
0.50341796875,
0.52978515625
],
[
0.07843017578125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.07843017578125,
0.53857421875
],
[
0.50244140625,
0.53076171875,
0.890625,
0.52978515625,
0.890625,
0.54443359375,
0.50244140625,
0.54541015625
],
[
0.080078125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.50244140625,
0.546875,
0.890625,
0.546875,
0.890625,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.080078125,
0.5546875,
0.4697265625,
0.5546875,
0.4697265625,
0.568359375,
0.080078125,
0.568359375
],
[
0.50341796875,
0.5625,
0.890625,
0.5625,
0.890625,
0.57275390625,
0.50341796875,
0.57275390625
],
[
0.080078125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.5771484375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.02777099609375,
0.58349609375,
0.053924560546875,
0.58349609375,
0.053924560546875,
0.6015625,
0.02777099609375,
0.6015625
],
[
0.080078125,
0.58544921875,
0.468017578125,
0.58544921875,
0.468017578125,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50244140625,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.60498046875,
0.50244140625,
0.60498046875
],
[
0.02532958984375,
0.60302734375,
0.056365966796875,
0.60302734375,
0.056365966796875,
0.642578125,
0.02532958984375,
0.642578125
],
[
0.080078125,
0.60107421875,
0.4697265625,
0.60107421875,
0.4697265625,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.50244140625,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.50244140625,
0.62060546875
],
[
0.07843017578125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.07843017578125,
0.62890625
],
[
0.50244140625,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.035125732421875,
0.6337890625,
0.051483154296875,
0.6337890625,
0.051483154296875,
0.66015625,
0.035125732421875,
0.66015625
],
[
0.080078125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.63720703125,
0.66259765625,
0.63720703125,
0.66259765625,
0.65087890625,
0.50244140625,
0.65087890625
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.0228729248046875,
0.65771484375,
0.052276611328125,
0.65478515625,
0.060455322265625,
0.7021484375,
0.0310516357421875,
0.705078125
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.521484375,
0.65576171875,
0.890625,
0.65576171875,
0.890625,
0.66943359375,
0.521484375,
0.66943359375
],
[
0.035125732421875,
0.66845703125,
0.0400390625,
0.66845703125,
0.0400390625,
0.67529296875,
0.035125732421875,
0.67529296875
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.50341796875,
0.671875,
0.8935546875,
0.671875,
0.8935546875,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68603515625,
0.890625,
0.68603515625,
0.890625,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.5009765625,
0.69970703125,
0.89208984375,
0.7021484375,
0.89208984375,
0.71533203125,
0.5009765625,
0.71337890625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.08087158203125,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.73291015625,
0.08087158203125,
0.73291015625
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.73291015625,
0.8896484375,
0.73291015625,
0.8896484375,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.7470703125,
0.89208984375,
0.7470703125,
0.89208984375,
0.76123046875,
0.50341796875,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5048828125,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.7734375,
0.5048828125,
0.7734375
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50244140625,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.08087158203125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50244140625,
0.79248046875,
0.890625,
0.79248046875,
0.890625,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.806640625,
0.890625,
0.806640625,
0.890625,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82275390625,
0.89208984375,
0.82275390625,
0.89208984375,
0.8359375,
0.5009765625,
0.8359375
],
[
0.08087158203125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.8369140625,
0.89208984375,
0.8369140625,
0.89208984375,
0.85107421875,
0.50341796875,
0.85107421875
],
[
0.080078125,
0.85693359375,
0.2132568359375,
0.85693359375,
0.2132568359375,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.8515625,
0.73779296875,
0.8515625,
0.73779296875,
0.86572265625,
0.50341796875,
0.86572265625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.47802734375,
0.9296875,
0.486083984375,
0.9248046875,
0.491943359375,
0.9287109375,
0.484375,
0.93505859375
]
] |
[
"Annotation-Free Group Robustness via Loss-Based Resampling",
"Mahdi Ghaznavi",
"Hesam Asadollahzadeh",
"mahdi.ghaznavi@ce.sharif.edu",
"hesam.asadzadeh26@researcher.sharif.edu",
"HamidReza Yaghoubi Araghi",
"Fahimeh HosseiniNoohdani",
"hamidreza.yaghoubiaraghi@sharif.edu",
"fhosseini@ce.sharif.edu",
" ]",
"Mohammad Hossein Rohban",
"Mahdieh SoleymaniBaghshah",
"rohban@sharif.edu",
"soleymani@sharif.edu",
"Sharif University of Technology",
"Tehran, Iran",
"Abstract",
"1. Introduction",
"Neural networks have the tendency to rely on patterns in",
"input data that are highly correlated with the target class",
"in image classification. If the correlations that a model",
"relies on are spurious, i.e. non-causal, the model's per-",
"It is well-known that training neural networks for image",
"formance will drop under spurious correlation shift in the",
"classification with empirical risk minimization (ERM)",
"test time [21, 10, 18]. In such situations, the model per-",
"108",
"makes them vulnerable to relying on spurious attributes",
"forms well on majority groups of training data which have",
"instead of causal ones for prediction.Previously, deep",
"the spurious pattern, while the performance is poor on mi-",
"feature re-weighting (DFR) has proposed retraining the",
"nority groups [1, 2]. As an example, in Waterbirds [16]",
"last layer of a pre-trained network on balanced data con-",
"dataset, there is a spurious correlation between the water-",
"cerning spurious attributes, making it robust to spurious",
"bird/landbirdlabelandthewater/landbackground.which",
"correlation. However, spurious attribute annotations are",
"implies waterbirds on water and landbirds on land as the",
"2",
"not always available. In order to provide group robustness",
"majority groups and the other two combinations as minor-",
"1:",
"without such annotations, we propose a new method, called",
"ity groups. The final goal is to achieve the best worst group",
"loss-based feature re-weighting (LFR), in which we infer",
"accuracy (WGA) among all groups for the sake of robust-",
".A",
"a grouping of the data by evaluating an ERM-pre-trained",
"ness to correlation shift.",
"model on a small left-out split of the training data. Then,",
"IXe",
"a balanced number of samples is chosen by selecting",
"Some earlier methods consider accessing group anno-",
"A",
"high-loss samples from misclassified data points and",
"tations for the whole data [16, 7] or a small portion of",
"low-loss samples from correctly-classified ones. Finally",
"it [20, 9]. [16] proposes group distributionally robust op-",
"timization (GDRO), which softly optimizes the worst-case",
"we retrain the last layer on the selected balanced groups",
"loss among groups during training.[7] suggests that re-",
"to make the model robust to spurious correlation.For a",
"complete assessment, we evaluate LFR on various versions",
"training the last layer of a classification model on a bal-",
"of Waterbirds and CelebA datasets with different spurious",
"anced dataset concerning group annotation and label is suf-",
"correlations, which is a novel technique for observing",
"ficientformakingthemodelrobusttospuriouscorrelation",
"the model's performance in a wide range of spuriosity",
"shift. Their deep feature re-weighting (DFR) method has",
"rates. While LFR is extremely fast and straightforward,",
"achieved good performance while being simple and fast.",
"it outperforms the previous methods that do not assume",
"Nonetheless, group annotations are not always available.",
"group label availability, as well as the DFR with group",
"Some works do not assume such annotation requirement",
"annotations provided, in cases of high spurious correlation",
"[11, 14, 19, 3, 12], but they need to train or fine-tune the",
"in the training data.",
"whole model to achieve robustness.",
"As stated in [7], models trained with empirical risk min-",
"imization (ERM) learn core (causal) features besides the",
"1"
] |
{
"arxiv_id": "2312.04893",
"language": "en",
"page_idx": 0
}
|
||
2312.04893-en-1
|
2312.04893-en-1.jpg
|
<FigureHere>
Figure 1. Overview of LFR method. After training a model on unbalanced data, we evaluate it on a small portion of the dataset (about 20% of the total). Then, we select images with low loss values in correctly-classified images and high loss in misclassified images and select an equal number of them in each class. At last, retraining the last layer of the model on the balanced dataset improves the model's robustness.
spurious ones, regardless of how strong the spurious correlation is. Thus, it seems that all we need is to find the proper linear combination of features to achieve the desired WGA. Recently, [15] proposed automatic feature re-weighting (AFR), which re-weights each sample while training the model's last layer, based on the model's probability of predicting the correct label on a given sample. However, as their weighting does not consider the size of minority groups, it does not work in situations where the size of the minority group is too small. Our work is focused on addressing this issue.
In this work, we propose a new method for robustness to spurious correlation shifts. First, we infer groups based on the model's correctness of prediction on a set of data. We then select a few number of high-loss samples from the misclassified ones and low-loss samples from the correctlyclassified images in each group. Finally, we apply DFR for feature re-weighting based on the selected groups. Our method, called loss-based feature re-weighting (LFR) is a simple yet effective method that achieves state-of-the-art (SOTA) on various famous datasets with high spurious correlation. The overview of LFR is shown in Figure 1.
Using LFR at the end of any training pipeline can improve WGA free of charge since it is fast, straightforward, and doesn't assume any inductive bias. LFR works well on a wide range of high to low spurious correlations by using only a small subset of training data for last layer retraining. By visualizing the saliency map of the model before and after last-layer retraining, we observe that the model's attention regions shift from spurious to core attributes in both majority and minority samples.
tions [16,7]. On the other hand, some works have tried to make a model robust to group shift without the availability of group annotations [11,14,22,15].
There are also some works that only need a small amount of group-annotated data.[20] used these group annotations to train a classifier for predicting group labels of the remaining data points, which are then fed into GDRO. [9] utilizes a set of diverse classifiers on a dataset without using group annotations, and then annotates a few samples for selecting the favorable model among this set.
Scope of Training: Different methods have different scopes of training, which leads to different amounts of required computation and effectiveness. Some works make a model robust during initial training. GDRO[16] solves a two-level optimization to minimize the maximum loss among all groups.[14] proposes learning from failure (LfF), which trains two models: a super-biased model to majority groups, and another model that upweights images with high loss value in a biased model. Thus, the second model goes in a way to be robust.
On the other hand, some works fine-tune a biased pretrained model that had been trained by ERM.[11] proposes just training twice (JTT), which upweights misclassified data points during the retraining.[22] uses a contrastive method to make samples of a class that are correctly or misclassified by a pre-trained model close together. Both aforementioned methods need validation data for tuning weighting hyperparameters.
However, some methods just retrain only the last layer of pre-trained models. These methods are based on DFR [7], which suggest retraining the last layer of the model on group-balanced data to make the model robust to spurious correlation. Other methods were proposed to solve the requirement of group annotation by different techniques for grouping or weighting. [8] suggests putting samples of each class whose prediction changed after dropout in one group and the rest in another for using DFR. AFR [15] is a recently proposed method in this line, which upweights
## $2. \textbf{ Related Work}$
We can divide the related work in the field by different
means.
Availability of group annotations: Some works in the literature have an assumption of access to group annota-
2
|
[
[
0.4306640625,
0.10101318359375,
0.49267578125,
0.10101318359375,
0.49267578125,
0.111083984375,
0.4306640625,
0.111083984375
],
[
0.73779296875,
0.10479736328125,
0.8486328125,
0.10479736328125,
0.8486328125,
0.11871337890625,
0.73779296875,
0.11871337890625
],
[
0.250732421875,
0.1060791015625,
0.306396484375,
0.1060791015625,
0.306396484375,
0.11614990234375,
0.250732421875,
0.11614990234375
],
[
0.098876953125,
0.1060791015625,
0.1707763671875,
0.1060791015625,
0.1707763671875,
0.117431640625,
0.098876953125,
0.117431640625
],
[
0.60302734375,
0.1060791015625,
0.67236328125,
0.1060791015625,
0.67236328125,
0.11614990234375,
0.60302734375,
0.11614990234375
],
[
0.4345703125,
0.11236572265625,
0.488525390625,
0.11236572265625,
0.488525390625,
0.12249755859375,
0.4345703125,
0.12249755859375
],
[
0.55224609375,
0.212158203125,
0.61767578125,
0.212158203125,
0.61767578125,
0.2197265625,
0.55224609375,
0.2197265625
],
[
0.82666015625,
0.210205078125,
0.8857421875,
0.210205078125,
0.8857421875,
0.2203369140625,
0.82666015625,
0.2203369140625
],
[
0.080078125,
0.255126953125,
0.890625,
0.255126953125,
0.890625,
0.268310546875,
0.080078125,
0.268310546875
],
[
0.351318359375,
0.2139892578125,
0.408447265625,
0.2139892578125,
0.408447265625,
0.2183837890625,
0.351318359375,
0.2183837890625
],
[
0.08087158203125,
0.26953125,
0.8896484375,
0.26953125,
0.8896484375,
0.27978515625,
0.08087158203125,
0.27978515625
],
[
0.080078125,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.080078125,
0.322509765625,
0.468017578125,
0.322509765625,
0.468017578125,
0.3359375,
0.080078125,
0.3359375
],
[
0.50244140625,
0.322509765625,
0.89208984375,
0.322509765625,
0.89208984375,
0.3359375,
0.50244140625,
0.3359375
],
[
0.080078125,
0.33837890625,
0.4697265625,
0.33837890625,
0.4697265625,
0.3515625,
0.080078125,
0.3515625
],
[
0.50244140625,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.3515625,
0.50244140625,
0.3515625
],
[
0.07843017578125,
0.352783203125,
0.468017578125,
0.3515625,
0.468017578125,
0.3662109375,
0.07843017578125,
0.367431640625
],
[
0.5009765625,
0.352783203125,
0.7509765625,
0.3515625,
0.7509765625,
0.3662109375,
0.5009765625,
0.367431640625
],
[
0.07843017578125,
0.3662109375,
0.4697265625,
0.367431640625,
0.4697265625,
0.3818359375,
0.07843017578125,
0.380615234375
],
[
0.521484375,
0.369384765625,
0.89208984375,
0.369384765625,
0.89208984375,
0.38330078125,
0.521484375,
0.38330078125
],
[
0.08087158203125,
0.385009765625,
0.46728515625,
0.385009765625,
0.46728515625,
0.395751953125,
0.08087158203125,
0.395751953125
],
[
0.50244140625,
0.386474609375,
0.89208984375,
0.386474609375,
0.89208984375,
0.399658203125,
0.50244140625,
0.399658203125
],
[
0.080078125,
0.3984375,
0.468017578125,
0.3984375,
0.468017578125,
0.412353515625,
0.080078125,
0.412353515625
],
[
0.50244140625,
0.401611328125,
0.890625,
0.401611328125,
0.890625,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.080078125,
0.41357421875,
0.468017578125,
0.41357421875,
0.468017578125,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.416748046875,
0.89208984375,
0.416748046875,
0.89208984375,
0.429931640625,
0.50244140625,
0.429931640625
],
[
0.08087158203125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.5009765625,
0.43115234375,
0.89208984375,
0.431884765625,
0.89208984375,
0.44580078125,
0.5009765625,
0.4443359375
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.50244140625,
0.460205078125
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.50244140625,
0.4609375,
0.73779296875,
0.4609375,
0.73779296875,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.08087158203125,
0.47607421875,
0.240234375,
0.47607421875,
0.240234375,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.521484375,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.521484375,
0.492431640625
],
[
0.09722900390625,
0.490478515625,
0.4697265625,
0.490478515625,
0.4697265625,
0.5048828125,
0.09722900390625,
0.5048828125
],
[
0.5009765625,
0.49365234375,
0.89208984375,
0.492431640625,
0.89208984375,
0.5068359375,
0.5009765625,
0.50830078125
],
[
0.07843017578125,
0.505859375,
0.4697265625,
0.505859375,
0.4697265625,
0.51953125,
0.07843017578125,
0.51953125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.080078125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.08087158203125,
0.53515625,
0.4697265625,
0.53515625,
0.4697265625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50341796875,
0.54150390625,
0.890625,
0.54150390625,
0.890625,
0.5517578125,
0.50341796875,
0.5517578125
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.080078125,
0.564453125
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.07843017578125,
0.56494140625,
0.468017578125,
0.564453125,
0.468017578125,
0.5791015625,
0.07843017578125,
0.580078125
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.08087158203125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.5966796875,
0.471435546875,
0.5966796875,
0.471435546875,
0.60986328125,
0.080078125,
0.60986328125
],
[
0.50244140625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.080078125,
0.611328125,
0.4697265625,
0.611328125,
0.4697265625,
0.625,
0.080078125,
0.625
],
[
0.50244140625,
0.61572265625,
0.72216796875,
0.61572265625,
0.72216796875,
0.62890625,
0.50244140625,
0.62890625
],
[
0.08087158203125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.08087158203125,
0.64013671875
],
[
0.5009765625,
0.62890625,
0.890625,
0.6298828125,
0.890625,
0.64453125,
0.5009765625,
0.64306640625
],
[
0.080078125,
0.64013671875,
0.420654296875,
0.64013671875,
0.420654296875,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64453125,
0.89208984375,
0.6591796875,
0.50244140625,
0.65771484375
],
[
0.09722900390625,
0.6572265625,
0.468017578125,
0.65771484375,
0.468017578125,
0.67236328125,
0.09722900390625,
0.671875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.080078125,
0.6875
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.07843017578125,
0.68798828125,
0.4697265625,
0.68798828125,
0.4697265625,
0.7021484375,
0.07843017578125,
0.7021484375
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.0767822265625,
0.7021484375,
0.4697265625,
0.7041015625,
0.4697265625,
0.71826171875,
0.0767822265625,
0.71630859375
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.08087158203125,
0.72021484375,
0.465576171875,
0.72021484375,
0.465576171875,
0.73095703125,
0.08087158203125,
0.73095703125
],
[
0.5009765625,
0.7197265625,
0.8935546875,
0.72021484375,
0.8935546875,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50341796875,
0.736328125,
0.685546875,
0.736328125,
0.685546875,
0.75,
0.50341796875,
0.75
],
[
0.080078125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7646484375,
0.4697265625,
0.7646484375,
0.4697265625,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.07843017578125,
0.77978515625,
0.285888671875,
0.77978515625,
0.285888671875,
0.79345703125,
0.07843017578125,
0.79345703125
],
[
0.5048828125,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.8125,
0.21728515625,
0.8125,
0.21728515625,
0.8271484375,
0.08087158203125,
0.8271484375
],
[
0.50341796875,
0.8125,
0.5947265625,
0.8125,
0.5947265625,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.58642578125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.58642578125,
0.82470703125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.098876953125,
0.8388671875,
0.4697265625,
0.8388671875,
0.4697265625,
0.85302734375,
0.098876953125,
0.85302734375
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.12744140625,
0.85693359375,
0.12744140625,
0.8671875,
0.080078125,
0.8671875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.09722900390625,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.8935546875,
0.87060546875,
0.8935546875,
0.88525390625,
0.5009765625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"Loss Based",
"Feature Reweighting",
"TrainERM",
"Training Data",
"RetrainData",
"Selection",
"HighLossLa",
"w Loss, Label 1",
"Figure 1. Overview of LFR method. After training a model on unbalanced data, we evaluate it on a small portion of the dataset (about 20%",
"ctedLastLave",
"ofthetota).Then.we select imageswithlowlossvalues incorrectlv-classified images and highloss inmisclassified imagesand selectan",
"equal number of them in each class. At last, retraining the last layer of the model on the balanced dataset improves the model's robustness.",
"spurious ones, regardless of how strong the spurious cor-",
"tions [16, 7]. On the other hand, some works have tried to",
"relation is. Thus, it seems that all we need is to find the",
"make a model robust to group shift without the availability",
"proper linear combination of features to achieve the de-",
"of group annotations [11, 14, 22, 15].",
"sired WGA. Recently, [15] proposed automatic feature",
"There are also some works that only need a small amount",
"re-weighting(AFR),whichre-weightseachsamplewhile",
"of group-annotated data. [20] used these group annotations",
"training the model's last layer, based on the model's prob-",
"to train a classifier for predicting group labels of the remain-",
"ability of predicting the correct label on a given sample.",
"ing data points, which are then fed into GDRO. [9] utilizes",
"However,astheirweightingdoesnotconsiderthesize of",
"a set of diverse classifiers on a dataset without using group",
"minority groups, it does not work in situations where the",
"annotations, and then annotates a few samples for selecting",
"size of the minority group is too small. Our work is focused",
"the favorable model among this set.",
"on addressing this issue.",
"Scope of Training: Different methods have different",
"In this work, we propose a new method for robustness",
"scopes of training, which leads to different amounts of",
"to spurious correlation shifts. First, we infer groups based",
"required computation and effectiveness. Some works make",
"on the model's correctness of prediction on a set of data.",
"a model robust during initial training. GDRO [16] solves",
"We then select a few number of high-loss samples from the",
"a two-level optimization to minimize the maximum loss",
"misclassified ones and low-loss samples from the correctly-",
"among all groups. [14] proposes learning from failure",
"classified images in each group.Finally, we apply DFR",
"(LfF), which trains two models: a super-biased model to",
"forfeaturere-weightingbased onthe selectedgroups.Our",
"majority groups, and another model that upweights images",
"method, called loss-based feature re-weighting (LFR) is a",
"with high loss value in a biased model. Thus, the second",
"simple yet effective method that achieves state-of-the-art",
"model goes in a way to be robust.",
"(SOTA) on various famous datasets with high spurious cor-",
"On the other hand, some works fine-tune a biased pre-",
"relation. The overview of LFR is shown in Figure 1.",
"trained model that had been trained by ERM. [11] proposes",
"Using LFR at the end of any training pipeline can im-",
"just training twice (JTT), which upweights misclassified",
"prove WGA free of charge since it is fast, straightforward,",
"data points during the retraining.[22] uses a contrastive",
"and doesn't assume any inductive bias. LFR works well on",
"method to make samples of a class that are correctly or",
"a wide range of high to low spurious correlations by using",
"misclassified by a pre-trained model close together. Both",
"onlya small subset of training dataforlastlayerretraining",
"aforementioned methods need validation data for tuning",
"By visualizing the saliency map of the model before and",
"weighting hyperparameters.",
"after last-layer retraining, we observe that the model's at-",
"However, some methods just retrain only the last layer of",
"tention regions shift from spurious to core attributes in both",
"pre-trained models. These methods are based on DFR [7],",
"majority and minority samples.",
"which suggest retraining the last layer of the model on",
"group-balanced data to make the model robust to spurious",
"2. Related Work",
"correlation.",
"Other methods were proposed to solve the",
"requirement of group annotation by different techniques",
"We can divide the related work in the field by different",
"for grouping or weighting. [8] suggests putting samples of",
"means.",
"each class whose prediction changed after dropout in one",
"Availability of group annotations: Some works in the",
"group and the rest in another for using DFR. AFR [15] is",
"literature have an assumption of access to group annota-",
"a recently proposed method in this line, which upweights",
"2"
] |
{
"arxiv_id": "2312.04893",
"language": "en",
"page_idx": 1
}
|
||
2312.04893-en-2
|
2312.04893-en-2.jpg
|
samples with lower confdence on the ground truth label
during retraining the last layer.
Our LFR method does not need access to group annotations and retrains only the last layer of a model which is pre-trained with ERM. Our setting is similar to AFR, and we compare it to LFR.
<FigureHere>
Figure 2. An abstraction of our LFR method in 2D space. The majority groups of data make the initial classifier biased towards themselves (black line). Selecting an equal number of samples from high-loss value misclassified and low-loss classified samples mitigates the bias towards a more robust classifier (green line)
## $3. \textbf{Method}$
# 3.1. Setting
Consider an image classification task among two classes on a dataset $\mathcal{D}\subset\mathbb{R}^{3\times H\times W}$ where $H$ and $W$ are the height and width of input images. Suppose images of each class $i$ $\in$ $\{ 1, 2\}$ in the dataset are split into unknown majority $Maj_i$ and $Min_i$ groups such that $|Maj_i|\gg|Min_i|$ and images in $Maj_i$ have predictive attributes which are simpler than the core attribute of class $i.$ We first train a deep neural network $f:\mathbb{R}^{3\times H\times W}\to[0,1]^2$ with ERM and loss function $\mathcal{L}:[0,1]^2\to\mathbb{R}$ on a randomly selected 80% of $\mathcal{D}$ with class balancing. Class balancing is a conventional necessary technique for training a good model in any classification task, even without strong spurious corre- $\operatorname*{lation.}$The function $f=softmax(h\circ g)$ is composed of a feature extractor $g:\mathbb{R}^{3\times\bar{H}\times W}\to\mathbb{R}^{\bar{d}}$ and a linear last layer function $h:\mathbb{R}^d\to\mathbb{R}^2$ with a softmax function on top of them. Then we are going to retrain the last layer $h$ of the model on the remaining 20% of $\mathcal{D}$, which refers to it by $\mathcal{D}_ll.$ We hope to have a higher WGA in groups $\{Maj_{1},Maj_{2}\}\cup\{\hat{M}in_{1},Min_{2}\}$ than before retraining.
## 3.2. Loss- Based Feature Reweighting
As it is illustrated in Fig. 1, after initial training, we evaluate the model $f$ on $\mathcal{D}_ll.$ We refer to the correctly-classified and misclassified samples of class $i$ by $C_i$ and $M_i.$ As we explain later in this section, these sets are good representatives of $Maj_i$ and $Min_i$ groups respectively.
After that, as has been suggested by [7], we want to select an equal number $s$ of data from each group in the set $\mathcal{G}=\bigcup_{i=1}^{2}\{M_{i},C_{i}\}.~s$,the number of groups' samples, is a hyperparameter and can be any number, but we usually select it approximately as $s=min(|M_{1}|,|M_{2}|).$ The optimal amount $s$ is found by hyperparameter selection based on worst group accuracy on the validation set. In each majority group like $C_i$, we select $s$ samples with the highest loss value $\mathcal{L}(f(x))$, whereas, for a minority group $M_i$, we select $s$ images with the lowest value of loss function. Then, as has been suggested by [7], we reinitialize the last layer $h$ of the model and then train it on a balanced subset of $\mathcal{D} _{ll}$ which consists $s$ samples from each group in $\mathcal{G}.$
It is important to select samples based on loss value, not randomly. If a model is pre-trained with ERM on a biased dataset, it probably relies on spurious correlation for pre-
diction. Thus, the loss value of the model is a good sign of belonging to majority or minority groups. So, in classified samples, if a sample has low loss, there is a great chance that the sample has a spurious attribute which has made the model biased. On the other hand, in a misclassified group, if a sample has a very high loss, it probably belongs to the minority group without the spurious attribute. The success of loss-based selection in comparison to random selection in the ablation study also suggests that among correctlyclassified samples, the ones that rely on spurious attributes have lower loss value than the ones that attend to core attributes. It is true as well for misclassified samples, as those with higher loss values are more likely to have not the spurious attribute. Figure 2 shows an abstraction of LFR in 2D space, which is compatible with the above explanations. Practical experiments also confırm our descriptions.
# 4. Experiments
4.1. Dataset
We evaluate LFR on two popular datasets.
Waterbirds: This dataset has been introduced by [16]. It is a classification between waterbirds and landbirds, where most of the waterbirds are in the sea and most of the landbirds are seen in the land background. Here, the background scene is considered a spurious feature in both classes.
CelebA: Using the dataset which has been proposed by [13], CelebA is a hair color classification of celebrities
3
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.08087158203125,
0.106689453125,
0.28515625,
0.10797119140625,
0.28515625,
0.121826171875,
0.08087158203125,
0.12060546875
],
[
0.81396484375,
0.10101318359375,
0.861328125,
0.10101318359375,
0.861328125,
0.11236572265625,
0.81396484375,
0.11236572265625
],
[
0.814453125,
0.11297607421875,
0.861328125,
0.11297607421875,
0.861328125,
0.12371826171875,
0.814453125,
0.12371826171875
],
[
0.080078125,
0.12249755859375,
0.46728515625,
0.12249755859375,
0.46728515625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.07843017578125,
0.1673583984375,
0.2296142578125,
0.1666259765625,
0.2296142578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.07843017578125,
0.2108154296875,
0.1690673828125,
0.2108154296875,
0.1690673828125,
0.2291259765625,
0.07843017578125,
0.2291259765625
],
[
0.07843017578125,
0.23486328125,
0.1683349609375,
0.2374267578125,
0.1683349609375,
0.25244140625,
0.07843017578125,
0.250732421875
],
[
0.09722900390625,
0.261962890625,
0.4697265625,
0.261962890625,
0.4697265625,
0.27587890625,
0.09722900390625,
0.27587890625
],
[
0.07843017578125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.291748046875,
0.07843017578125,
0.291748046875
],
[
0.080078125,
0.292236328125,
0.468017578125,
0.292236328125,
0.468017578125,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.080078125,
0.30810546875,
0.468017578125,
0.30810546875,
0.468017578125,
0.3212890625,
0.080078125,
0.3212890625
],
[
0.080078125,
0.322509765625,
0.46728515625,
0.322509765625,
0.46728515625,
0.3359375,
0.080078125,
0.3359375
],
[
0.080078125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.3515625,
0.080078125,
0.3515625
],
[
0.080078125,
0.3515625,
0.4697265625,
0.3515625,
0.4697265625,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.0767822265625,
0.3623046875,
0.471435546875,
0.3623046875,
0.471435546875,
0.38330078125,
0.0767822265625,
0.38330078125
],
[
0.50244140625,
0.371826171875,
0.89208984375,
0.371826171875,
0.89208984375,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.07843017578125,
0.3818359375,
0.4697265625,
0.380615234375,
0.4697265625,
0.395751953125,
0.07843017578125,
0.396484375
],
[
0.5048828125,
0.38818359375,
0.890625,
0.38818359375,
0.890625,
0.3984375,
0.5048828125,
0.3984375
],
[
0.080078125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.41357421875,
0.4697265625,
0.41357421875,
0.4697265625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.50244140625,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.50244140625,
0.427978515625
],
[
0.080078125,
0.4287109375,
0.46728515625,
0.4287109375,
0.46728515625,
0.442626953125,
0.080078125,
0.442626953125
],
[
0.50341796875,
0.4287109375,
0.869140625,
0.4287109375,
0.869140625,
0.442626953125,
0.50341796875,
0.442626953125
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.07598876953125,
0.453857421875,
0.47216796875,
0.45263671875,
0.47216796875,
0.473388671875,
0.07598876953125,
0.474853515625
],
[
0.080078125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.50341796875,
0.4697265625,
0.89208984375,
0.4697265625,
0.89208984375,
0.48291015625,
0.50341796875,
0.48291015625
],
[
0.080078125,
0.4892578125,
0.4697265625,
0.4892578125,
0.4697265625,
0.50244140625,
0.080078125,
0.50244140625
],
[
0.50244140625,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.080078125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50341796875,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.5146484375,
0.50341796875,
0.5146484375
],
[
0.080078125,
0.51806640625,
0.468017578125,
0.51806640625,
0.468017578125,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50244140625,
0.51513671875,
0.89208984375,
0.51513671875,
0.89208984375,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.08087158203125,
0.53271484375,
0.455810546875,
0.53271484375,
0.455810546875,
0.54736328125,
0.08087158203125,
0.54736328125
],
[
0.5009765625,
0.52978515625,
0.890625,
0.53076171875,
0.890625,
0.54443359375,
0.5009765625,
0.54345703125
],
[
0.5048828125,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.55810546875,
0.5048828125,
0.55810546875
],
[
0.080078125,
0.56005859375,
0.365234375,
0.56005859375,
0.365234375,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.50341796875,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.50244140625,
0.57568359375,
0.89208984375,
0.57568359375,
0.89208984375,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.09722900390625,
0.58203125,
0.468017578125,
0.58349609375,
0.468017578125,
0.59765625,
0.09722900390625,
0.5966796875
],
[
0.50341796875,
0.59228515625,
0.8896484375,
0.59228515625,
0.8896484375,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.080078125,
0.5986328125,
0.4697265625,
0.5986328125,
0.4697265625,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.5009765625,
0.60498046875,
0.890625,
0.60498046875,
0.890625,
0.61865234375,
0.5009765625,
0.61865234375
],
[
0.08087158203125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62548828125,
0.08087158203125,
0.62548828125
],
[
0.50244140625,
0.62060546875,
0.89208984375,
0.62060546875,
0.89208984375,
0.6337890625,
0.50244140625,
0.6337890625
],
[
0.07843017578125,
0.62744140625,
0.468017578125,
0.62890625,
0.468017578125,
0.64306640625,
0.07843017578125,
0.642578125
],
[
0.5009765625,
0.63525390625,
0.89208984375,
0.63623046875,
0.89208984375,
0.6494140625,
0.5009765625,
0.6484375
],
[
0.07843017578125,
0.64453125,
0.374267578125,
0.64453125,
0.374267578125,
0.65771484375,
0.07843017578125,
0.65771484375
],
[
0.50244140625,
0.65087890625,
0.890625,
0.6513671875,
0.890625,
0.66552734375,
0.50244140625,
0.6640625
],
[
0.1004638671875,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.1004638671875,
0.673828125
],
[
0.50244140625,
0.66650390625,
0.89208984375,
0.66650390625,
0.89208984375,
0.6796875,
0.50244140625,
0.6796875
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.681640625,
0.890625,
0.681640625,
0.890625,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.07843017578125,
0.68798828125,
0.4697265625,
0.689453125,
0.4697265625,
0.7041015625,
0.07843017578125,
0.70263671875
],
[
0.5009765625,
0.69580078125,
0.8408203125,
0.69580078125,
0.8408203125,
0.708984375,
0.5009765625,
0.708984375
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.734375,
0.080078125,
0.734375
],
[
0.5009765625,
0.72607421875,
0.63232421875,
0.72802734375,
0.63232421875,
0.74560546875,
0.5009765625,
0.74365234375
],
[
0.080078125,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.5048828125,
0.7548828125,
0.5947265625,
0.7548828125,
0.5947265625,
0.7666015625,
0.5048828125,
0.7666015625
],
[
0.08087158203125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.521484375,
0.7763671875,
0.79833984375,
0.77783203125,
0.79833984375,
0.79248046875,
0.521484375,
0.79150390625
],
[
0.080078125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.52294921875,
0.79443359375,
0.89208984375,
0.79443359375,
0.89208984375,
0.80810546875,
0.52294921875,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.08087158203125,
0.82763671875,
0.46728515625,
0.82763671875,
0.46728515625,
0.83837890625,
0.08087158203125,
0.83837890625
],
[
0.50341796875,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.8369140625,
0.50341796875,
0.8369140625
],
[
0.08087158203125,
0.84228515625,
0.390625,
0.84228515625,
0.390625,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.50244140625,
0.84033203125,
0.89208984375,
0.84033203125,
0.89208984375,
0.85400390625,
0.50244140625,
0.85400390625
],
[
0.098876953125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.098876953125,
0.86865234375
],
[
0.5009765625,
0.85595703125,
0.85302734375,
0.85595703125,
0.85302734375,
0.869140625,
0.5009765625,
0.869140625
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"samples with lower confidence on the ground truth label",
"during retraining the last layer.",
"Labelo",
"Label1",
"Our LFR method does not need access to group annota-",
"tions and retrains only the last layer of a model which is",
"pre-trained with ERM. Our setting is similar to AFR, and",
"we compare it to LFR.",
"3. Method",
"3.1. Setting",
"Consider an image classification task among two classes",
"On a dataset D c IR3xHxW where H and W are the height",
"and width of input images. Suppose images of each class",
"z (1,2] in the dataset are split into unknown major-",
"ity Maji and Mini groups such that Majil 》 JMini",
"and images in Maj; have predictive attributes which are",
"simpler than the core attribute of class i.. We first train a",
"deep neural network f : R3xHxW = [0,1]? with ERM",
"Figure 2. An abstraction of our LFR method in 2D space. The",
"and loss function C : [o, 1]j2 -→ IR on a randomly selected",
"majority groups of datamake theinitial classifierbiased towards",
"80% of D with class balancing. Class balancing is a con-",
"themselves (black line). Selecting an equal number of samples",
"ventional necessary technique for training a good model in",
"from high-loss value misclassified and low-loss classified samples",
"any classification task, even without strong spurious corre-",
"mitigates the bias towards a more robust classifier (green line)",
"lation. The function f = softmar(h o g) is composed of",
"a feature extractor g : R3xxW → Rd and a linear last",
"layer function h : IRd → R? with a softmax function on",
"diction. Thus, the loss value of the model is a good sign of",
"top of them. Then we are going to retrain the last layer",
"belonging to majority or minority groups. So, in classified",
"h of themodel ontheremaining20%of D,whichrefers",
"samples, if a sample has low loss, there is a great chance",
"to it by Du. We hope to have a higher WGA in groups",
"that the sample has a spurious attribute which has made the",
"[Maj1, Maj2] U [Mini, Min2] than before retraining.",
"model biased. On the other hand, in a misclassified group,",
"ifasamplehasaveryhighloss,itprobablybelongstothe",
"3.2.Loss-Based Feature Reweighting",
"minority group without the spurious attribute. The success",
"of loss-based selection in comparison to random selection",
"As it is illustrated in Fig. I, after initial training, we eval-",
"in theablation study also suggests that among correctly-",
"uate the model f on Du. We refer to the correctly-classified",
"classified samples, the ones that rely on spurious attributes",
"and misclassified samples of class i by C,and M;.As we",
"have lower loss value than the ones that attend to core at-",
"explain later in this section, these sets are good representa-",
"tributes. It is true as well for misclassified samples, as those",
"tives of Maji and Mini groups respectively.",
"with higherloss values are more likely to have not the spu-",
"After that, as has been suggested by [7], we want to se-",
"rious attribute. Figure 2 shows an abstraction of LFR in",
"lect an equal number s of data from each group in the set",
"2D space,which is compatible with the above explanations.",
"G = U?-,{ Mi, Ci]. s, the number of groups' samples, is",
"Practical experiments also confirm our descriptions.",
"a hyperparameter and can be any number, but we usually",
"select it approximately as s = min(Mil, M2). The op-",
"4. Experiments",
"timal amount s is found by hyperparameter selection based",
"on worst group accuracy on the validation set. In each ma-",
"4.1.Dataset",
"joritygrouplike Ci.we selects samples withthehighest",
"loss value L(f(α)), whereas, for a minority group M,, we",
"We evaluate LFR on two popular datasets.",
"select s images with the lowest value of loss function. Then,",
"Waterbirds: This dataset has been introduced by [16]. It",
"as has been suggested by [7], we reinitialize the last layer",
"is a classification between waterbirds and landbirds, where",
"h of the model and then train it on a balanced subset of Du",
"most of the waterbirds are in the sea and most of the land-",
"whichconsistsssamplesfromeachgrouping",
"birds are seen in the land background. Here, the background",
"It is important to select samples based on loss value, not",
"scene is considered a spurious feature in both classes.",
"randomly. If a model is pre-trained with ERM on a biased",
"CelebA: Using the dataset which has been proposed",
"dataset, it probably relies on spurious correlation for pre-",
"by [13], CelebA is a hair color classification of celebrities",
"3"
] |
{
"arxiv_id": "2312.04893",
"language": "en",
"page_idx": 2
}
|
||
2312.04893-en-4
|
2312.04893-en-4.jpg
|
Table 3. Effect of loss-based selection and random selection in each spurious correlation of Waterbirds dataset on worst and mean (in parentheses) group accuracy. Results are in percentage. CR-ML and CL-MR stand for $Correctlyclassiffedwith\boldsymbol{R}andom$ selection- Misclassiffed $with Loss- based selection $and $Misclassifed with Loss- based selection- Correctly classified with Random selection respectively. While Loss- based selection ande tion tespectively. While th Loss- based selection and with Loss- based selection- Correctly clas$ when the spurious correlation is not lower than 90% LFR has the best performance, when the correlation is 70% or 60% random selection of misclassified samples leads to a higher WGA
<table>
<tbody>
<tr>
<th>Method</th>
<th>$70\%$</th>
<th>$80\%$</th>
<th>$90\%$</th>
<th>$95\%$</th>
<th>$99\%$</th>
<th>$100\%$</th>
</tr>
<tr>
<td>$LFR$</td>
<td>$79.67\left(85.25\right)$</td>
<td>(87.88) 83.81</td>
<td>$\mathbf{5}\left(91.02\right)$ 86.45 1</td>
<td>$86.85\left(91.25\right)$</td>
<td>85.31 (89.24)</td>
<td>86.29 $\bullet(90.54$</td>
</tr>
<tr>
<td>$CFR$</td>
<td>75.45 :(87.15) </td>
<td>78.62 (87.57)</td>
<td>$75.09\left(84.93\right)$</td>
<td>67.45 (80.12)</td>
<td>78.76 (87.08)</td>
<td>82.09 (89.48</td>
</tr>
<tr>
<td>LFR- CR- ML</td>
<td>$81.09\left(89.61\right)$</td>
<td>83.36 (88.65)</td>
<td>82.68 (88.44)</td>
<td>$83.80\left(90.39\right)$</td>
<td>$80.84\left(87.93\right)$</td>
<td>83.80 (90.07</td>
</tr>
<tr>
<td>LFR- CL- MR</td>
<td>84.65 :(90.27) 1</td>
<td>84.27 (91.21)</td>
<td>$80.37\left(88.89\right)$</td>
<td>86.45 (92.03)</td>
<td>$83.96~(88.87)$</td>
<td>83.96 (89.35</td>
</tr>
</tbody>
</table>
and baselines, even works with the assumption of group annotations availability, in addition to better performance in attending core attributes in input images.
# References
<FigureHere>
Figure 3. Saliency map of models before (ERM) and after last layer retraining with AFR and LFR. LFR has a better quality in terms of pushing the model to attend to core attributes.
[1] Joy Buolamwini and Timnit Gebru. Gender shades: Inter-
sectional accuracy disparities in commercial gender classification. In Sorelle A. Friedler and Christo Wilson, editors, Proceedings ofthe Ist Conference on Fairness, Accountability and Transparency, volume 81 of Proceedings of Machine Learning Research, pages 77-91. PMLR, 23-24 Feb 2018.
[2] Alexandra Chouldechova. Fair prediction with disparate im-
pact: A study of bias in recidivism prediction instruments.
Big $Data$, 5(2):153-163, 2017. PMID: 28632438
[3] Elliot Creager, Jörn-Henrik Jacobsen, and Richard S. Zemel.
Environment inference for invariant learning. In Interna-
tional Conference on Machine Learning, 2020.
[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, $2009.CVPR2009.IEEEConferenceon$, pages 248-255. IEEE, 2009
[5] Jacob Gildenblat and contributors. Pytorch library for
cam methods.
https://github.com/jacobgil/
pytorch-grad-cam,2021.
[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition $(CVPR)$,pages $770-\bar{778},2016.$
[7] P. Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson.
Last layer re-training is sufficient for robustness to spurious
correlations. $ArXi\nu$, abs/2204.02937, 2022.
[8] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar.
Dropout disagreement: A recipe for group robustness with fewer annotations. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022.
[9] Yoonho Lee, Huaxiu Yao, and Chelsea Finn. Diversify and
disambiguate: Out-of-distribution robustness via disagreement. In The Eleventh International Conference on Learning Representations, 2023.
[10] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas,
Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come
underperforms in high spurious correlation settings because they do not have any consideration for the number of examples from the majority group. So, even if the underweight samples with lower loss are probably from the majority group, the overall effect of the majority samples still remains high. On the other hand, as LFR selects an equal number of samples from both the majority and minority, it does not have the aforementioned problem.
We also visualize our method's attention map with GradCam [5] to compare it with other models. As it can be seen in figure 3, using LFR, the model attends more to core attributes, in contrast to the pre-trained ERM model and AFR. More results are available in C.
## 5. Conclusion
We proposed loss-based feature re-weighting, a method for selecting a few images from a small proportion of training data to retrain the last linear of the model which improves robustness to spurious correlation. While LFR is fast and straightforward to use, it achieved SOTA in many various datasets in comparison to previously proposed methods
5
|
[
[
0.080078125,
0.08587646484375,
0.890625,
0.08709716796875,
0.890625,
0.10162353515625,
0.080078125,
0.10101318359375
],
[
0.07843017578125,
0.10162353515625,
0.890625,
0.10162353515625,
0.890625,
0.11553955078125,
0.07843017578125,
0.11553955078125
],
[
0.08087158203125,
0.117431640625,
0.890625,
0.117431640625,
0.890625,
0.1275634765625,
0.08087158203125,
0.1275634765625
],
[
0.080078125,
0.1300048828125,
0.89208984375,
0.1300048828125,
0.89208984375,
0.143310546875,
0.080078125,
0.143310546875
],
[
0.07843017578125,
0.143310546875,
0.365234375,
0.143310546875,
0.365234375,
0.1572265625,
0.07843017578125,
0.1572265625
],
[
0.11358642578125,
0.1590576171875,
0.17236328125,
0.1610107421875,
0.1707763671875,
0.175537109375,
0.11273193359375,
0.173583984375
],
[
0.25,
0.160400390625,
0.28515625,
0.160400390625,
0.28515625,
0.175537109375,
0.25,
0.175537109375
],
[
0.35693359375,
0.160400390625,
0.393798828125,
0.160400390625,
0.393798828125,
0.175537109375,
0.35693359375,
0.175537109375
],
[
0.57177734375,
0.160400390625,
0.609375,
0.160400390625,
0.609375,
0.175537109375,
0.57177734375,
0.175537109375
],
[
0.7607421875,
0.1610107421875,
0.80224609375,
0.1610107421875,
0.80224609375,
0.175537109375,
0.7607421875,
0.175537109375
],
[
0.68115234375,
0.1622314453125,
0.7158203125,
0.1622314453125,
0.7158203125,
0.1749267578125,
0.68115234375,
0.1749267578125
],
[
0.465576171875,
0.1622314453125,
0.5,
0.1622314453125,
0.5,
0.1749267578125,
0.465576171875,
0.1749267578125
],
[
0.11517333984375,
0.186279296875,
0.148681640625,
0.186279296875,
0.148681640625,
0.1988525390625,
0.11517333984375,
0.1988525390625
],
[
0.22216796875,
0.186279296875,
0.310546875,
0.186279296875,
0.310546875,
0.1976318359375,
0.22216796875,
0.1976318359375
],
[
0.330078125,
0.186279296875,
0.419189453125,
0.186279296875,
0.419189453125,
0.1976318359375,
0.330078125,
0.1976318359375
],
[
0.438720703125,
0.186279296875,
0.52783203125,
0.186279296875,
0.52783203125,
0.1976318359375,
0.438720703125,
0.1976318359375
],
[
0.5458984375,
0.186279296875,
0.6357421875,
0.186279296875,
0.6357421875,
0.1976318359375,
0.5458984375,
0.1976318359375
],
[
0.65283203125,
0.186279296875,
0.74267578125,
0.186279296875,
0.74267578125,
0.1976318359375,
0.65283203125,
0.1976318359375
],
[
0.7607421875,
0.186279296875,
0.85107421875,
0.186279296875,
0.85107421875,
0.1976318359375,
0.7607421875,
0.1976318359375
],
[
0.11358642578125,
0.199462890625,
0.1519775390625,
0.199462890625,
0.1519775390625,
0.21533203125,
0.11358642578125,
0.21533203125
],
[
0.221435546875,
0.2020263671875,
0.310546875,
0.2020263671875,
0.310546875,
0.21337890625,
0.221435546875,
0.21337890625
],
[
0.330078125,
0.2020263671875,
0.420654296875,
0.2020263671875,
0.420654296875,
0.21337890625,
0.330078125,
0.21337890625
],
[
0.438720703125,
0.2020263671875,
0.52783203125,
0.2020263671875,
0.52783203125,
0.21337890625,
0.438720703125,
0.21337890625
],
[
0.5439453125,
0.2020263671875,
0.6357421875,
0.2020263671875,
0.6357421875,
0.21337890625,
0.5439453125,
0.21337890625
],
[
0.65283203125,
0.2020263671875,
0.744140625,
0.2020263671875,
0.744140625,
0.21337890625,
0.65283203125,
0.21337890625
],
[
0.7607421875,
0.2020263671875,
0.85107421875,
0.2020263671875,
0.85107421875,
0.21337890625,
0.7607421875,
0.21337890625
],
[
0.5439453125,
0.21533203125,
0.63623046875,
0.21533203125,
0.63623046875,
0.2298583984375,
0.5439453125,
0.2298583984375
],
[
0.7587890625,
0.21533203125,
0.85107421875,
0.21533203125,
0.85107421875,
0.2298583984375,
0.7587890625,
0.2298583984375
],
[
0.221435546875,
0.2171630859375,
0.31201171875,
0.2171630859375,
0.31201171875,
0.2291259765625,
0.221435546875,
0.2291259765625
],
[
0.11517333984375,
0.216552734375,
0.205078125,
0.216552734375,
0.205078125,
0.2279052734375,
0.11517333984375,
0.2279052734375
],
[
0.330078125,
0.2171630859375,
0.420654296875,
0.2171630859375,
0.420654296875,
0.2291259765625,
0.330078125,
0.2291259765625
],
[
0.438720703125,
0.2171630859375,
0.52783203125,
0.2171630859375,
0.52783203125,
0.2291259765625,
0.438720703125,
0.2291259765625
],
[
0.654296875,
0.2171630859375,
0.74267578125,
0.2171630859375,
0.74267578125,
0.2291259765625,
0.654296875,
0.2291259765625
],
[
0.43701171875,
0.2310791015625,
0.53125,
0.2310791015625,
0.53125,
0.24560546875,
0.43701171875,
0.24560546875
],
[
0.65185546875,
0.2310791015625,
0.74609375,
0.2310791015625,
0.74609375,
0.24560546875,
0.65185546875,
0.24560546875
],
[
0.75732421875,
0.2291259765625,
0.85302734375,
0.2310791015625,
0.85302734375,
0.24560546875,
0.75732421875,
0.24365234375
],
[
0.221435546875,
0.231689453125,
0.31201171875,
0.231689453125,
0.31201171875,
0.24365234375,
0.221435546875,
0.24365234375
],
[
0.11517333984375,
0.231689453125,
0.2059326171875,
0.231689453125,
0.2059326171875,
0.24365234375,
0.11517333984375,
0.24365234375
],
[
0.330078125,
0.231689453125,
0.420654296875,
0.231689453125,
0.420654296875,
0.24365234375,
0.330078125,
0.24365234375
],
[
0.5439453125,
0.231689453125,
0.63623046875,
0.231689453125,
0.63623046875,
0.24365234375,
0.5439453125,
0.24365234375
],
[
0.160888671875,
0.279052734375,
0.393798828125,
0.279052734375,
0.393798828125,
0.2890625,
0.160888671875,
0.2890625
],
[
0.50244140625,
0.276611328125,
0.8896484375,
0.276611328125,
0.8896484375,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.1422119140625,
0.2890625,
0.16259765625,
0.291748046875,
0.160888671875,
0.2998046875,
0.1405029296875,
0.29736328125
],
[
0.2279052734375,
0.290283203125,
0.2442626953125,
0.290283203125,
0.2442626953125,
0.298583984375,
0.2279052734375,
0.298583984375
],
[
0.313720703125,
0.292236328125,
0.325927734375,
0.292236328125,
0.325927734375,
0.298583984375,
0.313720703125,
0.298583984375
],
[
0.38232421875,
0.291748046875,
0.42236328125,
0.291748046875,
0.42236328125,
0.298583984375,
0.38232421875,
0.298583984375
],
[
0.50244140625,
0.292236328125,
0.89208984375,
0.292236328125,
0.89208984375,
0.30615234375,
0.50244140625,
0.30615234375
],
[
0.50244140625,
0.30810546875,
0.771484375,
0.30810546875,
0.771484375,
0.3212890625,
0.50244140625,
0.3212890625
],
[
0.50341796875,
0.337158203125,
0.5947265625,
0.337158203125,
0.5947265625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.5107421875,
0.361083984375,
0.890625,
0.361083984375,
0.890625,
0.375,
0.5107421875,
0.375
],
[
0.53369140625,
0.373779296875,
0.890625,
0.373779296875,
0.890625,
0.386962890625,
0.53369140625,
0.386962890625
],
[
0.53369140625,
0.38818359375,
0.890625,
0.38818359375,
0.890625,
0.401611328125,
0.53369140625,
0.401611328125
],
[
0.53369140625,
0.401611328125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.53369140625,
0.415283203125
],
[
0.53515625,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.43115234375,
0.53515625,
0.43115234375
],
[
0.53515625,
0.43115234375,
0.8837890625,
0.43115234375,
0.8837890625,
0.4443359375,
0.53515625,
0.4443359375
],
[
0.5107421875,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.458984375,
0.5107421875,
0.458984375
],
[
0.53515625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.53515625,
0.473388671875
],
[
0.53515625,
0.473388671875,
0.8310546875,
0.473388671875,
0.8310546875,
0.4873046875,
0.53515625,
0.4873046875
],
[
0.5087890625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.50244140625,
0.5087890625,
0.50244140625
],
[
0.53515625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.53515625,
0.5166015625
],
[
0.080078125,
0.5087890625,
0.4697265625,
0.5087890625,
0.4697265625,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.53369140625,
0.5166015625,
0.81396484375,
0.51513671875,
0.81396484375,
0.52978515625,
0.53369140625,
0.53076171875
],
[
0.080078125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.080078125,
0.537109375
],
[
0.080078125,
0.5361328125,
0.40283203125,
0.5361328125,
0.40283203125,
0.5498046875,
0.080078125,
0.5498046875
],
[
0.5107421875,
0.5322265625,
0.89208984375,
0.5322265625,
0.89208984375,
0.54541015625,
0.5107421875,
0.54541015625
],
[
0.53369140625,
0.54443359375,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.53369140625,
0.55810546875
],
[
0.53515625,
0.5595703125,
0.890625,
0.5595703125,
0.890625,
0.57275390625,
0.53515625,
0.57275390625
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.53369140625,
0.5712890625,
0.890625,
0.57275390625,
0.890625,
0.5869140625,
0.53369140625,
0.5859375
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.5361328125,
0.58837890625,
0.611328125,
0.58837890625,
0.611328125,
0.5986328125,
0.5361328125,
0.5986328125
],
[
0.5107421875,
0.6015625,
0.76953125,
0.6015625,
0.76953125,
0.61572265625,
0.5107421875,
0.61572265625
],
[
0.765625,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.765625,
0.6142578125
],
[
0.53369140625,
0.61572265625,
0.8896484375,
0.61572265625,
0.8896484375,
0.62890625,
0.53369140625,
0.62890625
],
[
0.07843017578125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.07843017578125,
0.62060546875
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.53369140625,
0.6298828125,
0.72119140625,
0.62890625,
0.72119140625,
0.64306640625,
0.53369140625,
0.64453125
],
[
0.080078125,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.65087890625,
0.080078125,
0.65087890625
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50732421875,
0.64599609375,
0.890625,
0.64453125,
0.890625,
0.6591796875,
0.50732421875,
0.66015625
],
[
0.53369140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.53369140625,
0.673828125
],
[
0.080078125,
0.66748046875,
0.4697265625,
0.66748046875,
0.4697265625,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.53515625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.53515625,
0.6875
],
[
0.08087158203125,
0.6845703125,
0.36279296875,
0.6845703125,
0.36279296875,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.53515625,
0.6875,
0.72216796875,
0.6875,
0.72216796875,
0.70068359375,
0.53515625,
0.70068359375
],
[
0.098876953125,
0.69775390625,
0.468017578125,
0.69775390625,
0.468017578125,
0.7109375,
0.098876953125,
0.7109375
],
[
0.5107421875,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71630859375,
0.5107421875,
0.71630859375
],
[
0.08087158203125,
0.7119140625,
0.4697265625,
0.7119140625,
0.4697265625,
0.72607421875,
0.08087158203125,
0.72607421875
],
[
0.53515625,
0.71533203125,
0.89208984375,
0.71533203125,
0.89208984375,
0.72900390625,
0.53515625,
0.72900390625
],
[
0.080078125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.53515625,
0.73193359375,
0.7900390625,
0.73193359375,
0.7900390625,
0.7421875,
0.53515625,
0.7421875
],
[
0.080078125,
0.7421875,
0.468017578125,
0.7421875,
0.468017578125,
0.755859375,
0.080078125,
0.755859375
],
[
0.5107421875,
0.74560546875,
0.89208984375,
0.74560546875,
0.89208984375,
0.7587890625,
0.5107421875,
0.7587890625
],
[
0.080078125,
0.7568359375,
0.285888671875,
0.7568359375,
0.285888671875,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.53369140625,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.53369140625,
0.7734375
],
[
0.53515625,
0.7734375,
0.89208984375,
0.7734375,
0.89208984375,
0.787109375,
0.53515625,
0.787109375
],
[
0.07843017578125,
0.78515625,
0.1953125,
0.7861328125,
0.1953125,
0.80029296875,
0.07843017578125,
0.7998046875
],
[
0.53369140625,
0.787109375,
0.87158203125,
0.7861328125,
0.87158203125,
0.7998046875,
0.53369140625,
0.80029296875
],
[
0.51123046875,
0.8037109375,
0.890625,
0.8037109375,
0.890625,
0.814453125,
0.51123046875,
0.814453125
],
[
0.1004638671875,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.1004638671875,
0.82470703125
],
[
0.53515625,
0.81640625,
0.8896484375,
0.81640625,
0.8896484375,
0.830078125,
0.53515625,
0.830078125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.53369140625,
0.8291015625,
0.89208984375,
0.830078125,
0.89208984375,
0.8447265625,
0.53369140625,
0.84326171875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84033203125,
0.468017578125,
0.85400390625,
0.080078125,
0.85498046875
],
[
0.53515625,
0.8447265625,
0.6708984375,
0.8447265625,
0.6708984375,
0.85791015625,
0.53515625,
0.85791015625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.859375,
0.890625,
0.859375,
0.890625,
0.87255859375,
0.50244140625,
0.87255859375
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.53369140625,
0.8720703125,
0.89208984375,
0.87255859375,
0.89208984375,
0.88720703125,
0.53369140625,
0.88623046875
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.53369140625,
0.88623046875,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.53369140625,
0.8994140625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"Table 3. Effect of loss-based selection and random selection in each spurious correlation of Waterbirds dataset on worst and mean (in paren-",
"theses) group accuracy. Results are in percentage.CR-ML and CL-MR stand for Correctly classified with Random selection-Misclassified",
"withLoss-based selectionandMisclassified withLoss-based selection-Correctlyclassified withRandom selection respectively.While",
"when the spurious correlation is not lower than 90% LFR has the best performance, when the correlation is 70% or 60% random selection",
"of misclassified samples leads to a higher WGA",
"Method",
"70%",
"80%",
"95%",
"100%",
"99%",
"90%",
"LFR",
"79.67 (85.25)",
"83.81 (87.88)",
"86.45 (91.02)",
"86.85 (91.25)",
"85.31 (89.24)",
"86.29 (90.54)",
"CFR",
"75.45 (87.15)",
"78.62 (87.57)",
"75.09 (84.93)",
"67.45 (80.12)",
"78.76 (87.08)",
"82.09 (89.48)",
"83.80 (90.39)",
"83.80 (90.07)",
"81.09 (89.61)",
"LFR-CR-ML",
"83.36 (88.65)",
"82.68 (88.44)",
"80.84 (87.93)",
"80.37 (88.89)",
"83.96 (88.87)",
"83.96 (89.35)",
"84.65 (90.27)",
"LFR-CL-MR",
"84.27 (91.21)",
"86.45 (92.03)",
"Attention Maps on Minority Groups of Waterbirds and CelebA",
"and baselines, even works with the assumption of group an-",
"mage",
"ERM",
"AFR",
"Loss Based",
"notations availability, in addition to better performance in",
"attending core attributes in input images.",
"References",
"[1] Joy Buolamwini and Timnit Gebru. Gender shades: Inter-",
"sectional accuracy disparities in commercial gender classifi-",
"cation. In Sorelle A. Friedler and Christo Wilson, editors,",
"Proceedings of the Ist Conference on Fairness, Accountabil-",
"ity and Transparency, volume 81 of Proceedings of Machine",
"Learning Research, pages 77-91. PMLR, 23-24 Feb 2018.",
"[2] Alexandra Chouldechova. Fair prediction with disparate im-",
"pact: A study of bias in recidivism prediction instruments.",
"Big Data, 5(2):153-163, 2017. PMID: 28632438.",
"[3] Elliot Creager, Jorn-Henrik Jacobsen, and Richard S. Zemel.",
"Environment inference for invariant learning. In Interna-",
"Figure 3. Saliency map of models before (ERM) and after last",
"tional Conference on Machine Learning, 2020.",
"layer retraining with AFR and LFR. LFR has a better quality in",
"terms of pushing the model to attend to core attributes.",
"[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,",
"and Li Fei-Fei. Imagenet: A large-scale hierarchical im-",
"age database. In Computer Vision and Pattern Recognition,",
"underperforms in high spurious correlation settings because",
"2009.CVPR 2009.IEEE Conference on,pages 248-255.",
"they do not have any consideration for the number of ex-",
"IEEE.2009.",
"[5] Jacob Gildenblat and contributors.",
"Pytorch library for",
"cam methods.https://github.com/jacobgil/",
"amples from the majority group. So, even if the under-",
"weight samples with lower loss are probably from the ma-",
"pytorch-grad-cam, 2021.",
"jority group, the overall effect of the majority samples still",
"remains high. On the other hand, as LFR selects an equal",
"[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.",
"Deep residual learning for image recognition. In 2016 IEEE",
"number of samples from both the majority and minority, it",
"Conference on Computer Vision and Pattern Recognition",
"doesnothavetheaforementionedproblem.",
"(CVPR), pages 770-778, 2016.",
"We also visualize our method's attention map with Grad-",
"[7] P. Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson.",
"Cam [5] to compare it with other models. As it can be seen",
"Last layer re-training is sufficient for robustness to spurious",
"in figure 3, using LFR, the model attends more to core at-",
"correlations.ArXiv,abs/2204.029372022.",
"tributes, in contrast to the pre-trained ERM model and AFR.",
"[8] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar.",
"More results are available in C.",
"Dropout disagreement: A recipe for group robustness with",
"fewer annotations. In NeurlPs 2022 Workshop on Distribu-",
"5. Conclusion",
"tion Shifts: Connecting Methods and Applications,2022.",
"[9] Yoonho Lee, Huaxiu Yao, and Chelsea Finn. Diversify and",
"We proposed loss-based feature re-weighting, a method",
"disambiguate: Out-of-distribution robustness via disagree-",
"for selecting a few images from a small proportion of train-",
"ment. In The Eleventh International Conference on Learning",
"ing data to retrain the last linear of the model which im-",
"Representations, 2023.",
"proves robustness to spurious correlation. While LFR is fast",
"[10] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas,",
"and straightforward to use, it achieved SOTA in many vari-",
"Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and",
"ous datasets in comparison to previously proposed methods",
"Mark Ibrahim.A whac-a-mole dilemma: Shortcuts come",
"5"
] |
{
"arxiv_id": "2312.04893",
"language": "en",
"page_idx": 4
}
|
||
2302.02080-en-1
|
2302.02080-en-1.jpg
|
power of the old model to the new one on potential regression instances (Hinton et al.,2015). However, given the huge complexity of current neural architectures and relatively scarce training data in downstream tasks, models could have insufficient data to reliably estimate the probable regression cases and carry out the transfer on them (Xie et al., 2021;Cai et al., 2022). On the other hand, model ensemble aggregates predictions from differentlytrained new models but bears no connection with the legacy version (Yan et al., 2021). These limitations reveal the two major challenges when striving to ensure backward compatibility. First, the new model could have distinct inductive bias and prediction behavior than the old system, rooted from inherent differences such as architecture, model size, and pretraining procedure (Liu et al., 2021). Second, during new model training, a reliable mechanism is needed in place to bridge the gap between two models and mitigate potential inconsistencies.
Inspired by the strength and weakness of prior approaches, we propose Gated Fusion to integrate old and new models via gating mechanism (Hochreiter and Schmidhuber, 1997; Chung et al., 2014; Gu et al., 2016), essentially a light-weight ensemble of legacy and upgrade models connected via a learned fusion gate. Specifically, we add a learned gate on top of the new model. We combine the logits from old and new models according to the weight from the gate. We train our Gated Fusion model by minimizing the standard cross-entropy error. The intuition is that the gate could learn to put more weights on the old model when the new model cannot produce correct predictions, effectively doing fall-backs that optimizes backward compatibility. Empirical results demonstrate that our proposed
approach outperforms other competing methods significantly, where we can obtain on average $62\%$ reduction of total negative flips, i.e. new errors caused by the model upgrade, without any degradation in accuracy performance. The effectiveness of Gated Fusion is validated across three diverse clasSification tasks and two distinct model upgrade sce- this work, we take the development set as our $\mathcal{D}_{reg}$
narios (a) upgrade to larger model size (b) upgrade to advanced pretrained model, where consistent results are attained across the board.
Our main contributions are as follows:
$\cdot$ We propose Gated Fusion that integrates old and new models via gating mechanism for backward-compatible model upgrade;
$•$ We evaluate competing methods on two dis-
tinct and challenging model upgrade scenarios
across three diverse classification tasks;
$\bullet$ Empirical results show that our proposed approach significantly outperforms competing methods and achieves regression reductions by a large margin across the board.
# 2 The Backward-Compatible Model Upgrade Problem
The goal of backward-compatible model upgrade is to minimize regression errors without compromising the accuracy performance during model upgrade (Yan et al., 2021; Xie et al., 2021). In this work, we aim to improve the backward compatibility of model predictions in the NLP classification tasks. Following Xie et al. (2021), we study the scenario where the underlying pretrained language model (LM) is being upgraded.
Let $x$ be a natural language input with a class label $y\in\{1,2,...,C\}.~\tilde{\mathcal{D}}=\{x_i,y_i\}_{i=1}^N$ denotes a set of $N$ examples with corresponding labels. A classifer $f_\to$estimates the class probabilities given the input $\vec{f}(x)=(p(y=1|x),...,p(y=C|x))^\top.$ When upgrading from an old model $f_old$ to a new model $f_{new}$, normally with distinct architectures and trained on the same data, an improved model $f^*$ is produced based on $f_old$ and $f_new.$ Our goal is for $f^*$ to minimize regression errors as an additional objective, while still achieving comparable performance to $f_{new}^{o}$, the new model trained in the vanilla setting. Note that $f^*$ could be multiple times larger than $f_{new}^o$, with model ensemble of $f_{new}^o$ as one example (Yan et al., 2021).
Measuring Backward Compatibility. The backward compatibility is measured via quantifying regression errors on a given regression measurement set $\mathcal{D}_{reg}=\{x_{i},y_{i}\}_{i=1}^{M}.~\mathcal{D}_{reg}$ could be a hidden customer test set comprising critical use cases, a set of behavioral testing examples for targeted evaluation (Ribeiro et al., 2020), or the development split from the dataset of interest. In for evaluation.
For classification, regression errors are characterized by negative flips, denoted as $\mathcal{R}_NF-$the portion of samples in $\mathcal{D}_{reg}$ that flip from correct prediction $f_{old}(x_i)=y_i$ to incorrect output $f_{new}(x_i)\neq y_i$ during model upgrade:
$$\mathcal{R}_{NF}(\mathcal{D}_{reg},\vec{f}_{old},\vec{f}_{new})=\frac{|\{x|f_{old}=y,f_{new}\neq y\}|}{|\mathcal{D}_{reg}|}.$$
|
[
[
0.1158447265625,
0.08489990234375,
0.484375,
0.08489990234375,
0.484375,
0.09857177734375,
0.1158447265625,
0.09857177734375
],
[
0.54736328125,
0.08489990234375,
0.87890625,
0.08489990234375,
0.87890625,
0.09857177734375,
0.54736328125,
0.09857177734375
],
[
0.1175537109375,
0.10394287109375,
0.484375,
0.10394287109375,
0.484375,
0.114013671875,
0.1175537109375,
0.114013671875
],
[
0.54833984375,
0.10394287109375,
0.8388671875,
0.10394287109375,
0.8388671875,
0.114013671875,
0.54833984375,
0.114013671875
],
[
0.1175537109375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.53076171875,
0.1217041015625,
0.880859375,
0.122314453125,
0.880859375,
0.13720703125,
0.53076171875,
0.135986328125
],
[
0.1175537109375,
0.1353759765625,
0.48193359375,
0.1353759765625,
0.48193359375,
0.1455078125,
0.1175537109375,
0.1455078125
],
[
0.54736328125,
0.1395263671875,
0.880859375,
0.1395263671875,
0.880859375,
0.152587890625,
0.54736328125,
0.152587890625
],
[
0.1158447265625,
0.1484375,
0.484375,
0.1484375,
0.484375,
0.162109375,
0.1158447265625,
0.162109375
],
[
0.54833984375,
0.155029296875,
0.8818359375,
0.155029296875,
0.8818359375,
0.169189453125,
0.54833984375,
0.169189453125
],
[
0.1175537109375,
0.1650390625,
0.484375,
0.1650390625,
0.484375,
0.1787109375,
0.1175537109375,
0.1787109375
],
[
0.5458984375,
0.17041015625,
0.80615234375,
0.169189453125,
0.80615234375,
0.18408203125,
0.5458984375,
0.185302734375
],
[
0.118408203125,
0.18408203125,
0.486083984375,
0.18408203125,
0.486083984375,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.1175537109375,
0.19775390625,
0.48779296875,
0.19775390625,
0.48779296875,
0.21142578125,
0.1175537109375,
0.21142578125
],
[
0.5146484375,
0.201904296875,
0.5283203125,
0.201904296875,
0.5283203125,
0.2078857421875,
0.5146484375,
0.2078857421875
],
[
0.5400390625,
0.19775390625,
0.8388671875,
0.19775390625,
0.8388671875,
0.2120361328125,
0.5400390625,
0.2120361328125
],
[
0.1175537109375,
0.21435546875,
0.489501953125,
0.21435546875,
0.489501953125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.54150390625,
0.21435546875,
0.69873046875,
0.213134765625,
0.69873046875,
0.22802734375,
0.54150390625,
0.2286376953125
],
[
0.1158447265625,
0.23095703125,
0.486083984375,
0.23095703125,
0.486083984375,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.509765625,
0.2369384765625,
0.880859375,
0.2381591796875,
0.880859375,
0.2529296875,
0.509765625,
0.251708984375
],
[
0.1158447265625,
0.2464599609375,
0.490234375,
0.2452392578125,
0.490234375,
0.260009765625,
0.1158447265625,
0.26123046875
],
[
0.509765625,
0.254638671875,
0.88330078125,
0.254638671875,
0.88330078125,
0.26953125,
0.509765625,
0.26953125
],
[
0.1158447265625,
0.26318359375,
0.48779296875,
0.26318359375,
0.48779296875,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.51123046875,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.28515625,
0.51123046875,
0.28515625
],
[
0.1158447265625,
0.279052734375,
0.48779296875,
0.279052734375,
0.48779296875,
0.292236328125,
0.1158447265625,
0.292236328125
],
[
0.509765625,
0.2861328125,
0.8818359375,
0.28515625,
0.8818359375,
0.2998046875,
0.509765625,
0.300537109375
],
[
0.1175537109375,
0.295654296875,
0.490234375,
0.295654296875,
0.490234375,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51318359375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51318359375,
0.31640625
],
[
0.1175537109375,
0.311279296875,
0.489501953125,
0.311279296875,
0.489501953125,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.5087890625,
0.317626953125,
0.880859375,
0.317626953125,
0.880859375,
0.330810546875,
0.5087890625,
0.330810546875
],
[
0.1175537109375,
0.328857421875,
0.486083984375,
0.328857421875,
0.486083984375,
0.339111328125,
0.1175537109375,
0.339111328125
],
[
0.509765625,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.509765625,
0.3486328125
],
[
0.11419677734375,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.11419677734375,
0.355712890625
],
[
0.51123046875,
0.349853515625,
0.880859375,
0.350830078125,
0.880859375,
0.365234375,
0.51123046875,
0.364013671875
],
[
0.1158447265625,
0.359375,
0.489501953125,
0.358154296875,
0.489501953125,
0.372314453125,
0.1158447265625,
0.37353515625
],
[
0.51318359375,
0.36767578125,
0.744140625,
0.36767578125,
0.744140625,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.1158447265625,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.1158447265625,
0.3896484375
],
[
0.52880859375,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.52880859375,
0.39794921875
],
[
0.1158447265625,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.1158447265625,
0.40625
],
[
0.51318359375,
0.399658203125,
0.880859375,
0.399658203125,
0.880859375,
0.41455078125,
0.51318359375,
0.41455078125
],
[
0.133544921875,
0.4091796875,
0.486083984375,
0.4091796875,
0.486083984375,
0.4228515625,
0.133544921875,
0.4228515625
],
[
0.51123046875,
0.4150390625,
0.8818359375,
0.41455078125,
0.8818359375,
0.4287109375,
0.51123046875,
0.429931640625
],
[
0.1158447265625,
0.42578125,
0.484375,
0.42578125,
0.484375,
0.439453125,
0.1158447265625,
0.439453125
],
[
0.51123046875,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.1158447265625,
0.441162109375,
0.48779296875,
0.441162109375,
0.48779296875,
0.454833984375,
0.1158447265625,
0.454833984375
],
[
0.50732421875,
0.4453125,
0.884765625,
0.446533203125,
0.884765625,
0.46435546875,
0.50732421875,
0.463134765625
],
[
0.1158447265625,
0.457275390625,
0.483642578125,
0.457275390625,
0.483642578125,
0.47021484375,
0.1158447265625,
0.47021484375
],
[
0.509765625,
0.463134765625,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.509765625,
0.47802734375
],
[
0.1175537109375,
0.473876953125,
0.48779296875,
0.473876953125,
0.48779296875,
0.48681640625,
0.1175537109375,
0.48681640625
],
[
0.51123046875,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.49462890625,
0.51123046875,
0.49462890625
],
[
0.1158447265625,
0.4892578125,
0.48779296875,
0.488037109375,
0.48779296875,
0.5029296875,
0.1158447265625,
0.50439453125
],
[
0.51123046875,
0.495361328125,
0.8818359375,
0.495361328125,
0.8818359375,
0.5087890625,
0.51123046875,
0.5087890625
],
[
0.1158447265625,
0.505859375,
0.486083984375,
0.505859375,
0.486083984375,
0.51953125,
0.1158447265625,
0.51953125
],
[
0.515625,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.52685546875,
0.515625,
0.52685546875
],
[
0.1175537109375,
0.521484375,
0.48779296875,
0.521484375,
0.48779296875,
0.53515625,
0.1175537109375,
0.53515625
],
[
0.509765625,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.5419921875,
0.509765625,
0.5419921875
],
[
0.1175537109375,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.55029296875,
0.1175537109375,
0.55029296875
],
[
0.51318359375,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.51318359375,
0.55810546875
],
[
0.1158447265625,
0.552734375,
0.48779296875,
0.55419921875,
0.48779296875,
0.56787109375,
0.1158447265625,
0.56640625
],
[
0.51123046875,
0.56005859375,
0.8818359375,
0.55859375,
0.8818359375,
0.57373046875,
0.51123046875,
0.57470703125
],
[
0.1158447265625,
0.568359375,
0.48779296875,
0.5693359375,
0.48779296875,
0.58447265625,
0.1158447265625,
0.5830078125
],
[
0.51318359375,
0.57666015625,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.1158447265625,
0.5849609375,
0.486083984375,
0.5849609375,
0.486083984375,
0.5986328125,
0.1158447265625,
0.5986328125
],
[
0.509765625,
0.5927734375,
0.8818359375,
0.59130859375,
0.8818359375,
0.60546875,
0.509765625,
0.60693359375
],
[
0.1175537109375,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.1175537109375,
0.615234375
],
[
0.5146484375,
0.609375,
0.8037109375,
0.60693359375,
0.8037109375,
0.62158203125,
0.5146484375,
0.62353515625
],
[
0.1175537109375,
0.6181640625,
0.486083984375,
0.6181640625,
0.486083984375,
0.6318359375,
0.1175537109375,
0.6318359375
],
[
0.118408203125,
0.63427734375,
0.480224609375,
0.63427734375,
0.480224609375,
0.64794921875,
0.118408203125,
0.64794921875
],
[
0.51123046875,
0.6298828125,
0.8818359375,
0.6298828125,
0.8818359375,
0.64453125,
0.51123046875,
0.64453125
],
[
0.1351318359375,
0.6494140625,
0.48779296875,
0.65087890625,
0.48779296875,
0.6650390625,
0.1351318359375,
0.66455078125
],
[
0.51123046875,
0.646484375,
0.88330078125,
0.646484375,
0.88330078125,
0.6611328125,
0.51123046875,
0.6611328125
],
[
0.1158447265625,
0.66748046875,
0.48779296875,
0.66650390625,
0.48779296875,
0.68115234375,
0.1158447265625,
0.681640625
],
[
0.509765625,
0.662109375,
0.880859375,
0.662109375,
0.880859375,
0.67578125,
0.509765625,
0.67578125
],
[
0.509765625,
0.677734375,
0.88330078125,
0.677734375,
0.88330078125,
0.6953125,
0.509765625,
0.6953125
],
[
0.118408203125,
0.68505859375,
0.484375,
0.68505859375,
0.484375,
0.6953125,
0.118408203125,
0.6953125
],
[
0.1175537109375,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.1175537109375,
0.71337890625
],
[
0.51123046875,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.51123046875,
0.71044921875
],
[
0.509765625,
0.71044921875,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.509765625,
0.724609375
],
[
0.1175537109375,
0.71630859375,
0.490234375,
0.71630859375,
0.490234375,
0.72998046875,
0.1175537109375,
0.72998046875
],
[
0.509765625,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.74072265625,
0.509765625,
0.74072265625
],
[
0.1175537109375,
0.732421875,
0.489501953125,
0.732421875,
0.489501953125,
0.7451171875,
0.1175537109375,
0.7451171875
],
[
0.1175537109375,
0.74755859375,
0.490234375,
0.74755859375,
0.490234375,
0.76123046875,
0.1175537109375,
0.76123046875
],
[
0.51123046875,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.51123046875,
0.755859375
],
[
0.1158447265625,
0.76318359375,
0.48779296875,
0.76318359375,
0.48779296875,
0.77685546875,
0.1158447265625,
0.77685546875
],
[
0.509765625,
0.75830078125,
0.8818359375,
0.76025390625,
0.8818359375,
0.77490234375,
0.509765625,
0.7724609375
],
[
0.1175537109375,
0.78076171875,
0.489501953125,
0.78076171875,
0.489501953125,
0.79443359375,
0.1175537109375,
0.79443359375
],
[
0.51318359375,
0.775390625,
0.62060546875,
0.775390625,
0.62060546875,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.5283203125,
0.79052734375,
0.8818359375,
0.79150390625,
0.8818359375,
0.8056640625,
0.5283203125,
0.8046875
],
[
0.1158447265625,
0.79638671875,
0.48779296875,
0.79638671875,
0.48779296875,
0.81005859375,
0.1158447265625,
0.81005859375
],
[
0.1158447265625,
0.8125,
0.38623046875,
0.8125,
0.38623046875,
0.826171875,
0.1158447265625,
0.826171875
],
[
0.51123046875,
0.80712890625,
0.8818359375,
0.8056640625,
0.8818359375,
0.82080078125,
0.51123046875,
0.82177734375
],
[
0.509765625,
0.82373046875,
0.88330078125,
0.822265625,
0.88330078125,
0.83740234375,
0.509765625,
0.83837890625
],
[
0.1351318359375,
0.830078125,
0.420654296875,
0.830078125,
0.420654296875,
0.84326171875,
0.1351318359375,
0.84326171875
],
[
0.51318359375,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.1376953125,
0.8515625,
0.48779296875,
0.8515625,
0.48779296875,
0.865234375,
0.1376953125,
0.865234375
],
[
0.51123046875,
0.85693359375,
0.79443359375,
0.85546875,
0.79443359375,
0.87060546875,
0.51123046875,
0.87158203125
],
[
0.15283203125,
0.8681640625,
0.48779296875,
0.8681640625,
0.48779296875,
0.8818359375,
0.15283203125,
0.8818359375
],
[
0.509765625,
0.87744140625,
0.873046875,
0.8681640625,
0.875,
0.8935546875,
0.51123046875,
0.90283203125
],
[
0.151123046875,
0.88427734375,
0.43408203125,
0.88427734375,
0.43408203125,
0.89794921875,
0.151123046875,
0.89794921875
],
[
0.76318359375,
0.8935546875,
0.8095703125,
0.8935546875,
0.8095703125,
0.9091796875,
0.76318359375,
0.9091796875
],
[
0.1351318359375,
0.90087890625,
0.490234375,
0.90185546875,
0.490234375,
0.91943359375,
0.1351318359375,
0.91845703125
],
[
0.85791015625,
0.90478515625,
0.88330078125,
0.90478515625,
0.88330078125,
0.91943359375,
0.85791015625,
0.91943359375
]
] |
[
"power of the old model to the new one on potential",
"tinct and challenging model upgrade scenarios",
"regressioninstances (Hinton et al.,2015).How",
"acrossthreediverseclassificationtasks",
"ever, given the huge complexity of current neural",
".Empirical results show that our proposed ap",
"architectures and relatively scarce training data ir",
"proach significantly outperforms competing",
"downstream tasks, models could have insufficien",
"methods and achieves regression reductions",
"data to reliably estimate the probable regression",
"by a large margin across the board.",
"cases and carry out the transfer on them (Xie et al..",
"2021; Cai et al., 2022). On the other hand, model",
"",
"The Backward-Compatible Model",
"ensemble aggregates predictions from differently-",
"UpgradeProblem",
"trained new models but bears no connection with",
"The goal of backward-compatible model upgrade",
"the legacy version (Yan et al., 2021). These limita-",
"is to minimize regression errors without compro-",
"tions reveal the two major challenges when striving",
"mising the accuracy performance during model up-",
"to ensure backward compatibility. First, the new",
"grade (Yan et al., 2021; Xie et al., 2021). In this",
"model could have distinct inductive bias and predic-",
"work, we aim to improve the backward compatibil-",
"tion behavior than the old system, rooted from in-",
"ity of model predictions in the NLP classification",
"herentdifferencessuchasarchitecture,modelsize.",
"tasks. Following Xie et al. (2021), we study the",
"and pretraining procedure (Liu et al., 2021). Sec-",
"scenario where the underlying pretrained language",
"ond, during new model training, a reliable mecha-",
"model (LM) is being upgraded.",
"nism is needed in place to bridge the gap between",
"Let c be a natural language input with a class",
"two models and mitigate potential inconsistencies.",
"label y E [1, 2, ..,C]. D = [ci, yili denotes",
"Inspired by the strength and weakness of prior",
"a set of N examples with corresponding labels. A",
"approaches, we propose Gated Fusion to integrate",
"classifier f estimates the class probabilities given",
"old and new models via gating mechanism (Hochre-",
"the input f(α) = (p(y = 1|c),..,p(y = Clc)T.",
"iter and Schmidhuber, 1997; Chung et al., 2014; Gu",
"When upgrading from an old model fold to a new",
"et al., 2016), essentially a light-weight ensemble of",
"model fnew, normally with distinct architectures",
"legacy and upgrade models connected via a learned",
"and trained on the same data, an improved model",
"fusion gate. Specifically, we add a learned gate",
"f* is produced based on fold and fnew. Our goal",
"on top of the new model. We combine the logits",
"is for f* to minimize regression errors as an addi-",
"from old and new models according to the weight",
"tional objective, while still achieving comparable",
"from the gate. We train our Gated Fusion model by",
"performance to fnew, the new model trained in",
"minimizing the standard cross-entropy error: The",
"the vanilla setting. Note that f* could be multiple",
"intuition is that the gate could learn to put more",
"times larger than fnew, with model ensemble of",
"weights on the old model when the new model can-",
"fnew as one example (Yan et al., 2021).",
"not produce correct predictions, effectively doing",
"fall-backs that optimizes backward compatibility.",
"Measuring Backward Compatibility. The",
"Empirical results demonstrate that our proposed",
"backward compatibility is measured via quan-",
"approach outperforms other competing methods",
"tifying regression errors on a given regression",
"measurement set Dreg = [ci, Yil1. Dreg could",
"significantly.wherewecanobtainonaverage62%",
"reduction of total negative flips, i.e. new errors",
"be a hidden customer test set comprising critical",
"use cases, a set of behavioral testing examples for",
"caused by the model upgrade, without any degrada-",
"targeted evaluation (Ribeiro et al., 2020), or the",
"tion in accuracy performance. The effectiveness of",
"Gated Fusion is validated across three diverse clas-",
"development split from the dataset of interest. In",
"sification tasks and two distinct model upgrade sce-",
"this work, we take the development set as our Dreg",
"narios (a) upgrade to larger model size (b) upgrade",
"forevaluation.",
"For classification, regression errors are charac-",
"to advanced pretrained model, where consistent",
"results are attained across the board.",
"terized by negative flips, denoted as RNF - the",
"portion of samples in Dreg that flip from cor-",
"Our main contributions are as follows:",
"rect prediction fold(ci) = yi to incorrect output",
": We propose Gated Fusion that integrates old",
"fnew(ci) ≠ yi during model upgrade:",
"and new models via gating mechanism for",
"RNF(Dreg fald, fnew)=alfold=y, fnew y",
"backward-compatible model upgrade;",
"[Dregl",
". We evaluate competing methods on two dis-",
"(1)"
] |
{
"arxiv_id": "2302.02080",
"language": "en",
"page_idx": 1
}
|
||
2302.02080-en-2
|
2302.02080-en-2.jpg
|
<FigureHere>
Figure 2: Methods to improve prediction backward compatibility during model upgrade.(a) Distillation-based approach to align predicted logits on potential regression instances (Xie et al. 2021). (b) Ensemble of old and new models via weighted sum of either predicted logits or probabilities. (c) Our proposed Gated Fusion that learns a gate as a soft switch to dynamically determine whether to fall back to previous predictions.
One thing to emphasize is that maximizing classifier performance does not necessarily help in minimizing $\mathcal{R}_NF$ (Yan et al., 2021; Xie et al., 2021).
models as our final outputs:
$$l_{GF}^*(y|x)=(1-\alpha_{gate})\cdot\frac{l_{old}(y|x)}{T}+\alpha_{gate}\cdot l_{new}(y|x),$$
(3)
# 3 Gated Fusion: Methodology
### 3.1 Method Overview
where $l(y|x)$ denotes predicted logits from models and $T$ is the temperature scaling to regularize the magnitude of old model's logits. $f_{new}$ and $g_\theta$ are then jointly trained end-to-end with cross-entropy loss between our output logits $l_GF^*(y|x)$ and label distributions on downstream tasks.
The intuition behind Gated Fusion is that when $f_{new}$ makes a mistake while $f_old$ produces the correct output, the gate $g_\theta$ will learn to put more weight on $f_{old}$ in order to minimize the final classification loss. This process effectively mitigates potential negative flips introduced by the model upgrade and thus improves the backward compatibility of final predictions.
To improve backward compatibility in model upgrade, it's crucial to have a mechanism that detects potential regression errors and mitigates them when making predictions. We propose Gated Fusion (GF) to achieve this by learning a gate as a soft switch to choose between generating predictions by the new model or resorting to outputs of the old model. Gated Fusion is inspired by the gating mechanism widely used in other applications. For example, mixing word copying mode with word generation mode for language modeling (Merity et al., 2016) and summarization (See et al., 2017). Our proposed Gated Fusion $f_{GF}^*$ consists of the old model $f_{old}$,the new model $f_new$, and a gating network $g_\theta.$ The old model $f_old$ is the legacy version before upgrade where the parameters are fixed. The new model $f_{new}$ has the same architecture as $f_{new}^o$ and is randomly initialized. The gating network $g_\theta$ is a multi-layer feed-forward network with sigmoid function. It produces a scalar weight $\alpha_{gate}$ in the range [0,1] from the output layer of $f_{new}$, denoted as $\mathcal{E}_{new}$:
## 3.2 Training and Inference
In practice, training Gated Fusion with randomly initialized $f_{new}$ would make the shallow gating network quickly converge to favor the fully-trained $f_{old}.$ To prevent this, we only train $f_new$ for the first few epochs to ensure its competence before jointly training $g_\theta$ and $f_{new}$ using $l_GF^*(x).$ In addition, we found that stopping gradient flow from $g_{\theta}$ to $f_{new}$ can further prevent the performance decrease of the new model within Gated Fusion:
$$\alpha_{gate}(x)=g_{\theta}(stop_grad(\mathcal{E}_{new}(x))).$$
(4)
(2)
$$\alpha_{gate}(x)=g_{\theta}(\mathcal{E}_{new}(x)).$$
At inference time, Gated Fusion produces logits from $f_{old}$ and $f_{new}$ as well as the gate value $\alpha_gate$
We use $\alpha_{gate}$ to combine the logits of old and new
|
[
[
0.7001953125,
0.0843505859375,
0.767578125,
0.0843505859375,
0.767578125,
0.09857177734375,
0.7001953125,
0.09857177734375
],
[
0.189697265625,
0.09979248046875,
0.35107421875,
0.09979248046875,
0.35107421875,
0.114013671875,
0.189697265625,
0.114013671875
],
[
0.44580078125,
0.1015625,
0.55322265625,
0.1015625,
0.55322265625,
0.1134033203125,
0.44580078125,
0.1134033203125
],
[
0.7255859375,
0.09979248046875,
0.73876953125,
0.09979248046875,
0.73876953125,
0.105712890625,
0.7255859375,
0.105712890625
],
[
0.67822265625,
0.105712890625,
0.7900390625,
0.105712890625,
0.7900390625,
0.12054443359375,
0.67822265625,
0.12054443359375
],
[
0.30322265625,
0.11639404296875,
0.322509765625,
0.11639404296875,
0.322509765625,
0.12353515625,
0.30322265625,
0.12353515625
],
[
0.497802734375,
0.11224365234375,
0.5087890625,
0.12054443359375,
0.50146484375,
0.1258544921875,
0.489501953125,
0.1181640625
],
[
0.444091796875,
0.1312255859375,
0.47265625,
0.1312255859375,
0.47265625,
0.1395263671875,
0.444091796875,
0.1395263671875
],
[
0.2493896484375,
0.1175537109375,
0.26123046875,
0.1175537109375,
0.26123046875,
0.12054443359375,
0.2493896484375,
0.12054443359375
],
[
0.189697265625,
0.1318359375,
0.1956787109375,
0.1318359375,
0.1956787109375,
0.13720703125,
0.189697265625,
0.13720703125
],
[
0.28955078125,
0.133056640625,
0.30908203125,
0.133056640625,
0.30908203125,
0.13720703125,
0.28955078125,
0.13720703125
],
[
0.32666015625,
0.133056640625,
0.339111328125,
0.133056640625,
0.339111328125,
0.13720703125,
0.32666015625,
0.13720703125
],
[
0.5283203125,
0.1312255859375,
0.5498046875,
0.1312255859375,
0.5498046875,
0.141357421875,
0.5283203125,
0.141357421875
],
[
0.6572265625,
0.133056640625,
0.70947265625,
0.133056640625,
0.70947265625,
0.1436767578125,
0.6572265625,
0.1436767578125
],
[
0.74072265625,
0.1353759765625,
0.78271484375,
0.1353759765625,
0.78271484375,
0.146728515625,
0.74072265625,
0.146728515625
],
[
0.20654296875,
0.142578125,
0.214111328125,
0.142578125,
0.214111328125,
0.1455078125,
0.20654296875,
0.1455078125
],
[
0.43994140625,
0.142578125,
0.45166015625,
0.142578125,
0.45166015625,
0.1455078125,
0.43994140625,
0.1455078125
],
[
0.5458984375,
0.142578125,
0.55908203125,
0.142578125,
0.55908203125,
0.1455078125,
0.5458984375,
0.1455078125
],
[
0.6767578125,
0.146728515625,
0.68701171875,
0.146728515625,
0.68701171875,
0.149658203125,
0.6767578125,
0.149658203125
],
[
0.172119140625,
0.221435546875,
0.246826171875,
0.221435546875,
0.246826171875,
0.236328125,
0.172119140625,
0.236328125
],
[
0.411376953125,
0.223876953125,
0.48193359375,
0.223876953125,
0.48193359375,
0.235107421875,
0.411376953125,
0.235107421875
],
[
0.51318359375,
0.223876953125,
0.58837890625,
0.223876953125,
0.58837890625,
0.235107421875,
0.51318359375,
0.235107421875
],
[
0.6455078125,
0.221435546875,
0.7197265625,
0.221435546875,
0.7197265625,
0.236328125,
0.6455078125,
0.236328125
],
[
0.276123046875,
0.235107421875,
0.355224609375,
0.235107421875,
0.355224609375,
0.2464599609375,
0.276123046875,
0.2464599609375
],
[
0.75,
0.235107421875,
0.82861328125,
0.235107421875,
0.82861328125,
0.2464599609375,
0.75,
0.2464599609375
],
[
0.420654296875,
0.236328125,
0.47265625,
0.236328125,
0.47265625,
0.25048828125,
0.420654296875,
0.25048828125
],
[
0.183837890625,
0.236328125,
0.2359619140625,
0.236328125,
0.2359619140625,
0.25048828125,
0.183837890625,
0.25048828125
],
[
0.5224609375,
0.2369384765625,
0.58447265625,
0.2369384765625,
0.58447265625,
0.248779296875,
0.5224609375,
0.248779296875
],
[
0.6572265625,
0.2369384765625,
0.7080078125,
0.2369384765625,
0.7080078125,
0.248779296875,
0.6572265625,
0.248779296875
],
[
0.313232421875,
0.251708984375,
0.32080078125,
0.254638671875,
0.316650390625,
0.260009765625,
0.30908203125,
0.257080078125
],
[
0.444091796875,
0.248779296875,
0.456787109375,
0.2529296875,
0.44921875,
0.26318359375,
0.4375,
0.2587890625
],
[
0.54736328125,
0.254150390625,
0.55517578125,
0.254150390625,
0.55517578125,
0.25830078125,
0.54736328125,
0.25830078125
],
[
0.67919921875,
0.254150390625,
0.6884765625,
0.254150390625,
0.6884765625,
0.25830078125,
0.67919921875,
0.25830078125
],
[
0.2418212890625,
0.270751953125,
0.2822265625,
0.270751953125,
0.2822265625,
0.281982421875,
0.2418212890625,
0.281982421875
],
[
0.202392578125,
0.2529296875,
0.21240234375,
0.2493896484375,
0.2166748046875,
0.257080078125,
0.20654296875,
0.260009765625
],
[
0.474365234375,
0.270751953125,
0.5205078125,
0.270751953125,
0.5205078125,
0.281982421875,
0.474365234375,
0.281982421875
],
[
0.7138671875,
0.270751953125,
0.755859375,
0.270751953125,
0.755859375,
0.281982421875,
0.7138671875,
0.281982421875
],
[
0.426513671875,
0.29150390625,
0.57275390625,
0.29150390625,
0.57275390625,
0.3046875,
0.426513671875,
0.3046875
],
[
0.67822265625,
0.29150390625,
0.7900390625,
0.289306640625,
0.7900390625,
0.303955078125,
0.67822265625,
0.305908203125
],
[
0.1175537109375,
0.317626953125,
0.8818359375,
0.317626953125,
0.8818359375,
0.330810546875,
0.1175537109375,
0.330810546875
],
[
0.219970703125,
0.292236328125,
0.306396484375,
0.292236328125,
0.306396484375,
0.303955078125,
0.219970703125,
0.303955078125
],
[
0.11419677734375,
0.330810546875,
0.8818359375,
0.330078125,
0.8818359375,
0.344482421875,
0.11419677734375,
0.345703125
],
[
0.118408203125,
0.347412109375,
0.880859375,
0.347412109375,
0.880859375,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.1175537109375,
0.35986328125,
0.73046875,
0.35986328125,
0.73046875,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.118408203125,
0.399658203125,
0.495361328125,
0.399658203125,
0.495361328125,
0.413330078125,
0.118408203125,
0.413330078125
],
[
0.496337890625,
0.399658203125,
0.71728515625,
0.399658203125,
0.71728515625,
0.413330078125,
0.496337890625,
0.413330078125
],
[
0.1175537109375,
0.416259765625,
0.489501953125,
0.416259765625,
0.489501953125,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.67822265625,
0.419189453125,
0.74658203125,
0.419189453125,
0.74658203125,
0.43408203125,
0.67822265625,
0.43408203125
],
[
0.51318359375,
0.427490234375,
0.67529296875,
0.427490234375,
0.67529296875,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.748046875,
0.427490234375,
0.880859375,
0.427490234375,
0.880859375,
0.441162109375,
0.748046875,
0.441162109375
],
[
0.70947265625,
0.435791015625,
0.7197265625,
0.439453125,
0.71533203125,
0.44775390625,
0.705078125,
0.44482421875
],
[
0.85791015625,
0.4453125,
0.8818359375,
0.4453125,
0.8818359375,
0.457763671875,
0.85791015625,
0.457763671875
],
[
0.11419677734375,
0.431640625,
0.470947265625,
0.43115234375,
0.470947265625,
0.4453125,
0.11419677734375,
0.446533203125
],
[
0.11419677734375,
0.457763671875,
0.3896484375,
0.460205078125,
0.3896484375,
0.474365234375,
0.11419677734375,
0.47265625
],
[
0.51318359375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.51318359375,
0.473876953125
],
[
0.509765625,
0.474365234375,
0.8818359375,
0.4755859375,
0.8818359375,
0.490478515625,
0.509765625,
0.4892578125
],
[
0.1175537109375,
0.486328125,
0.301513671875,
0.486328125,
0.301513671875,
0.49951171875,
0.1175537109375,
0.49951171875
],
[
0.51123046875,
0.4912109375,
0.8818359375,
0.4921875,
0.8818359375,
0.50732421875,
0.51123046875,
0.505859375
],
[
0.1175537109375,
0.50732421875,
0.49365234375,
0.50732421875,
0.49365234375,
0.52099609375,
0.1175537109375,
0.52099609375
],
[
0.51123046875,
0.5087890625,
0.880859375,
0.5087890625,
0.880859375,
0.5224609375,
0.51123046875,
0.5224609375
],
[
0.1175537109375,
0.52392578125,
0.491943359375,
0.52392578125,
0.491943359375,
0.53759765625,
0.1175537109375,
0.53759765625
],
[
0.5087890625,
0.5224609375,
0.88330078125,
0.521484375,
0.88330078125,
0.5390625,
0.5087890625,
0.54052734375
],
[
0.1175537109375,
0.54052734375,
0.48779296875,
0.54052734375,
0.48779296875,
0.55419921875,
0.1175537109375,
0.55419921875
],
[
0.51318359375,
0.54052734375,
0.767578125,
0.54052734375,
0.767578125,
0.55419921875,
0.51318359375,
0.55419921875
],
[
0.11419677734375,
0.5546875,
0.48779296875,
0.5546875,
0.48779296875,
0.568359375,
0.11419677734375,
0.568359375
],
[
0.53076171875,
0.55810546875,
0.880859375,
0.55810546875,
0.880859375,
0.568359375,
0.53076171875,
0.568359375
],
[
0.1175537109375,
0.57177734375,
0.489501953125,
0.57177734375,
0.489501953125,
0.5859375,
0.1175537109375,
0.5859375
],
[
0.51123046875,
0.57373046875,
0.88330078125,
0.57177734375,
0.88330078125,
0.5859375,
0.51123046875,
0.58837890625
],
[
0.1175537109375,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6025390625,
0.1175537109375,
0.6025390625
],
[
0.5087890625,
0.59033203125,
0.8818359375,
0.58837890625,
0.8818359375,
0.6015625,
0.5087890625,
0.60400390625
],
[
0.1158447265625,
0.6025390625,
0.489501953125,
0.60400390625,
0.489501953125,
0.6181640625,
0.1158447265625,
0.61767578125
],
[
0.509765625,
0.60546875,
0.88330078125,
0.60546875,
0.88330078125,
0.619140625,
0.509765625,
0.619140625
],
[
0.1158447265625,
0.6181640625,
0.48779296875,
0.62060546875,
0.48779296875,
0.63427734375,
0.1158447265625,
0.6318359375
],
[
0.509765625,
0.62158203125,
0.8818359375,
0.62158203125,
0.8818359375,
0.63525390625,
0.509765625,
0.63525390625
],
[
0.1175537109375,
0.6357421875,
0.489501953125,
0.6357421875,
0.489501953125,
0.6494140625,
0.1175537109375,
0.6494140625
],
[
0.509765625,
0.63720703125,
0.8818359375,
0.6357421875,
0.8818359375,
0.65087890625,
0.509765625,
0.65185546875
],
[
0.1158447265625,
0.65185546875,
0.48779296875,
0.65087890625,
0.48779296875,
0.6650390625,
0.1158447265625,
0.66650390625
],
[
0.51123046875,
0.65380859375,
0.8818359375,
0.65380859375,
0.8818359375,
0.66748046875,
0.51123046875,
0.66748046875
],
[
0.1175537109375,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.51123046875,
0.66943359375,
0.7021484375,
0.66943359375,
0.7021484375,
0.68310546875,
0.51123046875,
0.68310546875
],
[
0.1158447265625,
0.68408203125,
0.483642578125,
0.68408203125,
0.483642578125,
0.69775390625,
0.1158447265625,
0.69775390625
],
[
0.509765625,
0.69140625,
0.73876953125,
0.6923828125,
0.73876953125,
0.71044921875,
0.509765625,
0.708984375
],
[
0.1368408203125,
0.70068359375,
0.486083984375,
0.70068359375,
0.486083984375,
0.71533203125,
0.1368408203125,
0.71533203125
],
[
0.1175537109375,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.732421875,
0.1175537109375,
0.732421875
],
[
0.509765625,
0.71435546875,
0.880859375,
0.71533203125,
0.880859375,
0.72998046875,
0.509765625,
0.728515625
],
[
0.1158447265625,
0.73291015625,
0.489501953125,
0.73291015625,
0.489501953125,
0.74755859375,
0.1158447265625,
0.74755859375
],
[
0.509765625,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.74658203125,
0.509765625,
0.74658203125
],
[
0.1175537109375,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76318359375,
0.1175537109375,
0.76318359375
],
[
0.5087890625,
0.74755859375,
0.8818359375,
0.74658203125,
0.8818359375,
0.76123046875,
0.5087890625,
0.7626953125
],
[
0.11419677734375,
0.7626953125,
0.490234375,
0.76318359375,
0.490234375,
0.78076171875,
0.11419677734375,
0.77978515625
],
[
0.51318359375,
0.76416015625,
0.8818359375,
0.76416015625,
0.8818359375,
0.779296875,
0.51318359375,
0.779296875
],
[
0.118408203125,
0.78076171875,
0.491943359375,
0.779296875,
0.491943359375,
0.79443359375,
0.118408203125,
0.79638671875
],
[
0.5087890625,
0.779296875,
0.880859375,
0.779296875,
0.880859375,
0.7919921875,
0.5087890625,
0.7919921875
],
[
0.1158447265625,
0.79736328125,
0.489501953125,
0.79736328125,
0.489501953125,
0.81103515625,
0.1158447265625,
0.81103515625
],
[
0.509765625,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81103515625,
0.509765625,
0.81103515625
],
[
0.11419677734375,
0.81005859375,
0.490234375,
0.8125,
0.490234375,
0.830078125,
0.11419677734375,
0.82763671875
],
[
0.509765625,
0.80859375,
0.8818359375,
0.81103515625,
0.8818359375,
0.8291015625,
0.509765625,
0.82666015625
],
[
0.1158447265625,
0.82763671875,
0.48779296875,
0.830078125,
0.48779296875,
0.84423828125,
0.1158447265625,
0.8427734375
],
[
0.51318359375,
0.8291015625,
0.8818359375,
0.8291015625,
0.8818359375,
0.8427734375,
0.51318359375,
0.8427734375
],
[
0.1158447265625,
0.8427734375,
0.246826171875,
0.8466796875,
0.246826171875,
0.86083984375,
0.1158447265625,
0.85693359375
],
[
0.51318359375,
0.84423828125,
0.755859375,
0.84423828125,
0.755859375,
0.85791015625,
0.51318359375,
0.85791015625
],
[
0.5634765625,
0.86767578125,
0.82861328125,
0.865234375,
0.82861328125,
0.8798828125,
0.5634765625,
0.8818359375
],
[
0.4609375,
0.87353515625,
0.490234375,
0.87353515625,
0.490234375,
0.89013671875,
0.4609375,
0.89013671875
],
[
0.85791015625,
0.869140625,
0.880859375,
0.869140625,
0.880859375,
0.8798828125,
0.85791015625,
0.8798828125
],
[
0.210693359375,
0.8759765625,
0.387939453125,
0.87451171875,
0.387939453125,
0.8896484375,
0.210693359375,
0.89013671875
],
[
0.52880859375,
0.88720703125,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.52880859375,
0.90185546875
],
[
0.11419677734375,
0.90185546875,
0.490234375,
0.90283203125,
0.490234375,
0.92041015625,
0.11419677734375,
0.91943359375
],
[
0.509765625,
0.90185546875,
0.8818359375,
0.90380859375,
0.8818359375,
0.92138671875,
0.509765625,
0.91943359375
]
] |
[
"CELOSS",
"Distill LossCE Loss",
"Weighted Sum",
"A",
"Weighted Sum",
"本",
"大",
"100",
"",
"0",
"",
"®",
"",
"000",
"Gate",
"",
"",
"",
"",
"Old Model",
"Old Model",
"New Model",
"Old Model",
"New Model",
"New Model",
"(Fixed)",
"(Fixed)",
"(Trained)",
"(Fixed)",
"大",
"元",
"A",
"",
"Input",
"A",
"Input",
"Input",
"(b) Old-New Ensemble",
"(c) Gated Fusion",
"Figure 2: Methods to improve prediction backward compatibility during model upgrade. (a) Distillation-based",
"(a) Distillation",
"approach to align predicted logits on potential regression instances (Xie et al., 2021). (b) Ensemble of old and new",
"models via weighted sum of either predicted logits or probabilities. (c)Our proposed Gated Fusion that learns a",
"gate as a soft switch to dynamically determine whether to fall back to previous predictions.",
"One thing to emphasize is that maximizing classi-",
"models as our final outputs:",
"fier performance does not necessarily help in mini-",
"lold(yla)",
"lGF(yl) = (1-Qgate)",
"+agate-Inew(y|r),",
"1",
"(3)",
"mizing RNF (Yan et al., 2021; Xie et al., 2021)",
"3Gated Fusion:Methodology",
"where l(y|α) denotes predicted logits from models",
"and T is the temperature scaling to regularize the",
"3.1 Method Overview",
"magnitude of old model's logits. fnew and ge are",
"To improve backward compatibility in model up-",
"then jointly trained end-to-end with cross-entropy",
"grade, it's crucial to have a mechanism that de-",
"loss between our output logits IGF(yla) and label",
"tects potential regression errors and mitigates them",
"distributions on downstream tasks.",
"when making predictions. We propose Gated Fu-",
"TheintuitionbehindGatedFusionisthatwhen",
"sion (GF) to achieve this by learning a gate as a soft",
"fnew makes a mistake while fold produces the cor-",
"switch to choose between generating predictions",
"rect output, the gate ge will learn to put more",
"by the new model or resorting to outputs of the",
"weight on fold in order to minimize the final clas-",
"old model. Gated Fusion is inspired by the gating",
"sification loss. This process effectively mitigates",
"mechanism widely used in other applications. For",
"potential negative flips introduced by the model",
"example,mixing word copying mode with word",
"upgrade and thus improves the backward compati-",
"generation mode for language modeling (Merity",
"bility of final predictions.",
"et al., 2016) and summarization (See et al., 2017)",
"3.2Training and Inference",
"Our proposed Gated Fusion fcF consists of the",
"old model fold, the new model fnew, and a gating",
"In practice, training Gated Fusion with randomly",
"network ge. The old model fold is the legacy ver-",
"initialized fnew would make the shallow gating",
"sion before upgrade where the parameters are fixed.",
"network quickly converge to favor the fully-trained",
"The new model fnew has the same architecture as",
"fold. To prevent this, we only train fnew for the first",
"fnew and is randomly initialized. The gating net-",
"few epochs to ensure its competence before jointly",
"work ge is a multi-layer feed-forward network with",
"training ge and fnew using lGF(α). In addition, we",
"sigmoid function. It produces a scalar weight agate",
"found that stopping gradient flow from ge to fnew",
"in the range [0, 1] from the output layer of fnew.",
"can further prevent the performance decrease of the",
"denoted as Enew:",
"new model within Gated Fusion:",
"Qgate(α) = ge(stop_grad(Enew(r).",
"(2)",
"(4)",
"Qgate(r) = go(Enew(α)).",
"At inference time, Gated Fusion produces logits",
"We use Qgate to combine the logits of old and new",
"from fold and fnew as well as the gate value agate"
] |
{
"arxiv_id": "2302.02080",
"language": "en",
"page_idx": 2
}
|
||
2302.02080-en-3
|
2302.02080-en-3.jpg
|
to make output predictions:
$$f_{GF}^{*}(x)=Softmax\Big((1-\alpha_{gate})\cdot\frac{l_{old}}{T}+\alpha_{gate}\cdot l_{new}\Big).$$
### 33 Inference with Cache
Our proposed Gated Fusion requires $f_old$ to be hosted together with the new model. In reality, one could have a resource-constrained setting and request the old model to be discarded at inference. We note that in real applications, repetitive inputs are commonly seen in live traffic (Batrinca and Treleaven, 2015) and the backward compatibility of model upgrade entails that correct predictions can be preserved on the legacy instances already seen and predicted by the old model.
To simulate real scenarios, we randomly cache old model's logits on a portion of test inputs. When getting out-of-cache instances, we use new model's output embedding $\mathcal{E}_new(x)$ as key and euclidean distance as metric to search for the nearest cached instance. The cached old-model logits can then be used for Gated Fusion to make predictions without hosting $f_{old}$ at inference.
# 4 Experiments Setup
4.1 Model Upgrade Scenarios
We conduct experiments on two representative model upgrade scenarios: (a) upgrade to a larger pretrained model of the same type, where we use tinct pretrained model with the same size. We use BERT$_base$ to ELECTRA$_base$ (Clark et al., 2020) as this challenging model upgrade scenario for backward-compatibility, as they are pretrained under different self-supervised learning paradigms. The former uses masked language modeling (MLM) with reconstruction loss, while the latter is pretrained in generative-contrastive (adversarial) fashion with distributional divergence as the loss (Liu et al., 2021).
## 4.2 Datasets and Implementation
We evaluate our approach across three datasets. They represent different sentence-level classification tasks, from single-sentence to sentence-pair classification, with varying dataset sizes. We use: (a) Stanford Sentiment Treebank (SST-2), a singlesentence task to classify movie review sentiment, with 67k train and 0.9k dev set (Socher et al., 2013). (b) Microsoft Research Paraphrase Corpus (MRPC)
(Dolan and Brockett, 2005), a sentence-pair classification task for identifying paraphrases, with 3.7k train and 0.4k dev set. (c) Question Natural Language Inference (QNLI), a question-paragraph pair task to determine whether the paragraph contains the answer to the question, with 100k train and 5.5k dev set. Datasets are taken from GLUE Benchmark (Wang et al., 2018) and processed with scripts from Hugging Face$^2.$
For implementation, we use the sequence classification and pre-trained model parameters from Hugging Face Transformers$^3.$ Experiments are done in PyTorch (Paszke et al., 2019) with Tesla V100 GPUs and results are averaged over 5 random seeds. Learning rate, batch size, and train epoch are tuned during training new model alone on given tasks and then fixed for all backward-compatible solutions. In Gated Fusion, we frst train $f_{new}$ alone for first $(N-1)$ epochs and then jointly train $g_{\theta}$ and $f_new$ with Gated Fusion logits $l_GF^*$ in the last epoch. Further implementation details can be found in the Appendix.
## 4.3 Baselines
We compare our approach with several strong baselines. (a) Train the new model directly on the target task without any adjustment, i.e. $f_new^o.$ $( \mathbf{b} )$ The specialized distillation method proposed in Xie et al. (2021), where the KL-divergence of prediction probabilities between old and new mod-
pretrained model of the same type, where we useels is applied when $p_{old}( y= y_i| x_i) > p_{new}( y=$ $\text{BERT}_{base}\textit{ to BERT}_{large. ( b) upgradetoadis- }y_i| x_i) .$ ( c) Model ensemble via maiority- voting
$y_{i}|x_{i}).$(c) Model ensemble via majority-voting that was shown to be very effective (Yan et al., 2021; Xie et al., 2021). Similarly, we use 5-seed new model ensemble as a strong baseline. (d) The ensemble of the old and new models probabilities, $p^{*}(y|x)=(1-\alpha)\cdot p_{old}(y|x)+\alpha\cdot p_{new}(y|x)$,as well as ensemble of the old and new models logits, $l^{* }( y| x) = ( 1- \alpha ) \cdot l_{old}( y| x) + \alpha \cdot l_{new}( y| x) $, where $\alpha$ is searched among [0.5,0.6,0.7,0.8,0.9] to maximize backward-compatibility while achieving accuracy on par with the vanilla $f_{new}^o.$
## 5 Results and Analysis
5.1 Upgrade to a Larger Pretrained Model
Our first model upgrade scenario scales up the size of underlying pretrained language models. We experiment with BERT$_base$ to BERT$_large$, where
$2{\mathrm{https: }/ / \mathrm{huggi}}$ngface. co/ datasets/ glue 3
$^{3}$https://huggingface.co/docs/transformers/index
|
[
[
0.1175537109375,
0.08612060546875,
0.32080078125,
0.08612060546875,
0.32080078125,
0.1009521484375,
0.1175537109375,
0.1009521484375
],
[
0.51318359375,
0.08489990234375,
0.8818359375,
0.08489990234375,
0.8818359375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.51123046875,
0.1009521484375,
0.8818359375,
0.1015625,
0.8818359375,
0.11639404296875,
0.51123046875,
0.11517333984375
],
[
0.34765625,
0.11224365234375,
0.382080078125,
0.1181640625,
0.3779296875,
0.1300048828125,
0.343505859375,
0.12353515625
],
[
0.1158447265625,
0.1181640625,
0.2822265625,
0.11639404296875,
0.2822265625,
0.1341552734375,
0.1158447265625,
0.135986328125
],
[
0.280517578125,
0.122314453125,
0.340087890625,
0.122314453125,
0.340087890625,
0.1341552734375,
0.280517578125,
0.1341552734375
],
[
0.38037109375,
0.122314453125,
0.4609375,
0.122314453125,
0.4609375,
0.1341552734375,
0.38037109375,
0.1341552734375
],
[
0.51318359375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.51318359375,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.466796875,
0.14013671875,
0.489501953125,
0.14013671875,
0.489501953125,
0.15087890625,
0.466796875,
0.15087890625
],
[
0.51318359375,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.51318359375,
0.1644287109375
],
[
0.1175537109375,
0.163330078125,
0.328369140625,
0.163330078125,
0.328369140625,
0.177001953125,
0.1175537109375,
0.177001953125
],
[
0.51318359375,
0.166259765625,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1175537109375,
0.185302734375,
0.48779296875,
0.185302734375,
0.48779296875,
0.1995849609375,
0.1175537109375,
0.1995849609375
],
[
0.51318359375,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.51318359375,
0.1953125
],
[
0.1175537109375,
0.201904296875,
0.48779296875,
0.201904296875,
0.48779296875,
0.215576171875,
0.1175537109375,
0.215576171875
],
[
0.509765625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.21728515625,
0.48779296875,
0.21728515625,
0.48779296875,
0.23095703125,
0.1175537109375,
0.23095703125
],
[
0.51123046875,
0.21435546875,
0.62890625,
0.21435546875,
0.62890625,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.1175537109375,
0.2340087890625,
0.489501953125,
0.2340087890625,
0.489501953125,
0.2476806640625,
0.1175537109375,
0.2476806640625
],
[
0.52880859375,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.52880859375,
0.2452392578125
],
[
0.1175537109375,
0.2493896484375,
0.489501953125,
0.2493896484375,
0.489501953125,
0.26318359375,
0.1175537109375,
0.26318359375
],
[
0.51123046875,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.1175537109375,
0.2666015625,
0.490234375,
0.2666015625,
0.490234375,
0.27978515625,
0.1175537109375,
0.27978515625
],
[
0.51123046875,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.27783203125,
0.51123046875,
0.27783203125
],
[
0.1158447265625,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.1158447265625,
0.29443359375
],
[
0.5087890625,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.5087890625,
0.292236328125
],
[
0.1158447265625,
0.297607421875,
0.48779296875,
0.29638671875,
0.48779296875,
0.311279296875,
0.1158447265625,
0.312255859375
],
[
0.51123046875,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.31005859375,
0.51123046875,
0.31005859375
],
[
0.11419677734375,
0.3134765625,
0.486083984375,
0.3134765625,
0.486083984375,
0.32666015625,
0.11419677734375,
0.32666015625
],
[
0.51123046875,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.32666015625,
0.51123046875,
0.32666015625
],
[
0.11419677734375,
0.330078125,
0.3525390625,
0.328857421875,
0.3525390625,
0.34326171875,
0.11419677734375,
0.344482421875
],
[
0.51123046875,
0.328857421875,
0.8818359375,
0.328857421875,
0.8818359375,
0.342529296875,
0.51123046875,
0.342529296875
],
[
0.1351318359375,
0.3466796875,
0.489501953125,
0.3466796875,
0.489501953125,
0.35986328125,
0.1351318359375,
0.35986328125
],
[
0.5087890625,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.35693359375,
0.5087890625,
0.35693359375
],
[
0.1175537109375,
0.3623046875,
0.48779296875,
0.3623046875,
0.48779296875,
0.377197265625,
0.1175537109375,
0.377197265625
],
[
0.509765625,
0.359375,
0.880859375,
0.361083984375,
0.880859375,
0.3759765625,
0.509765625,
0.37353515625
],
[
0.1158447265625,
0.37890625,
0.489501953125,
0.377685546875,
0.489501953125,
0.392578125,
0.1158447265625,
0.393798828125
],
[
0.51123046875,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.39013671875,
0.51123046875,
0.39013671875
],
[
0.1175537109375,
0.3955078125,
0.48779296875,
0.3955078125,
0.48779296875,
0.4091796875,
0.1175537109375,
0.4091796875
],
[
0.509765625,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.40625,
0.509765625,
0.40625
],
[
0.1158447265625,
0.410888671875,
0.486083984375,
0.410888671875,
0.486083984375,
0.424560546875,
0.1158447265625,
0.424560546875
],
[
0.509765625,
0.4091796875,
0.8818359375,
0.4091796875,
0.8818359375,
0.4228515625,
0.509765625,
0.4228515625
],
[
0.1175537109375,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.1175537109375,
0.440673828125
],
[
0.509765625,
0.424560546875,
0.68603515625,
0.424560546875,
0.68603515625,
0.439453125,
0.509765625,
0.439453125
],
[
0.1158447265625,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.1158447265625,
0.457275390625
],
[
0.1158447265625,
0.458984375,
0.301513671875,
0.458984375,
0.301513671875,
0.473876953125,
0.1158447265625,
0.473876953125
],
[
0.51318359375,
0.454833984375,
0.63037109375,
0.454833984375,
0.63037109375,
0.468505859375,
0.51318359375,
0.468505859375
],
[
0.51318359375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.1158447265625,
0.48681640625,
0.312255859375,
0.48681640625,
0.312255859375,
0.501953125,
0.1158447265625,
0.501953125
],
[
0.51318359375,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.509765625,
0.50732421875,
0.88330078125,
0.50830078125,
0.88330078125,
0.525390625,
0.509765625,
0.52490234375
],
[
0.1158447265625,
0.51318359375,
0.36279296875,
0.51318359375,
0.36279296875,
0.52685546875,
0.1158447265625,
0.52685546875
],
[
0.51318359375,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.1158447265625,
0.533203125,
0.48779296875,
0.53369140625,
0.48779296875,
0.548828125,
0.1158447265625,
0.54736328125
],
[
0.51123046875,
0.54150390625,
0.88330078125,
0.54150390625,
0.88330078125,
0.5546875,
0.51123046875,
0.5546875
],
[
0.1175537109375,
0.55029296875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.51318359375,
0.55712890625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.118408203125,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.5087890625,
0.57177734375,
0.8818359375,
0.572265625,
0.8818359375,
0.59033203125,
0.5087890625,
0.5888671875
],
[
0.11419677734375,
0.5791015625,
0.491943359375,
0.580078125,
0.491943359375,
0.60107421875,
0.11419677734375,
0.599609375
],
[
0.509765625,
0.59033203125,
0.8818359375,
0.5888671875,
0.8818359375,
0.60400390625,
0.509765625,
0.60498046875
],
[
0.1175537109375,
0.5986328125,
0.489501953125,
0.5986328125,
0.489501953125,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.51123046875,
0.60693359375,
0.88330078125,
0.60693359375,
0.88330078125,
0.62060546875,
0.51123046875,
0.62060546875
],
[
0.1175537109375,
0.61328125,
0.48779296875,
0.61328125,
0.48779296875,
0.62744140625,
0.1175537109375,
0.62744140625
],
[
0.509765625,
0.62158203125,
0.880859375,
0.62158203125,
0.880859375,
0.63525390625,
0.509765625,
0.63525390625
],
[
0.1175537109375,
0.63134765625,
0.489501953125,
0.63134765625,
0.489501953125,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.509765625,
0.63720703125,
0.8818359375,
0.63818359375,
0.8818359375,
0.65283203125,
0.509765625,
0.65185546875
],
[
0.11419677734375,
0.6455078125,
0.490234375,
0.646484375,
0.490234375,
0.6611328125,
0.11419677734375,
0.66015625
],
[
0.51123046875,
0.65478515625,
0.88330078125,
0.65478515625,
0.88330078125,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.11419677734375,
0.662109375,
0.48779296875,
0.6630859375,
0.48779296875,
0.677734375,
0.11419677734375,
0.6767578125
],
[
0.509765625,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68505859375,
0.509765625,
0.68505859375
],
[
0.1175537109375,
0.6796875,
0.490234375,
0.6796875,
0.490234375,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.5087890625,
0.68505859375,
0.8818359375,
0.68603515625,
0.8818359375,
0.70068359375,
0.5087890625,
0.69970703125
],
[
0.1175537109375,
0.6953125,
0.489501953125,
0.6953125,
0.489501953125,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.509765625,
0.703125,
0.8818359375,
0.703125,
0.8818359375,
0.71728515625,
0.509765625,
0.71728515625
],
[
0.118408203125,
0.71337890625,
0.489501953125,
0.71337890625,
0.489501953125,
0.7236328125,
0.118408203125,
0.7236328125
],
[
0.509765625,
0.71728515625,
0.87744140625,
0.71728515625,
0.87744140625,
0.73095703125,
0.509765625,
0.73095703125
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.51123046875,
0.7353515625,
0.87890625,
0.7353515625,
0.87890625,
0.7451171875,
0.51123046875,
0.7451171875
],
[
0.1158447265625,
0.7421875,
0.27783203125,
0.7421875,
0.27783203125,
0.755859375,
0.1158447265625,
0.755859375
],
[
0.5087890625,
0.7490234375,
0.8212890625,
0.75048828125,
0.8212890625,
0.7685546875,
0.5087890625,
0.7666015625
],
[
0.1158447265625,
0.7685546875,
0.3896484375,
0.76953125,
0.3896484375,
0.78369140625,
0.1158447265625,
0.783203125
],
[
0.51318359375,
0.783203125,
0.5439453125,
0.783203125,
0.5439453125,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.5400390625,
0.7822265625,
0.72314453125,
0.7822265625,
0.72314453125,
0.7958984375,
0.5400390625,
0.7958984375
],
[
0.1158447265625,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.1158447265625,
0.80810546875,
0.491943359375,
0.80810546875,
0.491943359375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.51318359375,
0.80859375,
0.85986328125,
0.80859375,
0.85986328125,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.1158447265625,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.51123046875,
0.83056640625,
0.8818359375,
0.83203125,
0.8818359375,
0.8466796875,
0.51123046875,
0.845703125
],
[
0.1175537109375,
0.84033203125,
0.490234375,
0.84033203125,
0.490234375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.8486328125,
0.8818359375,
0.8486328125,
0.8818359375,
0.86328125,
0.509765625,
0.86328125
],
[
0.1175537109375,
0.85693359375,
0.490234375,
0.85693359375,
0.490234375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5087890625,
0.8623046875,
0.88330078125,
0.86328125,
0.88330078125,
0.88134765625,
0.5087890625,
0.8798828125
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.1175537109375,
0.90185546875
],
[
0.53564453125,
0.8955078125,
0.75244140625,
0.8955078125,
0.75244140625,
0.90625,
0.53564453125,
0.90625
],
[
0.1175537109375,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.5322265625,
0.90380859375,
0.81201171875,
0.90625,
0.81201171875,
0.91943359375,
0.5322265625,
0.91748046875
]
] |
[
"to make output predictions:",
"(Dolan and Brockett, 2005), a sentence-pair classi-",
"fication task for identifying paraphrases, with 3.7k",
"lold",
"fGF(a) = Softmar ((1",
"-αgate)",
"+agateIne",
"train and 0.4k dev set. () Question Natural Lan-",
"guage Inference (QNLID), a question-paragraph pair",
"(5)",
"task to determine whether the paragraph contains",
"3.3 Inference with Cache",
"the answer to the question, with 100k train and 5.5k",
"Our proposed Gated Fusion requires fold to be",
"dev set. Datasets are taken from GLUE Benchmark",
"hosted together with the new model. In reality,",
"(Wang et al., 2018) and processed with scripts from",
"one could have a resource-constrained setting and",
"Hugging Face?.",
"request the old model to be discarded at inference.",
"For implementation, we use the sequence clas-",
"We note that in real applications, repetitive inputs",
"sification and pre-trained model parameters from",
"are commonly seen in live traffic (Batrinca and Tre-",
"Hugging Face Transformers3. Experiments are",
"leaven, 2015) and the backward compatibility of",
"done in PyTorch (Paszke et al., 2019) with Tesla",
"model upgrade entails that correct predictions can",
"V100 GPUs and results are averaged over 5 random",
"be preserved on the legacy instances already seen",
"seeds. Learning rate, batch size, and train epoch",
" and predicted by the old model.",
"are tuned during training new model alone on given",
"To simulate real scenarios, we randomly cache",
"tasks and then fixed for all backward-compatible",
"old model's logits on a portion of test inputs. When",
"solutions. In Gated Fusion, we first train fnew",
"getting out-of-cache instances, we use new model's",
"alone for first (N-1) epochs and then jointly train",
"output embedding Enew(α) as key and euclidean",
"ge and fnew with Gated Fusion logits I'ce in the",
"distance as metric to search for the nearest cached",
"last epoch. Further implementation details can be",
"instance. The cached old-model logits can then be",
"found in the Appendix.",
"used for Gated Fusion to make predictions without",
"hosting fold at inference.",
"4.3Baselines",
"We compare our approach with several strong base-",
"4Experiments Setup",
"lines. (a) Train the new model directly on the tar-",
"get task without any adjustment, i.e. fnew. (b)",
"4.1 Model Upgrade Scenarios",
"The specialized distillation method proposed in",
"We conduct experiments on two representative",
"Xie et al. (2021), where the KL-divergence of pre-",
"model upgrade scenarios: (a) upgrade to a larger",
"diction probabilities between old and new mod-",
"pretrained model of the same type, where we use",
"els is applied when Pold(y = yi|ci) > pnew(y =",
"BERTbase to BERTlarge- (b) upgrade to a dis-",
"Yi|ci). (c) Model ensemble via majority-voting",
"tinct pretrained model with the same size. We",
"that was shown to be very effective (Yan et al.,",
"use BERTbase to ELECTRAhase (Clark et al.,",
"2021; Xie et al., 2021). Similarly, we use 5-seed",
"2020) as this challenging model upgrade sce-",
"new model ensemble as a strong baseline. (d) The",
"nario for backward-compatibility,as they are pre-",
"ensemble of the old and new models probabilities,",
"trained under different self-supervised learning",
"p*(yl) = (1 -α) pold(ylα) +Q pnew(ylα), as",
"paradigms. The former uses masked language mod-",
"well as ensemble of the old and new models log",
"eling (MLM) with reconstruction loss, while the",
"its, I*(ylα) = (1 -a) - lold(ylc) +Q-lnew(yla),",
"latteris pretrained ingenerative-contrastive(adver-",
"where q is searched among [0.5, 0.6. 0.7. 0.8, 0.9",
"sarial) fashion with distributional divergence as the",
"tomaximizebackward-compatibility while achiey",
"loss (Liu et al., 2021),",
"ing accuracy on par with the vanilla fnew.",
"4.2Datasets and Implementation",
"5",
"Results and Analysis",
"We evaluate our approach across three datasets.",
"They represent different sentence-level classifica-",
"5.1Upgrade to a Larger Pretrained Model",
"tion tasks, from single-sentence to sentence-pair",
"Our first model upgrade scenario scales up the size",
"classification, with varying dataset sizes. We use:",
"of underlying pretrained language models. We",
"(a) Stanford Sentiment Treebank (SST-2), a single-",
"experiment with BERTbase to BERTlarge, where",
"sentence task to classify movie review sentiment",
"with 67k train and 0.9k dev set (Socher et al., 2013).",
"https://huggingface.co/datasets/glue",
"(b) Microsoft Research Paraphrase Corpus (MRPC)",
"3https://huggingface.co/docs/transformers/index"
] |
{
"arxiv_id": "2302.02080",
"language": "en",
"page_idx": 3
}
|
||
2302.02080-en-4
|
2302.02080-en-4.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{BERT}_{base}$ $\to\mathbf{BERT}_{large}$</th>
<th colspan="2">$SST-2$</th>
<th colspan="2">$\mathbf{MRPC}$</th>
<th colspan="2">$QNLI$</th>
</tr>
<tr>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
</tr>
<tr>
<td>OldModel NewModel</td>
<td>$2.18_{0.21}$</td>
<td>$92.00_{0.27}$ $93.12_{0.29}$</td>
<td>$4.12_{1.04}$</td>
<td>$85.69_{0.90}$ $87.40_{1.02}$</td>
<td>$2.72_{0.13}$</td>
<td>$90.74_{0.09}$ $92.22_{0.16}$</td>
</tr>
<tr>
<td>OldModel</td>
<td> </td>
<td>$92.00_{0.27}$</td>
<td> </td>
<td>$85.69_{0.90}$</td>
<td> </td>
<td>$90.74_{0.09}$</td>
</tr>
<tr>
<td>Distillation ( Xie et al. , 2021) </td>
<td>$1.97_{0.22}$</td>
<td>$93.33_{0.20}$</td>
<td>$3.53_{0.77}$</td>
<td>$87.70_{1.34}$</td>
<td>$2.31_{0.14}$</td>
<td>$92.60_{0.19}$</td>
</tr>
<tr>
<td>New Model Ensemble</td>
<td>$2.00_{0.31}$</td>
<td>$93.30_{0.24}$</td>
<td>$2.25_{0.61}$</td>
<td>$88.87_{0.77}$</td>
<td>$1.98_{0.21}$</td>
<td>$92.97_{0.22}$</td>
</tr>
<tr>
<td>Old-New Probs Ensemble</td>
<td>$1.06_{0.27}$</td>
<td>$93.12_{0.38}$</td>
<td>$1.67_{0.78}$</td>
<td>$87.16_{1.12}$</td>
<td>$1.04_{0.26}$</td>
<td>$92.44_{0.23}$</td>
</tr>
<tr>
<td>Old-New Logits Ensemble</td>
<td>$1.06_{0.27}$</td>
<td>$93.12_{0.38}$</td>
<td>$1.67_{0.78}$</td>
<td>$87.16_{1.12}$</td>
<td>$1.04_{0.26}$</td>
<td>$92.44_{0.23}$</td>
</tr>
<tr>
<td>Gated Fusion</td>
<td>$0.78_{0.20}$</td>
<td>$93.05_{0.09}$</td>
<td>$1.18_{0.52}$</td>
<td>$87.45_{0.52}$</td>
<td>$0.73_{0.13}$</td>
<td>$92.24_{0.24}$</td>
</tr>
</tbody>
</table>
Table 1: Negative flip rate $\mathcal{R}_{NF}$ and model accuracy (%) of competing methods to optimize backward-
compatibility without performance degradation during BERT$_base\to\mathrm{BERT}_{large}$ model upgrade.
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{BERT}_{base}$ ${}_{2}\rightarrow\mathbf{ELECTRA}_{base}$</th>
<th colspan="2">$\mathbf{SST-2}$</th>
<th colspan="2">$\mathbf{MRPC}$</th>
<th colspan="2">$QNLI$</th>
</tr>
<tr>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
<th>$\mathcal{R}_{NF}$</th>
<th>Accuracy</th>
</tr>
<tr>
<td>OldModel NewModel</td>
<td>$1.63_{0.20}$</td>
<td>$92.00_{0.27}$ $95.00_{0.06}$</td>
<td>$3.73_{0.36}$</td>
<td>$85.69_{0.90}$ $88.58_{0.57}$</td>
<td>$2.82_{0.32}$</td>
<td>$90.74_{0.09}$ $92.90_{0.26}$</td>
</tr>
<tr>
<td>Distillation ( Xie et al. , 2021) </td>
<td>$1.49_{0.24}$</td>
<td>$95.02_{0.21}$</td>
<td>$3.68_{0.79}$</td>
<td>$88.82_{0.94}$</td>
<td>$2.58_{0.17}$</td>
<td>$93.03_{0.16}$</td>
</tr>
<tr>
<td>New Model Ensemble</td>
<td>$1.12_{0.09}$</td>
<td>$95.39_{0.09}$</td>
<td>$3.24_{0.24}$</td>
<td>$89.02_{0.48}$</td>
<td>$2.26_{0.08}$</td>
<td>$93.49_{0.07}$</td>
</tr>
<tr>
<td>Old-New Probs Ensemble</td>
<td>$1.40_{0.17}$</td>
<td>$95.07_{0.15}$</td>
<td>$3.14_{0.42}$</td>
<td>$88.53_{0.48}$</td>
<td>$0.98_{0.20}$</td>
<td>$93.04_{0.21}$</td>
</tr>
<tr>
<td>Old-New Logits Ensemble</td>
<td>$0.89_{0.17}$</td>
<td>$94.95_{0.13}$</td>
<td>$3.28_{0.43}$</td>
<td>$88.48_{0.51}$</td>
<td>$0.98_{0.20}$</td>
<td>$93.04_{0.21}$</td>
</tr>
<tr>
<td>Gated Fusion</td>
<td>$0.71_{0.18}$</td>
<td>$95.02_{0.16}$</td>
<td>$2.40_{0.50}$</td>
<td>$88.68_{0.68}$</td>
<td>$0.81_{0.16}$</td>
<td>$92.98_{0.17}$</td>
</tr>
</tbody>
</table>
Table 2: Negative flip rate $\mathcal{R}_{NF}$ and model accuracy (%) of competing methods to optimize backward-
compatibility without performance degradation during BERT$_base\to\mathrm{ELECTRA}_{base}$ model upgrade.
the model size is tripled (110M vs 340M) and the
model depth is doubled (12 vs 24 layers).
-0.31% on QNLI, which translate to further reducing $12.8\%,11.9\%$, and $11.4\%$ of negative flip
Table 1 shows the results. For $f_{new}^o$, we can cases. These results show the effectiveness of our
method to mitigate a significant amount of regres-
sion errors during model upgrade.
5.2 Upgrade to a Different Pretrained Model
observe that the negative flip rates $\mathcal{R}_NF$ are usually much larger than the accuracy gains across tasks, which could be the reason to hinder new model adoptions in real-world applications. Besides, when dividing $\mathcal{R}_NF$ over the error rate $(1-accuracy)$, we can observe that around $30\%$ to 40% of all $f_new^o$ prediction errors are in fact the new errors introduced during model upgrade. For improving prediction backward-compatibility, our proposed Gated Fusion outperforms other competing methods to considerably reduce $\mathcal{R}_NF$ without degradation on accuracy. Note that best $\alpha$ values found for the two variants of old-new ensemble are both 0.5, hence producing identical results.
A more challenging upgrade scenario is when old and new models are pretrained under distinctive paradigms, producing two representation spaces of fairly different characteristics (Meng et al., 2021b). We experiment with BERT$_{base}$ to ELECTRA$_base$ in this scenario, where two models have the same size but are pretrained under utterly different schemes, i.e. generative versus adversarial.
Table 2 shows the results. For $f_new^o$,compared with upgrading to BERT$_large$,we observe larger accuracy gains and lower $\mathcal{R}_{NF}$ on SST-2 and MRPC.
Compared to the vanilla new model, gated fu- However, on QNLI, upgrading to ELECTRA$_{base}$
achieves a higher accuracy gain but an even a higher $\mathcal{R}_NF.$ This implies that boosting accuracy and improving backward compatibility could be two related but different objectives.
For mitigation strategies, Gated Fusion achieves the lowest negative flip rates across datasets without any accuracy loss. We obtain absolute $\mathcal{R}_{NF}$ re-
sion obtains absolute $\mathcal{R}_NF$ reductions of -1.40% on SST-2, -2.94% on MRPC, and -1.99% on QNLI. These translate to reducing the total negative flip cases by 64.2%, 71.4%, 73.2%, respectively. Compared to the strongest baseline (old-new ensemble), we obtain further absolute $\mathcal{R}_NF$ reductions of -0.28% on SST-2, -0.49% on MRPC, and
|
[
[
0.426513671875,
0.0872802734375,
0.47607421875,
0.0872802734375,
0.47607421875,
0.1015625,
0.426513671875,
0.1015625
],
[
0.57666015625,
0.0872802734375,
0.634765625,
0.0872802734375,
0.634765625,
0.1015625,
0.57666015625,
0.1015625
],
[
0.736328125,
0.0872802734375,
0.78662109375,
0.0872802734375,
0.78662109375,
0.1015625,
0.736328125,
0.1015625
],
[
0.16455078125,
0.09857177734375,
0.35107421875,
0.1009521484375,
0.35107421875,
0.119384765625,
0.16455078125,
0.1175537109375
],
[
0.450927734375,
0.1009521484375,
0.52294921875,
0.10272216796875,
0.52294921875,
0.1175537109375,
0.450927734375,
0.11517333984375
],
[
0.60546875,
0.1009521484375,
0.67822265625,
0.10272216796875,
0.67822265625,
0.1175537109375,
0.60546875,
0.11517333984375
],
[
0.759765625,
0.1009521484375,
0.8330078125,
0.10272216796875,
0.8330078125,
0.1175537109375,
0.759765625,
0.11517333984375
],
[
0.69873046875,
0.1015625,
0.7421875,
0.10394287109375,
0.74072265625,
0.1181640625,
0.69775390625,
0.11639404296875
],
[
0.3896484375,
0.1015625,
0.432373046875,
0.10394287109375,
0.4306640625,
0.1181640625,
0.387939453125,
0.11639404296875
],
[
0.5458984375,
0.10394287109375,
0.587890625,
0.10394287109375,
0.587890625,
0.1175537109375,
0.5458984375,
0.1175537109375
],
[
0.453369140625,
0.12054443359375,
0.5205078125,
0.12353515625,
0.51904296875,
0.1395263671875,
0.45166015625,
0.135986328125
],
[
0.60693359375,
0.12054443359375,
0.67529296875,
0.12353515625,
0.67333984375,
0.1395263671875,
0.60546875,
0.135986328125
],
[
0.76171875,
0.1217041015625,
0.8310546875,
0.12469482421875,
0.82958984375,
0.1395263671875,
0.759765625,
0.135986328125
],
[
0.4013671875,
0.1258544921875,
0.414794921875,
0.1258544921875,
0.414794921875,
0.1353759765625,
0.4013671875,
0.1353759765625
],
[
0.16796875,
0.12469482421875,
0.2451171875,
0.12469482421875,
0.2451171875,
0.135986328125,
0.16796875,
0.135986328125
],
[
0.71533203125,
0.127685546875,
0.72314453125,
0.127685546875,
0.72314453125,
0.1341552734375,
0.71533203125,
0.1341552734375
],
[
0.3779296875,
0.135986328125,
0.438232421875,
0.1383056640625,
0.4375,
0.15380859375,
0.376220703125,
0.1519775390625
],
[
0.45166015625,
0.135986328125,
0.5224609375,
0.1395263671875,
0.5205078125,
0.155029296875,
0.450927734375,
0.1519775390625
],
[
0.5341796875,
0.135986328125,
0.59375,
0.1383056640625,
0.591796875,
0.15380859375,
0.5322265625,
0.1519775390625
],
[
0.60693359375,
0.135986328125,
0.6767578125,
0.1383056640625,
0.67529296875,
0.15380859375,
0.60546875,
0.1519775390625
],
[
0.6884765625,
0.135986328125,
0.748046875,
0.1383056640625,
0.74658203125,
0.15380859375,
0.68701171875,
0.1519775390625
],
[
0.76171875,
0.13720703125,
0.8330078125,
0.14013671875,
0.8310546875,
0.155029296875,
0.759765625,
0.1519775390625
],
[
0.53076171875,
0.1561279296875,
0.59423828125,
0.1591796875,
0.59375,
0.174560546875,
0.52880859375,
0.171630859375
],
[
0.166259765625,
0.1383056640625,
0.252685546875,
0.1383056640625,
0.252685546875,
0.152587890625,
0.166259765625,
0.152587890625
],
[
0.378662109375,
0.15673828125,
0.438232421875,
0.1591796875,
0.4375,
0.174560546875,
0.3779296875,
0.1728515625
],
[
0.45166015625,
0.15673828125,
0.5224609375,
0.1602783203125,
0.5205078125,
0.174560546875,
0.450927734375,
0.171630859375
],
[
0.60693359375,
0.15673828125,
0.67529296875,
0.1591796875,
0.67333984375,
0.174560546875,
0.60546875,
0.1728515625
],
[
0.6884765625,
0.15673828125,
0.748046875,
0.1591796875,
0.74658203125,
0.174560546875,
0.68701171875,
0.1728515625
],
[
0.76171875,
0.1561279296875,
0.8310546875,
0.1602783203125,
0.82958984375,
0.17578125,
0.759765625,
0.171630859375
],
[
0.378662109375,
0.171630859375,
0.438232421875,
0.174560546875,
0.4375,
0.190673828125,
0.3779296875,
0.18701171875
],
[
0.16796875,
0.1591796875,
0.370361328125,
0.1591796875,
0.370361328125,
0.1728515625,
0.16796875,
0.1728515625
],
[
0.45166015625,
0.171630859375,
0.5224609375,
0.17578125,
0.5205078125,
0.190673828125,
0.450927734375,
0.1859130859375
],
[
0.5341796875,
0.171630859375,
0.59423828125,
0.177001953125,
0.591796875,
0.191162109375,
0.53076171875,
0.1859130859375
],
[
0.60791015625,
0.1728515625,
0.67529296875,
0.17578125,
0.67333984375,
0.190673828125,
0.60693359375,
0.18701171875
],
[
0.69189453125,
0.1728515625,
0.748046875,
0.17578125,
0.74658203125,
0.190673828125,
0.68994140625,
0.18701171875
],
[
0.76171875,
0.171630859375,
0.8310546875,
0.17578125,
0.82958984375,
0.190673828125,
0.759765625,
0.1859130859375
],
[
0.45166015625,
0.1859130859375,
0.5205078125,
0.190673828125,
0.51904296875,
0.204833984375,
0.450927734375,
0.20068359375
],
[
0.1695556640625,
0.17578125,
0.31982421875,
0.17578125,
0.31982421875,
0.1859130859375,
0.1695556640625,
0.1859130859375
],
[
0.60791015625,
0.1859130859375,
0.67529296875,
0.189453125,
0.67333984375,
0.204833984375,
0.60693359375,
0.201904296875
],
[
0.76171875,
0.1859130859375,
0.8310546875,
0.190673828125,
0.82958984375,
0.204833984375,
0.759765625,
0.20068359375
],
[
0.53564453125,
0.18701171875,
0.59375,
0.190673828125,
0.591796875,
0.2060546875,
0.5341796875,
0.203125
],
[
0.38037109375,
0.18701171875,
0.438232421875,
0.190673828125,
0.4375,
0.204833984375,
0.378662109375,
0.201904296875
],
[
0.68994140625,
0.18701171875,
0.748046875,
0.190673828125,
0.74658203125,
0.204833984375,
0.6884765625,
0.201904296875
],
[
0.76171875,
0.1995849609375,
0.8330078125,
0.207275390625,
0.82958984375,
0.2203369140625,
0.75830078125,
0.213134765625
],
[
0.16796875,
0.189453125,
0.34765625,
0.189453125,
0.34765625,
0.203125,
0.16796875,
0.203125
],
[
0.45166015625,
0.20068359375,
0.5205078125,
0.204833984375,
0.51904296875,
0.2197265625,
0.450927734375,
0.215576171875
],
[
0.68994140625,
0.20068359375,
0.75146484375,
0.2060546875,
0.748046875,
0.2197265625,
0.68701171875,
0.21435546875
],
[
0.38037109375,
0.201904296875,
0.438232421875,
0.204833984375,
0.4375,
0.2197265625,
0.378662109375,
0.2161865234375
],
[
0.166259765625,
0.201904296875,
0.3525390625,
0.203125,
0.3525390625,
0.21728515625,
0.166259765625,
0.2161865234375
],
[
0.53564453125,
0.201904296875,
0.59423828125,
0.204833984375,
0.59375,
0.2197265625,
0.5341796875,
0.2161865234375
],
[
0.60791015625,
0.201904296875,
0.67529296875,
0.204833984375,
0.67333984375,
0.2197265625,
0.60693359375,
0.2161865234375
],
[
0.378662109375,
0.215576171875,
0.438232421875,
0.218505859375,
0.4375,
0.2340087890625,
0.3779296875,
0.23095703125
],
[
0.453369140625,
0.2161865234375,
0.5224609375,
0.2197265625,
0.5205078125,
0.2340087890625,
0.45166015625,
0.23095703125
],
[
0.53564453125,
0.2161865234375,
0.59375,
0.2197265625,
0.591796875,
0.2340087890625,
0.5341796875,
0.23095703125
],
[
0.6884765625,
0.215576171875,
0.748046875,
0.218505859375,
0.74658203125,
0.2340087890625,
0.68701171875,
0.23095703125
],
[
0.76318359375,
0.2161865234375,
0.83447265625,
0.2197265625,
0.8330078125,
0.2340087890625,
0.76171875,
0.23095703125
],
[
0.60791015625,
0.218505859375,
0.6767578125,
0.218505859375,
0.6767578125,
0.2327880859375,
0.60791015625,
0.2327880859375
],
[
0.16796875,
0.2197265625,
0.26123046875,
0.2197265625,
0.26123046875,
0.23095703125,
0.16796875,
0.23095703125
],
[
0.1158447265625,
0.2476806640625,
0.88330078125,
0.2476806640625,
0.88330078125,
0.262451171875,
0.1158447265625,
0.262451171875
],
[
0.11419677734375,
0.26123046875,
0.771484375,
0.26123046875,
0.771484375,
0.279052734375,
0.11419677734375,
0.279052734375
],
[
0.4375,
0.29345703125,
0.486083984375,
0.29345703125,
0.486083984375,
0.308837890625,
0.4375,
0.308837890625
],
[
0.58251953125,
0.29345703125,
0.6396484375,
0.29345703125,
0.6396484375,
0.30810546875,
0.58251953125,
0.30810546875
],
[
0.73876953125,
0.29443359375,
0.78662109375,
0.29443359375,
0.78662109375,
0.308837890625,
0.73876953125,
0.308837890625
],
[
0.16455078125,
0.305908203125,
0.38037109375,
0.30810546875,
0.38037109375,
0.325927734375,
0.16455078125,
0.32373046875
],
[
0.399658203125,
0.30810546875,
0.444091796875,
0.31005859375,
0.443359375,
0.32470703125,
0.39794921875,
0.322509765625
],
[
0.5517578125,
0.30810546875,
0.59375,
0.31005859375,
0.591796875,
0.325927734375,
0.5498046875,
0.32373046875
],
[
0.7021484375,
0.30810546875,
0.7421875,
0.31005859375,
0.74072265625,
0.325927734375,
0.7001953125,
0.32373046875
],
[
0.76171875,
0.30810546875,
0.8330078125,
0.31005859375,
0.8330078125,
0.32470703125,
0.76171875,
0.322509765625
],
[
0.611328125,
0.308837890625,
0.6826171875,
0.308837890625,
0.6826171875,
0.32373046875,
0.611328125,
0.32373046875
],
[
0.462646484375,
0.308837890625,
0.53076171875,
0.308837890625,
0.53076171875,
0.32373046875,
0.462646484375,
0.32373046875
],
[
0.4609375,
0.327880859375,
0.52880859375,
0.330078125,
0.52880859375,
0.344482421875,
0.4609375,
0.342529296875
],
[
0.61279296875,
0.32666015625,
0.67822265625,
0.328857421875,
0.6767578125,
0.344482421875,
0.611328125,
0.342529296875
],
[
0.76318359375,
0.327880859375,
0.8330078125,
0.330078125,
0.8330078125,
0.344482421875,
0.76318359375,
0.342529296875
],
[
0.3896484375,
0.341552734375,
0.45166015625,
0.344482421875,
0.450927734375,
0.35986328125,
0.387939453125,
0.35693359375
],
[
0.166259765625,
0.328857421875,
0.2451171875,
0.328857421875,
0.2451171875,
0.34326171875,
0.166259765625,
0.34326171875
],
[
0.459228515625,
0.342529296875,
0.52880859375,
0.344482421875,
0.52880859375,
0.359375,
0.459228515625,
0.35693359375
],
[
0.54150390625,
0.342529296875,
0.60107421875,
0.344482421875,
0.599609375,
0.359375,
0.5400390625,
0.35693359375
],
[
0.611328125,
0.342529296875,
0.67919921875,
0.344482421875,
0.67919921875,
0.359375,
0.611328125,
0.35693359375
],
[
0.69189453125,
0.342529296875,
0.75146484375,
0.345703125,
0.75,
0.35986328125,
0.68994140625,
0.35693359375
],
[
0.76416015625,
0.341552734375,
0.8330078125,
0.345703125,
0.8310546875,
0.361083984375,
0.76318359375,
0.35693359375
],
[
0.3896484375,
0.361083984375,
0.45166015625,
0.364013671875,
0.450927734375,
0.380126953125,
0.387939453125,
0.377197265625
],
[
0.166259765625,
0.34326171875,
0.2493896484375,
0.34326171875,
0.2493896484375,
0.358154296875,
0.166259765625,
0.358154296875
],
[
0.4609375,
0.3623046875,
0.5283203125,
0.365234375,
0.5263671875,
0.380126953125,
0.459228515625,
0.377197265625
],
[
0.5400390625,
0.3623046875,
0.60107421875,
0.365234375,
0.599609375,
0.380126953125,
0.5380859375,
0.377197265625
],
[
0.61279296875,
0.361083984375,
0.67919921875,
0.364013671875,
0.67822265625,
0.380126953125,
0.611328125,
0.377197265625
],
[
0.76318359375,
0.3623046875,
0.8330078125,
0.365234375,
0.8310546875,
0.380126953125,
0.76171875,
0.377197265625
],
[
0.69189453125,
0.364013671875,
0.75244140625,
0.364013671875,
0.75244140625,
0.37890625,
0.69189453125,
0.37890625
],
[
0.166259765625,
0.363525390625,
0.36279296875,
0.363525390625,
0.36279296875,
0.377197265625,
0.166259765625,
0.377197265625
],
[
0.4609375,
0.3759765625,
0.52880859375,
0.380126953125,
0.5283203125,
0.394287109375,
0.459228515625,
0.39013671875
],
[
0.54150390625,
0.374755859375,
0.60205078125,
0.38134765625,
0.59765625,
0.396728515625,
0.5380859375,
0.39013671875
],
[
0.61376953125,
0.3759765625,
0.67919921875,
0.38134765625,
0.6767578125,
0.394287109375,
0.611328125,
0.3896484375
],
[
0.69287109375,
0.3759765625,
0.75244140625,
0.38134765625,
0.75,
0.394287109375,
0.68994140625,
0.3896484375
],
[
0.76416015625,
0.3759765625,
0.8310546875,
0.380126953125,
0.82958984375,
0.394287109375,
0.76318359375,
0.39013671875
],
[
0.391357421875,
0.377197265625,
0.450927734375,
0.380126953125,
0.44921875,
0.394287109375,
0.3896484375,
0.391357421875
],
[
0.16796875,
0.37890625,
0.31982421875,
0.37890625,
0.31982421875,
0.392578125,
0.16796875,
0.392578125
],
[
0.391357421875,
0.391357421875,
0.450927734375,
0.393798828125,
0.44921875,
0.407958984375,
0.3896484375,
0.40625
],
[
0.4609375,
0.391357421875,
0.52880859375,
0.394287109375,
0.5283203125,
0.4091796875,
0.459228515625,
0.40625
],
[
0.54150390625,
0.391357421875,
0.60107421875,
0.393798828125,
0.599609375,
0.407958984375,
0.5400390625,
0.40625
],
[
0.61279296875,
0.391357421875,
0.67822265625,
0.393798828125,
0.6767578125,
0.407958984375,
0.611328125,
0.40625
],
[
0.69189453125,
0.391357421875,
0.75244140625,
0.393798828125,
0.75146484375,
0.407958984375,
0.68994140625,
0.40625
],
[
0.76416015625,
0.391357421875,
0.8310546875,
0.394287109375,
0.82958984375,
0.4091796875,
0.76318359375,
0.40625
],
[
0.4609375,
0.405029296875,
0.52880859375,
0.4091796875,
0.5283203125,
0.42333984375,
0.459228515625,
0.419189453125
],
[
0.16796875,
0.393798828125,
0.343505859375,
0.393798828125,
0.343505859375,
0.40673828125,
0.16796875,
0.40673828125
],
[
0.61279296875,
0.405029296875,
0.6767578125,
0.407958984375,
0.67529296875,
0.42333984375,
0.611328125,
0.42041015625
],
[
0.76416015625,
0.405029296875,
0.8310546875,
0.4091796875,
0.82958984375,
0.42333984375,
0.76318359375,
0.419189453125
],
[
0.54150390625,
0.40625,
0.60205078125,
0.4091796875,
0.60107421875,
0.42333984375,
0.5400390625,
0.42041015625
],
[
0.3896484375,
0.40625,
0.450927734375,
0.407958984375,
0.44921875,
0.4228515625,
0.387939453125,
0.42041015625
],
[
0.69189453125,
0.40625,
0.75146484375,
0.4091796875,
0.75,
0.42333984375,
0.68994140625,
0.42041015625
],
[
0.69189453125,
0.419189453125,
0.75146484375,
0.4228515625,
0.75,
0.438232421875,
0.68994140625,
0.435302734375
],
[
0.16796875,
0.407958984375,
0.345947265625,
0.407958984375,
0.345947265625,
0.421630859375,
0.16796875,
0.421630859375
],
[
0.166259765625,
0.42041015625,
0.26025390625,
0.4228515625,
0.26025390625,
0.43701171875,
0.166259765625,
0.435302734375
],
[
0.387939453125,
0.42041015625,
0.450927734375,
0.4228515625,
0.44921875,
0.43701171875,
0.38623046875,
0.435302734375
],
[
0.459228515625,
0.42041015625,
0.52880859375,
0.4228515625,
0.52880859375,
0.43701171875,
0.459228515625,
0.435302734375
],
[
0.54150390625,
0.42041015625,
0.60107421875,
0.4228515625,
0.599609375,
0.43701171875,
0.5400390625,
0.435302734375
],
[
0.61279296875,
0.42041015625,
0.67822265625,
0.42333984375,
0.6767578125,
0.43701171875,
0.611328125,
0.43408203125
],
[
0.76318359375,
0.419189453125,
0.8330078125,
0.42333984375,
0.8310546875,
0.438232421875,
0.76171875,
0.43408203125
],
[
0.11419677734375,
0.449462890625,
0.8818359375,
0.45068359375,
0.8818359375,
0.465576171875,
0.11419677734375,
0.46435546875
],
[
0.1158447265625,
0.465576171875,
0.80029296875,
0.465576171875,
0.80029296875,
0.479736328125,
0.1158447265625,
0.479736328125
],
[
0.1175537109375,
0.505859375,
0.489501953125,
0.505859375,
0.489501953125,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.486083984375,
0.51123046875,
0.50146484375,
0.51123046875,
0.50146484375,
0.515625,
0.486083984375,
0.515625
],
[
0.50390625,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.50390625,
0.51953125
],
[
0.1158447265625,
0.5224609375,
0.424072265625,
0.5224609375,
0.424072265625,
0.5361328125,
0.1158447265625,
0.5361328125
],
[
0.509765625,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.5361328125,
0.509765625,
0.5361328125
],
[
0.1318359375,
0.5380859375,
0.490234375,
0.54150390625,
0.490234375,
0.5625,
0.1318359375,
0.55859375
],
[
0.509765625,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.509765625,
0.5517578125
],
[
0.5087890625,
0.552734375,
0.8818359375,
0.55419921875,
0.8818359375,
0.568359375,
0.5087890625,
0.56787109375
],
[
0.1175537109375,
0.56005859375,
0.489501953125,
0.56005859375,
0.489501953125,
0.57373046875,
0.1175537109375,
0.57373046875
],
[
0.51123046875,
0.57080078125,
0.76416015625,
0.57080078125,
0.76416015625,
0.5849609375,
0.51123046875,
0.5849609375
],
[
0.1175537109375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.1175537109375,
0.59033203125
],
[
0.1175537109375,
0.59326171875,
0.48779296875,
0.59326171875,
0.48779296875,
0.60693359375,
0.1175537109375,
0.60693359375
],
[
0.51318359375,
0.59423828125,
0.8759765625,
0.59423828125,
0.8759765625,
0.609375,
0.51318359375,
0.609375
],
[
0.1158447265625,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62255859375,
0.1158447265625,
0.62255859375
],
[
0.51123046875,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.51123046875,
0.62890625
],
[
0.1158447265625,
0.62353515625,
0.48779296875,
0.62451171875,
0.48779296875,
0.6396484375,
0.1158447265625,
0.63818359375
],
[
0.509765625,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.509765625,
0.64453125
],
[
0.1158447265625,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.1158447265625,
0.65478515625
],
[
0.509765625,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.66015625,
0.509765625,
0.66015625
],
[
0.11419677734375,
0.65625,
0.489501953125,
0.65625,
0.489501953125,
0.67333984375,
0.11419677734375,
0.67333984375
],
[
0.509765625,
0.662109375,
0.88671875,
0.6630859375,
0.88671875,
0.677734375,
0.509765625,
0.6767578125
],
[
0.1158447265625,
0.67333984375,
0.489501953125,
0.67333984375,
0.489501953125,
0.68701171875,
0.1158447265625,
0.68701171875
],
[
0.51318359375,
0.6787109375,
0.88330078125,
0.6787109375,
0.88330078125,
0.693359375,
0.51318359375,
0.693359375
],
[
0.1158447265625,
0.689453125,
0.48779296875,
0.689453125,
0.48779296875,
0.703125,
0.1158447265625,
0.703125
],
[
0.51123046875,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.51123046875,
0.71044921875
],
[
0.11419677734375,
0.70361328125,
0.489501953125,
0.70361328125,
0.489501953125,
0.71875,
0.11419677734375,
0.71875
],
[
0.51318359375,
0.7119140625,
0.8818359375,
0.7119140625,
0.8818359375,
0.7255859375,
0.51318359375,
0.7255859375
],
[
0.1158447265625,
0.7216796875,
0.489501953125,
0.7216796875,
0.489501953125,
0.736328125,
0.1158447265625,
0.736328125
],
[
0.51123046875,
0.72802734375,
0.875,
0.72802734375,
0.875,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1158447265625,
0.73681640625,
0.48779296875,
0.736328125,
0.48779296875,
0.75048828125,
0.1158447265625,
0.751953125
],
[
0.5263671875,
0.7412109375,
0.88330078125,
0.7421875,
0.88330078125,
0.76025390625,
0.5263671875,
0.7587890625
],
[
0.1175537109375,
0.75439453125,
0.48779296875,
0.75439453125,
0.48779296875,
0.76708984375,
0.1175537109375,
0.76708984375
],
[
0.50732421875,
0.755859375,
0.884765625,
0.75830078125,
0.884765625,
0.779296875,
0.50732421875,
0.77685546875
],
[
0.1158447265625,
0.7685546875,
0.43408203125,
0.76953125,
0.43408203125,
0.78369140625,
0.1158447265625,
0.783203125
],
[
0.51123046875,
0.775390625,
0.88330078125,
0.775390625,
0.88330078125,
0.79052734375,
0.51123046875,
0.79052734375
],
[
0.1351318359375,
0.79052734375,
0.491943359375,
0.79052734375,
0.491943359375,
0.80419921875,
0.1351318359375,
0.80419921875
],
[
0.509765625,
0.7880859375,
0.8818359375,
0.79052734375,
0.8818359375,
0.80810546875,
0.509765625,
0.8056640625
],
[
0.1175537109375,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.822265625,
0.1175537109375,
0.822265625
],
[
0.509765625,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1175537109375,
0.82373046875,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.51123046875,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1158447265625,
0.8388671875,
0.490234375,
0.84033203125,
0.490234375,
0.85498046875,
0.1158447265625,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.8388671875,
0.8818359375,
0.85400390625,
0.509765625,
0.85498046875
],
[
0.1158447265625,
0.85693359375,
0.490234375,
0.85693359375,
0.490234375,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.509765625,
0.85693359375,
0.77685546875,
0.85693359375,
0.77685546875,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.11419677734375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.52880859375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.52880859375,
0.8857421875
],
[
0.1158447265625,
0.8896484375,
0.491943359375,
0.8896484375,
0.491943359375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.509765625,
0.88720703125,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.88330078125,
0.90478515625,
0.88330078125,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"SST-2",
"MRPC",
"QNLI",
"BERTbase → BERTiarge",
"Accuracy",
"Accuracy",
"Accuracy",
"RNF",
"RNF",
"RNF",
"92.000.27",
"85.690.90",
"90.740.09",
"-",
"Old Model",
"-",
"2.180.21",
"93.120.29",
"4.121.04",
"87.401.02",
"2.720.13",
"92.220.16",
"3.530.77",
"New Model",
"1.970.22",
"93.330.20",
"87.701.34",
"2.310.14",
"92.600.19",
"2.000.31",
"Distillation (Xie et al., 2021)",
"93.300.24",
"2.250.61",
"88.870.77",
"1.980.21",
"92.970.22",
"93.120.38",
"NewModelEnsemble",
"87.161.12",
"92.440.23",
"1.670.78",
"1.060.27",
"1.040.26",
"92.440.23",
"Old-New Probs Ensemble",
"93.120.38",
"1.040.26",
"1.060.27",
"Old-New Logits Ensemble",
"1.670.78",
"87.161.12",
"0.780.20",
"93.050.09",
"1.180.52",
"0.730.13",
"92.240.24",
"87.450.52",
"GatedFusion",
"Table 1: Negative flip rate RnF and model accuracy (%) of competing methods to optimize backward-",
"compatibility without performance degradation during BERTbase -→ BERTiarge model upgrade.",
"SST-2",
"MRPC",
"QNLI",
"BERTbase → ELECTRAbase",
"RNF",
"RNF",
"RNF",
"Accuracy",
"Accuracy",
"Accuracy",
"92.000.27",
"85.690.90",
"90.740.09",
"1.630.20",
"Old Model",
"95.000.06",
"3.730.36",
"88.580.57",
"2.820.32",
"92.900.26",
"1.490.24",
"New Model",
"95.020.21",
"3.680.79",
"88.820.94",
"93.030.16",
"2.580.17",
"Distillation (Xie et al., 2021)",
"95.390.09",
"3.240.24",
"89.020.48",
"2.260.08",
"93.490.07",
"1.120.09",
"New Model Ensemble",
"1.400.17",
"95.070.15",
"3.140.42",
"88.530.48",
"0.980.20",
"93.040.21",
"94.950.13",
"Old-New Probs Ensemble",
"88.480.51",
"93.040.21",
"3.280.43",
"0.890.17",
"0.980.20",
"0.810.16",
"Old-New Logits Ensemble",
"Gated Fusion",
"0.710.18",
"95.020.16",
"2.400.50",
"88.680.68",
"92.980.17",
"Table 2: Negative flip rate RNF and model accuracy (%) of competing methods to optimize backward",
"compatibility without performance degradation during BERTbase -→ ELECTRAbase model upgrade.",
"the model size is tripled (110M vs 340M) and the",
"",
"-0.31% on QNLI, which translate to further re-",
"model depth is doubled (12 vs 24 layers).",
"ducing 12.8%, 11.9%, and 11.4% of negative flip",
"Table I shows the results. For frew, we can",
"cases. These results show the effectiveness of our",
"method to mitigate a significant amount of regres-",
"observe that the negative flip rates RNF are usu-",
"sion errors during model upgrade.",
"ally much larger than the accuracy gains across",
"tasks, which could be the reason to hinder new",
"5.2Upgrade to a Different Pretrained Model",
"model adoptions in real-world applications.Be-",
"A more challenging upgrade scenario is when old",
"sides, when dividing RNF over the error rate",
"and new models are pretrained under distinctive",
"(1 - accuracy), we can observe that around 30%",
"paradigms, producing two representation spaces",
"to 40% of all fnew prediction errors are in fact the",
"of fairly different characteristics (Meng et al.,",
"new errors introduced during model upgrade. For",
"2021b). We experiment with BERTbase to ELEC-",
"improving prediction backward-compatibility, our",
"TRAbase in this scenario, where two models have",
"proposed Gated Fusion outperforms other compet-",
"the same size but are pretrained under utterly dif-",
"ing methods to considerably reduce RNF without",
"ferent schemes, i.e. generative versus adversarial.",
"degradation on accuracy. Note that best α values",
"Table 2 shows the results. For fhew, compared",
"found for the two variants of old-new ensemble are",
"with upgrading to BERTlarge, we observe larger ac-",
"both 0.5, hence producing identical results.",
"curacy gains and lower RNF on SST-2 and MRPC.",
"Compared to the vanilla new model, gated fu-",
"However, on QNLI, upgrading to ELECTRAbase",
"sion obtains absolute RNF reductions of -1.40%",
"achieves a higher accuracy gain but an even a",
"on SST-2, -2.94% on MRPC, and -1.99% on",
"higher RNF. This implies that boosting accuracy",
"QNLI. These translate to reducing the total neg-",
"and improving backward compatibility could be",
"ative flip cases by 64.2%, 71.4%, 73.2%, respec-",
"two related but different objectives.",
"tively. Compared to the strongest baseline (old-new",
"For mitigation strategies, Gated Fusion achieves",
"ensemble), we obtain further absolute RNF reduc-",
"the lowest negative flip rates across datasets with-",
"tions of -0.28% on SST-2, -0.49% on MRPC, and",
" out any accuracy loss. We obtain absolute RNF re-"
] |
{
"arxiv_id": "2302.02080",
"language": "en",
"page_idx": 4
}
|
||
2110.02444-en-0
|
2110.02444-en-0.jpg
|
Influence-Balanced Loss for Imbalanced Visual Classification
Seulki Park Jongin Lim Younghan Jeon Jin Young Choi
ASRI, Dept. of Electrical and Computer Engineering, Seoul National University
$\{$seulki.park, ljin0429, yh1992, jychoi$\}$@snu.ac.kr
## $\mathbf{Abstract}$
In this paper, we propose a balancing training method to address problems in imbalanced data learning. To this end, we derive a new loss used in the balancing training phase that alleviates the influence of samples that cause an overftted decision boundary. The proposed loss efficiently improves the performance of any type of imbalance learning methods. In experiments on multiple benchmark data sets, we demonstrate the validity of our method and reveal that the proposed loss outperforms the state-of-the-art $cost- sensitive\textit{ loss methods. Furthermore, since our loss is}$ not restricted to a specific task, model, or training method, it can be easily used in combination with other recent resampling, meta-learning, and cost- sensitive learning meth- $odsforclass-imbalance$ problems. Our code is made avail- $ableathttps://github.com/pseulki/IB-Loss.$
## $1. \textbf{Introduction}$
Despite the remarkable success of deep neural networks (DNNs) these days, many areas of computer vision suffer from highly imbalanced datasets. Many real-world data exhibit skewed distributions [23, 16, 11, 24, 10], in which the number of samples per class differs greatly. This imbalance between classes can be problematic, since the model trained on such imbalanced data tends to overfit the dominant (majority) classes [18, 14, 4]. That is, while the overall performance appears to be satisfactory, the model performs poorly on minority classes. To overcome the class imbalance problem, extensive research has recently been conducted to improve the generalization performance by reducing the overwhelming influence of the dominant class on the model.
The research on imbalanced learning can be divided into three approaches: data-level approach, cost-sensitive re-weighting approach, and meta-learning approach. The data-level approach aims to directly balance the training data distributions via re-sampling (i.e., under-sampling or over-sampling) [6, 32] or by generating synthetic samples [28]. Meanwhile, the cost-sensitive re-weighting approach aims to design new loss functions to re-weight samples by
considering their importance [33, 17, 22]. Finally, the metalearning approach enhances the performance of the datalevel and/or cost-sensitive re-weighting approach via metalearning [31, 25, 30]. Most recent data-level approaches require a heavy computational burden. Moreover, undersampling can lose some valuable information, and oversampling or data generation can cause overfitting on certain repetitive samples. The meta-learning approach requires additional unbiased data[31] or a meta-sampler[30], which is computationally expensive in practice. Therefore, our work focuses on the cost-sensitive re-weighting approach to design a new loss function that is simple but efficient.
The cost-sensitive re-weighting approach aims to assign class penalties to shift the decision boundary in a way that reduces the bias induced by the data imbalance. For this purpose, the most commonly adopted method is to re-weight samples inversely to the number of training samples in each class to assign more weights for the minority classes [17, 33, 8]. These methods have focused on only globallevel class distribution and assign the same fixed weight to all samples belonging to the same class. However, not all samples in a dataset play an equal role in determining the model parameters[7].That is, some samples have greater influences on forming a decision boundary. Hence, each sample needs to be re-weighted differently according to its impact on the model.
Recently, numerous studies have been conducted in which each sample is considered to design sample-wise loss functions [9,22,27]. Specifically, these methods downweight well-classified samples and assign more weights to hard examples, which yield high errors. This re-weighting might lead to the complete training when the high capacity of DNNs is sufficient to fınally memorize the whole training data[34, 3]. This implies that DNN is overfitted to hard samples, which are located at the overlapping region between the majority and minority classes. In the imbalanced data, most hard samples are majority samples that enforce the decision boundary to be complex and shift to the minority region.
To address the aforementioned problem, in this paper, we propose a loss-sensitive method to down-weight sam-
|
[
[
0.1805419921875,
0.1331787109375,
0.78759765625,
0.1331787109375,
0.78759765625,
0.146484375,
0.1805419921875,
0.146484375
],
[
0.22216796875,
0.18310546875,
0.71826171875,
0.18310546875,
0.71826171875,
0.1976318359375,
0.22216796875,
0.1976318359375
],
[
0.1544189453125,
0.1976318359375,
0.78515625,
0.1988525390625,
0.78515625,
0.216552734375,
0.1544189453125,
0.21533203125
],
[
0.2548828125,
0.2183837890625,
0.6845703125,
0.2197265625,
0.6845703125,
0.2342529296875,
0.2548828125,
0.23291015625
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.50244140625,
0.270751953125,
0.890625,
0.270751953125,
0.890625,
0.2841796875,
0.50244140625,
0.2841796875
],
[
0.5009765625,
0.285888671875,
0.89208984375,
0.285400390625,
0.89208984375,
0.2998046875,
0.5009765625,
0.300537109375
],
[
0.098876953125,
0.3017578125,
0.4697265625,
0.3017578125,
0.4697265625,
0.31494140625,
0.098876953125,
0.31494140625
],
[
0.5009765625,
0.300537109375,
0.8896484375,
0.300537109375,
0.8896484375,
0.314453125,
0.5009765625,
0.314453125
],
[
0.080078125,
0.317626953125,
0.4697265625,
0.317626953125,
0.4697265625,
0.330810546875,
0.080078125,
0.330810546875
],
[
0.50244140625,
0.317626953125,
0.89208984375,
0.317626953125,
0.89208984375,
0.330810546875,
0.50244140625,
0.330810546875
],
[
0.080078125,
0.33203125,
0.468017578125,
0.33203125,
0.468017578125,
0.34521484375,
0.080078125,
0.34521484375
],
[
0.50244140625,
0.33203125,
0.890625,
0.33203125,
0.890625,
0.34521484375,
0.50244140625,
0.34521484375
],
[
0.07843017578125,
0.34521484375,
0.4697265625,
0.3466796875,
0.4697265625,
0.361083984375,
0.07843017578125,
0.35986328125
],
[
0.50244140625,
0.347900390625,
0.89208984375,
0.347900390625,
0.89208984375,
0.361083984375,
0.50244140625,
0.361083984375
],
[
0.080078125,
0.3623046875,
0.468017578125,
0.3623046875,
0.468017578125,
0.375732421875,
0.080078125,
0.375732421875
],
[
0.50244140625,
0.3623046875,
0.89208984375,
0.3623046875,
0.89208984375,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.080078125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.50244140625,
0.37744140625,
0.890625,
0.37744140625,
0.890625,
0.391357421875,
0.50244140625,
0.391357421875
],
[
0.080078125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.50244140625,
0.392578125,
0.89208984375,
0.392578125,
0.89208984375,
0.406005859375,
0.50244140625,
0.406005859375
],
[
0.08087158203125,
0.4091796875,
0.46728515625,
0.4091796875,
0.46728515625,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50341796875,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.419189453125,
0.50341796875,
0.419189453125
],
[
0.080078125,
0.42236328125,
0.468017578125,
0.42236328125,
0.468017578125,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.4375,
0.50244140625,
0.4375
],
[
0.080078125,
0.4375,
0.468017578125,
0.4375,
0.468017578125,
0.45068359375,
0.080078125,
0.45068359375
],
[
0.5009765625,
0.4375,
0.8408203125,
0.436279296875,
0.8408203125,
0.45068359375,
0.5009765625,
0.451904296875
],
[
0.080078125,
0.45263671875,
0.468017578125,
0.45263671875,
0.468017578125,
0.466552734375,
0.080078125,
0.466552734375
],
[
0.521484375,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.521484375,
0.4697265625
],
[
0.08087158203125,
0.46728515625,
0.46728515625,
0.46728515625,
0.46728515625,
0.48095703125,
0.08087158203125,
0.48095703125
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.080078125,
0.48291015625,
0.46728515625,
0.48291015625,
0.46728515625,
0.496826171875,
0.080078125,
0.496826171875
],
[
0.5009765625,
0.48486328125,
0.890625,
0.486083984375,
0.890625,
0.50048828125,
0.5009765625,
0.5
],
[
0.08087158203125,
0.49755859375,
0.468017578125,
0.49755859375,
0.468017578125,
0.51123046875,
0.08087158203125,
0.51123046875
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.080078125,
0.51318359375,
0.468017578125,
0.51318359375,
0.468017578125,
0.5263671875,
0.080078125,
0.5263671875
],
[
0.5009765625,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.5009765625,
0.53076171875
],
[
0.50244140625,
0.53076171875,
0.890625,
0.53076171875,
0.890625,
0.54443359375,
0.50244140625,
0.54443359375
],
[
0.50341796875,
0.54541015625,
0.890625,
0.546875,
0.890625,
0.56103515625,
0.50341796875,
0.56005859375
],
[
0.08087158203125,
0.55810546875,
0.2059326171875,
0.55810546875,
0.2059326171875,
0.5712890625,
0.08087158203125,
0.5712890625
],
[
0.50341796875,
0.5625,
0.89208984375,
0.5625,
0.89208984375,
0.57568359375,
0.50341796875,
0.57568359375
],
[
0.09722900390625,
0.583984375,
0.4697265625,
0.58349609375,
0.4697265625,
0.59765625,
0.09722900390625,
0.5986328125
],
[
0.5048828125,
0.5791015625,
0.8896484375,
0.5791015625,
0.8896484375,
0.5888671875,
0.5048828125,
0.5888671875
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.080078125,
0.5986328125,
0.468017578125,
0.599609375,
0.468017578125,
0.6142578125,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.60693359375,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.62060546875
],
[
0.08087158203125,
0.6162109375,
0.46728515625,
0.6162109375,
0.46728515625,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.5009765625,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63525390625,
0.5009765625,
0.63525390625
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.63623046875,
0.89208984375,
0.63623046875,
0.89208984375,
0.65087890625,
0.50244140625,
0.65087890625
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.5009765625,
0.6513671875,
0.6435546875,
0.65087890625,
0.6435546875,
0.66552734375,
0.5009765625,
0.66650390625
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.52294921875,
0.67041015625,
0.89208984375,
0.67041015625,
0.89208984375,
0.68505859375,
0.52294921875,
0.68505859375
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.68603515625,
0.89208984375,
0.68603515625,
0.89208984375,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50341796875,
0.71533203125
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71533203125,
0.89208984375,
0.71630859375,
0.89208984375,
0.73095703125,
0.50244140625,
0.72998046875
],
[
0.5009765625,
0.73095703125,
0.890625,
0.73193359375,
0.890625,
0.7470703125,
0.5009765625,
0.74560546875
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.08251953125,
0.75244140625,
0.46728515625,
0.75244140625,
0.46728515625,
0.76318359375,
0.08251953125,
0.76318359375
],
[
0.50244140625,
0.74755859375,
0.890625,
0.74755859375,
0.890625,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.4453125,
0.7666015625,
0.4453125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.76123046875,
0.890625,
0.76220703125,
0.890625,
0.77587890625,
0.5009765625,
0.7744140625
],
[
0.098876953125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.098876953125,
0.79443359375
],
[
0.50244140625,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79248046875,
0.890625,
0.79248046875,
0.890625,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.806640625,
0.89208984375,
0.806640625,
0.89208984375,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82275390625,
0.89208984375,
0.82275390625,
0.89208984375,
0.8359375,
0.50244140625,
0.8359375
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.83837890625,
0.890625,
0.83837890625,
0.890625,
0.8515625,
0.50341796875,
0.8515625
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85498046875,
0.57177734375,
0.85498046875,
0.57177734375,
0.86572265625,
0.50244140625,
0.86572265625
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.51953125,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.51953125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Infuence-Balanced LossforImbalancedVisual Classificatior",
"SeulkiParkJonginLimYounghan JeonJinYoungChoi",
"ASRI, Dept. of Electrical and Computer Engineering, Seoul National University",
"(seulki.park,1jin0429,yh1992,jychoi)@snu.ac.kr",
"Abstract",
"considering their importance [33, 17, 22]. Finally, the meta-",
"learning approach enhances the performance of the data-",
"In this paper, we propose a balancing training method",
"level and/or cost-sensitive re-weighting approach via meta-",
"to address problems in imbalanced data learning. To this",
"learning [31, 25, 30]. Most recent data-level approaches",
"end, we derive a new loss used in the balancing training",
"require a heavy computational burden. Moreover, under-",
"phase that alleviates the influence of samples that cause an",
"sampling can lose some valuable information, and over-",
"overfitted decision boundary. The proposed loss eficiently",
"sampling or data generation can cause overfitting on certain",
"improves the performance of any type of imbalance learn-",
"repetitive samples. The meta-learning approach requires ad-",
"ing methods. In experiments on multiple benchmark data",
"ditional unbiased data [31] or a meta-sampler [30], which is",
"sets. we demonstrate the validity of our method and re-",
"computationallyexpensiveinpractice.Therefore,ourwork",
"veal that the proposed loss outperforms the state-of-the-art",
"focuses on the cost-sensitive re-weighting approach to de-",
"cost-sensitive loss methods. Furthermore, since our loss is",
"sign a new loss function that is simple but efficient.",
"not restricted to a specific task, model, or training method,",
"The cost-sensitive re-weighting approach aims to assign",
"it can be easily used in combination with other recent re-",
"class penalties to shift the decision boundary in a way that",
"sampling, meta-learning, and cost-sensitive learning meth-",
"reduces the bias induced by the data imbalance.For this pur-",
"ods for class-imbalance problems. Our code is made avail-",
"pose, the most commonly adopted method is to re-weight",
"ableathttps://github.com/pseulki/IB-Loss.",
"samples inversely to the number of training samples in",
"each class to assign more weights for the minority classes",
"[17, 33, 8]. These methods have focused on only global-",
"1. Introduction",
"level class distribution and assign the same fixed weight to",
"Despite the remarkable success of deep neural networks",
"allsamplesbelonging tothesameclass.However,notall",
"samples in a dataset play an equal role in determining the",
"(DNNs) these days, many areas of computer vision suffer",
"model parameters [7]. That is, some samples have greater",
"fromhighly imbalanced datasets.Many real-world data ex-",
"influences on forming a decision boundary.Hence, each",
"hibit skewed distributions [23, 16, 11, 24, 10], in which the",
"sample needs to be re-weighted differently according to its",
"number of samples per class differs greatly. This imbalance",
"impact on the model.",
"between classes can be problematic, since the model trained",
"on such imbalanced data tends to overfit the dominant (ma-",
"Recently, numerous studies have been conducted in",
"jority) classes [18, 14, 4]. That is, while the overall perfor-",
"which each sample is considered to design sample-wise loss",
"mance appears to be satisfactory, the model performs poorly",
"functions [9, 22, 27]. Specifically, these methods down-",
"on minority classes. To overcome the class imbalance prob-",
"weight well-classified samples and assign more weights to",
"hard examples, which yield high errors. This re-weighting",
"lem,extensiveresearchhasrecentlybeenconductedtoim-",
"provethegeneralizationperformancebyreducingthe over-",
"might lead to the complete training when the high capacity",
"whelming influence of the dominant class on the model.",
"of DNNs is sufficient to finallymemorize the whole train-",
"The research on imbalanced learning can be divided",
"ing data [34, 3]. This implies that DNN is overfitted to hard",
"into three approaches: data-level approach, cost-sensitive",
"samples, which are located at the overlapping region be-",
"re-weighting approach, and meta-learning approach. The",
"tween the majority and minority classes. In the imbalanced",
"data-level approach aims to directly balance the training",
"data, most hard samples are majority samples that enforce",
"data distributions via re-sampling (i.e., under-sampling or",
"the decision boundary to be complex and shift to the minor-",
"over-sampling) [6, 32] or by generating synthetic samples",
"ityregion.",
"[28]. Meanwhile, the cost-sensitive re-weighting approach",
"To address the aforementioned problem, in this paper,",
"aims to design new loss functions to re-weight samples by",
"we propose a loss-sensitive method to down-weight sam-"
] |
{
"arxiv_id": "2110.02444",
"language": "en",
"page_idx": 0
}
|
||
2110.02444-en-1
|
2110.02444-en-1.jpg
|
ples that cause overfitting of a DNN trained with highly imbalanced data. To this end, we derive a formula that measures how much each sample influences the complex and biased decision boundary. To derive the formula, we utilize the influence function [7], which has been widely used in robust statistics. Using the derived formula, we design a novel loss function, called influence-balanced (IB) loss, that adaptively assigns different weights to samples according to their influence on a decision boundary. Specifically, we re-weight the loss proportionally to the inverse of the influence of each sample. Our method is divided into two phases: standard training and fine-tuning for influence balancing. During the fine-tuning phase, the proposed IB loss alleviates the influence of the samples that cause overfitting of the decision boundary.
Through extensive experiments on multiple benchmark data sets, we demonstrate the validity of our method, and show that the proposed method outperforms the state-ofthe-art cost-sensitive re-weighting methods. Furthermore, since our IB loss is not restricted to a specific task, model, or training method, it can be easily utilized in combination with other recent data-level algorithms and hybrid methods for class-imbalance problems.
The main contributions of this paper are as follows:
$\cdot$ We discover that the existing loss-based loss methods can lead a decision boundary of DNNs to eventually overfit to the majority classes.
· We design a novel influence-balanced loss function to re-weight samples more effectively in such a way that the overfitting of the decision boundary can be alleviated.
· We demonstrate that simply substituting our proposed
loss for the standard cross-entropy loss significantly improves the generalization performance on highly imbalanced data.
## $2. \textbf{Related Work}$
2.1. Class Imbalance Learning
To solve the imbalanced learning problem, numerous studies have been conducted. The research can be divided into three approaches: data-level, cost-sensitive reweighting, and meta-learning approaches.
Data-level approach. The data-level approach aims to directly balance the training data distributions by resampling (e.g., under-sampling the majority classes or oversampling the minority classes) [6,32] or generating synthetic samples [28]. However, under-sampling can lose some valuable information, and it is not applicable when the data imbalance between classes is significant. Although over-sampling or data generation could be effective, these methods are susceptible to overfitting to certain repetitive samples, and often require a longer training time.
Re-weighting approach. Cost-sensitive re-weighting methods assign different weights to samples to adjust their importance. Commonly used methods include re-weighting samples inversely proportional to the number of the class [17, 33] or the square root of class frequency [26]. Instead of heuristically using the number of classes, Cui et al.[8] proposed using the effective number of samples. While these methods can successfully assign more weights to the minority samples, they assign the same weights to all samples belonging to the same class, regardless of each importance.
To assign different weights to each sample according to its importance on the model, numerous methods were proposed for re-weighting samples based on their difficulties or losses [22, 9, 27]. That is, these methods down-weight well-classified samples and assign more weights to hard examples. These re-weighting methods might cause DNNs to be overfitted to the hard examples, since the high capacity of DNNs is sufficient to memorize the training data in the end [3]. In class imbalanced data, the hard examples are likely generated from the majority classes. As such, the minority samples are assigned smaller weights. Therefore, we need a more elaborate mean of re-weighting samples that can alleviate the overfitting to the majority samples. Meanwhile, Cao et al.[5] proposed label-distribution-aware margin loss to solve the overfitting to the minority classes by regularizing the margins.
Meta-learning approach. Recently, the meta-learningbased approach [31,25,30] has emerged to enhance the performance of both approaches. Shu et al.[31] proposed a meta-learning process to learn a weighting function, while Liu et al.[25] proposed a re-sampling method by combining the advantage of ensemble learning and meta-learning. Furthermore, Ren et al.[30] proposed the meta-sampler and a balanced softmax that accommodates the shift of the distributions between the training data and test data. Although these methods can achieve satisfactory performance, these methods are somewhat difficult to implement in practice. For example, meta-weight-net[31] requires additional unbiased data for learning, and the meta-sampler in [30] is computationally expensive in practice. On the other hand, our proposed loss is simple to implement because it does not require a hyperparameter, a specially designed architecture, or additional learning for data re-sampling. Therefore, it is easy to use in collaboration with other methods.
$2. 2. \textbf{ Influence function. }$
The influence function was proposed to find the influential instance of a sample to a model, which has been studied for decades in robust statistics [13,7]. Recently, attempts have been made to use influence function in deep neural networks [1, 19]. For example, Koh and Liang [19] employed the influence function to understand DNNs. While the influence function is primarily used as a diagnostic tool after the
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.521484375,
0.0889892578125,
0.89208984375,
0.09027099609375,
0.89208984375,
0.10797119140625,
0.521484375,
0.106689453125
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.12371826171875,
0.890625,
0.12371826171875,
0.890625,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.08087158203125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.08087158203125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.5009765625,
0.1666259765625,
0.890625,
0.1673583984375,
0.890625,
0.1817626953125,
0.5009765625,
0.18115234375
],
[
0.07843017578125,
0.1817626953125,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.07843017578125,
0.1956787109375
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1988525390625,
0.46728515625,
0.1988525390625,
0.46728515625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50341796875,
0.199462890625,
0.8896484375,
0.199462890625,
0.8896484375,
0.210205078125,
0.50341796875,
0.210205078125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50341796875,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.5009765625,
0.2266845703125,
0.890625,
0.2291259765625,
0.890625,
0.24365234375,
0.5009765625,
0.2412109375
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.521484375,
0.24365234375,
0.89208984375,
0.24560546875,
0.89208984375,
0.260009765625,
0.521484375,
0.25830078125
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.2734375,
0.50341796875,
0.2734375
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.27587890625,
0.89208984375,
0.27587890625,
0.89208984375,
0.2890625,
0.50341796875,
0.2890625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.290283203125,
0.89208984375,
0.290283203125,
0.89208984375,
0.3037109375,
0.50244140625,
0.3037109375
],
[
0.07843017578125,
0.302978515625,
0.25,
0.3037109375,
0.25,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.5009765625,
0.3037109375,
0.89208984375,
0.304931640625,
0.89208984375,
0.3193359375,
0.5009765625,
0.318115234375
],
[
0.1004638671875,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.1004638671875,
0.33203125
],
[
0.50244140625,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.50244140625,
0.333984375
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50244140625,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50341796875,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.50341796875,
0.380126953125
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50244140625,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.39453125,
0.50244140625,
0.39453125
],
[
0.080078125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.080078125,
0.40771484375,
0.4697265625,
0.4091796875,
0.4697265625,
0.423583984375,
0.080078125,
0.42236328125
],
[
0.50341796875,
0.412353515625,
0.890625,
0.412353515625,
0.890625,
0.425537109375,
0.50341796875,
0.425537109375
],
[
0.08087158203125,
0.4248046875,
0.27685546875,
0.4248046875,
0.27685546875,
0.438232421875,
0.08087158203125,
0.438232421875
],
[
0.5009765625,
0.4267578125,
0.8896484375,
0.4267578125,
0.8896484375,
0.43994140625,
0.5009765625,
0.43994140625
],
[
0.098876953125,
0.439453125,
0.43701171875,
0.438232421875,
0.43701171875,
0.45263671875,
0.098876953125,
0.453857421875
],
[
0.5009765625,
0.44140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.455078125,
0.5009765625,
0.455810546875
],
[
0.098876953125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.474853515625,
0.098876953125,
0.474853515625
],
[
0.50244140625,
0.45703125,
0.890625,
0.45703125,
0.890625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.1094970703125,
0.474853515625,
0.468017578125,
0.47607421875,
0.468017578125,
0.490478515625,
0.1094970703125,
0.4892578125
],
[
0.50244140625,
0.4716796875,
0.61279296875,
0.4716796875,
0.61279296875,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.111083984375,
0.4912109375,
0.3095703125,
0.4912109375,
0.3095703125,
0.5048828125,
0.111083984375,
0.5048828125
],
[
0.521484375,
0.488037109375,
0.890625,
0.488037109375,
0.890625,
0.501953125,
0.521484375,
0.501953125
],
[
0.09722900390625,
0.5068359375,
0.468017578125,
0.50830078125,
0.468017578125,
0.52294921875,
0.09722900390625,
0.521484375
],
[
0.50244140625,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.50244140625,
0.517578125
],
[
0.50341796875,
0.51806640625,
0.89208984375,
0.51806640625,
0.89208984375,
0.5322265625,
0.50341796875,
0.5322265625
],
[
0.11273193359375,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.11273193359375,
0.53857421875
],
[
0.111083984375,
0.5390625,
0.4697265625,
0.5390625,
0.4697265625,
0.55322265625,
0.111083984375,
0.55322265625
],
[
0.50244140625,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.5009765625,
0.54736328125,
0.890625,
0.548828125,
0.890625,
0.56298828125,
0.5009765625,
0.5625
],
[
0.11273193359375,
0.55615234375,
0.148681640625,
0.55615234375,
0.148681640625,
0.56640625,
0.11273193359375,
0.56640625
],
[
0.50244140625,
0.564453125,
0.8896484375,
0.564453125,
0.8896484375,
0.57763671875,
0.50244140625,
0.57763671875
],
[
0.09722900390625,
0.57080078125,
0.468017578125,
0.5712890625,
0.468017578125,
0.5859375,
0.09722900390625,
0.58544921875
],
[
0.1094970703125,
0.5859375,
0.468017578125,
0.5869140625,
0.468017578125,
0.6015625,
0.1094970703125,
0.60107421875
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.5048828125,
0.59521484375,
0.88818359375,
0.59521484375,
0.88818359375,
0.60595703125,
0.5048828125,
0.60595703125
],
[
0.11273193359375,
0.60302734375,
0.46728515625,
0.60302734375,
0.46728515625,
0.6162109375,
0.11273193359375,
0.6162109375
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.111083984375,
0.6162109375,
0.226318359375,
0.6171875,
0.226318359375,
0.6318359375,
0.111083984375,
0.630859375
],
[
0.5048828125,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.63623046875,
0.5048828125,
0.63623046875
],
[
0.50341796875,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.08087158203125,
0.64599609375,
0.2181396484375,
0.64599609375,
0.2181396484375,
0.66015625,
0.08087158203125,
0.66015625
],
[
0.50244140625,
0.65283203125,
0.89208984375,
0.65380859375,
0.89208984375,
0.66845703125,
0.50244140625,
0.66748046875
],
[
0.080078125,
0.66845703125,
0.314453125,
0.67041015625,
0.314453125,
0.68505859375,
0.080078125,
0.68310546875
],
[
0.50341796875,
0.66943359375,
0.89208984375,
0.66943359375,
0.89208984375,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.09722900390625,
0.68798828125,
0.4697265625,
0.68994140625,
0.4697265625,
0.705078125,
0.09722900390625,
0.70263671875
],
[
0.50244140625,
0.68505859375,
0.890625,
0.68505859375,
0.890625,
0.69873046875,
0.50244140625,
0.69873046875
],
[
0.5009765625,
0.69970703125,
0.89208984375,
0.69873046875,
0.89208984375,
0.71337890625,
0.5009765625,
0.7138671875
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71533203125,
0.890625,
0.71533203125,
0.890625,
0.72900390625,
0.50244140625,
0.72900390625
],
[
0.5009765625,
0.72900390625,
0.890625,
0.72900390625,
0.890625,
0.74365234375,
0.5009765625,
0.74365234375
],
[
0.08087158203125,
0.7373046875,
0.3544921875,
0.7373046875,
0.3544921875,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.1004638671875,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.1004638671875,
0.76513671875
],
[
0.50244140625,
0.74560546875,
0.8447265625,
0.74560546875,
0.8447265625,
0.7587890625,
0.50244140625,
0.7587890625
],
[
0.080078125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.77294921875,
0.6806640625,
0.7744140625,
0.6806640625,
0.7890625,
0.50244140625,
0.787109375
],
[
0.080078125,
0.7822265625,
0.46728515625,
0.7822265625,
0.46728515625,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.521484375,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.404296875,
0.88720703125,
0.404296875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"ples that cause overfitting of a DNN trained with highly im-",
"Re-weighting approach. Cost-sensitive re-weighting",
"balanced data.To this end, we derive a formula that mea-",
"methods assign different weights to samples to adjust their",
"sures how mucheach sampleinfluences the complex and",
"importance. Commonly used methods include re-weighting",
"biased decision boundary. To derive the formula, we uti-",
"samples inversely proportional to the number of the class",
"lize the influence function [7], which has been widely used",
"[17, 33] or the square root of class frequency [26]. Instead of",
"in robust statistics. Using the derived formula, we design",
"heuristically using the number of classes, Cui et al. [8 pro-",
"a novel loss function,called influence-balanced(IB)loss,",
"posed using the effective number of samples. While these",
"that adaptively assigns different weights to samples accord-",
"methods can successfullyassign more weightstothemi-",
"ing to their influence on a decision boundary. Specifically,",
"nority samples, they assign the same weights to all samples",
"we re-weight the loss proportionally to the inverse of the",
"belonging to the same class, regardless of each importance.",
"influence of each sample. Our method is divided into two",
"To assign different weights to each sample according to",
"phases: standard training and fine-tuning for influence bal-",
"its importance on the model, numerous methods were pro-",
"ancing. During the fine-tuning phase, the proposed IB loss",
"posed for re-weighting samples based on their difficulties",
"alleviates the influence of the samples that cause overfitting",
"or losses [22, 9, 27]. That is, these methods down-weight",
"of the decision boundary.",
"well-classified samples and assign more weights to hard ex-",
"Through extensive experiments on multiple benchmark",
"amples. These re-weighting methods might cause DNNs to",
"data sets, we demonstrate the validity of our method, and",
"be overfitted to the hard examples, since the high capacity of",
"show that the proposed method outperforms the state-of-",
"DNNs is sufficient to memorize the training data in the end",
"the-art cost-sensitive re-weighting methods. Furthermore,",
"[3]. In class imbalanced data, the hard examples are likely",
"sinceour IB loss isnotrestricted toa specific task,model.",
"generated from the majority classes. As such, the minority",
"or training method, it can be easily utilized in combination",
"samples are assigned smaller weights. Therefore, we need",
"with other recent data-level algorithms and hybrid methods",
"a more elaborate mean of re-weighting samples that can al-",
"forclass-imbalance problems.",
"leviate the overfitting to the majority samples. Meanwhile,",
"The main contributions of this paper are as follows:",
"Cao et al. [5] proposed label-distribution-aware margin loss",
". We discover that the existing loss-based loss methods",
"to solve the overfitting to the minority classes by regulariz-",
"can lead a decision boundary of DNNs to eventually",
"ing the margins.",
"overfit to the majority classes.",
"Meta-learning approach. Recently, the meta-learning-",
". We design a novel influence-balanced loss function to",
"based approach [31, 25, 30] has emerged to enhance the",
"performance of both approaches. Shu et al. [31] proposed a",
"re-weight samples more effectively in such a way that",
"the overfitting of the decision boundary can be allevi-",
"meta-learning process to learn a weighting function, while",
"Liu et al. [25] proposed a re-sampling method by combin-",
"ated.",
"ing the advantage of ensemble learning and meta-learning",
".We demonstrate that simply substituting our proposed",
"loss for the standard cross-entropy loss significantly",
"Furthermore, Ren et al. [30l proposed the meta-sampler and",
"abalanced softmaxthataccommodatesthe shiftof the dis-",
"improves the generalization performance on highly",
"tributions between the training data and test data. Although",
"imbalanced data.",
"these methods can achieve satisfactoryperformance,these",
"methods are somewhat difficult to implement in practice.",
"2. Related Work",
"For example, meta-weight-net [31] requires additional un-",
"2.1.Class Imbalance Learning",
"biased data for learning, and the meta-sampler in [30] is",
"To solve the imbalanced learning problem, numerous",
"computationally expensive in practice. On the other hand",
"our proposed loss is simple to implement because it does",
"studies have been conducted. The research can be di-",
"vided into three approaches: data-level, cost-sensitive re-",
"not require a hyperparameter, a specially designed architec-",
"ture, or additional learning for data re-sampling. Therefore,",
"weighting,andmeta-learningapproaches.",
"Data-level approach. The data-level approach aims",
"it is easy to use in collaboration with other methods.",
"to directly balance the training data distributions by re-",
"2.2. Influence function.",
"sampling (e.g., under-sampling the majority classes or over-",
"sampling the minority classes) [6, 32] or generating syn-",
"The influence function was proposed to find the influen-",
"thetic samples [28]. However, under-sampling can lose",
"tial instance of a sample to a model, which has been studied",
"some valuable information, and it is not applicable when",
"for decades in robust statistics [13, 7]. Recently, attempts",
"the data imbalance between classes is significant. Although",
"have been made to use influence function in deep neural net-",
"over-sampling or data generation could be effective, these",
"works [1, 19]. For example, Koh and Liang [19] employed",
"methods are susceptible to overfitting to certain repetitive",
"the influence function to understand DNNs. While the influ-",
"samples, and often require a longer training time.",
"ence function is primarily used as a diagnostic tool after the"
] |
{
"arxiv_id": "2110.02444",
"language": "en",
"page_idx": 1
}
|
||
2110.02444-en-4
|
2110.02444-en-4.jpg
|
$4. \textbf{Experiments}$
# 4.1. Experimental Settings
Datasets. We verified the effectiveness of our method on three commonly used benchmark datasets: CIFAR-10, CIFAR-100[20], Tiny ImageNet[21], and iNaturalist 2018 [16]. The CIFAR-10 and CIFAR-100 datasets consist of 50,000 training images and 10,000 test images with 10 and 100 classes, respectively. Meanwhile, Tiny ImageNet contains 200 classes for training, in which each class has 500 images. Its test set contains 10,000 images. Since CIFAR and Tiny ImageNet are evenly distributed, we have made these datasets imbalanced according to [8,4], respectively. Primarily, we investigate two common types of imbalance: (i) long-tailed imbalance [8] and (ii) step imbalance [4]. In long-tailed imbalance, the number of training samples for each class decreases exponentially from the largest majority class to the smallest minority class. To construct long-tailed imbalanced datasets, the number of selected samples in the $k$-th class was set to $n_k\mu^k(\mu\in(0,1))$,where $n_k$ is the original number of the $k$-th class. Meanwhile, in step imbalance, the classes are divided into two groups: the majority class group and minority class group. Every class within a group contains the same number of samples, and the class in the majority class group has many more samples than that in the minority class group. For evaluation, we used the original test set. The imbalance ratio $\rho$ is defined by $\rho=\frac{\max_{k}\left\{n_{k}\right\}}{\min_{k}\left\{n_{k}\right\}}.$ Thus, the imbalance ratio represents the degree of imbalance in the dataset. We evaluated the performance of our method under various imbalance ratios from 10 to 200.
The iNaturalist 2018 dataset is a large-scale real-world dataset containing 437,513 training images and 24,426 test images with 8,142 classes. iNaturalist 2018 exhibits longtailed imbalance, whose imbalance ratio is 500. We used the official training and test splits in our experiments.
Baselines. We compared our algorithm with the follow-
ing cost-sensitive loss methods:(1)Our baseline model, which is trained on the standard cross-entropy loss. Comparing our model with this baseline enables us to clearly understand how much our training scheme has improved the performance; (2) focal loss [22], which increases the relative loss for hard samples and down-weights well-classified samples; (3) CB loss [8], which re-weights the loss inversely proportional to the effective number of samples; (4) LDAM loss[5], which regularizes the minority classes to have larger margins.
Since our IB loss can be easily combined with other methods, we employee two further variants. First, IB+CB uses the effective number in CB loss, instead of using $\lambda_k$ in IB. Second, IB + focal uses focal loss during the finetuning phase, instead of using the cross-entropy loss. We demonstrate that combination with other methods can fur-
ther improve the performance.
Implementation Details. We used PyTorch[29] to implement and train all the models in the paper, and we used ResNet architecture[15] for all datasets. For CIFAR datasets, we used randomly initialized ResNet-32. The networks were trained for 200 epochs with stochastic gradient descent (SGD) (momentum=0.9). Following the training strategy in [8, 5], the initial learning rate was set to 0.1 and then decayed by 0.01 at 160 epochs and again at 180 epochs. Furthermore, we used a linear warm-up of the learning rate [12] in the first five epochs. Since our method uses a twophase training schedule, we trained for the first 100 epochs with the standard cross-entropy loss, then fine-tuned the networks using the IB loss for the next 100 epochs. We trained the models for CIFAR on a single NVIDIA GTX 1080Ti with a batch size of 128. For Tiny ImageNet, we employed ResNet-18 and used the stochastic gradient descent with a momentum of 0.9, and weight decay of 2e-4 for training. The networks were initially trained for 50 epochs, and then fine-tuned for the subsequent 50 epochs with IB loss. The learning rate at the start was set to 0.1 and was dropped by a factor of 0.1 after 50 and 90 epochs. For iNaturalist 2018, we trained ResNet-50 with four GTX 1080Ti GPUs. The networks were initially trained for 50 epochs and then fine-tuned for the subsequent 150 epochs with IB loss. The learning rate at the start was set to 0.01 and was decreased by a factor of 0.1 after 30 and 180 epochs.
As a simple but important implementation trick, we added $\epsilon=0.001$ to $\mathcal{IB}(x;w)$ to prevent numerical instability in inversion when the influence approaches zero. We discuss the influence of the hyperparameter $(\epsilon)$ in the following section.
## 4.2. Analysis
To validate the proposed method, we conducted exten-
sive experiments.
Is influence meaningful for re-weighting? First, to confirm whether influence can act as a meaningful clue of re-weighting for class imbalance learning, we compared the influences between a balanced dataset and an imbalanced dataset. For an imbalanced CIFAR-10, we used the long-tailed version of CIFAR-10 with the imbalance ratio $\rho=100$, in which the largest class, ‘plane’(i.e.,class index 0), contains 5,000 samples, while the smallest class, `truck'(i.e., class index 9), contains only 50 samples. We trained ResNet-32 with a standard cross-entropy loss for 200 epochs, as described in Implementation Details, on both the balanced (original) and imbalanced CIFAR-10. We plotted the influences of both classes in Figure 2. We scaled the influences to between 0 and 1 for each dataset. Since the minority class contains only 50 samples, we selected the highest 50 samples for comparison. As illustrated in Figure 2, there was little difference in the distributions of the in-
|
[
[
0.080078125,
0.0889892578125,
0.2091064453125,
0.091552734375,
0.2091064453125,
0.1060791015625,
0.080078125,
0.10418701171875
],
[
0.50244140625,
0.0921630859375,
0.70166015625,
0.0921630859375,
0.70166015625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.521484375,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.121826171875,
0.521484375,
0.121826171875
],
[
0.08087158203125,
0.11871337890625,
0.283447265625,
0.11871337890625,
0.283447265625,
0.1319580078125,
0.08087158203125,
0.1319580078125
],
[
0.50341796875,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50341796875,
0.136962890625
],
[
0.098876953125,
0.143310546875,
0.4697265625,
0.143310546875,
0.4697265625,
0.1572265625,
0.098876953125,
0.1572265625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.07843017578125,
0.1590576171875,
0.468017578125,
0.1578369140625,
0.468017578125,
0.17236328125,
0.07843017578125,
0.173583984375
],
[
0.50341796875,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.080078125,
0.173583984375,
0.468017578125,
0.173583984375,
0.468017578125,
0.1875,
0.080078125,
0.1875
],
[
0.50341796875,
0.17041015625,
0.890625,
0.17041015625,
0.890625,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.5009765625,
0.18310546875,
0.89208984375,
0.184326171875,
0.89208984375,
0.1988525390625,
0.5009765625,
0.1976318359375
],
[
0.080078125,
0.189453125,
0.471435546875,
0.189453125,
0.471435546875,
0.20263671875,
0.080078125,
0.20263671875
],
[
0.080078125,
0.2052001953125,
0.4697265625,
0.2052001953125,
0.4697265625,
0.2183837890625,
0.080078125,
0.2183837890625
],
[
0.50244140625,
0.199462890625,
0.89208984375,
0.199462890625,
0.89208984375,
0.21337890625,
0.50244140625,
0.21337890625
],
[
0.08087158203125,
0.2197265625,
0.468017578125,
0.2183837890625,
0.468017578125,
0.23291015625,
0.08087158203125,
0.2342529296875
],
[
0.50244140625,
0.2139892578125,
0.890625,
0.2139892578125,
0.890625,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.080078125,
0.23486328125,
0.4697265625,
0.23486328125,
0.4697265625,
0.248779296875,
0.080078125,
0.248779296875
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.50244140625,
0.24365234375
],
[
0.07843017578125,
0.25,
0.468017578125,
0.25,
0.468017578125,
0.26318359375,
0.07843017578125,
0.26318359375
],
[
0.50341796875,
0.2442626953125,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.50341796875,
0.25830078125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.26513671875,
0.4697265625,
0.26513671875,
0.4697265625,
0.279052734375,
0.080078125,
0.279052734375
],
[
0.080078125,
0.27978515625,
0.468017578125,
0.27978515625,
0.468017578125,
0.29345703125,
0.080078125,
0.29345703125
],
[
0.50244140625,
0.274658203125,
0.890625,
0.274658203125,
0.890625,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.29541015625,
0.468017578125,
0.29541015625,
0.468017578125,
0.309326171875,
0.080078125,
0.309326171875
],
[
0.08087158203125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.323974609375,
0.08087158203125,
0.323974609375
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.50341796875,
0.3193359375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.080078125,
0.32568359375,
0.4697265625,
0.32568359375,
0.4697265625,
0.339111328125,
0.080078125,
0.339111328125
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.341552734375,
0.468017578125,
0.341552734375,
0.468017578125,
0.354736328125,
0.080078125,
0.354736328125
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.5009765625,
0.349853515625,
0.890625,
0.35107421875,
0.890625,
0.3642578125,
0.5009765625,
0.363037109375
],
[
0.080078125,
0.37060546875,
0.468017578125,
0.37060546875,
0.468017578125,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.50341796875,
0.365478515625,
0.89208984375,
0.365478515625,
0.89208984375,
0.37890625,
0.50341796875,
0.37890625
],
[
0.080078125,
0.386474609375,
0.468017578125,
0.386474609375,
0.468017578125,
0.399658203125,
0.080078125,
0.399658203125
],
[
0.50244140625,
0.380615234375,
0.89208984375,
0.380615234375,
0.89208984375,
0.39453125,
0.50244140625,
0.39453125
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.4140625,
0.080078125,
0.4140625
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.50244140625,
0.410400390625,
0.8935546875,
0.410400390625,
0.8935546875,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.416748046875,
0.4697265625,
0.416748046875,
0.4697265625,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.43115234375,
0.4697265625,
0.43115234375,
0.4697265625,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.50341796875,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50341796875,
0.453857421875
],
[
0.080078125,
0.4462890625,
0.4697265625,
0.4462890625,
0.4697265625,
0.460205078125,
0.080078125,
0.460205078125
],
[
0.080078125,
0.4609375,
0.468017578125,
0.4609375,
0.468017578125,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.50341796875,
0.47021484375,
0.89208984375,
0.47021484375,
0.89208984375,
0.484130859375,
0.50341796875,
0.484130859375
],
[
0.080078125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.0743408203125,
0.4912109375,
0.468017578125,
0.486083984375,
0.468017578125,
0.5068359375,
0.0743408203125,
0.51220703125
],
[
0.50244140625,
0.486083984375,
0.779296875,
0.486083984375,
0.779296875,
0.5,
0.50244140625,
0.5
],
[
0.398681640625,
0.501953125,
0.4599609375,
0.501953125,
0.4599609375,
0.5087890625,
0.398681640625,
0.5087890625
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.52392578125,
0.50390625,
0.890625,
0.50390625,
0.890625,
0.5146484375,
0.52392578125,
0.5146484375
],
[
0.50244140625,
0.5166015625,
0.890625,
0.5166015625,
0.890625,
0.52978515625,
0.50244140625,
0.52978515625
],
[
0.07843017578125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.07843017578125,
0.537109375
],
[
0.50244140625,
0.5322265625,
0.89208984375,
0.5322265625,
0.89208984375,
0.54541015625,
0.50244140625,
0.54541015625
],
[
0.080078125,
0.5390625,
0.439453125,
0.5390625,
0.439453125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.5009765625,
0.546875,
0.890625,
0.546875,
0.890625,
0.56005859375,
0.5009765625,
0.56005859375
],
[
0.09722900390625,
0.556640625,
0.468017578125,
0.556640625,
0.468017578125,
0.57080078125,
0.09722900390625,
0.57080078125
],
[
0.5009765625,
0.56103515625,
0.60693359375,
0.5625,
0.60693359375,
0.5771484375,
0.5009765625,
0.57568359375
],
[
0.080078125,
0.57275390625,
0.4697265625,
0.57275390625,
0.4697265625,
0.5859375,
0.080078125,
0.5859375
],
[
0.07843017578125,
0.5859375,
0.468017578125,
0.5869140625,
0.468017578125,
0.6015625,
0.07843017578125,
0.60107421875
],
[
0.50244140625,
0.5888671875,
0.60107421875,
0.5888671875,
0.60107421875,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.08087158203125,
0.60400390625,
0.468017578125,
0.60400390625,
0.468017578125,
0.6142578125,
0.08087158203125,
0.6142578125
],
[
0.51953125,
0.6123046875,
0.89208984375,
0.61279296875,
0.89208984375,
0.62744140625,
0.51953125,
0.626953125
],
[
0.080078125,
0.61865234375,
0.4306640625,
0.61865234375,
0.4306640625,
0.6318359375,
0.080078125,
0.6318359375
],
[
0.50341796875,
0.630859375,
0.62158203125,
0.630859375,
0.62158203125,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.09722900390625,
0.64013671875,
0.468017578125,
0.6416015625,
0.468017578125,
0.65576171875,
0.09722900390625,
0.65478515625
],
[
0.521484375,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.521484375,
0.65771484375
],
[
0.08087158203125,
0.6591796875,
0.46728515625,
0.6591796875,
0.46728515625,
0.66943359375,
0.08087158203125,
0.66943359375
],
[
0.5009765625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.5009765625,
0.673828125
],
[
0.080078125,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70263671875,
0.4697265625,
0.70263671875,
0.4697265625,
0.71630859375,
0.080078125,
0.71630859375
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73291015625,
0.468017578125,
0.73291015625,
0.468017578125,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.5048828125,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.080078125,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.5009765625,
0.75,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.5009765625,
0.7646484375
],
[
0.080078125,
0.76318359375,
0.46728515625,
0.76318359375,
0.46728515625,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.07843017578125,
0.7763671875,
0.468017578125,
0.77783203125,
0.468017578125,
0.79248046875,
0.07843017578125,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.07843017578125,
0.79345703125,
0.214111328125,
0.79345703125,
0.214111328125,
0.806640625,
0.07843017578125,
0.806640625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.09722900390625,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.09722900390625,
0.82373046875
],
[
0.5009765625,
0.81103515625,
0.8896484375,
0.81103515625,
0.8896484375,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.07843017578125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.080078125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.080078125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88818359375,
0.890625,
0.88818359375,
0.890625,
0.8984375,
0.50341796875,
0.8984375
]
] |
[
"4.Experiments",
"ther improve the performance.",
"Implementation Details. We used PyTorch [29] to im-",
"4.1.ExperimentalSettings",
"plement and train all the models in the paper, and we",
"Datasets. We verified the effectiveness of our method",
"used ResNet architecture [15] for all datasets. For CIFAR",
"on three commonly used benchmark datasets: CIFAR-10,",
"datasets, we used randomly initialized ResNet-32.The net-",
"CIFAR-100 [20], Tiny ImageNet [21], and iNaturalist 2018",
"worksweretrainedfor200epochswithstochasticgradient",
"descent (SGD) (momentum = 0.9). Following the training",
"[16]. The CIFAR-10 and CIFAR-100 datasets consist of",
"50,000 training images and 10,000 test images with 10 and",
"strategy in [8, 5], the initial learning rate was set to 0.1 and",
"100 classes, respectively. Meanwhile, Tiny ImageNet con-",
"then decayed by 0.01 at 160 epochs and again at 180 epochs.",
"tains 200 classes for training, in which each class has 500",
"Furthermore, we used a linear warm-up of the learning rate",
"images. Its test set contains 10,000 images. Since CIFAR",
"[12] in the first five epochs. Since our method uses a two-",
"phase training schedule, we trained for the first 100 epochs",
"and Tiny ImageNet are evenly distributed, we have made",
"these datasets imbalanced according to [8, 4], respectively.",
"with the standard cross-entropy loss, then fine-tuned the net-",
"works using the IB loss for the next 100 epochs. We trained",
"Primarily, we investigate two common types of imbalance:",
"(i) long-tailed imbalance [8] and (ii) step imbalance [4]. In",
"the models for CIFAR on a single NVIDIA GTX 1080Ti",
"with a batch size of 128. For Tiny ImageNet, we employed",
"long-tailed imbalance, the number of training samples for",
"ResNet-18 and used the stochastic gradient descent with a",
"each class decreases exponentially from the largest majority",
"class to the smallest minority class. To construct long-tailed",
"momentum of 0.9,and weight decay of 2e-4 for training.",
"imbalanced datasets, the number of selected samples in the",
"The networks were initially trained for 50 epochs, and then",
"k-th class was set to nkμk(μ E (O, 1)), where nk is the orig-",
"fine-tuned for the subsequent 50 epochs with IB loss. The",
"inal number of the k-th class. Meanwhile, in step imbalance,",
"learning rate at the start was set to 0.1 and was dropped",
"by a factor of 0.1 after 50 and 90 epochs.For iNaturalist",
"the classes are divided into two groups: the majority class",
"2018, we trained ResNet-50 with four GTX 1080Ti GPUs.",
"group and minority class group. Every class within a group",
"The networks were initially trained for 50 epochs and then",
"contains the same number of samples, and the class in the",
"majority class group has many more samples than that in the",
"fine-tuned for the subsequent 150 epochs with IB loss. The",
"learning rate at the start was set to 0.01 and was decreased",
"minority class group. For evaluation, we used the original",
"test set. The imbalance ratio p is defined by p = maxx ma).",
"by a factor of 0.1 after 30 and 180 epochs.",
"minnk",
"Thus, the imbalance ratio represents the degree of imbal-",
"As a simplebut important implementation trick,we",
"added e = 0.001 to IB(c; w) to prevent numerical insta-",
"ance in the dataset. We evaluated the performance of our",
"bility in inversion when the influence approaches zero. We",
"method under various imbalance ratios from 10 to 200.",
"discuss the influence of the hyperparameter (e) in the fol-",
"The iNaturalist 2018 dataset is a large-scale real-world",
"lowing section.",
"dataset containing 437,513 training images and 24,426 test",
"images with 8,142 classes. iNaturalist 2018 exhibits long-",
"4.2. Analysis",
"tailed imbalance,whose imbalance ratiois 500.We used",
"To validate the proposed method, we conducted exten-",
"the official training and test splits in our experiments.",
"sive experiments.",
"Baselines. We compared our algorithm with the follow-",
"Is influence meaningful for re-weighting? First, to",
"ing cost-sensitive loss methods:(1)Our baseline model.",
"confirm whether influence can act as a meaningful clue",
"which is trained on the standard cross-entropy loss. Com-",
"of re-weighting for class imbalance learning, we compared",
"paring our model with this baseline enables us to clearly",
"the influences between a balanced dataset and an imbal-",
"understand how much our training scheme has improved the",
"anced dataset. For an imbalanced CIFAR-10, we used the",
"performance; (2) focal loss [22], which increases the rela-",
"long-tailed version of CIFAR-10 with the imbalance ratio",
"tive loss for hard samples and down-weights well-classified",
"o = 100.in which the largest class. “plane'(i.e.. class in-",
"samples; (3) CB loss [8], which re-weights the loss in-",
"dex 0), contains 5,000 samples, while the smallest class,",
"verselyproportional to the effective number of samples; (4)",
"\"truck' (i.e., class index 9), contains only 50 samples. We",
"LDAM loss [5], which regularizes the minority classes to",
"trained ResNet-32 with a standard cross-entropy loss for",
"have larger margins.",
"200 epochs, as described in Implementation Details, on both",
"Since our IB loss can be easily combined with other",
"the balanced (original) and imbalanced CIFAR-10. We plot-",
"methods, we employee two further variants. First, IB + CB",
"ted the influences of both classes in Figure 2. We scaled the",
"uses the effective number in CB loss, instead of using Ak",
"influences to between O and 1for each dataset.Since the",
"in IB.Second,IB +focal uses focal loss during the fine-",
"minority class contains only 50 samples,we selected the",
"tuning phase, instead of using the cross-entropy loss. We",
"highest 50 samples for comparison. As illustrated in Figure",
"demonstrate that combination with other methods can fur-",
"2.there was little difference in the distributions of the in-"
] |
{
"arxiv_id": "2110.02444",
"language": "en",
"page_idx": 4
}
|
||
2309.17390-en-1
|
2309.17390-en-1.jpg
|
backward flow for moving the 3D points from the current time step back to the canonical time step. Although the canonical-based representation with the backward flow is easy to implement, the backward flow field is non-smooth and discontinuous. As shown in Fig. 1(b), for a fixed 3D position along the timeline, there will be different types of points covering the position p, which needs discontinuous flows to map them back to canonical space (Fig. 1(d)). So the backward flow cannot be fitted well with commonly used smooth motion models (MLP, for example). Also, distortions are introduced to the static canonical space geometry due to the failure of the motion model, as shown in Fig. 8.
To solve the problem of backward flow, we propose using forward flow as the deformation model. Instead of warping the sampled points on image rays at other time steps back to the canonical time and rendering at the canonical space, we propose to warp the whole canonical radiance field from the canonical time to other time steps using forward deformation flow and render at corresponding time steps. In this way, the forward flow estimated by the deformation model will be smooth and continuous for the same 3D position along the timeline (Fig. 1(c) and (e)). Note that SNARF[7] has also used forward warping based on an inverse skinning model, but it is designed for dynamic human modeling and cannot be used in general scenes. In this paper, we aim to achieve forward warping for general scenes, which means we must warp the whole space.
However, introducing forward warping into the canonical space based NeRF methods is not straightforward as there are three main problems to be solved. First, the traditional canonical radiance field in existing methods cannot be warped explicitly, since the radiance field is represented as a continuous function parameterized by an MLP. To solve this problem, we propose to use the voxel grid to represent the canonical radiance field as it is finite and discrete. Recent voxel-based methods[27,44,62] have proven the effectiveness of this representation. The other two problems are the many-to-one and one-to-many mapping issues brought by the inherent property of the forward warping operation. To address them, we propose a differentiable forward warping method consisting of an average splatting operation and an inpaint network to solve the many-to-one and one-tomany issue, respectively. Extensive experiments have been conducted to verify the effectiveness of our method.
Our key contributions can be summarized as follows:
$\bullet$ To the best of our knowledge, we are the first to investigate forward warping in dynamic view synthesis for general scenes. We propose a novel canonical based NeRF with forward flow motion modeling for dynamic view synthesis. Thanks to the forward flow field, our method can better represent the object motions, and explicitly recover the trajectory of a surface point.
$\cdot$ We introduce voxel grid based canonical radiance field to enable reasonable computation of forward warping, and propose a differentiable forward warping method, including an average splatting operation and an inpaint network, to solve the many-to-one and one-to-many issues of forward warping.
$\bullet$Experiments on multiple datasets show that our method outperforms existing methods on the DNeRF[36] dataset, Hypernerf [33] dataset, NHR [54] dataset and our proposed dataset.
## $2. \textbf{Related Works}$
Novel View Synthesis (NVS) is a long-standing task in both computer vision and graphics [4,6,14,18], and surveys of recent methods can be found in [42,46,47]. Some methods require to reconstruct an explicit 3D model to represent the scene, such as point clouds [1], voxels or meshes [16,40,41,48]. Then novel images from arbitrary views can be rendered from this geometry. Another methods try to estimate depth map through multi-view geometry, and then aggregate features from multiple frames through the co-visibility, such as [8,11,17,34,40,41,57]. Recently, neural implicit representations have shown great promise for novel view synthesis and 3D modeling. In this section, we focus more on the neural implicit rendering methods and mainly summarize these schemes according to whether they can handle dynamic scenes.
NVS for Static Scenes NeRF[26], as a seminal work, uses MLPs to model a 5D radiance field, which can render impressive view synthesis for static scenes captured. Numerous subsequent works have extended NeRF to kinds of scenarios, such as larger and unbounded scenes [25,39, 45,56,64], relighting [3,43,58,66], incorporating antialiasing for multi-scale rendering [2], and generalization ability [5,50,52,63]. In addition, some methods are devoted to more efficient neural rendering and optimization in NeRF-like framework, such as [20,22,24,28,35,62] focus more on efficient sampling along each ray for color accumulation, [37, 38] subdivide the scene into multiple cells for efficient processing, and [27,44,61] exploit voxelgrid representation to speed up the optimization of radiance field. However, these methods are mainly applicable to static scenes, leaving out the scenes with dynamic objects, which are actually more extensive and practical.
NVS for Dynamic Scenes There are several works that extend NeRF from static scenes to dynamic scenes with nonrigid deformable objects. One feasible way is to build a 4D spatial-temporal representation. For example, Yoon $et$ $al.[60]$ combine single-view depth and depth from multiview stereo to render virtual views with 3D warping. Gao et al.[13] use a time-invariant model (static) and a time-
varying model (dynamic) to represent the scenes, and reg-
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.521484375,
0.09027099609375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.521484375,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.53369140625,
0.106689453125,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.53369140625,
0.12060546875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.53515625,
0.12371826171875,
0.890625,
0.12371826171875,
0.890625,
0.136962890625,
0.53515625,
0.136962890625
],
[
0.07843017578125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.07843017578125,
0.15087890625
],
[
0.53369140625,
0.136962890625,
0.89208984375,
0.136962890625,
0.89208984375,
0.15087890625,
0.53369140625,
0.15087890625
],
[
0.07843017578125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.07843017578125,
0.1666259765625
],
[
0.5361328125,
0.1546630859375,
0.890625,
0.1546630859375,
0.890625,
0.1654052734375,
0.5361328125,
0.1654052734375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.53515625,
0.1685791015625,
0.69921875,
0.1685791015625,
0.69921875,
0.1817626953125,
0.53515625,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.521484375,
0.1817626953125,
0.8935546875,
0.1817626953125,
0.8935546875,
0.1956787109375,
0.521484375,
0.1956787109375
],
[
0.08087158203125,
0.199462890625,
0.465576171875,
0.199462890625,
0.465576171875,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.53515625,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.53515625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.53515625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.53515625,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.53515625,
0.2291259765625,
0.75341796875,
0.2291259765625,
0.75341796875,
0.242431640625,
0.53515625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.465576171875,
0.24365234375,
0.465576171875,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.5009765625,
0.2587890625,
0.650390625,
0.260009765625,
0.650390625,
0.27783203125,
0.5009765625,
0.276611328125
],
[
0.08087158203125,
0.274658203125,
0.125,
0.274658203125,
0.125,
0.285888671875,
0.08087158203125,
0.285888671875
],
[
0.09722900390625,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.3037109375,
0.09722900390625,
0.3037109375
],
[
0.521484375,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.521484375,
0.302978515625
],
[
0.080078125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.5,
0.302978515625,
0.8935546875,
0.3037109375,
0.8935546875,
0.318115234375,
0.5,
0.317626953125
],
[
0.07843017578125,
0.32080078125,
0.471435546875,
0.32080078125,
0.471435546875,
0.333984375,
0.07843017578125,
0.333984375
],
[
0.50244140625,
0.3193359375,
0.89208984375,
0.3193359375,
0.89208984375,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.08087158203125,
0.337158203125,
0.46728515625,
0.337158203125,
0.46728515625,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.50244140625,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.080078125,
0.35107421875,
0.4697265625,
0.35107421875,
0.4697265625,
0.3642578125,
0.080078125,
0.3642578125
],
[
0.50244140625,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.50244140625,
0.365478515625,
0.89208984375,
0.365478515625,
0.89208984375,
0.37890625,
0.50244140625,
0.37890625
],
[
0.08087158203125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.5009765625,
0.37890625,
0.890625,
0.37890625,
0.890625,
0.392578125,
0.5009765625,
0.392578125
],
[
0.08087158203125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.50244140625,
0.395751953125,
0.89208984375,
0.395751953125,
0.89208984375,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.410400390625,
0.89208984375,
0.410400390625,
0.89208984375,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50244140625,
0.425537109375,
0.89208984375,
0.425537109375,
0.89208984375,
0.439453125,
0.50244140625,
0.439453125
],
[
0.080078125,
0.44140625,
0.4697265625,
0.44140625,
0.4697265625,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.08087158203125,
0.458251953125,
0.46728515625,
0.458251953125,
0.46728515625,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.50244140625,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.07843017578125,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.48486328125,
0.07843017578125,
0.48486328125
],
[
0.50341796875,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.48291015625,
0.50341796875,
0.48291015625
],
[
0.07843017578125,
0.4873046875,
0.4697265625,
0.4873046875,
0.4697265625,
0.50048828125,
0.07843017578125,
0.50048828125
],
[
0.50244140625,
0.486083984375,
0.890625,
0.486083984375,
0.890625,
0.5,
0.50244140625,
0.5
],
[
0.080078125,
0.501953125,
0.374267578125,
0.501953125,
0.374267578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.501953125,
0.68701171875,
0.501953125,
0.68701171875,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.098876953125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.5322265625,
0.098876953125,
0.5322265625
],
[
0.5009765625,
0.51953125,
0.8896484375,
0.51953125,
0.8896484375,
0.53271484375,
0.5009765625,
0.53271484375
],
[
0.080078125,
0.5341796875,
0.4697265625,
0.5341796875,
0.4697265625,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50341796875,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50341796875,
0.548828125
],
[
0.08087158203125,
0.55029296875,
0.46728515625,
0.55029296875,
0.46728515625,
0.56103515625,
0.08087158203125,
0.56103515625
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.08087158203125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.57568359375,
0.08087158203125,
0.57568359375
],
[
0.50244140625,
0.56494140625,
0.89208984375,
0.56494140625,
0.89208984375,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.5791015625,
0.4697265625,
0.5791015625,
0.4697265625,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.5009765625,
0.580078125,
0.8896484375,
0.580078125,
0.8896484375,
0.59326171875,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5947265625,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.50244140625,
0.60791015625
],
[
0.07843017578125,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62255859375,
0.07843017578125,
0.62255859375
],
[
0.50341796875,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.50341796875,
0.625
],
[
0.080078125,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.6396484375,
0.4697265625,
0.6396484375,
0.4697265625,
0.65283203125,
0.080078125,
0.65283203125
],
[
0.50341796875,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.080078125,
0.65576171875,
0.468017578125,
0.65576171875,
0.468017578125,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.50244140625,
0.65576171875,
0.89208984375,
0.65576171875,
0.89208984375,
0.66943359375,
0.50244140625,
0.66943359375
],
[
0.080078125,
0.66943359375,
0.468017578125,
0.66943359375,
0.468017578125,
0.68310546875,
0.080078125,
0.68310546875
],
[
0.50341796875,
0.67236328125,
0.890625,
0.67236328125,
0.890625,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.080078125,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.50341796875,
0.6875,
0.890625,
0.6875,
0.890625,
0.69775390625,
0.50341796875,
0.69775390625
],
[
0.08087158203125,
0.70068359375,
0.468017578125,
0.70068359375,
0.468017578125,
0.7138671875,
0.08087158203125,
0.7138671875
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.080078125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.50341796875,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50341796875,
0.72998046875
],
[
0.080078125,
0.72998046875,
0.4697265625,
0.72998046875,
0.4697265625,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.50244140625,
0.73193359375,
0.890625,
0.73193359375,
0.890625,
0.74560546875,
0.50244140625,
0.74560546875
],
[
0.08087158203125,
0.7470703125,
0.46728515625,
0.7470703125,
0.46728515625,
0.7568359375,
0.08087158203125,
0.7568359375
],
[
0.5009765625,
0.74560546875,
0.89208984375,
0.7470703125,
0.89208984375,
0.76123046875,
0.5009765625,
0.76025390625
],
[
0.080078125,
0.76025390625,
0.42236328125,
0.76025390625,
0.42236328125,
0.7734375,
0.080078125,
0.7734375
],
[
0.50341796875,
0.76220703125,
0.81884765625,
0.76220703125,
0.81884765625,
0.77587890625,
0.50341796875,
0.77587890625
],
[
0.098876953125,
0.77587890625,
0.44775390625,
0.77587890625,
0.44775390625,
0.7890625,
0.098876953125,
0.7890625
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.098876953125,
0.79443359375,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.098876953125,
0.8095703125
],
[
0.50341796875,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.111083984375,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.111083984375,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.11273193359375,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.11273193359375,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.111083984375,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.111083984375,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.11273193359375,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.11273193359375,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.1094970703125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.1094970703125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.111083984375,
0.88720703125,
0.441162109375,
0.88720703125,
0.441162109375,
0.90087890625,
0.111083984375,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"backward fow for moving the 3D points from the current",
":We introduce voxel grid based canonical radiance field",
"time stepbackto the canonicaltime step.Althoughthe",
"to enable reasonable computation of forward warping,",
"canonical-based representation with the backward flow is",
"and propose a differentiable forward warping method,",
"easy to implement, the backward flow field is non-smooth",
"including an average splatting operation and an inpaint",
"and discontinuous. As shown in Fig. 1(b), for a fixed 3D",
"network,tosolvethemany-to-oneand one-to-manyis-",
"position along the timeline, there will be different types",
"sues of forward warping.",
"of points covering the position p, which needs discontin-",
".Experiments on multiple datasets show that our",
"uousflowstomapthembacktocanonical space(Fig.l(d))",
"method outperforms existing methods on the D-",
"So the backward flow cannot be fitted well with commonly",
"NeRF [36] dataset, Hypernerf [33] dataset, NHR [54]",
"used smooth motion models (MLP, for example). Also, dis-",
"dataset and our proposed dataset.",
"tortions are introduced to the static canonical space geom-",
"etry due to the failure of the motion model, as shown in",
"2.Related Works",
"Fig. 8.",
"To solve the problem of backward flow,we propose us-",
"Novel View Synthesis (NVS) is a long-standing task in",
"ing forward flow as the deformation model. Instead of warp-",
"both computer vision and graphics [4, 6, 14, 18], and sur-",
"ing the sampled points on image rays at other time steps",
"veys of recent methods can be found in [42, 46, 47]. Some",
"back to the canonical time and rendering atthe canonical",
"methods require to reconstruct an explicit 3D model to",
"space, we propose to warp the whole canonical radiance",
"represent the scene, such as point clouds [1l, voxels or",
"field from the canonical time to other time steps using for-",
"meshes [16, 40,41, 48]. Then novel images from arbitrary",
"ward deformation flow and render at corresponding time",
"views can be rendered from this geometry.Another meth-",
"steps.Inthisway.theforward flow estimatedby the defor-",
"ods try to estimate depth map through multi-view geometry,",
"mation model will be smooth and continuous for the same",
"and then aggregate features from multiple frames through",
"3D position along the timeline (Fig. I(c) and (e)). Note that",
"the co-visibility, such as [8, 11,17,34,40,41,57]. Recently,",
"SNARF [7] has also used forward warping based on an in-",
"neural implicit representations have shown great promise",
"verseskinningmodel,butitisdesignedfordynamichuman",
"for novel view synthesis and 3D modeling. In this section",
"modeling and cannot be used in general scenes. In this pa-",
"wefocusmoreontheneuralimplicitrenderingmethodsand",
"per, we aim to achieve forward warping for general scenes,",
"mainly summarize these schemes according to whether they",
"which means we must warp the whole space.",
"can handle dynamic scenes.",
"However, introducing forward warping into the canon-",
"NVS for Static Scenes NeRF [26], as a seminal work,",
"ical space based NeRF methods is not straightforward as",
"uses MLPs to model a 5D radiance field, which can ren-",
"there are three main problems to be solved. First, the tra-",
"der impressive view synthesis for static scenes captured.",
"ditional canonical radiancefield in existing methods cannot",
"Numerous subsequent works have extended NeRF to kinds",
"be warped explicitly, since the radiance field is represented",
"of scenarios, such as larger and unbounded scenes [25,39.",
"as a continuous function parameterized by an MLP. To solve",
"45.56,64],relighting[3,43,58,66],incorporating anti-",
"this problem, we propose to use the voxel grid to represent",
"aliasing for multi-scale rendering [2], and generalization",
"the canonical radiance field as it is finite and discrete. Re-",
"ability [5, 50, 52,63]. In addition, some methods are de-",
"cent voxel-based methods [27,44,62] have proven the effec-",
"votedtomoreefficientneuralrenderingandoptimizationin",
"tiveness of this representation.The other two problems are",
"NeRF-like framework, such as [20,22,24, 28, 35, 62] fo-",
"the many-to-one and one-to-many mapping issues brought",
"cus more on efficient sampling along each ray for color",
"by the inherent property of the forward warping operation.",
"accumulation,37,38lsubdivide the sceneintomultiple",
"To address them, we propose a differentiable forward warp-",
"cells for efficient processing, and [27,44,61] exploit voxel-",
"ing method consisting of an average splatting operation and",
"grid representation to speed up the optimization of radiance",
"an inpaint network to solve the many-to-one and one-to-",
"field.However, these methods are mainly applicable to",
"manyissue,respectively.Extensiveexperimentshavebeen",
"static scenes, leaving out the scenes with dynamic objects,",
"conducted to verify the effectiveness of our method.",
"which are actually more extensive and practical.",
"Our key contributions can be summarized as follows:",
"NVS for Dynamic Scenes There are several works that ex-",
". To the best of our knowledge, we are the first to inves-",
"tendNeRF from static scenesto dynamic scenes with non-",
"tigate forward warping in dynamic view synthesis for",
"rigid deformable objects. One feasible way is to build a",
"general scenes. We propose a novel canonical based",
"4D spatial-temporal representation. For example, Yoon et",
"NeRF with forward flow motion modeling for dynamic",
"al. [60] combine single-view depth and depth from multi-",
"view synthesis. Thanks to the forward flow field, our",
"view stereo to render virtual views with 3D warping. Gao",
"method can better represent the object motions, and",
"et al. [13] use a time-invariant model (static) and a time-",
"explicitly recover the trajectory of a surface point.",
"varying model (dynamic) to represent the scenes, and reg-"
] |
{
"arxiv_id": "2309.17390",
"language": "en",
"page_idx": 1
}
|
||
2309.17390-en-4
|
2309.17390-en-4.jpg
|
<FigureHere>
Figure 4. Inpaint network. It consists of a 3D U-Net structure and an up-sampling layer, which could inpaint and upsample the input voxel grid.
warping. To resolve this issue, we propose building an inpaint network to fill these holes. As shown in Fig.4, we modify a 3D U-Net structure network [53] followed by an $\begin{array}{l}{\mathrm{up-sampling~layer,~which~is~V_{R_{\mathrm{Inp}}}^{t}=F_{\theta_{3}}(\mathbf{V_{R}^{t}}),~where~}}\\{\mathrm{is~the~inpainted~and~up-sampled~voxel~grid.}}\end{array}\dot{\mathbf{V_{R_{\mathrm{Inp}}}^{t}}}$
This structure helps encoding layers to learn information from local neighboring voxels and decoding layers to recover the original resolution with filled content. We only feed the warped voxel grid to the inpaint network, which has no access to temporal information. Thus, the inpaint network can not learn to warp, limiting its ability only to inpainting. Since warping operations and the inpaint network are relatively expensive to compute, we propose to attach an up-sample layer after the U-Net structure. In this way, we could warp the grids with a low resolution to save time and memory, while rendering with a higher resolution for better image quality. The up-sample layer consists of a interpolation layer, convolution layers, and activation layers.
# 3.3. Volume Rendering
After we get the radiance voxel grid at time $t$,the pixel colors of the image rays can be rendered using volume rendering techniques[26]. Given a ray r$(w)=\mathbf{o}+w\mathbf{d}$ emitted from the camera center o with view direction d through a given pixel on the image plane, we render the correspond- $\operatorname{ing}$pixel color $\mathbf{C} _{\mathrm{Inp}}( \mathbf{r} ) = F_{\mathrm{render}}( \mathbf{V} _{\mathrm{R} _{\mathrm{Inp}}}^{t}, \mathbf{r} ) .$To do this,we get all 3D points p that the ray intersects with voxel grids. Next, trilinear interpolation is applied to obtain the density and color feature of each 3D point.
(8)
$$(\sigma,\mathbf{c_f})=F_{\mathrm{inter}}(\mathbf{V_{R_{Inp}}^t},\mathbf{p}).$$
Then, we get the color of this 3D point p by
$$\mathbf{c}=\begin{cases}\mathbf{c_f}&\mathrm{if}\:\mathbf{c_f}\in\mathbb{R}^3,\\F_{\theta_4}(\mathbf{c_f},\mathbf{d})&\mathrm{otherwise},\end{cases}$$
(9)
where $F_{\theta_{4}}$ is a small MLP network that is used to produce
colors that depend on the direction of the ray.
Finally, the pixel color could be rendered by
(10)
$$\mathbf{C}(\mathbf{r})=\sum_{k=1}^{K}T(w_{k})\:\alpha\:(\sigma(w_{k})\delta_{k})\:\mathbf{c}(w_{k})\:,\\T(w_{k})=\exp\left(-\sum_{j=1}^{k-1}\sigma(w_{j})\delta_{j}\right)\:,$$
(11)
where $K$ is the number of sampled points along the ray, $\delta_k$ is the distance between adjacent samples along the ray, and $\alpha\left(\sigma(w_{k})\delta_{k}\right)=1-\exp(-\sigma(w_{k})\delta_{k}).$
Note that the warped radiance grid $\mathbf{V}_{\mathrm{R}}^{t}$ should also render reasonable images. So we directly up-sample $\mathbf{V}_{\mathrm{R}}^{t}$ to get $\mathbf{V}_{\mathrm{R_{Up}}}^{t}$without inpainting, and render the corresponding pixel $\operatorname { color} \mathbf{C} _{\mathrm{Up}}( \mathbf{r} ) = F_{\mathrm{render}}( \mathbf{V} _{\mathrm{R} _{\mathrm{Up}}}^{t}, \mathbf{r} ) .$
## 3.4. Model Optimization
To optimize the model, we design a series of loss functions. Photometric Loss First, the most important loss is the photometric loss, which is the mean square error (MSE) of the rendered color and ground truth color $\mathbf{C} _{\mathrm{gt}}( \mathbf{r} ) .$
(12)
$$\begin{aligned}\mathcal{L}^{\mathrm{photo}}=&\frac{1}{|\mathcal{R}|}\sum_{r\in\mathcal{R}}\left\|\mathbf{C}_{\mathrm{Inp}}(\mathbf{r})-\mathbf{C}_{\mathrm{gt}}(\mathbf{r})\right\|_{2}^{2}\\&+\frac{1}{|\mathcal{R}|}\sum_{r\in\mathcal{R}}\left\|\mathbf{C}_{\mathrm{Up}}(\mathbf{r})-\mathbf{C}_{\mathrm{gt}}(\mathbf{r})\right\|_{2}^{2},\end{aligned}$$
where $\mathcal{R}$ is the set of rays sampled in one batch.
Following DVGO[44], we use $\mathcal{L}^\mathrm{ptc}$ to directly supervise the color of sampled points. The intuition is that sampled points with larger weights contribute more to the rendered color. Also, we use the background entropy loss $\mathcal{L}^\mathrm{bg}$ to encourage the densities concentrating on either the foreground or the background.
Inpaint Network Loss As the inpaint network is used to complete the warped voxel grids, the output of the inpaint $\hat{\text{network should be close to the input. Therefore, we propose}}$ to use the L1 norm of the difference between the input and output of the inpaint network.
$$\mathcal{L}^{\mathrm{vdiff}}=\left\|\mathbf{V}_{\mathrm{R_{Inp}}}^t-\mathbf{V}_{\mathrm{R_{Up}}}^t\right\|_1.$$
(13)
Regularization Terms Since modeling the appearance and motion of a dynamic scene from monocular images is a nontrivial problem, we propose a series of regularization terms.
First, we encourage most 3D points in canonical space to be static by introducing $\mathcal{L}^\mathrm{flow}$ to be the L1 norm of $\mathbf{V}_\mathrm{flow}^t.$ Second, denoting $\mathcal{L}^\mathrm{tv}$ as the total variation function. We compute $\mathcal{L}^\mathrm{tv}(\tilde{\mathbf{V}_{\sigma}^\mathrm{Can}})$ to ensure the spatial smoothness of the $\hat{\text{density in canonical space, and introduce }\mathcal{L}}^{\mathrm{v}}(\mathbf{V}_{\mathrm{flow}}^{t})$toencourage the spatial smoothness of the motions. Finally, inspired by RegNeRF[29], we render depths D of image patches sampled from random views and minimize their total variations $\mathcal{L} ^\mathrm{tv}( \mathbf{D} ) .$
|
[
[
0.52392578125,
0.09027099609375,
0.810546875,
0.09027099609375,
0.810546875,
0.10479736328125,
0.52392578125,
0.10479736328125
],
[
0.5400390625,
0.125,
0.59326171875,
0.125,
0.59326171875,
0.1363525390625,
0.5400390625,
0.1363525390625
],
[
0.63818359375,
0.12371826171875,
0.8310546875,
0.12371826171875,
0.8310546875,
0.1383056640625,
0.63818359375,
0.1383056640625
],
[
0.859375,
0.125,
0.890625,
0.125,
0.890625,
0.136962890625,
0.859375,
0.136962890625
],
[
0.7236328125,
0.1514892578125,
0.73388671875,
0.1514892578125,
0.73388671875,
0.1546630859375,
0.7236328125,
0.1546630859375
],
[
0.72802734375,
0.1514892578125,
0.8037109375,
0.1514892578125,
0.8037109375,
0.17041015625,
0.72802734375,
0.17041015625
],
[
0.861328125,
0.1558837890625,
0.890625,
0.1558837890625,
0.890625,
0.1685791015625,
0.861328125,
0.1685791015625
],
[
0.50244140625,
0.1875,
0.88818359375,
0.189453125,
0.88818359375,
0.20263671875,
0.50244140625,
0.2008056640625
],
[
0.57177734375,
0.15283203125,
0.6640625,
0.15283203125,
0.6640625,
0.17041015625,
0.57177734375,
0.17041015625
],
[
0.11273193359375,
0.1932373046875,
0.136474609375,
0.1932373046875,
0.136474609375,
0.2108154296875,
0.11273193359375,
0.2108154296875
],
[
0.41015625,
0.1956787109375,
0.438720703125,
0.205810546875,
0.43310546875,
0.21533203125,
0.40283203125,
0.2052001953125
],
[
0.50244140625,
0.20263671875,
0.89208984375,
0.203857421875,
0.89208984375,
0.2183837890625,
0.50244140625,
0.2171630859375
],
[
0.5048828125,
0.2203369140625,
0.74853515625,
0.2203369140625,
0.74853515625,
0.2310791015625,
0.5048828125,
0.2310791015625
],
[
0.08087158203125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.08087158203125,
0.2412109375
],
[
0.521484375,
0.2342529296875,
0.890625,
0.2342529296875,
0.890625,
0.2474365234375,
0.521484375,
0.2474365234375
],
[
0.08087158203125,
0.242431640625,
0.468017578125,
0.242431640625,
0.468017578125,
0.255615234375,
0.08087158203125,
0.255615234375
],
[
0.50634765625,
0.250732421875,
0.890625,
0.250732421875,
0.890625,
0.261474609375,
0.50634765625,
0.261474609375
],
[
0.08087158203125,
0.2568359375,
0.176513671875,
0.2568359375,
0.176513671875,
0.268310546875,
0.08087158203125,
0.268310546875
],
[
0.50341796875,
0.264404296875,
0.89208984375,
0.264404296875,
0.89208984375,
0.279052734375,
0.50341796875,
0.279052734375
],
[
0.07843017578125,
0.2822265625,
0.46728515625,
0.281005859375,
0.46728515625,
0.29541015625,
0.07843017578125,
0.296630859375
],
[
0.5,
0.279052734375,
0.71337890625,
0.27978515625,
0.71337890625,
0.29736328125,
0.5,
0.296630859375
],
[
0.080078125,
0.29736328125,
0.468017578125,
0.29736328125,
0.468017578125,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.50244140625,
0.30615234375,
0.69140625,
0.306884765625,
0.69140625,
0.3212890625,
0.50244140625,
0.32080078125
],
[
0.080078125,
0.3125,
0.468017578125,
0.3125,
0.468017578125,
0.32568359375,
0.080078125,
0.32568359375
],
[
0.07843017578125,
0.323974609375,
0.461669921875,
0.326904296875,
0.461669921875,
0.3447265625,
0.07843017578125,
0.341552734375
],
[
0.50244140625,
0.330078125,
0.890625,
0.330078125,
0.890625,
0.3447265625,
0.50244140625,
0.3447265625
],
[
0.080078125,
0.34228515625,
0.361083984375,
0.34228515625,
0.361083984375,
0.35595703125,
0.080078125,
0.35595703125
],
[
0.50341796875,
0.34521484375,
0.89208984375,
0.34521484375,
0.89208984375,
0.359130859375,
0.50341796875,
0.359130859375
],
[
0.098876953125,
0.35986328125,
0.468017578125,
0.35986328125,
0.468017578125,
0.373779296875,
0.098876953125,
0.373779296875
],
[
0.50244140625,
0.363037109375,
0.8935546875,
0.363037109375,
0.8935546875,
0.376953125,
0.50244140625,
0.376953125
],
[
0.080078125,
0.375732421875,
0.468017578125,
0.375732421875,
0.468017578125,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.5,
0.375,
0.82666015625,
0.376953125,
0.82666015625,
0.39453125,
0.5,
0.392578125
],
[
0.080078125,
0.39013671875,
0.468017578125,
0.39013671875,
0.468017578125,
0.404052734375,
0.080078125,
0.404052734375
],
[
0.080078125,
0.406005859375,
0.4697265625,
0.406005859375,
0.4697265625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.5595703125,
0.4091796875,
0.59716796875,
0.406005859375,
0.5986328125,
0.41845703125,
0.560546875,
0.421630859375
],
[
0.650390625,
0.415283203125,
0.66259765625,
0.415283203125,
0.66259765625,
0.419189453125,
0.650390625,
0.419189453125
],
[
0.66259765625,
0.410400390625,
0.810546875,
0.410400390625,
0.810546875,
0.427978515625,
0.66259765625,
0.427978515625
],
[
0.07843017578125,
0.42041015625,
0.4697265625,
0.419189453125,
0.4697265625,
0.434326171875,
0.07843017578125,
0.43505859375
],
[
0.61669921875,
0.419189453125,
0.64208984375,
0.419189453125,
0.64208984375,
0.43310546875,
0.61669921875,
0.43310546875
],
[
0.6396484375,
0.4267578125,
0.6708984375,
0.4267578125,
0.6708984375,
0.439453125,
0.6396484375,
0.439453125
],
[
0.080078125,
0.436279296875,
0.465576171875,
0.436279296875,
0.465576171875,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.861328125,
0.434326171875,
0.89208984375,
0.434326171875,
0.89208984375,
0.4462890625,
0.861328125,
0.4462890625
],
[
0.62109375,
0.4443359375,
0.8388671875,
0.434326171875,
0.84228515625,
0.4697265625,
0.62353515625,
0.479736328125
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.6689453125,
0.468505859375,
0.689453125,
0.468505859375,
0.689453125,
0.4765625,
0.6689453125,
0.4765625
],
[
0.080078125,
0.48095703125,
0.4697265625,
0.48095703125,
0.4697265625,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.50341796875,
0.488037109375,
0.81787109375,
0.488037109375,
0.81787109375,
0.501953125,
0.50341796875,
0.501953125
],
[
0.080078125,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.521484375,
0.50390625,
0.890625,
0.50390625,
0.890625,
0.517578125,
0.521484375,
0.517578125
],
[
0.080078125,
0.51123046875,
0.4697265625,
0.51123046875,
0.4697265625,
0.5244140625,
0.080078125,
0.5244140625
],
[
0.080078125,
0.5263671875,
0.46728515625,
0.5263671875,
0.46728515625,
0.54052734375,
0.080078125,
0.54052734375
],
[
0.5048828125,
0.52099609375,
0.8896484375,
0.52099609375,
0.8896484375,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.50341796875,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.080078125,
0.54150390625,
0.418212890625,
0.54150390625,
0.418212890625,
0.5546875,
0.080078125,
0.5546875
],
[
0.50341796875,
0.548828125,
0.88818359375,
0.548828125,
0.88818359375,
0.5625,
0.50341796875,
0.5625
],
[
0.50244140625,
0.564453125,
0.89208984375,
0.564453125,
0.89208984375,
0.57763671875,
0.50244140625,
0.57763671875
],
[
0.07843017578125,
0.5712890625,
0.256591796875,
0.57373046875,
0.256591796875,
0.58837890625,
0.07843017578125,
0.5859375
],
[
0.50244140625,
0.580078125,
0.62841796875,
0.580078125,
0.62841796875,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.1004638671875,
0.5986328125,
0.4697265625,
0.5986328125,
0.4697265625,
0.6123046875,
0.1004638671875,
0.6123046875
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.080078125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.07843017578125,
0.62890625,
0.4697265625,
0.62744140625,
0.4697265625,
0.642578125,
0.07843017578125,
0.64306640625
],
[
0.5009765625,
0.62548828125,
0.89208984375,
0.626953125,
0.89208984375,
0.6416015625,
0.5009765625,
0.64013671875
],
[
0.08087158203125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.65576171875,
0.08087158203125,
0.65576171875
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.08087158203125,
0.6591796875,
0.465576171875,
0.6591796875,
0.465576171875,
0.67236328125,
0.08087158203125,
0.67236328125
],
[
0.50244140625,
0.65771484375,
0.701171875,
0.65771484375,
0.701171875,
0.671875,
0.50244140625,
0.671875
],
[
0.0767822265625,
0.67236328125,
0.4697265625,
0.671875,
0.4697265625,
0.689453125,
0.0767822265625,
0.68994140625
],
[
0.609375,
0.6845703125,
0.66015625,
0.68115234375,
0.6611328125,
0.6962890625,
0.611328125,
0.69970703125
],
[
0.70166015625,
0.68603515625,
0.76123046875,
0.68603515625,
0.76123046875,
0.7021484375,
0.70166015625,
0.7021484375
],
[
0.859375,
0.68603515625,
0.8935546875,
0.68603515625,
0.8935546875,
0.70263671875,
0.859375,
0.70263671875
],
[
0.68115234375,
0.6923828125,
0.7060546875,
0.6923828125,
0.7060546875,
0.69580078125,
0.68115234375,
0.69580078125
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.07843017578125,
0.68798828125,
0.46728515625,
0.689453125,
0.46728515625,
0.7041015625,
0.07843017578125,
0.70263671875
],
[
0.07843017578125,
0.71826171875,
0.31201171875,
0.7197265625,
0.31201171875,
0.734375,
0.07843017578125,
0.73291015625
],
[
0.50244140625,
0.71826171875,
0.89208984375,
0.7197265625,
0.89208984375,
0.734375,
0.50244140625,
0.73291015625
],
[
0.5048828125,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.188720703125,
0.75048828125,
0.358642578125,
0.75048828125,
0.358642578125,
0.76806640625,
0.188720703125,
0.76806640625
],
[
0.44677734375,
0.75244140625,
0.4697265625,
0.75244140625,
0.4697265625,
0.7666015625,
0.44677734375,
0.7666015625
],
[
0.5009765625,
0.75,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.5009765625,
0.7646484375
],
[
0.52294921875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.52294921875,
0.77978515625
],
[
0.5,
0.7763671875,
0.890625,
0.77978515625,
0.890625,
0.79736328125,
0.5,
0.79443359375
],
[
0.08087158203125,
0.78515625,
0.369384765625,
0.78515625,
0.369384765625,
0.798828125,
0.08087158203125,
0.798828125
],
[
0.50244140625,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.50341796875,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.2017822265625,
0.8193359375,
0.2254638671875,
0.8193359375,
0.2254638671875,
0.83349609375,
0.2017822265625,
0.83349609375
],
[
0.28759765625,
0.818359375,
0.365234375,
0.818359375,
0.365234375,
0.83349609375,
0.28759765625,
0.83349609375
],
[
0.44677734375,
0.8291015625,
0.4697265625,
0.8291015625,
0.4697265625,
0.84228515625,
0.44677734375,
0.84228515625
],
[
0.50244140625,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84130859375,
0.50244140625,
0.84130859375
],
[
0.2034912109375,
0.8369140625,
0.366943359375,
0.83544921875,
0.366943359375,
0.8515625,
0.2034912109375,
0.85400390625
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50244140625,
0.869140625
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.378173828125,
0.88720703125,
0.378173828125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.64453125,
0.88623046875,
0.64453125,
0.8994140625,
0.50244140625,
0.8994140625
]
] |
[
"Finally, the pixel color could be rendered by",
"C(r) =",
"T(wk)α (o(wk)ok) c(wk),",
"(10)",
"",
", o(wj)5j",
"(11)",
"where K is the number of sampled points along the ray, ",
"T(wk) = exp",
"VR",
"VRIn",
"is the distance between adjacent samples along the ray, and",
"α(a(wh)o)=1-exp(-a(w)o)",
"Figure 4. Inpaint network. It consists of a 3D U-Net structure",
"Note that the warped radiance grid V should also ren-",
"and an up-sampling layer, which could inpaint and upsample the",
"derreasonableimages.Sowedirectlyup-sampleVttoget",
"input voxel grid",
"Vku,without inpainting, and render the corresponding pixel",
"warping.To resolve this issue, we propose building an in-",
"color Cup(r)= Frender(VRup,r).",
"paint network to fill these holes. As shown in Fig. 4, we",
"3.4. Model Optimization",
"modify a 3D U-Net structure network [53] followed by an",
"up-sampling layer, which is VR.. - Foa (VR), where VR",
"To optimize the model, we design a series of loss functions.",
"is the inpainted and up-sampled voxel grid.",
"Photometric Loss First, the most important loss is the",
"This structure helps encoding layers to learn information",
"photometric loss, which is the mean square error (MSE) of",
"from local neighboring voxels and decoding layers to re-",
"the rendered color and ground truth color Cgt(r).",
"cover the original resolution with filled content. We only",
"feed the warped voxel grid to the inpaint network, which",
"Cphoto",
"",
"C ICimp(r) -Cgt(r)/2",
"has no access to temporal information. Thus, the inpaint",
"[R]",
"TER",
"network can not learn to warp, limiting its ability only to in-",
"(12)",
"+Cup(t) Ca()/,",
"painting.Since warping operations and the inpaint network",
"are relatively expensive to compute, we propose to attach an",
"ER",
"up-sample layer after the U-Net structure. In this way, we",
"where R is the set of rays sampled in one batch.",
"could warp the grids with a low resolution to save time and",
"Following DVGO [44], we use Cptc to directly supervise",
"memory, while rendering with a higher resolution for better",
"image quality. The up-sample layer consists of a interpola-",
"the colorof sampled points.Theintuitionis thatsampled",
"points with larger weights contribute more to the rendered",
"tion layer, convolution layers, and activation layers.",
"color. Also, we use the background entropy loss Cbg to en-",
"courage the densities concentrating on either the foreground",
"3.3.VolumeRendering",
"or the background.",
"After we get the radiance voxel grid at time t, the pixel",
"Inpaint Network LossAs the inpaint network is used to",
"colors of the image rays can be rendered using volume ren-",
"complete the warped voxel grids, the output of the inpaint",
"dering techniques [26]. Given a ray r(w) = o+ wd emitted",
"network should be close to the input. Therefore, we propose",
"from the camera center o with view direction d through a",
"to use the L1 norm of the difference between the input and",
"given pixel on the image plane, we render the correspond-",
"output of the inpaint network.",
"ing pixel color Cimp(r) = Frender(Vkim,r). To do this, we",
"Cvdiff",
"-VRU",
"(13)",
"",
"Next, trilinear interpolation is applied to obtain the density",
"get all 3D points p that the ray intersects with voxel grids.",
"and color feature of each 3D point.",
"Regularization Terms Since modeling the appearance and",
"motionofadynamicscenefrommonocularimagesisanon-",
"(o, Cr) = Finter(VRmp, P).",
"(8)",
"trivial problem, we propose a series of regularization terms.",
"First, we encourage most 3D points in canonical space",
"to be static by introducing Cfow to be the LI norm of Vtow",
"Then, we get the color of this 3D point p by",
"Second, denoting Ctv as the total variation function. We",
"compute Ctv(VCan) to ensure the spatial smoothness of the",
"[cf",
"if c ER3,",
"(9)",
"density in canonical space, and introduce Ctv(Vtow) to en-",
"(Fo(cr,d)otherwise,",
"courage the spatial smoothness of the motions.Finally,",
"inspired by RegNeRF [29], we render depths D of image",
"where Fe, is a small MLP network that is used to produce",
"patches sampled from random views and minimize their to-",
"colors that depend on the direction of the ray.",
"tal variations Cty(D)."
] |
{
"arxiv_id": "2309.17390",
"language": "en",
"page_idx": 4
}
|
||
2401.08206-en-2
|
2401.08206-en-2.jpg
|
<FigureHere>
Figure 2: Overall architecture of our GeMKR.
Liu 2022), adapter methods (Zhang et al. 2023b; Diao et al. 2023), and the low-rank (LoRA) method (Hu et al. 2021).
# Methodology
We propose GeMKR, an end-to-end generative framework for multi-modal knowledge retrieval. GeMKR consists of three components, as depicted in Fig. 2: Object-aware prefix-tuning for fine-tuning the visual backbone, Multi- $\dot{\text{Modal Alignment using LLMs to capture cross-modal in-}}$ teractions, and Knowledge-guided Constraint Decoding for generating informative knowledge clues.
### Problem Formulation Formally, let $\mathcal{D}=\{D_{1},D_{2},...,D_{i}\}_{i=1}^{N}$ denotes a knowl-
edge base used for the multi-modal knowledge retrieval task. As a generative retrieval model, our goal is to generate the relevant knowledge clues $\{C\}$, which can be definitivelymapped to documents in $D.$ Our model takes the multi- $\begin{array}{l}{{\mathrm{modal~query~}Q=\{T,V\}~as~input~and~generate~the~relevant}}\\{{\mathrm{knowledge~clue~with~an~auto-regressive~score,~as~Eq.~1.}}}\end{array}$
(1)
$$\mathrm{score}(C,Q)=\mathcal{F}_{\Theta}(C|Q)=\prod_{j=1}\mathcal{F}_{\Theta}(c_{j}|\boldsymbol{c}_{<j},Q).$$
where $\mathcal{F}_\Theta$ is the generative retriever with parameters $\Theta$, and $c_{j}$ is the $j_th$ token of the knowledge clue. During inference, the model employs a constrained strategy to guide the decoder in generating content within a limited token space at each step, which ensures that each knowledge clue $C_j$ can be definitively mapped to a document $D_i$ as Eq. 2.
$$\varphi:C_j\to D_i,\mathrm{where}\:D_i\in\mathcal{D}.$$
(2)
$\begin{array}{l}{\mathrm{Finally,~we~sort~the~document~set~\{D_{i}\}~based~on~auto-}}\\{\mathrm{regressive~scores~score(C_{i},Q)~to~obtain~the~final~retrieval}}\end{array}$
results.
## Object-aware Prefix Tuning
To efficiently fine-tune the visual backbone, we present the Object-aware Prefix Tuning method that explicitly guides the visual understanding using objects (i.e., visual entities) as the learnable prefix. As shown in the bottom of Fig. 2, we utilize the CLIP model with $N$ Transformer layers as our visual backbone, and we feed two groups of features, $X_P$ and $X_I$, into CLIP using a prefix tuning approach as Eq. 3. Here, $X_P$ denotes the learnable prefix prompts, which are mixed with fine-grained visual object information, while $X_I$ represents the embeddings of each visual token encoded by the patch embedding layer. To mitigate catastrophic forgetting of the visual backbone, we freeze the parameters $\Theta_v$ of the visual backbone and only make the prefix prompts $X_P$ learnable.
(3)
$$H_v=\mathcal{F}_{\mathrm{visual}}(\boldsymbol{X}_P,\boldsymbol{X}_I;\boldsymbol{\Theta}_v).$$
For the preparation of $X_P$, we frst randomly initialize the $N$ prefix prompts $X_P_{r}$ for $N$ layers, with the parameter matrix in a dimension of $l\times d$,where $l$ denotes the length of the prefix and $d$ is the visual dimension. Note that the prefix prompts are not shared across layers. To obtain object features, we crop the objects from images and transform them into a fixed resolution. Next, we extract their
features as $X_R\in\Re^{r\times d}$ using a frozen visual encoder, CLIP. Lastly, we feed $X_R$ into a learnable projection layer as $\boldsymbol{H}_R=$Linear$(\boldsymbol X_R)$ and pad them with zero vectors to maintain the same dimension as the prefix. Taking the i$^\mathrm{th}$ layer in the visual backbone as an example, we denote the predefined prefix vector as $X_{Pr}^{i}\in \Re ^{l\times d}$ and the visual fea- tures obtained from the $( i- 1) ^{th}$ layer is $H_{I}^{i- 1}\in \Re ^{l\times d}$,we acquire the object-aware prefix vector via simply addition as $Eq.4$,
$$X_P^i=X_{Pr}^i+H_R.$$
(4)
After that, we concatenate $X_P^i$ in front of $X_I^i$, as $X=$ $[X_{P}^{i};H_{I}^{i}]$,and feed them into the self-attention module. This method allows the fine-grained object features within $X_{P}^{i}$ to better guide the visual backbone during the adaptation process.
However, the distribution discrepancy between the object feature $H_R$ and the immediate output $\tilde{H}_I^i$ may lead to a significant loss at the early training stages, potentially disturbing the fine-tuning process. Similar to (Zhang et al.2023b), we introduce a dual-flow attention mechanism to address this issue by independently computing attention weights for prompt vectors and hidden states. Specifically, We com- $\begin{array}{l}{\mathrm{pute~the~query~vector~on~}}{\boldsymbol{H}_{I}^{i},\mathrm{by~applying~Linear}_{\mathrm{Q}_{i}}(\boldsymbol{H}_{I}^{i}),}\\{\mathrm{while~the~key~and~value~vectors~are~independently~computed}}\end{array}$ for $X_P^i$ and $H_I^i$ using the linear layers Linear$_K_i(\cdot)$ and $\begin{array}{l}{\mathrm{Linear}_{V_{i}}(\cdot),\mathrm{~denoted~as~}K_{p}^{i},K_{h}^{i},V_{p}^{i},\mathrm{~and~}V_{h}^{i}.\mathrm{~We~follow}}\\{\mathrm{the~vanilla~attention~module,~as~Eq.~5,~but~individually~ap-}}\end{array}$ ply attention maps for two components and multiply value vectors to obtain outputs, as Eq 6,
(5)
$$\mathrm{Att}(\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V})=\mathrm{Softmax}(\frac{\boldsymbol{Q}^{i}(\boldsymbol{K}^{i})^{T}}{\sqrt{d}})\boldsymbol{V},$$
$$H_O^i=\mathrm{Att}(\boldsymbol{Q}_h^i,\boldsymbol{K}_p^i,\boldsymbol{V}_p^i)*\sigma+\mathrm{Att}(\boldsymbol{Q}_h^i,\boldsymbol{K}_h^i,\boldsymbol{V}_h^i).$$
(6)
|
[
[
0.08660888671875,
0.0694580078125,
0.476318359375,
0.0694580078125,
0.476318359375,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.67626953125,
0.07135009765625,
0.79833984375,
0.07135009765625,
0.79833984375,
0.07891845703125,
0.67626953125,
0.07891845703125
],
[
0.82177734375,
0.0706787109375,
0.896484375,
0.0706787109375,
0.896484375,
0.080810546875,
0.82177734375,
0.080810546875
],
[
0.0849609375,
0.08209228515625,
0.463134765625,
0.083984375,
0.463134765625,
0.09783935546875,
0.0849609375,
0.0953369140625
],
[
0.66845703125,
0.08209228515625,
0.79736328125,
0.08331298828125,
0.79736328125,
0.09783935546875,
0.66845703125,
0.09661865234375
],
[
0.6611328125,
0.10101318359375,
0.6650390625,
0.10101318359375,
0.6650390625,
0.1060791015625,
0.6611328125,
0.1060791015625
],
[
0.2254638671875,
0.10797119140625,
0.338134765625,
0.111083984375,
0.338134765625,
0.126953125,
0.2254638671875,
0.12371826171875
],
[
0.7197265625,
0.1029052734375,
0.79345703125,
0.1029052734375,
0.79345703125,
0.11297607421875,
0.7197265625,
0.11297607421875
],
[
0.6357421875,
0.10919189453125,
0.68115234375,
0.10919189453125,
0.68115234375,
0.11614990234375,
0.6357421875,
0.11614990234375
],
[
0.78173828125,
0.111083984375,
0.86767578125,
0.11297607421875,
0.86767578125,
0.1275634765625,
0.78173828125,
0.1256103515625
],
[
0.63134765625,
0.117431640625,
0.67626953125,
0.117431640625,
0.67626953125,
0.1275634765625,
0.63134765625,
0.1275634765625
],
[
0.53369140625,
0.12060546875,
0.5498046875,
0.12060546875,
0.5498046875,
0.1275634765625,
0.53369140625,
0.1275634765625
],
[
0.08660888671875,
0.1287841796875,
0.47802734375,
0.1287841796875,
0.47802734375,
0.143310546875,
0.08660888671875,
0.143310546875
],
[
0.78515625,
0.1300048828125,
0.8857421875,
0.1300048828125,
0.8857421875,
0.14013671875,
0.78515625,
0.14013671875
],
[
0.529296875,
0.1307373046875,
0.5537109375,
0.1307373046875,
0.5537109375,
0.1395263671875,
0.529296875,
0.1395263671875
],
[
0.5556640625,
0.1142578125,
0.6005859375,
0.1142578125,
0.6005859375,
0.125,
0.5556640625,
0.125
],
[
0.6640625,
0.1363525390625,
0.77783203125,
0.1363525390625,
0.77783203125,
0.149658203125,
0.6640625,
0.149658203125
],
[
0.08660888671875,
0.143310546875,
0.4794921875,
0.143310546875,
0.4794921875,
0.1572265625,
0.08660888671875,
0.1572265625
],
[
0.78515625,
0.14013671875,
0.89794921875,
0.14013671875,
0.89794921875,
0.15087890625,
0.78515625,
0.15087890625
],
[
0.0849609375,
0.1558837890625,
0.47802734375,
0.1558837890625,
0.47802734375,
0.169189453125,
0.0849609375,
0.169189453125
],
[
0.62158203125,
0.1685791015625,
0.6435546875,
0.1685791015625,
0.6435546875,
0.18115234375,
0.62158203125,
0.18115234375
],
[
0.64208984375,
0.1685791015625,
0.775390625,
0.1685791015625,
0.775390625,
0.1817626953125,
0.64208984375,
0.1817626953125
],
[
0.86767578125,
0.1634521484375,
0.89453125,
0.1634521484375,
0.89453125,
0.175537109375,
0.86767578125,
0.175537109375
],
[
0.869140625,
0.173583984375,
0.89208984375,
0.173583984375,
0.89208984375,
0.1849365234375,
0.869140625,
0.1849365234375
],
[
0.08416748046875,
0.18310546875,
0.47802734375,
0.184326171875,
0.47802734375,
0.1988525390625,
0.08416748046875,
0.1976318359375
],
[
0.0849609375,
0.169189453125,
0.47802734375,
0.169189453125,
0.47802734375,
0.18310546875,
0.0849609375,
0.18310546875
],
[
0.08416748046875,
0.1976318359375,
0.4794921875,
0.1988525390625,
0.4794921875,
0.21337890625,
0.08416748046875,
0.212158203125
],
[
0.5595703125,
0.1925048828125,
0.896484375,
0.1925048828125,
0.896484375,
0.205810546875,
0.5595703125,
0.205810546875
],
[
0.0849609375,
0.21337890625,
0.37744140625,
0.21337890625,
0.37744140625,
0.2266845703125,
0.0849609375,
0.2266845703125
],
[
0.52685546875,
0.208984375,
0.89892578125,
0.208984375,
0.89892578125,
0.2235107421875,
0.52685546875,
0.2235107421875
],
[
0.0849609375,
0.238037109375,
0.253173828125,
0.238037109375,
0.253173828125,
0.251953125,
0.0849609375,
0.251953125
],
[
0.78173828125,
0.23291015625,
0.84716796875,
0.23291015625,
0.84716796875,
0.24365234375,
0.78173828125,
0.24365234375
],
[
0.53369140625,
0.2392578125,
0.63134765625,
0.2392578125,
0.63134765625,
0.25,
0.53369140625,
0.25
],
[
0.53369140625,
0.250732421875,
0.65673828125,
0.250732421875,
0.65673828125,
0.261474609375,
0.53369140625,
0.261474609375
],
[
0.72607421875,
0.2474365234375,
0.765625,
0.2474365234375,
0.765625,
0.255126953125,
0.72607421875,
0.255126953125
],
[
0.0849609375,
0.255126953125,
0.47802734375,
0.255126953125,
0.47802734375,
0.26953125,
0.0849609375,
0.26953125
],
[
0.53369140625,
0.261474609375,
0.56201171875,
0.261474609375,
0.56201171875,
0.272705078125,
0.53369140625,
0.272705078125
],
[
0.70751953125,
0.261962890625,
0.75244140625,
0.261962890625,
0.75244140625,
0.272705078125,
0.70751953125,
0.272705078125
],
[
0.08660888671875,
0.271484375,
0.4755859375,
0.271484375,
0.4755859375,
0.2822265625,
0.08660888671875,
0.2822265625
],
[
0.08660888671875,
0.2841796875,
0.4794921875,
0.2841796875,
0.4794921875,
0.29736328125,
0.08660888671875,
0.29736328125
],
[
0.0849609375,
0.29736328125,
0.47802734375,
0.29736328125,
0.47802734375,
0.311279296875,
0.0849609375,
0.311279296875
],
[
0.71337890625,
0.29736328125,
0.77392578125,
0.29736328125,
0.77392578125,
0.30810546875,
0.71337890625,
0.30810546875
],
[
0.7998046875,
0.298583984375,
0.85546875,
0.298583984375,
0.85546875,
0.309326171875,
0.7998046875,
0.309326171875
],
[
0.08416748046875,
0.311279296875,
0.4755859375,
0.31005859375,
0.4755859375,
0.324462890625,
0.08416748046875,
0.32568359375
],
[
0.70751953125,
0.31640625,
0.76953125,
0.31640625,
0.76953125,
0.32763671875,
0.70751953125,
0.32763671875
],
[
0.78125,
0.31640625,
0.875,
0.317626953125,
0.875,
0.328857421875,
0.78125,
0.32763671875
],
[
0.08416748046875,
0.323974609375,
0.4794921875,
0.32568359375,
0.4794921875,
0.339111328125,
0.08416748046875,
0.337158203125
],
[
0.71826171875,
0.333251953125,
0.81396484375,
0.333251953125,
0.81396484375,
0.343505859375,
0.71826171875,
0.343505859375
],
[
0.81640625,
0.333251953125,
0.875,
0.333251953125,
0.875,
0.343505859375,
0.81640625,
0.343505859375
],
[
0.564453125,
0.359130859375,
0.865234375,
0.359130859375,
0.865234375,
0.37255859375,
0.564453125,
0.37255859375
],
[
0.0849609375,
0.339111328125,
0.44775390625,
0.339111328125,
0.44775390625,
0.352783203125,
0.0849609375,
0.352783203125
],
[
0.1094970703125,
0.361083984375,
0.441162109375,
0.361083984375,
0.441162109375,
0.37890625,
0.1094970703125,
0.37890625
],
[
0.44677734375,
0.3642578125,
0.47802734375,
0.3642578125,
0.47802734375,
0.375732421875,
0.44677734375,
0.375732421875
],
[
0.2998046875,
0.380615234375,
0.328369140625,
0.380615234375,
0.328369140625,
0.392578125,
0.2998046875,
0.392578125
],
[
0.0849609375,
0.397705078125,
0.4794921875,
0.397705078125,
0.4794921875,
0.410888671875,
0.0849609375,
0.410888671875
],
[
0.51953125,
0.399658203125,
0.91259765625,
0.399658203125,
0.91259765625,
0.417236328125,
0.51953125,
0.417236328125
],
[
0.08660888671875,
0.412353515625,
0.47802734375,
0.412353515625,
0.47802734375,
0.425537109375,
0.08660888671875,
0.425537109375
],
[
0.51953125,
0.415283203125,
0.91259765625,
0.416748046875,
0.91259765625,
0.43115234375,
0.51953125,
0.429931640625
],
[
0.08660888671875,
0.425537109375,
0.47802734375,
0.425537109375,
0.47802734375,
0.439453125,
0.08660888671875,
0.439453125
],
[
0.51953125,
0.43115234375,
0.9140625,
0.43115234375,
0.9140625,
0.4443359375,
0.51953125,
0.4443359375
],
[
0.08660888671875,
0.44140625,
0.47802734375,
0.44140625,
0.47802734375,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.5185546875,
0.4443359375,
0.91259765625,
0.442626953125,
0.91259765625,
0.45703125,
0.5185546875,
0.458984375
],
[
0.08416748046875,
0.451904296875,
0.4794921875,
0.453857421875,
0.4794921875,
0.468505859375,
0.08416748046875,
0.466552734375
],
[
0.521484375,
0.460205078125,
0.91259765625,
0.460205078125,
0.91259765625,
0.47021484375,
0.521484375,
0.47021484375
],
[
0.08416748046875,
0.466552734375,
0.4150390625,
0.46728515625,
0.4150390625,
0.481689453125,
0.08416748046875,
0.48095703125
],
[
0.5185546875,
0.4697265625,
0.9140625,
0.4697265625,
0.9140625,
0.4873046875,
0.5185546875,
0.4873046875
],
[
0.181396484375,
0.488037109375,
0.381591796875,
0.4873046875,
0.381591796875,
0.501953125,
0.181396484375,
0.50244140625
],
[
0.455810546875,
0.488037109375,
0.4794921875,
0.488037109375,
0.4794921875,
0.501953125,
0.455810546875,
0.501953125
],
[
0.5185546875,
0.48486328125,
0.9150390625,
0.48486328125,
0.9150390625,
0.50244140625,
0.5185546875,
0.50244140625
],
[
0.08416748046875,
0.50830078125,
0.47802734375,
0.5087890625,
0.47802734375,
0.5234375,
0.08416748046875,
0.52294921875
],
[
0.51953125,
0.50244140625,
0.9140625,
0.50244140625,
0.9140625,
0.5166015625,
0.51953125,
0.5166015625
],
[
0.51953125,
0.5166015625,
0.56396484375,
0.5166015625,
0.56396484375,
0.53076171875,
0.51953125,
0.53076171875
],
[
0.08416748046875,
0.5234375,
0.4794921875,
0.5234375,
0.4794921875,
0.537109375,
0.08416748046875,
0.537109375
],
[
0.64892578125,
0.52978515625,
0.78369140625,
0.53076171875,
0.78369140625,
0.546875,
0.64892578125,
0.54541015625
],
[
0.890625,
0.5322265625,
0.9140625,
0.5322265625,
0.9140625,
0.54541015625,
0.890625,
0.54541015625
],
[
0.0849609375,
0.53857421875,
0.1380615234375,
0.53857421875,
0.1380615234375,
0.5498046875,
0.0849609375,
0.5498046875
],
[
0.521484375,
0.55322265625,
0.8935546875,
0.55322265625,
0.8935546875,
0.56640625,
0.521484375,
0.56640625
],
[
0.08416748046875,
0.56005859375,
0.30224609375,
0.56103515625,
0.30224609375,
0.5791015625,
0.08416748046875,
0.57763671875
],
[
0.88818359375,
0.55615234375,
0.90625,
0.55615234375,
0.90625,
0.56298828125,
0.88818359375,
0.56298828125
],
[
0.52294921875,
0.568359375,
0.9111328125,
0.568359375,
0.9111328125,
0.58203125,
0.52294921875,
0.58203125
],
[
0.08660888671875,
0.580078125,
0.4794921875,
0.580078125,
0.4794921875,
0.59326171875,
0.08660888671875,
0.59326171875
],
[
0.52294921875,
0.58349609375,
0.91259765625,
0.58349609375,
0.91259765625,
0.59326171875,
0.52294921875,
0.59326171875
],
[
0.0849609375,
0.5947265625,
0.47802734375,
0.5947265625,
0.47802734375,
0.60791015625,
0.0849609375,
0.60791015625
],
[
0.521484375,
0.59521484375,
0.91259765625,
0.59521484375,
0.91259765625,
0.609375,
0.521484375,
0.609375
],
[
0.08660888671875,
0.60986328125,
0.476318359375,
0.60986328125,
0.476318359375,
0.62060546875,
0.08660888671875,
0.62060546875
],
[
0.51953125,
0.611328125,
0.6044921875,
0.611328125,
0.6044921875,
0.62255859375,
0.51953125,
0.62255859375
],
[
0.0849609375,
0.62255859375,
0.4794921875,
0.62255859375,
0.4794921875,
0.63623046875,
0.0849609375,
0.63623046875
],
[
0.53515625,
0.62353515625,
0.9140625,
0.62353515625,
0.9140625,
0.63720703125,
0.53515625,
0.63720703125
],
[
0.0849609375,
0.63623046875,
0.4794921875,
0.63623046875,
0.4794921875,
0.6494140625,
0.0849609375,
0.6494140625
],
[
0.5185546875,
0.63623046875,
0.91259765625,
0.63720703125,
0.91259765625,
0.6513671875,
0.5185546875,
0.65087890625
],
[
0.0849609375,
0.65087890625,
0.47802734375,
0.65087890625,
0.47802734375,
0.6640625,
0.0849609375,
0.6640625
],
[
0.51953125,
0.6513671875,
0.91259765625,
0.6513671875,
0.91259765625,
0.66552734375,
0.51953125,
0.66552734375
],
[
0.08416748046875,
0.662109375,
0.47802734375,
0.66357421875,
0.47802734375,
0.67822265625,
0.08416748046875,
0.6767578125
],
[
0.51953125,
0.66650390625,
0.91259765625,
0.66650390625,
0.91259765625,
0.6796875,
0.51953125,
0.6796875
],
[
0.0849609375,
0.67822265625,
0.4794921875,
0.67822265625,
0.4794921875,
0.69140625,
0.0849609375,
0.69140625
],
[
0.521484375,
0.68115234375,
0.91259765625,
0.68115234375,
0.91259765625,
0.69140625,
0.521484375,
0.69140625
],
[
0.0849609375,
0.693359375,
0.4755859375,
0.693359375,
0.4755859375,
0.7041015625,
0.0849609375,
0.7041015625
],
[
0.5185546875,
0.6923828125,
0.91259765625,
0.6943359375,
0.91259765625,
0.70849609375,
0.5185546875,
0.70556640625
],
[
0.0849609375,
0.70556640625,
0.47802734375,
0.70556640625,
0.47802734375,
0.7197265625,
0.0849609375,
0.7197265625
],
[
0.5185546875,
0.70849609375,
0.9111328125,
0.70703125,
0.9111328125,
0.72021484375,
0.5185546875,
0.7216796875
],
[
0.08660888671875,
0.72021484375,
0.4755859375,
0.72021484375,
0.4755859375,
0.73095703125,
0.08660888671875,
0.73095703125
],
[
0.5185546875,
0.7197265625,
0.91015625,
0.7216796875,
0.91015625,
0.736328125,
0.5185546875,
0.734375
],
[
0.0849609375,
0.73291015625,
0.4794921875,
0.73291015625,
0.4794921875,
0.7470703125,
0.0849609375,
0.7470703125
],
[
0.51953125,
0.73486328125,
0.91259765625,
0.73486328125,
0.91259765625,
0.74853515625,
0.51953125,
0.74853515625
],
[
0.0849609375,
0.7470703125,
0.47802734375,
0.7470703125,
0.47802734375,
0.76025390625,
0.0849609375,
0.76025390625
],
[
0.51953125,
0.74853515625,
0.9140625,
0.74853515625,
0.9140625,
0.76318359375,
0.51953125,
0.76318359375
],
[
0.08660888671875,
0.76123046875,
0.15283203125,
0.76123046875,
0.15283203125,
0.77294921875,
0.08660888671875,
0.77294921875
],
[
0.51953125,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.77783203125,
0.51953125,
0.77783203125
],
[
0.1829833984375,
0.771484375,
0.3798828125,
0.77294921875,
0.3798828125,
0.79052734375,
0.1829833984375,
0.7890625
],
[
0.45751953125,
0.7744140625,
0.47802734375,
0.7744140625,
0.47802734375,
0.78857421875,
0.45751953125,
0.78857421875
],
[
0.51953125,
0.77783203125,
0.9111328125,
0.77783203125,
0.9111328125,
0.79150390625,
0.51953125,
0.79150390625
],
[
0.101318359375,
0.79150390625,
0.4794921875,
0.79150390625,
0.4794921875,
0.80615234375,
0.101318359375,
0.80615234375
],
[
0.51953125,
0.79248046875,
0.9140625,
0.79248046875,
0.9140625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.80615234375,
0.4794921875,
0.80615234375,
0.4794921875,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.51953125,
0.80615234375,
0.744140625,
0.80615234375,
0.744140625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.08660888671875,
0.8212890625,
0.47802734375,
0.8212890625,
0.47802734375,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.08660888671875,
0.83544921875,
0.47802734375,
0.83544921875,
0.47802734375,
0.84521484375,
0.08660888671875,
0.84521484375
],
[
0.7607421875,
0.830078125,
0.82861328125,
0.82763671875,
0.82861328125,
0.84228515625,
0.7607421875,
0.8447265625
],
[
0.576171875,
0.83837890625,
0.76318359375,
0.83837890625,
0.76318359375,
0.85302734375,
0.576171875,
0.85302734375
],
[
0.82177734375,
0.83837890625,
0.85693359375,
0.83837890625,
0.85693359375,
0.85107421875,
0.82177734375,
0.85107421875
],
[
0.8896484375,
0.8388671875,
0.9140625,
0.8388671875,
0.9140625,
0.8515625,
0.8896484375,
0.8515625
],
[
0.08660888671875,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.78369140625,
0.8486328125,
0.8095703125,
0.8486328125,
0.8095703125,
0.8623046875,
0.78369140625,
0.8623046875
],
[
0.08660888671875,
0.86279296875,
0.476318359375,
0.86279296875,
0.476318359375,
0.87353515625,
0.08660888671875,
0.87353515625
],
[
0.0849609375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.5390625,
0.875,
0.9140625,
0.875,
0.9140625,
0.8896484375,
0.5390625,
0.8896484375
]
] |
[
"Liu 2022), adapter methods (Zhang et al. 2023b; Diao et al.",
"ValidDistinert)&GerNexte",
"FM-Index DB",
"2023),and thelow-rank (LoRA)method (Hu et al.2021)",
"Falise& (those,..chopper)",
"",
"Methodology",
"ValidDistinct(*)",
"ondafury",
"LookupDoc(*)",
"resemble",
"dool",
"We propose GeMKR, an end-to-end generative framework",
"Doc Id:69046...of the",
"LLM",
"honda",
"honda fury ... a chopper",
"for multi-modal knowledge retrieval. GeMKR consists of",
"honda fury...a chopper. a..",
"three components, as depicted in Fig. 2: Object-aware",
"?",
"Large Language Model",
"人",
"A",
"Modal Alignment using LLMs to capture cross-modal in-",
"prefix-tuning for fine-tuning the visual backbone, Multi-",
"teractions, and Knowledge-guided Constraint Decoding",
"Generateknowledgecluesabout theQuery:[What]",
"for generating informative knowledge clues.",
"[style][motor][pictured]IMG][CLS]][OBJ][OBJ2]Response",
"Problem Formulation",
"Proj.Layer",
"Text: What style of",
"motorcycles are pictured",
"LaverN",
"Formally, let D = {D1, D2, .., D;, denotes a knowl-",
"here?",
"Layer i-th",
"edgebaseusedforthemulti-modalknowledgeretrievaltask",
"As a generative retrieval model, our goal is to generate the",
"relevant knowledge clues [C), which can be definitively",
"PrefixVector",
"Visual Token",
"mapped to documents in D. Our model takes the multi-",
"Proj.Layer",
"Patch Linear Layer",
"modal queryQ={T,V as input andgenerate therelevant",
"国",
"",
"Figure 2: Overall architecture of our GeMKR",
"knowledge clue with an auto-regressive score, as Eq. 1.",
"score(C, Q) = Fe(C|Q) = Fe(cilc<j, Q).",
"(1)",
"j=1",
"where Fe is the generative retriever with parameters , and",
"features as XR E Rrxd using a frozen visual encoder,",
"C, is the Jth token of the knowledge clue. During inference,",
"CLIP. Lastly, we feed X R into a learnable projection layer",
"the model employs a constrained strategy to guide the de-",
"as HR = Linear(XR) and pad them with zero vectors to",
"coderingeneratingcontentwithina limitedtoken spaceat",
"maintain the same dimension as the prefix. Taking the ith",
"each step, which ensures that each knowledge clue C, can",
"layer inthe visual backbone as an example.wedenotethe",
"be definitively mapped to a document D, as Eq. 2.",
"predefined prefix vector as Xp, e R/xd and the visual fea-",
"p: Ci-→ Di, where D, E D.",
"(2)",
"tures obtained from the (i- 1)th layer is Hi-1 e R/xd, we",
"Finally, we sort the document set {D,) based on auto-",
"acquire the object-aware prefix vector via simply addition as",
"Eq. 4,",
"regressive scores score(Ci, Q) to obtain the final retrieval",
"X= Xpr +HR.",
"(4)",
"results.",
"After that, we concatenate Xp in front of Xi, as X",
"Object-aware Prefix Tuning",
"",
"[Xp; Hil, and feed them into the self-attention module",
"To efficiently fine-tune the visual backbone, we present the",
"This method allowsthe fine-grained objectfeatures within",
"Object-aware Prefix Tuning method that explicitly guides",
"X to better guide the visual backbone during the adapta-",
"the visual understanding using objects (i.e.,visual entities)",
"tion process.",
"as the learnable prefix. As shown in the bottom of Fig. 2, we",
"However, the distribution discrepancy between the object",
"utilize the CLIP model with N Transformer layers as our",
"feature Hr and the immediate output Hi may lead to a sig-",
"visual backbone, and we feed two groups of features, Xp",
"nificant loss at the early training stages, potentially disturb-",
"and Xi, into CLIP using a prefix tuning approach as Eq. 3.",
"ing the fine-tuning process. Similar to (Zhang et al. 2023b),",
"Here, Xp denotes the learnable prefix prompts, which are",
"we introduce a dual-flow attention mechanism to address",
"mixedwithfine-grainedvisual objectinformation,whileX",
"this issue by independently computing attention weights for",
"represents the embeddings of each visual token encoded by",
"prompt vectors and hidden states.Specifically,We com-",
"the patch embeddinglayer.Tomitigate catastrophicforget-",
"pute the query vector on Hi, by applying LinearQ: (Hi)",
"ting of the visual backbone, we freeze the parameters u of",
"while the key and value vectors are independently computed",
"the visual backbone and only make the prefix prompts X p",
"for Xp and Hi using the linear layers Lineark () and",
"learnable.",
"Linearv(), denoted as Kr, Ki, Vi, and V. We follow",
"Hu = Fvisual(Xp, Xr; Ou).",
"(3)",
"the vanila attention module, as Eq. 3, but individually ap-",
"For the preparation of X p, we first randomly initialize",
"ply attention maps for two components and multiply value",
"the N prefix prompts X p for N layers, with the parameter",
"vectors to obtain outputs, as Eq 6,",
"matrix in a dimension of I x d, where l denotes the length",
"of the prefix and d is the visual dimension. Note that the",
"Q\"(K')T",
"Att(Q,K, V) = Softmax(",
"-)V.",
"(5)",
"prefix prompts are not shared across layers. To obtain ob-",
"Va",
"ject features, we crop the objects from images and trans-",
"form them into a fixed resolution. Next, we extract their",
"Ho = Att(Qh,Ki, Vi)*a + Att(Qh,Ki, Vi). (6)"
] |
{
"arxiv_id": "2401.08206",
"language": "en",
"page_idx": 2
}
|
||
1808.09308-en-1
|
1808.09308-en-1.jpg
|
ing datasets; (2) word pairs are selected manually from a wide range of domains and, unlike existing datasets, are not bound to a specific resource; (3) word pairs in the dataset are balanced across the similarity scale; and (4) the huge gap between state of the art and IAA (more than 0.50 in terms of Spearman correlation) promises a challenging dataset with lots of potential for future research.
The paper is structured as follows. The following Section covers the related work, highlighting some of the issues with the RW dataset. Section 3 details the construction procedure for CARD-660. In Section 4, we analyse the dataset from different aspects, showing how it improves existing benchmarks. Section 5 reports our evaluation of mainstream word embeddings and recent word representation techniques on the dataset. Finally, concluding remarks are mentioned in Section 6.
### 2 $\textbf{Related Work}$
Word similarity datasets have been one of the oldest, still most prominent, benchmarks for the evaluation and comparison of semantic representation techniques. As a result, several word similarity datasets have been constructed during the past few decades; to name a few: RG-65 (Rubenstein and Goodenough, 1965), WordSim-353 (Finkelstein et al., 2002), YP-130 (Yang and Powers, 2005), MEN-3K (Bruni et al., 2014), SimLex-999 (Hill et al., 2015), and SimVerb-3500(Gerz et al., 2016). Many of these English word similarity datasets have been translated to other languages to create frameworks for multilingual (Leviant and Reichart, 2015) or crosslingual (CamachoCollados et al., 2017) semantic representation techniques. However, these datasets mostly target words that occur frequently in generic texts and, as a result, are not suitable for the evaluation of subword or rare word representation models.
One may opt for transforming a frequent-word benchmarks into an artificial rare word dataset by downsampling the dataset's words in the underlying training corpus (Sergienya and Schütze, 2015). However, this benchmark might not properly simulate a real-world rare word representation scenario (cf. Section 3.1).
### 2.1 $\textbf{Stanford RW Dataset}$
The Stanford Rare Word Similarity (RW) dataset is an exception as it is dedicated to evaluating infrequent word representations. The dataset has
<FigureHere>
Figure 1: Distribution of relation types (“hypernymy”, $“similar\_to”, and others) across four quartiles (sorted$ by gold similarity scores) of the Stanford Rare Word Similarity dataset. The distribution of pairs with hypernymy relation is almost uniform across the quartiles, whereas one would expect many more pairs in the top quartiles (Q4 and Q3), given the high semantic similarity of hypernym-hyponyms.
been regarded as the de facto standard evaluation benchmark for subword and rare word representation techniques. However, our analysis shows that RW suffers from multiple issues:(1) skewed distribution of the scores, (2) low-quality and inconsistent scores, and as a consequence, (3) low inter-annotator agreement.
The RW dataset comprises 2034 word pairs (i.e., word$_1- word_2) .$ Candidates for word$_1$ were randomly sampled from Wikipedia documents, distributed across a wide range of frequencies (from 5 to 10,000) to ensure the inclusion of infrequent words. Given this automatic sampling, a measure was required to avoid noisy or junk words. To this end, a sampled word was checked in WordNet (Fellbaum, 1998) and was included only if it appeared in at least one synset. Hence, the vocabulary of the dataset is bound to that of WordNet. Words for $word_2$ were randomly picked from synsets that were directly connected to a synset of word 1, through various relations, such as hypernymy, holonymy, and attributes.
## 2.1.1 Distribution of scores
These word pairs were assigned similarity scores in [0,10]. Given that all word pairs in the dataset are semantically-related according to WordNet, the scores form a skewed distribution biased towards the upper bound (see Figure 2 and Section 4.1 for more details).
### 2.1.2 Consistency of annotations
The scoring of the pairs has been carried out through crowdsourcing: (Amazon Mechanical) Turkers have provided ten scores for each word pair. The raters were restricted to only US-based
|
[
[
0.12005615234375,
0.07659912109375,
0.486083984375,
0.07659912109375,
0.486083984375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.56103515625,
0.07244873046875,
0.5859375,
0.07244873046875,
0.5859375,
0.0843505859375,
0.56103515625,
0.0843505859375
],
[
0.52880859375,
0.08074951171875,
0.5458984375,
0.08074951171875,
0.5439453125,
0.1478271484375,
0.5283203125,
0.1478271484375
],
[
0.12005615234375,
0.0926513671875,
0.486083984375,
0.0926513671875,
0.486083984375,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.55908203125,
0.0872802734375,
0.587890625,
0.0872802734375,
0.587890625,
0.1051025390625,
0.55908203125,
0.1051025390625
],
[
0.118408203125,
0.10687255859375,
0.489501953125,
0.10809326171875,
0.489501953125,
0.122314453125,
0.118408203125,
0.1217041015625
],
[
0.56103515625,
0.10809326171875,
0.58447265625,
0.10809326171875,
0.58447265625,
0.1217041015625,
0.56103515625,
0.1217041015625
],
[
0.79443359375,
0.10394287109375,
0.873046875,
0.10394287109375,
0.873046875,
0.11517333984375,
0.79443359375,
0.11517333984375
],
[
0.79443359375,
0.1181640625,
0.87158203125,
0.1181640625,
0.87158203125,
0.1300048828125,
0.79443359375,
0.1300048828125
],
[
0.12005615234375,
0.12469482421875,
0.48779296875,
0.12469482421875,
0.48779296875,
0.1383056640625,
0.12005615234375,
0.1383056640625
],
[
0.56103515625,
0.1270751953125,
0.58447265625,
0.1270751953125,
0.58447265625,
0.14013671875,
0.56103515625,
0.14013671875
],
[
0.791015625,
0.1318359375,
0.84814453125,
0.1318359375,
0.84814453125,
0.146728515625,
0.791015625,
0.146728515625
],
[
0.12005615234375,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.52880859375,
0.15087890625,
0.53662109375,
0.1455078125,
0.54248046875,
0.149658203125,
0.53564453125,
0.155029296875
],
[
0.5634765625,
0.146728515625,
0.58251953125,
0.146728515625,
0.58251953125,
0.1561279296875,
0.5634765625,
0.1561279296875
],
[
0.118408203125,
0.1561279296875,
0.48779296875,
0.1561279296875,
0.48779296875,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.1175537109375,
0.171630859375,
0.48779296875,
0.1728515625,
0.48779296875,
0.18701171875,
0.1175537109375,
0.1859130859375
],
[
0.609375,
0.174560546875,
0.62890625,
0.174560546875,
0.62890625,
0.18408203125,
0.609375,
0.18408203125
],
[
0.6484375,
0.17333984375,
0.67333984375,
0.17333984375,
0.67333984375,
0.185302734375,
0.6484375,
0.185302734375
],
[
0.69873046875,
0.174560546875,
0.71728515625,
0.174560546875,
0.71728515625,
0.18408203125,
0.69873046875,
0.18408203125
],
[
0.74658203125,
0.177001953125,
0.759765625,
0.177001953125,
0.759765625,
0.182861328125,
0.74658203125,
0.182861328125
],
[
0.118408203125,
0.189453125,
0.47265625,
0.189453125,
0.47265625,
0.203125,
0.118408203125,
0.203125
],
[
0.5146484375,
0.188232421875,
0.8818359375,
0.188232421875,
0.8818359375,
0.203125,
0.5146484375,
0.203125
],
[
0.1368408203125,
0.2060546875,
0.48779296875,
0.2060546875,
0.48779296875,
0.2197265625,
0.1368408203125,
0.2197265625
],
[
0.51708984375,
0.204833984375,
0.8818359375,
0.204833984375,
0.8818359375,
0.215576171875,
0.51708984375,
0.215576171875
],
[
0.118408203125,
0.22265625,
0.489501953125,
0.22265625,
0.489501953125,
0.236328125,
0.118408203125,
0.236328125
],
[
0.51318359375,
0.218505859375,
0.884765625,
0.218505859375,
0.884765625,
0.232177734375,
0.51318359375,
0.232177734375
],
[
0.51318359375,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.51318359375,
0.2452392578125
],
[
0.118408203125,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.51123046875,
0.2464599609375,
0.88330078125,
0.24462890625,
0.88330078125,
0.25830078125,
0.51123046875,
0.260009765625
],
[
0.118408203125,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.515625,
0.262451171875,
0.8818359375,
0.262451171875,
0.8818359375,
0.2724609375,
0.515625,
0.2724609375
],
[
0.12005615234375,
0.271484375,
0.48779296875,
0.271484375,
0.48779296875,
0.281982421875,
0.12005615234375,
0.281982421875
],
[
0.51318359375,
0.273681640625,
0.880859375,
0.273681640625,
0.880859375,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.12005615234375,
0.2861328125,
0.489501953125,
0.2861328125,
0.489501953125,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.51318359375,
0.289306640625,
0.70361328125,
0.289306640625,
0.70361328125,
0.302734375,
0.51318359375,
0.302734375
],
[
0.118408203125,
0.300537109375,
0.486083984375,
0.300537109375,
0.486083984375,
0.314208984375,
0.118408203125,
0.314208984375
],
[
0.12005615234375,
0.318359375,
0.484375,
0.318359375,
0.484375,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.118408203125,
0.333251953125,
0.486083984375,
0.333251953125,
0.486083984375,
0.3466796875,
0.118408203125,
0.3466796875
],
[
0.5146484375,
0.328857421875,
0.88330078125,
0.328857421875,
0.88330078125,
0.34326171875,
0.5146484375,
0.34326171875
],
[
0.118408203125,
0.350830078125,
0.44580078125,
0.350830078125,
0.44580078125,
0.364013671875,
0.118408203125,
0.364013671875
],
[
0.5146484375,
0.345703125,
0.88330078125,
0.345703125,
0.88330078125,
0.359375,
0.5146484375,
0.359375
],
[
0.5146484375,
0.3623046875,
0.88330078125,
0.3623046875,
0.88330078125,
0.3759765625,
0.5146484375,
0.3759765625
],
[
0.118408203125,
0.380126953125,
0.273681640625,
0.380126953125,
0.273681640625,
0.393798828125,
0.118408203125,
0.393798828125
],
[
0.51318359375,
0.377197265625,
0.8818359375,
0.377197265625,
0.8818359375,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.515625,
0.3955078125,
0.880859375,
0.3955078125,
0.880859375,
0.40625,
0.515625,
0.40625
],
[
0.118408203125,
0.407958984375,
0.486083984375,
0.407958984375,
0.486083984375,
0.421630859375,
0.118408203125,
0.421630859375
],
[
0.51318359375,
0.410400390625,
0.88330078125,
0.410400390625,
0.88330078125,
0.42333984375,
0.51318359375,
0.42333984375
],
[
0.118408203125,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.118408203125,
0.438232421875
],
[
0.51318359375,
0.427001953125,
0.70947265625,
0.427001953125,
0.70947265625,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.1175537109375,
0.439453125,
0.490234375,
0.440673828125,
0.490234375,
0.454833984375,
0.1175537109375,
0.45361328125
],
[
0.53076171875,
0.4423828125,
0.88330078125,
0.4423828125,
0.88330078125,
0.457275390625,
0.53076171875,
0.457275390625
],
[
0.118408203125,
0.457275390625,
0.489501953125,
0.457275390625,
0.489501953125,
0.47021484375,
0.118408203125,
0.47021484375
],
[
0.509765625,
0.457763671875,
0.8818359375,
0.458984375,
0.8818359375,
0.473876953125,
0.509765625,
0.47265625
],
[
0.118408203125,
0.47265625,
0.490234375,
0.47265625,
0.490234375,
0.486328125,
0.118408203125,
0.486328125
],
[
0.51123046875,
0.474365234375,
0.8818359375,
0.473876953125,
0.8818359375,
0.488037109375,
0.51123046875,
0.4892578125
],
[
0.118408203125,
0.488037109375,
0.490234375,
0.488037109375,
0.490234375,
0.501953125,
0.118408203125,
0.501953125
],
[
0.5146484375,
0.4912109375,
0.8818359375,
0.4912109375,
0.8818359375,
0.5048828125,
0.5146484375,
0.5048828125
],
[
0.118408203125,
0.5048828125,
0.489501953125,
0.5048828125,
0.489501953125,
0.5185546875,
0.118408203125,
0.5185546875
],
[
0.51123046875,
0.505859375,
0.88330078125,
0.50732421875,
0.88330078125,
0.521484375,
0.51123046875,
0.52099609375
],
[
0.1175537109375,
0.52099609375,
0.489501953125,
0.52099609375,
0.489501953125,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.51318359375,
0.52392578125,
0.88330078125,
0.52392578125,
0.88330078125,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.12005615234375,
0.5361328125,
0.489501953125,
0.5361328125,
0.489501953125,
0.5498046875,
0.12005615234375,
0.5498046875
],
[
0.51318359375,
0.54052734375,
0.88330078125,
0.54052734375,
0.88330078125,
0.55419921875,
0.51318359375,
0.55419921875
],
[
0.12005615234375,
0.552734375,
0.48779296875,
0.552734375,
0.48779296875,
0.56640625,
0.12005615234375,
0.56640625
],
[
0.51318359375,
0.5556640625,
0.88330078125,
0.5556640625,
0.88330078125,
0.5693359375,
0.51318359375,
0.5693359375
],
[
0.12005615234375,
0.568359375,
0.48779296875,
0.568359375,
0.48779296875,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.51123046875,
0.57080078125,
0.8818359375,
0.57080078125,
0.8818359375,
0.58447265625,
0.51123046875,
0.58447265625
],
[
0.12005615234375,
0.5849609375,
0.489501953125,
0.5849609375,
0.489501953125,
0.5986328125,
0.12005615234375,
0.5986328125
],
[
0.51318359375,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6015625,
0.51318359375,
0.6015625
],
[
0.118408203125,
0.6015625,
0.490234375,
0.6015625,
0.490234375,
0.615234375,
0.118408203125,
0.615234375
],
[
0.51318359375,
0.60400390625,
0.884765625,
0.60400390625,
0.884765625,
0.61767578125,
0.51318359375,
0.61767578125
],
[
0.118408203125,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.118408203125,
0.6298828125
],
[
0.51318359375,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.12005615234375,
0.6328125,
0.490234375,
0.6328125,
0.490234375,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.5146484375,
0.6357421875,
0.88330078125,
0.6357421875,
0.88330078125,
0.6494140625,
0.5146484375,
0.6494140625
],
[
0.12005615234375,
0.6494140625,
0.489501953125,
0.6494140625,
0.489501953125,
0.6630859375,
0.12005615234375,
0.6630859375
],
[
0.51123046875,
0.65087890625,
0.88330078125,
0.65185546875,
0.88330078125,
0.66650390625,
0.51123046875,
0.6650390625
],
[
0.12005615234375,
0.66650390625,
0.489501953125,
0.66650390625,
0.489501953125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.5146484375,
0.66943359375,
0.8505859375,
0.66943359375,
0.8505859375,
0.68310546875,
0.5146484375,
0.68310546875
],
[
0.118408203125,
0.681640625,
0.490234375,
0.681640625,
0.490234375,
0.6953125,
0.118408203125,
0.6953125
],
[
0.118408203125,
0.69873046875,
0.44921875,
0.69873046875,
0.44921875,
0.7119140625,
0.118408203125,
0.7119140625
],
[
0.5146484375,
0.69677734375,
0.74072265625,
0.69677734375,
0.74072265625,
0.71044921875,
0.5146484375,
0.71044921875
],
[
0.1368408203125,
0.71533203125,
0.490234375,
0.71533203125,
0.490234375,
0.728515625,
0.1368408203125,
0.728515625
],
[
0.5146484375,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.1175537109375,
0.72998046875,
0.489501953125,
0.73095703125,
0.489501953125,
0.7451171875,
0.1175537109375,
0.74462890625
],
[
0.51318359375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.51318359375,
0.74755859375
],
[
0.118408203125,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.51318359375,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.118408203125,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.118408203125,
0.775390625
],
[
0.51318359375,
0.765625,
0.8818359375,
0.765625,
0.8818359375,
0.779296875,
0.51318359375,
0.779296875
],
[
0.1175537109375,
0.77783203125,
0.490234375,
0.779296875,
0.490234375,
0.79345703125,
0.1175537109375,
0.7919921875
],
[
0.51318359375,
0.7822265625,
0.884765625,
0.7822265625,
0.884765625,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.7958984375,
0.49365234375,
0.7958984375,
0.49365234375,
0.80859375,
0.118408203125,
0.80859375
],
[
0.51318359375,
0.79736328125,
0.6708984375,
0.79736328125,
0.6708984375,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.118408203125,
0.81103515625,
0.2880859375,
0.81103515625,
0.2880859375,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51318359375,
0.82470703125,
0.7841796875,
0.82470703125,
0.7841796875,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.118408203125,
0.8388671875,
0.33154296875,
0.8388671875,
0.33154296875,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.5146484375,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.1175537109375,
0.8623046875,
0.491943359375,
0.8623046875,
0.491943359375,
0.8759765625,
0.1175537109375,
0.8759765625
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87890625,
0.49365234375,
0.87890625,
0.49365234375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.118408203125,
0.89453125,
0.491943359375,
0.8935546875,
0.491943359375,
0.90771484375,
0.118408203125,
0.9091796875
],
[
0.51123046875,
0.89453125,
0.88330078125,
0.8935546875,
0.88330078125,
0.90771484375,
0.51123046875,
0.9091796875
]
] |
[
"ing datasets; (2) word pairs are selected manually",
"500",
"i",
"from a wide range of domains and, unlike exist-",
"400",
"ing datasets, are not bound to a specific resource;",
"300",
"Hypernym",
"Similar_to",
"(3) word pairs in the dataset are balanced across",
"200",
"Other",
"the similarity scale; and (4) the huge gap between",
"2",
"100",
"state of the art and IAA (more than 0.50 in terms",
"of Spearman correlation) promises a challenging",
"Q1",
"Q2",
"Q3",
"04",
"dataset with lots of potential for future research.",
"Figure l: Distribution of relation types (hypernymy\"",
"The paper is structured as follows. The follow-",
"\"similar_to\",and others)across fourquartiles(sorted",
"ing Section covers the related work, highlighting",
"by gold similarity scores) of the Stanford Rare Word",
"Similarity dataset. The distribution of pairs with hy-",
"some of the issues with the RW dataset. Section 3",
" pernymy relation is almost uniform across the quartiles.",
"details the construction procedure for CARD-660.",
"whereasonewouldexpectmanymorepairsinthetop",
"InSection4,weanalysethedatasetfromdifferent",
"quartiles (Q4 and Q3), given the high semantic similar",
"aspects, showing how it improves existing bench-",
"ity of hypernym-hyponyms.",
"marks. Section 5 reports our evaluation of main-",
"stream word embeddings and recent word repre-",
"sentation techniques on the dataset. Finally, con-",
"been regarded as the de facto standard evaluation",
"cluding remarks are mentioned in Section 6.",
"benchmark for subword and rare word represen-",
"tation techniques. However, our analysis shows",
"2Related Work",
"that RW suffers from multiple issues: (1) skewed",
"distribution of the scores,(2)low-quality and in-",
"Word similarity datasets have been one of the old-",
"consistent scores, and as a consequence, (3) low",
"est, still most prominent, benchmarks for the eval-",
"inter-annotator agreement.",
"uation and comparison of semantic representation",
"The RW dataset comprises 2034 word pairs",
"techniques.As a result, several word similar-",
"(i.e., wordi - word2).Candidates for wordi",
"ity datasets have been constructed during the past",
"were randomly sampled from Wikipedia docu-",
"few decades; to name a few: RG-65 (Rubenstein",
"ments, distributed across a wide range of frequen-",
"and Goodenough, 1965), WordSim-353 (Finkel-",
"cies (from 5 to 10,000) to ensure the inclusion",
"stein et al., 2002), YP-130 (Yang and Powers,",
"of infrequent words. Given this automatic sam-",
"2005), MEN-3K (Bruni et al., 2014), SimLex-999",
"pling, a measure was required to avoid noisy or",
"(Hill et al., 2015), and Sim Verb-3500 (Gerz et al.,",
"junk words.To this end, a sampled word was",
"2016). Many of these English word similarity",
"checked in WordNet (Fellbaum, 1998) and was in-",
"datasets have been translated to other languages",
"cluded only if it appeared in at least one synset.",
"to create frameworks for multilingual (Leviant",
"Hence, the vocabulary of the dataset is bound to",
"and Reichart, 2015) or crosslingual (Camacho-",
"that of WordNet. Words for word were randomly",
"Collados et al., 2017) semantic representation",
"picked from synsets that were directly connected",
"techniques. However, these datasets mostly target",
"to a synset of wordi, through various relations,",
"words that occur frequently in generic texts and,",
"such as hypernymy, holonymy, and attributes.",
"as a result, are not suitable for the evaluation of",
"subword or rare word representation models.",
"2.1.1Distribution of scores",
"One may opt for transforming a frequent-word",
"These word pairs were assigned similarity scores",
"benchmarks into an artificial rare word dataset by",
"in [0, 10]. Given that all word pairs in the dataset",
"downsampling the dataset's words in the underly-",
"are semantically-related according to WordNet,",
"ing training corpus (Sergienya and Schuitze, 2015).",
"the scores form a skewed distribution biased to-",
"However, this benchmark might not properly sim-",
"wards the upper bound (see Figure 2 and Section",
"ulate a real-world rare word representation sce-",
"4.1 for more details).",
"nario (cf. Section 3.1).",
"2.1.2 Consistency of annotations",
"2.1 Stanford RW Dataset",
"The scoring of the pairs has been carried out",
"The Stanford Rare Word Similarity (RW) dataset",
"through crowdsourcing: (Amazon Mechanical)",
"is an exception as it is dedicated to evaluating in-",
"Turkers have provided ten scores for each word",
"frequent word representations.The dataset has",
"pair. The raters were restricted to only Us-based"
] |
{
"arxiv_id": "1808.09308",
"language": "en",
"page_idx": 1
}
|
||
1808.09308-en-3
|
1808.09308-en-3.jpg
|
benchmark that can represent the challenging nature of the task: (1) it is unable to control the impact of second-order associations (words that frequently co-occur with the downsampled word) and cannot represent a real-world setting with novel rare usages; (2) given that morphological variations of a word (such as plural forms) are kept intact in this procedure, a subword technique can easily resort to these forms to compute the embedding for downsampled words; and (3) a constrained evaluation configuration in which the task is to estimate the embedding for a (rare) word using one or few occurrences (contexts) of it, limits the benchmark to a subset of corpus-based rare word representation techniques only. Moreover, the evaluation would require the comparison of the computed embeddings for rare words with a set of reference embeddings (computed on the full data). This dependency limits the ability of the benchmark in providing a direct evaluation of the rare word representation technique, independently from the impact of the model used to compute the reference embeddings.
The CARD-660 dataset aims at filling the gap for rigorous generic evaluation of rare word and subword representation models. In what follows in this section, we will detail the construction procedure of the dataset which was carefully planned to guarantee a challenging and reliable dataset.
## 3.2 Construction Procedure
The following four-phase procedure was used to
construct the dataset:
(1) A set of 660 rare words were carefully se-
lected from a wide range of domains;
(2) For each of these initial words, a pairing
word was manually selected according to a randomly sampled score from the similarity scale (Section 3.2.2);
(3) All pairs were scored by 8 annotators;
(4) A final adjudication was performed to ad-
dress disagreements (Section 3.2.3).
## 3.2.1 Similarity scale
We adopted the five-point Likert scale used for the annotation of the datasets in SemEval-2017 Task 2 (Camacho-Collados et al., 2017). The task reported high IAA scores which reflects the welldefinedness and clarity of the scale. We provided
$^{2}$http://www.fakenewschallenge.org
annotators with the concise guideline shown in Table 1, along with several examples. Given the continuity of the scale, the annotators were given flexibility to select values in between the five points, whenever appropriate, with a step size of 0.5.
The annotators were asked in the guidelines to make sure they were familiar with all common meanings of the word (as defined by WordNet or other online dictionaries). To facilitate the annotation, the annotators were provided with the definitions of some of the words that were defined in WordNet or named entities that had Wikipedia pages. For others, we asked the annotators to check the word in online dictionaries, such as WordNet browser$^3$ and Wiktionary$^4$, or encyclopediae, such as Wikipedia.
## 3.2.2 Word pair selection
Unlike previous work (Luong et al., 2013), we did not rely on random sampling (pruned by frequency) of initial words from a specific dictionary, to prevent the dataset from being restricted to a specific resource or vocabulary. Instead, we carefully hand-picked word pairs from a wide range of domains. To construct the 660 pairs of the dataset (each pair is denoted as $w_1-w_2)$, we first picked 660 $w_1$ words. Our aim was to have a dataset that can ideally reflect the performance of rare word representation techniques in downstream NLP tasks. To this end, we picked initial words $(w_1$s) from different common NLP datasets and resources, listed in Table 2. For each textbased resource, a frequency list was obtained and rare words were carefully picked from the long tail of the list, cross-checking the frequency of words in the Google News dataset. For the other resources (such as Wiktionary), we checked a word against a large frequency list to ensure they are not frequent words. The list was computed on the 2.8B token ukWaC+WaCkypedia corpus (Baroni et al., 2009) and comprised 16.5M unique words.
In order to have a balanced distribution of scores in the dataset, we first assigned random integer scores in [0-4] to the 660 initial $w_1$s. Then, with the corresponding score in mind, a pairing word $(w_2)$ was selected for each $w_1.$ We show in Section 4.1 that this strategy resulted in a uniformly distributed set of scores in the dataset.
The dataset comprises words from a wide range of genres and domains, including slang in so-
$^3$http://wordnetweb.princeton.edu/perl/webwn
$^{4}{\mathrm{www.wiktionary.org}}$
|
[
[
0.118408203125,
0.07598876953125,
0.48779296875,
0.07598876953125,
0.48779296875,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.5146484375,
0.07659912109375,
0.880859375,
0.07659912109375,
0.880859375,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.12005615234375,
0.0926513671875,
0.489501953125,
0.0926513671875,
0.489501953125,
0.105712890625,
0.12005615234375,
0.105712890625
],
[
0.5146484375,
0.0926513671875,
0.8818359375,
0.0926513671875,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.118408203125,
0.1092529296875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.51318359375,
0.10687255859375,
0.88330078125,
0.10809326171875,
0.88330078125,
0.122314453125,
0.51318359375,
0.1217041015625
],
[
0.118408203125,
0.12469482421875,
0.489501953125,
0.12353515625,
0.489501953125,
0.1383056640625,
0.118408203125,
0.1395263671875
],
[
0.51123046875,
0.12469482421875,
0.880859375,
0.12469482421875,
0.880859375,
0.1383056640625,
0.51123046875,
0.1383056640625
],
[
0.118408203125,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.118408203125,
0.155029296875
],
[
0.51123046875,
0.14013671875,
0.84814453125,
0.1395263671875,
0.84814453125,
0.15380859375,
0.51123046875,
0.155029296875
],
[
0.1175537109375,
0.1561279296875,
0.486083984375,
0.1561279296875,
0.486083984375,
0.169189453125,
0.1175537109375,
0.169189453125
],
[
0.53076171875,
0.1561279296875,
0.8818359375,
0.1561279296875,
0.8818359375,
0.169189453125,
0.53076171875,
0.169189453125
],
[
0.118408203125,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.51318359375,
0.1728515625,
0.88330078125,
0.1728515625,
0.88330078125,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.1175537109375,
0.188232421875,
0.489501953125,
0.188232421875,
0.489501953125,
0.201904296875,
0.1175537109375,
0.201904296875
],
[
0.51318359375,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.51318359375,
0.203125
],
[
0.118408203125,
0.204833984375,
0.489501953125,
0.204833984375,
0.489501953125,
0.218505859375,
0.118408203125,
0.218505859375
],
[
0.509765625,
0.204833984375,
0.880859375,
0.204833984375,
0.880859375,
0.218505859375,
0.509765625,
0.218505859375
],
[
0.118408203125,
0.22265625,
0.48779296875,
0.22265625,
0.48779296875,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.51318359375,
0.221435546875,
0.88330078125,
0.221435546875,
0.88330078125,
0.235107421875,
0.51318359375,
0.235107421875
],
[
0.118408203125,
0.2369384765625,
0.489501953125,
0.2369384765625,
0.489501953125,
0.25048828125,
0.118408203125,
0.25048828125
],
[
0.51123046875,
0.2369384765625,
0.88330078125,
0.2369384765625,
0.88330078125,
0.25048828125,
0.51123046875,
0.25048828125
],
[
0.118408203125,
0.254150390625,
0.490234375,
0.254150390625,
0.490234375,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.51318359375,
0.254150390625,
0.88330078125,
0.254150390625,
0.88330078125,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.12005615234375,
0.26953125,
0.489501953125,
0.26953125,
0.489501953125,
0.283203125,
0.12005615234375,
0.283203125
],
[
0.51123046875,
0.26953125,
0.88330078125,
0.268310546875,
0.88330078125,
0.283203125,
0.51123046875,
0.283935546875
],
[
0.12005615234375,
0.287353515625,
0.48779296875,
0.287353515625,
0.48779296875,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.51318359375,
0.28515625,
0.884765625,
0.28515625,
0.884765625,
0.298583984375,
0.51318359375,
0.298583984375
],
[
0.12005615234375,
0.3017578125,
0.489501953125,
0.3017578125,
0.489501953125,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.5146484375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.5146484375,
0.3154296875
],
[
0.12005615234375,
0.318359375,
0.489501953125,
0.318359375,
0.489501953125,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.51318359375,
0.317626953125,
0.7109375,
0.317626953125,
0.7109375,
0.33203125,
0.51318359375,
0.33203125
],
[
0.12005615234375,
0.334228515625,
0.489501953125,
0.334228515625,
0.489501953125,
0.347412109375,
0.12005615234375,
0.347412109375
],
[
0.5146484375,
0.341552734375,
0.7255859375,
0.341552734375,
0.7255859375,
0.355712890625,
0.5146484375,
0.355712890625
],
[
0.118408203125,
0.3486328125,
0.48779296875,
0.3486328125,
0.48779296875,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.5146484375,
0.361083984375,
0.884765625,
0.361083984375,
0.884765625,
0.374755859375,
0.5146484375,
0.374755859375
],
[
0.12176513671875,
0.36767578125,
0.48779296875,
0.36767578125,
0.48779296875,
0.377685546875,
0.12176513671875,
0.377685546875
],
[
0.51318359375,
0.3759765625,
0.8818359375,
0.377197265625,
0.8818359375,
0.391357421875,
0.51318359375,
0.39013671875
],
[
0.118408203125,
0.3818359375,
0.490234375,
0.3818359375,
0.490234375,
0.3955078125,
0.118408203125,
0.3955078125
],
[
0.118408203125,
0.3984375,
0.48779296875,
0.3984375,
0.48779296875,
0.412109375,
0.118408203125,
0.412109375
],
[
0.5146484375,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.40673828125,
0.5146484375,
0.40673828125
],
[
0.1175537109375,
0.413330078125,
0.490234375,
0.41455078125,
0.490234375,
0.4287109375,
0.1175537109375,
0.427490234375
],
[
0.51318359375,
0.4091796875,
0.884765625,
0.4091796875,
0.884765625,
0.42333984375,
0.51318359375,
0.42333984375
],
[
0.1175537109375,
0.43115234375,
0.287109375,
0.43115234375,
0.287109375,
0.44482421875,
0.1175537109375,
0.44482421875
],
[
0.51318359375,
0.42578125,
0.8818359375,
0.42578125,
0.8818359375,
0.439453125,
0.51318359375,
0.439453125
],
[
0.1351318359375,
0.4453125,
0.489501953125,
0.446533203125,
0.489501953125,
0.46142578125,
0.1351318359375,
0.460205078125
],
[
0.51123046875,
0.440673828125,
0.88330078125,
0.441162109375,
0.88330078125,
0.4560546875,
0.51123046875,
0.454833984375
],
[
0.118408203125,
0.463134765625,
0.490234375,
0.463134765625,
0.490234375,
0.47802734375,
0.118408203125,
0.47802734375
],
[
0.51318359375,
0.457763671875,
0.88330078125,
0.457763671875,
0.88330078125,
0.471435546875,
0.51318359375,
0.471435546875
],
[
0.118408203125,
0.479736328125,
0.490234375,
0.479736328125,
0.490234375,
0.493408203125,
0.118408203125,
0.493408203125
],
[
0.51123046875,
0.474365234375,
0.88330078125,
0.474365234375,
0.88330078125,
0.488037109375,
0.51123046875,
0.488037109375
],
[
0.118408203125,
0.495361328125,
0.490234375,
0.495361328125,
0.490234375,
0.5087890625,
0.118408203125,
0.5087890625
],
[
0.51318359375,
0.490478515625,
0.884765625,
0.490478515625,
0.884765625,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.51318359375,
0.505859375,
0.88330078125,
0.505859375,
0.88330078125,
0.51953125,
0.51318359375,
0.51953125
],
[
0.118408203125,
0.5126953125,
0.490234375,
0.5126953125,
0.490234375,
0.525390625,
0.118408203125,
0.525390625
],
[
0.51123046875,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.5361328125,
0.51123046875,
0.5361328125
],
[
0.1175537109375,
0.529296875,
0.465087890625,
0.52685546875,
0.465087890625,
0.54150390625,
0.1175537109375,
0.54345703125
],
[
0.51318359375,
0.5380859375,
0.884765625,
0.5380859375,
0.884765625,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.12005615234375,
0.5556640625,
0.35107421875,
0.5556640625,
0.35107421875,
0.57080078125,
0.12005615234375,
0.57080078125
],
[
0.51123046875,
0.55419921875,
0.884765625,
0.552734375,
0.884765625,
0.56787109375,
0.51123046875,
0.568359375
],
[
0.51123046875,
0.5693359375,
0.88330078125,
0.568359375,
0.88330078125,
0.5830078125,
0.51123046875,
0.58447265625
],
[
0.1175537109375,
0.57763671875,
0.489501953125,
0.5791015625,
0.489501953125,
0.59326171875,
0.1175537109375,
0.5927734375
],
[
0.5146484375,
0.58740234375,
0.88330078125,
0.58740234375,
0.88330078125,
0.60107421875,
0.5146484375,
0.60107421875
],
[
0.118408203125,
0.5966796875,
0.274658203125,
0.5966796875,
0.274658203125,
0.60693359375,
0.118408203125,
0.60693359375
],
[
0.5146484375,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.5146484375,
0.6162109375
],
[
0.124267578125,
0.62451171875,
0.489501953125,
0.62353515625,
0.489501953125,
0.63818359375,
0.124267578125,
0.6396484375
],
[
0.5146484375,
0.62060546875,
0.8818359375,
0.62060546875,
0.8818359375,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.15283203125,
0.64013671875,
0.432373046875,
0.64111328125,
0.432373046875,
0.65625,
0.15283203125,
0.65478515625
],
[
0.51318359375,
0.63525390625,
0.880859375,
0.63525390625,
0.880859375,
0.6484375,
0.51318359375,
0.6484375
],
[
0.51318359375,
0.65087890625,
0.88330078125,
0.65087890625,
0.88330078125,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.1259765625,
0.65673828125,
0.489501953125,
0.65673828125,
0.489501953125,
0.67138671875,
0.1259765625,
0.67138671875
],
[
0.51123046875,
0.66650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6796875,
0.51123046875,
0.68115234375
],
[
0.154541015625,
0.67333984375,
0.490234375,
0.67333984375,
0.490234375,
0.68701171875,
0.154541015625,
0.68701171875
],
[
0.5146484375,
0.68310546875,
0.8818359375,
0.68310546875,
0.8818359375,
0.69677734375,
0.5146484375,
0.69677734375
],
[
0.154541015625,
0.689453125,
0.48779296875,
0.689453125,
0.48779296875,
0.70361328125,
0.154541015625,
0.70361328125
],
[
0.51318359375,
0.69775390625,
0.88330078125,
0.69873046875,
0.88330078125,
0.71337890625,
0.51318359375,
0.7119140625
],
[
0.154541015625,
0.7060546875,
0.31494140625,
0.7060546875,
0.31494140625,
0.7197265625,
0.154541015625,
0.7197265625
],
[
0.51123046875,
0.71435546875,
0.8759765625,
0.71337890625,
0.8759765625,
0.72802734375,
0.51123046875,
0.728515625
],
[
0.1259765625,
0.7216796875,
0.438232421875,
0.7216796875,
0.438232421875,
0.736328125,
0.1259765625,
0.736328125
],
[
0.52880859375,
0.72998046875,
0.88330078125,
0.72998046875,
0.88330078125,
0.74365234375,
0.52880859375,
0.74365234375
],
[
0.1234130859375,
0.73681640625,
0.486083984375,
0.73681640625,
0.486083984375,
0.75048828125,
0.1234130859375,
0.75048828125
],
[
0.154541015625,
0.7529296875,
0.42236328125,
0.75439453125,
0.42236328125,
0.7685546875,
0.154541015625,
0.76708984375
],
[
0.5146484375,
0.74755859375,
0.88330078125,
0.74755859375,
0.88330078125,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.5146484375,
0.76318359375,
0.88330078125,
0.76318359375,
0.88330078125,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.118408203125,
0.7822265625,
0.298828125,
0.783203125,
0.298828125,
0.79736328125,
0.118408203125,
0.79638671875
],
[
0.51318359375,
0.779296875,
0.88330078125,
0.77978515625,
0.88330078125,
0.79443359375,
0.51318359375,
0.79345703125
],
[
0.5146484375,
0.79638671875,
0.88330078125,
0.79638671875,
0.88330078125,
0.81005859375,
0.5146484375,
0.81005859375
],
[
0.12005615234375,
0.80419921875,
0.489501953125,
0.80419921875,
0.489501953125,
0.81787109375,
0.12005615234375,
0.81787109375
],
[
0.5146484375,
0.8125,
0.8818359375,
0.8125,
0.8818359375,
0.826171875,
0.5146484375,
0.826171875
],
[
0.118408203125,
0.82080078125,
0.489501953125,
0.82080078125,
0.489501953125,
0.83447265625,
0.118408203125,
0.83447265625
],
[
0.5146484375,
0.82763671875,
0.8486328125,
0.82763671875,
0.8486328125,
0.84130859375,
0.5146484375,
0.84130859375
],
[
0.118408203125,
0.8359375,
0.486083984375,
0.8359375,
0.486083984375,
0.849609375,
0.118408203125,
0.849609375
],
[
0.53076171875,
0.8427734375,
0.88330078125,
0.84423828125,
0.88330078125,
0.85791015625,
0.53076171875,
0.85546875
],
[
0.12005615234375,
0.8525390625,
0.484375,
0.8525390625,
0.484375,
0.8662109375,
0.12005615234375,
0.8662109375
],
[
0.51318359375,
0.85986328125,
0.88330078125,
0.85986328125,
0.88330078125,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.118408203125,
0.8681640625,
0.489501953125,
0.8681640625,
0.489501953125,
0.8828125,
0.118408203125,
0.8828125
],
[
0.53564453125,
0.8818359375,
0.8154296875,
0.8818359375,
0.8154296875,
0.8955078125,
0.53564453125,
0.8955078125
],
[
0.139404296875,
0.8955078125,
0.358642578125,
0.8955078125,
0.358642578125,
0.9091796875,
0.139404296875,
0.9091796875
],
[
0.5341796875,
0.89453125,
0.6708984375,
0.8955078125,
0.6708984375,
0.91015625,
0.5341796875,
0.9091796875
]
] |
[
"benchmark that can represent the challenging na-",
"annotators with the concise guideline shown in Ta-",
"ture of the task: (1) it is unable to control the im-",
"ble 1, along with several examples. Given the con-",
"pact of second-order associations (words that fre-",
"tinuity of the scale, the annotators were given flex-",
"quently co-occur with the downsampled word) and",
"ibility to select values in between the five points.",
"cannot represent a real-world setting with novel",
"whenever appropriate, with a step size of 0.5.",
"rare usages: (2) given that morphological varia-",
"The annotators were asked in the guidelines to",
"tions of a word (such as plural forms) are kept",
"make sure they were familiar with all common",
"intact in this procedure, a subword technique can",
"meanings of the word (as defined by WordNet or",
"easily resort to these forms to compute the em-",
"other online dictionaries). To facilitate the anno",
"bedding for downsampled words: and (3) a con-",
"tation, the annotators were provided with the def-",
"strained evaluation configuration in which the task",
"initions of some of the words that were defined",
"is to estimate the embedding for a (rare) word us-",
"in WordNet or named entities that had Wikipedia",
"ing one or few occurrences (contexts) of it, lim-",
"pages.For others, we asked the annotators to",
"itsthebenchmarktoasubsetofcorpus-basedrare",
"check the word in online dictionaries, such as",
"word representation techniques only. Moreover,",
"WordNet browser3 and Wiktionary4, or encyclo-",
"the evaluation would require the comparison of",
"pediae, such as Wikipedia.",
"the computed embeddings for rare words with a",
"3.2.2 Word pair selection",
"set of reference embeddings (computed on the full",
"Unlike previous work (Luong et al., 2013), we",
"data).Thisdependency limits the ability of the",
"did not rely on random sampling (pruned by fre-",
"benchmark in providing a direct evaluation of the",
"rare word representation technique, independently",
"quency) of initial words from a specific dictionary.",
"from the impact of the model used to compute the",
"to prevent the dataset from being restricted to a",
"reference embeddings.",
"specific resource or vocabulary. Instead, we care-",
"The CARD-660 dataset aims at filling the gap",
"fully hand-picked word pairs from a wide range",
"for rigorous generic evaluation of rare word and",
"of domains. To construct the 660 pairs of the",
"subword representation models. In what follows",
"dataset (each pair is denoted as w1 - w2), we",
"in this section, we will detail the construction pro-",
"first picked 660 w1 words. Our aim was to have",
"a dataset that can ideally reflect the performance",
"cedure of the dataset which was carefully planned",
"of rare word representation techniques in down-",
"to guarantee a challenging and reliable dataset.",
"stream NLP tasks. To this end, we picked initial",
"3.2Construction Procedure",
"words (wis) from different common NLP datasets",
"and resources, listed in Table 2. For each text-",
"The following four-phase procedure was used to",
"based resource, a frequency list was obtained and",
"construct the dataset:",
"rare words were carefully picked from the long tail",
"(1) A set of 660 rare words were carefully se-",
"of thelist,cross-checkingthefrequencyof words",
"lected from a wide range of domains;",
"in the Google News dataset. For the other re-",
"sources (such as Wiktionary), we checked a word",
"(2) For each of these initial words, a pairing",
"against a large frequency list to ensure they are",
"word was manually selected according to a",
"not frequent words. The list was computed on the",
"randomly sampled score from the similarity",
"2.8B token ukWaC+WaCkypedia corpus (Baroni",
"scale (Section 3.2.2);",
"et al., 2009) and comprised 16.5M unique words",
"(3) All pairs were scored by 8 annotators;",
"In order to have a balanced distribution of",
"(4) A final adjudication was performed to ad-",
"dress disagreements (Section 3.2.3).",
"scores in the dataset, we first assigned random in-",
"teger scores in [0 - 4] to the 660 initial wis. Then,",
"3.2.1 Similarity scale",
"with the corresponding score in mind, a pairing",
"word (w2) was selected for each wi. We show",
"We adopted the five-point Likert scale used for the",
"in Section 4.1 that this strategy resulted in a uni-",
"annotation of the datasets in SemEval-2017 Task",
"formly distributed set of scores in the dataset.",
"2 (Camacho-Collados et al., 2017). The task re-",
"The dataset comprises words from a wide range",
"ported high IAA scores which reflects the well",
"of genres and domains, including slang in so-",
"definedness and clarity of the scale. We provided",
"3http://wordnetweb.princeton.edu/perl/webwn",
"2http://www.fakenewschallenge.org",
"4www.wiktionary.org"
] |
{
"arxiv_id": "1808.09308",
"language": "en",
"page_idx": 3
}
|
||
1805.02262-en-1
|
1805.02262-en-1.jpg
|
2016), and assume that entity types in the test set match those labeled in the training set (including work on domain adaptation, e.g., Daumé, 2007). These assumptions, while useful for developing and benchmarking new methods, are unrealistic for many domains and applications. The paper also serves as an overview of the approach we adopt at www.semanticscholar.org in a step towards more intelligent academic search engines (Etzioni, 2011).
In the next section, we start by describing our symbolic representation of the literature. Then, we discuss how we extract metadata associated with a paper such as authors and references, then how we extract the entities mentioned in paper text. Before we conclude, we briefly describe other research challenges we are actively working on in order to improve the quality of the literature graph.
2 $\textbf{Structure of The Literature Graph}$
The literature graph is a property graph with directed edges. Unlike Resource Description Framework (RDF) graphs, nodes and edges in property graphs have an internal structure which is more suitable for representing complex data types such as papers and entities. In this section, we describe the attributes associated with nodes and edges of different types in the literature graph.
# 2.1 Node Types
Papers. We obtain metadata and PDF files of papers via partnerships with publishers (e.g., Springer, Nature), catalogs (e.g., DBLP, MEDLINE), pre-publishing services (e.g., arXiv, bioRxive), as well as web-crawling. Paper nodes are associated with a set of attributes such as `title', 'abstract',“full text’,“venues’and‘publication year'. While some of the paper sources provide these attributes as metadata, it is often necessary to extract them from the paper PDF (details in §3). We deterministically remove duplicate papers based on string similarity of their metadata, resulting in 37M unique paper nodes. Papers in the literature graph cover a variety of scientific disciplines, including computer science, molecular biology, microbiology and neuroscience.
Authors. Each node of this type represents a unique author, with attributes such as ‘first name’ and ‘last name’. The literature graph has 12M nodes of this type.
Entities. Each node of this type represents a unique scientific concept discussed in the literature, with attributes such as ‘canonical name’,‘aliases’ and ‘description’. Our literature graph has 0.4M nodes of this type. We describe how we populate entity nodes in §4.3.
Entity mentions. Each node of this type represents a textual reference of an entity in one of the papers, with attributes such as‘mention text’, 'context', and‘confidence’. We describe how we populate the 237M mentions in the literature graph $in$ $\S 4. 1.$
# 2.2 $\mathbf{EdgeTypes}$
Citations. We instantiate a directed citation edge from paper nodes $p_1\longrightarrow p_2$ for each $p_2$ referenced in $p_1.$ Citation edges have attributes such as `from paper id', 'to paper id' and‘contexts’(the textual contexts where $p_2$ is referenced in $p_1).$ While some of the paper sources provide these attributes as metadata, it is often necessary to extract them from the paper PDF as detailed in §3.
Authorship. We instantiate a directed authorship edge between an author node and a paper node $a\longrightarrow p$ for each author of that paper.
Entity linking edges. We instantiate a directed edge from an extracted entity mention node to the entity it refers to.
Mention-mention relations. We instantiate a directed edge between a pair of mentions in the same sentential context if the textual relation extraction model predicts one of a predefined list of relation types between them in a sentential context.$^{1}$ We encode a symmetric relation between $m_1$ and $m_2$ as two directed edges $m_1\longrightarrow m_2$ and $m_2\longrightarrow m_1.$
While mention-
Entity-entity relations.
mention edges represent relations between mentions in a particular context, entity-entity edges represent relations between abstract entities. These relations may be imported from an existing knowledge base (KB) or inferred from other edges in the graph.
# 3 Extracting Metadata
In the previous section, we described the overall structure of the literature graph. Next, we discuss how we populate paper nodes, author nodes, authorship edges, and citation edges.
${^{1}\text{Due to spac}}$e constraints, we opted not to discuss our
relation extraction models in this draft.
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.5341796875,
0.0748291015625,
0.88330078125,
0.07659912109375,
0.88330078125,
0.0914306640625,
0.5341796875,
0.08905029296875
],
[
0.12005615234375,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51123046875,
0.0926513671875,
0.8818359375,
0.0926513671875,
0.8818359375,
0.105712890625,
0.51123046875,
0.105712890625
],
[
0.118408203125,
0.1092529296875,
0.490234375,
0.1092529296875,
0.490234375,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.51123046875,
0.10986328125,
0.88330078125,
0.10986328125,
0.88330078125,
0.12054443359375,
0.51123046875,
0.12054443359375
],
[
0.118408203125,
0.12469482421875,
0.489501953125,
0.12469482421875,
0.489501953125,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.51123046875,
0.12469482421875,
0.8818359375,
0.12469482421875,
0.8818359375,
0.1395263671875,
0.51123046875,
0.1395263671875
],
[
0.118408203125,
0.141357421875,
0.491943359375,
0.141357421875,
0.491943359375,
0.155029296875,
0.118408203125,
0.155029296875
],
[
0.51318359375,
0.14013671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.118408203125,
0.15673828125,
0.490234375,
0.15673828125,
0.490234375,
0.17041015625,
0.118408203125,
0.17041015625
],
[
0.51318359375,
0.15673828125,
0.66650390625,
0.15673828125,
0.66650390625,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.118408203125,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.5341796875,
0.17333984375,
0.88330078125,
0.17333984375,
0.88330078125,
0.188232421875,
0.5341796875,
0.188232421875
],
[
0.1175537109375,
0.188232421875,
0.48779296875,
0.188232421875,
0.48779296875,
0.201904296875,
0.1175537109375,
0.201904296875
],
[
0.51123046875,
0.190673828125,
0.88671875,
0.190673828125,
0.88671875,
0.2037353515625,
0.51123046875,
0.2037353515625
],
[
0.1175537109375,
0.2037353515625,
0.490234375,
0.204833984375,
0.490234375,
0.2197265625,
0.1175537109375,
0.218505859375
],
[
0.51318359375,
0.207275390625,
0.88671875,
0.207275390625,
0.88671875,
0.2203369140625,
0.51318359375,
0.2203369140625
],
[
0.12005615234375,
0.22265625,
0.1695556640625,
0.22265625,
0.1695556640625,
0.2340087890625,
0.12005615234375,
0.2340087890625
],
[
0.51318359375,
0.22265625,
0.884765625,
0.22265625,
0.884765625,
0.236328125,
0.51318359375,
0.236328125
],
[
0.1351318359375,
0.2369384765625,
0.490234375,
0.2381591796875,
0.490234375,
0.2529296875,
0.1351318359375,
0.251708984375
],
[
0.509765625,
0.236328125,
0.88671875,
0.2369384765625,
0.88671875,
0.254638671875,
0.509765625,
0.254150390625
],
[
0.118408203125,
0.254638671875,
0.49365234375,
0.254638671875,
0.49365234375,
0.268310546875,
0.118408203125,
0.268310546875
],
[
0.51123046875,
0.254638671875,
0.5751953125,
0.254638671875,
0.5751953125,
0.26953125,
0.51123046875,
0.26953125
],
[
0.118408203125,
0.271484375,
0.490234375,
0.271484375,
0.490234375,
0.28515625,
0.118408203125,
0.28515625
],
[
0.51318359375,
0.28076171875,
0.6484375,
0.28076171875,
0.6484375,
0.295654296875,
0.51318359375,
0.295654296875
],
[
0.118408203125,
0.287353515625,
0.489501953125,
0.2861328125,
0.489501953125,
0.300537109375,
0.118408203125,
0.3017578125
],
[
0.12005615234375,
0.303955078125,
0.489501953125,
0.303955078125,
0.489501953125,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.53564453125,
0.300537109375,
0.88330078125,
0.302734375,
0.88330078125,
0.317626953125,
0.53564453125,
0.3154296875
],
[
0.118408203125,
0.318359375,
0.48779296875,
0.318359375,
0.48779296875,
0.33203125,
0.118408203125,
0.33203125
],
[
0.51318359375,
0.319580078125,
0.884765625,
0.319580078125,
0.884765625,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.12176513671875,
0.336181640625,
0.48779296875,
0.336181640625,
0.48779296875,
0.3466796875,
0.12176513671875,
0.3466796875
],
[
0.51318359375,
0.3349609375,
0.884765625,
0.3349609375,
0.884765625,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.1175537109375,
0.350830078125,
0.4306640625,
0.3515625,
0.4306640625,
0.366455078125,
0.1175537109375,
0.365234375
],
[
0.51123046875,
0.350830078125,
0.88330078125,
0.349853515625,
0.88330078125,
0.364013671875,
0.51123046875,
0.365234375
],
[
0.51318359375,
0.36767578125,
0.88330078125,
0.36767578125,
0.88330078125,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.1175537109375,
0.383056640625,
0.44921875,
0.38427734375,
0.44921875,
0.3984375,
0.1175537109375,
0.39794921875
],
[
0.51123046875,
0.3818359375,
0.88330078125,
0.3818359375,
0.88330078125,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.5146484375,
0.40087890625,
0.88330078125,
0.40087890625,
0.88330078125,
0.410888671875,
0.5146484375,
0.410888671875
],
[
0.1175537109375,
0.413330078125,
0.490234375,
0.413330078125,
0.490234375,
0.427490234375,
0.1175537109375,
0.427490234375
],
[
0.51318359375,
0.4150390625,
0.791015625,
0.4150390625,
0.791015625,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.1158447265625,
0.4287109375,
0.491943359375,
0.429931640625,
0.491943359375,
0.44482421875,
0.1158447265625,
0.443603515625
],
[
0.5322265625,
0.432861328125,
0.884765625,
0.432861328125,
0.884765625,
0.446533203125,
0.5322265625,
0.446533203125
],
[
0.118408203125,
0.446533203125,
0.489501953125,
0.446533203125,
0.489501953125,
0.460205078125,
0.118408203125,
0.460205078125
],
[
0.51318359375,
0.449462890625,
0.88330078125,
0.449462890625,
0.88330078125,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.12005615234375,
0.463134765625,
0.489501953125,
0.463134765625,
0.489501953125,
0.476806640625,
0.12005615234375,
0.476806640625
],
[
0.515625,
0.468505859375,
0.53662109375,
0.468505859375,
0.53662109375,
0.4755859375,
0.515625,
0.4755859375
],
[
0.5322265625,
0.465576171875,
0.7900390625,
0.465576171875,
0.7900390625,
0.478515625,
0.5322265625,
0.478515625
],
[
0.1175537109375,
0.478515625,
0.489501953125,
0.47802734375,
0.489501953125,
0.4921875,
0.1175537109375,
0.493408203125
],
[
0.52880859375,
0.482177734375,
0.884765625,
0.482177734375,
0.884765625,
0.496337890625,
0.52880859375,
0.496337890625
],
[
0.1175537109375,
0.495361328125,
0.489501953125,
0.49462890625,
0.489501953125,
0.50830078125,
0.1175537109375,
0.5087890625
],
[
0.51318359375,
0.498779296875,
0.884765625,
0.498779296875,
0.884765625,
0.5126953125,
0.51318359375,
0.5126953125
],
[
0.118408203125,
0.51123046875,
0.490234375,
0.51123046875,
0.490234375,
0.52490234375,
0.118408203125,
0.52490234375
],
[
0.51318359375,
0.515625,
0.64208984375,
0.515625,
0.64208984375,
0.529296875,
0.51318359375,
0.529296875
],
[
0.1175537109375,
0.525390625,
0.393798828125,
0.52685546875,
0.393798828125,
0.54150390625,
0.1175537109375,
0.54052734375
],
[
0.53662109375,
0.5322265625,
0.88330078125,
0.5322265625,
0.88330078125,
0.5458984375,
0.53662109375,
0.5458984375
],
[
0.51318359375,
0.548828125,
0.884765625,
0.548828125,
0.884765625,
0.5625,
0.51318359375,
0.5625
],
[
0.1175537109375,
0.55712890625,
0.25439453125,
0.55859375,
0.25439453125,
0.572265625,
0.1175537109375,
0.57080078125
],
[
0.51318359375,
0.56396484375,
0.88330078125,
0.56396484375,
0.88330078125,
0.57763671875,
0.51318359375,
0.57763671875
],
[
0.145263671875,
0.58203125,
0.490234375,
0.58203125,
0.490234375,
0.5966796875,
0.145263671875,
0.5966796875
],
[
0.51318359375,
0.58056640625,
0.88671875,
0.58056640625,
0.88671875,
0.59423828125,
0.51318359375,
0.59423828125
],
[
0.118408203125,
0.5986328125,
0.491943359375,
0.5986328125,
0.491943359375,
0.61328125,
0.118408203125,
0.61328125
],
[
0.51318359375,
0.5966796875,
0.88671875,
0.5966796875,
0.88671875,
0.60986328125,
0.51318359375,
0.60986328125
],
[
0.1175537109375,
0.6142578125,
0.49365234375,
0.61328125,
0.49365234375,
0.62744140625,
0.1175537109375,
0.62890625
],
[
0.51318359375,
0.6123046875,
0.88330078125,
0.6123046875,
0.88330078125,
0.6259765625,
0.51318359375,
0.6259765625
],
[
0.1175537109375,
0.6298828125,
0.49365234375,
0.63134765625,
0.49365234375,
0.6455078125,
0.1175537109375,
0.64453125
],
[
0.51123046875,
0.62890625,
0.884765625,
0.62890625,
0.884765625,
0.642578125,
0.51123046875,
0.642578125
],
[
0.1175537109375,
0.646484375,
0.490234375,
0.646484375,
0.490234375,
0.66015625,
0.1175537109375,
0.66015625
],
[
0.5146484375,
0.646484375,
0.60107421875,
0.646484375,
0.60107421875,
0.65771484375,
0.5146484375,
0.65771484375
],
[
0.52880859375,
0.662109375,
0.71728515625,
0.662109375,
0.71728515625,
0.67578125,
0.52880859375,
0.67578125
],
[
0.765625,
0.662109375,
0.88671875,
0.662109375,
0.88671875,
0.67578125,
0.765625,
0.67578125
],
[
0.51318359375,
0.677734375,
0.88671875,
0.677734375,
0.88671875,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.118408203125,
0.6630859375,
0.490234375,
0.6630859375,
0.490234375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.12005615234375,
0.6796875,
0.49365234375,
0.6796875,
0.49365234375,
0.693359375,
0.12005615234375,
0.693359375
],
[
0.12005615234375,
0.6953125,
0.49365234375,
0.6953125,
0.49365234375,
0.708984375,
0.12005615234375,
0.708984375
],
[
0.51318359375,
0.69482421875,
0.884765625,
0.69482421875,
0.884765625,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.12176513671875,
0.71337890625,
0.490234375,
0.71337890625,
0.490234375,
0.7236328125,
0.12176513671875,
0.7236328125
],
[
0.51123046875,
0.71142578125,
0.88330078125,
0.71044921875,
0.88330078125,
0.7236328125,
0.51123046875,
0.724609375
],
[
0.1175537109375,
0.72705078125,
0.491943359375,
0.7255859375,
0.491943359375,
0.74072265625,
0.1175537109375,
0.7412109375
],
[
0.51123046875,
0.7255859375,
0.8818359375,
0.7255859375,
0.8818359375,
0.7392578125,
0.51123046875,
0.7392578125
],
[
0.1175537109375,
0.74365234375,
0.490234375,
0.7421875,
0.490234375,
0.75732421875,
0.1175537109375,
0.75830078125
],
[
0.51318359375,
0.7421875,
0.884765625,
0.7421875,
0.884765625,
0.755859375,
0.51318359375,
0.755859375
],
[
0.12005615234375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.12005615234375,
0.77392578125
],
[
0.51318359375,
0.7587890625,
0.5634765625,
0.7587890625,
0.5634765625,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.118408203125,
0.7890625
],
[
0.515625,
0.7890625,
0.53662109375,
0.7890625,
0.53662109375,
0.79736328125,
0.515625,
0.79736328125
],
[
0.53662109375,
0.7861328125,
0.724609375,
0.7861328125,
0.724609375,
0.80029296875,
0.53662109375,
0.80029296875
],
[
0.118408203125,
0.80859375,
0.489501953125,
0.80859375,
0.489501953125,
0.822265625,
0.118408203125,
0.822265625
],
[
0.118408203125,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51318359375,
0.8125,
0.88330078125,
0.8125,
0.88330078125,
0.82666015625,
0.51318359375,
0.82666015625
],
[
0.1175537109375,
0.82470703125,
0.25439453125,
0.82470703125,
0.25439453125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51318359375,
0.8291015625,
0.884765625,
0.8291015625,
0.884765625,
0.8427734375,
0.51318359375,
0.8427734375
],
[
0.1368408203125,
0.845703125,
0.490234375,
0.8466796875,
0.490234375,
0.86083984375,
0.1368408203125,
0.85986328125
],
[
0.51318359375,
0.845703125,
0.8818359375,
0.845703125,
0.8818359375,
0.859375,
0.51318359375,
0.859375
],
[
0.118408203125,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.86083984375,
0.76904296875,
0.86083984375,
0.76904296875,
0.87451171875,
0.51318359375,
0.87451171875
],
[
0.1175537109375,
0.87890625,
0.48779296875,
0.87890625,
0.48779296875,
0.8935546875,
0.1175537109375,
0.8935546875
],
[
0.5341796875,
0.8818359375,
0.884765625,
0.8828125,
0.884765625,
0.89794921875,
0.5341796875,
0.896484375
],
[
0.1175537109375,
0.8935546875,
0.255126953125,
0.8955078125,
0.255126953125,
0.91015625,
0.1175537109375,
0.90771484375
],
[
0.5146484375,
0.89794921875,
0.748046875,
0.89794921875,
0.748046875,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"2016), and assume that entity types in the test set",
"Entities.Each node of this type represents a",
"match those labeled in the training set (including",
"unique scientific concept discussed in the literature",
"work on domain adaptation, e.g., Daumé, 2007).",
"withattributessuchas\"canonicalname',aliases",
"These assumptions, while useful for developing",
" and 'description. Our literature graph has 0.4M",
"and benchmarking new methods, are unrealistic for",
"nodes of this type. We describe how we populate",
"many domains and applications. The paper also",
"entity nodes in $4.3.",
"serves as an overview of the approach we adopt",
"Entity mentions. Each node of this type rep-",
"at www.semanticscholar.org in a step towards",
"resents a textual reference of an entity in one of",
"more intelligent academic search engines (Etzioni,",
"the papers, with attributes such as 'mention text',",
"2011).",
"‘context', and confidence'. We describe how we",
"In the next section, we start by describing our",
"populate the 237M mentions in the literature graph",
"symbolic representation of the literature. Then, we",
"in $4.1.",
"discuss how we extract metadata associated with a",
"2.2 Edge Types",
"paper such as authors and references, then how we",
"extract the entities mentioned in paper text. Before",
"Citations. We instantiate a directed citation",
"we conclude, we briefly describe other research",
"edgefrom paper nodes pi p2 for each p2 refer-",
"challenges we are actively working on in order to",
"enced in pi. Citation edges have attributes such as",
"improve the quality of the literature graph.",
"\"from paper id', ‘to paper id' and ‘contexts\" (the tex-",
"tual contexts where p2 is referenced in pi). While",
"2Structure of The Literature Graph",
" some of the paper sources provide these attributes",
"asmetadataitisoftennecessarytoextractthem",
"The literature graph is a property graph with di-",
"from the paper PDF as detailed in s3.",
"rected edges. Unlike Resource Description Frame-",
"Authorship.We instantiate a directed author-",
"work (RDF) graphs, nodes and edges in property",
"ship edge between an author node and a paper node",
"graphs have an internal structure which is more",
"a-",
"→ p for each author of that paper.",
"suitable for representing complex data types such",
"Entity linking edges.We instantiate a directed",
"as papers and entities. In this section, we describe",
"edge from an extracted entity mention node to the",
"the attributes associated with nodes and edges of",
"entity it refers to.",
"different types in the literature graph.",
"Mention-mention relations.We instantiate",
"a directed edge between a pair of mentions in the",
"2.1 Node Types",
"same sentential context if the textual relation ex-",
"Papers. We obtain metadata and PDF files",
"traction model predicts one of a predefined list of",
"of papers via partnerships with publishers (e.g.,",
"relation types between them in a sentential con-",
"Springer, Nature), catalogs (e.g., DBLP, MED-",
"text.1 We encode a symmetric relation between",
"LINE), pre-publishing services (e.g., arXiv, bioRx-",
"mi and m2 as two directed edges mi m2 and",
"ive), as well as web-crawling.Paper nodes are",
"m2 - mi.",
"Entity-entity relations.",
"While mention-",
"mention edges represent relations between men-",
"associated with a set of attributes such as ‘title', 'ab-",
"stract', ‘full text', venues' and 'publication year.",
"While some of the paper sources provide these at-",
"tions in a particular context, entity-entity edges",
"tributesasmetadata,itis oftennecessarytoextract",
"represent relations between abstract entities. These",
"them from the paper PDF (details in s3). We de-",
"relations may be imported from an existing knowl",
"terministically remove duplicate papers based on",
"edge base (KB) or inferred from other edges in the",
"string similarity of their metadata, resulting in 37M",
"graph.",
"unique paper nodes. Papers in the literature graph",
"3",
"Extracting Metadata",
"computer science, molecular biology, microbiology",
"cover a variety of scientific disciplines, including",
"In the previous section, we described the overall",
"and neuroscience.",
"structure of the literature graph. Next, we discuss",
"Authors.Each node of this type represents a",
"how we populate paper nodes, author nodes, au-",
"unique author, with attributes such as 'first name'",
"thorship edges, and citation edges.",
"and ‘last name'. The literature graph has 12M",
"'Due to space constraints, we opted not to discuss our",
"nodes of this type.",
"relation extraction models in this draft."
] |
{
"arxiv_id": "1805.02262",
"language": "en",
"page_idx": 1
}
|
||
1805.02262-en-3
|
1805.02262-en-3.jpg
|
documents. The results are detailed in Table 1. We use a conservative evaluation where an instance is correct if it exactly matches the gold annotation, with no credit for partial matching.
To give an example for the type of errors our model makes, consider the paper (Wang et al., 2013) titled “Clinical review: Efficacy of antimicrobial-impregnated catheters in external ventricular drainage- a systematic review and metaanalysis." The title we extract for this paper omits the first part “Clinical review:". This is likely to be a result of the pattern “Foo: Bar Baz" appearing in many training examples with only “Bar Baz” labeled as the title.
# 4 Entity Extraction and Linking
In the previous section, we described how we populate the backbone of the literature graph, i.e., paper nodes, author nodes and citation edges. Next, we discuss how we populate mentions and entities in the literature graph using entity extraction and linking on the paper text. In order to focus on more salient entities in a given paper, we only use the title and abstract.
# 4.1 Approaches
## We experiment with three approaches for entity extraction and linking:
I. Statistical: uses one or more statistical models for predicting mention spans, then uses another statistical model to link mentions to candidate entities in a KB.
II. Hybrid: defines a small number of handengineered, deterministic rules for string-based matching of the input text to candidate entities in the KB, then uses a statistical model to disambiguate the mentions.$^{4}$
III. Off-the-shelf: uses existing libraries, namely (Ferragina and Scaiella, 2010, TagMe)$^{5}$ and
(Demner-Fushman et al., 2017, MetaMap Lite)$^6$, with minimal post-processing to extract and link entities to the KB
We evaluate the performance of each approach in two broad scientific areas: computer science (CS)
$^{4}$We also experimented with a“pure”rules-based approach which disambiguates deterministically but the hybrid approach consistently gave better results.
$^{5}$The TagMe APIs are described at https://sobigdata.
d4science.org/web/tagme/tagme-help
$^{6}$We use v3.4 (L0) of MetaMap Lite, available at https:
//metamap.nlm.nih.gov/MetaMapLite.shtml
<table>
<tbody>
<tr>
<th rowspan="2">Approach</th>
<th> </th>
<th>$S$</th>
<th colspan="2">Bio</th>
</tr>
<tr>
<th>$prec.$</th>
<th>vield</th>
<th>$prec.$</th>
<th>yield</th>
</tr>
<tr>
<td>Statistical</td>
<td>98.4</td>
<td>712</td>
<td>94.4</td>
<td>928</td>
</tr>
<tr>
<td>Hybrid</td>
<td>91.5</td>
<td>1990</td>
<td>92.1</td>
<td>3126</td>
</tr>
<tr>
<td>Off- the- shelf</td>
<td>97.4</td>
<td>873</td>
<td>77.5</td>
<td>1206</td>
</tr>
</tbody>
</table>
Table 2: Document-level evaluation of three approaches in two scientific areas: computer science (CS) and biomedical (Bio).
and biomedical research (Bio). For each unique (paper ID, entity ID) pair predicted by one of the approaches, we ask human annotators to label each mention extracted for this entity in the paper. We use CrowdFlower to manage human annotations and only include instances where three or more annotators agree on the label. If one or more of the entity mentions in that paper is judged to be correct, the pair (paper ID, entity ID) counts as one correct instance. Otherwise, it counts as an incorrect instance. We report 'yield’in lieu of `recall' due to the difficulty of doing a scalable comprehensive annotation.
Table 2 shows the results based on 500 papers using v1.1.2 of our entity extraction and linking components. In both domains, the statistical approach gives the highest precision and the lowest yield. The hybrid approach consistently gives the highest yield, but sacrifices precision. The TagMe off-the-shelf library used for the CS domain gives surprisingly good results, with precision within 1 point from the statistical models. However, the MetaMap Lite off-the-shelf library we used for the biomedical domain suffered a huge loss in precision. Our error analysis showed that each of the approaches is able to predict entities not predicted by the other approaches so we decided to pool their outputs in our deployed system, which gives significantly higher yield than any individual approach while maintaining reasonably high precision.
### 4.2 Entity Extraction Models
Given the token sequence $t_1,\ldots,t_N$ in a sentence, we need to identify spans which correspond to entity mentions. We use the BIILOU scheme to encode labels at the token level. Unlike most formulations of named entity recognition problems (NER), we do not identify the entity type (e.g., protein, drug, chemical, disease) for each mention since the output mentions are further grounded in a KB with
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.5576171875,
0.07781982421875,
0.6337890625,
0.07781982421875,
0.6337890625,
0.0926513671875,
0.5576171875,
0.0926513671875
],
[
0.68701171875,
0.0789794921875,
0.7138671875,
0.0789794921875,
0.7138671875,
0.0926513671875,
0.68701171875,
0.0926513671875
],
[
0.791015625,
0.07781982421875,
0.8251953125,
0.07781982421875,
0.8251953125,
0.09320068359375,
0.791015625,
0.09320068359375
],
[
0.118408203125,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.118408203125,
0.10687255859375
],
[
0.70361328125,
0.0955810546875,
0.75,
0.0955810546875,
0.75,
0.10986328125,
0.70361328125,
0.10986328125
],
[
0.76318359375,
0.0968017578125,
0.8076171875,
0.0968017578125,
0.8076171875,
0.11102294921875,
0.76318359375,
0.11102294921875
],
[
0.8154296875,
0.0955810546875,
0.8623046875,
0.09320068359375,
0.86376953125,
0.1092529296875,
0.81689453125,
0.11102294921875
],
[
0.1175537109375,
0.1092529296875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.122314453125,
0.1175537109375,
0.122314453125
],
[
0.6513671875,
0.097412109375,
0.68994140625,
0.097412109375,
0.68994140625,
0.1092529296875,
0.6513671875,
0.1092529296875
],
[
0.64990234375,
0.11639404296875,
0.69189453125,
0.11639404296875,
0.69189453125,
0.1318359375,
0.64990234375,
0.1318359375
],
[
0.7138671875,
0.11639404296875,
0.75,
0.11639404296875,
0.75,
0.1318359375,
0.7138671875,
0.1318359375
],
[
0.76318359375,
0.11639404296875,
0.80419921875,
0.11639404296875,
0.80419921875,
0.1318359375,
0.76318359375,
0.1318359375
],
[
0.82861328125,
0.11639404296875,
0.86376953125,
0.11639404296875,
0.86376953125,
0.1318359375,
0.82861328125,
0.1318359375
],
[
0.1158447265625,
0.12353515625,
0.376220703125,
0.12469482421875,
0.376220703125,
0.1395263671875,
0.1158447265625,
0.1383056640625
],
[
0.5576171875,
0.1175537109375,
0.634765625,
0.1175537109375,
0.634765625,
0.1318359375,
0.5576171875,
0.1318359375
],
[
0.7080078125,
0.1318359375,
0.75,
0.1318359375,
0.75,
0.1478271484375,
0.7080078125,
0.1478271484375
],
[
0.76416015625,
0.1312255859375,
0.80419921875,
0.1341552734375,
0.8017578125,
0.1484375,
0.76318359375,
0.146728515625
],
[
0.64990234375,
0.133056640625,
0.69189453125,
0.133056640625,
0.69189453125,
0.1484375,
0.64990234375,
0.1484375
],
[
0.57666015625,
0.133056640625,
0.634765625,
0.133056640625,
0.634765625,
0.1484375,
0.57666015625,
0.1484375
],
[
0.8193359375,
0.133056640625,
0.86376953125,
0.133056640625,
0.86376953125,
0.1484375,
0.8193359375,
0.1484375
],
[
0.1351318359375,
0.14013671875,
0.489501953125,
0.141357421875,
0.489501953125,
0.1561279296875,
0.1351318359375,
0.155029296875
],
[
0.7138671875,
0.1484375,
0.75,
0.1484375,
0.75,
0.1644287109375,
0.7138671875,
0.1644287109375
],
[
0.76416015625,
0.1484375,
0.80419921875,
0.1484375,
0.80419921875,
0.1644287109375,
0.76416015625,
0.1644287109375
],
[
0.6513671875,
0.149658203125,
0.68994140625,
0.149658203125,
0.68994140625,
0.162109375,
0.6513671875,
0.162109375
],
[
0.5380859375,
0.15087890625,
0.634765625,
0.15087890625,
0.634765625,
0.162109375,
0.5380859375,
0.162109375
],
[
0.82275390625,
0.15087890625,
0.861328125,
0.15087890625,
0.861328125,
0.163330078125,
0.82275390625,
0.163330078125
],
[
0.118408203125,
0.15673828125,
0.489501953125,
0.15673828125,
0.489501953125,
0.171630859375,
0.118408203125,
0.171630859375
],
[
0.118408203125,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.51123046875,
0.179931640625,
0.88330078125,
0.18115234375,
0.88330078125,
0.1953125,
0.51123046875,
0.19482421875
],
[
0.118408203125,
0.190673828125,
0.489501953125,
0.190673828125,
0.489501953125,
0.2037353515625,
0.118408203125,
0.2037353515625
],
[
0.51318359375,
0.19775390625,
0.884765625,
0.19775390625,
0.884765625,
0.21142578125,
0.51318359375,
0.21142578125
],
[
0.1175537109375,
0.2060546875,
0.490234375,
0.2060546875,
0.490234375,
0.2197265625,
0.1175537109375,
0.2197265625
],
[
0.51318359375,
0.2120361328125,
0.71533203125,
0.213134765625,
0.71533203125,
0.22802734375,
0.51318359375,
0.226806640625
],
[
0.1175537109375,
0.2203369140625,
0.489501953125,
0.221435546875,
0.489501953125,
0.236328125,
0.1175537109375,
0.235107421875
],
[
0.1175537109375,
0.2369384765625,
0.489501953125,
0.236328125,
0.489501953125,
0.25048828125,
0.1175537109375,
0.251708984375
],
[
0.118408203125,
0.254150390625,
0.490234375,
0.254150390625,
0.490234375,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.51318359375,
0.25830078125,
0.88330078125,
0.25830078125,
0.88330078125,
0.271484375,
0.51318359375,
0.271484375
],
[
0.118408203125,
0.26953125,
0.490234375,
0.268310546875,
0.490234375,
0.283203125,
0.118408203125,
0.283935546875
],
[
0.5146484375,
0.27490234375,
0.884765625,
0.27490234375,
0.884765625,
0.2880859375,
0.5146484375,
0.2880859375
],
[
0.1175537109375,
0.28515625,
0.2568359375,
0.28515625,
0.2568359375,
0.298583984375,
0.1175537109375,
0.298583984375
],
[
0.51318359375,
0.29150390625,
0.884765625,
0.29150390625,
0.884765625,
0.3046875,
0.51318359375,
0.3046875
],
[
0.51318359375,
0.30712890625,
0.884765625,
0.30712890625,
0.884765625,
0.320556640625,
0.51318359375,
0.320556640625
],
[
0.1175537109375,
0.3134765625,
0.4130859375,
0.3154296875,
0.4130859375,
0.330078125,
0.1175537109375,
0.327880859375
],
[
0.51318359375,
0.322509765625,
0.884765625,
0.322509765625,
0.884765625,
0.336181640625,
0.51318359375,
0.336181640625
],
[
0.1175537109375,
0.34033203125,
0.49365234375,
0.341552734375,
0.49365234375,
0.355712890625,
0.1175537109375,
0.355224609375
],
[
0.51318359375,
0.339111328125,
0.884765625,
0.339111328125,
0.884765625,
0.352783203125,
0.51318359375,
0.352783203125
],
[
0.118408203125,
0.358154296875,
0.491943359375,
0.358154296875,
0.491943359375,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51318359375,
0.355224609375,
0.88671875,
0.355224609375,
0.88671875,
0.3681640625,
0.51318359375,
0.3681640625
],
[
0.118408203125,
0.37353515625,
0.490234375,
0.37353515625,
0.490234375,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.51318359375,
0.371826171875,
0.884765625,
0.371826171875,
0.884765625,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.118408203125,
0.39013671875,
0.490234375,
0.39013671875,
0.490234375,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.51123046875,
0.385986328125,
0.88330078125,
0.385986328125,
0.88330078125,
0.399658203125,
0.51123046875,
0.399658203125
],
[
0.1175537109375,
0.40673828125,
0.491943359375,
0.40673828125,
0.491943359375,
0.42041015625,
0.1175537109375,
0.42041015625
],
[
0.51123046875,
0.402587890625,
0.884765625,
0.402099609375,
0.884765625,
0.416259765625,
0.51123046875,
0.41748046875
],
[
0.1175537109375,
0.4228515625,
0.491943359375,
0.4228515625,
0.491943359375,
0.435791015625,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.419189453125,
0.884765625,
0.419189453125,
0.884765625,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.1175537109375,
0.438232421875,
0.49365234375,
0.43701171875,
0.49365234375,
0.451904296875,
0.1175537109375,
0.453125
],
[
0.51123046875,
0.435302734375,
0.8818359375,
0.435302734375,
0.8818359375,
0.448974609375,
0.51123046875,
0.448974609375
],
[
0.1175537109375,
0.454833984375,
0.2476806640625,
0.454833984375,
0.2476806640625,
0.468505859375,
0.1175537109375,
0.468505859375
],
[
0.51318359375,
0.451904296875,
0.7138671875,
0.451904296875,
0.7138671875,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.5283203125,
0.46728515625,
0.88330078125,
0.468505859375,
0.88330078125,
0.482666015625,
0.5283203125,
0.482177734375
],
[
0.118408203125,
0.482177734375,
0.2568359375,
0.482177734375,
0.2568359375,
0.496337890625,
0.118408203125,
0.496337890625
],
[
0.51123046875,
0.48388671875,
0.88330078125,
0.485107421875,
0.88330078125,
0.49951171875,
0.51123046875,
0.498779296875
],
[
0.118408203125,
0.501953125,
0.489501953125,
0.5029296875,
0.489501953125,
0.51708984375,
0.118408203125,
0.5166015625
],
[
0.51123046875,
0.501953125,
0.8818359375,
0.501953125,
0.8818359375,
0.515625,
0.51123046875,
0.515625
],
[
0.12005615234375,
0.52099609375,
0.28955078125,
0.52099609375,
0.28955078125,
0.53369140625,
0.12005615234375,
0.53369140625
],
[
0.51318359375,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.51318359375,
0.53369140625,
0.88330078125,
0.53369140625,
0.88330078125,
0.54736328125,
0.51318359375,
0.54736328125
],
[
0.12005615234375,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5546875,
0.12005615234375,
0.5546875
],
[
0.51318359375,
0.5498046875,
0.88330078125,
0.5498046875,
0.88330078125,
0.5634765625,
0.51318359375,
0.5634765625
],
[
0.12005615234375,
0.55859375,
0.489501953125,
0.55859375,
0.489501953125,
0.5693359375,
0.12005615234375,
0.5693359375
],
[
0.51318359375,
0.5654296875,
0.88330078125,
0.5654296875,
0.88330078125,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.1175537109375,
0.572265625,
0.48779296875,
0.572265625,
0.48779296875,
0.5859375,
0.1175537109375,
0.5859375
],
[
0.5146484375,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.5146484375,
0.595703125
],
[
0.1158447265625,
0.5888671875,
0.183837890625,
0.58740234375,
0.183837890625,
0.6015625,
0.1158447265625,
0.60400390625
],
[
0.51318359375,
0.59716796875,
0.884765625,
0.59716796875,
0.884765625,
0.61083984375,
0.51318359375,
0.61083984375
],
[
0.118408203125,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.51123046875,
0.6123046875,
0.884765625,
0.61328125,
0.884765625,
0.62744140625,
0.51123046875,
0.626953125
],
[
0.118408203125,
0.62744140625,
0.489501953125,
0.62744140625,
0.489501953125,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.51123046875,
0.62890625,
0.88671875,
0.6298828125,
0.88671875,
0.64453125,
0.51123046875,
0.6435546875
],
[
0.118408203125,
0.6435546875,
0.490234375,
0.6435546875,
0.490234375,
0.65673828125,
0.118408203125,
0.65673828125
],
[
0.51318359375,
0.646484375,
0.884765625,
0.646484375,
0.884765625,
0.66015625,
0.51318359375,
0.66015625
],
[
0.118408203125,
0.66015625,
0.490234375,
0.66015625,
0.490234375,
0.67333984375,
0.118408203125,
0.67333984375
],
[
0.51318359375,
0.662109375,
0.884765625,
0.6611328125,
0.884765625,
0.67578125,
0.51318359375,
0.6767578125
],
[
0.12005615234375,
0.67578125,
0.287109375,
0.67578125,
0.287109375,
0.689453125,
0.12005615234375,
0.689453125
],
[
0.51123046875,
0.677734375,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.51123046875,
0.69140625
],
[
0.118408203125,
0.6953125,
0.48779296875,
0.69677734375,
0.48779296875,
0.71142578125,
0.118408203125,
0.71044921875
],
[
0.51318359375,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.1175537109375,
0.71337890625,
0.489501953125,
0.7119140625,
0.489501953125,
0.72705078125,
0.1175537109375,
0.72802734375
],
[
0.51123046875,
0.708984375,
0.88330078125,
0.71044921875,
0.88330078125,
0.724609375,
0.51123046875,
0.7236328125
],
[
0.118408203125,
0.72802734375,
0.489501953125,
0.72802734375,
0.489501953125,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.51318359375,
0.72705078125,
0.84619140625,
0.72705078125,
0.84619140625,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.1175537109375,
0.74658203125,
0.489501953125,
0.74658203125,
0.489501953125,
0.76025390625,
0.1175537109375,
0.76025390625
],
[
0.118408203125,
0.7626953125,
0.255126953125,
0.7626953125,
0.255126953125,
0.775390625,
0.118408203125,
0.775390625
],
[
0.5146484375,
0.75830078125,
0.75732421875,
0.75830078125,
0.75732421875,
0.771484375,
0.5146484375,
0.771484375
],
[
0.1351318359375,
0.7822265625,
0.489501953125,
0.78369140625,
0.489501953125,
0.79736328125,
0.1351318359375,
0.7958984375
],
[
0.51318359375,
0.78076171875,
0.88671875,
0.7822265625,
0.88671875,
0.79638671875,
0.51318359375,
0.7958984375
],
[
0.1175537109375,
0.7998046875,
0.490234375,
0.7998046875,
0.490234375,
0.8134765625,
0.1175537109375,
0.8134765625
],
[
0.51318359375,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51318359375,
0.8125
],
[
0.51318359375,
0.81396484375,
0.88671875,
0.81396484375,
0.88671875,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.1368408203125,
0.82373046875,
0.490234375,
0.82373046875,
0.490234375,
0.83740234375,
0.1368408203125,
0.83740234375
],
[
0.1175537109375,
0.8359375,
0.489501953125,
0.83740234375,
0.489501953125,
0.8486328125,
0.1175537109375,
0.84716796875
],
[
0.51318359375,
0.83056640625,
0.88671875,
0.83056640625,
0.88671875,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.118408203125,
0.849609375,
0.307373046875,
0.849609375,
0.307373046875,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51318359375,
0.84716796875,
0.88671875,
0.84716796875,
0.88671875,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.1376953125,
0.85986328125,
0.49365234375,
0.85986328125,
0.49365234375,
0.87353515625,
0.1376953125,
0.87353515625
],
[
0.51318359375,
0.86328125,
0.88671875,
0.86328125,
0.88671875,
0.87744140625,
0.51318359375,
0.87744140625
],
[
0.1175537109375,
0.87060546875,
0.393798828125,
0.87158203125,
0.393798828125,
0.88525390625,
0.1175537109375,
0.88427734375
],
[
0.14111328125,
0.88427734375,
0.495361328125,
0.88427734375,
0.495361328125,
0.89794921875,
0.14111328125,
0.89794921875
],
[
0.51318359375,
0.8798828125,
0.884765625,
0.8798828125,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.89794921875,
0.42822265625,
0.89794921875,
0.42822265625,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"documents. The results are detailed in Table 1. We",
"Approach",
"CS",
"Bio",
"use a conservative evaluation where an instance is",
"yield",
"prec.",
"yield",
"correct if it exactly matches the gold annotation",
"prec.",
"98.4",
"712",
"94.4",
"928",
"with no credit for partial matching.",
"Statistical",
"1990",
"92.1",
"91.5",
"Hybrid",
"3126",
"To give an example for the type of errors",
"873",
"77.5",
"97.4",
"Off-the-shelf",
"1206",
"our model makes, consider the paper (Wang",
"et al., 2013) titled \"Clinical review: Efficacy of",
"Table 2: Document-level evaluation of three ap",
"antimicrobial-impregnated catheters in external",
"proaches in two scientific areas: computer science",
"ventricular drainage - a systematic review and meta-",
"(CS) and biomedical (Bio).",
"analysis.\" The title we extract for this paper omits",
"the first part \"Clinical review:\" This is likely to",
"be a result of the pattern “Foo: Bar Baz\" appear-",
"and biomedical research (Bio). For each unique",
"ing in many training examples with only “Bar Baz\"",
"(paper ID, entity ID) pair predicted by one of the",
"labeled as the title.",
"approaches, we ask human annotators to label each",
"mention extracted for this entity in the paper. We",
"4Entity Extraction and Linking",
"use CrowdFlower to manage human annotations",
"In the previous section, we described how we popu-",
"and only include instances where three or more",
"late the backbone of the literature graph, i.e., paper",
"annotators agree on the label. If one or more of",
"nodes, author nodes and citation edges. Next, we",
"the entity mentions in that paper is judged to be",
"discuss how we populate mentions and entities in",
"correct, the pair (paper ID, entity ID) counts as",
"the literature graph using entity extraction and link-",
"one correct instance. Otherwise, it counts as an",
"ing on the paper text. In order to focus on more",
"incorrect instance.We reportyield'in lieu of",
"salient entities in a given paper, we only use the",
"'recall' due to the difficulty of doing a scalable",
"title and abstract.",
"comprehensive annotation.",
"Table 2 shows the results based on 500 papers",
"4.1Approaches",
"using v1.1.2 of our entity extraction and linking",
"We experiment with three approaches for entity",
"components. In both domains, the statistical ap.",
"extraction and linking:",
"proach gives the highest precision and the lowest",
"yield. The hybrid approach consistently gives the",
"I. Statistical: uses one or more statistical models",
"highest yield, but sacrifices precision. The TagMe",
"for predicting mention spans, then uses another sta-",
"off-the-shelf library used for the CS domain gives",
"tistical model to link mentions to candidate entities",
"surprisingly good results, with precision within 1",
"in a KB.",
"point from the statistical models. However, the",
"II. Hybrid: defines a small number of hand",
"MetaMap Lite off-the-shelf library we used for the",
"engineered, deterministic rules for string-based",
"biomedical domain suffered a huge loss in preci-",
"matching of the input text to candidate entities",
"sion. Our error analysis showed that each of the",
"in the KB, then uses a statistical model to disam-",
"approaches is able to predict entities not predicted",
"biguate the mentions.4",
"by the other approaches so we decided to pool their",
"III. Off-the-shelf: uses existing libraries, namely",
"outputs in our deployed system, which gives signif",
"(Ferragina and Scaiella, 2010, TagMe)5 and",
"icantly higher yield than any individual approach",
"(Demner-Fushman et al., 2017. MetaMap Lite)6,",
"while maintaining reasonably high precision.",
"with minimal post-processing to extract and link",
"entities to the KB.",
"4.2Entity Extraction Models",
"We evaluate theperformanceof each approach in",
"Given the token sequence ti,..., t in a sentence,",
"two broad scientific areas: computer science (CS)",
"we need to identify spans which correspond to en-",
"tity mentions. We use the BILOU scheme to en-",
"4We also experimented with a “pure\" rules-based approach",
"which disambiguates deterministically but the hybrid approach",
"code labels at the token level. Unlike most formula-",
"consistently gave better results",
"tions of named entity recognition problems (NER),",
"5The TagMe APIs are described at https: //sobigdata.",
"we do not identify the entity type (e.g., protein,",
"d4science.org/web/tagme/tagme-help",
"6We use v3.4 (L0) of MetaMap Lite, available at https:",
"drug, chemical, disease) for each mention since the",
"//metamap.nlm.nih.gov/MetaMapLite.shtml",
"output mentions are further grounded in a KB with"
] |
{
"arxiv_id": "1805.02262",
"language": "en",
"page_idx": 3
}
|
||
1805.02262-en-4
|
1805.02262-en-4.jpg
|
further information about the entity (including its
type), using an entity linking module.
Model. First, we construct the token embed- $\operatorname{ding}\mathbf{x}_{k}=[\mathbf{c}_{k};\mathbf{w}_{k}]$ for each token $t_{k}$ in the input sequence, where c$_k$ is a character-based representation computed using a convolutional neural network (CNN) with flter of size 3 characters, and $\mathbf{w}_k$ are learned word embeddings initialized with the GloVe embeddings (Pennington et al., 2014).
We also compute context-sensitive word embed- $\mathrm{dings,~denoted~as~lm}_{k}=[\mathbf{lm}_{k}^{\rightarrow};\mathbf{lm}_{k}^{\leftarrow}]$,by concatenating the projected outputs of forward and backward recurrent neural network language models $(\mathbf{RNN-LM})$ at position $k.$ The language model (LM) for each direction is trained independently and consists of a single layer long short-term memory (LSTM) network followed by a linear project layer. While training the LM parameters, $\mathbf{m}_k^{\to}$ is used to predict $t_k+1$ and Im$_k^\leftarrow$ is used to predict $t_k-1.$ We fix the LM parameters during training of the entity extraction model. See Peters et al.(2017) and Ammar et al. (2017) for more details.
Given the $\mathbf{x}_k$ and Im$_k$ embeddings for each token $k\in\{1,\ldots,N\}$, we use a two-layer bidirectional LSTM to encode the sequence with $\mathbf{x}_k$ and $\mathbf{lm}_k$ feeding into the first and second layer, respectively. That is,
$$\mathbf{g}_{k}^{\to}=\mathrm{LSTM}(\mathbf{x}_{k},\mathbf{g}_{k-1}^{\to}),\mathbf{g}_{k}=[\mathbf{g}_{k}^{\to};\mathbf{g}_{k}^{\leftarrow}],$$
$\mathbf{h}_{k}^{\rightarrow}=$LSTM$([\mathbf{g}_{k};\mathbf{lm}_{k}],\mathbf{h}_{k-1}^{\rightarrow}),\mathbf{h}_{k}=[\mathbf{h}_{k}^{\rightarrow};\mathbf{h}_{k}^{\leftarrow}]$,
where $\mathbf{g}_k^{\leftarrow}$ and $\mathbf{h}_k^\leftarrow$ are defined similarly to $\mathbf{g}_k^\rightarrow$ and $\mathbf{h}_k^{\to}$ but process token sequences in the opposite direction.
Similar to the model described in §3, we feed the output of the second LSTM into a dense layer to predict unnormalized label weights for each token and learn label bigram feature weights to account for dependencies between labels.
Results. We use the standard data splits of the SemEval-2017 Task 10 on entity (and relation) extraction from scientific papers (Augenstein et al., 2017). Table 3 compares three variants of our entity extraction model. The first line omits the LM embeddings $\mathbf{lm}_k$, while the second line is the full model (including LM embeddings) showing a large improvement of 4.2 F1 points. The third line shows that creating an ensemble of 15 models further improves the results by 1.1 F1 points.
Model instances. In the deployed system, we use three instances of the entity extraction model with a similar architecture, but trained on different datasets. Two instances are trained on the
<table>
<tbody>
<tr>
<th> </th>
<th>Description</th>
<th>$F1$</th>
</tr>
<tr>
<td> </td>
<td>${\mathrm{WithoutLM}}$</td>
<td>49.9</td>
</tr>
<tr>
<td> </td>
<td>$\mathbf{WithLM}$</td>
<td>54.1</td>
</tr>
<tr>
<td>Avg. of 15 mc</td>
<td>models with $1LM$</td>
<td>55.2</td>
</tr>
</tbody>
</table>
Table 3: Results of the entity extraction model on
the development set of SemEval-2017 task 10.
BC5CDR (Li et al., 2016) and the CHEMDNER datasets (Krallinger et al., 2015) to extract key entity mentions in the biomedical domain such as diseases, drugs and chemical compounds. The third instance is trained on mention labels induced from Wikipedia articles in the computer science domain. The output of all model instances are pooled together and combined with the rule-based entity extraction module, then fed into the entity linking model (described below).
## 4.3 Knowledge Bases
In this section, we describe the construction of entity nodes and entity-entity edges. Unlike other knowledge extraction systems such as the NeverEnding Language Learner (NELL)$^{7}$ and OpenIE 4,$^{8}$ we use existing knowledge bases (KBs) of entities to reduce the burden of identifying coherent concepts. Grounding the entity mentions in a manually-curated KB also increases user confidence in automated predictions. We use two KBs: UMLS: The UMLS metathesaurus integrates information about concepts in specialized ontologies in several biomedical domains, and is funded by the U.S. National Library of Medicine.
DBpedia: DBpedia provides access to structured information in Wikipedia. Rather than including all Wikipedia pages, we used a short list of Wikipedia categories about CS and included all pages up to depth four in their trees in order to exclude irrelevant entities, e.g., “Lord of the Rings" in DBpedia.
### 4.4 Entity Linking Models
Given a text span s identified by the entity extraction model in §4.2 (or with heuristics) and a reference KB, the goal of the entity linking model is to associate the span with the entity it refers to. A span and its surrounding words are collectively referred to as a mention. We first identify a set of candidate entities that a given mention may refer
$^7$http://rtw.ml.cmu.edu/rtw/ $^{8}$https://github.com/allenai/
openie-standalone
|
[
[
0.12005615234375,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.68701171875,
0.0789794921875,
0.77734375,
0.0789794921875,
0.77734375,
0.09320068359375,
0.68701171875,
0.09320068359375
],
[
0.79833984375,
0.0789794921875,
0.82373046875,
0.0789794921875,
0.82373046875,
0.0926513671875,
0.79833984375,
0.0926513671875
],
[
0.12005615234375,
0.0926513671875,
0.397216796875,
0.0926513671875,
0.397216796875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.6826171875,
0.1009521484375,
0.77734375,
0.1009521484375,
0.77734375,
0.11517333984375,
0.6826171875,
0.11517333984375
],
[
0.791015625,
0.09979248046875,
0.8330078125,
0.09979248046875,
0.8330078125,
0.11517333984375,
0.791015625,
0.11517333984375
],
[
0.139404296875,
0.10986328125,
0.489501953125,
0.10986328125,
0.489501953125,
0.12469482421875,
0.139404296875,
0.12469482421875
],
[
0.7060546875,
0.1175537109375,
0.77734375,
0.1175537109375,
0.77734375,
0.1318359375,
0.7060546875,
0.1318359375
],
[
0.791015625,
0.1175537109375,
0.82958984375,
0.1175537109375,
0.82958984375,
0.133056640625,
0.791015625,
0.133056640625
],
[
0.1175537109375,
0.1258544921875,
0.489501953125,
0.1270751953125,
0.489501953125,
0.141357421875,
0.1175537109375,
0.14013671875
],
[
0.568359375,
0.1341552734375,
0.77734375,
0.133056640625,
0.77734375,
0.1478271484375,
0.568359375,
0.1484375
],
[
0.791015625,
0.133056640625,
0.82958984375,
0.133056640625,
0.82958984375,
0.149658203125,
0.791015625,
0.149658203125
],
[
0.12005615234375,
0.1436767578125,
0.489501953125,
0.1436767578125,
0.489501953125,
0.15673828125,
0.12005615234375,
0.15673828125
],
[
0.1175537109375,
0.1602783203125,
0.489501953125,
0.157958984375,
0.489501953125,
0.171630859375,
0.1175537109375,
0.17333984375
],
[
0.51318359375,
0.163330078125,
0.88330078125,
0.1644287109375,
0.88330078125,
0.1787109375,
0.51318359375,
0.1776123046875
],
[
0.1175537109375,
0.17333984375,
0.48779296875,
0.174560546875,
0.48779296875,
0.189453125,
0.1175537109375,
0.188232421875
],
[
0.51123046875,
0.179931640625,
0.85400390625,
0.179931640625,
0.85400390625,
0.193603515625,
0.51123046875,
0.193603515625
],
[
0.118408203125,
0.191162109375,
0.489501953125,
0.191162109375,
0.489501953125,
0.204833984375,
0.118408203125,
0.204833984375
],
[
0.12005615234375,
0.207275390625,
0.450927734375,
0.207275390625,
0.450927734375,
0.221435546875,
0.12005615234375,
0.221435546875
],
[
0.1351318359375,
0.223876953125,
0.489501953125,
0.223876953125,
0.489501953125,
0.2369384765625,
0.1351318359375,
0.2369384765625
],
[
0.5146484375,
0.22265625,
0.88330078125,
0.22265625,
0.88330078125,
0.236328125,
0.5146484375,
0.236328125
],
[
0.118408203125,
0.240478515625,
0.49365234375,
0.240478515625,
0.49365234375,
0.254150390625,
0.118408203125,
0.254150390625
],
[
0.51123046875,
0.2381591796875,
0.884765625,
0.2381591796875,
0.884765625,
0.2529296875,
0.51123046875,
0.2529296875
],
[
0.1175537109375,
0.255859375,
0.49365234375,
0.254638671875,
0.49365234375,
0.26953125,
0.1175537109375,
0.270751953125
],
[
0.51123046875,
0.254638671875,
0.884765625,
0.254638671875,
0.884765625,
0.268310546875,
0.51123046875,
0.268310546875
],
[
0.118408203125,
0.2724609375,
0.490234375,
0.2724609375,
0.490234375,
0.2861328125,
0.118408203125,
0.2861328125
],
[
0.51123046875,
0.271484375,
0.884765625,
0.271484375,
0.884765625,
0.28515625,
0.51123046875,
0.28515625
],
[
0.118408203125,
0.2880859375,
0.490234375,
0.2880859375,
0.490234375,
0.3017578125,
0.118408203125,
0.3017578125
],
[
0.51318359375,
0.287353515625,
0.884765625,
0.287353515625,
0.884765625,
0.300537109375,
0.51318359375,
0.300537109375
],
[
0.118408203125,
0.303955078125,
0.489501953125,
0.303955078125,
0.489501953125,
0.317626953125,
0.118408203125,
0.317626953125
],
[
0.51123046875,
0.3017578125,
0.88671875,
0.302734375,
0.88671875,
0.317626953125,
0.51123046875,
0.31640625
],
[
0.118408203125,
0.320556640625,
0.491943359375,
0.320556640625,
0.491943359375,
0.334228515625,
0.118408203125,
0.334228515625
],
[
0.51318359375,
0.319580078125,
0.884765625,
0.319580078125,
0.884765625,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.12005615234375,
0.336181640625,
0.490234375,
0.336181640625,
0.490234375,
0.350830078125,
0.12005615234375,
0.350830078125
],
[
0.51318359375,
0.336181640625,
0.88330078125,
0.336181640625,
0.88330078125,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.12005615234375,
0.352783203125,
0.490234375,
0.352783203125,
0.490234375,
0.366455078125,
0.12005615234375,
0.366455078125
],
[
0.51123046875,
0.350830078125,
0.88330078125,
0.3515625,
0.88330078125,
0.366455078125,
0.51123046875,
0.365234375
],
[
0.118408203125,
0.3681640625,
0.489501953125,
0.3681640625,
0.489501953125,
0.3818359375,
0.118408203125,
0.3818359375
],
[
0.51123046875,
0.366455078125,
0.7021484375,
0.36767578125,
0.7021484375,
0.3818359375,
0.51123046875,
0.38134765625
],
[
0.12005615234375,
0.385498046875,
0.489501953125,
0.385498046875,
0.489501953125,
0.3984375,
0.12005615234375,
0.3984375
],
[
0.51123046875,
0.392578125,
0.6943359375,
0.393798828125,
0.6943359375,
0.407958984375,
0.51123046875,
0.40673828125
],
[
0.12005615234375,
0.40087890625,
0.490234375,
0.40087890625,
0.490234375,
0.41455078125,
0.12005615234375,
0.41455078125
],
[
0.12005615234375,
0.416259765625,
0.4248046875,
0.416259765625,
0.4248046875,
0.429931640625,
0.12005615234375,
0.429931640625
],
[
0.5146484375,
0.4150390625,
0.884765625,
0.4150390625,
0.884765625,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.1376953125,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.446533203125,
0.1376953125,
0.446533203125
],
[
0.51318359375,
0.43115234375,
0.884765625,
0.429931640625,
0.884765625,
0.44482421875,
0.51318359375,
0.4453125
],
[
0.118408203125,
0.449462890625,
0.489501953125,
0.449462890625,
0.489501953125,
0.463134765625,
0.118408203125,
0.463134765625
],
[
0.51318359375,
0.44775390625,
0.884765625,
0.44775390625,
0.884765625,
0.46142578125,
0.51318359375,
0.46142578125
],
[
0.118408203125,
0.465576171875,
0.48779296875,
0.465576171875,
0.48779296875,
0.478515625,
0.118408203125,
0.478515625
],
[
0.51318359375,
0.463134765625,
0.88330078125,
0.463134765625,
0.88330078125,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.118408203125,
0.482177734375,
0.490234375,
0.482177734375,
0.490234375,
0.495361328125,
0.118408203125,
0.495361328125
],
[
0.5146484375,
0.478515625,
0.88671875,
0.478515625,
0.88671875,
0.493408203125,
0.5146484375,
0.493408203125
],
[
0.1175537109375,
0.49755859375,
0.17626953125,
0.49755859375,
0.17626953125,
0.5126953125,
0.1175537109375,
0.5126953125
],
[
0.5146484375,
0.495361328125,
0.88671875,
0.495361328125,
0.88671875,
0.5087890625,
0.5146484375,
0.5087890625
],
[
0.1368408203125,
0.51318359375,
0.4189453125,
0.51318359375,
0.4189453125,
0.53076171875,
0.1368408203125,
0.53076171875
],
[
0.5146484375,
0.5126953125,
0.884765625,
0.5126953125,
0.884765625,
0.525390625,
0.5146484375,
0.525390625
],
[
0.1368408203125,
0.53076171875,
0.468505859375,
0.53076171875,
0.468505859375,
0.548828125,
0.1368408203125,
0.548828125
],
[
0.5146484375,
0.52783203125,
0.88330078125,
0.52783203125,
0.88330078125,
0.54150390625,
0.5146484375,
0.54150390625
],
[
0.51318359375,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.51318359375,
0.5556640625
],
[
0.118408203125,
0.5517578125,
0.489501953125,
0.5517578125,
0.489501953125,
0.56640625,
0.118408203125,
0.56640625
],
[
0.5146484375,
0.56005859375,
0.88330078125,
0.56005859375,
0.88330078125,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.1175537109375,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.1175537109375,
0.58056640625
],
[
0.1175537109375,
0.58203125,
0.19140625,
0.58447265625,
0.19140625,
0.5986328125,
0.1175537109375,
0.5966796875
],
[
0.5146484375,
0.57666015625,
0.88330078125,
0.57666015625,
0.88330078125,
0.59033203125,
0.5146484375,
0.59033203125
],
[
0.5146484375,
0.5927734375,
0.88330078125,
0.5927734375,
0.88330078125,
0.60546875,
0.5146484375,
0.60546875
],
[
0.1376953125,
0.60107421875,
0.48779296875,
0.60107421875,
0.48779296875,
0.6142578125,
0.1376953125,
0.6142578125
],
[
0.5146484375,
0.60791015625,
0.8017578125,
0.60791015625,
0.8017578125,
0.62158203125,
0.5146484375,
0.62158203125
],
[
0.12005615234375,
0.6181640625,
0.48779296875,
0.6181640625,
0.48779296875,
0.62890625,
0.12005615234375,
0.62890625
],
[
0.5146484375,
0.62353515625,
0.884765625,
0.62353515625,
0.884765625,
0.63818359375,
0.5146484375,
0.63818359375
],
[
0.12005615234375,
0.6328125,
0.490234375,
0.6328125,
0.490234375,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.51318359375,
0.64013671875,
0.88330078125,
0.64013671875,
0.88330078125,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.118408203125,
0.64794921875,
0.490234375,
0.6484375,
0.490234375,
0.6630859375,
0.118408203125,
0.662109375
],
[
0.51318359375,
0.65625,
0.884765625,
0.65625,
0.884765625,
0.67041015625,
0.51318359375,
0.67041015625
],
[
0.118408203125,
0.6650390625,
0.36279296875,
0.6650390625,
0.36279296875,
0.6787109375,
0.118408203125,
0.6787109375
],
[
0.51123046875,
0.67138671875,
0.884765625,
0.6728515625,
0.884765625,
0.68701171875,
0.51123046875,
0.68603515625
],
[
0.145263671875,
0.68310546875,
0.489501953125,
0.68310546875,
0.489501953125,
0.69775390625,
0.145263671875,
0.69775390625
],
[
0.51318359375,
0.6884765625,
0.88330078125,
0.6884765625,
0.88330078125,
0.703125,
0.51318359375,
0.703125
],
[
0.12005615234375,
0.69970703125,
0.490234375,
0.69970703125,
0.490234375,
0.71337890625,
0.12005615234375,
0.71337890625
],
[
0.51318359375,
0.705078125,
0.87744140625,
0.705078125,
0.87744140625,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.12005615234375,
0.71630859375,
0.490234375,
0.71630859375,
0.490234375,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.118408203125,
0.72998046875,
0.49365234375,
0.73095703125,
0.49365234375,
0.7451171875,
0.118408203125,
0.74462890625
],
[
0.51318359375,
0.73095703125,
0.736328125,
0.73095703125,
0.736328125,
0.7451171875,
0.51318359375,
0.7451171875
],
[
0.118408203125,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.51123046875,
0.75048828125,
0.880859375,
0.75048828125,
0.880859375,
0.76416015625,
0.51123046875,
0.76416015625
],
[
0.118408203125,
0.76416015625,
0.489501953125,
0.76416015625,
0.489501953125,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51318359375,
0.7685546875,
0.88330078125,
0.7685546875,
0.88330078125,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.118408203125,
0.77978515625,
0.490234375,
0.77978515625,
0.490234375,
0.79345703125,
0.118408203125,
0.79345703125
],
[
0.51318359375,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51318359375,
0.798828125
],
[
0.118408203125,
0.79638671875,
0.490234375,
0.79638671875,
0.490234375,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.51318359375,
0.80029296875,
0.88330078125,
0.80029296875,
0.88330078125,
0.81396484375,
0.51318359375,
0.81396484375
],
[
0.1175537109375,
0.8125,
0.490234375,
0.81103515625,
0.490234375,
0.826171875,
0.1175537109375,
0.82666015625
],
[
0.51318359375,
0.818359375,
0.87890625,
0.818359375,
0.87890625,
0.8291015625,
0.51318359375,
0.8291015625
],
[
0.1175537109375,
0.8291015625,
0.376220703125,
0.8291015625,
0.376220703125,
0.8427734375,
0.1175537109375,
0.8427734375
],
[
0.51318359375,
0.83203125,
0.88330078125,
0.83203125,
0.88330078125,
0.845703125,
0.51318359375,
0.845703125
],
[
0.1351318359375,
0.845703125,
0.490234375,
0.8466796875,
0.490234375,
0.86083984375,
0.1351318359375,
0.85986328125
],
[
0.51318359375,
0.85107421875,
0.8818359375,
0.85107421875,
0.8818359375,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.1175537109375,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.1175537109375,
0.876953125
],
[
0.53564453125,
0.87158203125,
0.748046875,
0.87158203125,
0.748046875,
0.88525390625,
0.53564453125,
0.88525390625
],
[
0.118408203125,
0.87890625,
0.491943359375,
0.87890625,
0.491943359375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.5341796875,
0.88427734375,
0.75732421875,
0.88427734375,
0.75732421875,
0.89794921875,
0.5341796875,
0.89794921875
],
[
0.12005615234375,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5146484375,
0.89794921875,
0.6533203125,
0.89794921875,
0.6533203125,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"further information about the entity (including its",
"Description",
"F1",
"type), using an entity linking module.",
"Without LM",
"49.9",
"Model.First, we construct the token embed-",
"With LM",
"54.1",
"ding xk = [ck; wk] for each token tk in the input",
"Avg. of 15 models with LM",
"55.2",
"sequence, where ck is a character-based represen-",
"tation computed using a convolutional neural net-",
"Table 3: Results of the entity extraction model on",
"work (CNN) with filter of size 3 characters, and wk",
"the development set of SemEval-2017 task 10",
"are learned word embeddings initialized with the",
"GloVe embeddings (Pennington et al., 2014).",
"We also compute context-sensitive word embed-",
"BC5CDR (Li et al., 2016) and the CHEMDNER",
"dings, denoted as Imk = [lm,; lm, ], by concate-",
"datasets (Krallinger et al., 2015) to extract key en-",
"nating the projected outputs of forward and back-",
"tity mentions in the biomedical domain such as dis-",
"ward recurrent neural network language models",
" eases, drugs and chemical compounds. The third",
"(RNN-LM) at position k. The language model",
"instance is trained on mention labels induced from",
"(LM) for each direction is trained independently",
"Wikipedia articles in the computer science domain.",
"and consists of a single layer long short-term mem-",
"The output of all model instances are pooled to-",
"ory (LSTM) network followed by a linear project",
"gether and combined with the rule-based entity",
"layer. While training the LM parameters, lm, is",
"extraction module, then fed into the entity linking",
"used to predict tk+1 and Im, is used to predict tk-1.",
"model (described below).",
"We fix the LM parameters during training of the",
"4.3Knowledge Bases",
"entity extraction model. See Peters et al. (2017)",
"and Ammar et al. (2017) for more details.",
"In this section, we describe the construction of en-",
"Given the Xk and Imk embeddings for each token",
"tity nodes and entity-entity edges. Unlike other",
"k e (l,..., N), we use a two-layer bidirectional",
"knowledge extraction systems such as the Never-",
"LSTM to encode the sequence with xk and lmk",
"Ending Language Learner (NELL)7 and OpenIE",
"feeding into the first and second layer, respectively.",
"4,8 we use existing knowledge bases (KBs) of en-",
"That is,",
"tities to reduce the burden of identifying coher-",
"g = LSTM(xk, gk-1), gk = [gk ; g],",
"ent concepts. Grounding the entity mentions in",
"h_ = LSTM([gk; Imi], h=-,),hk = [h;h],",
"a manually-curated KB also increases user confi-",
"dence in automated predictions. We use two KBs:",
"where g and h, are defined similarly to gk and",
"UMLS: The UMLS metathesaurus integrates in",
"h, but process token sequences in the opposite",
"direction.",
"formation about concepts in specialized ontologies",
"in several biomedical domains, and is funded by",
"Similar to the model described in 3, we feed the",
"the U.S. National Library of Medicine.",
"outputof thesecondLSTMintoadenselayerto",
"DBpedia: DBpedia provides access to structured",
"predict unnormalized label weights for each token",
"information in Wikipedia. Rather than including all",
"and learn label bigram feature weights to account",
"Wikipedia pages, we used a short list of Wikipedia",
"for dependencies between labels.",
"categories about CS and included all pages up to",
"Results. We use the standard data splits of",
"depth four in their trees in order to exclude irrele",
"the SemEval-2017 Task 10 on entity (and relation)",
"vant entities, e.g., \"Lord of the Rings\" in DBpedia.",
"extraction from scientific papers (Augenstein et al.,",
"2017). Table 3 compares three variants of our en-",
"4.4Entity Linking Models",
"tity extraction model. The first line omits the LM",
"Given a text span s identified by the entity extrac",
"embeddings lmk, while the second line is the full",
"tion model in s4.2 (or with heuristics) and a ref-",
"model (including LM embeddings) showing a large",
"erence KB, the goal of the entity linking model",
"improvement of 4.2 F1 points. The third line shows",
"is to associate the span with the entity it refers to.",
"that creating an ensemble of 15 models further im-",
"Aspananditssurroundingwordsarecollectively",
"proves the results by 1.1 F1 points.",
"referred to as a mention. We first identify a set of",
"Model instances.In the deployed system, we",
"candidate entitiesthatagivenmentionmay refer",
"use three instances of the entity extraction model",
"7http://rtw.ml.cmu.edu/rtw/",
"with a similar architecture, but trained on differ-",
"\"https://github.com/allenai/",
"ent datasets. Two instances are trained on the",
"openie-standalone"
] |
{
"arxiv_id": "1805.02262",
"language": "en",
"page_idx": 4
}
|
||
1512.05986-en-0
|
1512.05986-en-0.jpg
|
$\textbf{Can Pretrained Neural Networks Detect Anatomy? }$
Vlado Menkovski Philips Research
Eindhoven, Netherlands
vlado.menkovski@philips.com
$\mathbf{Axel}$ Saalbach Philips Research Hamburg, Germany
axel.saalbach@philips.coom
Zharko Aleksovski Philips Research
Eindhoven, Netherlands
zharko.aleksovski@philips.com
Hannes Nickisch Philips Research Hamburg, Germany
hannes.nickisch@philips.com
## $\mathbf{Abstract}$
Convolutional neural networks demonstrated outstanding empirical results in computer vision and speech recognition tasks where labeled training data is abundant. In medical imaging, there is a huge variety of possible imaging modalities and contrasts, where annotated data is usually very scarce. We present two approaches to deal with this challenge. A network pretrained in a different domain $\hat{\text{with abundant data is used as a feature extractor, while a subsequent classifier is}}$ trained on a small target dataset; and a deep architecture trained with heavy augmentation and equipped with sophisticated regularization methods. We test the approaches on a corpus of X-ray images to design an anatomy detection system.
### 1 Introduction and Motivation
Deep learning [9] (DL) is rapidly gaining momentum in computer vision [5], speech recognition [3] and medical imaging applications due to their record-breaking empirical performance and their promise to jointly learn the low-level features and the high-level decisions directly from raw data.
One major open research challenge is the transfer of the excellent predictive performance of DL methods, in particular convolutional neural networks (CNN), from the “data-laden”regime of computer vision$^{\mathrm{i}}$ to the“data-bounded”regime of medical imaging.
Since the typical number of parameters in the DL models is very big the major concern for operating in“data-bounded”regimes is inability to generalize and form robust features because the small number of examples do not capture welí the variability in the input space. There are various approaches to deal with this challenge. Data augmentation can be used to increase the effective size of the training set and drive the training to build invariances to specific transformations. Network architecture design can be used to hard-code prior knowledge (such as translation invariance) by design. Sophisticated regularization methods such as dropout[8] and batch normalization (BN)[4] can work to counter overfitting. Finally, global training of the feature extraction layers on abundant data domains (i.e. natural images) and then training the “classification”higher layers from the limited data domains can be used to ensure that the lower-level features are invariant to the particularities of the training dataset.
We choose two directions to investigate these approaches for the goal of anatomy detection in X-ray images. The first using a network trained on a large corpus of natural images for feature extraction
in
the
ILSVRC
challenge,
$10^3$
$^{1}$There
$10^6$
in
classes
are
images
http://www.image-net.org/challenges/LSVRC.
1
|
[
[
0.193603515625,
0.1395263671875,
0.8056640625,
0.1395263671875,
0.8056640625,
0.1572265625,
0.193603515625,
0.1572265625
],
[
0.27783203125,
0.2139892578125,
0.40283203125,
0.2139892578125,
0.40283203125,
0.2291259765625,
0.27783203125,
0.2291259765625
],
[
0.583984375,
0.2139892578125,
0.7197265625,
0.2139892578125,
0.7197265625,
0.2291259765625,
0.583984375,
0.2291259765625
],
[
0.281982421875,
0.2291259765625,
0.39453125,
0.2291259765625,
0.39453125,
0.24365234375,
0.281982421875,
0.24365234375
],
[
0.5947265625,
0.2291259765625,
0.70751953125,
0.2291259765625,
0.70751953125,
0.24365234375,
0.5947265625,
0.24365234375
],
[
0.25732421875,
0.242431640625,
0.418212890625,
0.242431640625,
0.418212890625,
0.255615234375,
0.25732421875,
0.255615234375
],
[
0.0187835693359375,
0.231689453125,
0.060455322265625,
0.231689453125,
0.063720703125,
0.37890625,
0.021240234375,
0.37890625
],
[
0.57177734375,
0.242431640625,
0.73193359375,
0.242431640625,
0.73193359375,
0.255615234375,
0.57177734375,
0.255615234375
],
[
0.205078125,
0.2568359375,
0.473876953125,
0.2568359375,
0.473876953125,
0.270751953125,
0.205078125,
0.270751953125
],
[
0.5087890625,
0.2568359375,
0.79736328125,
0.2568359375,
0.79736328125,
0.270751953125,
0.5087890625,
0.270751953125
],
[
0.291748046875,
0.291748046875,
0.39453125,
0.291748046875,
0.39453125,
0.30615234375,
0.291748046875,
0.30615234375
],
[
0.59228515625,
0.291748046875,
0.7099609375,
0.291748046875,
0.7099609375,
0.30615234375,
0.59228515625,
0.30615234375
],
[
0.285888671875,
0.30615234375,
0.400390625,
0.30615234375,
0.400390625,
0.3193359375,
0.285888671875,
0.3193359375
],
[
0.5947265625,
0.30615234375,
0.708984375,
0.30615234375,
0.708984375,
0.3193359375,
0.5947265625,
0.3193359375
],
[
0.273681640625,
0.318115234375,
0.408447265625,
0.3193359375,
0.408447265625,
0.333984375,
0.273681640625,
0.333251953125
],
[
0.58251953125,
0.318115234375,
0.71728515625,
0.3193359375,
0.71728515625,
0.333984375,
0.58251953125,
0.333251953125
],
[
0.2156982421875,
0.333984375,
0.47216796875,
0.333984375,
0.47216796875,
0.347900390625,
0.2156982421875,
0.347900390625
],
[
0.51708984375,
0.333984375,
0.7861328125,
0.333984375,
0.7861328125,
0.347900390625,
0.51708984375,
0.347900390625
],
[
0.021240234375,
0.371826171875,
0.06451416015625,
0.371826171875,
0.06451416015625,
0.46533203125,
0.021240234375,
0.46533203125
],
[
0.043304443359375,
0.3818359375,
0.053924560546875,
0.3818359375,
0.053924560546875,
0.386474609375,
0.043304443359375,
0.386474609375
],
[
0.4599609375,
0.3818359375,
0.5400390625,
0.3837890625,
0.5400390625,
0.401611328125,
0.4599609375,
0.399658203125
],
[
0.233642578125,
0.416748046875,
0.765625,
0.416748046875,
0.765625,
0.43115234375,
0.233642578125,
0.43115234375
],
[
0.233642578125,
0.43115234375,
0.7646484375,
0.43115234375,
0.7646484375,
0.4443359375,
0.233642578125,
0.4443359375
],
[
0.233642578125,
0.4443359375,
0.765625,
0.4443359375,
0.765625,
0.458984375,
0.233642578125,
0.458984375
],
[
0.233642578125,
0.458984375,
0.765625,
0.458984375,
0.765625,
0.472900390625,
0.233642578125,
0.472900390625
],
[
0.0196075439453125,
0.46728515625,
0.058013916015625,
0.46728515625,
0.058013916015625,
0.6875,
0.0196075439453125,
0.6875
],
[
0.233642578125,
0.4716796875,
0.76708984375,
0.4716796875,
0.76708984375,
0.486083984375,
0.233642578125,
0.486083984375
],
[
0.233642578125,
0.48486328125,
0.765625,
0.486083984375,
0.765625,
0.5,
0.233642578125,
0.498779296875
],
[
0.233642578125,
0.50048828125,
0.76318359375,
0.50048828125,
0.76318359375,
0.5146484375,
0.233642578125,
0.5146484375
],
[
0.233642578125,
0.5146484375,
0.765625,
0.5146484375,
0.765625,
0.52783203125,
0.233642578125,
0.52783203125
],
[
0.2344970703125,
0.52880859375,
0.75634765625,
0.52880859375,
0.75634765625,
0.54248046875,
0.2344970703125,
0.54248046875
],
[
0.17724609375,
0.57275390625,
0.2017822265625,
0.57275390625,
0.2017822265625,
0.58056640625,
0.17724609375,
0.58056640625
],
[
0.1993408203125,
0.57080078125,
0.4453125,
0.57080078125,
0.4453125,
0.583984375,
0.1993408203125,
0.583984375
],
[
0.1748046875,
0.6015625,
0.826171875,
0.6015625,
0.826171875,
0.61572265625,
0.1748046875,
0.61572265625
],
[
0.1748046875,
0.6162109375,
0.82421875,
0.6162109375,
0.82421875,
0.6298828125,
0.1748046875,
0.6298828125
],
[
0.1748046875,
0.6298828125,
0.81201171875,
0.6298828125,
0.81201171875,
0.64306640625,
0.1748046875,
0.64306640625
],
[
0.1748046875,
0.65087890625,
0.82275390625,
0.65087890625,
0.82275390625,
0.6640625,
0.1748046875,
0.6640625
],
[
0.1748046875,
0.66552734375,
0.82275390625,
0.66552734375,
0.82275390625,
0.6787109375,
0.1748046875,
0.6787109375
],
[
0.1748046875,
0.6787109375,
0.59228515625,
0.6787109375,
0.59228515625,
0.693359375,
0.1748046875,
0.693359375
],
[
0.1732177734375,
0.69873046875,
0.82177734375,
0.69970703125,
0.82177734375,
0.7138671875,
0.1732177734375,
0.71337890625
],
[
0.1732177734375,
0.71337890625,
0.82421875,
0.7119140625,
0.82421875,
0.7265625,
0.1732177734375,
0.72802734375
],
[
0.1748046875,
0.72802734375,
0.82275390625,
0.72802734375,
0.82275390625,
0.7412109375,
0.1748046875,
0.7412109375
],
[
0.1748046875,
0.7421875,
0.82421875,
0.7421875,
0.82421875,
0.755859375,
0.1748046875,
0.755859375
],
[
0.1732177734375,
0.7548828125,
0.82421875,
0.7548828125,
0.82421875,
0.76953125,
0.1732177734375,
0.76953125
],
[
0.1748046875,
0.76953125,
0.82275390625,
0.76953125,
0.82275390625,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1748046875,
0.78271484375,
0.82421875,
0.78271484375,
0.82421875,
0.796875,
0.1748046875,
0.796875
],
[
0.1732177734375,
0.796875,
0.82275390625,
0.796875,
0.82275390625,
0.81005859375,
0.1732177734375,
0.81005859375
],
[
0.176513671875,
0.8125,
0.82275390625,
0.8125,
0.82275390625,
0.82275390625,
0.176513671875,
0.82275390625
],
[
0.1732177734375,
0.82470703125,
0.826171875,
0.82470703125,
0.826171875,
0.83837890625,
0.1732177734375,
0.83837890625
],
[
0.1748046875,
0.8388671875,
0.306396484375,
0.8388671875,
0.306396484375,
0.85302734375,
0.1748046875,
0.85302734375
],
[
0.1732177734375,
0.859375,
0.82177734375,
0.859375,
0.82177734375,
0.87353515625,
0.1732177734375,
0.87353515625
],
[
0.1748046875,
0.87353515625,
0.82421875,
0.87353515625,
0.82421875,
0.88720703125,
0.1748046875,
0.88720703125
],
[
0.1968994140625,
0.8984375,
0.240234375,
0.8984375,
0.240234375,
0.91357421875,
0.1968994140625,
0.91357421875
],
[
0.310546875,
0.8984375,
0.3349609375,
0.8984375,
0.3349609375,
0.9111328125,
0.310546875,
0.9111328125
],
[
0.359375,
0.8984375,
0.408447265625,
0.8984375,
0.408447265625,
0.91357421875,
0.359375,
0.91357421875
],
[
0.51953125,
0.8984375,
0.56884765625,
0.8984375,
0.56884765625,
0.91357421875,
0.51953125,
0.91357421875
],
[
0.7607421875,
0.8984375,
0.826171875,
0.8984375,
0.826171875,
0.91357421875,
0.7607421875,
0.91357421875
],
[
0.5908203125,
0.8994140625,
0.611328125,
0.8994140625,
0.611328125,
0.91357421875,
0.5908203125,
0.91357421875
],
[
0.4306640625,
0.8994140625,
0.450927734375,
0.8994140625,
0.450927734375,
0.912109375,
0.4306640625,
0.912109375
],
[
0.63232421875,
0.8994140625,
0.65625,
0.8994140625,
0.65625,
0.912109375,
0.63232421875,
0.912109375
],
[
0.6806640625,
0.8994140625,
0.736328125,
0.8994140625,
0.736328125,
0.9111328125,
0.6806640625,
0.9111328125
],
[
0.468017578125,
0.90087890625,
0.495849609375,
0.89599609375,
0.5,
0.90771484375,
0.47216796875,
0.91357421875
],
[
0.26318359375,
0.90087890625,
0.285888671875,
0.90087890625,
0.285888671875,
0.912109375,
0.26318359375,
0.912109375
],
[
0.17236328125,
0.912109375,
0.5400390625,
0.912109375,
0.5400390625,
0.92529296875,
0.17236328125,
0.92529296875
],
[
0.494384765625,
0.94873046875,
0.5048828125,
0.94873046875,
0.5048828125,
0.9609375,
0.494384765625,
0.9609375
]
] |
[
"Can Pretrained Neural Networks Detect Anatomy?",
"Vlado Menkovski",
"Zharko Aleksovski",
"Philips Research",
"Philips Research",
"Eindhoven, Netherlands",
"8",
"Eindhoven, Netherlands",
"vlado.menkovski@philips.com",
"zharko.aleksovski@philips.com",
"Axel Saalbach",
"Hannes Nickisch",
"Philips Research",
"Philips Research",
"Hamburg, Germany",
"Hamburg, Germany",
"axel.saalbach@philips.coom",
"hannes.nickisch@philips.com",
"[As]",
"",
"Abstract",
"Convolutional neural networks demonstrated outstanding empirical results in",
"computer vision and speech recognition tasks where labeled training data is abun-",
"dant. In medical imaging, there is a huge variety of possible imaging modalities",
"and contrasts, where annotated data is usually very scarce. We present two ap-",
"1986502S",
"proaches to deal with this challenge. A network pretrained in a different domain",
"with abundant data is used as a feature extractor, while a subsequent classifier is",
"trained on a small target dataset; and a deep architecture trained with heavy aug-",
"mentation and equipped with sophisticated regularization methods. We test the",
"approaches on a corpus of X-ray images to design an anatomy detection system.",
"",
"Introduction and Motivation",
"Deep learning [9] (DL) is rapidly gaining momentum in computer vision [5], speech recognition",
"[3] and medical imaging applications due to their record-breaking empirical performance and their",
"promise to jointly learn the low-level features and the high-level decisions directly from raw data.",
"One major open research challenge is the transfer of the excellent predictive performance of DL",
"methods,in particular convolutional neural networks (CNN),from the\"data-ladenregime of com-",
"puter vision' to the“data-bounded\"regime of medical imaging.",
"Since the typical number of parameters in the DL models is very big the major concern for operat",
"ing in “data-bounded\" regimes is inability to generalize and form robust features because the small",
"number of examples do not capture well thevariability in the input space.There are various ap-",
"proaches to deal with this challenge. Data augmentation can be used to increase the effective size",
"of the training set and drive the training to build invariances to specific transformations. Network",
"architecture design can be used to hard-code prior knowledge (such as translation invariance) by de-",
"sign. Sophisticated regularization methods such as dropout [8] and batch normalization (BN) [4] can",
"work to counteroverfitting.Finally,global trainingof thefeature extraction layers on abundant data",
"domains (i.e.naturalimages)andthen trainingtheclassification\"higherlayersfromthelimited",
"data domains can be used to ensure that the lower-level features are invariant to the particularities of",
"the training dataset.",
"We choose two directions to investigate these approaches for the goal of anatomy detection in X-ray",
"images. The first using a network trained on a large corpus of natural images for feature extraction",
"'There",
"106",
"images",
"classes",
"challenge,",
"in",
"in",
"the",
"ILSVRC",
"103",
"are",
"http://www.image-net.org/challenges/LsvRc",
"1"
] |
{
"arxiv_id": "1512.05986",
"language": "en",
"page_idx": 0
}
|
||
1512.05986-en-1
|
1512.05986-en-1.jpg
|
combined with a SVM [1] classifier trained on the medical images. And a second approach using heavy augmentation and regularization of a deep neural network trained only on the small-sized medical image dataset.
# 2 Experiments and Results
In order to assess the performance of specifically trained and off-the-shelf networks, radiographs from the ImageClef 2009- Medical Image Annotation task$^{2}$ were used. This challenge database consists of a broad range of X-ray images from clinical routine, with a detailed anatomical classiffcation. For convenience, we flattened the hierarchical class representation and disregarded classes with fewer than 50 example images. In doing so we ended up with 24 unique classes. For evaluation purposes, the entire data corpus (of 14676 images) was divided into a training and test set, covering $\hat{9}0\%$ and 10% of the data respectively.
As a first baseline, the OverFeat network (see [6]) in combination with a (linear) multi-class SVM was employed. OverFeat is a convolutional neural network that was trained on 1.2 million nonmedical images form the ImageNet2012 database with 1000 classes, and has been successfully employed as a generic feature extractor in different applications.
For the purpose of this experiment, we selected the $f\alpha st$ OverFeat instantiation and used the default setting for feature extraction. This results in a 4096 dimensional representation of the images, which was subsequently used as input for the SVM classier. The regularization parameter C of the linear SVM was estimated on the training set using exhaustive search and 5-fold cross-validation.
For the second direction we designed a network 2 inspired by the work of Razavian et al. [7], using dropout and batch normalization with leaky ReLU activation functions [2]. In this case for training we used images of 128 by 128 pixels with various augmentations ranging from resizing and cropping, rotation, translation, shearing, stretching and flipping. The results of the two directions are given in table 1.
<table>
<tbody>
<tr>
<th>OverFeat+ $SVN$ ,</th>
<th>$M$ Custom $CNN$ </th>
</tr>
<tr>
<td>92.42</td>
<td>95.14</td>
</tr>
</tbody>
</table>
# 3 Conclusions and Perspectives
We investigated when, where, and how to (re)train a convolutional neural network used for anatomy detection. Our results suggest that pretrained networks are good image descriptors even outside their training domain and that retraining only the last layer is a viable alternative to perform transfer learning in light of limited training data. We also demonstrated that sufficiently expressive models such as a very deep neural network can be trained even on relatively small number of annotations if proper augmentation and regularization is implemented. Our work is a first step towards a multimodality anatomical expert system with components trained to maximize data efficiency.
# References
[1] C. Cortes and V. Vapnik. Support-vector networks. Machine learning, 20(3):273-297, 1995.
[2] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level
performance on imagenet classification. $arXi\nu preprint$ arXiv:1502.01852,2015.
[3] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke,
P. Nguyen, T. N. Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. Signal Processing Magazine, IEEE, 29(6):82- 97,2012.
$^{2}$http://www.imageclef.org/2009/medanno
2
|
[
[
0.1748046875,
0.10418701171875,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11932373046875,
0.1748046875,
0.11871337890625
],
[
0.1748046875,
0.11932373046875,
0.82421875,
0.11932373046875,
0.82421875,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.1732177734375,
0.1337890625,
0.325927734375,
0.1337890625,
0.325927734375,
0.147705078125,
0.1732177734375,
0.147705078125
],
[
0.1732177734375,
0.1685791015625,
0.414306640625,
0.1685791015625,
0.414306640625,
0.18310546875,
0.1732177734375,
0.18310546875
],
[
0.1732177734375,
0.2008056640625,
0.82421875,
0.2008056640625,
0.82421875,
0.2139892578125,
0.1732177734375,
0.2139892578125
],
[
0.1732177734375,
0.2139892578125,
0.82275390625,
0.21533203125,
0.82275390625,
0.2291259765625,
0.1732177734375,
0.2279052734375
],
[
0.176513671875,
0.2310791015625,
0.8203125,
0.2310791015625,
0.8203125,
0.2412109375,
0.176513671875,
0.2412109375
],
[
0.1748046875,
0.242431640625,
0.82421875,
0.242431640625,
0.82421875,
0.255615234375,
0.1748046875,
0.255615234375
],
[
0.1748046875,
0.2568359375,
0.82421875,
0.2568359375,
0.82421875,
0.270751953125,
0.1748046875,
0.270751953125
],
[
0.176513671875,
0.271484375,
0.826171875,
0.271484375,
0.826171875,
0.285400390625,
0.176513671875,
0.285400390625
],
[
0.1732177734375,
0.28271484375,
0.4248046875,
0.2841796875,
0.4248046875,
0.298583984375,
0.1732177734375,
0.29736328125
],
[
0.1748046875,
0.304931640625,
0.82421875,
0.304931640625,
0.82421875,
0.3193359375,
0.1748046875,
0.3193359375
],
[
0.1748046875,
0.3193359375,
0.82177734375,
0.3193359375,
0.82177734375,
0.333251953125,
0.1748046875,
0.333251953125
],
[
0.1748046875,
0.333984375,
0.82275390625,
0.333984375,
0.82275390625,
0.347900390625,
0.1748046875,
0.347900390625
],
[
0.1748046875,
0.347900390625,
0.5947265625,
0.347900390625,
0.5947265625,
0.361083984375,
0.1748046875,
0.361083984375
],
[
0.1732177734375,
0.367431640625,
0.82275390625,
0.367431640625,
0.82275390625,
0.380615234375,
0.1732177734375,
0.380615234375
],
[
0.176513671875,
0.3837890625,
0.82275390625,
0.3837890625,
0.82275390625,
0.39453125,
0.176513671875,
0.39453125
],
[
0.1748046875,
0.396484375,
0.82421875,
0.396484375,
0.82421875,
0.410400390625,
0.1748046875,
0.410400390625
],
[
0.1748046875,
0.410400390625,
0.76953125,
0.410400390625,
0.76953125,
0.423583984375,
0.1748046875,
0.423583984375
],
[
0.17236328125,
0.429931640625,
0.82275390625,
0.43115234375,
0.82275390625,
0.44580078125,
0.17236328125,
0.4443359375
],
[
0.1748046875,
0.44580078125,
0.82275390625,
0.44580078125,
0.82275390625,
0.458984375,
0.1748046875,
0.458984375
],
[
0.1748046875,
0.458984375,
0.82421875,
0.458984375,
0.82421875,
0.472900390625,
0.1748046875,
0.472900390625
],
[
0.1748046875,
0.473388671875,
0.82421875,
0.473388671875,
0.82421875,
0.4873046875,
0.1748046875,
0.4873046875
],
[
0.1732177734375,
0.4873046875,
0.306396484375,
0.486083984375,
0.306396484375,
0.50048828125,
0.1732177734375,
0.501953125
],
[
0.373291015625,
0.52587890625,
0.62353515625,
0.52587890625,
0.62353515625,
0.54052734375,
0.373291015625,
0.54052734375
],
[
0.384033203125,
0.54345703125,
0.5048828125,
0.54345703125,
0.5048828125,
0.556640625,
0.384033203125,
0.556640625
],
[
0.5048828125,
0.54345703125,
0.61279296875,
0.54345703125,
0.61279296875,
0.556640625,
0.5048828125,
0.556640625
],
[
0.42236328125,
0.56103515625,
0.464111328125,
0.56103515625,
0.464111328125,
0.5771484375,
0.42236328125,
0.5771484375
],
[
0.54345703125,
0.56103515625,
0.5849609375,
0.56103515625,
0.5849609375,
0.57568359375,
0.54345703125,
0.57568359375
],
[
0.1748046875,
0.6123046875,
0.1993408203125,
0.61572265625,
0.1968994140625,
0.62548828125,
0.17236328125,
0.62255859375
],
[
0.19775390625,
0.611328125,
0.449462890625,
0.611328125,
0.449462890625,
0.62548828125,
0.19775390625,
0.62548828125
],
[
0.1748046875,
0.64453125,
0.82275390625,
0.64453125,
0.82275390625,
0.65771484375,
0.1748046875,
0.65771484375
],
[
0.1748046875,
0.6591796875,
0.82275390625,
0.6591796875,
0.82275390625,
0.67236328125,
0.1748046875,
0.67236328125
],
[
0.1748046875,
0.67236328125,
0.82275390625,
0.67236328125,
0.82275390625,
0.68603515625,
0.1748046875,
0.68603515625
],
[
0.1748046875,
0.68798828125,
0.82275390625,
0.68798828125,
0.82275390625,
0.69873046875,
0.1748046875,
0.69873046875
],
[
0.1748046875,
0.70068359375,
0.82421875,
0.70068359375,
0.82421875,
0.7138671875,
0.1748046875,
0.7138671875
],
[
0.1748046875,
0.7138671875,
0.82275390625,
0.7138671875,
0.82275390625,
0.72802734375,
0.1748046875,
0.72802734375
],
[
0.1732177734375,
0.7265625,
0.75634765625,
0.72802734375,
0.75634765625,
0.7421875,
0.1732177734375,
0.7412109375
],
[
0.1748046875,
0.76318359375,
0.26708984375,
0.76318359375,
0.26708984375,
0.77783203125,
0.1748046875,
0.77783203125
],
[
0.1748046875,
0.79150390625,
0.8056640625,
0.79150390625,
0.8056640625,
0.8046875,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.81005859375,
0.82421875,
0.81005859375,
0.82421875,
0.82470703125,
0.1748046875,
0.82470703125
],
[
0.2017822265625,
0.82568359375,
0.73046875,
0.82568359375,
0.73046875,
0.8388671875,
0.2017822265625,
0.8388671875
],
[
0.176513671875,
0.8447265625,
0.82177734375,
0.8447265625,
0.82177734375,
0.85791015625,
0.176513671875,
0.85791015625
],
[
0.2017822265625,
0.859375,
0.82177734375,
0.859375,
0.82177734375,
0.87255859375,
0.2017822265625,
0.87255859375
],
[
0.2017822265625,
0.8720703125,
0.82421875,
0.8720703125,
0.82421875,
0.88623046875,
0.2017822265625,
0.88623046875
],
[
0.2017822265625,
0.88720703125,
0.2646484375,
0.88720703125,
0.2646484375,
0.8984375,
0.2017822265625,
0.8984375
],
[
0.1953125,
0.9111328125,
0.52783203125,
0.9111328125,
0.52783203125,
0.92529296875,
0.1953125,
0.92529296875
],
[
0.491943359375,
0.9462890625,
0.50732421875,
0.9462890625,
0.50732421875,
0.96337890625,
0.491943359375,
0.96337890625
]
] |
[
"combined with a SVM [1] classifier trained on the medical images. And a second approach using",
"heavy augmentation and regularization of a deep neural network trained only on the small-sized",
"medical image dataset.",
"2Experiments and Results",
"In order to assess the performance of specifically trained and off-the-shelf networks, radiographs",
"from the ImageClef 2009 -Medical Image Annotation task2 were used. This challenge database",
"consistsofabroadrangeofX-rayimagesfromclinicalroutine,withadetailedanatomicalclassifi",
"cation. For convenience, we flattened the hierarchical class representation and disregarded classes",
"with fewer than 50 example images. In doing so we ended up with 24 unique classes. For evaluation",
"purposes, the entire data corpus (of 14676 images) was divided into a training and test set, covering",
"90%and10%ofthedatarespectively",
"As a first baseline, the OverFeat network (see [6]) in combination with a (linear) multi-class SVM",
"was employed. OverFeat is a convolutional neural network that was trained on 1.2 million non-",
"medical images form the ImageNet2012 database with 1000 classes, and has been successfully",
"employedas agenericfeatureextractorindifferentapplications.",
"For the purpose of this experiment, we selected the fast OverFeat instantiation and used the default",
"settingforfeatureextraction.Thisresultsina4o96dimensionalrepresentationoftheimages,which",
"was subsequently used as input for the SVM classier. The regularization parameter C of the linear",
"SvM was estimated on the training set using exhaustive search and 5-fold cross-validation.",
"For the second direction we designed a network 2 inspired by the work of Razavian et al. [7],",
"using dropout and batch normalization with leaky ReLU activation functions [2]. In this case for",
"training we used images of 128 by 128 pixels with various augmentations ranging from resizing and",
"cropping, rotation, translation, shearing, stretching and flipping. The results of the two directions",
"are given in table l.",
"Table 1: Results: Accuracy in percent",
"OverFeat + SVM",
"CustomCNN",
"92.42",
"95.14",
"3",
"ConclusionsandPerspectives",
"We investigated when, where, and how to (re)train a convolutional neural network used for anatomy",
"detection. Our results suggest that pretrained networks are good image descriptors even outside",
"their training domain and that retraining only the last layer is a viable alternative to perform transfer",
"learninginlightoflimitedtrainingdata.Wealsodemonstratedthat sufficientlyexpressivemodels",
"such as a very deep neural network can be trained even on relatively small number of annotations",
"if proper augmentation and regularization is implemented. Our work is a first step towards a multi-",
"modality anatomical expert system with components trained to maximize data efficiency.",
"References",
"[1] C. Cortes and V.Vapnik. Support-vector networks. Machine learning, 20(3):273-297, 1995.",
"[2] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level",
"performance on imagenet classification. arXiv preprint arXiv:1502.01852,2015.",
"[3] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke.",
"P. Nguyen, T. N. Sainath, et al. Deep neural networks for acoustic modeling in speech recogni-",
"tion: The shared views of four research groups. Signal Processing Magazine, IEEE, 29(6):82-",
"97,2012.",
"2http://www.imageclef.org/2009/medanno",
"2"
] |
{
"arxiv_id": "1512.05986",
"language": "en",
"page_idx": 1
}
|
||
2403.13683-en-0
|
2403.13683-en-0.jpg
|
$\textbf{DVMNet: Computing Relative Pose for Unseen Objects Beyond Hypotheses}$
${\mathrm{Chen~Zhao}^{1}\:\mathrm{Tong~Zhang}^{1*}\:\mathrm{Zheng~Dang}^{1}\:\mathrm{Mathieu~Salzmann}^{1}}$
$\{$chen.zhao, tong.zhang, zheng.dang, mathieu.salzmann$\}$@epfl.ch
## $\mathbf{Abstract}$
Determining the relative pose of an object between two images is pivotal to the success of generalizable object pose estimation. Existing approaches typically approximate the continuous pose representation with a large number of discrete pose hypotheses, which incurs a computationally expensive process of scoring each hypothesis at test time. By contrast, we present a Deep Voxel Matching Network $(DVMNet)$ that eliminates the need for pose hypotheses and computes the relative object pose in a single pass. To this end, we map the two input RGB images, reference and query, to their respective voxelized 3D representations. We then pass the resulting voxels through a pose estimation module, where the voxels are aligned and the pose is computed in an end-to-end fashion by solving a least-squares problem. To enhance robustness, we introduce a weighted closest voxel algorithm capable of mitigating the impact $of$ noisy voxels. We conduct extensive experiments on the CO3D, LINEMOD, and Objaverse datasets, demonstrating that our method delivers more accurate relative pose esti-
$\begin{array}{cc}{\mathrm{mates~for~novel~objects~at~a~lower~computational~cost~com-}\\{\mathrm{C}}&{\mathrm{pared~to~state-of-the-art~methods.~Our~code~is~released~at:}\\{\mathrm{T}}&{\mathrm{https::\|github.com/sailor-z/DVMNet/.}}\end{array}}\\\\&{\mathrm{Thered~}}\\&{\mathrm{Thered~}}\\&{\mathrm{Thered~}}\\&{\mathrm{Thered~}}\\&{\mathrm{Thered~}}\\\\&$
<FigureHere>
Figure 1. Advantages of our DVMNet compared to hypothesisbased methods. Hypothesis-based techniques approximate the relative object pose by scoring numerous pose hypotheses, leading to a high computational cost. By contrast, our DVMNet computes the pose in a hypothesis-free fashion by robustly matching voxelized 3D representations of the reference and query images via a Weighted Closest Voxel (WCV) algorithm. Our method strikes a favorable balance between computational cost and accuracy in relative object pose estimation, as measured by multiply-accumulate operations (MACs) and angular error.
depicting unseen objects in diverse poses, serving as references. Object pose estimation is then carried out through template matching [21,29] or by establishing 2D-3D correspondences [14, 32]. Unfortunately, the effectiveness of these methods strongly depends on the references densely covering the viewpoints of unseen objects, making them inapplicable to practical scenarios where only sparse reference views are available.
In this context, a few methods[18,47,52] highlight the importance of relative object pose estimation. Unlike previous approaches in generalizable object pose estimation, these methods focus on estimating the relative pose of an unseen object between two images, i.e., a single reference image of the object and the query image. In this
## $1. \textbf{Introduction}$
Object pose estimation plays a crucial role in 3D computer vision and robotics tasks[2,11,23,33], aiming to estimate the 3D rotation and 3D translation of an object depicted in an RGB image. The vast majority of existing methods work under the assumption that the training and testing data include the same object instances, thereby limiting their applicability to scenarios that involve previously unseen objects. Recently, generalizable object pose estimation [21,29,32] has received growing attention, showcasing the potential to generalize to unseen objects from new categories without retraining the network. In pursuit of this generalization capability, existing methods leverage densely sampled images
$^{*}$Corresponding author.
|
[
[
0.1070556640625,
0.1307373046875,
0.86376953125,
0.1331787109375,
0.86376953125,
0.15087890625,
0.1070556640625,
0.1483154296875
],
[
0.21728515625,
0.18310546875,
0.313720703125,
0.18310546875,
0.313720703125,
0.1976318359375,
0.21728515625,
0.1976318359375
],
[
0.334228515625,
0.18310546875,
0.57421875,
0.18310546875,
0.57421875,
0.1976318359375,
0.334228515625,
0.1976318359375
],
[
0.58251953125,
0.18115234375,
0.75634765625,
0.1787109375,
0.75634765625,
0.1976318359375,
0.58251953125,
0.199462890625
],
[
0.580078125,
0.1875,
0.58642578125,
0.1875,
0.58642578125,
0.1925048828125,
0.580078125,
0.1925048828125
],
[
0.384033203125,
0.1976318359375,
0.44287109375,
0.199462890625,
0.441162109375,
0.21533203125,
0.38232421875,
0.21337890625
],
[
0.455078125,
0.1976318359375,
0.5908203125,
0.199462890625,
0.5908203125,
0.2171630859375,
0.455078125,
0.21533203125
],
[
0.2156982421875,
0.2197265625,
0.7509765625,
0.2197265625,
0.7509765625,
0.2342529296875,
0.2156982421875,
0.2342529296875
],
[
0.0228729248046875,
0.2587890625,
0.056365966796875,
0.2587890625,
0.056365966796875,
0.37255859375,
0.0228729248046875,
0.37255859375
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.5498046875,
0.268310546875,
0.63232421875,
0.268310546875,
0.63232421875,
0.27978515625,
0.5498046875,
0.27978515625
],
[
0.098876953125,
0.300537109375,
0.468017578125,
0.3017578125,
0.468017578125,
0.31640625,
0.098876953125,
0.31494140625
],
[
0.814453125,
0.298583984375,
0.8203125,
0.298583984375,
0.8203125,
0.306884765625,
0.814453125,
0.306884765625
],
[
0.81201171875,
0.31005859375,
0.82275390625,
0.31005859375,
0.82275390625,
0.328857421875,
0.81201171875,
0.328857421875
],
[
0.82861328125,
0.306884765625,
0.859375,
0.306884765625,
0.859375,
0.330078125,
0.82861328125,
0.330078125
],
[
0.08087158203125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.8095703125,
0.324462890625,
0.82861328125,
0.32568359375,
0.826171875,
0.354736328125,
0.80615234375,
0.354248046875
],
[
0.080078125,
0.33203125,
0.468017578125,
0.333251953125,
0.468017578125,
0.347900390625,
0.080078125,
0.3466796875
],
[
0.82421875,
0.34033203125,
0.82861328125,
0.34033203125,
0.82861328125,
0.34521484375,
0.82421875,
0.34521484375
],
[
0.080078125,
0.348388671875,
0.46728515625,
0.348388671875,
0.46728515625,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.0269622802734375,
0.363037109375,
0.0555419921875,
0.363037109375,
0.0555419921875,
0.395751953125,
0.0269622802734375,
0.395751953125
],
[
0.07843017578125,
0.363037109375,
0.46728515625,
0.3623046875,
0.46728515625,
0.376953125,
0.07843017578125,
0.37744140625
],
[
0.08087158203125,
0.37890625,
0.46728515625,
0.37890625,
0.46728515625,
0.392578125,
0.08087158203125,
0.392578125
],
[
0.74267578125,
0.375,
0.78125,
0.375,
0.78125,
0.4697265625,
0.74267578125,
0.4697265625
],
[
0.5908203125,
0.38330078125,
0.59716796875,
0.38330078125,
0.59716796875,
0.389404296875,
0.5908203125,
0.389404296875
],
[
0.609375,
0.386474609375,
0.6650390625,
0.386474609375,
0.6650390625,
0.397705078125,
0.609375,
0.397705078125
],
[
0.08087158203125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.08087158203125,
0.4072265625
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.60888671875,
0.40087890625,
0.67626953125,
0.40087890625,
0.67626953125,
0.415283203125,
0.60888671875,
0.415283203125
],
[
0.080078125,
0.4072265625,
0.468017578125,
0.40771484375,
0.468017578125,
0.42236328125,
0.080078125,
0.421630859375
],
[
0.57861328125,
0.4072265625,
0.59326171875,
0.4072265625,
0.59326171875,
0.4248046875,
0.57861328125,
0.4248046875
],
[
0.5908203125,
0.416748046875,
0.59716796875,
0.416748046875,
0.59716796875,
0.42041015625,
0.5908203125,
0.42041015625
],
[
0.62109375,
0.417236328125,
0.66650390625,
0.417236328125,
0.66650390625,
0.431884765625,
0.62109375,
0.431884765625
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.080078125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.615234375,
0.434326171875,
0.67236328125,
0.434326171875,
0.67236328125,
0.448974609375,
0.615234375,
0.448974609375
],
[
0.0318603515625,
0.448974609375,
0.058013916015625,
0.448974609375,
0.058013916015625,
0.48095703125,
0.0318603515625,
0.48095703125
],
[
0.08087158203125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.08087158203125,
0.46728515625
],
[
0.08087158203125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.48095703125,
0.08087158203125,
0.48095703125
],
[
0.61767578125,
0.46533203125,
0.65625,
0.46728515625,
0.654296875,
0.479736328125,
0.61669921875,
0.477783203125
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.72607421875,
0.4873046875,
0.73779296875,
0.4892578125,
0.736328125,
0.494384765625,
0.7255859375,
0.492431640625
],
[
0.76953125,
0.488037109375,
0.78125,
0.4912109375,
0.779296875,
0.49560546875,
0.7685546875,
0.49365234375
],
[
0.810546875,
0.4873046875,
0.82861328125,
0.4873046875,
0.82861328125,
0.49560546875,
0.810546875,
0.49560546875
],
[
0.6845703125,
0.490478515625,
0.6953125,
0.490478515625,
0.6953125,
0.494384765625,
0.6845703125,
0.494384765625
],
[
0.640625,
0.4892578125,
0.650390625,
0.4892578125,
0.650390625,
0.494384765625,
0.640625,
0.494384765625
],
[
0.68310546875,
0.492431640625,
0.73876953125,
0.494384765625,
0.73779296875,
0.5087890625,
0.68115234375,
0.5068359375
],
[
0.02532958984375,
0.50244140625,
0.056365966796875,
0.50244140625,
0.056365966796875,
0.5517578125,
0.02532958984375,
0.5517578125
],
[
0.080078125,
0.5,
0.4697265625,
0.5,
0.4697265625,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.07843017578125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.07843017578125,
0.52783203125
],
[
0.50244140625,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.07843017578125,
0.52978515625,
0.471435546875,
0.52978515625,
0.471435546875,
0.54345703125,
0.07843017578125,
0.54345703125
],
[
0.5009765625,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54736328125,
0.5009765625,
0.54736328125
],
[
0.02777099609375,
0.54345703125,
0.053924560546875,
0.54345703125,
0.053924560546875,
0.5625,
0.02777099609375,
0.5625
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50341796875,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.56103515625,
0.50341796875,
0.56103515625
],
[
0.0269622802734375,
0.5595703125,
0.0555419921875,
0.5595703125,
0.0555419921875,
0.5771484375,
0.0269622802734375,
0.5771484375
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50341796875,
0.564453125,
0.890625,
0.564453125,
0.890625,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.080078125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57568359375,
0.890625,
0.57568359375,
0.890625,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.0196075439453125,
0.58544921875,
0.058013916015625,
0.58544921875,
0.060455322265625,
0.7109375,
0.0228729248046875,
0.7109375
],
[
0.08087158203125,
0.59130859375,
0.46728515625,
0.59130859375,
0.46728515625,
0.6015625,
0.08087158203125,
0.6015625
],
[
0.50341796875,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.6015625,
0.50341796875,
0.6015625
],
[
0.080078125,
0.60498046875,
0.4697265625,
0.60498046875,
0.4697265625,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50341796875,
0.60498046875,
0.89208984375,
0.60498046875,
0.89208984375,
0.61572265625,
0.50341796875,
0.61572265625
],
[
0.080078125,
0.619140625,
0.325927734375,
0.61865234375,
0.325927734375,
0.63330078125,
0.080078125,
0.6337890625
],
[
0.50244140625,
0.6171875,
0.890625,
0.6171875,
0.890625,
0.630859375,
0.50244140625,
0.630859375
],
[
0.50244140625,
0.6318359375,
0.8935546875,
0.6318359375,
0.8935546875,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.50244140625,
0.64453125,
0.7255859375,
0.64453125,
0.7255859375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.08087158203125,
0.66650390625,
0.2091064453125,
0.66650390625,
0.2091064453125,
0.68115234375,
0.08087158203125,
0.68115234375
],
[
0.08087158203125,
0.693359375,
0.4697265625,
0.693359375,
0.4697265625,
0.70703125,
0.08087158203125,
0.70703125
],
[
0.5048828125,
0.6875,
0.8896484375,
0.6875,
0.8896484375,
0.69775390625,
0.5048828125,
0.69775390625
],
[
0.5009765625,
0.70068359375,
0.8896484375,
0.70068359375,
0.8896484375,
0.7138671875,
0.5009765625,
0.7138671875
],
[
0.08251953125,
0.71044921875,
0.46728515625,
0.71044921875,
0.46728515625,
0.72021484375,
0.08251953125,
0.72021484375
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.080078125,
0.7236328125,
0.468017578125,
0.7236328125,
0.468017578125,
0.7373046875,
0.080078125,
0.7373046875
],
[
0.50244140625,
0.73095703125,
0.8935546875,
0.73095703125,
0.8935546875,
0.744140625,
0.50244140625,
0.744140625
],
[
0.080078125,
0.7392578125,
0.4697265625,
0.7392578125,
0.4697265625,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.50244140625,
0.7470703125,
0.890625,
0.7470703125,
0.890625,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.08087158203125,
0.755859375,
0.46728515625,
0.755859375,
0.46728515625,
0.7666015625,
0.08087158203125,
0.7666015625
],
[
0.50244140625,
0.76123046875,
0.890625,
0.76123046875,
0.890625,
0.7744140625,
0.50244140625,
0.7744140625
],
[
0.080078125,
0.76953125,
0.468017578125,
0.76953125,
0.468017578125,
0.78271484375,
0.080078125,
0.78271484375
],
[
0.50341796875,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50341796875,
0.79052734375
],
[
0.08087158203125,
0.78515625,
0.46728515625,
0.78515625,
0.46728515625,
0.798828125,
0.08087158203125,
0.798828125
],
[
0.5009765625,
0.79052734375,
0.6650390625,
0.79052734375,
0.6650390625,
0.8037109375,
0.5009765625,
0.8037109375
],
[
0.08087158203125,
0.7998046875,
0.468017578125,
0.7998046875,
0.468017578125,
0.81298828125,
0.08087158203125,
0.81298828125
],
[
0.080078125,
0.814453125,
0.4697265625,
0.814453125,
0.4697265625,
0.82763671875,
0.080078125,
0.82763671875
],
[
0.51953125,
0.8095703125,
0.89208984375,
0.81005859375,
0.89208984375,
0.82470703125,
0.51953125,
0.82373046875
],
[
0.080078125,
0.830078125,
0.468017578125,
0.830078125,
0.468017578125,
0.84326171875,
0.080078125,
0.84326171875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.8447265625,
0.4697265625,
0.8447265625,
0.4697265625,
0.85791015625,
0.080078125,
0.85791015625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85986328125,
0.4697265625,
0.85986328125,
0.4697265625,
0.87353515625,
0.080078125,
0.87353515625
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88623046875,
0.2254638671875,
0.88720703125,
0.2254638671875,
0.9013671875,
0.09722900390625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"DVMNet: Computing Relative Pose for Unseen Objects Beyond Hypotheses",
"Chen Zhaol",
"Tong Zhangl*Zheng Dang!",
"Mathieu Salzmann12",
"",
"1EPFL",
" 2ClearSpace SA",
"(chen.zhao, tong.zhang, zheng.dang, mathieu.salzmann)@epfl.ch",
"",
"Abstract",
"Hypothesis-Fre",
"Determining the relative pose of an object between two",
"IL.",
"a",
"",
"images is pivotal to the success of generalizable object pose",
"AOM",
"estimation.Existing approaches typically approximate the",
"",
"continuous pose representation with a large number of dis-",
"2",
"crete pose hypotheses, which incurs a computationally ex-",
"pensive process of scoring each hypothesis at test time.",
"pao",
"50",
"RelPose",
"By contrast, we present a Deep Voxel Matching Network",
"LAD",
"RelPose++",
"(DVMNet) that eliminates the need for pose hypotheses and",
"TOI",
"T",
"3DAHV",
"computes the relative object pose in a single pass. To this",
"end, we map the two input RGB images, reference and",
"DVMNet",
"SO",
"query, to their respective voxelized 3D representations. We",
"then pass the resulting voxels through a pose estimation",
"Ours",
"module, where the voxels are aligned and the pose is com-",
"0",
"qu",
"100",
"1",
"60",
"MACs(G)",
"1A68",
"puted in an end-to-end fashion by solving a least-squares",
"problem. To enhance robustness, we introduce a weighted",
"Figure 1. Advantages of our DVMNet compared to hypothesis-",
"closest voxel algorithm capable of mitigating the impact",
"based methods. Hypothesis-based techniques approximate the",
"6",
"of noisy voxels.We conduct extensive experiments on the",
"relativeobjectposeby scoring numerousposehypotheses,leading",
"3",
"CO3D, LINEMOD, and Objaverse datasets, demonstrating",
"to a high computational cost.By contrast,ourDVMNet computes",
"that our method delivers more accurate relative pose esti-",
"the pose in a hypothesis-free fashion by robustly matching vox-",
"C",
"matesfornovel objects atalowercomputationalcostcom-",
"elized 3Drepresentations of thereferenceandqueryimagesviaa",
"pared to state-of-the-art methods. Our code is released at:",
"Weighted Closest Voxel (WCV)algorithm.Our method strikes a",
"https://github.com/sailor-z/DVMNet/.",
"favorable balance between computational cost and accuracy in rel-",
"ative object pose estimation, as measured by multiply-accumulate",
"operations (MACs) and angular error.",
"1. Introduction",
"Object pose estimation plays a crucial role in 3D computer",
"depicting unseen objectsindiverse poses.servingasrefer-",
"ences. Object pose estimation is then carried out through",
"visionandroboticstasks[2.11.23.331,aimingtoestimate",
"template matching [21, 29] or by establishing 2D-3D cor-",
"the 3D rotation and 3D translation of an object depicted in",
"respondences [14, 32]. Unfortunately, the effectiveness of",
"an RGB image. The vast majority of existing methods work",
"these methods strongly depends on the references densely",
"under the assumption that the training and testing data in-",
"covering the viewpoints of unseen objects, making them in-",
"clude the same object instances, thereby limiting their appli-",
"applicable to practical scenarios where only sparse refer-",
"cability to scenarios that involve previously unseen objects.",
"ence views are available.",
"Recently, generalizable object pose estimation [21, 29, 32]",
"has received growing attention, showcasing the potential to",
"In this context, a few methods [18, 47, 52] highlight",
"generalize to unseen objects from new categories without",
"the importance of relative object pose estimation. Unlike",
"retraining the network. In pursuit of this generalization ca-",
"previous approaches in generalizable object pose estima-",
"pability,existing methods leverage densely sampled images",
"tion, these methods focus on estimating the relative pose",
"of an unseen object between two images, i.e., a single ref-",
"*Corresponding author.",
"erence image of the object and the query image. In this"
] |
{
"arxiv_id": "2403.13683",
"language": "en",
"page_idx": 0
}
|
||
2403.13683-en-1
|
2403.13683-en-1.jpg
|
paper, we also work in this setting, motivated by the practical ease of obtaining a single reference image for a new object. One plausible solution is to compute the relative pose based on 2D-2D correspondences [13]. However, the single-reference scenario tends to yield a significant viewpoint gap between the reference and the query. Existing studies [18, 52] have shown that image-matching techniques [26, 31] are sensitive to such pose differences. To handle this issue, the prior methods [18,47,52] follow an alternative strategy of scoring multiple pose hypotheses for the input reference-query pair, and predicting the pose as the hypothesis with the highest score. However, this strategy comes with the drawback of requiring numerous pose hypotheses to achieve reasonable accuracy, e.g., 500,000 in [18], which thus induces a computational burden. Moreover, we empirically found that these approaches occasionally produce unnaturally large errors. One plausible explanation is their failure to model the continuous nature of the object pose space, as they primarily concentrate on learning to score discrete hypotheses.
To overcome these drawbacks, we present a novel Deep Voxel Matching Network (DVMNet) that computes the relative pose of unseen objects in a $hypothesis-free$ and endto-end manner. Our approach starts by voxelizing the query and reference images in a dedicated autoencoder. The encoder network lifts 2D image features to 3D voxels, leveraging cross-view 3D information. The decoder network reconstructs a masked object image from the voxels, encouraging the learned voxels to account for the object. Subsequently, we align the query and reference voxels based on a score matrix that measures the voxel similarities. To handle unreliable voxels due to background, varying illumination, and other potential nuisances, we present a Weighted Closest Voxel ($\hat{\mathrm{WCV}})$ algorithm to facilitate robust pose estimation. In this algorithm, each voxel-voxel correspondence is assigned a confidence score computed by utilizing both the 3D voxel objectness map and the 2D object mask learned by the autoencoder. The relative object pose is then computed by solving a weighted least-squares problem. Such an end-to-end learning mechanism eliminates the necessity for voxel-wise annotations and allows the network to directly learn pose-aware features from RGB images. As illustrated in Fig. 1, our DVMNet requires significantly fewer multiply-accumulate operations (MACs) while achieving smaller angular errors than its hypothesis-based competitors.
We perform comprehensive experiments on the CO3D[25], LINEMOD [15], and Objaverse[9] datasets. Our method yields more accurate and robust relative pose estimates for previously unseen objects than both image-matching and hypothesis-based competitors. We also conduct ablation studies where the results demonstrate the effectiveness of the key components in our framework.
In short, our contributions are twofolds: First, we tackle the problem of relative pose estimation for unseen objects in a hypothesis-free manner by introducing a deep voxel matching network. Second, we present a weighted closest voxel algorithm that robustly and end-to-end computes the object pose from voxel-voxel correspondences.
## $2. \textbf{ Related Work}$
Instance-Level Object Pose Estimation. The majority of previous deep learning approaches to object pose estimation [24, 30, 34, 35, 43] tackle the problem at an instance level, holding an assumption that the training and testing data depict the same object instances. Since the appearance of an object instance in different poses typically exhibits limited variations, these methods provide highly accurate object pose estimates. Nevertheless, they struggle to generalize to previously unseen objects during testing without retraining the network, as has been observed in the literature [21, 29, 32]. This limitation constrains their applicability in real-world scenarios that often involve diverse object instances. This has been remedied to a degree by categorylevel object pose estimation methods [7,19,36]. In this scenario, the testing images comprise new object instances from specific categories already included in the training data. Although these methods have achieved promising generalization ability within the predefined object categories, they become ineffective when facing objects from entirely new categories.
Generalizable Object Pose Estimation. To tackle the scenario of unseen objects from new categories, there has been growing interest in generalizable object pose estimation. When a textured 3D mesh is available for an unseen object, some approaches [29, 42, 51] suggest generating synthetic images as references by rendering the 3D mesh from various viewpoints. Given a query image that depicts this object, a template matching paradigm is utilized to identify the most similar reference and approximate the object pose in the query as that of the selected reference. Some methods bypass the need for 3D meshes by assuming the availability of multiple real reference images. Object pose estimation is then carried out by employing either a template matching strategy [21] or a 3D object reconstruction technique [14, 32]. Nevertheless, all of these methods rely on having access to dense-view reference images, which limits their applicability in scenarios where only sparse views of references are available.
Relative Object Pose Estimation. Recently, several studies [18, 37, 47, 52] have highlighted the importance of relative object pose estimation. These methods stand out in generalizable object pose estimation due to their key advantage of requiring only a single reference image. The objective of these methods is to estimate the relative object pose between the input query image and the reference. Since the
|
[
[
0.07843017578125,
0.0921630859375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.07843017578125,
0.106689453125
],
[
0.50244140625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08251953125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.8935546875,
0.1383056640625,
0.8935546875,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.07843017578125,
0.1685791015625,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.07843017578125,
0.18310546875
],
[
0.50244140625,
0.1685791015625,
0.81201171875,
0.1685791015625,
0.81201171875,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.07843017578125,
0.1976318359375,
0.4697265625,
0.1962890625,
0.4697265625,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.50341796875,
0.1962890625,
0.640625,
0.1962890625,
0.640625,
0.2108154296875,
0.50341796875,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.2203369140625,
0.89208984375,
0.2215576171875,
0.89208984375,
0.236083984375,
0.5009765625,
0.23486328125
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.471435546875,
0.24365234375,
0.471435546875,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.238037109375,
0.89208984375,
0.238037109375,
0.89208984375,
0.251953125,
0.50244140625,
0.251953125
],
[
0.07843017578125,
0.25830078125,
0.468017578125,
0.2568359375,
0.468017578125,
0.271484375,
0.07843017578125,
0.272705078125
],
[
0.50244140625,
0.25244140625,
0.89208984375,
0.25244140625,
0.89208984375,
0.266357421875,
0.50244140625,
0.266357421875
],
[
0.50244140625,
0.267578125,
0.89208984375,
0.267578125,
0.89208984375,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.08087158203125,
0.27587890625,
0.468017578125,
0.27587890625,
0.468017578125,
0.285888671875,
0.08087158203125,
0.285888671875
],
[
0.50341796875,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.50341796875,
0.296630859375
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.29736328125,
0.89208984375,
0.29736328125,
0.89208984375,
0.311279296875,
0.50244140625,
0.311279296875
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50341796875,
0.314453125,
0.890625,
0.314453125,
0.890625,
0.324462890625,
0.50341796875,
0.324462890625
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.5009765625,
0.326904296875,
0.890625,
0.32763671875,
0.890625,
0.34228515625,
0.5009765625,
0.341552734375
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.343505859375,
0.89208984375,
0.343505859375,
0.89208984375,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50244140625,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.50244140625,
0.371826171875
],
[
0.08087158203125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.50341796875,
0.373779296875,
0.890625,
0.373779296875,
0.890625,
0.386962890625,
0.50341796875,
0.386962890625
],
[
0.080078125,
0.380126953125,
0.26953125,
0.380126953125,
0.26953125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.38818359375,
0.8935546875,
0.38818359375,
0.8935546875,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.098876953125,
0.3984375,
0.468017578125,
0.3984375,
0.468017578125,
0.412353515625,
0.098876953125,
0.412353515625
],
[
0.50341796875,
0.404052734375,
0.89208984375,
0.404052734375,
0.89208984375,
0.417236328125,
0.50341796875,
0.417236328125
],
[
0.080078125,
0.412353515625,
0.468017578125,
0.41357421875,
0.468017578125,
0.427978515625,
0.080078125,
0.4267578125
],
[
0.50341796875,
0.42041015625,
0.890625,
0.42041015625,
0.890625,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.080078125,
0.4287109375,
0.468017578125,
0.4287109375,
0.468017578125,
0.442626953125,
0.080078125,
0.442626953125
],
[
0.5048828125,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.44580078125,
0.5048828125,
0.44580078125
],
[
0.080078125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.458251953125,
0.080078125,
0.458251953125
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.5009765625,
0.462158203125,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.5009765625,
0.4765625
],
[
0.08087158203125,
0.47607421875,
0.46728515625,
0.47607421875,
0.46728515625,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.477783203125,
0.89208984375,
0.492431640625,
0.50244140625,
0.49365234375
],
[
0.07843017578125,
0.4892578125,
0.468017578125,
0.488037109375,
0.468017578125,
0.50244140625,
0.07843017578125,
0.50390625
],
[
0.50244140625,
0.494384765625,
0.8935546875,
0.494384765625,
0.8935546875,
0.50830078125,
0.50244140625,
0.50830078125
],
[
0.080078125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.50244140625,
0.5087890625,
0.6611328125,
0.5087890625,
0.6611328125,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.07843017578125,
0.51953125,
0.468017578125,
0.51806640625,
0.468017578125,
0.53271484375,
0.07843017578125,
0.5341796875
],
[
0.50244140625,
0.5234375,
0.890625,
0.52294921875,
0.890625,
0.537109375,
0.50244140625,
0.53857421875
],
[
0.08087158203125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50244140625,
0.54052734375,
0.89208984375,
0.54052734375,
0.89208984375,
0.5537109375,
0.50244140625,
0.5537109375
],
[
0.08087158203125,
0.5517578125,
0.46728515625,
0.5517578125,
0.46728515625,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50341796875,
0.568359375
],
[
0.08087158203125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.07843017578125,
0.59326171875,
0.4697265625,
0.5947265625,
0.4697265625,
0.609375,
0.07843017578125,
0.60791015625
],
[
0.50244140625,
0.60107421875,
0.89208984375,
0.60107421875,
0.89208984375,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.08087158203125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.63720703125,
0.08087158203125,
0.63720703125
],
[
0.50244140625,
0.6298828125,
0.890625,
0.6298828125,
0.890625,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50341796875,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.080078125,
0.65576171875,
0.468017578125,
0.65576171875,
0.468017578125,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66943359375,
0.46728515625,
0.66943359375,
0.46728515625,
0.68310546875,
0.080078125,
0.68310546875
],
[
0.50341796875,
0.6767578125,
0.8896484375,
0.6767578125,
0.8896484375,
0.6875,
0.50341796875,
0.6875
],
[
0.080078125,
0.68603515625,
0.4697265625,
0.68603515625,
0.4697265625,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.70263671875,
0.46728515625,
0.70263671875,
0.46728515625,
0.71337890625,
0.08087158203125,
0.71337890625
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73095703125,
0.468017578125,
0.73095703125,
0.468017578125,
0.744140625,
0.080078125,
0.744140625
],
[
0.50341796875,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50341796875,
0.75
],
[
0.080078125,
0.7470703125,
0.4697265625,
0.7470703125,
0.4697265625,
0.76025390625,
0.080078125,
0.76025390625
],
[
0.50341796875,
0.75244140625,
0.8896484375,
0.75244140625,
0.8896484375,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76123046875,
0.4697265625,
0.76123046875,
0.4697265625,
0.7744140625,
0.080078125,
0.7744140625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.77783203125,
0.16259765625,
0.77783203125,
0.16259765625,
0.7890625,
0.08087158203125,
0.7890625
],
[
0.50244140625,
0.77978515625,
0.66015625,
0.77978515625,
0.66015625,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.098876953125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.098876953125,
0.81005859375
],
[
0.50341796875,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.8095703125,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.080078125,
0.82373046875
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82470703125,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.8388671875
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82470703125,
0.89208984375,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.5009765625,
0.8994140625
]
] |
[
"paper, we also work in this setting, motivated by the prac-",
"In short, our contributions are twofolds: First, we tackle",
"tical ease of obtaining a single reference image for a new",
"the problem of relative pose estimation for unseen objects",
"object.Oneplausible solutionisto compute therelative",
"in a hypothesis-free manner by introducing a deep voxel",
"pose based on 2D-2D correspondences [13]. However, the",
"matching network. Second, we present a weighted closest",
"single-reference scenario tends to yield a significant view-",
"voxel algorithm that robustly and end-to-end computes the",
"point gap between the reference and the query.Exist-",
"object pose from voxel-voxel correspondences.",
"ing studies [18, 52] have shown that image-matching tech-",
"niques [26, 31] are sensitive to such pose differences. To",
"2.Related Work",
"handle this issue, the prior methods [18, 47, 52] follow an",
"Instance-Level Object Pose Estimation. The majority of",
"alternative strategy of scoring multiple pose hypotheses for",
"the input reference-query pair, and predicting the pose as",
"previous deep learning approaches to object pose estima-",
"the hypothesis with the highest score. However, this strat-",
"tion [24, 30, 34, 35, 43] tackle the problem at an instance",
"level, holding an assumption that the training and testing",
"egycomeswiththedrawbackofrequiringnumerouspose",
"data depict the same object instances. Since the appearance",
"hypotheses to achieve reasonable accuracy, e.g., 500,000",
"of an object instance in different poses typically exhibits",
"in [18], which thus induces a computational burden. More-",
"limited variations,thesemethodsprovidehighlyaccurate",
"over,weempiricallyfoundthattheseapproachesoccasion-",
"object pose estimates. Nevertheless, they struggle to gen-",
"ally produce unnaturally large errors. One plausible expla-",
"eralize to previously unseen objects during testing without",
"nationistheirfailuretomodel thecontinuousnature of the",
"retraining the network, as has been observed in the litera-",
"object pose space, as they primarily concentrate on learning",
"ture [21, 29, 32]. This limitation constrains their applicabil-",
"to score discrete hypotheses.",
"ity in real-world scenarios that often involve diverse object",
"To overcome these drawbacks, we present a novel Deep",
"instances. This has been remedied to a degree by category-",
"Voxel Matching Network (DVMNet) that computes the rel-",
"level objectpose estimationmethods[7,19,36].Inthis",
"ative pose of unseen objects in a hypothesis-free and end-",
"scenario,the testing images comprise new objectinstances",
"to-end manner. Our approach starts by voxelizing the query",
"from specific categories already included in the training",
"and reference images in a dedicated autoencoder. The en-",
"data. Although these methods have achieved promising",
"codernetworklifts2Dimagefeaturesto3Dvoxels,lever-",
"generalization ability within the predefined object cate-",
"aging cross-view 3D information. The decoder network re-",
"gories, they become ineffective when facing objects from",
"constructs a masked object image from the voxels, encour-",
"entirely new categories.",
"aging the learned voxels to account for the object. Subse-",
"Generalizable Object Pose Estimation. To tackle the sce-",
"quently, we align the query and reference voxels based on",
"nario of unseen objects from new categories, there has been",
"a score matrix thatmeasures the voxel similarities. Tohan-",
"growing interest in generalizable object pose estimation.",
"dle unreliable voxels due to background, varying illumina-",
"When a textured 3D mesh is available for an unseen ob-",
"tion, and other potential nuisances, we present a Weighted",
"ject, some approaches [29, 42, 51] suggest generating syn-",
"Closest Voxel (WCV) algorithm to facilitate robust pose es-",
"thetic images as references by rendering the 3D mesh from",
"timation. In this algorithm, each voxel-voxel correspon-",
"various viewpoints. Given a query image that depicts this",
"dence is assigned a confidence score computed by utiliz-",
"object, a template matching paradigm is utilized to identify",
"ing both the 3D voxel objectness map and the 2D object",
"the most similar reference and approximate the object pose",
"mask learned by the autoencoder. The relative object pose",
"in the query as that of the selected reference. Some meth-",
"is then computed by solving a weighted least-squares prob-",
"odsbypass theneedfor3Dmeshesby assumingtheavail-",
"lem. Such an end-to-end learning mechanism eliminates",
"ability of multiple real reference images. Object pose esti-",
"thenecessityfor voxel-wise annotations and allows the net-",
"mation is then carried out by employing either a template",
"work to directly learn pose-aware features from RGB im-",
"matching strategy [21] or a 3D object reconstruction tech-",
"ages. As illustrated in Fig. 1, our DVMNet requires signifi-",
"nique [14, 32]. Nevertheless, all of these methods rely on",
"cantly fewer multiply-accumulate operations (MACs) while",
"having accessto dense-view reference images,which limits",
"achieving smaller angular errors than its hypothesis-based",
"their applicability in scenarios where only sparse views of",
"competitors.",
"references areavailable",
"We perform comprehensive experiments on the",
"Relative Object Pose Estimation. Recently, several stud-",
"CO3D [25], LINEMOD [15], and Objaverse [9] datasets.",
"ies [18, 37, 47, 52] have highlighted the importance of rel-",
"Our method yields more accurate and robust relative",
"ative object pose estimation. These methods stand out in",
"pose estimates for previously unseen objects than both",
"generalizable object pose estimation due to their key advan-",
"image-matching and hypothesis-based competitors.We",
"tage of requiring only a single reference image. The objec-",
"also conduct ablation studies where the results demonstrate",
"tive of these methods is to estimate the relative object pose",
"the effectiveness of the key components in our framework.",
"between the input query image and the reference. Since the"
] |
{
"arxiv_id": "2403.13683",
"language": "en",
"page_idx": 1
}
|
||
2403.13683-en-2
|
2403.13683-en-2.jpg
|
single-reference assumption tends to result in a large object pose difference between the query and the reference, unseen object pose estimation becomes more challenging. Intuitively, one could establish pixel-pixel correspondences between the two images and compute the relative object pose based on multi-view geometry[13]. However, as reported in the literature [18,52] and also in our experiments, even the state-of-the-art image-matching techniques [12,26,31] cannot deliver accurate pose estimates when confronted with large object pose differences. To address this, existing methods [18,47,52] suggest approximating the relative object pose via a discrete set of pose hypotheses, and learning to maximize scores of the positive hypotheses. Since object pose lies in a continuous space [53], accurately approximating the pose necessitates a vast number of pose hypotheses, which makes the hypothesis-based approaches computationally expensive. Moreover, scoring discrete samples lacks an understanding of the continuous pose distribution, leading to failure cases with unnaturally high pose estimation errors. By contrast, we present a hypothesis-free technique that is capable of computing the relative object pose in a single pass via deep voxel matching.
## $3. \textbf{Method}$
## $3. 1. \textbf{Problem Formulation}$
We tackle the problem of estimating the relative pose $\Delta\mathbf{P}$ $\begin{array}{l}\text{of a previously-unseen object between a query image I}_{q}\mathrm{~and}\\\text{a reference image I}_{r.~In~this~scenario,~the~objects~present~in}\end{array}$ the testing set $\Omega_\mathrm{test}$ differ from those in the training set $\Omega_\mathrm{train}$, and the goal is to handle the unseen objects without retraining the network. This requires addressing the challenges of generalization to unseen objects in $\Omega_\mathrm{test}$ and robustness to the large object pose difference between $\mathbf{I}_q$ and $\mathbf{I}_r.$
Since the relative object translation estimation can be facilitated by using zero-shot object detectors [17,39,50], our focus is on predicting the relative object rotation, i.e., $\Delta\mathbf{R}\in SO(3)$,which is more challenging. This choice aligns with existing literature [42,44,51,52]. Previous hypothesis-based approaches[18,47,52] approximate $\Delta\mathbf{R}$ by sampling discrete pose hypotheses and maximizing the scores of positive samples. The problem is formulated as
# 3.2. Motivation
Drawing inspiration from the success of pixel-pixel correspondences in image matching [26,45,49] a natural approach to avoiding the use of pose hypotheses would be to compute the relative object pose based on 2D correspondences. However, recent studies [18,52] have observed that such an image-matching strategy is unreliable in the scenario of object pose estimation. We trace this limitation back to the fact that image-matching methods are not fully differentiable w.r.t. the pose. Specifically, some approaches [45,46,48] encode a notion of consistency among the pixel-pixel correspondences utilizing the essential matrix. However, computing the pose from the essential matrix leads to multiple solutions [13]. The pose estimation is thus detached from the learning process as a post-processing step. Notably, in the context of object pose estimation, the pre-generated correspondences tend to be unreliable in the presence of challenges such as large object pose differences and textureless objects. Therefore, the isolated pose estimation step in the two-stage design becomes less effective.
To address this issue, we propose to lift the input images to voxelized 3D representations [40] and perform the matching process in 3D space. Therefore, the computation of the relative object pose from the resulting voxel-voxel matches becomes a differentiable operation. This characteristic enables us to directly supervise the entire framework with the actual quantity we aim to predict, i.e., the relative object pose. Subsequently, we elaborate on the steps involved in the presented hypothesis-free mechanism.
# 3.3. Image Voxelization
To achieve the pose estimation from voxel-voxel correspondences with only RGB images as input, we first need to lift each RGB image to a set of 3D voxels. To enable such a voxelization, we introduce an autoencoder network depicted in Fig. 2, which includes a 2D-3D encoder and a 3D- 2D decoder. Specifically, we employ a pretrained vision transformer [41] to convert the query and reference images to 2D feature embeddings denoted as $\mathbf{F}_q$ and $\mathbf{F}_r$, respectively. Considering the difficulty of lifting 2D images to 3D representations, we incorporate a cross-attention module to capture cross-view 3D information. We take the feature embedding $\mathbf{F}_q$ as an example (a symmetric process is carried out for $\mathbf{F}_{r}).$ The cross-attention module [41] is defined as
(2)
(3)
$$\begin{aligned}&\tilde{\mathbf{F}}_{q}^{l}=\mathrm{MHSA}(\mathrm{LN}(\mathbf{F}_{q}^{l-1}))+\mathbf{F}_{q}^{l-1},\\&\hat{\mathbf{F}}_{q}^{l}=\mathrm{MHCA}(\mathrm{LN}(\tilde{\mathbf{F}}_{q}^{l}),\mathrm{LN}(\mathbf{F}_{r}^{l-1}))+\tilde{\mathbf{F}}_{q}^{l},\\&\mathbf{F}_{q}^{l}=\mathrm{FFN}(\mathrm{LN}(\hat{\mathbf{F}}_{q}^{l}))+\hat{\mathbf{F}}_{q}^{l},\end{aligned}$$
(4)
where MHSA stands for a multi-head self-attention layer, MHCA represents a multi-head cross-attention layer that
(1)
$$\Delta\hat{\mathbf{R}}=\arg\max_{\Delta\mathbf{R}_{i}\in\mathcal{R}}f(\mathbf{I}_{q},\mathbf{I}_{r},\Delta\mathbf{R}_{i}),$$
where $\mathcal{R}$ denotes the set of discrete pose hypotheses. Achieving a decent approximation accuracy requires a large number of hypotheses, e.g., 500,000 in [18]. By contrast, we present a hypothesis-free technique that computes $\Delta\mathbf{R}$ ${\hat{\text{in a single pass as }}\Delta\hat{\mathbf{R}}}=g({\mathbf{I}}_{q},{\mathbf{I}}_{r}).$
|
[
[
0.080078125,
0.0921630859375,
0.47216796875,
0.0921630859375,
0.47216796875,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.62158203125,
0.091552734375,
0.62158203125,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.11553955078125,
0.890625,
0.11553955078125,
0.890625,
0.1287841796875,
0.50244140625,
0.1287841796875
],
[
0.080078125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.5009765625,
0.1307373046875,
0.890625,
0.1300048828125,
0.890625,
0.14453125,
0.5009765625,
0.1458740234375
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.07843017578125,
0.1514892578125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.50244140625,
0.146484375,
0.89208984375,
0.146484375,
0.89208984375,
0.160400390625,
0.50244140625,
0.160400390625
],
[
0.50341796875,
0.1622314453125,
0.8896484375,
0.1622314453125,
0.8896484375,
0.17236328125,
0.50341796875,
0.17236328125
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.5009765625,
0.175537109375,
0.89208984375,
0.175537109375,
0.89208984375,
0.189453125,
0.5009765625,
0.189453125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1912841796875,
0.89208984375,
0.1912841796875,
0.89208984375,
0.2052001953125,
0.50244140625,
0.2052001953125
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.205810546875,
0.890625,
0.205810546875,
0.890625,
0.2197265625,
0.50244140625,
0.2197265625
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.07843017578125,
0.2279052734375,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.50244140625,
0.23486328125,
0.890625,
0.236083984375,
0.890625,
0.250732421875,
0.50244140625,
0.25
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.251953125,
0.89208984375,
0.251953125,
0.89208984375,
0.26513671875,
0.50341796875,
0.26513671875
],
[
0.07843017578125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.07843017578125,
0.272705078125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.268310546875,
0.8896484375,
0.268310546875,
0.8896484375,
0.279052734375,
0.50341796875,
0.279052734375
],
[
0.50244140625,
0.2822265625,
0.89208984375,
0.2822265625,
0.89208984375,
0.29541015625,
0.50244140625,
0.29541015625
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.29736328125,
0.89208984375,
0.29736328125,
0.89208984375,
0.311279296875,
0.50244140625,
0.311279296875
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50244140625,
0.311279296875,
0.89208984375,
0.3125,
0.89208984375,
0.326904296875,
0.50244140625,
0.32568359375
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50341796875,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.50244140625,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.50244140625,
0.35595703125
],
[
0.080078125,
0.348388671875,
0.4697265625,
0.348388671875,
0.4697265625,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.50244140625,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.50244140625,
0.37060546875
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.080078125,
0.37744140625,
0.46728515625,
0.37744140625,
0.46728515625,
0.392578125,
0.080078125,
0.392578125
],
[
0.50244140625,
0.37255859375,
0.890625,
0.37255859375,
0.890625,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.50244140625,
0.386962890625,
0.87109375,
0.386962890625,
0.87109375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.52294921875,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.52294921875,
0.416748046875
],
[
0.080078125,
0.410400390625,
0.34716796875,
0.410400390625,
0.34716796875,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50244140625,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.50341796875,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.44580078125,
0.50341796875,
0.44580078125
],
[
0.080078125,
0.448974609375,
0.1683349609375,
0.448974609375,
0.1683349609375,
0.46337890625,
0.080078125,
0.46337890625
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.50244140625,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.080078125,
0.4765625,
0.281005859375,
0.4765625,
0.281005859375,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.50341796875,
0.49365234375,
0.8935546875,
0.49365234375,
0.8935546875,
0.5068359375,
0.50341796875,
0.5068359375
],
[
0.07843017578125,
0.50244140625,
0.46728515625,
0.501953125,
0.46728515625,
0.5166015625,
0.07843017578125,
0.517578125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.08087158203125,
0.52099609375,
0.46728515625,
0.52099609375,
0.46728515625,
0.53076171875,
0.08087158203125,
0.53076171875
],
[
0.50244140625,
0.5234375,
0.84326171875,
0.5234375,
0.84326171875,
0.537109375,
0.50244140625,
0.537109375
],
[
0.080078125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.07843017578125,
0.54736328125,
0.468017578125,
0.548828125,
0.468017578125,
0.564453125,
0.07843017578125,
0.56298828125
],
[
0.50244140625,
0.548828125,
0.68310546875,
0.5498046875,
0.68310546875,
0.564453125,
0.50244140625,
0.56298828125
],
[
0.080078125,
0.564453125,
0.46728515625,
0.564453125,
0.46728515625,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.50244140625,
0.57373046875,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.080078125,
0.5947265625,
0.4697265625,
0.5947265625,
0.4697265625,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50244140625,
0.5888671875,
0.89208984375,
0.5888671875,
0.89208984375,
0.60302734375,
0.50244140625,
0.60302734375
],
[
0.07843017578125,
0.60986328125,
0.41259765625,
0.60986328125,
0.41259765625,
0.62353515625,
0.07843017578125,
0.62353515625
],
[
0.5009765625,
0.60400390625,
0.8935546875,
0.60302734375,
0.8935546875,
0.6171875,
0.5009765625,
0.61865234375
],
[
0.50341796875,
0.62060546875,
0.88818359375,
0.62060546875,
0.88818359375,
0.630859375,
0.50341796875,
0.630859375
],
[
0.098876953125,
0.626953125,
0.465576171875,
0.626953125,
0.465576171875,
0.64013671875,
0.098876953125,
0.64013671875
],
[
0.50244140625,
0.6337890625,
0.890625,
0.63330078125,
0.890625,
0.64794921875,
0.50244140625,
0.6484375
],
[
0.080078125,
0.64306640625,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.5009765625,
0.64794921875,
0.89208984375,
0.6484375,
0.89208984375,
0.66357421875,
0.5009765625,
0.662109375
],
[
0.080078125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.080078125,
0.671875
],
[
0.50244140625,
0.66552734375,
0.89208984375,
0.66552734375,
0.89208984375,
0.6787109375,
0.50244140625,
0.6787109375
],
[
0.080078125,
0.671875,
0.4697265625,
0.67236328125,
0.4697265625,
0.6875,
0.080078125,
0.68603515625
],
[
0.50244140625,
0.6796875,
0.89208984375,
0.6796875,
0.89208984375,
0.693359375,
0.50244140625,
0.693359375
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.5009765625,
0.69580078125,
0.89208984375,
0.693359375,
0.89208984375,
0.70703125,
0.5009765625,
0.708984375
],
[
0.07843017578125,
0.70263671875,
0.468017578125,
0.7021484375,
0.468017578125,
0.71630859375,
0.07843017578125,
0.71728515625
],
[
0.5009765625,
0.708984375,
0.890625,
0.708984375,
0.890625,
0.72265625,
0.5009765625,
0.72265625
],
[
0.08087158203125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.50244140625,
0.724609375,
0.890625,
0.724609375,
0.890625,
0.73779296875,
0.50244140625,
0.73779296875
],
[
0.080078125,
0.734375,
0.4697265625,
0.734375,
0.4697265625,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50244140625,
0.74072265625,
0.89208984375,
0.74072265625,
0.89208984375,
0.75390625,
0.50244140625,
0.75390625
],
[
0.50244140625,
0.755859375,
0.89208984375,
0.755859375,
0.89208984375,
0.76953125,
0.50244140625,
0.76953125
],
[
0.1650390625,
0.77978515625,
0.38232421875,
0.78076171875,
0.38232421875,
0.798828125,
0.1650390625,
0.79736328125
],
[
0.443603515625,
0.7822265625,
0.47216796875,
0.7822265625,
0.47216796875,
0.798828125,
0.443603515625,
0.798828125
],
[
0.2197265625,
0.796875,
0.271240234375,
0.796875,
0.271240234375,
0.806640625,
0.2197265625,
0.806640625
],
[
0.55810546875,
0.79443359375,
0.78759765625,
0.79541015625,
0.78759765625,
0.81298828125,
0.55810546875,
0.8125
],
[
0.86767578125,
0.79736328125,
0.89453125,
0.79736328125,
0.89453125,
0.814453125,
0.86767578125,
0.814453125
],
[
0.55810546875,
0.81640625,
0.83251953125,
0.8173828125,
0.83251953125,
0.8359375,
0.55810546875,
0.83544921875
],
[
0.86669921875,
0.8193359375,
0.89453125,
0.8193359375,
0.89453125,
0.8359375,
0.86669921875,
0.8359375
],
[
0.080078125,
0.82568359375,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5556640625,
0.8359375,
0.74169921875,
0.8388671875,
0.74169921875,
0.85986328125,
0.5556640625,
0.85693359375
],
[
0.86767578125,
0.84033203125,
0.89453125,
0.84033203125,
0.89453125,
0.85693359375,
0.86767578125,
0.85693359375
],
[
0.080078125,
0.85595703125,
0.46728515625,
0.85595703125,
0.46728515625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.08087158203125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50341796875,
0.869140625,
0.890625,
0.87060546875,
0.890625,
0.88525390625,
0.50341796875,
0.8837890625
],
[
0.0767822265625,
0.88330078125,
0.316162109375,
0.88525390625,
0.316162109375,
0.90283203125,
0.0767822265625,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.5009765625,
0.8994140625
]
] |
[
"single-reference assumption tends to result in a large object",
"3.2.Motivation",
"pose difference between the query and the reference, unseen",
"Drawing inspiration from the success of pixel-pixel corre-",
"object pose estimation becomes more challenging. Intu-",
"spondences in image matching [26, 45, 49] a natural ap-",
"itively, one could establish pixel-pixel correspondences be-",
"tween the two images and compute the relative object pose",
"proach to avoiding the use of pose hypotheses would be to",
"compute the relative objectpose based on 2D correspon-",
"based on multi-view geometry [13]. However, as reported",
"dences. However, recent studies [18, 52] have observed",
"in the literature [18, 52] and also in our experiments, even",
"that such an image-matching strategy is unreliable in the",
"the state-of-the-artimage-matchingtechniques[12,26,31]",
"scenario of object pose estimation.We trace this limita-",
"cannot deliver accurate pose estimates when confronted",
"tion back to the fact that image-matching methods are not",
"withlargeobjectposedifferences.Toaddress this, existing",
"fully differentiable w.r.t. the pose. Specifically, some ap-",
"methods [18, 47, 52] suggest approximating the relative ob-",
"proaches [45, 46, 48] encode a notion of consistency among",
"ject pose via a discrete set of pose hypotheses, and learning",
"to maximize scores of the positive hypotheses. Since object",
"the pixel-pixel correspondences utilizing the essential ma-",
"trix. However, computing the pose from the essential matrix",
"pose lies in a continuous space [53], accurately approximat-",
"leads to multiple solutions [13]. The pose estimation is thus",
"ing the pose necessitates a vast number of pose hypothe-",
"detached from the learning process as a post-processing",
"ses.whichmakesthehypothesis-basedapproachescompu-",
"step. Notably, in the context of object pose estimation, the",
"tationally expensive. Moreover, scoring discrete samples",
"pre-generated correspondences tend to be unreliable in the",
"lacks an understanding of the continuous pose distribution,",
"presence of challenges such as large object pose differences",
"leading to failure cases with unnaturally high pose estima-",
"tion errors. By contrast, we present a hypothesis-free tech-",
"and textureless objects. Therefore, the isolated pose estima-",
"tion step in the two-stage design becomes less effective.",
"nique that is capable of computing the relative object pose",
"To address this issue, we propose to lift the input im-",
"in a single pass via deep voxel matching.",
"ages to voxelized 3D representations [40] and perform the",
"matchingprocessin3Dspace.Therefore,thecomputation",
"3. Method",
"of the relative object pose from the resulting voxel-voxel",
"matches becomes a differentiable operation. This charac-",
"3.1.Problem Formulation",
"teristic enables us to directly supervise the entire framework",
"with the actual quantity we aim to predict, i.e., the relative",
"We tackle the problem of estimating the relative pose △P",
"object pose. Subsequently, we elaborate on the steps in-",
"ofapreviously-unseenobjectbetweenaqueryimageIaand",
"volved in the presented hypothesis-free mechanism.",
"a reference image I. In this scenario, the objects present in",
"the testing set Ztest differ from those in the training set Zurain",
"3.3. Image Voxelization",
"and the goal is to handle the unseen objects without retrain-",
"To achieve the pose estimation from voxel-voxel correspon-",
"ing the network. This requires addressing the challenges of",
"generalization to unseen objects in 2test and robustness to",
"dences with only RGB images as input, we first need to lift",
"the large object pose difference between Ig and Ir.",
"each RGB image to a set of 3D voxels. To enable such",
"a voxelization, we introduce an autoencoder network de-",
"Since the relative obiject translation estimation can be fa-",
"picted in Fig. 2, which includes a 2D-3D encoder and a 3D-",
"cilitated by using zero-shot object detectors [17, 39, 50]",
"2D decoder.Specifically,we employ a pretrained vision",
"our focus is on predicting the relative object rotation, i.e.",
"transformer [41] to convert the query and reference images",
"R E SO(3), which is more challenging. This choice",
"to 2D feature embeddings denoted as F. and Fr, respec-",
"aligns with existing literature [42, 44, 51, 52].Previous",
"tively.Consideringthedifficultyof lifting2Dimagesto3D",
"hypothesis-based approaches [18,47,52] approximate △R",
"representations, we incorporate a cross-attention module to",
"by sampling discrete pose hypotheses and maximizing the",
"capture cross-view 3D information. We take the feature em-",
"scores of positive samples. The problem is formulated as",
"bedding F. as an example (a symmetric process is carried",
"out for F,). The cross-attention module [41] is defined as",
"AR = arg max f(Ig, Ir, AR,),",
"(1)",
"ARER",
"F, = MHSA(LN(Fl-1)) + F}-1,",
"(2)",
"F, = MHCA(LN(F), LN(F}-1)) + Fa",
"(3)",
"where R denotes the set of discrete pose hypotheses.",
"F% = FFN(LN(F)) + Fb,",
"(4)",
"number of hypotheses, e.g., 500,000 in [18]. By contrast,",
"Achieving a decent approximation accuracy requires a large",
"we present a hypothesis-free technique that computes R",
"where MHSA stands for a multi-head self-attention layer,",
"in a single pass as △R = g(Iq, I,).",
"MHCA represents a multi-head cross-attention layer that"
] |
{
"arxiv_id": "2403.13683",
"language": "en",
"page_idx": 2
}
|
||
2403.13683-en-3
|
2403.13683-en-3.jpg
|
<FigureHere>
Figure 2. Network architecture of our autoencoder. The encoder takes two RGB images, query and reference, as input and lifts their 2D feature embeddings to 3D voxels by leveraging cross-view 3D information. $\mathbf{O}_q$ and $\mathbf{O}_r$ represent the learned 3D objectness maps account for robust object pose estimation. The decoder then reconstructs the masked object images from the voxels, allowing the voxels to encode the object patterns.
takes $\tilde{\mathbf{F}}_q^l$ as query and $\mathbf{F}_r^l-1$ as key and value, LN denotes layer normalization[3], and FFN is a feed-forward network that includes MLPs. The resulting $\hat{\mathbf{F}}_q^l$ then serves as the input to the next cross-attention moduile. Consequently, the output of the last cross-attention module contains object features depicted from two different viewpoints, thus incorporating 3D information.
Benefiting from such a 3D-aware encoding process, we voxelize the image feature embeddings via a simple reshaping process. Note that, to facilitate the robust pose estimation that will be introduced in Sec. 3.5, we predict an objectness score for each voxel, which reflects the significance of the voxel to the relative object pose estimation. Therefore, the actual reshaping process is conducted as $\mathbb{R}^{C\times H\times W}\to\mathbb{R}^{(C^{\prime}+1)\times D\times H\times W}$,where $C=(C^{\prime}+1)\times D.$ As shown in Fig. 2, we denote the resulting 3D objectness maps and 3D volumes as $\mathbf{O}_q,\mathbf{O}_r\in\mathbb{R}^{1\times D\times H\times W}$and $\mathbf{V}_{q},\mathbf{V}_{r}\in\mathbb{R}^{C^{\prime}\times D\times H\times W}$,respectively.
As our approach does not rely on object segmentation, the learned voxel representations might be affected by the background of the query and reference images. To alleviate this issue, we introduce an object-aware decoding process over $\mathbf{V}_q$ and $\mathbf{V}_r.$ Concretely, $\mathbf{V}_q$ and $\mathbf{V}_r$ are projected to 2D space by aggregating the voxels along the depth direction as $\mathbb{R}^{C^{\prime}\times D\times H\times W}\to\mathbb{R}^{C^{*}\times H\times W}$,where $C^*=C^{\prime}\times D.$ The resulting 2D feature embeddings are then fed into a decoder that contains several self-attention modules[10] from which the object images $\hat{\mathbf{I}}_q$ and $\hat{\mathbf{I}}_r$ without background are produced. The object masks $\hat{\mathbf{M}}_q$ and $\hat{\mathbf{M}}_r$ are additionally predicted to provide auxiliary information that benefits the following robust object pose estimation.
We supervise the training of the autoencoder with an
image-level loss function defined as
(5)
$$\begin{aligned}&L_{ae}=L_{img}+L_{mask},\\&L_{img}=L_{mse}(\hat{\mathbf{I}}_{q},\hat{\mathbf{I}}_{q}^{gt})+L_{mse}(\hat{\mathbf{I}}_{r},\hat{\mathbf{I}}_{r}^{gt}),\\&L_{mask}=L_{bce}(\hat{\mathbf{M}}_{q},\hat{\mathbf{M}}_{q}^{gt})+L_{bce}(\hat{\mathbf{M}}_{r},\hat{\mathbf{M}}_{r}^{gt}),\end{aligned}$$
(6)
(7)
where $L_{mse}$ is the mean squared error loss, $L_bce$ indicates the binary cross entropy loss, $(\hat{\mathbf{I}}_q^{gt},\hat{\mathbf{I}}_r^{gt})$ denote the ground-truth foreground images, and $(\hat{\mathbf{M}}_{q}^{gt},\hat{\mathbf{M}}_{r}^{gt})$ represent the ground-truth object masks.
3.4. Object Pose from Deep Voxel Matching
According to multi-view geometry [4,13,38], the relative object pose can be computed by solving a leastsquares problem expressed in terms of voxel-voxel correspondences. Specifically, as we focus on 3D rotation estimation, the least-squares problem is formulated as
$$E(\Delta\mathbf{R})=\frac{1}{N}\sum_{i=1}^{N}||\Delta\mathbf{R}\mathbf{x}_{r}^{i}-\mathbf{x}_{q}^{i}||_{2},$$
(8)
where $\mathbf{x}_r^i\in\mathbf{X}_r$ and $\mathbf{x}_q^i\in\mathbf{X}_r$ stand for the 3D coordinates of the $i$-th reference and query voxels, respectively. The optimal $\Delta\hat{\mathbf{R}}$ is then determined as
$$\Delta\hat{\mathbf{R}}=\underset{\Delta\mathbf{R}_i\in SO(3)}{\arg\min}-2\sum_{i=1}^N\mathbf{x}_q^i{}^T\Delta\mathbf{R}_i\mathbf{x}_r^i,$$
(9)
As suggested in [4], this problem can be solved by performing a singular value decomposition (SVD) of a covariance matrix as
(10)
$$\begin{aligned}\mathbf{H}&=\sum_{i=1}^{N}\mathbf{x}_{r}^{i}\mathbf{x}_{q}^{i}{}^{T},\\\mathbf{H}&=\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{T},\end{aligned}$$
(11)
|
[
[
0.365234375,
0.091552734375,
0.44287109375,
0.091552734375,
0.44287109375,
0.0985107421875,
0.365234375,
0.0985107421875
],
[
0.775390625,
0.12060546875,
0.81787109375,
0.12060546875,
0.81787109375,
0.160400390625,
0.775390625,
0.160400390625
],
[
0.50341796875,
0.173583984375,
0.51708984375,
0.173583984375,
0.51708984375,
0.1849365234375,
0.50341796875,
0.1849365234375
],
[
0.5048828125,
0.17041015625,
0.515625,
0.17041015625,
0.515625,
0.1749267578125,
0.5048828125,
0.1749267578125
],
[
0.49267578125,
0.1912841796875,
0.515625,
0.1912841796875,
0.515625,
0.1956787109375,
0.49267578125,
0.1956787109375
],
[
0.76708984375,
0.1912841796875,
0.8349609375,
0.1912841796875,
0.8349609375,
0.2568359375,
0.76708984375,
0.2568359375
],
[
0.64794921875,
0.231689453125,
0.70361328125,
0.231689453125,
0.70361328125,
0.24365234375,
0.64794921875,
0.24365234375
],
[
0.366943359375,
0.2412109375,
0.381591796875,
0.2412109375,
0.381591796875,
0.248779296875,
0.366943359375,
0.248779296875
],
[
0.080078125,
0.290283203125,
0.890625,
0.291748046875,
0.890625,
0.30615234375,
0.080078125,
0.304931640625
],
[
0.080078125,
0.304931640625,
0.890625,
0.304931640625,
0.890625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.080078125,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.07843017578125,
0.333984375,
0.193603515625,
0.333984375,
0.193603515625,
0.347900390625,
0.07843017578125,
0.347900390625
],
[
0.08087158203125,
0.373779296875,
0.471435546875,
0.373779296875,
0.471435546875,
0.38818359375,
0.08087158203125,
0.38818359375
],
[
0.5009765625,
0.373779296875,
0.73779296875,
0.373779296875,
0.73779296875,
0.386962890625,
0.5009765625,
0.386962890625
],
[
0.080078125,
0.389404296875,
0.468017578125,
0.389404296875,
0.468017578125,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.5458984375,
0.396484375,
0.69921875,
0.396484375,
0.69921875,
0.4140625,
0.5458984375,
0.4140625
],
[
0.86669921875,
0.396484375,
0.89453125,
0.396484375,
0.89453125,
0.41357421875,
0.86669921875,
0.41357421875
],
[
0.080078125,
0.40478515625,
0.468017578125,
0.40478515625,
0.468017578125,
0.41845703125,
0.080078125,
0.41845703125
],
[
0.5439453125,
0.417236328125,
0.80810546875,
0.416748046875,
0.80810546875,
0.434326171875,
0.5439453125,
0.43505859375
],
[
0.86767578125,
0.419189453125,
0.89208984375,
0.419189453125,
0.89208984375,
0.43310546875,
0.86767578125,
0.43310546875
],
[
0.080078125,
0.436279296875,
0.4697265625,
0.436279296875,
0.4697265625,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.07843017578125,
0.42041015625,
0.468017578125,
0.419189453125,
0.468017578125,
0.434326171875,
0.07843017578125,
0.43505859375
],
[
0.5458984375,
0.439453125,
0.8935546875,
0.439453125,
0.8935546875,
0.45703125,
0.5458984375,
0.45703125
],
[
0.07843017578125,
0.449462890625,
0.468017578125,
0.45068359375,
0.468017578125,
0.46533203125,
0.07843017578125,
0.464111328125
],
[
0.080078125,
0.466552734375,
0.2459716796875,
0.466552734375,
0.2459716796875,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50341796875,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50341796875,
0.477783203125
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.49365234375,
0.50244140625,
0.49365234375
],
[
0.09722900390625,
0.4892578125,
0.468017578125,
0.4912109375,
0.468017578125,
0.505859375,
0.09722900390625,
0.50390625
],
[
0.50244140625,
0.49755859375,
0.89208984375,
0.49755859375,
0.89208984375,
0.51220703125,
0.50244140625,
0.51220703125
],
[
0.080078125,
0.5068359375,
0.46728515625,
0.5068359375,
0.46728515625,
0.52099609375,
0.080078125,
0.52099609375
],
[
0.50244140625,
0.51220703125,
0.705078125,
0.51220703125,
0.705078125,
0.52587890625,
0.50244140625,
0.52587890625
],
[
0.080078125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.080078125,
0.537109375,
0.4697265625,
0.537109375,
0.4697265625,
0.55029296875,
0.080078125,
0.55029296875
],
[
0.50341796875,
0.53515625,
0.83642578125,
0.53515625,
0.83642578125,
0.548828125,
0.50341796875,
0.548828125
],
[
0.08087158203125,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.08087158203125,
0.56494140625
],
[
0.50341796875,
0.55810546875,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.50341796875,
0.5712890625
],
[
0.080078125,
0.56640625,
0.46728515625,
0.56640625,
0.46728515625,
0.580078125,
0.080078125,
0.580078125
],
[
0.50244140625,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.50244140625,
0.5859375
],
[
0.08087158203125,
0.58203125,
0.4697265625,
0.58203125,
0.4697265625,
0.59521484375,
0.08087158203125,
0.59521484375
],
[
0.0767822265625,
0.59228515625,
0.468017578125,
0.59521484375,
0.468017578125,
0.6162109375,
0.0767822265625,
0.61279296875
],
[
0.50244140625,
0.5888671875,
0.890625,
0.5888671875,
0.890625,
0.60302734375,
0.50244140625,
0.60302734375
],
[
0.50244140625,
0.60400390625,
0.89208984375,
0.60400390625,
0.89208984375,
0.6171875,
0.50244140625,
0.6171875
],
[
0.080078125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.61865234375,
0.833984375,
0.61865234375,
0.833984375,
0.63330078125,
0.50244140625,
0.63330078125
],
[
0.07598876953125,
0.62744140625,
0.47216796875,
0.625,
0.47216796875,
0.64453125,
0.07598876953125,
0.64794921875
],
[
0.07843017578125,
0.64306640625,
0.339111328125,
0.64306640625,
0.339111328125,
0.66552734375,
0.07843017578125,
0.66552734375
],
[
0.65673828125,
0.6484375,
0.66845703125,
0.64599609375,
0.67333984375,
0.65478515625,
0.66259765625,
0.65771484375
],
[
0.68310546875,
0.64453125,
0.69384765625,
0.64453125,
0.69384765625,
0.6484375,
0.68310546875,
0.6484375
],
[
0.580078125,
0.65283203125,
0.6435546875,
0.65283203125,
0.6435546875,
0.66748046875,
0.580078125,
0.66748046875
],
[
0.64794921875,
0.6494140625,
0.81787109375,
0.64453125,
0.81884765625,
0.67236328125,
0.64892578125,
0.67822265625
],
[
0.6396484375,
0.65771484375,
0.646484375,
0.65771484375,
0.646484375,
0.66357421875,
0.6396484375,
0.66357421875
],
[
0.869140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.869140625,
0.66748046875
],
[
0.096435546875,
0.6748046875,
0.468017578125,
0.673828125,
0.468017578125,
0.68798828125,
0.096435546875,
0.689453125
],
[
0.68310546875,
0.673828125,
0.69384765625,
0.673828125,
0.69384765625,
0.67822265625,
0.68310546875,
0.67822265625
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50341796875,
0.68798828125,
0.8935546875,
0.68798828125,
0.8935546875,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.5009765625,
0.71826171875,
0.73193359375,
0.71728515625,
0.73193359375,
0.73193359375,
0.5009765625,
0.73291015625
],
[
0.07843017578125,
0.734375,
0.468017578125,
0.73486328125,
0.468017578125,
0.75,
0.07843017578125,
0.74853515625
],
[
0.68310546875,
0.7470703125,
0.7685546875,
0.7548828125,
0.765625,
0.7734375,
0.6806640625,
0.76513671875
],
[
0.72119140625,
0.7421875,
0.73388671875,
0.73779296875,
0.73876953125,
0.74853515625,
0.72607421875,
0.75244140625
],
[
0.07843017578125,
0.75048828125,
0.46728515625,
0.75,
0.46728515625,
0.7646484375,
0.07843017578125,
0.76513671875
],
[
0.56640625,
0.75244140625,
0.68701171875,
0.75244140625,
0.68701171875,
0.767578125,
0.56640625,
0.767578125
],
[
0.7548828125,
0.75244140625,
0.76708984375,
0.75244140625,
0.76708984375,
0.7587890625,
0.7548828125,
0.7587890625
],
[
0.7587890625,
0.74853515625,
0.82861328125,
0.75048828125,
0.82666015625,
0.76953125,
0.75732421875,
0.767578125
],
[
0.869140625,
0.75390625,
0.89208984375,
0.75390625,
0.89208984375,
0.767578125,
0.869140625,
0.767578125
],
[
0.0767822265625,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7841796875,
0.0767822265625,
0.7841796875
],
[
0.61669921875,
0.767578125,
0.69287109375,
0.767578125,
0.69287109375,
0.77783203125,
0.61669921875,
0.77783203125
],
[
0.7197265625,
0.771484375,
0.740234375,
0.771484375,
0.740234375,
0.77978515625,
0.7197265625,
0.77978515625
],
[
0.08087158203125,
0.7841796875,
0.464111328125,
0.7841796875,
0.464111328125,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.50244140625,
0.7861328125,
0.890625,
0.787109375,
0.890625,
0.8017578125,
0.50244140625,
0.80029296875
],
[
0.08087158203125,
0.80029296875,
0.46728515625,
0.80029296875,
0.46728515625,
0.81103515625,
0.08087158203125,
0.81103515625
],
[
0.50244140625,
0.80322265625,
0.8935546875,
0.80322265625,
0.8935546875,
0.81640625,
0.50244140625,
0.81640625
],
[
0.080078125,
0.814453125,
0.4697265625,
0.814453125,
0.4697265625,
0.82763671875,
0.080078125,
0.82763671875
],
[
0.50341796875,
0.8193359375,
0.56787109375,
0.8193359375,
0.56787109375,
0.83056640625,
0.50341796875,
0.83056640625
],
[
0.07843017578125,
0.83056640625,
0.46728515625,
0.830078125,
0.46728515625,
0.8447265625,
0.07843017578125,
0.84521484375
],
[
0.68115234375,
0.8369140625,
0.69775390625,
0.8369140625,
0.69775390625,
0.8466796875,
0.68115234375,
0.8466796875
],
[
0.08087158203125,
0.8466796875,
0.468017578125,
0.8466796875,
0.468017578125,
0.85986328125,
0.08087158203125,
0.85986328125
],
[
0.64208984375,
0.849609375,
0.67333984375,
0.849609375,
0.67333984375,
0.8623046875,
0.64208984375,
0.8623046875
],
[
0.859375,
0.8486328125,
0.89453125,
0.8486328125,
0.89453125,
0.8642578125,
0.859375,
0.8642578125
],
[
0.07843017578125,
0.85986328125,
0.3447265625,
0.861328125,
0.3447265625,
0.87548828125,
0.07843017578125,
0.875
],
[
0.69775390625,
0.85595703125,
0.734375,
0.85595703125,
0.734375,
0.86279296875,
0.69775390625,
0.86279296875
],
[
0.68115234375,
0.8662109375,
0.69921875,
0.8662109375,
0.69921875,
0.875,
0.68115234375,
0.875
],
[
0.6396484375,
0.880859375,
0.73388671875,
0.87890625,
0.73388671875,
0.89306640625,
0.6396484375,
0.89599609375
],
[
0.859375,
0.880859375,
0.89453125,
0.880859375,
0.89453125,
0.896484375,
0.859375,
0.896484375
],
[
0.098876953125,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.098876953125,
0.8994140625
]
] |
[
"2Dto3DEncod",
"",
"一",
"",
"",
"",
"Projecting",
"",
"Figure 2. Network architecture of our autoencoder. The encoder takes two RGB images, query and reference, as input and lifts their 2D",
"feature embeddings to 3D voxels by leveraging cross-view 3D information. O. and O, represent the learned 3D objectness maps account",
"for robust object pose estimation. The decoder then reconstructs the masked object images from the voxels, allowing the voxels to encode",
"the object patterns.",
"takes F! as query and Fl-1 as key and value, LN denotes",
"image-level loss function defined as",
"layer normalization [3],and FFN is a feed-forward network",
"Lae = Limg + Lmask,",
"(5)",
"that includes MLPs. The resulting F. then serves as the in-",
"Limg = Lmse(Iq, Igt) + Lmse(Ir, igt),",
"(6)",
"output of the last cross-attention module contains object",
"put to the next cross-attention module. Consequently, the",
"Lmash = Lbe(Mg, Mgt)+ Lbe(M,, Mgt), (7)",
"features depicted from two different viewpoints, thus incor-",
"porating 3D information.",
"where Lmse is the mean squared error loss, Lbce indi-",
"cates the binary cross entropy loss, (Igt, Igt) denote the",
"Benefiting from such a 3D-aware encoding process, we",
"ground-truth foreground images, and (Mgt, Mgt) represent",
"voxelize the image feature embeddings via a simple reshap-",
"the ground-truth object masks.",
"ing process. Note that, to facilitate the robust pose esti-",
"mation that will be introduced in Sec. 3.5, we predict an",
"3.4. Object Pose from Deep Voxel Matching",
"objectness score for each voxel, which reflects the signif-",
"According to multi-view geometry [4,13,38], the rel-",
"icance of the voxel to the relative object pose estimation.",
"ative object pose can be computed by solving a least-",
"Therefore, the actual reshaping process is conducted as",
"RC×HxW→ IR(c'+1)×Dx×H×W, where C = (C' +1)×D.",
"squares problem expressed in terms of voxel-voxel corre-",
"spondences. Specifically, as we focus on 3D rotation esti-",
"As shown in Fig. 2, we denote the resulting 3D object-",
"mation, the least-squares problem is formulated as",
"ness maps and 3D volumes as Oq, O, E IR1xDxHxW and",
"Vg, V, E RC xDxHxw,respectively.",
"1",
"T",
"E(△R) :",
"IARx, -x:ll,",
"",
"(8)",
"As our approach does not rely on object segmentation,",
"",
"the learned voxel representations might be affected by the",
"where x e Xr and x, e X, stand for the 3D coordinates",
"background of the query and reference images. To alleviate",
"of the i-th reference and query voxels, respectively. The",
"this issue, we introduce an object-aware decoding process",
"optimal △R is then determined as",
"over V.and Vr.Concretely,V.and V, are projected to",
"-2x",
"N",
"2D space by aggregating the voxels along the depth direc-",
"AR= argmin",
"21",
"TARx,",
"(9)",
"tion as RC'×DxHxW→ RC*xHxW, where C* = C ×D.",
"△R,ESO(3)",
"i=1",
"The resulting 2D feature embeddings are then fed into a de",
"As suggested in [4], this problem can be solved by perform-",
"coderthatcontainsseveralself-attentionmodules[1olfrom",
"ing a singular value decomposition (SvD) of a covariance",
"which the object images I and I, without background are",
"matrix as",
"produced. The object masks M, and M, are additionally",
"N",
"predicted to provide auxiliary information that benefits the",
"H=",
"(10)",
"following robust object pose estimation.",
"XrX",
"-1",
"H= UZVT",
"(11)",
"We supervise the training of the autoencoder with an"
] |
{
"arxiv_id": "2403.13683",
"language": "en",
"page_idx": 3
}
|
||
2108.04204-en-0
|
2108.04204-en-0.jpg
|
Meta Gradient Adversarial Attack
Zheng Yuan$^{1,2}$, Jie Zhang$^{1,2}$, Yunpei Jia$^{1,2}$, Chuanqi Tan$^3$, Tao Xue$^3$, Shiguang Shan$^{1,2}$
$^{1}$Institute of Computing Technology, Chinese Academy of Sciences
$^{2}$University of Chinese Academy of Sciences$^{3}$Tencent
$\{$zheng.yuan, yunpei.jia}@vipl.ict.ac.cn;{zhangjie,sgshan}]@ict.ac.cn;
$\{$jamestan,emmaxue$\}$btencent.com
## Abstract
els, such as FGSM [15], C&W [2], PGD [31], BIM [21], DeepFool[32], etc. Based on the access to model parameters, these methods can make the model misclassify the input images by only adding human-imperceptible perturbations, which is named as the white-box attack. However, in reality, a more practical scenario is that the attacker cannot obtain any information of the target model, that is, the black-box attack. Therefore, some methods turn to utilize the transferability of adversarial examples to conduct black-box attacks, such as MIM[8], DIM[50], TIM[9], NISI[27],etc. Although most of these methods have achieved promising results under the scenario of black-box attacks, the transferability of adversarial examples is still limited due to the discrepancy between the white-box models and unseen black-box models.
In recent years, research on adversarial attacks has become a hot spot. Although current literature on the transferbased adversarial attack has achieved promising results for improving the transferability to unseen black-box models, it still leaves a long way to go. Inspired by the idea of meta-
$\sum_{\begin{array}{cc}&\text{sunt itaves u ong wuy io go. msprtea oy inte ueu oj meur}\\&\text{learning, this paper proposes a novel architecture called}\\&\text{Meta Gradient Adversarial Attack (MGAA),which is plug-}\\&\text{and-play and can be integrated with any existing gradient-}\end{array}}$ $\begin{array}{lllllllll}&and-play\:and\:can\:be\:integrated\:wit\end{array}$
els from a model zoo to compose different tasks and iteratively simulate a white-box attack and a black-box attack in each task. By narrowing the gap between the gradient directions in white-box and black-box attacks, the transferability of adversarial examples on the black-box setting can be improved. Extensive experiments on the CIFAR10 and ImageNet datasets show that our architecture outperforms the state-of-the-art methods for both black-box and whitebox attack settings.
Inspired by the philosophy of meta-learning, we propose a novel architecture named Meta Gradient Adversarial Attack (MGAA), which is plug-and-play and can be incorporated with any gradient-based adversarial attack method. The main idea of MGAA is to generate the adversarial examples by iteratively simulating white-box and black-box attacks to improve the transferability. Specifically, as shown in Fig. l, in each iteration, multiple models are randomly sampled from the model zoo to compose a task, which can be divided into the meta-train and the meta-test step. The meta-train step first uses an ensemble of multiple models to simulate white-box attacks to obtain temporary adversarial examples, which are then used as a basis by the metatest step to obtain the perturbation by simulating a blackbox attack scenario. Finally, the perturbation obtained during the meta-test step is added to the adversarial examples generated in the previous iteration. In Section 3.3, more theoretical analyses demonstrate that our proposed architecture can gradually narrow the gap of gradient directions between white-box attack and black-box attack settings, thus improving the transferability of generated adversarial examples. Different from vanilla meta-learning methods which enhance the generalization through model training, our proposed MGAA directly utilizes the gradient informa-
### $1. \textbf{Introduction}$
With the rapid development of neural networks in recent years, the reliability of neural networks has gradually attracted more and more attention. The neural networks are exceedingly sensitive to adversarial examples, $i.e.$, the imperceptible perturbation on the input can easily fool the model, leading to unexpected mistakes. For example, when employing face recognition technology for payment, a slight perturbation on the face image may trick the face recognition model into recognizing as someone else. Since attack and defense are two complementary aspects, the researches on adversarial attacks can ultimately improve the robustness of the model, thereby making the model more reliable.
In recent years, many methods have been proposed to improve the success rate of attacks against white-box mod-
|
[
[
0.31201171875,
0.1337890625,
0.65869140625,
0.1337890625,
0.65869140625,
0.1483154296875,
0.31201171875,
0.1483154296875
],
[
0.131591796875,
0.179931640625,
0.8388671875,
0.177978515625,
0.8388671875,
0.1988525390625,
0.131591796875,
0.2008056640625
],
[
0.22216796875,
0.2008056640625,
0.7490234375,
0.2008056640625,
0.7490234375,
0.21533203125,
0.22216796875,
0.21533203125
],
[
0.2607421875,
0.2171630859375,
0.7138671875,
0.2171630859375,
0.7138671875,
0.231689453125,
0.2607421875,
0.231689453125
],
[
0.1895751953125,
0.2374267578125,
0.78369140625,
0.2374267578125,
0.78369140625,
0.251953125,
0.1895751953125,
0.251953125
],
[
0.0196075439453125,
0.260009765625,
0.063720703125,
0.260009765625,
0.066162109375,
0.375,
0.0228729248046875,
0.375
],
[
0.35302734375,
0.255126953125,
0.62158203125,
0.25244140625,
0.62158203125,
0.267578125,
0.35302734375,
0.26953125
],
[
0.236083984375,
0.302978515625,
0.314453125,
0.302978515625,
0.314453125,
0.32080078125,
0.236083984375,
0.32080078125
],
[
0.50244140625,
0.304931640625,
0.890625,
0.30615234375,
0.890625,
0.32080078125,
0.50244140625,
0.3193359375
],
[
0.50244140625,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50244140625,
0.335205078125
],
[
0.098876953125,
0.3359375,
0.4697265625,
0.3359375,
0.4697265625,
0.349853515625,
0.098876953125,
0.349853515625
],
[
0.50341796875,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.08087158203125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.08087158203125,
0.365478515625
],
[
0.50244140625,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.0269622802734375,
0.363037109375,
0.053924560546875,
0.363037109375,
0.053924560546875,
0.391357421875,
0.0269622802734375,
0.391357421875
],
[
0.080078125,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.5,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.5,
0.380126953125
],
[
0.08087158203125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50244140625,
0.3837890625,
0.890625,
0.3837890625,
0.890625,
0.39453125,
0.50244140625,
0.39453125
],
[
0.08087158203125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.08087158203125,
0.410888671875
],
[
0.50244140625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.41357421875,
0.471435546875,
0.41357421875,
0.471435546875,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.080078125,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.080078125,
0.44140625
],
[
0.50341796875,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.439453125,
0.50341796875,
0.439453125
],
[
0.08087158203125,
0.442626953125,
0.468017578125,
0.442626953125,
0.468017578125,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.50244140625,
0.442626953125,
0.890625,
0.442626953125,
0.890625,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.0310516357421875,
0.448974609375,
0.056365966796875,
0.448974609375,
0.056365966796875,
0.48095703125,
0.0310516357421875,
0.48095703125
],
[
0.080078125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.50341796875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.07843017578125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.486083984375,
0.07843017578125,
0.486083984375
],
[
0.50244140625,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.080078125,
0.501953125
],
[
0.50244140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50244140625,
0.4697265625,
0.50244140625,
0.4697265625,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.02532958984375,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.5888671875,
0.02532958984375,
0.5888671875
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.080078125,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50244140625,
0.51806640625,
0.67333984375,
0.51806640625,
0.67333984375,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.08087158203125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.54541015625,
0.08087158203125,
0.54541015625
],
[
0.521484375,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.521484375,
0.55322265625
],
[
0.080078125,
0.54736328125,
0.46728515625,
0.54736328125,
0.46728515625,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.5009765625,
0.568359375,
0.890625,
0.5693359375,
0.890625,
0.583984375,
0.5009765625,
0.58349609375
],
[
0.0196075439453125,
0.58349609375,
0.058013916015625,
0.58349609375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.08087158203125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.08087158203125,
0.59228515625
],
[
0.03350830078125,
0.58544921875,
0.049835205078125,
0.58544921875,
0.049835205078125,
0.59228515625,
0.03350830078125,
0.59228515625
],
[
0.50341796875,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5986328125,
0.890625,
0.599609375,
0.890625,
0.6142578125,
0.50244140625,
0.61279296875
],
[
0.080078125,
0.609375,
0.20751953125,
0.609375,
0.20751953125,
0.62255859375,
0.080078125,
0.62255859375
],
[
0.50341796875,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50341796875,
0.62890625
],
[
0.50244140625,
0.62890625,
0.89208984375,
0.6298828125,
0.89208984375,
0.64453125,
0.50244140625,
0.64306640625
],
[
0.08087158203125,
0.6494140625,
0.20751953125,
0.6494140625,
0.20751953125,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.50341796875,
0.646484375,
0.890625,
0.646484375,
0.890625,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.098876953125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.098876953125,
0.68798828125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.7041015625,
0.4697265625,
0.705078125,
0.4697265625,
0.7197265625,
0.07843017578125,
0.71826171875
],
[
0.5009765625,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50244140625,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50341796875,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.5009765625,
0.77978515625
],
[
0.07843017578125,
0.78076171875,
0.4697265625,
0.77978515625,
0.4697265625,
0.79443359375,
0.07843017578125,
0.79541015625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85791015625,
0.2197265625,
0.85791015625,
0.2197265625,
0.86865234375,
0.080078125,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.09722900390625,
0.88623046875
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Meta Gradient Adversarial Attack",
"Zheng Yuan1,2, Jie Zhangl,2, Yunpei Jial,2, Chuanqi Tan', Tao Xue3, Shiguang Shanl,2",
"1Institute of Computing Technology, Chinese Academy of Sciences",
"2Universityof ChineseAcademyof Sciences3Tencent",
"[zheng.yuan, yunpei.jia]@vipl.ict.ac.cn;[zhangjie,sgshan]@ict.ac.cn;",
"12 3n (",
"(jamestan,emmaxue)@tencent.com",
"Abstract",
"els, such as FGSM [15], C&W [2], PGD [31], BIM [2]],",
"DeepFool [32], etc. Based on the access to model param-",
"In recent years, research on adversarial attacks has be-",
"eters,these methods canmakethe model misclassifythe",
"come a hot spot. Although current literature on the transfer-",
"input images by only adding human-imperceptible pertur-",
"0",
"based adversarial attack has achieved promising results for",
"bations, which is named as the white-box attack. How-",
"improvingthetransferabilitytounseenblack-boxmodels,it",
"ever,in reality,a more practical scenario is that the attacker",
"still leaves a long way to go. Inspired by the idea of meta-",
"cannot obtain any information of the target model, that is,",
"LAD",
"learning, this paper proposes a novel architecture called",
"the black-box attack. Therefore, some methods turn to uti-",
"Meta Gradient Adversarial Attack (MGAA), which is plug-",
"lizethetransferability of adversarial examples to conduct",
"and-play and can be integrated with any existing gradient-",
"black-box attacks, such as MIM[8], DIM[50], TIM [9], NI-",
"SO",
"based attack method for improving the cross-model trans-",
"SI[27l,etc.Althoughmost ofthesemethods haveachieved",
"ferability Specifically, we randomly sample multiple mod-",
"promising results under the scenario of black-box attacks,",
"els from a model zoo to compose different tasks and iter-",
"the transferability of adversarial examples is still limited",
"atively simulate a white-box attack and a black-box attack",
"",
"due to the discrepancy between the white-box models and",
"in each task. By narrowing the gap between the gradient",
"unseen black-box models.",
"directions inwhite-box and black-box attacks,the transfer-",
"Inspired by the philosophy of meta-learning, we propose",
"ability of adversarial examples on the black-box setting can",
"a novel architecture named Meta Gradient Adversarial At-",
"be improved. Extensive experiments on the CIFAR1O and",
"tack (MGAA), which is plug-and-play and can be incor-",
"8017A0",
"ImageNet datasets show that our architecture outperforms",
"",
"porated with any gradient-based adversarial attack method.",
"the state-of-the-art methods for both black-box and white-",
"The main idea of MGAA is to generate the adversarial ex-",
"box attack settings.",
"amples by iteratively simulating white-box and black-box",
"attacks to improve the transferability. Specifically, as shown",
"1. Introduction",
"in Fig. l, in each iteration, multiple models are randomly",
"sampled from the model zoo to compose a task, which can",
"With the rapid development of neural networks in re-",
"be divided into the meta-train and the meta-test step. The",
"cent years, the reliability of neural networks has gradu-",
"meta-train step first uses an ensemble of multiple models",
"ally attracted more and more attention. The neural net-",
"to simulate white-box attacks to obtain temporary adversar-",
"works are exceedingly sensitive to adversarial examples,",
"ial examples, which are then used as a basis by the meta-",
"i.e.the imperceptible perturbation on the input can eas-",
"test stepto obtain the perturbation by simulating a black-",
"ily fool the model, leading to unexpected mistakes. For",
"box attack scenario. Finally, the perturbation obtained dur-",
"example, when employing face recognition technology for",
"ing the meta-test step is added to the adversarial examples",
"payment, a slight perturbation on the face image may trick",
"generated in the previous iteration.In Section 3.3, more",
"the face recognition model into recognizing as someone",
"theoretical analyses demonstrate that our proposed archi-",
"else. Since attack and defense are two complementary as-",
"tecture can gradually narrow the gap of gradient directions",
"pects, the researches on adversarial attacks can ultimately",
"between white-box attack and black-box attack settings,",
"improve the robustness of the model, thereby making the",
"thus improving the transferability of generated adversarial",
"model more reliable.",
"examples. Different from vanilla meta-learning methods,",
"In recent years, many methods have been proposed to",
"which enhance the generalization through model training,",
"improve the success rate of attacks against white-box mod-",
"our proposed MGAA directly utilizes the gradient informa-"
] |
{
"arxiv_id": "2108.04204",
"language": "en",
"page_idx": 0
}
|
||
2108.04204-en-1
|
2108.04204-en-1.jpg
|
<FigureHere>
Figure l: Overview of our Meta Gradient Adversarial Attack (MGAA). MGAA consists of multiple $(T$ in the figure) itera-
tions. In each iteration, $n+1$ models are randomly sampled from the model zoo to compose a meta-task. Each task is divided into two steps: the meta-train step and the meta-test step. In the meta-train step, an ensemble of the $n$ sampled models is utilized to conduct the gradient-based attack to generate perturbations, which can be repeated by $K$ times. The meta-test step uses the adversarial example $x_{i,K}$ obtained from the meta-train step as the basis, and utilizes the last sampled model to generate perturbations for adversarial attacks. Finally, the perturbation generated in the meta-test step $x_{i,mt}-x_{i,K}$ is added to $x_i$,the final adversarial example after the $i$-th task.
### $2. \textbf{Related Work}$
In this section, we will give a brief introduction to the related works, $i.e.$, the adversarial attack, the adversarial defense, and the meta-learning.
### $2. 1. \textbf{Adversarial Attack}$
tion to improve the transferability of the adversarial exam-
ples without the need of training an extra model.
Extensive experiments on the CIFAR10[20] and ImageNet[38] dataset demonstrate that our proposed architecture significantly improves the success rates of both whitebox and black-box attacks. Especially, by integrating TIDIM[9] method into our proposed architecture, the average attack success rate under the targeted attack setting against 10 white-box models increases by 27.67%, and the attack success rate against 6 black-box models increases by $28.52\%$ on ImageNet, which clearly shows the superiority of our method
The main contributions of this paper are as follows:
1. We propose the Meta Gradient Adversarial Attack architecture inspired by the philosophy of meta-learning, which iteratively simulates the white-box and the black-box attack to narrow the gap of the gradient directions when generating adversarial examples, thereby further improving the transferability of adversarial examples.
2. The proposed architecture can be combined with any existing gradient-based attack method in a plug-and-play mode.
3. Extensive experiments show that the proposed architecture can significantly improve the attack success rates under both white-box and black-box settings.
The task of adversarial attack is generally classified into four categories according to the amount of target model information we can access: white-box attack, scorebased black-box attack, decision-based black-box attack, and transfer-based black-box attack.
White-box Attack. The white-box attack can obtain all the information of the target model, including model parameters, model structure, gradient, $etc.$ FGSM [15] utilizes gradient information to update the adversarial example in one step along the direction of maximum classification loss. This method is extended by BIM [21] to propose an iterative method to generate adversarial examples through multi-step updates. PGD [31] is similar to BIM, except that it randomly selects an initial point in the neighborhood of the benign example as the starting point of the iterative attack. Some methods consider the task from the perspective of optimization. In DeepFool [32], an optimized method is employed to generate the smallest perturbation while meet-
|
[
[
0.3095703125,
0.0985107421875,
0.359375,
0.0985107421875,
0.359375,
0.10919189453125,
0.3095703125,
0.10919189453125
],
[
0.419189453125,
0.0985107421875,
0.468017578125,
0.0985107421875,
0.468017578125,
0.10919189453125,
0.419189453125,
0.10919189453125
],
[
0.57666015625,
0.0985107421875,
0.625,
0.0985107421875,
0.625,
0.10919189453125,
0.57666015625,
0.10919189453125
],
[
0.740234375,
0.09783935546875,
0.7890625,
0.09783935546875,
0.7890625,
0.10919189453125,
0.740234375,
0.10919189453125
],
[
0.16259765625,
0.10418701171875,
0.1690673828125,
0.0985107421875,
0.1748046875,
0.1029052734375,
0.1683349609375,
0.10797119140625
],
[
0.146240234375,
0.11553955078125,
0.1519775390625,
0.11553955078125,
0.1519775390625,
0.121826171875,
0.146240234375,
0.121826171875
],
[
0.19287109375,
0.11297607421875,
0.1993408203125,
0.11297607421875,
0.1993408203125,
0.117431640625,
0.19287109375,
0.117431640625
],
[
0.233642578125,
0.11614990234375,
0.2607421875,
0.11614990234375,
0.2607421875,
0.125,
0.233642578125,
0.125
],
[
0.6396484375,
0.125,
0.67626953125,
0.126953125,
0.6748046875,
0.136962890625,
0.63818359375,
0.1351318359375
],
[
0.7958984375,
0.125,
0.84228515625,
0.126953125,
0.84228515625,
0.136962890625,
0.7958984375,
0.1351318359375
],
[
0.48779296875,
0.126953125,
0.50634765625,
0.126953125,
0.50634765625,
0.1337890625,
0.48779296875,
0.1337890625
],
[
0.374267578125,
0.126953125,
0.39453125,
0.126953125,
0.39453125,
0.1337890625,
0.374267578125,
0.1337890625
],
[
0.53759765625,
0.126953125,
0.5537109375,
0.126953125,
0.5537109375,
0.1351318359375,
0.53759765625,
0.1351318359375
],
[
0.69921875,
0.126953125,
0.71826171875,
0.126953125,
0.71826171875,
0.1351318359375,
0.69921875,
0.1351318359375
],
[
0.1380615234375,
0.136962890625,
0.1446533203125,
0.141357421875,
0.1405029296875,
0.146484375,
0.13232421875,
0.141357421875
],
[
0.21728515625,
0.1427001953125,
0.275390625,
0.1427001953125,
0.275390625,
0.154052734375,
0.21728515625,
0.154052734375
],
[
0.310546875,
0.149658203125,
0.324462890625,
0.15087890625,
0.324462890625,
0.1546630859375,
0.310546875,
0.154052734375
],
[
0.328369140625,
0.149658203125,
0.3544921875,
0.149658203125,
0.3544921875,
0.1572265625,
0.328369140625,
0.1572265625
],
[
0.449462890625,
0.1514892578125,
0.465576171875,
0.1514892578125,
0.465576171875,
0.1558837890625,
0.449462890625,
0.1558837890625
],
[
0.61376953125,
0.15087890625,
0.625,
0.15087890625,
0.625,
0.154052734375,
0.61376953125,
0.154052734375
],
[
0.19287109375,
0.1572265625,
0.200927734375,
0.1572265625,
0.200927734375,
0.1610107421875,
0.19287109375,
0.1610107421875
],
[
0.439453125,
0.1622314453125,
0.449462890625,
0.1622314453125,
0.449462890625,
0.1673583984375,
0.439453125,
0.1673583984375
],
[
0.1422119140625,
0.1546630859375,
0.1519775390625,
0.1622314453125,
0.143798828125,
0.1673583984375,
0.1356201171875,
0.1610107421875
],
[
0.1683349609375,
0.1654052734375,
0.1748046875,
0.1654052734375,
0.1748046875,
0.1685791015625,
0.1683349609375,
0.1685791015625
],
[
0.136474609375,
0.1749267578125,
0.2034912109375,
0.1749267578125,
0.2034912109375,
0.186279296875,
0.136474609375,
0.186279296875
],
[
0.34228515625,
0.205810546875,
0.404296875,
0.205810546875,
0.404296875,
0.21337890625,
0.34228515625,
0.21337890625
],
[
0.69921875,
0.2052001953125,
0.75341796875,
0.2052001953125,
0.75341796875,
0.212158203125,
0.69921875,
0.212158203125
],
[
0.1192626953125,
0.21533203125,
0.1953125,
0.2171630859375,
0.1953125,
0.231689453125,
0.1192626953125,
0.2298583984375
],
[
0.306396484375,
0.2197265625,
0.325927734375,
0.2197265625,
0.325927734375,
0.2266845703125,
0.306396484375,
0.2266845703125
],
[
0.5986328125,
0.216552734375,
0.63232421875,
0.2203369140625,
0.6298828125,
0.2310791015625,
0.59619140625,
0.2266845703125
],
[
0.8037109375,
0.2374267578125,
0.83251953125,
0.2374267578125,
0.83251953125,
0.24560546875,
0.8037109375,
0.24560546875
],
[
0.1356201171875,
0.2822265625,
0.1690673828125,
0.2822265625,
0.1690673828125,
0.29345703125,
0.1356201171875,
0.29345703125
],
[
0.50341796875,
0.27978515625,
0.53125,
0.28271484375,
0.52783203125,
0.29345703125,
0.5009765625,
0.290283203125
],
[
0.339111328125,
0.2890625,
0.361083984375,
0.291748046875,
0.359375,
0.298583984375,
0.338134765625,
0.296630859375
],
[
0.488525390625,
0.2890625,
0.54150390625,
0.2890625,
0.54150390625,
0.2998046875,
0.488525390625,
0.2998046875
],
[
0.31201171875,
0.29345703125,
0.338134765625,
0.29345703125,
0.338134765625,
0.296630859375,
0.31201171875,
0.296630859375
],
[
0.1192626953125,
0.290283203125,
0.178955078125,
0.2890625,
0.178955078125,
0.2998046875,
0.1192626953125,
0.300537109375
],
[
0.188720703125,
0.294189453125,
0.2115478515625,
0.294189453125,
0.2115478515625,
0.29736328125,
0.188720703125,
0.29736328125
],
[
0.373291015625,
0.29345703125,
0.396240234375,
0.29345703125,
0.396240234375,
0.29736328125,
0.373291015625,
0.29736328125
],
[
0.498291015625,
0.304931640625,
0.52685546875,
0.304931640625,
0.52685546875,
0.30810546875,
0.498291015625,
0.30810546875
],
[
0.13232421875,
0.311279296875,
0.1641845703125,
0.311279296875,
0.1641845703125,
0.318115234375,
0.13232421875,
0.318115234375
],
[
0.3095703125,
0.313232421875,
0.358642578125,
0.313232421875,
0.358642578125,
0.31640625,
0.3095703125,
0.31640625
],
[
0.48779296875,
0.311279296875,
0.5400390625,
0.31005859375,
0.5400390625,
0.317626953125,
0.48779296875,
0.318115234375
],
[
0.476318359375,
0.330810546875,
0.50634765625,
0.333251953125,
0.5048828125,
0.34521484375,
0.4755859375,
0.343505859375
],
[
0.39453125,
0.34033203125,
0.423095703125,
0.343505859375,
0.420654296875,
0.354248046875,
0.393798828125,
0.3515625
],
[
0.5537109375,
0.343505859375,
0.5888671875,
0.34521484375,
0.58837890625,
0.35986328125,
0.55224609375,
0.35791015625
],
[
0.60302734375,
0.35791015625,
0.646484375,
0.3623046875,
0.6435546875,
0.375732421875,
0.6005859375,
0.371826171875
],
[
0.314453125,
0.365478515625,
0.326904296875,
0.365478515625,
0.326904296875,
0.371826171875,
0.314453125,
0.371826171875
],
[
0.358642578125,
0.37744140625,
0.564453125,
0.37744140625,
0.564453125,
0.392578125,
0.358642578125,
0.392578125
],
[
0.080078125,
0.406005859375,
0.890625,
0.406005859375,
0.890625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.080078125,
0.421630859375,
0.89208984375,
0.421630859375,
0.89208984375,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.080078125,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.080078125,
0.45068359375
],
[
0.080078125,
0.451904296875,
0.89208984375,
0.451904296875,
0.89208984375,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.080078125,
0.46728515625,
0.89208984375,
0.46728515625,
0.89208984375,
0.48095703125,
0.080078125,
0.48095703125
],
[
0.07843017578125,
0.48095703125,
0.8935546875,
0.48095703125,
0.8935546875,
0.498779296875,
0.07843017578125,
0.498779296875
],
[
0.07843017578125,
0.49755859375,
0.42724609375,
0.49755859375,
0.42724609375,
0.51123046875,
0.07843017578125,
0.51123046875
],
[
0.080078125,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50341796875,
0.5166015625,
0.64208984375,
0.5166015625,
0.64208984375,
0.52978515625,
0.50341796875,
0.52978515625
],
[
0.080078125,
0.53271484375,
0.39794921875,
0.53271484375,
0.39794921875,
0.546875,
0.080078125,
0.546875
],
[
0.521484375,
0.54345703125,
0.89208984375,
0.54345703125,
0.89208984375,
0.556640625,
0.521484375,
0.556640625
],
[
0.098876953125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.56298828125,
0.098876953125,
0.56298828125
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50341796875,
0.56005859375,
0.8896484375,
0.56005859375,
0.8896484375,
0.57080078125,
0.50341796875,
0.57080078125
],
[
0.5009765625,
0.5712890625,
0.69287109375,
0.57373046875,
0.69287109375,
0.58837890625,
0.5009765625,
0.5859375
],
[
0.08087158203125,
0.58056640625,
0.46728515625,
0.58056640625,
0.46728515625,
0.59130859375,
0.08087158203125,
0.59130859375
],
[
0.080078125,
0.5947265625,
0.46728515625,
0.5947265625,
0.46728515625,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50244140625,
0.6015625,
0.68115234375,
0.6015625,
0.68115234375,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.07843017578125,
0.62353515625,
0.4697265625,
0.625,
0.4697265625,
0.6396484375,
0.07843017578125,
0.63818359375
],
[
0.52294921875,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.52294921875,
0.642578125
],
[
0.08087158203125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.6513671875,
0.08087158203125,
0.6513671875
],
[
0.50341796875,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.07843017578125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.07843017578125,
0.66845703125
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.67041015625,
0.50341796875,
0.67041015625
],
[
0.07843017578125,
0.66845703125,
0.468017578125,
0.66943359375,
0.468017578125,
0.6845703125,
0.07843017578125,
0.68310546875
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.08087158203125,
0.68603515625,
0.17724609375,
0.68603515625,
0.17724609375,
0.69775390625,
0.08087158203125,
0.69775390625
],
[
0.50244140625,
0.68994140625,
0.74169921875,
0.68994140625,
0.74169921875,
0.70068359375,
0.50244140625,
0.70068359375
],
[
0.098876953125,
0.7021484375,
0.438720703125,
0.7021484375,
0.438720703125,
0.71533203125,
0.098876953125,
0.71533203125
],
[
0.521484375,
0.705078125,
0.8935546875,
0.705078125,
0.8935546875,
0.71826171875,
0.521484375,
0.71826171875
],
[
0.1004638671875,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.1004638671875,
0.73193359375
],
[
0.50244140625,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.07843017578125,
0.7646484375,
0.4697265625,
0.7646484375,
0.4697265625,
0.77783203125,
0.07843017578125,
0.77783203125
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.07843017578125,
0.77783203125,
0.4697265625,
0.77880859375,
0.4697265625,
0.79345703125,
0.07843017578125,
0.79248046875
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.79345703125,
0.359375,
0.79345703125,
0.359375,
0.806640625,
0.080078125,
0.806640625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.09722900390625,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.09722900390625,
0.82373046875
],
[
0.5009765625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.080078125,
0.82470703125,
0.468017578125,
0.82470703125,
0.468017578125,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.1217041015625,
0.84130859375,
0.1217041015625,
0.85302734375,
0.080078125,
0.85302734375
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.09722900390625,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.09722900390625,
0.869140625
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.36279296875,
0.88720703125,
0.36279296875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Task#1",
"Task#2",
"Task#i",
"Task#T",
"福",
"福",
"",
"Task",
"T1",
"CT+1",
"a",
"ro",
"c",
"CT",
"",
"Sampling",
"Me",
"aTrain",
"a",
"1",
"",
"",
"福",
"",
"Model Zoo",
"MetaTrair",
"MetaTest",
"Cio=Ci",
"03",
"Ti.K",
"i.mt",
"..",
"·",
"·",
"...",
"",
"...",
"oth",
"Tathn",
"",
"ainMo",
"me",
"Train Model",
"Ti.2",
"Ti,1",
"Ci,K",
"ai.mt",
"",
"ri+1=ai+(aimt-aik)",
"Figure l: Overview of our Meta Gradient Adversarial Attack (MGAA). MGAA consists of multiple (T in the figure) itera-",
"tions. In each iteration, n +1 models are randomly sampled from the model zoo to compose a meta-task. Each task is divided",
"into two steps: the meta-train step and the meta-test step. In the meta-train step, an ensemble of the n sampled models is",
"utilized to conduct the gradient-based attack to generate perturbations, which can be repeated by K times. The meta-test",
"step uses the adversarial example c;,K obtained from the meta-train step as the basis, and utilizes the last sampled model to",
"generate perturbations for adversarial attacks. Finally, the perturbation generated in the meta-test step ai,mt - ri,k is added",
"to ci, the final adversarial example after the i-th task",
"tion to improve the transferability of the adversarial exam-",
"2. Related Work",
"ples without the need of training an extra model.",
"In this section, we will give a brief introduction to the",
"Extensive experiments on the CIFAR10 [20] and Ima-",
"geNet [38] dataset demonstrate that our proposed architec-",
"related works,i.e.,the adversarial attack, the adversarial de-",
"fense, and the meta-learning.",
"ture significantlyimproves the success rates of both white-",
"box and black-box attacks. Especially, by integrating TI-",
"2.1. Adversarial Attack",
"DIM [9] method into our proposed architecture, the av-",
"erage attack success rate under the targeted attack setting",
"The task of adversarial attack is generally classified",
"against 10 white-box models increasesby 27.67%,and the",
"into fourcategories according totheamount of target",
" attack success rate against 6 black-box models increases by",
"modelinformationwecanaccess:white-boxattack,score-",
"28.52% on ImageNet, which clearly shows the superiority",
"based black-box attack, decision-based black-box attack,",
"of our method.",
"andtransfer-based black-boxattack.",
"The main contributions of this paper are as follows:",
"White-box Attack. The white-box attack can obtain all",
"1. We propose the Meta Gradient Adversarial Attack",
"the information of the target model, including model pa-",
"architecture inspired by the philosophy of meta-learning,",
"rameters, model structure, gradient, etc. FGSM [15] uti-",
"which iteratively simulates the white-box and the black-box",
"lizesgradientinformationtoupdatetheadversarialexample",
"attack to narrow the gap of the gradient directions when",
"in one step along the direction of maximum classification",
"generating adversarial examples, thereby further improving",
"loss. This method is extended by BIM [21] to propose an",
"the transferability of adversarial examples.",
"iterative method to generate adversarial examples through",
"2. The proposed architecture can be combined with any",
"multi-step updates. PGD [31]J is similar to BIM, except that",
"existing gradient-based attack method in a plug-and-play",
"it randomly selects an initial point in the neighborhood of",
"mode.",
"the benign example as the starting point of the iterative at-",
"3. Extensive experiments show that the proposed archi-",
"tack.Some methods considerthetaskfromtheperspective",
"tecture can significantly improve the attack success rates un-",
"of optimization. In DeepFool [32], an optimized method is",
"der both white-box and black-box settings.",
"employed to generate the smallest perturbation while meet-"
] |
{
"arxiv_id": "2108.04204",
"language": "en",
"page_idx": 1
}
|
||
2108.04204-en-2
|
2108.04204-en-2.jpg
|
ing the target of a successful attack. C&W[2] transforms the task into a constrained optimization problem and compares the effects of multiple objective functions.
Score-based Black-box Attack. This category of attack methods assumes they can obtain the classification probabilities of a given input image from the target model. ZOO[3] proposes a zeroth-order optimizationbased method, which directly estimates the gradients of the target model for generating adversarial examples. However, it suffers from a low attack success rate and poor query efficiency since it is non-trivial to estimate the gradient with limited information. To address these problems, P-RGR[5] utilizes the transfer-based prior and query information to improve query efficiency. $\hat{\mathcal{N}}$ ATTACK[25] further uses a probability density distribution over a small region centered around the input to generate adversarial examples, which defeats both vanilla DNNs and those generated by various defense techniques developed.
Decision-based Black-box Attack. Under the decisionbased black-box attack setting, only the predicted class of a given input image from the target model is available, which seems more difficult than the score-based black-box attack. Boundary Attack[1] is firstly proposed to this problem, which gradually minimizes the adversarial perturbation by the approach of random walking while maintaining the aggressiveness of the adversarial example. Cheng et al.[4] transforms the problem into a continuous real-valued optimization problem that can be solved by any zeroth-order optimization algorithm. Different from previous methods, Dong et al.[10] proposes an evolutionary attack algorithm, which reduces the dimension of the search space and improves the query efficiency.
Transfer-based Black-box Attack. The transfer-based black-box attack cannot obtain any information of the target model, which is the most challenging setting. Existing methods [8, 50, 9, 54, 27] mainly endeavor to improve the attack success rates by the transferability of the adversarial examples generated from the substitute model. We will review these methods in Sec. 3.1 in details.
Although these methods have been proposed to gradually improve the transferability of adversarial examples, the results in black-box attacks still leave a lot of room for improvement. Our work solves this problem from the perspective of meta-learning, which reduces the gap between whitebox and black-box attacks, thus achieving better results.
### $2. 2. \textbf{Adversarial Defense}$
The task of adversarial defense is to improve the robustness of the model so that the model can correctly classify the perturbed adversarial examples. Defense methods can be divided into five categories: adversarial training, input transformation, randomization, model ensemble, and certified defenses. Adversarial training [45, 31, 41] conducts
the model training based on the generated adversarial examples. Input transformation utilizes JPEG compression[12], denoising [26], and extra GAN model [39] before feeding the images to the model. Randomization refers to adding random noise into the input examples [49] or models [7] to make the model more robust to adversarial examples. The model ensemble means the ensemble of multiple models in the output layer. Compared with a single model, it can reduce the impact of distributions in the adversarial examples [22,29]. In addition, some work [37,46] proves that under a specific target model, a certified defense model can ensure robustness to adversarial examples.
# 2.3. Meta-learning
Meta-learning is a concept of learning how to learn. It mainly focuses on how to use a large amount of data to learn a learning mode. When encountering a new task later, only a small amount of additional data is needed to quickly complete the new task by fine-tuning the model[13,34]. As a newly proposed method in recent years, meta-learning has been widely used in various tasks. Meta-Attack[1] uses the meta-learning framework to train the gradient estimator. In the case of the score-based black-box attack, only a small number of queries are needed to fine-tune the gradient estimator to complete the adversarial attack.
Different from Meta-Attack, our proposed MGAA is actually not a meta-learning framework but only inspired by its philosophy. Specifically, we do not need to explicitly train an extra model but directly generate the adversarial examples in the process of iteration.
# 3. Methodology
Since our proposed architecture is plug-and-play and can be integrated with any existing gradient-based attack method, we first introduce some typical gradient-based attack methods in Sec. 3.1. And then, we introduce our proposed Meta Gradient Adversarial Attack architecture in detail in Sec. 3.2. Moreover, we give a detailed analysis of the effectiveness of our proposed architecture theoretically in Sec. 3.3. Finally, a discussion on the differences between our proposed MGAA and some related works is given in Sec. 3. 4.
Let $x$ denote the benign example and $y$ denote the true label corresponding to the benign example. $f(x):\mathcal{X}\to\mathcal{Y}$ denotes the target classifier, $J$ denotes the loss function used for training the classifier, which usually means cross entropy loss. $x^{adv}$ denotes the adversarial example that needs to be optimized. The optimization objective can be described as:
(1)
$$\arg\max_{x^{adv}}J(f(x^{adv}),y),\quad\mathrm{s.t.}\|x^{adv}-x\|_{\infty}\leq\epsilon,$$
where $\epsilon$ is the maximum perturbation per pixel allowed to
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.12249755859375,
0.398681640625,
0.121826171875,
0.398681640625,
0.1363525390625,
0.07843017578125,
0.136962890625
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.09722900390625,
0.1383056640625,
0.47216796875,
0.1383056640625,
0.47216796875,
0.15283203125,
0.09722900390625,
0.15283203125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.468017578125,
0.1546630859375,
0.468017578125,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50244140625,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.080078125,
0.1685791015625,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50244140625,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.2008056640625,
0.46728515625,
0.2008056640625,
0.46728515625,
0.2108154296875,
0.08087158203125,
0.2108154296875
],
[
0.50341796875,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50341796875,
0.210205078125
],
[
0.080078125,
0.2139892578125,
0.468017578125,
0.2139892578125,
0.468017578125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2279052734375,
0.46728515625,
0.2291259765625,
0.46728515625,
0.242431640625,
0.080078125,
0.2412109375
],
[
0.50244140625,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.8935546875,
0.24365234375,
0.8935546875,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.2587890625,
0.78173828125,
0.2587890625,
0.78173828125,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.274658203125,
0.4697265625,
0.274658203125,
0.4697265625,
0.28857421875,
0.080078125,
0.28857421875
],
[
0.50244140625,
0.281005859375,
0.64794921875,
0.28271484375,
0.64794921875,
0.29736328125,
0.50244140625,
0.29541015625
],
[
0.07843017578125,
0.2890625,
0.4697265625,
0.28857421875,
0.4697265625,
0.302978515625,
0.07843017578125,
0.3037109375
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.51953125,
0.3037109375,
0.89208984375,
0.304931640625,
0.89208984375,
0.3193359375,
0.51953125,
0.318115234375
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50244140625,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.50244140625,
0.333984375
],
[
0.080078125,
0.335205078125,
0.471435546875,
0.335205078125,
0.471435546875,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50244140625,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.08087158203125,
0.3515625,
0.279296875,
0.3515625,
0.279296875,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.50244140625,
0.35107421875,
0.8896484375,
0.35107421875,
0.8896484375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.098876953125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.098876953125,
0.37890625
],
[
0.50244140625,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.50244140625,
0.380126953125
],
[
0.080078125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.080078125,
0.39453125
],
[
0.5009765625,
0.380615234375,
0.890625,
0.380126953125,
0.890625,
0.39453125,
0.5009765625,
0.395751953125
],
[
0.08087158203125,
0.396484375,
0.4697265625,
0.396484375,
0.4697265625,
0.410400390625,
0.08087158203125,
0.410400390625
],
[
0.50341796875,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.40771484375,
0.50341796875,
0.40771484375
],
[
0.080078125,
0.410888671875,
0.468017578125,
0.410888671875,
0.468017578125,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.410888671875,
0.890625,
0.410888671875,
0.890625,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.080078125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50244140625,
0.4267578125,
0.89208984375,
0.4267578125,
0.89208984375,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.080078125,
0.455078125
],
[
0.5009765625,
0.43994140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455810546875,
0.5009765625,
0.455078125
],
[
0.07843017578125,
0.455810546875,
0.46728515625,
0.45703125,
0.46728515625,
0.47021484375,
0.07843017578125,
0.4697265625
],
[
0.5009765625,
0.45703125,
0.79345703125,
0.45703125,
0.79345703125,
0.47021484375,
0.5009765625,
0.47021484375
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.521484375,
0.4716796875,
0.890625,
0.4716796875,
0.890625,
0.486083984375,
0.521484375,
0.486083984375
],
[
0.08087158203125,
0.488037109375,
0.465576171875,
0.488037109375,
0.465576171875,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50244140625,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.5166015625,
0.468017578125,
0.53076171875,
0.080078125,
0.52978515625
],
[
0.50244140625,
0.5166015625,
0.89208984375,
0.5166015625,
0.89208984375,
0.52978515625,
0.50244140625,
0.52978515625
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.5009765625,
0.5322265625,
0.740234375,
0.5322265625,
0.740234375,
0.54541015625,
0.5009765625,
0.54541015625
],
[
0.07843017578125,
0.54541015625,
0.468017578125,
0.546875,
0.468017578125,
0.56103515625,
0.07843017578125,
0.56005859375
],
[
0.080078125,
0.56103515625,
0.2607421875,
0.56103515625,
0.2607421875,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.50341796875,
0.56005859375,
0.63232421875,
0.56005859375,
0.63232421875,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.1004638671875,
0.57763671875,
0.468017578125,
0.57763671875,
0.468017578125,
0.59130859375,
0.1004638671875,
0.59130859375
],
[
0.521484375,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.07843017578125,
0.59130859375,
0.468017578125,
0.59228515625,
0.468017578125,
0.60693359375,
0.07843017578125,
0.60595703125
],
[
0.07843017578125,
0.60693359375,
0.4697265625,
0.60791015625,
0.4697265625,
0.62255859375,
0.07843017578125,
0.6220703125
],
[
0.50244140625,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.5009765625,
0.61572265625,
0.88818359375,
0.61572265625,
0.88818359375,
0.62890625,
0.5009765625,
0.62890625
],
[
0.080078125,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.080078125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.50341796875,
0.646484375,
0.890625,
0.646484375,
0.890625,
0.66015625,
0.50341796875,
0.66015625
],
[
0.07843017578125,
0.65283203125,
0.4697265625,
0.6513671875,
0.4697265625,
0.66650390625,
0.07843017578125,
0.66748046875
],
[
0.50341796875,
0.6611328125,
0.89208984375,
0.6611328125,
0.89208984375,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.080078125,
0.66845703125,
0.36279296875,
0.66845703125,
0.36279296875,
0.681640625,
0.080078125,
0.681640625
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.101318359375,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69873046875,
0.101318359375,
0.69873046875
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.69970703125,
0.468017578125,
0.69970703125,
0.468017578125,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.72021484375,
0.50244140625,
0.72021484375
],
[
0.080078125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72900390625,
0.080078125,
0.72900390625
],
[
0.50341796875,
0.72265625,
0.560546875,
0.72265625,
0.560546875,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.07843017578125,
0.72900390625,
0.468017578125,
0.72998046875,
0.468017578125,
0.744140625,
0.07843017578125,
0.74365234375
],
[
0.521484375,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.521484375,
0.75
],
[
0.080078125,
0.744140625,
0.468017578125,
0.744140625,
0.468017578125,
0.75830078125,
0.080078125,
0.75830078125
],
[
0.5009765625,
0.75,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.5009765625,
0.7646484375
],
[
0.07843017578125,
0.76025390625,
0.44677734375,
0.76025390625,
0.44677734375,
0.7734375,
0.07843017578125,
0.7734375
],
[
0.50341796875,
0.76806640625,
0.890625,
0.76806640625,
0.890625,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.78515625,
0.268798828125,
0.7861328125,
0.268798828125,
0.80029296875,
0.080078125,
0.7998046875
],
[
0.50341796875,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.50341796875,
0.79541015625
],
[
0.5009765625,
0.79443359375,
0.8935546875,
0.79443359375,
0.8935546875,
0.8125,
0.5009765625,
0.8125
],
[
0.098876953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.098876953125,
0.82373046875
],
[
0.5009765625,
0.81103515625,
0.89208984375,
0.8125,
0.89208984375,
0.8271484375,
0.5009765625,
0.82568359375
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50341796875,
0.82763671875,
0.5888671875,
0.82763671875,
0.5888671875,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.0767822265625,
0.85498046875,
0.4697265625,
0.85595703125,
0.4697265625,
0.87060546875,
0.0767822265625,
0.869140625
],
[
0.52685546875,
0.85107421875,
0.8935546875,
0.85107421875,
0.8935546875,
0.86865234375,
0.52685546875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.8994140625
]
] |
[
"ing the target of a successful attack. C&w [2] transforms",
"the model training based on the generated adversarial exam-",
"thetaskintoa constrainedoptimizationproblemand com-",
"ples. Input transformation utilizes JPEG compression [12]",
"pares the effects of multiple objective functions.",
"denoising [26], and extra GAN model [39] before feeding",
"Score-based Black-box Attack.This category of",
"the images to the model. Randomization refers to adding",
"attack methods assumes they can obtain the classifica-",
"random noise into the input examples [49] or models [7] to",
"tion probabilities of a given input image from the target",
"make the model more robust to adversarial examples. The",
"model. zoo [3] proposes a zeroth-order optimization-",
"model ensemble means the ensemble of multiple models",
"basedmethod.whichdirectlyestimatesthegradientsofthe",
"intheoutput layer.Compared witha singlemodel,itcan",
"target model for generating adversarial examples. However,",
"reduce the impact of distributions in the adversarial exam-",
"it suffers from a low attack success rate and poor query ef-",
"ples [22, 29]. In addition, some work [37, 46] proves that",
"ficiency since it is non-trivial to estimate the gradient with",
"under a specific target model, a certified defense model can",
"limited information. To address these problems, P-RGR [5]",
"ensure robustness to adversarial examples.",
"utilizes the transfer-based prior and query information to",
"2.3. Meta-learning",
"improve query efficiency.NATTACK [25] further uses a",
"probability density distribution over a small region centered",
"Meta-learning is a concept of learning how to learn. It",
"around the input to generate adversarial examples, which",
"mainly focuses on how to use a large amount of data to learn",
"defeats both vanilla DNNs and those generated by various",
"a learning mode. When encountering a new task later, only",
"defensetechniques developed.",
"a small amount of additional data is needed to quickly com-",
"Decision-based Black-box Attack. Under the decision-",
"plete the new task by fine-tuning the model [13, 34]. As a",
"based black-box attack setting, only the predicted class of a",
"newly proposed method in recent years, meta-learning has",
"given input image from the target model is available, which",
"been widely used in various tasks. Meta-Attack [lll uses",
"seems more difficult than the score-based black-box attack.",
"the meta-learning framework to train the gradient estima-",
"Boundary Attack [l] is firstly proposed to this problem,",
"tor. In the case of the score-based black-box attack, only a",
"which gradually minimizes the adversarial perturbation by",
"small number of queries are needed to fine-tune the gradient",
"the approach of random walking while maintaining the ag-",
"estimator to complete the adversarial attack.",
"gressiveness of the adversarial example. Cheng et al. [4]",
"Different from Meta-Attack, our proposed MGAA is ac-",
"transforms the problem into a continuous real-valued op-",
"tually not a meta-learning framework but only inspired by",
"timization problem that can be solved by any zeroth-order",
"its philosophy.Specifically, we do not need to explicitly",
"optimization algorithm.Different from previous methods,",
"train an extra model but directly generate the adversarial",
"Dong et al. [10] proposes an evolutionary attack algorithm,",
"examples in the process of iteration.",
"which reduces the dimension of the search space and im-",
"proves the query efficiency.",
"3.Methodology",
"Transfer-based Black-box Attack. The transfer-based",
"Since our proposed architecture is plug-and-play and",
"black-box attack cannot obtain any information of the tar-",
"get model, which is the most challenging setting. Existing",
"can be integrated with any existing gradient-based attack",
"method, we first introduce some typical gradient-based at-",
"methods [8, 50, 9, 54, 27] mainly endeavor to improve the",
"tack methods in Sec. 3.1. And then, we introduce our pro-",
"attack success rates by the transferability of the adversar-",
"posed Meta Gradient Adversarial Attack architecture in de-",
"ial examples generated from the substitute model.We will",
"tail in Sec. 3.2. Moreover, we give a detailed analysis of",
"review these methods in Sec. 3.1 in details.",
"the effectiveness of our proposed architecture theoretically",
"Although these methods have been proposed to gradu-",
"in Sec. 3.3. Finally, a discussion on the differences between",
"ally improve the transferability of adversarial examples, the",
"our proposed MGAA and some related works is given in",
"results in black-box attacks still leave a lot of room for im-",
"Sec.3.4.",
"provement. Our work solves this problem from the perspec-",
"Let r denote the benign example and y denote the true",
"tive of meta-learning, which reduces the gap between white-",
"label corresponding to the benign example. f() : X → )",
"box and black-box attacks, thus achieving better results.",
"denotes the target classifier,J denotes the loss function",
"2.2. Adversarial Defense",
"used for training the classifier, which usually means cross",
"entropy loss. radu denotes the adversarial example that",
"The task of adversarial defense is to improve the robust-",
"needs to be optimized. The optimization objective can be",
"ness of the model so that the model can correctly classify",
"described as:",
"the perturbed adversarial examples. Defense methods can",
"be divided into five categories: adversarial training, input",
"argmax J(f(radu),y), s.t:llradu-allo ≤e, (1)",
"transformation, randomization, model ensemble, and cer-",
"tified defenses. Adversarial training [45, 31, 41] conducts",
"where e is the maximum perturbation per pixel allowed to"
] |
{
"arxiv_id": "2108.04204",
"language": "en",
"page_idx": 2
}
|
||
2108.04204-en-3
|
2108.04204-en-3.jpg
|
add in the benign example. The goal of the adversarial example is to mislead the classifier into misclassification $(f(x^{\hat{adv}}\neq y))$ with less perturbation added in compared to the benign example.
### 3.1. Gradient-Based Attack Method
We briefly introduce several gradient-based attack methods in this section, which focus on solving the problem of the white-box or transfer-based black-box attack.
Fast Gradient Sign Method. FGSM[15] is a one-step method for white-box attack, which update the adversarial example in the direction of maximizing the loss function:
$$x^{adv}=x+\epsilon\cdot\mathrm{sign}(\nabla_{x}J(f(x),y)).$$
(2)
Basic Iterative Method. Extended by FGSM, BIM[21] proposes an iteration method to improve the success rate of white-box attack, which can be described as:
(3)
$$x_{t+1}^{adv}=x_{t}^{adv}+\alpha\cdot\mathrm{sign}(\nabla_{x}J(f(x_{t}^{adv}),y)),$$
where $x_0^{adv}= x$, $\alpha = \epsilon / T$ and $T$ is the number of iterations.
Momentum Iterative Fast Gradient Sign Method. A momentum term is further proposed in MIM [8] to accumulate the update direction of previous steps in the process of iteration:
(4)
$$g_{t+1}=\mu\cdot g_{t}+\frac{\nabla_{x}J(f(x_{t}^{adv}),y)}{\|\nabla_{x}J(f(x_{t}^{adv}),y)\|_{1}},\\x_{t+1}^{adv}=x_{t}^{adv}+\alpha\cdot\mathrm{sign}(g_{t+1}),$$
(5)
where $g_t$ denotes the momentum term of gradient in the $t$-th iteration and $\mu$ is a decay factor. Compared with BIM, MIM achieves significant progress in the black-box attack.
Diverse Inputs Method. DIM [50] proposes to apply random resizing and padding transformations in the benign example to improve the transferability of adversarial examples. Although this method is simple and easy to implement, it also brings great improvements.
Translation-Invariant Attack Method. Different from previous methods, TIM[9] proposes to convolve the gradient with a Gaussian kernel, which can be combined with $MIM:$
(6)
$$g_{t+1}=\mu\cdot g_{t}+\frac{W*\nabla_{x}J(f(x_{t}^{adv}),y)}{\|W*\nabla_{x}J(f(x_{t}^{adv}),y)\|_{1}},\\x_{t+1}^{adv}=x_{t}^{adv}+\alpha\cdot\mathrm{sign}(g_{t+1}),$$
(7)
where $W$ is a Gaussian kernel and *denotes the operator of
convolution.
SI-NI. Moreover, SI-NI[27] proposes two approaches to further improve the transferability of adversarial examples,
$i.e.$, NI-FGSM and SIM. NI-FGSM integrates Nesterov Accelerated Gradient[33] into the iterative gradient-based attack, e.g. MIM, to conduct a robust adversarial attack:
(8)
$$x_{t}^{nes}=x_{t}^{adv}+\alpha\cdot\mu\cdot g_{t},\\g_{t+1}=\mu\cdot g_{t}+\frac{\nabla_{x}J(f(x_{t}^{nes}),y)}{\|\nabla_{x}J(f(x_{t}^{nes}),y)\|_{1}},\\x_{t+1}^{adv}=x_{t}^{adv}+\alpha\cdot\mathrm{sign}(g_{t+1}).$$
(9)
(10)
SIM optimizes the adversarial examples over the scale
copies of the benign example:
(11)
$$\arg\max\limits_{x^{adv}}\frac{1}{m}\sum_{i=0}^{m}J(f(S_{i}(x^{adv})),y),$$
$$\mathrm{s.t.}\|x^{adv}-x\|_{\infty}\leq\epsilon,$$
where $S_i(x)=x/2^i$ denotes the scaled image and $m$ de-
notes the number of scale copies.
3.2. Meta Gradient Adversarial Attack
Most of the existing transfer-based black-box attack methods use gradient-based methods to generate adversarial examples against the white-box models, and then directly utilize the transferability of the adversarial examples to conduct the attack against the black-box model. However, due to the differences in the structures and the parameters between different models, relying only on the transferability of adversarial examples to conduct the black-box attack cannot achieve perfect results, which leaves a lot of room for improvement.
In this paper, we solve the problem of the black-box attack from a different perspective. Inspired by the philosophy of meta-learning, we propose an architecture called Meta Gradient Adversarial Attack (MGAA), which is plugand-play and can be integrated with any existing gradientbased attack method. As shown in Fig. 1, $T$ tasks are sampled from a model zoo iteratively. In each task, we simulate a white-box attack in the meta-train step and a black-box attack in the meta-test step, narrowing the gap of gradient directions between white-box and black-box settings to improve the transferability
Specifically, suppose there are a total of $N$ models of $M_1,M_2,\cdots,M_N$ in the model zoo, we randomly select $n+1$ models to compose a task in each iteration, which consists of two steps, $i.e.$, meta-train and meta-test. At the beginning of each iteration (take $i$-th iteration as an example), we take the adversarial example generated in the previous iteration $x_i$ as the input, denoted as $x_i,0.$
$\textbf{Meta- train. A total of }n$ models of $M_k_1,M_{k_2},\cdots,M_{k_n}$ are utilized to simulate the white-box attack. We employ the same approach as [8] when doing the adversarial attack by the ensemble of multiple models. Specifically, to attack the ensemble of $n$ models, we fuse the logits as:
$$l(x_{i,0})=\sum_{s=1}^nw_sl_{k_s}(x_{i,0}),$$
(12)
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08251953125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08251953125,
0.11932373046875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.5009765625,
0.12371826171875,
0.85546875,
0.12371826171875,
0.85546875,
0.136962890625,
0.5009765625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.214111328125,
0.1383056640625,
0.214111328125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.69384765625,
0.143310546875,
0.78173828125,
0.146484375,
0.78173828125,
0.1610107421875,
0.69384765625,
0.1578369140625
],
[
0.86767578125,
0.143310546875,
0.89453125,
0.143310546875,
0.89453125,
0.160400390625,
0.86767578125,
0.160400390625
],
[
0.689453125,
0.1634521484375,
0.8056640625,
0.1634521484375,
0.8056640625,
0.177978515625,
0.689453125,
0.177978515625
],
[
0.61279296875,
0.14453125,
0.685546875,
0.14453125,
0.685546875,
0.1590576171875,
0.61279296875,
0.1590576171875
],
[
0.080078125,
0.1666259765625,
0.35546875,
0.1666259765625,
0.35546875,
0.179931640625,
0.080078125,
0.179931640625
],
[
0.56787109375,
0.17041015625,
0.68310546875,
0.173583984375,
0.68310546875,
0.189453125,
0.56787109375,
0.186279296875
],
[
0.86767578125,
0.17041015625,
0.89453125,
0.17041015625,
0.89453125,
0.1875,
0.86767578125,
0.1875
],
[
0.67236328125,
0.177978515625,
0.82177734375,
0.1787109375,
0.82177734375,
0.1962890625,
0.67236328125,
0.1956787109375
],
[
0.09722900390625,
0.189453125,
0.46728515625,
0.1900634765625,
0.46728515625,
0.2052001953125,
0.09722900390625,
0.203857421875
],
[
0.59326171875,
0.1962890625,
0.80126953125,
0.1962890625,
0.80126953125,
0.2171630859375,
0.59326171875,
0.2171630859375
],
[
0.859375,
0.199462890625,
0.89453125,
0.199462890625,
0.89453125,
0.21533203125,
0.859375,
0.21533203125
],
[
0.080078125,
0.205810546875,
0.4697265625,
0.205810546875,
0.4697265625,
0.2197265625,
0.080078125,
0.2197265625
],
[
0.080078125,
0.2215576171875,
0.40185546875,
0.2215576171875,
0.40185546875,
0.23486328125,
0.080078125,
0.23486328125
],
[
0.50341796875,
0.222900390625,
0.89208984375,
0.222900390625,
0.89208984375,
0.236083984375,
0.50341796875,
0.236083984375
],
[
0.096435546875,
0.236083984375,
0.468017578125,
0.2374267578125,
0.468017578125,
0.251953125,
0.096435546875,
0.250732421875
],
[
0.50244140625,
0.2374267578125,
0.70361328125,
0.2374267578125,
0.70361328125,
0.250732421875,
0.50244140625,
0.250732421875
],
[
0.08087158203125,
0.25390625,
0.46728515625,
0.25390625,
0.46728515625,
0.264404296875,
0.08087158203125,
0.264404296875
],
[
0.5703125,
0.266357421875,
0.63232421875,
0.266357421875,
0.63232421875,
0.27783203125,
0.5703125,
0.27783203125
],
[
0.69384765625,
0.26318359375,
0.82275390625,
0.26318359375,
0.82275390625,
0.281005859375,
0.69384765625,
0.281005859375
],
[
0.861328125,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.2890625,
0.861328125,
0.2890625
],
[
0.080078125,
0.267578125,
0.45751953125,
0.267578125,
0.45751953125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.63232421875,
0.279052734375,
0.646484375,
0.279052734375,
0.646484375,
0.28271484375,
0.63232421875,
0.28271484375
],
[
0.68310546875,
0.27783203125,
0.689453125,
0.27783203125,
0.689453125,
0.281005859375,
0.68310546875,
0.281005859375
],
[
0.67333984375,
0.2890625,
0.826171875,
0.2890625,
0.826171875,
0.31005859375,
0.67333984375,
0.31005859375
],
[
0.583984375,
0.27978515625,
0.60888671875,
0.27783203125,
0.609375,
0.285888671875,
0.5849609375,
0.28857421875
],
[
0.1502685546875,
0.29345703125,
0.39453125,
0.296630859375,
0.39453125,
0.314453125,
0.1502685546875,
0.311279296875
],
[
0.44677734375,
0.29736328125,
0.4697265625,
0.29736328125,
0.4697265625,
0.311279296875,
0.44677734375,
0.311279296875
],
[
0.50341796875,
0.31640625,
0.890625,
0.31640625,
0.890625,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.098876953125,
0.328857421875,
0.46728515625,
0.328857421875,
0.46728515625,
0.34228515625,
0.098876953125,
0.34228515625
],
[
0.50244140625,
0.330810546875,
0.72119140625,
0.33203125,
0.72119140625,
0.34521484375,
0.50244140625,
0.3447265625
],
[
0.08087158203125,
0.3447265625,
0.4697265625,
0.3447265625,
0.4697265625,
0.35791015625,
0.08087158203125,
0.35791015625
],
[
0.50341796875,
0.352783203125,
0.80126953125,
0.354248046875,
0.80126953125,
0.36865234375,
0.50341796875,
0.367431640625
],
[
0.080078125,
0.35986328125,
0.374267578125,
0.35986328125,
0.374267578125,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.51953125,
0.375732421875,
0.89208984375,
0.376953125,
0.89208984375,
0.391357421875,
0.51953125,
0.39013671875
],
[
0.12744140625,
0.386474609375,
0.423095703125,
0.3837890625,
0.423095703125,
0.406005859375,
0.12744140625,
0.40771484375
],
[
0.146240234375,
0.392578125,
0.1568603515625,
0.392578125,
0.1568603515625,
0.395751953125,
0.146240234375,
0.395751953125
],
[
0.4453125,
0.391357421875,
0.468017578125,
0.391357421875,
0.468017578125,
0.40283203125,
0.4453125,
0.40283203125
],
[
0.50244140625,
0.393310546875,
0.89208984375,
0.393310546875,
0.89208984375,
0.4072265625,
0.50244140625,
0.4072265625
],
[
0.50341796875,
0.40771484375,
0.890625,
0.40771484375,
0.890625,
0.421630859375,
0.50341796875,
0.421630859375
],
[
0.08087158203125,
0.419189453125,
0.465576171875,
0.419189453125,
0.465576171875,
0.43310546875,
0.08087158203125,
0.43310546875
],
[
0.50341796875,
0.42236328125,
0.890625,
0.42236328125,
0.890625,
0.436279296875,
0.50341796875,
0.436279296875
],
[
0.098876953125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.449462890625,
0.098876953125,
0.449462890625
],
[
0.50244140625,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.451904296875,
0.50244140625,
0.451904296875
],
[
0.080078125,
0.451904296875,
0.46728515625,
0.451904296875,
0.46728515625,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.50244140625,
0.45263671875,
0.890625,
0.45263671875,
0.890625,
0.466552734375,
0.50244140625,
0.466552734375
],
[
0.080078125,
0.46728515625,
0.4697265625,
0.46728515625,
0.4697265625,
0.48095703125,
0.080078125,
0.48095703125
],
[
0.50244140625,
0.468505859375,
0.890625,
0.468505859375,
0.890625,
0.481689453125,
0.50244140625,
0.481689453125
],
[
0.080078125,
0.48291015625,
0.143798828125,
0.48291015625,
0.143798828125,
0.494384765625,
0.080078125,
0.494384765625
],
[
0.5048828125,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.49560546875,
0.5048828125,
0.49560546875
],
[
0.50244140625,
0.498779296875,
0.89208984375,
0.498779296875,
0.89208984375,
0.51220703125,
0.50244140625,
0.51220703125
],
[
0.2646484375,
0.5048828125,
0.3857421875,
0.5068359375,
0.3857421875,
0.5244140625,
0.2646484375,
0.52294921875
],
[
0.5009765625,
0.51220703125,
0.619140625,
0.51318359375,
0.619140625,
0.52783203125,
0.5009765625,
0.5263671875
],
[
0.146240234375,
0.52099609375,
0.253173828125,
0.52099609375,
0.253173828125,
0.5341796875,
0.146240234375,
0.5341796875
],
[
0.44677734375,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.44677734375,
0.5322265625
],
[
0.253173828125,
0.5263671875,
0.398681640625,
0.5263671875,
0.398681640625,
0.54150390625,
0.253173828125,
0.54150390625
],
[
0.521484375,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54248046875,
0.521484375,
0.54248046875
],
[
0.1707763671875,
0.54248046875,
0.37744140625,
0.54248046875,
0.37744140625,
0.56298828125,
0.1707763671875,
0.56298828125
],
[
0.4453125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.4453125,
0.56005859375
],
[
0.50244140625,
0.54443359375,
0.8896484375,
0.54443359375,
0.8896484375,
0.55810546875,
0.50244140625,
0.55810546875
],
[
0.50244140625,
0.5595703125,
0.89208984375,
0.5595703125,
0.89208984375,
0.57275390625,
0.50244140625,
0.57275390625
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57470703125,
0.468017578125,
0.5888671875,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.57373046875,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50341796875,
0.5888671875,
0.890625,
0.5888671875,
0.890625,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.080078125,
0.60693359375,
0.4248046875,
0.60693359375,
0.4248046875,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50341796875,
0.60400390625,
0.89208984375,
0.60400390625,
0.89208984375,
0.6171875,
0.50341796875,
0.6171875
],
[
0.096435546875,
0.6220703125,
0.468017578125,
0.62255859375,
0.468017578125,
0.63720703125,
0.096435546875,
0.63623046875
],
[
0.50341796875,
0.61865234375,
0.89208984375,
0.61865234375,
0.89208984375,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.07843017578125,
0.6396484375,
0.4697265625,
0.6396484375,
0.4697265625,
0.65283203125,
0.07843017578125,
0.65283203125
],
[
0.50244140625,
0.6337890625,
0.890625,
0.6337890625,
0.890625,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.080078125,
0.65380859375,
0.468017578125,
0.65380859375,
0.468017578125,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.50244140625,
0.6494140625,
0.89208984375,
0.6494140625,
0.89208984375,
0.66357421875,
0.50244140625,
0.66357421875
],
[
0.080078125,
0.66943359375,
0.468017578125,
0.66943359375,
0.468017578125,
0.68310546875,
0.080078125,
0.68310546875
],
[
0.50244140625,
0.6640625,
0.890625,
0.6640625,
0.890625,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.080078125,
0.6845703125,
0.343017578125,
0.6845703125,
0.343017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.6787109375,
0.66650390625,
0.6787109375,
0.66650390625,
0.6923828125,
0.50244140625,
0.6923828125
],
[
0.098876953125,
0.69970703125,
0.468017578125,
0.69970703125,
0.468017578125,
0.71337890625,
0.098876953125,
0.71337890625
],
[
0.521484375,
0.69580078125,
0.89208984375,
0.693359375,
0.89208984375,
0.70703125,
0.521484375,
0.708984375
],
[
0.080078125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72900390625,
0.080078125,
0.72900390625
],
[
0.50341796875,
0.71044921875,
0.89208984375,
0.71044921875,
0.89208984375,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.080078125,
0.72998046875,
0.4697265625,
0.72998046875,
0.4697265625,
0.74365234375,
0.080078125,
0.74365234375
],
[
0.5048828125,
0.7265625,
0.890625,
0.7265625,
0.890625,
0.7373046875,
0.5048828125,
0.7373046875
],
[
0.080078125,
0.74560546875,
0.1217041015625,
0.74560546875,
0.1217041015625,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.50341796875,
0.74072265625,
0.89208984375,
0.74072265625,
0.89208984375,
0.75390625,
0.50341796875,
0.75390625
],
[
0.50244140625,
0.7548828125,
0.89208984375,
0.7548828125,
0.89208984375,
0.76806640625,
0.50244140625,
0.76806640625
],
[
0.250732421875,
0.771484375,
0.398681640625,
0.771484375,
0.398681640625,
0.7861328125,
0.250732421875,
0.7861328125
],
[
0.50341796875,
0.77099609375,
0.89208984375,
0.77099609375,
0.89208984375,
0.7841796875,
0.50341796875,
0.7841796875
],
[
0.1298828125,
0.7841796875,
0.2344970703125,
0.7841796875,
0.2344970703125,
0.796875,
0.1298828125,
0.796875
],
[
0.44677734375,
0.7822265625,
0.4697265625,
0.7822265625,
0.4697265625,
0.79541015625,
0.44677734375,
0.79541015625
],
[
0.50244140625,
0.7841796875,
0.80615234375,
0.78515625,
0.80615234375,
0.7998046875,
0.50244140625,
0.798828125
],
[
0.51953125,
0.798828125,
0.88623046875,
0.7998046875,
0.88623046875,
0.814453125,
0.51953125,
0.81298828125
],
[
0.236083984375,
0.787109375,
0.41259765625,
0.7890625,
0.41259765625,
0.806640625,
0.236083984375,
0.8046875
],
[
0.1690673828125,
0.80615234375,
0.37744140625,
0.80615234375,
0.37744140625,
0.8271484375,
0.1690673828125,
0.8271484375
],
[
0.44677734375,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.44677734375,
0.82373046875
],
[
0.5048828125,
0.8173828125,
0.890625,
0.8173828125,
0.890625,
0.82763671875,
0.5048828125,
0.82763671875
],
[
0.50244140625,
0.83056640625,
0.890625,
0.83056640625,
0.890625,
0.8447265625,
0.50244140625,
0.8447265625
],
[
0.080078125,
0.83837890625,
0.4697265625,
0.8388671875,
0.4697265625,
0.85400390625,
0.080078125,
0.85302734375
],
[
0.50341796875,
0.84765625,
0.890625,
0.84765625,
0.890625,
0.85791015625,
0.50341796875,
0.85791015625
],
[
0.080078125,
0.85693359375,
0.1650390625,
0.85693359375,
0.1650390625,
0.8671875,
0.080078125,
0.8671875
],
[
0.50244140625,
0.861328125,
0.7958984375,
0.861328125,
0.7958984375,
0.875,
0.50244140625,
0.875
],
[
0.09722900390625,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.09722900390625,
0.88623046875
],
[
0.59716796875,
0.880859375,
0.689453125,
0.8837890625,
0.68896484375,
0.90283203125,
0.59619140625,
0.8994140625
],
[
0.6953125,
0.8837890625,
0.7958984375,
0.8837890625,
0.7958984375,
0.9013671875,
0.6953125,
0.9013671875
],
[
0.859375,
0.8837890625,
0.8935546875,
0.8837890625,
0.8935546875,
0.8994140625,
0.859375,
0.8994140625
],
[
0.69287109375,
0.89599609375,
0.70361328125,
0.89599609375,
0.70361328125,
0.8994140625,
0.69287109375,
0.8994140625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
]
] |
[
"add in the benign example.The goal of the adversarial",
"i.e., NI-FGSM and SIM. NI-FGSM integrates Nesterov Ac-",
"exampleis tomislead the classifierinto misclassification",
"celerated Gradient [33] into the iterative gradient-based at-",
"(f(radu + y) with less perturbation added in compared to",
"tack, e.g. MIM, to conduct a robust adversarial attack:",
"the benign example.",
"+a.μ.gt,",
"(8)",
"VrJ(f(cnes),y)",
"rnes = a",
"3.1. Gradient-Based Attack Method",
"gt+1=μ·9t+",
"(9)",
"J(f(rnes),y)l",
"We briefly introduce several gradient-based attack meth-",
"atd = aad + Q- sign(gt+1).",
"(10)",
"ods in this section, which focus on solving the problem of",
"the white-box or transfer-based black-box attack.",
"SIM optimizes the adversarial examples over the scale",
"Fast Gradient Sign Method. FGSM [15] is a one-step",
"copies of the benign example:",
"methodforwhite-box attack,whichupdatetheadversarial",
"argmax",
"J(f(S;(rad)), y),",
"(11)",
"example in the direction of maximizing the loss function:",
"",
"",
"Ss.t.radu - all ≤e,",
"radt",
"aadu = r +E· sign(VaJ(f(r),y).",
"(2)",
"where S;(r) = /2\" denotes the scaled image and m de-",
"Basic Iterative Method. Extended by FGSM, BIM [2]]",
"notes the number of scale copies.",
"proposes an iteration method to improve the success rate of",
"3.2. Meta Gradient Adversarial Attack",
"white-box attack, which can be described as:",
"Most of the existing transfer-based black-box attack",
" = aadu +a sign(VaJ(f(agdr),g),",
"",
"(3)",
"methods use gradient-based methods to generate adversarial",
"examples against the white-box models, and then directly",
"where radu=r, Q=e/T and T is the number of iterations.",
"utilize the transferability of the adversarial examples to con-",
"Momentum Iterative Fast Gradient Sign Method.A",
"duct the attack against the black-box model. However, due",
"momentum term is further proposed in MIM [8] to accumu-",
"to the differences in the structures and the parameters be-",
"late the update direction of previous steps in the process of",
"tween different models, relying only on the transferabil-",
"iteration:",
"ity of adversarial examplesto conduct theblack-box attack",
"cannot achieve perfect results, which leaves a lot of room",
"VJ(f(radv), y)",
"for improvement.",
"9t+1=μ:9t +",
"(4)",
"/VJ(f(aadu),y)ll1",
"In this paper, we solve the problem of the black-box at-",
"a1 = agdu +a sign(gt+1),",
"(5)",
"tack from a different perspective.: Inspired by the philos-",
"ophy of meta-learning, we propose an architecture called",
"where gt denotes the momentum term of gradient in the t-th",
"Meta Gradient Adversarial Attack (MGAA), which is plug-",
"iteration and μ is a decay factor. Compared with BIM, MIM",
"and-play and can be integrated with any existing gradient-",
"achieves significant progress in the black-box attack.",
"based attack method. As shown in Fig. 1, T tasks are sam-",
"Diverse Inputs Method. DIM [5O] proposes to apply",
"pled from a model zoo iteratively. In each task, we simulate",
"random resizing and padding transformations in the benign",
"a white-box attack in the meta-train step and a black-box",
"example to improve the transferability of adversarial exam-",
"attack in the meta-test step, narrowing the gap of gradient",
"ples. Although this method is simple and easy to imple-",
"directions between white-box and black-box settings to im-",
"ment, it also brings great improvements.",
"prove the transferability.",
"Translation-Invariant Attack Method. Different from",
"Specifically,suppose there are a total of N models of",
"previous methods, TIM [9] proposes to convolve the gra-",
"Mi, M2,*.., Mn in the model zoo, we randomly select",
"dient with a Gaussian kernel, which can be combined with",
"n + 1models to composeatask in eachiteration,which",
"MIM:",
"consists of two steps, i.e., meta-train and meta-test. At the",
"beginning of each iteration (take i-th iteration as an exam-",
"W *VrJ(f(radu),y)",
"ple), we take the adversarial example generated in the pre-",
"gt+1 = μ: 9t +",
"(6)",
"vious iteration ri as the input, denoted as &i,o.",
"Meta-train.Atotal of nmodels of Mki,Mka,., Mk,",
"IW*J(f(ad),y)l1",
"a = aqdu + α sign(g+1),",
"(7)",
"areutilizedtosimulatethewhite-boxattack.Weemploythe",
"same approach as [8] when doing the adversarial attack by",
"where W is a Gaussian kernel and * denotes the operator of",
"theensembleofmultiplemodels.Specifically,toattackthe",
"convolution.",
"ensemble of n models, we fuse the logits as:",
"SI-NI. Moreover, SI-NI [27] proposes two approaches to",
"l(ri0) =",
"-, wslk (aio),",
"(12)",
"",
"further improve the transferability of adversarial examples,"
] |
{
"arxiv_id": "2108.04204",
"language": "en",
"page_idx": 3
}
|
||
2012.07657-en-0
|
2012.07657-en-0.jpg
|
Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection
Alexandros Haliassos$^{1, \dagger }$ Konstantinos Vougioukas$^{1}$
Stavros Petridis$^{1}$
Maja Pantic$^{1,2}$
$^{1}$Imperial College London
$^{2}$Facebook London
$ন্ন$
### $\mathbf{Abstract}$
$\begin{array}{ccc}&Although¤t&deep&learning-based&face&forgery&de-1\\5&tectors&achieve&impressive&performance&in&constrained&sce-1\\7&narios,&they&are&vulnerable&to&samples&created&by&unseen\end{array}$
manipulation methods. Some recent works show improvements in generalisation but rely on cues that are easily corrupted by common post-processing operations such as compression. In this paper, we propose LipForensics, a detection approach capable of both generalising to novel manipulations and withstanding various distortions. LipForensics targets high-level semantic irregularities in mouth movements, which are common in many generated videos. It consists in first pretraining a spatio-temporal network to perform visual speech recognition (lipreading), thus learning rich internal representations related to natural mouth motion. A temporal network is subsequently finetuned on fixed mouth embeddings of real and forged data in order to detect fake videos based on mouth movements without overfuting to low-level, manipulation-specific artefacts. Extensive experiments show that this simple approach significantly surpasses the state-of-the-art in terms of generalisation to unseen manipulations and robustness to perturbations, as well as shed light on the factors responsible for its performance. Code is available on GitHub.
## $1. \textbf{Introduction}$
Recent advances in deep generative models, especially Generative Adversarial Networks (GANs)[26], have enabled the use of off-the-shelf models that can produce ultrarealistic fake videos with little human effort or expertise. Face manipulation methods in particular have raised considerable concerns due to their ability to alter a person's identity [42,34,38], expression [57, 56, 48], or lip movements [55,62,61] to match the face in a given target video. The misuse of such technology can spread political propaganda, defame individuals, or damage trust in journalism.
$^{\dagger}$Corresponding author.
$^{\mathrm{l}}$https://github.com/ahaliassos/LipForensics
<FigureHere>
Figure 1. Illustration of our training and testing objectives
By leveraging semantically high-level, spatio-temporal representations learned via the task of lipreading, our face forgery detector can handle both novel manipulations and common corruptions at test-time.
With the aid of recent releases of large-scale face forgery datasets [50,42,34,18], it is possible to train deep convolutional neural networks (CNNs) to detect fake videos [3,27,69,4,15,49,46,40]. Despite excellent accuracy on samples that are independent and identically distributed to the training distribution, it is known that dramatic drops in performance may be experienced when this stringent criterion is not met [39, 9, 47]. For example, detectors often significantly underperform on novel forgery types. This understandably sparks concerns, as a deployed detector is unlikely to be exposed only to forgeries seen during training. Moreover, they are often sensitive to common perturbations such as compression and are, as a result, vulnerable to the image processing pipelines on social networks.
Recent attempts to boost generalisation to novel forgeries include simple data augmentations [64], a two-branch network that amplifies multi-band frequencies [45], an autoencoder-like structure to serve as an anomaly detector [14, 19, 47], and patch-based classification to model local patterns [9]. However, these methods still substantially overfit to seen manipulations. A particularly effective method is Face X-ray [39], which proposes to predict the blending boundary between the background image and the
1
|
[
[
0.08087158203125,
0.1331787109375,
0.8896484375,
0.1331787109375,
0.8896484375,
0.15087890625,
0.08087158203125,
0.15087890625
],
[
0.111083984375,
0.18115234375,
0.3056640625,
0.179931640625,
0.3056640625,
0.1976318359375,
0.111083984375,
0.1988525390625
],
[
0.330810546875,
0.18310546875,
0.5390625,
0.18310546875,
0.5390625,
0.1976318359375,
0.330810546875,
0.1976318359375
],
[
0.576171875,
0.18310546875,
0.70751953125,
0.18310546875,
0.70751953125,
0.1976318359375,
0.576171875,
0.1976318359375
],
[
0.74169921875,
0.1817626953125,
0.861328125,
0.1787109375,
0.861328125,
0.1962890625,
0.74169921875,
0.199462890625
],
[
0.27685546875,
0.2083740234375,
0.483642578125,
0.2083740234375,
0.483642578125,
0.222900390625,
0.27685546875,
0.222900390625
],
[
0.54150390625,
0.2052001953125,
0.69775390625,
0.20703125,
0.69775390625,
0.2247314453125,
0.54150390625,
0.222900390625
],
[
0.13232421875,
0.2279052734375,
0.833984375,
0.2279052734375,
0.833984375,
0.2412109375,
0.13232421875,
0.2412109375
],
[
0.0236968994140625,
0.26318359375,
0.05963134765625,
0.26318359375,
0.05963134765625,
0.317626953125,
0.0236968994140625,
0.317626953125
],
[
0.23779296875,
0.276611328125,
0.313720703125,
0.276611328125,
0.313720703125,
0.291748046875,
0.23779296875,
0.291748046875
],
[
0.57177734375,
0.279052734375,
0.6435546875,
0.279052734375,
0.6435546875,
0.290283203125,
0.57177734375,
0.290283203125
],
[
0.76123046875,
0.279052734375,
0.82666015625,
0.279052734375,
0.82666015625,
0.290283203125,
0.76123046875,
0.290283203125
],
[
0.5107421875,
0.296630859375,
0.68701171875,
0.296630859375,
0.68701171875,
0.31005859375,
0.5107421875,
0.31005859375
],
[
0.736328125,
0.296630859375,
0.84619140625,
0.296630859375,
0.84619140625,
0.31005859375,
0.736328125,
0.31005859375
],
[
0.0310516357421875,
0.30810546875,
0.06451416015625,
0.311279296875,
0.056365966796875,
0.3662109375,
0.0228729248046875,
0.363037109375
],
[
0.037567138671875,
0.311279296875,
0.043304443359375,
0.311279296875,
0.043304443359375,
0.31640625,
0.037567138671875,
0.31640625
],
[
0.098876953125,
0.309326171875,
0.46728515625,
0.309326171875,
0.46728515625,
0.323974609375,
0.098876953125,
0.323974609375
],
[
0.080078125,
0.323974609375,
0.465576171875,
0.323974609375,
0.465576171875,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.080078125,
0.34033203125,
0.468017578125,
0.34033203125,
0.468017578125,
0.354248046875,
0.080078125,
0.354248046875
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.02777099609375,
0.365478515625,
0.0555419921875,
0.365478515625,
0.0555419921875,
0.385009765625,
0.02777099609375,
0.385009765625
],
[
0.080078125,
0.37060546875,
0.468017578125,
0.37060546875,
0.468017578125,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.515625,
0.371826171875,
0.66845703125,
0.371826171875,
0.66845703125,
0.3818359375,
0.515625,
0.3818359375
],
[
0.7509765625,
0.37060546875,
0.8349609375,
0.37060546875,
0.8349609375,
0.3818359375,
0.7509765625,
0.3818359375
],
[
0.08251953125,
0.386962890625,
0.465576171875,
0.386962890625,
0.465576171875,
0.397705078125,
0.08251953125,
0.397705078125
],
[
0.080078125,
0.40087890625,
0.468017578125,
0.40087890625,
0.468017578125,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.0269622802734375,
0.4072265625,
0.052276611328125,
0.4072265625,
0.052276611328125,
0.44140625,
0.0269622802734375,
0.44140625
],
[
0.080078125,
0.416748046875,
0.46728515625,
0.416748046875,
0.46728515625,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.080078125,
0.43115234375,
0.471435546875,
0.43115234375,
0.471435546875,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.080078125,
0.44580078125,
0.4697265625,
0.44580078125,
0.4697265625,
0.458984375,
0.080078125,
0.458984375
],
[
0.50244140625,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.4609375,
0.468017578125,
0.4609375,
0.468017578125,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.03350830078125,
0.45068359375,
0.0555419921875,
0.45068359375,
0.0555419921875,
0.479248046875,
0.03350830078125,
0.479248046875
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.08087158203125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.08087158203125,
0.4892578125
],
[
0.50244140625,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.080078125,
0.4912109375,
0.4697265625,
0.4912109375,
0.4697265625,
0.5048828125,
0.080078125,
0.5048828125
],
[
0.5009765625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.5009765625,
0.50390625
],
[
0.02777099609375,
0.496826171875,
0.0555419921875,
0.496826171875,
0.0555419921875,
0.5146484375,
0.02777099609375,
0.5146484375
],
[
0.080078125,
0.505859375,
0.468017578125,
0.505859375,
0.468017578125,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.505859375,
0.56201171875,
0.505859375,
0.56201171875,
0.517578125,
0.50244140625,
0.517578125
],
[
0.080078125,
0.521484375,
0.4697265625,
0.521484375,
0.4697265625,
0.53515625,
0.080078125,
0.53515625
],
[
0.035125732421875,
0.5087890625,
0.051483154296875,
0.5087890625,
0.051483154296875,
0.5263671875,
0.035125732421875,
0.5263671875
],
[
0.02532958984375,
0.52880859375,
0.056365966796875,
0.52880859375,
0.056365966796875,
0.5673828125,
0.02532958984375,
0.5673828125
],
[
0.03350830078125,
0.5234375,
0.037567138671875,
0.5234375,
0.037567138671875,
0.53271484375,
0.03350830078125,
0.53271484375
],
[
0.080078125,
0.53515625,
0.468017578125,
0.5361328125,
0.468017578125,
0.55029296875,
0.080078125,
0.5498046875
],
[
0.080078125,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.521484375,
0.548828125,
0.890625,
0.5498046875,
0.890625,
0.564453125,
0.521484375,
0.56298828125
],
[
0.0318603515625,
0.55810546875,
0.0482177734375,
0.5625,
0.037567138671875,
0.5859375,
0.021240234375,
0.58203125
],
[
0.07843017578125,
0.56494140625,
0.46728515625,
0.56640625,
0.46728515625,
0.58056640625,
0.07843017578125,
0.580078125
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.59521484375,
0.080078125,
0.59521484375
],
[
0.50244140625,
0.580078125,
0.89208984375,
0.580078125,
0.89208984375,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.02777099609375,
0.5888671875,
0.052276611328125,
0.5888671875,
0.052276611328125,
0.611328125,
0.02777099609375,
0.611328125
],
[
0.037567138671875,
0.58837890625,
0.0482177734375,
0.58837890625,
0.0482177734375,
0.59130859375,
0.037567138671875,
0.59130859375
],
[
0.07843017578125,
0.5966796875,
0.468017578125,
0.59521484375,
0.468017578125,
0.60986328125,
0.07843017578125,
0.611328125
],
[
0.50341796875,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.50341796875,
0.609375
],
[
0.02532958984375,
0.60498046875,
0.058013916015625,
0.60498046875,
0.058013916015625,
0.71044921875,
0.02532958984375,
0.71044921875
],
[
0.08087158203125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.62353515625,
0.08087158203125,
0.62353515625
],
[
0.50341796875,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.080078125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.6416015625,
0.27685546875,
0.64013671875,
0.27685546875,
0.65478515625,
0.080078125,
0.65576171875
],
[
0.5009765625,
0.6416015625,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.5009765625,
0.65478515625
],
[
0.5009765625,
0.65478515625,
0.89208984375,
0.65576171875,
0.89208984375,
0.67041015625,
0.5009765625,
0.66943359375
],
[
0.50341796875,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.08251953125,
0.6875,
0.20751953125,
0.6875,
0.20751953125,
0.7021484375,
0.08251953125,
0.7021484375
],
[
0.50341796875,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.5009765625,
0.70068359375,
0.8896484375,
0.70068359375,
0.8896484375,
0.7138671875,
0.5009765625,
0.7138671875
],
[
0.098876953125,
0.71337890625,
0.468017578125,
0.71337890625,
0.468017578125,
0.72802734375,
0.098876953125,
0.72802734375
],
[
0.50244140625,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.08087158203125,
0.72802734375,
0.46728515625,
0.72802734375,
0.46728515625,
0.7412109375,
0.08087158203125,
0.7412109375
],
[
0.50341796875,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.080078125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.5009765625,
0.7470703125,
0.81396484375,
0.74560546875,
0.81396484375,
0.76025390625,
0.5009765625,
0.76123046875
],
[
0.07843017578125,
0.7568359375,
0.468017578125,
0.75830078125,
0.468017578125,
0.77294921875,
0.07843017578125,
0.771484375
],
[
0.51953125,
0.7646484375,
0.8896484375,
0.76513671875,
0.8896484375,
0.77978515625,
0.51953125,
0.77880859375
],
[
0.080078125,
0.7734375,
0.468017578125,
0.7734375,
0.468017578125,
0.787109375,
0.080078125,
0.787109375
],
[
0.5009765625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.7890625,
0.468017578125,
0.7890625,
0.468017578125,
0.80322265625,
0.080078125,
0.80322265625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.80615234375,
0.468017578125,
0.80615234375,
0.468017578125,
0.81640625,
0.08087158203125,
0.81640625
],
[
0.5048828125,
0.8125,
0.88623046875,
0.8125,
0.88623046875,
0.82275390625,
0.5048828125,
0.82275390625
],
[
0.080078125,
0.8193359375,
0.4697265625,
0.818359375,
0.4697265625,
0.83203125,
0.080078125,
0.83349609375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.07843017578125,
0.83349609375,
0.468017578125,
0.833984375,
0.468017578125,
0.8486328125,
0.07843017578125,
0.84765625
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.849609375,
0.41015625,
0.849609375,
0.41015625,
0.86279296875,
0.080078125,
0.86279296875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.87353515625,
0.226318359375,
0.875,
0.226318359375,
0.8896484375,
0.09722900390625,
0.88818359375
],
[
0.50341796875,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.10455322265625,
0.88720703125,
0.438720703125,
0.88720703125,
0.438720703125,
0.9013671875,
0.10455322265625,
0.9013671875
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.5009765625,
0.8994140625
],
[
0.48193359375,
0.92822265625,
0.490234375,
0.92822265625,
0.490234375,
0.93310546875,
0.48193359375,
0.93310546875
]
] |
[
"Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection",
"Alexandros Haliassosl.t",
"Konstantinos Vougioukas!",
"Stavros Petridis!",
"Maja Panticl,2",
"'Imperial College London",
" 2Facebook London",
"[alexandros.haliassosl4,k.vougioukas,stavros.petridis04,m.panticj@imperial.ac.uk",
"12",
"Abstract",
"Train-time",
"Test-time",
"Stage 1: Which words are spoken?",
"Unseen forgery types",
"Sn",
"",
"Although current deep learning-based face forgery de-",
"tectors achieve impressive performance in constrained sce-",
"narios, they are vulnerable to samples created by unseen",
"manipulation methods.Some recent works show improve-",
"5",
"ments in generalisation but rely on cues that are easily cor-",
"Stage2:Arethesereal orfake?",
"Corrupted videos",
"ruptedbycommonpost-processingoperationssuchascom-",
"pression. In this paper, we propose LipForensics, a detec-",
"LAD",
"tion approach capable of both generalising to novel manip-",
"ulations and withstanding various distortions. LipForensics",
"targets high-level semantic irregularities in mouth move-",
"Figure 1. Ilustration of our training and testing objectives.",
"ments, which are common in many generated videos. It con-",
"S",
"By leveraging semantically high-level, spatio-temporal represen-",
"sists in first pretraining a spatio-temporal network to per-",
"tations learned via the task of lipreading,our face forgery detector",
"form visual speech recognition (lipreading), thus learning",
"can handle both novel manipulations and common corruptions at",
"3",
"rich internal representations related to natural mouth mo-",
"test-time.",
"tion. A temporal network is subsequently finetuned on fixed",
"V",
"LS9",
"",
"mouth embeddings of real and forged data in order to detect",
"fake videos based on mouth movements without overfitting",
"With the aid of recent releases of large-scale face forgery",
"LO",
"to low-level, manipulation-specific artefacts. Extensive ex-",
"datasets [50, 42, 34, 18], it is possible to train deep con-",
"periments show that this simple approach significantly sur-",
"volutional neural networks (CNNs) to detect fake videos",
"2",
"",
"passes the state-of-the-art in terms of generalisation to un-",
"[3, 27, 69, 4, 15, 49, 46, 40]. Despite excellent accuracy",
"10Z:AIXI",
"seenmanipulationsandrobustnesstoperturbations,aswell",
"onsamplesthatareindependentandidenticallydistributed",
"as shed light on the factors responsible for its performance.",
"to the training distribution, it is known that dramatic drops",
"Code is available on GitHub.!",
"in performance may be experienced when this stringent cri-",
"terion is not met [39, 9, 47]. For example, detectors often",
"significantly underperform on novel forgery types: This un-",
"1.Introduction",
"derstandably sparks concerns, as a deployed detector is un-",
"likely to be exposed only to forgeries seen during training.",
"Recent advances in deep generative models, especially",
"Moreover, they are often sensitive to common perturbations",
"Generative Adversarial Networks (GANs) [26], have en-",
"such as compression and are,as a result,vulnerableto the",
"abled the use of off-the-shelf models that can produce ultra-",
"image processing pipelines on social networks.",
"realistic fake videos with little human effort or expertise.",
"Recent attempts to boost generalisation to novel forg-",
"Face manipulation methods in particular have raised consid-",
"eries include simple data augmentations [64], a two-branch",
"erable concerns due to their ability to alter a person's iden-",
"network that amplifies multi-band frequencies [45], an",
"tity[42,34,381,expression[57,56,48l,orlipmovements",
"autoencoder-like structure to serve as an anomaly detec",
"[55, 62, 61] to match the face in a given target video. The",
"tor [14, 19, 47], and patch-based classification to model",
"misuse of such technology can spread political propaganda,",
"local patterns [9]. However, these methods still substan-",
"defame individuals, or damage trust in journalism.",
"tially overfit to seen manipulations. A particularly effective",
"tCorresponding author.",
"method is Face X-ray [39], which proposes to predict the",
"Ihttps://github.com/ahaliassos/LipForensics",
"blending boundary between the background image and the",
"1"
] |
{
"arxiv_id": "2012.07657",
"language": "en",
"page_idx": 0
}
|
||
2012.07657-en-1
|
2012.07657-en-1.jpg
|
inserted, altered face. Although it attains impressive generalisation in cross-manipulation experiments, it relies on often imperceptible patterns which are susceptible to lowlevel post-processing operations.
It is natural to ask: Are there semantically high-level inconsistencies across manipulation methods, which are thus more robust to routine perturbations? We observe that most face video forgeries alter the mouth in some way, to match it with someone else's identity, speech, or expression. Due to the intricate motion of the mouth, current manipulation methods find it difficult to generate movements without falling into the “uncanny valley?" For example, fake mouths often do not adequately close when they pronounce certain phonemes [4]. We also notice unnatural fluctuations in the speed of movements, as well as alterations in the shape of the mouth or its interior $(e.g.$,teeth) from frame-toframe, even when there is no speech (see Appendix C.1 for examples). These irregularities provide a precious opportunity for detectors to capitalise on; yet, to the best of our knowledge, previous works do not specifically target mouth motion using spatio-temporal neural networks.
The human visual system can perceive subtle anomalies in forged oral movements as a result of extensive experience in observing real mouths move. In order to endow a network with such experience, we propose to pretrain it on a large corpus of real videos to perform the difficult task of visual speech recognition, also known as lipreading. To be capable of disambiguating similar words, the network must learn rich spatio-temporal representations related to the mouth as well as the teeth and tongue[13]. Next, we transfer the acquired knowledge to face forgery detection. Crucially, we $\hat{\text{treat the first part of the network as a frozen feature extrac-}}$ tor that outputs an embedding per frame, and only finetune the temporal convolutional network that takes these embeddings as input. This prevents the network from learning to discriminate the data based on unstable, low-level patterns that may not be generated by other forgery methods
We dub our approach LipForensics (see Figure 1). We conduct extensive experiments to compare its performance with the state-of-the-art in various challenging scenarios. We find that, in most cases, it significantly outperforms previous methods with respect to generalisation to unseen forgeries, while exhibiting remarkable robustness to common corruptions which degrade other models' performance. Further, in-distribution experiments reveal that LipForensics can effectively learn even on heavily compressed data, unlike other detectors. Finally, we validate our design choices through ablation studies, and compare with other large-scale pretraining tasks to demonstrate the superiority of lipreading for achieving generalisable and robust face forgery detection.
$^{2}$The“uncanny valley”refers to the unease experienced by humans
when observing a realistic computer-generated face.
## $2. \textbf{Related Work}$
Face forgery detection. Some earlier face forgery detection works bias the network away from learning highlevel features by constraining CNN filters [6] or using relatively shallow networks[3] to focus on mesoscopic features. Rössler $etal.[50]$ showed, however, that these are outperformed by a deep, unconstrained Xception [11] network. Some methods [5, 54] predict the manipulated region along with the label in a multi-task fashion. Another line of research relates to the observation that fully CNN-generated images may exhibit anomalous patterns in the frequency spectrum [21,20,68,49]. In [27,51], a CNN followed by an LSTM [30] are used to capture spatio-temporal features. In [63], fake frames are detected based on anomalies in the neuron behaviour of a face recognition network.
Some works exploit the correspondence between the visual and auditory modalities [4,46,12,37]. For example,[4] targets inconsistencies between the “M”,“B”, and "P" phonemes and their corresponding visemes, requiring voice-to-text transcriptions and viseme-phoneme alignment. We note that, in contrast, our method is visual-only, $i.e.$, does not take as input the audio during training or inference. A concurrent work[66] also exploits lip movements but is a biometric approach, hence requiring a reference video at inference.
Generalisation to novel manipulations. Although current detectors tend to perform well when the training and test distributions are similar, cross-dataset experiments reveal that they underperform on unseen forgeries [39,9,42,47]. To address this issue, ForensicTransfer [14] proposes an autoencoder-like structure to prevent the network from discarding frame information. Follow-up works LAE[19] and MTDS [47] additionally exploit binary forgery segmentation masks. In [45], a two-branch recurrent neural network amplifes multi-band frequencies. Wang et al.[64] apply blurring and compression augmentations, which help generalisation across fully-synthetic fake images. Chai et al.[9] instead hypothesise that local patterns generalise better and truncate image classifiers to reduce their receptive felds. We challenge this hypothesis by showing that high-level temporal inconsistencies in the mouth region can generalise very well (and lead to more robust detectors)
A particularly successful self-supervised approach is to generate fake videos on-the-fly that contain only the specific irregularities that one may want to target. For instance, FWA[41] targets the common affine face warping step in forgery pipelines. More recently, Li et al.[39] observe that many forgery algorithms depend on a blending step to realistically insert an altered face into a background image. They predict the blending boundary between the two images to achieve the current state-of-the-art generalisation
2
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50341796875,
0.09027099609375,
0.640625,
0.09027099609375,
0.640625,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.5009765625,
0.11553955078125,
0.89208984375,
0.11614990234375,
0.89208984375,
0.1307373046875,
0.5009765625,
0.1300048828125
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.1319580078125,
0.890625,
0.1319580078125,
0.890625,
0.1458740234375,
0.50244140625,
0.1458740234375
],
[
0.07843017578125,
0.1383056640625,
0.29736328125,
0.1383056640625,
0.29736328125,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.5009765625,
0.1458740234375,
0.89208984375,
0.146484375,
0.89208984375,
0.1610107421875,
0.5009765625,
0.160400390625
],
[
0.098876953125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.098876953125,
0.1673583984375
],
[
0.5009765625,
0.1610107421875,
0.89208984375,
0.1622314453125,
0.89208984375,
0.175537109375,
0.5009765625,
0.1749267578125
],
[
0.08087158203125,
0.17041015625,
0.468017578125,
0.17041015625,
0.468017578125,
0.18115234375,
0.08087158203125,
0.18115234375
],
[
0.5009765625,
0.175537109375,
0.8896484375,
0.175537109375,
0.8896484375,
0.189453125,
0.5009765625,
0.189453125
],
[
0.08087158203125,
0.1849365234375,
0.4697265625,
0.1849365234375,
0.4697265625,
0.1956787109375,
0.08087158203125,
0.1956787109375
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50341796875,
0.1932373046875,
0.890625,
0.1932373046875,
0.890625,
0.203857421875,
0.50341796875,
0.203857421875
],
[
0.50244140625,
0.20703125,
0.89208984375,
0.20703125,
0.89208984375,
0.2203369140625,
0.50244140625,
0.2203369140625
],
[
0.080078125,
0.2139892578125,
0.4697265625,
0.2139892578125,
0.4697265625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.2215576171875,
0.89208984375,
0.2215576171875,
0.89208984375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50341796875,
0.238037109375,
0.890625,
0.238037109375,
0.890625,
0.248779296875,
0.50341796875,
0.248779296875
],
[
0.08087158203125,
0.24560546875,
0.465576171875,
0.24560546875,
0.465576171875,
0.255615234375,
0.08087158203125,
0.255615234375
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.25390625,
0.8896484375,
0.25390625,
0.8896484375,
0.264404296875,
0.50341796875,
0.264404296875
],
[
0.50244140625,
0.267578125,
0.890625,
0.267578125,
0.890625,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.080078125,
0.274658203125,
0.468017578125,
0.274658203125,
0.468017578125,
0.28857421875,
0.080078125,
0.28857421875
],
[
0.50244140625,
0.2822265625,
0.890625,
0.2822265625,
0.890625,
0.29541015625,
0.50244140625,
0.29541015625
],
[
0.080078125,
0.290283203125,
0.4697265625,
0.290283203125,
0.4697265625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.50244140625,
0.29736328125,
0.89208984375,
0.29736328125,
0.89208984375,
0.311279296875,
0.50244140625,
0.311279296875
],
[
0.07843017578125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.313232421875,
0.81884765625,
0.313232421875,
0.81884765625,
0.326904296875,
0.50244140625,
0.326904296875
],
[
0.08087158203125,
0.3212890625,
0.46728515625,
0.3212890625,
0.46728515625,
0.33203125,
0.08087158203125,
0.33203125
],
[
0.080078125,
0.335205078125,
0.4697265625,
0.335205078125,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.521484375,
0.330810546875,
0.890625,
0.330810546875,
0.890625,
0.34521484375,
0.521484375,
0.34521484375
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.5009765625,
0.3466796875,
0.89208984375,
0.34521484375,
0.89208984375,
0.35986328125,
0.5009765625,
0.361083984375
],
[
0.080078125,
0.365478515625,
0.4697265625,
0.365478515625,
0.4697265625,
0.37890625,
0.080078125,
0.37890625
],
[
0.50244140625,
0.3623046875,
0.89208984375,
0.35986328125,
0.89208984375,
0.373779296875,
0.50244140625,
0.375732421875
],
[
0.07843017578125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.07843017578125,
0.393310546875
],
[
0.5009765625,
0.375,
0.89208984375,
0.376953125,
0.89208984375,
0.391357421875,
0.5009765625,
0.389404296875
],
[
0.07843017578125,
0.395751953125,
0.3857421875,
0.395751953125,
0.3857421875,
0.4091796875,
0.07843017578125,
0.4091796875
],
[
0.50341796875,
0.393310546875,
0.890625,
0.393310546875,
0.890625,
0.404052734375,
0.50341796875,
0.404052734375
],
[
0.09722900390625,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.09722900390625,
0.4248046875
],
[
0.50244140625,
0.4072265625,
0.89208984375,
0.4072265625,
0.89208984375,
0.42041015625,
0.50244140625,
0.42041015625
],
[
0.07843017578125,
0.425537109375,
0.4697265625,
0.425537109375,
0.4697265625,
0.439453125,
0.07843017578125,
0.439453125
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.08087158203125,
0.442626953125,
0.468017578125,
0.442626953125,
0.468017578125,
0.45263671875,
0.08087158203125,
0.45263671875
],
[
0.50244140625,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.080078125,
0.45703125,
0.468017578125,
0.45703125,
0.468017578125,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50244140625,
0.45263671875,
0.890625,
0.45263671875,
0.890625,
0.466552734375,
0.50244140625,
0.466552734375
],
[
0.07843017578125,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.48486328125,
0.07843017578125,
0.48486328125
],
[
0.50244140625,
0.46728515625,
0.6611328125,
0.46728515625,
0.6611328125,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.080078125,
0.486083984375,
0.4697265625,
0.486083984375,
0.4697265625,
0.5,
0.080078125,
0.5
],
[
0.50341796875,
0.4892578125,
0.8486328125,
0.4892578125,
0.8486328125,
0.50390625,
0.50341796875,
0.50390625
],
[
0.8486328125,
0.4912109375,
0.890625,
0.4912109375,
0.890625,
0.50244140625,
0.8486328125,
0.50244140625
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.5009765625,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.5009765625,
0.51953125
],
[
0.080078125,
0.517578125,
0.4697265625,
0.517578125,
0.4697265625,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.5009765625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.5009765625,
0.5341796875
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5322265625,
0.468017578125,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.50341796875,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.080078125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.50244140625,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.50244140625,
0.564453125
],
[
0.080078125,
0.5625,
0.4697265625,
0.5625,
0.4697265625,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.50341796875,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.50341796875,
0.580078125
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.5,
0.580078125,
0.89208984375,
0.58056640625,
0.89208984375,
0.59521484375,
0.5,
0.5947265625
],
[
0.08087158203125,
0.59326171875,
0.46728515625,
0.59326171875,
0.46728515625,
0.60400390625,
0.08087158203125,
0.60400390625
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.60986328125,
0.50244140625,
0.60986328125
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50244140625,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.50244140625,
0.625
],
[
0.080078125,
0.62255859375,
0.4697265625,
0.62255859375,
0.4697265625,
0.63623046875,
0.080078125,
0.63623046875
],
[
0.50244140625,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.6396484375,
0.50244140625,
0.6396484375
],
[
0.080078125,
0.63818359375,
0.4248046875,
0.63818359375,
0.4248046875,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.50244140625,
0.6396484375,
0.89208984375,
0.64013671875,
0.89208984375,
0.65478515625,
0.50244140625,
0.65380859375
],
[
0.1004638671875,
0.65283203125,
0.468017578125,
0.65283203125,
0.468017578125,
0.66748046875,
0.1004638671875,
0.66748046875
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.07843017578125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.07843017578125,
0.681640625
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.07843017578125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.07843017578125,
0.69775390625
],
[
0.5009765625,
0.68603515625,
0.890625,
0.68603515625,
0.890625,
0.69970703125,
0.5009765625,
0.69970703125
],
[
0.080078125,
0.69873046875,
0.4697265625,
0.69873046875,
0.4697265625,
0.7119140625,
0.080078125,
0.7119140625
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7021484375,
0.890625,
0.71533203125,
0.50244140625,
0.71533203125
],
[
0.080078125,
0.7138671875,
0.4697265625,
0.7138671875,
0.4697265625,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.72900390625,
0.50341796875,
0.72900390625
],
[
0.080078125,
0.72900390625,
0.4697265625,
0.72900390625,
0.4697265625,
0.7421875,
0.080078125,
0.7421875
],
[
0.5009765625,
0.73193359375,
0.89208984375,
0.73193359375,
0.89208984375,
0.74560546875,
0.5009765625,
0.74560546875
],
[
0.080078125,
0.74365234375,
0.46728515625,
0.74365234375,
0.46728515625,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.50244140625,
0.74755859375,
0.61767578125,
0.74755859375,
0.61767578125,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.76025390625,
0.46728515625,
0.76025390625,
0.46728515625,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.521484375,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.521484375,
0.77978515625
],
[
0.07843017578125,
0.7734375,
0.468017578125,
0.7734375,
0.468017578125,
0.787109375,
0.07843017578125,
0.787109375
],
[
0.50244140625,
0.7822265625,
0.89208984375,
0.7822265625,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.7890625,
0.468017578125,
0.7890625,
0.468017578125,
0.80322265625,
0.080078125,
0.80322265625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.80615234375,
0.468017578125,
0.80615234375,
0.468017578125,
0.81640625,
0.08087158203125,
0.81640625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8193359375,
0.468017578125,
0.8193359375,
0.468017578125,
0.83349609375,
0.080078125,
0.83349609375
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.83544921875,
0.4697265625,
0.83544921875,
0.4697265625,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.849609375,
0.1993408203125,
0.849609375,
0.1993408203125,
0.86279296875,
0.080078125,
0.86279296875
],
[
0.50341796875,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.1004638671875,
0.87353515625,
0.471435546875,
0.875,
0.471435546875,
0.8896484375,
0.1004638671875,
0.88818359375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.35546875,
0.88720703125,
0.35546875,
0.9013671875,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.47802734375,
0.92236328125,
0.494384765625,
0.92236328125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"inserted, altered face. Although it attains impressive gen-",
"2. Related Work",
"eralisation in cross-manipulation experiments.it relies on",
"Face forgery detection. Some earlier face forgery detec-",
"often imperceptible patterns which are susceptible to low-",
"tion works bias the network away from learning high-",
"level post-processing operations.",
"level features by constraining CNN filters [6] or using rela-",
"It is natural to ask: Are there semantically high-level in-",
"tively shallow networks [3]tofocuson mesoscopic features.",
"consistencies across manipulation methods,which are thus",
"Rossler et al. [50] showed, however, that these are outper-",
"more robusttoroutine perturbations? Weobservethatmost",
"face video forgeries alter the mouth in some way, to match",
"formed by a deep, unconstrained Xception Jl1l network.",
"Some methods [5, 54] predict the manipulated region along",
"it with someone else's identity, speech, or expression. Due",
"with the label in a multi-task fashion. Another line of re-",
"to the intricate motion of the mouth, current manipula-",
"searchrelatestothe observation thatfullyCNN-generated",
"tion methodsfinditdifficultto generatemovements with-",
"out falling into the“uncanny valley?\"For example,fake",
"imagesmay exhibit anomalous patterns in thefrequency",
"spectrum [21, 20, 68, 49]. In [27, 51], a CNN followed by",
"mouths often do not adequately close when they pronounce",
"an LSTM [30] are used to capture spatio-temporal features.",
"certain phonemes [4]. We also notice unnatural fluctuations",
"In [63], fake frames are detected based on anomalies in the",
"in the speed of movements, as well as alterations in the",
"neuron behaviour of a face recognition network.",
"shapeofthemouthorits interior(e.g.teeth)fromframe-to-",
"frame, even when there is no speech (see Appendix C.I for",
"Some works exploit the correspondence between the vi-",
"examples). These irregularities provide a precious oppor-",
"sual and auditory modalities [4, 46, 12, 37]. For exam-",
"tunity for detectors to capitalise on; yet, to the best of our",
"ple,[4] targets inconsistencies between the\"M\",\"B\",and",
"knowledge, previous works do not specifically target mouth",
"\"p\" phonemes and their corresponding visemes, requir-",
"motion using spatio-temporal neural networks.",
"ing voice-to-text transcriptions and viseme-phoneme align-",
"The human visual system can perceive subtle anomalies",
"ment. We note that, in contrast, our method is visual-only,",
"in forged oral movements as a result of extensive experience",
"i.e., does not take as input the audio during training or in-",
"inobservingrealmouthsmove.Inordertoendowanetwork",
"ference. A concurrent work [66] also exploits lip move-",
"with such experience, we propose to pretrain it on a large",
"ments but is a biometric approach, hence requiring a refer-",
"corpus of real videos to perform the difficult task of visual",
"ence video at inference.",
"speech recognition, also known as lipreading. To be capa-",
"Generalisation to novel manipulations. Although",
"cur-",
"ble of disambiguating similar words, the network must learn",
"rent detectors tend to perform well when the training",
"rich spatio-temporal representations related to the mouth as",
"and test distributions are similar, cross-dataset experi-",
"well as the teeth and tongue [13]. Next, we transfer the ac-",
"mentsreveal that they underperform on unseen forgeries",
"quired knowledge to face forgery detection. Crucially, we",
"[39, 9, 42, 47]. To address this issue, ForensicTransfer",
"treat the first part of the network as a frozen feature extrac-",
"[14] proposes an autoencoder-like structure to prevent the",
"tor that outputs an embedding per frame, and only finetune",
"network from discarding frame information.Follow-up",
"thetemporal convolutional networkthattakestheseembed-",
"works LAE [19] and MTDS [47] additionally exploit",
"dings as input. This prevents the network from learning to",
"binary forgery segmentation masks. In [45], a two-branch",
"discriminate the data based on unstable, low-level patterns",
"recurrent neural network amplifies multi-band frequencies.",
"that may not be generated by other forgery methods.",
"Wang et al. [64] apply blurring and compression augmen-",
"We dub our approach LipForensics (see Figure 1). We",
"tations, which help generalisation across fully-synthetic",
"conduct extensive experiments to compare its performance",
"fake images. Chai et al. [9] instead hypothesise that local",
"with the state-of-the-art in various challenging scenarios.",
"patterns generalise better and truncate image classifiers to",
"We find that, in most cases, it significantly outperforms",
"reduce their receptive fields. We challenge this hypothesis",
"previous methods with respect to generalisation to unseen",
"by showing thathigh-leveltemporal inconsistenciesin the",
"forgeries, while exhibiting remarkable robustness to com-",
"mouth region can generalise very well (and lead to more",
"mon corruptions which degrade other models'performance.",
"robust detectors).",
"Further,in-distribution experiments reveal that LipForen-",
"A particularly successful self-supervised approach is to",
"sics can effectively learn even on heavily compressed data,",
"generate fake videos on-the-fly that contain only the spe-",
"unlike other detectors.Finally, we validate our design",
"cific irregularities that one may want to target. For instance,",
"choices through ablation studies,and comparewithother",
"FWA [41] targets the common affine face warping step in",
"large-scale pretraining tasks to demonstrate the superior-",
"forgery pipelines. More recently, Li et al. [39] observe that",
"ity of lipreading for achieving generalisable and robust face",
"many forgery algorithms depend on a blending step to re-",
"forgery detection.",
"alistically insert an altered face into a background image.",
"2The “uncanny valley\" refers to the unease experienced by humans",
"They predict the blending boundary between the two im-",
"when observing a realistic computer-generated face.",
"ages to achieve the current state-of-the-art generalisation",
"2"
] |
{
"arxiv_id": "2012.07657",
"language": "en",
"page_idx": 1
}
|
||
2012.07657-en-3
|
2012.07657-en-3.jpg
|
net is a multi-scale temporal convolutional network (MSTCN) [44], combining short- and long-term temporal information at every layer by concatenating outputs of multiple branches, each with a different temporal receptive field. After a temporal global average pooling layer, a task-specific linear classifier outputs the estimated class probabilities. Architecture details are in Appendix A.1.
Lipreading pretraining. The model is pretrained on Lipreading in the Wild (LRW) [13], a dataset containing over 500,000 utterances spanning hundreds of speakers in various poses. It is trained using the approach proposed in [43], which employs born-again distillation [22]. We use the student classifier in the third generation of teacherstudent training. Unless stated otherwise, we use the publicly available, pretrained model found here$^3.$
Preprocessing. The faces are detected and then aligned to the mean face, after which a scaled $96\times96$ region around the mouth is cropped and transformed to grayscale. Our clips comprise 25 frames; thus, following random cropping, our input tensor is of size $25\times88\times88\times1.$ Note that this size corresponds to a similar number of entries as the standard RGB frame input to many of the forgery detectors in the literature (commonly of size $1\times256\times256\times3)[39].$ More preprocessing details are in Appendix A.3. We consider the effect of using the full face as input rather than mouth crops in Appendix B
Training the forgery detector. We use a batch size of 32 and Adam [36] optimisation with a learning rate of $2\times10^{-4}.$ To address any data imbalance, we oversample the minority class. We terminate training when there is negligible improvement to the validation loss for 10 epochs. As data augmentation, we randomly crop the clips with size 88×88 and horizontally flip with probability 0.5.
## $4. \textbf{Experiments}$
4.1. Setup
Datasets. As in [39, 45], we mostly use FaceForensics+ (FF++)[50] as our training dataset due to its forgery diversity. It contains 1.8 million manipulated frames and 4,000 fake videos generated using two face swapping algorithms, DeepFakes [1] and FaceSwap [2], and two face reenactment methods, Face2Face [57] and NeuralTextures [56]. As recommended in [50], we only use the first 270 frames for each training video, and the first 110 frames for each validation/testing video. Other datasets in our experiments include DeeperForensics [34] and FaceShifter [38], each featuring an improved face swapping algorithm applied to the real videos from FF++. Further, we use the test set of the
$^3$https://github.com/mpc001/Lipreading\_using\_
Temporal\_Convolutional\_Networks
face swapping Celeb-DF-v2 [42] dataset. Finally, we use 3,215 test set videos from the DeepFake Detection Challenge (DFDC) [18], where subjects were filmed in extreme conditions, such as large poses and low lighting. More information is in Appendix A.2.
Metrics. We report results using accuracy (following [50]) and/or area under the receiver operating characteristic curve (AUC; following [42, 39]). Since most existing models use a single frame as input, we compute video-level measures for fair comparison: we average the model predictions (each prediction corresponding to a frame or a video clip) across the entire video, as in [45]. As a result, all models'video predictions are based on the same number offrames.
4.2. Generalisation to unseen manipulations
Given the rapid advancements in forgery generation, we desire a detector that can correctly classify samples from novel manipulation methods after it is trained on a set of known forgeries. We simulate this scenario below.
Models for comparison. For comparison on our experiments using video-level metrics, we train various state-ofthe-art detectors designed to improve cross-dataset generalisation as well as some other popular baselines, including: (1) Face X-ray [39]: we train an HRNet-W48 [53] both with constructed blended images and fake samples from the considered datasets.(2) CNN-aug [64]: we employ a ResNet-50 with JPEG compression and Gaussian blurring augmentations, both with probability of 0.1. $( 3) \textbf{Patch- }$ based[9]: we train an Xception classifier truncated after block 3 and average the predictions across the patches. We align the frames to remove rotation variation. (4) Xception [50]: we train the popular Xception baseline.(5) $\dot{\mathbf{CNN-}}$ GRU[51]: we train a DenseNet-161 [32] followed by a GRU[10], to compare with a temporal model. All these models take as input a single RGB frame of the full face, except for CNN-GRU, which takes a 5-frame RGB clip. All but Patch-based use ImageNet pretrained weights. See Appendix A.4 for more details. We also evaluate the effects of (1) training our spatio-temporal detector from random initialisation (“Scratch”),(2) finetuning the whole network after pretraining on LRW (“Ft whole”), and (3) finetuning only the temporal net (“LipForensics”), all of which are trained on grayscale mouth crops.
Cross-manipulation generalisation. To directly assess generalisation to unseen manipulations, without introducing confounders such as variations in pose or illumination, we experiment here with fake data that were created from the same source videos. Specifically, we test on each of the four methods in FF++ after we train on the remaining three. Rather than using raw videos (as in [39]), we use the high quality (HQ) subset of the dataset (as in $e.g.,[45,9,47])$,
4
|
[
[
0.07843017578125,
0.091552734375,
0.468017578125,
0.09027099609375,
0.468017578125,
0.10479736328125,
0.07843017578125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.080078125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.5009765625,
0.15283203125,
0.70166015625,
0.1514892578125,
0.70166015625,
0.1666259765625,
0.5009765625,
0.1673583984375
],
[
0.07843017578125,
0.1666259765625,
0.46728515625,
0.1673583984375,
0.46728515625,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.5009765625,
0.1767578125,
0.890625,
0.177978515625,
0.890625,
0.1925048828125,
0.5009765625,
0.1912841796875
],
[
0.080078125,
0.18310546875,
0.351318359375,
0.18310546875,
0.351318359375,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.1932373046875,
0.890625,
0.1932373046875,
0.890625,
0.20703125,
0.50244140625,
0.20703125
],
[
0.08087158203125,
0.208984375,
0.44287109375,
0.208984375,
0.44287109375,
0.222900390625,
0.08087158203125,
0.222900390625
],
[
0.435546875,
0.210205078125,
0.468017578125,
0.210205078125,
0.468017578125,
0.2215576171875,
0.435546875,
0.2215576171875
],
[
0.50244140625,
0.208984375,
0.89208984375,
0.208984375,
0.89208984375,
0.222900390625,
0.50244140625,
0.222900390625
],
[
0.07843017578125,
0.222900390625,
0.468017578125,
0.2235107421875,
0.468017578125,
0.238037109375,
0.07843017578125,
0.2374267578125
],
[
0.50244140625,
0.2235107421875,
0.89208984375,
0.2235107421875,
0.89208984375,
0.2374267578125,
0.50244140625,
0.2374267578125
],
[
0.07843017578125,
0.2392578125,
0.468017578125,
0.2392578125,
0.468017578125,
0.25244140625,
0.07843017578125,
0.25244140625
],
[
0.50244140625,
0.2392578125,
0.89208984375,
0.2392578125,
0.89208984375,
0.25244140625,
0.50244140625,
0.25244140625
],
[
0.08087158203125,
0.255615234375,
0.468017578125,
0.255615234375,
0.468017578125,
0.266357421875,
0.08087158203125,
0.266357421875
],
[
0.50244140625,
0.255126953125,
0.8935546875,
0.255126953125,
0.8935546875,
0.268310546875,
0.50244140625,
0.268310546875
],
[
0.080078125,
0.270751953125,
0.468017578125,
0.270751953125,
0.468017578125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.50341796875,
0.270751953125,
0.890625,
0.270751953125,
0.890625,
0.281005859375,
0.50341796875,
0.281005859375
],
[
0.08087158203125,
0.285888671875,
0.465576171875,
0.285888671875,
0.465576171875,
0.296630859375,
0.08087158203125,
0.296630859375
],
[
0.5009765625,
0.28271484375,
0.84716796875,
0.28271484375,
0.84716796875,
0.296630859375,
0.5009765625,
0.296630859375
],
[
0.080078125,
0.298583984375,
0.465576171875,
0.298583984375,
0.465576171875,
0.3125,
0.080078125,
0.3125
],
[
0.080078125,
0.314453125,
0.378173828125,
0.314453125,
0.378173828125,
0.328857421875,
0.080078125,
0.328857421875
],
[
0.50341796875,
0.31005859375,
0.8388671875,
0.311279296875,
0.8388671875,
0.32568359375,
0.50341796875,
0.324462890625
],
[
0.521484375,
0.333251953125,
0.89208984375,
0.333984375,
0.89208984375,
0.348388671875,
0.521484375,
0.347900390625
],
[
0.08087158203125,
0.34033203125,
0.468017578125,
0.34033203125,
0.468017578125,
0.354248046875,
0.08087158203125,
0.354248046875
],
[
0.5009765625,
0.348388671875,
0.89208984375,
0.349853515625,
0.89208984375,
0.3642578125,
0.5009765625,
0.363037109375
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354736328125,
0.468017578125,
0.36865234375,
0.080078125,
0.36865234375
],
[
0.5009765625,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.37744140625,
0.5009765625,
0.37744140625
],
[
0.080078125,
0.371826171875,
0.4697265625,
0.371826171875,
0.4697265625,
0.385009765625,
0.080078125,
0.385009765625
],
[
0.080078125,
0.386474609375,
0.468017578125,
0.386474609375,
0.468017578125,
0.399658203125,
0.080078125,
0.399658203125
],
[
0.50244140625,
0.380615234375,
0.8349609375,
0.380615234375,
0.8349609375,
0.39453125,
0.50244140625,
0.39453125
],
[
0.080078125,
0.401611328125,
0.4697265625,
0.401611328125,
0.4697265625,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.50244140625,
0.40771484375,
0.890625,
0.40771484375,
0.890625,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.080078125,
0.416748046875,
0.468017578125,
0.416748046875,
0.468017578125,
0.429931640625,
0.080078125,
0.429931640625
],
[
0.50244140625,
0.423583984375,
0.890625,
0.423583984375,
0.890625,
0.4375,
0.50244140625,
0.4375
],
[
0.080078125,
0.43115234375,
0.4697265625,
0.43115234375,
0.4697265625,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.5009765625,
0.4375,
0.89208984375,
0.438232421875,
0.89208984375,
0.45263671875,
0.5009765625,
0.451904296875
],
[
0.08087158203125,
0.447509765625,
0.468017578125,
0.447509765625,
0.468017578125,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.50244140625,
0.453857421875,
0.89208984375,
0.453857421875,
0.89208984375,
0.46728515625,
0.50244140625,
0.46728515625
],
[
0.080078125,
0.4609375,
0.468017578125,
0.4609375,
0.468017578125,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.50341796875,
0.468505859375,
0.89208984375,
0.468505859375,
0.89208984375,
0.481689453125,
0.50341796875,
0.481689453125
],
[
0.080078125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.50341796875,
0.484130859375,
0.89208984375,
0.484130859375,
0.89208984375,
0.49755859375,
0.50341796875,
0.49755859375
],
[
0.08087158203125,
0.49365234375,
0.181396484375,
0.49365234375,
0.181396484375,
0.5048828125,
0.08087158203125,
0.5048828125
],
[
0.50341796875,
0.498779296875,
0.89208984375,
0.498779296875,
0.89208984375,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.5146484375,
0.89208984375,
0.5146484375,
0.89208984375,
0.52783203125,
0.50244140625,
0.52783203125
],
[
0.07843017578125,
0.5322265625,
0.468017578125,
0.53076171875,
0.468017578125,
0.54541015625,
0.07843017578125,
0.546875
],
[
0.50244140625,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54248046875,
0.50244140625,
0.54248046875
],
[
0.08087158203125,
0.54736328125,
0.4697265625,
0.54736328125,
0.4697265625,
0.56103515625,
0.08087158203125,
0.56103515625
],
[
0.50244140625,
0.54345703125,
0.89208984375,
0.54345703125,
0.89208984375,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.5009765625,
0.55810546875,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.5009765625,
0.5712890625
],
[
0.080078125,
0.57763671875,
0.4697265625,
0.57763671875,
0.4697265625,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.50341796875,
0.57568359375,
0.890625,
0.57568359375,
0.890625,
0.5859375,
0.50341796875,
0.5859375
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50341796875,
0.5888671875,
0.89208984375,
0.5888671875,
0.89208984375,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.07843017578125,
0.60693359375,
0.351318359375,
0.60693359375,
0.351318359375,
0.62060546875,
0.07843017578125,
0.62060546875
],
[
0.50244140625,
0.60302734375,
0.8935546875,
0.60498046875,
0.8935546875,
0.61865234375,
0.50244140625,
0.6162109375
],
[
0.50341796875,
0.61865234375,
0.8935546875,
0.61865234375,
0.8935546875,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.07843017578125,
0.63525390625,
0.2099609375,
0.63720703125,
0.2099609375,
0.65478515625,
0.07843017578125,
0.65283203125
],
[
0.50341796875,
0.63525390625,
0.890625,
0.63525390625,
0.890625,
0.6484375,
0.50341796875,
0.6484375
],
[
0.50341796875,
0.6494140625,
0.890625,
0.6494140625,
0.890625,
0.66357421875,
0.50341796875,
0.66357421875
],
[
0.080078125,
0.6591796875,
0.158447265625,
0.662109375,
0.1568603515625,
0.67822265625,
0.07843017578125,
0.6748046875
],
[
0.5009765625,
0.6640625,
0.890625,
0.66552734375,
0.890625,
0.6787109375,
0.5009765625,
0.67822265625
],
[
0.080078125,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69873046875,
0.080078125,
0.69873046875
],
[
0.50341796875,
0.6796875,
0.89208984375,
0.6796875,
0.89208984375,
0.693359375,
0.50341796875,
0.693359375
],
[
0.080078125,
0.69970703125,
0.46728515625,
0.70068359375,
0.46728515625,
0.71533203125,
0.080078125,
0.7138671875
],
[
0.50341796875,
0.6962890625,
0.890625,
0.6962890625,
0.890625,
0.70703125,
0.50341796875,
0.70703125
],
[
0.07843017578125,
0.71533203125,
0.468017578125,
0.7138671875,
0.468017578125,
0.72900390625,
0.07843017578125,
0.72998046875
],
[
0.50244140625,
0.71044921875,
0.89208984375,
0.71044921875,
0.89208984375,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.50244140625,
0.724609375,
0.89208984375,
0.724609375,
0.89208984375,
0.73779296875,
0.50244140625,
0.73779296875
],
[
0.080078125,
0.73095703125,
0.46728515625,
0.73095703125,
0.46728515625,
0.744140625,
0.080078125,
0.744140625
],
[
0.50244140625,
0.74072265625,
0.89208984375,
0.74072265625,
0.89208984375,
0.75390625,
0.50244140625,
0.75390625
],
[
0.080078125,
0.7470703125,
0.468017578125,
0.7470703125,
0.468017578125,
0.76025390625,
0.080078125,
0.76025390625
],
[
0.07843017578125,
0.76025390625,
0.46728515625,
0.76123046875,
0.46728515625,
0.77587890625,
0.07843017578125,
0.7744140625
],
[
0.50244140625,
0.755859375,
0.7255859375,
0.755859375,
0.7255859375,
0.76953125,
0.50244140625,
0.76953125
],
[
0.08087158203125,
0.7763671875,
0.468017578125,
0.7763671875,
0.468017578125,
0.79052734375,
0.08087158203125,
0.79052734375
],
[
0.50341796875,
0.78076171875,
0.8935546875,
0.78076171875,
0.8935546875,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.080078125,
0.79150390625,
0.468017578125,
0.79150390625,
0.468017578125,
0.8046875,
0.080078125,
0.8046875
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.806640625,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.080078125,
0.82080078125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8212890625,
0.4697265625,
0.8212890625,
0.4697265625,
0.83544921875,
0.080078125,
0.83544921875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.8369140625,
0.468017578125,
0.8369140625,
0.468017578125,
0.85107421875,
0.080078125,
0.85107421875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.8515625,
0.4697265625,
0.8515625,
0.4697265625,
0.86572265625,
0.080078125,
0.86572265625
],
[
0.5009765625,
0.85498046875,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5009765625,
0.86865234375
],
[
0.1004638671875,
0.87353515625,
0.468017578125,
0.87548828125,
0.468017578125,
0.89013671875,
0.1004638671875,
0.88818359375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88818359375,
0.328369140625,
0.88818359375,
0.328369140625,
0.9013671875,
0.080078125,
0.9013671875
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"net is a multi-scale temporal convolutional network (Ms-",
"face swapping Celeb-DF-v2 [42] dataset. Finally, we use",
"TCN) [44], combining short- and long-term temporal infor-",
"3,215 test set videos from the DeepFake Detection Chal-",
"mation at every layer by concatenating outputs of multiple",
"lenge (DFDC) [18], where subjects were filmed in extreme",
"branches, each with a different temporal receptive field. Af-",
"conditions, such as large poses and low lighting. More in-",
"ter a temporal global average pooling layer, a task-specific",
"formation is in Appendix A.2.",
"linear classifier outputs the estimated class probabilities.",
"Metrics. We report results using accuracy (following [50])",
"Architecture details are in Appendix A.1.",
"and/or area under the receiver operating characteristic curve",
"Lipreading pretraining. The model is pretrained",
"1on",
"(AUC; following [42, 39]). Since most existing models use",
"Lipreading in the Wild (LRW) [13], a dataset containing",
"a single frame as input, we compute video-level measures",
"over 500,000 utterances spanning hundreds of speakers in",
"forfair comparison: we average the model predictions (each",
"various poses.It is trained using the approach proposed",
"prediction corresponding to a frame or a video clip) across",
"in[43], which employsborn-againdistillation [22l.We",
"the entire video, as in [45l. As aresult, all models'video",
"usethe student classifierin the third generation of teacher",
"predictions are based on the same number of frames.",
"student training. Unless stated otherwise, we use the pub-",
"licly available, pretrained model found here3.",
"4.2. Generalisation to unseen manipulations",
"Given the rapid advancements in forgery generation, we",
"Preprocessing. The faces are detected and then aligned to",
"desire a detector that can correctly classify samples from",
"the mean face, after which a scaled 96 × 96 region around",
"novel manipulation methods after it is trained on a set of",
"the mouth is cropped and transformed to grayscale. Our",
"clips comprise 25 frames; thus, following random cropping,",
"known forgeries. We simulate this scenario below.",
"our input tensor is of size 25 × 88× 88 ×1. Note that this size",
"Models for comparison. For comparison on our experi-",
"corresponds to a similar number of entries as the standard",
"ments using video-level metrics, we train various state-of-",
"RGB frame input to many of the forgery detectors in the",
"the-art detectors designed to improve cross-dataset general-",
"literature(commonlyofsize1×256×256×3)[39l.More",
"isation as well as some other popular baselines, including:",
"preprocessing details are in Appendix A.3. We consider the",
"() Face X-ray [39]: we train an HRNet-W48 [53] both",
"effect of using the full face as input rather than mouth crops",
"with constructed blended images and fake samples from",
"in Appendix B.",
"the considered datasets. (2) CNN-aug [64]: we employ a",
"Training the forgery detector. We use a batch size of 32",
"ResNet-50 with JPEG compression and Gaussian blurring",
"and Adam[36] optimisation with a learning rate of 2×10-4.",
"augmentations, both with probability of 0.1. (3) Patch-",
"To address any data imbalance, we oversample the minor-",
"based [9]: we train an Xception classifier truncated after",
"ity class. We terminate training when there is negligible",
"block 3 and average the predictions across the patches. We",
"improvement to the validation loss for 10 epochs. As data",
"aligntheframestoremoverotationvariation.(4)Xception",
"augmentation, we randomly crop the clips with size 88x 88",
"[50]: we train the popular Xception baseline. (5) CNN-",
"and horizontally flip with probability 0.5.",
"GRU[51]:wetrain aDenseNet-161 [32] followed by a",
"GRU [10], to compare with a temporal model. All these",
"4. Experiments",
"models take as input a single RGB frame of the full face,",
"except for CNN-GRU, which takes a 5-frame RGB clip. All",
"4.1.Setup",
"but Patch-based use ImageNet pretrained weights. See Ap-",
"Datasets. As in [39, 45], we mostly use FaceForensics++",
"pendix A.4 for more details. We also evaluate the effects",
"(FF++) [50] as our training dataset due to its forgery diver-",
"of (1)training our spatio-temporal detectorfrom random",
"sity. It contains 1.8 million manipulated frames and 4,000",
"initialisation (\"Scratch\"), (2) finetuning the whole network",
"after pretraining on LRW (Ft whole\"), and (3) finetuning",
"fake videos generated using two face swapping algorithms,",
"only the temporal net (LipForensics\"), all of which are",
"DeepFakes [1] and FaceSwap [2], and two face reenact-",
"ment methods, Face2Face [57] and NeuralTextures [56]",
"trained on grayscale mouth crops.",
"As recommended in [50], we only use the first 270 frames",
"Cross-manipulation generalisation. To directly assess",
"for each training video, and the first 110 frames for each",
"generalisation to unseen manipulations, without introduc-",
"validation/testing video. Other datasets in our experiments",
"ing confounders such as variations in pose or illumination,",
"include DeeperForensics [34] and FaceShifter [38], each",
"we experiment here with fake data that were created from",
"featuring an improved face swapping algorithm applied to",
"the same source videos. Specifically, we test on each of the",
"the real videos from FF++. Further, we use the test set of the",
"four methods in FF++ after we train on the remaining three.",
"3https://github.com/mpcoo1/Lipreading_using",
"Rather than using raw videos (as in [39]), we use the high",
"Temporal_convolutional_Networks",
"quality (HQ) subset of the dataset (as in e.g., [45, 9, 47]),",
"4"
] |
{
"arxiv_id": "2012.07657",
"language": "en",
"page_idx": 3
}
|
||
2208.04378-en-0
|
2208.04378-en-0.jpg
|
# Contrast-Phys: Unsupervised Video-based Remote Physiological Measurement via $\operatorname{Spatiotemporal}$ Contrast
Zhaodong Sun and Xiaobai Li
Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland
$\{$zhaodong.sun, xiaobai.li$\}$@oulu.fi
$\mathbf{Abstract.~Video-based~remote~physiological~measurement~utilizes~face}$ videos to measure the blood volume change signal, which is also called remote photoplethysmography (rPPG). Supervised methods for rPPG measurements achieve state-of-the-art performance. However, supervised rPPG methods require face videos and ground truth physiological signals for model training. In this paper, we propose an unsupervised rPPG measurement method that does not require ground truth signals for training. We use a 3DCNN model to generate multiple rPPG signals from each video in different spatiotemporal locations and train the model with a contrastive loss where rPPG signals from the same video are pulled together while those from different videos are pushed away. We test on five public datasets, including RGB videos and NIR videos. The results show that our method outperforms the previous unsupervised baseline and achieves accuracies very close to the current best supervised rPPG methods on all five datasets. Furthermore, we also demonstrate that our approach can run at a much faster speed and is more robust to noises than the previous unsupervised baseline. Our code is available at https://github. com/zhaodongsun/contrast-phys.
Keywords: Remote Photoplethysmography, Face Video, Unsupervised
Learning, Contrastive Learning
## 1 Introduction
Traditional physiological measurement requires skin-contact sensors to measure physiological signals such as contact photoplethysmography (PPG) and electrocardiography (ECG). Some physiological parameters like heart rate (HR), respiration frequency (RF), and heart rate variability (HRV) can be derived from PPG signals for healthcare [42,55] and emotion analysis [57,29,40]. However, the skin-contact physiological measurement requires specific biomedical equipment like pulse oximeters, and contact sensors may cause discomfort and skin irritation. Remote physiological measurement uses a camera to record face videos for measuring remote photoplethysmography (rPPG). The weak color change im faces can be captured by cameras to obtain the rPPG signal from which several
|
[
[
0.259033203125,
0.1458740234375,
0.74658203125,
0.1458740234375,
0.74658203125,
0.1634521484375,
0.259033203125,
0.1634521484375
],
[
0.275390625,
0.1685791015625,
0.7294921875,
0.1685791015625,
0.7294921875,
0.18310546875,
0.275390625,
0.18310546875
],
[
0.358642578125,
0.1900634765625,
0.646484375,
0.189453125,
0.646484375,
0.20703125,
0.358642578125,
0.2083740234375
],
[
0.39453125,
0.238037109375,
0.61279296875,
0.238037109375,
0.61279296875,
0.251953125,
0.39453125,
0.251953125
],
[
0.0236968994140625,
0.264404296875,
0.05963134765625,
0.264404296875,
0.05963134765625,
0.326904296875,
0.0236968994140625,
0.326904296875
],
[
0.2303466796875,
0.264404296875,
0.775390625,
0.264404296875,
0.775390625,
0.27783203125,
0.2303466796875,
0.27783203125
],
[
0.371826171875,
0.279052734375,
0.6337890625,
0.279052734375,
0.6337890625,
0.29345703125,
0.371826171875,
0.29345703125
],
[
0.02777099609375,
0.3125,
0.066162109375,
0.31494140625,
0.056365966796875,
0.399658203125,
0.0187835693359375,
0.396484375
],
[
0.265625,
0.335205078125,
0.740234375,
0.335205078125,
0.740234375,
0.348388671875,
0.265625,
0.348388671875
],
[
0.265625,
0.349853515625,
0.740234375,
0.349853515625,
0.740234375,
0.363037109375,
0.265625,
0.363037109375
],
[
0.2646484375,
0.363037109375,
0.740234375,
0.363037109375,
0.740234375,
0.376953125,
0.2646484375,
0.376953125
],
[
0.265625,
0.37890625,
0.73876953125,
0.37890625,
0.73876953125,
0.389404296875,
0.265625,
0.389404296875
],
[
0.021240234375,
0.391357421875,
0.06451416015625,
0.392578125,
0.062103271484375,
0.4873046875,
0.0187835693359375,
0.486083984375
],
[
0.2646484375,
0.391357421875,
0.740234375,
0.391357421875,
0.740234375,
0.40478515625,
0.2646484375,
0.40478515625
],
[
0.2646484375,
0.40478515625,
0.73876953125,
0.40478515625,
0.73876953125,
0.41845703125,
0.2646484375,
0.41845703125
],
[
0.2646484375,
0.41845703125,
0.73876953125,
0.41845703125,
0.73876953125,
0.431884765625,
0.2646484375,
0.431884765625
],
[
0.2646484375,
0.431884765625,
0.740234375,
0.431884765625,
0.740234375,
0.44580078125,
0.2646484375,
0.44580078125
],
[
0.2646484375,
0.4462890625,
0.740234375,
0.4462890625,
0.740234375,
0.460205078125,
0.2646484375,
0.460205078125
],
[
0.2646484375,
0.460205078125,
0.73876953125,
0.460205078125,
0.73876953125,
0.473388671875,
0.2646484375,
0.473388671875
],
[
0.2646484375,
0.473388671875,
0.740234375,
0.473388671875,
0.740234375,
0.4873046875,
0.2646484375,
0.4873046875
],
[
0.0228729248046875,
0.48486328125,
0.05963134765625,
0.48486328125,
0.05963134765625,
0.70703125,
0.0228729248046875,
0.70703125
],
[
0.2646484375,
0.488037109375,
0.740234375,
0.488037109375,
0.740234375,
0.501953125,
0.2646484375,
0.501953125
],
[
0.2646484375,
0.501953125,
0.740234375,
0.501953125,
0.740234375,
0.51513671875,
0.2646484375,
0.51513671875
],
[
0.26318359375,
0.51513671875,
0.740234375,
0.5146484375,
0.740234375,
0.52783203125,
0.26318359375,
0.52880859375
],
[
0.2646484375,
0.52978515625,
0.740234375,
0.52978515625,
0.740234375,
0.54345703125,
0.2646484375,
0.54345703125
],
[
0.2646484375,
0.54345703125,
0.740234375,
0.54345703125,
0.740234375,
0.556640625,
0.2646484375,
0.556640625
],
[
0.26318359375,
0.556640625,
0.740234375,
0.55615234375,
0.740234375,
0.57080078125,
0.26318359375,
0.5712890625
],
[
0.261474609375,
0.5693359375,
0.60888671875,
0.57080078125,
0.60888671875,
0.58544921875,
0.261474609375,
0.583984375
],
[
0.2646484375,
0.5986328125,
0.740234375,
0.5986328125,
0.740234375,
0.61279296875,
0.2646484375,
0.61279296875
],
[
0.2646484375,
0.6123046875,
0.4755859375,
0.6123046875,
0.4755859375,
0.62548828125,
0.2646484375,
0.62548828125
],
[
0.221435546875,
0.6572265625,
0.233642578125,
0.6572265625,
0.233642578125,
0.66845703125,
0.221435546875,
0.66845703125
],
[
0.25,
0.65576171875,
0.373291015625,
0.65576171875,
0.373291015625,
0.67041015625,
0.25,
0.67041015625
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2156982421875,
0.70556640625,
0.78369140625,
0.705078125,
0.78369140625,
0.7197265625,
0.2156982421875,
0.72021484375
],
[
0.2181396484375,
0.7216796875,
0.78369140625,
0.7216796875,
0.78369140625,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.2181396484375,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.751953125,
0.7861328125,
0.751953125,
0.7861328125,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.78076171875,
0.2181396484375,
0.78076171875
],
[
0.21728515625,
0.78076171875,
0.78369140625,
0.78076171875,
0.78369140625,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.7861328125,
0.8125,
0.7861328125,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Contrast-Phys: Unsupervised Video-based",
"Remote Physiological Measurement via",
"Spatiotemporal Contrast",
"Zhaodong Sun and Xiaobai Li",
"",
"Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland",
"(zhaodong.sun, xiaobai.li)@oulu.fi",
"Sn8",
"Abstract. Video-based remote physiological measurement utilizes face",
"videos to measure the blood volume change signal, which is also called",
"remote photoplethysmography (rPPG). Supervised methods for rPPG",
"measurements achieve state-of-the-artperformance.However,supervised",
"[A's]",
"rPPG methods require face videos and ground truth physiological signals",
"for model training. In this paper, we propose an unsupervised rPPG mea-",
"surement method that does not require ground truth signals for training.",
"We use a 3DCNN model to generate multiple rPPG signals from each",
"video in different spatiotemporal locations and train the model with a",
"contrastive loss where rPPG signals from the same video are pulled to-",
"gether while those from different videos are pushed away. We test on",
"1880",
"five public datasets, including RGB videos and NIR videos. The results",
"show that our method outperforms the previous unsupervised baseline",
"and achieves accuracies very close to the current best supervised rPPG",
"methods on all five datasets. Furthermore, we also demonstrate that",
"our approach can run at a much faster speed and is more robust to",
"noises than the previous unsupervised baseline. Our code is available at",
"https://github.com/zhaodongsun/contrast-phys.",
"Keywords: Remote Photoplethysmography, Face Video, Unsupervised",
"Learning, Contrastive Learning",
"1",
"Introduction",
"Traditional physiological measurement requires skin-contact sensors to measure",
"physiological signals such as contact photoplethysmography (PPG) and electro-",
"cardiography (ECG). Some physiological parameters like heart rate (HR), res-",
"piration frequency (RF), and heart rate variability (HRV) can be derived from",
"PPG signals for healthcare [42,55] and emotion analysis [57,29,40].However, the",
"skin-contact physiological measurement requires specific biomedical equipment",
"like pulse oximeters, and contact sensors may cause discomfort and skin irri-",
"tation. Remote physiological measurement uses a camera to record face videos",
"for measuring remote photoplethysmography (rPPG).The weak color change in",
"faces can be captured by cameras to obtain the rPPG signal from which several"
] |
{
"arxiv_id": "2208.04378",
"language": "en",
"page_idx": 0
}
|
||
2208.04378-en-1
|
2208.04378-en-1.jpg
|
2 Z. Sun and X. Li
physiological parameters such as HR, RF, and HRV [38] can be measured. Videobased physiological measurement only requires off-the-shelf cameras rather than professional biomedical sensors, and is not constrained by physical distance, which has a great potential for remote healthcare [42,55] and emotion analysis applications [57,29,40].
In earlier rPPG studies [49,38,8,52], researchers proposed handcrafted features to extract rPPG signals. Later, some deep learning (DL)-based methods [7,45,56,58,18,23,32,33,25,35] were proposed, which employed supervised approaches with various network architectures for measuring rPPG signals. On one side, under certain circumstances, e.g., when head motions are involved or the videos are heterogeneous, DL-based methods could be more robust than the traditional handcrafted approaches. On the other side, DL-based rPPG methods require a large-scale dataset including face videos and ground truth physiological signals. Although face videos are comparatively easy to obtain in large amount, it is expensive to get the ground truth physiological signals which are measured by contact sensors, and synchronized with the face videos.
Can we only use face videos without ground truth physiological signals to train models for rPPG measurement? Gideon and Stent [10] proposed a selfsupervised method to train rPPG measurement models without labels. They first downsample a video to get the downsampled rPPG (negative sample) and upsample the downsampled rPPG to get the reconstructed rPPG (positive sample). They also used the original video to get the anchor rPPG. Then a triplet loss is used to pull together the positive and anchor samples, and push away negative and anchor samples. However, their method has three problems. 1) They have to forward one video into their backbone model twice, which causes an extra computation burden. 2) There is still a significant gap between the performance of their unsupervised method and the state-of-the-art supervised rPPG methods [45,56,34]. 3) They showed in their paper [10] that their method is easily impacted by external periodic noise.
We propose a new unsupervised method (Contrast-Phys) to tackle the problems above. Our method was built on four observations about rPPG. 1) rPPG spatial similarity: rPPG signals measured from different facial areas have similar power spectrum densities (PSDs) 2) rPPG temporal similarity: In a short time, two rPPG signals (e.g., two consecutive 5s clips) usually present similar PSDs as the HR tends to take smooth transits in most cases. 3) Cross-video rPPG dissimilarity: The PSDs of rPPG signals from different videos are different. 4) HR range constraint: The HR should fall between 40 and 250 beats per minute (bpm), so we only care about the PSD in the frequency interval between 0.66 Hz and 4.16 Hz.
We propose to use a 3D convolutional neural network (3DCNN) to process an input video to get a spatiotemporal rPPG (ST-rPPG) block. The ST-rPPG block contains multiple rPPG signals along three dimensions of height, width, and time. According to the rPPG spatiotemporal similarity, we can randomly sample rPPG signals from the same video in different spatiotemporal locations and pull them together. According to the cross-video rPPG dissimilarity, the sampled
|
[
[
0.221435546875,
0.11297607421875,
0.233642578125,
0.117431640625,
0.226318359375,
0.1287841796875,
0.214111328125,
0.125
],
[
0.26953125,
0.11553955078125,
0.38818359375,
0.11553955078125,
0.38818359375,
0.1287841796875,
0.26953125,
0.1287841796875
],
[
0.21728515625,
0.147705078125,
0.78173828125,
0.146484375,
0.78173828125,
0.1610107421875,
0.21728515625,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.1767578125,
0.2181396484375,
0.1767578125
],
[
0.2181396484375,
0.177978515625,
0.78515625,
0.177978515625,
0.78515625,
0.1912841796875,
0.2181396484375,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.21728515625,
0.2083740234375,
0.381591796875,
0.2083740234375,
0.381591796875,
0.222900390625,
0.21728515625,
0.222900390625
],
[
0.24267578125,
0.2266845703125,
0.78369140625,
0.2266845703125,
0.78369140625,
0.2412109375,
0.24267578125,
0.2412109375
],
[
0.2181396484375,
0.242431640625,
0.78369140625,
0.242431640625,
0.78369140625,
0.255615234375,
0.2181396484375,
0.255615234375
],
[
0.2181396484375,
0.25830078125,
0.78369140625,
0.25830078125,
0.78369140625,
0.271484375,
0.2181396484375,
0.271484375
],
[
0.2181396484375,
0.272705078125,
0.7861328125,
0.272705078125,
0.7861328125,
0.285888671875,
0.2181396484375,
0.285888671875
],
[
0.2181396484375,
0.287353515625,
0.78759765625,
0.287353515625,
0.78759765625,
0.300537109375,
0.2181396484375,
0.300537109375
],
[
0.221435546875,
0.3037109375,
0.78173828125,
0.3037109375,
0.78173828125,
0.314453125,
0.221435546875,
0.314453125
],
[
0.21728515625,
0.317626953125,
0.7861328125,
0.317626953125,
0.7861328125,
0.33203125,
0.21728515625,
0.33203125
],
[
0.21728515625,
0.33203125,
0.78515625,
0.33203125,
0.78515625,
0.34521484375,
0.21728515625,
0.34521484375
],
[
0.2181396484375,
0.347900390625,
0.78369140625,
0.347900390625,
0.78369140625,
0.361083984375,
0.2181396484375,
0.361083984375
],
[
0.2181396484375,
0.363037109375,
0.7861328125,
0.363037109375,
0.7861328125,
0.376953125,
0.2181396484375,
0.376953125
],
[
0.21728515625,
0.376953125,
0.63623046875,
0.37744140625,
0.63623046875,
0.392578125,
0.21728515625,
0.391357421875
],
[
0.2442626953125,
0.39453125,
0.78515625,
0.39453125,
0.78515625,
0.4091796875,
0.2442626953125,
0.4091796875
],
[
0.2156982421875,
0.410400390625,
0.78515625,
0.410888671875,
0.78515625,
0.425537109375,
0.2156982421875,
0.4248046875
],
[
0.2181396484375,
0.4267578125,
0.78515625,
0.4267578125,
0.78515625,
0.43994140625,
0.2181396484375,
0.43994140625
],
[
0.21728515625,
0.442626953125,
0.78759765625,
0.442626953125,
0.78759765625,
0.455810546875,
0.21728515625,
0.455810546875
],
[
0.2181396484375,
0.45703125,
0.78369140625,
0.45703125,
0.78369140625,
0.47021484375,
0.2181396484375,
0.47021484375
],
[
0.2181396484375,
0.4716796875,
0.7861328125,
0.4716796875,
0.7861328125,
0.486083984375,
0.2181396484375,
0.486083984375
],
[
0.2181396484375,
0.4873046875,
0.7861328125,
0.4873046875,
0.7861328125,
0.50048828125,
0.2181396484375,
0.50048828125
],
[
0.2181396484375,
0.50244140625,
0.78515625,
0.50244140625,
0.78515625,
0.5166015625,
0.2181396484375,
0.5166015625
],
[
0.2197265625,
0.51953125,
0.78515625,
0.51953125,
0.78515625,
0.52978515625,
0.2197265625,
0.52978515625
],
[
0.21728515625,
0.53076171875,
0.7861328125,
0.5322265625,
0.7861328125,
0.546875,
0.21728515625,
0.54541015625
],
[
0.2181396484375,
0.54736328125,
0.7861328125,
0.54736328125,
0.7861328125,
0.56103515625,
0.2181396484375,
0.56103515625
],
[
0.2156982421875,
0.56103515625,
0.78759765625,
0.56298828125,
0.78759765625,
0.5771484375,
0.2156982421875,
0.57470703125
],
[
0.21728515625,
0.57763671875,
0.5400390625,
0.57763671875,
0.5400390625,
0.59130859375,
0.21728515625,
0.59130859375
],
[
0.24267578125,
0.59521484375,
0.78369140625,
0.59521484375,
0.78369140625,
0.60986328125,
0.24267578125,
0.60986328125
],
[
0.21728515625,
0.6123046875,
0.78515625,
0.6123046875,
0.78515625,
0.62548828125,
0.21728515625,
0.62548828125
],
[
0.2181396484375,
0.626953125,
0.78369140625,
0.626953125,
0.78369140625,
0.64013671875,
0.2181396484375,
0.64013671875
],
[
0.21728515625,
0.6416015625,
0.78515625,
0.6416015625,
0.78515625,
0.65478515625,
0.21728515625,
0.65478515625
],
[
0.2181396484375,
0.6572265625,
0.7861328125,
0.6572265625,
0.7861328125,
0.67041015625,
0.2181396484375,
0.67041015625
],
[
0.2181396484375,
0.671875,
0.7861328125,
0.671875,
0.7861328125,
0.68505859375,
0.2181396484375,
0.68505859375
],
[
0.2181396484375,
0.68603515625,
0.78515625,
0.68603515625,
0.78515625,
0.69970703125,
0.2181396484375,
0.69970703125
],
[
0.2181396484375,
0.7021484375,
0.78515625,
0.7021484375,
0.78515625,
0.71533203125,
0.2181396484375,
0.71533203125
],
[
0.2181396484375,
0.71728515625,
0.78759765625,
0.71728515625,
0.78759765625,
0.73095703125,
0.2181396484375,
0.73095703125
],
[
0.2197265625,
0.734375,
0.4306640625,
0.734375,
0.4306640625,
0.744140625,
0.2197265625,
0.744140625
],
[
0.2442626953125,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2442626953125,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.7861328125,
0.7666015625,
0.7861328125,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.78076171875,
0.78759765625,
0.78076171875,
0.78759765625,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.21728515625,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.21728515625,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82470703125,
0.21728515625,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"2",
"Z. Sun and X. Li",
"physiological parameters such as HR, RF, and HRV [38] can be measured. Video-",
"based physiological measurement only requires off-the-shelf cameras rather than",
"professional biomedical sensors, and is not constrained by physical distance,",
"which has a great potential for remote healthcare [42,55] and emotion analysis",
"applications [57,29,40].",
"In earlier rPPG studies [49,38,8,52], researchers proposed handcrafted fea-",
"tures to extract rPPG signals. Later, some deep learning (DL)-based meth-",
"ods [7,45,56,58,18,23,32,33,25,35] were proposed, which employed supervised ap-",
"proaches with various network architectures for measuring rPPG signals. On one",
"side, under certain circumstances, e.g., when head motions are involved or the",
"videos areheterogeneous, DL-based methods could bemorerobust than the tra",
"ditional handcrafted approaches. On the other side, DL-based rPPG methods",
"require a large-scale dataset including face videos and ground truth physiological",
"signals. Although face videos are comparatively easy to obtain in large amount.",
"it is expensive to get the ground truth physiological signals which are measured",
"by contact sensors, and synchronized with the face videos.",
"Can we only use face videos without ground truth physiological signals to",
"train models for rPPG measurement? Gideon and Stent [1o] proposed a self-",
"supervised method to train rPPG measurement models without labels. They",
"first downsample a video to get the downsampled rPPG (negative sample) and",
"upsample the dowmsampled rPPG to get the reconstructed rPPG (positive sam-",
"ple). They also used the original video to get the anchor rPPG. Then a triplet",
"loss is used to pull together the positive and anchor samples, and push away",
"negative and anchor samples. However, their method has three problems. 1)",
"Theyhave toforward one video into their backbone model twice, which causes",
"an extra computation burden. 2) There is still a significant gap between the",
"performance of their unsupervised method and the state-of-the-art supervised",
"rPPGmethods[45,56,34].3)Theyshowed intheirpaper[10] thattheirmethod",
"is easily impacted by external periodic noise.",
"We propose a new unsupervised method (Contrast-Phys) to tackle the prob-",
"lems above. Our method was built on four observations about rPPG. 1)rPPG",
"spatial similarity: rPPG signals measured from different facial areas have simi-",
"lar power spectrum densities (PSDs) 2) rPPG temporal similarity: In a short",
"time, two rPPG signals (e.g., two consecutive 5s clips) usually present similar",
"PSDs as the HR tends to take smooth transits in most cases. 3) Cross-video",
"rPPG dissimilarity: The PSDs of rPPG signals from different videos are dif-",
"ferent. 4) HR range constraint: The HR should fall between 40 and 250 beats",
"per minute (bpm), so we only care about the PSD in the frequency interval",
"between0.66Hzand4.16Hz.",
"We propose to use a 3D convolutional neural network (3DCNN) to process an",
"input video to get a spatiotemporal rPPG (ST-rPPG) block. The ST-rPPG block",
"contains multiple rPPG signals along three dimensions of height, width, and",
"time. According to the rPPG spatiotemporal similarity, we can randomly sample",
"rPPG signals from the same video in different spatiotemporal locations and pull",
"them together. According to the cross-video rPPG dissimilarity, the sampled"
] |
{
"arxiv_id": "2208.04378",
"language": "en",
"page_idx": 1
}
|
||
2208.04378-en-2
|
2208.04378-en-2.jpg
|
Contrast- Phys 3
rPPG signals from different videos are pushed away. The whole procedures are
shown in Fig. 4 and Fig. 5.
The contributions of this work are 1) We propose a novel rPPG representation called spatiotemporal rPPG (ST-rPPG) block to obtain rPPG signals in spatiotemporal dimensions. 2) Based on four observations about rPPG, including rPPG spatiotemporal similarity and cross-video rPPG dissimilarity, we propose an unsupervised method based on contrastive learning. 3) We conduct experiments on five rPPG datasets (PURE [46], UBFC-rPPG [4], OBF [19], MRNIRP [26], and MMSE-HR [59]) including RGB and NIR videos under various scenarios. Our method outperforms the previous unsupervised baseline [10] and achieves very close performance to supervised rPPG methods. Contrast-Phys also shows significant advantages with fast running speed and noise robustness compared to the previous unsupervised baseline.
# 2 Related Work
Video-Based Remote Physiological Measurement. Verkruysse et al.[49] first proposed that rPPG can be measured from face videos from the green channel. Several traditional handcraft methods [38,8,9,52,17,48,53] were proposed to further improve rPPG signal quality. Most rPPG methods proposed in earlier years used handcrafted procedures and did not need datasets for training, which are referred to as traditional methods. Deep learning (DL) methods for rPPG measurement are rapidly emerging. Several studies [7,45,23,35] used a 2D convolutional neural network (2DCNN) with two consecutive video frames as the input for rPPG measurement. Another type of DL-based methods [32,33,25] used a spatial-temporal signal map extracted from different facial areas as the input to feed into a 2DCNN model. Recently, 3DCNN-based methods [58,56,10] were proposed to achieved good performance on compressed videos [58].The DL-based methods require both face videos and ground truth physiological signals, so we refer to them as supervised methods. Recently, Gideon and Stent [10] proposed an unsupervised method to train a DL model without ground truth physiological signals. However, their method falls behind some supervised methods and is not robust to external noise. In addition, the running speed is not satisfactory.
Contrastive Learning. Contrastive learning is a self-supervised learning method widely used in video and image feature embedding, which facilitates downstream task training and small dataset fne-tuning [12,41,36,6,47,13,11,30,39]. A DL model working as a feature extractor maps a high-dimensional image/video into a low dimensional feature vector. To train this DL feature extractor, features from different views of the same sample (positive pairs) are pulled together while features from views of different samples (negative pairs) are pushed away Data augmentations (such as cropping, blurring [6], and temporal sampling [39]) are used to obtain different views of the same sample so that the learned features are invariant to some augmentations. Previous works mentioned above use
|
[
[
0.63623046875,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1275634765625,
0.63623046875,
0.1275634765625
],
[
0.77392578125,
0.11297607421875,
0.7890625,
0.11614990234375,
0.78515625,
0.1300048828125,
0.76953125,
0.126953125
],
[
0.2181396484375,
0.147705078125,
0.7861328125,
0.147705078125,
0.7861328125,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.21728515625,
0.1634521484375,
0.41259765625,
0.1634521484375,
0.41259765625,
0.177978515625,
0.21728515625,
0.177978515625
],
[
0.2442626953125,
0.1767578125,
0.78173828125,
0.1767578125,
0.78173828125,
0.1912841796875,
0.2442626953125,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.21728515625,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.21728515625,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.7861328125,
0.2235107421875,
0.7861328125,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.214111328125,
0.2374267578125,
0.7861328125,
0.236083984375,
0.7861328125,
0.25390625,
0.214111328125,
0.255126953125
],
[
0.2181396484375,
0.25390625,
0.78369140625,
0.25390625,
0.78369140625,
0.267578125,
0.2181396484375,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.78515625,
0.26953125,
0.78515625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2197265625,
0.285888671875,
0.78515625,
0.285888671875,
0.78515625,
0.296630859375,
0.2197265625,
0.296630859375
],
[
0.2181396484375,
0.2998046875,
0.78515625,
0.2998046875,
0.78515625,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.2181396484375,
0.314453125,
0.7861328125,
0.314453125,
0.7861328125,
0.328857421875,
0.2181396484375,
0.328857421875
],
[
0.2181396484375,
0.328857421875,
0.56640625,
0.328857421875,
0.56640625,
0.34228515625,
0.2181396484375,
0.34228515625
],
[
0.2197265625,
0.36865234375,
0.2344970703125,
0.36865234375,
0.2344970703125,
0.37890625,
0.2197265625,
0.37890625
],
[
0.2459716796875,
0.3662109375,
0.386474609375,
0.3662109375,
0.386474609375,
0.380615234375,
0.2459716796875,
0.380615234375
],
[
0.2197265625,
0.396484375,
0.78515625,
0.396484375,
0.78515625,
0.410888671875,
0.2197265625,
0.410888671875
],
[
0.2181396484375,
0.412353515625,
0.78369140625,
0.412353515625,
0.78369140625,
0.425537109375,
0.2181396484375,
0.425537109375
],
[
0.21728515625,
0.427978515625,
0.7861328125,
0.427978515625,
0.7861328125,
0.44140625,
0.21728515625,
0.44140625
],
[
0.21728515625,
0.442626953125,
0.7861328125,
0.442626953125,
0.7861328125,
0.45703125,
0.21728515625,
0.45703125
],
[
0.21728515625,
0.458251953125,
0.7861328125,
0.458251953125,
0.7861328125,
0.4716796875,
0.21728515625,
0.4716796875
],
[
0.21728515625,
0.472900390625,
0.7861328125,
0.4716796875,
0.7861328125,
0.486083984375,
0.21728515625,
0.4873046875
],
[
0.21728515625,
0.4873046875,
0.78369140625,
0.4873046875,
0.78369140625,
0.501953125,
0.21728515625,
0.501953125
],
[
0.2181396484375,
0.50390625,
0.7861328125,
0.50390625,
0.7861328125,
0.517578125,
0.2181396484375,
0.517578125
],
[
0.21728515625,
0.51806640625,
0.78515625,
0.51806640625,
0.78515625,
0.5322265625,
0.21728515625,
0.5322265625
],
[
0.2181396484375,
0.53271484375,
0.78515625,
0.53271484375,
0.78515625,
0.54736328125,
0.2181396484375,
0.54736328125
],
[
0.21728515625,
0.548828125,
0.78515625,
0.548828125,
0.78515625,
0.5625,
0.21728515625,
0.5625
],
[
0.2181396484375,
0.56298828125,
0.7861328125,
0.56298828125,
0.7861328125,
0.5771484375,
0.2181396484375,
0.5771484375
],
[
0.2181396484375,
0.5791015625,
0.78369140625,
0.5791015625,
0.78369140625,
0.59228515625,
0.2181396484375,
0.59228515625
],
[
0.2197265625,
0.59521484375,
0.78369140625,
0.59521484375,
0.78369140625,
0.60595703125,
0.2197265625,
0.60595703125
],
[
0.2181396484375,
0.609375,
0.7861328125,
0.609375,
0.7861328125,
0.62255859375,
0.2181396484375,
0.62255859375
],
[
0.21728515625,
0.62353515625,
0.78369140625,
0.62255859375,
0.78369140625,
0.63720703125,
0.21728515625,
0.63818359375
],
[
0.2156982421875,
0.63818359375,
0.78515625,
0.6396484375,
0.78515625,
0.65380859375,
0.2156982421875,
0.65283203125
],
[
0.2181396484375,
0.65576171875,
0.3056640625,
0.65576171875,
0.3056640625,
0.66748046875,
0.2181396484375,
0.66748046875
],
[
0.2197265625,
0.68994140625,
0.79736328125,
0.68994140625,
0.79736328125,
0.7041015625,
0.2197265625,
0.7041015625
],
[
0.2181396484375,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.21728515625,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.734375,
0.21728515625,
0.734375
],
[
0.2181396484375,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.21728515625,
0.76513671875,
0.78173828125,
0.76513671875,
0.78173828125,
0.77880859375,
0.21728515625,
0.77880859375
],
[
0.2181396484375,
0.77978515625,
0.78369140625,
0.7822265625,
0.78369140625,
0.79541015625,
0.2181396484375,
0.79345703125
],
[
0.2181396484375,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81103515625,
0.2181396484375,
0.81103515625
],
[
0.2181396484375,
0.8125,
0.78369140625,
0.8125,
0.78369140625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Contrast-Phys",
"3",
"rPPG signals from different videos are pushed away. The whole procedures are",
"shown in Fig. 4 and Fig. 5.",
"The contributions of this work are 1) We propose a novel rPPG represen",
"tation called spatiotemporal rPPG (ST-rPPG) block to obtain rPPG signals",
"in spatiotemporal dimensions. 2) Based on four observations about rPPG, in-",
"cluding rPPG spatiotemporal similarity and cross-video rPPG dissimilarity, we",
" propose an unsupervised method based on contrastive learning. 3) We conduct",
"experiments on five rPPG datasets (PURE [46], UBFC-rPPG [4], OBF [19], MR-",
"NIRP [26], and MMSE-HR [59] including RGB and NIR videos under various",
"scenarios.Ourmethod outperformsthepreviousunsupervised baselineioand",
"achieves very close performance to supervised rPPG methods. Contrast-Phys",
"also shows significant advantages with fast running speed and noise robustness",
"compared to the previous unsupervised baseline.",
"2",
" Related Work",
"Video-Based Remote Physiological Measurement. Verkruysse et al. [49]",
"first proposed that rPPG can be measured from face videos from the green chan-",
"nel. Several traditional handcraft methods [38,8,9,52,17,48,53] were proposed to",
"further improve rPPG signal quality. Most rPPG methods proposed in earlier",
"years used handcrafted procedures and did not need datasets for training, which",
"are referred to as traditional methods. Deep learning (DL) methods for rPPG",
"measurement are rapidly emerging. Several studies [7,45,23,35] used a 2D con-",
"volutional neural network (2DCNN) with two consecutive video frames as the",
"input for rPPG measurement. Another type of DL-based methods [32,33,25]",
"used a spatial-temporal signal map extracted from different facial areas as the",
"input to feed into a 2DCNN model. Recently, 3DCNN-based methods [58,56,10]",
"were proposed to achieved good performance on compressed videos [58]. The",
"DL-based methods require both face videos and ground truth physiological sig-",
"nals,so we refer to them as supervised methods.Recently,Gideon and Stent10",
"proposed an unsupervised method to train a DL model without ground truth",
"physiological signals. However, their method falls behind some supervised meth-",
"ods and is not robust to external noise. In addition, the running speed is not",
"satisfactory.",
"Contrastive Learning. Contrastive learning is a self-supervised learning method",
"widely used in video and image feature embedding, which facilitates downstream",
"task training and small dataset fine-tuning [12,41,36,6,47,13,11,30,39]. A DL",
"model working as a feature extractor maps a high-dimensional image/video into",
"a low dimensional feature vector. To train this DL feature extractor, features",
"from different views of the same sample (positive pairs) are pulled together",
"while features from views of different samples (negative pairs)are pushed away",
"Data augmentations (such as cropping, blurring [6], and temporal sampling [39])",
"are used to obtain different views of the same sample so that the learned fea-",
"tures are invariant to some augmentations. Previous works mentioned above use"
] |
{
"arxiv_id": "2208.04378",
"language": "en",
"page_idx": 2
}
|
||
2208.04378-en-3
|
2208.04378-en-3.jpg
|
4 Z. Sun and X. Li
contrastive learning to let a DL model produce abstract features for downstream tasks such as image classification [6], video classification [39], face recognition [41]. On the other hand, our work uses contrastive learning to directly let a DL model produce rPPG signals, enabling unsupervised learning without ground truth physiological signals.
# 3 Observations about rPPG
This section describes four observations about rPPG, which are the precondition
to design our method and enable unsupervised learning.
rPPG Spatial Similarity. rPPG signals from different facial areas have similar waveforms, and their PSDs are also similar. Several works [17,48,16,53,51,22,21] also exploited rPPG spatial similarity to design their methods. There might be small phase and amplitude differences between two rPPG signals from two different body skin areas [14,15]. However, when rPPG waveforms are transformed to PSDs, the phase information is erased, and the amplitude can be normalized to cancel the amplitude difference. In Fig. 1, the rPPG waveforms from four spatial areas are similar, and they have the same peaks in PSDs.
<FigureHere>
Fig.1: Illustration of rPPG spatial similarity. The rPPG signals from four facial areas (A, B, C, D) have similar waveforms and power spectrum densities (PSDs)
rPPG Temporal Similarity. The HR does not change rapidly in a short term [10]. Stricker et al. [46] also found that the HR varies slightly in a short time interval in their dataset. Since the HR has a dominant peak in PSD, the PSD does not change rapidly, either. If we randomly sample several small windows from a short rPPG clip (e.g., 10s), the PSDs of these windows should be similar. In Fig. 2, we sample two 5s windows from a short 10s rPPG signal and get the PSDs of these two windows. The two PSDs are similar and have sharp peaks at the same frequency. Since this observation is only valid under the condition of short-term rPPG signals, in the following sensitivity analysis part, we will discuss the influence of the signal length on our model performance. Overall,
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.126953125,
0.21728515625,
0.126953125
],
[
0.26953125,
0.11553955078125,
0.38818359375,
0.11553955078125,
0.38818359375,
0.1287841796875,
0.26953125,
0.1287841796875
],
[
0.2156982421875,
0.147705078125,
0.78515625,
0.146484375,
0.78515625,
0.1610107421875,
0.2156982421875,
0.1622314453125
],
[
0.2181396484375,
0.1634521484375,
0.78369140625,
0.1634521484375,
0.78369140625,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.2197265625,
0.177978515625,
0.78369140625,
0.177978515625,
0.78369140625,
0.1912841796875,
0.2197265625,
0.1912841796875
],
[
0.2181396484375,
0.1925048828125,
0.78759765625,
0.1925048828125,
0.78759765625,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.21728515625,
0.208984375,
0.410888671875,
0.208984375,
0.410888671875,
0.222900390625,
0.21728515625,
0.222900390625
],
[
0.2181396484375,
0.2474365234375,
0.25244140625,
0.2474365234375,
0.25244140625,
0.2587890625,
0.2181396484375,
0.2587890625
],
[
0.2459716796875,
0.246826171875,
0.50244140625,
0.246826171875,
0.50244140625,
0.261474609375,
0.2459716796875,
0.261474609375
],
[
0.21728515625,
0.27587890625,
0.78369140625,
0.276611328125,
0.78369140625,
0.291748046875,
0.21728515625,
0.290283203125
],
[
0.21728515625,
0.291748046875,
0.619140625,
0.292236328125,
0.619140625,
0.306884765625,
0.21728515625,
0.30615234375
],
[
0.2181396484375,
0.328857421875,
0.78515625,
0.328857421875,
0.78515625,
0.34228515625,
0.2181396484375,
0.34228515625
],
[
0.2197265625,
0.3447265625,
0.78515625,
0.3447265625,
0.78515625,
0.35791015625,
0.2197265625,
0.35791015625
],
[
0.2181396484375,
0.359130859375,
0.78515625,
0.359130859375,
0.78515625,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.2181396484375,
0.373779296875,
0.78173828125,
0.373779296875,
0.78173828125,
0.386962890625,
0.2181396484375,
0.386962890625
],
[
0.2181396484375,
0.39013671875,
0.7861328125,
0.39013671875,
0.7861328125,
0.404052734375,
0.2181396484375,
0.404052734375
],
[
0.21728515625,
0.40478515625,
0.7861328125,
0.40478515625,
0.7861328125,
0.41845703125,
0.21728515625,
0.41845703125
],
[
0.2181396484375,
0.42041015625,
0.78515625,
0.42041015625,
0.78515625,
0.434326171875,
0.2181396484375,
0.434326171875
],
[
0.2181396484375,
0.43505859375,
0.6806640625,
0.43505859375,
0.6806640625,
0.448974609375,
0.2181396484375,
0.448974609375
],
[
0.2484130859375,
0.472900390625,
0.26953125,
0.472900390625,
0.26953125,
0.486083984375,
0.2484130859375,
0.486083984375
],
[
0.365234375,
0.474853515625,
0.38818359375,
0.474853515625,
0.38818359375,
0.486083984375,
0.365234375,
0.486083984375
],
[
0.62841796875,
0.481689453125,
0.6396484375,
0.481689453125,
0.6396484375,
0.48486328125,
0.62841796875,
0.48486328125
],
[
0.736328125,
0.4765625,
0.75634765625,
0.4765625,
0.75634765625,
0.4873046875,
0.736328125,
0.4873046875
],
[
0.301513671875,
0.4892578125,
0.56640625,
0.4892578125,
0.56640625,
0.52587890625,
0.301513671875,
0.52587890625
],
[
0.5361328125,
0.4912109375,
0.7099609375,
0.4912109375,
0.7099609375,
0.5234375,
0.5361328125,
0.5234375
],
[
0.303955078125,
0.52978515625,
0.60693359375,
0.52978515625,
0.60693359375,
0.5693359375,
0.303955078125,
0.5693359375
],
[
0.2197265625,
0.59130859375,
0.78515625,
0.59130859375,
0.78515625,
0.60595703125,
0.2197265625,
0.60595703125
],
[
0.2181396484375,
0.60791015625,
0.78369140625,
0.60791015625,
0.78369140625,
0.6220703125,
0.2181396484375,
0.6220703125
],
[
0.2197265625,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.68798828125,
0.2197265625,
0.68798828125
],
[
0.2197265625,
0.68994140625,
0.78515625,
0.68994140625,
0.78515625,
0.7041015625,
0.2197265625,
0.7041015625
],
[
0.2197265625,
0.705078125,
0.78515625,
0.705078125,
0.78515625,
0.7197265625,
0.2197265625,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.21728515625,
0.734375,
0.78369140625,
0.73486328125,
0.78369140625,
0.75,
0.21728515625,
0.74853515625
],
[
0.2181396484375,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"Z. Sun and X. Li",
"contrastive learning to let a DL model produce abstract features for downstream",
"tasks such as image classification [6], video classification [39], face recognition",
"[4i]. On the other hand, our work uses contrastive learning to directly let a DL",
"model produce rPPG signals, enabling unsupervised learning without ground",
"truth physiological signals.",
"3",
" Observations about rPPG",
"This section describes four observations about rPPG,which are the precondition",
"to design our method and enable unsupervised learning.",
"rPPG Spatial Similarity. rPPG signals from different facial areas have similar",
"waveforms, and their PSDs are also similar. Several works [17,48,16,53,51,22,21]",
"also exploited rPPG spatial similarity to design their methods. There might be",
"small phase and amplitude differences between two rPPG signals from two dif",
"ferent body skin areas [14,15]. However, when rPPG waveforms are transformed",
"to PsDs, the phase information is erased, and the amplitude can be normalized",
"to cancel the amplitude difference. In Fig. 1, the rPPG waveforms from four",
"spatial areas are similar, and they have the same peaks in PsDs.",
"PSD",
"TPPG",
"P",
"PSD",
"--",
"--",
"---",
"Fig.1: Illustration of rPPG spatial similarity. The rPPG signals from four facial",
"areas (A, B, C, D) have similar waveforms and power spectrum densities (PSDs)",
"rPPG Temporal Similarity. The HR does not change rapidly in a short term",
"[10]. Stricker et al. [46] also found that the HR varies slightly in a short time",
"interval in their dataset. Since the HR has a dominant peak in PSD, the PSD",
"does not change rapidly, either. If we randomly sample several small windows",
"from a short rPPG clip (e.g., 10s), the PSDs of these windows should be similar.",
"In Fig. 2, we sample two 5s windows from a short 10s rPPG signal and get the",
"PSDs of these two windows. The two PSDs are similar and have sharp peaks",
"at the same frequency. Since this observation is only valid under the condition",
"of short-term rPPG signals, in the following sensitivity analysis part, we will",
"discuss the influence of the signal length on our model performance. Overall.",
"we can use the equation PSD,G(u(ti →ti + △t, Hi, W))) ~ PSDJG(u(t2 -)"
] |
{
"arxiv_id": "2208.04378",
"language": "en",
"page_idx": 3
}
|
||
2208.04378-en-4
|
2208.04378-en-4.jpg
|
Contrast-Phys 5
$\left.\begin{array}{ll}{t_{2}+\Delta t,\mathcal{H}_{2},\mathcal{W}_{2})}\\{t_{2}+\Delta t,\mathcal{H}_{2},\mathcal{W}_{2})}\\{t_{2}+\Delta t,\mathcal{H}_{2},\mathcal{W}_{2})}\\\end{array}\right\}$ to describe spatiotemporal rPPG similarity. $v\in\mathbb{R}^{T\times H\times W\times3}$ is a facial video, $G$ is an rPPG measurement algorithm. We can choose one facial area with a set of height $\mathcal{H}_1$ and width $\mathcal{W}_1$, and a time interval $t_1\to t_1+\Delta t$ from video $v$ to achieve one rPPG signal. We can achieve another rPPG signal similarly from the same video with $\mathcal{H}_2,\mathcal{W}_2$, and $t_2\to t_2+\Delta t.|t_1-t_2|$ should be small to satisfy the condition of short-term rPPG signals.
<FigureHere>
Fig.2: Illustration of rPPG temporal similarity. The rPPG signals from two
$\bar{\text{temporal windows (A, B) have similar PSDs}}$
<FigureHere>
Cross-video rPPG Dissimilarity. We assume rPPG signals from different face videos have different PSDs. Each video is recorded with different people and different physiological states (such as exercises and emotion status), so the HRs across different videos are likely different [1].Even though the HRs might be similar between two videos, the PSDs might still different (right) cross-video PSD pairs in
be different since PSD also contains the OBF dataset.
other physiological factors such as respiration rate [5] and HRV [37] which are unlikely to be all the same between two videos. To further validate the obser-
Fig.3: The most similar (left) and most
vation, we calculate the mean squared error for all cross-video PSD pairs in the OBF dataset [19] and show the most similar and most different cross-video PSD pairs in Fig. 3. It can be observed that the main cross-video PSD difference is the heart rate peak. The following equation describes cross-video rPPG dissimilar- $\mathop{\mathrm{ity.~PSD}}\{\dot{G}\big(v(t_{1}\rightarrow t_{1}+\Delta t,\bar{\mathcal H}_{1},\bar{\mathcal W}_{1})\big)\big\}\neq\mathop{\mathrm{PSD}}\big\{G\big(v^{\prime}(t_{2}\rightarrow t_{2}+\Delta t,\mathcal{H}_{2},\mathcal{W}_{2})\big)\big\}$ where $v$ and $v^\prime$ are two different videos. We can choose facial areas and time intervals from these two videos. The PSDs of the two rPPG signals should be different.
HR Range Constraint. The HR range for most people is between 40 and 250 bpm [2]. Most works [38,20] use this HR range for rPPG signal filtering and find
|
[
[
0.6357421875,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1275634765625,
0.6357421875,
0.1275634765625
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.214111328125,
0.1458740234375,
0.7890625,
0.141357421875,
0.7890625,
0.1622314453125,
0.214111328125,
0.1666259765625
],
[
0.21728515625,
0.1610107421875,
0.7861328125,
0.1622314453125,
0.7861328125,
0.1767578125,
0.21728515625,
0.175537109375
],
[
0.2181396484375,
0.177978515625,
0.78515625,
0.177978515625,
0.78515625,
0.1912841796875,
0.2181396484375,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.2083740234375,
0.7861328125,
0.2083740234375,
0.7861328125,
0.2215576171875,
0.2181396484375,
0.2215576171875
],
[
0.21728515625,
0.2235107421875,
0.65185546875,
0.2235107421875,
0.65185546875,
0.2374267578125,
0.21728515625,
0.2374267578125
],
[
0.295654296875,
0.26318359375,
0.32275390625,
0.26513671875,
0.3203125,
0.279052734375,
0.294189453125,
0.276611328125
],
[
0.435546875,
0.26513671875,
0.491943359375,
0.26513671875,
0.491943359375,
0.276611328125,
0.435546875,
0.276611328125
],
[
0.51708984375,
0.26513671875,
0.57275390625,
0.26513671875,
0.57275390625,
0.276611328125,
0.51708984375,
0.276611328125
],
[
0.685546875,
0.26318359375,
0.71142578125,
0.26513671875,
0.708984375,
0.279052734375,
0.68310546875,
0.276611328125
],
[
0.431396484375,
0.28857421875,
0.57177734375,
0.28857421875,
0.57177734375,
0.333984375,
0.431396484375,
0.333984375
],
[
0.494384765625,
0.313232421875,
0.50732421875,
0.313232421875,
0.50732421875,
0.317626953125,
0.494384765625,
0.317626953125
],
[
0.50732421875,
0.313232421875,
0.51708984375,
0.313232421875,
0.51708984375,
0.317626953125,
0.50732421875,
0.317626953125
],
[
0.2181396484375,
0.3642578125,
0.7861328125,
0.3642578125,
0.7861328125,
0.37890625,
0.2181396484375,
0.37890625
],
[
0.2156982421875,
0.380126953125,
0.53759765625,
0.37890625,
0.53759765625,
0.393310546875,
0.2156982421875,
0.39453125
],
[
0.2197265625,
0.45263671875,
0.484375,
0.45263671875,
0.484375,
0.466552734375,
0.2197265625,
0.466552734375
],
[
0.2181396484375,
0.466552734375,
0.484375,
0.46728515625,
0.484375,
0.481689453125,
0.2181396484375,
0.48095703125
],
[
0.2181396484375,
0.48291015625,
0.484375,
0.48291015625,
0.484375,
0.496826171875,
0.2181396484375,
0.496826171875
],
[
0.6357421875,
0.4873046875,
0.67626953125,
0.4873046875,
0.67626953125,
0.55615234375,
0.6357421875,
0.55615234375
],
[
0.50341796875,
0.49560546875,
0.52392578125,
0.49560546875,
0.52392578125,
0.505859375,
0.50341796875,
0.505859375
],
[
0.65625,
0.490478515625,
0.66259765625,
0.490478515625,
0.66259765625,
0.49365234375,
0.65625,
0.49365234375
],
[
0.2197265625,
0.5,
0.486083984375,
0.5,
0.486083984375,
0.51025390625,
0.2197265625,
0.51025390625
],
[
0.50341796875,
0.5068359375,
0.51708984375,
0.5068359375,
0.51708984375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.21728515625,
0.51318359375,
0.486083984375,
0.51318359375,
0.486083984375,
0.52783203125,
0.21728515625,
0.52783203125
],
[
0.50341796875,
0.51513671875,
0.521484375,
0.5166015625,
0.51953125,
0.52587890625,
0.50244140625,
0.5244140625
],
[
0.2181396484375,
0.52978515625,
0.483642578125,
0.52978515625,
0.483642578125,
0.54052734375,
0.2181396484375,
0.54052734375
],
[
0.2181396484375,
0.54345703125,
0.484375,
0.54345703125,
0.484375,
0.556640625,
0.2181396484375,
0.556640625
],
[
0.50341796875,
0.54150390625,
0.521484375,
0.54150390625,
0.521484375,
0.548828125,
0.50341796875,
0.548828125
],
[
0.54736328125,
0.5498046875,
0.62158203125,
0.54736328125,
0.62158203125,
0.56298828125,
0.54736328125,
0.56494140625
],
[
0.62158203125,
0.5498046875,
0.64453125,
0.5498046875,
0.64453125,
0.55810546875,
0.62158203125,
0.55810546875
],
[
0.689453125,
0.54736328125,
0.73876953125,
0.55029296875,
0.73779296875,
0.56640625,
0.68896484375,
0.56298828125
],
[
0.7646484375,
0.5498046875,
0.78369140625,
0.5498046875,
0.78369140625,
0.556640625,
0.7646484375,
0.556640625
],
[
0.66650390625,
0.5517578125,
0.6748046875,
0.5517578125,
0.6748046875,
0.556640625,
0.66650390625,
0.556640625
],
[
0.525390625,
0.5517578125,
0.53515625,
0.5517578125,
0.53515625,
0.556640625,
0.525390625,
0.556640625
],
[
0.73388671875,
0.5517578125,
0.765625,
0.548828125,
0.76708984375,
0.55615234375,
0.734375,
0.5595703125
],
[
0.2181396484375,
0.55810546875,
0.48779296875,
0.55810546875,
0.48779296875,
0.5712890625,
0.2181396484375,
0.5712890625
],
[
0.2181396484375,
0.57373046875,
0.488525390625,
0.57373046875,
0.488525390625,
0.5869140625,
0.2181396484375,
0.5869140625
],
[
0.498291015625,
0.57568359375,
0.7861328125,
0.57568359375,
0.7861328125,
0.5888671875,
0.498291015625,
0.5888671875
],
[
0.2156982421875,
0.58837890625,
0.490234375,
0.5869140625,
0.490234375,
0.6015625,
0.2156982421875,
0.60302734375
],
[
0.48779296875,
0.5888671875,
0.7861328125,
0.59033203125,
0.7861328125,
0.60498046875,
0.48779296875,
0.60400390625
],
[
0.214111328125,
0.599609375,
0.62841796875,
0.60302734375,
0.62841796875,
0.62060546875,
0.214111328125,
0.6171875
],
[
0.2197265625,
0.62060546875,
0.484375,
0.62060546875,
0.484375,
0.630859375,
0.2197265625,
0.630859375
],
[
0.21728515625,
0.6337890625,
0.486083984375,
0.63330078125,
0.486083984375,
0.64794921875,
0.21728515625,
0.6484375
],
[
0.2197265625,
0.65087890625,
0.78173828125,
0.65087890625,
0.78173828125,
0.6611328125,
0.2197265625,
0.6611328125
],
[
0.2181396484375,
0.6640625,
0.7861328125,
0.6640625,
0.7861328125,
0.67822265625,
0.2181396484375,
0.67822265625
],
[
0.2197265625,
0.68115234375,
0.78515625,
0.68115234375,
0.78515625,
0.69140625,
0.2197265625,
0.69140625
],
[
0.2181396484375,
0.693359375,
0.7861328125,
0.6923828125,
0.7861328125,
0.70703125,
0.2181396484375,
0.70849609375
],
[
0.21728515625,
0.708984375,
0.78369140625,
0.70849609375,
0.78369140625,
0.72265625,
0.21728515625,
0.7236328125
],
[
0.2181396484375,
0.7236328125,
0.78173828125,
0.7236328125,
0.78173828125,
0.73779296875,
0.2181396484375,
0.73779296875
],
[
0.2181396484375,
0.7392578125,
0.78515625,
0.7392578125,
0.78515625,
0.75244140625,
0.2181396484375,
0.75244140625
],
[
0.2181396484375,
0.7548828125,
0.78759765625,
0.7548828125,
0.78759765625,
0.76953125,
0.2181396484375,
0.76953125
],
[
0.2156982421875,
0.76806640625,
0.28515625,
0.77099609375,
0.28515625,
0.78515625,
0.2156982421875,
0.78271484375
],
[
0.21728515625,
0.8095703125,
0.78515625,
0.81005859375,
0.78515625,
0.82470703125,
0.21728515625,
0.82373046875
],
[
0.2181396484375,
0.82568359375,
0.7861328125,
0.82568359375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Contrast-Phys",
"5",
"t2 + At, H2, W2)) to describe spatiotemporal rPPG similarity. E RTxHxW×3",
"is a facial video, G is an rPPG measurement algorithm. We can choose one facial",
"area with a set of height Hi and width Wi, and a time interval ti → ti + △t",
"from video u to achieve one rPPG signal. We can achieve another rPPG signal",
"similarly from the same video with H2,W2,and t2→t2+△t.ti -t2l should",
"be small to satisfy the condition of short-term rPPG signals.",
"PSD",
"WindowA",
"WindowB",
"PSD",
"m",
"",
"",
"Fig.2:Illustration of rPPG temporal similarity. The rPPG signals from two",
"temporal windows (A, B) have similar PSDs",
"Cross-video rPPG Dissimilarity.",
"We assume rPPG signals from differ-",
"ent face videos have different PSDs.",
"",
"0.8",
"",
"Eachvideo is recorded with differ-",
"0",
"ent people and different physiologi-",
"80",
"cal states (such as exercises and emo-",
"tion status), so the HRs across differ-",
"200",
"100HR(150m,200",
"250",
"100HR(150m)",
"250",
"50",
"50",
"1200",
"ent videos are likely different [1]. Even",
"though the HRs might be similar be-",
"Fig.3: The most similar (left) and most",
"tween two videos, the PSDs might still",
"different (right) cross-video PSD pairs in",
"be different since PSD also contains the OBF dataset.",
"otherphvsiologicalfactorssuchasres-",
"piration rate [5] and HRV [37] which",
"are unlikelvto beallthe samebetweentwo videos.Tofurthervalidatetheobser",
"vation, we calculate the mean squared error for all cross-video PSD pairs in the",
"OBFdataset[19landshowthemostsimilarandmostdifferentcross-videoPSD",
"pairs in Fig. 3. It can be observed that the main cross-video PSD difference is the",
"heart rate peak. The following equation describes cross-video rPPG dissimilar-",
"ity. PSD,G(v(ti -→ ti + △t, H1, Wi))/ + PSD[G(u'(t2 → t2 + t, H2, W2)) !",
"where u and y' are two different videos. We can choose facial areas and time",
"intervals from these two videos. The PSDs of the two rPPG signals should be",
"different.",
"HR Range Constraint. The HR range for most people is between 40 and 250",
"bpm [2]. Most works [38,20] use this HR range for rPPG signal filtering and find"
] |
{
"arxiv_id": "2208.04378",
"language": "en",
"page_idx": 4
}
|
||
2302.00609-en-0
|
2302.00609-en-0.jpg
|
## $\textbf{Zero- shot Transfer of Article- aware Legal Outcome Classification for}$ $\textbf{European Court of Human Rights Cases}$
Santosh T.Y.S.S$^1,\textbf{Oana Ichim}^2,\textbf{Matthias Grabmair}^1$ $^{1}$School of Computation, Information, and Technology;
$\hat{\text{Technical University of Munich, Germany}}$
$^{2}$Graduate Institute of International and Development Studies, Geneva, Switzerland
$\{$santosh.tokala, matthias.grabmair}etum.de
oana. ichim@graduateinstitute. ch
## Abstract
the sole input. This reliance on the model learning statistical correspondences from case fact descriptions directly to outcomes neglects the role of legal sources in this relationship. As a consequence, the model may learn sub-optimal fact-outcome patterns that are informed by the case distribution in the data rather than learning to align facts with the legal source text containing applicable rules. The models may also attend to outcome-correlating distractors present in the dataset rather than engage in the legal fact-vs-law reasoning that is required of legal practitioners for a proper justification of the outcome (Santosh et al., 2022).
In this paper, we cast Legal Judgment Prediction on European Court of Human Rights cases into an article-aware classification task, where the case outcome is classified from a combined input of case facts and convention articles. This configuration facilitates the model learning some legal reasoning ability in mapping article text to specific case fact text. It also provides an opportunity to evaluate the model's ability to generalize to zero-shot settings when asked to classify the case outcome with respect to articles not seen during training. We devise zero-shot experiments and apply domain adaptation methods based on domain discrimination and Wasserstein distance. Our results demonstrate that the article-aware architecture outperforms straightforward fact classification. We also find that domain adaptation methods improve zero-shot transfer performance, with article relatedness and encoder pre-training influencing the effect.
This work seeks to remedy this incomplete inference and enable the model to learn more authentic reasoning between rules and case facts by casting LJP into an article-aware classification setting and subjecting it to a zero-shot transfer challenge. Article-aware classification has been explored on Chinese criminal case corpora (Wang et al., 2018, 2019b; Yue et al., 2021; Chen et al., 2022). Similarly, Holzenberger et al. 2020 has modeled statutory reasoning by classifying US tax law provisions concatenated with textual case descriptions. We build on this prior work in two ways. First, we develop and evaluate our model on a public dataset (Chalkidis et al., 2022b) of cases by the European Court of Human Rights (ECtHR), which hears complaints by individuals about possible infringements of their rights enshrined in the European Convention on Human Rights (ECHR) by states. To the best of our knowledge, this is the first work applying article-aware case outcome prediction setting to human rights adjudication. Our approach pairs case fact descriptions with candidate ECHR articles and assigns a binary target label depending on whether the article has been alleged/deemed to have been violated, or not. Our results show that the article-aware classification model outperforms the traditional classification setup by a small but consistent margin.
### 1 Introduction
Legal Judgment Prediction (LJP) has recently gained considerable attention in the mainstream NLP community (e.g., Aletras et al.2016; Chalkidis et al. 2019, 2021, 2022b; Santosh et al. 2022, 2023). In LJP, the outcome of a case should be classified/predicted based on a textual description of case facts. In actual legal reasoning, legal practitioners (e.g., advocates, judges) determine relevant rules from the sources of law (e.g., statutes, regulations, precedent) that are relevant to the case at hand. They then carry out an analysis to determine which rules apply to the case at hand, and deduce the outcome of the case by applying them. Subsuming case facts under elements of rules given in legal sources plays a critical role in this process. Many current LJP approaches (e.g., Aletras et al. 2016; Chalkidis et al. 2019, 2022b; Santosh et al. 2023) tackle this as a straightforward classification problem with the textual descriptions of case fact as
|
[
[
0.1435546875,
0.08612060546875,
0.85546875,
0.08612060546875,
0.85546875,
0.1009521484375,
0.1435546875,
0.1009521484375
],
[
0.29296875,
0.105712890625,
0.7080078125,
0.105712890625,
0.7080078125,
0.12054443359375,
0.29296875,
0.12054443359375
],
[
0.279541015625,
0.1318359375,
0.72705078125,
0.1318359375,
0.72705078125,
0.146728515625,
0.279541015625,
0.146728515625
],
[
0.27783203125,
0.1478271484375,
0.7255859375,
0.149658203125,
0.7255859375,
0.1644287109375,
0.27783203125,
0.162109375
],
[
0.328369140625,
0.1650390625,
0.67333984375,
0.166259765625,
0.67333984375,
0.18115234375,
0.328369140625,
0.179931640625
],
[
0.16455078125,
0.18115234375,
0.8388671875,
0.1817626953125,
0.8388671875,
0.196533203125,
0.16455078125,
0.1953125
],
[
0.28955078125,
0.198974609375,
0.7138671875,
0.1995849609375,
0.7138671875,
0.21435546875,
0.28955078125,
0.213134765625
],
[
0.343505859375,
0.2161865234375,
0.6591796875,
0.2161865234375,
0.6591796875,
0.2298583984375,
0.343505859375,
0.2298583984375
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51123046875,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51123046875,
0.267333984375
],
[
0.51123046875,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.145263671875,
0.279052734375,
0.459228515625,
0.279052734375,
0.459228515625,
0.29345703125,
0.145263671875,
0.29345703125
],
[
0.51318359375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.145263671875,
0.29345703125,
0.45751953125,
0.29345703125,
0.45751953125,
0.30712890625,
0.145263671875,
0.30712890625
],
[
0.145263671875,
0.30712890625,
0.459228515625,
0.30810546875,
0.459228515625,
0.32177734375,
0.145263671875,
0.320556640625
],
[
0.51318359375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51318359375,
0.31640625
],
[
0.0235137939453125,
0.31005859375,
0.060455322265625,
0.31005859375,
0.060455322265625,
0.43408203125,
0.0235137939453125,
0.43408203125
],
[
0.1435546875,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.1435546875,
0.336181640625
],
[
0.509765625,
0.317626953125,
0.8818359375,
0.318359375,
0.8818359375,
0.333251953125,
0.509765625,
0.33203125
],
[
0.1485595703125,
0.33837890625,
0.45751953125,
0.33837890625,
0.45751953125,
0.3486328125,
0.1485595703125,
0.3486328125
],
[
0.51318359375,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.1485595703125,
0.3515625,
0.456787109375,
0.3515625,
0.456787109375,
0.3623046875,
0.1485595703125,
0.3623046875
],
[
0.51318359375,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.14697265625,
0.366455078125,
0.456787109375,
0.366455078125,
0.456787109375,
0.377197265625,
0.14697265625,
0.377197265625
],
[
0.51318359375,
0.36767578125,
0.880859375,
0.36767578125,
0.880859375,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.14697265625,
0.377685546875,
0.45751953125,
0.377685546875,
0.45751953125,
0.391357421875,
0.14697265625,
0.391357421875
],
[
0.51123046875,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.145263671875,
0.393798828125,
0.45751953125,
0.393798828125,
0.45751953125,
0.40673828125,
0.145263671875,
0.40673828125
],
[
0.51123046875,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.51123046875,
0.413330078125
],
[
0.145263671875,
0.40673828125,
0.459228515625,
0.40673828125,
0.459228515625,
0.42041015625,
0.145263671875,
0.42041015625
],
[
0.51123046875,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.145263671875,
0.421630859375,
0.45751953125,
0.421630859375,
0.45751953125,
0.435302734375,
0.145263671875,
0.435302734375
],
[
0.03863525390625,
0.431640625,
0.0487060546875,
0.431640625,
0.0487060546875,
0.43701171875,
0.03863525390625,
0.43701171875
],
[
0.51123046875,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.14697265625,
0.43701171875,
0.45751953125,
0.43701171875,
0.45751953125,
0.44775390625,
0.14697265625,
0.44775390625
],
[
0.0218353271484375,
0.435791015625,
0.06463623046875,
0.43701171875,
0.060455322265625,
0.5185546875,
0.01763916015625,
0.51708984375
],
[
0.1485595703125,
0.451904296875,
0.45751953125,
0.451904296875,
0.45751953125,
0.4619140625,
0.1485595703125,
0.4619140625
],
[
0.509765625,
0.446533203125,
0.7421875,
0.4453125,
0.7421875,
0.460205078125,
0.509765625,
0.46142578125
],
[
0.14697265625,
0.46435546875,
0.4609375,
0.46435546875,
0.4609375,
0.47802734375,
0.14697265625,
0.47802734375
],
[
0.52880859375,
0.46728515625,
0.8818359375,
0.4697265625,
0.8818359375,
0.48388671875,
0.52880859375,
0.482177734375
],
[
0.1485595703125,
0.479736328125,
0.45751953125,
0.479736328125,
0.45751953125,
0.490478515625,
0.1485595703125,
0.490478515625
],
[
0.51318359375,
0.48681640625,
0.880859375,
0.48681640625,
0.880859375,
0.49755859375,
0.51318359375,
0.49755859375
],
[
0.14697265625,
0.49462890625,
0.456787109375,
0.49462890625,
0.456787109375,
0.5048828125,
0.14697265625,
0.5048828125
],
[
0.14697265625,
0.50830078125,
0.45751953125,
0.50830078125,
0.45751953125,
0.5185546875,
0.14697265625,
0.5185546875
],
[
0.51318359375,
0.50439453125,
0.880859375,
0.50439453125,
0.880859375,
0.51416015625,
0.51318359375,
0.51416015625
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.145263671875,
0.52099609375,
0.45751953125,
0.521484375,
0.45751953125,
0.533203125,
0.145263671875,
0.5322265625
],
[
0.51123046875,
0.5185546875,
0.87890625,
0.5185546875,
0.87890625,
0.5322265625,
0.51123046875,
0.5322265625
],
[
0.145263671875,
0.53369140625,
0.45751953125,
0.53369140625,
0.45751953125,
0.54736328125,
0.145263671875,
0.54736328125
],
[
0.51318359375,
0.53515625,
0.8818359375,
0.53515625,
0.8818359375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.145263671875,
0.5498046875,
0.459228515625,
0.5498046875,
0.459228515625,
0.5634765625,
0.145263671875,
0.5634765625
],
[
0.51318359375,
0.55029296875,
0.8818359375,
0.55029296875,
0.8818359375,
0.56396484375,
0.51318359375,
0.56396484375
],
[
0.1435546875,
0.5634765625,
0.3779296875,
0.5625,
0.3779296875,
0.57666015625,
0.1435546875,
0.57763671875
],
[
0.51123046875,
0.5654296875,
0.8818359375,
0.56640625,
0.8818359375,
0.58056640625,
0.51123046875,
0.580078125
],
[
0.51123046875,
0.58056640625,
0.880859375,
0.58056640625,
0.880859375,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.12176513671875,
0.59326171875,
0.130126953125,
0.59326171875,
0.130126953125,
0.5986328125,
0.12176513671875,
0.5986328125
],
[
0.1351318359375,
0.59326171875,
0.1427001953125,
0.59326171875,
0.1427001953125,
0.5986328125,
0.1351318359375,
0.5986328125
],
[
0.1427001953125,
0.59326171875,
0.1485595703125,
0.59326171875,
0.1485595703125,
0.5986328125,
0.1427001953125,
0.5986328125
],
[
0.1435546875,
0.5888671875,
0.26025390625,
0.5888671875,
0.26025390625,
0.60400390625,
0.1435546875,
0.60400390625
],
[
0.51318359375,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.1175537109375,
0.6142578125,
0.48779296875,
0.6142578125,
0.48779296875,
0.62744140625,
0.1175537109375,
0.62744140625
],
[
0.509765625,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.509765625,
0.62890625
],
[
0.1175537109375,
0.63134765625,
0.48779296875,
0.63134765625,
0.48779296875,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.509765625,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.509765625,
0.6435546875
],
[
0.1175537109375,
0.646484375,
0.490234375,
0.646484375,
0.490234375,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.509765625,
0.6455078125,
0.88330078125,
0.646484375,
0.88330078125,
0.6611328125,
0.509765625,
0.66015625
],
[
0.118408203125,
0.6630859375,
0.490234375,
0.6630859375,
0.490234375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.509765625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.118408203125,
0.6787109375,
0.489501953125,
0.6787109375,
0.489501953125,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.5087890625,
0.677734375,
0.8818359375,
0.6787109375,
0.8818359375,
0.693359375,
0.5087890625,
0.6923828125
],
[
0.1175537109375,
0.6953125,
0.490234375,
0.6953125,
0.490234375,
0.708984375,
0.1175537109375,
0.708984375
],
[
0.51318359375,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1158447265625,
0.71044921875,
0.48779296875,
0.71142578125,
0.48779296875,
0.7255859375,
0.1158447265625,
0.724609375
],
[
0.51318359375,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.51318359375,
0.724609375
],
[
0.1175537109375,
0.72802734375,
0.490234375,
0.72802734375,
0.490234375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.509765625,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.509765625,
0.7412109375
],
[
0.1158447265625,
0.74365234375,
0.490234375,
0.7421875,
0.490234375,
0.75732421875,
0.1158447265625,
0.75830078125
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51318359375,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51123046875,
0.7880859375
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1175537109375,
0.8056640625
],
[
0.5087890625,
0.79052734375,
0.8818359375,
0.79150390625,
0.8818359375,
0.8056640625,
0.5087890625,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51318359375,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51318359375,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1175537109375,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51123046875,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.8681640625,
0.51123046875,
0.8681640625
],
[
0.1175537109375,
0.87158203125,
0.490234375,
0.87158203125,
0.490234375,
0.88525390625,
0.1175537109375,
0.88525390625
],
[
0.5087890625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.5087890625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90185546875,
0.1175537109375,
0.90185546875
],
[
0.5087890625,
0.88720703125,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.5087890625,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.6484375,
0.90478515625,
0.6484375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"Zero-shot Transfer of Article-aware Legal Outcome Classification for",
"European Court of Human Rights Cases",
"Santosh T.Y.S.S', Oana Ichim?, Matthias Grabmairl",
"1School of Computation, Information, and Technology;",
"Technical University of Munich, Germany",
"2Graduate Institute of International and Development Studies, Geneva, Switzerland",
"(santosh.tokala,matthias.grabmairj@tum.de",
"oana.ichim@graduateinstitute.ch",
"Abstract",
"the sole input. This reliance on the model learning",
"statistical correspondences from case fact descrip",
"In this paper, we cast Legal Judgment Pre-",
"tions directly to outcomes neglects the role of legal",
"diction on European Court of Human Rights",
"cases into an article-aware classification task,",
"sources in this relationship. As a consequence",
"",
"where the case outcome is classified from",
"the model may learn sub-optimal fact-outcome pat-",
"acombinedinputof casefacts andconven-",
"terns that are informed by the case distribution in",
"tion articles.This configuration facilitates the",
"the data rather than learning to align facts with the",
"model learning somelegal reasoning abilityin",
"legal source text containing applicable rules. The",
"mapping article text to specific case fact text.",
"models may also attend to outcome-correlating dis-",
"It also provides an opportunity to evaluate the",
"tractors present in the dataset rather than engage in",
"model's ability to generalize to zero-shot set-",
"the legal fact-vs-law reasoning that is required of",
"tings when asked to classify the case outcome",
"",
"legal practitioners for a proper justification of the",
"with respect to articles not seen during train-",
"[T's]",
"ing.Wedevise zero-shotexperiments and ap-",
"outcome (Santosh et al., 2022).",
"ply domain adaptation methods based on do-",
"This work seeks to remedy this incomplete infer-",
"main discrimination and Wasserstein distance",
"enceandenablethemodeltolearnmoreauthentic",
"Ourresultsdemonstratethatthearticle-aware",
"architectureoutperforms straightforward fact",
"reasoning between rules and case facts by cast-",
"C60900206210",
"classification.We also find that domain adap-",
"ing LJP into an article-aware classification setting",
"tation methods improve zero-shot transfer per",
"and subjecting it to a zero-shot transfer challenge",
"formance, with article relatedness and encoder",
"Article-aware classification has been explored on",
"pre-training influencing the effect.",
"Chinese criminal case corpora (Wang et al., 2018,",
"2019b: Yue et al., 2021: Chen et al., 2022). Simi-",
"",
"",
"",
"Introduction",
"larly, Holzenberger et al. 2020 has modeled statu",
"Legal Judgment Prediction (LJP) has recently",
"tory reasoning by classifying US tax law provisions",
"gained considerable attention in the mainstream",
"concatenated with textual case descriptions. We",
"NLP community (e.g., Aletras et al. 2016;",
"build on this prior work in two ways. First, we",
"Chalkidis et al. 2019, 2021, 2022b; Santosh et al.",
" develop and evaluate our model on a public dataset",
"2022, 2023). In LJP, the outcome of a case should",
"(Chalkidis et al., 2022b) of cases by the European",
"be classified/predicted based on a textual descrip-",
"Court of Human Rights (ECtHR), which hears com-",
"tion of case facts. In actual legal reasoning, legal",
"plaints by individuals about possible infringements",
"practitioners (e.g., advocates, judges) determine rel-",
"of their rights enshrined in the European Conven-",
"evant rules from the sources of law (e.g., statutes,",
"tion on Human Rights (ECHR) by states. To the",
"regulations, precedent) that are relevant to the case",
"best of our knowledge, this is the first work apply-",
"at hand. They then carry out an analysis to deter-",
"ing article-aware case outcome prediction setting",
"mine which rules apply to the case at hand, and",
"to human rights adjudication. Our approach pairs",
"deduce the outcome of the case by applying them.",
"case fact descriptions with candidate ECHR arti",
"Subsuming case facts under elements of rules given",
"cles and assigns a binary target label depending",
"in legal sources plays a critical role in this process.",
"on whether the article has been alleged/deemed to",
"Many current LJP approaches (e.g., Aletras et al.",
"havebeen violated.ornot. Ourresults show that",
"2016; Chalkidis et al. 2019, 2022b; Santosh et al.",
" the article-aware classification model outperforms",
"2023) tackle this as a straightforward classification",
"the traditional classification setup by a small but",
"problem with the textual descriptions of case fact as",
"consistent margin."
] |
{
"arxiv_id": "2302.00609",
"language": "en",
"page_idx": 0
}
|
||
2302.00609-en-1
|
2302.00609-en-1.jpg
|
Second, we subject the model to a zero-shot transfer task. Models trained on case facts alone cannot produce inferences about convention articles they did not observe during training. By contrast, human judges can conduct outcome analysis with new/amended legal provisions because they are trained to understand the rules they contain and apply them to case facts in an expertise-informed way, even in the absence of secondary sources (e.g., commentaries to the rule, etc.). Article-aware classification allows an emulation of this process by means of a zero-shot benchmarking task on articles unseen at training time. We compare two conditions where (1) the model either has no access to the target articles, or (2) it is allowed to `read' the target articles but is not given any prediction outcome labels for case-target article pairs.
We experiment with domain adaptation by means of a domain discriminator (Ganin et al., 2016) and Wasserstein distance (Shen et al., 2018). Our results show that this improves performance on unseen articles compared to a vanilla model. We study the impact of law-specific pre-trained encoders on this zero-shot transferability compared to the standard language pre-trained one. Intuitively, we observe that our models perform better in zeroshot transfer if the target/unseen articles are semantically related to articles seen at training time.
It should be noted that, despite these tasks being typically referred to as instances of 'legal judgment prediction', ECtHR fact statements are typically not finalized until the decision outcome is known, making the task effectively one of retrospective classification rather than prediction (Medvedeva et al., 2021). While this does lead to distracting and confounding phenomena (see our prior work in Santosh et al. 2022), the dataset remains a useful resource for the development of NLP models that analyze these fact statements for text patterns that correspond to specific convention articles as drafted by the court. Consequently, in this paper we hence speak of our models as engaging in case outcome $classification\left(COC\right).$
Our main contributions in this paper are$^{1}:$
·We cast LJP/COC on ECtHR cases as an article-aware classification task by pairing case fact descriptions with candidate articles. Assuming a frozen pre-trained encoder network, our article-aware prediction model out-
at
$is$
$^{1}Our$
code
available
https://github.com/TUMLegalTech/zeroshotLJP
performs straightforward fact classification.
· We conduct zero-shot transfer benchmarking of article-aware COC models. We find this to be a difficult testing task for the generalization of COC models. We show that domain adaptation using a domain discriminator and a Wasserstein distance method improves generalization.
· We conduct auxiliary experiments validating that article relatedness positively affects transfer performance and show an interaction between domain adaptation and domain specific encoder pre-training.
## 2 $\textbf{Related Work}$
Legal Judgement Prediction: LJP/COC as an NLP task has been studied using corpora from different jurisdictions, such as the ECtHR (Chalkidis et al., 2019, 2021, 2022b; Aletras et al., 2016; Liu and Chen, 2017; Medvedeva et al., 2020; SAYS, 2020; Medvedeva et al., 2021; Santosh et al., 2023) Chinese Criminal Courts (Luo et al., 2017; Zhong et al., 2018; Yang et al., 2019; Yue et al., 2021; Zhong et al., 2020), US Supreme Court (Katz et al., 2017; Kaufman et al., 2019), Indian Supreme Court (Malik et al., 2021; Shaikh et al., 2020) the French court of Cassation (Sylea et al., 2017b,a), Brazilian courts (Lage-Freitas et al., 2022), the Federal Supreme Court of Switzerland (Niklaus et al., 2021), UK courts (Strickson and De La Iglesia, 2020) and German courts (Waltl et al., 2017)
Early works (Aletras et al., 2016; Syulea et al., 2017a,b; Virtucio et al., 2018; Shaikh et al., 2020; Medvedeva et al., 2020) used bag-of-words features. More recent approaches use deep learning (Zhong et al., 2018, 2020; Yang et al., 2019). Large pre-trained transformer models have since become the dominant model family in COC/LJP (Chalkidis et al., 2019; Niklaus et al., 2021), including legal-domain specific pre-trained variants (Chalkidis et al., 2020; Zheng et al., 2021) that have been employed for the benchmark ECtHR corpus we use in this paper (Chalkidis et al. 2021, 2022b).
Prior work on Chinese criminal case corpora case extends fact-based classification by providing the text of legal source articles as additional input. Luo et al. 2017 used an attention-based neural network which jointly models charge prediction and relevant article extraction in a unified framework whose input includes the text of legal articles. Sim-
|
[
[
0.1351318359375,
0.08612060546875,
0.48779296875,
0.08612060546875,
0.48779296875,
0.09979248046875,
0.1351318359375,
0.09979248046875
],
[
0.54736328125,
0.08612060546875,
0.87158203125,
0.08489990234375,
0.87158203125,
0.09979248046875,
0.54736328125,
0.1009521484375
],
[
0.1175537109375,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.5341796875,
0.11224365234375,
0.8818359375,
0.1134033203125,
0.8818359375,
0.127685546875,
0.5341796875,
0.1270751953125
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1484375,
0.1158447265625,
0.1478271484375
],
[
0.54736328125,
0.1300048828125,
0.8818359375,
0.1300048828125,
0.8818359375,
0.1436767578125,
0.54736328125,
0.1436767578125
],
[
0.54736328125,
0.144287109375,
0.88330078125,
0.1455078125,
0.88330078125,
0.1602783203125,
0.54736328125,
0.1591796875
],
[
0.1175537109375,
0.15087890625,
0.486083984375,
0.15087890625,
0.486083984375,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.1158447265625,
0.166259765625,
0.484375,
0.166259765625,
0.484375,
0.179931640625,
0.1158447265625,
0.179931640625
],
[
0.54833984375,
0.162109375,
0.8818359375,
0.162109375,
0.8818359375,
0.17578125,
0.54833984375,
0.17578125
],
[
0.1175537109375,
0.18408203125,
0.484375,
0.18408203125,
0.484375,
0.19482421875,
0.1175537109375,
0.19482421875
],
[
0.54833984375,
0.1787109375,
0.8818359375,
0.1787109375,
0.8818359375,
0.1923828125,
0.54833984375,
0.1923828125
],
[
0.11419677734375,
0.198974609375,
0.48779296875,
0.19775390625,
0.48779296875,
0.2120361328125,
0.11419677734375,
0.213134765625
],
[
0.54736328125,
0.193603515625,
0.88330078125,
0.19482421875,
0.88330078125,
0.208984375,
0.54736328125,
0.2078857421875
],
[
0.11419677734375,
0.21435546875,
0.489501953125,
0.215576171875,
0.489501953125,
0.2298583984375,
0.11419677734375,
0.2286376953125
],
[
0.5498046875,
0.2120361328125,
0.63232421875,
0.2120361328125,
0.63232421875,
0.223876953125,
0.5498046875,
0.223876953125
],
[
0.1158447265625,
0.23095703125,
0.489501953125,
0.23095703125,
0.489501953125,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.5341796875,
0.236328125,
0.8818359375,
0.2381591796875,
0.8818359375,
0.2529296875,
0.5341796875,
0.25048828125
],
[
0.1175537109375,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.54736328125,
0.2529296875,
0.88330078125,
0.254150390625,
0.88330078125,
0.268310546875,
0.54736328125,
0.267333984375
],
[
0.1158447265625,
0.26318359375,
0.48779296875,
0.26318359375,
0.48779296875,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.54833984375,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.283935546875,
0.54833984375,
0.283935546875
],
[
0.1158447265625,
0.279052734375,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.54736328125,
0.2861328125,
0.8818359375,
0.28515625,
0.8818359375,
0.2998046875,
0.54736328125,
0.300537109375
],
[
0.118408203125,
0.29638671875,
0.486083984375,
0.29638671875,
0.486083984375,
0.30712890625,
0.118408203125,
0.30712890625
],
[
0.54833984375,
0.302734375,
0.70361328125,
0.302734375,
0.70361328125,
0.31640625,
0.54833984375,
0.31640625
],
[
0.1158447265625,
0.31005859375,
0.486083984375,
0.31005859375,
0.486083984375,
0.32373046875,
0.1158447265625,
0.32373046875
],
[
0.1158447265625,
0.327880859375,
0.490234375,
0.327880859375,
0.490234375,
0.341552734375,
0.1158447265625,
0.341552734375
],
[
0.5146484375,
0.333251953125,
0.53076171875,
0.333251953125,
0.53076171875,
0.341552734375,
0.5146484375,
0.341552734375
],
[
0.5400390625,
0.330078125,
0.6650390625,
0.330078125,
0.6650390625,
0.344482421875,
0.5400390625,
0.344482421875
],
[
0.1158447265625,
0.34326171875,
0.41064453125,
0.34326171875,
0.41064453125,
0.35693359375,
0.1158447265625,
0.35693359375
],
[
0.1351318359375,
0.359375,
0.48779296875,
0.359375,
0.48779296875,
0.37353515625,
0.1351318359375,
0.37353515625
],
[
0.509765625,
0.355224609375,
0.8818359375,
0.355224609375,
0.8818359375,
0.369384765625,
0.509765625,
0.369384765625
],
[
0.1158447265625,
0.3759765625,
0.489501953125,
0.3759765625,
0.489501953125,
0.3896484375,
0.1158447265625,
0.3896484375
],
[
0.509765625,
0.372314453125,
0.88330078125,
0.372314453125,
0.88330078125,
0.385986328125,
0.509765625,
0.385986328125
],
[
0.1158447265625,
0.39013671875,
0.48779296875,
0.39013671875,
0.48779296875,
0.40380859375,
0.1158447265625,
0.40380859375
],
[
0.509765625,
0.388427734375,
0.8818359375,
0.38720703125,
0.8818359375,
0.402099609375,
0.509765625,
0.402587890625
],
[
0.1175537109375,
0.407958984375,
0.48779296875,
0.407958984375,
0.48779296875,
0.421630859375,
0.1175537109375,
0.421630859375
],
[
0.509765625,
0.405029296875,
0.8818359375,
0.405029296875,
0.8818359375,
0.418701171875,
0.509765625,
0.418701171875
],
[
0.1158447265625,
0.42333984375,
0.48779296875,
0.42333984375,
0.48779296875,
0.43701171875,
0.1158447265625,
0.43701171875
],
[
0.509765625,
0.42041015625,
0.8818359375,
0.42041015625,
0.8818359375,
0.43408203125,
0.509765625,
0.43408203125
],
[
0.11419677734375,
0.438232421875,
0.49365234375,
0.439453125,
0.49365234375,
0.45361328125,
0.11419677734375,
0.453125
],
[
0.51318359375,
0.43701171875,
0.88330078125,
0.43701171875,
0.88330078125,
0.45068359375,
0.51318359375,
0.45068359375
],
[
0.1175537109375,
0.4560546875,
0.490234375,
0.4560546875,
0.490234375,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.51123046875,
0.451904296875,
0.8818359375,
0.453125,
0.8818359375,
0.46728515625,
0.51123046875,
0.466064453125
],
[
0.1158447265625,
0.47265625,
0.490234375,
0.47265625,
0.490234375,
0.486328125,
0.1158447265625,
0.486328125
],
[
0.51123046875,
0.4697265625,
0.88330078125,
0.4697265625,
0.88330078125,
0.482666015625,
0.51123046875,
0.482666015625
],
[
0.11419677734375,
0.488037109375,
0.491943359375,
0.488037109375,
0.491943359375,
0.501953125,
0.11419677734375,
0.501953125
],
[
0.5087890625,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.5087890625,
0.498779296875
],
[
0.1158447265625,
0.50439453125,
0.49365234375,
0.50439453125,
0.49365234375,
0.51708984375,
0.1158447265625,
0.51708984375
],
[
0.51123046875,
0.501953125,
0.87890625,
0.501953125,
0.87890625,
0.515625,
0.51123046875,
0.515625
],
[
0.11419677734375,
0.52099609375,
0.456787109375,
0.52099609375,
0.456787109375,
0.53369140625,
0.11419677734375,
0.53369140625
],
[
0.51318359375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.133544921875,
0.53759765625,
0.486083984375,
0.53759765625,
0.486083984375,
0.55029296875,
0.133544921875,
0.55029296875
],
[
0.51318359375,
0.53369140625,
0.8818359375,
0.53369140625,
0.8818359375,
0.54736328125,
0.51318359375,
0.54736328125
],
[
0.1158447265625,
0.552734375,
0.48779296875,
0.5517578125,
0.48779296875,
0.56640625,
0.1158447265625,
0.56787109375
],
[
0.51123046875,
0.548828125,
0.880859375,
0.548828125,
0.880859375,
0.5625,
0.51123046875,
0.5625
],
[
0.1158447265625,
0.568359375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58203125,
0.1158447265625,
0.5830078125
],
[
0.51318359375,
0.5654296875,
0.88330078125,
0.5654296875,
0.88330078125,
0.580078125,
0.51318359375,
0.580078125
],
[
0.1158447265625,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.1158447265625,
0.5986328125
],
[
0.51318359375,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.51318359375,
0.595703125
],
[
0.1175537109375,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.61328125,
0.1175537109375,
0.61328125
],
[
0.51123046875,
0.59716796875,
0.8447265625,
0.59716796875,
0.8447265625,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.118408203125,
0.6181640625,
0.486083984375,
0.6181640625,
0.486083984375,
0.62890625,
0.118408203125,
0.62890625
],
[
0.5283203125,
0.61328125,
0.8818359375,
0.6142578125,
0.8818359375,
0.62890625,
0.5283203125,
0.62744140625
],
[
0.1158447265625,
0.6318359375,
0.486083984375,
0.6318359375,
0.486083984375,
0.6455078125,
0.1158447265625,
0.6455078125
],
[
0.51123046875,
0.6298828125,
0.88330078125,
0.62890625,
0.88330078125,
0.6435546875,
0.51123046875,
0.64453125
],
[
0.1175537109375,
0.6494140625,
0.489501953125,
0.6494140625,
0.489501953125,
0.6630859375,
0.1175537109375,
0.6630859375
],
[
0.509765625,
0.646484375,
0.88330078125,
0.646484375,
0.88330078125,
0.66015625,
0.509765625,
0.66015625
],
[
0.1158447265625,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.1158447265625,
0.677734375
],
[
0.50732421875,
0.6611328125,
0.8818359375,
0.662109375,
0.8818359375,
0.6767578125,
0.50732421875,
0.67578125
],
[
0.1158447265625,
0.681640625,
0.489501953125,
0.681640625,
0.489501953125,
0.6953125,
0.1158447265625,
0.6953125
],
[
0.5087890625,
0.6787109375,
0.88330078125,
0.677734375,
0.88330078125,
0.6923828125,
0.5087890625,
0.693359375
],
[
0.1158447265625,
0.69873046875,
0.489501953125,
0.69873046875,
0.489501953125,
0.7119140625,
0.1158447265625,
0.7119140625
],
[
0.51318359375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1158447265625,
0.71337890625,
0.486083984375,
0.71337890625,
0.486083984375,
0.72705078125,
0.1158447265625,
0.72705078125
],
[
0.51123046875,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.72265625,
0.51123046875,
0.72265625
],
[
0.1158447265625,
0.72998046875,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.1158447265625,
0.74462890625
],
[
0.51123046875,
0.72705078125,
0.88330078125,
0.72705078125,
0.88330078125,
0.74072265625,
0.51123046875,
0.74072265625
],
[
0.1175537109375,
0.7451171875,
0.489501953125,
0.7451171875,
0.489501953125,
0.76025390625,
0.1175537109375,
0.76025390625
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.11419677734375,
0.76123046875,
0.273681640625,
0.76123046875,
0.273681640625,
0.77490234375,
0.11419677734375,
0.77490234375
],
[
0.51318359375,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1318359375,
0.775390625,
0.444091796875,
0.77490234375,
0.444091796875,
0.7919921875,
0.1318359375,
0.79345703125
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.509765625,
0.79150390625,
0.8818359375,
0.79052734375,
0.8818359375,
0.8046875,
0.509765625,
0.8056640625
],
[
0.1368408203125,
0.80419921875,
0.48779296875,
0.8046875,
0.48779296875,
0.8193359375,
0.1368408203125,
0.818359375
],
[
0.52880859375,
0.80712890625,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.52880859375,
0.82080078125
],
[
0.151123046875,
0.82177734375,
0.48779296875,
0.82177734375,
0.48779296875,
0.8349609375,
0.151123046875,
0.8349609375
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.15283203125,
0.83740234375,
0.490234375,
0.83740234375,
0.490234375,
0.85107421875,
0.15283203125,
0.85107421875
],
[
0.51123046875,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.15283203125,
0.85400390625,
0.49365234375,
0.85400390625,
0.49365234375,
0.86767578125,
0.15283203125,
0.86767578125
],
[
0.509765625,
0.85693359375,
0.88330078125,
0.85693359375,
0.88330078125,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.15283203125,
0.869140625,
0.49365234375,
0.869140625,
0.49365234375,
0.8828125,
0.15283203125,
0.8828125
],
[
0.50537109375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.50537109375,
0.8857421875
],
[
0.1376953125,
0.8935546875,
0.172119140625,
0.8935546875,
0.172119140625,
0.90478515625,
0.1376953125,
0.90478515625
],
[
0.36279296875,
0.8935546875,
0.42236328125,
0.8935546875,
0.42236328125,
0.90478515625,
0.36279296875,
0.90478515625
],
[
0.509765625,
0.89013671875,
0.8818359375,
0.89013671875,
0.8818359375,
0.90087890625,
0.509765625,
0.90087890625
],
[
0.300537109375,
0.8955078125,
0.316650390625,
0.8955078125,
0.316650390625,
0.90283203125,
0.300537109375,
0.90283203125
],
[
0.214111328125,
0.89453125,
0.255126953125,
0.892578125,
0.2568359375,
0.90380859375,
0.2149658203125,
0.90625
],
[
0.47265625,
0.8955078125,
0.491943359375,
0.8955078125,
0.491943359375,
0.90478515625,
0.47265625,
0.90478515625
],
[
0.11419677734375,
0.90478515625,
0.411376953125,
0.90478515625,
0.411376953125,
0.91845703125,
0.11419677734375,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.88330078125,
0.90478515625,
0.88330078125,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Second, we subject the model to a zero-shot",
"performs straightforward fact classification.",
"transfer task. Models trained on case facts alone",
". We conduct zero-shot transfer benchmarking",
"cannot produce inferences about convention arti-",
"cles they did not observe during training. By con-",
"of article-aware COC models. We find this to",
"be a difficult testing task for the generaliza-",
"trast, human judges can conduct outcome analysis",
"with new/amended legal provisions because they",
"tion of COC models. We show that domain",
"are trained to understand the rules they contain and",
"adaptation using a domain discriminator and",
"apply them to case facts in an expertise-informed",
"a Wasserstein distance method improves gen-",
"way, even in the absence of secondary sources (e.g.,",
"eralization.",
"commentaries to the rule, etc.). Article-aware clas-",
". We conduct auxiliary experiments validating",
"sification allows an emulation of this process by",
"that article relatedness positively affects trans-",
"means of a zero-shot benchmarking task on articles",
"fer performance and show an interaction be-",
"unseen at training time. We compare two condi-",
"tween domain adaptation and domain specific",
"tions where (1)the modeleitherhas no accessto",
"encoder pre-training.",
"the target articles, or (2) it is allowed to ‘read' the",
"target articles but is not given any prediction out-",
"2",
"Related Work",
"come labels for case-target article pairs.",
"We experiment with domain adaptation by",
"Legal Judgement Prediction: LJP/COC as an",
"means of a domain discriminator (Ganin et al.,",
"NLP task has been studied using corpora from dif-",
"2016) and Wasserstein distance (Shen et al.. 2018)",
"ferent jurisdictions, such as the ECtHR (Chalkidis",
"Our results show that this improves performance",
"et al., 2019, 2021, 2022b; Aletras et al., 2016; Liu",
"on unseen articles compared to a vanilla model",
"and Chen, 2017; Medvedeva et al., 2020; SAYS",
"We study the impact of law-specific pre-trained en-",
"2020; Medvedeva et al., 2021; Santosh et al., 2023)",
"coders on this zero-shot transferability compared to",
"Chinese Criminal Courts (Luo et al., 2017; Zhong",
"the standard language pre-trained one. Intuitively,",
"et al., 2018; Yang et al., 2019; Yue et al., 2021;",
" we observe that our models perform better in zero-",
"Zhong et al., 2020), US Supreme Court (Katz",
"shot transfer if the target/unseen articles are seman-",
"et al., 2017; Kaufman et al., 2019), Indian Supreme",
"tically related to articles seen at training time.",
"Court (Malik et al., 2021; Shaikh et al., 2020) the",
"It should be noted that, despite these tasks being",
"French court of Cassation (Sulea et al., 2017b,a)",
"typically referred to as instances of 'legal judgment",
"Brazilian courts (Lage-Freitas et al., 2022), the Fed-",
"prediction', ECtHR fact statements are typically",
"eral Supreme Court of Switzerland (Niklaus et al.,",
"not finalized until the decision outcome is known,",
"2021), UK courts (Strickson and De La Iglesia,",
"making the task effectively one of retrospective",
"2020) and German courts (Waltl et al., 2017)",
"classificationratherthan prediction(Medvedeva",
"Early works (Aletras et al., 2016; Sulea et al.,",
"et al., 2021). While this does lead to distracting",
"2017a,b; Virtucio et al., 2018; Shaikh et al., 2020;",
"and confounding phenomena (see our prior work",
"Medvedeva et al., 2020) used bag-of-words fea-",
"in Santosh et al. 2022), the dataset remains a useful",
"tures. More recent approaches use deep learn-",
"resource for the development of NLP models that",
"ing (Zhong et al., 2018, 2020; Yang et al., 2019).",
"analyze these fact statements for text patterns that",
"Large pre-trained transformer models have since",
"correspond to specific convention articles as drafted",
"become the dominant model family in COC/LJP",
"by the court. Consequently, in this paper we hence",
"(Chalkidis et al., 2019; Niklaus et al., 2021), in-",
"speak of our models as engaging in case outcome",
"cluding legal-domain specific pre-trained variants",
"classification (CoC).",
"(Chalkidis et al., 2020; Zheng et al., 2021) that have",
"Our main contributions in this paper are':",
"been employed for the benchmark ECtHR corpus",
"we use in this paper (Chalkidis et al. 2021, 2022b)",
": We cast LJP/COC on ECtHR cases as an",
"Prior work on Chinese criminal case corpora",
"article-aware classification task by pairing",
"case extends fact-based classification by providing",
"case fact descriptions with candidate articles.",
"the text of legal source articles as additional input.",
"Assuming a frozen pre-trained encoder net-",
"Luo et al. 2017 used an attention-based neural net-",
"work, our article-aware prediction model out-",
"work which jointly models charge prediction and",
"lOur",
"available",
"relevant articleextraction in a unified framework",
"is",
"code",
"at",
"https://github.com/TUMLegalTech/zeroshotLJP",
"whose input includes the text of legal articles. Sim-"
] |
{
"arxiv_id": "2302.00609",
"language": "en",
"page_idx": 1
}
|
||
2302.00609-en-2
|
2302.00609-en-2.jpg
|
ilarly, Wang et al. 2018, 2019b; Chen et al. 2022; Yue et al. 2021 employ matching mechanism between case facts and article texts. To the best of our knowledge, ours is the first work to adapt articleaware prediction to the ECtHR corpus, which is situated in the in human rights litigation domain. Going beyond previous works, we further benchmark the zero-shot transfer performance of such models, providing a test bed to evaluate their capability to process article texts they have not seen during training time and applying them to case facts towards classifying allegations/outcomes.
Domain Adaptation (DA): In transfer learning, the field of domain adaptation (DA) addresses the covariate shift between source and target data distributions (Ruder, 2019). It is tackled under three different settings: (1) Semi-supervised DA (Bollegala et al., 2011; Daume III and Marcu, 2006) where labels for the source and a small set of labels for the target domain are available, (2) unsupervised DA (Ganin et al., 2016; Blitzer et al., 2006) where only labels for the source domain and unlabelled target data are given, and (3) Any Domain Adaptation / Out of Distribution generalization (Ben-David et al., 2022; Volk et al., 2022) where only labeled source data is given. In this work, we distill the existing public LexGLUE ECtHR dataset into a new benchmark on more challenging unsupervised and any domain adaptation settings for COC to emulate legal reasoning involving previously unseen convention articles.
DA variants have been benchmarked for various NLP tasks, such as Question answering (Yu et al., 2018), duplicate question detection (Shah et al., 2018), sentiment analysis (Li et al., 2017; Ganin et al., 2016), dependency parsing (Sato et al., 2017), relation extraction (Wu et al., 2017), POS tagging (Yasunaga et al., 2018), named entity recognition (Jia et al., 2019), event trigger identification (Naik and Rose, 2020), machine reading comprehension (Wang et al., 2019a), and machine translation ( Yang et al., 2018). To the best of our knowledge, this work is the first to benchmark domain adaptation in COC/LJP. While previous works typically involve short text, COC on ECtHR data involves case facts and articles, both of which typically are long documents.
Methods proposed for domain adaptation can be categorized into four types: (a) Instance-based data selection methods (Jiang and Zhai, 2007; Remus, 2012) which employ similarity metrics to
sample source data points to match the distribution of the target domain and train models based on obtained subsamples from the source domain, (b) Pseudo-labeling approaches (Ruder and Plank, 2018; Rotman and Reichart, 2019) which train a classifier based on source data initially and use it to predict labels on unlabeled target data towards further adapting the model, (c) Pivot-based methods (Blitzer et al., 2006; Ziser and Reichart, 2017) which aim to map different domains to a common latent space (where the feature distributions are close) by employing auto encoders and structural correspondence learning, and (d) Loss-based methods (Ganin and Lempitsky, 2015; Shen et al., 2018) which employ domain adversaries aiming to minimize the discrepancies between source and target data distributions. In this woork, we employ lossbased approaches using a domain discriminator (Ganin et al., 2016) and Wasserstein distance (Shen et al., 2018) to enable domain adaptation for our COC models.
# 3 Dataset, Tasks & Settings
We use the LexGLUE ECtHR dataset provided by (Chalkidis et al., 2022b), which consists of 11k case fact descriptions along with target label information about which convention articles have been alleged to be violated (task B), and which the court has eventually found to have been violated (task A). The dataset is chronologically split into training (2001-2016),validation (2016-2017), and test set (2017-2019) with 9k, 1k, and 1k cases, respectively. The label set includes 10 prominent ECHR articles, which forms a subset of all the rights contained in the convention and its protocols. In both the ECtHR A and B benchmarks, it is assumed that the model classified the target from the fact description alone, which we refer to as the fact classification variant.
For our article-aware classification settings, we augment the dataset with the texts of the 10 articles copied from the publicly available ECHR convention document$^2.$ We formulate the article-aware prediction variant for both tasks: Given both the case fact statements and a particular article information, the model should classify the binary outcome of whether an article has been alleged to be violated by the claimant (task B) or found to have been violated by the court (task A).
$^2$https: / / www. echr. coe. int/ documents/
convention_eng.pdf
|
[
[
0.1158447265625,
0.08489990234375,
0.490234375,
0.0843505859375,
0.490234375,
0.09857177734375,
0.1158447265625,
0.09979248046875
],
[
0.486083984375,
0.09027099609375,
0.490234375,
0.09027099609375,
0.490234375,
0.0955810546875,
0.486083984375,
0.0955810546875
],
[
0.48779296875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.48779296875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.491943359375,
0.1015625,
0.491943359375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.51123046875,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.114013671875,
0.51123046875,
0.114013671875
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.490234375,
0.1341552734375,
0.490234375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51318359375,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.1175537109375,
0.149658203125,
0.489501953125,
0.149658203125,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51318359375,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.163330078125,
0.51318359375,
0.163330078125
],
[
0.1175537109375,
0.1650390625,
0.489501953125,
0.1650390625,
0.489501953125,
0.1787109375,
0.1175537109375,
0.1787109375
],
[
0.51318359375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.118408203125,
0.18408203125,
0.489501953125,
0.18408203125,
0.489501953125,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.509765625,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.198974609375,
0.8818359375,
0.198974609375,
0.8818359375,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.215576171875,
0.490234375,
0.215576171875,
0.490234375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.51123046875,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.118408203125,
0.2327880859375,
0.48779296875,
0.2327880859375,
0.48779296875,
0.243408203125,
0.118408203125,
0.243408203125
],
[
0.509765625,
0.23095703125,
0.880859375,
0.23095703125,
0.880859375,
0.24462890625,
0.509765625,
0.24462890625
],
[
0.1175537109375,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.51123046875,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.1158447265625,
0.26318359375,
0.426513671875,
0.26318359375,
0.426513671875,
0.27783203125,
0.1158447265625,
0.27783203125
],
[
0.51123046875,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.1175537109375,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.1175537109375,
0.297607421875
],
[
0.51123046875,
0.27978515625,
0.88330078125,
0.27978515625,
0.88330078125,
0.29345703125,
0.51123046875,
0.29345703125
],
[
0.1158447265625,
0.300537109375,
0.484375,
0.300537109375,
0.484375,
0.311279296875,
0.1158447265625,
0.311279296875
],
[
0.51318359375,
0.295654296875,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.51318359375,
0.308837890625
],
[
0.1158447265625,
0.3154296875,
0.490234375,
0.314208984375,
0.490234375,
0.328857421875,
0.1158447265625,
0.330078125
],
[
0.51123046875,
0.31005859375,
0.880859375,
0.31005859375,
0.880859375,
0.32373046875,
0.51123046875,
0.32373046875
],
[
0.509765625,
0.325927734375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.509765625,
0.339111328125
],
[
0.1175537109375,
0.33203125,
0.490234375,
0.33203125,
0.490234375,
0.345703125,
0.1175537109375,
0.345703125
],
[
0.51318359375,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.118408203125,
0.349853515625,
0.48779296875,
0.349853515625,
0.48779296875,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.1158447265625,
0.363525390625,
0.48779296875,
0.3623046875,
0.48779296875,
0.377197265625,
0.1158447265625,
0.377685546875
],
[
0.51123046875,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.51123046875,
0.37353515625
],
[
0.1158447265625,
0.37890625,
0.486083984375,
0.37890625,
0.486083984375,
0.392578125,
0.1158447265625,
0.392578125
],
[
0.51318359375,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.38720703125,
0.51318359375,
0.38720703125
],
[
0.509765625,
0.39013671875,
0.8818359375,
0.391357421875,
0.8818359375,
0.40625,
0.509765625,
0.405029296875
],
[
0.1158447265625,
0.396728515625,
0.48779296875,
0.396728515625,
0.48779296875,
0.410400390625,
0.1158447265625,
0.410400390625
],
[
0.509765625,
0.40625,
0.61376953125,
0.407958984375,
0.61376953125,
0.4228515625,
0.509765625,
0.42041015625
],
[
0.1175537109375,
0.412109375,
0.48779296875,
0.412109375,
0.48779296875,
0.42578125,
0.1175537109375,
0.42578125
],
[
0.1158447265625,
0.427001953125,
0.48779296875,
0.4287109375,
0.48779296875,
0.443603515625,
0.1158447265625,
0.441162109375
],
[
0.5146484375,
0.441162109375,
0.5283203125,
0.441162109375,
0.5283203125,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.53662109375,
0.441162109375,
0.54150390625,
0.441162109375,
0.54150390625,
0.446533203125,
0.53662109375,
0.446533203125
],
[
0.5380859375,
0.43701171875,
0.76171875,
0.43701171875,
0.76171875,
0.451904296875,
0.5380859375,
0.451904296875
],
[
0.1158447265625,
0.44482421875,
0.48779296875,
0.443603515625,
0.48779296875,
0.457763671875,
0.1158447265625,
0.458984375
],
[
0.1158447265625,
0.46142578125,
0.48779296875,
0.46142578125,
0.48779296875,
0.474365234375,
0.1158447265625,
0.474365234375
],
[
0.509765625,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.47802734375,
0.509765625,
0.47802734375
],
[
0.1175537109375,
0.476806640625,
0.48779296875,
0.476806640625,
0.48779296875,
0.490478515625,
0.1175537109375,
0.490478515625
],
[
0.5087890625,
0.478515625,
0.8818359375,
0.479736328125,
0.8818359375,
0.49462890625,
0.5087890625,
0.493408203125
],
[
0.1158447265625,
0.4921875,
0.48779296875,
0.4912109375,
0.48779296875,
0.505859375,
0.1158447265625,
0.50732421875
],
[
0.51123046875,
0.496337890625,
0.8818359375,
0.496337890625,
0.8818359375,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.1175537109375,
0.5087890625,
0.489501953125,
0.5087890625,
0.489501953125,
0.5224609375,
0.1175537109375,
0.5224609375
],
[
0.51123046875,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52490234375,
0.51123046875,
0.52490234375
],
[
0.1175537109375,
0.52685546875,
0.486083984375,
0.52685546875,
0.486083984375,
0.53759765625,
0.1175537109375,
0.53759765625
],
[
0.509765625,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.509765625,
0.5419921875
],
[
0.11419677734375,
0.54052734375,
0.489501953125,
0.54052734375,
0.489501953125,
0.55419921875,
0.11419677734375,
0.55419921875
],
[
0.51123046875,
0.54443359375,
0.8818359375,
0.5458984375,
0.8818359375,
0.55859375,
0.51123046875,
0.55810546875
],
[
0.1158447265625,
0.5556640625,
0.48779296875,
0.55712890625,
0.48779296875,
0.57177734375,
0.1158447265625,
0.57080078125
],
[
0.509765625,
0.56005859375,
0.8818359375,
0.56103515625,
0.8818359375,
0.576171875,
0.509765625,
0.57470703125
],
[
0.1158447265625,
0.57373046875,
0.264404296875,
0.57373046875,
0.264404296875,
0.58740234375,
0.1158447265625,
0.58740234375
],
[
0.51318359375,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.133544921875,
0.5927734375,
0.486083984375,
0.5927734375,
0.486083984375,
0.60546875,
0.133544921875,
0.60546875
],
[
0.51123046875,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.51123046875,
0.60693359375
],
[
0.1158447265625,
0.60791015625,
0.489501953125,
0.609375,
0.489501953125,
0.62353515625,
0.1158447265625,
0.62255859375
],
[
0.5087890625,
0.60791015625,
0.88330078125,
0.609375,
0.88330078125,
0.62353515625,
0.5087890625,
0.62255859375
],
[
0.1158447265625,
0.62451171875,
0.489501953125,
0.6259765625,
0.489501953125,
0.64013671875,
0.1158447265625,
0.6396484375
],
[
0.51123046875,
0.626953125,
0.880859375,
0.626953125,
0.880859375,
0.63720703125,
0.51123046875,
0.63720703125
],
[
0.1158447265625,
0.64111328125,
0.486083984375,
0.64111328125,
0.486083984375,
0.65478515625,
0.1158447265625,
0.65478515625
],
[
0.5087890625,
0.64013671875,
0.880859375,
0.64013671875,
0.880859375,
0.65380859375,
0.5087890625,
0.65380859375
],
[
0.1158447265625,
0.65673828125,
0.489501953125,
0.65771484375,
0.489501953125,
0.6728515625,
0.1158447265625,
0.67138671875
],
[
0.51318359375,
0.65673828125,
0.8818359375,
0.65673828125,
0.8818359375,
0.67041015625,
0.51318359375,
0.67041015625
],
[
0.1158447265625,
0.67333984375,
0.48779296875,
0.6748046875,
0.48779296875,
0.689453125,
0.1158447265625,
0.6884765625
],
[
0.509765625,
0.67333984375,
0.8818359375,
0.67333984375,
0.8818359375,
0.68701171875,
0.509765625,
0.68701171875
],
[
0.1175537109375,
0.6904296875,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.1175537109375,
0.70361328125
],
[
0.5087890625,
0.6884765625,
0.8818359375,
0.689453125,
0.8818359375,
0.70361328125,
0.5087890625,
0.703125
],
[
0.1158447265625,
0.705078125,
0.48779296875,
0.7060546875,
0.48779296875,
0.72021484375,
0.1158447265625,
0.7197265625
],
[
0.509765625,
0.7060546875,
0.5751953125,
0.7060546875,
0.5751953125,
0.72021484375,
0.509765625,
0.72021484375
],
[
0.1175537109375,
0.72265625,
0.486083984375,
0.72265625,
0.486083984375,
0.736328125,
0.1175537109375,
0.736328125
],
[
0.5283203125,
0.7216796875,
0.8818359375,
0.72265625,
0.8818359375,
0.73681640625,
0.5283203125,
0.736328125
],
[
0.1175537109375,
0.7392578125,
0.48779296875,
0.7392578125,
0.48779296875,
0.7529296875,
0.1175537109375,
0.7529296875
],
[
0.51123046875,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.1175537109375,
0.7548828125,
0.489501953125,
0.7548828125,
0.489501953125,
0.7685546875,
0.1175537109375,
0.7685546875
],
[
0.509765625,
0.7548828125,
0.88330078125,
0.7548828125,
0.88330078125,
0.7685546875,
0.509765625,
0.7685546875
],
[
0.1158447265625,
0.77099609375,
0.489501953125,
0.77099609375,
0.489501953125,
0.78369140625,
0.1158447265625,
0.78369140625
],
[
0.509765625,
0.77099609375,
0.8818359375,
0.77099609375,
0.8818359375,
0.78369140625,
0.509765625,
0.78369140625
],
[
0.1158447265625,
0.78759765625,
0.49365234375,
0.78759765625,
0.49365234375,
0.80029296875,
0.1158447265625,
0.80029296875
],
[
0.509765625,
0.78759765625,
0.8818359375,
0.78759765625,
0.8818359375,
0.80029296875,
0.509765625,
0.80029296875
],
[
0.1175537109375,
0.802734375,
0.489501953125,
0.802734375,
0.489501953125,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.51123046875,
0.8046875,
0.880859375,
0.8046875,
0.880859375,
0.8154296875,
0.51123046875,
0.8154296875
],
[
0.1158447265625,
0.818359375,
0.48779296875,
0.8193359375,
0.48779296875,
0.83447265625,
0.1158447265625,
0.8330078125
],
[
0.51123046875,
0.82080078125,
0.880859375,
0.82080078125,
0.880859375,
0.83056640625,
0.51123046875,
0.83056640625
],
[
0.1158447265625,
0.8349609375,
0.2032470703125,
0.83740234375,
0.2032470703125,
0.8486328125,
0.1158447265625,
0.8466796875
],
[
0.51318359375,
0.8349609375,
0.8818359375,
0.8349609375,
0.8818359375,
0.8486328125,
0.51318359375,
0.8486328125
],
[
0.133544921875,
0.85546875,
0.48779296875,
0.85546875,
0.48779296875,
0.869140625,
0.133544921875,
0.869140625
],
[
0.51318359375,
0.8515625,
0.8818359375,
0.8515625,
0.8818359375,
0.865234375,
0.51318359375,
0.865234375
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.8681640625,
0.73291015625,
0.8681640625,
0.73291015625,
0.8818359375,
0.509765625,
0.8818359375
],
[
0.1158447265625,
0.88818359375,
0.489501953125,
0.88720703125,
0.489501953125,
0.90185546875,
0.1158447265625,
0.90283203125
],
[
0.52880859375,
0.89111328125,
0.80419921875,
0.892578125,
0.80419921875,
0.90673828125,
0.52880859375,
0.90625
],
[
0.11419677734375,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.11419677734375,
0.91845703125
],
[
0.509765625,
0.90625,
0.6533203125,
0.90625,
0.6533203125,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"ilarly, Wang et al. 2018, 2019b; Chen et al. 2022:",
"-",
"sample sourcedata points tomatchthedistribu",
"Yue et al. 2021 employ matching mechanism be-",
"tion of the target domain and train models based",
"tween case facts and article texts. To the best of our",
"on obtained subsamples from the source domain.",
"knowledge, ours is the first work to adapt article-",
"(b) Pseudo-labeling approaches (Ruder and Plank,",
"aware prediction to the ECtHR corpus, which is",
"2018; Rotman and Reichart, 2019) which train a",
"situated in the in human rights litigation domain.",
"classifier based on source data initially and use it",
"Going beyond previousworks,wefurtherbench-",
" to predict labels on unlabeled target data towards",
"mark the zero-shot transfer performance of such",
"further adapting the model, (c) Pivot-based meth-",
"models, providing a test bed to evaluate their ca-",
"ods (Blitzer et al., 2006; Ziser and Reichart, 2017)",
"pability to process articletexts theyhave not seen",
"which aim to map different domains to a common",
"during training time and applying them to case facts",
"latent space (where the feature distributions are",
"towards classifying allegations/outcomes.",
"close) by employing auto encoders and structural",
"Domain Adaptation (DA): In transfer learning,",
"correspondence learning, and (d) Loss-based meth-",
"thefield of domain adaptation(DA)addresses the",
"ods (Ganin and Lempitsky, 2015; Shen et al., 2018)",
"covariate shift between source and target data distri-",
"which employ domain adversaries aiming to mini-",
"mize the discrepancies between source and target",
"butions (Ruder, 2019). It is tackled under three dif-",
"data distributions. In this woork, we employ loss",
"ferentsettings:(1)Semi-supervisedDA(Bollegala",
"et al., 2011; Daume III and Marcu, 2006) where",
"based approaches using a domain discriminator",
"labels for the source and a small set of labels for the",
"(Ganin et al.,2016)and Wasserstein distance (Shen",
"et al., 2018) to enable domain adaptation for our",
"target domain are available, (2) unsupervised DA",
"COC models.",
"(Ganin et al., 2016; Blitzer et al., 2006) where only",
"labels for the source domain and unlabelled target",
"3",
"",
"Dataset, Tasks & Settings",
"data are given, and (3) Any Domain Adaptation",
"/ Out of Distribution generalization (Ben-David",
"We use the LexGLUE ECtHR dataset provided by",
"et al., 2022; Volk et al., 2022) where only labeled",
"(Chalkidis et al., 2022b), which consists of 11k case",
"source data is given. In this work, we distill the",
"fact descriptions along with target label information",
"existing public LexGLUE ECtHR dataset into a",
"aboutwhichconventionarticleshavebeenalleged",
"newbenchmark onmore challenging unsupervised",
"to be violated (task B), and which the court has",
"and any domain adaptation settings for COC to em-",
"eventually found to have been violated (task A)",
"ulate legal reasoning involving previously unseen",
"The dataset is chronologically split into training",
"convention articles.",
"(2001-2016), validation (2016-2017), and test set",
"DA variants have been benchmarked for various",
"(2017-2019) with 9k, 1k, and 1k cases, respectively.",
"NLP tasks, such as Question answering (Yu et al.,",
"The label set includes 10 prominent ECHR articles,",
"2018), duplicate question detection (Shah et al.,",
"which forms a subset of all the rights contained",
"2018), sentiment analysis (Li et al., 2017: Ganin",
"in the convention and its protocols. In both the",
"et al., 2016), dependency parsing (Sato et al., 2017),",
"ECtHR A and B benchmarks, it is assumed that the",
"relation extraction (Wu et al., 2017), POS tagging",
"model classified the target from the fact description",
"(Yasunaga et al., 2018), named entity recognition",
"alone, which we refer to as the fact classification",
"(Jia et al., 2019), event trigger identification (Naik",
"variant.",
"and Rose,2020),machine reading comprehension",
"For our article-aware classification settings, we",
"(Wang et al., 2019a), and machine translation (Yang",
"augment the dataset with the texts of the 10 articles",
"et al., 2018). To the best of our knowledge, this",
"copied from the publicly available ECHR conven-",
"work is the first tobenchmark domain adaptation",
"tion document?. We formulate the article-aware",
"in COC/LJP. While previous works typically in-",
"prediction variant for both tasks: Given both the",
"volve short text, COC on ECtHR data involves case",
"casefact statements and aparticulararticleinforma-",
"facts and articles, both of which typically are long",
"tion,themodel should classify thebinaryoutcome",
"documents.",
"of whether an article has been alleged to be vio",
"Methods proposed for domain adaptation can",
"lated by the claimant (task B) or found to have been",
"be categorized into four types: (a) Instance-based",
"violated by the court (task A).",
"data selection methods (Jiang and Zhai, 2007; Re-",
"2https://www.echr.coe.int/documents/",
"mus, 2012) which employ similarity metrics to",
"convention_eng.pdf"
] |
{
"arxiv_id": "2302.00609",
"language": "en",
"page_idx": 2
}
|
||
2302.00609-en-3
|
2302.00609-en-3.jpg
|
Our zero-shot transfer task then involves determining violation/allegation from case facts with respect to articles which are not seen during training time. We consider a 'domain’to be a particular convention article (i.e., 10 convention articles form 10 domains). The objective is to train a model on a source domain (seen articles) with the goal of performing well at test-time on a target domain (unseen articles). Following (Yin et al., 2019; Ramponi and Plank, 2020), we propose two settings under zero-shot COC:
Zero-Shot Restrictive / Unsupervised Domain Adaptation (UDA): In this setting, the model is given a pair of case facts and the text of training set articles (i.e., the source domain) along with their corresponding violation/allegation outcome label. In the target domain, it is provided with case facts and article text pairs as well, but the outcome label is withheld. The goal of UDA is to learn an outcome classifter from the outcome labelled source domain which should generalize well on the target domain by leveraging outcome-unlabeled target data. This setting is legally realistic, as the text of new or modified written legal sources is typically known for a given task and available for domain adaptation (e.g., a public administration decision support tool receives an update after relevant legislation has changed).
Zero-shot Wild / Any Domain Adaptation
### $(\mathbf{ADA})/\mathbf{Out~of~Distribution~Generalization:~In}$ this setting, the model never sees any article data
from the target domain during training, yet should be able to generalize to it. In the legal setting, this corresponds to a model which is required to work with texts of sources only available at query time (e.g., complex retrieval settings where multiple legal sources potentially apply).
We reorganize the dataset to evaluate our zeroshot transfer/adaptation models by splitting the 10 ECHR articles into two non-overlapping groups, such that both contain articles of various frequencies (common, moderate, rare).
$\bullet$ $split\_ 0\nobreak {: } 6, 8$,P1-1,2,9
$\bullet$ $split\_ 1\nobreak {: } 3, 5, 10, 14, 11$
We evaluate UDA and ADA on $split\_0$ as source
and $split\_1$ as target, and vice-versa.
## 4 $\mathbf{Method}$
We employ a hierarchical neural model which takes the case fact description $x$ along with the article
$a$ as input and outputs a binary outcome (allegation in Task B and violation in Task A) for case $x$ with respect to article $a.$ Our architecture is a modified version of the Enhanced Sequential Inference Model (ESIM) (Chen et al., 2017) incorporating conditional encoding (Augenstein et al., 2016; Rocktäschel et al., 2016) that has been adapted to deal with long input sequences following hierarchical attention networks (Yang et al., 2016). We experiment with two domain adaptation components based on adversarial training: (1) a classificationbased domain discriminator and (2) a Wassersteindistance based method which aims to reduce the difference between the source and the target domain distributions.
4.1 Article-aware prediction Model
Given the facts of the case $x=\{x_1,x_2,\ldots,x_m\}$ $\mathop{\mathrm{where~}}x_{i}=\{x_{i1},x_{i2},\ldots,x_{in}\}\mathop{\mathrm{and~the~article~}}a=$ $\{a_{1},a_{2},\ldots,a_{k}\}$ where $a_j=\{a_{j1},a_{j2},\ldots,a_{jl}\}$, the model outputs a binary label. $x_i/a_i$ and $x_jp$ $Ia_{jp}$ denote the $i^\mathrm{th}$ sentence and $p^\mathrm{th}$ token of the $j^{\mathrm{th}}$sentence of the case facts / article, respectively. $m/ k\textbf{ and }n/ l$ denote the number of sentences and tokens in the $i^\mathrm{th}$ sentence of case facts/ article, respectively. Our model contains an encoding layer, followed by an interaction layer, a post-interaction encoding layer, and a classification header. See Fig. 1 for an overview of our architecture.
## 4.1.1 Pre-interaction Encoding Layer
Our model encodes the facts of the case $x$ sentencewise with LegalBERT (Chalkidis et al., 2020) to obtain token level representations $\{z_{i1},z_{i2},\ldots,z_{in}\}.$ These are aggregated into sentence level representations using token attention:
(1)
$$u_{it}=\tanh(W_wz_{it}+b_w)$$
(2)
$$\alpha_{it}=\frac{\exp(u_{it}u_w)}{\sum_t\exp(u_{it}u_w)}\:\&\:f_i=\sum_{t=1}^n\alpha_{it}z_{it}$$
where $W_w,b_w$ and $u_w$ are trainable parameters. The sentence level representations $\{f_1,\ldots,f_n\}$ are passed through a GRU encoder to obtain contextaware sentence representations of the facts $h=$ $\{h_1,h_2,\ldots,h_m\}.$ The analogous article encoder takes $a$ as input and outputs $s=\{s_1,s_2,\ldots,s_k\}.$
## 4.1.2 Interaction Layer
Interaction between the sentences of the case facts and articles is done via dot product attention be-
|
[
[
0.1368408203125,
0.08612060546875,
0.49365234375,
0.08612060546875,
0.49365234375,
0.09979248046875,
0.1368408203125,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.489501953125,
0.1015625,
0.489501953125,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.5087890625,
0.1015625,
0.87890625,
0.1015625,
0.87890625,
0.11517333984375,
0.5087890625,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.5087890625,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.5087890625,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.486083984375,
0.133056640625,
0.486083984375,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.5087890625,
0.133056640625,
0.880859375,
0.133056640625,
0.880859375,
0.146728515625,
0.5087890625,
0.146728515625
],
[
0.118408203125,
0.15087890625,
0.484375,
0.15087890625,
0.484375,
0.160888671875,
0.118408203125,
0.160888671875
],
[
0.509765625,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.509765625,
0.1644287109375
],
[
0.1175537109375,
0.166259765625,
0.486083984375,
0.166259765625,
0.486083984375,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.5087890625,
0.1650390625,
0.880859375,
0.1650390625,
0.880859375,
0.1787109375,
0.5087890625,
0.1787109375
],
[
0.11419677734375,
0.182861328125,
0.48779296875,
0.1817626953125,
0.48779296875,
0.1953125,
0.11419677734375,
0.196533203125
],
[
0.51123046875,
0.18115234375,
0.880859375,
0.18115234375,
0.880859375,
0.19482421875,
0.51123046875,
0.19482421875
],
[
0.1158447265625,
0.198974609375,
0.486083984375,
0.198974609375,
0.486083984375,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.5087890625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.5087890625,
0.2120361328125
],
[
0.1175537109375,
0.21435546875,
0.489501953125,
0.21435546875,
0.489501953125,
0.22802734375,
0.1175537109375,
0.22802734375
],
[
0.51123046875,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.1175537109375,
0.23095703125,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.51318359375,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.51318359375,
0.2452392578125
],
[
0.1175537109375,
0.2464599609375,
0.28369140625,
0.2464599609375,
0.28369140625,
0.260009765625,
0.1175537109375,
0.260009765625
],
[
0.51318359375,
0.2476806640625,
0.88330078125,
0.2476806640625,
0.88330078125,
0.26123046875,
0.51318359375,
0.26123046875
],
[
0.1158447265625,
0.262451171875,
0.48779296875,
0.262451171875,
0.48779296875,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.509765625,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.509765625,
0.276611328125
],
[
0.1158447265625,
0.27783203125,
0.489501953125,
0.279052734375,
0.489501953125,
0.29345703125,
0.1158447265625,
0.292236328125
],
[
0.51123046875,
0.27783203125,
0.880859375,
0.27783203125,
0.880859375,
0.29150390625,
0.51123046875,
0.29150390625
],
[
0.1175537109375,
0.295654296875,
0.489501953125,
0.295654296875,
0.489501953125,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.51123046875,
0.295654296875,
0.88330078125,
0.295654296875,
0.88330078125,
0.308837890625,
0.51123046875,
0.308837890625
],
[
0.1175537109375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.325927734375,
0.1175537109375,
0.325927734375
],
[
0.509765625,
0.311279296875,
0.6513671875,
0.311279296875,
0.6513671875,
0.32470703125,
0.509765625,
0.32470703125
],
[
0.1175537109375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.51318359375,
0.339111328125,
0.8037109375,
0.339111328125,
0.8037109375,
0.352783203125,
0.51318359375,
0.352783203125
],
[
0.1158447265625,
0.359375,
0.48779296875,
0.358154296875,
0.48779296875,
0.372314453125,
0.1158447265625,
0.37353515625
],
[
0.51123046875,
0.35986328125,
0.880859375,
0.361083984375,
0.880859375,
0.3759765625,
0.51123046875,
0.374755859375
],
[
0.11419677734375,
0.37353515625,
0.490234375,
0.374755859375,
0.490234375,
0.3896484375,
0.11419677734375,
0.388427734375
],
[
0.51123046875,
0.377685546875,
0.875,
0.377685546875,
0.875,
0.392578125,
0.51123046875,
0.392578125
],
[
0.1175537109375,
0.391357421875,
0.48779296875,
0.391357421875,
0.48779296875,
0.405029296875,
0.1175537109375,
0.405029296875
],
[
0.51318359375,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.407958984375,
0.51318359375,
0.407958984375
],
[
0.1158447265625,
0.40673828125,
0.48779296875,
0.407958984375,
0.48779296875,
0.4228515625,
0.1158447265625,
0.421630859375
],
[
0.509765625,
0.40625,
0.8818359375,
0.407958984375,
0.8818359375,
0.42578125,
0.509765625,
0.42333984375
],
[
0.1175537109375,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.5087890625,
0.424560546875,
0.8818359375,
0.42333984375,
0.8818359375,
0.438232421875,
0.5087890625,
0.439453125
],
[
0.1175537109375,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.45361328125,
0.1175537109375,
0.45361328125
],
[
0.51318359375,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.11419677734375,
0.454833984375,
0.48779296875,
0.4560546875,
0.48779296875,
0.47021484375,
0.11419677734375,
0.4697265625
],
[
0.50537109375,
0.457275390625,
0.8818359375,
0.4560546875,
0.8818359375,
0.47021484375,
0.50537109375,
0.471435546875
],
[
0.11419677734375,
0.471435546875,
0.489501953125,
0.47021484375,
0.489501953125,
0.485107421875,
0.11419677734375,
0.486328125
],
[
0.509765625,
0.47265625,
0.8818359375,
0.47265625,
0.8818359375,
0.48681640625,
0.509765625,
0.48681640625
],
[
0.1158447265625,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.5029296875,
0.1158447265625,
0.5029296875
],
[
0.5087890625,
0.488037109375,
0.88330078125,
0.4892578125,
0.88330078125,
0.50439453125,
0.5087890625,
0.5029296875
],
[
0.1175537109375,
0.5048828125,
0.491943359375,
0.5048828125,
0.491943359375,
0.5185546875,
0.1175537109375,
0.5185546875
],
[
0.5087890625,
0.5048828125,
0.8818359375,
0.50732421875,
0.8818359375,
0.52099609375,
0.5087890625,
0.5185546875
],
[
0.1175537109375,
0.52099609375,
0.268798828125,
0.52099609375,
0.268798828125,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.51123046875,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.5361328125,
0.51123046875,
0.5361328125
],
[
0.118408203125,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.509765625,
0.5380859375,
0.78662109375,
0.5380859375,
0.78662109375,
0.5517578125,
0.509765625,
0.5517578125
],
[
0.118408203125,
0.552734375,
0.486083984375,
0.552734375,
0.486083984375,
0.56640625,
0.118408203125,
0.56640625
],
[
0.1175537109375,
0.568359375,
0.48779296875,
0.568359375,
0.48779296875,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.51318359375,
0.56396484375,
0.8193359375,
0.56396484375,
0.8193359375,
0.57763671875,
0.51318359375,
0.57763671875
],
[
0.118408203125,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.118408203125,
0.5986328125
],
[
0.509765625,
0.5830078125,
0.8818359375,
0.58447265625,
0.8818359375,
0.5986328125,
0.509765625,
0.59716796875
],
[
0.118408203125,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.61328125,
0.118408203125,
0.61328125
],
[
0.509765625,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.6142578125,
0.509765625,
0.6142578125
],
[
0.1175537109375,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.1175537109375,
0.6298828125
],
[
0.5087890625,
0.615234375,
0.8818359375,
0.6162109375,
0.8818359375,
0.63134765625,
0.5087890625,
0.6298828125
],
[
0.1158447265625,
0.6318359375,
0.489501953125,
0.6328125,
0.489501953125,
0.64794921875,
0.1158447265625,
0.646484375
],
[
0.509765625,
0.6328125,
0.8818359375,
0.6328125,
0.8818359375,
0.646484375,
0.509765625,
0.646484375
],
[
0.1175537109375,
0.6494140625,
0.490234375,
0.6494140625,
0.490234375,
0.6630859375,
0.1175537109375,
0.6630859375
],
[
0.509765625,
0.6484375,
0.72900390625,
0.6484375,
0.72900390625,
0.662109375,
0.509765625,
0.662109375
],
[
0.1158447265625,
0.66455078125,
0.340087890625,
0.6650390625,
0.340087890625,
0.6796875,
0.1158447265625,
0.6787109375
],
[
0.1351318359375,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.693359375,
0.1351318359375,
0.693359375
],
[
0.62060546875,
0.6796875,
0.7734375,
0.6796875,
0.7734375,
0.693359375,
0.62060546875,
0.693359375
],
[
0.85546875,
0.677734375,
0.884765625,
0.677734375,
0.884765625,
0.69482421875,
0.85546875,
0.69482421875
],
[
0.1175537109375,
0.69775390625,
0.486083984375,
0.69775390625,
0.486083984375,
0.71142578125,
0.1175537109375,
0.71142578125
],
[
0.1158447265625,
0.7119140625,
0.489501953125,
0.71337890625,
0.489501953125,
0.72802734375,
0.1158447265625,
0.72705078125
],
[
0.623046875,
0.7216796875,
0.6943359375,
0.7216796875,
0.6943359375,
0.736328125,
0.623046875,
0.736328125
],
[
0.77099609375,
0.72265625,
0.77734375,
0.72265625,
0.77734375,
0.72802734375,
0.77099609375,
0.72802734375
],
[
0.1158447265625,
0.728515625,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.1158447265625,
0.74365234375
],
[
0.76416015625,
0.728515625,
0.8212890625,
0.73095703125,
0.8193359375,
0.7451171875,
0.76318359375,
0.74365234375
],
[
0.85791015625,
0.728515625,
0.8818359375,
0.728515625,
0.8818359375,
0.74365234375,
0.85791015625,
0.74365234375
],
[
0.7109375,
0.73095703125,
0.75146484375,
0.73095703125,
0.75146484375,
0.7421875,
0.7109375,
0.7421875
],
[
0.57275390625,
0.732421875,
0.60693359375,
0.732421875,
0.60693359375,
0.74365234375,
0.57275390625,
0.74365234375
],
[
0.61376953125,
0.7392578125,
0.705078125,
0.7392578125,
0.705078125,
0.75048828125,
0.61376953125,
0.75048828125
],
[
0.1158447265625,
0.7451171875,
0.345947265625,
0.7451171875,
0.345947265625,
0.7587890625,
0.1158447265625,
0.7587890625
],
[
0.76416015625,
0.74462890625,
0.78076171875,
0.74462890625,
0.78076171875,
0.75439453125,
0.76416015625,
0.75439453125
],
[
0.51318359375,
0.76123046875,
0.8818359375,
0.76123046875,
0.8818359375,
0.775390625,
0.51318359375,
0.775390625
],
[
0.1376953125,
0.77099609375,
0.33154296875,
0.7685546875,
0.33154296875,
0.783203125,
0.1376953125,
0.78515625
],
[
0.51123046875,
0.77685546875,
0.8818359375,
0.779296875,
0.8818359375,
0.7919921875,
0.51123046875,
0.79052734375
],
[
0.1368408203125,
0.7958984375,
0.333251953125,
0.79443359375,
0.333251953125,
0.80859375,
0.1368408203125,
0.81005859375
],
[
0.51123046875,
0.79345703125,
0.8818359375,
0.7919921875,
0.8818359375,
0.80712890625,
0.51123046875,
0.80810546875
],
[
0.51318359375,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.1158447265625,
0.818359375,
0.486083984375,
0.8193359375,
0.486083984375,
0.83447265625,
0.1158447265625,
0.8330078125
],
[
0.5146484375,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.5146484375,
0.84033203125
],
[
0.11419677734375,
0.8349609375,
0.3896484375,
0.8359375,
0.3896484375,
0.85107421875,
0.11419677734375,
0.849609375
],
[
0.5087890625,
0.8388671875,
0.8759765625,
0.8427734375,
0.8759765625,
0.859375,
0.5087890625,
0.85546875
],
[
0.1175537109375,
0.8623046875,
0.21826171875,
0.8623046875,
0.21826171875,
0.876953125,
0.1175537109375,
0.876953125
],
[
0.51123046875,
0.86767578125,
0.70947265625,
0.8681640625,
0.70947265625,
0.8828125,
0.51123046875,
0.8818359375
],
[
0.11419677734375,
0.8857421875,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.11419677734375,
0.90087890625
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.1158447265625,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"Our zero-shot transfer task then involves deter-",
"a as input and outputs a binary outcome (allega-",
"mining violation/allegation from case facts with",
"tion in Task B and violation in Task A) for case",
"respect to articles which are not seen during train-",
"& with respect to article a. Our architecture is a",
"ing time. We consider a 'domain' to be a particular",
"modified version of the Enhanced Sequential Infer",
"convention article(i.e..lO conventionarticles form",
"ence Model (ESIM) (Chen et al., 2017) incorporat-",
"10 domains). The objective is to train a model on",
"ing conditional encoding (Augenstein et al., 2016:",
"a source domain (seen articles) with the goal of",
"Rocktaschel et al., 2016) that has been adapted to",
"performing well at test-time on a target domain",
"deal with long input sequences following hierarchi-",
"(unseen articles). Following (Yin et al., 2019; Ram-",
"cal attention networks (Yang et al., 2016). We ex-",
"poni and Plank, 2020), we propose two settings",
"periment with two domain adaptation components",
"under zero-shot COC:",
"based on adversarial training: (1) a classification-",
"Zero-Shot Restrictive / Unsupervised Domain",
"based domain discriminator and (2) a Wasserstein-",
"Adaptation (UDA): In this setting, the model is",
"distance based method which aims to reduce the",
"given a pair of case facts and the text of training set",
"difference between the source and the target do-",
"articles (i.e., the source domain) along with their",
"main distributions.",
"corresponding violation/allegation outcome label.",
"In the target domain, it is provided with case facts",
"4.1 Article-aware prediction Model",
"and article text pairs as well, but the outcome label",
"Given the facts of the case = {α1, &2, .. ., &m]",
"is withheld. The goal of UDA is to learn an out-",
"where i = [cil, Ci2, ..., Cin] and the article a -",
"come classifier from the outcome labelled source",
"[a1,a2,...,ak] where aj - [ajl,aj2,...,ajt],",
"domain which should generalize well on the target",
"the model outputs a binary label. c, / ai and ajp",
"domain by leveraging outcome-unlabeled target",
"/ aip denote the ith sentence and pth token of the",
"data. This setting is legally realistic, as the text of",
"jth sentence of the case facts / article, respectively",
"new or modified written legal sources is typically",
"m/k and n/l denote the number of sentences and",
"known for a given task and available for domain",
"tokens in the ith sentence of case facts / article,",
"adaptation (e.g., a public administration decision",
"respectively. Our model contains an encoding layer,",
"support tool receives an update after relevant legis-",
"followed by an interaction layer, a post-interaction",
"lation has changed).",
"encoding layer, and a classification header. See Fig.",
"Zero-shot Wild / Any Domain Adaptation",
"1 for an overview of our architecture.",
"(ADA) / Out of Distribution Generalization: In",
"this setting, the model never sees any article data",
"4.1.1 Pre-interaction Encoding Layer",
"from the target domain during training, yet should",
"Our model encodes the facts of the case α sentence-",
"be able to generalize to it. In the legal setting, this",
"wise with LegalBERT (Chalkidis et al., 2020) to ob-",
"corresponds to a model which is required to work",
"tain token level representations [ zi1, zi2, .. , zin]",
"with texts of sources only available at query time",
"These are aggregated into sentence level represen-",
"(e.g., complex retrieval settings where multiple le-",
"tations using token attention:",
"gal sources potentially apply).",
"We reorganize the dataset to evaluate our zero-",
"uit = tanh(Wwzit +bu)",
"(1)",
"shot transfer/adaptation models by splitting the 10",
"ECHR articles into two non-overlapping groups,",
"exp(uituw)",
"n",
"such that both contain articles of various frequen-",
"Eaitzit",
"(2)",
"&fi:",
"Qit=",
"Ztexp(uituw)",
"cies (common, moderate, rare)",
"t",
"where Wu,bw and uw are trainable parameters.",
"· split_0: 6, 8, P1-1,2,9",
"The sentence levelrepresentations {fi,...,fn] are",
": split_1: 3, 5, 10, 14, 11",
"passed through a GRU encoder to obtain context-",
"aware sentence representations of the facts h =-",
"We evaluate UDA and ADA on split_0 as source",
"[hi, h2, ..., hm]. The analogous article encoder",
"and split_1 as target, and vice-versa.",
"takes a as input and outputs s = {si, $2, .. , Sh].",
"4 Method",
"4.1.2 Interaction Layer",
"We employ a hierarchical neural model which takes",
"Interaction between the sentences of the case facts",
"the case fact description along with the article",
" and articles is done via dot product attention be"
] |
{
"arxiv_id": "2302.00609",
"language": "en",
"page_idx": 3
}
|
||
2302.00609-en-4
|
2302.00609-en-4.jpg
|
tween the two sequences of sentences as follows:
(3)
$$e_{ij}=h_i^Ts_j\:\&\:h_i'=\sum_{j=1}^k\frac{\exp(e_{ij})}{\sum_{l=1}^k\exp(e_{il})}s_j$$
(4)
$$\begin{aligned}s_j'&=\sum_{i=1}^m\frac{\exp(e_{ij})}{\sum_{l=1}^m\exp(e_{lj})}h_i\end{aligned}$$
where $e_ij$ represents the dot product interaction score between the context-aware representations of the $i^\mathrm{th}$ sentence of the case facts and the $j^\mathrm{th}$ sentence of the article. $h_i^{\prime}$ and $s_j^{\prime}$ represent articleaware representations corresponding to the $i^\mathrm{th}$ sentence of the case facts and the fact-aware representation corresponding to the $j^\mathrm{th}$ sentence of the article, respectively. Finally, we obtain interactionaware sentence representations of the facts $h^{\prime}=$ $\{h_{1}^{\prime},h_{2}^{\prime},\ldots,h_{m}^{\prime}\}.$ Similarly for the article, we ob- $\tan s^{\prime}=\{s_1^{\prime},s_2^{\prime},\ldots,s_k^{\prime}\}$
4.1.3 Post-Interaction Encoding Layer
The article-dependent final representation of the case facts is obtained in two steps: (i) we compute the final representation of the article text and (ii) use it as a conditional encoding (Augenstein et al., 2016; Rocktäschel et al., 2016) to obtain the final article-dependent fact representation.
Final representation of article: We first combine the pre-interaction sentence encodings and factaware sentence representations of the article:
(5)
$$p_i=[s_i,s_i',s_i-s_i',s_i\odot s_i']$$
where $\odot$ denotes element-wise product. This representation aims to capture high-order interaction between the pre- and post- interaction elements (Chen et al., 2017). The sentence representations $p_i$ are passed over a non-linear projection and a GRU (as in the pre-interaction encoder) to perform context-level modelling among sentence sequences. The final article representation $A$ is obtained via sentence attention analogous to eq. 2.
Final Representation of Case Facts: Similarly, we pass the combined representation of case facts using pre- and post- interaction similar to Eq. 5 over a non linear projection, a GRU layer, and sentence level attention to the obtain article-dependent final representation of case facts. To ensure conditioning, we initialize the GRU hidden state with the final representation of the articles $A.$ This facilitates capturing the salient case fact information with respect to the specified article.
<FigureHere>
Figure 1: Our article-aware prediction model architec-
ture
### 4.1.4 Classification Layer
We pass the article-dependent final representation of the case facts through a nonlinear projection to classify the outcome.
### 4.2 Domain Adaptation Components
Domain Adaptation aims to make models generalize well from a source to a target domain. Both domains are mapped to a common latent space, reducing differences between their distributions and facilitating domain invariant feature representations. In our case of article-aware COC, we regard reasoning with respect to every ECHR article as a domain and seek to learn article-invariant case facts representations. Put differently, we want our model to learn how to read two texts and interrelate them towards an outcome determination (as lawyers do) with minimal encoding of the information contained in the texts into the model itself. This way, the models can achieve generalization capability to adapt and perform reasoning with regard to articles not seen during training time.
## 4.2.1 Domain Discriminator
We employ a two layer feed forward network as a discriminator which takes the article-dependent case fact representation as input to predict the article (i.e., the domain). We train the discriminator in an adversarial fashion to maximize the model's ability to capture information required for the outcome task while minimizing its ability to
|
[
[
0.1175537109375,
0.08612060546875,
0.480224609375,
0.08612060546875,
0.480224609375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.58251953125,
0.10986328125,
0.66162109375,
0.10986328125,
0.66162109375,
0.1175537109375,
0.58251953125,
0.1175537109375
],
[
0.30908203125,
0.12054443359375,
0.318115234375,
0.12054443359375,
0.318115234375,
0.12469482421875,
0.30908203125,
0.12469482421875
],
[
0.334228515625,
0.119384765625,
0.397216796875,
0.119384765625,
0.397216796875,
0.1341552734375,
0.334228515625,
0.1341552734375
],
[
0.172119140625,
0.1258544921875,
0.300537109375,
0.12353515625,
0.300537109375,
0.1395263671875,
0.172119140625,
0.141357421875
],
[
0.462646484375,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.462646484375,
0.1395263671875
],
[
0.297119140625,
0.1341552734375,
0.41650390625,
0.1318359375,
0.41650390625,
0.15087890625,
0.297119140625,
0.152587890625
],
[
0.62646484375,
0.160888671875,
0.6337890625,
0.160888671875,
0.6337890625,
0.166259765625,
0.62646484375,
0.166259765625
],
[
0.258544921875,
0.16748046875,
0.276123046875,
0.168701171875,
0.274658203125,
0.17578125,
0.2568359375,
0.174560546875
],
[
0.285400390625,
0.16748046875,
0.35107421875,
0.16748046875,
0.35107421875,
0.1817626953125,
0.285400390625,
0.1817626953125
],
[
0.4609375,
0.174560546875,
0.48779296875,
0.174560546875,
0.48779296875,
0.188232421875,
0.4609375,
0.188232421875
],
[
0.58251953125,
0.17041015625,
0.67919921875,
0.17041015625,
0.67919921875,
0.18115234375,
0.58251953125,
0.18115234375
],
[
0.2225341796875,
0.1776123046875,
0.2392578125,
0.174560546875,
0.2435302734375,
0.18408203125,
0.2275390625,
0.18701171875
],
[
0.267822265625,
0.179931640625,
0.3720703125,
0.1817626953125,
0.3720703125,
0.1995849609375,
0.267822265625,
0.19775390625
],
[
0.366943359375,
0.179931640625,
0.376220703125,
0.179931640625,
0.376220703125,
0.185302734375,
0.366943359375,
0.185302734375
],
[
0.80419921875,
0.17578125,
0.8271484375,
0.17578125,
0.8271484375,
0.203125,
0.80419921875,
0.203125
],
[
0.59033203125,
0.18408203125,
0.66015625,
0.182861328125,
0.66015625,
0.191162109375,
0.59033203125,
0.1923828125
],
[
0.6650390625,
0.1859130859375,
0.67919921875,
0.1859130859375,
0.67919921875,
0.190673828125,
0.6650390625,
0.190673828125
],
[
0.7421875,
0.18408203125,
0.748046875,
0.185302734375,
0.744140625,
0.193603515625,
0.73779296875,
0.191162109375
],
[
0.69775390625,
0.196533203125,
0.791015625,
0.196533203125,
0.791015625,
0.207275390625,
0.69775390625,
0.207275390625
],
[
0.1158447265625,
0.208984375,
0.486083984375,
0.2078857421875,
0.486083984375,
0.22265625,
0.1158447265625,
0.223876953125
],
[
0.6962890625,
0.208984375,
0.77685546875,
0.2078857421875,
0.77685546875,
0.2197265625,
0.6962890625,
0.2203369140625
],
[
0.58447265625,
0.2161865234375,
0.62890625,
0.2161865234375,
0.62890625,
0.226806640625,
0.58447265625,
0.226806640625
],
[
0.8076171875,
0.21142578125,
0.81787109375,
0.21142578125,
0.81787109375,
0.21435546875,
0.8076171875,
0.21435546875
],
[
0.7001953125,
0.2197265625,
0.74462890625,
0.2197265625,
0.74462890625,
0.2298583984375,
0.7001953125,
0.2298583984375
],
[
0.1175537109375,
0.2257080078125,
0.486083984375,
0.2257080078125,
0.486083984375,
0.2392578125,
0.1175537109375,
0.2392578125
],
[
0.765625,
0.22265625,
0.7734375,
0.22265625,
0.7734375,
0.22802734375,
0.765625,
0.22802734375
],
[
0.57861328125,
0.2298583984375,
0.66015625,
0.2298583984375,
0.66015625,
0.240478515625,
0.57861328125,
0.240478515625
],
[
0.11248779296875,
0.2381591796875,
0.489501953125,
0.2369384765625,
0.489501953125,
0.254638671875,
0.11248779296875,
0.255859375
],
[
0.7119140625,
0.2327880859375,
0.76904296875,
0.2327880859375,
0.76904296875,
0.240478515625,
0.7119140625,
0.240478515625
],
[
0.580078125,
0.2452392578125,
0.6396484375,
0.2452392578125,
0.6396484375,
0.257080078125,
0.580078125,
0.257080078125
],
[
0.6533203125,
0.2476806640625,
0.6767578125,
0.2476806640625,
0.6767578125,
0.255859375,
0.6533203125,
0.255859375
],
[
0.6962890625,
0.2476806640625,
0.73046875,
0.2476806640625,
0.73046875,
0.2587890625,
0.6962890625,
0.2587890625
],
[
0.11419677734375,
0.255859375,
0.489501953125,
0.257080078125,
0.489501953125,
0.271484375,
0.11419677734375,
0.270751953125
],
[
0.759765625,
0.2529296875,
0.76904296875,
0.2529296875,
0.76904296875,
0.255859375,
0.759765625,
0.255859375
],
[
0.63037109375,
0.26123046875,
0.705078125,
0.26123046875,
0.705078125,
0.2724609375,
0.63037109375,
0.2724609375
],
[
0.11419677734375,
0.27490234375,
0.489501953125,
0.273681640625,
0.489501953125,
0.2880859375,
0.11419677734375,
0.289306640625
],
[
0.58447265625,
0.276611328125,
0.63037109375,
0.276611328125,
0.63037109375,
0.287353515625,
0.58447265625,
0.287353515625
],
[
0.640625,
0.27783203125,
0.6689453125,
0.27783203125,
0.6689453125,
0.28515625,
0.640625,
0.28515625
],
[
0.705078125,
0.27783203125,
0.72705078125,
0.27783203125,
0.72705078125,
0.2861328125,
0.705078125,
0.2861328125
],
[
0.1158447265625,
0.29150390625,
0.48779296875,
0.29150390625,
0.48779296875,
0.3046875,
0.1158447265625,
0.3046875
],
[
0.58447265625,
0.292236328125,
0.6650390625,
0.292236328125,
0.6650390625,
0.302734375,
0.58447265625,
0.302734375
],
[
0.69873046875,
0.292236328125,
0.77099609375,
0.292236328125,
0.77099609375,
0.302734375,
0.69873046875,
0.302734375
],
[
0.58837890625,
0.302734375,
0.67529296875,
0.302734375,
0.67529296875,
0.3134765625,
0.58837890625,
0.3134765625
],
[
0.6884765625,
0.302734375,
0.7734375,
0.302734375,
0.7734375,
0.3134765625,
0.6884765625,
0.3134765625
],
[
0.8037109375,
0.3046875,
0.81201171875,
0.3046875,
0.81201171875,
0.30810546875,
0.8037109375,
0.30810546875
],
[
0.58447265625,
0.3134765625,
0.6572265625,
0.3134765625,
0.6572265625,
0.32373046875,
0.58447265625,
0.32373046875
],
[
0.1158447265625,
0.30712890625,
0.48779296875,
0.30712890625,
0.48779296875,
0.320556640625,
0.1158447265625,
0.320556640625
],
[
0.70361328125,
0.3154296875,
0.76318359375,
0.3154296875,
0.76318359375,
0.322509765625,
0.70361328125,
0.322509765625
],
[
0.1158447265625,
0.32373046875,
0.48779296875,
0.32373046875,
0.48779296875,
0.33740234375,
0.1158447265625,
0.33740234375
],
[
0.58447265625,
0.330810546875,
0.6337890625,
0.330810546875,
0.6337890625,
0.341552734375,
0.58447265625,
0.341552734375
],
[
0.646484375,
0.33203125,
0.66748046875,
0.33203125,
0.66748046875,
0.34033203125,
0.646484375,
0.34033203125
],
[
0.69873046875,
0.330810546875,
0.744140625,
0.330810546875,
0.744140625,
0.341552734375,
0.69873046875,
0.341552734375
],
[
0.75732421875,
0.33203125,
0.767578125,
0.3349609375,
0.76416015625,
0.341552734375,
0.75244140625,
0.33837890625
],
[
0.1158447265625,
0.339111328125,
0.48193359375,
0.339111328125,
0.48193359375,
0.352783203125,
0.1158447265625,
0.352783203125
],
[
0.599609375,
0.342529296875,
0.64990234375,
0.342529296875,
0.64990234375,
0.349853515625,
0.599609375,
0.349853515625
],
[
0.71875,
0.341552734375,
0.75244140625,
0.341552734375,
0.75244140625,
0.352783203125,
0.71875,
0.352783203125
],
[
0.1158447265625,
0.35400390625,
0.490234375,
0.3515625,
0.490234375,
0.369384765625,
0.1158447265625,
0.371826171875
],
[
0.11248779296875,
0.369384765625,
0.30908203125,
0.37060546875,
0.30908203125,
0.388427734375,
0.11248779296875,
0.38720703125
],
[
0.51123046875,
0.364013671875,
0.88330078125,
0.364013671875,
0.88330078125,
0.377685546875,
0.51123046875,
0.377685546875
],
[
0.51123046875,
0.380126953125,
0.54248046875,
0.380126953125,
0.54248046875,
0.391357421875,
0.51123046875,
0.391357421875
],
[
0.1175537109375,
0.399658203125,
0.432373046875,
0.399658203125,
0.432373046875,
0.413330078125,
0.1175537109375,
0.413330078125
],
[
0.1175537109375,
0.42041015625,
0.489501953125,
0.42041015625,
0.489501953125,
0.43408203125,
0.1175537109375,
0.43408203125
],
[
0.51318359375,
0.419189453125,
0.7255859375,
0.419189453125,
0.7255859375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.1175537109375,
0.43701171875,
0.489501953125,
0.43701171875,
0.489501953125,
0.45068359375,
0.1175537109375,
0.45068359375
],
[
0.509765625,
0.439453125,
0.8818359375,
0.439453125,
0.8818359375,
0.453125,
0.509765625,
0.453125
],
[
0.1175537109375,
0.453125,
0.48779296875,
0.453125,
0.48779296875,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.5087890625,
0.45361328125,
0.8818359375,
0.454833984375,
0.8818359375,
0.4697265625,
0.5087890625,
0.468505859375
],
[
0.118408203125,
0.47021484375,
0.486083984375,
0.47021484375,
0.486083984375,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.509765625,
0.471435546875,
0.6708984375,
0.471435546875,
0.6708984375,
0.485107421875,
0.509765625,
0.485107421875
],
[
0.1175537109375,
0.485107421875,
0.486083984375,
0.485107421875,
0.486083984375,
0.498779296875,
0.1175537109375,
0.498779296875
],
[
0.1175537109375,
0.50048828125,
0.39208984375,
0.50048828125,
0.39208984375,
0.51416015625,
0.1175537109375,
0.51416015625
],
[
0.51123046875,
0.49755859375,
0.8154296875,
0.498779296875,
0.8154296875,
0.51318359375,
0.51123046875,
0.5126953125
],
[
0.1175537109375,
0.51708984375,
0.489501953125,
0.51708984375,
0.489501953125,
0.53076171875,
0.1175537109375,
0.53076171875
],
[
0.509765625,
0.52099609375,
0.88330078125,
0.52099609375,
0.88330078125,
0.53369140625,
0.509765625,
0.53369140625
],
[
0.1175537109375,
0.533203125,
0.489501953125,
0.533203125,
0.489501953125,
0.54638671875,
0.1175537109375,
0.54638671875
],
[
0.509765625,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.55029296875,
0.509765625,
0.55029296875
],
[
0.1158447265625,
0.55029296875,
0.44921875,
0.55029296875,
0.44921875,
0.56396484375,
0.1158447265625,
0.56396484375
],
[
0.51318359375,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.51318359375,
0.56787109375
],
[
0.51318359375,
0.57080078125,
0.87890625,
0.57080078125,
0.87890625,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.2166748046875,
0.5830078125,
0.385498046875,
0.5830078125,
0.385498046875,
0.5966796875,
0.2166748046875,
0.5966796875
],
[
0.462646484375,
0.5830078125,
0.48779296875,
0.5830078125,
0.48779296875,
0.595703125,
0.462646484375,
0.595703125
],
[
0.5087890625,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.5087890625,
0.59716796875
],
[
0.51123046875,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.51123046875,
0.615234375
],
[
0.11419677734375,
0.61328125,
0.48779296875,
0.61328125,
0.48779296875,
0.626953125,
0.11419677734375,
0.626953125
],
[
0.5087890625,
0.61767578125,
0.880859375,
0.6162109375,
0.880859375,
0.63134765625,
0.5087890625,
0.6318359375
],
[
0.11419677734375,
0.6298828125,
0.486083984375,
0.6298828125,
0.486083984375,
0.6435546875,
0.11419677734375,
0.6435546875
],
[
0.509765625,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64794921875,
0.509765625,
0.64794921875
],
[
0.11419677734375,
0.6455078125,
0.48779296875,
0.6455078125,
0.48779296875,
0.6591796875,
0.11419677734375,
0.6591796875
],
[
0.509765625,
0.6484375,
0.880859375,
0.6484375,
0.880859375,
0.662109375,
0.509765625,
0.662109375
],
[
0.1175537109375,
0.662109375,
0.489501953125,
0.662109375,
0.489501953125,
0.67578125,
0.1175537109375,
0.67578125
],
[
0.51123046875,
0.66748046875,
0.880859375,
0.66748046875,
0.880859375,
0.677734375,
0.51123046875,
0.677734375
],
[
0.11419677734375,
0.6787109375,
0.489501953125,
0.6767578125,
0.489501953125,
0.69140625,
0.11419677734375,
0.693359375
],
[
0.509765625,
0.681640625,
0.8818359375,
0.681640625,
0.8818359375,
0.6953125,
0.509765625,
0.6953125
],
[
0.1175537109375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.509765625,
0.69873046875,
0.88330078125,
0.69873046875,
0.88330078125,
0.7119140625,
0.509765625,
0.7119140625
],
[
0.1175537109375,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.509765625,
0.71435546875,
0.88330078125,
0.71435546875,
0.88330078125,
0.72802734375,
0.509765625,
0.72802734375
],
[
0.1158447265625,
0.72705078125,
0.489501953125,
0.72705078125,
0.489501953125,
0.74072265625,
0.1158447265625,
0.74072265625
],
[
0.5087890625,
0.728515625,
0.88330078125,
0.72998046875,
0.88330078125,
0.74462890625,
0.5087890625,
0.74365234375
],
[
0.1175537109375,
0.7421875,
0.393798828125,
0.7421875,
0.393798828125,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.51123046875,
0.74658203125,
0.8818359375,
0.74658203125,
0.8818359375,
0.76025390625,
0.51123046875,
0.76025390625
],
[
0.1175537109375,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.509765625,
0.76318359375,
0.80615234375,
0.76318359375,
0.80615234375,
0.77685546875,
0.509765625,
0.77685546875
],
[
0.11419677734375,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.11419677734375,
0.7880859375
],
[
0.51123046875,
0.7861328125,
0.74658203125,
0.78759765625,
0.74658203125,
0.8017578125,
0.51123046875,
0.80029296875
],
[
0.1158447265625,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.1158447265625,
0.80810546875,
0.491943359375,
0.80810546875,
0.491943359375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.509765625,
0.8056640625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.509765625,
0.82080078125
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.826171875,
0.880859375,
0.826171875,
0.880859375,
0.8359375,
0.51123046875,
0.8359375
],
[
0.1175537109375,
0.84033203125,
0.491943359375,
0.84033203125,
0.491943359375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51123046875,
0.8427734375,
0.8818359375,
0.8427734375,
0.8818359375,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.118408203125,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.8681640625,
0.118408203125,
0.8681640625
],
[
0.5087890625,
0.85546875,
0.8818359375,
0.85498046875,
0.8818359375,
0.869140625,
0.5087890625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.490234375,
0.87255859375,
0.490234375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.3779296875,
0.90478515625,
0.3779296875,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"tween the two sequences of sentences as follows:",
"Classification Laye",
"",
"exp(eii)",
"ej =hTs,& h,=",
"(3)",
", exp(eil)",
"C",
"m",
"exp(ei)",
"(4)",
"Sentence Attention",
"",
"exp(etu)",
"h",
"oe Bu",
"Sentence GRU",
"",
"a",
"Sentence Attention",
"where eij represents the dot product interaction",
"Sentence GRU",
"91.9293",
"",
"PuP2...",
"score between the context-aware representations",
"P",
"Combination",
"of the zth sentence of the case facts and the jth",
"Combination",
"h',h'2,h's,.!",
"..h'.",
"$'i,s'2,",
"sentence of the article. h, and s, represent article-",
"S",
"Interaction Layer",
"aware representations corresponding to the ith sen-",
"h,h, hg.",
"...h.",
"S1,S2",
"tence of the case facts and the fact-aware repre-",
"Sentence GRU",
"Sentence GRU",
"Token Attention",
"Token Attention",
"",
"LEGALBERT",
"sentation corresponding to the jth sentence of the",
"LEGAL BERT",
"article, respectively. Finally, we obtain interaction-",
"Xf, X2, Xgt -.",
"...X.",
"a,a,...",
"a",
"aware sentence representations of the facts h'",
"CaseFacts",
"Article",
"[h'i, h2,..., h'm]. Similarly for the article, we ob-",
"tain s - [$i, $2,. $]",
"Figure I: Our article-aware prediction model architec-",
"ture",
"4.1.3Post-Interaction Encoding Layer",
"The article-dependent final representation of the",
"4.1.4 Classification Layer",
"case facts is obtained in two steps: (i) we compute",
" We pass the article-dependent final representation",
"the final representation of the article text and (ii)",
"of the case facts through a nonlinear projection to",
"use it as a conditional encoding (Augenstein et al.",
"classify the outcome.",
"2016; Rocktaschel et al., 2016) to obtain the final",
"article-dependent fact representation.",
"4.2Domain Adaptation Components",
"Final representation of article: We first combine",
"Domain Adaptation aims to make models gener-",
"the pre-interaction sentence encodings and fact-",
"alize well from a source to a target domain. Both",
"aware sentence representations of the article:",
"domains are mapped to a common latent space,",
"reducingdifferences between theirdistributions",
"pi = [si, st,si - st,sios]",
"(5)",
"and facilitating domain invariant feature representa",
"tions. In our case of article-aware COC, we regard",
"where denotes element-wise product. This rep-",
"reasoning with respect to every ECHR article as",
"resentation aims to capture high-order interaction",
" a domain and seek to learn article-invariant case",
"between the pre- and post- interaction elements",
"facts representations. Put differently, we want our",
"(Chen et al., 2017). The sentence representations",
"modelto learn how toread two texts and interre",
"pi are passed over a non-linear projection and a",
"late them towards an outcome determination (as",
"GRU (as in the pre-interaction encoder) to perform",
"lawyers do)with minimal encoding of the infor-",
"context-level modelling among sentence sequences.",
"mation contained in the texts into the model itself.",
"The final article representation A is obtained via",
"This way, the models can achieve generalization ca-",
"sentence attention analogous to eg. 2.",
"pability to adapt and perform reasoning with regard",
"Final Representation of Case Facts: Similarly,",
" to articles not seen during training time.",
"we pass the combined representation of case facts",
"4.2.1 Domain Discriminator",
"using pre- and post- interaction similar to Eq. 5",
"over a non linear projection, a GRU layer, and sen-",
"We employ a two layer feed forward network as",
"tence level attention to the obtain article-dependent",
"a discriminator whichtakesthearticle-dependent",
"final representation of case facts. To ensure con-",
"casefactrepresentationas inputtopredictthear-",
"ditioning,weinitializethe GRU hidden state with",
"ticle (i.e., the domain).We train the discrimi-",
"the final representation of the articles A. This fa-",
" nator in an adversarial fashion to maximize the",
"cilitates capturing the salient case fact information",
"model's ability to capture information required for",
"with respect to the specified article.",
"the outcome task while minimizing its ability to"
] |
{
"arxiv_id": "2302.00609",
"language": "en",
"page_idx": 4
}
|
||
1704.03925-en-0
|
1704.03925-en-0.jpg
|
Provable Self-Representation Based Outlier Detection in a Union of Subspaces
Chong You, Daniel P. Robinson, René Vidal
Johns Hopkins University, Baltimore, MD, 21218, USA
$ਨਨੇ$
## $\mathbf{Abstract}$
Many computer vision tasks involve processing large amounts of data contaminated by outliers, which need to be detected and rejected. While outlier detection methods
$\sum_{cently have methods based on robust statistics have existed for decades, only re-}\\S cently have methods based on sparse and low-rank repre-$
sentation been developed along with guarantees of correct outlier detection when the inliers lie in one or more lowdimensional subspaces. This paper proposes a new outlier detection method that combines tools from sparse representation with random walks on a graph. By exploiting the property that data points can be expressed as sparse linear combinations of each other, we obtain an asymmetric affinity matrix among data points, which we use to construct a weighted directed graph. By defning a suitable $Markov\textit{Chain from this graph, we establish a connection}$ between inliers/outliers and essential/inessential states of the Markov chain, which allows us to detect outliers by using random walks. We provide a theoretical analysis that justifres the correctness of our method under geometric and connectivity assumptions. Experimental results on image databases demonstrate its superiority with respect to state- $of$-the-art sparse and low-rank outlier detection methods.
## $1. \textbf{Introduction}$
In many applications in computer vision, including motion estimation and segmentation [19] and face recognition [2], high-dimensional datasets can be well approximated by a union of low-dimensional subspaces. Such applications have motivated a lot of research on the problems of learning one or more subspaces from data, a.k.a. subspace learning and subspace clustering, respectively. In practice, datasets are often contaminated by points that do not lie in the subspaces, i.e. outliers. In such situations, it is often essential to detect and reject these outliers before any subsequent processing/analysis is performed.
Prior work. We address the problem of outlier detection in the setting when the inlier data are assumed to lie close to a union of unknown low-dimensional subspaces (low relative to the dimension of the ambient space). A traditional method for solving this problem is RANSAC[13], which is
based on randomly selecting a subset of points, fitting a subspace to them, and counting the number of points that are well fit by this subspace; this process is repeated for suffciently many trials and the best fit is chosen. RANSAC is intrinsically combinatorial and the number of trials needed to find a good estimate of the subspace grows exponentially with the subspace dimension. Consequently, the methods of choice have been to robustly learn the subspaces by penalizing the sum of unsquared distances (in lieu of squared distances used in classical methods such as PCA) of points to the closest subspace [9,22,62,61]. Such a penalty is robust to outliers because it reduces the contributions from large residuals arising from outliers. However, the optimization problem is usually nonconvex and a good initialization is extremely important for finding the optimal solution.
The groundbreaking work of Wright et al.[54] and Candès et al. [4] on using convex optimization techniques to solve the PCA problem with robustness to corrupted entries has led to many recent methods for PCA with robustness to outliers [55, 32, 25, 60, 21]. For example, Outlier Pursuit [55] uses the nuclear norm $\|\cdot\|_*$ to seek low-rank solutions by solving the problem $\min_L\|X-L\|_{2,1}+\lambda\|L\|_*$ for some $\lambda>0.$ A prominent advantage of convex optimization techniques is that they are guaranteed to correctly identify outliers under certain conditions. Very recently, several nonconvex outlier detection methods have also been developed with guaranteed correctness [20, 6]. Nonetheless, these methods typically model a unique inlier subspace, e.g., by a low rank matrix $L$ in Outlier Pursuit, and therefore cannot deal with multiple inlier subspaces since the union of multiple subspaces could be high-dimensional.
Another class of methods with theoretical guarantees for correctness utilizes the fact that outliers are expected to have low similarities with other data points. In [5,1],a multiway similarity is introduced that is defined from the polar curvature, which has the advantage of exploiting the subspace structure. However, the number of combinations in multi-way similarity can be prohibitively large. Some recent works have explored using inner products between data points for outlier detection [17,40]. Although computationally very efficient, these methods require the inliers to be well distributed and densely sampled within the subspaces.
|
[
[
0.096435546875,
0.1337890625,
0.875,
0.1337890625,
0.875,
0.1483154296875,
0.096435546875,
0.1483154296875
],
[
0.314453125,
0.18310546875,
0.65869140625,
0.18310546875,
0.65869140625,
0.1976318359375,
0.314453125,
0.1976318359375
],
[
0.265625,
0.199462890625,
0.70361328125,
0.199462890625,
0.70361328125,
0.2139892578125,
0.265625,
0.2139892578125
],
[
0.236083984375,
0.250732421875,
0.314453125,
0.250732421875,
0.314453125,
0.268310546875,
0.236083984375,
0.268310546875
],
[
0.0187835693359375,
0.2587890625,
0.062103271484375,
0.25830078125,
0.066162109375,
0.367431640625,
0.0228729248046875,
0.36865234375
],
[
0.50341796875,
0.25390625,
0.890625,
0.25390625,
0.890625,
0.267578125,
0.50341796875,
0.267578125
],
[
0.5048828125,
0.270751953125,
0.890625,
0.270751953125,
0.890625,
0.281005859375,
0.5048828125,
0.281005859375
],
[
0.09722900390625,
0.2822265625,
0.4697265625,
0.28271484375,
0.4697265625,
0.29736328125,
0.09722900390625,
0.296630859375
],
[
0.50244140625,
0.2841796875,
0.890625,
0.2841796875,
0.890625,
0.29736328125,
0.50244140625,
0.29736328125
],
[
0.08251953125,
0.300537109375,
0.468017578125,
0.300537109375,
0.468017578125,
0.311279296875,
0.08251953125,
0.311279296875
],
[
0.50341796875,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.311279296875,
0.50341796875,
0.311279296875
],
[
0.080078125,
0.314453125,
0.4697265625,
0.314453125,
0.4697265625,
0.32763671875,
0.080078125,
0.32763671875
],
[
0.5009765625,
0.314453125,
0.89208984375,
0.314453125,
0.89208984375,
0.32763671875,
0.5009765625,
0.32763671875
],
[
0.080078125,
0.328857421875,
0.4697265625,
0.328857421875,
0.4697265625,
0.34228515625,
0.080078125,
0.34228515625
],
[
0.50244140625,
0.330078125,
0.890625,
0.330078125,
0.890625,
0.343505859375,
0.50244140625,
0.343505859375
],
[
0.080078125,
0.3447265625,
0.4697265625,
0.3447265625,
0.4697265625,
0.35791015625,
0.080078125,
0.35791015625
],
[
0.50341796875,
0.3447265625,
0.89208984375,
0.3447265625,
0.89208984375,
0.35791015625,
0.50341796875,
0.35791015625
],
[
0.080078125,
0.35986328125,
0.4697265625,
0.35986328125,
0.4697265625,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.5009765625,
0.359130859375,
0.8896484375,
0.359130859375,
0.8896484375,
0.37255859375,
0.5009765625,
0.37255859375
],
[
0.080078125,
0.375,
0.4697265625,
0.375,
0.4697265625,
0.38818359375,
0.080078125,
0.38818359375
],
[
0.0269622802734375,
0.3642578125,
0.0555419921875,
0.3642578125,
0.0555419921875,
0.392578125,
0.0269622802734375,
0.392578125
],
[
0.50244140625,
0.375,
0.890625,
0.375,
0.890625,
0.38818359375,
0.50244140625,
0.38818359375
],
[
0.08087158203125,
0.39013671875,
0.4697265625,
0.39013671875,
0.4697265625,
0.404052734375,
0.08087158203125,
0.404052734375
],
[
0.50244140625,
0.39013671875,
0.89208984375,
0.39013671875,
0.89208984375,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.07843017578125,
0.404052734375,
0.4697265625,
0.40478515625,
0.4697265625,
0.419189453125,
0.07843017578125,
0.41845703125
],
[
0.50244140625,
0.40478515625,
0.8935546875,
0.40478515625,
0.8935546875,
0.41845703125,
0.50244140625,
0.41845703125
],
[
0.5009765625,
0.419189453125,
0.89208984375,
0.42041015625,
0.89208984375,
0.434326171875,
0.5009765625,
0.43310546875
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.434326171875,
0.080078125,
0.434326171875
],
[
0.08087158203125,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.448974609375,
0.08087158203125,
0.448974609375
],
[
0.50244140625,
0.43505859375,
0.89208984375,
0.43505859375,
0.89208984375,
0.448974609375,
0.50244140625,
0.448974609375
],
[
0.03350830078125,
0.449462890625,
0.0555419921875,
0.449462890625,
0.0555419921875,
0.479736328125,
0.03350830078125,
0.479736328125
],
[
0.08087158203125,
0.451904296875,
0.46728515625,
0.451904296875,
0.46728515625,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.50244140625,
0.449462890625,
0.89208984375,
0.448974609375,
0.89208984375,
0.46337890625,
0.50244140625,
0.464111328125
],
[
0.080078125,
0.46533203125,
0.468017578125,
0.46533203125,
0.468017578125,
0.479248046875,
0.080078125,
0.479248046875
],
[
0.50244140625,
0.46533203125,
0.8486328125,
0.46533203125,
0.8486328125,
0.479248046875,
0.50244140625,
0.479248046875
],
[
0.080078125,
0.479736328125,
0.471435546875,
0.479736328125,
0.471435546875,
0.494384765625,
0.080078125,
0.494384765625
],
[
0.52294921875,
0.486083984375,
0.89208984375,
0.486083984375,
0.89208984375,
0.5,
0.52294921875,
0.5
],
[
0.0236968994140625,
0.50048828125,
0.058013916015625,
0.50048828125,
0.058013916015625,
0.5888671875,
0.0236968994140625,
0.5888671875
],
[
0.080078125,
0.49560546875,
0.4697265625,
0.49560546875,
0.4697265625,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.08087158203125,
0.51220703125,
0.468017578125,
0.51220703125,
0.468017578125,
0.52294921875,
0.08087158203125,
0.52294921875
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.080078125,
0.52587890625,
0.468017578125,
0.52587890625,
0.468017578125,
0.5390625,
0.080078125,
0.5390625
],
[
0.50341796875,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.080078125,
0.54150390625,
0.4697265625,
0.54150390625,
0.4697265625,
0.5546875,
0.080078125,
0.5546875
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.08087158203125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.568359375,
0.08087158203125,
0.568359375
],
[
0.50244140625,
0.56103515625,
0.89208984375,
0.5625,
0.89208984375,
0.5771484375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.5712890625,
0.4697265625,
0.5712890625,
0.4697265625,
0.58544921875,
0.080078125,
0.58544921875
],
[
0.50244140625,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.035125732421875,
0.58544921875,
0.049835205078125,
0.58544921875,
0.049835205078125,
0.59228515625,
0.035125732421875,
0.59228515625
],
[
0.08087158203125,
0.5859375,
0.46728515625,
0.5859375,
0.46728515625,
0.599609375,
0.08087158203125,
0.599609375
],
[
0.50341796875,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50341796875,
0.60693359375
],
[
0.08087158203125,
0.6015625,
0.451904296875,
0.6015625,
0.451904296875,
0.61572265625,
0.08087158203125,
0.61572265625
],
[
0.02532958984375,
0.5888671875,
0.0555419921875,
0.5888671875,
0.0555419921875,
0.63818359375,
0.02532958984375,
0.63818359375
],
[
0.50341796875,
0.60791015625,
0.890625,
0.60791015625,
0.890625,
0.6220703125,
0.50341796875,
0.6220703125
],
[
0.5048828125,
0.625,
0.8896484375,
0.625,
0.8896484375,
0.63525390625,
0.5048828125,
0.63525390625
],
[
0.0228729248046875,
0.63330078125,
0.056365966796875,
0.63330078125,
0.056365966796875,
0.708984375,
0.0228729248046875,
0.708984375
],
[
0.08087158203125,
0.6318359375,
0.20751953125,
0.6318359375,
0.20751953125,
0.64599609375,
0.08087158203125,
0.64599609375
],
[
0.50341796875,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.6494140625,
0.50341796875,
0.6494140625
],
[
0.09722900390625,
0.65576171875,
0.46728515625,
0.65576171875,
0.46728515625,
0.66943359375,
0.09722900390625,
0.66943359375
],
[
0.50341796875,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.50341796875,
0.66650390625
],
[
0.07843017578125,
0.67041015625,
0.4697265625,
0.671875,
0.4697265625,
0.68603515625,
0.07843017578125,
0.68505859375
],
[
0.50244140625,
0.66845703125,
0.890625,
0.66845703125,
0.890625,
0.681640625,
0.50244140625,
0.681640625
],
[
0.08087158203125,
0.6875,
0.468017578125,
0.6875,
0.468017578125,
0.70068359375,
0.08087158203125,
0.70068359375
],
[
0.5048828125,
0.68505859375,
0.88818359375,
0.68505859375,
0.88818359375,
0.69580078125,
0.5048828125,
0.69580078125
],
[
0.08416748046875,
0.7041015625,
0.46728515625,
0.7041015625,
0.46728515625,
0.7138671875,
0.08416748046875,
0.7138671875
],
[
0.50341796875,
0.69873046875,
0.8935546875,
0.69873046875,
0.8935546875,
0.7119140625,
0.50341796875,
0.7119140625
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.50244140625,
0.7138671875,
0.81201171875,
0.7138671875,
0.81201171875,
0.72802734375,
0.50244140625,
0.72802734375
],
[
0.08087158203125,
0.73291015625,
0.468017578125,
0.73291015625,
0.468017578125,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.521484375,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.521484375,
0.74853515625
],
[
0.08087158203125,
0.74755859375,
0.4697265625,
0.74755859375,
0.4697265625,
0.76123046875,
0.08087158203125,
0.76123046875
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.080078125,
0.76220703125,
0.468017578125,
0.76220703125,
0.468017578125,
0.77587890625,
0.080078125,
0.77587890625
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.77783203125,
0.468017578125,
0.7763671875,
0.468017578125,
0.79052734375,
0.080078125,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.07843017578125,
0.79150390625,
0.46728515625,
0.79248046875,
0.46728515625,
0.806640625,
0.07843017578125,
0.80615234375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.80810546875,
0.27685546875,
0.80810546875,
0.27685546875,
0.8212890625,
0.080078125,
0.8212890625
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Provable Self-Representation Based Outlier Detection in a Union of Subspaces",
"Chong You, Daniel P. Robinson, René Vidal",
"Johns Hopkins University, Baltimore, MD, 21218, USA",
"Abstract",
"L10Id",
"based on randomly selecting a subset of points, fitting a sub-",
"spacetothem,and countingthenumberof pointsthatare",
"Many computer vision tasks involve processing large",
"well fit by this subspace; this process is repeated for suffi-",
"amounts of data contaminated by outliers,which need to",
"cientlymanytrialsandthebestfit ischosen.RANSAC is",
"be detected and rejected. While outlier detection methods",
"intrinsically combinatorial and the number of trials needed",
"based on robust statistics have existed for decades, only re-",
"to find a good estimate of the subspace grows exponentially",
"cently have methods based on sparse and low-rank repre-",
"with the subspace dimension. Consequently, the methods of",
"sentation been developed along with guarantees of correct",
"choice have been to robustly learn the subspaces by penaliz-",
"outlier detection when the inliers lie in one or more low-",
"2",
"ing the sum of unsquared distances (in lieu of squared dis-",
"dimensional subspaces.This paper proposes a new outlier",
"tances used in classical methods such as PCA) of points to",
"detection method that combines tools from sparse represen-",
"the closest subspace [9, 22, 62, 61]. Such a penalty is robust",
"to outliers because it reduces the contributions from large",
"LAD",
"tation with random walks on a graph. By exploiting the",
"property that data points can be expressed as sparse lin-",
"residuals arising from outliers. However, the optimization",
"'SO",
"ear combinations of each otherwe obtainanasymmetric",
"problem is usually nonconvex and a good initialization is",
"affinity matrix among data points, which we use to con-",
"extremely important for finding the optimal solution.",
"struct a weighted directed graph. By defining a suitable",
"The groundbreaking work of Wright et al. [54] and",
"152600",
"Markov Chain from this graph, we establish a connection",
"Candes et al. [4] on using convex optimization techniques to",
"betweeninliers/outliersand essential/inessential states of",
"solve the PCA problem with robustness to corrupted entries",
"the Markov chain, which allows us to detect outliers by us-",
"hasledtomanyrecentmethodsforPCAwithrobustnessto",
"ing random walks.We provide a theoretical analysis that",
"outliers [55, 32, 25, 60, 21]. For example, Outlier Pursuit",
"justifiesthecorrectnessofourmethodundergeometricand",
"[55] uses the nuclear norm I + to seek low-rank solutions",
"connectivity assumptions.Experimental results on image",
"by solving the problem minL JX-L2.1+入Lfor some",
"",
"databases demonstrate its superiority with respect to state-",
"> 0. A prominent advantage of convex optimization tech-",
"of-the-art sparse and low-rank outlier detection methods.",
"1",
"niques is that they are guaranteed to correctly identify out-",
"liersunder certain conditions.Very recently,several non-",
"LAIXI",
"1.Introduction",
"convexoutlierdetectionmethodshavealsobeendeveloped",
"In many applications in computer vision, including mo-",
"with guaranteed correctness [20, 6]. Nonetheless, these",
"tion estimation and segmentation [19] and face recognition",
"methods typically model a unique inlier subspace, e.g., by",
"[2], high-dimensional datasets can be well approximated by",
"alow rank matrix Lin OutlierPursuit,and therefore can-",
"aunionoflow-dimensionalsubspaces.Suchapplications",
"not deal with multiple inlier subspaces since the union of",
"have motivated a lot of research on the problems of learning",
"multiple subspaces could be high-dimensional.",
"one or more subspaces from data, a.k.a. subspace learning",
"Another class of methods with theoretical guarantees for",
"and subspace clustering, respectively. In practice, datasets",
"correctness utilizes the fact that outliers are expected to have",
"are often contaminated by points that do not lie in the sub-",
"low similarities with other data points. In [5, I], a multi-",
"spaces, i.e. outliers. In such situations, it is often essential to",
"way similarity is introduced thatis defined fromthepolar",
"detect and reject these outliers before any subsequent pro-",
"curvature, which has the advantage of exploiting the sub-",
"cessing/analysis is performed.",
"space structure. However, the number of combinations in",
"Prior work. We address the problem of outlier detection in",
"multi-way similarity can be prohibitively large. Some re-",
"the setting when the inlier data are assumed to lie close to",
"cent works have explored using inner products between data",
"a union of unknownlow-dimensional subspaces (lowrela-",
"points for outlier detection [17, 40]. Although computation-",
"tive to the dimension of the ambient space). A traditional",
"ally very efficient, these methods require the inliers to be",
"method for solving this problem is RANSAC[13], which is",
"well distributed and densely sampled within the subspaces."
] |
{
"arxiv_id": "1704.03925",
"language": "en",
"page_idx": 0
}
|
||
1704.03925-en-2
|
1704.03925-en-2.jpg
|
makes it more robust to outliers. However, LRR requires the subspaces to be independent and the sum of the union of subspaces to be low-dimensional [29].
Outlier detection by maximum consensus. In a diverse range of contexts such as maximum consensus [63,7] and robust linear regression [33, 49], people have studied problems of the form
(2)
$$\min\limits_{\mathbf{b}}\sum\limits_{i=1}^N\mathbb{I}(|\boldsymbol{x}_i^\top\mathbf{b}-y_i|\geq\epsilon),$$
in which $\mathbb{I}(\cdot)$ is the indicator function. Note that if we set $y_i$ = 1 for all $i$, then (2)can be interpreted as detecting outliers in data $X$ where the inliers lie close to an affine hyperplane. A problem closely related to (2) is
$$\min\limits_{\mathbf{b}}\sum\limits_{i=1}^{N}\mathbb{I}(|\boldsymbol{x}_{i}^{\top}\mathbf{b}|\geq\epsilon)\:\mathrm{s.t.}\:\mathbf{b}\neq0,$$
(3)
which appears in many applications (e.g. see[39]). In particular, (3) can be used to learn a linear hyperplane from data corrupted by outliers. To detect outliers in a general low-dimensional subspace, one can apply (2) and (3) recursively to find a basis for the orthogonal complement of the subspace [46]. However, such an approach is limited because there can be only one inlier subspace and the dimension of that subspace must be known in advance.
Outlier detection by random walk. Perhaps the most well-known random walk based algorithm is PageRank[3]. Originally introduced to determine the authority of website pages from web graphs, PageRank and its variants have been used in different contexts for ranking the centrality of the vertices in a graph. In particular,[34,35] propose the OutRank, which ranks the“outlierness”of points in a dataset by applying PageRank to an undirected graph in which the weight of an edge is the cosine similarity or RBF similarity between the two connected data points. Then, points that have low centrality are regarded as outliers. The outliers returned by OutRank are those that have low similarity to other data points. Therefore, OutRank does not work if points in a subspace are not dense enough.
$3. \textbf{Outlier detection by self- representation}$
In this section, we present our data self-representation based outlier detection method. We first describe the data self-representation and its associated properties for inliers and outliers. We then design a random walk algorithm on the representation graph whose limiting behavior allows us to identify the sets of inliers and outliers.
## 3.1. Data self-representation
Given an unlabeled dataset $X=[x_1,\cdots,x_N]$ containing inliers and outliers, the first step of our algorithm is
to construct the data self-representation matrix denoted by $R=[\boldsymbol{r}_1,\cdots,\boldsymbol{r}_N].$ As briefly discussed in the introduction (see also Figure 1), a self-representation matrix $R$ computed from (1) is observed to have different properties for inliers and outliers. Specifically, inliers usually use only other inliers for self-representation, i.e. for an inlier $x_j$, the representation is such that $r_{ij}\neq0$ only if $x_i$ is also an inlier, where $r_{ij}$ is the $(i,j)$-th entry of $R.$ This property is expected to hold if the inliers lie in a union of low dimensional subspaces, as evidenced from the works [12,43,59,52,50]. As an intuitive explanation, if the inliers lie in a low dimensional subspace, then any inlier has a sparse representation using other points in this subspace. Thus such a representation can be found by using sparsity-inducing regularization as seen in (1). In contrast, outliers are generally randomly distributed in the ambient space, so that a selfrepresentation usually contains both inliers and outliers.
Since the representation $R$ computed from (1) is sparse, there are potentially connectivity issues in the representation graph, i.e. an inlier that is not well-connected to other inliers may be detected as an outlier, and an outlier that is not well connected may be detected as an inlier. To address the connectivity issue, we compute the data selfrepresentation matrix $R$ by the elastic net problem [64,56]:
$$\min\limits_{\boldsymbol{r}_{j}}\lambda\|\boldsymbol{r}_{j}\|_{1}+\frac{1-\lambda}{2}\|\boldsymbol{r}_{j}\|_{2}^{2}+\frac{\gamma}{2}\|\boldsymbol{x}_{j}-X\boldsymbol{r}_{j}\|_{2}^{2}\:\mathrm{s.t.}\:r_{jj}=0,$$
(4)
in which $\lambda\in[0,1]$ controls the balance between sparse-
ness (via $\ell_1$ regularization) and connectivity (via $\ell_2$ regularization). Specifically, if $\lambda$ is chosen close to 1, we can still expect that the computed representation for an inlier will only use inliers. The $\ell_2$ regularization has been introduced to promote more connections between data points, i.e. if $\lambda\in[0,1)$, then one expects more nonzero entries in $R.$ A detailed discussion of the representation computed from (4) and the connectivity issue is provided in Section 4.
3.2. Representation graph and random walk
We use a directed graph $G$, which we call a representation graph, to capture the behavior of inliers and outliers from the representation matrix $R.$ The vertices of $G$ correspond to the data points $X$, and the edges are given by the (weighted) adjacency matrix $A:=\left|R\right|^{\breve{\top}}\in R^{N\times N}$ with the absolute value taken elementwise, i.e., the weight of the edge from $x_i$ to $x_j$ is given by $a_{ij}=|r_{ji}|.$ In the representation graph, we expect that vertices corresponding to inliers will have edges that only lead to inliers, while vertices that are outliers will have edges that lead to both inliers and outliers. In other words, we do not expect to have any edges that lead from an inlier to an outlier.
Using the previous paragraph as motivation, we design a random walk procedure to identify the outliers. A random walk on the representation graph $G$ is a discrete time
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.1060791015625,
0.89208984375,
0.106689453125,
0.89208984375,
0.121826171875,
0.50244140625,
0.12060546875
],
[
0.080078125,
0.12371826171875,
0.3505859375,
0.12371826171875,
0.3505859375,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.08087158203125,
0.14013671875,
0.468017578125,
0.14013671875,
0.468017578125,
0.154052734375,
0.08087158203125,
0.154052734375
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.0767822265625,
0.1558837890625,
0.468017578125,
0.1546630859375,
0.468017578125,
0.169189453125,
0.0767822265625,
0.17041015625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.1717529296875,
0.468017578125,
0.1717529296875,
0.468017578125,
0.1849365234375,
0.080078125,
0.1849365234375
],
[
0.5009765625,
0.1666259765625,
0.89208984375,
0.1673583984375,
0.89208984375,
0.1817626953125,
0.5009765625,
0.18115234375
],
[
0.080078125,
0.186279296875,
0.193603515625,
0.186279296875,
0.193603515625,
0.199462890625,
0.080078125,
0.199462890625
],
[
0.50341796875,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.50244140625,
0.1976318359375,
0.8896484375,
0.1962890625,
0.8896484375,
0.2108154296875,
0.50244140625,
0.212158203125
],
[
0.2197265625,
0.2083740234375,
0.2344970703125,
0.2083740234375,
0.2344970703125,
0.216552734375,
0.2197265625,
0.216552734375
],
[
0.178955078125,
0.216552734375,
0.369384765625,
0.216552734375,
0.369384765625,
0.2374267578125,
0.178955078125,
0.2374267578125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.44677734375,
0.2203369140625,
0.4697265625,
0.2203369140625,
0.4697265625,
0.2342529296875,
0.44677734375,
0.2342529296875
],
[
0.5048828125,
0.2298583984375,
0.8896484375,
0.2298583984375,
0.8896484375,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.2156982421875,
0.2392578125,
0.23779296875,
0.2392578125,
0.23779296875,
0.246826171875,
0.2156982421875,
0.246826171875
],
[
0.50244140625,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.2568359375,
0.468017578125,
0.2568359375,
0.468017578125,
0.270751953125,
0.080078125,
0.270751953125
],
[
0.50341796875,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.270751953125,
0.50341796875,
0.270751953125
],
[
0.07843017578125,
0.272705078125,
0.4697265625,
0.271484375,
0.4697265625,
0.285888671875,
0.07843017578125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.07843017578125,
0.285888671875,
0.4697265625,
0.287353515625,
0.4697265625,
0.3017578125,
0.07843017578125,
0.300537109375
],
[
0.50244140625,
0.2890625,
0.890625,
0.2890625,
0.890625,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.302978515625,
0.3896484375,
0.302978515625,
0.3896484375,
0.31640625,
0.080078125,
0.31640625
],
[
0.50244140625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.50341796875,
0.32080078125,
0.8896484375,
0.32080078125,
0.8896484375,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.1993408203125,
0.326904296875,
0.2132568359375,
0.326904296875,
0.2132568359375,
0.3359375,
0.1993408203125,
0.3359375
],
[
0.160888671875,
0.337158203125,
0.3896484375,
0.337158203125,
0.3896484375,
0.354736328125,
0.160888671875,
0.354736328125
],
[
0.50244140625,
0.333984375,
0.86767578125,
0.333984375,
0.86767578125,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.4453125,
0.34033203125,
0.4697265625,
0.34033203125,
0.4697265625,
0.352783203125,
0.4453125,
0.352783203125
],
[
0.51953125,
0.348388671875,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.51953125,
0.3623046875
],
[
0.19287109375,
0.35595703125,
0.2197265625,
0.35595703125,
0.2197265625,
0.367431640625,
0.19287109375,
0.367431640625
],
[
0.50244140625,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.080078125,
0.375732421875,
0.468017578125,
0.375732421875,
0.468017578125,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.5009765625,
0.37890625,
0.89208984375,
0.37744140625,
0.89208984375,
0.392578125,
0.5009765625,
0.393310546875
],
[
0.080078125,
0.391357421875,
0.4697265625,
0.391357421875,
0.4697265625,
0.40478515625,
0.080078125,
0.40478515625
],
[
0.50244140625,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.080078125,
0.406005859375,
0.468017578125,
0.406005859375,
0.468017578125,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.50244140625,
0.4091796875,
0.89208984375,
0.4091796875,
0.89208984375,
0.42236328125,
0.50244140625,
0.42236328125
],
[
0.080078125,
0.421630859375,
0.4697265625,
0.421630859375,
0.4697265625,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.50244140625,
0.4248046875,
0.890625,
0.4248046875,
0.890625,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.080078125,
0.436279296875,
0.468017578125,
0.436279296875,
0.468017578125,
0.45068359375,
0.080078125,
0.45068359375
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.60302734375,
0.46533203125,
0.64208984375,
0.46728515625,
0.64208984375,
0.47607421875,
0.60302734375,
0.473388671875
],
[
0.640625,
0.468505859375,
0.72119140625,
0.468505859375,
0.72119140625,
0.486083984375,
0.640625,
0.486083984375
],
[
0.70751953125,
0.4716796875,
0.88818359375,
0.4716796875,
0.88818359375,
0.486083984375,
0.70751953125,
0.486083984375
],
[
0.0767822265625,
0.48095703125,
0.398681640625,
0.481689453125,
0.398681640625,
0.49560546875,
0.0767822265625,
0.494384765625
],
[
0.5048828125,
0.4697265625,
0.6044921875,
0.4697265625,
0.6044921875,
0.4873046875,
0.5048828125,
0.4873046875
],
[
0.51123046875,
0.48486328125,
0.52685546875,
0.48486328125,
0.52685546875,
0.492431640625,
0.51123046875,
0.492431640625
],
[
0.87109375,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.505859375,
0.87109375,
0.505859375
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.5009765625,
0.505859375,
0.89208984375,
0.5068359375,
0.89208984375,
0.521484375,
0.5009765625,
0.52099609375
],
[
0.080078125,
0.51513671875,
0.46728515625,
0.51513671875,
0.46728515625,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.50244140625,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.08087158203125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.50244140625,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.50244140625,
0.55029296875
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50341796875,
0.5537109375,
0.890625,
0.5537109375,
0.890625,
0.564453125,
0.50341796875,
0.564453125
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.57568359375,
0.4697265625,
0.57568359375,
0.4697265625,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.07843017578125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.07843017578125,
0.60400390625
],
[
0.50341796875,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.50341796875,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.07843017578125,
0.6220703125,
0.468017578125,
0.619140625,
0.468017578125,
0.63330078125,
0.07843017578125,
0.63525390625
],
[
0.50341796875,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.50341796875,
0.642578125
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50341796875,
0.6513671875,
0.84228515625,
0.6513671875,
0.84228515625,
0.66650390625,
0.50341796875,
0.66650390625
],
[
0.08087158203125,
0.66748046875,
0.46728515625,
0.66748046875,
0.46728515625,
0.67822265625,
0.08087158203125,
0.67822265625
],
[
0.52294921875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.52294921875,
0.68798828125
],
[
0.080078125,
0.68115234375,
0.4697265625,
0.68115234375,
0.4697265625,
0.6943359375,
0.080078125,
0.6943359375
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.080078125,
0.6962890625,
0.41015625,
0.6962890625,
0.41015625,
0.71044921875,
0.080078125,
0.71044921875
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.07843017578125,
0.72265625,
0.4306640625,
0.7236328125,
0.4306640625,
0.73779296875,
0.07843017578125,
0.7373046875
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.50341796875,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.1004638671875,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76123046875,
0.1004638671875,
0.76123046875
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.78076171875,
0.50244140625,
0.78076171875
],
[
0.080078125,
0.77978515625,
0.471435546875,
0.77978515625,
0.471435546875,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.79443359375,
0.4697265625,
0.79443359375,
0.4697265625,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.080078125,
0.82373046875
],
[
0.50341796875,
0.8125,
0.890625,
0.8125,
0.890625,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.07843017578125,
0.82568359375,
0.351318359375,
0.82568359375,
0.351318359375,
0.8388671875,
0.07843017578125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.78125,
0.84130859375,
0.78125,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.8486328125,
0.298095703125,
0.8486328125,
0.298095703125,
0.8623046875,
0.080078125,
0.8623046875
],
[
0.51953125,
0.85498046875,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.51953125,
0.869140625
],
[
0.09722900390625,
0.87060546875,
0.468017578125,
0.87060546875,
0.468017578125,
0.88525390625,
0.09722900390625,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"makes it more robust to outliers. However, LRR requires",
"to construct the data self-representation matrix denoted by",
"the subspaces to be independent and the sum of the union",
"R=ri,...,rn].As briefly discussed in the introduction",
"of subspaces to be low-dimensional [29].",
"(seealsoFigureI),aself-representationmatrixRcomputed",
"Outlier detection by maximum consensus. In a diverse",
"from (1) is observed to have different properties for inliers",
"range of contexts such as maximum consensus [63, 7] and",
"and outliers. Specifically, inliers usually use only other in-",
"robust linear regression [33, 49], people have studied prob-",
"liers for self-representation, i.e. for an inlier j, the repre-",
"lems of the form",
"sentation is such that rij + O only if c; is also an inlier,",
"where rij is the (i, )-th entry of R. This property is ex-",
"N",
"min I(laeTb - yil ≥e),",
"pected to hold if the inliers lie in a union of low dimensional",
"(2)",
"subspaces,asevidencedfromtheworks[12.4359,52,50].",
"=1",
"As an intuitive explanation, if the inliers lie in a low di-",
"in which I() is the indicator function. Note that if we set",
"mensional subspace,thenany inlierhas a sparserepresen-",
"Yi = 1 for all i, then (2) can be interpreted as detecting",
"tation using other points in this subspace. Thus such a rep-",
"outliers in data X where the inliers lie close to an affine",
"resentation can be found by using sparsity-inducing regu-",
"hyperplane. A problem closely related to (2) is",
"larization as seen in (1). In contrast, outliers are generally",
"randomly distributed in the ambient space.so that a self-",
"N",
"min I(laTbl ≥e) s.t. b + 0,",
"representation usually contains both inliers and outliers.",
"(3)",
"Sincethe representation Rcomputedfrom (1)is sparse,",
"2=1",
"there are potentially connectivity issues in the representa-",
"which appears in many applications (e.g. see [39]). In par-",
"tion graph, i.e. an inlier that is not well-connected to other",
"ticular, (3) can be used to learn a linear hyperplane from",
"inliers may be detected as an outlier, and an outlier that",
"data corrupted by outliers. To detect outliers in a general",
"is not well connected may be detected as an inlier.To",
"low-dimensional subspace, one can apply (2) and (3) recur-",
"address the connectivity issue, we compute the data self-",
"sively to find a basis for the orthogonal complement of the",
"representation matrix R by the elastic net problem [64, 56]:",
"subspace [46]. However, such an approach is limited be-",
"cause there can be only one inlier subspace and the dimen-",
"1-入",
"r照+",
"la,-Xr,l2s.t.rij=0",
"sion ofthat subspacemustbeknown in advance.",
"min >/r;ll1 +",
"T",
"(4)",
"Outlier detection by random walk.Perhaps the most",
"in which > E [o,1] controls the balance between sparse-",
"well-known random walk based algorithm is PageRank [3]",
"ness (via li regularization) and connectivity (via l2 regu-",
"Originally introduced to determine the authority of website",
"larization). Specifically, if is chosen close to 1, we can",
"pages from web graphs, PageRank and its variants have",
"still expect that the computed representation for an inlier",
"been used in different contexts for ranking the centrality",
"will only use inliers. The l2 regularization has been intro-",
"of the vertices in a graph. In particular, [34, 35] propose",
"duced to promote more connections between data points,",
"the OutRank, which ranks the “outlierness\" of points in",
"i.e.if X E[o,1).then oneexpectsmorenonzero entriesin",
"a dataset by applying PageRank to an undirected graph in",
"R. A detailed discussion of the representation computed",
"which the weight of an edge is the cosine similarity or RBF",
"from (4) and the connectivity issue is provided in Section 4.",
"similarity between the two connected data points.Then,",
"points that have low centrality are regarded as outliers. The",
"3.2. Representation graph and random walk",
"outliers returnedbyOutRankarethosethathavelow sim-",
"We use a directed graph G, which we call a represen-",
"ilarity to other data points. Therefore, OutRank does not",
"tation graph, to capture the behavior of inliers and outliers",
"work if points in a subspace are not dense enough.",
"from the representation matrix R. The vertices of G cor-",
"3.Outlier detection by self-representation",
"respond to the data points X, and the edges are given by",
"the (weighted) adjacency matrix A := R|T E IRNxN with",
"In this section,wepresent ourdata self-representation",
"the absolute value taken elementwise, i.e., the weight of the",
"based outlier detection method. We first describe the data",
"edge from c, to &, is given by ai = rizl. In the rep-",
"self-representation and its associated properties for inliers",
"resentationgraph, we expectthat vertices corresponding to",
"and outliers. We then design a random walk algorithm on",
"inliers will have edges that only lead to inliers, while ver-",
"the representation graph whose limiting behavior allows us",
"ticesthatare outliers will haveedges that lead to both inliers",
"to identify the sets of inliers and outliers.",
"and outliers. In other words, we do not expect to have any",
"edges that lead from an inlier to an outlier.",
"3.1.Data self-representation",
"Using the previous paragraph as motivation, we design",
"Given an unlabeled dataset X = [αi, ::: , &nl contain-",
"a random walk procedure to identify the outliers. A ran-",
"ing inliers and outliers, the first step of our algorithm is",
"dom walk on the representation graph G is a discrete time"
] |
{
"arxiv_id": "1704.03925",
"language": "en",
"page_idx": 2
}
|
||
1704.03925-en-4
|
1704.03925-en-4.jpg
|
Definition 4.1 (Essential and inessential state[23]). A state $i\in \Omega \textit{ is essential iffor all j such that i }\to j\textit{ it is also true}$ that $j\to i.$ $A$ state $is$ inessential $if$ $it$ $is$ not $essential.$
Our aim in this section is to establish that if inliers connect to themselves, i.e. they are subspace-preserving (Section 4.1), and the representation $R$ satisfies certain connectivity conditions (Section 4.2), then inliers are essential states of the Markov chain and outliers are inessential states. Subsequently, in Section 4.3 we show that the Cesàro mean (6) identifies essential and inessential states, thus establishing the correctness of Algorithm 1 for outlier detection.
$4. 1. \textbf{ Subspace- preserving representation}$
We first establish that inliers express themselves with only other inliers when they lie in a union of low dimensional subspaces. This property is well-studied in the subspace clustering literature. We will borrow terminologies and results from prior work and modify them for our current task of outlier detection.
Definition 4.2 (Subspace-preserving representation [47]).
$Ifx_j\in\mathcal{S}_\ell$ is an inlier, then the representation $r_j\in R^N$
$is\textit{called subspace-preserving if the nonzero entries of r}_{j}$
$\begin{array}{l}{correspond~to~points~in~S_{\ell},~i.e.~r_{ij}~\neq~0~only~if~x_{i}\in\mathcal{S}_{\ell}.}\\{The~representation~matrix~R=~[r_{1},\cdots,r_{N}]\in~R^{N\times N}~is}\end{array}$
$calledsubspace-preservingif\boldsymbol{r}_jissubspace-preservingfor$
every inlier $x_j.$
A representation matrix $R$ is subspace-preserving if each
inlier uses points in its own subspace for representation.
Given $X$, a subspace-preserving representation $R$ can be
obtained by solving (4) when certain geometric conditions
hold. The following result is modified from [56]. It assumes
that columns of $X$ are normalized to have unit $\ell_{2}$-norm.
$\textbf{Theorem 4. 1. Let }x_{j}\in \mathcal{S} _{\ell }$ $be$ $an$ $inlier.$ Define the oracle point $of$ $x_{j}$ $to$ $be$ $\delta _{j}: = \gamma \cdot ( x_{j}- X_{- j}^{\ell }\cdot r_{j}^{\ell }) $, where $X_{- j}^{\ell }$ $is$ the matrix containing all points $in$ $S_{\ell }$ except $x_{j}$ and
$$r_j^\ell:=\arg\min\limits_{\boldsymbol{r}}\lambda\|\boldsymbol{r}\|_1+\frac{1-\lambda}{2}\|\boldsymbol{r}\|_2^2+\frac{\gamma}{2}\|\boldsymbol{x}_j-X_{-j}^\ell\boldsymbol{r}\|_2^2.$$
The solution $r_j\textit{ to ( 4) is subspace- preserving if}$
(7)
$$\max_{k\neq j,\boldsymbol{x}_k\in\mathcal{S}_\ell}|\langle\boldsymbol{x}_k,\bar{\boldsymbol{\delta}}_j\rangle|-\max_{k:\boldsymbol{x}_k\notin\mathcal{S}_\ell}|\langle\boldsymbol{x}_k,\bar{\boldsymbol{\delta}}_j\rangle|>\frac{1-\lambda}{\lambda},$$
where $\bar{\boldsymbol{\delta }} _{j}: = \boldsymbol{\delta }_{j}/ \| \boldsymbol{\delta }_{j}\| _{2}.$
An outline of the proof is given in the appendix. Note that the oracle point $\delta_j$ lies in $S_\ell$ and that its definition only depends on points in $\mathcal{S}_\ell.$ The first term in condition (7) captures the distribution of points in $\mathcal{S}_\ell$ near $\bar{\delta}_j$, and is expected to be large if the neighborhood of $\bar{\delta}_j$ is well-covered by points from $\mathcal{S}_\ell.$ The second term characterizes the similarity between the oracle point $\bar{\delta}_j$ and all other data points,
which includes the outliers and the inliers from other subspaces. The condition requires the former to be larger than the latter by a margin of $\frac{1-\lambda}\lambda$, which is close to zero if $\lambda$ is close to 1. Overall, condition (7) requires that points in $S_{\ell}$ are dense around $\delta_j$, which is itself in $\mathcal{S}_\ell$, and that outliers and inliers from other subspaces do not lie close to $\bar{\delta}_j.$
Even if (7) holds for all $j$ so that the representation $R$ is subspace-preserving, we cannot automatically establish an equivalence between inliers/outliers and essential/inessential states because of potential complications related to the graph's connectivity. This is addressed next.
### 4.2. Connectivity considerations
In the context of sparse subspace clustering, the wellknown connectivity issue[36,53,30,56,51] refers to the problem that points in the same subspace may not be wellconnected in the representation graph, which may cause oversegmentation of the true clusters. Thus, one has to make the assumption that each true cluster is connected to guarantee correct clustering. For the outlier detection problem, it may happen that an inlier is inessential and thus classified as an outlier when the inliers are not well-connected; similarly, an outlier may be essential and thus classified as an inlier if it is not connected to at least one inlier. In fact, the situation is even more involved since the representation graph is directed and inliers and outliers behave differently.
Suppose, as a first example, that there exists an inlier that is never used to express any other inliers. This is equivalent to saying that there is no edge going into this point from any other inliers. Note that the subspace-preserving property can still hold if this inlier expresses itself using other inliers. Yet, since a random walk leaving this point would never return it can not be identified as an inlier. To avoid such cases, we need the following assumption.
Assumption 4.1. For any inlier subspace Se, the vertices $\{x_j\in\mathcal{S}_\ell\}$ of the representation graph are strongly connected, i.e. there is a path in each direction between each pair of vertices.
Assumption 4.1 requires good connectivity between points from the same inlier subspace. We also need good connectivity between outliers and inliers. Consider the example when there is a subset of outliers for which all of their outgoing edges lead only to points within that same subset. In this case, the subset of points can not be detected as outliers since their representation pattern is the same as for the inliers. The next assumption rules out this case.
Assumption 4.2. For each subset of outliers there exists an edge in the representation graph that goes from a point in this subset to an inlier or to an outlier outside this subset.
|
[
[
0.08087158203125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.09027099609375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.10479736328125
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08087158203125,
0.125,
0.42236328125,
0.125,
0.42236328125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.50244140625,
0.121826171875,
0.89208984375,
0.121826171875,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.50244140625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.09722900390625,
0.146484375,
0.468017578125,
0.146484375,
0.468017578125,
0.160400390625,
0.09722900390625,
0.160400390625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.08087158203125,
0.1634521484375,
0.46728515625,
0.1634521484375,
0.46728515625,
0.173583984375,
0.08087158203125,
0.173583984375
],
[
0.5009765625,
0.1666259765625,
0.85888671875,
0.1685791015625,
0.85888671875,
0.1817626953125,
0.5009765625,
0.179931640625
],
[
0.080078125,
0.175537109375,
0.46728515625,
0.175537109375,
0.46728515625,
0.189453125,
0.080078125,
0.189453125
],
[
0.52294921875,
0.1849365234375,
0.890625,
0.1849365234375,
0.890625,
0.1988525390625,
0.52294921875,
0.1988525390625
],
[
0.080078125,
0.1912841796875,
0.468017578125,
0.1912841796875,
0.468017578125,
0.2052001953125,
0.080078125,
0.2052001953125
],
[
0.50244140625,
0.1988525390625,
0.890625,
0.199462890625,
0.890625,
0.2139892578125,
0.50244140625,
0.21337890625
],
[
0.08087158203125,
0.20703125,
0.468017578125,
0.20703125,
0.468017578125,
0.2203369140625,
0.08087158203125,
0.2203369140625
],
[
0.080078125,
0.2215576171875,
0.4697265625,
0.2215576171875,
0.4697265625,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.50341796875,
0.2171630859375,
0.8896484375,
0.2171630859375,
0.8896484375,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.50244140625,
0.2310791015625,
0.890625,
0.2310791015625,
0.890625,
0.2442626953125,
0.50244140625,
0.2442626953125
],
[
0.08251953125,
0.238037109375,
0.46728515625,
0.238037109375,
0.46728515625,
0.248779296875,
0.08251953125,
0.248779296875
],
[
0.080078125,
0.251953125,
0.44287109375,
0.251953125,
0.44287109375,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50244140625,
0.246826171875,
0.87109375,
0.246826171875,
0.87109375,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.080078125,
0.274658203125,
0.38818359375,
0.27587890625,
0.38818359375,
0.290283203125,
0.080078125,
0.2890625
],
[
0.50244140625,
0.27587890625,
0.7490234375,
0.276611328125,
0.7490234375,
0.291748046875,
0.50244140625,
0.290283203125
],
[
0.09722900390625,
0.29736328125,
0.4697265625,
0.298583984375,
0.4697265625,
0.313232421875,
0.09722900390625,
0.3125
],
[
0.521484375,
0.302978515625,
0.890625,
0.302978515625,
0.890625,
0.31640625,
0.521484375,
0.31640625
],
[
0.080078125,
0.314453125,
0.468017578125,
0.314453125,
0.468017578125,
0.32763671875,
0.080078125,
0.32763671875
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50244140625,
0.33203125
],
[
0.080078125,
0.330078125,
0.468017578125,
0.330078125,
0.468017578125,
0.343505859375,
0.080078125,
0.343505859375
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.080078125,
0.3447265625,
0.4697265625,
0.3447265625,
0.4697265625,
0.35791015625,
0.080078125,
0.35791015625
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.359130859375,
0.468017578125,
0.359130859375,
0.468017578125,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.50244140625,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.080078125,
0.375,
0.271240234375,
0.375,
0.271240234375,
0.38818359375,
0.080078125,
0.38818359375
],
[
0.50244140625,
0.37890625,
0.89208984375,
0.37890625,
0.89208984375,
0.392578125,
0.50244140625,
0.392578125
],
[
0.080078125,
0.396484375,
0.46728515625,
0.397705078125,
0.46728515625,
0.412353515625,
0.080078125,
0.410888671875
],
[
0.50341796875,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.40771484375,
0.50341796875,
0.40771484375
],
[
0.07843017578125,
0.41357421875,
0.465576171875,
0.412353515625,
0.465576171875,
0.4267578125,
0.07843017578125,
0.427978515625
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.07843017578125,
0.427978515625,
0.46728515625,
0.429931640625,
0.46728515625,
0.443115234375,
0.07843017578125,
0.44140625
],
[
0.5048828125,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.436279296875,
0.5048828125,
0.436279296875
],
[
0.08087158203125,
0.4443359375,
0.46728515625,
0.4443359375,
0.46728515625,
0.458984375,
0.08087158203125,
0.458984375
],
[
0.50244140625,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.07598876953125,
0.455810546875,
0.47216796875,
0.453857421875,
0.47216796875,
0.474853515625,
0.07598876953125,
0.4765625
],
[
0.50341796875,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.466552734375,
0.50341796875,
0.466552734375
],
[
0.080078125,
0.474853515625,
0.4697265625,
0.474853515625,
0.4697265625,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.50341796875,
0.47021484375,
0.890625,
0.47021484375,
0.890625,
0.48095703125,
0.50341796875,
0.48095703125
],
[
0.07843017578125,
0.488037109375,
0.181396484375,
0.4892578125,
0.181396484375,
0.50390625,
0.07843017578125,
0.50244140625
],
[
0.50341796875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.51953125,
0.501953125,
0.89208984375,
0.50048828125,
0.89208984375,
0.51513671875,
0.51953125,
0.5166015625
],
[
0.1004638671875,
0.51220703125,
0.468017578125,
0.51220703125,
0.468017578125,
0.5263671875,
0.1004638671875,
0.5263671875
],
[
0.50244140625,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.52783203125,
0.468017578125,
0.52783203125,
0.468017578125,
0.54150390625,
0.080078125,
0.54150390625
],
[
0.50244140625,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.546875,
0.50244140625,
0.546875
],
[
0.08087158203125,
0.54345703125,
0.4697265625,
0.54345703125,
0.4697265625,
0.556640625,
0.08087158203125,
0.556640625
],
[
0.5009765625,
0.546875,
0.890625,
0.54736328125,
0.890625,
0.5625,
0.5009765625,
0.56103515625
],
[
0.080078125,
0.55810546875,
0.4697265625,
0.556640625,
0.4697265625,
0.5712890625,
0.080078125,
0.57275390625
],
[
0.50341796875,
0.564453125,
0.8896484375,
0.564453125,
0.8896484375,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.080078125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.080078125,
0.5869140625
],
[
0.50244140625,
0.5771484375,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59033203125
],
[
0.080078125,
0.5888671875,
0.44677734375,
0.5888671875,
0.44677734375,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.08087158203125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.08087158203125,
0.62548828125
],
[
0.5009765625,
0.60693359375,
0.775390625,
0.60693359375,
0.775390625,
0.62060546875,
0.5009765625,
0.62060546875
],
[
0.080078125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.6416015625,
0.080078125,
0.6416015625
],
[
0.080078125,
0.642578125,
0.42236328125,
0.642578125,
0.42236328125,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.5048828125,
0.64013671875,
0.890625,
0.64013671875,
0.890625,
0.65380859375,
0.5048828125,
0.65380859375
],
[
0.50341796875,
0.65576171875,
0.890625,
0.65478515625,
0.890625,
0.66943359375,
0.50341796875,
0.67041015625
],
[
0.253173828125,
0.67041015625,
0.279296875,
0.67041015625,
0.279296875,
0.67822265625,
0.253173828125,
0.67822265625
],
[
0.371826171875,
0.671875,
0.459228515625,
0.673828125,
0.459228515625,
0.69140625,
0.371826171875,
0.689453125
],
[
0.50341796875,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50341796875,
0.68505859375
],
[
0.28759765625,
0.67529296875,
0.330810546875,
0.67529296875,
0.330810546875,
0.68994140625,
0.28759765625,
0.68994140625
],
[
0.125,
0.67529296875,
0.2459716796875,
0.67529296875,
0.2459716796875,
0.68994140625,
0.125,
0.68994140625
],
[
0.50244140625,
0.68603515625,
0.609375,
0.68603515625,
0.609375,
0.69970703125,
0.50244140625,
0.69970703125
],
[
0.08087158203125,
0.705078125,
0.390625,
0.705078125,
0.390625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.52294921875,
0.71728515625,
0.890625,
0.71728515625,
0.890625,
0.73193359375,
0.52294921875,
0.73193359375
],
[
0.158447265625,
0.734375,
0.43310546875,
0.7236328125,
0.4345703125,
0.75244140625,
0.16015625,
0.76318359375
],
[
0.50341796875,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.11517333984375,
0.74072265625,
0.1544189453125,
0.74072265625,
0.1544189453125,
0.75048828125,
0.11517333984375,
0.75048828125
],
[
0.4453125,
0.7392578125,
0.468017578125,
0.7392578125,
0.468017578125,
0.751953125,
0.4453125,
0.751953125
],
[
0.098876953125,
0.74853515625,
0.1478271484375,
0.74853515625,
0.1478271484375,
0.76025390625,
0.098876953125,
0.76025390625
],
[
0.406005859375,
0.751953125,
0.41259765625,
0.751953125,
0.41259765625,
0.7568359375,
0.406005859375,
0.7568359375
],
[
0.50341796875,
0.74853515625,
0.8896484375,
0.74853515625,
0.8896484375,
0.76220703125,
0.50341796875,
0.76220703125
],
[
0.5009765625,
0.76318359375,
0.89208984375,
0.76220703125,
0.89208984375,
0.7763671875,
0.5009765625,
0.77783203125
],
[
0.253173828125,
0.75244140625,
0.2607421875,
0.75244140625,
0.2607421875,
0.7568359375,
0.253173828125,
0.7568359375
],
[
0.080078125,
0.76953125,
0.2344970703125,
0.771484375,
0.2344970703125,
0.78857421875,
0.080078125,
0.7861328125
],
[
0.50341796875,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50341796875,
0.79248046875
],
[
0.1004638671875,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.1004638671875,
0.81005859375
],
[
0.5048828125,
0.79541015625,
0.8896484375,
0.79541015625,
0.8896484375,
0.80615234375,
0.5048828125,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.8095703125,
0.89208984375,
0.8095703125,
0.89208984375,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.080078125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.82470703125,
0.81787109375,
0.82470703125,
0.81787109375,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.08251953125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08251953125,
0.86865234375
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50341796875,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.88134765625,
0.88720703125,
0.88134765625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Definition 4.1 (Essential and inessential state [23]). A state",
"which includes the outliers and the inliers from other sub-",
"iE Qisessential iffor all isuchthati iitisalsotrue",
"spaces. The condition requires the former to be larger than",
"that i i.A state is inessential ifit is not essential.",
"the latter by a margin of l-, which is close to zero if > is",
"close to 1. Overall, condition (7) requires that points in Se",
"Our aim in this section is to establish that if inliers con-",
"are dense around ,, which is itself in Se, and that outliers",
"necttothemselves,i.e.they aresubspace-preserving (Sec-",
"and inliers from other subspaces do not lie close to d,:",
"tion 4.1), and the representation R satisfies certain con-",
"Even if (7) holds for all j so that the representation",
"nectivity conditions (Section 4.2), then inliers are essential",
"R is subspace-preserving, we cannot automatically es-",
"states of the Markov chain and outliers are inessential states.",
"Subsequently, in Section 4.3 we show that the Cesaro mean",
"tablish an equivalence between inliers/outliers and essen-",
"tial/inessential states because of potential complications re-",
"(6)identifiesessential and inessential states,thusestablish-",
"ing the correctness of Algorithm I for outlier detection.",
"lated to the graph's connectivity. This is addressed next.",
"4.1. Subspace-preserving representation",
"4.2.Connectivity considerations",
"We first establish that inliers express themselves with",
"In the context of sparse subspace clustering, the well-",
"only other inliers when they lie in a union of low dimen-",
"known connectivity issue [36, 53, 30, 56, 51] refers to the",
"sional subspaces. This property is well-studied in the sub-",
"problem that points in the same subspace may not be well-",
"space clustering literature.We will borrow terminologies",
"connected in the representation graph, which may cause",
"and results from prior work and modify them for our cur-",
"oversegmentation of the true clusters.Thus, one has to",
"rent task of outlier detection.",
"make the assumption that each true cluster is connected to",
"Definition 4.2 (Subspace-preserving representation [47]).",
"guarantee correct clustering. For the outlier detection prob-",
"If &g E Se is an inlier then the representation r E RN",
"lem, it may happen that an inlier is inessential and thus clas-",
"is called subspace-preserving if the nonzero entries of rj",
"sified as an outlier when the inliers arenot well-connected:",
"correspond to points in Se, i.e. Tij / O only if &i E Se.",
"similarly, an outlier may be essential and thus classified as",
"The representation matrix R = [ri,...,rN] e IRNxN is",
"an inlier if it is not connected to at least one inlier.In fact,",
"called subspace-preserving if r, is subspace-preserving for",
"thesituationisevenmoreinvolvedsincetherepresentation",
"every inlier aj.",
"graph is directed and inliers and outliers behave differently.",
"Suppose, as a first example, that there exists an inlier that",
"A representation matrix R is subspace-preserving if each",
"is never used to express any other inliers. This is equivalent",
"inlier uses points in its own subspace for representation.",
"to saying that there is no edge going into this point from any",
"Given X, a subspace-preserving representation R can be",
"other inliers. Note that the subspace-preserving property",
"obtained by solving (4) when certain geometric conditions",
"canstillhold if thisinlierexpressesitselfusing otherinliers.",
"hold. The following result is modified from [56]. It assumes",
"Yet, since a random walk leaving this point would never",
"that columns of X are normalized to have unit l2-norm.",
"return it can not be identified as an inlier. To avoid such",
"Theorem 4.1. Let ri E Se be an inlier Define the oracle",
"cases, we need the following assumption.",
"point of a, to be d, :=-(cj- X-, -r'), where X-, is",
"the matrix containing all points in Se except , and",
"Assumption 4.1. For any inlier subspace Se, the vertices",
"[aj E Se) of the representation graph are strongly con-",
"-",
"-X-,rll2",
"nected, i.e. there is a path in each direction between each",
"rl2 -",
" arg min ^rl1 +",
"pair of vertices.",
"The solution r; to (4) is subspace-preserving if",
"Assumption 4.1 requires good connectivity between",
"[Kak, 0,) - Lmax, (<a,0,)/> 1二,",
"points from the same inlier subspace. We also need good",
"max",
"(7)",
"ktj,ak",
"A",
"connectivity between outliers and inliers. Consider the ex-",
"ample when there is a subset of outliers for which all of their",
"-.",
"where8,:=8;/18;ll2",
"outgoing edges lead only to points within that same subset.",
"An outline of the proof is given in the appendix. Note",
"In this case,the subset of points can not be detected as out-",
"that the oracle point , lies in Se and that its definition only",
"liers since their representation pattern is the same as for the",
"depends on points in Se. The first term in condition (7)",
"inliers. The next assumption rules out this case.",
"captures the distribution of points in Se near , and is ex-",
"pectedtobelargeiftheneighborhood of ,is well-covered",
"Assumption 4.2. For each subset of outliers there exists an",
"by points from Se. The second term characterizes the simi-",
"edge in the representation graph that goes from a point in",
"larity between the oracle point , and all other data points,",
"this subset to an inlier or to an outlier outside this subset."
] |
{
"arxiv_id": "1704.03925",
"language": "en",
"page_idx": 4
}
|
||
1703.08603-en-0
|
1703.08603-en-0.jpg
|
Adversarial Examples for Semantic Segmentation and Object Detection
Cihang Xie$^{1*}$, Jianyu Wang$^{2*}$, Zhishuai Zhang$^{1*}$, Yuyin Zhou$^1$, Lingxi Xie$^1$, Alan Yuille$^1$
$^{1}\ddot{\text{Department of Computer Science, The Johns Hopkins University, Baltimore, MD 21218 USA}}$
$^{2}$Baidu Research USA, Sunnyvale, CA 94089 USA
$\{$cihangxie306, wjyouch, zhshuai.zhang, zhouyuyiner, 198808xc, alan.l.yuille$\}$@gmail.com
## $\mathbf{Abstract}$
It has been well demonstrated that adversarial examples, $i.e.$, natural images with visually imperceptible perturbations added, cause deep networks to fail on image classification. In this paper, we extend adversarial examples to semantic segmentation and object detection which are much more difficult. Our observation is that both segmentation and detection are based on classifying multiple targets on an image (e.g., the target is a pixel or a receptive freld in segmentation, and an object proposal in detection). This inspires us to optimize a loss function over a set of pixels/proposals for generating adversarial perturbations. Based on this idea, we propose a novel algorithm named Dense Adversary Generation (DAG), which generates a large family of adversarial examples, and applies to a wide range of state-of-the-art deep networks for segmentation and detection. We also find that the adversarial perturbations can be transferred across networks with different training data, based on different architectures, and even for different recognition tasks. In particular, the transferability across networks with the same architecture is more significant than in other cases. Besides, summing up heterogeneous perturbations often leads to better transfer performance, which provides an effective method of blackbox adversarial attack.
## $1. \textbf{Introduction}$
Convolutional Neural Networks (CNN) [14][31][32][13] have become the state-of-the-art solution for a wide range of visual recognition problems. Based on a large-scale labeled dataset such as ImageNet [6] and powerful computational resources like modern GPUs, it is possible to train a hierarchical deep network to capture different levels of visual patterns. A deep network is also capable of generating $\hat{\text{transferrable features for different tasks such as image clas-}}$ sification [7] and instance retrieval [29], or being fine-tuned
$^*$The first three authors contributed equally to this work. This work was
done when Jianyu Wang was a Ph.D. student at UCLA
<FigureHere>
Figure 1: An Adversarial example for semantic segmen-
tation and object detection. FCN[19] is used for segmentation, and Faster-RCNN[28] is used for detection. Left column: the original image (top row) with the normal segmentation (the purple region is predicted as $dog)$ and detection results. Right column: after the adversarial perturbation (top row, magnified by 10) is added to the original image, both segmentation (the light green region as train and the pink region as person) and detection results are completely wrong. Note that, though the added perturbation can confuse both networks, it is visually imperceptible (the maximal absolute intensity in each channel is less than 10). to deal with a wide range of vision tasks, including object detection [11][10], visual concept discovery [35], semantic segmentation [19][3][38], boundary detection [36][30], etc.
1
|
[
[
0.1298828125,
0.1337890625,
0.8408203125,
0.1337890625,
0.8408203125,
0.1483154296875,
0.1298828125,
0.1483154296875
],
[
0.1380615234375,
0.18115234375,
0.8310546875,
0.179931640625,
0.8310546875,
0.1976318359375,
0.1380615234375,
0.1988525390625
],
[
0.11517333984375,
0.1988525390625,
0.85693359375,
0.1988525390625,
0.85693359375,
0.216552734375,
0.11517333984375,
0.216552734375
],
[
0.28515625,
0.2183837890625,
0.685546875,
0.2183837890625,
0.685546875,
0.231689453125,
0.28515625,
0.231689453125
],
[
0.1282958984375,
0.236083984375,
0.84619140625,
0.236083984375,
0.84619140625,
0.250732421875,
0.1282958984375,
0.250732421875
],
[
0.0236968994140625,
0.271484375,
0.058013916015625,
0.271484375,
0.058013916015625,
0.335205078125,
0.0236968994140625,
0.335205078125
],
[
0.238525390625,
0.287353515625,
0.313720703125,
0.287353515625,
0.313720703125,
0.3017578125,
0.238525390625,
0.3017578125
],
[
0.1004638671875,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.33203125,
0.1004638671875,
0.33203125
],
[
0.0310516357421875,
0.333251953125,
0.053924560546875,
0.333251953125,
0.053924560546875,
0.356689453125,
0.0310516357421875,
0.356689453125
],
[
0.08087158203125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.08087158203125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.03594970703125,
0.367431640625,
0.05963134765625,
0.385009765625,
0.0400390625,
0.399658203125,
0.01715087890625,
0.3818359375
],
[
0.08251953125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.376953125,
0.08251953125,
0.376953125
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.080078125,
0.392578125
],
[
0.08087158203125,
0.396484375,
0.465576171875,
0.396484375,
0.465576171875,
0.4072265625,
0.08087158203125,
0.4072265625
],
[
0.0269622802734375,
0.4072265625,
0.052276611328125,
0.4072265625,
0.052276611328125,
0.4375,
0.0269622802734375,
0.4375
],
[
0.080078125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.080078125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.080078125,
0.439453125
],
[
0.0318603515625,
0.43310546875,
0.037567138671875,
0.43310546875,
0.037567138671875,
0.4443359375,
0.0318603515625,
0.4443359375
],
[
0.080078125,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.03350830078125,
0.451904296875,
0.0555419921875,
0.451904296875,
0.0555419921875,
0.477783203125,
0.03350830078125,
0.477783203125
],
[
0.07843017578125,
0.453857421875,
0.4697265625,
0.455078125,
0.4697265625,
0.4697265625,
0.07843017578125,
0.468505859375
],
[
0.08087158203125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.08087158203125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.0318603515625,
0.496826171875,
0.053924560546875,
0.50048828125,
0.0482177734375,
0.51513671875,
0.0269622802734375,
0.51123046875
],
[
0.08087158203125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.02777099609375,
0.5068359375,
0.053924560546875,
0.5068359375,
0.053924560546875,
0.54150390625,
0.02777099609375,
0.54150390625
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.0310516357421875,
0.546875,
0.049835205078125,
0.546875,
0.049835205078125,
0.56103515625,
0.0310516357421875,
0.56103515625
],
[
0.080078125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.02777099609375,
0.5546875,
0.053924560546875,
0.5546875,
0.053924560546875,
0.5869140625,
0.02777099609375,
0.5869140625
],
[
0.037567138671875,
0.5595703125,
0.04412841796875,
0.5595703125,
0.04412841796875,
0.5625,
0.037567138671875,
0.5625
],
[
0.080078125,
0.56103515625,
0.4697265625,
0.56103515625,
0.4697265625,
0.57470703125,
0.080078125,
0.57470703125
],
[
0.080078125,
0.5771484375,
0.4697265625,
0.5771484375,
0.4697265625,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.0318603515625,
0.5888671875,
0.0555419921875,
0.59765625,
0.0400390625,
0.6220703125,
0.01715087890625,
0.61279296875
],
[
0.080078125,
0.59033203125,
0.46728515625,
0.59033203125,
0.46728515625,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.035125732421875,
0.60302734375,
0.049835205078125,
0.60302734375,
0.049835205078125,
0.62255859375,
0.035125732421875,
0.62255859375
],
[
0.080078125,
0.60791015625,
0.4697265625,
0.60791015625,
0.4697265625,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.03350830078125,
0.6162109375,
0.037567138671875,
0.6162109375,
0.037567138671875,
0.62890625,
0.03350830078125,
0.62890625
],
[
0.08087158203125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.08087158203125,
0.63525390625
],
[
0.03350830078125,
0.63720703125,
0.053924560546875,
0.63720703125,
0.053924560546875,
0.67236328125,
0.03350830078125,
0.67236328125
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.080078125,
0.6513671875,
0.46728515625,
0.6513671875,
0.46728515625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.0294189453125,
0.66552734375,
0.0555419921875,
0.66552734375,
0.0555419921875,
0.70849609375,
0.0294189453125,
0.70849609375
],
[
0.080078125,
0.66650390625,
0.2320556640625,
0.66650390625,
0.2320556640625,
0.6796875,
0.080078125,
0.6796875
],
[
0.50341796875,
0.662109375,
0.890625,
0.662109375,
0.890625,
0.67529296875,
0.50341796875,
0.67529296875
],
[
0.50341796875,
0.6787109375,
0.6953125,
0.67822265625,
0.6953125,
0.689453125,
0.50341796875,
0.68994140625
],
[
0.6953125,
0.6767578125,
0.8896484375,
0.67822265625,
0.8896484375,
0.69140625,
0.6953125,
0.68994140625
],
[
0.50244140625,
0.68994140625,
0.8896484375,
0.68994140625,
0.8896484375,
0.705078125,
0.50244140625,
0.705078125
],
[
0.04412841796875,
0.681640625,
0.049835205078125,
0.681640625,
0.049835205078125,
0.6923828125,
0.04412841796875,
0.6923828125
],
[
0.08251953125,
0.70849609375,
0.2091064453125,
0.70849609375,
0.2091064453125,
0.72265625,
0.08251953125,
0.72265625
],
[
0.50244140625,
0.70703125,
0.890625,
0.70703125,
0.890625,
0.72021484375,
0.50244140625,
0.72021484375
],
[
0.50244140625,
0.72265625,
0.89208984375,
0.72265625,
0.89208984375,
0.736328125,
0.50244140625,
0.736328125
],
[
0.098876953125,
0.73291015625,
0.468017578125,
0.73291015625,
0.468017578125,
0.7470703125,
0.098876953125,
0.7470703125
],
[
0.50341796875,
0.7392578125,
0.890625,
0.7392578125,
0.890625,
0.75,
0.50341796875,
0.75
],
[
0.080078125,
0.74853515625,
0.4697265625,
0.74853515625,
0.4697265625,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50244140625,
0.75244140625,
0.89208984375,
0.75244140625,
0.89208984375,
0.7666015625,
0.50244140625,
0.7666015625
],
[
0.08087158203125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.08087158203125,
0.7763671875
],
[
0.50244140625,
0.76806640625,
0.890625,
0.76806640625,
0.890625,
0.7822265625,
0.50244140625,
0.7822265625
],
[
0.080078125,
0.77880859375,
0.468017578125,
0.77880859375,
0.468017578125,
0.79248046875,
0.080078125,
0.79248046875
],
[
0.50341796875,
0.78515625,
0.8896484375,
0.78515625,
0.8896484375,
0.79541015625,
0.50341796875,
0.79541015625
],
[
0.08087158203125,
0.79541015625,
0.465576171875,
0.79541015625,
0.465576171875,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.50244140625,
0.798828125,
0.89208984375,
0.798828125,
0.89208984375,
0.8125,
0.50244140625,
0.8125
],
[
0.080078125,
0.8095703125,
0.468017578125,
0.8095703125,
0.468017578125,
0.82275390625,
0.080078125,
0.82275390625
],
[
0.5009765625,
0.8125,
0.89208984375,
0.81298828125,
0.89208984375,
0.82763671875,
0.5009765625,
0.8271484375
],
[
0.080078125,
0.82568359375,
0.1396484375,
0.82470703125,
0.1396484375,
0.8359375,
0.080078125,
0.8369140625
],
[
0.1405029296875,
0.82470703125,
0.4697265625,
0.82470703125,
0.4697265625,
0.83837890625,
0.1405029296875,
0.83837890625
],
[
0.50244140625,
0.8291015625,
0.89208984375,
0.8291015625,
0.89208984375,
0.84228515625,
0.50244140625,
0.84228515625
],
[
0.080078125,
0.8388671875,
0.46728515625,
0.8388671875,
0.46728515625,
0.85302734375,
0.080078125,
0.85302734375
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85498046875,
0.468017578125,
0.86865234375,
0.080078125,
0.86865234375
],
[
0.5009765625,
0.85498046875,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.5009765625,
0.869140625
],
[
0.098876953125,
0.876953125,
0.4697265625,
0.876953125,
0.4697265625,
0.88720703125,
0.098876953125,
0.88720703125
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.07843017578125,
0.88623046875,
0.370849609375,
0.88818359375,
0.370849609375,
0.9013671875,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Adversarial Examples for Semantic Segmentation and Obiect Detection",
"Cihang Xiel*, Jianyu Wang?*, Zhishuai Zhangl*, Yuyin Zhou', Lingxi Xie', Alan Yuille!",
"'Department of Computer Science, The Johns Hopkins University, Baltimore, MD 21218 USA",
"2BaiduResearchUSA,Sunnyvale,CA94089USA",
"[ cihangxie306, wjyouch, zhshuai.zhang, zhouyuyiner, 198808xc, alan.l.yuille↓ @gmail.com",
"110",
"Abstract",
"It has been well demonstrated that adversarial examples,",
"E",
"i.e., natural images with visually imperceptible perturba-",
"tions added, cause deep networks to fail on image clas-",
"C",
"sification.In this paper, we extend adversarial examples",
"to semantic segmentation and object detection which are",
"muchmore difficult.Our observationisthat both seg",
"LA",
"mentation and detection are based on classifying multiple",
"targets on an image (e.g., the target is a pixel or a receptive",
"",
"field in segmentation, and an object proposal in detection).",
"O",
"This inspires us to optimize a loss function over a set of",
"pixels/proposals for generating adversarial perturbations.",
"Based on this idea, we propose a novel algorithm named",
"3",
"Dense Adversary Generation (DAG), which generates a",
"ACO",
"large family of adversarial examples, and applies to a wide",
"range of state-of-the-art deep networks for segmentation",
"6",
"and detection. We also find that the adversarial pertur-",
"980",
"",
"bations can be transferred across networks with diferent",
"training data, based on different architectures, and even",
"8",
"for different recognition tasks.In particular, the trans-",
"0",
"ferability across networks with the same architecture is",
"",
"more significant than in other cases. Besides, summing up",
":AI",
"heterogeneous perturbations often leads to better transfer",
"performance, which provides an effective method of black-",
"",
"box adversarial attack.",
"Figure I: An Adversarial example for semantic segmen-",
"tation and object detection.",
"FCN[19] isusedfor seg-",
"mentation, and Faster-RCNN [28] is used for detection.",
"",
"1.Introduction",
"Left column:the original image (top row) with the nor-",
"mal segmentation (the purple region is predicted as dog)",
"Convolutional Neural Networks (CNN) [14][31][32][13]",
"and detection results. Right column: after the adversarial",
"have become the state-of-the-art solutionfora wide range of",
"perturbation (top row, magnified by 10) is added to the",
"visual recognition problems.Based on a large-scale labeled",
"original image, both segmentation (the light green region as",
"dataset such as ImageNet [6] and powerful computational",
"trainandthepinkregionasperson)anddetectionresultsare",
"resourceslikemodernGPUs,itispossibletotrain ahier-",
"completely wrong. Note that, though the added perturbation",
"archical deep network to capture different levels of visual",
"can confuse both networks, it is visually imperceptible (the",
"patterns.",
"A deep network is also capable of generating",
"maximal absolute intensity in each channel is less than 10).",
"transferrable features for different tasks such as image clas-",
"sification [7] and instance retrieval [29], or being fine-tuned",
"to deal with a wide range of vision tasks, including object",
"*Thefirst three authors contributed equallyto this work.This work was",
"detection [11][10], visual concept discovery [35], semantic",
"done when Jianyu Wang was a Ph.D. student at UCLA",
"segmentation [19][3][38], boundary detection [36][30], etc."
] |
{
"arxiv_id": "1703.08603",
"language": "en",
"page_idx": 0
}
|
||
1703.08603-en-1
|
1703.08603-en-1.jpg
|
Despite their success in visual recognition and feature representation, deep networks are often sensitive to small perturbations to the input image. In [33], it was shown that adding visually imperceptible perturbations can result in failures for image classification. These perturbed images, often called adversarial examples, are considered to fall on some areas in the large, high-dimensional feature space which are not explored in the training process. Thus, investigating this not only helps understand the working mechanism of deep networks, but also provides opportunities to improve the robustness of network training.
In this paper, we go one step further by generating adversarial examples for semantic segmentation and object detection, and showing the transferability of them. To the best of our knowledge, this topic has not been systematically studied $(e.g.$, on a large dataset) before. Note that these tasks are much more difficult, as we need to consider orders of magnitude more targets (e.g., pixels or proposals). Motivated by the fact that each target undergoes a separate classification process, we propose the Dense Adversary Generation (DAG) algorithm, which considers all the targets simultaneously and optimizes the overall loss function. The implementation of DAG is simple, as it only involves specifying an adversarial label for each target and performing iterative gradient back-propagation. In practice, the algorithm often comes to an end after a reasonable number of, say, 150 to 200, iterations. Figure 1 shows an adversarial example which can confuse both deep segmentation and detection networks.
We point out that generating an adversarial example is more difficult in detection than in segmentation, as the number of targets is orders of magnitude larger in the former case, $e.g.$, for an image with $K$ pixels, the number of possible proposals is $O(\tilde{K}^{2})$ while the number of pixels is only $O(K)$,where $O(\cdot)$ is the big-O notation. In addition, if only a subset of proposals are considered, the perturbed image may still be correctly recognized after a new set of proposals are extracted (note that DAG aims at generating recognition failures on the original proposals). To increase the robustness of adversarial attack, we change the intersection-overunion (IOU) rate to preserve an increased but still reasonable number of proposals in optimization. In experiments, we verify that when the proposals are dense enough on the original image, it is highly likely that incorrect recognition results are also produced on the new proposals generated on the perturbed image. We also study the effectiveness and efficiency of the algorithm with respect to the denseness of the considered proposals.
Following [33], we investigate the transferability of the generated perturbations. To this end, we use the adversarial perturbation computed on one network to attack another network. Three situations are considered:(1) networks with the same architecture but trained with different data;
(2) networks with different architectures but trained for the same task; and (3) networks for different tasks. Although the difficulty increases as the difference goes more significant, the perturbations generated by DAG is able to transfer to some extent. Interestingly, adding two or more heterogeneous perturbations significantly increases the transferability, which provides an effective way of performing blackbox adversarial attack[26] to some networks with unknown structures and/or properties.
The remainder of this paper is organized as follows. Section 2 briefly introduces prior work related to our research. Section 3 describes our algorithm for generating adversarial perturbations, and Section 4 investigates the transferability of the perturbations. Conclusions are drawn in Section 5.
## $2. \textbf{Related Work}$
$2. 1. \textbf{ Deep Learning for Detection and Segmentation}$
Deep learning approaches, especially deep convolutional neural networks, have been very successful in object detection [28][5][17] and semantic segmentation [19][3] tasks. Currently, one of the most popular object detection pipeline [28][5][17] involves first generating a number of proposals of different scales and positions, classifying each of them, and performing post-processing such as nonmaximal suppression (NMS). On the other hand, the dominating segmentation pipeline[19] works by first predicting a class-dependent score map at a reduced resolution, and performing up-sampling to obtain high-resolution segmentation.[3] incorporates the “atrous”algorithm and the conditional random feld (CRF) to this pipeline to improve the segmentation performance further.
## 2.2. Adversarial Attack and Defense
Generating adversarial examples for classification has been extensively studied in many different ways recently. [33] first showed that adversarial examples, computed by adding visually imperceptible perturbations to the original images, make CNNs predict a wrong label with high confidence.[12] proposed a simple and fast gradient sign method to generate adversarial examples based on the linear nature of CNNs.[24] proposed a simple algorithm to compute the minimal adversarial perturbation by assuming that the loss function can be linearized around the current data point at each iteration. [23] showed the existence of universal (image-agnostic) adversarial perturbations. [1] trained a network to generate adversarial examples for a particular target model (without using gradients).[15] showed the adversarial examples for machine learning systems also exist in the physical world. [18] studied the transferability of both non-targeted and targeted adversarial examples, and proposed an ensemble-based approaches to generate adversarial examples with stronger transferability. [25] generated
2
|
[
[
0.098876953125,
0.0921630859375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.1060791015625,
0.098876953125,
0.1060791015625
],
[
0.50341796875,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.07843017578125,
0.10797119140625,
0.468017578125,
0.106689453125,
0.468017578125,
0.12060546875,
0.07843017578125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50341796875,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.07843017578125,
0.1673583984375,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.5009765625,
0.1673583984375,
0.890625,
0.1666259765625,
0.890625,
0.18115234375,
0.5009765625,
0.1817626953125
],
[
0.08087158203125,
0.1817626953125,
0.468017578125,
0.1817626953125,
0.468017578125,
0.1956787109375,
0.08087158203125,
0.1956787109375
],
[
0.50244140625,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08251953125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08251953125,
0.210205078125
],
[
0.50341796875,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50341796875,
0.210205078125
],
[
0.08087158203125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.68896484375,
0.21337890625,
0.68896484375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.471435546875,
0.2291259765625,
0.471435546875,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.52294921875,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.24365234375,
0.52294921875,
0.24365234375
],
[
0.07843017578125,
0.242431640625,
0.365234375,
0.24365234375,
0.365234375,
0.25830078125,
0.07843017578125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.09722900390625,
0.261474609375,
0.46728515625,
0.261474609375,
0.46728515625,
0.27587890625,
0.09722900390625,
0.27587890625
],
[
0.50244140625,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.276611328125,
0.4697265625,
0.276611328125,
0.4697265625,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.50244140625,
0.274658203125,
0.890625,
0.274658203125,
0.890625,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.08087158203125,
0.29345703125,
0.468017578125,
0.29345703125,
0.468017578125,
0.3037109375,
0.08087158203125,
0.3037109375
],
[
0.50341796875,
0.291748046875,
0.87744140625,
0.291748046875,
0.87744140625,
0.3017578125,
0.50341796875,
0.3017578125
],
[
0.080078125,
0.306884765625,
0.468017578125,
0.306884765625,
0.468017578125,
0.32080078125,
0.080078125,
0.32080078125
],
[
0.07843017578125,
0.3212890625,
0.4697265625,
0.32080078125,
0.4697265625,
0.335205078125,
0.07843017578125,
0.3359375
],
[
0.50341796875,
0.3193359375,
0.640625,
0.3193359375,
0.640625,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.08087158203125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50341796875,
0.3447265625,
0.890625,
0.3447265625,
0.890625,
0.359130859375,
0.50341796875,
0.359130859375
],
[
0.07843017578125,
0.35107421875,
0.4697265625,
0.3515625,
0.4697265625,
0.3662109375,
0.07843017578125,
0.365478515625
],
[
0.080078125,
0.367431640625,
0.4697265625,
0.367431640625,
0.4697265625,
0.380615234375,
0.080078125,
0.380615234375
],
[
0.52294921875,
0.36865234375,
0.88818359375,
0.36865234375,
0.88818359375,
0.3818359375,
0.52294921875,
0.3818359375
],
[
0.080078125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.396484375,
0.080078125,
0.396484375
],
[
0.5009765625,
0.3818359375,
0.89208984375,
0.38330078125,
0.89208984375,
0.397705078125,
0.5009765625,
0.396484375
],
[
0.08087158203125,
0.397705078125,
0.4697265625,
0.397705078125,
0.4697265625,
0.410888671875,
0.08087158203125,
0.410888671875
],
[
0.50341796875,
0.40087890625,
0.8896484375,
0.40087890625,
0.8896484375,
0.410888671875,
0.50341796875,
0.410888671875
],
[
0.080078125,
0.412353515625,
0.468017578125,
0.412353515625,
0.468017578125,
0.425537109375,
0.080078125,
0.425537109375
],
[
0.5009765625,
0.4140625,
0.8896484375,
0.4140625,
0.8896484375,
0.427978515625,
0.5009765625,
0.427978515625
],
[
0.07843017578125,
0.425537109375,
0.4697265625,
0.4267578125,
0.4697265625,
0.44140625,
0.07843017578125,
0.43994140625
],
[
0.5009765625,
0.4287109375,
0.89208984375,
0.427978515625,
0.89208984375,
0.442626953125,
0.5009765625,
0.443115234375
],
[
0.07843017578125,
0.442626953125,
0.4697265625,
0.44140625,
0.4697265625,
0.455810546875,
0.07843017578125,
0.45703125
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.080078125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.5009765625,
0.458251953125,
0.890625,
0.458984375,
0.890625,
0.473388671875,
0.5009765625,
0.472900390625
],
[
0.080078125,
0.473388671875,
0.4697265625,
0.473388671875,
0.4697265625,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.50244140625,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.50244140625,
0.488037109375
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.080078125,
0.501953125
],
[
0.50244140625,
0.490478515625,
0.890625,
0.490478515625,
0.890625,
0.50390625,
0.50244140625,
0.50390625
],
[
0.07843017578125,
0.501953125,
0.468017578125,
0.50244140625,
0.468017578125,
0.517578125,
0.07843017578125,
0.5166015625
],
[
0.5009765625,
0.5048828125,
0.890625,
0.5048828125,
0.890625,
0.51806640625,
0.5009765625,
0.51806640625
],
[
0.080078125,
0.51806640625,
0.330078125,
0.51806640625,
0.330078125,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50341796875,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.1004638671875,
0.5361328125,
0.468017578125,
0.5361328125,
0.468017578125,
0.5498046875,
0.1004638671875,
0.5498046875
],
[
0.50341796875,
0.5361328125,
0.5595703125,
0.5361328125,
0.5595703125,
0.54736328125,
0.50341796875,
0.54736328125
],
[
0.5537109375,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.5537109375,
0.54736328125
],
[
0.080078125,
0.55029296875,
0.46728515625,
0.55029296875,
0.46728515625,
0.564453125,
0.080078125,
0.564453125
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.08087158203125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.5009765625,
0.564453125,
0.75244140625,
0.564453125,
0.75244140625,
0.57763671875,
0.5009765625,
0.57763671875
],
[
0.080078125,
0.58056640625,
0.46728515625,
0.58056640625,
0.46728515625,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.50244140625,
0.5888671875,
0.78173828125,
0.59130859375,
0.78173828125,
0.60595703125,
0.50244140625,
0.60400390625
],
[
0.080078125,
0.5966796875,
0.468017578125,
0.59521484375,
0.468017578125,
0.60986328125,
0.080078125,
0.611328125
],
[
0.08087158203125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.08087158203125,
0.62548828125
],
[
0.51953125,
0.61279296875,
0.89208984375,
0.6142578125,
0.89208984375,
0.62890625,
0.51953125,
0.62744140625
],
[
0.080078125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.5009765625,
0.62890625,
0.890625,
0.6298828125,
0.890625,
0.64453125,
0.5009765625,
0.64306640625
],
[
0.07843017578125,
0.642578125,
0.4697265625,
0.642578125,
0.4697265625,
0.65576171875,
0.07843017578125,
0.65576171875
],
[
0.5048828125,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.080078125,
0.6572265625,
0.4697265625,
0.6572265625,
0.4697265625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.07843017578125,
0.67041015625,
0.468017578125,
0.671875,
0.468017578125,
0.68603515625,
0.07843017578125,
0.68505859375
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.6875,
0.4697265625,
0.6875,
0.4697265625,
0.70068359375,
0.080078125,
0.70068359375
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.7041015625,
0.465576171875,
0.7041015625,
0.465576171875,
0.7138671875,
0.08087158203125,
0.7138671875
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.08087158203125,
0.73193359375,
0.46728515625,
0.73193359375,
0.46728515625,
0.74560546875,
0.08087158203125,
0.74560546875
],
[
0.5048828125,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.08087158203125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76025390625,
0.08087158203125,
0.76025390625
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.76318359375,
0.4697265625,
0.76318359375,
0.4697265625,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.77978515625,
0.46728515625,
0.77978515625,
0.46728515625,
0.79052734375,
0.08087158203125,
0.79052734375
],
[
0.50341796875,
0.78271484375,
0.89208984375,
0.78271484375,
0.89208984375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79345703125,
0.471435546875,
0.79345703125,
0.471435546875,
0.806640625,
0.080078125,
0.806640625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.8095703125,
0.2484130859375,
0.8095703125,
0.2484130859375,
0.82275390625,
0.080078125,
0.82275390625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.098876953125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.098876953125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.07843017578125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.08087158203125,
0.88818359375,
0.468017578125,
0.88818359375,
0.468017578125,
0.8984375,
0.08087158203125,
0.8984375
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"Despite their success in visual recognition and feature",
"(2) networks with different architectures but trained for the",
"representation, deep networks are often sensitive to small",
"same task; and (3) networks for different tasks. Although",
"perturbations to theinput image.In[33l,it was shown",
"the difficulty increases as the difference goes more signifi-",
"that adding visually imperceptible perturbations can result",
"cant, the perturbations generated by DAG is able to transfer",
"infailuresforimageclassification.Theseperturbedimages.",
"to some extent.Interestingly,addingtwo ormoreheteroge-",
"often called adversarial examples, are considered to fall",
"neous perturbations significantly increases the transferabil-",
"on some areas in the large, high-dimensional feature space",
"ity, which provides an effective way of performing black-",
"whicharenotexploredinthetrainingprocess.Thus,inves-",
"boxadversarialattack[26]tosomenetworkswithunknown",
"tigating this not only helps understand the working mech-",
"structures and/or properties.",
"anism of deep networks, but also provides opportunities to",
"The remainder of this paper is organized as follows. Sec-",
"improve the robustness of network training.",
"tion 2 briefly introduces prior work related to our research",
"In this paper, we go one step further by generating ad-",
"Section 3 describes our algorithm for generating adversarial",
"versarial examples for semantic segmentation and object",
"perturbations, and Section 4 investigates the transferability",
"detection, and showing the transferability of them.To",
"oftheperturbations.Conclusions are drawn in Section 5.",
"the best of our knowledge, this topic has not been sys-",
"tematically studied (e.g., on a large dataset) before. Note",
"2.Related Work",
"that these tasks are much more difficult. as we need to",
"2.1.Deep Learning for Detection and Segmentation",
"consider orders of magnitude more targets (e.g., pixels or",
"proposals). Motivated by the fact that each target undergoes",
"Deep learning approaches, especially deep convolu-",
"a separate classification process, we propose the Dense",
"tional neural networks, have been very successful in object",
"Adversary Generation (DAG) algorithm, which considers",
"detection[281[51[17]andsemanticsegmentation[191[3]",
"all the targets simultaneously and optimizes the overall",
"tasks.Currently, one of the most popular object detec-",
"loss function. The implementation of DAG is simple, as",
"tion pipeline [28][5][17] involves first generating a number",
"it only involves specifying an adversarial label for each",
"of proposals of different scales and positions, classifying",
"target and performing iterative gradient back-propagation.",
"each of them, and performing post-processing such as non-",
"In practice, the algorithm often comes to an end after a",
"maximal suppression (NMS). On the other hand, the domi-",
"reasonable number of, say, 150 to 200, iterations. Figure 1",
"nating segmentation pipeline [19] works by first predicting",
"shows an adversarial example which can confuse both deep",
"a class-dependent score map at a reduced resolution, and",
"segmentation and detection networks.",
"performing up-sampling to obtain high-resolution segmen-",
"We point out that generating an adversarial example is",
"tation.",
"[3] incorporates the “atrous\" algorithm and the",
"more difficult in detection than in segmentation, as the",
"conditional random field (CRF) to this pipeline to improve",
"numberoftargetsisorders ofmagnitudelargerintheformer",
"the segmentation performance further.",
"case, e.g.,for an image with K pixels, the number of possi-",
"2.2. Adversarial Attack and Defense",
"ble proposals is O(K2) while the number of pixels is only",
"O(K), where O()is the big-O notation. In addition, if only",
"Generating adversarial examples for classification has",
"a subset of proposals are considered, the perturbed image",
"been extensively studied in many different ways recently.",
"may still be correctlyrecognized after a new set of proposals",
"[33]first showed that adversarial examples, computed by",
"are extracted (note that DAG aims at generating recognition",
"adding visually imperceptible perturbations to the original",
"failures on the original proposals). To increase the robust-",
"images, make CNNs predict a wrong label with high confi-",
"ness of adversarial attack, we change the intersection-over-",
"dence. [12] proposed a simple and fast gradient sign method",
"union(IOU)ratetopreserveanincreasedbutstillreason-",
"to generate adversarial examples based on the linear nature",
"able number of proposals in optimization. In experiments,",
"of CNNs. [24] proposed a simple algorithm to compute",
"we verify that when the proposals are dense enough on the",
"the minimal adversarialperturbation byassumingthat the",
"original image,itishighlylikelythatincorrectrecognition",
"loss function can be linearized around the current data point",
"results are also produced on the new proposals generated on",
"at each iteration. [23] showed the existence of universal",
"the perturbed image.Wealso studytheeffectivenessand",
"(image-agnostic)adversarial perturbations.[1]trained a",
"efficiency of the algorithm with respect to the denseness of",
"network to generate adversarial examples for a particular",
"the considered proposals.",
"target model (without using gradients). [15] showed the",
"Following [33], we investigate the transferability of the",
"adversarial examples for machine learning systems also",
"generated perturbations. To this end, we use the adversarial",
"exist in the physical world. [18] studied the transferability",
"perturbation computed on one network to attack another",
"of both non-targeted and targeted adversarial examples, and",
"network. Three situations are considered: (1) networks",
"proposed an ensemble-based approaches to generate adver-",
"with the same architecture but trained with different data:",
"sarial examples with stronger transferability. [25] generated",
"2"
] |
{
"arxiv_id": "1703.08603",
"language": "en",
"page_idx": 1
}
|
||
1703.08603-en-2
|
1703.08603-en-2.jpg
|
images using evolutionary algorithms that are unrecognizable to humans, but cause CNNs to output very confident (incorrect) predictions. This can be thought of as in the opposite direction of above works.
In contrast to generating adversarial examples, there are some works trying to reduce the effect of adversarial examples. [20] proposed a forveation-based mechanism to alleviate adversarial examples. [27] showed networks trained using defensive distillation can effectively against adversarial examples, while [2] developed stronger attacks which are unable to defend by defensive distillation. [16] trained the network on adversarial examples using the largescale ImageNet, and showed that this brings robustness to adversarial attack. This is imporved by[34],which proposed an ensemble adversarial training method to increase the network robustness to black-box attacks.[21] trained a detector on the inner layer of the classifier to detect adversarial examples.
There are two concurrent works [9] and [22] that studied adversarial examples in semantic segmentation on the Cityscapes dataset[4], where[9] showed the existence of adversarial examples, and [22] showed the existence of universal perturbations. We refer interested readers to their papers for details.
## 3. Generating Adversarial Examples
In this section, we introduce DAG algorithm. Given an image and the recognition targets (proposals and/or pixels), DAG generates an adversarial perturbation which is aimed at confusing as many targets as possible.
## $3. 1. \textbf{ Dense Adversary Generation}$
Let X be an image which contains $N$ recognition targets $\mathcal{T}$ = $\{ t_{1}, t_{2}, \ldots , t_{N}\} .$ Each target $t_n, n$ = $1, 2, \ldots , N$, is assigned a ground-truth class label $l_n\in\{1,2,\ldots,C\}$, where $C$ is the number of classes, e.g., $C=21$ (including the background class) in the PascalVOC dataset[8]. Denote $\mathcal{L}=\{l_{1},l_{2},\ldots,l_{n}\}.$ The detailed form of $\mathcal{T}$ varies among different tasks. In image classification, $\mathcal{T}$ only contains one element, $i.e.$, the entire image. Conversely, $\mathcal{T}$ is composed of all pixels (or the corresponding receptive fields) in semantic segmentation, and all proposals in object detection. We will discuss how to construct $\mathcal{T}$ in Section 3.2.
Given a deep network for a specific task, we use $\mathbf{f} ( \mathbf{X} , t_{n})$ $\in$ $\mathbb{R} ^{\hat{C} }$ to denote the classification score vector (before softmax normalization) on the $n$-th recognition target of X. To generate an adversarial example, the goal is to make the predictions of all targets go wrong, $i.e.$, $\forall n$, arg$\operatorname*{max}_c\left\{f_{c}(\mathbf{X}+\mathbf{r},t_{n})\right\}\neq l_{n}.$ Here r denotes an adversarial perturbation added to $\mathbf{X} .$ To this end, we specify an adversarial label $l_n^{\prime}$ for each target, in which $l_n^{\prime}$ is randomly sampled from other incorrect classes, i.e.,
Algorithm 1: Dense Adversary Generation ( DAG)
${\overline{\mathrm{Input~:~input~image~X;}}}$
$\mathop{\text{the classifier f}(}\cdot,\cdot)\in\mathbb{R}^{C};$
the target set $\mathcal{T}=\{t_1,t_2,\ldots,t_N\};$
the original label set $\mathcal{L}=\{l_1,l_2,\ldots,l_N\};$
the adversarial label set $\mathcal{L}^\prime=\{l_1^\prime,l_2^\prime,\ldots,l_N^\prime\};$
the maximal iterations $M_0;$
Output: the adversarial perturbation r;
1 $\mathbf{X}_0\leftarrow\mathbf{X},\mathbf{r}\leftarrow\mathbf{0},m\leftarrow0,\mathcal{T}_0\leftarrow\mathcal{T};$
$2\textbf{ while }m< M_{0}\textbf{and }\mathcal{T} _{m}\neq \varnothing \textbf{ do}$
3
$\mathcal{T}_m=\{t_n\mid\arg\max_c\{f_c(\mathbf{X}_m,t_n)\}=l_n\};$
$\mathbf{r}_{m}\leftarrow$
$\sum_{t_n\in\mathcal{T}_m}\left[\nabla\mathbf{x}_{m}f_{l_n^{\prime}}(\mathbf{X}_{m},t_{n})-\nabla\mathbf{x}_{m}f_{l_n}(\mathbf{X}_{m},t_{n})\right];$
5 $\mathbf{r} _{m}^{\prime }\leftarrow \frac \gamma {\left \| \mathbf{r} _{m}\right \| _{\infty }}\mathbf{r} _{m};$
$\mathbf{r}\leftarrow\mathbf{r}+\mathbf{r}_{m}^{\prime};$
7
$\mathbf{X}_{m+1}\leftarrow\mathbf{X}_{m}+\mathbf{r}_{m}^{\prime};$
8
$m\leftarrow m+1;$
9 end
Return: r
$l_{n}^{\prime}\in\{1,2,\ldots,C\}\setminus\{l_{n}\}.$ Denote $\mathcal{L}^\prime=\{l_{1}^{\prime},l_{2}^{\prime},\ldots,l_{n}^{\prime}\}.$ In practice, we define a random permutation function $\pi:$
$\{1,2,\ldots,C\}\to\{1,2,\ldots,C\}$ for every image indepen- $\mathop{\mathrm{dently,~in~which~\pi}(c)\neq c\mathrm{~for~}c=1,2,\ldots,C,\mathrm{and~generate}}$ $\mathcal{L}^{\prime}$ by setting $l_n^\prime=\pi(l_n)$ for all $n.$ Under this setting, the loss function covering all targets can be written as:
$$L(\mathbf{X},\mathcal{T},\mathcal{L},\mathcal{L}')=\sum_{n=1}^{N}\begin{bmatrix}f_{l_n}(\mathbf{X},t_n)-f_{l_n'}(\mathbf{X},t_n)\end{bmatrix}$$
(1)
Minimizing $L$ can be achieved via making every target to be incorrectly predicted, $i.e.$, suppressing the confidence of the original correct class $f_{l_n}(\mathbf{X}+$r$,t_n)$,while increasing that of the desired (adversarial) incorrect class $f_{l_n^{\prime}}(\mathbf{X}+$r$,t_n).$
We apply a gradient descent algorithm for optimization. At the $m$-th iteration, denote the current image (possibly after adding several perturbations) as $\mathbf{X}_m.$ We find the set of correctly predicted targets, named the active target $set\nobreak {: } \mathcal{T} _{m}$ = $\{ t_{n}$ | $\arg \max _{c}$ $\{ f_{c}( \mathbf{X} _{m}, t_{n}) \} = l_{n}\} .$Then we compute the gradient with respect to the input data and then accumulate all these perturbations:
$$\mathbf{r}_{m}=\sum_{t_{n}\in\mathcal{T}_{m}}\left[\nabla_{\mathbf{X}_{m}}f_{l_{n}^{\prime}}(\mathbf{X}_{m},t_{n})-\nabla_{\mathbf{X}_{m}}f_{l_{n}}(\mathbf{X}_{m},t_{n})\right]\:(2)$$
Note that $|\mathcal{T}_m|\ll|\mathcal{T}|$ when $m$ gets large, thus this strategy considerably reduces the computational overhead. To avoid numerical instability, we normalize r$_m$ as
(3)
$$\mathbf{r}_m'=\frac{\gamma}{\left\|\mathbf{r}_m\right\|_\infty}\cdot\mathbf{r}_m$$
where $\gamma=0.5$ is a fixed hyper-parameter. We then add $\mathbf{r}_m^{\prime}$ to the current image $\mathbf{X}_m$ and proceed to the next iteration.
3
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5107421875,
0.0953369140625,
0.85302734375,
0.0953369140625,
0.85302734375,
0.10919189453125,
0.5107421875,
0.10919189453125
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.51318359375,
0.11297607421875,
0.68310546875,
0.11297607421875,
0.68310546875,
0.126953125,
0.51318359375,
0.126953125
],
[
0.08087158203125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.57666015625,
0.1275634765625,
0.7509765625,
0.1256103515625,
0.7509765625,
0.14013671875,
0.57666015625,
0.1427001953125
],
[
0.080078125,
0.1383056640625,
0.3095703125,
0.1383056640625,
0.3095703125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.57666015625,
0.1427001953125,
0.81201171875,
0.143310546875,
0.81201171875,
0.1578369140625,
0.57666015625,
0.1572265625
],
[
0.09722900390625,
0.15283203125,
0.468017578125,
0.1514892578125,
0.468017578125,
0.1666259765625,
0.09722900390625,
0.1673583984375
],
[
0.57666015625,
0.1572265625,
0.85693359375,
0.1578369140625,
0.85693359375,
0.17236328125,
0.57666015625,
0.1717529296875
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.57666015625,
0.17041015625,
0.8798828125,
0.17236328125,
0.8798828125,
0.1881103515625,
0.57666015625,
0.186279296875
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.57666015625,
0.1881103515625,
0.76123046875,
0.1881103515625,
0.76123046875,
0.2020263671875,
0.57666015625,
0.2020263671875
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.51318359375,
0.2020263671875,
0.77685546875,
0.203857421875,
0.77685546875,
0.2171630859375,
0.51318359375,
0.21533203125
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2183837890625,
0.75244140625,
0.2183837890625,
0.75244140625,
0.231689453125,
0.50244140625,
0.231689453125
],
[
0.07843017578125,
0.2266845703125,
0.468017578125,
0.2279052734375,
0.468017578125,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.5,
0.231689453125,
0.7294921875,
0.23291015625,
0.7294921875,
0.2474365234375,
0.5,
0.246826171875
],
[
0.080078125,
0.24365234375,
0.46728515625,
0.24365234375,
0.46728515625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5400390625,
0.248779296875,
0.826171875,
0.248779296875,
0.826171875,
0.26318359375,
0.5400390625,
0.26318359375
],
[
0.08087158203125,
0.260009765625,
0.46728515625,
0.260009765625,
0.46728515625,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.26953125,
0.5087890625,
0.267578125,
0.51123046875,
0.2734375,
0.5048828125,
0.27587890625
],
[
0.5390625,
0.266357421875,
0.57861328125,
0.264404296875,
0.580078125,
0.276611328125,
0.5400390625,
0.279052734375
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.5458984375,
0.279052734375,
0.619140625,
0.2822265625,
0.61767578125,
0.300537109375,
0.5439453125,
0.29736328125
],
[
0.609375,
0.281005859375,
0.8837890625,
0.281005859375,
0.8837890625,
0.29541015625,
0.609375,
0.29541015625
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.53759765625,
0.298583984375,
0.6611328125,
0.298583984375,
0.6611328125,
0.317626953125,
0.53759765625,
0.317626953125
],
[
0.080078125,
0.318115234375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.080078125,
0.33203125
],
[
0.53759765625,
0.31494140625,
0.62744140625,
0.317626953125,
0.62744140625,
0.330810546875,
0.53759765625,
0.328857421875
],
[
0.50244140625,
0.3212890625,
0.5087890625,
0.32080078125,
0.5107421875,
0.326904296875,
0.50341796875,
0.32763671875
],
[
0.07843017578125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.07843017578125,
0.347900390625
],
[
0.50634765625,
0.333984375,
0.51318359375,
0.339111328125,
0.50634765625,
0.3447265625,
0.5,
0.339111328125
],
[
0.5361328125,
0.328857421875,
0.68310546875,
0.330078125,
0.68310546875,
0.347900390625,
0.5361328125,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.223876953125,
0.349853515625,
0.223876953125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.35107421875,
0.5107421875,
0.35107421875,
0.5107421875,
0.35791015625,
0.50341796875,
0.35791015625
],
[
0.5390625,
0.348388671875,
0.62841796875,
0.348388671875,
0.62841796875,
0.35986328125,
0.5390625,
0.35986328125
],
[
0.096435546875,
0.3623046875,
0.468017578125,
0.363037109375,
0.468017578125,
0.37744140625,
0.096435546875,
0.376953125
],
[
0.5,
0.3623046875,
0.5439453125,
0.3623046875,
0.5439453125,
0.37744140625,
0.5,
0.37744140625
],
[
0.07843017578125,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.07843017578125,
0.392578125
],
[
0.51318359375,
0.376953125,
0.5908203125,
0.376953125,
0.5908203125,
0.391357421875,
0.51318359375,
0.391357421875
],
[
0.080078125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5009765625,
0.417236328125,
0.89208984375,
0.41845703125,
0.89208984375,
0.436279296875,
0.5009765625,
0.43505859375
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.080078125,
0.43994140625,
0.1993408203125,
0.43994140625,
0.1993408203125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.436279296875,
0.88818359375,
0.436279296875,
0.88818359375,
0.449462890625,
0.50341796875,
0.449462890625
],
[
0.5048828125,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.46533203125,
0.5048828125,
0.46533203125
],
[
0.08087158203125,
0.46728515625,
0.3857421875,
0.46728515625,
0.3857421875,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50341796875,
0.46533203125,
0.89208984375,
0.46533203125,
0.89208984375,
0.479736328125,
0.50341796875,
0.479736328125
],
[
0.5048828125,
0.479736328125,
0.89208984375,
0.479736328125,
0.89208984375,
0.494384765625,
0.5048828125,
0.494384765625
],
[
0.098876953125,
0.492431640625,
0.4697265625,
0.492431640625,
0.4697265625,
0.505859375,
0.098876953125,
0.505859375
],
[
0.50244140625,
0.496826171875,
0.83642578125,
0.496826171875,
0.83642578125,
0.51025390625,
0.50244140625,
0.51025390625
],
[
0.080078125,
0.50830078125,
0.46728515625,
0.50830078125,
0.46728515625,
0.521484375,
0.080078125,
0.521484375
],
[
0.080078125,
0.52294921875,
0.4697265625,
0.52294921875,
0.4697265625,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.650390625,
0.517578125,
0.6708984375,
0.517578125,
0.6708984375,
0.52978515625,
0.650390625,
0.52978515625
],
[
0.52783203125,
0.5322265625,
0.64453125,
0.5322265625,
0.64453125,
0.54541015625,
0.52783203125,
0.54541015625
],
[
0.64453125,
0.52880859375,
0.8505859375,
0.52978515625,
0.8505859375,
0.54736328125,
0.64453125,
0.546875
],
[
0.07843017578125,
0.53857421875,
0.34716796875,
0.537109375,
0.34716796875,
0.5517578125,
0.07843017578125,
0.55322265625
],
[
0.86669921875,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.54541015625,
0.86669921875,
0.54541015625
],
[
0.64794921875,
0.548828125,
0.67626953125,
0.548828125,
0.67626953125,
0.56005859375,
0.64794921875,
0.56005859375
],
[
0.07843017578125,
0.56005859375,
0.336669921875,
0.56103515625,
0.336669921875,
0.57568359375,
0.07843017578125,
0.57470703125
],
[
0.50244140625,
0.56494140625,
0.89208984375,
0.56640625,
0.89208984375,
0.58056640625,
0.50244140625,
0.580078125
],
[
0.09722900390625,
0.58349609375,
0.468017578125,
0.583984375,
0.468017578125,
0.5986328125,
0.09722900390625,
0.59765625
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.080078125,
0.5986328125,
0.468017578125,
0.599609375,
0.468017578125,
0.6142578125,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.5009765625,
0.611328125,
0.88232421875,
0.6123046875,
0.88232421875,
0.626953125,
0.5009765625,
0.62548828125
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.080078125,
0.64453125
],
[
0.52294921875,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.642578125,
0.52294921875,
0.642578125
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.5009765625,
0.6416015625,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.65576171875,
0.89208984375,
0.67041015625,
0.50244140625,
0.671875
],
[
0.080078125,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.6875,
0.080078125,
0.6875
],
[
0.50244140625,
0.67236328125,
0.8935546875,
0.67236328125,
0.8935546875,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.70556640625,
0.46728515625,
0.70556640625,
0.46728515625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.50244140625,
0.71630859375
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.07843017578125,
0.73486328125
],
[
0.50244140625,
0.71826171875,
0.73388671875,
0.71826171875,
0.73388671875,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.080078125,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50732421875,
0.7470703125,
0.55126953125,
0.7470703125,
0.55126953125,
0.75830078125,
0.50732421875,
0.75830078125
],
[
0.55810546875,
0.7470703125,
0.583984375,
0.7470703125,
0.583984375,
0.75830078125,
0.55810546875,
0.75830078125
],
[
0.5810546875,
0.7412109375,
0.890625,
0.7421875,
0.890625,
0.76025390625,
0.5810546875,
0.7587890625
],
[
0.080078125,
0.751953125,
0.16015625,
0.751953125,
0.16015625,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.54833984375,
0.7587890625,
0.5908203125,
0.76123046875,
0.5888671875,
0.7744140625,
0.54736328125,
0.77294921875
],
[
0.098876953125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.098876953125,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5009765625,
0.77978515625,
0.890625,
0.7822265625,
0.890625,
0.796875,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5,
0.796875,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5,
0.81103515625
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81103515625,
0.779296875,
0.8125,
0.779296875,
0.8271484375,
0.5009765625,
0.82568359375
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.63134765625,
0.8369140625,
0.67333984375,
0.84326171875,
0.6689453125,
0.859375,
0.62744140625,
0.85302734375
],
[
0.69677734375,
0.83349609375,
0.70751953125,
0.83349609375,
0.70751953125,
0.84326171875,
0.69677734375,
0.84326171875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.73046875,
0.84228515625,
0.76123046875,
0.84228515625,
0.76123046875,
0.85400390625,
0.73046875,
0.85400390625
],
[
0.86767578125,
0.83837890625,
0.89453125,
0.83837890625,
0.89453125,
0.85498046875,
0.86767578125,
0.85498046875
],
[
0.67626953125,
0.8466796875,
0.72802734375,
0.849609375,
0.72607421875,
0.86572265625,
0.6748046875,
0.8623046875
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.88818359375,
0.8720703125,
0.88818359375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"images using evolutionary algorithms that are unrecogniz-",
"Algorithm 1:Dense Adversary Generation (DAG)",
"ableto humans,but cause CNNs tooutputvery confident",
"Input :input imageX;",
"(incorrect) predictions.This can be thought of as in the",
"the classifier f(, ) e IRC:",
"opposite direction of above works.",
"the target set T = {ti,t2,...,t];",
"In contrastto generating adversarial examples,there",
"the original label set C = {li, l2,...,ln];",
"are some works trying to reduce the effect of adversarial",
"the adversarial label set C' - {'i, l2, .., I'];",
"examples.[20] proposed a forveation-based mechanism",
"the maximal iterations Mo;",
"to alleviateadversarial examples.[27]showednetworks",
"Output: the adversarial perturbation r;",
"trained using defensive distillation can effectively against",
"1 Xo ←X, r ←0, m←0, T ←T;",
"adversarial examples, while [2] developed stronger attacks",
"2 while m< Mo and Tm ± @ do",
"which are unable to defend by defensive distillation. [16]",
"Tm = (tn / argmaxe (f(Xm,tn)) = In);",
"trained thenetwork on adversarial examples usingthelarge-",
"4",
"rm←",
"scale ImageNet, and showed that this brings robustness to",
"EteTm[",
"[Vxmfu,(Xm,tn)-Vxmfi.(Xm,tn)l;",
"adversarial attack. This is imporved by [34], which pro-",
"posed an ensemble adversarial training method to increase",
"rm←rmrm;",
"the network robustness to black-box attacks. [21] trained",
"r←r+rm",
"6",
"a detector on the inner layer of the classifier to detect",
"A",
"Xm+1←Xm+rm;",
"adversarial examples.",
"8",
"m<m+1;",
"There are two concurrent works [9] and [22] that stud-",
"9 end",
"ied adversarial examples in semantic segmentation on the",
"Return: r",
"Cityscapes dataset [4], where [9] showed the existence of",
"adversarial examples, and [22] showed the existence of",
"I' e {1, 2,...,C)I fn]. Denote c' = {li, l?,..., '.",
"universal perturbations. We refer interested readers to their",
"papers for details.",
"In practice, we define a random permutation function ",
"[1, 2, .., C] → [1, 2, ...,C] for every image indepen-",
"3. Generating Adversarial Examples",
"dently, in which (c) + c for c = 1, 2, ..., C, and generate",
"C' by setting I' = (ln) for all n. Under this setting, the",
"In this section, we introduce DAG algorithm. Given an",
"loss function covering all targets can be written as:",
"image and the recognition targets (proposals and/or pixels),",
"DAG generates an adversarial perturbation which is aimed",
"N",
"L(X, T, C, C') =",
"[fi,(X,tn) - fus (x,tn)]",
"at confusing as many targets as possible.",
"(1)",
"n=1",
"3.1.Dense Adversary Generation",
"Minimizing L can be achieved via making every target to be",
"Let X be an image which contains N recognition targets",
"incorrectly predicted, i.e.,suppressingthe confidence of the",
"T= {ti,t2,...,tl. Each target tn, n = 1,2,..,N,",
"original correct class fi, (X + r,tn), while increasing that",
"is assigned a ground-truth class label In E {1, 2,..., C],",
"of the desired (adversarial) incorrect class fu (X + r,tn).",
"where C is the number of classes, e.g., C = 21 (including",
"We apply a gradient descent algorithm for optimization.",
"the background class) in the PascalVOC dataset [8]. De-",
"At the m-th iteration, denote the current image (possibly",
"note L = {li, l2,..., ln]. The detailed form of T varies",
"after adding several perturbations) as Xm: We find the",
"among different tasks.In image classification,T only",
"set of correctly predicted targets, named the active target",
"contains one element, i.e., the entire image. Conversely,",
"set: Tm = (tn | argmaxc (fe(Xm,tn)) = ln). Then we",
"T is composed of all pixels (or the corresponding recep-",
"compute the gradient with respect to the input data and then",
"tive fields) in semantic segmentation, and all proposals in",
"accumulate all these perturbations:",
"object detection. We will discuss how to construct T in",
"rm=",
">",
"[Vxmfuz (Xm,tn)-Vxmfi.(Xm,tn)] (2)",
"Section 3.2.",
"tnETm",
"Given a deep network for a specific task, we use",
"f(X,tn) E R to denote the classification score vector",
"Note that |Tm < |T| when m gets large, thus this strategy",
"(before softmax normalization) on the n-th recognition tar-",
"considerably reduces the computational overhead. To avoid",
"get of X. To generate an adversarial example, the goal",
"numerical instability, we normalize rm as",
"is to make the predictions of all targets go wrong, i.e.,",
"r'm=",
"2",
"Vn, argmax (fe(X+r,tn)) + In. Here r denotes an",
".rm",
"(3)",
"Ilrmlloo",
"adversarial perturbation added to X.To this end, we",
"specify an adversarial label I', for each target, in which",
"where = 0.5 is a fixed hyper-parameter. We then add rm",
"I'n is randomly sampled from other incorrect classes, i.e.,",
"to the current image Xm and proceed to the next iteration.",
"3"
] |
{
"arxiv_id": "1703.08603",
"language": "en",
"page_idx": 2
}
|
||
1703.08603-en-3
|
1703.08603-en-3.jpg
|
The algorithm terminates if either all the targets are predicted as desired, $i.e.,\mathcal{T}_m=\varnothing$, or it reaches the maximum iteration number, which is set to be 200 in segmentation and 150 in detection.
The final adversarial perturbation is computed as r = $\sum_{m}\mathbf{r}_{m}^{\prime}.$ Note that, in practice, we often obtain the input image X after subtracting the mean image $\widehat{\mathbf{X}}.$ In this case, the adversarial image is Trunc$\left(\mathbf{X}+\mathbf{r}+\widehat{\mathbf{X}}\right)$,where Trunc$(\cdot)$ denotes the function that truncates every pixel value by [0,255]. Although truncation may harm the adversarial perturbation, we observed little effect in experiments, mainly because the magnitude of perturbation r is very small (see Section 3.5.3). The overall pipeline of DAG algorithm is illustrated in Algorithm 1.
3.2. Selecting Input Proposals for Detection
A critical issue in DAG is to select a proper set $\mathcal{T}$ of targets. This is relatively easy in the semantic segmentation task, because the goal is to produce incorrect classification on all pixels, and thus we can set each of them as a separate target, $i.e.$, performing dense sampling on the image lattice. This is tractable, $i.e.$, the computational complexity is proportional to the total number of pixels.
In the scenario of object detection, target selection becomes a lot more difficult, as the total number of possible targets (bounding box proposals) is orders of magnitudes larger than that in semantic segmentation. A straightforward choice is to only consider the proposals generated by a sideway network, e.g., the regional proposal network (RPN)[28], but we find that when the adversarial perturbation r is added to the original image $\mathbf{X}$, a different set of proposals may be generated according to the new input $\mathbf{X+r,~and~the~network~may~still~be~able~to~correctly~classify}$ these new proposals[20]. To overcome this problem, we make the proposals very dense by increasing the threshold of NMS in RPN. In practice, when the intersection-overunion (IOU) goes up from 0.70 to 0.90,the average number of proposals on each image increases from around 300 to around 3000. Using this denser target set $\mathcal{T}$,most probable object bounding boxes are only pixels away from at least one of the selected input proposals, and we can expect the classification error transfers among neighboring bounding boxes. As shown in experiments, this heuristic idea works very well, and the effect of adversarial perturbations is positively correlated to the number of proposals considered $\operatorname{in}$DAG.
Technically, given the proposals generated by RPN, we preserve all positive proposals and discard the remaining. Here, a positive proposal satisfes the following two con- ditions: 1) the IOU with the closest ground-truth object is greater than 0.1, and 2) the confdence score for the corresponding ground-truth class is greater than 0.1. If both conditions hold on multiple ground-truth objects, we select
<table>
<tbody>
<tr>
<th>Network</th>
<th>$ORIG$</th>
<th>ADVR</th>
<th>PERM</th>
</tr>
<tr>
<td>$\mathbf{FCN-Alex}$</td>
<td>48.04</td>
<td>3.98</td>
<td>48.04</td>
</tr>
<tr>
<td>$\mathbf{FCN-Alex}^{*}$</td>
<td>48.92</td>
<td>3.98</td>
<td>48.91</td>
</tr>
<tr>
<td>$\mathbf{FCN-VGG}$</td>
<td>65.49</td>
<td>4.09</td>
<td>65.47</td>
</tr>
<tr>
<td>$\mathbf{FCN-VGG}^{*}$</td>
<td>67.09</td>
<td>4.18</td>
<td>67.08</td>
</tr>
<tr>
<td>$\mathbf{FR-ZF-07}$</td>
<td>58.70</td>
<td>3.61</td>
<td>58.33</td>
</tr>
<tr>
<td>$FR-ZF-0712$</td>
<td>61.07</td>
<td>1.95</td>
<td>60.94</td>
</tr>
<tr>
<td>$FR-VGG-07$</td>
<td>69.14</td>
<td>5.92</td>
<td>68.68</td>
</tr>
<tr>
<td>$FR-VGG-0712$</td>
<td>72.07</td>
<td>3.36</td>
<td>71.97</td>
</tr>
</tbody>
</table>
Table 1: Semantic segmentation (measured by mIOU, %)
and object detection (measured by mAP, %) results of different networks. Here, ORIG represents the accuracy obtained on the original image set, ADVR is obtained on the set after the adversarial perturbations are added, and PERM is obtained after the randomly permuted perturbations are added. Please see Section 3.3 for details.
the one with the maximal IOU. The label of the proposal is defined as the corresponding confident class. This strategy aims at selecting high-quality targets for Algorithm 1.
## 3.3. Quantitative Evaluation
Following some previous work [33][24], we evaluate our approach by measuring the drop in recognition accuracy, $i.e.$, mean intersection-over-union (mIOU) for semantic segmentation and mean average precision (mAP) for object detection, using the original test images and the ones after adding adversarial perturbations$^{1}.$
$\bullet$ For semantic segmentation, we study two network ar-
chitectures based on the FCN [19] framework. One of them is based on the AlexNet[14] and the other one is based on the 16-layer VGGNet[31]. Both networks have two variants. We use FCN-Alex and FCN-VGG, which are publicly available, to denote the networks that are trained on the original FCN[19] training set which has 9610 images, and use FCNAlex* and FCN-VGG* to denote the networks that are trained on the DeepLab [3] training set which has 10582 images. We use the validation set in [19] which has 736 images as our semantic segmentation test set.
$\bullet$ For object detection, based on the Faster-RCNN[28]
framework, we study two network architectures, $i.e.$, the ZFNet [37] and the 16-layer VGGNet [31]. Both networks have two variants, which are either trained on the PascalVOC-2007 trainval set, or the combined PascalVOC-2007 and PascalVOC-2012 trainval sets. These four models are publicly available, and are denoted as FR-ZF-07, FR-ZF-0712, FR-VGG-07 and
$^{1}$For implementation simplicity, we keep targets with ground-truth class
label background unchanged when generating adversarial examples.
4
|
[
[
0.56396484375,
0.09027099609375,
0.62353515625,
0.09027099609375,
0.62353515625,
0.10479736328125,
0.56396484375,
0.10479736328125
],
[
0.66845703125,
0.09027099609375,
0.7138671875,
0.09027099609375,
0.7138671875,
0.10479736328125,
0.66845703125,
0.10479736328125
],
[
0.73388671875,
0.09027099609375,
0.78369140625,
0.09027099609375,
0.78369140625,
0.10479736328125,
0.73388671875,
0.10479736328125
],
[
0.79833984375,
0.09027099609375,
0.85302734375,
0.09027099609375,
0.85302734375,
0.10479736328125,
0.79833984375,
0.10479736328125
],
[
0.5537109375,
0.1060791015625,
0.63134765625,
0.10797119140625,
0.63134765625,
0.12371826171875,
0.5537109375,
0.121826171875
],
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.6689453125,
0.10919189453125,
0.71337890625,
0.10919189453125,
0.71337890625,
0.12371826171875,
0.6689453125,
0.12371826171875
],
[
0.740234375,
0.10797119140625,
0.775390625,
0.10797119140625,
0.775390625,
0.12371826171875,
0.740234375,
0.12371826171875
],
[
0.80224609375,
0.10919189453125,
0.8486328125,
0.10919189453125,
0.8486328125,
0.12371826171875,
0.80224609375,
0.12371826171875
],
[
0.080078125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.6708984375,
0.125,
0.71337890625,
0.125,
0.71337890625,
0.1395263671875,
0.6708984375,
0.1395263671875
],
[
0.740234375,
0.125,
0.77392578125,
0.125,
0.77392578125,
0.14013671875,
0.740234375,
0.14013671875
],
[
0.80224609375,
0.125,
0.84716796875,
0.125,
0.84716796875,
0.1395263671875,
0.80224609375,
0.1395263671875
],
[
0.08087158203125,
0.1395263671875,
0.19287109375,
0.1395263671875,
0.19287109375,
0.149658203125,
0.08087158203125,
0.149658203125
],
[
0.5537109375,
0.1256103515625,
0.63232421875,
0.1256103515625,
0.63232421875,
0.136962890625,
0.5537109375,
0.136962890625
],
[
0.5537109375,
0.14013671875,
0.63232421875,
0.14013671875,
0.63232421875,
0.1546630859375,
0.5537109375,
0.1546630859375
],
[
0.6708984375,
0.14013671875,
0.7138671875,
0.14013671875,
0.7138671875,
0.1546630859375,
0.6708984375,
0.1546630859375
],
[
0.74267578125,
0.141357421875,
0.77294921875,
0.141357421875,
0.77294921875,
0.154052734375,
0.74267578125,
0.154052734375
],
[
0.80126953125,
0.14013671875,
0.84619140625,
0.1383056640625,
0.84716796875,
0.154052734375,
0.80224609375,
0.1558837890625
],
[
0.098876953125,
0.15283203125,
0.46728515625,
0.15283203125,
0.46728515625,
0.1666259765625,
0.098876953125,
0.1666259765625
],
[
0.5498046875,
0.1558837890625,
0.6357421875,
0.1558837890625,
0.6357421875,
0.17041015625,
0.5498046875,
0.17041015625
],
[
0.740234375,
0.1558837890625,
0.775390625,
0.1558837890625,
0.775390625,
0.1717529296875,
0.740234375,
0.1717529296875
],
[
0.8037109375,
0.1558837890625,
0.84716796875,
0.1558837890625,
0.84716796875,
0.1717529296875,
0.8037109375,
0.1717529296875
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1666259765625,
0.468017578125,
0.18115234375,
0.080078125,
0.18310546875
],
[
0.6708984375,
0.1572265625,
0.71337890625,
0.1572265625,
0.71337890625,
0.1717529296875,
0.6708984375,
0.1717529296875
],
[
0.5556640625,
0.1749267578125,
0.6298828125,
0.1749267578125,
0.6298828125,
0.189453125,
0.5556640625,
0.189453125
],
[
0.6689453125,
0.1749267578125,
0.71337890625,
0.1749267578125,
0.71337890625,
0.1900634765625,
0.6689453125,
0.1900634765625
],
[
0.740234375,
0.1749267578125,
0.77294921875,
0.1749267578125,
0.77294921875,
0.1900634765625,
0.740234375,
0.1900634765625
],
[
0.80224609375,
0.1749267578125,
0.8486328125,
0.1749267578125,
0.8486328125,
0.1900634765625,
0.80224609375,
0.1900634765625
],
[
0.0767822265625,
0.18310546875,
0.4697265625,
0.1817626953125,
0.4697265625,
0.199462890625,
0.0767822265625,
0.2008056640625
],
[
0.54833984375,
0.189453125,
0.63818359375,
0.189453125,
0.63818359375,
0.203857421875,
0.54833984375,
0.203857421875
],
[
0.6708984375,
0.1912841796875,
0.71337890625,
0.1912841796875,
0.71337890625,
0.205810546875,
0.6708984375,
0.205810546875
],
[
0.74169921875,
0.1900634765625,
0.775390625,
0.1900634765625,
0.775390625,
0.205810546875,
0.74169921875,
0.205810546875
],
[
0.8037109375,
0.1912841796875,
0.84716796875,
0.1912841796875,
0.84716796875,
0.205810546875,
0.8037109375,
0.205810546875
],
[
0.07843017578125,
0.20263671875,
0.46728515625,
0.20263671875,
0.46728515625,
0.2171630859375,
0.07843017578125,
0.2171630859375
],
[
0.6708984375,
0.205810546875,
0.71337890625,
0.205810546875,
0.71337890625,
0.2203369140625,
0.6708984375,
0.2203369140625
],
[
0.74169921875,
0.20703125,
0.77392578125,
0.20703125,
0.77392578125,
0.2197265625,
0.74169921875,
0.2197265625
],
[
0.8037109375,
0.205810546875,
0.84716796875,
0.205810546875,
0.84716796875,
0.2215576171875,
0.8037109375,
0.2215576171875
],
[
0.080078125,
0.2215576171875,
0.468017578125,
0.2215576171875,
0.468017578125,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.5498046875,
0.2083740234375,
0.6357421875,
0.2083740234375,
0.6357421875,
0.2197265625,
0.5498046875,
0.2197265625
],
[
0.5390625,
0.222900390625,
0.64794921875,
0.222900390625,
0.64794921875,
0.236083984375,
0.5390625,
0.236083984375
],
[
0.74169921875,
0.222900390625,
0.77392578125,
0.222900390625,
0.77392578125,
0.23486328125,
0.74169921875,
0.23486328125
],
[
0.8037109375,
0.2215576171875,
0.84619140625,
0.2215576171875,
0.84619140625,
0.2374267578125,
0.8037109375,
0.2374267578125
],
[
0.080078125,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.6708984375,
0.2235107421875,
0.71142578125,
0.2235107421875,
0.71142578125,
0.23486328125,
0.6708984375,
0.23486328125
],
[
0.50341796875,
0.2392578125,
0.89208984375,
0.2392578125,
0.89208984375,
0.25390625,
0.50341796875,
0.25390625
],
[
0.080078125,
0.251953125,
0.468017578125,
0.251953125,
0.468017578125,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50244140625,
0.255126953125,
0.8935546875,
0.255126953125,
0.8935546875,
0.26953125,
0.50244140625,
0.26953125
],
[
0.07843017578125,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.281005859375,
0.07843017578125,
0.281005859375
],
[
0.50244140625,
0.270751953125,
0.89208984375,
0.270751953125,
0.89208984375,
0.2841796875,
0.50244140625,
0.2841796875
],
[
0.080078125,
0.2822265625,
0.4697265625,
0.2822265625,
0.4697265625,
0.29541015625,
0.080078125,
0.29541015625
],
[
0.50341796875,
0.287353515625,
0.890625,
0.287353515625,
0.890625,
0.29736328125,
0.50341796875,
0.29736328125
],
[
0.07843017578125,
0.29541015625,
0.253173828125,
0.296630859375,
0.253173828125,
0.311279296875,
0.07843017578125,
0.31005859375
],
[
0.50244140625,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.314453125,
0.50244140625,
0.314453125
],
[
0.080078125,
0.3193359375,
0.410888671875,
0.3193359375,
0.410888671875,
0.333984375,
0.080078125,
0.333984375
],
[
0.50244140625,
0.31640625,
0.89208984375,
0.31640625,
0.89208984375,
0.330078125,
0.50244140625,
0.330078125
],
[
0.50244140625,
0.330810546875,
0.771484375,
0.330810546875,
0.771484375,
0.3447265625,
0.50244140625,
0.3447265625
],
[
0.098876953125,
0.343505859375,
0.4697265625,
0.343505859375,
0.4697265625,
0.356689453125,
0.098876953125,
0.356689453125
],
[
0.08087158203125,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.37060546875,
0.08087158203125,
0.37060546875
],
[
0.50244140625,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.50244140625,
0.37060546875
],
[
0.080078125,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.5009765625,
0.371826171875,
0.890625,
0.373779296875,
0.890625,
0.386962890625,
0.5009765625,
0.385009765625
],
[
0.080078125,
0.389404296875,
0.468017578125,
0.389404296875,
0.468017578125,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.50244140625,
0.38818359375,
0.85693359375,
0.38818359375,
0.85693359375,
0.401611328125,
0.50244140625,
0.401611328125
],
[
0.07843017578125,
0.40283203125,
0.468017578125,
0.404052734375,
0.468017578125,
0.41845703125,
0.07843017578125,
0.417236328125
],
[
0.080078125,
0.419189453125,
0.468017578125,
0.419189453125,
0.468017578125,
0.43310546875,
0.080078125,
0.43310546875
],
[
0.50341796875,
0.415283203125,
0.72119140625,
0.415283203125,
0.72119140625,
0.429931640625,
0.50341796875,
0.429931640625
],
[
0.080078125,
0.434326171875,
0.369384765625,
0.434326171875,
0.369384765625,
0.447509765625,
0.080078125,
0.447509765625
],
[
0.521484375,
0.439453125,
0.89208984375,
0.439453125,
0.89208984375,
0.453857421875,
0.521484375,
0.453857421875
],
[
0.098876953125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.46337890625,
0.098876953125,
0.46337890625
],
[
0.50244140625,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.080078125,
0.464111328125,
0.4697265625,
0.464111328125,
0.4697265625,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50244140625,
0.4697265625,
0.890625,
0.4697265625,
0.890625,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.07843017578125,
0.479248046875,
0.4697265625,
0.477783203125,
0.4697265625,
0.492431640625,
0.07843017578125,
0.49365234375
],
[
0.50244140625,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.08087158203125,
0.49560546875,
0.46728515625,
0.49560546875,
0.46728515625,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.50244140625,
0.50048828125,
0.89208984375,
0.50048828125,
0.89208984375,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.50244140625,
0.51513671875,
0.72607421875,
0.51513671875,
0.72607421875,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.080078125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.07843017578125,
0.53857421875,
0.468017578125,
0.5390625,
0.468017578125,
0.5537109375,
0.07843017578125,
0.55322265625
],
[
0.5185546875,
0.54541015625,
0.89208984375,
0.546875,
0.89208984375,
0.56103515625,
0.5185546875,
0.56005859375
],
[
0.07843017578125,
0.5546875,
0.4697265625,
0.5546875,
0.4697265625,
0.568359375,
0.07843017578125,
0.568359375
],
[
0.53515625,
0.5625,
0.89208984375,
0.5625,
0.89208984375,
0.57568359375,
0.53515625,
0.57568359375
],
[
0.08087158203125,
0.57080078125,
0.468017578125,
0.57080078125,
0.468017578125,
0.583984375,
0.08087158203125,
0.583984375
],
[
0.53515625,
0.5771484375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59033203125,
0.53515625,
0.59033203125
],
[
0.08087158203125,
0.58544921875,
0.46728515625,
0.58544921875,
0.46728515625,
0.5986328125,
0.08087158203125,
0.5986328125
],
[
0.53515625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.53515625,
0.60595703125
],
[
0.080078125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.080078125,
0.6142578125
],
[
0.53515625,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.53515625,
0.62060546875
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.53515625,
0.62255859375,
0.89208984375,
0.62255859375,
0.89208984375,
0.63623046875,
0.53515625,
0.63623046875
],
[
0.07843017578125,
0.62890625,
0.468017578125,
0.6298828125,
0.468017578125,
0.64453125,
0.07843017578125,
0.64306640625
],
[
0.53515625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.53515625,
0.6513671875
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.5361328125,
0.65283203125,
0.890625,
0.65283203125,
0.890625,
0.66748046875,
0.5361328125,
0.66748046875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.53515625,
0.66650390625,
0.89208984375,
0.66748046875,
0.89208984375,
0.681640625,
0.53515625,
0.68115234375
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5361328125,
0.68505859375,
0.890625,
0.68505859375,
0.890625,
0.69580078125,
0.5361328125,
0.69580078125
],
[
0.5361328125,
0.69873046875,
0.89208984375,
0.69873046875,
0.89208984375,
0.7119140625,
0.5361328125,
0.7119140625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.53515625,
0.71337890625,
0.88818359375,
0.71337890625,
0.88818359375,
0.7265625,
0.53515625,
0.7265625
],
[
0.07843017578125,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.07843017578125,
0.734375
],
[
0.080078125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.51953125,
0.7421875,
0.890625,
0.7421875,
0.890625,
0.755859375,
0.51953125,
0.755859375
],
[
0.080078125,
0.75244140625,
0.468017578125,
0.75244140625,
0.468017578125,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.53515625,
0.75830078125,
0.890625,
0.75830078125,
0.890625,
0.771484375,
0.53515625,
0.771484375
],
[
0.07843017578125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.07843017578125,
0.77978515625
],
[
0.53515625,
0.77294921875,
0.89208984375,
0.77294921875,
0.89208984375,
0.7861328125,
0.53515625,
0.7861328125
],
[
0.080078125,
0.7822265625,
0.1356201171875,
0.7822265625,
0.1356201171875,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.53515625,
0.787109375,
0.89208984375,
0.787109375,
0.89208984375,
0.80029296875,
0.53515625,
0.80029296875
],
[
0.09722900390625,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.09722900390625,
0.81005859375
],
[
0.53515625,
0.80322265625,
0.89208984375,
0.80322265625,
0.89208984375,
0.81640625,
0.53515625,
0.81640625
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.53369140625,
0.818359375,
0.8896484375,
0.818359375,
0.8896484375,
0.83203125,
0.53369140625,
0.83203125
],
[
0.07843017578125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.53515625,
0.833984375,
0.890625,
0.833984375,
0.890625,
0.84765625,
0.53515625,
0.84765625
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.53515625,
0.8486328125,
0.89208984375,
0.8486328125,
0.89208984375,
0.8623046875,
0.53515625,
0.8623046875
],
[
0.07843017578125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.0767822265625,
0.8720703125,
0.4697265625,
0.87060546875,
0.4697265625,
0.88525390625,
0.0767822265625,
0.88623046875
],
[
0.5185546875,
0.87353515625,
0.89208984375,
0.875,
0.89208984375,
0.8896484375,
0.5185546875,
0.88818359375
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.86279296875,
0.88720703125,
0.86279296875,
0.9013671875,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"Network",
"ORIG",
"ADVR",
"PERM",
"FCN-Alex",
"The algorithm terminates if either all the targets are pre-",
"dicted as desired.i.e.,Tm =O,or itreachesthe maximum",
"48.04",
"3.98",
"48.04",
"iteration number, which is set to be 200 in segmentation and",
"48.92",
"3.98",
"48.91",
"150indetection.",
"FCN-Alex*",
"FCN-VGG",
"65.49",
"4.09",
"65.47",
"The final adversarial perturbation is computed as r =",
"FCN-VGG*",
"4.18",
"67.08",
"mr'm.Note that, in practice, we often obtain the input im-",
"67.09",
"FR-ZF-07",
"58.70",
"3.61",
"58.33",
" age X after subtracting the mean image X. In this case, the",
"FR-ZF-0712",
"61.07",
"1.95",
"60.94",
"adversarial image is Trunc(x + r + X), where Trunc()",
"69.14",
"5.92",
"68.68",
"denotes the function that truncates every pixel value by",
"FR-VGG-07",
"FR-VGG-0712",
"3.36",
"71.97",
"[0, 255]. Although truncation may harm the adversarial per-",
"72.07",
"Table 1: Semantic segmentation (measured by mIOU, %)",
"turbation, we observed little effect in experiments, mainly",
"and object detection (measured by mAP, %) results of",
"because the magnitude of perturbation r is very small (see",
"different networks. Here, ORIG represents the accuracy",
"Section 3.5.3). The overall pipeline of DAG algorithm is",
"obtained ontheoriginal imageset,ADVRisobtainedonthe",
"illustrated in Algorithm 1.",
"set after the adversarial perturbations are added, and PERM",
"3.2. Selecting Input Proposals for Detection",
"is obtained after the randomly permuted perturbations are",
"added. Please see Section 3.3 for details.",
"A critical issue in DAG is to select a proper set T of",
"targets.This isrelatively easyinthe semantic segmentation",
"the one with the maximal IOU. The label of the proposal is",
"task, because the goal is to produce incorrect classification",
"defined as the corresponding confident class. This strategy",
"on all pixels, and thus we can set each of them as a sep-",
"aims at selecting high-quality targets for Algorithm 1.",
"arate target, i.e., performing dense sampling on the image",
"lattice. This is tractable, i.e., the computational complexity",
"3.3.Quantitative Evaluation",
"is proportional to the total number of pixels.",
"Following some previous work [33][24], we evaluate our",
"In the scenario of object detection, target selection be-",
"approach by measuring the drop in recognition accuracy,",
"comes a lot more difficult, as the total number of possible",
"i.e., mean intersection-over-union (mIOU) for semantic seg-",
"targets (bounding box proposals) is orders of magnitudes",
"mentation and mean average precision (mAP) for object",
"larger than that in semantic segmentation.A straightfor-",
"detection, using the original test images and the ones after",
"ward choice is to only consider the proposals generated",
"adding adversarial perturbations'.",
"by a sideway network, e.g., the regional proposal network",
"(RPN) [28], but we find that when the adversarial pertur-",
". For semantic segmentation, we study two network ar-",
"bation r is added to the original image X, a different set",
"chitectures based on the FCN [19] framework. One",
"of proposals may be generated according to the new input",
"of them is based on the AlexNet [14] and the other",
"X +r, and the network may still be able to correctly classify",
"one is based on the 16-layer VGGNet [31]. Both",
"these new proposals [20]. To overcome this problem, we",
"networks have two variants.We use FCN-Alex and",
"make the proposals very dense by increasing the threshold",
"FCN-VGG, which are publicly available, to denote",
"of NMS in RPN. In practice, when the intersection-over-",
"the networks that are trained on the original FCN [19]",
"union (1OU) goes up from 0.70 to 0.90, the average number",
"training set which has 9610 images, and use FCN-",
"of proposals on each image increases from around 300 to",
"Alex* and FCN-VGG* to denote the networks that",
"around 3000. Using this denser target set T, most probable",
"object bounding boxes are only pixels away from at least",
"aretrainedontheDeepLab[3ltraining setwhichhas",
"10582 images. We use the validation set in [19] which",
"one of the selected input proposals, and we can expect the",
"has 736 images as our semantic segmentation test set.",
"classification error transfers among neighboring bounding",
"boxes.As shown in experiments. this heuristic idea works",
". For object detection, based on the Faster-RCNN [28]",
"very well, and the effect of adversarial perturbations is",
"framework, we study two network architectures, i.e",
"positively correlated to the number of proposals considered",
"the ZFNet [37] and the 16-layer VGGNet [31]. Both",
"in DAG.",
"networks have two variants, which are either trained",
"Technically, given the proposals generated by RPN, we",
"on the PascalVOC-2007 trainval set, or the combined",
"preserve all positive proposals and discard the remaining.",
"PascalVOC-2007 and PascalVOC-2012 trainval sets.",
"Here, a positive proposal satisfies the following two con-",
"These four models are publicly available, and are de-",
"ditions: 1) the IOU with the closest ground-truth object",
"noted as FR-ZF-07, FR-ZF-0712, FR-VGG-07 and",
"is greater than 0.1, and 2) the confidence score for the",
" corresponding ground-truth class is greater than 0.1. If both",
"'Forimplementation simplicity,wekeep targets with ground-truth class",
"conditions hold on multiple ground-truth objects, we select",
"label background unchanged when generating adversarial examples.",
"4"
] |
{
"arxiv_id": "1703.08603",
"language": "en",
"page_idx": 3
}
|
||
1703.08603-en-4
|
1703.08603-en-4.jpg
|
FR-VGG-0712, respectively. We use the PascalVOC- 2007 test set which has 4952 images as our object detection test set.
Results are summarized in Table 1. We can observe that the accuracy (mIOU for segmentation and mAP for detection) drops significantly after the adversarial perturbations are added, demonstrating the effectiveness of DAG algorithm. Moreover, for detection, the networks with more training data are often more sensitive to the adversarial perturbation. This is verifed by the fact that FR-ZF-07 (from $58.70\%$ to 3.61%) has a smaller performance drop than FR- $\textbf{ZF- 0712( from61. 07\% to 1. 95\% ) , and that FR- VGG- 07}$ (from 69.14% to 5.92%) has a smaller performance drop than FR-VGG-0712 (from 72.04% to 3.36%).
To verify the importance of the spatial structure of adversarial perturbations, we evaluate the accuracy after randomly permuting the rows and/or columns of r. In Table 1, we find that permuted perturbations cause negligible accuracy drop, indicating that it is the spatial structure of r, instead of its magnitude, that indeed contributes in generating adversarial examples. For permutation results, we randomly permute r for three times and take the average.
## 3.4. Adversarial Examples
Figure 1 shows an adversarial example that fails in both detection and segmentation networks. In addition, we show that DAG is able to control the output of adversarial images very well. In Figure 2, we apply DAG to generating one adversarial image (which humans can recognize but deep networks cannot) and one fooling image [25] (which is completely unrecognizable to humans but deep networks produce false positives). This suggests that deep networks only cover a limited area in the high-dimensional feature space, and that we can easily find adversarial and/or fooling examples that fall in the unexplored parts.
## 3.5. Diagnostics
3.5.1 $\textbf{The Denseness of Proposals}$
We first observe the impact on adversarial generation of the denseness of the proposals. To this end, we use different IOU rates in the NMS process after the RPN. This directly affects the number of proposals preserved in Algorithm 1. As we can see in Figure 3, the mAP value goes down $(i.e.$, stronger adversarial perturbations are generated) as the IOU rate increases, which means that fewer proposals are filtered out and thus the set of targets $\mathcal{T}$ becomes larger. This is in line of our expectation, since DAG only guarantees misclassification on the targets in $\mathcal{T}.$ The denser sampling on proposals allows the recognition error to propagate to other possible object positions better. Therefore, we choose a large IOU value (0.90) which produces good results.
<FigureHere>
Figure 2: Fancy examples generated by DAG for semantic
segmentation. The adversarial image is on the left and the fooling image is on the right. From top to bottom: the original image, the perturbation (magnified by 10), the adversarial image after adding perturbation, and the segmentation results. The red, blue and black regions are predicted as airplane, bus and background, respectively.
# 3.5.2 Convergence
We then investigate the convergence of DAG, $i.e.$, how many iterations are needed to find the desired adversarial perturbation. Figure 4 shows the number of active targets, $i.e.,|\mathcal{T}_m|$, with respect to the number of iterations $m.$ In general, the training process goes smoothly in the early rounds, in which we find that the number of active proposals is significantly reduced. After the algorithm reaches the maximal number of iterations, $i.e.,200$ in segmentation and 150 in detection, only few (less than 1%) image fail to converge. Even on these cases, DAG is able to produce reasonable adversarial perturbations.
Another interesting observation is the difficulty in generating adversarial examples. In general, the detection networks are more difficult to attack than the segmentation
5
|
[
[
0.11358642578125,
0.091552734375,
0.46728515625,
0.091552734375,
0.46728515625,
0.1060791015625,
0.11358642578125,
0.1060791015625
],
[
0.11273193359375,
0.1060791015625,
0.468017578125,
0.106689453125,
0.468017578125,
0.121826171875,
0.11273193359375,
0.12060546875
],
[
0.11358642578125,
0.125,
0.2279052734375,
0.125,
0.2279052734375,
0.1351318359375,
0.11358642578125,
0.1351318359375
],
[
0.098876953125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.098876953125,
0.1666259765625
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.08087158203125,
0.2008056640625,
0.46728515625,
0.2008056640625,
0.46728515625,
0.2108154296875,
0.08087158203125,
0.2108154296875
],
[
0.080078125,
0.2139892578125,
0.4697265625,
0.2139892578125,
0.4697265625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.080078125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.24365234375,
0.080078125,
0.24365234375
],
[
0.080078125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.080078125,
0.25830078125
],
[
0.08087158203125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.08087158203125,
0.2734375
],
[
0.08087158203125,
0.274658203125,
0.468017578125,
0.274658203125,
0.468017578125,
0.28857421875,
0.08087158203125,
0.28857421875
],
[
0.080078125,
0.28857421875,
0.468017578125,
0.2890625,
0.468017578125,
0.3037109375,
0.080078125,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.38818359375,
0.304931640625,
0.38818359375,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.09722900390625,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.335205078125,
0.09722900390625,
0.335205078125
],
[
0.080078125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.080078125,
0.349853515625
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.080078125,
0.3662109375,
0.468017578125,
0.3662109375,
0.468017578125,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.080078125,
0.380615234375,
0.46728515625,
0.380615234375,
0.46728515625,
0.39453125,
0.080078125,
0.39453125
],
[
0.080078125,
0.396484375,
0.468017578125,
0.396484375,
0.468017578125,
0.410400390625,
0.080078125,
0.410400390625
],
[
0.080078125,
0.410888671875,
0.468017578125,
0.410888671875,
0.468017578125,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.080078125,
0.4267578125,
0.386474609375,
0.4267578125,
0.386474609375,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.51318359375,
0.43994140625,
0.69384765625,
0.43994140625,
0.69384765625,
0.52099609375,
0.51318359375,
0.52099609375
],
[
0.69921875,
0.43994140625,
0.8857421875,
0.43994140625,
0.8857421875,
0.52099609375,
0.69921875,
0.52099609375
],
[
0.080078125,
0.451904296875,
0.283447265625,
0.45263671875,
0.283447265625,
0.46728515625,
0.080078125,
0.466552734375
],
[
0.098876953125,
0.4765625,
0.46728515625,
0.4765625,
0.46728515625,
0.490478515625,
0.098876953125,
0.490478515625
],
[
0.08087158203125,
0.494384765625,
0.46728515625,
0.494384765625,
0.46728515625,
0.5048828125,
0.08087158203125,
0.5048828125
],
[
0.07843017578125,
0.5068359375,
0.4697265625,
0.50830078125,
0.4697265625,
0.521484375,
0.07843017578125,
0.52099609375
],
[
0.07843017578125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.07843017578125,
0.537109375
],
[
0.0767822265625,
0.537109375,
0.4697265625,
0.53857421875,
0.4697265625,
0.55322265625,
0.0767822265625,
0.5517578125
],
[
0.50341796875,
0.54052734375,
0.89208984375,
0.54052734375,
0.89208984375,
0.5537109375,
0.50341796875,
0.5537109375
],
[
0.080078125,
0.5537109375,
0.4697265625,
0.5537109375,
0.4697265625,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50341796875,
0.55615234375,
0.8935546875,
0.55615234375,
0.8935546875,
0.5693359375,
0.50341796875,
0.5693359375
],
[
0.080078125,
0.568359375,
0.4697265625,
0.568359375,
0.4697265625,
0.58203125,
0.080078125,
0.58203125
],
[
0.50244140625,
0.5712890625,
0.890625,
0.5712890625,
0.890625,
0.58544921875,
0.50244140625,
0.58544921875
],
[
0.080078125,
0.583984375,
0.4697265625,
0.583984375,
0.4697265625,
0.59765625,
0.080078125,
0.59765625
],
[
0.50341796875,
0.5859375,
0.890625,
0.5859375,
0.890625,
0.599609375,
0.50341796875,
0.599609375
],
[
0.08251953125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.611328125,
0.08251953125,
0.611328125
],
[
0.50244140625,
0.6015625,
0.89208984375,
0.6015625,
0.89208984375,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.08087158203125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.08087158203125,
0.62744140625
],
[
0.50244140625,
0.6162109375,
0.89208984375,
0.6162109375,
0.89208984375,
0.6298828125,
0.50244140625,
0.6298828125
],
[
0.07843017578125,
0.62744140625,
0.35693359375,
0.62890625,
0.35693359375,
0.64306640625,
0.07843017578125,
0.642578125
],
[
0.50244140625,
0.6318359375,
0.869140625,
0.6318359375,
0.869140625,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.080078125,
0.65576171875,
0.2017822265625,
0.65576171875,
0.2017822265625,
0.67041015625,
0.080078125,
0.67041015625
],
[
0.50244140625,
0.66357421875,
0.64453125,
0.66357421875,
0.64453125,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.07843017578125,
0.67822265625,
0.324462890625,
0.6787109375,
0.324462890625,
0.693359375,
0.07843017578125,
0.6923828125
],
[
0.50341796875,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50341796875,
0.70263671875
],
[
0.08087158203125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.08087158203125,
0.76513671875
],
[
0.50341796875,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08087158203125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50244140625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.5009765625,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.74267578125,
0.84228515625,
0.74267578125,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.51953125,
0.85498046875,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.51953125,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.43701171875,
0.88720703125,
0.43701171875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"FR-VGG-0712, respectively. We use the PascalVOC-",
"2007 test set which has 4952 images as our object",
"detection test set.",
"Results are summarized in Table 1.We can observe",
"that the accuracy (mIOU for segmentation and mAP for",
"detection) drops significantly after the adversarial pertur-",
"bationsareadded.demonstratingtheeffectivenessofDAG",
"algorithm. Moreover, for detection, the networks with more",
"training data are often more sensitive to the adversarial per-",
"turbation. This is verified by the fact that FR-ZF-07 (from",
"58.70% to 3.61%) has a smaller performance drop than FR-",
"ZF-0712 (from 61.07% to 1.95%), and that FR-VGG-07",
"(from 69.14% to 5.92%) has a smaller performance drop",
"than FR-VGG-0712 (from 72.04% to 3.36%).",
"To verify the importance of the spatial structure of ad-",
"versarial perturbations, we evaluate the accuracy after ran-",
"domly permuting the rows and/or columns of r. In Table 1,",
"we find that permuted perturbations cause negligible accu-",
"racy drop, indicating that it is the spatial structure of r, in-",
"stead of its magnitude, that indeed contributes in generating",
"adversarial examples. For permutation results, we randomly",
"permute r for three times and take the average.",
"ICCV",
"2017",
"3.4.AdversarialExamples",
"Figure 1 shows an adversarial example that fails in both",
"detection and segmentationnetworks.In addition,we show",
"that DAG is able to control the output of adversarial images",
"very well. In Figure 2, we apply DAG to generating one",
"adversarial image (which humans can recognize but deep",
"Figure 2: Fancy examples generated by DAG for semantic",
"networks cannot) and one fooling image [25] (which is",
"segmentation.The adversarial image is on the left and",
"completely unrecognizable to humans but deep networks",
"the fooling image is on the right.From top to bottom:",
"produce false positives). This suggests that deep networks",
"the original image, the perturbation (magnified by 10),",
"only cover a limited area in the high-dimensional feature",
"the adversarial image after adding perturbation, and the",
"space, and that we can easily find adversarial and/or fooling",
"segmentation results. The red, blue and black regions are",
"examples that fall in the unexplored parts.",
"predicted as airplane, bus and background, respectively",
"3.5.Diagnostics",
"3.5.2Convergence",
"3.5.1 The Denseness of Proposals",
"We then investigate the convergence of DAG, i.e., how",
"We first observe the impact on adversarial generation of the",
"many iterations are needed to find the desired adversarial",
"denseness of the proposals. To this end, we use different",
"perturbation. Figure 4 shows the number of active targets,",
"IOU rates in the NMS process after the RPN. This directly",
"i.e., ITml, with respect to the number of iterations m. In",
"affects the number of proposals preserved in Algorithm 1.",
"general, the training process goes smoothly in the early",
"As we can see in Figure 3, the mAP value goes down (i.e.,",
"rounds, in which we find that the number of active proposals",
"stronger adversarial perturbations are generated) as the IOU",
"is significantly reduced. After the algorithm reaches the",
"rate increases, which means that fewer proposals are filtered",
"maximal number of iterations, i.e., 200 in segmentation",
"out and thus the set of targets T becomes larger. This",
"and 150 in detection, only few (less than 1%) image fail",
"is in line of our expectation, since DAG only guarantees",
"to converge. Even on these cases, DAG is able to produce",
"misclassification on the targets in T. The denser sampling",
"reasonableadversarial perturbations.",
"on proposals allows the recognition error to propagate to",
"Another interesting observation is the difficulty in gen-",
"other possible object positions better. Therefore, we choose",
"erating adversarial examples. In general, the detection",
"a large IOU value (0.90) which produces good results.",
"networks are more difficult to attack than the segmentation",
"5"
] |
{
"arxiv_id": "1703.08603",
"language": "en",
"page_idx": 4
}
|
||
1512.00570-en-2
|
1512.00570-en-2.jpg
|
3
Attribute2Image: Conditional Image Generation from Visual Attributes
is unknown, we propose a general optimization-based approach for posterior
inference using image generation models and latent priors (Section 4).
We evaluate the proposed model on two datasets, the Labeled Faces in the Wild (LFW) dataset [10] and the Caltech-UCSD Birds-200-2011 (CUB) dataset [37]. In the LFW dataset, the attributes are 73-dimensional vectors describing age, gender, expressions, hair and many others [18]. In the CUB dataset, the 312-dimensional binary attribute vectors are converted from descriptions about bird parts and colors. We organize our experiments in the following two tasks. First, we demonstrate the quality of attribute-conditioned image generation with comparisons to nearest-neighbor search, and analyze the disentangling performance of latent space and corresponding foreground-background layers Second, we perform image reconstruction and completion on a set of novel test images by posterior inference with quantitative evaluation. Results from those experiments show the superior performance of the proposed model over previous art. The contributions of this paper are summarized as follows:
- We propose a novel problem of conditioned image generation from visual
attributes.
- We tackle this problem by learning conditional variational auto-encoders
and propose a novel layered foreground-background generative model that
significantly improves the generation quality of complex images.
-We propose a general optimization-based method for posterior inference on
novel images and use it to evaluate generative models in the context of image
reconstruction and completion.
## 2 Related Work
$Image\textit{ generation. In terms of generating realistic and novel images, there are}$ several recent work [4,9,17,8,3,25] that are relevant to ours. Dosovitskiy et al. [4] proposed to generate 3D chairs given graphics code using deep convolutional neural networks, and Kulkarni et al. [17] used variational auto-encoders [15] to model the rendering process of 3D objects. Both of these models [17,4] assume the existence of a graphics engine during training, from which they have 1) virtually infinite amount of training data and/or 2) pairs of rendered images that differ only in one factor of variation. Therefore, they are not directly applicable to natural image generation. While both work [17,4] studied generation of rendered images from complete description (e.g., object identity, view-point, color) trained from synthetic images (via graphics engine), generation of images from an incomplete description (e.g., class labels, visual attributes) is still under-explored. In fact, image generation from incomplete description is a more challenging task and the one-to-one mapping formulation of [4] is inherently limited. Gregor et al. [9] developed recurrent variational auto-encoders with spatial attention mechanism that allows iterative image generation by patches. This elegant algorithm mimics the process of human drawing but at the same time faces challenges when scaling up to large complex images. Recently, generative adversarial networks (GANs) [8,7,3,25] have been developed for image generation
|
[
[
0.138916015625,
0.017059326171875,
0.83837890625,
0.017059326171875,
0.83837890625,
0.0308685302734375,
0.138916015625,
0.0308685302734375
],
[
0.90087890625,
0.0162506103515625,
0.9228515625,
0.0162506103515625,
0.9228515625,
0.03167724609375,
0.90087890625,
0.03167724609375
],
[
0.0821533203125,
0.059295654296875,
0.91552734375,
0.059295654296875,
0.91552734375,
0.07391357421875,
0.0821533203125,
0.07391357421875
],
[
0.0784912109375,
0.0780029296875,
0.81982421875,
0.0780029296875,
0.81982421875,
0.09503173828125,
0.0784912109375,
0.09503173828125
],
[
0.11712646484375,
0.0966796875,
0.91650390625,
0.0966796875,
0.91650390625,
0.11456298828125,
0.11712646484375,
0.11456298828125
],
[
0.0809326171875,
0.1153564453125,
0.91552734375,
0.1153564453125,
0.91552734375,
0.1331787109375,
0.0809326171875,
0.1331787109375
],
[
0.0821533203125,
0.1373291015625,
0.9130859375,
0.1373291015625,
0.9130859375,
0.15185546875,
0.0821533203125,
0.15185546875
],
[
0.0784912109375,
0.156005859375,
0.91552734375,
0.156005859375,
0.91552734375,
0.173828125,
0.0784912109375,
0.173828125
],
[
0.0809326171875,
0.1746826171875,
0.91552734375,
0.1746826171875,
0.91552734375,
0.1893310546875,
0.0809326171875,
0.1893310546875
],
[
0.0821533203125,
0.194091796875,
0.91650390625,
0.194091796875,
0.91650390625,
0.208740234375,
0.0821533203125,
0.208740234375
],
[
0.0772705078125,
0.2120361328125,
0.91552734375,
0.212890625,
0.91552734375,
0.230712890625,
0.0772705078125,
0.2298583984375
],
[
0.0809326171875,
0.2340087890625,
0.91650390625,
0.2340087890625,
0.91650390625,
0.24853515625,
0.0809326171875,
0.24853515625
],
[
0.0821533203125,
0.25341796875,
0.9130859375,
0.25341796875,
0.9130859375,
0.267333984375,
0.0821533203125,
0.267333984375
],
[
0.0809326171875,
0.272216796875,
0.91650390625,
0.272216796875,
0.91650390625,
0.286865234375,
0.0809326171875,
0.286865234375
],
[
0.0809326171875,
0.29248046875,
0.91650390625,
0.29248046875,
0.91650390625,
0.30712890625,
0.0809326171875,
0.30712890625
],
[
0.0809326171875,
0.31201171875,
0.91796875,
0.31201171875,
0.91796875,
0.32666015625,
0.0809326171875,
0.32666015625
],
[
0.0809326171875,
0.332275390625,
0.7509765625,
0.332275390625,
0.7509765625,
0.346923828125,
0.0809326171875,
0.346923828125
],
[
0.09417724609375,
0.359130859375,
0.91796875,
0.359130859375,
0.91796875,
0.376953125,
0.09417724609375,
0.376953125
],
[
0.11834716796875,
0.378662109375,
0.23193359375,
0.378662109375,
0.23193359375,
0.39306640625,
0.11834716796875,
0.39306640625
],
[
0.09906005859375,
0.396484375,
0.91650390625,
0.396484375,
0.91650390625,
0.414306640625,
0.09906005859375,
0.414306640625
],
[
0.11956787109375,
0.41845703125,
0.91552734375,
0.41845703125,
0.91552734375,
0.432861328125,
0.11956787109375,
0.432861328125
],
[
0.12078857421875,
0.437744140625,
0.794921875,
0.437744140625,
0.794921875,
0.452392578125,
0.12078857421875,
0.452392578125
],
[
0.113525390625,
0.455810546875,
0.91650390625,
0.454833984375,
0.91650390625,
0.472900390625,
0.113525390625,
0.4736328125
],
[
0.11712646484375,
0.4736328125,
0.91650390625,
0.474365234375,
0.91650390625,
0.4921875,
0.11712646484375,
0.491455078125
],
[
0.11956787109375,
0.49560546875,
0.44921875,
0.49560546875,
0.44921875,
0.51025390625,
0.11956787109375,
0.51025390625
],
[
0.0821533203125,
0.54443359375,
0.1026611328125,
0.54443359375,
0.1026611328125,
0.55810546875,
0.0821533203125,
0.55810546875
],
[
0.11474609375,
0.54248046875,
0.327392578125,
0.54248046875,
0.327392578125,
0.5595703125,
0.11474609375,
0.5595703125
],
[
0.0821533203125,
0.58154296875,
0.91552734375,
0.58154296875,
0.91552734375,
0.59619140625,
0.0821533203125,
0.59619140625
],
[
0.0809326171875,
0.60205078125,
0.9130859375,
0.60205078125,
0.9130859375,
0.61669921875,
0.0809326171875,
0.61669921875
],
[
0.0821533203125,
0.62158203125,
0.91552734375,
0.62158203125,
0.91552734375,
0.63623046875,
0.0821533203125,
0.63623046875
],
[
0.0784912109375,
0.63916015625,
0.91650390625,
0.63916015625,
0.91650390625,
0.6572265625,
0.0784912109375,
0.6572265625
],
[
0.0784912109375,
0.658203125,
0.91650390625,
0.658203125,
0.91650390625,
0.67578125,
0.0784912109375,
0.67578125
],
[
0.0809326171875,
0.677734375,
0.91650390625,
0.677734375,
0.91650390625,
0.6953125,
0.0809326171875,
0.6953125
],
[
0.0821533203125,
0.69873046875,
0.91650390625,
0.69873046875,
0.91650390625,
0.71337890625,
0.0821533203125,
0.71337890625
],
[
0.0809326171875,
0.71728515625,
0.91162109375,
0.71728515625,
0.91162109375,
0.73193359375,
0.0809326171875,
0.73193359375
],
[
0.0784912109375,
0.73583984375,
0.92041015625,
0.7353515625,
0.92041015625,
0.7529296875,
0.0784912109375,
0.75390625
],
[
0.0784912109375,
0.75537109375,
0.91650390625,
0.75537109375,
0.91650390625,
0.7734375,
0.0784912109375,
0.7734375
],
[
0.0809326171875,
0.7763671875,
0.9130859375,
0.7763671875,
0.9130859375,
0.791015625,
0.0809326171875,
0.791015625
],
[
0.0784912109375,
0.79443359375,
0.91796875,
0.79443359375,
0.91796875,
0.8125,
0.0784912109375,
0.8125
],
[
0.0772705078125,
0.81298828125,
0.92041015625,
0.81396484375,
0.92041015625,
0.83203125,
0.0772705078125,
0.8310546875
],
[
0.0821533203125,
0.8349609375,
0.91064453125,
0.8349609375,
0.91064453125,
0.849609375,
0.0821533203125,
0.849609375
],
[
0.0772705078125,
0.85302734375,
0.91650390625,
0.8515625,
0.91650390625,
0.869140625,
0.0772705078125,
0.8701171875
],
[
0.0809326171875,
0.873046875,
0.9130859375,
0.873046875,
0.9130859375,
0.88720703125,
0.0809326171875,
0.88720703125
],
[
0.0784912109375,
0.89111328125,
0.91650390625,
0.89013671875,
0.91650390625,
0.908203125,
0.0784912109375,
0.9091796875
],
[
0.0809326171875,
0.91162109375,
0.91650390625,
0.91162109375,
0.91650390625,
0.9287109375,
0.0809326171875,
0.9287109375
],
[
0.07611083984375,
0.93017578125,
0.91796875,
0.93115234375,
0.91796875,
0.94873046875,
0.07611083984375,
0.947265625
]
] |
[
"Attribute2Image:ConditionalImageGenerationfromVisualAttributes",
"3",
"is unknown,we propose a general optimization-based approachfor posterior",
"inference using image generation models and latent priors (Section 4).",
"We evaluate the proposed model on two datasets, the Labeled Faces in",
"the Wild (LFW) dataset [10] and the Caltech-UCSD Birds-200-2011 (CUB)",
"dataset37].In the LFW dataset,the attributes are 73-dimensional vectors de-",
"scribing age, gender, expressions, hair and many others [18]. In the CUB dataset,",
"the 312-dimensional binary attribute vectors are converted from descriptions",
"about bird parts and colors. We organize our experiments in the following two",
"tasks. First, we demonstrate the quality of attribute-conditioned image genera-",
"tion with comparisons to nearest-neighbor search, and analyze the disentangling",
"performance of latent space and correspondingforeground-background layers",
"Second, we perform image reconstruction and completion on a set of novel test",
"images by posterior inference with quantitative evaluation. Results from those",
"experiments showthe superior performance of theproposed model over previous",
"art. The contributions of this paper are summarized as follows:",
"- We propose a novel problem of conditioned image generation from visual",
"attributes.",
"-We tackle this problem by learning conditional variational auto-encoders",
"and propose anovel layered foreground-background generative modelthat",
"significantlyimprovesthegeneration quality of complex images.",
"We propose a general optimization-based method for posterior inference on",
"novel images and use it to evaluate generative models in the context of image",
"reconstructionandcompletion.",
"2",
"RelatedWork",
"Image qeneration.In terms of generating realistic and novel images, there are",
"several recent work[4.9,17.8,3.25that are relevant to ours. Dosovitskiy et al.[4",
"proposed to generate 3D chairs given graphics code using deep convolutional",
"neural networks, and Kulkarni et al. [17] used variational auto-encoders [15] to",
"model the rendering process of 3D objects. Both of these models [17,4] assume",
"the existence of a graphics engine during training, from which they have 1)",
"virtually infinite amount of training data and/or 2) pairs of rendered images",
"that differ only in one factor of variation.Therefore,they are not directly ap",
"plicable to natural image generation. While both work [17,4] studied generation",
"of rendered images from complete description (e.g., object identity, view-point,",
"color)trained from synthetic images (via graphics engine),generation of im-",
"ages from an incomplete description (e.g., class labels, visual attributes) is still",
"under-explored. In fact, image generation from incomplete description is a more",
"challenging taskand the one-to-onemapping formulationof4lis inherentlylim",
"ited.Gregoretal.[9]developed recurrentvariational auto-encoderswith spatial",
"attentionmechanismthat allows iterative imagegeneration by patches.Thisele",
"gant algorithm mimics the process of human drawing but at the same time faces",
"challenges when scaling up to large complex images. Recently, generative adver-",
"sarial networks(GANs)[8,7,3,25] have been developedfor image generation."
] |
{
"arxiv_id": "1512.00570",
"language": "en",
"page_idx": 2
}
|
||
1512.00570-en-3
|
1512.00570-en-3.jpg
|
4 Xinchen Yan, Jimei Yang, Kihyuk Sohn and Honglak Lee
In the GAN, two models are trained to against each other: a generative model aims to capture the data distribution, while a discriminative model attempts to distinguish between generated samples and training data. The GAN training is based on a min-max objective, which is known to be challenging to optimize.
$Layered~modeling~of~images.~$Layered models or 2.1D representations of images have been studied in the context of moving or still object segmentation [38,23,39,41,11]. The layered structure is introduced into generative image modeling [20,35]. Tang et al. [35] modeled the occluded images with gated restricted Boltzmann machines and achieved good inpainting and denoising results on well cropped face images. Le Roux et al. [20] explicitly modeled the occlusion layer in a masked restricted Boltzmann machine for separating foreground and background and demonstrated promising results on small patches. Though similar to our proposed gating in the form, these models face challenges when applied to model large natural images due to its difficulty in learning hierarchical representation based on restricted Boltzmann machine.
$Multimodal\textit{ Learning. Generative models of image and text have been stud- }$ ied in multimodal learning to model joint distribution of multiple data modalities [22,33,31]. For example, Srivastava and Salakhutdinov [33] developed a multimodal deep Boltzmann machine that models joint distribution of image and text (e.g., image tag). Sohn et al. [31] proposed improved shared representation learning of multimodal data through bi-directional conditional prediction by deriving a conditional prediction model of one data modality given the other and vice versa. Both of these works focused more on shared representation learning using hand-crafted low-level image features and therefore have limited applications such as conditional image or text retrieval than actual generation of images.
3 Attribute-conditioned Generative Modeling of Images
In this section, we describe our proposed method for attribute-conditioned generative modeling of images. We frst describe a conditional variational autoencoder, followed by the formulation of layered generative model and its variational learning.
3.1 Base Model: Conditional Variational Auto-Encoder (CVAE)
Given the attribute $y\in \mathbb{R} ^{N_{y}}$ and latent variable $z\in \mathbb{R} ^{N_{z}}$,our goal is to build a model $p_{\theta }( x| y, z)$ that generates realistic image $x\in \mathbb{R} ^{N_{x}}$ conditioned on $y$ and $z.$ Here, we refer $p_{\theta }$ a generator ( or generation mo
Conditioned image generation is simply a two-step process in the following:
1. Randomly sample latent variable $z$ from prior distribution $p(z);$
2. Given $y$ and $z$ as conditioning variable, generate image $x$ from $p_\theta(x|y,z).$
|
[
[
0.0772705078125,
0.0162506103515625,
0.1014404296875,
0.0162506103515625,
0.1014404296875,
0.0308685302734375,
0.0772705078125,
0.0308685302734375
],
[
0.158203125,
0.0162506103515625,
0.72119140625,
0.0162506103515625,
0.72119140625,
0.0308685302734375,
0.158203125,
0.0308685302734375
],
[
0.0784912109375,
0.056060791015625,
0.91796875,
0.056854248046875,
0.91796875,
0.07470703125,
0.0784912109375,
0.07391357421875
],
[
0.0821533203125,
0.07879638671875,
0.91796875,
0.07879638671875,
0.91796875,
0.09344482421875,
0.0821533203125,
0.09344482421875
],
[
0.0809326171875,
0.0966796875,
0.92041015625,
0.0966796875,
0.92041015625,
0.11456298828125,
0.0809326171875,
0.11456298828125
],
[
0.0772705078125,
0.1153564453125,
0.89501953125,
0.11614990234375,
0.89501953125,
0.134033203125,
0.0772705078125,
0.1331787109375
],
[
0.08331298828125,
0.15673828125,
0.9130859375,
0.15673828125,
0.9130859375,
0.17138671875,
0.08331298828125,
0.17138671875
],
[
0.0821533203125,
0.17626953125,
0.91552734375,
0.17626953125,
0.91552734375,
0.19091796875,
0.0821533203125,
0.19091796875
],
[
0.08331298828125,
0.196533203125,
0.91552734375,
0.196533203125,
0.91552734375,
0.211181640625,
0.08331298828125,
0.211181640625
],
[
0.08331298828125,
0.216064453125,
0.91552734375,
0.216064453125,
0.91552734375,
0.230712890625,
0.08331298828125,
0.230712890625
],
[
0.0821533203125,
0.2347412109375,
0.91552734375,
0.2347412109375,
0.91552734375,
0.2493896484375,
0.0821533203125,
0.2493896484375
],
[
0.0809326171875,
0.255126953125,
0.91650390625,
0.255126953125,
0.91650390625,
0.269775390625,
0.0809326171875,
0.269775390625
],
[
0.0809326171875,
0.273681640625,
0.9130859375,
0.273681640625,
0.9130859375,
0.288330078125,
0.0809326171875,
0.288330078125
],
[
0.0809326171875,
0.291748046875,
0.91796875,
0.2900390625,
0.91796875,
0.307861328125,
0.0809326171875,
0.30859375
],
[
0.0809326171875,
0.31103515625,
0.91796875,
0.31103515625,
0.91796875,
0.3291015625,
0.0809326171875,
0.3291015625
],
[
0.0809326171875,
0.332275390625,
0.91552734375,
0.332275390625,
0.91552734375,
0.346923828125,
0.0809326171875,
0.346923828125
],
[
0.0809326171875,
0.350830078125,
0.5771484375,
0.351806640625,
0.5771484375,
0.366455078125,
0.0809326171875,
0.365478515625
],
[
0.0821533203125,
0.38916015625,
0.91552734375,
0.38818359375,
0.91552734375,
0.40625,
0.0821533203125,
0.406982421875
],
[
0.0821533203125,
0.4111328125,
0.9130859375,
0.4111328125,
0.9130859375,
0.42578125,
0.0821533203125,
0.42578125
],
[
0.0809326171875,
0.428955078125,
0.91552734375,
0.428955078125,
0.91552734375,
0.44677734375,
0.0809326171875,
0.44677734375
],
[
0.0821533203125,
0.44921875,
0.91650390625,
0.44921875,
0.91650390625,
0.4638671875,
0.0821533203125,
0.4638671875
],
[
0.0809326171875,
0.469482421875,
0.91552734375,
0.469482421875,
0.91552734375,
0.484130859375,
0.0809326171875,
0.484130859375
],
[
0.0809326171875,
0.489013671875,
0.9130859375,
0.489013671875,
0.9130859375,
0.50341796875,
0.0809326171875,
0.50341796875
],
[
0.0784912109375,
0.5068359375,
0.91796875,
0.5068359375,
0.91796875,
0.52490234375,
0.0784912109375,
0.52490234375
],
[
0.0784912109375,
0.525390625,
0.91650390625,
0.5263671875,
0.91650390625,
0.54443359375,
0.0784912109375,
0.54345703125
],
[
0.0809326171875,
0.54736328125,
0.9130859375,
0.54736328125,
0.9130859375,
0.56201171875,
0.0809326171875,
0.56201171875
],
[
0.0809326171875,
0.56640625,
0.91796875,
0.56640625,
0.91796875,
0.5810546875,
0.0809326171875,
0.5810546875
],
[
0.0784912109375,
0.5859375,
0.158203125,
0.5859375,
0.158203125,
0.60009765625,
0.0784912109375,
0.60009765625
],
[
0.0772705078125,
0.6337890625,
0.89599609375,
0.63671875,
0.89599609375,
0.65478515625,
0.0772705078125,
0.6513671875
],
[
0.0772705078125,
0.67431640625,
0.91552734375,
0.67578125,
0.91552734375,
0.6962890625,
0.0772705078125,
0.6953125
],
[
0.0809326171875,
0.69873046875,
0.9130859375,
0.69873046875,
0.9130859375,
0.71337890625,
0.0809326171875,
0.71337890625
],
[
0.0772705078125,
0.71484375,
0.91552734375,
0.71630859375,
0.91552734375,
0.734375,
0.0772705078125,
0.73291015625
],
[
0.0772705078125,
0.734375,
0.2403564453125,
0.73583984375,
0.2403564453125,
0.75390625,
0.0772705078125,
0.75146484375
],
[
0.0784912109375,
0.78466796875,
0.141357421875,
0.78466796875,
0.141357421875,
0.80078125,
0.0784912109375,
0.80078125
],
[
0.138916015625,
0.783203125,
0.873046875,
0.78466796875,
0.873046875,
0.802734375,
0.138916015625,
0.80078125
],
[
0.0772705078125,
0.8154296875,
0.91796875,
0.81640625,
0.91796875,
0.8369140625,
0.0772705078125,
0.8359375
],
[
0.07611083984375,
0.8359375,
0.91796875,
0.8349609375,
0.91796875,
0.85693359375,
0.07611083984375,
0.85791015625
],
[
0.0784912109375,
0.85791015625,
0.91552734375,
0.85791015625,
0.91552734375,
0.87548828125,
0.0784912109375,
0.87548828125
],
[
0.0772705078125,
0.875,
0.87939453125,
0.87646484375,
0.87939453125,
0.8974609375,
0.0772705078125,
0.89501953125
],
[
0.090576171875,
0.90966796875,
0.794921875,
0.90966796875,
0.794921875,
0.927734375,
0.090576171875,
0.927734375
],
[
0.090576171875,
0.931640625,
0.888671875,
0.931640625,
0.888671875,
0.9462890625,
0.090576171875,
0.9462890625
]
] |
[
"4",
"Xinchen Yan, Jimei Yang,Kihyuk Sohn and Honglak Lee",
"In the GAN, two models are trained to against each other: a generative model",
"aims to capture the data distribution, while a discriminative model attempts to",
"distinguish between generated samples and training data. The GAN training is",
"based on a min-max objective, which is known to be challenging to optimize.",
"Layered modeling of imaqes.Layered modelsor 2.iDrepresentations of im-",
"ages havebeen studied in the context of moving or still object segmentation",
"[38,23,39,41,11].Thelayeredstructureisintroduced intogenerativeimagemod-",
"eling[20,35].Tang etal.[35]modeledtheoccluded imageswithgatedrestricted",
"Boltzmann machines and achieved good inpainting and denoising results on well",
"cropped face images.LeRoux et al.[2o]explicitlymodeled the occlusionlayer",
"in a masked restricted Boltzmann machine for separating foreground and back-",
"ground and demonstrated promising results on small patches. Though similar to",
"our proposed gating in the form, these models face challenges when applied to",
"modellargenatural images due to its difficulty in learning hierarchical represen-",
"tation based on restricted Boltzmannmachine.",
"Multimodal Learning. Generative models of image and text have been stud-",
"ied inmultimodal learning tomodel joint distribution of multiple datamodali",
"ties [22,33,31]. For example, Srivastava and Salakhutdinov [33] developed a mul-",
"timodal deep Boltzmann machine that models joint distribution of image and",
"text (e.g.,imagetag).Sohn et al.[31lproposed improved shared representation",
"learningofmultimodaldatathroughbi-directional conditionalpredictionbyde",
"riving a conditional prediction model of one data modality given the other and",
"vice versa. Both of these works focused more on shared representation learning",
"usinghand-craftedlow-level imagefeatures and therefore have limited appli-",
"cations such as conditional image or text retrieval than actual generation of",
"images.",
"3Attribute-conditioned GenerativeModeling of Images",
"In this section, we describe our proposed method for attribute-conditioned gen-",
"erative modeling of images. We first describe a conditional variational auto",
"encoder, followed by the formulation of layered generative model and its varia-",
"tional learning.",
"3.1",
"Base Model: Conditional Variational Auto-Encoder (CVAE)",
"Given the attribute y e IRNv and latent variable e IRN-, our goal is to build",
"a model pe(αly,z) that generates realistic image r e RNa conditioned on y",
"and z. Here, we refer pg a generator (or generation model), parametrized by .",
"Conditioned image generation is simply a two-step process in the following:",
"1. Randomly sample latent variable z from prior distribution p(z);",
"2.Givenyandz as conditioning variable,generateimagefrom pe(r y,z)"
] |
{
"arxiv_id": "1512.00570",
"language": "en",
"page_idx": 3
}
|
||
1512.00570-en-4
|
1512.00570-en-4.jpg
|
5
Attribute2Image: Conditional Image Generation from Visual Attributes
Here, the purpose of learning is to find the best parameter $\theta$ that maximizes the log-likelihood $\log p_{\theta}(x|y).$ As proposed in [28,15], variational auto-encoders try to maximize the variational lower bound of the log-likelihood $\log p_{\theta}(x|y).$ Specifically, an auxiliary distribution $q_\phi(z|x,y)$ is introduced to approximate the true posterior $p_\theta(z|x,y).$ We refer the base model a conditional variational auto-encoder (CVAE) with the conditional log-likelihood
$$\log p_\theta(x|y)=KL(q_\phi(z|x,y)||p_\theta(z|x,y))+\mathcal{L}_{\mathrm{CVAE}}(x,y;\theta,\phi),$$
where the variational lower bound
$$\mathcal{L}_{\mathrm{CVAE}}(x,y;\theta,\phi)=-KL(q_{\phi}(z|x,y)||p_{\theta}(z))+\mathbb{E}_{q_{\phi}(z|x,y)}\big[\log p_{\theta}(x|y,z)\big]$$
is maximized for learning the model parameters.
Here, the prior $p_\theta(z)$ is assumed to follow isotropic multivariate Gaussian distribution, while two conditional distributions $p_\theta(x|y,z)$ and $q_\phi(z|x,y)$ are multivariate Gaussian distributions whose mean and covariance are parametrized by $\mathcal{N}\left(\mu_{\theta}(z,y),diag(\sigma_{\theta}^{2}(z,y))\right)$ and $\mathcal{N}\left(\mu_{\phi}(x,y),diag(\sigma_{\phi}^{2}(x,y))\right)$, respectively. We refer the auxiliary proposal distribution $q_\phi(z|x,y)$ a recognition model and the conditional data distribution $p_\theta(x|y,z)$ a generation model.
The first term $KL(q_\phi(z|x,y)||p_\theta(z))$ is a regularization term that reduces the gap between the prior $p(z)$ and the proposal distribution $q_\phi(z|x,y)$, while the second term $\log p_\theta(x|y,z)$ is the log likelihood of samples. In practice, we usually take as a deterministic generation function the mean $x=\mu_\theta(z,y)$ of conditional distribution $p_\theta(x|z,y)$ given $z$ and $y$, so it is convenient to assume the standard deviation function $\sigma_\theta(z,y)$ is a constant shared by all the pixels as the latent factors capture all the data variations. We will keep this assumption for the rest of the paper if not particularly mentioned. Thus, we can rewrite the second term in the variational lower bound as reconstruction loss $L(\cdot,\cdot)$ (e.g., $\ell_2$ loss):
$$\mathcal{L}_{\mathrm{CVAE}}=-\:KL(q_{\phi}(z|x,y)||p_{\theta}(z))-\mathbb{E}_{q_{\phi}(z|x,y)}L(\mu_{\theta}(y,z),x)$$
(2)
Note that the discriminator of GANs [8] can be used as the loss function $L(\cdot,\cdot)$ as well, especially when $\ell _2$ ( or $\ell _1)$ reconstruction loss may not capture the true image similarities. We leave it for future study.
3.2 Disentangling CVAE with a Layered Representation
An image $x$ can be interpreted as a composite of a foreground layer (or a foreground image $x_F)$ and a background layer (or a background image $x_B)$ via a ${\mathrm{matting~equation~[24]:}}$
$$x=x_F\odot(1-g)+x_B\odot g,$$
(3)
where $\odot$denotes the element-wise product. $g\in[0,1]^{N_{x}}$ is an occlusion layer or a gating function that determines the visibility of background pixels while $1-g$ defines the visibility of foreground pixels. However, the model based on
|
[
[
0.138916015625,
0.0162506103515625,
0.8369140625,
0.0162506103515625,
0.8369140625,
0.030059814453125,
0.138916015625,
0.030059814453125
],
[
0.90087890625,
0.0162506103515625,
0.9228515625,
0.0162506103515625,
0.9228515625,
0.03167724609375,
0.90087890625,
0.03167724609375
],
[
0.11712646484375,
0.056854248046875,
0.91650390625,
0.056854248046875,
0.91650390625,
0.07470703125,
0.11712646484375,
0.07470703125
],
[
0.0821533203125,
0.0780029296875,
0.91650390625,
0.0780029296875,
0.91650390625,
0.09503173828125,
0.0821533203125,
0.09503173828125
],
[
0.0784912109375,
0.09503173828125,
0.91552734375,
0.0966796875,
0.91552734375,
0.11456298828125,
0.0784912109375,
0.1129150390625
],
[
0.0809326171875,
0.11614990234375,
0.91650390625,
0.11614990234375,
0.91650390625,
0.134033203125,
0.0809326171875,
0.134033203125
],
[
0.0809326171875,
0.1356201171875,
0.91796875,
0.1356201171875,
0.91796875,
0.153564453125,
0.0809326171875,
0.153564453125
],
[
0.0784912109375,
0.1551513671875,
0.68359375,
0.1551513671875,
0.68359375,
0.1722412109375,
0.0784912109375,
0.1722412109375
],
[
0.1895751953125,
0.1893310546875,
0.80419921875,
0.1893310546875,
0.80419921875,
0.2071533203125,
0.1895751953125,
0.2071533203125
],
[
0.0821533203125,
0.2242431640625,
0.442138671875,
0.2242431640625,
0.442138671875,
0.23876953125,
0.0821533203125,
0.23876953125
],
[
0.11956787109375,
0.256591796875,
0.85498046875,
0.259033203125,
0.85498046875,
0.279541015625,
0.11956787109375,
0.27783203125
],
[
0.87548828125,
0.260009765625,
0.91552734375,
0.260009765625,
0.91552734375,
0.277099609375,
0.87548828125,
0.277099609375
],
[
0.0809326171875,
0.294189453125,
0.5927734375,
0.294189453125,
0.5927734375,
0.30859375,
0.0809326171875,
0.30859375
],
[
0.11474609375,
0.312744140625,
0.91552734375,
0.31103515625,
0.91552734375,
0.3291015625,
0.11474609375,
0.33056640625
],
[
0.0772705078125,
0.33056640625,
0.91552734375,
0.332275390625,
0.91552734375,
0.350830078125,
0.0772705078125,
0.349365234375
],
[
0.08331298828125,
0.353271484375,
0.91552734375,
0.353271484375,
0.91552734375,
0.367919921875,
0.08331298828125,
0.367919921875
],
[
0.0809326171875,
0.373779296875,
0.91796875,
0.373779296875,
0.91796875,
0.39404296875,
0.0809326171875,
0.39404296875
],
[
0.0809326171875,
0.399658203125,
0.91650390625,
0.399658203125,
0.91650390625,
0.41748046875,
0.0809326171875,
0.41748046875
],
[
0.0809326171875,
0.419189453125,
0.705078125,
0.419189453125,
0.705078125,
0.43701171875,
0.0809326171875,
0.43701171875
],
[
0.11474609375,
0.43701171875,
0.91650390625,
0.436279296875,
0.91650390625,
0.45654296875,
0.11474609375,
0.458251953125
],
[
0.0772705078125,
0.455810546875,
0.91796875,
0.45654296875,
0.91796875,
0.4775390625,
0.0772705078125,
0.476806640625
],
[
0.07611083984375,
0.47607421875,
0.91650390625,
0.474365234375,
0.91650390625,
0.49560546875,
0.07611083984375,
0.496337890625
],
[
0.0809326171875,
0.496337890625,
0.91650390625,
0.496337890625,
0.91650390625,
0.51416015625,
0.0809326171875,
0.51416015625
],
[
0.0821533203125,
0.517578125,
0.91552734375,
0.517578125,
0.91552734375,
0.5322265625,
0.0821533203125,
0.5322265625
],
[
0.0784912109375,
0.5361328125,
0.91650390625,
0.5361328125,
0.91650390625,
0.55419921875,
0.0784912109375,
0.55419921875
],
[
0.0772705078125,
0.5546875,
0.91650390625,
0.55419921875,
0.91650390625,
0.5712890625,
0.0772705078125,
0.57275390625
],
[
0.0809326171875,
0.576171875,
0.91552734375,
0.576171875,
0.91552734375,
0.5908203125,
0.0809326171875,
0.5908203125
],
[
0.0772705078125,
0.5947265625,
0.8515625,
0.5947265625,
0.8515625,
0.61181640625,
0.0772705078125,
0.61181640625
],
[
0.1968994140625,
0.62939453125,
0.8017578125,
0.62939453125,
0.8017578125,
0.6474609375,
0.1968994140625,
0.6474609375
],
[
0.8818359375,
0.62646484375,
0.92138671875,
0.62646484375,
0.92138671875,
0.6484375,
0.8818359375,
0.6484375
],
[
0.0772705078125,
0.65869140625,
0.91796875,
0.6611328125,
0.91796875,
0.681640625,
0.0772705078125,
0.68017578125
],
[
0.0809326171875,
0.68408203125,
0.91552734375,
0.68408203125,
0.91552734375,
0.69873046875,
0.0809326171875,
0.69873046875
],
[
0.0809326171875,
0.70361328125,
0.57470703125,
0.70361328125,
0.57470703125,
0.71826171875,
0.0809326171875,
0.71826171875
],
[
0.0809326171875,
0.75048828125,
0.1473388671875,
0.75048828125,
0.1473388671875,
0.76513671875,
0.0809326171875,
0.76513671875
],
[
0.138916015625,
0.75048828125,
0.775390625,
0.75048828125,
0.775390625,
0.76513671875,
0.138916015625,
0.76513671875
],
[
0.08331298828125,
0.783203125,
0.9130859375,
0.783203125,
0.9130859375,
0.7978515625,
0.08331298828125,
0.7978515625
],
[
0.0809326171875,
0.802734375,
0.91650390625,
0.802734375,
0.91650390625,
0.8173828125,
0.0809326171875,
0.8173828125
],
[
0.0784912109375,
0.8212890625,
0.320068359375,
0.8212890625,
0.320068359375,
0.83935546875,
0.0784912109375,
0.83935546875
],
[
0.3515625,
0.8564453125,
0.6435546875,
0.8564453125,
0.6435546875,
0.873046875,
0.3515625,
0.873046875
],
[
0.88427734375,
0.85400390625,
0.92138671875,
0.85400390625,
0.92138671875,
0.87548828125,
0.88427734375,
0.87548828125
],
[
0.0809326171875,
0.89013671875,
0.91552734375,
0.89013671875,
0.91552734375,
0.908203125,
0.0809326171875,
0.908203125
],
[
0.0772705078125,
0.91162109375,
0.91650390625,
0.9091796875,
0.91650390625,
0.9267578125,
0.0772705078125,
0.9287109375
],
[
0.0821533203125,
0.931640625,
0.91650390625,
0.931640625,
0.91650390625,
0.9462890625,
0.0821533203125,
0.9462890625
]
] |
[
"Attribute2Image:Conditional Image Generation fromVisual Attributes",
"5",
"Here, the purpose of learning is to find the best parameter that maximizes",
"the log-likelihood log pe(r|y). As proposed in [28,15], variational auto-encoders",
"try to maximize the variational lower bound of the log-likelihood log pe(aly).",
"Specifically, an auxiliary distribution q(zc, y) is introduced to approximate",
"the true posterior pe(zc, y). We refer the base model a conditional variational",
"auto-encoder (CVAE) with the conditional log-likelihood",
"logpe(ry) = KL(q(z, y)llpe(zα, y)) + LcVAE(,y; 0, Φ)",
"wherethevariational lower bound",
"LcVAE(,y; 0, ) = -KL(q(z, y)|le(z) +E(2,y)[log Pe(ry,z)]",
"(1)",
"ismaximized for learningthe model parameters.",
"Here, the prior pe(z) is assumed to follow isotropic multivariate Gaussian dis-",
"tribution, while two conditional distributions pe(ry, z) and q(zc,y) aremulti-",
"variate Gaussian distributions whose mean and covariance are parametrized by",
"N (μe(z, y), diag(o(z, y)) and N (μs(c, y),diag(o2(r, y))), respectively. We",
"refer the auxiliary proposal distribution qs(zr, ) a recognition model and the",
"conditional data distribution pe(rly, z) a generation model.",
"The first term KL(qs(zr, y)llpe(z)) is a regularization term that reduces the",
"gap between the prior p(z) and the proposal distribution q(zr, y), while the",
"second term log pe(rly, z) is the log likelihood of samples. In practice, we usually",
"take as a deterministic generation function the mean = μe(z, y) of conditional",
"distribution pe(r z,y) given z and y, so it is convenient to assume the standard",
"deviation function o(z, y) is a constant shared by all the pixels as the latent",
"factors capture all the data variations. We will keep this assumption for the rest",
"of the paperif not particularly mentioned.Thus,we can rewritethe second term",
"in the variational lower bound as reconstruction loss L(, ) (e.g., l2 loss):",
"LcVAE =-KL(q(z, y)lpe(z)) - Eg(zr,y)L(μe(y, z), r)",
"(2)",
"Note that the discriminator of GANs [8] can be used as the loss function L(, )",
"aswell,especiallywhenl2(orl)reconstructionlossmaynotcapturethetrue",
"image similarities.We leave it for future study",
"3.2",
"Disentangling CVAE with aLayeredRepresentation",
"An imagercan be interpretedas acompositeofaforeground layer(or afore",
"ground image F)and abackground layer(or abackground image rB)via a",
"matting equation [24]:",
"T=TFO(1-g)+aBOg,",
"(3)",
"where denotes the element-wise product. g E Jo,1jNa is an occlusion layer",
"or a gatingfunction that determinesthe visibilityof background pixels while",
"I -g defines the visibility of foreground pixels.However, the model based on"
] |
{
"arxiv_id": "1512.00570",
"language": "en",
"page_idx": 4
}
|
||
2112.06166-en-0
|
2112.06166-en-0.jpg
|
# Topic Detection and Tracking with Time-Aware Document Embeddings
Hang Jiang, Doug Beeferman, Weiquan Mao, Deb Roy
Massachusetts Institute of Technology, Stanford University
75 Amherst St, Cambridge, MA 02139
$\{$hjian42, dougb5, dkroy}@mit.edu, mwq@stanford.edu
Abstract
The time at which a message is communicated is a vital piece of metadata in many real-world natural language processing tasks such as Topic Detection and Tracking (TDT). TDT systems aim to cluster a corpus of news articles by event, and in that context, stories that describe the same event are likely to have been written at around the same time. Prior work on time modeling for TDT takes this into account, but does not well capture how time interacts with the semantic nature of the event. For example, stories about a tropical storm are likely to be written within a short time interval, while stories about a movie release may appear over weeks or months. In our work, we design a neural method that fuses temporal and textual information into a single representation of news documents for event detection. We fine-tune these time-aware document embeddings with a triplet loss architecture, integrate the model into downstream TDT systems, and evaluate the systems on two benchmark TDT data sets in English. In the retrospective setting, we apply clustering algorithms to the time-aware embeddings and show substantial improvements over baselines on the News2013 data set. In the online streaming setting, we add our document encoder to an existing state-of-the-art TDT pipeline and demonstrate that it can benefit the overall performance. We conduct ablation studies on the time representation and fusion algorithm strategies, showing that our proposed model outperforms alternative strategies. Finally, we probe the model to examine how it handles recurring events more effectively than previous TDT systems.
Keywords: Topic Detection and Tracking, Online Clustering, Document Embedding, Temporal Embedding
## 1. Introduction
TDT modelers use this metadata, at a minimum, to exploit the temporal locality inherent in news events. But beyond that, researchers tend to focus on improving text representations (Hu et al., 2017; Staykovski et al., 2019; Saravanakumar et al., 2021; Fan et al., 2021) instead of time representations. The time model is typically parameter-free or low capacity, and unconnected to the text. This overlooks important ways that the time may interact with the topic of the article to influence where it falls on the calendar and how long it lasts as an event. For example, stock reports happen daily, but corporate earnings releases may happen quarterly and be discussed over several days. Stories about a weather event are likely to be written over a short time period, while stories a movie release may be discussed over weeks or months.
In this paper, we address these gaps with a unified neural time-text model that fuses temporal and semantic embeddings to represent news documents. Our contributions are as follows:
Following emerging news stories is crucial to making real-time decisions on important political and public safety matters. Amid the COVID-19 pandemic, for instance, media platforms and government agencies need to identify the emergence of misinformation in time and take action to protect the safety of the public. As humans cannot read all of the articles produced by the news media, automatic clustering of news articles into real-world events is needed to make this work tractable. Computational tools for this task are useful for organizing not just news articles, but also scientific papers, microblogs, online reviews, forum messages, and social media posts.
Allan et al. (1998) first introduced the topic detection and tracking (TDT) framework to address these needs. In this paper, we focus on the event detection part of TDT, where events are defined as real-world news stories and we want to categorize news articles into these events. Researchers have proposed both topic modeling and clusteringbased methods for this task, in both retrospective and online streaming settings. In the retrospective setting, researchers process the news documents altogether, whereas in the online setting they are processed one by one as they appear in the stream. Most real-world documents have a timestamp, the time at which it was written, spoken, or shared.
·We propose a time-aware neural document embedding method that can be applied to topic detection and tracking and other NLP tasks.
·We build two TDT pipelines based on our time-aware model, for retrospective and online event clustering tasks respectively, achieving state-of-the-art performance in both settings for the corpus used as a benchmark by re-
|
[
[
0.175537109375,
0.097412109375,
0.82958984375,
0.097412109375,
0.82958984375,
0.11224365234375,
0.175537109375,
0.11224365234375
],
[
0.426513671875,
0.11517333984375,
0.57421875,
0.1181640625,
0.57421875,
0.135986328125,
0.426513671875,
0.133056640625
],
[
0.2435302734375,
0.15087890625,
0.75732421875,
0.152587890625,
0.75732421875,
0.16748046875,
0.2435302734375,
0.1650390625
],
[
0.28369140625,
0.16748046875,
0.71728515625,
0.16748046875,
0.71728515625,
0.1817626953125,
0.28369140625,
0.1817626953125
],
[
0.359375,
0.1817626953125,
0.64404296875,
0.1817626953125,
0.64404296875,
0.196533203125,
0.359375,
0.196533203125
],
[
0.298828125,
0.196533203125,
0.705078125,
0.196533203125,
0.705078125,
0.21142578125,
0.298828125,
0.21142578125
],
[
0.466796875,
0.2298583984375,
0.53662109375,
0.2298583984375,
0.53662109375,
0.2410888671875,
0.466796875,
0.2410888671875
],
[
0.118408203125,
0.24462890625,
0.88330078125,
0.24462890625,
0.88330078125,
0.25830078125,
0.118408203125,
0.25830078125
],
[
0.12176513671875,
0.2587890625,
0.88330078125,
0.2587890625,
0.88330078125,
0.26953125,
0.12176513671875,
0.26953125
],
[
0.12005615234375,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.281982421875,
0.12005615234375,
0.281982421875
],
[
0.118408203125,
0.283935546875,
0.884765625,
0.283935546875,
0.884765625,
0.297607421875,
0.118408203125,
0.297607421875
],
[
0.1158447265625,
0.295654296875,
0.884765625,
0.29638671875,
0.884765625,
0.31005859375,
0.1158447265625,
0.308837890625
],
[
0.0201568603515625,
0.302734375,
0.060455322265625,
0.302734375,
0.060455322265625,
0.438232421875,
0.0201568603515625,
0.438232421875
],
[
0.118408203125,
0.308837890625,
0.88330078125,
0.308837890625,
0.88330078125,
0.322509765625,
0.118408203125,
0.322509765625
],
[
0.118408203125,
0.322509765625,
0.88330078125,
0.322509765625,
0.88330078125,
0.336181640625,
0.118408203125,
0.336181640625
],
[
0.12005615234375,
0.33740234375,
0.88330078125,
0.33740234375,
0.88330078125,
0.347412109375,
0.12005615234375,
0.347412109375
],
[
0.118408203125,
0.3486328125,
0.884765625,
0.3486328125,
0.884765625,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.12005615234375,
0.3623046875,
0.8818359375,
0.3623046875,
0.8818359375,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.118408203125,
0.37353515625,
0.88330078125,
0.37353515625,
0.88330078125,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.12005615234375,
0.3896484375,
0.88330078125,
0.3896484375,
0.88330078125,
0.399658203125,
0.12005615234375,
0.399658203125
],
[
0.1175537109375,
0.399658203125,
0.884765625,
0.40087890625,
0.884765625,
0.4150390625,
0.1175537109375,
0.41455078125
],
[
0.12005615234375,
0.41455078125,
0.884765625,
0.41455078125,
0.884765625,
0.427490234375,
0.12005615234375,
0.427490234375
],
[
0.12005615234375,
0.4287109375,
0.411376953125,
0.4287109375,
0.411376953125,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.0218353271484375,
0.43701171875,
0.06463623046875,
0.43701171875,
0.06463623046875,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.118408203125,
0.45068359375,
0.8251953125,
0.451904296875,
0.8251953125,
0.466064453125,
0.118408203125,
0.465576171875
],
[
0.225830078125,
0.47802734375,
0.382080078125,
0.47802734375,
0.382080078125,
0.4921875,
0.225830078125,
0.4921875
],
[
0.5146484375,
0.478515625,
0.88330078125,
0.478515625,
0.88330078125,
0.4921875,
0.5146484375,
0.4921875
],
[
0.51318359375,
0.493408203125,
0.884765625,
0.493408203125,
0.884765625,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.118408203125,
0.50732421875,
0.490234375,
0.50830078125,
0.490234375,
0.5224609375,
0.118408203125,
0.521484375
],
[
0.5146484375,
0.5087890625,
0.8818359375,
0.5087890625,
0.8818359375,
0.51953125,
0.5146484375,
0.51953125
],
[
0.019317626953125,
0.51953125,
0.060455322265625,
0.51953125,
0.06378173828125,
0.73095703125,
0.0218353271484375,
0.73095703125
],
[
0.118408203125,
0.5224609375,
0.490234375,
0.5224609375,
0.490234375,
0.5361328125,
0.118408203125,
0.5361328125
],
[
0.51123046875,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.51123046875,
0.53515625
],
[
0.12005615234375,
0.53759765625,
0.490234375,
0.53759765625,
0.490234375,
0.55029296875,
0.12005615234375,
0.55029296875
],
[
0.515625,
0.53759765625,
0.88330078125,
0.53759765625,
0.88330078125,
0.54736328125,
0.515625,
0.54736328125
],
[
0.12005615234375,
0.552734375,
0.489501953125,
0.552734375,
0.489501953125,
0.5634765625,
0.12005615234375,
0.5634765625
],
[
0.51318359375,
0.548828125,
0.884765625,
0.5498046875,
0.884765625,
0.56396484375,
0.51318359375,
0.5634765625
],
[
0.118408203125,
0.5654296875,
0.490234375,
0.56396484375,
0.490234375,
0.57763671875,
0.118408203125,
0.5791015625
],
[
0.51318359375,
0.5634765625,
0.88330078125,
0.5634765625,
0.88330078125,
0.57666015625,
0.51318359375,
0.57666015625
],
[
0.12005615234375,
0.580078125,
0.489501953125,
0.580078125,
0.489501953125,
0.59326171875,
0.12005615234375,
0.59326171875
],
[
0.51123046875,
0.5791015625,
0.884765625,
0.57763671875,
0.884765625,
0.59130859375,
0.51123046875,
0.5927734375
],
[
0.12005615234375,
0.595703125,
0.489501953125,
0.595703125,
0.489501953125,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.5146484375,
0.59423828125,
0.88330078125,
0.59423828125,
0.88330078125,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.118408203125,
0.60791015625,
0.489501953125,
0.60791015625,
0.489501953125,
0.62158203125,
0.118408203125,
0.62158203125
],
[
0.51318359375,
0.60693359375,
0.884765625,
0.60693359375,
0.884765625,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.118408203125,
0.62060546875,
0.48779296875,
0.62060546875,
0.48779296875,
0.63427734375,
0.118408203125,
0.63427734375
],
[
0.5146484375,
0.62158203125,
0.88671875,
0.62158203125,
0.88671875,
0.63525390625,
0.5146484375,
0.63525390625
],
[
0.1175537109375,
0.63525390625,
0.489501953125,
0.63525390625,
0.489501953125,
0.6484375,
0.1175537109375,
0.6484375
],
[
0.5146484375,
0.63720703125,
0.8818359375,
0.63720703125,
0.8818359375,
0.64794921875,
0.5146484375,
0.64794921875
],
[
0.118408203125,
0.65087890625,
0.489501953125,
0.65087890625,
0.489501953125,
0.66455078125,
0.118408203125,
0.66455078125
],
[
0.51318359375,
0.6494140625,
0.884765625,
0.6494140625,
0.884765625,
0.6630859375,
0.51318359375,
0.6630859375
],
[
0.1175537109375,
0.66455078125,
0.490234375,
0.6650390625,
0.490234375,
0.6787109375,
0.1175537109375,
0.677734375
],
[
0.51318359375,
0.6630859375,
0.88330078125,
0.6630859375,
0.88330078125,
0.6767578125,
0.51318359375,
0.6767578125
],
[
0.118408203125,
0.677734375,
0.48779296875,
0.677734375,
0.48779296875,
0.69140625,
0.118408203125,
0.69140625
],
[
0.5146484375,
0.6787109375,
0.88330078125,
0.6787109375,
0.88330078125,
0.689453125,
0.5146484375,
0.689453125
],
[
0.118408203125,
0.693359375,
0.26123046875,
0.693359375,
0.26123046875,
0.70703125,
0.118408203125,
0.70703125
],
[
0.5146484375,
0.6923828125,
0.884765625,
0.6923828125,
0.884765625,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.1368408203125,
0.708984375,
0.490234375,
0.708984375,
0.490234375,
0.72265625,
0.1368408203125,
0.72265625
],
[
0.5146484375,
0.7080078125,
0.759765625,
0.7080078125,
0.759765625,
0.71875,
0.5146484375,
0.71875
],
[
0.1175537109375,
0.72265625,
0.48779296875,
0.72265625,
0.48779296875,
0.736328125,
0.1175537109375,
0.736328125
],
[
0.53076171875,
0.72021484375,
0.884765625,
0.72021484375,
0.884765625,
0.73388671875,
0.53076171875,
0.73388671875
],
[
0.118408203125,
0.736328125,
0.48779296875,
0.736328125,
0.48779296875,
0.75,
0.118408203125,
0.75
],
[
0.51318359375,
0.7353515625,
0.88330078125,
0.7353515625,
0.88330078125,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.12005615234375,
0.751953125,
0.490234375,
0.751953125,
0.490234375,
0.765625,
0.12005615234375,
0.765625
],
[
0.515625,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76123046875,
0.515625,
0.76123046875
],
[
0.12005615234375,
0.7666015625,
0.490234375,
0.7666015625,
0.490234375,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.5146484375,
0.76318359375,
0.8486328125,
0.76318359375,
0.8486328125,
0.77685546875,
0.5146484375,
0.77685546875
],
[
0.12005615234375,
0.77978515625,
0.489501953125,
0.77978515625,
0.489501953125,
0.79345703125,
0.12005615234375,
0.79345703125
],
[
0.5341796875,
0.78759765625,
0.884765625,
0.78759765625,
0.884765625,
0.80029296875,
0.5341796875,
0.80029296875
],
[
0.118408203125,
0.79443359375,
0.489501953125,
0.79443359375,
0.489501953125,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.54736328125,
0.80029296875,
0.884765625,
0.80029296875,
0.884765625,
0.8154296875,
0.54736328125,
0.8154296875
],
[
0.12005615234375,
0.80859375,
0.489501953125,
0.80859375,
0.489501953125,
0.822265625,
0.12005615234375,
0.822265625
],
[
0.54736328125,
0.8154296875,
0.873046875,
0.8154296875,
0.873046875,
0.8291015625,
0.54736328125,
0.8291015625
],
[
0.118408203125,
0.822265625,
0.489501953125,
0.822265625,
0.489501953125,
0.8359375,
0.118408203125,
0.8359375
],
[
0.12005615234375,
0.83740234375,
0.48779296875,
0.83740234375,
0.48779296875,
0.85107421875,
0.12005615234375,
0.85107421875
],
[
0.5341796875,
0.83837890625,
0.88330078125,
0.83837890625,
0.88330078125,
0.8515625,
0.5341796875,
0.8515625
],
[
0.118408203125,
0.85107421875,
0.489501953125,
0.85107421875,
0.489501953125,
0.8642578125,
0.118408203125,
0.8642578125
],
[
0.54736328125,
0.8525390625,
0.88330078125,
0.8525390625,
0.88330078125,
0.8662109375,
0.54736328125,
0.8662109375
],
[
0.12005615234375,
0.8662109375,
0.489501953125,
0.8662109375,
0.489501953125,
0.8798828125,
0.12005615234375,
0.8798828125
],
[
0.5458984375,
0.8662109375,
0.88330078125,
0.86767578125,
0.88330078125,
0.8818359375,
0.5458984375,
0.88134765625
],
[
0.133544921875,
0.8798828125,
0.489501953125,
0.88134765625,
0.489501953125,
0.8955078125,
0.133544921875,
0.89453125
],
[
0.54736328125,
0.8818359375,
0.88330078125,
0.8818359375,
0.88330078125,
0.8955078125,
0.54736328125,
0.8955078125
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.91015625,
0.118408203125,
0.91015625
],
[
0.54736328125,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.54736328125,
0.9091796875
]
] |
[
"TopicDetectionandTrackingwithTime-AwareDocument",
"Embeddings",
"HangJiang,DougBeeferman,WeiquanMao,DebRoy",
"Massachusetts Institute of Technology,Stanford University",
"75 Amherst St, Cambridge, MA 02139",
"[hjian42,dougb5,dkroy)@mit.edu,mwq@stanford.edu",
"Abstract",
"The time at which a message is communicated is a vital piece of metadata in many real-world natural language",
"processingtaskssuchasTopicDetectionandTracking(TDT).TDTsystemsaimtoclusteracorpusofnewsarticles",
"by event,and in that context,storiesthat describe the same event are likelyto have been written ataround the same",
"time. Prior work on time modeling for TDT takes this into account, but does not well capture how time interacts",
"withthesemanticnatureof theevent.Forexample,storiesaboutatropical stormarelikelytobewrittenwithina",
"0",
"short time interval, while stories about a movie release may appear over weeks or months. In our work, we design",
"a neural method that fuses temporal and textual information into a single representation of news documents for",
"event detection.We fine-tunethese time-aware document embeddings witha triplet lossarchitecture,integrate",
"the model into downstream TDT systems, and evaluate the systems on two benchmark TDT data sets in English.",
"Intheretrospectivesettingweapplyclusteringalgorithmstothetime-awareembeddingsandshowsubstantial",
"improvements over baselines on the News2013 data set. In the online streaming setting, we add our document",
"encodertoan existing state-of-the-artTDTpipelineanddemonstratethat it canbenefittheoverallperformance.We",
"conduct ablation studies on the time representation and fusion algorithm strategies, showing that our proposed",
"model outperforms alternative strategies. Finally, we probe the model to examine how it handles recurring events",
"moreeffectivelythanpreviousTDTsystems",
"[T']",
"Keywords: Topic Detection and Tracking, Online Clustering, Document Embedding, Temporal Embedding",
"1.Introduction",
"TDT modelers use this metadata, at a minimum,",
"to exploit the temporal locality inherent in news",
"Following emerging news stories is crucial to mak-",
"events.Butbeyondthat,researcherstendtofocus",
"ZA9990211Z:A1X11",
"ing real-time decisions on important political and",
"on improving text representations (Hu et al., 2017",
"public safety matters. Amid the COVID-19 pan-",
"Staykovskietal.,2019;Saravanakumaretal.,2021;",
"demic,forinstance,mediaplatformsandgovern-",
"Fan et al., 2021) instead of time representations.",
"ment agencies need to identify the emergence of",
"The time model is typically parameter-free or low",
"misinformation in timeandtake action to protect the",
"capacity, and unconnected to the text. This over-",
"safetyofthepublic.Ashumanscannotreadallof",
"looks important waysthatthetime may interact",
"the articles produced by the news media, automatic",
"with the topic of the article to influence where it falls",
"clustering of news articles into real-world events",
"on the calendar and how long it lasts as an event.",
"is needed to make this work tractable. Computa-",
"Forexample,stockreportshappendaily,butcorpo",
"tional tools for this task are useful for organizing",
"rate earnings releases may happen quarterly and",
"not just news articles,but also scientific papers,",
"be discussed over several days. Stories about a",
"microblogs, online reviews, forum messages, and",
"weather event are likelyto bewritten over a short",
"social media posts",
"time period, while stories a movie release may be",
"Allan et al. (1998) first introduced the topic de-",
"discussedoverweeksormonths",
"tection and tracking (TDT) framework to address",
"In this paper, we address these gaps with a",
"these needs. In this paper, we focus on the event",
"unified neural time-text model that fuses tempo-",
"detection part of TDT, where events are defined",
"ralandsemanticembeddingstorepresentnews",
"as real-world news stories and we want to catego-",
"documents. Our contributions are as follows:",
"rize news articles into these events. Researchers",
". We propose a time-aware neural document",
"have proposed both topic modeling and clustering-",
"embedding method that can be applied to topic",
"based methods for this task, in both retrospective",
"detection and tracking and other NLP tasks.",
"and online streaming settings. In the retrospective",
"setting, researchers process the news documents",
".We build two TDT pipelines based on our",
"altogether, whereas in the online setting they are",
"time-aware model, for retrospective and online",
"processed one by one as they appear in the stream.",
"eventclusteringtasks respectively,achieving",
"Most real-world documents have a timestamp,",
"state-of-the-art performance in both settings",
"the time at which it was written, spoken, or shared.",
"for the corpus used as a benchmark by re-"
] |
{
"arxiv_id": "2112.06166",
"language": "en",
"page_idx": 0
}
|
||
2112.06166-en-1
|
2112.06166-en-1.jpg
|
cent prior work. Importantly, our retrospective model is free of the TF-IDF features that are needed by similar systems, allowing it to be adapted to new domains more easily.
· We conduct an ablation analysis on our time representation. We find that sinusoidal positional embedding outperforms two alternatives, learned positional embedding and Date2Vec, to encode timestamps for TDT.
$\cdot$ We analyze the event predictions of our model and baseline methods and find that our model can better handle recurring events that pose challenges to previous TDT systems.
## 2. Related Work
Topic detection and tracking (TDT) program (Liu, 2009; Fiscus and Doddington, 2002) focuses on building algorithms to organize multilingual, news oriented textual materials from the Internet. Apart from news, TDT techniques are also widely used in processing social media data (Xiong et al., 2022). Traditionally, some researchers (Allan et al., 1998; Yang et al., 1998; Xu et al., 2019; Liu et al., 2020) have focused on applying different topic models (e.g., LDA) for TDT. Other researchers (Hatzivassiloglou et al., 2000; Allan et al., 2003; Dai et al., 2010; Li et al., 2020) explored the use of text clustering algorithms on sparse features and word embeddings for TDT.
Recent approaches to TDT have explored both sparse and dense features. Miranda et al. (2018) proposes an online clustering method that represents documents with TF-IDF features, and demonstrates high performance on a benchmark news article data set. Building on this work, Staykovski et al. (2019) compares sparse TF-IDF features with dense Doc2Vec representations, showing a sizeable improvement on the standard data set according to the BCubed evaluation metric. Saravanakumar et al. (2021) is the first to include BERT contextual representations for the task and achieves further improvement. Specifically, they fine-tune an entity-aware BERT model on an event similarity task with a triplet loss function. They generate triplets for each document using the batch-hard regime (Hermans et al., 2017). In each document in a mini-batch, they mark documents with the same label as positive examples and different labels as negative examples. The hardest positive (biggest positive-anchor document distance) and negative (smallest anchor-negative document distance) examples are picked per anchor document to form a triplet. The entity-aware BERT model is trained to make the embedding distance between anchor and positive documents closer than
anchor and negative documents. Overall, this finetuning process effectively improves the contextual embedding for the overall TDT system. Santos et al. (2022) simplified the multilingual news clustering process with multilingual document embeddings. Recent studies have also leveraged large language models (LLMs) for text clustering (Zhang et al., 2023; Viswanathan et al., 2023) and story discovery (Yoon et al.,2023a,b). This paper focuses on comparing our method against previous embedding-based methods with small language models such as BERT.
TDT systems vary in how they model the timestamps of the news stories. Some online approaches combine the time element implicitly by sorting documents in chronological order, dividing them with time slicing, and processing each slice (Allan et al., 1998; Yang et al., 1998; Dai et al., 2010; Hu et al., 2017). Other work uses decay functions to extract sparse time features (Yang et al., 1998; Brants et al., 2003; Li et al., 2005; He et al., 2010; Ribeiro et al., 2017; Miranda et al., 2018; Saravanakumar et al., 2021). However, none of the previous work has used temporal embeddings to represent time for the TDT task. This work aims to introduce a popular temporal embedding method from Devlin et al. (2018) to TDT such that the fused document embedding contains both temporal and semantic information for clustering.
# 3. Methodology
In this section we propose a novel method called T-E-BERT to encode news documents by fusing text and time information. We adopt a triplet loss function to train the model on the event similarity task and integrate the fine-tuned model into both retrospective and online TDT pipelines.
## 3.1. The Proposed Model
<FigureHere>
Figure 1: The proposed T-E-BERT model.
## 3.1.1. T-E-BERT Encoder
We design a simple time-text encoder (“T-E-BERT”) based on entity-aware BERT to combine textual
|
[
[
0.151123046875,
0.07598876953125,
0.489501953125,
0.07598876953125,
0.489501953125,
0.08905029296875,
0.151123046875,
0.08905029296875
],
[
0.51318359375,
0.07598876953125,
0.88330078125,
0.07598876953125,
0.88330078125,
0.08905029296875,
0.51318359375,
0.08905029296875
],
[
0.151123046875,
0.0885009765625,
0.48779296875,
0.0885009765625,
0.48779296875,
0.1015625,
0.151123046875,
0.1015625
],
[
0.5146484375,
0.0914306640625,
0.88330078125,
0.0914306640625,
0.88330078125,
0.1015625,
0.5146484375,
0.1015625
],
[
0.151123046875,
0.10394287109375,
0.489501953125,
0.10394287109375,
0.489501953125,
0.1175537109375,
0.151123046875,
0.1175537109375
],
[
0.5146484375,
0.10394287109375,
0.8310546875,
0.10394287109375,
0.8310546875,
0.1175537109375,
0.5146484375,
0.1175537109375
],
[
0.82275390625,
0.10394287109375,
0.88330078125,
0.105712890625,
0.88330078125,
0.1175537109375,
0.82275390625,
0.11517333984375
],
[
0.151123046875,
0.1181640625,
0.4306640625,
0.1181640625,
0.4306640625,
0.1318359375,
0.151123046875,
0.1318359375
],
[
0.5146484375,
0.1181640625,
0.88330078125,
0.1181640625,
0.88330078125,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.5146484375,
0.133056640625,
0.88330078125,
0.133056640625,
0.88330078125,
0.146728515625,
0.5146484375,
0.146728515625
],
[
0.1376953125,
0.1455078125,
0.489501953125,
0.1455078125,
0.489501953125,
0.1591796875,
0.1376953125,
0.1591796875
],
[
0.5146484375,
0.146728515625,
0.88330078125,
0.146728515625,
0.88330078125,
0.1602783203125,
0.5146484375,
0.1602783203125
],
[
0.151123046875,
0.1602783203125,
0.489501953125,
0.1602783203125,
0.489501953125,
0.17333984375,
0.151123046875,
0.17333984375
],
[
0.5146484375,
0.160888671875,
0.88330078125,
0.160888671875,
0.88330078125,
0.174560546875,
0.5146484375,
0.174560546875
],
[
0.151123046875,
0.174560546875,
0.489501953125,
0.174560546875,
0.489501953125,
0.188232421875,
0.151123046875,
0.188232421875
],
[
0.51318359375,
0.17333984375,
0.88330078125,
0.174560546875,
0.88330078125,
0.189453125,
0.51318359375,
0.188232421875
],
[
0.15283203125,
0.189453125,
0.490234375,
0.189453125,
0.490234375,
0.203125,
0.15283203125,
0.203125
],
[
0.5146484375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.5146484375,
0.203125
],
[
0.151123046875,
0.203125,
0.382080078125,
0.203125,
0.382080078125,
0.2161865234375,
0.151123046875,
0.2161865234375
],
[
0.5146484375,
0.2037353515625,
0.884765625,
0.2037353515625,
0.884765625,
0.21728515625,
0.5146484375,
0.21728515625
],
[
0.51318359375,
0.2161865234375,
0.88330078125,
0.21728515625,
0.88330078125,
0.232177734375,
0.51318359375,
0.23095703125
],
[
0.1376953125,
0.2298583984375,
0.489501953125,
0.2298583984375,
0.489501953125,
0.243408203125,
0.1376953125,
0.243408203125
],
[
0.5146484375,
0.232177734375,
0.68408203125,
0.232177734375,
0.68408203125,
0.2452392578125,
0.5146484375,
0.2452392578125
],
[
0.15283203125,
0.2452392578125,
0.48779296875,
0.2452392578125,
0.48779296875,
0.255859375,
0.15283203125,
0.255859375
],
[
0.52880859375,
0.2452392578125,
0.88671875,
0.24462890625,
0.88671875,
0.2587890625,
0.52880859375,
0.260009765625
],
[
0.15283203125,
0.2587890625,
0.489501953125,
0.2587890625,
0.489501953125,
0.2724609375,
0.15283203125,
0.2724609375
],
[
0.515625,
0.262451171875,
0.73876953125,
0.262451171875,
0.73876953125,
0.2724609375,
0.515625,
0.2724609375
],
[
0.748046875,
0.26123046875,
0.8818359375,
0.262451171875,
0.8818359375,
0.273681640625,
0.748046875,
0.2724609375
],
[
0.15283203125,
0.271484375,
0.4248046875,
0.271484375,
0.4248046875,
0.2861328125,
0.15283203125,
0.2861328125
],
[
0.51318359375,
0.273681640625,
0.88330078125,
0.27490234375,
0.88330078125,
0.289306640625,
0.51318359375,
0.2880859375
],
[
0.51318359375,
0.287353515625,
0.88330078125,
0.2880859375,
0.88330078125,
0.302734375,
0.51318359375,
0.3017578125
],
[
0.51318359375,
0.3017578125,
0.88330078125,
0.3017578125,
0.88330078125,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.219970703125,
0.31005859375,
0.387939453125,
0.31005859375,
0.387939453125,
0.32470703125,
0.219970703125,
0.32470703125
],
[
0.5146484375,
0.317626953125,
0.88330078125,
0.317626953125,
0.88330078125,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.51318359375,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.12005615234375,
0.34033203125,
0.490234375,
0.34033203125,
0.490234375,
0.35400390625,
0.12005615234375,
0.35400390625
],
[
0.51318359375,
0.345703125,
0.88330078125,
0.345703125,
0.88330078125,
0.359375,
0.51318359375,
0.359375
],
[
0.12005615234375,
0.355224609375,
0.489501953125,
0.355224609375,
0.489501953125,
0.3681640625,
0.12005615234375,
0.3681640625
],
[
0.5146484375,
0.35986328125,
0.884765625,
0.35986328125,
0.884765625,
0.37353515625,
0.5146484375,
0.37353515625
],
[
0.12176513671875,
0.37060546875,
0.48779296875,
0.37060546875,
0.48779296875,
0.38134765625,
0.12176513671875,
0.38134765625
],
[
0.5146484375,
0.37353515625,
0.88330078125,
0.37353515625,
0.88330078125,
0.38720703125,
0.5146484375,
0.38720703125
],
[
0.12005615234375,
0.383056640625,
0.489501953125,
0.383056640625,
0.489501953125,
0.396728515625,
0.12005615234375,
0.396728515625
],
[
0.51123046875,
0.38720703125,
0.8818359375,
0.38720703125,
0.8818359375,
0.40087890625,
0.51123046875,
0.40087890625
],
[
0.118408203125,
0.39794921875,
0.489501953125,
0.39794921875,
0.489501953125,
0.410888671875,
0.118408203125,
0.410888671875
],
[
0.51318359375,
0.402587890625,
0.88330078125,
0.402099609375,
0.88330078125,
0.4150390625,
0.51318359375,
0.416259765625
],
[
0.118408203125,
0.410888671875,
0.48779296875,
0.410888671875,
0.48779296875,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.51123046875,
0.41748046875,
0.88330078125,
0.416259765625,
0.88330078125,
0.429931640625,
0.51123046875,
0.43115234375
],
[
0.118408203125,
0.42578125,
0.490234375,
0.42578125,
0.490234375,
0.439453125,
0.118408203125,
0.439453125
],
[
0.51318359375,
0.43115234375,
0.88330078125,
0.43115234375,
0.88330078125,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.12005615234375,
0.441162109375,
0.48779296875,
0.441162109375,
0.48779296875,
0.451904296875,
0.12005615234375,
0.451904296875
],
[
0.51123046875,
0.44482421875,
0.8818359375,
0.44482421875,
0.8818359375,
0.457763671875,
0.51123046875,
0.457763671875
],
[
0.12176513671875,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.466064453125,
0.12176513671875,
0.466064453125
],
[
0.51123046875,
0.458984375,
0.88330078125,
0.458984375,
0.88330078125,
0.47265625,
0.51123046875,
0.47265625
],
[
0.118408203125,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51318359375,
0.473876953125,
0.77490234375,
0.473876953125,
0.77490234375,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.118408203125,
0.482666015625,
0.48779296875,
0.482666015625,
0.48779296875,
0.496337890625,
0.118408203125,
0.496337890625
],
[
0.12005615234375,
0.496337890625,
0.489501953125,
0.496337890625,
0.489501953125,
0.51025390625,
0.12005615234375,
0.51025390625
],
[
0.6171875,
0.50830078125,
0.6591796875,
0.50830078125,
0.6591796875,
0.521484375,
0.6171875,
0.521484375
],
[
0.6484375,
0.505859375,
0.779296875,
0.50830078125,
0.779296875,
0.5224609375,
0.6484375,
0.52099609375
],
[
0.1175537109375,
0.52490234375,
0.252685546875,
0.52392578125,
0.252685546875,
0.5380859375,
0.1175537109375,
0.5390625
],
[
0.118408203125,
0.51123046875,
0.489501953125,
0.51123046875,
0.489501953125,
0.52490234375,
0.118408203125,
0.52490234375
],
[
0.1351318359375,
0.54052734375,
0.489501953125,
0.54052734375,
0.489501953125,
0.55419921875,
0.1351318359375,
0.55419921875
],
[
0.5146484375,
0.53515625,
0.88330078125,
0.53515625,
0.88330078125,
0.548828125,
0.5146484375,
0.548828125
],
[
0.51123046875,
0.54736328125,
0.8818359375,
0.5498046875,
0.8818359375,
0.5634765625,
0.51123046875,
0.56103515625
],
[
0.118408203125,
0.5546875,
0.489501953125,
0.5546875,
0.489501953125,
0.568359375,
0.118408203125,
0.568359375
],
[
0.12005615234375,
0.568359375,
0.489501953125,
0.568359375,
0.489501953125,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.51318359375,
0.5634765625,
0.88330078125,
0.5634765625,
0.88330078125,
0.57666015625,
0.51318359375,
0.57666015625
],
[
0.51123046875,
0.57666015625,
0.8818359375,
0.57763671875,
0.8818359375,
0.59130859375,
0.51123046875,
0.59033203125
],
[
0.118408203125,
0.5830078125,
0.490234375,
0.5830078125,
0.490234375,
0.5966796875,
0.118408203125,
0.5966796875
],
[
0.51318359375,
0.59130859375,
0.88330078125,
0.59130859375,
0.88330078125,
0.60498046875,
0.51318359375,
0.60498046875
],
[
0.118408203125,
0.59716796875,
0.489501953125,
0.59716796875,
0.489501953125,
0.61083984375,
0.118408203125,
0.61083984375
],
[
0.118408203125,
0.60986328125,
0.48779296875,
0.60986328125,
0.48779296875,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.51318359375,
0.60546875,
0.80419921875,
0.60546875,
0.80419921875,
0.619140625,
0.51318359375,
0.619140625
],
[
0.12005615234375,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.12005615234375,
0.6396484375
],
[
0.515625,
0.63818359375,
0.55908203125,
0.63818359375,
0.55908203125,
0.6494140625,
0.515625,
0.6494140625
],
[
0.56103515625,
0.63720703125,
0.748046875,
0.63720703125,
0.748046875,
0.65185546875,
0.56103515625,
0.65185546875
],
[
0.12005615234375,
0.65283203125,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.12005615234375,
0.66650390625
],
[
0.12005615234375,
0.64013671875,
0.490234375,
0.64013671875,
0.490234375,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.12005615234375,
0.66748046875,
0.490234375,
0.66748046875,
0.490234375,
0.68115234375,
0.12005615234375,
0.68115234375
],
[
0.1175537109375,
0.68115234375,
0.48779296875,
0.68115234375,
0.48779296875,
0.69482421875,
0.1175537109375,
0.69482421875
],
[
0.62109375,
0.6904296875,
0.6689453125,
0.6904296875,
0.6689453125,
0.70068359375,
0.62109375,
0.70068359375
],
[
0.118408203125,
0.69775390625,
0.48779296875,
0.69775390625,
0.48779296875,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.77099609375,
0.6923828125,
0.79833984375,
0.6923828125,
0.79833984375,
0.70068359375,
0.77099609375,
0.70068359375
],
[
0.61279296875,
0.705078125,
0.68115234375,
0.705078125,
0.68115234375,
0.7119140625,
0.61279296875,
0.7119140625
],
[
0.12005615234375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.72265625,
0.12005615234375,
0.72265625
],
[
0.12005615234375,
0.724609375,
0.489501953125,
0.724609375,
0.489501953125,
0.73828125,
0.12005615234375,
0.73828125
],
[
0.6455078125,
0.728515625,
0.69873046875,
0.7353515625,
0.6943359375,
0.75048828125,
0.640625,
0.74462890625
],
[
0.748046875,
0.724609375,
0.8251953125,
0.724609375,
0.8251953125,
0.728515625,
0.748046875,
0.728515625
],
[
0.59765625,
0.73095703125,
0.6484375,
0.7353515625,
0.6455078125,
0.7490234375,
0.59423828125,
0.74462890625
],
[
0.118408203125,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.118408203125,
0.7529296875
],
[
0.12005615234375,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.765625,
0.12005615234375,
0.765625
],
[
0.771484375,
0.75048828125,
0.78271484375,
0.75048828125,
0.78271484375,
0.7548828125,
0.771484375,
0.7548828125
],
[
0.118408203125,
0.76708984375,
0.489501953125,
0.76708984375,
0.489501953125,
0.78076171875,
0.118408203125,
0.78076171875
],
[
0.59765625,
0.7666015625,
0.63623046875,
0.7666015625,
0.63623046875,
0.779296875,
0.59765625,
0.779296875
],
[
0.6591796875,
0.76708984375,
0.67236328125,
0.7724609375,
0.66748046875,
0.77978515625,
0.654296875,
0.77490234375
],
[
0.118408203125,
0.7822265625,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.118408203125,
0.7958984375
],
[
0.12005615234375,
0.79638671875,
0.489501953125,
0.79638671875,
0.489501953125,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.12005615234375,
0.81005859375,
0.489501953125,
0.81005859375,
0.489501953125,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.54248046875,
0.81103515625,
0.85400390625,
0.81005859375,
0.85400390625,
0.82470703125,
0.54248046875,
0.826171875
],
[
0.12005615234375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.118408203125,
0.83837890625,
0.489501953125,
0.83837890625,
0.489501953125,
0.8515625,
0.118408203125,
0.8515625
],
[
0.118408203125,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.118408203125,
0.8662109375
],
[
0.515625,
0.85986328125,
0.56494140625,
0.85986328125,
0.56494140625,
0.87060546875,
0.515625,
0.87060546875
],
[
0.5634765625,
0.8623046875,
0.57080078125,
0.8623046875,
0.57080078125,
0.86767578125,
0.5634765625,
0.86767578125
],
[
0.56689453125,
0.85791015625,
0.724609375,
0.85791015625,
0.724609375,
0.87158203125,
0.56689453125,
0.87158203125
],
[
0.12005615234375,
0.8681640625,
0.48779296875,
0.8681640625,
0.48779296875,
0.87890625,
0.12005615234375,
0.87890625
],
[
0.12005615234375,
0.8828125,
0.489501953125,
0.8828125,
0.489501953125,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.51318359375,
0.8798828125,
0.88330078125,
0.8798828125,
0.88330078125,
0.89453125,
0.51318359375,
0.89453125
],
[
0.118408203125,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"cent prior work. Importantly,our retrospective",
"anchor and negative documents.Overall, this fine",
"model is free of the TF-IDF features that are",
"tuningprocesseffectivelyimprovesthecontextual",
"needed by similar systems, allowing it to be",
"embedding for the overall TDT system.",
"Santos",
"adapted to new domains more easily.",
"et al. (2022) simplified the multilingual news clus-",
"tering process with multilingual document embed-",
". We conduct an ablation analysis on our time",
"dings. Recent studies have also leveraged large",
"representation. We find that sinusoidal posi-",
"language models (LLMs) for text clustering (Zhang",
"tional embedding outperforms two alternatives,",
"et al., 2023; Viswanathan et al., 2023) and story",
"learned positional embedding and Date2Vec,",
"discovery (Yoon et al., 2023a,b). This paper fo-",
"to encode timestamps for TDT.",
"cuses on comparing our method against previous",
"embedding-based methods with small language",
": We analyze the event predictions of our model",
"models such as BERT.",
"and baseline methods andfindthat ourmodel",
"TDT systems vary in how they model the times-",
"can better handle recurring events that pose",
"tamps of the newsstories.",
"Some online ap",
"challenges to previous TDT systems.",
"proaches combine the time element implicitly by",
"sorting documents in chronological order, dividing",
"them with time slicing, and processing each slice",
"2.Related Work",
"(Allan et al., 1998; Yang et al., 1998; Dai et al., 2010;",
"Hu et al., 2017). Other work uses decay functions",
"Topic detection and tracking (TDT) program (Liu,",
"to extract sparse time features (Yang et al., 1998;",
"2009; Fiscus and Doddington, 2002) focuses on",
"Brants et al., 2003; Li et al., 2005; He et al., 2010;",
"buildingalgorithmstoorganizemultilingual,news",
"Ribeiro et al., 2017; Miranda et al., 2018; Sara",
"oriented textual materials from the Internet. Apart",
"vanakumar et al., 2021). However, none of the",
"from news, TDT techniques are also widely used in",
"previous work has used temporal embeddings to",
"processing social media data (Xiong et al., 2022)",
"represent time for the TDT task. This work aims to",
"Traditionally, some researchers (Allan et al., 1998;",
"introduce a popular temporal embedding method",
"Yang et al.,1998; Xu et al.,2019; Liu et al.,2020)",
"from Devlin et al. (2018) to TDT such that the fused",
"havefocusedonapplyingdifferenttopicmodels",
" document embedding contains both temporal and",
"(e.g., LDA) for TDT. Other researchers (Hatzivas-",
"semantic information for clustering.",
"siloglou et al., 2000; Allan et al., 2003; Dai et al.",
"2010; Li et al., 2020) explored the use of text clus-",
"3.",
"Methodology",
"beddings for TDT.",
"tering algorithms on sparse features and word em-",
"Recent approaches to TDT have explored both",
"In this section we propose a novel method called",
"T-E-BERTto encode news documents by fusing",
"sparse and dense features. Miranda et al. (2018)",
"proposes an online clustering method that repre-",
"text and time information. We adopt a triplet loss",
"function to train the model on the event similarity",
"sents documents with TF-IDF features, and demon-",
"task and integrate the fine-tuned model into both",
"strates high performance on a benchmark news",
"article data set. Building on this work, Staykovski",
"retrospective and online TDT pipelines.",
"et al. (2019) compares sparse TF-IDF features with",
"3.1.",
"The Proposed Model",
"able improvement on the standard data set accord",
"dense Doc2Vec representations, showing a size-",
"ing to the BCubed evaluation metric. Saravanaku-",
"mar et al. (2021) is the first to include BERT con",
"Add&Nom",
"textualrepresentationsforthetaskandachieves",
"Pooling",
"Multi-Head Attention",
"furtherimprovement.Specifically,theyfine-tune",
"an entity-aware BERT model on an event similar-",
"Mtime",
"acDounneCnbed",
"Mtert",
"ity task with a triplet loss function. They generate",
"triplets for each document using the batch-hard",
"",
"regime (Hermans et al., 2017). In each document",
"ENBERTON",
"Co",
"in a mini-batch, they mark documents with the",
"same label as positive examples and different la-",
"bels as negative examples. The hardest positive",
"Figure 1: The proposed T-E-BERT model.",
"(biggest positive-anchor document distance) and",
"negative(smallestanchor-negativedocumentdis-",
"tance) examples are picked per anchor document",
"3.1.1.",
"",
"T-E-BERTEncoder",
"toforma triplet.The entity-aware BERTmodel",
"is trained to make the embedding distance be-",
"We design a simple time-text encoder (\"T-E-BERT\")",
"tween anchor and positive documents closer than",
"based on entity-aware BERT to combine textual"
] |
{
"arxiv_id": "2112.06166",
"language": "en",
"page_idx": 1
}
|
||
2112.06166-en-2
|
2112.06166-en-2.jpg
|
and temporal information to encode news documents (shown in Figure 1). Following Saravanakumar et al. (2021), we add an entity presenceabsence embedding layer to enhance BERT's entity awareness, which can improve the text representation of news events. We use all the hidden output from the last layer as the textual matrix $M_{text}^{1}.$ To represent time, we convert the date time in each document into a time step (e.g. the number of days from the earliest date in the data) and transform the time step into a temporal embedding with the sinusoidal position encoding method introduced by Vaswani et al. (2017). The temporal embedding is repeated "sequence length" times to generate $M_{time}$,which is the same shape as the $M_text. \textbf{Af- }$ terwards, we introduce a fusion module based on multi-head attention to transform the concatenation of text and time embeddings into a text-time matrix. At last, the fused matrix is fed into the pooling layer to generate a news document embedding. The model is trained on a event similarity task to learn how to combine text and time information before this is used for downstream TDT tasks.
## 3.1.2. Fine-tuning
We follow the fine-tuning procedure suggested by Saravanakumar et al. (2021) to adapt the triplet network structure (Hoffer and Ailon, 2015) and finetune our T-E-BERT model on the event similarity task. The task aims to tune the model such that it can make the temporal-semantic similarity between same events smaller than the temporal-semantic similarity between different events. In Figure 2, we demonstrate how the training paradigm works. Given an anchor document $d_a$,we sample a positive document $d_p$ ( from the same event as $d_{a})$ as well as a negative document $d_n$ (from a different event). We compute the triplet loss function as follows:
$$L_{triplet}=sim(d_a,d_n)-sim(d_a,d_p)+m$$
(1)
where sim is the cosine similarity function and $m$
is the hyper-parameter margin.
<FigureHere>
Figure 2: This figures demonstrates the fine-tuning
procedure of T-E-BERT on news event triplets.
## 3.2. Retrospective TDT Pipeline
The retrospective TDT pipeline is simple, consisting of a document encoder and a clustering module At the document encoding step, we concatenate the title and the body of the news articles to form the input text. We then replace the original TFIDF encoder (Yang et al., 1998; Allan et al., 1998; Schultz and Liberman, 1999) with the fine-tuned T-E-BERT to vectorize the news documents, considering both their timestamps and texts. These vectors are directly fed into a clustering algorithm to produce event clusters. We choose the HDBSCAN clustering algorithm2 (Mclnnes et al., 2017) for two reasons. First, HDBSCAN does not require the number of clusters as a hyperparameter, which is unknown to a TDT system in a real-world deployment. Second, HDBSCAN shows strong empirical performance in our experiments, even compared with K-Means and agglomerative clustering algorithms using the true number of clusters.
## 3.3. Online TDT Pipeline
We follow the previous work (Miranda et al., 2018; Saravanakumar et al., 2021) to adopt a variant of the streaming K-means algorithm (Figure 3) with a few key changes. This system consists of three main components: (1) a document encoder, (2) a document-cluster weighted similarity model, (3) a cluster creation model. At any point time $t$,let $n$ be the number of clusters in the cluster pool. For any input document, we assume it belongs to a single event cluster. We first represent this document with a set of vectors including 9 TF-IDF sparse vectors, one dense vector from T-E-BERT, and one time sparse vector (Saravanakumar et al., 2021). After the document representation is extracted, we use the document-cluster weighted similarity model to find the best matching event cluster $C^*$ from the cluster pool. Finally we use a cluster creation model to decide whether a new cluster is needed. If the cluster $C^*$ is predicted to be a good fit for the new document, we add the document to this cluster. Otherwise, we create a new cluster containing this document and add the cluster into the cluster pool. We make two major changes to previous methods (Miranda et al., 2018; Saravanakumar et al., 2021). First, we switch the weighted similarity model from SVM-rank to a linear model with MarginRankingLoss. This change boosts the performance of the weighted similarity model3. Second, we sample balanced positive and negative examples to train the cluster creation model. This alleviates the issue of data imbalance because only 5% of the
$^{2}$https://hdbscan.readthedocs.io/
We also experimented with SVM-triplet but our implementation performed worse than their reported performance (Saravanakumar et al., 2021).
$^{1}M_{text}$ has a shape of [sequence length, hidden size].
|
[
[
0.118408203125,
0.07598876953125,
0.49365234375,
0.07598876953125,
0.49365234375,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.5146484375,
0.07598876953125,
0.5615234375,
0.07598876953125,
0.5615234375,
0.0872802734375,
0.5146484375,
0.0872802734375
],
[
0.55908203125,
0.0789794921875,
0.56494140625,
0.0789794921875,
0.56494140625,
0.0843505859375,
0.55908203125,
0.0843505859375
],
[
0.56103515625,
0.0748291015625,
0.80419921875,
0.0748291015625,
0.80419921875,
0.0885009765625,
0.56103515625,
0.0885009765625
],
[
0.118408203125,
0.08905029296875,
0.491943359375,
0.08905029296875,
0.491943359375,
0.10272216796875,
0.118408203125,
0.10272216796875
],
[
0.51318359375,
0.0955810546875,
0.88330078125,
0.0968017578125,
0.88330078125,
0.11102294921875,
0.51318359375,
0.10986328125
],
[
0.12176513671875,
0.1051025390625,
0.489501953125,
0.1051025390625,
0.489501953125,
0.11517333984375,
0.12176513671875,
0.11517333984375
],
[
0.118408203125,
0.11639404296875,
0.489501953125,
0.1175537109375,
0.489501953125,
0.1318359375,
0.118408203125,
0.1312255859375
],
[
0.5146484375,
0.1134033203125,
0.884765625,
0.1134033203125,
0.884765625,
0.12353515625,
0.5146484375,
0.12353515625
],
[
0.51318359375,
0.1258544921875,
0.88330078125,
0.1258544921875,
0.88330078125,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.12176513671875,
0.1341552734375,
0.48779296875,
0.1341552734375,
0.48779296875,
0.144287109375,
0.12176513671875,
0.144287109375
],
[
0.51318359375,
0.14013671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.15380859375,
0.51318359375,
0.15380859375
],
[
0.12005615234375,
0.1478271484375,
0.48779296875,
0.1478271484375,
0.48779296875,
0.157958984375,
0.12005615234375,
0.157958984375
],
[
0.51318359375,
0.15380859375,
0.884765625,
0.15380859375,
0.884765625,
0.16748046875,
0.51318359375,
0.16748046875
],
[
0.118408203125,
0.160888671875,
0.489501953125,
0.160888671875,
0.489501953125,
0.174560546875,
0.118408203125,
0.174560546875
],
[
0.51318359375,
0.16748046875,
0.8818359375,
0.16748046875,
0.8818359375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.12005615234375,
0.177001953125,
0.48779296875,
0.177001953125,
0.48779296875,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.5146484375,
0.182861328125,
0.884765625,
0.182861328125,
0.884765625,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.118408203125,
0.189453125,
0.490234375,
0.189453125,
0.490234375,
0.203125,
0.118408203125,
0.203125
],
[
0.51318359375,
0.1953125,
0.884765625,
0.19775390625,
0.884765625,
0.21142578125,
0.51318359375,
0.208984375
],
[
0.12005615234375,
0.204833984375,
0.48779296875,
0.204833984375,
0.48779296875,
0.215576171875,
0.12005615234375,
0.215576171875
],
[
0.5146484375,
0.21142578125,
0.884765625,
0.21142578125,
0.884765625,
0.2244873046875,
0.5146484375,
0.2244873046875
],
[
0.12005615234375,
0.2197265625,
0.48779296875,
0.2197265625,
0.48779296875,
0.2298583984375,
0.12005615234375,
0.2298583984375
],
[
0.51318359375,
0.2257080078125,
0.88330078125,
0.2257080078125,
0.88330078125,
0.2392578125,
0.51318359375,
0.2392578125
],
[
0.118408203125,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.51318359375,
0.2392578125,
0.88330078125,
0.2392578125,
0.88330078125,
0.2529296875,
0.51318359375,
0.2529296875
],
[
0.12005615234375,
0.2452392578125,
0.489501953125,
0.2452392578125,
0.489501953125,
0.2587890625,
0.12005615234375,
0.2587890625
],
[
0.51123046875,
0.2529296875,
0.884765625,
0.251708984375,
0.884765625,
0.2666015625,
0.51123046875,
0.267333984375
],
[
0.12005615234375,
0.260009765625,
0.489501953125,
0.260009765625,
0.489501953125,
0.273681640625,
0.12005615234375,
0.273681640625
],
[
0.51318359375,
0.267333984375,
0.88330078125,
0.267333984375,
0.88330078125,
0.28076171875,
0.51318359375,
0.28076171875
],
[
0.118408203125,
0.273681640625,
0.490234375,
0.2724609375,
0.490234375,
0.2880859375,
0.118408203125,
0.289306640625
],
[
0.5146484375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.118408203125,
0.2880859375,
0.490234375,
0.2880859375,
0.490234375,
0.3017578125,
0.118408203125,
0.3017578125
],
[
0.51318359375,
0.295654296875,
0.884765625,
0.29638671875,
0.884765625,
0.31005859375,
0.51318359375,
0.308837890625
],
[
0.12176513671875,
0.303955078125,
0.48779296875,
0.303955078125,
0.48779296875,
0.314208984375,
0.12176513671875,
0.314208984375
],
[
0.5146484375,
0.31005859375,
0.88330078125,
0.31005859375,
0.88330078125,
0.32373046875,
0.5146484375,
0.32373046875
],
[
0.12005615234375,
0.317626953125,
0.490234375,
0.317626953125,
0.490234375,
0.330810546875,
0.12005615234375,
0.330810546875
],
[
0.1175537109375,
0.330078125,
0.490234375,
0.33203125,
0.490234375,
0.345703125,
0.1175537109375,
0.34326171875
],
[
0.51318359375,
0.32470703125,
0.884765625,
0.32470703125,
0.884765625,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.51123046875,
0.33837890625,
0.8818359375,
0.33837890625,
0.8818359375,
0.3515625,
0.51123046875,
0.3515625
],
[
0.118408203125,
0.3466796875,
0.48779296875,
0.345703125,
0.48779296875,
0.35693359375,
0.118408203125,
0.358154296875
],
[
0.51318359375,
0.3515625,
0.8134765625,
0.3515625,
0.8134765625,
0.365234375,
0.51318359375,
0.365234375
],
[
0.118408203125,
0.35986328125,
0.489501953125,
0.35986328125,
0.489501953125,
0.37353515625,
0.118408203125,
0.37353515625
],
[
0.118408203125,
0.37353515625,
0.489501953125,
0.37353515625,
0.489501953125,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.515625,
0.38427734375,
0.5615234375,
0.38427734375,
0.5615234375,
0.3955078125,
0.515625,
0.3955078125
],
[
0.55908203125,
0.3818359375,
0.73876953125,
0.3818359375,
0.73876953125,
0.3955078125,
0.55908203125,
0.3955078125
],
[
0.1175537109375,
0.388427734375,
0.408935546875,
0.388427734375,
0.408935546875,
0.402099609375,
0.1175537109375,
0.402099609375
],
[
0.55908203125,
0.38720703125,
0.56494140625,
0.38720703125,
0.56494140625,
0.392578125,
0.55908203125,
0.392578125
],
[
0.5146484375,
0.405029296875,
0.88330078125,
0.405029296875,
0.88330078125,
0.418701171875,
0.5146484375,
0.418701171875
],
[
0.12005615234375,
0.418701171875,
0.17041015625,
0.418701171875,
0.17041015625,
0.429931640625,
0.12005615234375,
0.429931640625
],
[
0.1695556640625,
0.419189453125,
0.270263671875,
0.419189453125,
0.270263671875,
0.43115234375,
0.1695556640625,
0.43115234375
],
[
0.5146484375,
0.419189453125,
0.884765625,
0.419189453125,
0.884765625,
0.432861328125,
0.5146484375,
0.432861328125
],
[
0.51123046875,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44775390625,
0.51123046875,
0.44775390625
],
[
0.118408203125,
0.440673828125,
0.48779296875,
0.440673828125,
0.48779296875,
0.45361328125,
0.118408203125,
0.45361328125
],
[
0.118408203125,
0.45361328125,
0.48779296875,
0.45361328125,
0.48779296875,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.51318359375,
0.448974609375,
0.88330078125,
0.448974609375,
0.88330078125,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.51123046875,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.51123046875,
0.4755859375
],
[
0.118408203125,
0.468505859375,
0.490234375,
0.468505859375,
0.490234375,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51318359375,
0.476806640625,
0.88330078125,
0.476806640625,
0.88330078125,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.12005615234375,
0.48388671875,
0.48779296875,
0.48388671875,
0.48779296875,
0.49462890625,
0.12005615234375,
0.49462890625
],
[
0.118408203125,
0.496337890625,
0.490234375,
0.496337890625,
0.490234375,
0.51025390625,
0.118408203125,
0.51025390625
],
[
0.5146484375,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.5029296875,
0.5146484375,
0.5029296875
],
[
0.12005615234375,
0.51025390625,
0.490234375,
0.51123046875,
0.490234375,
0.52490234375,
0.12005615234375,
0.52392578125
],
[
0.51318359375,
0.5048828125,
0.88330078125,
0.5048828125,
0.88330078125,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.51318359375,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.51318359375,
0.533203125
],
[
0.118408203125,
0.525390625,
0.489501953125,
0.525390625,
0.489501953125,
0.5390625,
0.118408203125,
0.5390625
],
[
0.5146484375,
0.53369140625,
0.88330078125,
0.53369140625,
0.88330078125,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.12005615234375,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.51123046875,
0.54638671875,
0.88330078125,
0.54736328125,
0.88330078125,
0.5625,
0.51123046875,
0.56103515625
],
[
0.1175537109375,
0.55419921875,
0.489501953125,
0.55419921875,
0.489501953125,
0.56787109375,
0.1175537109375,
0.56787109375
],
[
0.12005615234375,
0.56787109375,
0.490234375,
0.56787109375,
0.490234375,
0.58056640625,
0.12005615234375,
0.58056640625
],
[
0.5146484375,
0.5634765625,
0.88330078125,
0.5634765625,
0.88330078125,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.51318359375,
0.576171875,
0.88330078125,
0.576171875,
0.88330078125,
0.5888671875,
0.51318359375,
0.5888671875
],
[
0.118408203125,
0.58203125,
0.489501953125,
0.58203125,
0.489501953125,
0.595703125,
0.118408203125,
0.595703125
],
[
0.51318359375,
0.59033203125,
0.884765625,
0.59033203125,
0.884765625,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.1175537109375,
0.5966796875,
0.489501953125,
0.5966796875,
0.489501953125,
0.60986328125,
0.1175537109375,
0.60986328125
],
[
0.118408203125,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.5146484375,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.6162109375,
0.5146484375,
0.6162109375
],
[
0.5146484375,
0.6181640625,
0.884765625,
0.6181640625,
0.884765625,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.118408203125,
0.6259765625,
0.17626953125,
0.6259765625,
0.17626953125,
0.63720703125,
0.118408203125,
0.63720703125
],
[
0.51318359375,
0.6318359375,
0.8818359375,
0.6318359375,
0.8818359375,
0.6455078125,
0.51318359375,
0.6455078125
],
[
0.5146484375,
0.6484375,
0.8818359375,
0.6484375,
0.8818359375,
0.6591796875,
0.5146484375,
0.6591796875
],
[
0.145263671875,
0.662109375,
0.44921875,
0.6611328125,
0.44921875,
0.67578125,
0.145263671875,
0.6767578125
],
[
0.468505859375,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.67333984375,
0.468505859375,
0.67333984375
],
[
0.51318359375,
0.6611328125,
0.884765625,
0.6611328125,
0.884765625,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.5146484375,
0.67578125,
0.88671875,
0.67578125,
0.88671875,
0.689453125,
0.5146484375,
0.689453125
],
[
0.1175537109375,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.1175537109375,
0.69873046875
],
[
0.5146484375,
0.6904296875,
0.884765625,
0.6904296875,
0.884765625,
0.70361328125,
0.5146484375,
0.70361328125
],
[
0.1175537109375,
0.69873046875,
0.345947265625,
0.69873046875,
0.345947265625,
0.7119140625,
0.1175537109375,
0.7119140625
],
[
0.51318359375,
0.703125,
0.88671875,
0.70361328125,
0.88671875,
0.71728515625,
0.51318359375,
0.71630859375
],
[
0.52880859375,
0.71630859375,
0.884765625,
0.71630859375,
0.884765625,
0.72998046875,
0.52880859375,
0.72998046875
],
[
0.393798828125,
0.732421875,
0.40478515625,
0.732421875,
0.40478515625,
0.73681640625,
0.393798828125,
0.73681640625
],
[
0.5146484375,
0.732421875,
0.88671875,
0.732421875,
0.88671875,
0.7451171875,
0.5146484375,
0.7451171875
],
[
0.2333984375,
0.7392578125,
0.279541015625,
0.7392578125,
0.279541015625,
0.75048828125,
0.2333984375,
0.75048828125
],
[
0.2166748046875,
0.7421875,
0.2359619140625,
0.7421875,
0.2359619140625,
0.75,
0.2166748046875,
0.75
],
[
0.16796875,
0.73681640625,
0.2091064453125,
0.73681640625,
0.2091064453125,
0.7529296875,
0.16796875,
0.7529296875
],
[
0.124267578125,
0.74365234375,
0.1502685546875,
0.74365234375,
0.1502685546875,
0.751953125,
0.124267578125,
0.751953125
],
[
0.51318359375,
0.7451171875,
0.884765625,
0.74658203125,
0.884765625,
0.76025390625,
0.51318359375,
0.7587890625
],
[
0.345947265625,
0.755859375,
0.50146484375,
0.755859375,
0.50146484375,
0.7666015625,
0.345947265625,
0.7666015625
],
[
0.51318359375,
0.76025390625,
0.88671875,
0.76025390625,
0.88671875,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1695556640625,
0.771484375,
0.20654296875,
0.771484375,
0.20654296875,
0.78369140625,
0.1695556640625,
0.78369140625
],
[
0.2333984375,
0.77490234375,
0.279541015625,
0.77490234375,
0.279541015625,
0.78515625,
0.2333984375,
0.78515625
],
[
0.5146484375,
0.77490234375,
0.884765625,
0.77490234375,
0.884765625,
0.7880859375,
0.5146484375,
0.7880859375
],
[
0.5146484375,
0.7890625,
0.884765625,
0.7890625,
0.884765625,
0.802734375,
0.5146484375,
0.802734375
],
[
0.21826171875,
0.77783203125,
0.231689453125,
0.77783203125,
0.231689453125,
0.7822265625,
0.21826171875,
0.7822265625
],
[
0.4130859375,
0.7958984375,
0.459228515625,
0.7958984375,
0.459228515625,
0.802734375,
0.4130859375,
0.802734375
],
[
0.16796875,
0.802734375,
0.208251953125,
0.802734375,
0.208251953125,
0.818359375,
0.16796875,
0.818359375
],
[
0.235107421875,
0.80712890625,
0.268798828125,
0.80712890625,
0.268798828125,
0.8154296875,
0.235107421875,
0.8154296875
],
[
0.515625,
0.8046875,
0.88330078125,
0.8046875,
0.88330078125,
0.8154296875,
0.515625,
0.8154296875
],
[
0.5146484375,
0.81787109375,
0.88330078125,
0.81787109375,
0.88330078125,
0.83056640625,
0.5146484375,
0.83056640625
],
[
0.124267578125,
0.80859375,
0.1561279296875,
0.80859375,
0.1561279296875,
0.81640625,
0.124267578125,
0.81640625
],
[
0.12005615234375,
0.83740234375,
0.489501953125,
0.83740234375,
0.489501953125,
0.85107421875,
0.12005615234375,
0.85107421875
],
[
0.5146484375,
0.83203125,
0.884765625,
0.83203125,
0.884765625,
0.845703125,
0.5146484375,
0.845703125
],
[
0.12005615234375,
0.8515625,
0.46337890625,
0.8515625,
0.46337890625,
0.865234375,
0.12005615234375,
0.865234375
],
[
0.5322265625,
0.85693359375,
0.75390625,
0.85791015625,
0.75390625,
0.87255859375,
0.5322265625,
0.87158203125
],
[
0.5322265625,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.5322265625,
0.8857421875
],
[
0.51318359375,
0.8857421875,
0.884765625,
0.8857421875,
0.884765625,
0.8994140625,
0.51318359375,
0.8994140625
],
[
0.139404296875,
0.8935546875,
0.486083984375,
0.8955078125,
0.486083984375,
0.91015625,
0.139404296875,
0.90771484375
],
[
0.51123046875,
0.8984375,
0.78271484375,
0.8994140625,
0.78271484375,
0.9130859375,
0.51123046875,
0.912109375
]
] |
[
"and temporal information to encode news docu-",
"3.2.",
"",
"RetrospectiveTDTPipeline",
"ments (shown in Figure 1). Following Saravanaku-",
"The retrospective TDT pipeline is simple, consisting",
"mar et al.(2021),we add an entity presence-",
"absenceembeddinglayerto enhanceBERT's entity",
"of a documentencoder and a clustering module.",
"At the document encoding step, we concatenate",
"awareness,whichcanimprovethetextrepresenta",
"the title and the body of the news articles to form",
"tion of news events.We use all the hiddenoutput",
"the input text. We then replace the original TF-",
"from the last layer as the textual matrix Mtert1. To",
"IDF encoder (Yang et al., 1998; Allan et al., 1998",
"representtime,we convertthe datetimeineach",
"Schultz and Liberman, 1999) with the fine-tuned",
"document into a time step (e.g. the number of days",
"T-E-BERT to vectorize the news documents, con-",
"fromthe earliestdateinthedata)andtransform",
"sidering both their timestamps and texts. These",
"thetimestepintoatemporal embeddingwiththe",
"vectors are directly fed into a clustering algorithm to",
"sinusoidal position encoding method introduced by",
"produce event clusters. We choose the HDBSCAN",
"Vaswani et al. (2017). The temporal embedding",
"clustering algorithm? (Mclnnes et al., 2017) for two",
"is repeated \"sequence length\" times to generate",
"reasons. First, HDBSCAN does not require the",
"Mtime, which is the same shape as the Mtert. Af-",
"number of clusters as a hyperparameter, which is",
"terwards, we introduce a fusion module based on",
"unknowntoaTDTsystem inareal-worlddeploy-",
"multi-headattentiontotransformtheconcatenation",
"ment. Second, HDBSCAN shows strong empirical",
"of text and time embeddings into a text-time matrix.",
"At last, the fused matrix is fed into the pooling layer",
"performance in our experiments, even compared",
"with K-Means and agglomerative clustering algo",
"to generate a news document embedding.The",
"rithms using the true number of clusters.",
"model is trained on a event similarity task to learn",
"how to combine text and time information before",
"3.3.",
"Online TDT Pipeline",
"this is used for downstream TDT tasks.",
"",
"We follow the previous work (Miranda et al., 2018;",
"3.1.2.",
"Fine-tuning",
"Saravanakumar et al., 2021) to adopt a variant of",
"the streaming K-means algorithm (Figure 3) with",
"We follow the fine-tuning procedure suggested by",
"Saravanakumar et al. (2021) to adapt the triplet",
"a few key changes. This system consists of three",
"main components: (1) a document encoder, (2) a",
"network structure (Hoffer and Ailon, 2015) and fine-",
"document-cluster weighted similarity model, (3) a",
"tune our T-E-BERT model onthe event similarity",
"task. The task aims to tune the model such that it",
"cluster creation model.At any pointtimet,let nbe",
"can make the temporal-semantic similarity between",
"the number of clusters in the cluster pool. For any",
"input document, we assume it belongs to a single",
"same events smaller than the temporal-semantic",
"event cluster. We first represent this document with",
"similarity between different events.In Figure 2,",
"a set of vectors including 9 TF-IDF sparse vectors,",
"we demonstrate how the training paradigm works",
"Given an anchor document da, we sample a posi-",
"onedensevectorfromT-E-BERT,andonetime",
"sparse vector (Saravanakumar et al., 2021).After",
"tive document dp (from the same event as da) as",
"the document representation is extracted, we use",
"well as a negative document dn (from a different",
"event). We compute the triplet loss function as",
"thedocument-clusterweightedsimilaritymodelto",
"find the best matching event cluster C* from the",
"follows:",
"cluster pool. Finally we use a cluster creation model",
"todecidewhetheranewclusterisneeded.Ifthe",
"Ltriplet = sim(da, dn) - sim(da, dp) +m",
"(1)",
"cluster C* is predicted to be a good fit for the new",
"document, we add the document to this cluster.",
"where sim is the cosine similarity function and m",
"Otherwise, we create a new cluster containing this",
"is the hyper-parameter margin.",
"document and add the cluster into the cluster pool.",
"We make two major changes to previous meth-",
"MON",
"ods (Miranda et al., 2018; Saravanakumar et al.,",
"TE-BERT",
"",
"国",
"anchor",
"2021). First, we switch the weighted similarity",
"sim(da,dn)-sim(da,dp)+m",
"model from SvM-rank to a linear model with Margin-",
"",
"TE-BERT",
"RankingLoss.This changeboosts theperformance",
"of the weighted similarity mode/3. Second, we sam-",
"",
"therawayfron",
"",
"TE-BERT",
"plebalancedpositiveandnegative examplesto",
"train the cluster creation model. This alleviates the",
"negative",
"Figure 2: This figures demonstrates the fine-tuning",
"issue of data imbalance because only 5% of the",
"procedure of T-E-BERT on news event triplets.",
"2https://hdbscan.readthedocs.io/",
"3We also experimented with SVM-triplet but our im-",
"plementation performed worse than their reported per-",
"1 Mtert has a shape of [sequence length, hidden size]",
"formance (Saravanakumar et al., 2021)."
] |
{
"arxiv_id": "2112.06166",
"language": "en",
"page_idx": 2
}
|
||
2112.06166-en-3
|
2112.06166-en-3.jpg
|
training data points contain a positive label (create
a new cluster).
### 4. Experiments
## 4.1. Data Sets
We use two data sets for our experiments, which we refer to as News2013 and TDT-1. News2013 refers to the English portion of the multilingual news data set produced in Miranda et al. (2018), which in turn derives from Rupnik et al. (2016). The News2013 data set contains a train set from 2013-12-18 to 2014-11-02, and a temporally disjoint test set from 2014-11-02 to 2015-08-25. TDT-1 refers to the pilot data set for the TDT initiative. We follow Saravanakumar et al. (2021) to generate its train set from 1994-07-09 to 1995-06-30 and a temporally overlapping test set from 1994-07-04 to 1995-06- $28^4.$ Dataset statistics are shown in Table 1.
<table>
<tbody>
<tr>
<th rowspan="3">Dataset</th>
<th>Trai</th>
<th>n</th>
<th>Tes</th>
<th>st</th>
</tr>
<tr>
<th>Trai $|D|$</th>
<th>n $|E|$ 1 1</th>
<th>Tes $|D|$</th>
<th>st $|E|$</th>
</tr>
<tr>
<th>$|D|$</th>
<th>$E$</th>
<th>$|D|$</th>
<th>$E$</th>
</tr>
<tr>
<td>News2013 (Miranda et al., 2018)</td>
<td>12,233</td>
<td>593</td>
<td>8,726</td>
<td>222</td>
</tr>
<tr>
<td>TDT-1 (Allan et al., 1998)</td>
<td>899</td>
<td>13</td>
<td>654</td>
<td>12</td>
</tr>
</tbody>
</table>
Table 1: Characteristics of the two TDT datasets. $|D|$ and $|E|$ denote the number of documents and topic events, respectively.
### 4.2.1. Experiment Setup
We use the online BatchHardTriplet algorithm to fine-tune BERT models for 1-5 epochs and use the best model. The training adopts a batch size of 32 and a max sequence length 230 to best fit into an 11GB GTX 1080 Ti GPU.
## 4.2.2. Time Encoding Strategies
We explored three ways of encoding temporal information into dense embeddings: (1) Date2Vec; (2) learned position embedding (LearnPE); (3) sinusoidal position embedding (SinPE). The first method directly transforms a date-time string into a dense vector. The second and third methods transform a document timestamp into a position by calculating the time span (e.g., in days) between the target document and the earliest document, generating a position embedding from this relative position. Previous studies (Vaswani et al.,2017; Wang and Chen, 2020) show that (2) and (3) perform similarly for language modeling. To integrate these time encodings into the TDT system, we use the E-BERT text encoder and apply the concatenation + multi- head- attention fusion method described in the method section (Figure 1).
· Date2Vec is a pre-trained time encoder based
on Time2Vec (Kazemi et al., 2019) that transforms a date and time into a dense vector, while preserving the time-specific characteristics (progression, periodicity, scale, etc). Kazemi et al. (2019) shows that Date2Vec is able to improve downstream NLP tasks. We use the pretrained Date2Vec model$^{5}$ released by Kazemi et al. (2019) in our experiments. We decided to update the Date2Vec module during training because we also tried to freeze the time module, leading to 2% decrease in F1 on News2013.
· Learned position embedding (LearnPE) is
a learned position embedding method used by many Transformer-based models (Devlin et al. 2018; Liu et al., 2019; Radford et al., 2019). It randomly initializes an embedding layer and updates the look-up table during training.
4.2. Retrospective TDT Experiments
We designed three experiments in retrospective TDT setting. First, we compare different representations with clustering algorithms that need the gold number of clusters (Table 2). We include two known baseline algorithms based on TF-IDF features including (a) the K-Means algorithm and (b) the augmented Group Average Clustering (GAC) (Yang et al., 1998). We then compare them against different BERT representations on GAC to understand the effect of representations. Second, we conduct experiments to compare these representations (TFIDF, BERT, E- BERT, three variants of T- E- BERT) with HDBSCAN, which does not require the cluster number as an input. In Table 2, we also conduct experiments with three time encoding strategies (Date2Vec, LearnPE, SinPE). At last, we run some studies on the time-text fusion strategy and time granularity option. With the most performant time encoding method SinPE, we compare four time-text fusion methods (Table 3) and five time granularity choices on News2013 (Table 4).
$^{4}$We follow their event splits, but get different train and
test instances from Saravanakumar et al. (2021).
$\cdot$Sinusoidal position embedding (SinPE) uses the sine and cosine functions of different frequencies to encode positions (Vaswani et al., 2017). As shown in the following equations, $i$ is the position index and $j$ is the dimension index. Compared to the learned PE method, this approach assigns a fixed embedding to each position.
$^{5}$https://github.com/ojus1/Date2Vec
|
[
[
0.118408203125,
0.07598876953125,
0.490234375,
0.07598876953125,
0.490234375,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.51318359375,
0.07598876953125,
0.71728515625,
0.07598876953125,
0.71728515625,
0.08905029296875,
0.51318359375,
0.08905029296875
],
[
0.12005615234375,
0.0914306640625,
0.2283935546875,
0.0914306640625,
0.2283935546875,
0.10272216796875,
0.12005615234375,
0.10272216796875
],
[
0.51318359375,
0.09857177734375,
0.884765625,
0.09857177734375,
0.884765625,
0.11224365234375,
0.51318359375,
0.11224365234375
],
[
0.51318359375,
0.1134033203125,
0.88330078125,
0.1134033203125,
0.88330078125,
0.1270751953125,
0.51318359375,
0.1270751953125
],
[
0.2216796875,
0.1258544921875,
0.382080078125,
0.1258544921875,
0.382080078125,
0.14013671875,
0.2216796875,
0.14013671875
],
[
0.51123046875,
0.1270751953125,
0.8818359375,
0.1270751953125,
0.8818359375,
0.14013671875,
0.51123046875,
0.14013671875
],
[
0.51318359375,
0.14013671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.15380859375,
0.51318359375,
0.15380859375
],
[
0.162841796875,
0.155029296875,
0.255126953125,
0.15673828125,
0.255126953125,
0.171630859375,
0.162841796875,
0.169189453125
],
[
0.5146484375,
0.155029296875,
0.70361328125,
0.155029296875,
0.70361328125,
0.168701171875,
0.5146484375,
0.168701171875
],
[
0.1158447265625,
0.1787109375,
0.489501953125,
0.179931640625,
0.489501953125,
0.19482421875,
0.1158447265625,
0.193603515625
],
[
0.12005615234375,
0.15673828125,
0.1695556640625,
0.15673828125,
0.1695556640625,
0.168701171875,
0.12005615234375,
0.168701171875
],
[
0.51123046875,
0.189453125,
0.779296875,
0.189453125,
0.779296875,
0.203125,
0.51123046875,
0.203125
],
[
0.12005615234375,
0.1953125,
0.48779296875,
0.1953125,
0.48779296875,
0.2060546875,
0.12005615234375,
0.2060546875
],
[
0.118408203125,
0.2078857421875,
0.489501953125,
0.2078857421875,
0.489501953125,
0.221435546875,
0.118408203125,
0.221435546875
],
[
0.51123046875,
0.213134765625,
0.88330078125,
0.213134765625,
0.88330078125,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.118408203125,
0.22265625,
0.490234375,
0.22265625,
0.490234375,
0.236328125,
0.118408203125,
0.236328125
],
[
0.5146484375,
0.22802734375,
0.88330078125,
0.22802734375,
0.88330078125,
0.2410888671875,
0.5146484375,
0.2410888671875
],
[
0.12176513671875,
0.2381591796875,
0.48779296875,
0.2381591796875,
0.48779296875,
0.248779296875,
0.12176513671875,
0.248779296875
],
[
0.51318359375,
0.2410888671875,
0.88330078125,
0.2410888671875,
0.88330078125,
0.254638671875,
0.51318359375,
0.254638671875
],
[
0.1175537109375,
0.25048828125,
0.48779296875,
0.25048828125,
0.48779296875,
0.26416015625,
0.1175537109375,
0.26416015625
],
[
0.51318359375,
0.255859375,
0.884765625,
0.255859375,
0.884765625,
0.26953125,
0.51318359375,
0.26953125
],
[
0.118408203125,
0.265380859375,
0.489501953125,
0.265380859375,
0.489501953125,
0.279052734375,
0.118408203125,
0.279052734375
],
[
0.515625,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.281982421875,
0.515625,
0.281982421875
],
[
0.12176513671875,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29150390625,
0.12176513671875,
0.29150390625
],
[
0.51318359375,
0.283935546875,
0.88330078125,
0.283935546875,
0.88330078125,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.12005615234375,
0.29345703125,
0.490234375,
0.29345703125,
0.490234375,
0.30712890625,
0.12005615234375,
0.30712890625
],
[
0.5146484375,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.311279296875,
0.5146484375,
0.311279296875
],
[
0.12005615234375,
0.30810546875,
0.490234375,
0.30810546875,
0.490234375,
0.32177734375,
0.12005615234375,
0.32177734375
],
[
0.5146484375,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.32666015625,
0.5146484375,
0.32666015625
],
[
0.1175537109375,
0.32177734375,
0.489501953125,
0.322509765625,
0.489501953125,
0.336181640625,
0.1175537109375,
0.3349609375
],
[
0.5146484375,
0.32666015625,
0.884765625,
0.32666015625,
0.884765625,
0.34033203125,
0.5146484375,
0.34033203125
],
[
0.12005615234375,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.12005615234375,
0.349853515625
],
[
0.5146484375,
0.341552734375,
0.884765625,
0.341552734375,
0.884765625,
0.355224609375,
0.5146484375,
0.355224609375
],
[
0.1175537109375,
0.3486328125,
0.447509765625,
0.349853515625,
0.447509765625,
0.363525390625,
0.1175537109375,
0.3623046875
],
[
0.51318359375,
0.355712890625,
0.884765625,
0.355712890625,
0.884765625,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.5146484375,
0.369384765625,
0.884765625,
0.369384765625,
0.884765625,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.341796875,
0.38134765625,
0.3837890625,
0.38134765625,
0.3837890625,
0.394287109375,
0.341796875,
0.394287109375
],
[
0.4306640625,
0.38134765625,
0.466796875,
0.38134765625,
0.466796875,
0.393798828125,
0.4306640625,
0.393798828125
],
[
0.124267578125,
0.388427734375,
0.1822509765625,
0.388427734375,
0.1822509765625,
0.399658203125,
0.124267578125,
0.399658203125
],
[
0.5146484375,
0.38427734375,
0.88330078125,
0.38427734375,
0.88330078125,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.333251953125,
0.394287109375,
0.359375,
0.394287109375,
0.359375,
0.407958984375,
0.333251953125,
0.407958984375
],
[
0.3720703125,
0.394287109375,
0.4013671875,
0.394287109375,
0.4013671875,
0.40673828125,
0.3720703125,
0.40673828125
],
[
0.418212890625,
0.394287109375,
0.443359375,
0.394287109375,
0.443359375,
0.407958984375,
0.418212890625,
0.407958984375
],
[
0.4609375,
0.394287109375,
0.484375,
0.394287109375,
0.484375,
0.40673828125,
0.4609375,
0.40673828125
],
[
0.51318359375,
0.39794921875,
0.884765625,
0.39794921875,
0.884765625,
0.410888671875,
0.51318359375,
0.410888671875
],
[
0.12176513671875,
0.412109375,
0.48779296875,
0.412109375,
0.48779296875,
0.427001953125,
0.12176513671875,
0.427001953125
],
[
0.5146484375,
0.412109375,
0.884765625,
0.412109375,
0.884765625,
0.42578125,
0.5146484375,
0.42578125
],
[
0.1234130859375,
0.42578125,
0.27197265625,
0.42578125,
0.27197265625,
0.439453125,
0.1234130859375,
0.439453125
],
[
0.32666015625,
0.42578125,
0.359375,
0.42578125,
0.359375,
0.438232421875,
0.32666015625,
0.438232421875
],
[
0.3779296875,
0.42578125,
0.403076171875,
0.42578125,
0.403076171875,
0.438232421875,
0.3779296875,
0.438232421875
],
[
0.411376953125,
0.427001953125,
0.447509765625,
0.427001953125,
0.447509765625,
0.438232421875,
0.411376953125,
0.438232421875
],
[
0.459228515625,
0.427001953125,
0.484375,
0.427001953125,
0.484375,
0.439453125,
0.459228515625,
0.439453125
],
[
0.5146484375,
0.427490234375,
0.88330078125,
0.427490234375,
0.88330078125,
0.438232421875,
0.5146484375,
0.438232421875
],
[
0.51123046875,
0.440673828125,
0.75390625,
0.440673828125,
0.75390625,
0.45361328125,
0.51123046875,
0.45361328125
],
[
0.118408203125,
0.45361328125,
0.48779296875,
0.45361328125,
0.48779296875,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.12005615234375,
0.468505859375,
0.489501953125,
0.468505859375,
0.489501953125,
0.482177734375,
0.12005615234375,
0.482177734375
],
[
0.53076171875,
0.46728515625,
0.88330078125,
0.46728515625,
0.88330078125,
0.48095703125,
0.53076171875,
0.48095703125
],
[
0.1175537109375,
0.482666015625,
0.307373046875,
0.482666015625,
0.307373046875,
0.496337890625,
0.1175537109375,
0.496337890625
],
[
0.54736328125,
0.482177734375,
0.88330078125,
0.482177734375,
0.88330078125,
0.495361328125,
0.54736328125,
0.495361328125
],
[
0.54736328125,
0.496337890625,
0.884765625,
0.496337890625,
0.884765625,
0.51025390625,
0.54736328125,
0.51025390625
],
[
0.54736328125,
0.51123046875,
0.884765625,
0.51025390625,
0.884765625,
0.52392578125,
0.54736328125,
0.52490234375
],
[
0.54736328125,
0.52490234375,
0.884765625,
0.52490234375,
0.884765625,
0.5380859375,
0.54736328125,
0.5380859375
],
[
0.16796875,
0.54150390625,
0.450927734375,
0.54150390625,
0.450927734375,
0.5546875,
0.16796875,
0.5546875
],
[
0.54736328125,
0.5380859375,
0.88330078125,
0.5380859375,
0.88330078125,
0.5517578125,
0.54736328125,
0.5517578125
],
[
0.156982421875,
0.5458984375,
0.166259765625,
0.5458984375,
0.166259765625,
0.55029296875,
0.156982421875,
0.55029296875
],
[
0.54833984375,
0.552734375,
0.884765625,
0.552734375,
0.884765625,
0.56640625,
0.54833984375,
0.56640625
],
[
0.12176513671875,
0.5419921875,
0.154541015625,
0.5419921875,
0.154541015625,
0.55419921875,
0.12176513671875,
0.55419921875
],
[
0.1175537109375,
0.56396484375,
0.489501953125,
0.5654296875,
0.489501953125,
0.580078125,
0.1175537109375,
0.5791015625
],
[
0.54736328125,
0.56787109375,
0.884765625,
0.56787109375,
0.884765625,
0.58056640625,
0.54736328125,
0.58056640625
],
[
0.1175537109375,
0.5791015625,
0.48779296875,
0.5791015625,
0.48779296875,
0.5927734375,
0.1175537109375,
0.5927734375
],
[
0.54833984375,
0.58203125,
0.884765625,
0.58203125,
0.884765625,
0.595703125,
0.54833984375,
0.595703125
],
[
0.12005615234375,
0.595703125,
0.48779296875,
0.595703125,
0.48779296875,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.5458984375,
0.59423828125,
0.88330078125,
0.59423828125,
0.88330078125,
0.60791015625,
0.5458984375,
0.60791015625
],
[
0.118408203125,
0.60791015625,
0.489501953125,
0.60791015625,
0.489501953125,
0.62158203125,
0.118408203125,
0.62158203125
],
[
0.54833984375,
0.60986328125,
0.88330078125,
0.60986328125,
0.88330078125,
0.62353515625,
0.54833984375,
0.62353515625
],
[
0.1175537109375,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63525390625,
0.1175537109375,
0.63525390625
],
[
0.54736328125,
0.62353515625,
0.88330078125,
0.62353515625,
0.88330078125,
0.63720703125,
0.54736328125,
0.63720703125
],
[
0.118408203125,
0.63525390625,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.118408203125,
0.6494140625
],
[
0.54833984375,
0.6396484375,
0.63232421875,
0.6396484375,
0.63232421875,
0.65087890625,
0.54833984375,
0.65087890625
],
[
0.118408203125,
0.6494140625,
0.489501953125,
0.65087890625,
0.489501953125,
0.6650390625,
0.118408203125,
0.66455078125
],
[
0.12005615234375,
0.6650390625,
0.489501953125,
0.6650390625,
0.489501953125,
0.6787109375,
0.12005615234375,
0.6787109375
],
[
0.53076171875,
0.6630859375,
0.88330078125,
0.66455078125,
0.88330078125,
0.6787109375,
0.53076171875,
0.677734375
],
[
0.12005615234375,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.6904296875,
0.12005615234375,
0.6904296875
],
[
0.54833984375,
0.68115234375,
0.8818359375,
0.68115234375,
0.8818359375,
0.69140625,
0.54833984375,
0.69140625
],
[
0.118408203125,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.118408203125,
0.70703125
],
[
0.54736328125,
0.693359375,
0.88330078125,
0.693359375,
0.88330078125,
0.70703125,
0.54736328125,
0.70703125
],
[
0.12005615234375,
0.7080078125,
0.489501953125,
0.7080078125,
0.489501953125,
0.7216796875,
0.12005615234375,
0.7216796875
],
[
0.54833984375,
0.70703125,
0.884765625,
0.70703125,
0.884765625,
0.72021484375,
0.54833984375,
0.72021484375
],
[
0.1175537109375,
0.72021484375,
0.486083984375,
0.72021484375,
0.486083984375,
0.73388671875,
0.1175537109375,
0.73388671875
],
[
0.5458984375,
0.72021484375,
0.884765625,
0.7216796875,
0.884765625,
0.736328125,
0.5458984375,
0.7353515625
],
[
0.1158447265625,
0.7353515625,
0.489501953125,
0.736328125,
0.489501953125,
0.75,
0.1158447265625,
0.7490234375
],
[
0.5458984375,
0.7353515625,
0.8505859375,
0.736328125,
0.8505859375,
0.75048828125,
0.5458984375,
0.75
],
[
0.12005615234375,
0.751953125,
0.48779296875,
0.751953125,
0.48779296875,
0.7626953125,
0.12005615234375,
0.7626953125
],
[
0.118408203125,
0.76416015625,
0.48779296875,
0.76416015625,
0.48779296875,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.5341796875,
0.7626953125,
0.88330078125,
0.7626953125,
0.88330078125,
0.775390625,
0.5341796875,
0.775390625
],
[
0.118408203125,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.5458984375,
0.775390625,
0.88330078125,
0.775390625,
0.88330078125,
0.7890625,
0.5458984375,
0.7890625
],
[
0.12005615234375,
0.7919921875,
0.489501953125,
0.7919921875,
0.489501953125,
0.8056640625,
0.12005615234375,
0.8056640625
],
[
0.5498046875,
0.7919921875,
0.88330078125,
0.7919921875,
0.88330078125,
0.802734375,
0.5498046875,
0.802734375
],
[
0.118408203125,
0.80712890625,
0.486083984375,
0.80712890625,
0.486083984375,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.54736328125,
0.8046875,
0.88671875,
0.8046875,
0.88671875,
0.818359375,
0.54736328125,
0.818359375
],
[
0.12005615234375,
0.82177734375,
0.489501953125,
0.82177734375,
0.489501953125,
0.8349609375,
0.12005615234375,
0.8349609375
],
[
0.54736328125,
0.8193359375,
0.884765625,
0.8193359375,
0.884765625,
0.8330078125,
0.54736328125,
0.8330078125
],
[
0.1175537109375,
0.83447265625,
0.48779296875,
0.8349609375,
0.48779296875,
0.849609375,
0.1175537109375,
0.8486328125
],
[
0.54736328125,
0.83447265625,
0.88330078125,
0.83447265625,
0.88330078125,
0.84716796875,
0.54736328125,
0.84716796875
],
[
0.118408203125,
0.849609375,
0.359375,
0.849609375,
0.359375,
0.86328125,
0.118408203125,
0.86328125
],
[
0.54736328125,
0.8486328125,
0.884765625,
0.8486328125,
0.884765625,
0.8623046875,
0.54736328125,
0.8623046875
],
[
0.54833984375,
0.8642578125,
0.70947265625,
0.8642578125,
0.70947265625,
0.87451171875,
0.54833984375,
0.87451171875
],
[
0.1376953125,
0.8798828125,
0.489501953125,
0.8818359375,
0.489501953125,
0.896484375,
0.1376953125,
0.89453125
],
[
0.12005615234375,
0.89794921875,
0.444091796875,
0.89794921875,
0.444091796875,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.5341796875,
0.8935546875,
0.77490234375,
0.89453125,
0.77490234375,
0.9091796875,
0.5341796875,
0.90771484375
]
] |
[
"training data points contain a positive label (create",
"4.2.1.ExperimentSetup",
"a new cluster).",
"We use the online BatchHardTriplet algorithm to",
"fine-tune BERT models for 1-5 epochs and use the",
"4.Experiments",
"best model. The training adopts a batch size of 32",
"and a max sequence length 230 to best fit into an",
"DataSets",
"11GB GTX 1080Ti GPU.",
"We use two data sets for our experiments, which we",
"4.1.",
"4.2.2.TimeEncodingStrategies",
"refertoasNews2013andTDT-1.News2013refers",
"to the English portion of the multilingual news data",
"We explored three ways of encoding temporal in",
"set produced in Miranda et al. (2018), which in turn",
"formation into dense embeddings:(1)Date2Vec;",
"derivesfromRupniketal.(2016).TheNews2013",
"(2) learned position embedding (LearnPE); (3) si-",
"data set contains a train set from 2013-12-18 to",
"nusoidal position embedding (SinPE). The first",
"2014-11-02, and a temporally disjoint test set from",
"method directly transforms a date-time string into",
"2014-11-02to2015-08-25.TDT-1referstothe",
"a dense vector. The second and third methods",
"pilot data set for the TDT initiative. We follow Sar-",
"transformadocumenttimestampintoapositionby",
"avanakumar et al. (2021) to generate its train set",
"calculating the time span (e.g., in days) between",
"from1994-07-09to1995-06-30andatemporally",
"the target document and the earliest document,",
"overlapping test set from 1994-07-04 to 1995-06-",
"generating a position embedding from this relative",
"284. Dataset statistics are shown in Table 1.",
"position. Previous studies (Vaswani et al., 2017;",
"Wang and Chen, 2020) show that (2) and (3) per-",
"Train",
"Test",
"Dataset",
"form similarly for language modeling. To integrate",
"[DI ",
"[E",
"[D]",
"[E]",
"these time encodings into the TDT system, we use",
"News2013 (Miranda etal., 2018) 12,233 593 8,726 222",
"the E-BERT text encoder and apply the concatena-",
"TDT-1 (Allan et al., 1998)",
"899",
"13",
"654",
"12",
"tion+multi-head-attentionfusionmethoddescribed",
"in the method section (Figure 1)",
"Table 1: Characteristics of the two TDT datasets.",
"[D| and E| denote the number of documents and",
"·Date2Vec is a pre-trained time encoder based",
"topic events,respectively",
"on Time2Vec (Kazemi et al., 2019) that trans-",
"forms a date and time into a dense vector, while",
"preserving the time-specific characteristics",
"(progression,periodicity,scale,etc).Kazemi",
"RetrospectiveTDTExperiments",
"et al. (2019) shows that Date2Vec is able to im",
"",
"prove downstream NLP tasks. We use the pre-",
"4.2.",
"We designed three experiments in retrospective",
"trainedDate2Vecmodel5 released byKazemi",
"TDT setting.First, we compare different represen",
"et al. (2019) in our experiments. We decided",
"tationswithclusteringalgorithmsthatneedthegold",
"to update the Date2Vec module during train",
"number of clusters (Table 2). We include two known",
"ing because we also tried to freeze the time",
"baseline algorithms based on TF-IDF features in-",
"module, leading to 2% decrease in F1 on",
"cluding (a) the K-Means algorithm and (b) the aug",
"News2013.",
"mented Group Average Clustering(GAC)(Yang",
"et al., 1998). We then compare them against differ-",
".Learned position embedding (LearnPE) is",
"entBERTrepresentationsonGACtounderstand",
"alearnedpositionembeddingmethodusedby",
"the effect of representations. Second, we conduct",
"many Transformer-based models (Devlin et al.",
"experiments to comparethese representations (TF-",
"2018; Liu et al., 2019; Radford et al., 2019). It",
"IDF, BERT, E-BERT, three variants of T-E-BERT)",
"randomly initializes an embedding layer and",
"with HDBSCAN, which does not require the cluster",
"updates the look-up table during training.",
"numberas aninput.In Table2,wealso conduct",
"experiments with three time encoding strategies",
"·Sinusoidal position embedding (SinPE)",
"(Date2Vec, LearnPE, SinPE). At last, we run some",
"uses the sine and cosine functions of differ-",
"studies on the time-text fusion strategy and time",
"entfrequenciestoencodepositions(Vaswani",
"granularity option. With the most performant time",
"et al., 2017). As shown in the following equa-",
"encoding method SinPE, we compare four time-text",
"tions, iis the position index and j is the di-",
"fusion methods (Table 3) and five time granularity",
"mension index. Compared to the learned PE",
"choices on News2013 (Table 4).",
"method, this approach assigns a fixed embed-",
"dingto eachposition.",
"4We follow their event splits, but get different train and",
"testinstancesfromSaravanakumaretal.(2021).",
"5https://github.com/ojus1/Date2Vec"
] |
{
"arxiv_id": "2112.06166",
"language": "en",
"page_idx": 3
}
|
||
2112.06166-en-4
|
2112.06166-en-4.jpg
|
<FigureHere>
Figure 3: Overview of the online TDT pipeline.
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th colspan="4">News2013</th>
<th colspan="4">$TDT-1$</th>
</tr>
<tr>
<th>Precision</th>
<th>Recall</th>
<th>$F1$</th>
<th>$CN$</th>
<th>Precision</th>
<th>Recall</th>
<th>$F1$</th>
<th>$CN$</th>
</tr>
<tr>
<td>TF-IDF+K-Means</td>
<td>87.00</td>
<td>53.94</td>
<td>66.60</td>
<td>222</td>
<td>84.23</td>
<td>69.01</td>
<td>75.87</td>
<td>12</td>
</tr>
<tr>
<td>TF-IDF+GAC</td>
<td>71.12</td>
<td>96.06</td>
<td>81.73</td>
<td>222</td>
<td>42.70</td>
<td>97.35</td>
<td>59.36</td>
<td>12</td>
</tr>
<tr>
<td>BERT+GAC</td>
<td>69.35</td>
<td>88.59</td>
<td>77.79</td>
<td>222</td>
<td>79.27</td>
<td>89.65</td>
<td>84.14</td>
<td>12</td>
</tr>
<tr>
<td>$E-BERT+GAC$</td>
<td>71.82</td>
<td>86.82</td>
<td>78.61</td>
<td>222</td>
<td>76.34</td>
<td>89.84</td>
<td>82.54</td>
<td>12</td>
</tr>
<tr>
<td>SinPE-E-BERT+CM+GAC</td>
<td>83.09</td>
<td>95.04</td>
<td>88.66</td>
<td>222</td>
<td>80.79</td>
<td>91.96</td>
<td>86.01</td>
<td>12</td>
</tr>
<tr>
<td>TF-IDF+HDBSCAN</td>
<td>88.45</td>
<td>58.58</td>
<td>70.48</td>
<td>301</td>
<td>93.05</td>
<td>89.34</td>
<td>91.16</td>
<td>1 1 </td>
</tr>
<tr>
<td>BERT+HDBSCAN</td>
<td>81.93</td>
<td>69.22</td>
<td>75.04</td>
<td>208</td>
<td>83.57</td>
<td>90.95</td>
<td>87.10</td>
<td>12</td>
</tr>
<tr>
<td>E-BERT+HDBSCAN</td>
<td>82.07</td>
<td>70.41</td>
<td>75.79</td>
<td>210</td>
<td>82.23</td>
<td>90.50</td>
<td>86.17</td>
<td>11 1 1 1</td>
</tr>
<tr>
<td>Date2Vec-E-BERT+CM+HDBSCAN</td>
<td>78.68</td>
<td>88.99</td>
<td>83.52</td>
<td>156</td>
<td>62.83</td>
<td>58.32</td>
<td>60.49</td>
<td>10</td>
</tr>
<tr>
<td>LearnPE-E-BERT+CM+HDBSCAN</td>
<td>83.33</td>
<td>54.24</td>
<td>65.71</td>
<td>252</td>
<td>79.51</td>
<td>59.36</td>
<td>67.97</td>
<td>13</td>
</tr>
<tr>
<td>SinPE-E-BERT+CM+HDBSCAN</td>
<td>90.18</td>
<td>89.90</td>
<td>90.04</td>
<td>186</td>
<td>86.79</td>
<td>90.96</td>
<td>88.83</td>
<td>9</td>
</tr>
</tbody>
</table>
Table 2: Retrospective TDT performance comparison of baselines with our T-E-BERT variants. CN stands for the predicted cluster number. We adopt the B-Cubed F measure for precision, recall and F1. Note that CN is automatically determined by HDBSCAN but set to be the gold number of clusters for the K-Means and Group Agglomerative Clustering (GAC) algorithms. GAC and HDBSCAN are deterministic, whereas K-Means is stochastic. We run K-Means five times and pick the best run for the table. The average F1 and standard deviation of the K-Means algorithm on two data sets are 67.86±1.00 (News2013) and $71. 06\pm 4. 81$ (TDT-1).
(2)
$$PE_{(i,2j)}=sin(i/10000^{2j/d_{model}})\\PE_{(i,2j+1)}=cos(i/10000^{2j/d_{model}})$$
achieves the best results relative to the other three methods (A, AM, and ACM). As a result, we decide to adopt the SinPE-E-BERT+ CM implementation for T-E-BERT in the later experiments
$\cdot$ additive (A):
(3)
## 4.2.3. Time-text Fusion Methods
$$e_f=\mathrm{POOL}(M_{text}\oplus M_{time})$$
$\cdot$ additive + multi-head attention (AM):
$$e_f=\mathrm{POOL}(\mathrm{ATT}(M_{text}\oplus M_{time}))$$
$\cdot$ concatenate + multi-head attention (CM):
We experiment with four methods to fuse the time and content embeddings. We repeat each document's position embedding "sequence length" times to form the position matrix $M_{time}.M_{text}$ indicates the text matrix from the last layer of the BERT model. $e_f$ indicates the fused document embedding. $\oplus$ indicates addition operation of two matrices. $[M_{1},M_{2}]$ indicates a concatenation operation of two matrices. POOL indicates the mean pooling layer. ATT indicates the multi-head attention layer. The third CM equation corresponds to the T-E-BERT diagram in Figure 1. Based on the results with the following four time-text fusion methods on News2013 (Table 3), the CM fusion method
$$e_f=\text{POOL}(\text{ATT}([M_{text},M_{time}]))$$
$\cdot$ additive + concatenate + multi-head atten-
## tion (ACM):
$$e_f=\text{POOL}(\text{ATT}([M_{text},M_{time}]\oplus M_{time}))$$
|
[
[
0.46337890625,
0.0843505859375,
0.5263671875,
0.0843505859375,
0.5263671875,
0.09442138671875,
0.46337890625,
0.09442138671875
],
[
0.55908203125,
0.08489990234375,
0.6767578125,
0.0843505859375,
0.6767578125,
0.0955810546875,
0.55908203125,
0.0968017578125
],
[
0.333251953125,
0.097412109375,
0.3720703125,
0.097412109375,
0.3720703125,
0.10986328125,
0.333251953125,
0.10986328125
],
[
0.465087890625,
0.0968017578125,
0.5263671875,
0.0968017578125,
0.5263671875,
0.10809326171875,
0.465087890625,
0.10809326171875
],
[
0.32421875,
0.1092529296875,
0.3779296875,
0.1092529296875,
0.3779296875,
0.12054443359375,
0.32421875,
0.12054443359375
],
[
0.476806640625,
0.10986328125,
0.5146484375,
0.10986328125,
0.5146484375,
0.1175537109375,
0.476806640625,
0.1175537109375
],
[
0.490234375,
0.1258544921875,
0.496337890625,
0.12469482421875,
0.49951171875,
0.1318359375,
0.49365234375,
0.133056640625
],
[
0.204833984375,
0.133056640625,
0.246826171875,
0.133056640625,
0.246826171875,
0.144287109375,
0.204833984375,
0.144287109375
],
[
0.58447265625,
0.1383056640625,
0.6337890625,
0.1383056640625,
0.6337890625,
0.1484375,
0.58447265625,
0.1484375
],
[
0.6650390625,
0.1383056640625,
0.79248046875,
0.1383056640625,
0.79248046875,
0.1484375,
0.6650390625,
0.1484375
],
[
0.32421875,
0.1455078125,
0.3837890625,
0.1455078125,
0.3837890625,
0.15673828125,
0.32421875,
0.15673828125
],
[
0.580078125,
0.15087890625,
0.63818359375,
0.15087890625,
0.63818359375,
0.162109375,
0.580078125,
0.162109375
],
[
0.58837890625,
0.163330078125,
0.62744140625,
0.163330078125,
0.62744140625,
0.17578125,
0.58837890625,
0.17578125
],
[
0.32080078125,
0.1787109375,
0.382080078125,
0.1787109375,
0.382080078125,
0.193603515625,
0.32080078125,
0.193603515625
],
[
0.57861328125,
0.179931640625,
0.599609375,
0.179931640625,
0.599609375,
0.193603515625,
0.57861328125,
0.193603515625
],
[
0.75732421875,
0.174560546875,
0.78076171875,
0.174560546875,
0.78076171875,
0.18701171875,
0.75732421875,
0.18701171875
],
[
0.20654296875,
0.1817626953125,
0.270263671875,
0.1817626953125,
0.270263671875,
0.193603515625,
0.20654296875,
0.193603515625
],
[
0.32568359375,
0.191162109375,
0.37451171875,
0.189453125,
0.376220703125,
0.201904296875,
0.32666015625,
0.2037353515625
],
[
0.366943359375,
0.1859130859375,
0.378662109375,
0.18701171875,
0.3779296875,
0.1923828125,
0.365234375,
0.191162109375
],
[
0.60546875,
0.1923828125,
0.609375,
0.1923828125,
0.609375,
0.19775390625,
0.60546875,
0.19775390625
],
[
0.432373046875,
0.204833984375,
0.49951171875,
0.204833984375,
0.49951171875,
0.2161865234375,
0.432373046875,
0.2161865234375
],
[
0.50537109375,
0.204833984375,
0.5810546875,
0.204833984375,
0.5810546875,
0.2161865234375,
0.50537109375,
0.2161865234375
],
[
0.61279296875,
0.204833984375,
0.68115234375,
0.204833984375,
0.68115234375,
0.2161865234375,
0.61279296875,
0.2161865234375
],
[
0.6884765625,
0.204833984375,
0.70361328125,
0.2060546875,
0.7021484375,
0.215576171875,
0.68701171875,
0.213134765625
],
[
0.7119140625,
0.201904296875,
0.7958984375,
0.2037353515625,
0.7958984375,
0.21728515625,
0.7119140625,
0.215576171875
],
[
0.298828125,
0.2286376953125,
0.40478515625,
0.2286376953125,
0.40478515625,
0.240478515625,
0.298828125,
0.240478515625
],
[
0.5517578125,
0.2286376953125,
0.62451171875,
0.2286376953125,
0.62451171875,
0.240478515625,
0.5517578125,
0.240478515625
],
[
0.330078125,
0.257080078125,
0.6708984375,
0.257080078125,
0.6708984375,
0.270751953125,
0.330078125,
0.270751953125
],
[
0.1259765625,
0.2880859375,
0.177978515625,
0.2880859375,
0.177978515625,
0.302734375,
0.1259765625,
0.302734375
],
[
0.480224609375,
0.2880859375,
0.55908203125,
0.2880859375,
0.55908203125,
0.302734375,
0.480224609375,
0.302734375
],
[
0.74072265625,
0.287353515625,
0.78857421875,
0.287353515625,
0.78857421875,
0.302734375,
0.74072265625,
0.302734375
],
[
0.403076171875,
0.2998046875,
0.470947265625,
0.3017578125,
0.47021484375,
0.317626953125,
0.4013671875,
0.3154296875
],
[
0.64990234375,
0.2998046875,
0.71875,
0.3017578125,
0.71728515625,
0.317626953125,
0.6484375,
0.3154296875
],
[
0.8017578125,
0.2998046875,
0.82861328125,
0.2998046875,
0.82861328125,
0.317626953125,
0.8017578125,
0.317626953125
],
[
0.8505859375,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.317626953125,
0.8505859375,
0.317626953125
],
[
0.5556640625,
0.3017578125,
0.57861328125,
0.3017578125,
0.57861328125,
0.3154296875,
0.5556640625,
0.3154296875
],
[
0.484375,
0.3017578125,
0.5341796875,
0.3017578125,
0.5341796875,
0.31640625,
0.484375,
0.31640625
],
[
0.60791015625,
0.3017578125,
0.634765625,
0.3017578125,
0.634765625,
0.3154296875,
0.60791015625,
0.3154296875
],
[
0.7314453125,
0.3017578125,
0.78076171875,
0.3017578125,
0.78076171875,
0.31640625,
0.7314453125,
0.31640625
],
[
0.54736328125,
0.319580078125,
0.591796875,
0.319580078125,
0.591796875,
0.334228515625,
0.54736328125,
0.334228515625
],
[
0.66162109375,
0.319580078125,
0.7060546875,
0.319580078125,
0.7060546875,
0.3349609375,
0.66162109375,
0.3349609375
],
[
0.79443359375,
0.319580078125,
0.8369140625,
0.319580078125,
0.8369140625,
0.3349609375,
0.79443359375,
0.3349609375
],
[
0.4130859375,
0.320556640625,
0.45751953125,
0.318359375,
0.459228515625,
0.333251953125,
0.414794921875,
0.3349609375
],
[
0.1275634765625,
0.320556640625,
0.2568359375,
0.320556640625,
0.2568359375,
0.334228515625,
0.1275634765625,
0.334228515625
],
[
0.489501953125,
0.320556640625,
0.53076171875,
0.320556640625,
0.53076171875,
0.334228515625,
0.489501953125,
0.334228515625
],
[
0.60546875,
0.320556640625,
0.63623046875,
0.320556640625,
0.63623046875,
0.333251953125,
0.60546875,
0.333251953125
],
[
0.736328125,
0.320556640625,
0.77490234375,
0.320556640625,
0.77490234375,
0.333251953125,
0.736328125,
0.333251953125
],
[
0.85546875,
0.320556640625,
0.87744140625,
0.320556640625,
0.87744140625,
0.334228515625,
0.85546875,
0.334228515625
],
[
0.1275634765625,
0.333251953125,
0.2301025390625,
0.333251953125,
0.2301025390625,
0.347412109375,
0.1275634765625,
0.347412109375
],
[
0.414794921875,
0.333251953125,
0.459228515625,
0.333251953125,
0.459228515625,
0.347412109375,
0.414794921875,
0.347412109375
],
[
0.486083984375,
0.333251953125,
0.5322265625,
0.333251953125,
0.5322265625,
0.347412109375,
0.486083984375,
0.347412109375
],
[
0.5458984375,
0.333251953125,
0.591796875,
0.333251953125,
0.591796875,
0.347412109375,
0.5458984375,
0.347412109375
],
[
0.603515625,
0.333251953125,
0.63818359375,
0.333251953125,
0.63818359375,
0.3486328125,
0.603515625,
0.3486328125
],
[
0.66162109375,
0.333251953125,
0.7060546875,
0.333251953125,
0.7060546875,
0.3486328125,
0.66162109375,
0.3486328125
],
[
0.73291015625,
0.333251953125,
0.779296875,
0.333251953125,
0.779296875,
0.347412109375,
0.73291015625,
0.347412109375
],
[
0.79248046875,
0.333251953125,
0.8388671875,
0.333251953125,
0.8388671875,
0.3486328125,
0.79248046875,
0.3486328125
],
[
0.85400390625,
0.333251953125,
0.8759765625,
0.333251953125,
0.8759765625,
0.347412109375,
0.85400390625,
0.347412109375
],
[
0.414794921875,
0.3466796875,
0.4609375,
0.3466796875,
0.4609375,
0.361083984375,
0.414794921875,
0.361083984375
],
[
0.48779296875,
0.3466796875,
0.5322265625,
0.3466796875,
0.5322265625,
0.361083984375,
0.48779296875,
0.361083984375
],
[
0.54736328125,
0.3466796875,
0.591796875,
0.3466796875,
0.591796875,
0.361083984375,
0.54736328125,
0.361083984375
],
[
0.66162109375,
0.3466796875,
0.7060546875,
0.3466796875,
0.7060546875,
0.361083984375,
0.66162109375,
0.361083984375
],
[
0.73291015625,
0.3466796875,
0.779296875,
0.3466796875,
0.779296875,
0.361083984375,
0.73291015625,
0.361083984375
],
[
0.79248046875,
0.3466796875,
0.84033203125,
0.3466796875,
0.84033203125,
0.361083984375,
0.79248046875,
0.361083984375
],
[
0.85400390625,
0.3466796875,
0.8759765625,
0.3466796875,
0.8759765625,
0.361083984375,
0.85400390625,
0.361083984375
],
[
0.60546875,
0.347412109375,
0.63623046875,
0.347412109375,
0.63623046875,
0.35986328125,
0.60546875,
0.35986328125
],
[
0.1292724609375,
0.347412109375,
0.2216796875,
0.347412109375,
0.2216796875,
0.359375,
0.1292724609375,
0.359375
],
[
0.414794921875,
0.359375,
0.459228515625,
0.359375,
0.459228515625,
0.37353515625,
0.414794921875,
0.37353515625
],
[
0.48779296875,
0.359375,
0.5322265625,
0.359375,
0.5322265625,
0.374755859375,
0.48779296875,
0.374755859375
],
[
0.603515625,
0.359375,
0.63818359375,
0.359375,
0.63818359375,
0.374755859375,
0.603515625,
0.374755859375
],
[
0.66162109375,
0.359375,
0.7060546875,
0.359375,
0.7060546875,
0.374755859375,
0.66162109375,
0.374755859375
],
[
0.73291015625,
0.359375,
0.779296875,
0.359375,
0.779296875,
0.374755859375,
0.73291015625,
0.374755859375
],
[
0.79248046875,
0.359375,
0.8388671875,
0.359375,
0.8388671875,
0.374755859375,
0.79248046875,
0.374755859375
],
[
0.85546875,
0.35986328125,
0.8759765625,
0.35986328125,
0.8759765625,
0.374755859375,
0.85546875,
0.374755859375
],
[
0.54833984375,
0.35986328125,
0.59033203125,
0.35986328125,
0.59033203125,
0.37353515625,
0.54833984375,
0.37353515625
],
[
0.1292724609375,
0.361083984375,
0.235107421875,
0.361083984375,
0.235107421875,
0.372314453125,
0.1292724609375,
0.372314453125
],
[
0.1259765625,
0.372314453125,
0.32421875,
0.371826171875,
0.32421875,
0.385986328125,
0.1259765625,
0.38720703125
],
[
0.414794921875,
0.372314453125,
0.459228515625,
0.372314453125,
0.459228515625,
0.38720703125,
0.414794921875,
0.38720703125
],
[
0.54736328125,
0.372314453125,
0.591796875,
0.372314453125,
0.591796875,
0.38720703125,
0.54736328125,
0.38720703125
],
[
0.66162109375,
0.372314453125,
0.7060546875,
0.372314453125,
0.7060546875,
0.388427734375,
0.66162109375,
0.388427734375
],
[
0.73291015625,
0.372314453125,
0.77734375,
0.372314453125,
0.77734375,
0.388427734375,
0.73291015625,
0.388427734375
],
[
0.79248046875,
0.372314453125,
0.8369140625,
0.372314453125,
0.8369140625,
0.388427734375,
0.79248046875,
0.388427734375
],
[
0.85546875,
0.372314453125,
0.87744140625,
0.372314453125,
0.87744140625,
0.38720703125,
0.85546875,
0.38720703125
],
[
0.60546875,
0.37353515625,
0.63623046875,
0.37353515625,
0.63623046875,
0.388427734375,
0.60546875,
0.388427734375
],
[
0.489501953125,
0.37353515625,
0.53076171875,
0.37353515625,
0.53076171875,
0.38720703125,
0.489501953125,
0.38720703125
],
[
0.1275634765625,
0.391357421875,
0.270263671875,
0.391357421875,
0.270263671875,
0.40625,
0.1275634765625,
0.40625
],
[
0.414794921875,
0.391357421875,
0.459228515625,
0.391357421875,
0.459228515625,
0.40625,
0.414794921875,
0.40625
],
[
0.54736328125,
0.391357421875,
0.591796875,
0.391357421875,
0.591796875,
0.40625,
0.54736328125,
0.40625
],
[
0.66162109375,
0.391357421875,
0.7060546875,
0.391357421875,
0.7060546875,
0.40625,
0.66162109375,
0.40625
],
[
0.73486328125,
0.391357421875,
0.779296875,
0.391357421875,
0.779296875,
0.40625,
0.73486328125,
0.40625
],
[
0.79248046875,
0.391357421875,
0.8388671875,
0.391357421875,
0.8388671875,
0.40625,
0.79248046875,
0.40625
],
[
0.60546875,
0.392578125,
0.63623046875,
0.392578125,
0.63623046875,
0.40673828125,
0.60546875,
0.40673828125
],
[
0.489501953125,
0.392578125,
0.53076171875,
0.392578125,
0.53076171875,
0.40625,
0.489501953125,
0.40625
],
[
0.85400390625,
0.392578125,
0.8759765625,
0.392578125,
0.8759765625,
0.40625,
0.85400390625,
0.40625
],
[
0.1275634765625,
0.405029296875,
0.2626953125,
0.405029296875,
0.2626953125,
0.418701171875,
0.1275634765625,
0.418701171875
],
[
0.414794921875,
0.405029296875,
0.4609375,
0.405029296875,
0.4609375,
0.419189453125,
0.414794921875,
0.419189453125
],
[
0.54736328125,
0.405029296875,
0.591796875,
0.405029296875,
0.591796875,
0.419189453125,
0.54736328125,
0.419189453125
],
[
0.66162109375,
0.405029296875,
0.7060546875,
0.405029296875,
0.7060546875,
0.419189453125,
0.66162109375,
0.419189453125
],
[
0.73291015625,
0.405029296875,
0.779296875,
0.405029296875,
0.779296875,
0.419189453125,
0.73291015625,
0.419189453125
],
[
0.79248046875,
0.405029296875,
0.8388671875,
0.405029296875,
0.8388671875,
0.419189453125,
0.79248046875,
0.419189453125
],
[
0.85400390625,
0.405029296875,
0.8759765625,
0.405029296875,
0.8759765625,
0.419189453125,
0.85400390625,
0.419189453125
],
[
0.60546875,
0.40625,
0.63623046875,
0.40625,
0.63623046875,
0.418701171875,
0.60546875,
0.418701171875
],
[
0.489501953125,
0.40625,
0.53076171875,
0.40625,
0.53076171875,
0.419189453125,
0.489501953125,
0.419189453125
],
[
0.414794921875,
0.41748046875,
0.459228515625,
0.41748046875,
0.459228515625,
0.431640625,
0.414794921875,
0.431640625
],
[
0.54736328125,
0.41748046875,
0.591796875,
0.41748046875,
0.591796875,
0.431640625,
0.54736328125,
0.431640625
],
[
0.603515625,
0.41748046875,
0.63818359375,
0.41748046875,
0.63818359375,
0.432861328125,
0.603515625,
0.432861328125
],
[
0.66162109375,
0.41748046875,
0.7060546875,
0.41748046875,
0.7060546875,
0.431640625,
0.66162109375,
0.431640625
],
[
0.73291015625,
0.41748046875,
0.77734375,
0.41748046875,
0.77734375,
0.431640625,
0.73291015625,
0.431640625
],
[
0.79248046875,
0.41748046875,
0.8388671875,
0.41748046875,
0.8388671875,
0.431640625,
0.79248046875,
0.431640625
],
[
0.85400390625,
0.418701171875,
0.8759765625,
0.418701171875,
0.8759765625,
0.431640625,
0.85400390625,
0.431640625
],
[
0.489501953125,
0.418701171875,
0.53076171875,
0.418701171875,
0.53076171875,
0.431640625,
0.489501953125,
0.431640625
],
[
0.1292724609375,
0.419189453125,
0.27783203125,
0.419189453125,
0.27783203125,
0.429931640625,
0.1292724609375,
0.429931640625
],
[
0.486083984375,
0.429931640625,
0.5322265625,
0.429931640625,
0.5322265625,
0.44482421875,
0.486083984375,
0.44482421875
],
[
0.54736328125,
0.429931640625,
0.591796875,
0.429931640625,
0.591796875,
0.44482421875,
0.54736328125,
0.44482421875
],
[
0.603515625,
0.429931640625,
0.63818359375,
0.429931640625,
0.63818359375,
0.4453125,
0.603515625,
0.4453125
],
[
0.66162109375,
0.429931640625,
0.7060546875,
0.429931640625,
0.7060546875,
0.4453125,
0.66162109375,
0.4453125
],
[
0.73291015625,
0.429931640625,
0.779296875,
0.429931640625,
0.779296875,
0.4453125,
0.73291015625,
0.4453125
],
[
0.79248046875,
0.429931640625,
0.8388671875,
0.429931640625,
0.8388671875,
0.4453125,
0.79248046875,
0.4453125
],
[
0.418212890625,
0.43115234375,
0.459228515625,
0.43115234375,
0.459228515625,
0.44482421875,
0.418212890625,
0.44482421875
],
[
0.1275634765625,
0.43115234375,
0.3896484375,
0.43115234375,
0.3896484375,
0.44482421875,
0.1275634765625,
0.44482421875
],
[
0.85400390625,
0.43115234375,
0.8759765625,
0.43115234375,
0.8759765625,
0.44482421875,
0.85400390625,
0.44482421875
],
[
0.1275634765625,
0.443603515625,
0.382080078125,
0.443603515625,
0.382080078125,
0.457275390625,
0.1275634765625,
0.457275390625
],
[
0.414794921875,
0.443603515625,
0.459228515625,
0.443603515625,
0.459228515625,
0.457763671875,
0.414794921875,
0.457763671875
],
[
0.486083984375,
0.4423828125,
0.53076171875,
0.4423828125,
0.53076171875,
0.457763671875,
0.486083984375,
0.457763671875
],
[
0.73291015625,
0.443603515625,
0.779296875,
0.443603515625,
0.779296875,
0.457763671875,
0.73291015625,
0.457763671875
],
[
0.79248046875,
0.443603515625,
0.8388671875,
0.443603515625,
0.8388671875,
0.457763671875,
0.79248046875,
0.457763671875
],
[
0.85400390625,
0.443603515625,
0.8759765625,
0.443603515625,
0.8759765625,
0.457763671875,
0.85400390625,
0.457763671875
],
[
0.603515625,
0.44482421875,
0.63623046875,
0.44482421875,
0.63623046875,
0.457275390625,
0.603515625,
0.457275390625
],
[
0.54833984375,
0.44482421875,
0.59033203125,
0.44482421875,
0.59033203125,
0.457763671875,
0.54833984375,
0.457763671875
],
[
0.6630859375,
0.44482421875,
0.705078125,
0.44482421875,
0.705078125,
0.457763671875,
0.6630859375,
0.457763671875
],
[
0.1275634765625,
0.4560546875,
0.365234375,
0.4560546875,
0.365234375,
0.4697265625,
0.1275634765625,
0.4697265625
],
[
0.486083984375,
0.4560546875,
0.53076171875,
0.4560546875,
0.53076171875,
0.47021484375,
0.486083984375,
0.47021484375
],
[
0.54736328125,
0.4560546875,
0.591796875,
0.4560546875,
0.591796875,
0.47021484375,
0.54736328125,
0.47021484375
],
[
0.603515625,
0.4560546875,
0.63818359375,
0.4560546875,
0.63818359375,
0.471435546875,
0.603515625,
0.471435546875
],
[
0.66162109375,
0.4560546875,
0.7060546875,
0.4560546875,
0.7060546875,
0.471435546875,
0.66162109375,
0.471435546875
],
[
0.73291015625,
0.4560546875,
0.77734375,
0.4560546875,
0.77734375,
0.471435546875,
0.73291015625,
0.471435546875
],
[
0.79248046875,
0.4560546875,
0.8388671875,
0.4560546875,
0.8388671875,
0.471435546875,
0.79248046875,
0.471435546875
],
[
0.8564453125,
0.4560546875,
0.875,
0.4560546875,
0.875,
0.471435546875,
0.8564453125,
0.471435546875
],
[
0.1175537109375,
0.48388671875,
0.884765625,
0.485107421875,
0.884765625,
0.49951171875,
0.1175537109375,
0.498779296875
],
[
0.418212890625,
0.457275390625,
0.459228515625,
0.457275390625,
0.459228515625,
0.47021484375,
0.418212890625,
0.47021484375
],
[
0.118408203125,
0.49951171875,
0.884765625,
0.49951171875,
0.884765625,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.118408203125,
0.5126953125,
0.884765625,
0.51318359375,
0.884765625,
0.52783203125,
0.118408203125,
0.52685546875
],
[
0.118408203125,
0.52783203125,
0.88330078125,
0.52783203125,
0.88330078125,
0.54150390625,
0.118408203125,
0.54150390625
],
[
0.118408203125,
0.54150390625,
0.880859375,
0.54150390625,
0.880859375,
0.5546875,
0.118408203125,
0.5546875
],
[
0.1175537109375,
0.55712890625,
0.88330078125,
0.5556640625,
0.88330078125,
0.5693359375,
0.1175537109375,
0.57080078125
],
[
0.118408203125,
0.57080078125,
0.274658203125,
0.57080078125,
0.274658203125,
0.58447265625,
0.118408203125,
0.58447265625
],
[
0.5146484375,
0.60986328125,
0.884765625,
0.60986328125,
0.884765625,
0.62353515625,
0.5146484375,
0.62353515625
],
[
0.183837890625,
0.6259765625,
0.42236328125,
0.62158203125,
0.42236328125,
0.6396484375,
0.183837890625,
0.6435546875
],
[
0.466796875,
0.62744140625,
0.489501953125,
0.62744140625,
0.489501953125,
0.64013671875,
0.466796875,
0.64013671875
],
[
0.51123046875,
0.62353515625,
0.88330078125,
0.62353515625,
0.88330078125,
0.63720703125,
0.51123046875,
0.63720703125
],
[
0.51318359375,
0.63818359375,
0.884765625,
0.63818359375,
0.884765625,
0.65283203125,
0.51318359375,
0.65283203125
],
[
0.173828125,
0.6494140625,
0.432373046875,
0.6484375,
0.432373046875,
0.67041015625,
0.173828125,
0.67138671875
],
[
0.466796875,
0.65478515625,
0.490234375,
0.65478515625,
0.490234375,
0.66845703125,
0.466796875,
0.66845703125
],
[
0.51123046875,
0.65185546875,
0.85205078125,
0.65283203125,
0.85205078125,
0.66748046875,
0.51123046875,
0.66650390625
],
[
0.5341796875,
0.6767578125,
0.64404296875,
0.6767578125,
0.64404296875,
0.6884765625,
0.5341796875,
0.6884765625
],
[
0.12005615234375,
0.6884765625,
0.17041015625,
0.6884765625,
0.17041015625,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.173828125,
0.68701171875,
0.387939453125,
0.68701171875,
0.387939453125,
0.70068359375,
0.173828125,
0.70068359375
],
[
0.609375,
0.70068359375,
0.8193359375,
0.70068359375,
0.8193359375,
0.71533203125,
0.609375,
0.71533203125
],
[
0.12005615234375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.72265625,
0.12005615234375,
0.72265625
],
[
0.118408203125,
0.7255859375,
0.489501953125,
0.7255859375,
0.489501953125,
0.7392578125,
0.118408203125,
0.7392578125
],
[
0.5322265625,
0.72998046875,
0.8369140625,
0.72998046875,
0.8369140625,
0.74365234375,
0.5322265625,
0.74365234375
],
[
0.118408203125,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.7529296875,
0.118408203125,
0.7529296875
],
[
0.118408203125,
0.75439453125,
0.490234375,
0.75439453125,
0.490234375,
0.76708984375,
0.118408203125,
0.76708984375
],
[
0.5859375,
0.75439453125,
0.8427734375,
0.75439453125,
0.8427734375,
0.7685546875,
0.5859375,
0.7685546875
],
[
0.12005615234375,
0.76953125,
0.48779296875,
0.76953125,
0.48779296875,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.118408203125,
0.7822265625,
0.490234375,
0.7822265625,
0.490234375,
0.7958984375,
0.118408203125,
0.7958984375
],
[
0.5322265625,
0.78369140625,
0.87158203125,
0.78369140625,
0.87158203125,
0.79736328125,
0.5322265625,
0.79736328125
],
[
0.12005615234375,
0.79638671875,
0.489501953125,
0.79638671875,
0.489501953125,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.118408203125,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.587890625,
0.80859375,
0.84033203125,
0.80859375,
0.84033203125,
0.82373046875,
0.587890625,
0.82373046875
],
[
0.12176513671875,
0.826171875,
0.48779296875,
0.826171875,
0.48779296875,
0.8359375,
0.12176513671875,
0.8359375
],
[
0.118408203125,
0.83837890625,
0.490234375,
0.83837890625,
0.490234375,
0.8515625,
0.118408203125,
0.8515625
],
[
0.5322265625,
0.83740234375,
0.88671875,
0.83740234375,
0.88671875,
0.8515625,
0.5322265625,
0.8515625
],
[
0.118408203125,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.118408203125,
0.8662109375
],
[
0.5439453125,
0.849609375,
0.6396484375,
0.8515625,
0.6396484375,
0.8662109375,
0.5439453125,
0.8642578125
],
[
0.118408203125,
0.86767578125,
0.489501953125,
0.86767578125,
0.489501953125,
0.88134765625,
0.118408203125,
0.88134765625
],
[
0.118408203125,
0.88134765625,
0.490234375,
0.88134765625,
0.490234375,
0.89453125,
0.118408203125,
0.89453125
],
[
0.55517578125,
0.87744140625,
0.875,
0.87744140625,
0.875,
0.892578125,
0.55517578125,
0.892578125
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
]
] |
[
"Weighted",
"Choosea cluster C*",
"TF-IDF",
"Similarity",
"Encoders",
"Model",
"",
"",
"Cluster",
"Createa newcluster?",
"T-E-BERT",
"Creation",
"Mode",
"Timestamr",
"no",
"yes",
"News Doc",
"Extracto",
"m",
"",
"ClusterC",
"ClusterC,",
"ClusterC.",
"+",
"ClusterCn+1",
"Feature Extraction",
"ClusterPool",
"Figure 3:Overview of the online TDT pipeline.",
"Model",
"News2013",
"TDT-1",
"Precision",
"Precision",
"F1",
"CN",
"F1",
"Recall",
"CN",
"Recall",
"66.60",
"84.23",
"75.87",
"87.00",
"TF-IDF + K-Means",
"53.94",
"222",
"69.01",
"12",
"TF-IDF + GAC",
"71.12",
"96.06",
"81.73",
"222",
"42.70",
"97.35",
"59.36",
"12",
"69.35",
"88.59",
"77.79",
"79.27",
"89.65",
"84.14",
"12",
"222",
"BERT +GAC",
"71.82",
"86.82",
"222",
"76.34",
"89.84",
"82.54",
"12",
"78.61",
"E-BERT + GAC",
"SinPE-E-BERT + CM + GAC",
"83.09",
"88.66",
"80.79",
"91.96",
"86.01",
"12",
"222",
"95.04",
"TF-IDF + HDBSCAN",
"88.45",
"70.48",
"93.05",
"89.34",
"91.16",
"301",
"58.58",
"11",
"BERT + HDBSCAN",
"81.93",
"75.04",
"83.57",
"90.95",
"87.10",
"12",
"208",
"69.22",
"82.07",
"75.79",
"210",
"82.23",
"90.50",
"86.17",
"11",
"70.41",
"E-BERT +HDBSCAN",
"88.99",
"83.52",
"156",
"62.83",
"58.32",
"60.49",
"78.68",
"Date2Vec-E-BERT + CM + HDBSCAN",
"10",
"LearnPE-E-BERT + CM + HDBSCAN",
"83.33",
"54.24",
"59.36",
"67.97",
"13",
"252",
"65.71",
"79.51",
"SinPE-E-BERT + CM + HDBSCAN",
"89.90",
"90.04",
"186",
"86.79",
"90.96",
"88.83",
"9",
"Table2:Retrospective TDT performance comparison of baselines with our T-E-BERT variants.CN stands",
"90.18",
"for the predicted cluster number. We adopt the B-Cubed F measure for precision, recall and F1. Note that",
"CN is automatically determined by HDBSCAN but set to be the gold number of clusters for the K-Means",
"and Group Agglomerative Clustering (GAC) algorithms. GAC and HDBSCAN are deterministic, whereas",
"K-Means is stochastic. We run K-Means five times and pick the best run for the table.The average F1",
"and standard deviation of the K-Means algorithm on two data sets are 67.86 ± 1.00 (News2013) and",
"71.06 ± 4.81 (TDT-1).",
"achieves the best results relative to the other three",
"PE(c,25) = sin(i/100002j/dmodet)",
"(2)",
"methods (A, AM, and ACM). As a result, we decide",
"toadopt the SinPE-E-BERT+CM implementa-",
"PE(s,2j+1) = cs(i/1002i/dmodel)",
"(3)",
"tionforT-E-BERT inthelaterexperiments.",
".additive (A):",
"4.2.3.",
"Time-text Fusion Methods",
"ef = POOL(Mtert @ Mtime)",
"Weexperimentwithfourmethodstofusethetime",
"and content embeddings. We repeat each doc-",
": additive + multi-head attention (AM):",
"ument's position embedding \"sequence length\"",
"times to form the position matrix Mtime. Mtert in-",
"ef = POOL(ATT(Mtert @ Mtime))",
"dicatesthetextmatrixfromthelastlayerofthe",
"BERT model. ef indicates the fused document",
"concatenate +multi-head attention (CM):",
"embedding. @ indicates addition operation of two",
"matrices. [Mi, M2] indicates a concatenation oper",
"ef = POOL(ATT([Mtert, Mtimel))",
"ationoftwomatrices.POOLindicatesthemean",
"pooling layer. ATT indicates the multi-head atten-",
"· additive + concatenate + multi-head atten-",
"tion layer. The third CM equation corresponds to",
"tion (ACM):",
"the T-E-BERT diagram in Figure 1. Based on the",
"results withthefollowingfour time-textfusionmeth-",
"ef = POOL(ATT([Mtert, Mtimel @ Mtime))",
"ods on News2013 (Table 3), the CM fusion method"
] |
{
"arxiv_id": "2112.06166",
"language": "en",
"page_idx": 4
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.