id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2203.03691-en-1
|
2203.03691-en-1.jpg
|
$\mathbf{Model~Layer}$
<FigureHere>
Figure 1: The figure outlines a general model layer consisting of a token mixing component and a feature mixing component (MLP). For token mixing, MLPMixer uses an MLP with a $f$xed size, maximum input length $N$ and $position-specific$ weights. In contrast, HyperMixer generates an appropriately sized MLP based on the variable size of the input in a $\textit{position- invariant way, similar to the attention mechanism. When using attention as token mixing}$ the whole layer is equivalent to a Transformer encoder layer.
4. An ablation demonstrating that HyperMixer
learns attention patterns similar to Transform-
ers. (Section: 4.7)
### 2 $\textbf{Method}$
cially for long inputs; favorable performance in the low-resource regime (D); and efficient tuning for hyperparameters (H). We attribute HyperMixer's success to its ability to approximate an attentionlike function. Further experiments on a synthetic task demonstrate that HyperMixer indeed learns to attend to tokens in similar pattern to the attention mechanism.
In summary, our contributions can be enumer-
ated as follows:
## 2.1 Inductive Biases in NLP Models
1. A novel all-MLP model, HyperMixer, with in-
ductive biases similar to Transformers. (Sec-
tion: 2)
In machine learning, the inductive biases of a model reflect implicit modeling assumptions which are key to facilitate learning and improve generalization on specific tasks. In NLP, well-known models with strong inductive biases include: recurrent neural networks (Elman, 1990), which assume the input to be a sequence; and recursive neural networks (Socher et al., 2013), which assume a treestructure. While both these inductive biases are reasonable, empirically, Transformers have been more successful in recent years. Furthermore, we reiterate the arguments of Henderson (2020) for inductive biases in language and apply them to our model design. Henderson (2020) attributes the Transformer's success to two concepts: vari-
2.A performance analysis of HyperMixer against alternative token mixing methods based on controlled experiments on the GLUE benchmark. (Section: 4.3)
3.A comprehensive comparison of the cost
$Cost(R)$ of HyperMixer and Transformers.
(Sections: 4.4, 4.5, 4.6)
|
[
[
0.43408203125,
0.0831298828125,
0.5576171875,
0.08489990234375,
0.5576171875,
0.09857177734375,
0.43408203125,
0.0968017578125
],
[
0.28369140625,
0.114013671875,
0.391357421875,
0.114013671875,
0.391357421875,
0.1258544921875,
0.28369140625,
0.1258544921875
],
[
0.62060546875,
0.114013671875,
0.72900390625,
0.114013671875,
0.72900390625,
0.1258544921875,
0.62060546875,
0.1258544921875
],
[
0.62451171875,
0.146728515625,
0.66650390625,
0.146728515625,
0.66650390625,
0.157958984375,
0.62451171875,
0.157958984375
],
[
0.2392578125,
0.15673828125,
0.267822265625,
0.15673828125,
0.267822265625,
0.168701171875,
0.2392578125,
0.168701171875
],
[
0.26123046875,
0.15673828125,
0.28955078125,
0.15673828125,
0.28955078125,
0.168701171875,
0.26123046875,
0.168701171875
],
[
0.44921875,
0.15673828125,
0.51318359375,
0.15673828125,
0.51318359375,
0.168701171875,
0.44921875,
0.168701171875
],
[
0.7197265625,
0.157958984375,
0.7841796875,
0.157958984375,
0.7841796875,
0.169189453125,
0.7197265625,
0.169189453125
],
[
0.2283935546875,
0.1602783203125,
0.2392578125,
0.1602783203125,
0.2392578125,
0.1650390625,
0.2283935546875,
0.1650390625
],
[
0.1695556640625,
0.163330078125,
0.185546875,
0.163330078125,
0.185546875,
0.189453125,
0.1695556640625,
0.189453125
],
[
0.5498046875,
0.168701171875,
0.5615234375,
0.168701171875,
0.5615234375,
0.188232421875,
0.5498046875,
0.188232421875
],
[
0.344970703125,
0.1787109375,
0.387939453125,
0.1787109375,
0.387939453125,
0.190673828125,
0.344970703125,
0.190673828125
],
[
0.640625,
0.188232421875,
0.6484375,
0.188232421875,
0.6484375,
0.1953125,
0.640625,
0.1953125
],
[
0.744140625,
0.188232421875,
0.76171875,
0.188232421875,
0.76171875,
0.196533203125,
0.744140625,
0.196533203125
],
[
0.7314453125,
0.190673828125,
0.73779296875,
0.190673828125,
0.73779296875,
0.1953125,
0.7314453125,
0.1953125
],
[
0.17041015625,
0.191162109375,
0.185546875,
0.1923828125,
0.1822509765625,
0.218505859375,
0.16796875,
0.21728515625
],
[
0.343505859375,
0.190673828125,
0.391357421875,
0.190673828125,
0.391357421875,
0.201904296875,
0.343505859375,
0.201904296875
],
[
0.5517578125,
0.196533203125,
0.5634765625,
0.196533203125,
0.5634765625,
0.218505859375,
0.5517578125,
0.218505859375
],
[
0.240966796875,
0.2120361328125,
0.287109375,
0.2120361328125,
0.287109375,
0.2197265625,
0.240966796875,
0.2197265625
],
[
0.45166015625,
0.2120361328125,
0.468505859375,
0.2120361328125,
0.468505859375,
0.2197265625,
0.45166015625,
0.2197265625
],
[
0.466796875,
0.2120361328125,
0.49365234375,
0.2120361328125,
0.49365234375,
0.2197265625,
0.466796875,
0.2197265625
],
[
0.490234375,
0.2120361328125,
0.509765625,
0.2120361328125,
0.509765625,
0.2197265625,
0.490234375,
0.2197265625
],
[
0.72705078125,
0.213134765625,
0.73291015625,
0.21728515625,
0.72705078125,
0.221435546875,
0.72119140625,
0.21728515625
],
[
0.7421875,
0.2120361328125,
0.7841796875,
0.2120361328125,
0.7841796875,
0.22265625,
0.7421875,
0.22265625
],
[
0.316650390625,
0.232177734375,
0.418212890625,
0.232177734375,
0.418212890625,
0.2464599609375,
0.316650390625,
0.2464599609375
],
[
0.2301025390625,
0.21435546875,
0.2392578125,
0.21435546875,
0.2392578125,
0.21728515625,
0.2301025390625,
0.21728515625
],
[
0.225830078125,
0.2340087890625,
0.279541015625,
0.2340087890625,
0.279541015625,
0.248779296875,
0.225830078125,
0.248779296875
],
[
0.455078125,
0.232177734375,
0.50537109375,
0.2340087890625,
0.50390625,
0.248779296875,
0.453369140625,
0.2464599609375
],
[
0.611328125,
0.235107421875,
0.67529296875,
0.2327880859375,
0.67529296875,
0.24462890625,
0.611328125,
0.2464599609375
],
[
0.7255859375,
0.2340087890625,
0.779296875,
0.2340087890625,
0.779296875,
0.248779296875,
0.7255859375,
0.248779296875
],
[
0.73486328125,
0.365234375,
0.765625,
0.365234375,
0.765625,
0.377685546875,
0.73486328125,
0.377685546875
],
[
0.71728515625,
0.37890625,
0.78271484375,
0.37890625,
0.78271484375,
0.39013671875,
0.71728515625,
0.39013671875
],
[
0.20068359375,
0.410400390625,
0.219970703125,
0.410400390625,
0.219970703125,
0.41455078125,
0.20068359375,
0.41455078125
],
[
0.2476806640625,
0.410400390625,
0.26025390625,
0.410400390625,
0.26025390625,
0.41455078125,
0.2476806640625,
0.41455078125
],
[
0.50732421875,
0.410400390625,
0.5224609375,
0.410400390625,
0.5224609375,
0.41455078125,
0.50732421875,
0.41455078125
],
[
0.181396484375,
0.429931640625,
0.28955078125,
0.429931640625,
0.28955078125,
0.44482421875,
0.181396484375,
0.44482421875
],
[
0.4306640625,
0.429931640625,
0.5498046875,
0.429931640625,
0.5498046875,
0.44482421875,
0.4306640625,
0.44482421875
],
[
0.6962890625,
0.429931640625,
0.81787109375,
0.429931640625,
0.81787109375,
0.44482421875,
0.6962890625,
0.44482421875
],
[
0.1175537109375,
0.4619140625,
0.880859375,
0.4619140625,
0.880859375,
0.476806640625,
0.1175537109375,
0.476806640625
],
[
0.1175537109375,
0.476806640625,
0.880859375,
0.476806640625,
0.880859375,
0.490478515625,
0.1175537109375,
0.490478515625
],
[
0.1158447265625,
0.4912109375,
0.8818359375,
0.4912109375,
0.8818359375,
0.5048828125,
0.1158447265625,
0.5048828125
],
[
0.1175537109375,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.1175537109375,
0.52099609375,
0.5283203125,
0.52099609375,
0.5283203125,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.1175537109375,
0.56005859375,
0.48779296875,
0.56005859375,
0.48779296875,
0.57470703125,
0.1175537109375,
0.57470703125
],
[
0.52294921875,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.52294921875,
0.572265625
],
[
0.1175537109375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.1175537109375,
0.59033203125
],
[
0.54736328125,
0.576171875,
0.880859375,
0.576171875,
0.880859375,
0.5888671875,
0.54736328125,
0.5888671875
],
[
0.1158447265625,
0.5927734375,
0.48779296875,
0.59130859375,
0.48779296875,
0.60546875,
0.1158447265625,
0.60693359375
],
[
0.54736328125,
0.5927734375,
0.68603515625,
0.5927734375,
0.68603515625,
0.60693359375,
0.54736328125,
0.60693359375
],
[
0.1175537109375,
0.60791015625,
0.48779296875,
0.60791015625,
0.48779296875,
0.62158203125,
0.1175537109375,
0.62158203125
],
[
0.1175537109375,
0.62451171875,
0.48779296875,
0.62451171875,
0.48779296875,
0.63818359375,
0.1175537109375,
0.63818359375
],
[
0.51123046875,
0.62451171875,
0.61279296875,
0.62451171875,
0.61279296875,
0.6396484375,
0.51123046875,
0.6396484375
],
[
0.1175537109375,
0.64013671875,
0.486083984375,
0.64013671875,
0.486083984375,
0.65380859375,
0.1175537109375,
0.65380859375
],
[
0.1158447265625,
0.65673828125,
0.489501953125,
0.65673828125,
0.489501953125,
0.67041015625,
0.1158447265625,
0.67041015625
],
[
0.51123046875,
0.65380859375,
0.80615234375,
0.65380859375,
0.80615234375,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.1175537109375,
0.6748046875,
0.20654296875,
0.6748046875,
0.20654296875,
0.68505859375,
0.1175537109375,
0.68505859375
],
[
0.51123046875,
0.6787109375,
0.8818359375,
0.6787109375,
0.8818359375,
0.693359375,
0.51123046875,
0.693359375
],
[
0.133544921875,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.70361328125,
0.133544921875,
0.70361328125
],
[
0.51123046875,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.51123046875,
0.708984375
],
[
0.11419677734375,
0.70361328125,
0.2359619140625,
0.705078125,
0.2359619140625,
0.7197265625,
0.11419677734375,
0.71875
],
[
0.51123046875,
0.71044921875,
0.88330078125,
0.7119140625,
0.88330078125,
0.7255859375,
0.51123046875,
0.7236328125
],
[
0.51123046875,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1318359375,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.1318359375,
0.74755859375
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.15283203125,
0.75048828125,
0.490234375,
0.75048828125,
0.490234375,
0.76416015625,
0.15283203125,
0.76416015625
],
[
0.51123046875,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.15283203125,
0.7666015625,
0.21240234375,
0.7666015625,
0.21240234375,
0.78076171875,
0.15283203125,
0.78076171875
],
[
0.509765625,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.509765625,
0.7880859375
],
[
0.130126953125,
0.79443359375,
0.489501953125,
0.79443359375,
0.489501953125,
0.80859375,
0.130126953125,
0.80859375
],
[
0.509765625,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.509765625,
0.80419921875
],
[
0.15283203125,
0.81103515625,
0.48779296875,
0.81103515625,
0.48779296875,
0.826171875,
0.15283203125,
0.826171875
],
[
0.5087890625,
0.80712890625,
0.8818359375,
0.8056640625,
0.8818359375,
0.82080078125,
0.5087890625,
0.82177734375
],
[
0.15283203125,
0.82666015625,
0.48779296875,
0.82666015625,
0.48779296875,
0.84033203125,
0.15283203125,
0.84033203125
],
[
0.509765625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.509765625,
0.83837890625
],
[
0.15283203125,
0.84326171875,
0.349365234375,
0.84326171875,
0.349365234375,
0.85693359375,
0.15283203125,
0.85693359375
],
[
0.51318359375,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.51123046875,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.869140625,
0.51123046875,
0.869140625
],
[
0.130126953125,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.130126953125,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.151123046875,
0.88720703125,
0.491943359375,
0.88818359375,
0.491943359375,
0.90283203125,
0.151123046875,
0.90185546875
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.154541015625,
0.90380859375,
0.328369140625,
0.90380859375,
0.328369140625,
0.91845703125,
0.154541015625,
0.91845703125
],
[
0.509765625,
0.90283203125,
0.88330078125,
0.90380859375,
0.88330078125,
0.91845703125,
0.509765625,
0.91748046875
]
] |
[
"Model Layer",
"Skip connection",
"Skip connection",
"MLP",
".",
".",
"...",
"...",
"福",
"Layer",
"Laye",
"Token",
"S",
".",
"D",
"Norm",
"Mixing",
"Norn",
"100",
"O",
"·",
".",
"福",
".",
"RNxd-RNxd",
"",
"[Nxd]",
"[Nxd]",
"RdRd",
"[N × d]",
"Self",
"Attention",
"IV",
"",
"TD1",
"MLPMixer",
"HyperMixer",
"Transformer",
"Figure I: The figure outlines a general model layer consisting of a token mixing component and a feature mixing",
"component (MLP). For token mixing, MLPMixer uses an MLP with a fixed size, maximum input length N and",
"position-specific weights. In contrast, HyperMixer generates an appropriately sized MLP based on the variable size",
"of the input in a position-invariant way, similar to the attention mechanism. When using attention as token mixing",
"the whole layer is equivalent to a Transformer encoder layer.",
"cially for long inputs; favorable performance in the",
" 4. An ablation demonstrating that HyperMixer",
"low-resource regime (D); and efficient tuning for",
"learns attention patterns similar to Transform-",
"hyperparameters (H). We attribute HyperMixer's",
"ers. (Section: 4.7)",
"success to its ability to approximate an attention-",
"like function. Further experiments on a synthetic",
"2 Method",
"task demonstrate that HyperMixer indeed learns to",
"attend to tokens in similar pattern to the attention",
"2.1 Inductive Biases in NLP Models",
"mechanism.",
"In machine learning, the inductive biases of a",
"In summary, our contributions can be enumer-",
"model reflect implicit modeling assumptions which",
" ated as follows:",
"are key to facilitate learning and improve gener-",
"alization on specific tasks. In NLP, well-known",
"1. A novel all-MLP model, HyperMixer, with in-",
"models with strong inductive biases include: recur-",
"ductive biases similar to Transformers. (Sec-",
"rent neural networks (Elman, 1990), which assume",
"tion: 2)",
"the input to be a sequence; and recursive neural net",
"2. A performance analysis of HyperMixer",
"works (Socher et al., 2013), which assume a tree-",
"against alternative token mixing methods",
"structure. While both these inductive biases are",
"based on controlled experiments on the GLUE",
"reasonable, empirically, Transformers have been",
"benchmark. (Section: 4.3)",
"more successful in recent years. Furthermore, we",
"reiterate the arguments of Henderson (2020) for",
"3. A comprehensive comparison of the cost",
"inductive biases in language and apply them to",
"Cost(R) of HyperMixer and Transformers.",
"our model design. Henderson (2020) attributes",
"(Sections: 4.4, 4.5, 4.6)",
"the Transformer's success to two concepts: vari-"
] |
{
"arxiv_id": "2203.03691",
"language": "en",
"page_idx": 1
}
|
||
2203.03691-en-2
|
2203.03691-en-2.jpg
|
able binding and systematicity. Variable binding refers to the model's ability to represent multiple entities at once. This is arguably challenging in single-vector representations such as recurrent neural networks. However, Transformers represent each token with its own vector which accounts for variable binding as each token can be interpreted as an entity. Systematicity refers to the models ability to learn generalizable rules that reflect the structural relationship between entities (Fodor and Pylyshyn, 1988). Transformers achieve systematicity through the attention mechanism which is a learnable set of functions that determines the interaction between entities by matching query representations to key representations (as shown in Figure 1). The mechanism modulates, for every position in the sequence, how to functionally process any other position. Moreover, these function parameters are learnable and shared across all entities.
## 2.2 MLPMixer
A general layer of MLPMixer is shown in Figure 1. Similarly to Transformers, each token is represented as a vector of features, which undergo (nonlinear) transformations in multiple layers. MLPMixer employs two MLPs at each layer, one for feature mixing and one for token mixing. The feature mixing component is applied to each token vector independently, which models the interactions between features. The Token Mixing MLP (TM-MLP) is applied to each feature independently (i.e. its vector of values across tokens), which models the interactions between spatial locations or positions. This could be interpreted as a global attention mechanism which is static and position-modulated. Practically, this is achieved by transposing the dimension representing the features and the dimension representing the positions. Each vector $\mathbf{x}_i^T\in\mathbb{R}^N$, representing feature $i\leq d$, of some input of fxed length $N$, is input into TM-MLP, which has the following form:
$$\mathrm{TM-MLP}(\mathbf{x}_i^T)=\mathbf{W}_1(\sigma(\mathbf{W}_2^T\mathbf{x}_i^T)),$$
(1)
where $\mathbf{W} _1, \mathbf{W} _2$ $\in$ $\mathbb{R} ^{N\times d^{\prime }}$, and $\sigma$ represents the GELU non-linearity (Hendrycks and Gimpel, 2016). Finally, to facilitate learning, layer normalization (Ba et al., 2016) and skip connections (He et al., 2016) are added around each MLP, respectively. How to best arrange these components is still an open question (Wang et al., 2019; Bach-
lechner et al., 2021). We experiment with different
variants in Appendix F.
Considerations for NLP The token mixing MLP assumes an input of fxed dimension, which is necessary as the parameters need to be shared across all examples. However, unlike images, textual input is generally of a variable dimension. Therefore, to apply MLPMixer to texts of variable length, a simplistic approach is to assume a maximum length (e.g. the maximum in the dataset). Thereafter, all inputs are padded to the maximum length and masks are applied in the token mixing MLP. This model is able to do variable binding, since each token is represented by its own vector. However, this model lacks systematicity because the rules learned to model interactions between tokens (i.e. the MLP's weights) are not shared across positions.
## 2.3 HyperMixer
<table>
<tbody>
<tr>
<th>$\underline{\overline{\mathbf{lgor}}}$</th>
<th>rithm 1 HyperMixer pseudo-code</th>
</tr>
<tr>
<td>$\mathbf{lgor}$ A $1ass$ $H$ def def</td>
<td>rithm I HyperM1xer pseudo-code HyperMixing( nn. Module) : $:f$ $f\_\underline{init}\_($self,d,d'): # learnable parameters self. hypernetwork in = MLP( [ d, d, $d^{\prime }$] ) $self.hypernetwork\_out\_out\_MLP([d,~d,~d^{\prime}])$ # layer normalization improves training stability $self.layer\_norm=LayerNorm(d)$ forward(self. alleries keys, values):</td>
</tr>
<tr>
<td>$1ass$ $H$ def</td>
<td>HyperMixing( nn. Module) : f $E\_\underline{\mathrm{init}}\_($self, d, d'): # learnable parameters self. hypernetwork$\_$in = MLP( [ d, d, d, $d^{\prime }$] ) $self.hypernetwork\_out\_out\_MLP([d,~d,~d^{\prime}])$ # layer normalization improves training stability $self.layer\_norm=LayerNorm(d)$</td>
</tr>
<tr>
<td rowspan="5">def</td>
<td>f E forward(self, queries, keys, values): $\#$ queries: [B, M, d] # keys / values: [B, N, d]</td>
</tr>
<tr>
<td>$\mathbf{W2}$ = self. hypernetwork$\_$out( hyp$\_$out) $\#$ $[ B, M, d^{\prime }]$ $\begin{array}{ccc}\mathrm{TM-MLP(x)}&=W_-2(\mathrm{~GELU~(~W_-1^{\wedge}T~x)~)}\end{array}$ $\#$ maps [B,d,N] $\rightarrow[B,d,M]$ $token\_mixing\_mlp=compose\_TM\_MLP(W1,W2)$</td>
</tr>
<tr>
<td>values values.transpose :(1, 2) 上 $[B$, d NI</td>
</tr>
<tr>
<td>output = token\_mixing\_mlp(values) $\# [ B$, $\mathbf{d} $, $\mathbf{M} ]$</td>
</tr>
<tr>
<td># transpose back output = output.transpose $(1,2)\#[\mathcal{B},\mathcal{M},\mathcal{d}]$ # optionally apply LayerNorm return self.layer\_norm(output)</td>
</tr>
</tbody>
</table>
HyperMixer includes systematicity into the MLPMixer architecture by introducing a novel token mixing mechanism, $HyperMixing^{2}$, which can be regarded as a drop-in replacement for attention. For ease of understanding, we provide pseudo-code in Algorithm 1. While the queries, keys, and values in HyperMixing need not be the same, we will
$^2$HyperMixing is to HyperMixer what self-attention is to
Transformer encoders.
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.509765625,
0.08489990234375,
0.880859375,
0.08489990234375,
0.880859375,
0.09857177734375,
0.509765625,
0.09857177734375
],
[
0.1158447265625,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.50732421875,
0.1015625,
0.68701171875,
0.1009521484375,
0.68701171875,
0.11517333984375,
0.50732421875,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.12890625,
0.8818359375,
0.1300048828125,
0.8818359375,
0.144287109375,
0.51123046875,
0.1436767578125
],
[
0.118408203125,
0.135986328125,
0.490234375,
0.135986328125,
0.490234375,
0.146728515625,
0.118408203125,
0.146728515625
],
[
0.1158447265625,
0.1484375,
0.489501953125,
0.149658203125,
0.489501953125,
0.1644287109375,
0.1158447265625,
0.163330078125
],
[
0.509765625,
0.1478271484375,
0.88330078125,
0.1478271484375,
0.88330078125,
0.160888671875,
0.509765625,
0.160888671875
],
[
0.1175537109375,
0.166259765625,
0.489501953125,
0.166259765625,
0.489501953125,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.509765625,
0.163330078125,
0.8818359375,
0.163330078125,
0.8818359375,
0.177001953125,
0.509765625,
0.177001953125
],
[
0.1175537109375,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.51123046875,
0.179931640625,
0.88330078125,
0.179931640625,
0.88330078125,
0.193603515625,
0.51123046875,
0.193603515625
],
[
0.1158447265625,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.5087890625,
0.1953125,
0.8818359375,
0.19482421875,
0.8818359375,
0.208984375,
0.5087890625,
0.210205078125
],
[
0.1175537109375,
0.215576171875,
0.489501953125,
0.215576171875,
0.489501953125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.509765625,
0.21142578125,
0.88330078125,
0.21142578125,
0.88330078125,
0.2257080078125,
0.509765625,
0.2257080078125
],
[
0.1158447265625,
0.23095703125,
0.48779296875,
0.23095703125,
0.48779296875,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.51123046875,
0.22802734375,
0.8818359375,
0.22802734375,
0.8818359375,
0.2410888671875,
0.51123046875,
0.2410888671875
],
[
0.1175537109375,
0.2464599609375,
0.491943359375,
0.2464599609375,
0.491943359375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.5087890625,
0.243408203125,
0.88330078125,
0.2423095703125,
0.88330078125,
0.257080078125,
0.5087890625,
0.25830078125
],
[
0.11419677734375,
0.262451171875,
0.48779296875,
0.262451171875,
0.48779296875,
0.275634765625,
0.11419677734375,
0.275634765625
],
[
0.51318359375,
0.260009765625,
0.8818359375,
0.260009765625,
0.8818359375,
0.273681640625,
0.51318359375,
0.273681640625
],
[
0.1175537109375,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29150390625,
0.1175537109375,
0.29150390625
],
[
0.509765625,
0.275634765625,
0.8818359375,
0.275634765625,
0.8818359375,
0.289306640625,
0.509765625,
0.289306640625
],
[
0.11419677734375,
0.29345703125,
0.490234375,
0.295654296875,
0.490234375,
0.31005859375,
0.11419677734375,
0.30810546875
],
[
0.51123046875,
0.29345703125,
0.880859375,
0.29345703125,
0.880859375,
0.303955078125,
0.51123046875,
0.303955078125
],
[
0.1158447265625,
0.312255859375,
0.489501953125,
0.312255859375,
0.489501953125,
0.325927734375,
0.1158447265625,
0.325927734375
],
[
0.509765625,
0.30810546875,
0.884765625,
0.30810546875,
0.884765625,
0.32177734375,
0.509765625,
0.32177734375
],
[
0.1175537109375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.51123046875,
0.32373046875,
0.8818359375,
0.32373046875,
0.8818359375,
0.33740234375,
0.51123046875,
0.33740234375
],
[
0.1158447265625,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35693359375,
0.1158447265625,
0.35693359375
],
[
0.51123046875,
0.34033203125,
0.88330078125,
0.34033203125,
0.88330078125,
0.35400390625,
0.51123046875,
0.35400390625
],
[
0.1175537109375,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.509765625,
0.35693359375,
0.8818359375,
0.35693359375,
0.8818359375,
0.37060546875,
0.509765625,
0.37060546875
],
[
0.11419677734375,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.11419677734375,
0.388427734375
],
[
0.51318359375,
0.385986328125,
0.654296875,
0.385986328125,
0.654296875,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.1158447265625,
0.392578125,
0.1485595703125,
0.392578125,
0.1485595703125,
0.405029296875,
0.1158447265625,
0.405029296875
],
[
0.1158447265625,
0.419189453125,
0.2493896484375,
0.419189453125,
0.2493896484375,
0.432861328125,
0.1158447265625,
0.432861328125
],
[
0.51318359375,
0.42333984375,
0.8017578125,
0.42333984375,
0.8017578125,
0.43701171875,
0.51318359375,
0.43701171875
],
[
0.1175537109375,
0.440673828125,
0.48779296875,
0.440673828125,
0.48779296875,
0.454833984375,
0.1175537109375,
0.454833984375
],
[
0.51318359375,
0.448974609375,
0.6884765625,
0.448974609375,
0.6884765625,
0.458984375,
0.51318359375,
0.458984375
],
[
0.1158447265625,
0.457275390625,
0.490234375,
0.457763671875,
0.490234375,
0.471435546875,
0.1158447265625,
0.47021484375
],
[
0.580078125,
0.460205078125,
0.59375,
0.457763671875,
0.59423828125,
0.4619140625,
0.5810546875,
0.46435546875
],
[
0.603515625,
0.457763671875,
0.69287109375,
0.457763671875,
0.69287109375,
0.465576171875,
0.603515625,
0.465576171875
],
[
0.1175537109375,
0.473876953125,
0.489501953125,
0.473876953125,
0.489501953125,
0.48681640625,
0.1175537109375,
0.48681640625
],
[
0.5400390625,
0.46142578125,
0.55322265625,
0.46142578125,
0.55322265625,
0.46435546875,
0.5400390625,
0.46435546875
],
[
0.55908203125,
0.478515625,
0.79248046875,
0.479736328125,
0.79248046875,
0.49462890625,
0.55908203125,
0.493408203125
],
[
0.5615234375,
0.473876953125,
0.6962890625,
0.473876953125,
0.6962890625,
0.48388671875,
0.5615234375,
0.48388671875
],
[
0.1158447265625,
0.4892578125,
0.489501953125,
0.488037109375,
0.489501953125,
0.5029296875,
0.1158447265625,
0.50439453125
],
[
0.55908203125,
0.4892578125,
0.7958984375,
0.488037109375,
0.7958984375,
0.49951171875,
0.55908203125,
0.50048828125
],
[
0.1175537109375,
0.505859375,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.603515625,
0.50732421875,
0.6943359375,
0.50732421875,
0.6943359375,
0.51416015625,
0.603515625,
0.51416015625
],
[
0.7314453125,
0.5048828125,
0.8564453125,
0.505859375,
0.8564453125,
0.5166015625,
0.7314453125,
0.515625
],
[
0.6943359375,
0.51025390625,
0.7255859375,
0.51025390625,
0.7255859375,
0.51318359375,
0.6943359375,
0.51318359375
],
[
0.56494140625,
0.50830078125,
0.60205078125,
0.51025390625,
0.60205078125,
0.51416015625,
0.56494140625,
0.5126953125
],
[
0.5634765625,
0.515625,
0.74072265625,
0.515625,
0.74072265625,
0.5224609375,
0.5634765625,
0.5224609375
],
[
0.1158447265625,
0.5224609375,
0.48779296875,
0.5224609375,
0.48779296875,
0.5361328125,
0.1158447265625,
0.5361328125
],
[
0.5380859375,
0.53076171875,
0.78515625,
0.53076171875,
0.78515625,
0.54150390625,
0.5380859375,
0.54150390625
],
[
0.1175537109375,
0.5380859375,
0.489501953125,
0.5380859375,
0.489501953125,
0.5517578125,
0.1175537109375,
0.5517578125
],
[
0.56103515625,
0.54052734375,
0.67822265625,
0.54052734375,
0.67822265625,
0.54736328125,
0.56103515625,
0.54736328125
],
[
0.56103515625,
0.5458984375,
0.65576171875,
0.54736328125,
0.65576171875,
0.55810546875,
0.56103515625,
0.5556640625
],
[
0.64990234375,
0.54638671875,
0.71875,
0.54638671875,
0.71875,
0.55712890625,
0.64990234375,
0.55712890625
],
[
0.11419677734375,
0.55419921875,
0.489501953125,
0.552734375,
0.489501953125,
0.56787109375,
0.11419677734375,
0.568359375
],
[
0.1158447265625,
0.568359375,
0.48779296875,
0.5693359375,
0.48779296875,
0.58447265625,
0.1158447265625,
0.5830078125
],
[
0.56103515625,
0.5634765625,
0.79443359375,
0.56396484375,
0.79443359375,
0.57470703125,
0.56103515625,
0.57373046875
],
[
0.5634765625,
0.57470703125,
0.72705078125,
0.57470703125,
0.72705078125,
0.58203125,
0.5634765625,
0.58203125
],
[
0.724609375,
0.57470703125,
0.77490234375,
0.57470703125,
0.77490234375,
0.58203125,
0.724609375,
0.58203125
],
[
0.1175537109375,
0.5859375,
0.489501953125,
0.5859375,
0.489501953125,
0.599609375,
0.1175537109375,
0.599609375
],
[
0.56103515625,
0.58056640625,
0.8017578125,
0.58056640625,
0.8017578125,
0.59130859375,
0.56103515625,
0.59130859375
],
[
0.55517578125,
0.5927734375,
0.85546875,
0.5927734375,
0.85546875,
0.62060546875,
0.55517578125,
0.62060546875
],
[
0.1158447265625,
0.6025390625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.1158447265625,
0.6162109375
],
[
0.56494140625,
0.6015625,
0.57421875,
0.6015625,
0.57421875,
0.60498046875,
0.56494140625,
0.60498046875
],
[
0.1175537109375,
0.6181640625,
0.48779296875,
0.6181640625,
0.48779296875,
0.6318359375,
0.1175537109375,
0.6318359375
],
[
0.5634765625,
0.62353515625,
0.64404296875,
0.62353515625,
0.64404296875,
0.63134765625,
0.5634765625,
0.63134765625
],
[
0.6455078125,
0.62353515625,
0.7197265625,
0.62353515625,
0.7197265625,
0.63134765625,
0.6455078125,
0.63134765625
],
[
0.73046875,
0.62158203125,
0.7841796875,
0.62158203125,
0.7841796875,
0.6318359375,
0.73046875,
0.6318359375
],
[
0.1158447265625,
0.63427734375,
0.486083984375,
0.63427734375,
0.486083984375,
0.6484375,
0.1158447265625,
0.6484375
],
[
0.5634765625,
0.63134765625,
0.66748046875,
0.6318359375,
0.66748046875,
0.6396484375,
0.5634765625,
0.63818359375
],
[
0.5615234375,
0.63818359375,
0.8076171875,
0.63818359375,
0.8076171875,
0.6484375,
0.5615234375,
0.6484375
],
[
0.66650390625,
0.63427734375,
0.73291015625,
0.63427734375,
0.73291015625,
0.63818359375,
0.66650390625,
0.63818359375
],
[
0.11419677734375,
0.65087890625,
0.486083984375,
0.6484375,
0.486083984375,
0.662109375,
0.11419677734375,
0.66455078125
],
[
0.56494140625,
0.65771484375,
0.63232421875,
0.65771484375,
0.63232421875,
0.6650390625,
0.56494140625,
0.6650390625
],
[
0.6943359375,
0.66015625,
0.7314453125,
0.66015625,
0.7314453125,
0.66455078125,
0.6943359375,
0.66455078125
],
[
0.771484375,
0.66015625,
0.7958984375,
0.66015625,
0.7958984375,
0.66455078125,
0.771484375,
0.66455078125
],
[
0.1158447265625,
0.66748046875,
0.486083984375,
0.66650390625,
0.486083984375,
0.6796875,
0.1158447265625,
0.68115234375
],
[
0.69287109375,
0.66455078125,
0.8193359375,
0.66455078125,
0.8193359375,
0.67138671875,
0.69287109375,
0.67138671875
],
[
0.5634765625,
0.68115234375,
0.59033203125,
0.68115234375,
0.59033203125,
0.6884765625,
0.5634765625,
0.6884765625
],
[
0.66162109375,
0.6796875,
0.8330078125,
0.6796875,
0.8330078125,
0.6904296875,
0.66162109375,
0.6904296875
],
[
0.1175537109375,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.1175537109375,
0.71337890625
],
[
0.1175537109375,
0.68310546875,
0.489501953125,
0.68310546875,
0.489501953125,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.56689453125,
0.7060546875,
0.60546875,
0.7060546875,
0.60546875,
0.71337890625,
0.56689453125,
0.71337890625
],
[
0.609375,
0.7060546875,
0.66162109375,
0.7060546875,
0.66162109375,
0.71337890625,
0.609375,
0.71337890625
],
[
0.66015625,
0.7080078125,
0.67236328125,
0.7080078125,
0.67236328125,
0.71142578125,
0.66015625,
0.71142578125
],
[
0.68115234375,
0.7060546875,
0.81201171875,
0.7060546875,
0.81201171875,
0.71337890625,
0.68115234375,
0.71337890625
],
[
0.1175537109375,
0.71435546875,
0.48779296875,
0.71435546875,
0.48779296875,
0.728515625,
0.1175537109375,
0.728515625
],
[
0.5634765625,
0.7216796875,
0.609375,
0.7216796875,
0.609375,
0.732421875,
0.5634765625,
0.732421875
],
[
0.1158447265625,
0.73095703125,
0.48779296875,
0.73095703125,
0.48779296875,
0.74462890625,
0.1158447265625,
0.74462890625
],
[
0.591796875,
0.72998046875,
0.66015625,
0.73095703125,
0.66015625,
0.73828125,
0.591796875,
0.73681640625
],
[
0.68115234375,
0.73095703125,
0.74072265625,
0.73095703125,
0.74072265625,
0.73828125,
0.68115234375,
0.73828125
],
[
0.1158447265625,
0.7451171875,
0.426513671875,
0.74658203125,
0.426513671875,
0.76123046875,
0.1158447265625,
0.76025390625
],
[
0.17041015625,
0.77392578125,
0.435791015625,
0.775390625,
0.435791015625,
0.7919921875,
0.17041015625,
0.79052734375
],
[
0.52880859375,
0.7685546875,
0.8818359375,
0.7685546875,
0.8818359375,
0.7822265625,
0.52880859375,
0.7822265625
],
[
0.462646484375,
0.779296875,
0.48779296875,
0.779296875,
0.48779296875,
0.79052734375,
0.462646484375,
0.79052734375
],
[
0.51123046875,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51123046875,
0.798828125
],
[
0.11419677734375,
0.802734375,
0.491943359375,
0.802734375,
0.491943359375,
0.82373046875,
0.11419677734375,
0.82373046875
],
[
0.51318359375,
0.8017578125,
0.8818359375,
0.8017578125,
0.8818359375,
0.8154296875,
0.51318359375,
0.8154296875
],
[
0.1158447265625,
0.822265625,
0.489501953125,
0.82373046875,
0.489501953125,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.509765625,
0.81787109375,
0.8818359375,
0.81787109375,
0.8818359375,
0.83056640625,
0.509765625,
0.83056640625
],
[
0.1175537109375,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.51318359375,
0.8349609375,
0.880859375,
0.8349609375,
0.880859375,
0.845703125,
0.51318359375,
0.845703125
],
[
0.5087890625,
0.849609375,
0.880859375,
0.849609375,
0.880859375,
0.86328125,
0.5087890625,
0.86328125
],
[
0.1158447265625,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1158447265625,
0.87060546875
],
[
0.5087890625,
0.865234375,
0.880859375,
0.865234375,
0.880859375,
0.87890625,
0.5087890625,
0.87890625
],
[
0.1175537109375,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.1158447265625,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.52880859375,
0.89111328125,
0.8818359375,
0.8935546875,
0.8818359375,
0.90771484375,
0.52880859375,
0.90625
],
[
0.1158447265625,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.6533203125,
0.90478515625,
0.6533203125,
0.91845703125,
0.5087890625,
0.91845703125
]
] |
[
"able binding and systematicity. Variable binding",
"lechner et al., 2021). We experiment with different",
"refers to the model's ability to represent multiple",
"variants in Appendix F.",
"entities at once. This is arguably challenging in",
"Considerations for NLP The token mixing MLP",
"single-vector representations such as recurrent neu-",
"ral networks. However, Transformers represent",
"assumes an input of fixed dimension, which is nec-",
"each token with its own vector which accounts for",
" essary as the parameters need to be shared across all",
"variable binding as each token can be interpreted",
"examples. However, unlike images, textual input",
"as an entity. Systematicity refers to the models",
"is generally of a variable dimension. Therefore, to",
"ability to learn generalizable rules that reflect the",
" apply MLPMixer to texts of variable length, a sim-",
" structural relationship between entities (Fodor and",
"plistic approach is to assume a maximum length",
"Pylyshyn, 1988). Transformers achieve system-",
"(e.g. the maximum in the dataset). Thereafter,",
"aticity through the attention mechanism which is",
"all inputs are padded to the maximum length and",
"a learnable set of functions that determines the",
"masks are applied in the token mixing MLP. This",
"interaction between entities by matching query rep-",
"model is able to do variable binding, since each",
"resentations to key representations (as shown in",
"token is represented by its own vector. However,",
"Figure 1). The mechanism modulates, for every",
"this model lacks systematicity because the rules",
"position in the sequence, how to functionally pro-",
"learned to model interactions between tokens (i.e.",
"cess any other position. Moreover, these function",
"the MLP's weights) are not shared across positions",
"parameters are learnable and shared across all enti-",
"2.3 HyperMixer",
"ties.",
"2.2 MLPMixer",
"Algorithm1HyperMixerpseudo-code",
"A general layer of MLPMixer is shown in Figure 1.",
"class HyperMixing(nn.Module) :",
"Similarlyto Transformers,eachtoken isrepre-",
"nt",
"(self,d,d'):",
"sented as a vector of features, which undergo (non-",
"E",
"self.hypernetwork in - MLP([d, d, d'])",
"# learnable parameters",
"linear) transformations in multiple layers. MLP-",
"self.hypernetwork_out = MLP([d, d, d'])",
"Mixer employs two MLPs at each layer, one for",
"normalization",
"es training stability",
"mOrO",
"tlaver",
"self.layer norm = LayerNorm(d)",
"feature mixing and one for token mixing. The",
"def forward(self, queries, keys, values):",
"feature mixing component is applied to each to-",
"#queries:[B,M, d",
"#keys / values",
"! [B, N, d]",
"ken vector independently, which models the in-",
"teractions between features. The Token Mixing",
"# add token information (e.g.position",
"hyp_in = add_token informat",
"ion(keys)",
"MLP (TM-MLP) is applied to each feature inde-",
"hyp_out = add_token_information(queries)",
" ",
"pendently (i.e. its vector of values across tokens),",
"I",
"which models the interactions between spatial lo-",
"# TM-MLP(X) =",
"W2GELU",
"1^T ×))",
"cations or positions. This could be interpreted as",
"#maps[B,d,N]",
"token_mixing_mlp = compose_TM MLP (Wl, W2)",
">BM",
"a global attention mechanism which is static and",
"+transpose",
"applie",
"euen",
"position-modulated. Practically, this is achieved",
"ose(1,2)[B,d,N]",
"outpu",
"ixing_mlp(values)#[B,d,M]",
"tures and the dimension representing the positions.",
"by transposing the dimension representing the fea-",
"utput",
"output.t",
"T",
"spose(1,2)#[B,M,d",
"Each vector xT e RN, representing feature i ≤ d,",
"# option",
"of some input of fixed length N, is input into",
"nself.lave",
"rm(output)",
"TM-MLP, which has the following form:",
"TM-MLP(xT) = W1(α(WTxT),",
"HyperMixer includes systematicity into the",
"(1)",
"MLPMixer architecture by introducing a novel to-",
"where Wi,W2 e RNxd', and represents",
"ken mixing mechanism, HyperMixing², which can",
"the GELU non-linearity (Hendrycks and Gimpel,",
"be regarded as a drop-in replacement for attention",
"2016). Finally, to facilitate learning, layer normal-",
"Foreaseof understanding,weprovidepseudo-code",
"in Algorithm 1. While the queries, keys, and val",
"ization (Ba et al., 2016) and skip connections (He",
"ues in HyperMixing need not be the same, we will",
"et al., 2016) are added around each MLP, respec-",
"tively. How to best arrange these components is",
"2HyperMixing is to HyperMixer what self-attention is to",
"still an open question (Wang et al., 2019; Bach-",
"Transformer encoders."
] |
{
"arxiv_id": "2203.03691",
"language": "en",
"page_idx": 2
}
|
||
2203.03691-en-3
|
2203.03691-en-3.jpg
|
assume they are identical in the following formulation. HyperMixing relies on the use of hypernetworks, which are used to generate the weights $\mathbf{W}_1,\mathbf{W}_2$ of TM-MLP (Equation 1) dynamically as a function of the input. Let $\mathbf{x}_j\in\tilde{\mathbb{R}}^d,j\leq\tilde{N}$, where $N$ is the (variable) dimension of the input, represent token $j$ (i.e., query, key, and value). W$_1$ and $\mathbf{W}_2$ are generated by parameterized functions $h_{1},h_{2}:\mathbb{R}^{N\times d}\to\mathbb{R}^{N\times\tilde{d}^{\prime}}.$Theoretically$,h_1$ and$h_{2}$ could be any function, including sophisticated networks that consider non-linear interactions between tokens, such as the attention mechanism. However, this would defeat the purpose of our model, which is simplicity. Therefore, we choose to generate the rows of the weight matrices from each token independently via another MLP. Concretely, a hypernetwork function can be defined as
$$h_i(\mathbf{x})=\left(\begin{array}{c}\mathrm{MLP}^{\mathbf{W_i}}(\mathbf{x}_1+\mathbf{p}_1)\\\vdots\\\mathrm{MLP}^{\mathbf{W_i}}(\mathbf{x}_N+\mathbf{p}_N)\end{array}\right)\in\mathbb{R}^{N\times d'},$$
where MLP$^\mathbf{W}_1$,MLP$^\mathbf{W}_2:\mathbb{R}^d\to\mathbb{R}^{d^{\prime}}$are themselves multi-layer perceptrons with GELU nonlinearity. $\mathbf{p}_j\in\mathbb{R}^{\hat d}$ is a vector that can encode additional information such as the position via absolute position embeddings (Vaswani et al., 2017). Intuitively, for each token $\mathbf{x}_j,h_1$ decides which information to send to the hidden layer of TM-MLP, where the information from all tokens are mixed, and $h_2$ decides for each token how to extract information from the hidden layer. Note that, even though $h_1$ and $h_2$ only consider one token at once, non-linear interactions between tokens are still modeled through the hidden layer of TM- MLP.
Finally, layer normalization (Ba et al., 2016) can be applied to the output of TM-MLP. We found this helpful to facilitate training with a wide variety of Transformer layouts (Appendix F).
Tying $h_1\textbf{and}h_2$ In order to reduce the number
of parameters and operations in the model, and thereby the complexity, we found it useful to tie $h_1$ and $h_2$ by setting $\mathbf{W}_2=\mathbf{W}_1.$
Considerations for NLP In comparison to the MLPMixer defined in Section 2.2, the use of hypernetworks overcomes two challenges. Firstly, the input no longer has to be of fxed dimensionality. The hypernetwork generates a token mixing MLP of appropriate dimension as a function of the input. Secondly, the hypernetwork models the interaction
between tokens with shared weights across all positions in the input. Hence, systematicity is ensured.
### 3 $\textbf{Related Work}$
Research on all-MLP models like MLPMixer (Tolstikhin et al., 2021) is widespread in the computer vision community (Tu et al., 2022; Yu et al., 2022; Wang et al., 2022, among many others). However, they lack some desirable inductive biases for NLP, which we discuss in length in Appendix A.2. Specifically, in contrast to HyperMixer, none of the previously proposed methods simultaneously provide i) position invariance, which is important for generalization, ii) adaptive size for variable-length inputs, iii) a global receptive freld, which allows interactions to not be limited to small token neighborhoods, iv) learnabilty allowing for universal applicablility to various tasks, and v) dynamicity, which means that token mixing is a function of the input. Consequently, only a few works have used MLP-based models as their backbone in NLP tasks. gMLP (Liu et al., 2021) serves as one of our baselines and pnlp-mixer (Fusco et al., 2022) employs standard MLPMixer on top of a novel token embedding method.
Apart from all-MLP models, there is an abundance of research on efficient alternatives to standard attention layers (Katharopoulos et al., 2020; Bello, 2021, et cetera). While they don't qualify as all-MLP models, they have close connections to our work (see Appendix E) and aim at lowering the cost of AI, albeit it on fewer dimensions than our work (Appendix A.1). We employ FNet (Lee-Thorp et al., 2021) and Linear Transformers (Katharopoulos et al., 2020) as representatives of these as a baseline.
### 4 Experiments
Our experiments are designed to test the following three hypotheses. H1 (Section 4.3): Since HyperMixer reflects more inductive biases that are adequate for NLP, our hypothesis is that HyperMixer performs better at NLP tasks than MLPMixer and similar MLP-based alternatives, specifically at those tasks that require to model the interactions between tokens. H2: Since HyperMixer has similar inductive biases as transformers but is considerably simpler conceptually and in terms of computational complexity, it can be seen as a low cost alternative to Transformers, reducing the cost in terms of single example processing time
|
[
[
0.1175537109375,
0.08489990234375,
0.49365234375,
0.08489990234375,
0.49365234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1158447265625,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.5087890625,
0.1009521484375,
0.8818359375,
0.1015625,
0.8818359375,
0.11639404296875,
0.5087890625,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.486083984375,
0.1341552734375,
0.486083984375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51318359375,
0.1300048828125,
0.6650390625,
0.1300048828125,
0.6650390625,
0.1436767578125,
0.51318359375,
0.1436767578125
],
[
0.1175537109375,
0.149658203125,
0.489501953125,
0.149658203125,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.509765625,
0.155029296875,
0.880859375,
0.155029296875,
0.880859375,
0.168701171875,
0.509765625,
0.168701171875
],
[
0.1158447265625,
0.1650390625,
0.486083984375,
0.1650390625,
0.486083984375,
0.1787109375,
0.1158447265625,
0.1787109375
],
[
0.51123046875,
0.17333984375,
0.88330078125,
0.17333984375,
0.88330078125,
0.18701171875,
0.51123046875,
0.18701171875
],
[
0.1158447265625,
0.182861328125,
0.48779296875,
0.182861328125,
0.48779296875,
0.196533203125,
0.1158447265625,
0.196533203125
],
[
0.5087890625,
0.188232421875,
0.8818359375,
0.188232421875,
0.8818359375,
0.201904296875,
0.5087890625,
0.201904296875
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.5087890625,
0.2037353515625,
0.88330078125,
0.204833984375,
0.88330078125,
0.2197265625,
0.5087890625,
0.218505859375
],
[
0.1158447265625,
0.2120361328125,
0.489501953125,
0.2120361328125,
0.489501953125,
0.2298583984375,
0.1158447265625,
0.2298583984375
],
[
0.509765625,
0.221435546875,
0.88330078125,
0.221435546875,
0.88330078125,
0.235107421875,
0.509765625,
0.235107421875
],
[
0.1158447265625,
0.2298583984375,
0.490234375,
0.23095703125,
0.490234375,
0.2452392578125,
0.1158447265625,
0.24462890625
],
[
0.51123046875,
0.236328125,
0.8818359375,
0.2369384765625,
0.8818359375,
0.251708984375,
0.51123046875,
0.25048828125
],
[
0.1158447265625,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1158447265625,
0.26123046875
],
[
0.51123046875,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51123046875,
0.267333984375
],
[
0.1158447265625,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.51123046875,
0.26953125,
0.88330078125,
0.26953125,
0.88330078125,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.1175537109375,
0.279052734375,
0.48779296875,
0.279052734375,
0.48779296875,
0.29345703125,
0.1175537109375,
0.29345703125
],
[
0.51318359375,
0.2861328125,
0.87890625,
0.2861328125,
0.87890625,
0.29638671875,
0.51318359375,
0.29638671875
],
[
0.11419677734375,
0.29443359375,
0.484375,
0.29443359375,
0.484375,
0.30810546875,
0.11419677734375,
0.30810546875
],
[
0.51123046875,
0.300537109375,
0.87890625,
0.300537109375,
0.87890625,
0.314208984375,
0.51123046875,
0.314208984375
],
[
0.1175537109375,
0.31005859375,
0.486083984375,
0.31005859375,
0.486083984375,
0.32373046875,
0.1175537109375,
0.32373046875
],
[
0.51123046875,
0.317626953125,
0.8818359375,
0.31640625,
0.8818359375,
0.330810546875,
0.51123046875,
0.33203125
],
[
0.1158447265625,
0.325927734375,
0.48779296875,
0.32666015625,
0.48779296875,
0.341552734375,
0.1158447265625,
0.34033203125
],
[
0.51123046875,
0.333251953125,
0.880859375,
0.333251953125,
0.880859375,
0.3466796875,
0.51123046875,
0.3466796875
],
[
0.11419677734375,
0.34326171875,
0.399658203125,
0.342529296875,
0.399658203125,
0.35693359375,
0.11419677734375,
0.358154296875
],
[
0.51318359375,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.3623046875,
0.51318359375,
0.3623046875
],
[
0.2225341796875,
0.364013671875,
0.364501953125,
0.36767578125,
0.364501953125,
0.385498046875,
0.2225341796875,
0.3818359375
],
[
0.51123046875,
0.366455078125,
0.8818359375,
0.365234375,
0.8818359375,
0.380126953125,
0.51123046875,
0.38134765625
],
[
0.133544921875,
0.388427734375,
0.1822509765625,
0.388427734375,
0.1822509765625,
0.402587890625,
0.133544921875,
0.402587890625
],
[
0.51123046875,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.391357421875,
0.3896484375,
0.462646484375,
0.380126953125,
0.466796875,
0.3955078125,
0.3955078125,
0.405029296875
],
[
0.2149658203125,
0.402587890625,
0.376220703125,
0.40625,
0.376220703125,
0.424560546875,
0.2149658203125,
0.421630859375
],
[
0.51123046875,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.412109375,
0.51123046875,
0.412109375
],
[
0.51123046875,
0.41455078125,
0.8818359375,
0.41455078125,
0.8818359375,
0.427490234375,
0.51123046875,
0.427490234375
],
[
0.1158447265625,
0.432861328125,
0.490234375,
0.432861328125,
0.490234375,
0.45068359375,
0.1158447265625,
0.45068359375
],
[
0.51318359375,
0.429931640625,
0.88330078125,
0.429931640625,
0.88330078125,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.51123046875,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.1175537109375,
0.453125,
0.489501953125,
0.453125,
0.489501953125,
0.46728515625,
0.1175537109375,
0.46728515625
],
[
0.51318359375,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.1158447265625,
0.468505859375,
0.48779296875,
0.46728515625,
0.48779296875,
0.482177734375,
0.1158447265625,
0.482666015625
],
[
0.51318359375,
0.478515625,
0.66162109375,
0.478515625,
0.66162109375,
0.4921875,
0.51318359375,
0.4921875
],
[
0.1175537109375,
0.485107421875,
0.489501953125,
0.485107421875,
0.489501953125,
0.498779296875,
0.1175537109375,
0.498779296875
],
[
0.5283203125,
0.49462890625,
0.88330078125,
0.493408203125,
0.88330078125,
0.50830078125,
0.5283203125,
0.5087890625
],
[
0.1158447265625,
0.50048828125,
0.486083984375,
0.50048828125,
0.486083984375,
0.51416015625,
0.1158447265625,
0.51416015625
],
[
0.133544921875,
0.5166015625,
0.486083984375,
0.51708984375,
0.486083984375,
0.5322265625,
0.133544921875,
0.53076171875
],
[
0.51318359375,
0.51123046875,
0.88330078125,
0.51123046875,
0.88330078125,
0.52490234375,
0.51318359375,
0.52490234375
],
[
0.51318359375,
0.52685546875,
0.8818359375,
0.52685546875,
0.8818359375,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.1175537109375,
0.53369140625,
0.483642578125,
0.53369140625,
0.483642578125,
0.54443359375,
0.1175537109375,
0.54443359375
],
[
0.51318359375,
0.54345703125,
0.8818359375,
0.54345703125,
0.8818359375,
0.55712890625,
0.51318359375,
0.55712890625
],
[
0.1175537109375,
0.5498046875,
0.486083984375,
0.5498046875,
0.486083984375,
0.5634765625,
0.1175537109375,
0.5634765625
],
[
0.51318359375,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.118408203125,
0.56787109375,
0.486083984375,
0.56787109375,
0.486083984375,
0.57763671875,
0.118408203125,
0.57763671875
],
[
0.51123046875,
0.576171875,
0.8818359375,
0.57470703125,
0.8818359375,
0.5888671875,
0.51123046875,
0.59033203125
],
[
0.1175537109375,
0.58203125,
0.48779296875,
0.58203125,
0.48779296875,
0.595703125,
0.1175537109375,
0.595703125
],
[
0.51318359375,
0.59130859375,
0.8818359375,
0.59130859375,
0.8818359375,
0.60498046875,
0.51318359375,
0.60498046875
],
[
0.1175537109375,
0.5986328125,
0.489501953125,
0.5986328125,
0.489501953125,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.51123046875,
0.60791015625,
0.880859375,
0.60791015625,
0.880859375,
0.62158203125,
0.51123046875,
0.62158203125
],
[
0.1158447265625,
0.6142578125,
0.489501953125,
0.6142578125,
0.489501953125,
0.62744140625,
0.1158447265625,
0.62744140625
],
[
0.1158447265625,
0.62890625,
0.48779296875,
0.62890625,
0.48779296875,
0.642578125,
0.1158447265625,
0.642578125
],
[
0.51318359375,
0.62353515625,
0.8818359375,
0.62353515625,
0.8818359375,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.51123046875,
0.64013671875,
0.8818359375,
0.64013671875,
0.8818359375,
0.65380859375,
0.51123046875,
0.65380859375
],
[
0.118408203125,
0.646484375,
0.198974609375,
0.646484375,
0.198974609375,
0.65771484375,
0.118408203125,
0.65771484375
],
[
0.133544921875,
0.6611328125,
0.484375,
0.66015625,
0.484375,
0.6748046875,
0.133544921875,
0.67578125
],
[
0.51123046875,
0.65771484375,
0.580078125,
0.65771484375,
0.580078125,
0.66943359375,
0.51123046875,
0.66943359375
],
[
0.1158447265625,
0.677734375,
0.486083984375,
0.677734375,
0.486083984375,
0.69140625,
0.1158447265625,
0.69140625
],
[
0.5146484375,
0.689453125,
0.52294921875,
0.689453125,
0.52294921875,
0.69482421875,
0.5146484375,
0.69482421875
],
[
0.53662109375,
0.689453125,
0.54150390625,
0.689453125,
0.54150390625,
0.69482421875,
0.53662109375,
0.69482421875
],
[
0.5400390625,
0.68505859375,
0.654296875,
0.68505859375,
0.654296875,
0.69970703125,
0.5400390625,
0.69970703125
],
[
0.1158447265625,
0.693359375,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.1158447265625,
0.70703125
],
[
0.1175537109375,
0.71142578125,
0.393798828125,
0.71142578125,
0.393798828125,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.51123046875,
0.71044921875,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.51123046875,
0.724609375
],
[
0.51123046875,
0.72705078125,
0.8818359375,
0.7255859375,
0.8818359375,
0.74072265625,
0.51123046875,
0.7412109375
],
[
0.118408203125,
0.7353515625,
0.486083984375,
0.7353515625,
0.486083984375,
0.7490234375,
0.118408203125,
0.7490234375
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.7421875,
0.8818359375,
0.75732421875,
0.51123046875,
0.75830078125
],
[
0.1175537109375,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.765625,
0.1175537109375,
0.765625
],
[
0.51123046875,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.1175537109375,
0.76708984375,
0.48779296875,
0.76708984375,
0.48779296875,
0.78076171875,
0.1175537109375,
0.78076171875
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.1158447265625,
0.78369140625,
0.33154296875,
0.78369140625,
0.33154296875,
0.79736328125,
0.1158447265625,
0.79736328125
],
[
0.509765625,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.509765625,
0.80419921875
],
[
0.1175537109375,
0.80712890625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82080078125,
0.1175537109375,
0.82080078125
],
[
0.509765625,
0.80810546875,
0.88330078125,
0.80810546875,
0.88330078125,
0.82177734375,
0.509765625,
0.82177734375
],
[
0.1158447265625,
0.822265625,
0.490234375,
0.82373046875,
0.490234375,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.5087890625,
0.822265625,
0.8818359375,
0.82373046875,
0.8818359375,
0.83837890625,
0.5087890625,
0.83740234375
],
[
0.1158447265625,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.11419677734375,
0.85498046875,
0.490234375,
0.85546875,
0.490234375,
0.87060546875,
0.11419677734375,
0.869140625
],
[
0.51123046875,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.8681640625,
0.51123046875,
0.8681640625
],
[
0.1158447265625,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.5087890625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.5087890625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.5087890625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.5087890625,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.5087890625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.5087890625,
0.91845703125
]
] |
[
"assume they are identical in the following formu-",
"between tokens with shared weights across all posi",
"lation. HyperMixing relies on the use of hyper-",
"tions in the input. Hence, systematicity is ensured.",
"networks, which are used to generate the weights",
"W1, W2 of TM-MLP (Equation 1) dynamically",
"3Related Work",
"as a function of the input. Let x, E Rd, j ≤ N,",
"Research on all-MLP models like MLPMixer (Tol",
"where N is the (variable) dimension of the input.",
"stikhin et al., 2021) is widespread in the computer",
"represent token j (i.e., query, key, and value). Wi",
" vision community (Tu et al., 2022: Yu et al., 2022:",
"and W2 are generated by parameterized functions",
"Wang et al., 2022, among many others). How-",
"hi, h2 : RNxd → RNxd'. Theoretically, hi and h2",
" ever, they lack some desirable inductive biases for",
"could be any function, including sophisticated net-",
"NL, which we discuss in length in Appendix A.2",
"works that consider non-linear interactions between",
"Specifically, in contrast to HyperMixer, none of the",
"tokens, such as the attention mechanism. However,",
"previously proposed methods simultaneously pro-",
"this would defeat the purpose of our model, which",
"videipositioninvariancewhichisimportantfor",
"is simplicity. Therefore, we choose to generate",
"generalization, ii) adaptive size for variable-length",
"the rows of the weight matrices from each token",
"inputs, ii) a global receptive field, which allows",
"independently via another MLP. Concretely, a hy-",
"interactions to not be limited to small token neigh-",
"pernetwork function can be defined as",
"borhoods,iv)learnabilty allowing for universal",
"MLPWi(x1 + Pi)",
"applicablility to various tasks, and v) dynamicity,",
"hi(x) :",
"which means that token mixing is a function of",
"ERNxd'",
"MLPWi(xN +PN)",
"the input. Consequently, only a few works have",
"used MLP-based models as their backbone in NLP",
"where MLPWi, MLPW2 : Rd → Rd' are them-",
"tasks. gMLP (Liu et al., 2021) serves as one of our",
"baselines and pnlp-mixer (Fusco et al., 2022) em-",
"selves multi-layer perceptrons with GELU non-",
"ploys standard MLPMixer on top of a novel token",
"linearity. p, E IRd is a vector that can encode",
"embedding method.",
"additional information such as the position via ab-",
"Apart from all-MLP models, there is an abun-",
"solute position embeddings (Vaswani et al., 2017)",
"Intuitively,for each token Xj, hi decides",
"dance of research on efficient alternatives to stan-",
"dard attention layers (Katharopoulos et al., 2020",
"which information to send tothe hidden laver of",
"Bello, 2021, et cetera). While they don't qualify as",
"TM-MLP, where the information from all tokens",
"all-MLP models, they have close connections to our",
"aremixedandh2decidesforeachtokenhowto",
"work (see Appendix E) and aim at lowering the cost",
"extract information from the hidden layer. Note",
"of AI, albeit it on fewer dimensions than our work",
"that, even though hi and h2 only consider one to-",
"(Appendix A.1). We employ FNet (Lee-Thorp",
"ken at once, non-linear interactions between to-",
"kens are still modeled through the hidden layer of",
"et al., 2021) and Linear Transformers (Katharopou",
"los et al., 2020) as representatives of these as a",
"TM-MLP.",
"Finally, layer normalization (Ba et al., 2016) can",
"baseline.",
"be applied to the output of TM-MLP. We found",
"A",
"",
"Experiments",
"this helpful to facilitate training with a wide variety",
"of Transformer layouts (Appendix F)",
"Our experiments are designed to test the following",
"three hypotheses. H1 (Section 4.3): Since Hy",
"Tying hi and h2In order to reduce the number",
"perMixer reflects more inductive biases that are",
"of parameters and operations in the model, and",
"adequate for NLP, our hypothesis is that Hyper-",
"thereby the complexity, we found it useful to tie hi",
"Mixer performs better at NLP tasks than MLP-",
"and h2 by setting W2 = W1.",
"Mixer and similar MLP-based alternatives, specif",
"Considerations for NLP In comparison to the",
"ically at those tasks that require to model the in-",
"MLPMixer defined in Section 2.2, the use of hyper-",
"teractions between tokens. H2: Since HyperMixer",
"networks overcomes two challenges. Firstly, the",
"has similar inductive biases as transformers but",
"input no longer has to be of fixed dimensionality.",
"isconsiderablysimplerconceptuallyandinterms",
"The hypernetwork generates a token mixing MLP",
" of computational complexity, it can be seen as a",
"of appropriate dimension as a function of the input.",
"low cost alternative to Transformers, reducing the",
"Secondly, the hypernetwork models the interaction",
" cost in terms of single example processing time"
] |
{
"arxiv_id": "2203.03691",
"language": "en",
"page_idx": 3
}
|
||
2203.03691-en-4
|
2203.03691-en-4.jpg
|
(Section 4.4), required dataset size (Section 4.5), and hyperparameter tuning (Section 4.6). H3 (Section 4.7): Due to its inductive biases mirroring those of Transformers, HyperMixer also learns similar patterns as the attention mechanism.
## 4.1 Datasets
We evaluate on four sentence-pair classification tasks and one single-sentence classification task. The sentence-pair tasks are QQP (Iyer et al., 2017), QNLI (Rajpurkar et al., 2016), MNLI (Williams et al., 2018) and SNLI (Bowman et al., 2015). For uniformity, datasets are formatted as in the GLUE benchmark (Wang et al., 2018). We choose these tasks for two properties: firstly, they have large training datasets (Table 2, appendix) enabling reasonable performances without pretraining; secondly, solving these tasks requires good modeling of the interactions between tokens from different sentences, which is the main focus of this paper. As a control, we experiment on the single-input dataset SST2 (Socher et al., 2013), which is a sentiment classification task. Many examples in this dataset can be solved by identifying key sentiment words, rather than modeling the token interaction.
## 4.2 Baselines
The following baselines can be categorized into MLP-based (to support H1) and not MLP-based (e.g., Transformers, to support H2). Note that our study is about the design of the token mixing module. Therefore, we only compare to models that fit into the general framework displayed in Figure 1, where there is a feature mixing module and a token mixing module for textual inputs. As a result, models such as RNNs are excluded. To enable a controlled experiment, we use the same feature mixing module in all models; the models only differ in their token mixing module.
MLP-based The conceptually closest baseline is MLPMixer (Tolstikhin et al., 2021), which combines both token and feature mixing using fixed dimensional MLPs, as described in Section 2.2. Concurrently, (Liu et al., 2021) proposed gMLP, in which token mixing is achieved through weighted summation of all other inputs, similar to the attention mechanism. However, rather than computing weights as function of the inputs like in attention, in gMLP the weights are fixed learnable parameters. Additionally, linear gating initialized close to one is introduced to facilitate training. The origi-
nal gMLP method does not employ feature mixing modules, as their token mixing module is capable of modeling feature interactions as well in a single gMLP block. However, for comparability we inject gMLP blocks as token mixing modules in our general architecture and keep feature mixing modules as well.
Non MLP-based Transformers (Vaswani et al., 2017) are used in the current state of the art in virtually all NLP tasks. Their key component is the softmax-based self-attention module, which we use for token mixing. Linear Transformer(Katharopoulos et al., 2020) replaces softmax attention with a featuremap based dot-product attention.
Finally,
FNet (Yu et al., 2021) replaces the self-attention part of Transformers with a fixed, non-learnable set of Fourier transforms for token mixing.
## 4.3 Performance
Initially we compare the performance of HyperMixer in comparison to our baselines. Thereafter, we further explore the model's benefits with respects to its cost.
For comparability, we adjust the size of the token mixing components such that all models have the same number of parameters (11M). FNet is an exception since it has no learnable parameters in its token mixing component. We tune the learning rate of each model via grid-search, and report the performance of the best configuration. Further experimental details on all experiments can be found in Appendix B.
Results Validation and test set results are shown in Table 1. On the test and the validation set, HyperMixer performs the best among MLP-based models on all datasets, although for SST the difference on the validation set is smaller than one standard deviation. MLPMixer generally achieves good performances, outperforming Transformers on two datasets.
Comparing to non-MLP-based methods, HyperMixer also outperforms vanilla Transformers on all datasets. The differences are generally small $(\leq2$ points), except on QNLI, where the difference is 3.9 points. We suspect that this discrepancy is due to the relatively small training set of QNLI. We investigate low-resource behavior of Transformers in comparison to HyperMixer in Section 4.5. FNet performs substantially worse than the other methods, particularly on SNLI and QQP. Linear Trans-
|
[
[
0.118408203125,
0.08489990234375,
0.880859375,
0.08489990234375,
0.880859375,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.491943359375,
0.1015625,
0.491943359375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.509765625,
0.11517333984375
],
[
0.1158447265625,
0.1175537109375,
0.490234375,
0.1181640625,
0.490234375,
0.1318359375,
0.1158447265625,
0.1312255859375
],
[
0.509765625,
0.1181640625,
0.87890625,
0.1181640625,
0.87890625,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.491943359375,
0.1341552734375,
0.491943359375,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.51123046875,
0.133056640625,
0.880859375,
0.133056640625,
0.880859375,
0.146728515625,
0.51123046875,
0.146728515625
],
[
0.1158447265625,
0.15087890625,
0.411376953125,
0.15087890625,
0.411376953125,
0.160888671875,
0.1158447265625,
0.160888671875
],
[
0.51318359375,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.163330078125,
0.51318359375,
0.163330078125
],
[
0.509765625,
0.1650390625,
0.880859375,
0.1650390625,
0.880859375,
0.1787109375,
0.509765625,
0.1787109375
],
[
0.1175537109375,
0.1776123046875,
0.2283935546875,
0.1776123046875,
0.2283935546875,
0.1923828125,
0.1175537109375,
0.1923828125
],
[
0.509765625,
0.1817626953125,
0.57275390625,
0.1817626953125,
0.57275390625,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.1175537109375,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.1175537109375,
0.213134765625
],
[
0.5087890625,
0.2060546875,
0.880859375,
0.2060546875,
0.880859375,
0.2197265625,
0.5087890625,
0.2197265625
],
[
0.1175537109375,
0.2161865234375,
0.48779296875,
0.2161865234375,
0.48779296875,
0.2298583984375,
0.1175537109375,
0.2298583984375
],
[
0.509765625,
0.22265625,
0.88330078125,
0.22265625,
0.88330078125,
0.236328125,
0.509765625,
0.236328125
],
[
0.1158447265625,
0.232177734375,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.1158447265625,
0.2464599609375
],
[
0.5087890625,
0.2369384765625,
0.8818359375,
0.2381591796875,
0.8818359375,
0.2529296875,
0.5087890625,
0.251708984375
],
[
0.118408203125,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.26318359375,
0.118408203125,
0.26318359375
],
[
0.509765625,
0.254638671875,
0.88330078125,
0.254638671875,
0.88330078125,
0.268310546875,
0.509765625,
0.268310546875
],
[
0.1175537109375,
0.26416015625,
0.489501953125,
0.26416015625,
0.489501953125,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.5087890625,
0.26953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.28515625,
0.5087890625,
0.283935546875
],
[
0.1175537109375,
0.28076171875,
0.489501953125,
0.28076171875,
0.489501953125,
0.29443359375,
0.1175537109375,
0.29443359375
],
[
0.51123046875,
0.2861328125,
0.884765625,
0.287353515625,
0.884765625,
0.300537109375,
0.51123046875,
0.2998046875
],
[
0.1175537109375,
0.29638671875,
0.48779296875,
0.29638671875,
0.48779296875,
0.31005859375,
0.1175537109375,
0.31005859375
],
[
0.51123046875,
0.302734375,
0.88330078125,
0.302734375,
0.88330078125,
0.31640625,
0.51123046875,
0.31640625
],
[
0.1175537109375,
0.3134765625,
0.48779296875,
0.3134765625,
0.48779296875,
0.32666015625,
0.1175537109375,
0.32666015625
],
[
0.51123046875,
0.320556640625,
0.79443359375,
0.320556640625,
0.79443359375,
0.330810546875,
0.51123046875,
0.330810546875
],
[
0.8193359375,
0.318359375,
0.8818359375,
0.320556640625,
0.8818359375,
0.33203125,
0.8193359375,
0.330078125
],
[
0.1158447265625,
0.328857421875,
0.48779296875,
0.328857421875,
0.48779296875,
0.34326171875,
0.1158447265625,
0.34326171875
],
[
0.51123046875,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.51123046875,
0.3486328125
],
[
0.1175537109375,
0.345703125,
0.48779296875,
0.345703125,
0.48779296875,
0.359375,
0.1175537109375,
0.359375
],
[
0.51318359375,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.51318359375,
0.364013671875
],
[
0.1158447265625,
0.361083984375,
0.486083984375,
0.361083984375,
0.486083984375,
0.374755859375,
0.1158447265625,
0.374755859375
],
[
0.51123046875,
0.366455078125,
0.8037109375,
0.36767578125,
0.8037109375,
0.3818359375,
0.51123046875,
0.38134765625
],
[
0.1175537109375,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.1175537109375,
0.39013671875
],
[
0.1175537109375,
0.393798828125,
0.490234375,
0.393798828125,
0.490234375,
0.40673828125,
0.1175537109375,
0.40673828125
],
[
0.51123046875,
0.391357421875,
0.6591796875,
0.392578125,
0.6591796875,
0.40673828125,
0.51123046875,
0.40625
],
[
0.118408203125,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.51123046875,
0.413330078125,
0.88330078125,
0.413330078125,
0.88330078125,
0.427490234375,
0.51123046875,
0.427490234375
],
[
0.1175537109375,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.1175537109375,
0.439453125
],
[
0.51123046875,
0.429931640625,
0.88330078125,
0.429931640625,
0.88330078125,
0.44482421875,
0.51123046875,
0.44482421875
],
[
0.1158447265625,
0.440673828125,
0.489501953125,
0.441162109375,
0.489501953125,
0.4560546875,
0.1158447265625,
0.454833984375
],
[
0.51123046875,
0.446533203125,
0.88330078125,
0.446533203125,
0.88330078125,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.1175537109375,
0.457763671875,
0.48779296875,
0.457763671875,
0.48779296875,
0.471435546875,
0.1175537109375,
0.471435546875
],
[
0.51318359375,
0.46435546875,
0.6396484375,
0.46435546875,
0.6396484375,
0.474365234375,
0.51318359375,
0.474365234375
],
[
0.1158447265625,
0.474365234375,
0.484375,
0.474365234375,
0.484375,
0.488037109375,
0.1158447265625,
0.488037109375
],
[
0.5283203125,
0.47802734375,
0.8818359375,
0.476806640625,
0.8818359375,
0.4912109375,
0.5283203125,
0.4921875
],
[
0.509765625,
0.493408203125,
0.8818359375,
0.4921875,
0.8818359375,
0.50732421875,
0.509765625,
0.50830078125
],
[
0.1175537109375,
0.501953125,
0.235107421875,
0.501953125,
0.235107421875,
0.515625,
0.1175537109375,
0.515625
],
[
0.51123046875,
0.51025390625,
0.880859375,
0.51025390625,
0.880859375,
0.52392578125,
0.51123046875,
0.52392578125
],
[
0.1175537109375,
0.52392578125,
0.48779296875,
0.52392578125,
0.48779296875,
0.53759765625,
0.1175537109375,
0.53759765625
],
[
0.51318359375,
0.52685546875,
0.8818359375,
0.52685546875,
0.8818359375,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.1175537109375,
0.5390625,
0.489501953125,
0.5390625,
0.489501953125,
0.55419921875,
0.1175537109375,
0.55419921875
],
[
0.5087890625,
0.54150390625,
0.8818359375,
0.5419921875,
0.8818359375,
0.55712890625,
0.5087890625,
0.5556640625
],
[
0.1175537109375,
0.5556640625,
0.489501953125,
0.5556640625,
0.489501953125,
0.57080078125,
0.1175537109375,
0.57080078125
],
[
0.509765625,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.509765625,
0.572265625
],
[
0.1175537109375,
0.572265625,
0.490234375,
0.572265625,
0.490234375,
0.5859375,
0.1175537109375,
0.5859375
],
[
0.51123046875,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.1175537109375,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6025390625,
0.1175537109375,
0.6025390625
],
[
0.51123046875,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.60498046875,
0.51123046875,
0.60498046875
],
[
0.1158447265625,
0.60498046875,
0.49365234375,
0.60498046875,
0.49365234375,
0.6181640625,
0.1158447265625,
0.6181640625
],
[
0.5087890625,
0.60693359375,
0.62890625,
0.60546875,
0.62890625,
0.62060546875,
0.5087890625,
0.62158203125
],
[
0.1175537109375,
0.62060546875,
0.490234375,
0.62060546875,
0.490234375,
0.63427734375,
0.1175537109375,
0.63427734375
],
[
0.51123046875,
0.6298828125,
0.8818359375,
0.6298828125,
0.8818359375,
0.64453125,
0.51123046875,
0.64453125
],
[
0.1158447265625,
0.6357421875,
0.48779296875,
0.63525390625,
0.48779296875,
0.6494140625,
0.1158447265625,
0.65087890625
],
[
0.5087890625,
0.6455078125,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.5087890625,
0.66015625
],
[
0.1175537109375,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.1175537109375,
0.66650390625
],
[
0.51123046875,
0.6630859375,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.51123046875,
0.677734375
],
[
0.118408203125,
0.67041015625,
0.484375,
0.67041015625,
0.484375,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.1158447265625,
0.68310546875,
0.48779296875,
0.68310546875,
0.48779296875,
0.69677734375,
0.1158447265625,
0.69677734375
],
[
0.51123046875,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.51123046875,
0.693359375
],
[
0.1175537109375,
0.70068359375,
0.36279296875,
0.70068359375,
0.36279296875,
0.71435546875,
0.1175537109375,
0.71435546875
],
[
0.51318359375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.51123046875,
0.71044921875,
0.88330078125,
0.7119140625,
0.88330078125,
0.72705078125,
0.51123046875,
0.724609375
],
[
0.1175537109375,
0.72705078125,
0.489501953125,
0.72705078125,
0.489501953125,
0.74072265625,
0.1175537109375,
0.74072265625
],
[
0.51123046875,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51123046875,
0.7412109375
],
[
0.1158447265625,
0.74365234375,
0.489501953125,
0.74365234375,
0.489501953125,
0.75732421875,
0.1158447265625,
0.75732421875
],
[
0.51123046875,
0.7412109375,
0.57666015625,
0.74365234375,
0.5751953125,
0.75830078125,
0.509765625,
0.755859375
],
[
0.1158447265625,
0.75830078125,
0.48779296875,
0.7587890625,
0.48779296875,
0.77392578125,
0.1158447265625,
0.7724609375
],
[
0.5283203125,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.7724609375,
0.5283203125,
0.7724609375
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.509765625,
0.7880859375
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.5087890625,
0.80712890625,
0.8818359375,
0.8056640625,
0.8818359375,
0.82080078125,
0.5087890625,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.491943359375,
0.82470703125,
0.491943359375,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1158447265625,
0.8388671875,
0.489501953125,
0.84033203125,
0.489501953125,
0.85498046875,
0.1158447265625,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.1158447265625,
0.85546875,
0.49365234375,
0.85693359375,
0.49365234375,
0.87060546875,
0.1158447265625,
0.869140625
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.49365234375,
0.87255859375,
0.49365234375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.490234375,
0.88818359375,
0.490234375,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.884765625,
0.90478515625,
0.884765625,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"(Section 4.4), required dataset size (Section 4.5),nal gMLP method does not employ feature mixing",
"and hyperparameter tuning (Section 4.6). H3 (Sec-",
"modules, as their token mixing module is capable",
"tion 4.7): Due to its inductive biases mirroring",
"of modeling feature interactions as well in a single",
"those of Transformers, HyperMixer also learns sim-",
"gMLP block.However,for comparability we inject",
"ilarpatternsasthe.attention.mechanism",
"gMLP blocks as token mixing modules in our gen-",
"eral architecture and keep feature mixing modules",
"4.1 Datasets",
"as well.",
"We evaluate on four sentence-pair classification",
"Non MLP-based Transformers (Vaswani",
"tasks and one single-sentence classification task.",
"et al., 2017) are used in the current state of",
"The sentence-pair tasks are QQP (yer et al., 2017),",
"the art in virtually all NLP tasks. Their key",
"QNLI (Rajpurkar et al., 2016), MNLI (Williams",
"component is the softmax-based self-attention",
"et al., 2018) and SNLI (Bowman et al., 2015).",
"module,which we use for token mixing",
"For uniformity, datasets are formatted as in the",
"Linear Transformer (Katharopoulos et al.,",
"GLUE benchmark (Wang et al., 2018). We choose",
"2020) replaces softmax attention with a feature-",
"these tasks for two properties: firstly, they have",
"mapbased dot-product attention.",
"Finally",
"large training datasets (Table 2, appendix) enabling",
"FNet (Yu et al., 2021) replaces the self-attention",
"reasonable performances without pretraining; sec-",
"part of Transformers with a fixed, non-learnable set",
"ondly, solving these tasks requires good modeling",
"of Fourier transforms fortoken mixing.",
"of the interactions between tokens from different",
"sentences, which is the main focus of this paper.",
"4.3Performance",
"As a control, we experiment on the single-input",
"Initially we compare the performance of Hyper-",
"dataset SST2 (Socher et al., 2013), which is a sen-",
"Mixer in comparison to our baselines. Thereafter,",
"timent classification task. Many examples in this",
"we further explore the model's benefits with re-",
"dataset can be solved by identifying key sentiment",
"spectstoitscost.",
"words, rather than modeling the token interaction.",
"For comparability, we adjust the size of the to-",
"ken mixing components such that all models have",
"4.2Baselines",
"the same number of parameters (11M). FNet is an",
"The following baselines can be categorized into",
"exception since it has no learnable parameters in",
"MLP-based (to support H1) and not MLP-based",
"its token mixing component. We tune the learning",
"(e.g., Transformers, to support H2). Note that our",
"rate of each model via grid-search, and report the",
"study is about the design of the token mixing mod-",
"performance of the best configuration. Further ex-",
"ule. Therefore, we only compare to models that fit",
"perimental details on all experiments can be found",
"into the general framework displayed in Figure 1,",
"in Appendix B.",
"where there is a feature mixing module and a to-",
"ResultsValidation and test set results are shown",
"ken mixing module for textual inputs. As a result,",
"in Table 1. On the test and the validation set, Hyper",
"models such as RNNs are excluded. To enable",
"Mixer performs the best among MLP-based mod-",
"acontrolled experiment,weusethe samefeature",
"mixing module in all models: the models only dif-",
"els on all datasets, although for SST the difference",
"fer in their token mixing module.",
"on the validation set is smaller than one standard",
"deviation. MLPMixer generally achieves good per-",
"MLP-basedThe conceptually closest baseline is",
"formances, outperforming Transformers on two",
"MLPMixer (Tolstikhin et al., 2021), which com-",
"datasets.",
"bines both token and feature mixing using fixed",
"Comparing to non-MLP-based methods, Hyper-",
"dimensional MLPs, as described in Section 2.2.",
"Mixer also outperforms vanilla Transformers on all",
"Concurrently, (Liu et al., 2021) proposed gMLP, in",
"datasets. The differences are generally small (< 2",
"which token mixing is achieved through weighted",
"points), except on QNLI, where the difference is",
"summation of all other inputs, similar to the atten-",
"3.9 points. We suspect that this discrepancy is due",
"tion mechanism. However, rather than computing",
"to the relatively small training set of QNLI. We",
"weights as function of the inputs like in attention,",
"investigate low-resource behavior of Transformers",
"in gMLP the weights are fixed learnable parame-",
"in comparison to HyperMixer in Section 4.5. FNet",
"ters. Additionally, linear gating initialized close to",
"performs substantially worse than the other meth-",
"one is introduced to facilitate training. The origi-",
" ods, particularly on SNLI and QQP. Linear Trans-"
] |
{
"arxiv_id": "2203.03691",
"language": "en",
"page_idx": 4
}
|
||
2402.08869-en-0
|
2402.08869-en-0.jpg
|
ScamSpot: Fighting Financial Fraud in Instagram Comments
Stefan Erben$^{1,2}$,Andreas Waldis$^{1,3}$
$^{1}$Lucerne University of Applied Sciences and Arts $^{2}$University of Applied Sciences Technikum Vienna
3Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science and
Hessian Center for AI (hessian.AI), Technical University of Darmstadt
<officeßstefanerben.com>
## Abstract
The long-standing problem of spam and fraudulent messages in the comment sections of Instagram pages in the financial sector claims new victims every day. Instagram's current spam filter proves $inadequate$, and existing research approaches are primarily confined to theoretical concepts. Practical implementations with evaluated results are missing. To solve this problem, we propose ScamSpot, a comprehensive system that includes a browser extension, a fine-tuned BERT model and a REST API. This approach ensures public accessibility of our results for Instagram users using the Chrome browser. Furthermore, we conduct a data annotation study, shedding light on the reasons and causes of the problem and evaluate the system through user feedback and comparison with existing models. ScamSpot is an open-source project and is publicly
available
at
https://scamspot.github.io/.
## 1 Introduction
Financial fraud has switched its medium - from phone calls and emails to social media (Ramli et al., 2023; Soomro& Hussain, 2019). A recent report from the U.S. Federal Trade Commission shows that the number of social media scams has soared in recent years and especially cryptocurrency scams are initiated on Instagram$^{1}.$
Last year, the topic also gained political attendance$^2$, but so far, Instagram users have not seen any improvements (Table 8) and industry
experts continue to voice their concerns about the spam and scam problem (Kerr et al., 2023). While spam is defined as unsolicited or unwanted content/comments (Hayati et al.,2010), scam is characterized as deceptive or fraudulent activity resulting in mostly financial loss for the victim (Liebau & & Schueffel, 2019). New and inexperienced investors in particular fall victim to targeted attacks, often losing significant sums of money in the process$^1.$
Instagram's existing spam filter has a precision of 98.36%, but only a recall of 11.51% (Table 6). Existing research mentioned in Section 6 made its first success in theoretical concepts and general spam detection, yet no practical solutions to detect financial spam and scam comments have been published. The problem has not been solved, as 90% of our survey participants expressed their dissatisfaction (Table 8), showcasing the urgency for a solution.
To close this gap, we show a way to efficiently classify comments with high precision and communicate the results to the user in real-time to improve the user experience and reduce the likelihood of fraud incidents. The solution is ScamSpot, a system designed to remove spam and scam comments from Instagram. More specifically, we contribute as follows:
, Dataset & Data Annotation: We compile what we believe to be the first large dataset of over 100,000 comments focusing specifically on Instagram comments of the financial space (Sections 2 and 3). We annotate over 3,000 comments as part of a data annotation study, which shows that
2
https://www.feinstein.senate.gov/
public/index.cfm/press-releases?TD= C51E17BC-D39D-4913-AA6C-09AB6B259083
$^{1}$ https://www.ftc.gov/news-
events/data-visualizations/dataspotlight/2022/06/reports-showscammers-cashing-crypto-craze
|
[
[
0.1705322265625,
0.086181640625,
0.8310546875,
0.086181640625,
0.8310546875,
0.1004638671875,
0.1705322265625,
0.1004638671875
],
[
0.347900390625,
0.12066650390625,
0.6513671875,
0.1182861328125,
0.6513671875,
0.1361083984375,
0.347900390625,
0.1385498046875
],
[
0.294189453125,
0.135009765625,
0.705078125,
0.1361083984375,
0.705078125,
0.1539306640625,
0.294189453125,
0.15283203125
],
[
0.289794921875,
0.1551513671875,
0.71240234375,
0.1551513671875,
0.71240234375,
0.16943359375,
0.289794921875,
0.16943359375
],
[
0.1302490234375,
0.168212890625,
0.86962890625,
0.16943359375,
0.86962890625,
0.187255859375,
0.1302490234375,
0.18603515625
],
[
0.212646484375,
0.1884765625,
0.7900390625,
0.1884765625,
0.7900390625,
0.2015380859375,
0.212646484375,
0.2015380859375
],
[
0.37646484375,
0.2181396484375,
0.6220703125,
0.2181396484375,
0.6220703125,
0.2318115234375,
0.37646484375,
0.2318115234375
],
[
0.26123046875,
0.2705078125,
0.342041015625,
0.2705078125,
0.342041015625,
0.285400390625,
0.26123046875,
0.285400390625
],
[
0.51171875,
0.2705078125,
0.88232421875,
0.268798828125,
0.88232421875,
0.282958984375,
0.51171875,
0.285400390625
],
[
0.51025390625,
0.285888671875,
0.88232421875,
0.285400390625,
0.88232421875,
0.299560546875,
0.51025390625,
0.30078125
],
[
0.1512451171875,
0.2978515625,
0.435302734375,
0.2978515625,
0.435302734375,
0.31103515625,
0.1512451171875,
0.31103515625
],
[
0.51171875,
0.302001953125,
0.88232421875,
0.30078125,
0.88232421875,
0.315185546875,
0.51171875,
0.31640625
],
[
0.148681640625,
0.313232421875,
0.43701171875,
0.313232421875,
0.43701171875,
0.32763671875,
0.148681640625,
0.32763671875
],
[
0.51171875,
0.318603515625,
0.88232421875,
0.318603515625,
0.88232421875,
0.331787109375,
0.51171875,
0.331787109375
],
[
0.148681640625,
0.32763671875,
0.43701171875,
0.32763671875,
0.43701171875,
0.34130859375,
0.148681640625,
0.34130859375
],
[
0.51171875,
0.3330078125,
0.87890625,
0.3330078125,
0.87890625,
0.3466796875,
0.51171875,
0.3466796875
],
[
0.148681640625,
0.340087890625,
0.43701171875,
0.34130859375,
0.43701171875,
0.356201171875,
0.148681640625,
0.35498046875
],
[
0.5126953125,
0.349609375,
0.880859375,
0.349609375,
0.880859375,
0.36328125,
0.5126953125,
0.36328125
],
[
0.148681640625,
0.356201171875,
0.43701171875,
0.356201171875,
0.43701171875,
0.369140625,
0.148681640625,
0.369140625
],
[
0.51171875,
0.364501953125,
0.880859375,
0.364501953125,
0.880859375,
0.378173828125,
0.51171875,
0.378173828125
],
[
0.148681640625,
0.370361328125,
0.23193359375,
0.370361328125,
0.23193359375,
0.38232421875,
0.148681640625,
0.38232421875
],
[
0.2462158203125,
0.37158203125,
0.282470703125,
0.37158203125,
0.282470703125,
0.3828125,
0.2462158203125,
0.3828125
],
[
0.300048828125,
0.37158203125,
0.36376953125,
0.37158203125,
0.36376953125,
0.3828125,
0.300048828125,
0.3828125
],
[
0.361328125,
0.375732421875,
0.374755859375,
0.375732421875,
0.374755859375,
0.3798828125,
0.361328125,
0.3798828125
],
[
0.371337890625,
0.37158203125,
0.435302734375,
0.37158203125,
0.435302734375,
0.38232421875,
0.371337890625,
0.38232421875
],
[
0.148681640625,
0.384033203125,
0.438720703125,
0.384033203125,
0.438720703125,
0.39892578125,
0.148681640625,
0.39892578125
],
[
0.5126953125,
0.381103515625,
0.88232421875,
0.381103515625,
0.88232421875,
0.394775390625,
0.5126953125,
0.394775390625
],
[
0.1495361328125,
0.3994140625,
0.435302734375,
0.3994140625,
0.435302734375,
0.4130859375,
0.1495361328125,
0.4130859375
],
[
0.5126953125,
0.397705078125,
0.884765625,
0.397705078125,
0.884765625,
0.411376953125,
0.5126953125,
0.411376953125
],
[
0.148681640625,
0.412109375,
0.435302734375,
0.4130859375,
0.435302734375,
0.427978515625,
0.148681640625,
0.4267578125
],
[
0.51171875,
0.4130859375,
0.67822265625,
0.4130859375,
0.67822265625,
0.4267578125,
0.51171875,
0.4267578125
],
[
0.1495361328125,
0.4267578125,
0.43701171875,
0.427978515625,
0.43701171875,
0.4423828125,
0.1495361328125,
0.441162109375
],
[
0.53173828125,
0.4287109375,
0.88232421875,
0.4287109375,
0.88232421875,
0.443603515625,
0.53173828125,
0.443603515625
],
[
0.1495361328125,
0.4423828125,
0.435302734375,
0.4423828125,
0.435302734375,
0.4560546875,
0.1495361328125,
0.4560546875
],
[
0.51171875,
0.443603515625,
0.880859375,
0.443603515625,
0.880859375,
0.457763671875,
0.51171875,
0.457763671875
],
[
0.1512451171875,
0.457763671875,
0.435302734375,
0.457763671875,
0.435302734375,
0.468505859375,
0.1512451171875,
0.468505859375
],
[
0.5126953125,
0.460205078125,
0.88232421875,
0.460205078125,
0.88232421875,
0.474365234375,
0.5126953125,
0.474365234375
],
[
0.1495361328125,
0.47021484375,
0.43701171875,
0.47021484375,
0.43701171875,
0.48388671875,
0.1495361328125,
0.48388671875
],
[
0.5126953125,
0.476806640625,
0.88232421875,
0.476806640625,
0.88232421875,
0.490478515625,
0.5126953125,
0.490478515625
],
[
0.148681640625,
0.485107421875,
0.43701171875,
0.48388671875,
0.43701171875,
0.498779296875,
0.148681640625,
0.5
],
[
0.5126953125,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.505859375,
0.5126953125,
0.505859375
],
[
0.1495361328125,
0.5,
0.435302734375,
0.5,
0.435302734375,
0.51318359375,
0.1495361328125,
0.51318359375
],
[
0.51171875,
0.50732421875,
0.88232421875,
0.50830078125,
0.88232421875,
0.521484375,
0.51171875,
0.52099609375
],
[
0.148681640625,
0.51416015625,
0.43701171875,
0.51416015625,
0.43701171875,
0.52783203125,
0.148681640625,
0.52783203125
],
[
0.51171875,
0.5224609375,
0.88232421875,
0.5224609375,
0.88232421875,
0.53759765625,
0.51171875,
0.53759765625
],
[
0.1495361328125,
0.529296875,
0.43701171875,
0.529296875,
0.43701171875,
0.5419921875,
0.1495361328125,
0.5419921875
],
[
0.148681640625,
0.5419921875,
0.43701171875,
0.5419921875,
0.43701171875,
0.5556640625,
0.148681640625,
0.5556640625
],
[
0.51171875,
0.5380859375,
0.88232421875,
0.5380859375,
0.88232421875,
0.5517578125,
0.51171875,
0.5517578125
],
[
0.1495361328125,
0.55712890625,
0.43701171875,
0.55712890625,
0.43701171875,
0.57080078125,
0.1495361328125,
0.57080078125
],
[
0.51025390625,
0.55419921875,
0.88232421875,
0.5546875,
0.88232421875,
0.5693359375,
0.51025390625,
0.568359375
],
[
0.1495361328125,
0.5712890625,
0.435302734375,
0.5712890625,
0.435302734375,
0.5849609375,
0.1495361328125,
0.5849609375
],
[
0.51171875,
0.57080078125,
0.61572265625,
0.57080078125,
0.61572265625,
0.58203125,
0.51171875,
0.58203125
],
[
0.148681640625,
0.58642578125,
0.43701171875,
0.58642578125,
0.43701171875,
0.60009765625,
0.148681640625,
0.60009765625
],
[
0.53173828125,
0.58642578125,
0.88232421875,
0.58642578125,
0.88232421875,
0.6005859375,
0.53173828125,
0.6005859375
],
[
0.279052734375,
0.6005859375,
0.34619140625,
0.59765625,
0.347900390625,
0.6123046875,
0.28076171875,
0.615234375
],
[
0.51025390625,
0.6005859375,
0.880859375,
0.6005859375,
0.880859375,
0.6142578125,
0.51025390625,
0.6142578125
],
[
0.421875,
0.60302734375,
0.435302734375,
0.60302734375,
0.435302734375,
0.6123046875,
0.421875,
0.6123046875
],
[
0.1495361328125,
0.6015625,
0.2066650390625,
0.6015625,
0.2066650390625,
0.61279296875,
0.1495361328125,
0.61279296875
],
[
0.148681640625,
0.615234375,
0.417724609375,
0.615234375,
0.417724609375,
0.62890625,
0.148681640625,
0.62890625
],
[
0.5126953125,
0.6181640625,
0.88232421875,
0.6181640625,
0.88232421875,
0.6318359375,
0.5126953125,
0.6318359375
],
[
0.51025390625,
0.63330078125,
0.880859375,
0.63330078125,
0.880859375,
0.64697265625,
0.51025390625,
0.64697265625
],
[
0.1142578125,
0.6484375,
0.12939453125,
0.6455078125,
0.1336669921875,
0.65283203125,
0.11846923828125,
0.65673828125
],
[
0.1419677734375,
0.6484375,
0.148681640625,
0.6484375,
0.148681640625,
0.65380859375,
0.1419677734375,
0.65380859375
],
[
0.1453857421875,
0.64453125,
0.262939453125,
0.64453125,
0.262939453125,
0.6591796875,
0.1453857421875,
0.6591796875
],
[
0.51416015625,
0.64990234375,
0.88330078125,
0.64990234375,
0.88330078125,
0.66357421875,
0.51416015625,
0.66357421875
],
[
0.5126953125,
0.6650390625,
0.88232421875,
0.6650390625,
0.88232421875,
0.6787109375,
0.5126953125,
0.6787109375
],
[
0.13525390625,
0.64990234375,
0.1419677734375,
0.64990234375,
0.1419677734375,
0.65283203125,
0.13525390625,
0.65283203125
],
[
0.115966796875,
0.671875,
0.4873046875,
0.671875,
0.4873046875,
0.68505859375,
0.115966796875,
0.68505859375
],
[
0.50830078125,
0.68017578125,
0.88232421875,
0.6806640625,
0.88232421875,
0.69580078125,
0.50830078125,
0.6943359375
],
[
0.115966796875,
0.6875,
0.4873046875,
0.6875,
0.4873046875,
0.701171875,
0.115966796875,
0.701171875
],
[
0.1168212890625,
0.70263671875,
0.4873046875,
0.70263671875,
0.4873046875,
0.71630859375,
0.1168212890625,
0.71630859375
],
[
0.51171875,
0.697265625,
0.70068359375,
0.697265625,
0.70068359375,
0.7109375,
0.51171875,
0.7109375
],
[
0.1168212890625,
0.71923828125,
0.48486328125,
0.71923828125,
0.48486328125,
0.73291015625,
0.1168212890625,
0.73291015625
],
[
0.53515625,
0.72607421875,
0.88232421875,
0.72607421875,
0.88232421875,
0.73876953125,
0.53515625,
0.73876953125
],
[
0.1168212890625,
0.73486328125,
0.4873046875,
0.73486328125,
0.4873046875,
0.74853515625,
0.1168212890625,
0.74853515625
],
[
0.546875,
0.7412109375,
0.88232421875,
0.7412109375,
0.88232421875,
0.7548828125,
0.546875,
0.7548828125
],
[
0.1168212890625,
0.751953125,
0.48486328125,
0.751953125,
0.48486328125,
0.76513671875,
0.1168212890625,
0.76513671875
],
[
0.54541015625,
0.7548828125,
0.88232421875,
0.75634765625,
0.88232421875,
0.7705078125,
0.54541015625,
0.76953125
],
[
0.115966796875,
0.767578125,
0.477294921875,
0.76611328125,
0.477294921875,
0.77978515625,
0.115966796875,
0.78125
],
[
0.546875,
0.7705078125,
0.88232421875,
0.7705078125,
0.88232421875,
0.78515625,
0.546875,
0.78515625
],
[
0.1361083984375,
0.78173828125,
0.486572265625,
0.78125,
0.486572265625,
0.79541015625,
0.1361083984375,
0.796875
],
[
0.548828125,
0.787109375,
0.88232421875,
0.787109375,
0.88232421875,
0.80078125,
0.548828125,
0.80078125
],
[
0.1168212890625,
0.79833984375,
0.486572265625,
0.79833984375,
0.486572265625,
0.81201171875,
0.1168212890625,
0.81201171875
],
[
0.548828125,
0.802734375,
0.88232421875,
0.802734375,
0.88232421875,
0.81640625,
0.548828125,
0.81640625
],
[
0.115966796875,
0.814453125,
0.48486328125,
0.814453125,
0.48486328125,
0.82763671875,
0.115966796875,
0.82763671875
],
[
0.548828125,
0.81884765625,
0.88330078125,
0.81884765625,
0.88330078125,
0.83154296875,
0.548828125,
0.83154296875
],
[
0.115966796875,
0.85791015625,
0.371337890625,
0.85791015625,
0.371337890625,
0.87158203125,
0.115966796875,
0.87158203125
],
[
0.55859375,
0.85791015625,
0.87890625,
0.85791015625,
0.87890625,
0.87158203125,
0.55859375,
0.87158203125
],
[
0.115966796875,
0.87255859375,
0.42529296875,
0.87255859375,
0.42529296875,
0.88623046875,
0.115966796875,
0.88623046875
],
[
0.5126953125,
0.87255859375,
0.849609375,
0.87255859375,
0.849609375,
0.88623046875,
0.5126953125,
0.88623046875
],
[
0.115966796875,
0.88720703125,
0.416015625,
0.88720703125,
0.416015625,
0.90087890625,
0.115966796875,
0.90087890625
],
[
0.51171875,
0.88623046875,
0.86279296875,
0.88525390625,
0.86279296875,
0.8994140625,
0.51171875,
0.90087890625
],
[
0.115966796875,
0.90185546875,
0.401611328125,
0.90185546875,
0.401611328125,
0.91552734375,
0.115966796875,
0.91552734375
]
] |
[
"ScamSpot: Fighting Financial Fraud in Instagram Comments",
"Stefan Erben',2, Andreas Waldis1,3",
"'Lucerne University of Applied Sciences and Arts",
"2University of Applied Sciences Technikum Vienna",
"3Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science and",
"Hessian Center for AI (hessian.Al), Technical University of Darmstadt",
"<office@stefanerben.com>",
"Abstract",
"experts continue to voice their concerns about the",
"spam and scam problem (Kerr et al., 2023). While",
"The long-standing problem of spam and",
"spam is defined as unsolicited or unwanted",
"fraudulent messages in the comment",
"content/comments (Hayati et al., 2010), scam is",
" sections of Instagram pages in the financial",
"characterized as deceptive or fraudulent activity",
"sector claims new victims every day.",
"resulting in mostly financial loss for the victim",
"Instagram's current spam filter proves",
"(Liebau &Schueffel.2019).New and",
"inadequate,",
"and",
"existing",
"",
"research",
"approaches are primarily confined to",
"inexperienced investors in particular fall victim to",
"theoretical concepts. Practical implemen-",
"targeted attacks, often losing significant sums of",
"tations with evaluated results are missing.",
"money in the process'.",
"Tosolvethis problem,we propose",
"Instagram's existing spam filter has a precision",
"ScamSpot, a comprehensive system that",
"of 98.36%, but only a recall of 11.51% (Table 6).",
"includes a browser extension,a fine-tuned",
"Existing research mentioned in Section 6 made its",
"BERT model and a REST API. This",
"first success in theoretical concepts and general",
"approach ensures public accessibility of our",
"spam detection, yet no practical solutions to detect",
"results for Instagram users using the",
"financial spam and scam comments have been",
"Chrome browser. Furthermore, we conduct",
"published. The problem has not been solved, as",
"a data annotation study, shedding light on",
"the reasons and causes of the problem and",
"90% of our survey participants expressed their",
"evaluate the system through user feedback",
"dissatisfaction (Table 8), showcasing the urgency",
"and comparison with existing models.",
"for a solution.",
"ScamSpot is an open-source project and is",
"To close this gap, we show a way to efficiently",
"available",
"classify comments with high precision and",
"at",
"publicly",
"https://scamspot.github.io/.",
"communicate the results to the user in real-time to",
"improve the user experience and reduce the",
"1",
"",
"Introduction",
"likelihood of fraud incidents. The solution is",
"ScamSpot, a system designed to remove spam and",
"",
"Financial fraud has switched its medium - from",
"scam comments from Instagram.More specifically,",
"phone calls and emails to social media (Ramli et",
"al., 2023; Soomro & Hussain, 2019). A recent",
"we contribute as follows:",
"report from the U.S. Federal Trade Commission",
"Dataset & Data Annotation:We compile",
"shows that the number of social media scams has",
"what we believe to be the first large dataset",
"soaredinrecentyearsandespecially",
"of over100,000commentsfocusing",
"cryptocurrency scams are initiated on Instagram.",
"specifically on Instagram comments of the",
"Last year, the topic also gained political",
"financial space (Sections 2 and 3). We",
"attendance?, but so far, Instagram users have not",
"annotate over 3,000 comments as part of a",
"seen any improvements (Table 8)and industry",
"data annotation study, which shows that",
"1 https://www.ftc.gov/news-",
"https://www.feinstein.senate.gov/",
"events/data-visualizations/data-",
"public/index.cfm/press-releases?ID-",
"spot1ight/2022/06/reports-show",
"C51E17BC-D39D-4913-AA6C-09AB6B259083",
"scammers-cashing-crypto-craze"
] |
{
"arxiv_id": "2402.08869",
"language": "en",
"page_idx": 0
}
|
||
2402.08869-en-2
|
2402.08869-en-2.jpg
|
## 3 Model Considerations
The next objective is to demonstrate an effective system rather than conduct an exhaustive analysis of various models. Nevertheless, we test a variety of models to identify an effective solution for detecting fraudulent Instagram comments. The selection of models includes traditional statistical models, large language models and BERT as a representative of the transformer models. The following specific models are selected:
$\bullet$ Statistical Models: We start with a linear
regression, a decision tree and a random forest model which provide us with a baseline and show that they lack the sophistication needed for our complex requirements.
Models
$(\mathbf{LLMs}):$
$\cdot$ Large Language
We assess the advanced natural language processing capabilities of GPT-3 (“gpt- 3.5-turbo") and GPT-4 (“gpt-4-1106- preview"). Their ability to understand and generate human-like text is a key consideration in our analysis. However, the lack of transparency and control over the training data of these models is a major limitation. Despite the adjustment of seed and temperature parameters (Table 5), the results are not deterministic. Three test runs are made for each model and the results can be found in Table 5. The best results are documented in Table l.
$\bullet$ Transformer Models: As a representative,
we select BERT (“bert-base-cased”) and fine-tune it based on the annotated data mentioned in Section 2. Its capability to understand context makes it a strong candidate in our selection.
In this study, we adopt a zero-shot approach with LLMs like GPT-3 and GPT-4, contrasting them with the fine-tuned BERT model. This methodology is chosen to demonstrate the practical usability of general-purpose LLMs in their standard configuration. While models like GPT-3 and GPT-4 show proficiency in general tasks, our findings align with those of Yu et al. (2023), when illustrating their limitations in specific tasks compared to fine-tuned models. This highlights the necessity of model selection tailored to task specificity and resource availability.
<table>
<tbody>
<tr>
<th> </th>
<th>$\mathbf{Recall}$</th>
<th>$\mathbf{Precision}$</th>
<th>$F1$</th>
</tr>
<tr>
<td>$IG$</td>
<td>0.1151</td>
<td>0.9836</td>
<td>0.2061</td>
</tr>
<tr>
<td>$DT$</td>
<td>0.8032</td>
<td>0.7692</td>
<td>0.7859</td>
</tr>
<tr>
<td>$RF$</td>
<td>0.8093</td>
<td>0.9244</td>
<td>0.8631</td>
</tr>
<tr>
<td>$LR$</td>
<td>0.8353</td>
<td>0.9163</td>
<td>0.8740</td>
</tr>
<tr>
<td>$\mathbf{GPT-3}$</td>
<td>0.3739</td>
<td>0.3660</td>
<td>0.3699</td>
</tr>
<tr>
<td>GPT- 4</td>
<td>0.6348</td>
<td>0.5530</td>
<td>0.5911</td>
</tr>
<tr>
<td>BERT</td>
<td>0.9213</td>
<td>0.9286</td>
<td>0.9249</td>
</tr>
</tbody>
</table>
Table 1: Model metrics. IG: Instagram's current spam filter, DT: Decision Tree, RF: Random Forest, LR: Linear Regression, BERT: Fined-tuned BERT
model from ScamSpot
Despite the hype around large language models, the results in Table 1 show that even established models like a fine-tuned BERT model can drastically outperform newer models like GPT-4 or GPT-3. While both BERT and LLMs are transformer-based, our research also demonstrates that a fine-tuned BERT model is more successful in specific tasks, which is also shown by Yu et al. (2023).
This leads us to the decision to select a finetuned BERT model for ScamSpot. The reasons also include:
$\bullet$ Promising Results: Tests show that the fine-
tuned BERT model far outperforms the other models in detecting fraudulent Instagram comments (Table 1).
$\bullet$ Stability and Predictability: Our fine-tuned
BERT model demonstrates stable and predictable performance, a crucial factor for consistent user experience compared to the tested LLMs (Table 5).
Local Execution: Unlike LLMs like GPT- 3.5 and GPT-4, which often require external cloud services potentially exposing data to third parties, BERT can be deployed internally, ensuring data privacy. Moreover, BERT's smaller model size leads to lower
an
important
energy
consumption,
consideration in sustainable AI development.
Adaptability to New Data: BERT's ability to understand context and meanings beyond just keyword matching makes it superior to simpler models like TF-IDF vectors. This adaptability is vital for detecting variations in
|
[
[
0.1319580078125,
0.08502197265625,
0.342041015625,
0.08502197265625,
0.342041015625,
0.09869384765625,
0.1319580078125,
0.09869384765625
],
[
0.603515625,
0.086181640625,
0.65869140625,
0.0880126953125,
0.6572265625,
0.10284423828125,
0.6025390625,
0.1004638671875
],
[
0.70068359375,
0.0880126953125,
0.7763671875,
0.0880126953125,
0.7763671875,
0.10284423828125,
0.70068359375,
0.10284423828125
],
[
0.78076171875,
0.0880126953125,
0.8076171875,
0.0880126953125,
0.8076171875,
0.1016845703125,
0.78076171875,
0.1016845703125
],
[
0.60498046875,
0.1046142578125,
0.65869140625,
0.1046142578125,
0.65869140625,
0.1195068359375,
0.60498046875,
0.1195068359375
],
[
0.1168212890625,
0.08740234375,
0.1319580078125,
0.08740234375,
0.1319580078125,
0.0963134765625,
0.1168212890625,
0.0963134765625
],
[
0.70166015625,
0.1046142578125,
0.75390625,
0.1046142578125,
0.75390625,
0.1195068359375,
0.70166015625,
0.1195068359375
],
[
0.78466796875,
0.1058349609375,
0.8359375,
0.1058349609375,
0.8359375,
0.12127685546875,
0.78466796875,
0.12127685546875
],
[
0.5234375,
0.10699462890625,
0.546875,
0.10699462890625,
0.546875,
0.12066650390625,
0.5234375,
0.12066650390625
],
[
0.1361083984375,
0.1112060546875,
0.4873046875,
0.1112060546875,
0.4873046875,
0.1248779296875,
0.1361083984375,
0.1248779296875
],
[
0.52197265625,
0.1236572265625,
0.552734375,
0.1236572265625,
0.552734375,
0.1402587890625,
0.52197265625,
0.1402587890625
],
[
0.60498046875,
0.1236572265625,
0.65869140625,
0.1236572265625,
0.65869140625,
0.1385498046875,
0.60498046875,
0.1385498046875
],
[
0.70166015625,
0.1236572265625,
0.75390625,
0.1236572265625,
0.75390625,
0.1385498046875,
0.70166015625,
0.1385498046875
],
[
0.78466796875,
0.1236572265625,
0.8369140625,
0.1236572265625,
0.8369140625,
0.1385498046875,
0.78466796875,
0.1385498046875
],
[
0.52197265625,
0.1427001953125,
0.55126953125,
0.1427001953125,
0.55126953125,
0.1593017578125,
0.52197265625,
0.1593017578125
],
[
0.115966796875,
0.1278076171875,
0.4873046875,
0.1265869140625,
0.4873046875,
0.1414794921875,
0.115966796875,
0.1427001953125
],
[
0.60498046875,
0.1427001953125,
0.65869140625,
0.1427001953125,
0.65869140625,
0.156982421875,
0.60498046875,
0.156982421875
],
[
0.78466796875,
0.1427001953125,
0.8369140625,
0.1427001953125,
0.8369140625,
0.156982421875,
0.78466796875,
0.156982421875
],
[
0.703125,
0.143310546875,
0.75390625,
0.143310546875,
0.75390625,
0.1551513671875,
0.703125,
0.1551513671875
],
[
0.1168212890625,
0.143310546875,
0.486572265625,
0.143310546875,
0.486572265625,
0.156982421875,
0.1168212890625,
0.156982421875
],
[
0.1168212890625,
0.1593017578125,
0.4873046875,
0.1593017578125,
0.4873046875,
0.17236328125,
0.1168212890625,
0.17236328125
],
[
0.60498046875,
0.159912109375,
0.65869140625,
0.159912109375,
0.65869140625,
0.176025390625,
0.60498046875,
0.176025390625
],
[
0.70166015625,
0.159912109375,
0.75390625,
0.159912109375,
0.75390625,
0.176025390625,
0.70166015625,
0.176025390625
],
[
0.78466796875,
0.159912109375,
0.8388671875,
0.159912109375,
0.8388671875,
0.176025390625,
0.78466796875,
0.176025390625
],
[
0.1168212890625,
0.176025390625,
0.486572265625,
0.176025390625,
0.486572265625,
0.1890869140625,
0.1168212890625,
0.1890869140625
],
[
0.52197265625,
0.162353515625,
0.552734375,
0.162353515625,
0.552734375,
0.178955078125,
0.52197265625,
0.178955078125
],
[
0.60693359375,
0.18017578125,
0.65869140625,
0.18017578125,
0.65869140625,
0.1944580078125,
0.60693359375,
0.1944580078125
],
[
0.70166015625,
0.18017578125,
0.75537109375,
0.18017578125,
0.75537109375,
0.1944580078125,
0.70166015625,
0.1944580078125
],
[
0.7841796875,
0.18017578125,
0.8369140625,
0.18017578125,
0.8369140625,
0.1944580078125,
0.7841796875,
0.1944580078125
],
[
0.1168212890625,
0.1903076171875,
0.4873046875,
0.1903076171875,
0.4873046875,
0.2039794921875,
0.1168212890625,
0.2039794921875
],
[
0.5234375,
0.1807861328125,
0.57568359375,
0.1807861328125,
0.57568359375,
0.19677734375,
0.5234375,
0.19677734375
],
[
0.60498046875,
0.1973876953125,
0.65869140625,
0.1973876953125,
0.65869140625,
0.21337890625,
0.60498046875,
0.21337890625
],
[
0.70166015625,
0.1973876953125,
0.75390625,
0.1973876953125,
0.75390625,
0.21337890625,
0.70166015625,
0.21337890625
],
[
0.78466796875,
0.1973876953125,
0.8369140625,
0.1973876953125,
0.8369140625,
0.21337890625,
0.78466796875,
0.21337890625
],
[
0.115966796875,
0.2069091796875,
0.489013671875,
0.2069091796875,
0.489013671875,
0.2205810546875,
0.115966796875,
0.2205810546875
],
[
0.52197265625,
0.1986083984375,
0.57568359375,
0.1986083984375,
0.57568359375,
0.21337890625,
0.52197265625,
0.21337890625
],
[
0.60498046875,
0.2176513671875,
0.65869140625,
0.2176513671875,
0.65869140625,
0.2318115234375,
0.60498046875,
0.2318115234375
],
[
0.70068359375,
0.2176513671875,
0.75390625,
0.2176513671875,
0.75390625,
0.2318115234375,
0.70068359375,
0.2318115234375
],
[
0.7841796875,
0.2176513671875,
0.8369140625,
0.2176513671875,
0.8369140625,
0.2318115234375,
0.7841796875,
0.2318115234375
],
[
0.5234375,
0.2181396484375,
0.572265625,
0.2181396484375,
0.572265625,
0.2330322265625,
0.5234375,
0.2330322265625
],
[
0.1168212890625,
0.222412109375,
0.4873046875,
0.222412109375,
0.4873046875,
0.236083984375,
0.1168212890625,
0.236083984375
],
[
0.11846923828125,
0.2384033203125,
0.402587890625,
0.2384033203125,
0.402587890625,
0.252685546875,
0.11846923828125,
0.252685546875
],
[
0.53515625,
0.2467041015625,
0.86572265625,
0.2467041015625,
0.86572265625,
0.260498046875,
0.53515625,
0.260498046875
],
[
0.1319580078125,
0.266357421875,
0.48486328125,
0.266357421875,
0.48486328125,
0.280029296875,
0.1319580078125,
0.280029296875
],
[
0.52197265625,
0.260986328125,
0.87646484375,
0.260498046875,
0.87646484375,
0.274658203125,
0.52197265625,
0.27587890625
],
[
0.52783203125,
0.27587890625,
0.87158203125,
0.27587890625,
0.87158203125,
0.28955078125,
0.52783203125,
0.28955078125
],
[
0.1512451171875,
0.282958984375,
0.4873046875,
0.282958984375,
0.4873046875,
0.296630859375,
0.1512451171875,
0.296630859375
],
[
0.6220703125,
0.2900390625,
0.77880859375,
0.2900390625,
0.77880859375,
0.3037109375,
0.6220703125,
0.3037109375
],
[
0.1512451171875,
0.2978515625,
0.4873046875,
0.2978515625,
0.4873046875,
0.31103515625,
0.1512451171875,
0.31103515625
],
[
0.1512451171875,
0.31201171875,
0.486572265625,
0.31201171875,
0.486572265625,
0.326904296875,
0.1512451171875,
0.326904296875
],
[
0.529296875,
0.320556640625,
0.87890625,
0.320556640625,
0.87890625,
0.334228515625,
0.529296875,
0.334228515625
],
[
0.1512451171875,
0.32763671875,
0.4873046875,
0.32763671875,
0.4873046875,
0.342529296875,
0.1512451171875,
0.342529296875
],
[
0.51171875,
0.3359375,
0.880859375,
0.3359375,
0.880859375,
0.349609375,
0.51171875,
0.349609375
],
[
0.1529541015625,
0.345458984375,
0.257080078125,
0.345458984375,
0.257080078125,
0.356689453125,
0.1529541015625,
0.356689453125
],
[
0.5126953125,
0.3525390625,
0.88232421875,
0.3525390625,
0.88232421875,
0.3662109375,
0.5126953125,
0.3662109375
],
[
0.4150390625,
0.370361328125,
0.48486328125,
0.370361328125,
0.48486328125,
0.38525390625,
0.4150390625,
0.38525390625
],
[
0.5126953125,
0.36865234375,
0.88232421875,
0.36865234375,
0.88232421875,
0.38232421875,
0.5126953125,
0.38232421875
],
[
0.22607421875,
0.37158203125,
0.39990234375,
0.37158203125,
0.39990234375,
0.386474609375,
0.22607421875,
0.386474609375
],
[
0.1470947265625,
0.372802734375,
0.20751953125,
0.372802734375,
0.20751953125,
0.384033203125,
0.1470947265625,
0.384033203125
],
[
0.1512451171875,
0.38525390625,
0.4873046875,
0.386962890625,
0.4873046875,
0.40185546875,
0.1512451171875,
0.3994140625
],
[
0.5126953125,
0.38525390625,
0.88330078125,
0.38525390625,
0.88330078125,
0.39892578125,
0.5126953125,
0.39892578125
],
[
0.1529541015625,
0.40185546875,
0.483154296875,
0.40185546875,
0.483154296875,
0.416259765625,
0.1529541015625,
0.416259765625
],
[
0.51171875,
0.400634765625,
0.880859375,
0.400634765625,
0.880859375,
0.414306640625,
0.51171875,
0.414306640625
],
[
0.1512451171875,
0.41845703125,
0.48486328125,
0.41845703125,
0.48486328125,
0.43212890625,
0.1512451171875,
0.43212890625
],
[
0.5126953125,
0.416259765625,
0.88232421875,
0.416259765625,
0.88232421875,
0.429931640625,
0.5126953125,
0.429931640625
],
[
0.1512451171875,
0.43408203125,
0.486572265625,
0.43408203125,
0.486572265625,
0.448974609375,
0.1512451171875,
0.448974609375
],
[
0.51025390625,
0.432861328125,
0.88232421875,
0.43212890625,
0.88232421875,
0.4453125,
0.51025390625,
0.446533203125
],
[
0.1512451171875,
0.45068359375,
0.486572265625,
0.449462890625,
0.486572265625,
0.46435546875,
0.1512451171875,
0.465576171875
],
[
0.51171875,
0.44775390625,
0.56787109375,
0.44775390625,
0.56787109375,
0.4619140625,
0.51171875,
0.4619140625
],
[
0.1529541015625,
0.466064453125,
0.486572265625,
0.466064453125,
0.486572265625,
0.479736328125,
0.1529541015625,
0.479736328125
],
[
0.53125,
0.463134765625,
0.88232421875,
0.463134765625,
0.88232421875,
0.476806640625,
0.53125,
0.476806640625
],
[
0.1546630859375,
0.482666015625,
0.486572265625,
0.482666015625,
0.486572265625,
0.493408203125,
0.1546630859375,
0.493408203125
],
[
0.51025390625,
0.479736328125,
0.88232421875,
0.479736328125,
0.88232421875,
0.493408203125,
0.51025390625,
0.493408203125
],
[
0.1529541015625,
0.49755859375,
0.4873046875,
0.49755859375,
0.4873046875,
0.51123046875,
0.1529541015625,
0.51123046875
],
[
0.5126953125,
0.496337890625,
0.57373046875,
0.496337890625,
0.57373046875,
0.50830078125,
0.5126953125,
0.50830078125
],
[
0.1529541015625,
0.5126953125,
0.4873046875,
0.5126953125,
0.4873046875,
0.525390625,
0.1529541015625,
0.525390625
],
[
0.1529541015625,
0.52783203125,
0.4873046875,
0.52783203125,
0.4873046875,
0.54150390625,
0.1529541015625,
0.54150390625
],
[
0.548828125,
0.52392578125,
0.880859375,
0.52392578125,
0.880859375,
0.53759765625,
0.548828125,
0.53759765625
],
[
0.1512451171875,
0.54345703125,
0.486572265625,
0.54345703125,
0.486572265625,
0.55712890625,
0.1512451171875,
0.55712890625
],
[
0.546875,
0.5390625,
0.88232421875,
0.5390625,
0.88232421875,
0.552734375,
0.546875,
0.552734375
],
[
0.1529541015625,
0.55712890625,
0.486572265625,
0.55712890625,
0.486572265625,
0.57080078125,
0.1529541015625,
0.57080078125
],
[
0.54541015625,
0.55419921875,
0.88232421875,
0.5546875,
0.88232421875,
0.5693359375,
0.54541015625,
0.568359375
],
[
0.1512451171875,
0.57373046875,
0.486572265625,
0.57373046875,
0.486572265625,
0.58740234375,
0.1512451171875,
0.58740234375
],
[
0.546875,
0.5693359375,
0.705078125,
0.5693359375,
0.705078125,
0.583984375,
0.546875,
0.583984375
],
[
0.1529541015625,
0.58935546875,
0.326904296875,
0.58935546875,
0.326904296875,
0.60302734375,
0.1529541015625,
0.60302734375
],
[
0.53955078125,
0.60302734375,
0.546875,
0.60302734375,
0.546875,
0.60693359375,
0.53955078125,
0.60693359375
],
[
0.546875,
0.59765625,
0.88330078125,
0.59765625,
0.88330078125,
0.611328125,
0.546875,
0.611328125
],
[
0.1378173828125,
0.6171875,
0.48486328125,
0.6171875,
0.48486328125,
0.630859375,
0.1378173828125,
0.630859375
],
[
0.546875,
0.6123046875,
0.88232421875,
0.6123046875,
0.88232421875,
0.6259765625,
0.546875,
0.6259765625
],
[
0.54541015625,
0.6279296875,
0.88232421875,
0.6279296875,
0.88232421875,
0.642578125,
0.54541015625,
0.642578125
],
[
0.1546630859375,
0.63427734375,
0.48486328125,
0.63427734375,
0.48486328125,
0.64453125,
0.1546630859375,
0.64453125
],
[
0.1529541015625,
0.6484375,
0.486572265625,
0.6484375,
0.486572265625,
0.662109375,
0.1529541015625,
0.662109375
],
[
0.546875,
0.64453125,
0.88232421875,
0.64453125,
0.88232421875,
0.6591796875,
0.546875,
0.6591796875
],
[
0.1512451171875,
0.6640625,
0.486572265625,
0.6640625,
0.486572265625,
0.677734375,
0.1512451171875,
0.677734375
],
[
0.546875,
0.66015625,
0.72119140625,
0.66015625,
0.72119140625,
0.673828125,
0.546875,
0.673828125
],
[
0.1495361328125,
0.677734375,
0.4873046875,
0.68017578125,
0.4873046875,
0.6943359375,
0.1495361328125,
0.69287109375
],
[
0.53955078125,
0.69140625,
0.546875,
0.69140625,
0.546875,
0.69580078125,
0.53955078125,
0.69580078125
],
[
0.54296875,
0.68603515625,
0.87890625,
0.68603515625,
0.87890625,
0.69970703125,
0.54296875,
0.69970703125
],
[
0.1529541015625,
0.693359375,
0.350341796875,
0.693359375,
0.350341796875,
0.70703125,
0.1529541015625,
0.70703125
],
[
0.548828125,
0.70263671875,
0.88232421875,
0.70263671875,
0.88232421875,
0.7177734375,
0.548828125,
0.7177734375
],
[
0.13525390625,
0.7216796875,
0.48486328125,
0.7216796875,
0.48486328125,
0.73583984375,
0.13525390625,
0.73583984375
],
[
0.548828125,
0.71826171875,
0.88232421875,
0.71826171875,
0.88232421875,
0.73193359375,
0.548828125,
0.73193359375
],
[
0.115966796875,
0.73583984375,
0.486572265625,
0.7373046875,
0.486572265625,
0.751953125,
0.115966796875,
0.7509765625
],
[
0.548828125,
0.734375,
0.88232421875,
0.734375,
0.88232421875,
0.74853515625,
0.548828125,
0.74853515625
],
[
0.1142578125,
0.751953125,
0.486572265625,
0.751953125,
0.486572265625,
0.76611328125,
0.1142578125,
0.76611328125
],
[
0.548828125,
0.74951171875,
0.880859375,
0.74951171875,
0.880859375,
0.76318359375,
0.548828125,
0.76318359375
],
[
0.546875,
0.76318359375,
0.88232421875,
0.76318359375,
0.88232421875,
0.77685546875,
0.546875,
0.77685546875
],
[
0.115966796875,
0.76953125,
0.486572265625,
0.76953125,
0.486572265625,
0.783203125,
0.115966796875,
0.783203125
],
[
0.548828125,
0.78173828125,
0.603515625,
0.78173828125,
0.603515625,
0.79345703125,
0.548828125,
0.79345703125
],
[
0.62255859375,
0.78173828125,
0.734375,
0.78173828125,
0.734375,
0.79345703125,
0.62255859375,
0.79345703125
],
[
0.75390625,
0.78173828125,
0.7880859375,
0.78173828125,
0.7880859375,
0.79345703125,
0.75390625,
0.79345703125
],
[
0.79833984375,
0.77880859375,
0.88330078125,
0.78125,
0.88330078125,
0.79541015625,
0.79833984375,
0.79345703125
],
[
0.546875,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.546875,
0.80810546875
],
[
0.115966796875,
0.7841796875,
0.48486328125,
0.7841796875,
0.48486328125,
0.7978515625,
0.115966796875,
0.7978515625
],
[
0.1168212890625,
0.80078125,
0.486572265625,
0.80078125,
0.486572265625,
0.81494140625,
0.1168212890625,
0.81494140625
],
[
0.115966796875,
0.81494140625,
0.48486328125,
0.81640625,
0.48486328125,
0.8310546875,
0.115966796875,
0.830078125
],
[
0.546875,
0.82275390625,
0.880859375,
0.82275390625,
0.880859375,
0.8369140625,
0.546875,
0.8369140625
],
[
0.115966796875,
0.83154296875,
0.48486328125,
0.83154296875,
0.48486328125,
0.84521484375,
0.115966796875,
0.84521484375
],
[
0.546875,
0.83935546875,
0.88330078125,
0.83935546875,
0.88330078125,
0.8525390625,
0.546875,
0.8525390625
],
[
0.115966796875,
0.8466796875,
0.4873046875,
0.84765625,
0.4873046875,
0.8623046875,
0.115966796875,
0.86083984375
],
[
0.546875,
0.853515625,
0.88330078125,
0.853515625,
0.88330078125,
0.86865234375,
0.546875,
0.86865234375
],
[
0.115966796875,
0.8642578125,
0.4873046875,
0.8642578125,
0.4873046875,
0.8779296875,
0.115966796875,
0.8779296875
],
[
0.548828125,
0.869140625,
0.88232421875,
0.869140625,
0.88232421875,
0.8828125,
0.548828125,
0.8828125
],
[
0.115966796875,
0.8798828125,
0.4873046875,
0.8798828125,
0.4873046875,
0.8935546875,
0.115966796875,
0.8935546875
],
[
0.548828125,
0.88525390625,
0.88330078125,
0.88525390625,
0.88330078125,
0.89892578125,
0.548828125,
0.89892578125
],
[
0.115966796875,
0.8955078125,
0.37890625,
0.8955078125,
0.37890625,
0.91015625,
0.115966796875,
0.91015625
]
] |
[
"Model Considerations",
"Recall",
"Precision",
"F1",
"0.1151",
"3",
"0.9836",
"0.2061",
"IG",
"The next objective is to demonstrate an effective",
"DT",
"0.8032",
"0.7692",
"0.7859",
"RF",
"system rather than conduct an exhaustive analysis",
"0.8093",
"0.8631",
"0.9244",
"of various models. Nevertheless, we test a variety",
"of models to identify an effective solution for",
"0.8353",
"0.9163",
"0.8740",
"detecting fraudulent Instagram comments. The",
"LR",
"0.3739",
"0.3660",
"0.3699",
"selection of models includes traditional statistical",
"GPT-3",
"0.6348",
"0.5530",
"0.5911",
"models, large language models and BERT as a",
"GPT-4",
"0.9213",
"0.9286",
"0.9249",
"BERT",
"representative of the transformer models. The",
"following specific models are selected:",
"Table 1: Model metrics. IG: Instagram's current",
". Statistical Models: We start with a linear",
"spam filter, DT: Decision Tree, RF: Random Forest,",
"LR: Linear Regression, BERT: Fined-tuned BERT",
"regression, a decision tree and a random",
"model from ScamSpot",
"forest model which provide us with a",
"baseline and show that they lack the",
"Despite the hype around large language models",
"sophistication needed for our complex",
"the results in Table 1 show that even established",
"requirements.",
"models like a fine-tuned BERT model can",
"(LLMs):",
"drastically outperform newer models like GPT-4 or",
"Language Models",
"Large",
"We assess the advanced natural language",
"GPT-3.While both BERT and LLMs are",
"processing capabilities of GPT-3 (gpt-",
"transformer-based, our research also demonstrates",
"3.5-turbo\")and GPT-4 (gpt-4-1106-",
"that a fine-tuned BERT model is more successful in",
"preview\"). Their ability to understand and",
"specific tasks, which is also shown by Yu et al.",
"generate human-like text is a key",
"(2023).",
"consideration in our analysis. However, the",
"This leads us to the decision to select a fine-",
"lackof transparencyand control over the",
"tuned BERT model for ScamSpot. The reasons also",
"training data of these models is a major",
"include:",
"limitation. Despite the adjustment of seed",
"and temperature parameters (Table 5), the",
"Promising Results: Tests show that the fine-",
"results are not deterministic. Three test runs",
"tuned BERT model far outperforms the other",
"are made for each model and the results can",
"models in detecting fraudulent Instagram",
"be found in Table 5. The best results are",
"comments (Table 1).",
"documented in Table 1.",
"",
"Stability and Predictability: Our fine-tuned",
"Transformer Models: As a representative,",
"BERT model demonstrates stableand",
"predictable performance, a crucial factor for",
"weselectBERT(bert-base-cased)and",
"fine-tune it based on the annotated data",
"consistent user experience compared to the",
"mentioned in Section 2. Its capability to",
"tested LLMs (Table 5).",
"understand context makes it a strong",
"",
"Local Execution: Unlike LLMs like GPT",
"candidate in our selection.",
"3.5 and GPT-4, which often require external",
"In this study, we adopt a zero-shot approach with",
"cloud services potentially exposing data to",
"LLMs like GPT-3 and GPT-4, contrasting them",
"third parties, BERT can be deployed",
"with thefine-tuned BERT model.This",
"internally, ensuring data privacy. Moreover,",
"BERT's smaller model size leads to lower",
"methodology is chosen to demonstrate the practical",
"energy",
"consumption,",
"an",
"important",
"consideration in sustainable AI development.",
"usability of general-purpose LLMs in their",
"standard configuration. While models like GPT-3",
"and GPT-4 show proficiency in general tasks, our",
"Adaptability to New Data: BERT's ability",
"findings align with those of Yu et al. (2023), when",
"to understand context and meanings beyond",
"illustrating their limitations in specific tasks",
"just keyword matching makes it superior to",
"compared to fine-tuned models. This highlights the",
"simpler models like TF-IDF vectors. This",
"necessity of model selection tailored to task",
"adaptability is vital for detecting variations in",
"specificity and resource availability."
] |
{
"arxiv_id": "2402.08869",
"language": "en",
"page_idx": 2
}
|
||
2402.08869-en-4
|
2402.08869-en-4.jpg
|
in the HTML DOM are modified. The browser extension as well as its code is publicly available$^{5}.$
### 4.3 REST API
The REST API is developed using the Flask web framework and deployed utilizing Gunicorn, a Python Web Server Gateway Interface (WSGI) HTTP server. The server loads the fine-tuned BERT model and classifies the comment received on the /scam endpoint. The server's response indicates whether the comment is genuine or not. The code is available as a public repository$^{6}.$ Until now, the endpoints of the API have been accessible without user authentication/token access. This allows others to quickly use the API and prevents compliance issues. As usage increases, a user management system is planned, but as little data as possible should be stored.
### 4.4 Features
To use the application, the users need to install the Chrome browser extension and follow the brief installation guide$^5.$ After successful installation, the user can choose between two modes.
$\bullet$ Detecting Fraudulent Comments: The first
mode of the application marks spam/fraud comments with a red label (Figure 1). Users are warned about a potential fraudulent comment, reducing the likelihood of a fraud case. However, the evaluation surveys conducted as part of the research show that users do not want to see spam or scam comments at all (Table 8).
$\bullet$ Hiding Spam & Scam: The second mode
complies with these requests and hides all comments that are classified as spam or scam (Figure 2). The evaluation survey has clearly shown that users prefer the second mode, resulting in a drastically improved user experience.
$\bullet$ Dynamic System: Another valuable feature
enables users to report incorrectly classified
to
model
comments,
contributing
enhancement through subsequent data-
driven refinement.
### 5 System Evaluation
ScamSpot's effectiveness is evaluated in two cycles based on Hevner's Design Science Research Framework (Hevner et al., 2004), which emphasizes the iterative development and refinement of a system through multiple cycles of design, testing, and feedback. In each cycle, both qualitative and quantitative evaluation metrics are considered.
The quantitative results focus on the metrics and the effectiveness of the model itself. As can be seen in Table l, the model achieves both precision and recall of 92% after the second cycle.
Our model exhibits notable performance compared to other baseline models as well as OpenAI's GPT-3 and GPT-4, justifying its utilization in the project. The F1 scores can be found in Table 1, with the fine-tuned BERT model archiving a considerably higher F1 score.
When comparing the results with Instagram's existing spam filter, the model also performs well. While Instagram's spam filter only achieves a recall of 11.51%, our model achieves 92.13%. However, we must give credit to Instagram's spam filter as its precision was 98.36%, while our model only achieves 92.86%. This shows room for improvement.
Another notable result is that both GPT models fail to correctly categorise comments as spam/scam or genuine. Despite multiple approaches and different prompts, the results are surprisingly disappointing. Results can be found in Table 5.
Not only the quantitative but also the qualitative results improve drastically after the second cycle, in which the feedback from the first evaluation cycle was implemented.20 ScamSpot users are asked in a questionnaire about their experiences with the system and their feedback.
When using the prototype, almost all test users report that the user experience has improved and that they are more willing to interact with others in the comments section. With the browser extension, users report overall positive feelings of joy and
$^6$ https://github.com/ScamSpot/
scamspot\_api/
$^{5}$ https://github.com/ScamSpot/
scamspot\_chrome-extension/
|
[
[
0.1168212890625,
0.086181640625,
0.4873046875,
0.086181640625,
0.4873046875,
0.0963134765625,
0.1168212890625,
0.0963134765625
],
[
0.52001953125,
0.086181640625,
0.529296875,
0.0921630859375,
0.52197265625,
0.09747314453125,
0.5126953125,
0.091552734375
],
[
0.54296875,
0.08502197265625,
0.70947265625,
0.08502197265625,
0.70947265625,
0.09869384765625,
0.54296875,
0.09869384765625
],
[
0.115966796875,
0.1004638671875,
0.4814453125,
0.1004638671875,
0.4814453125,
0.1141357421875,
0.115966796875,
0.1141357421875
],
[
0.5126953125,
0.1112060546875,
0.88232421875,
0.1112060546875,
0.88232421875,
0.1248779296875,
0.5126953125,
0.1248779296875
],
[
0.1142578125,
0.1265869140625,
0.2529296875,
0.1248779296875,
0.2529296875,
0.1385498046875,
0.1142578125,
0.1402587890625
],
[
0.5126953125,
0.1278076171875,
0.88232421875,
0.1278076171875,
0.88232421875,
0.1414794921875,
0.5126953125,
0.1414794921875
],
[
0.51171875,
0.1427001953125,
0.880859375,
0.1427001953125,
0.880859375,
0.15576171875,
0.51171875,
0.15576171875
],
[
0.1168212890625,
0.148681640625,
0.486572265625,
0.148681640625,
0.486572265625,
0.162353515625,
0.1168212890625,
0.162353515625
],
[
0.51416015625,
0.1611328125,
0.60498046875,
0.1611328125,
0.60498046875,
0.17236328125,
0.51416015625,
0.17236328125
],
[
0.5966796875,
0.158203125,
0.88232421875,
0.1593017578125,
0.88232421875,
0.173583984375,
0.5966796875,
0.17236328125
],
[
0.115966796875,
0.1640625,
0.486572265625,
0.1640625,
0.486572265625,
0.177734375,
0.115966796875,
0.177734375
],
[
0.115966796875,
0.1807861328125,
0.4873046875,
0.1807861328125,
0.4873046875,
0.1944580078125,
0.115966796875,
0.1944580078125
],
[
0.51171875,
0.176025390625,
0.88330078125,
0.176025390625,
0.88330078125,
0.1890869140625,
0.51171875,
0.1890869140625
],
[
0.5126953125,
0.1903076171875,
0.88232421875,
0.1903076171875,
0.88232421875,
0.205078125,
0.5126953125,
0.205078125
],
[
0.1168212890625,
0.19677734375,
0.486572265625,
0.19677734375,
0.486572265625,
0.2098388671875,
0.1168212890625,
0.2098388671875
],
[
0.1142578125,
0.2122802734375,
0.486572265625,
0.2122802734375,
0.486572265625,
0.2259521484375,
0.1142578125,
0.2259521484375
],
[
0.51171875,
0.2069091796875,
0.88330078125,
0.2069091796875,
0.88330078125,
0.2205810546875,
0.51171875,
0.2205810546875
],
[
0.115966796875,
0.2288818359375,
0.4873046875,
0.2288818359375,
0.4873046875,
0.2425537109375,
0.115966796875,
0.2425537109375
],
[
0.5126953125,
0.2235107421875,
0.59765625,
0.2235107421875,
0.59765625,
0.23486328125,
0.5126953125,
0.23486328125
],
[
0.115966796875,
0.2431640625,
0.48486328125,
0.244384765625,
0.48486328125,
0.25927734375,
0.115966796875,
0.258056640625
],
[
0.53125,
0.2384033203125,
0.88232421875,
0.2384033203125,
0.88232421875,
0.252685546875,
0.53125,
0.252685546875
],
[
0.5126953125,
0.25390625,
0.88232421875,
0.25390625,
0.88232421875,
0.267578125,
0.5126953125,
0.267578125
],
[
0.1168212890625,
0.260498046875,
0.486572265625,
0.260498046875,
0.486572265625,
0.2734375,
0.1168212890625,
0.2734375
],
[
0.51171875,
0.269287109375,
0.88330078125,
0.269287109375,
0.88330078125,
0.282958984375,
0.51171875,
0.282958984375
],
[
0.1142578125,
0.27587890625,
0.4873046875,
0.274658203125,
0.4873046875,
0.28955078125,
0.1142578125,
0.2900390625
],
[
0.51025390625,
0.285888671875,
0.77880859375,
0.285888671875,
0.77880859375,
0.299560546875,
0.51025390625,
0.299560546875
],
[
0.115966796875,
0.29248046875,
0.48486328125,
0.29248046875,
0.48486328125,
0.302734375,
0.115966796875,
0.302734375
],
[
0.53125,
0.302001953125,
0.88232421875,
0.302001953125,
0.88232421875,
0.315185546875,
0.53125,
0.315185546875
],
[
0.1168212890625,
0.307861328125,
0.4873046875,
0.307861328125,
0.4873046875,
0.321533203125,
0.1168212890625,
0.321533203125
],
[
0.51171875,
0.3173828125,
0.88232421875,
0.3173828125,
0.88232421875,
0.3310546875,
0.51171875,
0.3310546875
],
[
0.115966796875,
0.32470703125,
0.489013671875,
0.32470703125,
0.489013671875,
0.33837890625,
0.115966796875,
0.33837890625
],
[
0.51171875,
0.331787109375,
0.880859375,
0.331787109375,
0.880859375,
0.345458984375,
0.51171875,
0.345458984375
],
[
0.115966796875,
0.33935546875,
0.4873046875,
0.33935546875,
0.4873046875,
0.3525390625,
0.115966796875,
0.3525390625
],
[
0.115966796875,
0.35498046875,
0.3076171875,
0.35498046875,
0.3076171875,
0.369140625,
0.115966796875,
0.369140625
],
[
0.51171875,
0.349609375,
0.88232421875,
0.347900390625,
0.88232421875,
0.36083984375,
0.51171875,
0.36328125
],
[
0.5126953125,
0.3662109375,
0.880859375,
0.3662109375,
0.880859375,
0.376953125,
0.5126953125,
0.376953125
],
[
0.155517578125,
0.381103515625,
0.236083984375,
0.381103515625,
0.236083984375,
0.395263671875,
0.155517578125,
0.395263671875
],
[
0.51171875,
0.381103515625,
0.8134765625,
0.381103515625,
0.8134765625,
0.394775390625,
0.51171875,
0.394775390625
],
[
0.1546630859375,
0.38525390625,
0.1605224609375,
0.389404296875,
0.1571044921875,
0.39111328125,
0.1512451171875,
0.386962890625
],
[
0.1168212890625,
0.38232421875,
0.1546630859375,
0.38232421875,
0.1546630859375,
0.394775390625,
0.1168212890625,
0.394775390625
],
[
0.53173828125,
0.396484375,
0.88232421875,
0.396484375,
0.88232421875,
0.411376953125,
0.53173828125,
0.411376953125
],
[
0.1168212890625,
0.403564453125,
0.48486328125,
0.403564453125,
0.48486328125,
0.41845703125,
0.1168212890625,
0.41845703125
],
[
0.5126953125,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.4267578125,
0.5126953125,
0.4267578125
],
[
0.1168212890625,
0.419677734375,
0.4873046875,
0.419677734375,
0.4873046875,
0.432861328125,
0.1168212890625,
0.432861328125
],
[
0.5126953125,
0.427978515625,
0.88232421875,
0.427978515625,
0.88232421875,
0.4423828125,
0.5126953125,
0.4423828125
],
[
0.1168212890625,
0.435302734375,
0.4873046875,
0.435302734375,
0.4873046875,
0.448974609375,
0.1168212890625,
0.448974609375
],
[
0.5126953125,
0.443603515625,
0.880859375,
0.443603515625,
0.880859375,
0.457275390625,
0.5126953125,
0.457275390625
],
[
0.115966796875,
0.45068359375,
0.38818359375,
0.45068359375,
0.38818359375,
0.46435546875,
0.115966796875,
0.46435546875
],
[
0.5126953125,
0.458984375,
0.88232421875,
0.458984375,
0.88232421875,
0.473876953125,
0.5126953125,
0.473876953125
],
[
0.127685546875,
0.479736328125,
0.486572265625,
0.479736328125,
0.486572265625,
0.493408203125,
0.127685546875,
0.493408203125
],
[
0.5126953125,
0.4755859375,
0.88232421875,
0.4755859375,
0.88232421875,
0.4892578125,
0.5126953125,
0.4892578125
],
[
0.1512451171875,
0.495361328125,
0.4873046875,
0.495361328125,
0.4873046875,
0.5087890625,
0.1512451171875,
0.5087890625
],
[
0.51171875,
0.490478515625,
0.88330078125,
0.490478515625,
0.88330078125,
0.5048828125,
0.51171875,
0.5048828125
],
[
0.1512451171875,
0.5087890625,
0.4873046875,
0.51025390625,
0.4873046875,
0.52490234375,
0.1512451171875,
0.52392578125
],
[
0.5126953125,
0.5087890625,
0.61572265625,
0.5087890625,
0.61572265625,
0.52099609375,
0.5126953125,
0.52099609375
],
[
0.1512451171875,
0.525390625,
0.4873046875,
0.52490234375,
0.4873046875,
0.5390625,
0.1512451171875,
0.54052734375
],
[
0.53125,
0.5224609375,
0.88232421875,
0.5224609375,
0.88232421875,
0.5361328125,
0.53125,
0.5361328125
],
[
0.1529541015625,
0.54150390625,
0.4873046875,
0.54150390625,
0.4873046875,
0.5546875,
0.1529541015625,
0.5546875
],
[
0.51025390625,
0.53759765625,
0.88330078125,
0.5380859375,
0.88330078125,
0.552734375,
0.51025390625,
0.5517578125
],
[
0.1529541015625,
0.5556640625,
0.4873046875,
0.5556640625,
0.4873046875,
0.5693359375,
0.1529541015625,
0.5693359375
],
[
0.51171875,
0.5546875,
0.88330078125,
0.5546875,
0.88330078125,
0.568359375,
0.51171875,
0.568359375
],
[
0.1512451171875,
0.5712890625,
0.489013671875,
0.5712890625,
0.489013671875,
0.58642578125,
0.1512451171875,
0.58642578125
],
[
0.5126953125,
0.57080078125,
0.88232421875,
0.57080078125,
0.88232421875,
0.583984375,
0.5126953125,
0.583984375
],
[
0.1495361328125,
0.58642578125,
0.489013671875,
0.58740234375,
0.489013671875,
0.6015625,
0.1495361328125,
0.6005859375
],
[
0.51025390625,
0.583984375,
0.8505859375,
0.583984375,
0.8505859375,
0.59765625,
0.51025390625,
0.59765625
],
[
0.1529541015625,
0.60302734375,
0.349609375,
0.60302734375,
0.349609375,
0.61669921875,
0.1529541015625,
0.61669921875
],
[
0.529296875,
0.6005859375,
0.880859375,
0.6005859375,
0.880859375,
0.6142578125,
0.529296875,
0.6142578125
],
[
0.51171875,
0.6171875,
0.880859375,
0.6171875,
0.880859375,
0.630859375,
0.51171875,
0.630859375
],
[
0.1470947265625,
0.630859375,
0.4873046875,
0.630859375,
0.4873046875,
0.64453125,
0.1470947265625,
0.64453125
],
[
0.51171875,
0.63330078125,
0.88232421875,
0.63330078125,
0.88232421875,
0.64697265625,
0.51171875,
0.64697265625
],
[
0.1512451171875,
0.64697265625,
0.4873046875,
0.64697265625,
0.4873046875,
0.66015625,
0.1512451171875,
0.66015625
],
[
0.13525390625,
0.63623046875,
0.148681640625,
0.63623046875,
0.148681640625,
0.64013671875,
0.13525390625,
0.64013671875
],
[
0.5126953125,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.6611328125,
0.5126953125,
0.6611328125
],
[
0.1529541015625,
0.662109375,
0.486572265625,
0.662109375,
0.486572265625,
0.67578125,
0.1529541015625,
0.67578125
],
[
0.51171875,
0.6650390625,
0.88232421875,
0.6650390625,
0.88232421875,
0.6787109375,
0.51171875,
0.6787109375
],
[
0.155517578125,
0.677734375,
0.48486328125,
0.677734375,
0.48486328125,
0.6884765625,
0.155517578125,
0.6884765625
],
[
0.51025390625,
0.68017578125,
0.7705078125,
0.68017578125,
0.7705078125,
0.693359375,
0.51025390625,
0.693359375
],
[
0.1529541015625,
0.69287109375,
0.4873046875,
0.69287109375,
0.4873046875,
0.70703125,
0.1529541015625,
0.70703125
],
[
0.529296875,
0.69580078125,
0.880859375,
0.69580078125,
0.880859375,
0.7099609375,
0.529296875,
0.7099609375
],
[
0.1512451171875,
0.70703125,
0.4873046875,
0.7080078125,
0.4873046875,
0.72216796875,
0.1512451171875,
0.7216796875
],
[
0.51171875,
0.71240234375,
0.88232421875,
0.71240234375,
0.88232421875,
0.7265625,
0.51171875,
0.7265625
],
[
0.1546630859375,
0.724609375,
0.2403564453125,
0.724609375,
0.2403564453125,
0.73583984375,
0.1546630859375,
0.73583984375
],
[
0.5126953125,
0.72900390625,
0.88232421875,
0.72900390625,
0.88232421875,
0.74267578125,
0.5126953125,
0.74267578125
],
[
0.5126953125,
0.7431640625,
0.880859375,
0.7431640625,
0.880859375,
0.7568359375,
0.5126953125,
0.7568359375
],
[
0.1336669921875,
0.7568359375,
0.1419677734375,
0.7568359375,
0.1419677734375,
0.76123046875,
0.1336669921875,
0.76123046875
],
[
0.1378173828125,
0.751953125,
0.4873046875,
0.751953125,
0.4873046875,
0.76513671875,
0.1378173828125,
0.76513671875
],
[
0.51171875,
0.76025390625,
0.88232421875,
0.76025390625,
0.88232421875,
0.7734375,
0.51171875,
0.7734375
],
[
0.1495361328125,
0.76611328125,
0.4873046875,
0.76513671875,
0.4873046875,
0.77978515625,
0.1495361328125,
0.78125
],
[
0.1529541015625,
0.783203125,
0.239501953125,
0.783203125,
0.239501953125,
0.79443359375,
0.1529541015625,
0.79443359375
],
[
0.26123046875,
0.783203125,
0.363037109375,
0.783203125,
0.363037109375,
0.79443359375,
0.26123046875,
0.79443359375
],
[
0.386474609375,
0.7841796875,
0.408447265625,
0.7841796875,
0.408447265625,
0.79443359375,
0.386474609375,
0.79443359375
],
[
0.432861328125,
0.78125,
0.4873046875,
0.78125,
0.4873046875,
0.79541015625,
0.432861328125,
0.79541015625
],
[
0.1529541015625,
0.7978515625,
0.4873046875,
0.7978515625,
0.4873046875,
0.81103515625,
0.1529541015625,
0.81103515625
],
[
0.1529541015625,
0.8134765625,
0.29248046875,
0.8134765625,
0.29248046875,
0.8271484375,
0.1529541015625,
0.8271484375
],
[
0.115966796875,
0.88623046875,
0.402587890625,
0.88623046875,
0.402587890625,
0.8994140625,
0.115966796875,
0.8994140625
],
[
0.51171875,
0.88525390625,
0.7998046875,
0.88623046875,
0.7998046875,
0.90087890625,
0.51171875,
0.8994140625
],
[
0.1142578125,
0.90185546875,
0.36962890625,
0.90185546875,
0.36962890625,
0.91552734375,
0.1142578125,
0.91552734375
],
[
0.51171875,
0.90185546875,
0.6396484375,
0.90185546875,
0.6396484375,
0.91552734375,
0.51171875,
0.91552734375
]
] |
[
"inthe HTMLDOMaremodified.Thebrowser",
"7",
"System Evaluation",
"extension as well as its code is publicly availables",
"ScamSpot's effectiveness is evaluated in two",
"4.3 RESTAPI",
"cycles based on Hevner's Design Science Research",
"Framework(Hevner et al.,2004),which",
"The REST API is developed using the Flask web",
"emphasizes",
"theiterativedevelopment and",
"framework and deployed utilizing Gunicorn, a",
"Python Web Server Gateway Interface (WSGI)",
"refinement of a system through multiple cycles of",
"design, testing, and feedback. In each cycle, both",
"HTTP server. The server loads the fine-tuned",
"BERT model and classifies the comment received",
"qualitative and quantitative evaluation metrics are",
"on the /scam endpoint. The server's response",
"considered.",
"indicates whether the comment is genuine or not.",
"The quantitative results focus on the metrics and",
"the effectiveness of the model itself. As can be seen",
"The code is available as a public repository. Until",
"in Table 1, the model achieves both precision and",
"now, the endpoints of the API have been accessible",
"recall of 92% after the second cycle.",
"without userauthentication/tokenaccess.This",
"Our model exhibitsnotable performance",
"allows others to quickly use the API and prevents",
"compared to other baseline models as well as",
"compliance issues. As usage increases, a user",
"OpenAI's GPT-3 and GPT-4, justifying its",
"management system is planned, but as little data as",
"possible should be stored.",
"utilization in the project. The F1 scores can be",
"found in Table 1, with the fine-tuned BERT model",
"Features",
"archiving a considerably higher F1 score.",
"",
"4.4",
"When comparing the results with Instagram's",
"To use the application, the users need to install the",
"existing spam filter, the model also performs well.",
"Chrome browser extension and follow the brief",
"While Instagram's spam filter only achieves a",
"installation guides. After successful installation, the",
"recall of 11.51%, our model achieves 92.13%.",
"user can choose between two modes.",
"However, we must give credit to Instagram's spam",
". Detecting Fraudulent Comments: The first",
"filter as its precision was 98.36%, while our model",
"mode of the application marks spam/fraud",
"onlyachieves 92.86%.This shows room for",
"comments with a red label (Figure 1). Users",
"improvement.",
"are warned about a potential fraudulent",
"Another notable result is that both GPT models",
"comment, reducing the likelihood of a fraud",
"fail to correctly categorise comments as spam/scam",
"case.However, the evaluation surveys",
"or genuine. Despite multiple approaches and",
"conducted as part of the research show that",
"different prompts, the results are surprisingly",
"users do not want to see spam or scam",
"disappointing. Results can be found in Table 5",
"comments at all (Table 8).",
"Not only the quantitative but also the qualitative",
"results improve drastically after the second cycle",
"Hiding Spam & Scam: The second mode",
"in which the feedback from the first evaluation",
"complies with these requests and hides all",
"",
"cycle was implemented.20 ScamSpot users are",
"comments that are classified as spam or scam",
"asked in a questionnaire about their experiences",
"(Figure2).Theevaluationsurveyhasclearly",
"with the system and their feedback.",
"shown that users prefer the second mode,",
" When using the prototype, almost all test users",
"resulting in a drastically improved user",
"report that the user experience has improved and",
"experience.",
"that they are more willing to interact with others in",
"the comments section. With the browser extension.",
"",
"Dynamic System:Another valuable feature",
"users report overall positive feelings of joy and",
"enables users to report incorrectly classified",
"comments.",
"contributing",
"to",
"model",
"enhancement through subsequent data-",
"driven refinement.",
"5https://github.com/scamSpot/",
"6https://github.com/scamSpot/",
"scamspot chrome-extension/",
"scamspot_api/"
] |
{
"arxiv_id": "2402.08869",
"language": "en",
"page_idx": 4
}
|
||
2310.16225-en-1
|
2310.16225-en-1.jpg
|
predictions that are falsely counted as errors due to noisy annotation. We argue that given the maturity of research into high-resource NER, the research community requires a dataset with highly consistent, nearly noise-free labels in order to correctly estimate the performance of state-of-the-art models and to better understand their remaining errors.
Adding and leveraging entity links. To address this issue, we present CLEANCONLL, a revised version of CoNLL-03 that significantly improves annotation quality and consistency. Our main idea is to introduce a second layer of annotation to the corpus that adds entity linking information. Entity linking is the task of disambiguating named entities to a unique ID in a knowledge base, typically the Wikipedia page representing a specific entity.
As Figure 1 illustrates with an example sentence, adding entity links serves two purposes: First, entity links allow us to define automatic consistency checks on NER labels, since prior work found that NER labels can be (partially) derived from Wikipedia page IDs (Tedeschi et al., 2021). Second, we find that the disambiguating information provided by entity links makes NER labels more immediately transparent to human inspection, thus making it easier for annotators to spot mistakes and discuss disagreements.
Contributions. In more detail, our contributions
are the following:
$\bullet$ We present CLEANCONLL, the result of a comprehensive relabeling effort of CoNLL-03 that includes updated, more consistent NER annotations in all three splits and integrates an additional layer of entity linking annotation.
· We detail a method for deriving NER annotation from Wikipedia links that leverages the category system in Wikidata, and discuss how this approach acts as consistency checks to safeguard annotation quality.
·We further describe an automatic approach to flag potential inconsistencies that we use throughout the manual re-annotation process, which we refer to as "cross-checking".
·We conduct extensive analyses of CLEANCONLL and present both quantitative results of manual assessment of annotation quality in comparison to previous versions and qualitative examples of error corrections. We train a set of current state-of-the-art approaches on
CLEANCONLL, and comparatively evaluate
the true error rate of these approaches.
We make the resource publicly available and encourage the community to use and further improve its quality.
## 2 Creating CLEANCoNLL
CLEANCONLL was created in three distinct phases of annotation, which we describe chronologically in the following subsections:
1. We first (see Sec. 2.1) added a layer of entity
linking annotation by merging two existing resources, and developed a heuristic to automatically correct NER errors using entity $links.$
2. We then (see Sec. 2.2) performed 3 iterative
rounds of cross-checking to identify potential inconsistencies, which we manually examined and corrected whenever necessary$^2.$
3. Finally (see Sec. 2.3), we developed spe-
cial handling for adjectival affiliation entity
names.
## 2.1 Phase 1: Adding Entity Links
Base corpus. We use the version from Reiss et al. (2020) as the starting point of our relabeling effort$^3.$ Reiss et al.(2020) already corrected 3.8% of all labels in CoNLL-03, including several major sentence and entity mention boundary problems. While this version has less noise than the original corpus, our own analysis (Sec. 3.4) and prior work (Wang and Mueller, 2022) found that large numbers of errors still remain.
Merging AIDA. For initial corpus creation, we take advantage of the fact that AIDA (Hoffart et al., 2011), a well-known entity linking dataset, shares the same text base as CoNLL-03. Since the entity links in AIDA are expert-annotated by a group distinct of the original CoNLL-03 annotators, they provide a layer of high quality annotation of an arguably more complex task.
However, merging these two corpora was not straightforward since there are discrepancies both in terms of sentence boundary definitions as well
$^2$Manual annotation in all phases of the process was
carried out by the authors of this work. See the limitations
section for a brief discussion.
$^3$Obtained by using their instructions from https://github.com/CODAIT/ Identifying- Incorrect- Labels- In- CoNLL- 2003
|
[
[
0.1175537109375,
0.08612060546875,
0.484375,
0.08612060546875,
0.484375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.5498046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.5498046875,
0.09979248046875
],
[
0.1175537109375,
0.10394287109375,
0.484375,
0.10394287109375,
0.484375,
0.114013671875,
0.1175537109375,
0.114013671875
],
[
0.54833984375,
0.10272216796875,
0.8330078125,
0.10272216796875,
0.8330078125,
0.11639404296875,
0.54833984375,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.52880859375,
0.1258544921875,
0.880859375,
0.1270751953125,
0.880859375,
0.141357421875,
0.52880859375,
0.14013671875
],
[
0.1158447265625,
0.1341552734375,
0.48779296875,
0.133056640625,
0.48779296875,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.51123046875,
0.1436767578125,
0.880859375,
0.1436767578125,
0.880859375,
0.15673828125,
0.51123046875,
0.15673828125
],
[
0.118408203125,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.162109375,
0.118408203125,
0.162109375
],
[
0.1175537109375,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51318359375,
0.160888671875,
0.59033203125,
0.160888671875,
0.59033203125,
0.1728515625,
0.51318359375,
0.1728515625
],
[
0.1175537109375,
0.1817626953125,
0.465087890625,
0.182861328125,
0.465087890625,
0.196533203125,
0.1175537109375,
0.1953125
],
[
0.5146484375,
0.189453125,
0.53662109375,
0.189453125,
0.53662109375,
0.19775390625,
0.5146484375,
0.19775390625
],
[
0.5380859375,
0.1859130859375,
0.76171875,
0.185302734375,
0.76171875,
0.1995849609375,
0.5380859375,
0.20068359375
],
[
0.1158447265625,
0.198974609375,
0.486083984375,
0.19775390625,
0.486083984375,
0.2120361328125,
0.1158447265625,
0.213134765625
],
[
0.1175537109375,
0.215576171875,
0.489501953125,
0.215576171875,
0.489501953125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.51123046875,
0.2120361328125,
0.8818359375,
0.2120361328125,
0.8818359375,
0.2257080078125,
0.51123046875,
0.2257080078125
],
[
0.1158447265625,
0.23095703125,
0.489501953125,
0.23095703125,
0.489501953125,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.509765625,
0.2286376953125,
0.8818359375,
0.2286376953125,
0.8818359375,
0.2423095703125,
0.509765625,
0.2423095703125
],
[
0.1175537109375,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.51318359375,
0.24462890625,
0.765625,
0.24462890625,
0.765625,
0.2587890625,
0.51318359375,
0.2587890625
],
[
0.1158447265625,
0.26318359375,
0.48779296875,
0.26318359375,
0.48779296875,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.5263671875,
0.26953125,
0.880859375,
0.270751953125,
0.880859375,
0.28515625,
0.5263671875,
0.283935546875
],
[
0.1158447265625,
0.27783203125,
0.48779296875,
0.279052734375,
0.48779296875,
0.29345703125,
0.1158447265625,
0.292236328125
],
[
0.54833984375,
0.287353515625,
0.8818359375,
0.287353515625,
0.8818359375,
0.3017578125,
0.54833984375,
0.3017578125
],
[
0.1158447265625,
0.29443359375,
0.486083984375,
0.29443359375,
0.486083984375,
0.30810546875,
0.1158447265625,
0.30810546875
],
[
0.54736328125,
0.303955078125,
0.88330078125,
0.303955078125,
0.88330078125,
0.317626953125,
0.54736328125,
0.317626953125
],
[
0.1158447265625,
0.31005859375,
0.486083984375,
0.31005859375,
0.486083984375,
0.32373046875,
0.1158447265625,
0.32373046875
],
[
0.54736328125,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.333251953125,
0.54736328125,
0.333251953125
],
[
0.1158447265625,
0.325927734375,
0.450927734375,
0.32666015625,
0.450927734375,
0.341552734375,
0.1158447265625,
0.34033203125
],
[
0.54736328125,
0.333251953125,
0.59375,
0.3349609375,
0.591796875,
0.350830078125,
0.5458984375,
0.3486328125
],
[
0.1351318359375,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.1351318359375,
0.355712890625
],
[
0.1175537109375,
0.359375,
0.48779296875,
0.359375,
0.48779296875,
0.372314453125,
0.1175537109375,
0.372314453125
],
[
0.5263671875,
0.361083984375,
0.880859375,
0.361083984375,
0.880859375,
0.374755859375,
0.5263671875,
0.374755859375
],
[
0.1158447265625,
0.374755859375,
0.484375,
0.3759765625,
0.484375,
0.39013671875,
0.1158447265625,
0.3896484375
],
[
0.54736328125,
0.37890625,
0.880859375,
0.37890625,
0.880859375,
0.392578125,
0.54736328125,
0.392578125
],
[
0.1158447265625,
0.39013671875,
0.486083984375,
0.391357421875,
0.486083984375,
0.40625,
0.1158447265625,
0.405029296875
],
[
0.54833984375,
0.394287109375,
0.8818359375,
0.394287109375,
0.8818359375,
0.407958984375,
0.54833984375,
0.407958984375
],
[
0.1175537109375,
0.407958984375,
0.486083984375,
0.407958984375,
0.486083984375,
0.421630859375,
0.1175537109375,
0.421630859375
],
[
0.54736328125,
0.4091796875,
0.81005859375,
0.410400390625,
0.81005859375,
0.424560546875,
0.54736328125,
0.42333984375
],
[
0.1175537109375,
0.42333984375,
0.48779296875,
0.42333984375,
0.48779296875,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.11419677734375,
0.440673828125,
0.486083984375,
0.440673828125,
0.486083984375,
0.45361328125,
0.11419677734375,
0.45361328125
],
[
0.52490234375,
0.435302734375,
0.880859375,
0.435791015625,
0.880859375,
0.45068359375,
0.52490234375,
0.449462890625
],
[
0.1175537109375,
0.457275390625,
0.48779296875,
0.457275390625,
0.48779296875,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.54833984375,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.54833984375,
0.466064453125
],
[
0.11419677734375,
0.471435546875,
0.48779296875,
0.47265625,
0.48779296875,
0.48681640625,
0.11419677734375,
0.486328125
],
[
0.54736328125,
0.468505859375,
0.60546875,
0.468505859375,
0.60546875,
0.482666015625,
0.54736328125,
0.482666015625
],
[
0.1175537109375,
0.4892578125,
0.48779296875,
0.4892578125,
0.48779296875,
0.5029296875,
0.1175537109375,
0.5029296875
],
[
0.51318359375,
0.49462890625,
0.7900390625,
0.49462890625,
0.7900390625,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.1175537109375,
0.5048828125,
0.28955078125,
0.5048828125,
0.28955078125,
0.5185546875,
0.1175537109375,
0.5185546875
],
[
0.1175537109375,
0.52099609375,
0.486083984375,
0.52099609375,
0.486083984375,
0.53369140625,
0.1175537109375,
0.53369140625
],
[
0.51123046875,
0.515625,
0.8818359375,
0.51416015625,
0.8818359375,
0.529296875,
0.51123046875,
0.52978515625
],
[
0.1158447265625,
0.53759765625,
0.252685546875,
0.53759765625,
0.252685546875,
0.55029296875,
0.1158447265625,
0.55029296875
],
[
0.51318359375,
0.5322265625,
0.8818359375,
0.5322265625,
0.8818359375,
0.54638671875,
0.51318359375,
0.54638671875
],
[
0.51123046875,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.51123046875,
0.56103515625
],
[
0.1376953125,
0.56396484375,
0.489501953125,
0.56396484375,
0.489501953125,
0.5791015625,
0.1376953125,
0.5791015625
],
[
0.509765625,
0.5634765625,
0.8818359375,
0.56396484375,
0.8818359375,
0.5791015625,
0.509765625,
0.57763671875
],
[
0.15283203125,
0.58056640625,
0.48779296875,
0.58056640625,
0.48779296875,
0.59423828125,
0.15283203125,
0.59423828125
],
[
0.509765625,
0.58056640625,
0.88330078125,
0.58056640625,
0.88330078125,
0.59423828125,
0.509765625,
0.59423828125
],
[
0.151123046875,
0.5966796875,
0.48779296875,
0.595703125,
0.48779296875,
0.60986328125,
0.151123046875,
0.61083984375
],
[
0.5087890625,
0.595703125,
0.8818359375,
0.5966796875,
0.8818359375,
0.61083984375,
0.5087890625,
0.60986328125
],
[
0.151123046875,
0.61328125,
0.489501953125,
0.61328125,
0.489501953125,
0.626953125,
0.151123046875,
0.626953125
],
[
0.509765625,
0.61328125,
0.8818359375,
0.61328125,
0.8818359375,
0.626953125,
0.509765625,
0.626953125
],
[
0.1502685546875,
0.62744140625,
0.483642578125,
0.62890625,
0.483642578125,
0.6435546875,
0.1502685546875,
0.642578125
],
[
0.5087890625,
0.62744140625,
0.8818359375,
0.62890625,
0.8818359375,
0.6435546875,
0.5087890625,
0.642578125
],
[
0.5146484375,
0.646484375,
0.736328125,
0.646484375,
0.736328125,
0.65673828125,
0.5146484375,
0.65673828125
],
[
0.1376953125,
0.65478515625,
0.489501953125,
0.65478515625,
0.489501953125,
0.66845703125,
0.1376953125,
0.66845703125
],
[
0.51318359375,
0.6611328125,
0.8818359375,
0.6611328125,
0.8818359375,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.15283203125,
0.6728515625,
0.489501953125,
0.6728515625,
0.489501953125,
0.68603515625,
0.15283203125,
0.68603515625
],
[
0.509765625,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.509765625,
0.6904296875
],
[
0.15283203125,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.15283203125,
0.703125
],
[
0.509765625,
0.693359375,
0.8818359375,
0.693359375,
0.8818359375,
0.70703125,
0.509765625,
0.70703125
],
[
0.15283203125,
0.705078125,
0.489501953125,
0.705078125,
0.489501953125,
0.7197265625,
0.15283203125,
0.7197265625
],
[
0.51123046875,
0.708984375,
0.880859375,
0.708984375,
0.880859375,
0.72265625,
0.51123046875,
0.72265625
],
[
0.15283203125,
0.7216796875,
0.36865234375,
0.7216796875,
0.36865234375,
0.7353515625,
0.15283203125,
0.7353515625
],
[
0.51123046875,
0.7255859375,
0.8818359375,
0.7255859375,
0.8818359375,
0.7392578125,
0.51123046875,
0.7392578125
],
[
0.1368408203125,
0.74658203125,
0.48779296875,
0.74755859375,
0.48779296875,
0.7626953125,
0.1368408203125,
0.76123046875
],
[
0.51123046875,
0.7412109375,
0.8818359375,
0.7412109375,
0.8818359375,
0.7548828125,
0.51123046875,
0.7548828125
],
[
0.51123046875,
0.75830078125,
0.8818359375,
0.75830078125,
0.8818359375,
0.771484375,
0.51123046875,
0.771484375
],
[
0.15283203125,
0.765625,
0.489501953125,
0.765625,
0.489501953125,
0.779296875,
0.15283203125,
0.779296875
],
[
0.51123046875,
0.77490234375,
0.72705078125,
0.77490234375,
0.72705078125,
0.7880859375,
0.51123046875,
0.7880859375
],
[
0.15283203125,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.79443359375,
0.15283203125,
0.79443359375
],
[
0.52880859375,
0.79052734375,
0.88330078125,
0.79052734375,
0.88330078125,
0.80419921875,
0.52880859375,
0.80419921875
],
[
0.15283203125,
0.79638671875,
0.438232421875,
0.79638671875,
0.438232421875,
0.81005859375,
0.15283203125,
0.81005859375
],
[
0.51318359375,
0.8056640625,
0.8818359375,
0.8056640625,
0.8818359375,
0.8193359375,
0.51318359375,
0.8193359375
],
[
0.139404296875,
0.82373046875,
0.49365234375,
0.82373046875,
0.49365234375,
0.83837890625,
0.139404296875,
0.83837890625
],
[
0.51123046875,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.51123046875,
0.8359375
],
[
0.15283203125,
0.8388671875,
0.490234375,
0.84033203125,
0.490234375,
0.85498046875,
0.15283203125,
0.85400390625
],
[
0.5283203125,
0.84326171875,
0.85400390625,
0.845703125,
0.85400390625,
0.859375,
0.5283203125,
0.85693359375
],
[
0.15283203125,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.15283203125,
0.87060546875
],
[
0.5087890625,
0.859375,
0.87158203125,
0.859375,
0.87158203125,
0.869140625,
0.5087890625,
0.869140625
],
[
0.15283203125,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.15283203125,
0.8857421875
],
[
0.5146484375,
0.87060546875,
0.68701171875,
0.87060546875,
0.68701171875,
0.88134765625,
0.5146484375,
0.88134765625
],
[
0.5341796875,
0.88134765625,
0.76318359375,
0.88134765625,
0.76318359375,
0.89453125,
0.5341796875,
0.89453125
],
[
0.15283203125,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.15283203125,
0.90283203125
],
[
0.51123046875,
0.8935546875,
0.74658203125,
0.8935546875,
0.74658203125,
0.90673828125,
0.51123046875,
0.90673828125
],
[
0.15283203125,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.15283203125,
0.91845703125
],
[
0.509765625,
0.90625,
0.8369140625,
0.90380859375,
0.8369140625,
0.91748046875,
0.509765625,
0.91943359375
]
] |
[
"predictions that are falsely counted as errors due to",
"CLEANCoNLL, and comparatively evaluate",
"noisy annotation. We arguethat given the maturity",
"the true error rate of these approaches.",
"of research into high-resource NER, the research",
"We make the resource publicly available and en-",
"community requires a dataset with highly consis-",
"courage the community to use and further improve",
"tent,nearlynoise-freelabels inordertocorrectly",
"estimate the performance of state-of-the-art models",
"its quality.",
"and to better understand their remaining errors.",
"2",
"Creating CLEANCoNLL",
"Adding and leveraging entity links. To address",
"this issue, we present CLEANCoNLL, a revised",
"CLEANCoNLL was created in three distinct pha-",
"version of CoNLL-03 that significantly improves",
"ses of annotation, which we describe chronologi",
"annotation quality and consistency. Our main idea",
"cally in the following subsections:",
"is to introduce a second layer of annotation to the",
"1. We first (see Sec. 2.1) added a layer of entity",
"corpus that adds entity linking information. Entity",
"linking annotation by merging two existing",
"linking is the task of disambiguating named entities",
"resources, and developed a heuristic to au-",
"to a unique ID in a knowledge base, typically the",
"tomatically correct NER errors using entity",
"Wikipedia page representing a specific entity.",
"links.",
"As Figure I illustrates with an example sentence.",
"adding entity links serves two purposes: First, en-",
"2. We then (see Sec. 2.2) performed 3 iterative",
"tity links allow us to define automatic consistency",
"rounds of cross-checking to identify potential",
"checks on NER labels, since prior work found",
"inconsistencies, which we manually examined",
"that NER labels can be (partially) derived from",
"and corrected whenever necessary?",
"Wikipedia page IDs (Tedeschi et al., 2021). Sec-",
" ond, we find that the disambiguating information",
"3.Finally (see Sec.2.3), we developed spe",
"provided by entity links makes NER labels more",
"cial handling for adjectival affiliation entity",
"immediately transparent to human inspection, thus",
"names.",
"making it easier for annotators to spot mistakes and",
"2.1 Phase 1: Adding Entity Links",
"discuss disagreements.",
"Contributions. In more detail, our contributions",
"Base corpus. We use the version from Reiss et al",
"are the following:",
"(2020) as the starting point of our relabeling ef-",
"fort3. Reiss et al. (2020) already corrected 3.8%",
". We present CLEANCoNLL, the result of a",
"of all labels in CoNLL-03, including several major",
"comprehensive relabeling effort of CoNLL-03",
"sentence and entity mention boundary problems.",
"that includes updated, more consistent NER",
"While this version has less noise than the origi-",
"annotations in all three splits and integrates an",
"nal corpus, our own analysis (Sec. 3.4) and prior",
"additional layer of entity linking annotation.",
"work (Wang and Mueller, 2022) found that large",
"numbers of errors stillremain.",
". We detail a method for deriving NER annota-",
"Merging AIDA. For initial corpus creation, we",
"tion from Wikipedia links that leverages the",
" take advantage of the fact that AIDA (Hoffart et al.,",
"category system in Wikidata, and discuss how",
"2011), a well-known entity linking dataset, shares",
"this approach acts as consistency checks to",
"the same text base as CoNLL-03. Since the entity",
"safeguard annotation quality.",
"links in AIDA are expert-annotated by a group",
": We further describe an automatic approach",
"distinct of the original CoNLL-03 annotators, they",
"provide a layer of high quality annotation of an",
"to flag potential inconsistencies that we use",
"arguably more complex task.",
"throughout the manual re-annotation process,",
"However, merging these two corpora was not",
"which we refer to as \"cross-checking\".",
"straightforward since there are discrepancies both",
": We conduct extensive analyses of CLEAN-",
"in terms of sentence boundary definitions as well",
"CoNLL and present both quantitative results",
"2Manual annotation in all phases of the process was",
"of manual assessment of annotation quality in",
"carried out by the authors of this work. See the limitations",
"comparison to previous versions and qualita-",
"sectionforabriefdiscussion",
"3Obtained by using their instructions",
"tive examples of error corrections. We train a",
"fromhttps://github.com/CODAIT/",
"set of current state-of-the-art approaches on",
"Identifying-Incorrect-Labels-In-CoNLL-2003"
] |
{
"arxiv_id": "2310.16225",
"language": "en",
"page_idx": 1
}
|
||
2310.16225-en-2
|
2310.16225-en-2.jpg
|
<FigureHere>
Figure 2: A good portion of named entities in CoNLL-03 have manually assigned Wikipedia URLs in AIDA. In the example sentence "JAPAN" and "CHINA" are linked to the article of the referred sports teams from which we derive their Wikidata items and categories.
ORG. A full description of our mappings is provided in Appendix A.1. Using these mappings, we automatically mapped 4,163 (74.4%) of the 5,595 unique Wikipedia labels in AIDA to their NER type. However, as we prioritized precision over recall in the development of our manual mappings, $1,432(25.6\%)$ could not be classified automatically because either none or multiple of our heuristics took effect.
Manual annotation. These remaining 1,432 entity links were manually assigned NER types. This annotation was carried out separately for all labels by two experts who were provided with guidelines as well as the respective Wikipedia and Wikidata information. The inter-annotator agreement was 94.9%, cases of disagreement were subsequently resolved in discussion.
Using the resulting mappings, we updated all NER labels in the corpus to be consistent with the entity link
as the coverage of entity annotations. We found that of all 34,941 entities in the Reiss CoNLL version, 78.9% have an entity link in AIDA, while the remaining 21.1% (7,369 entities) do not. We found three main reasons for lack of coverage: $(I)$ Rare entities that at time of labeling had no corresponding Wikipedia article are unlabeled in AIDA. (2) The definition of what constitutes a named entity differs in some instances, with CoNLL-03 for instance labeling the span "Boeing 707" as entity, while AIDA links only "Boeing". (3) Some common entities like "EU" for unclear reasons are consistently unlabeled in AIDA.
Deriving NER labels from entity links. We derive NER labels from Wikipedia URL links by leveraging Wikidata$^{4}$, a manually curated knowledge graph that connects entities (represented by Wikipedia pages) through a set of relations. We leveraged two of these relations: (1) instance\_of, which connects an instance to the conceptual class it instantiates, (2) subclass\_of, which connects a conceptual class to a more general conceptual class.
Refer to Figure 2 for an example of this: The mention "CHINA" is linked to the entity China\_national\_football\_team",which is a Wikidata instance of "national association football team", which in turn is a subclass of "national sports team". With these two relations, we automatically select conceptual classes for each linked entity in our corpus.
For the most common of these classes, we manually defined a mapping to CoNLL-03 entity types. For instance, we manually defined that the Wikidata category "sports team" maps to the entity type
$^{4}$https://www.wikidata.org/
## 2.2 Phase 2: Consistency Checking
After Phase 1 of our relabeling effort, a number of
potential error sources remained:
$\bullet$ AIDA entity links, from which we derive NER labels in Phase 1, were missing for a subset of entities, and may themselves contain noise. For instance, we qualitatively observed instances in which national sports teams were incorrectly linked to the Wikipedia article of the country instead of the team.
$\bullet$ Though developed to favor precision, our heuristics of deriving NER labels from Wikidata categories may lead to an incorrect NER label assignment.
|
[
[
0.318115234375,
0.09027099609375,
0.39208984375,
0.09027099609375,
0.39208984375,
0.1051025390625,
0.318115234375,
0.1051025390625
],
[
0.4189453125,
0.09027099609375,
0.47607421875,
0.09027099609375,
0.47607421875,
0.1051025390625,
0.4189453125,
0.1051025390625
],
[
0.51123046875,
0.0914306640625,
0.6572265625,
0.0914306640625,
0.6572265625,
0.1051025390625,
0.51123046875,
0.1051025390625
],
[
0.69873046875,
0.0872802734375,
0.765625,
0.0872802734375,
0.765625,
0.097412109375,
0.69873046875,
0.097412109375
],
[
0.162841796875,
0.10272216796875,
0.2435302734375,
0.10272216796875,
0.2435302734375,
0.114013671875,
0.162841796875,
0.114013671875
],
[
0.4189453125,
0.1092529296875,
0.497802734375,
0.1092529296875,
0.497802734375,
0.11639404296875,
0.4189453125,
0.11639404296875
],
[
0.759765625,
0.11224365234375,
0.8017578125,
0.11224365234375,
0.8017578125,
0.11639404296875,
0.759765625,
0.11639404296875
],
[
0.432373046875,
0.119384765625,
0.49951171875,
0.119384765625,
0.49951171875,
0.12353515625,
0.432373046875,
0.12353515625
],
[
0.181396484375,
0.1258544921875,
0.295654296875,
0.1258544921875,
0.295654296875,
0.135986328125,
0.181396484375,
0.135986328125
],
[
0.343505859375,
0.1318359375,
0.4013671875,
0.1300048828125,
0.403076171875,
0.144287109375,
0.344970703125,
0.146728515625
],
[
0.42236328125,
0.1318359375,
0.61376953125,
0.1318359375,
0.61376953125,
0.1455078125,
0.42236328125,
0.1455078125
],
[
0.355224609375,
0.152587890625,
0.4072265625,
0.152587890625,
0.4072265625,
0.1602783203125,
0.355224609375,
0.1602783203125
],
[
0.349365234375,
0.1644287109375,
0.411376953125,
0.1644287109375,
0.411376953125,
0.16748046875,
0.349365234375,
0.16748046875
],
[
0.5634765625,
0.1591796875,
0.66748046875,
0.1591796875,
0.66748046875,
0.17041015625,
0.5634765625,
0.17041015625
],
[
0.59619140625,
0.201904296875,
0.7109375,
0.201904296875,
0.7109375,
0.208984375,
0.59619140625,
0.208984375
],
[
0.3955078125,
0.2476806640625,
0.43994140625,
0.248779296875,
0.43994140625,
0.255859375,
0.3955078125,
0.254638671875
],
[
0.68701171875,
0.248779296875,
0.7314453125,
0.248779296875,
0.7314453125,
0.255859375,
0.68701171875,
0.255859375
],
[
0.76318359375,
0.2452392578125,
0.8193359375,
0.243408203125,
0.8193359375,
0.254638671875,
0.76318359375,
0.257080078125
],
[
0.1175537109375,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.1175537109375,
0.283935546875
],
[
0.1175537109375,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.1175537109375,
0.298583984375
],
[
0.11419677734375,
0.298583984375,
0.405517578125,
0.2998046875,
0.405517578125,
0.3134765625,
0.11419677734375,
0.312255859375
],
[
0.1158447265625,
0.330078125,
0.489501953125,
0.330078125,
0.489501953125,
0.34326171875,
0.1158447265625,
0.34326171875
],
[
0.509765625,
0.327880859375,
0.8818359375,
0.330078125,
0.8818359375,
0.344482421875,
0.509765625,
0.342529296875
],
[
0.1158447265625,
0.344482421875,
0.491943359375,
0.344482421875,
0.491943359375,
0.358154296875,
0.1158447265625,
0.358154296875
],
[
0.5087890625,
0.344482421875,
0.8818359375,
0.3466796875,
0.8818359375,
0.361083984375,
0.5087890625,
0.359375
],
[
0.1158447265625,
0.3623046875,
0.489501953125,
0.3623046875,
0.489501953125,
0.3759765625,
0.1158447265625,
0.3759765625
],
[
0.509765625,
0.3623046875,
0.880859375,
0.3623046875,
0.880859375,
0.377197265625,
0.509765625,
0.377197265625
],
[
0.1175537109375,
0.377685546875,
0.489501953125,
0.377685546875,
0.489501953125,
0.392578125,
0.1175537109375,
0.392578125
],
[
0.5087890625,
0.377685546875,
0.880859375,
0.377197265625,
0.880859375,
0.391357421875,
0.5087890625,
0.392578125
],
[
0.1158447265625,
0.393798828125,
0.48779296875,
0.393798828125,
0.48779296875,
0.40673828125,
0.1158447265625,
0.40673828125
],
[
0.5087890625,
0.394287109375,
0.88330078125,
0.393798828125,
0.88330078125,
0.407958984375,
0.5087890625,
0.4091796875
],
[
0.1158447265625,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.1158447265625,
0.42333984375
],
[
0.5087890625,
0.4091796875,
0.8818359375,
0.410400390625,
0.8818359375,
0.424560546875,
0.5087890625,
0.42333984375
],
[
0.1158447265625,
0.427001953125,
0.489501953125,
0.427001953125,
0.489501953125,
0.440673828125,
0.1158447265625,
0.440673828125
],
[
0.51123046875,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.51123046875,
0.440673828125
],
[
0.1175537109375,
0.4423828125,
0.486083984375,
0.4423828125,
0.486083984375,
0.4560546875,
0.1175537109375,
0.4560546875
],
[
0.51123046875,
0.4423828125,
0.8818359375,
0.4423828125,
0.8818359375,
0.4560546875,
0.51123046875,
0.4560546875
],
[
0.1175537109375,
0.457763671875,
0.491943359375,
0.457763671875,
0.491943359375,
0.471435546875,
0.1175537109375,
0.471435546875
],
[
0.51123046875,
0.458984375,
0.59765625,
0.458984375,
0.59765625,
0.47021484375,
0.51123046875,
0.47021484375
],
[
0.1175537109375,
0.474365234375,
0.490234375,
0.474365234375,
0.490234375,
0.488037109375,
0.1175537109375,
0.488037109375
],
[
0.5087890625,
0.473876953125,
0.8818359375,
0.474365234375,
0.8818359375,
0.4892578125,
0.5087890625,
0.488037109375
],
[
0.11419677734375,
0.4912109375,
0.491943359375,
0.4912109375,
0.491943359375,
0.5048828125,
0.11419677734375,
0.5048828125
],
[
0.509765625,
0.4912109375,
0.8818359375,
0.4912109375,
0.8818359375,
0.5048828125,
0.509765625,
0.5048828125
],
[
0.1158447265625,
0.505859375,
0.49365234375,
0.50732421875,
0.49365234375,
0.521484375,
0.1158447265625,
0.52099609375
],
[
0.509765625,
0.50732421875,
0.8818359375,
0.505859375,
0.8818359375,
0.52099609375,
0.509765625,
0.521484375
],
[
0.1175537109375,
0.5224609375,
0.330078125,
0.5224609375,
0.330078125,
0.5361328125,
0.1175537109375,
0.5361328125
],
[
0.51123046875,
0.52392578125,
0.8818359375,
0.52392578125,
0.8818359375,
0.53759765625,
0.51123046875,
0.53759765625
],
[
0.1175537109375,
0.5390625,
0.491943359375,
0.5390625,
0.491943359375,
0.552734375,
0.1175537109375,
0.552734375
],
[
0.509765625,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.552734375,
0.509765625,
0.552734375
],
[
0.11419677734375,
0.55419921875,
0.48779296875,
0.55419921875,
0.48779296875,
0.568359375,
0.11419677734375,
0.568359375
],
[
0.509765625,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.509765625,
0.5693359375
],
[
0.1158447265625,
0.57177734375,
0.49365234375,
0.57080078125,
0.49365234375,
0.5849609375,
0.1158447265625,
0.5859375
],
[
0.5087890625,
0.57080078125,
0.8818359375,
0.57177734375,
0.8818359375,
0.5859375,
0.5087890625,
0.5849609375
],
[
0.1175537109375,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6015625,
0.1175537109375,
0.6015625
],
[
0.509765625,
0.58837890625,
0.6826171875,
0.58837890625,
0.6826171875,
0.6015625,
0.509765625,
0.6015625
],
[
0.1175537109375,
0.60400390625,
0.48779296875,
0.60400390625,
0.48779296875,
0.6181640625,
0.1175537109375,
0.6181640625
],
[
0.52880859375,
0.60400390625,
0.8818359375,
0.60400390625,
0.8818359375,
0.6181640625,
0.52880859375,
0.6181640625
],
[
0.1175537109375,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63427734375,
0.1175537109375,
0.63427734375
],
[
0.5087890625,
0.619140625,
0.8818359375,
0.62060546875,
0.8818359375,
0.63525390625,
0.5087890625,
0.63427734375
],
[
0.118408203125,
0.63818359375,
0.48779296875,
0.63818359375,
0.48779296875,
0.6484375,
0.118408203125,
0.6484375
],
[
0.509765625,
0.6357421875,
0.59423828125,
0.6357421875,
0.59423828125,
0.65087890625,
0.509765625,
0.65087890625
],
[
0.1158447265625,
0.65283203125,
0.486083984375,
0.65283203125,
0.486083984375,
0.66650390625,
0.1158447265625,
0.66650390625
],
[
0.1158447265625,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.68310546875,
0.1158447265625,
0.68310546875
],
[
0.51123046875,
0.6630859375,
0.80029296875,
0.66455078125,
0.80029296875,
0.6787109375,
0.51123046875,
0.677734375
],
[
0.1175537109375,
0.68505859375,
0.156982421875,
0.68505859375,
0.156982421875,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.51318359375,
0.68505859375,
0.8818359375,
0.68505859375,
0.8818359375,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.133544921875,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.133544921875,
0.71337890625
],
[
0.51123046875,
0.703125,
0.75732421875,
0.703125,
0.75732421875,
0.71630859375,
0.51123046875,
0.71630859375
],
[
0.11419677734375,
0.71630859375,
0.483642578125,
0.71630859375,
0.483642578125,
0.72998046875,
0.11419677734375,
0.72998046875
],
[
0.118408203125,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.53076171875,
0.72998046875,
0.880859375,
0.72998046875,
0.880859375,
0.74365234375,
0.53076171875,
0.74365234375
],
[
0.118408203125,
0.75048828125,
0.484375,
0.75048828125,
0.484375,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.54736328125,
0.74658203125,
0.8818359375,
0.74658203125,
0.8818359375,
0.76025390625,
0.54736328125,
0.76025390625
],
[
0.1158447265625,
0.76416015625,
0.484375,
0.76416015625,
0.484375,
0.77783203125,
0.1158447265625,
0.77783203125
],
[
0.54833984375,
0.76416015625,
0.88330078125,
0.76416015625,
0.88330078125,
0.77783203125,
0.54833984375,
0.77783203125
],
[
0.11419677734375,
0.7822265625,
0.489501953125,
0.78076171875,
0.489501953125,
0.7958984375,
0.11419677734375,
0.79638671875
],
[
0.54833984375,
0.77978515625,
0.8818359375,
0.77978515625,
0.8818359375,
0.79345703125,
0.54833984375,
0.79345703125
],
[
0.1158447265625,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.1158447265625,
0.8125
],
[
0.54833984375,
0.79736328125,
0.87890625,
0.79736328125,
0.87890625,
0.80810546875,
0.54833984375,
0.80810546875
],
[
0.1158447265625,
0.8154296875,
0.267822265625,
0.8154296875,
0.267822265625,
0.8291015625,
0.1158447265625,
0.8291015625
],
[
0.54833984375,
0.8125,
0.8818359375,
0.8125,
0.8818359375,
0.826171875,
0.54833984375,
0.826171875
],
[
0.1351318359375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.1351318359375,
0.84423828125
],
[
0.54736328125,
0.82666015625,
0.78271484375,
0.82763671875,
0.78271484375,
0.8427734375,
0.54736328125,
0.84130859375
],
[
0.1175537109375,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.1175537109375,
0.86083984375
],
[
0.5341796875,
0.85546875,
0.88330078125,
0.85546875,
0.88330078125,
0.87060546875,
0.5341796875,
0.87060546875
],
[
0.1175537109375,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.1175537109375,
0.876953125
],
[
0.11419677734375,
0.876953125,
0.489501953125,
0.87744140625,
0.489501953125,
0.8955078125,
0.11419677734375,
0.89453125
],
[
0.54736328125,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.54736328125,
0.8857421875
],
[
0.54833984375,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.54833984375,
0.90283203125
],
[
0.1376953125,
0.90478515625,
0.334228515625,
0.90478515625,
0.334228515625,
0.91845703125,
0.1376953125,
0.91845703125
],
[
0.54736328125,
0.90478515625,
0.67822265625,
0.90478515625,
0.67822265625,
0.91845703125,
0.54736328125,
0.91845703125
]
] |
[
"SOCCER",
"JAPAN",
"GET LUCKY WIN,",
"WIKIPEDIA",
"WIKIPEDIA",
"Japannational",
"foothalte",
"notbaltea",
"China national football team",
"CHINA",
"IN SURPRISE DEFEAT.",
"anationa",
"oratea",
"CoNLL-AIDA",
"Japannationalfootballtear",
"football team",
"foothallteam",
"+ORG",
"Figure 2: A good portion of named entities in CoNLL-03 have manually assigned Wikipedia URLs in AIDA. In",
"the example sentence \"JAPAN\" and \"CHINA\" are linked to the article of the referred sports teams from which we",
"derive their Wikidata items and categories.",
"as the coverage of entity annotations.Wefound",
"ORG. A full description of our mappings is pro-",
"that of all 34.941 entities in the Reiss CoNLL ver-",
"vided in Appendix A.1. Using these mappings, we",
"sion, 78.9% have an entity link in AIDA, while the",
"automatically mapped 4,163 (74.4%) of the 5,595",
"remaining 21.1% (7,369 entities) do not. We found",
"unique Wikipedia labels in AIDA to their NER",
"three main reasons for lack of coverage: (l) Rare",
"type.However, as we prioritized precision over",
"entities that at time of labeling had nocorrespond-",
"recall in the development of our manual mappings,",
"ing Wikipedia article are unlabeled in AIDA. (2)",
"1,432 (25.6%) could not be classified automatically",
"The definition of what constitutes a named entity",
"because either none or multiple of our heuristics",
"differs in some instances, with CoNLL-03 for in-",
"took effect.",
"stance labeling the span \"Boeing 707\" as entity,",
"Manual annotation. These remaining 1,432 entity",
"while AIDA links only\"Boeing\". (3) Some com-",
"links were manually assigned NER types. This",
"mon entities like\"EU\" for unclear reasons are con-",
"annotation was carried out separately for all labels",
"sistently unlabeled in AIDA.",
"by two experts who were provided with guidelines",
"Deriving NER labels from entity links. We de-",
" as well as the respective Wikipedia and Wikidata",
"rive NER labels from Wikipedia URL links by",
"information. The inter-annotator agreement was",
"leveraging Wikidata, a manually curated knowl-",
"94.9%, cases of disagreement were subsequently",
"edge graph that connects entities (represented by",
"resolved in discussion.",
"Wikipedia pages) through a set of relations. We",
"Using the resulting mappings, we updated all",
"leveraged two of these relations: (1) instance_of,",
"NER labels in the corpus to be consistent with the",
"whichconnectsaninstancetotheconceptualclass",
"entity link.",
"it instantiates, (2) subclass_of, which connects",
"a conceptual class to a more general conceptual",
"2.2Phase 2: Consistency Checking",
"class.",
"After Phase 1 of our relabeling effort, a number of",
"Refer to Figure 2 for an example of this:",
"potential error sources remained:",
"The mention \"CHINA\" is linked to the entity",
"\"China_ national football team\", which is a Wiki-",
". AIDA entity links, from which we derive NER",
"data instance of\"national association football",
"labels in Phase I, were missing for a sub-",
"team\", which in turn is a subclass of \"national",
"set of entities, and may themselves contain",
"sports team\". With these two relations, we auto-",
"noise. For instance, we qualitatively observed",
"matically select conceptual classes for each linked",
"instancesinwhichnationalsportsteamswere",
"entity in our corpus.",
"incorrectly linked to the Wikipedia article of",
"For the most common of these classes, we man-",
"the country instead of the team.",
"ually defined a mapping to CoNLL-03 entity types.",
": Though developed to favor precision, our",
"For instance, we manually defined that the Wiki-",
"data category \"sports team\" maps to the entity type",
"heuristics of deriving NER labels from Wiki-",
"data categories may lead to an incorrect NER",
"thttps://www.wikidata.org/",
"label assignment."
] |
{
"arxiv_id": "2310.16225",
"language": "en",
"page_idx": 2
}
|
||
2310.16225-en-3
|
2310.16225-en-3.jpg
|
$\bullet$ Leveraging AIDA entities is ill-suited to correct wrongly spanned mention boundaries or tokenization problems.
## 2.2.1 Cross-Checking and Correction
To identify and correct these errors, we follow a cross-checking" approach as introduced in Wang et al. (2019). The main idea is to predict labels for a small part of the corpus using the entirety of the remaining corpus, but make sure that no entity names are memorized.
To accomplish this, we collect all unique surface forms $\mathcal{S}$ of mentions in the corpus and group them randomly into 20 distinct batches $S_i$ of around 500 surface forms each. For each of these 20 batches of surface forms, we create two splits: $\mathcal{D}_i^{test}$, which contains all sentences that have one of the surface forms $\mathcal{S}_i$ of the respective batch. $\mathcal{D}_i^train$, which contains all other sentences, i.e. the sentences that contain only $\mathcal{S}\setminus\mathcal{S}_i$ (in other words: do not contain any of the surface forms of the respective batch). For each batch, we train a state-of-the-art model on $\mathcal{D}_i^{train}$ and use it to predict on $\mathcal{D}_i^{test}.$ All predictions that differ from the annotated NER labels are flagged as potential errors.
Manual error resolution. All annotations flagged as potential errors were then manually inspected by annotators. They were given the entity span within document context and the different label possibilities including their source: the current gold annotation, the diverging model prediction, as well as the Wikipedia label and its automatically derived NER label. Based on this information, they were asked to a) decide which NER label to use, and b) add or change the current Wikipedia label if necessary. See Figure 3 for an illustration of this interface.
Annotation process. In the first round of annotation, all spans were independently labeled by two annotators (agreement was 79.3%), after which the guidelines were refined and the rest was done primarily by one. Annotators were given the option to mark cases of doubt $(2.4\%);$ these were given to the second annotator for independent annotation.
Annotators were further asked to take note of error classes that may go beyond a single instance. For instance, we noted general problems with annotation of time zones ("GMT", "EST"): Some were labeled as entities and others were not. In such cases, we clarified the annotation guidelines and manually corrected all instances in the corpus.
<table>
<tbody>
<tr>
<td>You 祥祥林</td>
<td>notated 75 out of 1020 spans! 赫科学材料林林林林</td>
</tr>
<tr>
<td>$spa:$</td>
<td>and $1abels:$ pred: ORG, gold: Loc, aida: England https://en.wikipedia.org/wiki/England</td>
</tr>
<tr>
<td>$(nr$ 祥祥林</td>
<td>with this mention and pattern) 赫科学材料林林林林</td>
</tr>
<tr>
<td>$In$</td>
<td>and marked: manager Glenn Hoddle called up uncapped Everton r Andy Hinchcliffe on Sunday to the national squad for ning World Cup qualifier against Moldova next weekend.</td>
</tr>
<tr>
<td>[?]</td>
<td>Pred: ORG $d:$ do with this $88pan?:use$ $LOC$</td>
</tr>
<tr>
<td> </td>
<td>2: URG like last: label: LOC, aida: England $("0^{\prime\prime})$</td>
</tr>
<tr>
<td> </td>
<td>ge url (NER can be changed afterwards) e context re) URLs nearby</td>
</tr>
</tbody>
</table>
Figure 3: Example of the annotator interface during cross-checking. The mention "England" is annotated as "LOC", but the model predicted "ORG", which is correct. The wrong annotation originates from an error in AIDA, where the mention is linked to the country's article ("England") instead of the football team. Our annotator decided on NER label "ORG" and corrected the entity link to the Wikipedia page “ England$\_$national$\_$football$\_$team" .
2.2.2 Iterative Cross-Checking in 3 Rounds
We executed 3 rounds of cross-checking. After each round, labels were manually corrected and the improved corpus used to train the next models:
1. Entity detection and classification. The first round of cross-checking directly followed Wang et al. (2019) to train a model to jointly detect mention boundaries as well as classify them using a BIOES sequence tagging approach (Schweter and Akbik, 2020). Over all batches, this resulted in 1,020 flagged annotations, resulting in 521 corrected labels after manual inspection.
$2. \textbf{ Only entity classification. The second round}$ of cross-checking instead trained a model to only classify already provided entity mentions. See Appendix A.7 for details of this model. This flagged a further 709 instances for inspection, resulting in 98 additional corrected NER labels.
3. Only entity detection. The third round of crosschecking focused only on consistent definitions of entity mentions. To this end, we trained a BIOES sequence labeler to predict only entity boundaries (but not their type). Over all batches, we collected 726 sentences with need of manual decision concerning mention boundaries (1,720 spans when counting the predicted and gold ones separately
|
[
[
0.1368408203125,
0.0843505859375,
0.489501953125,
0.08489990234375,
0.489501953125,
0.09979248046875,
0.1368408203125,
0.09857177734375
],
[
0.51904296875,
0.0885009765625,
0.55908203125,
0.0885009765625,
0.55908203125,
0.0926513671875,
0.51904296875,
0.0926513671875
],
[
0.609375,
0.08489990234375,
0.7314453125,
0.08489990234375,
0.7314453125,
0.0955810546875,
0.609375,
0.0955810546875
],
[
0.5205078125,
0.0968017578125,
0.6171875,
0.0968017578125,
0.6171875,
0.1009521484375,
0.5205078125,
0.1009521484375
],
[
0.15283203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.15283203125,
0.11639404296875
],
[
0.5205078125,
0.11224365234375,
0.52880859375,
0.11224365234375,
0.52880859375,
0.11639404296875,
0.5205078125,
0.11639404296875
],
[
0.6943359375,
0.11224365234375,
0.7138671875,
0.11224365234375,
0.7138671875,
0.11517333984375,
0.6943359375,
0.11517333984375
],
[
0.15283203125,
0.1181640625,
0.32568359375,
0.1181640625,
0.32568359375,
0.1318359375,
0.15283203125,
0.1318359375
],
[
0.6943359375,
0.119384765625,
0.7255859375,
0.119384765625,
0.7255859375,
0.122314453125,
0.6943359375,
0.122314453125
],
[
0.68408203125,
0.1258544921875,
0.6962890625,
0.1258544921875,
0.6962890625,
0.12890625,
0.68408203125,
0.12890625
],
[
0.67919921875,
0.133056640625,
0.76904296875,
0.133056640625,
0.76904296875,
0.14013671875,
0.67919921875,
0.14013671875
],
[
0.779296875,
0.1312255859375,
0.880859375,
0.1312255859375,
0.880859375,
0.141357421875,
0.779296875,
0.141357421875
],
[
0.1158447265625,
0.15087890625,
0.424072265625,
0.1519775390625,
0.424072265625,
0.166259765625,
0.1158447265625,
0.1650390625
],
[
0.51904296875,
0.1561279296875,
0.611328125,
0.1561279296875,
0.611328125,
0.163330078125,
0.51904296875,
0.163330078125
],
[
0.51904296875,
0.1728515625,
0.5283203125,
0.171630859375,
0.52880859375,
0.17578125,
0.5205078125,
0.177001953125
],
[
0.5322265625,
0.17041015625,
0.6337890625,
0.17041015625,
0.6337890625,
0.1776123046875,
0.5322265625,
0.1776123046875
],
[
0.57275390625,
0.177001953125,
0.83642578125,
0.177001953125,
0.83642578125,
0.18701171875,
0.57275390625,
0.18701171875
],
[
0.5400390625,
0.185302734375,
0.69189453125,
0.1859130859375,
0.69189453125,
0.193603515625,
0.5400390625,
0.1923828125
],
[
0.1175537109375,
0.174560546875,
0.48779296875,
0.174560546875,
0.48779296875,
0.188232421875,
0.1175537109375,
0.188232421875
],
[
0.7001953125,
0.189453125,
0.75,
0.189453125,
0.75,
0.1923828125,
0.7001953125,
0.1923828125
],
[
0.78515625,
0.18701171875,
0.8671875,
0.1859130859375,
0.8671875,
0.193603515625,
0.78515625,
0.19482421875
],
[
0.59765625,
0.193603515625,
0.8671875,
0.193603515625,
0.8671875,
0.20068359375,
0.59765625,
0.20068359375
],
[
0.54248046875,
0.1953125,
0.599609375,
0.1953125,
0.599609375,
0.1995849609375,
0.54248046875,
0.1995849609375
],
[
0.118408203125,
0.191162109375,
0.486083984375,
0.191162109375,
0.486083984375,
0.204833984375,
0.118408203125,
0.204833984375
],
[
0.1175537109375,
0.207275390625,
0.48779296875,
0.207275390625,
0.48779296875,
0.2203369140625,
0.1175537109375,
0.2203369140625
],
[
0.51708984375,
0.207275390625,
0.767578125,
0.207275390625,
0.767578125,
0.21728515625,
0.51708984375,
0.21728515625
],
[
0.5400390625,
0.218505859375,
0.587890625,
0.218505859375,
0.587890625,
0.22265625,
0.5400390625,
0.22265625
],
[
0.118408203125,
0.2244873046875,
0.486083984375,
0.2244873046875,
0.486083984375,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.5751953125,
0.2327880859375,
0.60107421875,
0.2327880859375,
0.60107421875,
0.2369384765625,
0.5751953125,
0.2369384765625
],
[
0.67529296875,
0.2327880859375,
0.68994140625,
0.2327880859375,
0.68994140625,
0.2369384765625,
0.67529296875,
0.2369384765625
],
[
0.7001953125,
0.232177734375,
0.77099609375,
0.232177734375,
0.77099609375,
0.2392578125,
0.7001953125,
0.2392578125
],
[
0.1158447265625,
0.2392578125,
0.486083984375,
0.2392578125,
0.486083984375,
0.2529296875,
0.1158447265625,
0.2529296875
],
[
0.53564453125,
0.2340087890625,
0.54736328125,
0.2340087890625,
0.54736328125,
0.2369384765625,
0.53564453125,
0.2369384765625
],
[
0.5341796875,
0.2476806640625,
0.57080078125,
0.24462890625,
0.57275390625,
0.2529296875,
0.53564453125,
0.255859375
],
[
0.1158447265625,
0.254638671875,
0.287109375,
0.254638671875,
0.287109375,
0.268310546875,
0.1158447265625,
0.268310546875
],
[
0.5341796875,
0.254150390625,
0.60546875,
0.254150390625,
0.60546875,
0.26123046875,
0.5341796875,
0.26123046875
],
[
0.5341796875,
0.26123046875,
0.58447265625,
0.26318359375,
0.58447265625,
0.270751953125,
0.5341796875,
0.268310546875
],
[
0.611328125,
0.262451171875,
0.779296875,
0.262451171875,
0.779296875,
0.26953125,
0.611328125,
0.26953125
],
[
0.133544921875,
0.271484375,
0.484375,
0.271484375,
0.484375,
0.28515625,
0.133544921875,
0.28515625
],
[
0.53662109375,
0.279052734375,
0.5517578125,
0.279052734375,
0.5517578125,
0.283203125,
0.53662109375,
0.283203125
],
[
0.5576171875,
0.27978515625,
0.58837890625,
0.27978515625,
0.58837890625,
0.283203125,
0.5576171875,
0.283203125
],
[
0.59375,
0.27978515625,
0.61279296875,
0.27978515625,
0.61279296875,
0.283203125,
0.59375,
0.283203125
],
[
0.63623046875,
0.27783203125,
0.6572265625,
0.27783203125,
0.6572265625,
0.28515625,
0.63623046875,
0.28515625
],
[
0.1158447265625,
0.287353515625,
0.48779296875,
0.2880859375,
0.48779296875,
0.302734375,
0.1158447265625,
0.3017578125
],
[
0.5322265625,
0.289306640625,
0.81005859375,
0.289306640625,
0.81005859375,
0.302734375,
0.5322265625,
0.302734375
],
[
0.5341796875,
0.297607421875,
0.599609375,
0.297607421875,
0.599609375,
0.30810546875,
0.5341796875,
0.30810546875
],
[
0.1158447265625,
0.3046875,
0.48779296875,
0.3046875,
0.48779296875,
0.318359375,
0.1158447265625,
0.318359375
],
[
0.5341796875,
0.30810546875,
0.59765625,
0.30810546875,
0.59765625,
0.3154296875,
0.5341796875,
0.3154296875
],
[
0.1158447265625,
0.320556640625,
0.48779296875,
0.320556640625,
0.48779296875,
0.334228515625,
0.1158447265625,
0.334228515625
],
[
0.11419677734375,
0.3349609375,
0.48779296875,
0.333251953125,
0.48779296875,
0.350830078125,
0.11419677734375,
0.352783203125
],
[
0.51123046875,
0.330810546875,
0.8671875,
0.33203125,
0.8671875,
0.3466796875,
0.51123046875,
0.345703125
],
[
0.1158447265625,
0.3515625,
0.486083984375,
0.3515625,
0.486083984375,
0.365234375,
0.1158447265625,
0.365234375
],
[
0.51123046875,
0.3466796875,
0.875,
0.3466796875,
0.875,
0.35986328125,
0.51123046875,
0.35986328125
],
[
0.51123046875,
0.361083984375,
0.85546875,
0.361083984375,
0.85546875,
0.374755859375,
0.51123046875,
0.374755859375
],
[
0.1175537109375,
0.36767578125,
0.48779296875,
0.36767578125,
0.48779296875,
0.3818359375,
0.1175537109375,
0.3818359375
],
[
0.509765625,
0.37353515625,
0.8671875,
0.374755859375,
0.8671875,
0.3896484375,
0.509765625,
0.388427734375
],
[
0.1175537109375,
0.385986328125,
0.484375,
0.385986328125,
0.484375,
0.396728515625,
0.1175537109375,
0.396728515625
],
[
0.509765625,
0.3896484375,
0.85205078125,
0.3896484375,
0.85205078125,
0.402587890625,
0.509765625,
0.402587890625
],
[
0.1158447265625,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.4150390625,
0.1158447265625,
0.4150390625
],
[
0.509765625,
0.40380859375,
0.86572265625,
0.40380859375,
0.86572265625,
0.41748046875,
0.509765625,
0.41748046875
],
[
0.1175537109375,
0.41748046875,
0.489501953125,
0.41748046875,
0.489501953125,
0.43115234375,
0.1175537109375,
0.43115234375
],
[
0.5087890625,
0.41748046875,
0.85986328125,
0.416259765625,
0.85986328125,
0.43115234375,
0.5087890625,
0.431640625
],
[
0.1175537109375,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.446533203125,
0.1175537109375,
0.446533203125
],
[
0.50732421875,
0.43115234375,
0.85986328125,
0.431640625,
0.85986328125,
0.446533203125,
0.50732421875,
0.4453125
],
[
0.11248779296875,
0.446533203125,
0.490234375,
0.446533203125,
0.490234375,
0.46435546875,
0.11248779296875,
0.46435546875
],
[
0.51318359375,
0.446533203125,
0.75146484375,
0.446533203125,
0.75146484375,
0.460205078125,
0.51318359375,
0.460205078125
],
[
0.1175537109375,
0.465576171875,
0.486083984375,
0.465576171875,
0.486083984375,
0.478515625,
0.1175537109375,
0.478515625
],
[
0.1175537109375,
0.482177734375,
0.313232421875,
0.482177734375,
0.313232421875,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.51318359375,
0.476806640625,
0.86376953125,
0.476806640625,
0.86376953125,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.1175537109375,
0.49951171875,
0.48779296875,
0.49951171875,
0.48779296875,
0.51318359375,
0.1175537109375,
0.51318359375
],
[
0.509765625,
0.49951171875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.509765625,
0.51318359375
],
[
0.1158447265625,
0.5166015625,
0.48779296875,
0.5166015625,
0.48779296875,
0.52978515625,
0.1158447265625,
0.52978515625
],
[
0.51123046875,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.1158447265625,
0.5322265625,
0.48779296875,
0.5322265625,
0.48779296875,
0.5458984375,
0.1158447265625,
0.5458984375
],
[
0.5087890625,
0.533203125,
0.85791015625,
0.533203125,
0.85791015625,
0.54638671875,
0.5087890625,
0.54638671875
],
[
0.11419677734375,
0.54638671875,
0.486083984375,
0.54638671875,
0.486083984375,
0.56005859375,
0.11419677734375,
0.56005859375
],
[
0.509765625,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.509765625,
0.56787109375
],
[
0.1158447265625,
0.56396484375,
0.48779296875,
0.56396484375,
0.48779296875,
0.57763671875,
0.1158447265625,
0.57763671875
],
[
0.5087890625,
0.5693359375,
0.8818359375,
0.57080078125,
0.8818359375,
0.5849609375,
0.5087890625,
0.58447265625
],
[
0.1158447265625,
0.580078125,
0.48779296875,
0.580078125,
0.48779296875,
0.59423828125,
0.1158447265625,
0.59423828125
],
[
0.51123046875,
0.5849609375,
0.88330078125,
0.5859375,
0.88330078125,
0.60107421875,
0.51123046875,
0.599609375
],
[
0.1158447265625,
0.595703125,
0.48779296875,
0.595703125,
0.48779296875,
0.609375,
0.1158447265625,
0.609375
],
[
0.51123046875,
0.6015625,
0.8818359375,
0.6025390625,
0.8818359375,
0.61767578125,
0.51123046875,
0.6162109375
],
[
0.1158447265625,
0.60986328125,
0.48779296875,
0.61083984375,
0.48779296875,
0.6259765625,
0.1158447265625,
0.62451171875
],
[
0.51123046875,
0.61767578125,
0.8818359375,
0.6181640625,
0.8818359375,
0.6328125,
0.51123046875,
0.6318359375
],
[
0.1175537109375,
0.62744140625,
0.48779296875,
0.62744140625,
0.48779296875,
0.64111328125,
0.1175537109375,
0.64111328125
],
[
0.51318359375,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.51318359375,
0.6484375
],
[
0.11419677734375,
0.6435546875,
0.486083984375,
0.6435546875,
0.486083984375,
0.65673828125,
0.11419677734375,
0.65673828125
],
[
0.51318359375,
0.65087890625,
0.88330078125,
0.65087890625,
0.88330078125,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.1158447265625,
0.6611328125,
0.48779296875,
0.6611328125,
0.48779296875,
0.6748046875,
0.1158447265625,
0.6748046875
],
[
0.509765625,
0.66748046875,
0.78857421875,
0.66748046875,
0.78857421875,
0.68115234375,
0.509765625,
0.68115234375
],
[
0.1175537109375,
0.677734375,
0.187255859375,
0.677734375,
0.187255859375,
0.689453125,
0.1175537109375,
0.689453125
],
[
0.1175537109375,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51318359375,
0.689453125,
0.8818359375,
0.689453125,
0.8818359375,
0.703125,
0.51318359375,
0.703125
],
[
0.1158447265625,
0.71044921875,
0.48779296875,
0.71044921875,
0.48779296875,
0.7236328125,
0.1158447265625,
0.7236328125
],
[
0.509765625,
0.7060546875,
0.8818359375,
0.7060546875,
0.8818359375,
0.7197265625,
0.509765625,
0.7197265625
],
[
0.1158447265625,
0.7255859375,
0.489501953125,
0.7255859375,
0.489501953125,
0.74072265625,
0.1158447265625,
0.74072265625
],
[
0.509765625,
0.72021484375,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.509765625,
0.7353515625
],
[
0.1175537109375,
0.7412109375,
0.489501953125,
0.7412109375,
0.489501953125,
0.7548828125,
0.1175537109375,
0.7548828125
],
[
0.51123046875,
0.73681640625,
0.8818359375,
0.73681640625,
0.8818359375,
0.751953125,
0.51123046875,
0.751953125
],
[
0.1158447265625,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.7724609375,
0.1158447265625,
0.7724609375
],
[
0.509765625,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76708984375,
0.509765625,
0.76708984375
],
[
0.11419677734375,
0.7724609375,
0.48779296875,
0.77392578125,
0.48779296875,
0.7880859375,
0.11419677734375,
0.78759765625
],
[
0.509765625,
0.7685546875,
0.77490234375,
0.7685546875,
0.77490234375,
0.7822265625,
0.509765625,
0.7822265625
],
[
0.118408203125,
0.7919921875,
0.47607421875,
0.7919921875,
0.47607421875,
0.802734375,
0.118408203125,
0.802734375
],
[
0.51318359375,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8046875,
0.51318359375,
0.8046875
],
[
0.1368408203125,
0.80859375,
0.486083984375,
0.80859375,
0.486083984375,
0.8193359375,
0.1368408203125,
0.8193359375
],
[
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
],
[
0.5087890625,
0.82373046875,
0.8818359375,
0.822265625,
0.8818359375,
0.83740234375,
0.5087890625,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.491943359375,
0.84033203125,
0.491943359375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.5087890625,
0.84130859375,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.5087890625,
0.85498046875
],
[
0.1158447265625,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1158447265625,
0.869140625
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.459228515625,
0.90478515625,
0.459228515625,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
". Leveraging AIDA entities is ill-suited to cor-",
"You.have",
"d 75 out of 1020 spans",
"生带#世花#带批花祛花带批花批花",
"rect wrongly spanned mention boundaries or",
"BD",
"",
"tokenization problems.",
"",
"",
"https://en.wikipe",
"La.org/wiki/England",
"2.2.1 Cross-Checking and Correction",
"###种营#批####批###垫",
"Tn",
"context and marked",
"d manager Glenn Hoddle called up uncapped Everton",
"defender Andy Hinchcliffe on",
"To identify and correct these errors, we follow a",
"aO",
"cional squadfor",
"World Cup qualifier against Moldova next weekend",
"heopening",
"\"cross-checking\" approach as introduced in Wang",
"et al. (2019). The main idea is to predict labels",
"[?] What to do with this span?: use Pred: ORG",
"eep cold:",
"forasmallpartofthecorpususingtheentiretyof",
"Tike",
"TOC",
"aida:England",
"the remaining corpus, but make sure that no entity",
"o:",
"USe PER",
"names are memorized.",
"no label(\"o\"",
"add/chang",
"(NER can be changed afterwards)",
"To accomplish this, we collect all unique surface",
"sho",
"Ffmore",
"RT",
"arby",
"forms S of mentions in the corpus and group them",
"mark: SECOND OPINION (still needs your annotation!)",
"<<<previous",
"randomly into 20 distinct batches S; of around 500",
"clear screen",
"surface forms each. For each of these 20 batches of",
"surface forms, we create two splits: Dtest, which",
"Figure 3: Example of the annotator interface during",
"contains all sentences that have one of the surface",
"cross-checking. The mention \"England\" is annotated",
"as\"LOC\", but the model predicted \"ORG\", which",
"forms S, of the respective batch. Dtrain, which",
"is correct. The wrong annotation originates from an",
"contains allother sentences.i.e.thesentencesthat",
"error in AIDA, where the mention is linked to the",
"contain only S S; (in other words: do not contain",
"country's article (\"England\") instead of the football",
"any of the surface forms of the respective batch).",
"team. Our annotator decided on NER label \"ORG\"",
"For each batch, we train a state-of-the-art model",
"and corrected the entity link to the Wikipedia page",
"on Dtrain and use it to predict on Dtest. All predic-",
"\"England_national_football_team\".",
"tions that differ from the annotated NER labels are",
"flagged as potential errors.",
"2.2.2Iterative Cross-Checking in 3 Rounds",
"Manual error resolution. All annotations flagged",
"We executed 3 rounds of cross-checking. After",
"as potential errors were then manually inspected",
"each round.labels were manually corrected and the",
"by annotators. They were given the entity span",
"improved corpus used to train the next models:",
"within document context and the different label",
"1. Entity detection and classification. The first",
"possibilities including their source: the current gold",
"round of cross-checking directly followed Wang",
"annotation, the diverging model prediction, as well",
"et al. (2019) to train a model to jointly detect men-",
"as the Wikipedia label and its automatically derived",
"tion boundaries as well as classify them using a",
"NER label. Based on this information, they were",
"BIOES sequence tagging approach (Schweter and",
"asked to a) decide which NER label to use, and",
"Akbik, 2020). Over all batches, this resulted in",
"b) add or change the current Wikipedia label if",
"1,020 flagged annotations, resulting in 521 cor-",
"necessary. See Figure 3 for an illustration of this",
"rected labels after manual inspection.",
"interface.",
"Annotation process. In the first round of annota-",
"2. Only entity classification. The second round",
"tion, all spans were independently labeled by two",
" of cross-checking instead trained a model to only",
"annotators (agreement was 79.3%), after which the",
"classify already provided entity mentions. See Ap",
"guidelines were refined and the rest was done pri-",
"pendix A.7 for details of this model. This flagged",
"marily by one. Annotators were given the option",
"a further 709 instances for inspection, resulting in",
"to mark cases of doubt (2.4%); these were given to",
"98 additional corrected NER labels.",
"the second annotatorforindependentannotation.",
"3. Only entity detection. The third round of cross-",
"Annotators werefurtherasked totakenoteof",
"checking focused only on consistent definitions of",
"error classes that may go beyond a single instance.",
"entity mentions. To this end, we trained a BIOES",
"For instance, we noted general problems with anno-",
" sequence labeler to predict only entity boundaries",
"tation of time zones (\"GMT\",\"EST\"): Some were",
"(but not their type). Over all batches, we collected",
"labeled as entities and others were not. In such",
"726 sentences with need of manual decision con-",
"cases, we clarified the annotation guidelines and",
"cerning mention boundaries (1,720 spans when",
"manually corrected all instances in the corpus.",
"counting the predicted and gold ones separately"
] |
{
"arxiv_id": "2310.16225",
"language": "en",
"page_idx": 3
}
|
||
2310.16225-en-4
|
2310.16225-en-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Sentence (part)</th>
<th rowspan="2">Mentions $[ \textbf{Italian}]$ [Roberto Baggio]</th>
<th rowspan="2">Entity Link (Wikipedia URL) Italy\_national\_football\_team Roberto\_Baggio</th>
<th rowspan="2">$NER$ before Phase 3 $0RG$ $PER$</th>
<th rowspan="2">$NER$ after $Phase:$ $MISC$ $PER$</th>
</tr>
<tr>
<th>$a)$</th>
<th>[...] while Italian team mate Roberto Baggio did not play due to injury.</th>
</tr>
<tr>
<td>$b)$</td>
<td>Dick Morris, the Republican politi- cal consultant who reshaped [...]</td>
<td>|Dick Morris $[ \textbf{Republican}]$</td>
<td>Dick\_Morris Republican\_Party\_(United\_States)</td>
<td>$PER$ $0RG$</td>
<td>$PER$ $MISC$</td>
</tr>
<tr>
<td rowspan="3">$c)$</td>
<td>Rosati will meet Serbian President</td>
<td>[Rosati]</td>
<td>Dariusz Rosati</td>
<td>$PER$</td>
<td>$PER$</td>
</tr>
<tr>
<td>Slobodan Milosevic</td>
<td>$|\mathbf{Serbian}|$</td>
<td>Serbia</td>
<td>$L0C$</td>
<td>$MISC$</td>
</tr>
<tr>
<td> </td>
<td>|Slobodan Milosevicl</td>
<td>Slobodan Milošević</td>
<td>$PER$</td>
<td>$PER$</td>
</tr>
</tbody>
</table>
Table 1: Example sentences to illustrate the differences in adjectival affiliation entities. In the original CoNLL-03, affiliations such as"Italian","Republican" and "Serbian" are marked MISC. However, since we derive NER tags from entity links, they are initially marked as ORG ("Italian" here for instance refers to the national football team of Italy) or LOC. Phase 3 of our relabeling reverts these back to MISC usage.
despite their overlaps) that were resolved by manual annotation. In 450 of these sentences, different boundaries of one (or more) mentions were set.
<table>
<tbody>
<tr>
<th> </th>
<th>CoNLI abs</th>
<th>$L-03$ $\%$</th>
<th>$Rei:$ abs</th>
<th>SS $\%$</th>
<th>$\mathbf{C}_{\mathrm{LEANC}}$ abs</th>
<th>ONLL $\%$</th>
</tr>
<tr>
<td>$LOC$</td>
<td>10,645</td>
<td>30.3</td>
<td>10,103</td>
<td>28.9</td>
<td>9,399</td>
<td>26.7</td>
</tr>
<tr>
<td>$ORG$</td>
<td>9,323</td>
<td>26.6</td>
<td>9,922</td>
<td>28.4</td>
<td>10,492</td>
<td>29.8</td>
</tr>
<tr>
<td>$PER$</td>
<td>10,059</td>
<td>28.7</td>
<td>9,983</td>
<td>28.6</td>
<td>9,947</td>
<td>28.2</td>
</tr>
<tr>
<td>$MISC$</td>
<td>5,062</td>
<td>14.4</td>
<td>4,933</td>
<td>14.1</td>
<td>5,419</td>
<td>15.4</td>
</tr>
<tr>
<td>entities 井</td>
<td>35,089</td>
<td> </td>
<td>34,941</td>
<td> </td>
<td>35,257</td>
<td> </td>
</tr>
<tr>
<td>井 $sent.$</td>
<td>20,744</td>
<td> </td>
<td>20,617</td>
<td> </td>
<td>20,617</td>
<td> </td>
</tr>
</tbody>
</table>
2.3 Phase 3: Handling Adjectival Affiliations
Table 2: Statistics of CLEANCONLL, in comparison
to CoNLL-03 and the Reiss version.
After Phase 2 of our relabeling effort, a major inconsistency to the original CoNLL-03 definitions became apparent: Adjectival affiliation entities are consistently labeled MISC in CoNLL-03, whereas our corpus assigns the NER tags of their nominal usage. As Table 1 illustrates, adjectival entities such as "Italian" may be labeled ORG in our corpus if linked to the Wikipedia page of the national football team of Italy.
To make our annotations consistent with original CoNLL-03 definitions, we semi-automatically revert such cases to the MISC label. Using two prediction models trained on OntoNotes (Hovy et al., 2006), we predict adjectives and entities of OntoNotes type "NORP" (affiliations to"nationalities or religious or political groups") to flag entities that should potentially be reverted to MISC. We automatically reverted flagged entities if they were originally labeled MISC in the Reiss version and manually checked all others.
tion across the four classes for CLEANCONLL, the
original CoNLL-03 and the Reiss variant.
We observe that CLEANCONLL has a slightly higher number of ORG and MISC entities than the base versions. Among other reasons, this stems from a more consistent use of ORG labels for sports teams referred to by their geographic name, and more consistent use of MISC for entity types such as named sports leagues and stocks. The appendix provides more information on such cases in Appendix A.6, and discusses qualitative examples in Appendix A.5.
Updated NER labels. Table 3 further examines the extent of label updates introduced by our approach. Compared to CoNLL-03, the labels of 1,832 entity mentions (5.2%) were updated, and $625(1.8\%)$ mentions were newly added$^5.$ In sum, we thus modified 7% of labels, affecting 10.3% of sentences$^{6}$. We specifically note slightly higher label update rates in the test split $( 9. 4\%$ modified labels, affecting 14.2% of sentences) and the dev
$^{5}$This includes both completely new ones as well as
## 3 $\textbf{Evaluation}$
We aim to get an understanding of annotation quality of CLEANCONLL. To this end, we discuss its properties (Sec. 3.1) and perform a manual evaluation of 100 sentences to assess and compare annotation quality with earlier variants (Sec. 3.2-3.4).
## 3.1 Dataset Statistics
Label distribution. Table 2 compares the total
mentions where boundaries were moved.
Refer to Appendix A.8.3 for the number of changed labels before reverting MISC to its original definition in ${\mathrm{Phase~3~(14.9\%)}}.$
count of annotated named entities and the distribu-
|
[
[
0.1502685546875,
0.08489990234375,
0.2476806640625,
0.0885009765625,
0.2476806640625,
0.10272216796875,
0.1502685546875,
0.09979248046875
],
[
0.72119140625,
0.0872802734375,
0.75830078125,
0.0872802734375,
0.75830078125,
0.09857177734375,
0.72119140625,
0.09857177734375
],
[
0.8037109375,
0.08489990234375,
0.8388671875,
0.0872802734375,
0.8369140625,
0.09979248046875,
0.8017578125,
0.097412109375
],
[
0.5517578125,
0.0885009765625,
0.623046875,
0.0885009765625,
0.623046875,
0.09979248046875,
0.5517578125,
0.09979248046875
],
[
0.39208984375,
0.0885009765625,
0.455078125,
0.0885009765625,
0.455078125,
0.09979248046875,
0.39208984375,
0.09979248046875
],
[
0.54833984375,
0.09979248046875,
0.6591796875,
0.09857177734375,
0.6591796875,
0.10986328125,
0.54833984375,
0.11102294921875
],
[
0.72119140625,
0.09979248046875,
0.767578125,
0.09979248046875,
0.767578125,
0.11102294921875,
0.72119140625,
0.11102294921875
],
[
0.8017578125,
0.09979248046875,
0.8369140625,
0.09979248046875,
0.8369140625,
0.11102294921875,
0.8017578125,
0.11102294921875
],
[
0.7197265625,
0.10986328125,
0.77490234375,
0.10986328125,
0.77490234375,
0.12469482421875,
0.7197265625,
0.12469482421875
],
[
0.8017578125,
0.10986328125,
0.85546875,
0.10986328125,
0.85546875,
0.12469482421875,
0.8017578125,
0.12469482421875
],
[
0.1292724609375,
0.12890625,
0.33154296875,
0.12890625,
0.33154296875,
0.142578125,
0.1292724609375,
0.142578125
],
[
0.393798828125,
0.1300048828125,
0.450927734375,
0.1300048828125,
0.450927734375,
0.141357421875,
0.393798828125,
0.141357421875
],
[
0.72119140625,
0.1300048828125,
0.755859375,
0.1300048828125,
0.755859375,
0.142578125,
0.72119140625,
0.142578125
],
[
0.8017578125,
0.1300048828125,
0.8388671875,
0.1300048828125,
0.8388671875,
0.142578125,
0.8017578125,
0.142578125
],
[
0.151123046875,
0.141357421875,
0.35693359375,
0.141357421875,
0.35693359375,
0.155029296875,
0.151123046875,
0.155029296875
],
[
0.5498046875,
0.1312255859375,
0.68994140625,
0.1312255859375,
0.68994140625,
0.141357421875,
0.5498046875,
0.141357421875
],
[
0.393798828125,
0.142578125,
0.5029296875,
0.142578125,
0.5029296875,
0.15380859375,
0.393798828125,
0.15380859375
],
[
0.5498046875,
0.141357421875,
0.63232421875,
0.1436767578125,
0.63232421875,
0.1561279296875,
0.5498046875,
0.15380859375
],
[
0.7197265625,
0.142578125,
0.75,
0.142578125,
0.75,
0.15380859375,
0.7197265625,
0.15380859375
],
[
0.8017578125,
0.142578125,
0.8310546875,
0.142578125,
0.8310546875,
0.155029296875,
0.8017578125,
0.155029296875
],
[
0.15283203125,
0.155029296875,
0.210693359375,
0.155029296875,
0.210693359375,
0.166259765625,
0.15283203125,
0.166259765625
],
[
0.1259765625,
0.169189453125,
0.38037109375,
0.171630859375,
0.38037109375,
0.185302734375,
0.1259765625,
0.182861328125
],
[
0.72119140625,
0.171630859375,
0.75,
0.171630859375,
0.75,
0.18408203125,
0.72119140625,
0.18408203125
],
[
0.8017578125,
0.171630859375,
0.8310546875,
0.171630859375,
0.8310546875,
0.18408203125,
0.8017578125,
0.18408203125
],
[
0.5498046875,
0.1728515625,
0.615234375,
0.1728515625,
0.615234375,
0.18408203125,
0.5498046875,
0.18408203125
],
[
0.39208984375,
0.1728515625,
0.480224609375,
0.1728515625,
0.480224609375,
0.18408203125,
0.39208984375,
0.18408203125
],
[
0.391357421875,
0.18115234375,
0.48193359375,
0.182861328125,
0.48193359375,
0.19775390625,
0.391357421875,
0.1953125
],
[
0.54833984375,
0.182861328125,
0.755859375,
0.182861328125,
0.755859375,
0.19775390625,
0.54833984375,
0.19775390625
],
[
0.8017578125,
0.18408203125,
0.8388671875,
0.18408203125,
0.8388671875,
0.196533203125,
0.8017578125,
0.196533203125
],
[
0.391357421875,
0.20068359375,
0.447509765625,
0.20068359375,
0.447509765625,
0.215576171875,
0.391357421875,
0.215576171875
],
[
0.154541015625,
0.185302734375,
0.349365234375,
0.185302734375,
0.349365234375,
0.1953125,
0.154541015625,
0.1953125
],
[
0.1292724609375,
0.201904296875,
0.372802734375,
0.201904296875,
0.372802734375,
0.215576171875,
0.1292724609375,
0.215576171875
],
[
0.5498046875,
0.203125,
0.62646484375,
0.203125,
0.62646484375,
0.21435546875,
0.5498046875,
0.21435546875
],
[
0.72119140625,
0.203125,
0.748046875,
0.203125,
0.748046875,
0.21435546875,
0.72119140625,
0.21435546875
],
[
0.8017578125,
0.203125,
0.82958984375,
0.203125,
0.82958984375,
0.21435546875,
0.8017578125,
0.21435546875
],
[
0.151123046875,
0.213134765625,
0.276123046875,
0.21435546875,
0.276123046875,
0.2257080078125,
0.151123046875,
0.2244873046875
],
[
0.393798828125,
0.21435546875,
0.455078125,
0.21435546875,
0.455078125,
0.2257080078125,
0.393798828125,
0.2257080078125
],
[
0.54833984375,
0.21435546875,
0.58837890625,
0.21435546875,
0.58837890625,
0.2257080078125,
0.54833984375,
0.2257080078125
],
[
0.7197265625,
0.21435546875,
0.75244140625,
0.21435546875,
0.75244140625,
0.226806640625,
0.7197265625,
0.226806640625
],
[
0.8017578125,
0.21435546875,
0.8388671875,
0.21435546875,
0.8388671875,
0.2257080078125,
0.8017578125,
0.2257080078125
],
[
0.39208984375,
0.226806640625,
0.52490234375,
0.226806640625,
0.52490234375,
0.2369384765625,
0.39208984375,
0.2369384765625
],
[
0.5498046875,
0.226806640625,
0.6513671875,
0.226806640625,
0.6513671875,
0.2381591796875,
0.5498046875,
0.2381591796875
],
[
0.72119140625,
0.226806640625,
0.75,
0.226806640625,
0.75,
0.2381591796875,
0.72119140625,
0.2381591796875
],
[
0.8017578125,
0.226806640625,
0.82958984375,
0.226806640625,
0.82958984375,
0.2381591796875,
0.8017578125,
0.2381591796875
],
[
0.1175537109375,
0.254150390625,
0.880859375,
0.254150390625,
0.880859375,
0.267333984375,
0.1175537109375,
0.267333984375
],
[
0.1175537109375,
0.267333984375,
0.8818359375,
0.267333984375,
0.8818359375,
0.28076171875,
0.1175537109375,
0.28076171875
],
[
0.1175537109375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.1175537109375,
0.295654296875
],
[
0.1158447265625,
0.29638671875,
0.63818359375,
0.29638671875,
0.63818359375,
0.31005859375,
0.1158447265625,
0.31005859375
],
[
0.1158447265625,
0.336181640625,
0.489501953125,
0.3349609375,
0.489501953125,
0.349853515625,
0.1158447265625,
0.350830078125
],
[
0.59423828125,
0.33740234375,
0.67236328125,
0.33740234375,
0.67236328125,
0.3515625,
0.59423828125,
0.3515625
],
[
0.705078125,
0.33837890625,
0.744140625,
0.33837890625,
0.744140625,
0.350830078125,
0.705078125,
0.350830078125
],
[
0.771484375,
0.33740234375,
0.8759765625,
0.33740234375,
0.8759765625,
0.3515625,
0.771484375,
0.3515625
],
[
0.61279296875,
0.350830078125,
0.6396484375,
0.350830078125,
0.6396484375,
0.363525390625,
0.61279296875,
0.363525390625
],
[
0.6572265625,
0.350830078125,
0.67919921875,
0.350830078125,
0.67919921875,
0.363525390625,
0.6572265625,
0.363525390625
],
[
0.70361328125,
0.350830078125,
0.72900390625,
0.350830078125,
0.72900390625,
0.363525390625,
0.70361328125,
0.363525390625
],
[
0.748046875,
0.350830078125,
0.771484375,
0.350830078125,
0.771484375,
0.363525390625,
0.748046875,
0.363525390625
],
[
0.79248046875,
0.350830078125,
0.8193359375,
0.350830078125,
0.8193359375,
0.364013671875,
0.79248046875,
0.364013671875
],
[
0.85986328125,
0.3515625,
0.875,
0.3515625,
0.875,
0.363525390625,
0.85986328125,
0.363525390625
],
[
0.67822265625,
0.36767578125,
0.7314453125,
0.36767578125,
0.7314453125,
0.3818359375,
0.67822265625,
0.3818359375
],
[
0.1175537109375,
0.352783203125,
0.48779296875,
0.352783203125,
0.48779296875,
0.366455078125,
0.1175537109375,
0.366455078125
],
[
0.77490234375,
0.36767578125,
0.8212890625,
0.36767578125,
0.8212890625,
0.3818359375,
0.77490234375,
0.3818359375
],
[
0.8427734375,
0.36767578125,
0.87890625,
0.36767578125,
0.87890625,
0.383056640625,
0.8427734375,
0.383056640625
],
[
0.64208984375,
0.3681640625,
0.6767578125,
0.3681640625,
0.6767578125,
0.38134765625,
0.64208984375,
0.38134765625
],
[
0.5224609375,
0.3681640625,
0.55908203125,
0.3681640625,
0.55908203125,
0.38134765625,
0.5224609375,
0.38134765625
],
[
0.59375,
0.369384765625,
0.6396484375,
0.369384765625,
0.6396484375,
0.38134765625,
0.59375,
0.38134765625
],
[
0.73291015625,
0.3681640625,
0.767578125,
0.3681640625,
0.767578125,
0.38134765625,
0.73291015625,
0.38134765625
],
[
0.59423828125,
0.380126953125,
0.67822265625,
0.380126953125,
0.67822265625,
0.394287109375,
0.59423828125,
0.394287109375
],
[
0.1175537109375,
0.37060546875,
0.46337890625,
0.37060546875,
0.46337890625,
0.38134765625,
0.1175537109375,
0.38134765625
],
[
0.52294921875,
0.38134765625,
0.56103515625,
0.38134765625,
0.56103515625,
0.393798828125,
0.52294921875,
0.393798828125
],
[
0.68408203125,
0.380126953125,
0.73291015625,
0.380126953125,
0.73291015625,
0.394287109375,
0.68408203125,
0.394287109375
],
[
0.7314453125,
0.38134765625,
0.767578125,
0.38134765625,
0.767578125,
0.393798828125,
0.7314453125,
0.393798828125
],
[
0.77099609375,
0.3818359375,
0.8193359375,
0.3818359375,
0.8193359375,
0.393798828125,
0.77099609375,
0.393798828125
],
[
0.8427734375,
0.38134765625,
0.87744140625,
0.38134765625,
0.87744140625,
0.393798828125,
0.8427734375,
0.393798828125
],
[
0.68701171875,
0.391357421875,
0.7314453125,
0.391357421875,
0.7314453125,
0.40625,
0.68701171875,
0.40625
],
[
0.73046875,
0.391357421875,
0.767578125,
0.393798828125,
0.765625,
0.40625,
0.72900390625,
0.40380859375
],
[
0.77490234375,
0.391357421875,
0.8193359375,
0.391357421875,
0.8193359375,
0.40625,
0.77490234375,
0.40625
],
[
0.8427734375,
0.392578125,
0.87744140625,
0.392578125,
0.87744140625,
0.405029296875,
0.8427734375,
0.405029296875
],
[
0.59375,
0.393798828125,
0.67333984375,
0.393798828125,
0.67333984375,
0.405029296875,
0.59375,
0.405029296875
],
[
0.1158447265625,
0.394287109375,
0.483642578125,
0.3955078125,
0.483642578125,
0.410400390625,
0.1158447265625,
0.4091796875
],
[
0.52490234375,
0.392578125,
0.5576171875,
0.392578125,
0.5576171875,
0.405029296875,
0.52490234375,
0.405029296875
],
[
0.52294921875,
0.40625,
0.56494140625,
0.40625,
0.56494140625,
0.41748046875,
0.52294921875,
0.41748046875
],
[
0.59619140625,
0.405029296875,
0.67919921875,
0.405029296875,
0.67919921875,
0.419189453125,
0.59619140625,
0.419189453125
],
[
0.6884765625,
0.40625,
0.73486328125,
0.40625,
0.73486328125,
0.41748046875,
0.6884765625,
0.41748046875
],
[
0.7314453125,
0.40625,
0.767578125,
0.40625,
0.767578125,
0.41748046875,
0.7314453125,
0.41748046875
],
[
0.77734375,
0.40625,
0.8193359375,
0.40625,
0.8193359375,
0.41748046875,
0.77734375,
0.41748046875
],
[
0.8447265625,
0.40625,
0.87890625,
0.40625,
0.87890625,
0.418701171875,
0.8447265625,
0.418701171875
],
[
0.1175537109375,
0.41748046875,
0.489501953125,
0.41748046875,
0.489501953125,
0.43115234375,
0.1175537109375,
0.43115234375
],
[
0.6767578125,
0.4228515625,
0.72705078125,
0.42041015625,
0.72900390625,
0.435791015625,
0.67822265625,
0.438232421875
],
[
0.767578125,
0.4228515625,
0.8193359375,
0.42041015625,
0.8212890625,
0.435791015625,
0.76904296875,
0.438232421875
],
[
0.1175537109375,
0.432861328125,
0.489501953125,
0.432861328125,
0.489501953125,
0.446533203125,
0.1175537109375,
0.446533203125
],
[
0.5205078125,
0.42333984375,
0.6396484375,
0.421630859375,
0.6396484375,
0.435791015625,
0.5205078125,
0.438232421875
],
[
0.52294921875,
0.435791015625,
0.568359375,
0.438232421875,
0.56689453125,
0.449462890625,
0.5224609375,
0.44775390625
],
[
0.59033203125,
0.43701171875,
0.63818359375,
0.43701171875,
0.63818359375,
0.448974609375,
0.59033203125,
0.448974609375
],
[
0.67919921875,
0.43701171875,
0.72705078125,
0.43701171875,
0.72705078125,
0.448974609375,
0.67919921875,
0.448974609375
],
[
0.76904296875,
0.435791015625,
0.8212890625,
0.435791015625,
0.8212890625,
0.45068359375,
0.76904296875,
0.45068359375
],
[
0.1175537109375,
0.449462890625,
0.48779296875,
0.449462890625,
0.48779296875,
0.463134765625,
0.1175537109375,
0.463134765625
],
[
0.1158447265625,
0.465576171875,
0.48779296875,
0.465576171875,
0.48779296875,
0.478515625,
0.1158447265625,
0.478515625
],
[
0.50732421875,
0.4619140625,
0.87744140625,
0.463134765625,
0.87744140625,
0.47802734375,
0.50732421875,
0.476806640625
],
[
0.1158447265625,
0.482177734375,
0.48779296875,
0.482177734375,
0.48779296875,
0.495361328125,
0.1158447265625,
0.495361328125
],
[
0.51123046875,
0.478515625,
0.75732421875,
0.478515625,
0.75732421875,
0.4921875,
0.51123046875,
0.4921875
],
[
0.11419677734375,
0.49755859375,
0.486083984375,
0.49755859375,
0.486083984375,
0.51123046875,
0.11419677734375,
0.51123046875
],
[
0.1158447265625,
0.51318359375,
0.490234375,
0.51318359375,
0.490234375,
0.52783203125,
0.1158447265625,
0.52783203125
],
[
0.51123046875,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51123046875,
0.525390625
],
[
0.1158447265625,
0.52978515625,
0.489501953125,
0.52978515625,
0.489501953125,
0.54443359375,
0.1158447265625,
0.54443359375
],
[
0.509765625,
0.52783203125,
0.8193359375,
0.52783203125,
0.8193359375,
0.54150390625,
0.509765625,
0.54150390625
],
[
0.1158447265625,
0.54638671875,
0.279541015625,
0.54638671875,
0.279541015625,
0.56005859375,
0.1158447265625,
0.56005859375
],
[
0.52880859375,
0.5419921875,
0.880859375,
0.54345703125,
0.880859375,
0.55810546875,
0.52880859375,
0.55712890625
],
[
0.1368408203125,
0.5625,
0.490234375,
0.5625,
0.490234375,
0.576171875,
0.1368408203125,
0.576171875
],
[
0.51123046875,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51123046875,
0.57373046875
],
[
0.1158447265625,
0.57763671875,
0.48779296875,
0.57763671875,
0.48779296875,
0.59130859375,
0.1158447265625,
0.59130859375
],
[
0.509765625,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.509765625,
0.59033203125
],
[
0.1158447265625,
0.59326171875,
0.48779296875,
0.59326171875,
0.48779296875,
0.60693359375,
0.1158447265625,
0.60693359375
],
[
0.509765625,
0.5927734375,
0.8818359375,
0.5927734375,
0.8818359375,
0.60546875,
0.509765625,
0.60546875
],
[
0.1158447265625,
0.609375,
0.486083984375,
0.609375,
0.486083984375,
0.62255859375,
0.1158447265625,
0.62255859375
],
[
0.509765625,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.509765625,
0.62255859375
],
[
0.1175537109375,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.509765625,
0.62353515625,
0.8818359375,
0.62353515625,
0.8818359375,
0.63818359375,
0.509765625,
0.63818359375
],
[
0.1158447265625,
0.64111328125,
0.490234375,
0.64013671875,
0.490234375,
0.65478515625,
0.1158447265625,
0.65625
],
[
0.509765625,
0.64111328125,
0.8818359375,
0.64111328125,
0.8818359375,
0.65478515625,
0.509765625,
0.65478515625
],
[
0.1158447265625,
0.6591796875,
0.489501953125,
0.6591796875,
0.489501953125,
0.6728515625,
0.1158447265625,
0.6728515625
],
[
0.51123046875,
0.65673828125,
0.88330078125,
0.65673828125,
0.88330078125,
0.67041015625,
0.51123046875,
0.67041015625
],
[
0.1175537109375,
0.6748046875,
0.48779296875,
0.6748046875,
0.48779296875,
0.6884765625,
0.1175537109375,
0.6884765625
],
[
0.51318359375,
0.67333984375,
0.8818359375,
0.67333984375,
0.8818359375,
0.68701171875,
0.51318359375,
0.68701171875
],
[
0.1158447265625,
0.69140625,
0.48779296875,
0.69140625,
0.48779296875,
0.705078125,
0.1158447265625,
0.705078125
],
[
0.5146484375,
0.6904296875,
0.62109375,
0.6904296875,
0.62109375,
0.70166015625,
0.5146484375,
0.70166015625
],
[
0.1175537109375,
0.70703125,
0.48779296875,
0.70703125,
0.48779296875,
0.72021484375,
0.1175537109375,
0.72021484375
],
[
0.51123046875,
0.705078125,
0.8818359375,
0.705078125,
0.8818359375,
0.71875,
0.51123046875,
0.71875
],
[
0.1158447265625,
0.7236328125,
0.328369140625,
0.7236328125,
0.328369140625,
0.73681640625,
0.1158447265625,
0.73681640625
],
[
0.509765625,
0.7197265625,
0.88330078125,
0.72021484375,
0.88330078125,
0.7353515625,
0.509765625,
0.73388671875
],
[
0.51123046875,
0.73681640625,
0.88330078125,
0.73681640625,
0.88330078125,
0.75048828125,
0.51123046875,
0.75048828125
],
[
0.11419677734375,
0.7490234375,
0.2435302734375,
0.75048828125,
0.2435302734375,
0.765625,
0.11419677734375,
0.76318359375
],
[
0.509765625,
0.751953125,
0.880859375,
0.751953125,
0.880859375,
0.765625,
0.509765625,
0.765625
],
[
0.51123046875,
0.76708984375,
0.8818359375,
0.7685546875,
0.8818359375,
0.783203125,
0.51123046875,
0.7822265625
],
[
0.1158447265625,
0.77490234375,
0.489501953125,
0.775390625,
0.489501953125,
0.79052734375,
0.1158447265625,
0.7890625
],
[
0.51318359375,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.51318359375,
0.798828125
],
[
0.1158447265625,
0.7919921875,
0.489501953125,
0.7919921875,
0.489501953125,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.51123046875,
0.8017578125,
0.880859375,
0.8017578125,
0.880859375,
0.8154296875,
0.51123046875,
0.8154296875
],
[
0.1175537109375,
0.80859375,
0.490234375,
0.80859375,
0.490234375,
0.822265625,
0.1175537109375,
0.822265625
],
[
0.51318359375,
0.818359375,
0.8818359375,
0.818359375,
0.8818359375,
0.83203125,
0.51318359375,
0.83203125
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.83447265625,
0.8818359375,
0.83447265625,
0.8818359375,
0.84716796875,
0.51123046875,
0.84716796875
],
[
0.1175537109375,
0.84033203125,
0.47607421875,
0.84033203125,
0.47607421875,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.5322265625,
0.85546875,
0.8505859375,
0.85693359375,
0.8505859375,
0.87060546875,
0.5322265625,
0.869140625
],
[
0.1158447265625,
0.86767578125,
0.2939453125,
0.86767578125,
0.2939453125,
0.88134765625,
0.1158447265625,
0.88134765625
],
[
0.51318359375,
0.87060546875,
0.75830078125,
0.87060546875,
0.75830078125,
0.88134765625,
0.51318359375,
0.88134765625
],
[
0.5322265625,
0.88134765625,
0.861328125,
0.88134765625,
0.861328125,
0.8955078125,
0.5322265625,
0.8955078125
],
[
0.1175537109375,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.8935546875,
0.85986328125,
0.8935546875,
0.85986328125,
0.90673828125,
0.509765625,
0.90673828125
],
[
0.11419677734375,
0.90478515625,
0.491943359375,
0.90478515625,
0.491943359375,
0.91845703125,
0.11419677734375,
0.91845703125
],
[
0.51123046875,
0.90625,
0.6171875,
0.90625,
0.6171875,
0.91748046875,
0.51123046875,
0.91748046875
]
] |
[
"Sentence (part)",
"NER",
"NER",
"Entity Link",
"Mentions",
"(Wikipedia URL)",
"before",
"after",
"Phase 3",
"Phase 3",
"a [...J while Italian team mate",
"[Italian]",
"ORG",
"MISC",
"Roberto Baggio did not play due",
"Italy_national_football_team",
"[Roberto Baggio]",
"Roberto_Baggio",
"PER",
"PER",
"to injury.",
"b)Dick Morris, the Republican politi-",
"PER",
"PER",
"Dick_Morris",
"[Dick Morris]",
"[Republican]",
"Republican_Party_(United_ States) ORG",
"MISC",
"[Rosati]",
"cal consultantwho reshaped [...]",
"c) Rosati will meet Serbian President",
"Dariusz_Rosati",
"PER",
"PER",
"Slobodan Milosevic",
"[Serbian]",
"Serbia",
"LOC",
"MISC",
"[SlobodanMilosevic]",
"Slobodan_Milosevie",
"PER",
"PER",
"Table 1: Example sentences to illustrate the differences in adjectival affiliation entities. In the original CoNLL-03.",
"affiliations such as\"Italian\",\"Republican\" and \"Serbian\"are marked MISC. However, since we derive NER tags",
"from entity links, they are initially marked as ORG (\"Italian\" here for instance refers to the national football team",
"of Italy) or LOC. Phase 3 of our relabeling reverts these back to MISC usage.",
"despite their overlaps) that were resolved by man-",
"CoNLL-03",
"Reiss",
"CLEANCONLL",
"abs",
"%",
"abs",
"% ",
"abs",
"%",
"10,103",
"ual annotation. In 450 of these sentences, different",
"9,399",
"26.7",
"30.3",
"LOC",
"10,645",
"28.9",
"9.32326.6",
"boundaries of one(ormore)mentionswereset",
"ORG",
"9,922",
"28.4",
"10,492",
"29.8",
"9,983",
"28.6",
"9,947",
"28.2",
"10.05928.7",
"2.3 Phase 3: Handling Adjectival Affiliations",
"PER",
"MISC",
"5,06214.4",
"4,933",
"14.1",
"5,419",
"15.4",
"After Phase 2 of our relabeling effort, a major in-",
"34,941",
"35,257",
"consistency to the original CoNLL-03 definitions",
"#entities35.089",
"# sent.",
"20,744",
"20,617",
"20,617",
"became apparent: Adjectival affiliation entities are",
"consistently labeled MISC in CoNLL-03, whereas",
" Table 2: Statistics of CLEANCoNLL, in comparison",
"our corpus assigns the NER tags of their nominal",
"to CoNLL-03 and the Reiss version.",
"usage. As Table 1 illustrates, adjectival entities",
"such as \"Italian\" may be labeled ORG in our cor-",
"tion across the four classes for CLEANCoNLL, the",
"pus if linked to the Wikipedia page of the national",
"original CoNLL-03 and the Reiss variant.",
"football team of Italy.",
"We observe that CLEANCoNLL has a slightly",
"To make our annotations consistent with origi-",
"higher number of ORG and MISC entities than",
"nal CoNLL-03 definitions, we semi-automatically",
"the base versions. Among other reasons, this stems",
"revert such cases to the MISC label. Using two",
"from a more consistent use of ORG labels for sports",
"prediction models trained on OntoNotes (Hovy",
"teams referred to by their geographic name, and",
"et al., 2006), we predict adjectives and entities of",
"more consistent use of MISC for entity types such",
"OntoNotes type\"NORP\"(affiliations to\"nationali-",
" as named sports leagues and stocks. The appendix",
"ties or religious or political groups\") to flag entities",
"provides more information on such cases in Ap-",
"that should potentially be reverted to MISC. We",
"pendix A.6, and discusses qualitative examples in",
"automatically reverted flagged entities if they were",
"Appendix A.5.",
"originally labeled MISC in the Reiss version and",
"Updated NER labels. Table 3 further examines",
"manually checked all others.",
"the extent of label updates introduced by our ap-",
"proach. Compared to CoNLL-03, the labels of",
"3Evaluation",
"1,832 entity mentions (5.2%) were updated, and",
"625 (1.8%) mentions were newly added5. In sum,",
"We aim to get an understanding of annotation qual-",
"we thus modified 7% of labels, affecting 10.3% of",
"ity of CLEANCoNLL. To this end, we discuss its",
"sentences. We specifically note slightly higher",
"properties (Sec. 3.1) and perform a manual evalua-",
"label update rates in the test split (9.4% modified",
"tion of 100 sentences to assess and compare anno-",
"labels, affecting 14.2% of sentences) and the dev",
"tation quality with earlier variants (Sec. 3.2-3.4).",
"5This includes both completely new ones as well as",
"3.1 Dataset Statistics",
"mentions where boundaries were moved",
"6Refer to Appendix A.8.3 for the number of changed",
"Label distribution. Table 2 compares the total",
"labels before reverting MISC to its original definition in",
" count of annotated named entities and the distribu-",
"Phase 3 (14.9%)."
] |
{
"arxiv_id": "2310.16225",
"language": "en",
"page_idx": 4
}
|
||
2401.11725-en-0
|
2401.11725-en-0.jpg
|
# $\textbf{Speak It Out: Solving Symbol- Related Problems with}$ $\textbf{Symbol- to- Language Conversion for Language Models}$
Yile Wang$^1$ $\textbf{Sijie Cheng}^{1, 2, 3}$ $\textbf{Zixin Sun}^2$ $\textbf{Peng Li}^{\otimes 1, 4}$ $\textbf{Yang Liu}^{\otimes 1, 2, 4}$ $^{1}$Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China $^{2}$Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China
$^{3}01.$AI$^{4}$Shanghai Artificial Intelligence Laboratory, Shanghai, China
$\{$wangyile,lipeng$\}@$air.tsinghua.edu.cn, liuyang2011@tsinghua.edu.cn
## $\mathbf{A}\mathbf{B}\mathbf{S}\mathbf{T}\mathbf{R}\mathbf{A}\mathbf{C}\mathbf{T}$
Symbols (or more broadly, non-natural language textual representations) such as numerical sequences, molecular formulas, and table delimiters widely exist, playing important roles in various tasks such as abstract reasoning, chemical property prediction, and table question answering. Despite the impressive natural language comprehension capabilities of large language models (LLMs), their reasoning abilities for symbols remain inadequate, which could attributed to the difference between symbol representations and general natural languages. We propose symbol-to-language (S2L), a tuning-free method that enables large language models to solve symbol-related problems with information expressed in natural $language.$ Specifically, S2L first converts the symbols involved to language-based representations, which can be implemented by prompting LLMs or leveraging external tools, then these language-based representations are integrated into the original problem via direct substitution or concatenation, serving as useful input information for LLMs. We evaluate the S2L method using both API-based (GPT-4, ČhatGPT) and open-source (OpenChat) models over eight symbol-related tasks, ranging from symbol-only abstract reasoning to sentiment analysis in social media. Experimental results show that S2L consistently leads to superior performance. For example, by employing S2L for GPT- 4, there can be average significant improvements of $\hat{+}21.9\tilde{\%}$ and $+9.5\%$ for subtasks in 1D-ARC and DycK language, respectively. Codes and data are available at https: //github.com/THUNLP-MT/symbol2language.
<FigureHere>
Figure 1: Top: This illustration presents a symbol-related problem (move 1 pixel forward) from the 1D-ARC benchmark, comparing the responses of large language models using both conventional symbol-based and our language-based representations with symbol-to-language (S2L) conversion. Bottom: The S2L conversion has broad applicability across various scenarios involving diverse types of symbols.
|
[
[
0.178955078125,
0.125,
0.81884765625,
0.125,
0.81884765625,
0.1427001953125,
0.178955078125,
0.1427001953125
],
[
0.1707763671875,
0.149658203125,
0.82861328125,
0.1483154296875,
0.82861328125,
0.1673583984375,
0.1707763671875,
0.1685791015625
],
[
0.2296142578125,
0.2310791015625,
0.76708984375,
0.2291259765625,
0.76708984375,
0.25,
0.2296142578125,
0.251953125
],
[
0.2442626953125,
0.25,
0.75634765625,
0.25,
0.75634765625,
0.264404296875,
0.2442626953125,
0.264404296875
],
[
0.0228729248046875,
0.261474609375,
0.05963134765625,
0.261474609375,
0.05963134765625,
0.4072265625,
0.0228729248046875,
0.4072265625
],
[
0.223876953125,
0.261962890625,
0.77392578125,
0.26318359375,
0.77392578125,
0.27783203125,
0.223876953125,
0.276611328125
],
[
0.2646484375,
0.279052734375,
0.318603515625,
0.279052734375,
0.318603515625,
0.290283203125,
0.2646484375,
0.290283203125
],
[
0.313720703125,
0.276611328125,
0.734375,
0.27783203125,
0.734375,
0.292236328125,
0.313720703125,
0.291748046875
],
[
0.22216796875,
0.292236328125,
0.779296875,
0.292236328125,
0.779296875,
0.30615234375,
0.22216796875,
0.30615234375
],
[
0.451904296875,
0.356689453125,
0.5439453125,
0.356689453125,
0.5439453125,
0.36865234375,
0.451904296875,
0.36865234375
],
[
0.205078125,
0.37890625,
0.7958984375,
0.37890625,
0.7958984375,
0.392578125,
0.205078125,
0.392578125
],
[
0.2034912109375,
0.392578125,
0.7958984375,
0.392578125,
0.7958984375,
0.406005859375,
0.2034912109375,
0.406005859375
],
[
0.0196075439453125,
0.404052734375,
0.05963134765625,
0.40283203125,
0.062103271484375,
0.4892578125,
0.0228729248046875,
0.490478515625
],
[
0.2034912109375,
0.406005859375,
0.79736328125,
0.406005859375,
0.79736328125,
0.419189453125,
0.2034912109375,
0.419189453125
],
[
0.2034912109375,
0.42041015625,
0.7958984375,
0.42041015625,
0.7958984375,
0.434326171875,
0.2034912109375,
0.434326171875
],
[
0.2034912109375,
0.434326171875,
0.7939453125,
0.434326171875,
0.7939453125,
0.447509765625,
0.2034912109375,
0.447509765625
],
[
0.2017822265625,
0.447509765625,
0.7958984375,
0.447509765625,
0.7958984375,
0.4609375,
0.2017822265625,
0.4609375
],
[
0.2017822265625,
0.4609375,
0.7939453125,
0.4609375,
0.7939453125,
0.474853515625,
0.2017822265625,
0.474853515625
],
[
0.2034912109375,
0.474853515625,
0.7958984375,
0.474853515625,
0.7958984375,
0.488037109375,
0.2034912109375,
0.488037109375
],
[
0.0228729248046875,
0.4873046875,
0.05963134765625,
0.4873046875,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.2034912109375,
0.488037109375,
0.7958984375,
0.488037109375,
0.7958984375,
0.50244140625,
0.2034912109375,
0.50244140625
],
[
0.2034912109375,
0.50244140625,
0.7939453125,
0.50244140625,
0.7939453125,
0.5166015625,
0.2034912109375,
0.5166015625
],
[
0.2034912109375,
0.5166015625,
0.7958984375,
0.5166015625,
0.7958984375,
0.52978515625,
0.2034912109375,
0.52978515625
],
[
0.2034912109375,
0.52978515625,
0.79736328125,
0.52978515625,
0.79736328125,
0.54443359375,
0.2034912109375,
0.54443359375
],
[
0.2034912109375,
0.54345703125,
0.7958984375,
0.54345703125,
0.7958984375,
0.556640625,
0.2034912109375,
0.556640625
],
[
0.2034912109375,
0.55810546875,
0.7958984375,
0.55810546875,
0.7958984375,
0.5712890625,
0.2034912109375,
0.5712890625
],
[
0.2034912109375,
0.5712890625,
0.7939453125,
0.5712890625,
0.7939453125,
0.58544921875,
0.2034912109375,
0.58544921875
],
[
0.2034912109375,
0.58544921875,
0.79736328125,
0.58544921875,
0.79736328125,
0.5986328125,
0.2034912109375,
0.5986328125
],
[
0.2017822265625,
0.5986328125,
0.7958984375,
0.59765625,
0.7958984375,
0.6123046875,
0.2017822265625,
0.61279296875
],
[
0.2017822265625,
0.611328125,
0.79736328125,
0.6123046875,
0.79736328125,
0.62548828125,
0.2017822265625,
0.625
],
[
0.2034912109375,
0.62548828125,
0.53515625,
0.626953125,
0.53515625,
0.6416015625,
0.2034912109375,
0.64013671875
],
[
0.2320556640625,
0.65478515625,
0.441162109375,
0.65478515625,
0.441162109375,
0.66845703125,
0.2320556640625,
0.66845703125
],
[
0.56201171875,
0.65478515625,
0.8310546875,
0.65478515625,
0.8310546875,
0.66943359375,
0.56201171875,
0.66943359375
],
[
0.340576171875,
0.67822265625,
0.431396484375,
0.67822265625,
0.431396484375,
0.689453125,
0.340576171875,
0.689453125
],
[
0.4453125,
0.67529296875,
0.51708984375,
0.67529296875,
0.51708984375,
0.68994140625,
0.4453125,
0.68994140625
],
[
0.55615234375,
0.67822265625,
0.7099609375,
0.67822265625,
0.7099609375,
0.69140625,
0.55615234375,
0.69140625
],
[
0.7060546875,
0.67822265625,
0.8388671875,
0.67822265625,
0.8388671875,
0.69140625,
0.7060546875,
0.69140625
],
[
0.44677734375,
0.69970703125,
0.515625,
0.69970703125,
0.515625,
0.7138671875,
0.44677734375,
0.7138671875
],
[
0.2099609375,
0.681640625,
0.221435546875,
0.681640625,
0.221435546875,
0.68603515625,
0.2099609375,
0.68603515625
],
[
0.279296875,
0.7021484375,
0.325927734375,
0.7021484375,
0.325927734375,
0.7119140625,
0.279296875,
0.7119140625
],
[
0.340576171875,
0.70068359375,
0.431396484375,
0.70068359375,
0.431396484375,
0.7119140625,
0.340576171875,
0.7119140625
],
[
0.5595703125,
0.7021484375,
0.67333984375,
0.7021484375,
0.67333984375,
0.7119140625,
0.5595703125,
0.7119140625
],
[
0.69677734375,
0.7021484375,
0.83837890625,
0.7021484375,
0.83837890625,
0.7119140625,
0.69677734375,
0.7119140625
],
[
0.1854248046875,
0.70263671875,
0.2320556640625,
0.70263671875,
0.2320556640625,
0.71044921875,
0.1854248046875,
0.71044921875
],
[
0.1707763671875,
0.705078125,
0.1846923828125,
0.705078125,
0.1846923828125,
0.70849609375,
0.1707763671875,
0.70849609375
],
[
0.318603515625,
0.7236328125,
0.414306640625,
0.7236328125,
0.414306640625,
0.73486328125,
0.318603515625,
0.73486328125
],
[
0.41259765625,
0.72265625,
0.50244140625,
0.7236328125,
0.50244140625,
0.73486328125,
0.41259765625,
0.734375
],
[
0.55615234375,
0.7236328125,
0.6748046875,
0.7236328125,
0.6748046875,
0.7373046875,
0.55615234375,
0.7373046875
],
[
0.2296142578125,
0.724609375,
0.279296875,
0.724609375,
0.279296875,
0.736328125,
0.2296142578125,
0.736328125
],
[
0.191162109375,
0.724609375,
0.2303466796875,
0.724609375,
0.2303466796875,
0.73486328125,
0.191162109375,
0.73486328125
],
[
0.74658203125,
0.72607421875,
0.78515625,
0.72607421875,
0.78515625,
0.736328125,
0.74658203125,
0.736328125
],
[
0.7900390625,
0.7265625,
0.82177734375,
0.7265625,
0.82177734375,
0.734375,
0.7900390625,
0.734375
],
[
0.2646484375,
0.76025390625,
0.51708984375,
0.76025390625,
0.51708984375,
0.7734375,
0.2646484375,
0.7734375
],
[
0.7255859375,
0.72802734375,
0.74609375,
0.72802734375,
0.74609375,
0.73486328125,
0.7255859375,
0.73486328125
],
[
0.1519775390625,
0.76123046875,
0.256591796875,
0.76123046875,
0.256591796875,
0.7734375,
0.1519775390625,
0.7734375
],
[
0.5595703125,
0.76318359375,
0.6845703125,
0.7646484375,
0.6845703125,
0.77783203125,
0.5595703125,
0.7763671875
],
[
0.69677734375,
0.771484375,
0.7607421875,
0.771484375,
0.7607421875,
0.7822265625,
0.69677734375,
0.7822265625
],
[
0.77392578125,
0.771484375,
0.83837890625,
0.771484375,
0.83837890625,
0.78271484375,
0.77392578125,
0.78271484375
],
[
0.423095703125,
0.7734375,
0.488525390625,
0.7734375,
0.488525390625,
0.78857421875,
0.423095703125,
0.78857421875
],
[
0.1846923828125,
0.77587890625,
0.21728515625,
0.77587890625,
0.21728515625,
0.78857421875,
0.1846923828125,
0.78857421875
],
[
0.306396484375,
0.7763671875,
0.348876953125,
0.7763671875,
0.348876953125,
0.78857421875,
0.306396484375,
0.78857421875
],
[
0.55810546875,
0.77587890625,
0.67724609375,
0.7744140625,
0.67724609375,
0.7861328125,
0.55810546875,
0.787109375
],
[
0.1732177734375,
0.79052734375,
0.343017578125,
0.79248046875,
0.343017578125,
0.806640625,
0.1732177734375,
0.8046875
],
[
0.42236328125,
0.79345703125,
0.488525390625,
0.79345703125,
0.488525390625,
0.8046875,
0.42236328125,
0.8046875
],
[
0.57275390625,
0.7998046875,
0.7060546875,
0.7998046875,
0.7060546875,
0.81005859375,
0.57275390625,
0.81005859375
],
[
0.73193359375,
0.7998046875,
0.82421875,
0.7998046875,
0.82421875,
0.81005859375,
0.73193359375,
0.81005859375
],
[
0.16259765625,
0.80810546875,
0.36767578125,
0.80810546875,
0.36767578125,
0.8212890625,
0.16259765625,
0.8212890625
],
[
0.40283203125,
0.8095703125,
0.50341796875,
0.8095703125,
0.50341796875,
0.82080078125,
0.40283203125,
0.82080078125
],
[
0.580078125,
0.8095703125,
0.70166015625,
0.8095703125,
0.70166015625,
0.82275390625,
0.580078125,
0.82275390625
],
[
0.740234375,
0.81005859375,
0.81396484375,
0.81005859375,
0.81396484375,
0.8212890625,
0.740234375,
0.8212890625
],
[
0.146240234375,
0.8388671875,
0.85302734375,
0.8388671875,
0.85302734375,
0.85302734375,
0.146240234375,
0.85302734375
],
[
0.143798828125,
0.85302734375,
0.8544921875,
0.85302734375,
0.8544921875,
0.8662109375,
0.143798828125,
0.8662109375
],
[
0.1446533203125,
0.8662109375,
0.8544921875,
0.8662109375,
0.8544921875,
0.8798828125,
0.1446533203125,
0.8798828125
],
[
0.1422119140625,
0.87890625,
0.6650390625,
0.8798828125,
0.6650390625,
0.89453125,
0.1422119140625,
0.89306640625
]
] |
[
"Speak It Out: Solving Symbol-Related Problems with",
"Symbol-to-Language Conversion for Language Models",
"Yile Wangl Sijie Cheng,2,3 Zixin Sun? Peng Li-1,4 Yang Liu1,2,4",
"'Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China",
"",
"2Dept.of Comp. Sci. &Tech., Institute for AI,Tsinghua University, Beijing, China",
"301.A1",
"4Shanghai Artificial Intelligence Laboratory, Shanghai, China",
"(wangyile,lipeng)@air.tsinghua.edu.cn,liuyang2011@tsinghua.edu.cn",
"ABSTRACT",
"Symbols (or more broadly, non-natural language textual representations) such as numerical",
"sequences, molecular formulas, and table delimiters widely exist, playing important roles in",
"[T's]",
"various tasks such as abstract reasoning, chemical property prediction, and table question",
"answering. Despite the impressive natural language comprehension capabilities of large",
"language models (LLMs), their reasoning abilities for symbols remain inadequate, which",
"could attributed to the difference between symbol representations and general natural",
"languages.We propose symbol-to-language (S2L), a tuning-free method that enables",
"large language models to solve symbol-related problems with information expressed in",
"A",
"natural language. Specifically, S2L first converts the symbols involved to language-based",
"representations, which can be implemented by prompting LLMs or leveraging external",
"tools, then these language-based representations are integrated into the original problem",
"via direct substitution or concatenation, serving as useful input information for LLMs.",
"We evaluate the S2L method using both API-based (GPT-4, ChatGPT) and open-source",
"(OpenChat) models over eight symbol-related tasks, ranging from symbol-only abstract",
"reasoning to sentiment analysis in social media. Experimental results show that S2L",
"consistently leads to superior performance. For example, by employing S2L for GPT-",
"4, there can be average significant improvements of +21.9% and +9.5% for subtasks",
"in 1D-ARC and Dyck language,respectively.Codes and data are available at https:",
"//github.com/THUNLP-MT/symbol2language.",
"Symbol-Based Representation",
"Language-BasedRepresentation (Ours)",
"01,0.0.0→",
"0,0,1,0,0",
"OneO:One 1Three0→",
"TwoO:One1;Two0",
"0,0,0,6,0",
"",
"",
"00.6.0.0→",
"TwoO:One6:Two0",
">ThreeO:One 6:One0",
"",
"",
")0,2,2,0,0",
"02,2,2",
"One O:Two 2;Two 0",
"→?",
"",
"O:Two",
"2:One",
"PropertyPredictionEmotionofEmojis",
"TwO",
"DyckLanguage",
"open square bracket;",
"Propionylol",
"crying face",
"U+1F62D",
"［]",
"CCCO",
"close square bracket",
"TableQuestion Answering",
"Sentiment",
"Sweden isranked first",
"laughingmy ass",
"Rank/Nation|Goldin1|Sweden|15ln",
"Imaooo all day !!",
"with 15 gold medals.",
"off all day!",
"Figure 1: Top: This illustration presents a symbol-related problem (move 1 pixel forward) from the 1D-ARC",
"benchmark, comparing the responses of large language models using both conventional symbol-based and our",
"language-based representations with symbol-to-language (S2L) conversion. Bottom: The S2L conversion has",
"broad applicability across various scenarios involving diverse types of symbols."
] |
{
"arxiv_id": "2401.11725",
"language": "en",
"page_idx": 0
}
|
||
2401.11725-en-2
|
2401.11725-en-2.jpg
|
Preprint.
<table>
<tbody>
<tr>
<th>$\mathbf{Symbol}$</th>
<th>Instance</th>
<th>Task</th>
<th>$\mathbf{Example}$</th>
</tr>
<tr>
<td>Sequence of Numbers</td>
<td>1,0,0</td>
<td>Abstract Reasoning (§4.1)</td>
<td>$1, 0, 0\to 0, 0, 1;$ $5, 0, 0\to ?$</td>
</tr>
<tr>
<td>String of Brackets</td>
<td>$[\left(\:\right)]$</td>
<td>Dyck Language ${\mathrm{e}}\left(\S\:4.2\right)$</td>
<td>$( [ ] \rightarrow ) ;$ $\{ ( < > ) \rightarrow \} ;$ $\{ \} [ \rightarrow$ ?</td>
</tr>
<tr>
<td>Molecular Formula</td>
<td>$CCCCO$</td>
<td>Property Prediction (§4.3)</td>
<td>$\operatorname{ccco}\left(\textit{Toxicity: Yes or No}?\right)$</td>
</tr>
<tr>
<td>Emoji ( Unicode) </td>
<td>$(\mathbf{U}+1$F62D) T</td>
<td>Emotion Analysis(§ 2<2 4.4)</td>
<td>心 $( Anger? Fear? \textit{ Joy? Sadness? })$</td>
</tr>
<tr>
<td>Table Delimiter</td>
<td>$1,\backslash$n</td>
<td>Question Answering (§ 4.5) Fact Verification $(\S4.5)$</td>
<td>rank|name|wins$\backslash$n1|jack$|3\backslash$n $(Statement:TrueorFalse?)$</td>
</tr>
<tr>
<td>Abbreviation</td>
<td>$LMAO$</td>
<td>Stance Detection (§ 5 4.6) Sentiment Classification (§ 4.6)</td>
<td>Imagine being that bold LMAO $(Text\colon Positive~or~Negative?)$</td>
</tr>
</tbody>
</table>
Table 1: Symbols, instances, and examples across eight different types of tasks in our experiments, varying
from symbol-only inductive abstract reasoning to traditional sentiment classification in social media.
abstract reasoning tasks based on the strong ability to generate executable codes. Qiu et al.(2023) assess LLMs on symbol-based inductive reasoning tasks, uncovering a range of counter-intuitive behaviors. Gendron et al. (2023) demonstrate that LLMs possess a constrained capacity for abstract reasoning compared with other natural language tasks. These works indicate that there is still room for improvement in the reasoning ability of LLMs for symbol-related problems. Our method tries to bridge the gap between symbol understanding and language models by replenishing language-based representation, eliciting the LLMs' capability of solving symbol-related problems without affecting their general abilities.
Reasoning by Chain-of-Thought Prompting. Chain-of-Thought style prompting (Wei et al., 2022; Kojima et al., 2022; Chen et al., 2022; Besta et al., 2023; Yao et al., 2023; Zhang et al., 2023b) have become integral in augmenting the reasoning capabilities of LLMs. Cheng et al.(2023) and Tang et al.(2023b) propose to generate readable explanations for solving commonsense and program translation problems, respectively. Deng et al. (2023) introduce rephrase-and-respond to tackle potentially ambiguous questions by using self-rephrased questions, which shares similarities with our symbol-to-language method in $rephrasing.$ However, we aim to address symbol-related problems, rephrasing symbols into their natural language equivalents to allow LLMs to engage with more accessible language-based information for reasoning.
Reasoning with Symbolic Methods. There is a series of studies on the integration of symbolic methods to solve general reasoning tasks. Wei et al. (2023) propose symbol-tuning to fine-tune LLMs using substituted labels, aiming to bolster their in-context learning abilities. Hu et al. (2023) propose chain-of-symbol to solve planning-based tasks. Wang et al. (2023c) introduce meta-reasoning as a means to construct generic symbolic representations for reasoning tasks. Fang et al.(2024) design symbolic module in LLM agent for solving text-based games. Trinh et al. (2024) combine LLMs and symbolic engines to solve geometry problems. As a comparison, our work focuses on symbol-related tasks and proposes eliciting the powerful comprehension and reasoning abilities of LLMs at the language level to help solve the problems
### 3 $\textbf{Symbol- to- Language}$
We consider a range of symbol-related problems depicted in Table 1, in which the interpretation of symbol meanings is crucial for accomplishing the associated tasks. We formalize this type of problem as a question $q_s$ that incorporates a set of symbols $s=\{s_1,s_2,...,s_m\}$ and try using current LLMs to solve $q_s.$ Vanilla zero-shot (Brown et al., 2020) and zero-shot-CoT (Kojima et al., 2022) method directly solve the original problem and the responses of these two methods by LLM $\mathcal{M}$ can be written as:
(1)
$$\mathcal{R}_{\mathrm{zs}}=\mathcal{M}(q_{s});\quad\mathcal{R}_{\mathrm{zsc}}=\mathcal{M}(q_{s}\oplus p),$$
where zs and zsc indicate zero-shot and zero-shot-CoT, respectively. $p$ is a prompt like“Let’s think step by
$step”$,and$\oplus$is the concatenation operation.
The above methods directly tackle the problem with symbol-based representations. Instead of designing external prompt $p$,we focus on the question $q_s$ and propose utilizing language-based representation to better leverage the LLMs’ strong capabilities of natural language for solving symbol-related problems. Specifically, the S2L framework first converts the symbols $s_i$ $(i=1,...,m)$ to its corresponding plain text $l_i$ with a conversion operation $f$,which can be implemented by either LLMs themselves or external tools. Then the
3
|
[
[
0.143798828125,
0.043548583984375,
0.205078125,
0.043548583984375,
0.205078125,
0.058074951171875,
0.143798828125,
0.058074951171875
],
[
0.1544189453125,
0.09344482421875,
0.2115478515625,
0.09344482421875,
0.2115478515625,
0.10797119140625,
0.1544189453125,
0.10797119140625
],
[
0.41015625,
0.0921630859375,
0.44677734375,
0.0921630859375,
0.44677734375,
0.10797119140625,
0.41015625,
0.10797119140625
],
[
0.619140625,
0.09344482421875,
0.68115234375,
0.09344482421875,
0.68115234375,
0.10797119140625,
0.619140625,
0.10797119140625
],
[
0.156005859375,
0.11236572265625,
0.295654296875,
0.11236572265625,
0.295654296875,
0.1256103515625,
0.156005859375,
0.1256103515625
],
[
0.324462890625,
0.0947265625,
0.38232421875,
0.0947265625,
0.38232421875,
0.1060791015625,
0.324462890625,
0.1060791015625
],
[
0.330078125,
0.11236572265625,
0.375732421875,
0.11236572265625,
0.375732421875,
0.126953125,
0.330078125,
0.126953125
],
[
0.410888671875,
0.11236572265625,
0.57666015625,
0.11236572265625,
0.57666015625,
0.126953125,
0.410888671875,
0.126953125
],
[
0.62109375,
0.11236572265625,
0.80810546875,
0.11236572265625,
0.80810546875,
0.1256103515625,
0.62109375,
0.1256103515625
],
[
0.156005859375,
0.1307373046875,
0.27294921875,
0.1307373046875,
0.27294921875,
0.1458740234375,
0.156005859375,
0.1458740234375
],
[
0.33251953125,
0.1307373046875,
0.373291015625,
0.1307373046875,
0.373291015625,
0.146484375,
0.33251953125,
0.146484375
],
[
0.410888671875,
0.1307373046875,
0.5556640625,
0.1307373046875,
0.5556640625,
0.1458740234375,
0.410888671875,
0.1458740234375
],
[
0.62158203125,
0.1307373046875,
0.826171875,
0.1307373046875,
0.826171875,
0.14453125,
0.62158203125,
0.14453125
],
[
0.156005859375,
0.149658203125,
0.27783203125,
0.149658203125,
0.27783203125,
0.1641845703125,
0.156005859375,
0.1641845703125
],
[
0.3349609375,
0.15087890625,
0.370849609375,
0.15087890625,
0.370849609375,
0.1634521484375,
0.3349609375,
0.1634521484375
],
[
0.408447265625,
0.149658203125,
0.57421875,
0.1483154296875,
0.57421875,
0.1634521484375,
0.408447265625,
0.1641845703125
],
[
0.62109375,
0.15087890625,
0.79150390625,
0.15087890625,
0.79150390625,
0.1641845703125,
0.62109375,
0.1641845703125
],
[
0.156005859375,
0.17041015625,
0.2646484375,
0.17041015625,
0.2646484375,
0.1849365234375,
0.156005859375,
0.1849365234375
],
[
0.310546875,
0.1717529296875,
0.396240234375,
0.1717529296875,
0.396240234375,
0.18310546875,
0.310546875,
0.18310546875
],
[
0.41015625,
0.1717529296875,
0.56787109375,
0.1717529296875,
0.56787109375,
0.1849365234375,
0.41015625,
0.1849365234375
],
[
0.62109375,
0.1717529296875,
0.82666015625,
0.1717529296875,
0.82666015625,
0.1849365234375,
0.62109375,
0.1849365234375
],
[
0.410888671875,
0.1900634765625,
0.58251953125,
0.1900634765625,
0.58251953125,
0.203857421875,
0.410888671875,
0.203857421875
],
[
0.619140625,
0.1900634765625,
0.830078125,
0.1900634765625,
0.830078125,
0.203857421875,
0.619140625,
0.203857421875
],
[
0.3349609375,
0.1956787109375,
0.373291015625,
0.1956787109375,
0.373291015625,
0.2108154296875,
0.3349609375,
0.2108154296875
],
[
0.156005859375,
0.1956787109375,
0.25732421875,
0.1956787109375,
0.25732421875,
0.210205078125,
0.156005859375,
0.210205078125
],
[
0.41015625,
0.2020263671875,
0.55810546875,
0.20263671875,
0.55810546875,
0.2171630859375,
0.41015625,
0.216552734375
],
[
0.62158203125,
0.2052001953125,
0.7900390625,
0.2052001953125,
0.7900390625,
0.21533203125,
0.62158203125,
0.21533203125
],
[
0.410888671875,
0.2215576171875,
0.560546875,
0.2215576171875,
0.560546875,
0.23486328125,
0.410888671875,
0.23486328125
],
[
0.62109375,
0.2215576171875,
0.8486328125,
0.2215576171875,
0.8486328125,
0.23486328125,
0.62109375,
0.23486328125
],
[
0.158447265625,
0.2291259765625,
0.2418212890625,
0.2291259765625,
0.2418212890625,
0.240478515625,
0.158447265625,
0.240478515625
],
[
0.33251953125,
0.2266845703125,
0.373291015625,
0.2266845703125,
0.373291015625,
0.2412109375,
0.33251953125,
0.2412109375
],
[
0.410888671875,
0.2342529296875,
0.60888671875,
0.2342529296875,
0.60888671875,
0.2474365234375,
0.410888671875,
0.2474365234375
],
[
0.61376953125,
0.2342529296875,
0.79736328125,
0.2342529296875,
0.79736328125,
0.2474365234375,
0.61376953125,
0.2474365234375
],
[
0.143798828125,
0.25830078125,
0.8544921875,
0.2587890625,
0.8544921875,
0.2734375,
0.143798828125,
0.272705078125
],
[
0.1446533203125,
0.2734375,
0.80126953125,
0.2734375,
0.80126953125,
0.287353515625,
0.1446533203125,
0.287353515625
],
[
0.1446533203125,
0.323974609375,
0.8544921875,
0.323974609375,
0.8544921875,
0.337158203125,
0.1446533203125,
0.337158203125
],
[
0.1446533203125,
0.33837890625,
0.8544921875,
0.33837890625,
0.8544921875,
0.3515625,
0.1446533203125,
0.3515625
],
[
0.1446533203125,
0.3515625,
0.8544921875,
0.3515625,
0.8544921875,
0.365478515625,
0.1446533203125,
0.365478515625
],
[
0.1446533203125,
0.3662109375,
0.8544921875,
0.3662109375,
0.8544921875,
0.380126953125,
0.1446533203125,
0.380126953125
],
[
0.1446533203125,
0.380126953125,
0.8544921875,
0.380126953125,
0.8544921875,
0.393310546875,
0.1446533203125,
0.393310546875
],
[
0.1446533203125,
0.393310546875,
0.8544921875,
0.393310546875,
0.8544921875,
0.4072265625,
0.1446533203125,
0.4072265625
],
[
0.1446533203125,
0.4072265625,
0.57275390625,
0.4072265625,
0.57275390625,
0.42041015625,
0.1446533203125,
0.42041015625
],
[
0.1446533203125,
0.425537109375,
0.8544921875,
0.4267578125,
0.8544921875,
0.44140625,
0.1446533203125,
0.43994140625
],
[
0.1446533203125,
0.44140625,
0.8544921875,
0.44140625,
0.8544921875,
0.455078125,
0.1446533203125,
0.455078125
],
[
0.1446533203125,
0.455078125,
0.8544921875,
0.455078125,
0.8544921875,
0.4697265625,
0.1446533203125,
0.4697265625
],
[
0.1446533203125,
0.4697265625,
0.8544921875,
0.4697265625,
0.8544921875,
0.48291015625,
0.1446533203125,
0.48291015625
],
[
0.1446533203125,
0.48291015625,
0.8544921875,
0.48291015625,
0.8544921875,
0.496826171875,
0.1446533203125,
0.496826171875
],
[
0.1446533203125,
0.496826171875,
0.8544921875,
0.496826171875,
0.8544921875,
0.51123046875,
0.1446533203125,
0.51123046875
],
[
0.1446533203125,
0.51123046875,
0.85302734375,
0.51123046875,
0.85302734375,
0.5244140625,
0.1446533203125,
0.5244140625
],
[
0.1446533203125,
0.5244140625,
0.61376953125,
0.5244140625,
0.61376953125,
0.5390625,
0.1446533203125,
0.5390625
],
[
0.146240234375,
0.54345703125,
0.85302734375,
0.54345703125,
0.85302734375,
0.556640625,
0.146240234375,
0.556640625
],
[
0.1446533203125,
0.5595703125,
0.8544921875,
0.5595703125,
0.8544921875,
0.57275390625,
0.1446533203125,
0.57275390625
],
[
0.1446533203125,
0.57275390625,
0.8544921875,
0.57275390625,
0.8544921875,
0.5859375,
0.1446533203125,
0.5859375
],
[
0.1446533203125,
0.5869140625,
0.8544921875,
0.5869140625,
0.8544921875,
0.60107421875,
0.1446533203125,
0.60107421875
],
[
0.1446533203125,
0.60107421875,
0.8544921875,
0.60107421875,
0.8544921875,
0.6142578125,
0.1446533203125,
0.6142578125
],
[
0.1446533203125,
0.6142578125,
0.8544921875,
0.6142578125,
0.8544921875,
0.62744140625,
0.1446533203125,
0.62744140625
],
[
0.1446533203125,
0.62744140625,
0.8544921875,
0.62744140625,
0.8544921875,
0.6416015625,
0.1446533203125,
0.6416015625
],
[
0.143798828125,
0.6416015625,
0.64892578125,
0.6416015625,
0.64892578125,
0.65478515625,
0.143798828125,
0.65478515625
],
[
0.1478271484375,
0.6787109375,
0.158447265625,
0.6787109375,
0.158447265625,
0.68798828125,
0.1478271484375,
0.68798828125
],
[
0.1707763671875,
0.6767578125,
0.351318359375,
0.6767578125,
0.351318359375,
0.69140625,
0.1707763671875,
0.69140625
],
[
0.143798828125,
0.70703125,
0.8544921875,
0.70703125,
0.8544921875,
0.72021484375,
0.143798828125,
0.72021484375
],
[
0.143798828125,
0.7197265625,
0.85546875,
0.72021484375,
0.85546875,
0.73486328125,
0.143798828125,
0.734375
],
[
0.1446533203125,
0.73486328125,
0.85546875,
0.73486328125,
0.85546875,
0.75,
0.1446533203125,
0.75
],
[
0.1446533203125,
0.74853515625,
0.85546875,
0.74853515625,
0.85546875,
0.76220703125,
0.1446533203125,
0.76220703125
],
[
0.1446533203125,
0.76318359375,
0.66650390625,
0.76318359375,
0.66650390625,
0.7763671875,
0.1446533203125,
0.7763671875
],
[
0.370849609375,
0.7861328125,
0.62353515625,
0.7861328125,
0.62353515625,
0.80029296875,
0.370849609375,
0.80029296875
],
[
0.83251953125,
0.7861328125,
0.85546875,
0.7861328125,
0.85546875,
0.7998046875,
0.83251953125,
0.7998046875
],
[
0.1446533203125,
0.80810546875,
0.8544921875,
0.80810546875,
0.8544921875,
0.82275390625,
0.1446533203125,
0.82275390625
],
[
0.1446533203125,
0.82275390625,
0.431396484375,
0.82275390625,
0.431396484375,
0.8359375,
0.1446533203125,
0.8359375
],
[
0.143798828125,
0.84130859375,
0.8544921875,
0.84228515625,
0.8544921875,
0.85693359375,
0.143798828125,
0.85595703125
],
[
0.1446533203125,
0.85791015625,
0.85546875,
0.85791015625,
0.85546875,
0.8720703125,
0.1446533203125,
0.8720703125
],
[
0.1446533203125,
0.87060546875,
0.8544921875,
0.87060546875,
0.8544921875,
0.8837890625,
0.1446533203125,
0.8837890625
],
[
0.1422119140625,
0.88330078125,
0.85546875,
0.8837890625,
0.85546875,
0.8984375,
0.1422119140625,
0.89794921875
],
[
0.143798828125,
0.8984375,
0.85302734375,
0.8984375,
0.85302734375,
0.912109375,
0.143798828125,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Preprint.",
"Symbol",
"Task",
"Example",
"Sequence of Numbers",
"Instance",
"1,0,0",
"Abstract Reasoning (§ 4.1)",
"1,0,0→0,0,1; 5,0,0→?",
"String of Brackets",
"[0]",
"Dyck Language ( 4.2)",
"( →); (<>) →}; 0[→?",
"Molecular Formula",
"CCCO",
"Property Prediction (g4.3)",
"CccO (Toxicity: Yes or No?)",
"Emoji (Unicode)",
"@(U+1F62D)",
"Emotion Analysis (§ 4.4)",
" (Anger? Fear? Joy? Sadness?)",
"Question Answering (§ 4.5)",
"rank/name/wins n1ljack/3|n",
"I, n",
"Table Delimiter",
"Fact Verification ( 4.5)",
"(Statement:True or False?)",
"Stance Detection (S 4.6)",
"Imagine being that bold LMAo",
"Abbreviation",
"LMAO",
"Sentiment Classification (§ 4.6)",
"(Text: Positive or Negative?)",
"Table l: Symbols, instances, and examples across eight different types of tasks in our experiments, varying",
"from symbol-only inductive abstract reasoning to traditional sentiment classification in social media.",
"abstract reasoning tasks based on the strong ability to generate executable codes. Qiu et al. (2023) assess",
"LLMs on symbol-based inductive reasoning tasks, uncovering a range of counter-intuitive behaviors. Gendron",
"et al. (2023) demonstrate that LLMs possess a constrained capacity for abstract reasoning compared with other",
"natural language tasks. These works indicate that there is still room for improvement in the reasoning ability",
"of LLMs for symbol-related problems. Our method tries to bridge the gap between symbol understanding and",
"language models by replenishing language-based representation, eliciting the LLMs' capability of solving",
"symbol-related problems without affecting their general abilities.",
"Reasoning by Chain-of-Thought Prompting. Chain-of-Thought style prompting (Wei et al., 2022; Kojima",
"et al., 2022; Chen et al., 2022; Besta et al., 2023; Yao et al., 2023; Zhang et al., 2023b) have become integral",
"in augmenting the reasoning capabilities of LLMs. Cheng et al. (2023) and Tang et al. (2023b) propose to",
"generate readable explanations for solving commonsense and program translation problems, respectively. Deng",
"et al. (2023) introduce rephrase-and-respond to tackle potentially ambiguous questions by using self-rephrased",
"questions, which shares similarities with our symbol-to-language method in rephrasing. However, we aim to",
"address symbol-related problems, rephrasing symbols into their natural language equivalents to allow LLMs to",
"engage with more accessible language-based information for reasoning.",
"Reasoning with Symbolic Methods. There is a series of studies on the integration of symbolic methods to",
"solve general reasoning tasks. Wei et al. (2023) propose symbol-tuning to fine-tune LLMs using substituted",
"labels, aiming to bolster their in-context learning abilities. Hu et al. (2023) propose chain-of-symbol to solve",
"planning-based tasks. Wang et al. (2023c) introduce meta-reasoning as a means to construct generic symbolic",
"representations for reasoning tasks. Fang et al. (2024) design symbolic module in LLM agent for solving",
"text-based games. Trinh et al. (2024) combine LLMs and symbolic engines to solve geometry problems. As a",
"comparison, our work focuses on symbol-related tasks and proposes eliciting the powerful comprehension and",
"reasoning abilities of LLMs at the language level to help solve the problems.",
"3",
"Symbol-to-Language",
"We consider a range of symbol-related problems depicted in Table l, in which the interpretation of symbol",
"meanings is crucial for accomplishing the associated tasks. We formalize this type of problem as a question",
"qs that incorporates a set of symbols s = [si, $2, .., Sm] and try using current LLMs to solve qs. Vanilla",
"zero-shot (Brown et al., 2020) and zero-shot-CoT (Kojima et al., 2022) method directly solve the original",
"problem and the responses of these two methods by LLM M can be written as:",
"Rzs = M(qs);Rzsc =M(qs ①p),",
"(1)",
"where zs and zsc indicate zero-shot and zero-shot-CoT, respectively p is a prompt like “Let's think step by",
"step\", and @ is the concatenation operation.",
"The above methods directly tackle the problem with symbol-based representations. Instead of designing",
"external prompt p, we focus on the question qs and propose utilizing language-based representation to better",
"leverage the LLMs' strong capabilities of natural language for solving symbol-related problems. Specifically,",
"the s2L framework first converts the symbols si (i = 1,..,m) to its corresponding plain text l with a",
"conversion operation f, which can be implemented by either LLMs themselves or external tools. Then the",
"3"
] |
{
"arxiv_id": "2401.11725",
"language": "en",
"page_idx": 2
}
|
||
2401.11725-en-3
|
2401.11725-en-3.jpg
|
Preprint.
<FigureHere>
Figure 2: Example of applying symbol-to-language for 1D abstract reasoning task. We convert every sequence
to its textual representation by prompting LLMs or using simple rules implemented in code, and then we
transform the symbolized problem to language-enhanced or language-only representations.
S2L framework incorporates the converted language-based representation $l_i$ into two alternative questions $q_l$
or $q_{s\oplus l}$ as the input for LLMs to generate answers. The details are discussed as follows.
## 3.1 Symbol-to-Language Conversion
Conversion with LLMs. We first employ the LLMs $\mathcal{M}$ to convert symbols $s_i$ to their corresponding natural
language descriptions $l_i^\mathrm{LLM}$ via zero-shot prompting, as expressed by:
(2)
$$l_{i}^{\mathrm{LLM}}=f_{\mathrm{LLM}}\circ s_{i}=\mathcal{M}(p_{\mathrm{s2l}}\oplus s_{i}),$$
where $f_\mathrm{LLM}\circ s_i$ denotes converting $s_i$ using the LLM $\mathcal{M} , p_\mathrm{s2l}$ is the task-specific prompt facilitating the S2L conversion. For instance, when the symbol-related question $q_s$ is about property prediction and $s_i$ is a molecular formula, $p_{\mathrm{s21}}$ could be “What does the following molecular formula represent?”.
Conversion with Tools. Considering there exist some constructed “symbol-language”pairs, we further propose using external tools for conversion, which can take on several forms. Rule-based codes, for example, can convert $s_i=$“ranklnation\n1|SwE” into $l_i^\mathrm{rule}=$“rank: 1; nation: SWE” according to the table delimiters“|”and“\n”. Translators can transform molecular formulas into their formal names, such as converting $s_i=“$CCCO" into $l_i^\text{translator}=“$Propionylo”. Unicode dictionaries can provide descriptions of emojis, like converting $s_i=“$U+1F62D”into $l_i^\mathrm{dict}=“$crying face". Despite having some limitations in terms of usage scenarios, conversion with tools offers two primary advantages: 1) it circumvents the costs associated with using LLMs; 2) it provides verified language-based information, which can help reduce potential errors in descriptions generated by LLMs.
3.2 Utilizing Language-Based Representations
We propose two alternative ways that incorporate the language-based representation $l_i$ into the final input.
4
|
[
[
0.143798828125,
0.043548583984375,
0.205078125,
0.043548583984375,
0.205078125,
0.058074951171875,
0.143798828125,
0.058074951171875
],
[
0.17724609375,
0.091552734375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.106689453125,
0.17724609375,
0.1060791015625
],
[
0.5498046875,
0.0921630859375,
0.7939453125,
0.0921630859375,
0.7939453125,
0.1060791015625,
0.5498046875,
0.1060791015625
],
[
0.1854248046875,
0.11297607421875,
0.281005859375,
0.11297607421875,
0.281005859375,
0.125,
0.1854248046875,
0.125
],
[
0.31787109375,
0.11297607421875,
0.33251953125,
0.11297607421875,
0.33251953125,
0.12249755859375,
0.31787109375,
0.12249755859375
],
[
0.371826171875,
0.11297607421875,
0.465576171875,
0.11297607421875,
0.465576171875,
0.125,
0.371826171875,
0.125
],
[
0.5185546875,
0.10986328125,
0.83251953125,
0.10986328125,
0.83251953125,
0.12371826171875,
0.5185546875,
0.12371826171875
],
[
0.1854248046875,
0.1300048828125,
0.281982421875,
0.1300048828125,
0.281982421875,
0.141357421875,
0.1854248046875,
0.141357421875
],
[
0.370849609375,
0.1300048828125,
0.46728515625,
0.1300048828125,
0.46728515625,
0.141357421875,
0.370849609375,
0.141357421875
],
[
0.521484375,
0.126953125,
0.82421875,
0.126953125,
0.82421875,
0.14013671875,
0.521484375,
0.14013671875
],
[
0.521484375,
0.143310546875,
0.7294921875,
0.143310546875,
0.7294921875,
0.1572265625,
0.521484375,
0.1572265625
],
[
0.31787109375,
0.1319580078125,
0.328369140625,
0.1287841796875,
0.33251953125,
0.1383056640625,
0.32177734375,
0.141357421875
],
[
0.1854248046875,
0.14453125,
0.281005859375,
0.14453125,
0.281005859375,
0.1558837890625,
0.1854248046875,
0.1558837890625
],
[
0.314453125,
0.147705078125,
0.330810546875,
0.147705078125,
0.330810546875,
0.1572265625,
0.314453125,
0.1572265625
],
[
0.410888671875,
0.14453125,
0.42724609375,
0.14453125,
0.42724609375,
0.1590576171875,
0.410888671875,
0.1590576171875
],
[
0.1871337890625,
0.1622314453125,
0.250732421875,
0.1622314453125,
0.250732421875,
0.173583984375,
0.1871337890625,
0.173583984375
],
[
0.25,
0.1654052734375,
0.348876953125,
0.1654052734375,
0.348876953125,
0.17236328125,
0.25,
0.17236328125
],
[
0.51953125,
0.1610107421875,
0.68701171875,
0.1610107421875,
0.68701171875,
0.1749267578125,
0.51953125,
0.1749267578125
],
[
0.2296142578125,
0.18115234375,
0.420654296875,
0.179931640625,
0.420654296875,
0.1944580078125,
0.2296142578125,
0.1956787109375
],
[
0.576171875,
0.1817626953125,
0.7685546875,
0.1817626953125,
0.7685546875,
0.1956787109375,
0.576171875,
0.1956787109375
],
[
0.17236328125,
0.199462890625,
0.2296142578125,
0.199462890625,
0.2296142578125,
0.2108154296875,
0.17236328125,
0.2108154296875
],
[
0.525390625,
0.1988525390625,
0.6640625,
0.1988525390625,
0.6640625,
0.212158203125,
0.525390625,
0.212158203125
],
[
0.1707763671875,
0.2108154296875,
0.4697265625,
0.212158203125,
0.4697265625,
0.2259521484375,
0.1707763671875,
0.2247314453125
],
[
0.1732177734375,
0.2247314453125,
0.26318359375,
0.2247314453125,
0.26318359375,
0.236083984375,
0.1732177734375,
0.236083984375
],
[
0.52685546875,
0.2247314453125,
0.7900390625,
0.2247314453125,
0.7900390625,
0.238037109375,
0.52685546875,
0.238037109375
],
[
0.52685546875,
0.2374267578125,
0.82177734375,
0.2374267578125,
0.82177734375,
0.250732421875,
0.52685546875,
0.250732421875
],
[
0.17236328125,
0.25,
0.298095703125,
0.25,
0.298095703125,
0.26318359375,
0.17236328125,
0.26318359375
],
[
0.1748046875,
0.26318359375,
0.4248046875,
0.26318359375,
0.4248046875,
0.27783203125,
0.1748046875,
0.27783203125
],
[
0.52783203125,
0.264404296875,
0.685546875,
0.264404296875,
0.685546875,
0.27783203125,
0.52783203125,
0.27783203125
],
[
0.1707763671875,
0.27783203125,
0.439453125,
0.27783203125,
0.439453125,
0.291748046875,
0.1707763671875,
0.291748046875
],
[
0.525390625,
0.276611328125,
0.6845703125,
0.276611328125,
0.6845703125,
0.290283203125,
0.525390625,
0.290283203125
],
[
0.1478271484375,
0.29736328125,
0.84619140625,
0.29736328125,
0.84619140625,
0.3125,
0.1478271484375,
0.3125
],
[
0.1732177734375,
0.313232421875,
0.46728515625,
0.313232421875,
0.46728515625,
0.326904296875,
0.1732177734375,
0.326904296875
],
[
0.5361328125,
0.31494140625,
0.74267578125,
0.31494140625,
0.74267578125,
0.328857421875,
0.5361328125,
0.328857421875
],
[
0.8349609375,
0.317626953125,
0.84326171875,
0.317626953125,
0.84326171875,
0.32080078125,
0.8349609375,
0.32080078125
],
[
0.1748046875,
0.32763671875,
0.4755859375,
0.32763671875,
0.4755859375,
0.341552734375,
0.1748046875,
0.341552734375
],
[
0.53515625,
0.330078125,
0.736328125,
0.330078125,
0.736328125,
0.343505859375,
0.53515625,
0.343505859375
],
[
0.1748046875,
0.343505859375,
0.259033203125,
0.34228515625,
0.259033203125,
0.354248046875,
0.1748046875,
0.354736328125
],
[
0.1732177734375,
0.354736328125,
0.441162109375,
0.354736328125,
0.441162109375,
0.369384765625,
0.1732177734375,
0.369384765625
],
[
0.53515625,
0.363037109375,
0.740234375,
0.363037109375,
0.740234375,
0.376953125,
0.53515625,
0.376953125
],
[
0.1732177734375,
0.376953125,
0.4345703125,
0.376953125,
0.4345703125,
0.39013671875,
0.1732177734375,
0.39013671875
],
[
0.5361328125,
0.37744140625,
0.7490234375,
0.37744140625,
0.7490234375,
0.391357421875,
0.5361328125,
0.391357421875
],
[
0.1732177734375,
0.38818359375,
0.441162109375,
0.389404296875,
0.441162109375,
0.404052734375,
0.1732177734375,
0.40283203125
],
[
0.1748046875,
0.412353515625,
0.431396484375,
0.412353515625,
0.431396484375,
0.425537109375,
0.1748046875,
0.425537109375
],
[
0.5361328125,
0.410888671875,
0.7509765625,
0.410400390625,
0.7509765625,
0.4248046875,
0.5361328125,
0.425537109375
],
[
0.17236328125,
0.425537109375,
0.4306640625,
0.4248046875,
0.4306640625,
0.439453125,
0.17236328125,
0.43994140625
],
[
0.5361328125,
0.4248046875,
0.814453125,
0.4248046875,
0.814453125,
0.438232421875,
0.5361328125,
0.438232421875
],
[
0.146240234375,
0.455078125,
0.8544921875,
0.455078125,
0.8544921875,
0.468505859375,
0.146240234375,
0.468505859375
],
[
0.1446533203125,
0.468505859375,
0.8544921875,
0.468505859375,
0.8544921875,
0.481689453125,
0.1446533203125,
0.481689453125
],
[
0.1446533203125,
0.48291015625,
0.74267578125,
0.48291015625,
0.74267578125,
0.496826171875,
0.1446533203125,
0.496826171875
],
[
0.1446533203125,
0.5341796875,
0.85302734375,
0.53515625,
0.85302734375,
0.5498046875,
0.1446533203125,
0.548828125
],
[
0.143798828125,
0.5498046875,
0.72119140625,
0.548828125,
0.72119140625,
0.56298828125,
0.143798828125,
0.564453125
],
[
0.1446533203125,
0.58349609375,
0.414306640625,
0.58349609375,
0.414306640625,
0.59765625,
0.1446533203125,
0.59765625
],
[
0.146240234375,
0.60791015625,
0.85302734375,
0.60791015625,
0.85302734375,
0.62255859375,
0.146240234375,
0.62255859375
],
[
0.143798828125,
0.62060546875,
0.6044921875,
0.62060546875,
0.6044921875,
0.63818359375,
0.143798828125,
0.63818359375
],
[
0.37744140625,
0.646484375,
0.61767578125,
0.6494140625,
0.61767578125,
0.66845703125,
0.37744140625,
0.66552734375
],
[
0.83251953125,
0.6513671875,
0.85546875,
0.6513671875,
0.85546875,
0.6640625,
0.83251953125,
0.6640625
],
[
0.146240234375,
0.67529296875,
0.8544921875,
0.67529296875,
0.8544921875,
0.68994140625,
0.146240234375,
0.68994140625
],
[
0.1446533203125,
0.68994140625,
0.8544921875,
0.68994140625,
0.8544921875,
0.705078125,
0.1446533203125,
0.705078125
],
[
0.1446533203125,
0.705078125,
0.73779296875,
0.705078125,
0.73779296875,
0.71826171875,
0.1446533203125,
0.71826171875
],
[
0.146240234375,
0.7236328125,
0.8544921875,
0.7236328125,
0.8544921875,
0.7373046875,
0.146240234375,
0.7373046875
],
[
0.143798828125,
0.73779296875,
0.85546875,
0.7373046875,
0.85546875,
0.751953125,
0.143798828125,
0.75244140625
],
[
0.1446533203125,
0.75048828125,
0.8544921875,
0.75048828125,
0.8544921875,
0.76513671875,
0.1446533203125,
0.76513671875
],
[
0.1446533203125,
0.7666015625,
0.8544921875,
0.7666015625,
0.8544921875,
0.77978515625,
0.1446533203125,
0.77978515625
],
[
0.1422119140625,
0.77978515625,
0.404296875,
0.7763671875,
0.404296875,
0.79443359375,
0.1422119140625,
0.79736328125
],
[
0.40283203125,
0.78271484375,
0.85302734375,
0.78271484375,
0.85302734375,
0.79345703125,
0.40283203125,
0.79345703125
],
[
0.1422119140625,
0.79443359375,
0.85546875,
0.79345703125,
0.85546875,
0.81103515625,
0.1422119140625,
0.8125
],
[
0.1446533203125,
0.81005859375,
0.8544921875,
0.81005859375,
0.8544921875,
0.82373046875,
0.1446533203125,
0.82373046875
],
[
0.146240234375,
0.82568359375,
0.85302734375,
0.82568359375,
0.85302734375,
0.8359375,
0.146240234375,
0.8359375
],
[
0.1446533203125,
0.83837890625,
0.363525390625,
0.83837890625,
0.363525390625,
0.8515625,
0.1446533203125,
0.8515625
],
[
0.1446533203125,
0.8720703125,
0.48046875,
0.8720703125,
0.48046875,
0.88623046875,
0.1446533203125,
0.88623046875
],
[
0.1446533203125,
0.8984375,
0.83642578125,
0.8984375,
0.83642578125,
0.912109375,
0.1446533203125,
0.912109375
],
[
0.488525390625,
0.939453125,
0.50244140625,
0.9345703125,
0.5087890625,
0.9443359375,
0.495849609375,
0.94970703125
]
] |
[
"Preprint.",
"(a)Underlying Visual Problem (Move-2p)",
"(b)Symbol-BasedRepresentation",
"",
"",
"",
"input 1: 0,0,1,1,0,0,0,0output 1: 0,0,0,0,1,1,0,0-",
"",
"",
"input 2: 0,6,6,6,6,0,0,0 output 2: 0,0,0,6,6,6,6,0",
"input 3: 0,0,2,2,2,0,0,0 output 3:",
"X",
"",
"",
"?",
"(Answer",
"",
"(Answer.0.0.0.0.2,2,2.2.0)",
"(c.1) Conversion with LLM",
"(d.1) Conversion with Rule",
"(Prompt)",
"(Input)0,0,1,1,0,0,0,0",
"Describe the sequence of digits using language",
"0,0,1,1,0,0,0,0",
"(Rules,amanually written python function",
"def SeqTransition(Input):...return Output",
"(ResponsebyLLM)",
"The sequence starts with two zeros,",
"(Output of SegTransition)",
"followed by two ones, then four zeros.",
"Twoo,Two1,Four 0.",
"(c.2) Language-Based Representation with LLM:(d.2)Language-Based Representation withRule",
"input 1: 0,0,1,1,0,0,0,0 (The sequence starts",
"input 1: Two 0, Two 1, Four 0.",
"",
"with two zeros,followed bytwo ones,then",
"output 1: Four O, Two 1, Two 0.",
"four zeros.)",
"output 1: 0,0.0,0,1,1,0,0 (The sequence ..)",
"input 2: One O, Four 6, Three 0.",
"input 2: 0,6,6,6,6,0,0.0 (The sequence ...)",
"output 2: Three O, Four 6, One 0.",
"output 2: 0,0,0,6,6,6,6,0 (The sequence ...)",
"input 3: 0,0,2,2,2,0,0,0 (The sequence ...)",
"input 3: Two 0, Three 2, Three 0.",
"output 3: (Answer.0.0.0,0,2,2,2,2.0 (.-.)",
"output 3: (Answer:Four 0, Three 2,One 0.)",
"Figure 2: Example of applying symbol-to-language for ID abstract reasoning task. We convert every sequence",
"to its textual representation by prompting LLMs or using simple rules implemented in code, and then we",
"transform the symbolized problem to language-enhanced or language-only representations.",
"S2L framework incorporates the converted language-based representation l, into two alternative questions qi",
"or qs@i as the input for LLMs to generate answers. The details are discussed as follows.",
"3.1 Symbol-to-Language Conversion",
"Conversion with LLMs. We first employ the LLMs M to convert symbols s, to their corresponding natural",
"language descriptions ILLM via zero-shot prompting, as expressed by:",
"ILM = fLLM 0 8i = M(Ps21 ④ si),",
"(2)",
"where fLLM o Si denotes converting Si using the LLM M, Ps21 is the task-specific prompt facilitating the",
"S2L conversion. For instance, when the symbol-related question qs is about property prediction and s is a",
"molecular formula, ps2i could be \"What does the following molecular formula represent?\"",
"Conversion with Tools. Considering there exist some constructed “symbol-language\"pairs, we further",
"propose using external tools for conversion, which can take on several forms. Rule-based codes, for example,",
"can convert si = “rank|nation niISwE\" into lrule = \"rank: l; nation: SWE\" according to the table",
"delimiters \"\"' and \"n\". Translators can transform molecular formulas into their formal names, such as",
"converting si = “CCCo\" into translator",
"=“Propionylo\"Unicode dictionariescanprovidedescriptions of",
"emojis, like converting si =“U+1F62D\" into Igict =“crying face\". Despite having some limitations in terms",
"of usage scenarios, conversion with tools offers two primary advantages: 1) it circumvents the costs associated",
"withusingLLMs;2)itprovidesverifiedlanguage-basedinformation,whichcanhelpreducepotentialerrorsin",
"descriptions generated by LLMs.",
"3.2Utilizing Language-Based Representations",
"We propose two alternative ways that incorporate the language-based representation l, into the final input.",
"4"
] |
{
"arxiv_id": "2401.11725",
"language": "en",
"page_idx": 3
}
|
||
2401.11725-en-4
|
2401.11725-en-4.jpg
|
Preprint.
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{Model}$</th>
<th colspan="2">$\mathbf{Move-1p}$</th>
<th colspan="2">$\mathbf{Move-2p}$</th>
<th colspan="2">$\mathbf{Move-3p}$</th>
<th rowspan="2">$AVG.$</th>
</tr>
<tr>
<th>$n=3$</th>
<th>$n=4$</th>
<th>$n=3$</th>
<th>$n=4$</th>
<th>$n=3$</th>
<th>$n=4$</th>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>$({\mathrm{Zero-Shc}}$</td>
<td>$\eta t)$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>gpt- 4</td>
<td>93.3</td>
<td>90.0</td>
<td>1 48.3</td>
<td>1 46.7</td>
<td>35.0</td>
<td>45.0</td>
<td>59.7</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>90.0 )(-3.3)</td>
<td>91.7 (+1.7)</td>
<td>$48.3\left(-\right)$</td>
<td>56.7 $^{*}(+10.0)$ 1</td>
<td>38.3 (+3.3)</td>
<td>48.3 (+3.3)</td>
<td>62.2 +2.5)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>96.7 (+3.4)</td>
<td>100.0 (+10.0)</td>
<td>88.3 (+40.0)</td>
<td>96.7 (+50.0)</td>
<td>48.3 (+13.3)</td>
<td>60.0 (+15.0)</td>
<td>81.6 (+21.9</td>
</tr>
<tr>
<td>gpt- 3. 5- turbo</td>
<td>68.3</td>
<td>71.7</td>
<td>11.7</td>
<td>25.0</td>
<td>23.3</td>
<td>26.7</td>
<td>37.8</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>80.0 (+11.7)</td>
<td>75.0 (+3.3)</td>
<td>31.6 (+1.9.9)</td>
<td>26.6 {:}(+1.6)</td>
<td>33.3 (+10.0)</td>
<td>31.7 (+5.0)</td>
<td>46.4 (+8.6)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>71.6 ${\tilde{J}}\left(+3.3\right)$ 1</td>
<td>88.3 (+16.6)</td>
<td>25.0 (+13.3)</td>
<td>31.7 ${\vec{7} }$ (+6.7)</td>
<td>25.0 (+1.7)</td>
<td>$26.7\left(-\right)$</td>
<td>44.7 (+6.9)</td>
</tr>
<tr>
<td>openchat- 3. 5- 7b</td>
<td>61.7</td>
<td>71.7</td>
<td>15.0</td>
<td>21.6</td>
<td>11.7</td>
<td>11.7</td>
<td>32.2</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>68.3 (+6.6)</td>
<td>78.3 (+6.6)</td>
<td>23.3 (+8.3)</td>
<td>25.0 (+3.4)</td>
<td>25.0 (+13.3)</td>
<td>21.7 (+10.0)</td>
<td>40.3 (+8.1)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>63.3 (+1.6)</td>
<td>75.0 (+3.3)</td>
<td>16.6 (+1.6) 1</td>
<td>$18.3\left(-3.3\right)$</td>
<td>15.0 (+3.3)</td>
<td>$11.7\left(-\right)$</td>
<td>33.3 (+1.1) $P_{1}=\frac{2}{2}$</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>(Zero- Shot- </td>
<td>$CoT)$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>gpt- 4</td>
<td>93.3</td>
<td>90.0</td>
<td>55.0</td>
<td>50.0</td>
<td>36.7</td>
<td>41.7</td>
<td>61.1</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>95.0 (+1.7)</td>
<td>91.7 (+1.7)</td>
<td>38.3 $3\left(-16.7\right)$</td>
<td>46.7 7 (-3.3)</td>
<td>43.3 (+6.6)</td>
<td>48.3 (+6.6)</td>
<td>60.5 (-0.6)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>96.7 (+3.4)</td>
<td>100.0 (+10.0)</td>
<td>86.7 (+31.7)</td>
<td>88.3 (+38.3)</td>
<td>50.0 (+13.3)</td>
<td>66.7 (+25.0)</td>
<td>$81.4\left(+20.3\right.$</td>
</tr>
<tr>
<td>gpt- 3. 5- turbo</td>
<td>76.7</td>
<td>78.3</td>
<td>15.0</td>
<td>25.0</td>
<td>26.7</td>
<td>25.0</td>
<td>41.1</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>$75.0\left(-1.7\right)$</td>
<td>$71.7\left(-6.6\right)$</td>
<td>30.0 +15.0)</td>
<td>35.0 0 (+10.0)</td>
<td>35.0 (+8.3)</td>
<td>30.0 (+5.0)</td>
<td>46.1 (+5.0)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>75.0 (-1.7)</td>
<td>86.7 (+8.4)</td>
<td>30.0 (+15.0)</td>
<td>36.7 7 (+11.7) 1 1</td>
<td>$21.7\left(-5.0\right)$</td>
<td>$25.0\left(-\right)$</td>
<td>45.8(+4.7)</td>
</tr>
<tr>
<td>openchat- 3. 5- 7b</td>
<td>61.7</td>
<td>71.7</td>
<td>15.0</td>
<td>21.6</td>
<td>11.7</td>
<td>11.7</td>
<td>32.2</td>
</tr>
<tr>
<td>+S2L w/ model</td>
<td>68.3 (+6.6)</td>
<td>78.3 (+6.6)</td>
<td>23.3 (+8.3)</td>
<td>25.0 (+3.4)</td>
<td>25.0 (+1.3.3)</td>
<td>21.7 (+10.0)</td>
<td>40.3 (+8.1)</td>
</tr>
<tr>
<td>+S2L w/ rule</td>
<td>63.3 (+1.6)</td>
<td>75.0 (+3.3)</td>
<td>16.6 (+1.6)</td>
<td>18.3 {:}(-3.3)</td>
<td>15.0 (+3.3)</td>
<td>$11.7\left(-\right)$</td>
<td>33.3 (+1.1)</td>
</tr>
</tbody>
</table>
Table 2: Results for three subtasks on 1D-ARC. $n$ indicates the number of given input-output pairs for finding
the patterns. The values in parentheses indicate the difference between the baseline and our results. w/ model denotes conversion with LLMs, and w/ rule denotes conversion with manually designed rules using codes.
Direct Substitution. The first way of utilization is to directly substitute the symbol-based representation $s_i$ with language-based representation $l_i.$ To some extent, $l_i$ can be regarded as the language-based equivalent of $s_i.$ Thus we can use them to replace the symbol-based representations for both the question and ground-truth label. The response by using S2L can be written as:
(3)
$$\mathcal{R}_{\mathrm{s2l}}=\mathcal{M}(q_{l}),\quad q_{l}=q_{f\circ s}=q_{f\circ\{s_{1},...,s_{m}\}}=q_{f\circ s_{1},...,f\circ s_{m}}.$$
Concatenation. However, in some other tasks, the generated $l_i$ by LLMs may not always be a perfect substitution or convey complete information of $s_i.$ This can occur for two reasons: i) $l_i$ might be incorrect due to the undesired output formats, misleading content, or noisy context; and $2)l_i$ could lose some information during S2L conversion. For instance, the ground truth might be a span-based abbreviation for table understanding $(e.g$, the“SWE”in table“rankInation\n1|SWE”), meaning that the converted $l_i$ as a full name (e.g.,“Sweden") may not exactly match the final answer. Therefore, the second method uses both the original symbol-based representation $s_i$ and the language-based representation $l_i$ as the combined input. This approach allows the LLMs $\mathcal{M}$ to reason on questions that include rich contextual information from two distinct perspectives:
$$\mathcal{R}_{\mathrm{s2l}}=\mathcal{M}(q_{s\oplus l}),\quad q_{s\oplus l}=q_{s_{1}\oplus l_{1},...,s_{m}\oplus l_{m}}=q_{s_{1}\oplus\{f\circ s_{1}\},...,s_{m}\oplus\{f\circ s_{m}\}}.$$
(4)
## 4 $\mathbf{Experiments}$
To assess the performance of the S2L framework, we conduct six categories of symbol-related problems with eight specific tasks as shown in Table 1, varying from symbol-only inductive abstract reasoning to traditional sentiment analysis in social media. As for LLMs, we evaluate both API-based and open-source models, including GPT-4 (OpenAI, 2023b), ChatGPT (OpenAI, 2022), and OpenChat-7b (Wang et al., 2023a)$^1.$ To ensure the reproducibility of the responses generated by LLMs, we set the decoding temperature as 0.
# 4.1 Abstract Reasoning
Abstract reasoning (Webb et al., 2023; Gendron et al., 2023; Wang et al., 2023b) is a type of task that involves
summarizing patterns from limited observations. We conduct experiments on subtasks from the 1D-ARC
$^{1}$Specifically, we use the versions of gpt-4-0613 and gpt-3.5-turbo-1106 released by OpenAI, and
openchat\_3.5 released in https://huggingface.co/openchat/openchat\_3.5.
5
|
[
[
0.143798828125,
0.043548583984375,
0.205078125,
0.043548583984375,
0.205078125,
0.058074951171875,
0.143798828125,
0.058074951171875
],
[
0.328369140625,
0.0889892578125,
0.384033203125,
0.091552734375,
0.38232421875,
0.106689453125,
0.326904296875,
0.10479736328125
],
[
0.495849609375,
0.09027099609375,
0.5498046875,
0.0921630859375,
0.54833984375,
0.10797119140625,
0.494384765625,
0.1060791015625
],
[
0.65869140625,
0.09027099609375,
0.7138671875,
0.0921630859375,
0.71337890625,
0.106689453125,
0.65673828125,
0.10479736328125
],
[
0.15283203125,
0.10101318359375,
0.19775390625,
0.10101318359375,
0.19775390625,
0.11553955078125,
0.15283203125,
0.11553955078125
],
[
0.7939453125,
0.10162353515625,
0.826171875,
0.10162353515625,
0.826171875,
0.1142578125,
0.7939453125,
0.1142578125
],
[
0.273681640625,
0.10797119140625,
0.314453125,
0.10797119140625,
0.314453125,
0.12249755859375,
0.273681640625,
0.12249755859375
],
[
0.35546875,
0.10797119140625,
0.396240234375,
0.10797119140625,
0.396240234375,
0.12249755859375,
0.35546875,
0.12249755859375
],
[
0.52392578125,
0.10797119140625,
0.564453125,
0.1060791015625,
0.56640625,
0.121826171875,
0.525390625,
0.12371826171875
],
[
0.60888671875,
0.10797119140625,
0.650390625,
0.10797119140625,
0.650390625,
0.12249755859375,
0.60888671875,
0.12249755859375
],
[
0.68896484375,
0.10797119140625,
0.73046875,
0.1060791015625,
0.73193359375,
0.121826171875,
0.689453125,
0.12371826171875
],
[
0.465576171875,
0.125,
0.5361328125,
0.125,
0.5361328125,
0.1395263671875,
0.465576171875,
0.1395263671875
],
[
0.44677734375,
0.10919189453125,
0.484375,
0.10919189453125,
0.484375,
0.12249755859375,
0.44677734375,
0.12249755859375
],
[
0.273681640625,
0.1363525390625,
0.3056640625,
0.1363525390625,
0.3056640625,
0.1483154296875,
0.273681640625,
0.1483154296875
],
[
0.35693359375,
0.1363525390625,
0.386474609375,
0.1363525390625,
0.386474609375,
0.1483154296875,
0.35693359375,
0.1483154296875
],
[
0.52783203125,
0.1363525390625,
0.55615234375,
0.1363525390625,
0.55615234375,
0.15087890625,
0.52783203125,
0.15087890625
],
[
0.60888671875,
0.1363525390625,
0.6396484375,
0.1363525390625,
0.6396484375,
0.1483154296875,
0.60888671875,
0.1483154296875
],
[
0.69287109375,
0.1363525390625,
0.72119140625,
0.1363525390625,
0.72119140625,
0.1483154296875,
0.69287109375,
0.1483154296875
],
[
0.77392578125,
0.1363525390625,
0.8037109375,
0.1363525390625,
0.8037109375,
0.1483154296875,
0.77392578125,
0.1483154296875
],
[
0.44677734375,
0.136962890625,
0.4755859375,
0.136962890625,
0.4755859375,
0.1483154296875,
0.44677734375,
0.1483154296875
],
[
0.1544189453125,
0.1383056640625,
0.193603515625,
0.1363525390625,
0.1953125,
0.1483154296875,
0.156005859375,
0.15087890625
],
[
0.158447265625,
0.1483154296875,
0.2459716796875,
0.1483154296875,
0.2459716796875,
0.160400390625,
0.158447265625,
0.160400390625
],
[
0.275390625,
0.1483154296875,
0.340576171875,
0.1483154296875,
0.340576171875,
0.160400390625,
0.275390625,
0.160400390625
],
[
0.35693359375,
0.1483154296875,
0.393798828125,
0.1483154296875,
0.393798828125,
0.160400390625,
0.35693359375,
0.160400390625
],
[
0.3896484375,
0.1483154296875,
0.428955078125,
0.1483154296875,
0.428955078125,
0.160400390625,
0.3896484375,
0.160400390625
],
[
0.44677734375,
0.1483154296875,
0.494384765625,
0.1483154296875,
0.494384765625,
0.160400390625,
0.44677734375,
0.160400390625
],
[
0.52783203125,
0.1483154296875,
0.5986328125,
0.1483154296875,
0.5986328125,
0.160400390625,
0.52783203125,
0.160400390625
],
[
0.609375,
0.1483154296875,
0.67333984375,
0.1483154296875,
0.67333984375,
0.160400390625,
0.609375,
0.160400390625
],
[
0.69287109375,
0.1483154296875,
0.75634765625,
0.1483154296875,
0.75634765625,
0.160400390625,
0.69287109375,
0.160400390625
],
[
0.77392578125,
0.1483154296875,
0.8408203125,
0.1483154296875,
0.8408203125,
0.160400390625,
0.77392578125,
0.160400390625
],
[
0.273681640625,
0.1590576171875,
0.340576171875,
0.1590576171875,
0.340576171875,
0.173583984375,
0.273681640625,
0.173583984375
],
[
0.69140625,
0.1590576171875,
0.76123046875,
0.1590576171875,
0.76123046875,
0.173583984375,
0.69140625,
0.173583984375
],
[
0.77392578125,
0.1590576171875,
0.85107421875,
0.1590576171875,
0.85107421875,
0.173583984375,
0.77392578125,
0.173583984375
],
[
0.358642578125,
0.160400390625,
0.43310546875,
0.160400390625,
0.43310546875,
0.1717529296875,
0.358642578125,
0.1717529296875
],
[
0.158447265625,
0.160400390625,
0.233642578125,
0.160400390625,
0.233642578125,
0.1717529296875,
0.158447265625,
0.1717529296875
],
[
0.44677734375,
0.160400390625,
0.515625,
0.160400390625,
0.515625,
0.1717529296875,
0.44677734375,
0.1717529296875
],
[
0.52783203125,
0.160400390625,
0.59716796875,
0.160400390625,
0.59716796875,
0.1717529296875,
0.52783203125,
0.1717529296875
],
[
0.611328125,
0.160400390625,
0.6787109375,
0.160400390625,
0.6787109375,
0.1717529296875,
0.611328125,
0.1717529296875
],
[
0.1519775390625,
0.175537109375,
0.25,
0.173583984375,
0.25,
0.1881103515625,
0.1519775390625,
0.1900634765625
],
[
0.275390625,
0.175537109375,
0.3056640625,
0.175537109375,
0.3056640625,
0.1881103515625,
0.275390625,
0.1881103515625
],
[
0.358642578125,
0.175537109375,
0.386474609375,
0.175537109375,
0.386474609375,
0.1881103515625,
0.358642578125,
0.1881103515625
],
[
0.44677734375,
0.175537109375,
0.473876953125,
0.175537109375,
0.473876953125,
0.1881103515625,
0.44677734375,
0.1881103515625
],
[
0.52783203125,
0.175537109375,
0.55615234375,
0.175537109375,
0.55615234375,
0.1881103515625,
0.52783203125,
0.1881103515625
],
[
0.609375,
0.175537109375,
0.6396484375,
0.175537109375,
0.6396484375,
0.1881103515625,
0.609375,
0.1881103515625
],
[
0.69287109375,
0.175537109375,
0.72119140625,
0.175537109375,
0.72119140625,
0.1881103515625,
0.69287109375,
0.1881103515625
],
[
0.77392578125,
0.175537109375,
0.80224609375,
0.175537109375,
0.80224609375,
0.1881103515625,
0.77392578125,
0.1881103515625
],
[
0.273681640625,
0.184326171875,
0.346435546875,
0.186279296875,
0.346435546875,
0.2008056640625,
0.273681640625,
0.1988525390625
],
[
0.35546875,
0.1849365234375,
0.423095703125,
0.1875,
0.423095703125,
0.2020263671875,
0.35546875,
0.199462890625
],
[
0.52685546875,
0.186279296875,
0.59326171875,
0.186279296875,
0.59326171875,
0.2008056640625,
0.52685546875,
0.2008056640625
],
[
0.60888671875,
0.186279296875,
0.6806640625,
0.186279296875,
0.6806640625,
0.2008056640625,
0.60888671875,
0.2008056640625
],
[
0.69140625,
0.186279296875,
0.75732421875,
0.186279296875,
0.75732421875,
0.2008056640625,
0.69140625,
0.2008056640625
],
[
0.77392578125,
0.184326171875,
0.84228515625,
0.186279296875,
0.84228515625,
0.2008056640625,
0.77392578125,
0.1988525390625
],
[
0.44677734375,
0.1881103515625,
0.51708984375,
0.1881103515625,
0.51708984375,
0.199462890625,
0.44677734375,
0.199462890625
],
[
0.158447265625,
0.1881103515625,
0.2459716796875,
0.1881103515625,
0.2459716796875,
0.1988525390625,
0.158447265625,
0.1988525390625
],
[
0.35693359375,
0.1962890625,
0.428955078125,
0.1988525390625,
0.428955078125,
0.21337890625,
0.35693359375,
0.2108154296875
],
[
0.52685546875,
0.1976318359375,
0.59326171875,
0.1976318359375,
0.59326171875,
0.212158203125,
0.52685546875,
0.212158203125
],
[
0.69140625,
0.1976318359375,
0.74267578125,
0.1976318359375,
0.74267578125,
0.212158203125,
0.69140625,
0.212158203125
],
[
0.77392578125,
0.1976318359375,
0.84326171875,
0.1976318359375,
0.84326171875,
0.212158203125,
0.77392578125,
0.212158203125
],
[
0.27685546875,
0.1988525390625,
0.34228515625,
0.1988525390625,
0.34228515625,
0.210205078125,
0.27685546875,
0.210205078125
],
[
0.158447265625,
0.1988525390625,
0.233642578125,
0.1988525390625,
0.233642578125,
0.210205078125,
0.158447265625,
0.210205078125
],
[
0.44677734375,
0.1988525390625,
0.51708984375,
0.1988525390625,
0.51708984375,
0.210205078125,
0.44677734375,
0.210205078125
],
[
0.611328125,
0.1988525390625,
0.6748046875,
0.1988525390625,
0.6748046875,
0.210205078125,
0.611328125,
0.210205078125
],
[
0.15283203125,
0.2139892578125,
0.265625,
0.2139892578125,
0.265625,
0.2279052734375,
0.15283203125,
0.2279052734375
],
[
0.273681640625,
0.2139892578125,
0.303955078125,
0.2139892578125,
0.303955078125,
0.2266845703125,
0.273681640625,
0.2266845703125
],
[
0.358642578125,
0.2139892578125,
0.386474609375,
0.2139892578125,
0.386474609375,
0.2266845703125,
0.358642578125,
0.2266845703125
],
[
0.44677734375,
0.2139892578125,
0.476318359375,
0.2139892578125,
0.476318359375,
0.2266845703125,
0.44677734375,
0.2266845703125
],
[
0.52783203125,
0.2139892578125,
0.55810546875,
0.2139892578125,
0.55810546875,
0.2266845703125,
0.52783203125,
0.2266845703125
],
[
0.611328125,
0.2139892578125,
0.63818359375,
0.2139892578125,
0.63818359375,
0.2266845703125,
0.611328125,
0.2266845703125
],
[
0.69287109375,
0.2139892578125,
0.7197265625,
0.2139892578125,
0.7197265625,
0.2266845703125,
0.69287109375,
0.2266845703125
],
[
0.77392578125,
0.2139892578125,
0.8037109375,
0.2139892578125,
0.8037109375,
0.2266845703125,
0.77392578125,
0.2266845703125
],
[
0.443603515625,
0.2235107421875,
0.51123046875,
0.2259521484375,
0.51123046875,
0.240478515625,
0.443603515625,
0.238037109375
],
[
0.525390625,
0.2235107421875,
0.59326171875,
0.2259521484375,
0.59326171875,
0.240478515625,
0.525390625,
0.238037109375
],
[
0.60888671875,
0.2247314453125,
0.6806640625,
0.2247314453125,
0.6806640625,
0.2392578125,
0.60888671875,
0.2392578125
],
[
0.689453125,
0.2235107421875,
0.76318359375,
0.2259521484375,
0.76318359375,
0.240478515625,
0.689453125,
0.238037109375
],
[
0.275390625,
0.2266845703125,
0.340576171875,
0.2266845703125,
0.340576171875,
0.238037109375,
0.275390625,
0.238037109375
],
[
0.158447265625,
0.2266845703125,
0.2442626953125,
0.2266845703125,
0.2442626953125,
0.238037109375,
0.158447265625,
0.238037109375
],
[
0.358642578125,
0.2266845703125,
0.42236328125,
0.2266845703125,
0.42236328125,
0.238037109375,
0.358642578125,
0.238037109375
],
[
0.77392578125,
0.2266845703125,
0.84228515625,
0.2266845703125,
0.84228515625,
0.238037109375,
0.77392578125,
0.238037109375
],
[
0.158447265625,
0.238037109375,
0.233642578125,
0.238037109375,
0.233642578125,
0.25,
0.158447265625,
0.25
],
[
0.27685546875,
0.238037109375,
0.339111328125,
0.238037109375,
0.339111328125,
0.25,
0.27685546875,
0.25
],
[
0.358642578125,
0.238037109375,
0.420654296875,
0.238037109375,
0.420654296875,
0.25,
0.358642578125,
0.25
],
[
0.44775390625,
0.238037109375,
0.5087890625,
0.238037109375,
0.5087890625,
0.25,
0.44775390625,
0.25
],
[
0.529296875,
0.238037109375,
0.59228515625,
0.238037109375,
0.59228515625,
0.25,
0.529296875,
0.25
],
[
0.611328125,
0.238037109375,
0.67333984375,
0.238037109375,
0.67333984375,
0.25,
0.611328125,
0.25
],
[
0.69384765625,
0.238037109375,
0.74169921875,
0.238037109375,
0.74169921875,
0.25,
0.69384765625,
0.25
],
[
0.77392578125,
0.236083984375,
0.84326171875,
0.236083984375,
0.84326171875,
0.250732421875,
0.77392578125,
0.250732421875
],
[
0.453369140625,
0.255126953125,
0.5498046875,
0.255126953125,
0.5498046875,
0.26953125,
0.453369140625,
0.26953125
],
[
0.35546875,
0.26513671875,
0.38818359375,
0.26513671875,
0.38818359375,
0.281005859375,
0.35546875,
0.281005859375
],
[
0.60693359375,
0.26513671875,
0.640625,
0.26513671875,
0.640625,
0.281005859375,
0.60693359375,
0.281005859375
],
[
0.69287109375,
0.266357421875,
0.72216796875,
0.266357421875,
0.72216796875,
0.279052734375,
0.69287109375,
0.279052734375
],
[
0.4453125,
0.267578125,
0.4755859375,
0.267578125,
0.4755859375,
0.27978515625,
0.4453125,
0.27978515625
],
[
0.273681640625,
0.267578125,
0.3056640625,
0.267578125,
0.3056640625,
0.279052734375,
0.273681640625,
0.279052734375
],
[
0.52783203125,
0.267578125,
0.55615234375,
0.267578125,
0.55615234375,
0.27978515625,
0.52783203125,
0.27978515625
],
[
0.77392578125,
0.267578125,
0.8037109375,
0.267578125,
0.8037109375,
0.279052734375,
0.77392578125,
0.279052734375
],
[
0.273681640625,
0.276611328125,
0.339111328125,
0.279052734375,
0.339111328125,
0.292236328125,
0.273681640625,
0.290283203125
],
[
0.1544189453125,
0.268310546875,
0.1953125,
0.266357421875,
0.1968994140625,
0.279052734375,
0.156005859375,
0.281005859375
],
[
0.35546875,
0.276611328125,
0.428955078125,
0.279052734375,
0.428955078125,
0.292236328125,
0.35546875,
0.290283203125
],
[
0.443603515625,
0.276611328125,
0.51708984375,
0.279052734375,
0.51708984375,
0.29345703125,
0.443603515625,
0.291748046875
],
[
0.689453125,
0.276611328125,
0.75732421875,
0.279052734375,
0.75732421875,
0.29345703125,
0.689453125,
0.291748046875
],
[
0.529296875,
0.279052734375,
0.59228515625,
0.279052734375,
0.59228515625,
0.290283203125,
0.529296875,
0.290283203125
],
[
0.158447265625,
0.27783203125,
0.2459716796875,
0.27783203125,
0.2459716796875,
0.2890625,
0.158447265625,
0.2890625
],
[
0.611328125,
0.279052734375,
0.67333984375,
0.279052734375,
0.67333984375,
0.290283203125,
0.611328125,
0.290283203125
],
[
0.77392578125,
0.279052734375,
0.83837890625,
0.279052734375,
0.83837890625,
0.290283203125,
0.77392578125,
0.290283203125
],
[
0.27294921875,
0.28857421875,
0.34228515625,
0.290283203125,
0.34228515625,
0.304931640625,
0.27294921875,
0.302978515625
],
[
0.443603515625,
0.28857421875,
0.5185546875,
0.290283203125,
0.5185546875,
0.304931640625,
0.443603515625,
0.302978515625
],
[
0.525390625,
0.28857421875,
0.5986328125,
0.290283203125,
0.5986328125,
0.304931640625,
0.525390625,
0.302978515625
],
[
0.609375,
0.290283203125,
0.6787109375,
0.290283203125,
0.6787109375,
0.3017578125,
0.609375,
0.3017578125
],
[
0.358642578125,
0.291748046875,
0.4345703125,
0.291748046875,
0.4345703125,
0.3017578125,
0.358642578125,
0.3017578125
],
[
0.16015625,
0.291748046875,
0.233642578125,
0.291748046875,
0.233642578125,
0.3017578125,
0.16015625,
0.3017578125
],
[
0.69287109375,
0.291748046875,
0.76123046875,
0.291748046875,
0.76123046875,
0.302978515625,
0.69287109375,
0.302978515625
],
[
0.77392578125,
0.291748046875,
0.8486328125,
0.291748046875,
0.8486328125,
0.302978515625,
0.77392578125,
0.302978515625
],
[
0.60888671875,
0.304931640625,
0.6396484375,
0.304931640625,
0.6396484375,
0.32080078125,
0.60888671875,
0.32080078125
],
[
0.69140625,
0.304931640625,
0.72216796875,
0.304931640625,
0.72216796875,
0.32080078125,
0.69140625,
0.32080078125
],
[
0.35693359375,
0.30615234375,
0.386474609375,
0.30615234375,
0.386474609375,
0.318115234375,
0.35693359375,
0.318115234375
],
[
0.275390625,
0.30615234375,
0.303955078125,
0.30615234375,
0.303955078125,
0.318115234375,
0.275390625,
0.318115234375
],
[
0.44677734375,
0.30615234375,
0.4755859375,
0.30615234375,
0.4755859375,
0.318115234375,
0.44677734375,
0.318115234375
],
[
0.52783203125,
0.30615234375,
0.55810546875,
0.30615234375,
0.55810546875,
0.318115234375,
0.52783203125,
0.318115234375
],
[
0.775390625,
0.30615234375,
0.80224609375,
0.30615234375,
0.80224609375,
0.318115234375,
0.775390625,
0.318115234375
],
[
0.273681640625,
0.317626953125,
0.34228515625,
0.317626953125,
0.34228515625,
0.33203125,
0.273681640625,
0.33203125
],
[
0.156005859375,
0.306884765625,
0.2484130859375,
0.306884765625,
0.2484130859375,
0.318115234375,
0.156005859375,
0.318115234375
],
[
0.35546875,
0.317626953125,
0.4248046875,
0.317626953125,
0.4248046875,
0.33203125,
0.35546875,
0.33203125
],
[
0.609375,
0.317626953125,
0.6748046875,
0.317626953125,
0.6748046875,
0.328857421875,
0.609375,
0.328857421875
],
[
0.4453125,
0.318115234375,
0.51708984375,
0.318115234375,
0.51708984375,
0.330078125,
0.4453125,
0.330078125
],
[
0.158447265625,
0.318115234375,
0.2459716796875,
0.318115234375,
0.2459716796875,
0.330078125,
0.158447265625,
0.330078125
],
[
0.52783203125,
0.318115234375,
0.59716796875,
0.318115234375,
0.59716796875,
0.330078125,
0.52783203125,
0.330078125
],
[
0.69287109375,
0.318115234375,
0.75634765625,
0.318115234375,
0.75634765625,
0.330078125,
0.69287109375,
0.330078125
],
[
0.77392578125,
0.318115234375,
0.84228515625,
0.318115234375,
0.84228515625,
0.330078125,
0.77392578125,
0.330078125
],
[
0.158447265625,
0.330078125,
0.233642578125,
0.330078125,
0.233642578125,
0.341552734375,
0.158447265625,
0.341552734375
],
[
0.275390625,
0.330078125,
0.339111328125,
0.330078125,
0.339111328125,
0.341552734375,
0.275390625,
0.341552734375
],
[
0.358642578125,
0.330078125,
0.420654296875,
0.330078125,
0.420654296875,
0.341552734375,
0.358642578125,
0.341552734375
],
[
0.44677734375,
0.330078125,
0.515625,
0.330078125,
0.515625,
0.341552734375,
0.44677734375,
0.341552734375
],
[
0.52783203125,
0.330078125,
0.59716796875,
0.330078125,
0.59716796875,
0.341552734375,
0.52783203125,
0.341552734375
],
[
0.609375,
0.330078125,
0.67333984375,
0.330078125,
0.67333984375,
0.341552734375,
0.609375,
0.341552734375
],
[
0.69287109375,
0.330078125,
0.74169921875,
0.330078125,
0.74169921875,
0.341552734375,
0.69287109375,
0.341552734375
],
[
0.775390625,
0.330078125,
0.84228515625,
0.330078125,
0.84228515625,
0.341552734375,
0.775390625,
0.341552734375
],
[
0.35546875,
0.343505859375,
0.386474609375,
0.343505859375,
0.386474609375,
0.359130859375,
0.35546875,
0.359130859375
],
[
0.52685546875,
0.343505859375,
0.5595703125,
0.343505859375,
0.5595703125,
0.359130859375,
0.52685546875,
0.359130859375
],
[
0.611328125,
0.3447265625,
0.6396484375,
0.3447265625,
0.6396484375,
0.359130859375,
0.611328125,
0.359130859375
],
[
0.69287109375,
0.3447265625,
0.72119140625,
0.3447265625,
0.72119140625,
0.359130859375,
0.69287109375,
0.359130859375
],
[
0.275390625,
0.34521484375,
0.303955078125,
0.34521484375,
0.303955078125,
0.356689453125,
0.275390625,
0.356689453125
],
[
0.1544189453125,
0.3447265625,
0.2646484375,
0.3447265625,
0.2646484375,
0.35791015625,
0.1544189453125,
0.35791015625
],
[
0.44677734375,
0.34521484375,
0.4755859375,
0.34521484375,
0.4755859375,
0.35791015625,
0.44677734375,
0.35791015625
],
[
0.77392578125,
0.34521484375,
0.80224609375,
0.34521484375,
0.80224609375,
0.356689453125,
0.77392578125,
0.356689453125
],
[
0.158447265625,
0.356689453125,
0.2459716796875,
0.356689453125,
0.2459716796875,
0.36865234375,
0.158447265625,
0.36865234375
],
[
0.275390625,
0.356689453125,
0.340576171875,
0.356689453125,
0.340576171875,
0.36865234375,
0.275390625,
0.36865234375
],
[
0.358642578125,
0.356689453125,
0.42236328125,
0.356689453125,
0.42236328125,
0.36865234375,
0.358642578125,
0.36865234375
],
[
0.44677734375,
0.356689453125,
0.5107421875,
0.356689453125,
0.5107421875,
0.36865234375,
0.44677734375,
0.36865234375
],
[
0.52783203125,
0.356689453125,
0.59228515625,
0.356689453125,
0.59228515625,
0.36865234375,
0.52783203125,
0.36865234375
],
[
0.609375,
0.356689453125,
0.6806640625,
0.356689453125,
0.6806640625,
0.36865234375,
0.609375,
0.36865234375
],
[
0.69287109375,
0.356689453125,
0.76123046875,
0.356689453125,
0.76123046875,
0.36865234375,
0.69287109375,
0.36865234375
],
[
0.77294921875,
0.35595703125,
0.8408203125,
0.35595703125,
0.8408203125,
0.367431640625,
0.77294921875,
0.367431640625
],
[
0.158447265625,
0.36865234375,
0.233642578125,
0.36865234375,
0.233642578125,
0.380126953125,
0.158447265625,
0.380126953125
],
[
0.275390625,
0.36865234375,
0.339111328125,
0.36865234375,
0.339111328125,
0.380126953125,
0.275390625,
0.380126953125
],
[
0.358642578125,
0.36865234375,
0.42236328125,
0.36865234375,
0.42236328125,
0.380126953125,
0.358642578125,
0.380126953125
],
[
0.44775390625,
0.36865234375,
0.5087890625,
0.36865234375,
0.5087890625,
0.380126953125,
0.44775390625,
0.380126953125
],
[
0.529296875,
0.36865234375,
0.59228515625,
0.36865234375,
0.59228515625,
0.380126953125,
0.529296875,
0.380126953125
],
[
0.611328125,
0.36865234375,
0.6748046875,
0.36865234375,
0.6748046875,
0.380126953125,
0.611328125,
0.380126953125
],
[
0.69384765625,
0.36865234375,
0.74169921875,
0.36865234375,
0.74169921875,
0.380126953125,
0.69384765625,
0.380126953125
],
[
0.77392578125,
0.367431640625,
0.8408203125,
0.367431640625,
0.8408203125,
0.37890625,
0.77392578125,
0.37890625
],
[
0.1446533203125,
0.391357421875,
0.8544921875,
0.391357421875,
0.8544921875,
0.40478515625,
0.1446533203125,
0.40478515625
],
[
0.1422119140625,
0.40478515625,
0.8544921875,
0.404052734375,
0.8544921875,
0.41845703125,
0.1422119140625,
0.419189453125
],
[
0.1446533203125,
0.41845703125,
0.8388671875,
0.41845703125,
0.8388671875,
0.431884765625,
0.1446533203125,
0.431884765625
],
[
0.143798828125,
0.466552734375,
0.85302734375,
0.46728515625,
0.85302734375,
0.481689453125,
0.143798828125,
0.48095703125
],
[
0.1446533203125,
0.481689453125,
0.8544921875,
0.481689453125,
0.8544921875,
0.49560546875,
0.1446533203125,
0.49560546875
],
[
0.146240234375,
0.49560546875,
0.8544921875,
0.49560546875,
0.8544921875,
0.5087890625,
0.146240234375,
0.5087890625
],
[
0.1446533203125,
0.5087890625,
0.486083984375,
0.5087890625,
0.486083984375,
0.52294921875,
0.1446533203125,
0.52294921875
],
[
0.295654296875,
0.5263671875,
0.701171875,
0.53076171875,
0.701171875,
0.548828125,
0.295654296875,
0.54443359375
],
[
0.83251953125,
0.52978515625,
0.85693359375,
0.52978515625,
0.85693359375,
0.54345703125,
0.83251953125,
0.54345703125
],
[
0.146240234375,
0.556640625,
0.85546875,
0.556640625,
0.85546875,
0.57080078125,
0.146240234375,
0.57080078125
],
[
0.1446533203125,
0.57080078125,
0.8544921875,
0.57080078125,
0.8544921875,
0.583984375,
0.1446533203125,
0.583984375
],
[
0.1446533203125,
0.58544921875,
0.8544921875,
0.58544921875,
0.8544921875,
0.5986328125,
0.1446533203125,
0.5986328125
],
[
0.143798828125,
0.59765625,
0.85546875,
0.5986328125,
0.85546875,
0.61279296875,
0.143798828125,
0.6123046875
],
[
0.1446533203125,
0.611328125,
0.8544921875,
0.611328125,
0.8544921875,
0.625,
0.1446533203125,
0.625
],
[
0.1446533203125,
0.62548828125,
0.8544921875,
0.62548828125,
0.8544921875,
0.6396484375,
0.1446533203125,
0.6396484375
],
[
0.1446533203125,
0.63818359375,
0.85302734375,
0.63818359375,
0.85302734375,
0.65283203125,
0.1446533203125,
0.65283203125
],
[
0.1446533203125,
0.65380859375,
0.830078125,
0.65380859375,
0.830078125,
0.66748046875,
0.1446533203125,
0.66748046875
],
[
0.25732421875,
0.671875,
0.73779296875,
0.67529296875,
0.73779296875,
0.693359375,
0.25732421875,
0.689453125
],
[
0.8310546875,
0.673828125,
0.85693359375,
0.673828125,
0.85693359375,
0.6875,
0.8310546875,
0.6875
],
[
0.146240234375,
0.708984375,
0.158447265625,
0.708984375,
0.158447265625,
0.71826171875,
0.146240234375,
0.71826171875
],
[
0.17236328125,
0.70849609375,
0.283447265625,
0.70849609375,
0.283447265625,
0.7216796875,
0.17236328125,
0.7216796875
],
[
0.143798828125,
0.736328125,
0.8544921875,
0.7373046875,
0.8544921875,
0.751953125,
0.143798828125,
0.75048828125
],
[
0.1446533203125,
0.751953125,
0.8544921875,
0.751953125,
0.8544921875,
0.76513671875,
0.1446533203125,
0.76513671875
],
[
0.1446533203125,
0.7666015625,
0.8544921875,
0.7666015625,
0.8544921875,
0.77978515625,
0.1446533203125,
0.77978515625
],
[
0.1446533203125,
0.77978515625,
0.8544921875,
0.77978515625,
0.8544921875,
0.79345703125,
0.1446533203125,
0.79345703125
],
[
0.143798828125,
0.79345703125,
0.8095703125,
0.79345703125,
0.8095703125,
0.806640625,
0.143798828125,
0.806640625
],
[
0.1446533203125,
0.82470703125,
0.1690673828125,
0.82470703125,
0.1690673828125,
0.8359375,
0.1446533203125,
0.8359375
],
[
0.178955078125,
0.82275390625,
0.32177734375,
0.82373046875,
0.32177734375,
0.83837890625,
0.178955078125,
0.8369140625
],
[
0.1446533203125,
0.8486328125,
0.8544921875,
0.8486328125,
0.8544921875,
0.86279296875,
0.1446533203125,
0.86279296875
],
[
0.143798828125,
0.8623046875,
0.85302734375,
0.861328125,
0.85302734375,
0.87548828125,
0.143798828125,
0.876953125
],
[
0.1683349609375,
0.88623046875,
0.85546875,
0.88623046875,
0.85546875,
0.8994140625,
0.1683349609375,
0.8994140625
],
[
0.146240234375,
0.90087890625,
0.650390625,
0.90087890625,
0.650390625,
0.9111328125,
0.146240234375,
0.9111328125
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Preprint.",
"Move-1p",
"Move-2p",
"Move-3p",
"Model",
"AVG.",
"n=3",
"n= 4",
"n=4",
"n=3",
"n=4",
"(Zero-Shot)",
"n=3",
"93.3",
"90.0",
"46.7",
"35.0",
"45.0",
"59.7",
"48.3",
"gpt-4",
"+S2L w/model",
"90.0 (3.3)",
"91.7",
"(+1.7)",
"48.3 (-)",
"56.7(+10.0)",
"38.3 (+3.3)",
"48.3 (+3.3)",
"62.2(+2.5)",
"96.7 (+3.4)",
"60.0 (+15.0)",
"81.6 (+21.9)",
"100.0(+10.0)",
"+S2Lw/rule",
"88.3 (+40.0)",
"96.7 (+50.0)",
"48.3(+13.3)",
"gpt-3.5-turbo",
"68.3",
"71.7",
"11.7",
"25.0",
"23.3",
"26.7",
"37.8",
"80.0 (+11.7)",
"75.0 (+3.3)",
"26.6 (+1.6)",
"33.3 (+10.0)",
"31.7(+5.0)",
"46.4(+8.6)",
"31.6 (+19.9)",
"+S2Lw/model",
"88.3(+16.6)",
"31.7(+6.7)",
"26.7 (-)",
"44.7(+6.9)",
"71.6 (+3.3)",
"+S2L w/rule",
"25.0(+13.3)",
"25.0 (+1.7)",
"openchat-3.5-7b",
"61.7",
"71.7",
"15.0",
"21.6",
"11.7",
"11.7",
"32.2",
"23.3(+8.3)",
"25.0(+3.4)",
"25.0 (+13.3)",
"21.7(+10.0)",
"68.3 (+6.6)",
"+S2L w/model",
"78.3(+6.6)",
"40.3(+8.1)",
"+S2L w/rule",
"63.3 (+1.6)",
"75.0 (+3.3)",
"16.6 (+1.6)",
"18.3 (-3.3)",
"15.0 (+3.3)",
"11.7 (-)",
"33.3(+1.1)",
"(Zero-Shot-CoT)",
"90.0",
"36.7",
"41.7",
"55.0",
"93.3",
"50.0",
"61.1",
"95.0(+1.7)",
"gpt-4",
"91.7(+1.7)",
"38.3(16.7)",
"48.3 (+6.6)",
"46.7 (3.3)",
"+S2Lw/model",
"43.3 (+6.6)",
"60.5 (0.6)",
"96.7(+3.4)",
"86.7(+31.7)",
"88.3(+38.3)",
"50.0(+13.3)",
"100.0(+10.0)",
"+S2Lw/rule",
"66.7 (+25.0)",
"81.4(+20.3)",
"26.7",
"25.0",
"78.3",
"76.7",
"15.0",
"25.0",
"41.1",
"75.0 (1.7)",
"gpt-3.5-turbo",
"71.7 (-6.6)",
"35.0(+8.3)",
"30.0 (+15.0)",
"+S2L w/model",
"35.0(+10.0)",
"30.0(+5.0)",
"46.1(+5.0)",
"+S2L w/ rule",
"75.0 (1.7)",
"86.7 (+8.4)",
"30.0 (+15.0)",
"36.7 (+11.7)",
"21.7 (-5.0)",
"25.0 (-)",
"45.8 (+4.7)",
"71.7",
"21.6",
"11.7",
"11.7",
"61.7",
"openchat-3.5-7b",
"15.0",
"32.2",
"+S2L w/model",
"68.3 (+6.6)",
"78.3 (+6.6)",
"23.3(+8.3)",
"25.0(+3.4)",
"25.0(+13.3)",
"21.7 (+10.0)",
"40.3 (+8.1)",
"+S2Lw/rule",
"63.3 (+1.6)",
"75.0 (+3.3)",
"16.6(+1.6)",
"18.3 (3.3)",
"15.0(+3.3)",
"11.7(-)",
"33.3(+1.1)",
"Table 2: Results for three subtasks on 1D-ARC. n indicates the number of given input-output pairs for finding",
"the patterns.The values in parentheses indicate the difference between the baseline and our results. w/ model",
"denotes conversion with LLMs, and w/ rule denotes conversion with manually designed rules using codes.",
"Direct Substitution. The first way of utilization is to directly substitute the symbol-based representation s;",
"with language-based representation l,. To some extent, l; can be regarded as the language-based equivalent of",
"Si. Thus we can use them to replace the symbol-based representations for both the question and ground-truth",
"label. The response by using S2L can be written as:",
"Rg21 = M(qu), ql = qfos = qfo(s...m) = qfos... fosm.",
"(3)",
"Concatenation. However, in some other tasks, the generated l, by LLMs may not always be a perfect substitu-",
"tion or convey complete information of si. This can occur for two reasons: 1) l, might be incorrect due to the",
"undesired output formats, misleading content, or noisy context; and 2) l; could lose some information during",
"S2L conversion. For instance, the ground truth might be a span-based abbreviation for table understanding (e.g.",
"the\"swE\"in table\"rank |nation niIswE\"),meaning that the converted l; as a full name (e.g.,\"Sweden\")",
"may not exactly match the final answer. Therefore, the second method uses both the original symbol-based",
"representation s; and the language-based representation l, as the combined input. This approach allows the",
"LLMs M to reason on questions that include rich contextual information from two distinct perspectives:",
"Rs21=M(qsot),qs0l=qsiΦl..sm@lm=qsio(fos1....sm@(fosm)",
"(4)",
"4",
"Experiments",
"To assess the performance of the S2L framework, we conduct six categories of symbol-related problems with",
"eight specific tasks as shown in Table 1, varying from symbol-only inductive abstract reasoning to traditional",
"sentiment analysis in social media. As for LLMs, we evaluate both API-based and open-source models,",
"including GPT-4 (OpenAI, 2023b), ChatGPT (OpenAI, 2022), and OpenChat-7b (Wang et al., 2023a). To",
"ensure the reproducibility of the responses generated by LLMs, we set the decoding temperature as O.",
"4.1",
"AbstractReasoning",
"Abstract reasoning (Webb et al., 2023; Gendron et al., 2023; Wang et al., 2023b) is a type of task that involves",
" summarizing patterns from limited observations. We conduct experiments on subtasks from the 1D-ARC",
"'Specifically,we use the versions of gpt-4-0613 and gpt-3.5-turbo-1106 released by OpenAl,and",
"openchat_3.5releasedinhttps://huggingface.co/openchat/openchat_3.5",
"5"
] |
{
"arxiv_id": "2401.11725",
"language": "en",
"page_idx": 4
}
|
||
1911.08348-en-0
|
1911.08348-en-0.jpg
|
$\textbf{Live Face De- Identification in Video}$
Yaniv Taigman
Facebook AI Research
yanivefb.com
Oran Gafni, Lior Wolf
Facebook AI Research and Tel-Aviv University
$\{$oran,wolf$\}\textcircled{\mathrm{efb.com}}$
$ଶ୍ନି$
## $\mathbf{Abstract}$
We propose a method for face de-identification that en-
$\begin{array}{lll}\geq&ablesfully&automatic&video&modification&at&highframe&rates.\\0&The&goal&is&to&maximally&decorrelate&the&identity,&while&hav-\\\sum&ing&the&perception&(pose,illumination&and&expression)fxed.\end{array}$
We achieve this by a novel feed-forward encoder-decoder network architecture that is conditioned on the high-level representation of a person's facial image. The network is global, in the sense that it does not need to be retrained for a given video or for a given identity, and it creates natural looking image sequences with little distortion in time.
## $1. \textbf{Introduction}$
In consumer image and video applications, the face has a unique importance that stands out from all other objects. For example, face recognition (detection followed by identification) is perhaps much more widely applicable than any other object recognition (categorization, detection, or instance identification) in consumer images. Similarly, putting aside image processing operators that are applied to the entire frame, face filters remain the most popular filters for consumer video. Since face technology is both useful and impactful, it also raises many ethical concerns. Face recognition can lead to loss of privacy and face replacement technology may be misused to create misleading videos.
In this work, we focus on video de-identification, which is a video filtering application that both requires a technological leap over the current state-of-the-art, and is benign in nature. This application requires the creation of a video of a similar looking person, such that the perceived identity is changed. This allows, for example, the user to leave a natural-looking video message in a public forum in an anonymous way, that would presumably prevent face recognition technology from recognizing them.
Video de-identification is a challenging task. The video needs to be modified in a seamless way, without causing flickering or other visual artifacts and distortions, such that the identity is changed, while all other factors remain identical, see Fig. 1. These factors include pose, expression, lip
<FigureHere>
Figure 1. De-identification video results demonstrated on a variety
of poses, expressions, illumination conditions and occlusions. Pairs of the source frame (first row) and the output frame (second row) are shown. The high-level features (e.g. nose, eyes, eyebrows and mouth) are altered, while the pose, expression, lip articulation, illumination, and skin tone are preserved.
positioning (for unaltered speech), occlusion, illumination
and shadow, and their dynamics.
In contrast to the literature methods, which are limited to still images and often swap a given face with a dataset face, our method handles video and generates de novo faces. Our experiments show convincing performance for unconstrained videos, producing natural looking videos. The person in the rendered video has a similar appearance to the person in the original video. However, a state-of-the-art face-recognition network fails to identify the person. A similar experiment shows that humans cannot identify the generated face, even without time constraints.
Our results would not have been possible, without a host of novelties. We introduce a novel encoder-decoder architecture, in which we concatenate to the latent space the activations of the representation layer of a network trained to perform face recognition. As far as we know, this is the first time that a representation from an existing classifier network is used to augment an autoencoder, which enables the feed-forward treatment of new persons, unseen during training. In addition, this is the fırst work to introduce a new
|
[
[
0.306396484375,
0.1337890625,
0.6640625,
0.1337890625,
0.6640625,
0.1483154296875,
0.306396484375,
0.1483154296875
],
[
0.271240234375,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1976318359375,
0.271240234375,
0.1976318359375
],
[
0.6396484375,
0.18310546875,
0.75732421875,
0.18310546875,
0.75732421875,
0.1976318359375,
0.6396484375,
0.1976318359375
],
[
0.1846923828125,
0.2008056640625,
0.5537109375,
0.2008056640625,
0.5537109375,
0.2139892578125,
0.1846923828125,
0.2139892578125
],
[
0.60693359375,
0.2008056640625,
0.78759765625,
0.2008056640625,
0.78759765625,
0.2139892578125,
0.60693359375,
0.2139892578125
],
[
0.293212890625,
0.2215576171875,
0.449462890625,
0.2215576171875,
0.449462890625,
0.231689453125,
0.293212890625,
0.231689453125
],
[
0.6435546875,
0.2203369140625,
0.75341796875,
0.2203369140625,
0.75341796875,
0.2342529296875,
0.6435546875,
0.2342529296875
],
[
0.0236968994140625,
0.25830078125,
0.058013916015625,
0.25830078125,
0.058013916015625,
0.322509765625,
0.0236968994140625,
0.322509765625
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.101318359375,
0.3017578125,
0.46728515625,
0.3017578125,
0.46728515625,
0.31494140625,
0.101318359375,
0.31494140625
],
[
0.0269622802734375,
0.311279296875,
0.058013916015625,
0.3125,
0.053924560546875,
0.371826171875,
0.0228729248046875,
0.37060546875
],
[
0.08087158203125,
0.317626953125,
0.4697265625,
0.317626953125,
0.4697265625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.08251953125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.3447265625,
0.08251953125,
0.3447265625
],
[
0.08087158203125,
0.347900390625,
0.471435546875,
0.347900390625,
0.471435546875,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.02777099609375,
0.365478515625,
0.053924560546875,
0.365478515625,
0.053924560546875,
0.386474609375,
0.02777099609375,
0.386474609375
],
[
0.08251953125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.375,
0.08251953125,
0.375
],
[
0.03350830078125,
0.38330078125,
0.037567138671875,
0.38330078125,
0.037567138671875,
0.389404296875,
0.03350830078125,
0.389404296875
],
[
0.08251953125,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.389404296875,
0.08251953125,
0.389404296875
],
[
0.08087158203125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.406005859375,
0.08087158203125,
0.406005859375
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.4462890625,
0.02532958984375,
0.4462890625
],
[
0.705078125,
0.401611328125,
0.71337890625,
0.401611328125,
0.71337890625,
0.40478515625,
0.705078125,
0.40478515625
],
[
0.08087158203125,
0.40771484375,
0.4697265625,
0.40771484375,
0.4697265625,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.07843017578125,
0.42236328125,
0.4697265625,
0.421630859375,
0.4697265625,
0.436279296875,
0.07843017578125,
0.4375
],
[
0.080078125,
0.438232421875,
0.431396484375,
0.438232421875,
0.431396484375,
0.451904296875,
0.080078125,
0.451904296875
],
[
0.0310516357421875,
0.4443359375,
0.056365966796875,
0.4443359375,
0.056365966796875,
0.48291015625,
0.0310516357421875,
0.48291015625
],
[
0.50244140625,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.4765625,
0.50244140625,
0.4765625
],
[
0.08251953125,
0.48291015625,
0.20751953125,
0.48291015625,
0.20751953125,
0.49755859375,
0.08251953125,
0.49755859375
],
[
0.50341796875,
0.479736328125,
0.890625,
0.479736328125,
0.890625,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.0236968994140625,
0.496826171875,
0.058013916015625,
0.496826171875,
0.058013916015625,
0.5888671875,
0.0236968994140625,
0.5888671875
],
[
0.50341796875,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.50390625,
0.50341796875,
0.50390625
],
[
0.09722900390625,
0.50830078125,
0.4697265625,
0.50830078125,
0.4697265625,
0.521484375,
0.09722900390625,
0.521484375
],
[
0.50244140625,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.5048828125,
0.521484375,
0.890625,
0.521484375,
0.890625,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.07843017578125,
0.537109375,
0.4697265625,
0.53857421875,
0.4697265625,
0.55322265625,
0.07843017578125,
0.5517578125
],
[
0.50244140625,
0.5341796875,
0.74853515625,
0.5341796875,
0.74853515625,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.07843017578125,
0.55322265625,
0.468017578125,
0.5537109375,
0.468017578125,
0.568359375,
0.07843017578125,
0.5673828125
],
[
0.08087158203125,
0.57080078125,
0.46728515625,
0.57080078125,
0.46728515625,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.5009765625,
0.5791015625,
0.89208984375,
0.57763671875,
0.89208984375,
0.59228515625,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.58544921875,
0.4697265625,
0.58544921875,
0.4697265625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.080078125,
0.599609375,
0.4697265625,
0.599609375,
0.4697265625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.5947265625,
0.71728515625,
0.5947265625,
0.71728515625,
0.60791015625,
0.50244140625,
0.60791015625
],
[
0.03350830078125,
0.611328125,
0.037567138671875,
0.611328125,
0.037567138671875,
0.6162109375,
0.03350830078125,
0.6162109375
],
[
0.080078125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.080078125,
0.626953125
],
[
0.52294921875,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.52294921875,
0.625
],
[
0.50244140625,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.02777099609375,
0.6142578125,
0.052276611328125,
0.6142578125,
0.052276611328125,
0.63623046875,
0.02777099609375,
0.63623046875
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.0310516357421875,
0.63623046875,
0.0555419921875,
0.63623046875,
0.0555419921875,
0.681640625,
0.0310516357421875,
0.681640625
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.66943359375,
0.50341796875,
0.66943359375
],
[
0.50244140625,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.0318603515625,
0.66357421875,
0.037567138671875,
0.66357421875,
0.037567138671875,
0.6787109375,
0.0318603515625,
0.6787109375
],
[
0.02532958984375,
0.6767578125,
0.051483154296875,
0.67236328125,
0.060455322265625,
0.705078125,
0.035125732421875,
0.708984375
],
[
0.080078125,
0.6748046875,
0.42236328125,
0.6748046875,
0.42236328125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.09722900390625,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.09722900390625,
0.7041015625
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.08087158203125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.08087158203125,
0.7216796875,
0.4697265625,
0.7216796875,
0.4697265625,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.08087158203125,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.50244140625,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.75,
0.6650390625,
0.75,
0.6650390625,
0.76025390625,
0.50341796875,
0.76025390625
],
[
0.080078125,
0.7646484375,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77783203125
],
[
0.521484375,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.521484375,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.5009765625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.1993408203125,
0.81103515625,
0.1993408203125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.09722900390625,
0.82470703125,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.09722900390625,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.471435546875,
0.85595703125,
0.471435546875,
0.869140625,
0.080078125,
0.869140625
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.5009765625,
0.88525390625
],
[
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
],
[
0.48193359375,
0.92822265625,
0.490234375,
0.92822265625,
0.490234375,
0.93310546875,
0.48193359375,
0.93310546875
]
] |
[
"Live Face De-Identification in Video",
"Oran Gafni,Lior Wolf",
"Yaniv Taigman",
"Facebook AI Research and Tel-Aviv University",
"Facebook AI Research",
"[oran,wolf]@fb.com",
"yaniv@fb.com",
"610",
"Abstract",
"We propose a method for face de-identification that en",
"AON",
"ables fully automatic video modification at high frame rates.",
"Thegoal isto maximallydecorrelatetheidentity,whilehav",
"ing the perception (pose, illumination and expression) fixed.",
"9",
"We achieve thisby a novel feed-forward encoder-decoder",
"",
"network architecture that is conditioned on the high-level",
"representation of a person's facial image. The network is",
"[DT",
"-",
"global, in the sense that it does not need to be retrained for",
"a given video or for a given identity, and it creates natural",
"looking image sequences with little distortion in time.",
"",
"Figure 1. De-identification video results demonstrated on a variety",
"1. Introduction",
"of poses,expressions,illumination conditions and occlusions.Pairs",
"78700",
"ofthesourceframe(firstrow)andtheoutputframe(secondrow)",
"In consumer image and video applications, the face has",
"are shown. The high-level features (e.g. nose, eyes, eyebrows",
"a unique importance that stands out from all other objects.",
"andmouth)arealtered,whilethepose,expression,lip articulation",
"For example, face recognition (detection followed by identi-",
"illumination, and skin tone are preserved.",
"fication) is perhaps much more widely applicable than any",
"other objectrecognition (categorization,detection, or in",
"positioning (for unaltered speech), occlusion, illumination",
"stance identification) in consumer images. Similarly, putting",
"aside image processing operators that are applied to the en-",
"and shadow, and their dynamics.",
"",
"tire frame, face filters remain the most popular filters for",
"In contrast to the literature methods, which are limited to",
"still images and often swap a given face with a dataset face,",
"9",
"consumer video. Since face technology is both useful and",
":AIX",
"impactful, it also raises many ethical concerns. Face recog.",
"our method handles video and generates de novo faces. Our",
"nition can lead to loss of privacy and face replacement tech-",
"experimentsshowconvincingperformanceforunconstrained",
"videos, producing natural looking videos. The person in the",
"",
"",
"nology may be misused to create misleading videos.",
"In this work, we focus on video de-identification, which is",
"rendered video has a similar appearance to the person in the",
"a video filtering application that both requires a technological",
"original video.However.a state-of-the-artface-recognition",
"leap over the current state-of-the-art.and is benign innature",
"network fails to identify the person. A similar experiment",
"This application requires the creation of a video of a similar",
"shows that humans cannot identify the generated face, even",
"looking person, such that the perceived identity is changed.",
"without time constraints.",
"This allows, for example, the user to leave a natural-looking",
"Our results would not have been possible, without a host",
"video message in a public forum in an anonymous way, that",
"of novelties. We introduce a novel encoder-decoder archi-",
"would presumably prevent face recognition technology from",
"tecture, in which we concatenate to the latent space the",
"recognizing them.",
"activations of the representation layer of a network trained",
"Video de-identification is a challenging task. The video",
"to perform face recognition. As far as we know, this is the",
"needs to be modified in a seamless way, without causing",
"first time that a representation from an existing classifier",
"flickering or other visual artifacts and distortions, such that",
"network is usedto augment an autoencoder,which enables",
"the identity is changed, while all other factors remain identi-",
"the feed-forward treatment of new persons, unseen during",
"cal, see Fig. 1. These factors include pose, expression, lip",
"training. In addition, this is the first work to introduce a new",
"1"
] |
{
"arxiv_id": "1911.08348",
"language": "en",
"page_idx": 0
}
|
||
1911.08348-en-1
|
1911.08348-en-1.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Newton, $^{,}05\left[32\right]$</th>
<th>$Gross$, ,08[10]</th>
<th>Samarzija, ,14[41]</th>
<th>Jourabloo, 15[16]</th>
<th>Meden, 17[31]</th>
<th>$Wu$, ,18[49]</th>
<th>$\operatorname{Sun'18}$ [43,44]</th>
<th>Our</th>
</tr>
<tr>
<td>Preserves expression</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>+</td>
</tr>
<tr>
<td>Preserves pose</td>
<td> </td>
<td>+</td>
<td>+</td>
<td> </td>
<td>+</td>
<td> </td>
<td>+</td>
<td>+</td>
</tr>
<tr>
<td>Generates new faces</td>
<td> </td>
<td>十</td>
<td> </td>
<td>十</td>
<td>+</td>
<td>+</td>
<td>+</td>
<td>+</td>
</tr>
<tr>
<td>Demonstrated on video</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>+</td>
</tr>
<tr>
<td>Demonstrated on a diverse dataset (gender, ethnicity, age, etc.)</td>
<td> </td>
<td>+</td>
<td> </td>
<td>+</td>
<td> </td>
<td> </td>
<td> </td>
<td>+</td>
</tr>
<tr>
<td>Reference to a comparison with ours</td>
<td> </td>
<td> </td>
<td>$Fig.7$</td>
<td> </td>
<td>$Fig.4$</td>
<td>$Fig.\xi$ 8</td>
<td>Fig. 5 14</td>
<td> </td>
</tr>
</tbody>
</table>
Table l. A comparison to the literature methods. The final row references comparison figures in this work. We compare to all methods that provide reasonable quality images in their manuscript, under conditions that are favorable to previous work (we crop the input images from the pdf files, except for the images received from the authors of [43, 44]). $^{\dagger}$The face is swapped with an average of a few dataset faces.
type of attractor-repeller perceptual loss term. This term distinguishes between low- and mid-level perceptual terms, and high-level ones. The former are used to tie the output frame to the input video frame, while the latter is used to distance the identity. In this novel architecture, the injection of the representation to the latent space enables the network to create an output that adheres to this complex criterion. Another unique feature is that the network outputs both an image and a mask, which are used, in tandem, to reconstruct the output frame. The method is trained with a specific data augmentation technique that encourages the mapping to be semantic. Additional terms include reconstruction losses, edge losses, and an adversarial loss.
faces from pre-existing datasets [32,16,31]. The experiments conducted by the aforementioned techniques are restricted, in most cases, to low-resolution, black and white results. Although it is possible to create eye-pleasing results, they are not robust to different poses, illuminations and facial structures, making them inadequate for video generation. The use of GANs for face de-identification has been suggested[49]. However, the experiments were restricted to a homogeneous dataset, with no apparent expression preservation within the results. In the GAN-based methods of [43, 44], face de-identification is employed for the related task of person obfuscation. The work of [43] conditions the output image based on both a blurred version of the input and the extracted facial pose information. The follow-up work[44] combines the GAN-based reconstruction with a parametric face generation network. As both methods are applied over full upper-body images, they result in low facial resolution outputs of $64\times64.$ These methods do not preserve expressions, are unsuitable for video, and occasionally provide unnatural outputs.
## $2. \textbf{Previous Work}$
Faces have been modeled by computer graphics systems for a long time. In machine learning, faces have been one of the key benchmarks for GAN-based generative models [9, 37,40] since their inception. High resolution natural looking faces were recently generated by training both the generator and the discriminator of the GAN progressively, starting with shallower networks and lower resolutions, and enlarging them gradually [17].
Conditional generation of faces has been a key task in various unsupervised domain translation contributions, where the task is to learn to map, e.g., a person without eyewear to a person with eyeglasses, without seeing matching samples from the two domains [20,51,1,27]. For more distant domain mapping, such as mapping between a face image and the matching computer graphics avatar, additional supervision in the form of a face descriptor network was used [45]. Our work uses these face descriptors, in order to distance the identity of the output from that of the input.
As far as we know, our work is the first de-identification work to present results on videos. In still images, several methods have been previously suggested. Earlier work implemented different types of image distortions for face deidentification [33, 10], while more recent works rely on techniques for selecting distant faces [41] or averaging/fusing
Tab. 1 provides a comparative view of the literature. The current literature on de-identification often involves face swapping (our method does not). Face swapping, i.e., the replacement of a person's face in an image with another person's face, has been an active research topic for some time, starting with the influential work of [3,2]. Recent contributions have shown a great deal of robustness to the source image, as well as for the properties of the image, from which the target face is taken [19, 34]. While these classical face swapping methods work in the pixel space and copy the expression of the target image, recent deep-learning based work swaps the identity, while maintaining the other aspects of the source image [23]. In comparison to our work, [23] requires training a new network for every target person, the transferred expression does not show subtleties (which would be critical, e.g., for a speaking person), and the results are not as natural as ours. These limitations are probably a result of capturing the appearance of the target, by restricting the output to be similar, patch by patch, to a collection of
|
[
[
0.455810546875,
0.0921630859375,
0.52392578125,
0.0947265625,
0.52392578125,
0.10919189453125,
0.455810546875,
0.106689453125
],
[
0.6435546875,
0.0921630859375,
0.69384765625,
0.0947265625,
0.69287109375,
0.10919189453125,
0.64208984375,
0.106689453125
],
[
0.7255859375,
0.0947265625,
0.75341796875,
0.0947265625,
0.75341796875,
0.106689453125,
0.7255859375,
0.106689453125
],
[
0.3896484375,
0.0953369140625,
0.4306640625,
0.0953369140625,
0.4306640625,
0.106689453125,
0.3896484375,
0.106689453125
],
[
0.318603515625,
0.0953369140625,
0.370849609375,
0.0953369140625,
0.370849609375,
0.106689453125,
0.318603515625,
0.106689453125
],
[
0.55224609375,
0.0953369140625,
0.61669921875,
0.0953369140625,
0.61669921875,
0.106689453125,
0.55224609375,
0.106689453125
],
[
0.77783203125,
0.0953369140625,
0.82421875,
0.0953369140625,
0.82421875,
0.106689453125,
0.77783203125,
0.106689453125
],
[
0.84326171875,
0.0953369140625,
0.87158203125,
0.0953369140625,
0.87158203125,
0.106689453125,
0.84326171875,
0.106689453125
],
[
0.384033203125,
0.1060791015625,
0.435546875,
0.10797119140625,
0.4345703125,
0.12371826171875,
0.38232421875,
0.121826171875
],
[
0.464111328125,
0.1060791015625,
0.51708984375,
0.10797119140625,
0.515625,
0.12371826171875,
0.463134765625,
0.121826171875
],
[
0.55615234375,
0.1060791015625,
0.609375,
0.10797119140625,
0.60888671875,
0.12371826171875,
0.5556640625,
0.121826171875
],
[
0.64208984375,
0.1060791015625,
0.6953125,
0.10797119140625,
0.69384765625,
0.12371826171875,
0.640625,
0.121826171875
],
[
0.7099609375,
0.1060791015625,
0.76123046875,
0.10797119140625,
0.7607421875,
0.12371826171875,
0.708984375,
0.121826171875
],
[
0.77392578125,
0.10797119140625,
0.82666015625,
0.10797119140625,
0.82666015625,
0.12249755859375,
0.77392578125,
0.12249755859375
],
[
0.0882568359375,
0.126953125,
0.21728515625,
0.1275634765625,
0.21728515625,
0.1427001953125,
0.0882568359375,
0.141357421875
],
[
0.318603515625,
0.10919189453125,
0.36767578125,
0.10919189453125,
0.36767578125,
0.12060546875,
0.318603515625,
0.12060546875
],
[
0.339111328125,
0.1307373046875,
0.3505859375,
0.1307373046875,
0.3505859375,
0.14453125,
0.339111328125,
0.14453125
],
[
0.484375,
0.1307373046875,
0.495849609375,
0.1307373046875,
0.495849609375,
0.141357421875,
0.484375,
0.141357421875
],
[
0.73193359375,
0.1307373046875,
0.744140625,
0.1307373046875,
0.744140625,
0.143310546875,
0.73193359375,
0.143310546875
],
[
0.7958984375,
0.1307373046875,
0.80810546875,
0.1307373046875,
0.80810546875,
0.143310546875,
0.7958984375,
0.143310546875
],
[
0.85302734375,
0.1307373046875,
0.865234375,
0.1307373046875,
0.865234375,
0.141357421875,
0.85302734375,
0.141357421875
],
[
0.6611328125,
0.1319580078125,
0.6708984375,
0.1287841796875,
0.67626953125,
0.1395263671875,
0.6650390625,
0.1427001953125
],
[
0.400390625,
0.1331787109375,
0.410888671875,
0.1287841796875,
0.418212890625,
0.1395263671875,
0.406982421875,
0.143310546875
],
[
0.400390625,
0.1395263671875,
0.420654296875,
0.1395263671875,
0.420654296875,
0.1685791015625,
0.400390625,
0.1685791015625
],
[
0.5810546875,
0.1363525390625,
0.58837890625,
0.1363525390625,
0.58837890625,
0.14013671875,
0.5810546875,
0.14013671875
],
[
0.79345703125,
0.14013671875,
0.8095703125,
0.14013671875,
0.8095703125,
0.1572265625,
0.79345703125,
0.1572265625
],
[
0.85888671875,
0.14013671875,
0.87109375,
0.149658203125,
0.85888671875,
0.1590576171875,
0.84619140625,
0.149658203125
],
[
0.66015625,
0.1427001953125,
0.67724609375,
0.1427001953125,
0.67724609375,
0.1572265625,
0.66015625,
0.1572265625
],
[
0.0906982421875,
0.143310546875,
0.181396484375,
0.143310546875,
0.181396484375,
0.1546630859375,
0.0906982421875,
0.1546630859375
],
[
0.48193359375,
0.1427001953125,
0.498291015625,
0.1427001953125,
0.498291015625,
0.1572265625,
0.48193359375,
0.1572265625
],
[
0.5810546875,
0.1458740234375,
0.5888671875,
0.1458740234375,
0.5888671875,
0.1546630859375,
0.5810546875,
0.1546630859375
],
[
0.73193359375,
0.143310546875,
0.744140625,
0.143310546875,
0.744140625,
0.1546630859375,
0.73193359375,
0.1546630859375
],
[
0.57861328125,
0.154052734375,
0.5888671875,
0.154052734375,
0.5888671875,
0.1666259765625,
0.57861328125,
0.1666259765625
],
[
0.334228515625,
0.146484375,
0.3447265625,
0.141357421875,
0.351318359375,
0.15087890625,
0.340576171875,
0.1558837890625
],
[
0.66015625,
0.1558837890625,
0.67724609375,
0.1558837890625,
0.67724609375,
0.1717529296875,
0.66015625,
0.1717529296875
],
[
0.7294921875,
0.1558837890625,
0.74609375,
0.1558837890625,
0.74609375,
0.1717529296875,
0.7294921875,
0.1717529296875
],
[
0.79345703125,
0.1558837890625,
0.8095703125,
0.1558837890625,
0.8095703125,
0.1717529296875,
0.79345703125,
0.1717529296875
],
[
0.8505859375,
0.1572265625,
0.86669921875,
0.1572265625,
0.86669921875,
0.1717529296875,
0.8505859375,
0.1717529296875
],
[
0.339111328125,
0.1578369140625,
0.348876953125,
0.1578369140625,
0.348876953125,
0.17041015625,
0.339111328125,
0.17041015625
],
[
0.08905029296875,
0.1572265625,
0.2156982421875,
0.1572265625,
0.2156982421875,
0.17041015625,
0.08905029296875,
0.17041015625
],
[
0.48779296875,
0.160400390625,
0.495849609375,
0.160400390625,
0.495849609375,
0.169189453125,
0.48779296875,
0.169189453125
],
[
0.73046875,
0.1673583984375,
0.74609375,
0.1673583984375,
0.74609375,
0.186279296875,
0.73046875,
0.186279296875
],
[
0.85888671875,
0.1685791015625,
0.87109375,
0.177978515625,
0.85888671875,
0.1875,
0.84619140625,
0.177978515625
],
[
0.404296875,
0.17236328125,
0.416748046875,
0.17236328125,
0.416748046875,
0.18310546875,
0.404296875,
0.18310546875
],
[
0.08905029296875,
0.17041015625,
0.2320556640625,
0.17041015625,
0.2320556640625,
0.184326171875,
0.08905029296875,
0.184326171875
],
[
0.484375,
0.17236328125,
0.495849609375,
0.17236328125,
0.495849609375,
0.18310546875,
0.484375,
0.18310546875
],
[
0.66259765625,
0.17236328125,
0.67333984375,
0.17236328125,
0.67333984375,
0.18310546875,
0.66259765625,
0.18310546875
],
[
0.7958984375,
0.17236328125,
0.80615234375,
0.17236328125,
0.80615234375,
0.18310546875,
0.7958984375,
0.18310546875
],
[
0.58251953125,
0.175537109375,
0.58837890625,
0.179931640625,
0.5849609375,
0.1817626953125,
0.580078125,
0.177978515625
],
[
0.34228515625,
0.1749267578125,
0.346435546875,
0.1749267578125,
0.346435546875,
0.1817626953125,
0.34228515625,
0.1817626953125
],
[
0.08905029296875,
0.184326171875,
0.295654296875,
0.184326171875,
0.295654296875,
0.1976318359375,
0.08905029296875,
0.1976318359375
],
[
0.404296875,
0.1849365234375,
0.416748046875,
0.1849365234375,
0.416748046875,
0.1976318359375,
0.404296875,
0.1976318359375
],
[
0.48779296875,
0.186279296875,
0.496826171875,
0.189453125,
0.49267578125,
0.1988525390625,
0.483642578125,
0.1956787109375
],
[
0.57861328125,
0.186279296875,
0.5908203125,
0.186279296875,
0.5908203125,
0.1962890625,
0.57861328125,
0.1962890625
],
[
0.6650390625,
0.1849365234375,
0.67626953125,
0.1881103515625,
0.6708984375,
0.1988525390625,
0.66015625,
0.1956787109375
],
[
0.73193359375,
0.186279296875,
0.744140625,
0.186279296875,
0.744140625,
0.1976318359375,
0.73193359375,
0.1976318359375
],
[
0.79833984375,
0.1881103515625,
0.80615234375,
0.1881103515625,
0.80615234375,
0.1976318359375,
0.79833984375,
0.1976318359375
],
[
0.85107421875,
0.186279296875,
0.865234375,
0.186279296875,
0.865234375,
0.1976318359375,
0.85107421875,
0.1976318359375
],
[
0.098876953125,
0.2008056640625,
0.26318359375,
0.2008056640625,
0.26318359375,
0.2108154296875,
0.098876953125,
0.2108154296875
],
[
0.4697265625,
0.210205078125,
0.5107421875,
0.210205078125,
0.5107421875,
0.2259521484375,
0.4697265625,
0.2259521484375
],
[
0.64892578125,
0.210205078125,
0.689453125,
0.210205078125,
0.689453125,
0.2259521484375,
0.64892578125,
0.2259521484375
],
[
0.71728515625,
0.2108154296875,
0.7587890625,
0.2108154296875,
0.7587890625,
0.2259521484375,
0.71728515625,
0.2259521484375
],
[
0.76953125,
0.2108154296875,
0.83251953125,
0.2108154296875,
0.83251953125,
0.2259521484375,
0.76953125,
0.2259521484375
],
[
0.080078125,
0.2342529296875,
0.89208984375,
0.2342529296875,
0.89208984375,
0.2474365234375,
0.080078125,
0.2474365234375
],
[
0.09234619140625,
0.21337890625,
0.3095703125,
0.21337890625,
0.3095703125,
0.2235107421875,
0.09234619140625,
0.2235107421875
],
[
0.080078125,
0.248779296875,
0.89208984375,
0.248779296875,
0.89208984375,
0.261962890625,
0.080078125,
0.261962890625
],
[
0.07843017578125,
0.261474609375,
0.87109375,
0.261962890625,
0.87109375,
0.276611328125,
0.07843017578125,
0.27587890625
],
[
0.080078125,
0.300537109375,
0.4697265625,
0.300537109375,
0.4697265625,
0.31494140625,
0.080078125,
0.31494140625
],
[
0.50244140625,
0.3017578125,
0.89208984375,
0.3017578125,
0.89208984375,
0.31494140625,
0.50244140625,
0.31494140625
],
[
0.08087158203125,
0.318115234375,
0.471435546875,
0.318115234375,
0.471435546875,
0.328857421875,
0.08087158203125,
0.328857421875
],
[
0.5009765625,
0.31640625,
0.89208984375,
0.31640625,
0.89208984375,
0.330078125,
0.5009765625,
0.330078125
],
[
0.080078125,
0.33203125,
0.4697265625,
0.33203125,
0.4697265625,
0.34521484375,
0.080078125,
0.34521484375
],
[
0.5009765625,
0.330810546875,
0.890625,
0.330810546875,
0.890625,
0.3447265625,
0.5009765625,
0.3447265625
],
[
0.080078125,
0.347900390625,
0.471435546875,
0.347900390625,
0.471435546875,
0.361083984375,
0.080078125,
0.361083984375
],
[
0.5009765625,
0.347900390625,
0.8935546875,
0.347900390625,
0.8935546875,
0.361083984375,
0.5009765625,
0.361083984375
],
[
0.080078125,
0.3623046875,
0.4697265625,
0.3623046875,
0.4697265625,
0.375732421875,
0.080078125,
0.375732421875
],
[
0.50244140625,
0.3623046875,
0.8935546875,
0.3623046875,
0.8935546875,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.080078125,
0.37744140625,
0.4697265625,
0.37744140625,
0.4697265625,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.50244140625,
0.376953125,
0.8935546875,
0.37744140625,
0.8935546875,
0.391357421875,
0.50244140625,
0.39013671875
],
[
0.080078125,
0.392578125,
0.47216796875,
0.392578125,
0.47216796875,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.5009765625,
0.391357421875,
0.89208984375,
0.392578125,
0.89208984375,
0.4072265625,
0.5009765625,
0.406005859375
],
[
0.08087158203125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50341796875,
0.40771484375,
0.8935546875,
0.40771484375,
0.8935546875,
0.421630859375,
0.50341796875,
0.421630859375
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.50244140625,
0.423583984375,
0.8935546875,
0.423583984375,
0.8935546875,
0.4375,
0.50244140625,
0.4375
],
[
0.07843017578125,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.451904296875,
0.07843017578125,
0.451904296875
],
[
0.5009765625,
0.438232421875,
0.8935546875,
0.438232421875,
0.8935546875,
0.451904296875,
0.5009765625,
0.451904296875
],
[
0.07843017578125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.07843017578125,
0.46728515625
],
[
0.5048828125,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.46533203125,
0.5048828125,
0.46533203125
],
[
0.0767822265625,
0.466552734375,
0.471435546875,
0.46728515625,
0.471435546875,
0.481689453125,
0.0767822265625,
0.48095703125
],
[
0.50341796875,
0.4697265625,
0.890625,
0.4697265625,
0.890625,
0.479736328125,
0.50341796875,
0.479736328125
],
[
0.080078125,
0.48291015625,
0.31787109375,
0.48291015625,
0.31787109375,
0.496826171875,
0.080078125,
0.496826171875
],
[
0.50244140625,
0.48291015625,
0.89208984375,
0.48291015625,
0.89208984375,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.5009765625,
0.49755859375,
0.89208984375,
0.498779296875,
0.89208984375,
0.51220703125,
0.5009765625,
0.51123046875
],
[
0.08087158203125,
0.51220703125,
0.226318359375,
0.51220703125,
0.226318359375,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.50244140625,
0.51318359375,
0.8935546875,
0.51318359375,
0.8935546875,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.50244140625,
0.52880859375,
0.89208984375,
0.52880859375,
0.89208984375,
0.54248046875,
0.50244140625,
0.54248046875
],
[
0.09722900390625,
0.537109375,
0.4697265625,
0.53857421875,
0.4697265625,
0.55322265625,
0.09722900390625,
0.5517578125
],
[
0.50244140625,
0.54345703125,
0.89208984375,
0.54345703125,
0.89208984375,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.5537109375,
0.4697265625,
0.5537109375,
0.4697265625,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50244140625,
0.5595703125,
0.89208984375,
0.5595703125,
0.89208984375,
0.57275390625,
0.50244140625,
0.57275390625
],
[
0.08087158203125,
0.57080078125,
0.46728515625,
0.57080078125,
0.46728515625,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50244140625,
0.57373046875,
0.89208984375,
0.57373046875,
0.89208984375,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50244140625,
0.5888671875,
0.67724609375,
0.5888671875,
0.67724609375,
0.60302734375,
0.50244140625,
0.60302734375
],
[
0.080078125,
0.599609375,
0.4697265625,
0.599609375,
0.4697265625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.51953125,
0.61279296875,
0.89208984375,
0.6142578125,
0.89208984375,
0.62890625,
0.51953125,
0.62744140625
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.630859375,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.64453125
],
[
0.07843017578125,
0.64453125,
0.2156982421875,
0.64453125,
0.2156982421875,
0.65771484375,
0.07843017578125,
0.65771484375
],
[
0.50244140625,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.09722900390625,
0.66015625,
0.471435546875,
0.66015625,
0.471435546875,
0.673828125,
0.09722900390625,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.08087158203125,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.6875,
0.08087158203125,
0.6875
],
[
0.5009765625,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.5009765625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.70556640625,
0.471435546875,
0.70556640625,
0.471435546875,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.5009765625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.5009765625,
0.7197265625
],
[
0.07843017578125,
0.71826171875,
0.4697265625,
0.7197265625,
0.4697265625,
0.734375,
0.07843017578125,
0.73291015625
],
[
0.50244140625,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.366943359375,
0.796875,
0.366943359375,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.8935546875,
0.796875,
0.8935546875,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.1004638671875,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.1004638671875,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.8935546875,
0.81103515625,
0.8935546875,
0.82470703125,
0.50341796875,
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,
0.471435546875,
0.84130859375,
0.471435546875,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5009765625,
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,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
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
]
] |
[
"Samarzija,",
"Meden,",
"Wu,",
"Gross,",
"Newton,",
"Jourabloo,",
"Sun'18",
"Our",
"°08 [10]",
"'14 [41]",
"'15 [16]",
"\"17 [31]",
"'18 [49]",
"[43, 44]",
"Preserves expression",
"'05[32]",
"",
"-",
"-",
"-",
"+",
".",
".",
"++",
"-",
"+",
"x",
"+",
"Preserves pose",
"+",
"-",
"-",
"+",
"、",
"+",
"+",
"+",
"+",
"-",
"Generates new faces",
"-",
"",
"x",
"-",
"Demonstrated on video",
"-",
"-",
"-",
"",
"",
"Demonstrated on a diverse dataset",
"+",
".",
"+",
".",
"-",
"-",
"+",
"(gender,ethnicity,age,etc.)",
"Fig.7",
"Fig. 4",
"Fig. 8",
"Fig. 5, 14",
"Table 1. A comparison to the literature methods. The final row references comparison figures in this work. We compare to all methods that",
"Referencetoacomparisonwithours",
"provide reasonable quality images in their manuscript, under conditions that are favorable to previous work (we crop the input images from",
"the pdf files, except for the images received from the authors of [43, 44]). t The face is swapped with an average of a few dataset faces.",
"type of attractor-repeller perceptual loss term. This term",
"faces from pre-existing datasets [32, 16, 31]. The experi-",
"distinguishes between low-andmid-level perceptual terms,",
"ments conducted by the aforementioned techniques are re-",
"and high-level ones. The former are used to tie the output",
"stricted, in most cases, to low-resolution, black and white",
"frame to the input video frame, while the latter is used to",
"results. Although it is possible to create eye-pleasing results,",
"distance the identity. In this novel architecture, the injection",
"they are not robust to different poses, illuminations and fa-",
"of the representation to the latent space enables the network",
"cial structures, making them inadequatefor video generation.",
"to create an output that adheres to this complex criterion.",
"The use of GANs for face de-identification has been sug-",
"Anotheruniquefeatureisthatthenetworkoutputsbothan",
"gested [49]. However, the experiments were restricted to a",
"imageandamask,whichareused,intandem,toreconstruct",
"homogeneous dataset, with no apparent expression preser-",
"the output frame. The method is trained with a specific data",
"vation within the results. In the GAN-based methods of",
" augmentation technique that encourages the mapping to be",
"[43,44l,facede-identificationisemployedfortherelated",
"semantic. Additional terms include reconstruction losses,",
"taskofpersonobfuscation.Theworkof43conditionsthe",
"edge losses, and an adversarial loss.",
"output image based on both a blurred version of the input",
"and the extracted facial pose information.The follow-up",
"2. Previous Work",
"work [44] combines the GAN-based reconstruction with a",
"parametric face generation network. As both methods are",
"Faces have been modeled by computer graphics systems",
"applied over full upper-body images, they result in low facial",
"for a long time. In machine learning, faces have been one of",
"resolution outputs of 64 × 64. These methods do not pre-",
"thekeybenchmarksforGAN-basedgenerativemodels[9",
"serve expressions, are unsuitable for video, and occasionally",
"37, 40] since their inception. High resolution natural looking",
"provide unnatural outputs.",
"faces were recently generated by training both the generator",
"and the discriminator of the GAN progressively, starting with",
"Tab. I provides a comparative view of the literature. The",
"shallower networks and lower resolutions, and enlarging",
"current literature on de-identification often involves face",
"them gradually [17].",
"swapping (our method does not). Face swapping, i.e., the",
"Conditional generation of faces has been a key task in var-",
"replacement of a person's face in an image with another",
"ious unsupervised domain translation contributions,where",
"person's face, has been an active research topic for some",
"the task is to learn to map, e.g., a person without eyewear to",
"time, starting with the influential work of [3, 2]. Recent",
"a person with eyeglasses, without seeing matching samples",
"contributions have shown a great deal of robustness to the",
"from the two domains [20, 51, 1, 27]. For more distant do-",
"source image, as well as for the properties of the image,",
"mainmapping,suchasmappingbetweenafaceimageand",
"from which the targetface is taken19, 34l. Whilethese",
"the matching computer graphics avatar, additional supervi-",
"classical face swapping methods work in the pixel space and",
"sionintheformofafacedescriptornetworkwasused45l",
"copy the expression of the target image, recent deep-learning",
"Our work uses these face descriptors, in order to distance the",
"based work swapstheidentity,whilemaintaining the other",
"identity of the output from that of the input.",
"aspects of the source image [23]. In comparison to our work,",
"As far as we know, our work is the first de-identification",
"[23] requires training a new network for every target person,",
"work to present results on videos. In still images, several",
"the transferred expression does not show subtleties (which",
"methods have been previously suggested. Earlier work im-",
"would be critical, e.g., for a speaking person), and the results",
"plemented different types of image distortions for face de-",
"are not as natural as ours. These limitations are probably a",
"identification [33, 10], while more recent works rely on tech-",
"result of capturing the appearance of the target, by restricting",
"niques for selecting distant faces [41] or averaging/fusing",
"the output to be similar, patch by patch, to a collection of"
] |
{
"arxiv_id": "1911.08348",
"language": "en",
"page_idx": 1
}
|
||
1911.08348-en-3
|
1911.08348-en-3.jpg
|
contributes to the applicability of the method, allowing it to be applied to live videos. In our experiments, we do not use an input frame in order to show the generality of the approach. To encode the target image, we use a pre-trained face classifier ResNet-50 network[12], trained over the VGGFace2 dataset[4].
The process during test time is similar to the steps taken in the face swapping literature and involves the following steps: (a) A square bounding box is extracted using the'dlib'[21] face detector. (b) 68 facial points are detected using [18]. (c) A transformation matrix is extracted, using an estimated similarity transformation (scale, rotation and translation) to an averaged face. (d) The estimated transformation is applied to the input face. (e) The transformed face is passed to our network, together with the representation of the target image, obtaining both an output image and a mask. (f) The output image and mask are projected back, using the inverse of the similarity transformation. (g) We generate an output frame by linearly mixing, per pixel, the input and the network's transformed output image, according to the weights of the transformed mask.(h) The outcome is merged into the original frame, in the region defined by the convex hull of the facial points.
At training time, we perform the following steps: (a) The face image is distorted and augmented. This is done by applying random scaling, rotation and elastic deformation. (b) The distorted image is fed into the network, together with the representation of a target image. During training, we select the same image, undistorted. (c) A linear combination of the masked output (computed as in step (g) above) and the undistorted input is fed to the discriminator. This is the mixup technique [52] discussed below. (d) Losses are applied on the network's mask and image output, as well as to the masked output, as detailed below.
Note that there is a discrepancy between how the network is trained and how it is applied. Not only do we not make any explicit effort to train on videos, the target images are selected in a different way. During training, we extract the identity from the training image itself and not from an independent target image. The method is still able to generalize to perform the real task on unconstrained videos.
3.1. Network architecture
The architecture is illustrated in Fig. 2(a). The encoder is composed of a convolutional layer, followed by five strided, depth-wise separable[6] convolutions with instance normalization [48]. Subsequently, a single fully connected layer is employed, and the target face representation is concatenated. The decoder is composed of a fully connected layer, followed by a lattice of upscale and residual [12] blocks, terminated with a tanh activated convolution for the output image, and a sigmoid activated convolution for the mask output. Each upscale block is comprised of a 2D convolu-
tion, with twice the number of filters as the input channel size. Following an instance normalization and a LReLU[1] activation, the activations are re-ordered, so that the width and height are doubled, while the channel size is halved. Each residual block input is summed with the output of a Conv2D-LReLU-Conv2D chain.
A low-capacity U-net connection [38] is employed $(32x32x1)$, thus relieving the autoencoder's bottleneck, allowing a stronger focus on the encoding of transfer-related information. The connection size does not exceed the bottleneck size (1024) and due to the distortion of the input image, a collapse into a simple reconstructing autoencoder in early training stages is averted.
The discriminator consists of four strided convolutions with LReLU activations, with instance normalization applied on all but the first one. A sigmoid activated convolution yields a single output.
The network has two versions: a lower resolution version generating 128x128 images, and a higher resolution version, generating 256x256 images. The higher resolution decoder is simplified and enlarged and consists of a lattice of 6x(Upscale block-> Residual block). Unless otherwise specified, the results presented in the experiments are done with the high-res model.
## 3.2. Training and the Losses Used
For training all networks, except for the discriminator $D$,we use a compound loss $\mathcal{L}$, which is a weighted sum of multiple parts:
$$\begin{aligned}\mathcal{L}=&\alpha_{0}\mathcal{L}_{G}+\alpha_{1}\mathcal{L}_{R}^{raw}+\alpha_{1}\mathcal{L}_{R}^{masked}+\alpha_{2}\mathcal{L}_{x}^{raw}\\&+\alpha_{2}\mathcal{L}_{y}^{raw}+\alpha_{2}\mathcal{L}_{x}^{masked}+\alpha_{2}\mathcal{L}_{y}^{masked}\\&+\alpha_{3}\mathcal{L}_{p}^{raw}+\alpha_{3}\mathcal{L}_{p}^{masked}+\alpha_{4}\mathcal{L}^{m}+\alpha_{5}\mathcal{L}_{x}^{m}+\alpha_{5}\mathcal{L}_{y}^{m},\end{aligned}$$
where $\mathcal{L}_G$ is the generator's loss, $\mathcal{L}_R^{raw}$ and $\mathcal{L}_R^{masked}$ are reconstruction losses for the output image of the decoder $z^{raw}$ and the version after applying the masking $z^masked$, $\mathcal{L}_x^*$ and $\mathcal{L}_y^*$ are reconstruction losses applied to the spatial images derivatives, $\mathcal{L}_p^*$ are the perceptual losses, and $\mathcal{L}_*^m$ are regularization losses on the mask. The discriminator network is trained using its own loss $\mathcal{L}_D.$ Throughout our experiments, we employ $\alpha_0=\alpha_1=\alpha_2=\alpha_3=0.5,\alpha_4=$ $3\cdot10^{-3},\alpha_{5}=10^{-2}.$
To maintain realistic looking generator outputs, an adversarial loss is used with a convex combination of example pairs (known as mixup)[52] over a Least Square GAN[30]:
$$\begin{aligned}&\mathcal{L}_{D}=\|D(\delta_{mx})-\lambda_{\beta}1\|_{2}^{2}\\&\mathcal{L}_{G}=\alpha_{0}\|D(\delta_{mx})-(1-\lambda_{\beta})1\|_{2}^{2}\end{aligned}$$
While, $\delta_mx=\lambda_\beta\cdot x+(1-\lambda_\beta)z^{masked}$ and $\lambda_\beta$ is sampled $\begin{array}{l}\text{out of a Beta distribution }\lambda_{\beta}\sim Beta(\alpha,\alpha),x\text{ is the undis-}\\\text{torted input“real”sample and }z^{masked}\text{ is the post masking}\end{array}$
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
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,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50341796875,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50341796875,
0.1395263671875,
0.89208984375,
0.1395263671875,
0.89208984375,
0.149658203125,
0.50341796875,
0.149658203125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.8935546875,
0.15283203125,
0.8935546875,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.20751953125,
0.1685791015625,
0.20751953125,
0.1817626953125,
0.07843017578125,
0.179931640625
],
[
0.50341796875,
0.1673583984375,
0.71728515625,
0.1673583984375,
0.71728515625,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.1004638671875,
0.184326171875,
0.4697265625,
0.184326171875,
0.4697265625,
0.1976318359375,
0.1004638671875,
0.1976318359375
],
[
0.521484375,
0.18310546875,
0.8935546875,
0.18310546875,
0.8935546875,
0.1962890625,
0.521484375,
0.1962890625
],
[
0.080078125,
0.199462890625,
0.4697265625,
0.199462890625,
0.4697265625,
0.21337890625,
0.080078125,
0.21337890625
],
[
0.50341796875,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.2108154296875,
0.50341796875,
0.2108154296875
],
[
0.08087158203125,
0.2139892578125,
0.46728515625,
0.2139892578125,
0.46728515625,
0.2279052734375,
0.08087158203125,
0.2279052734375
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.24365234375,
0.080078125,
0.24365234375
],
[
0.50341796875,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.50341796875,
0.240478515625
],
[
0.08087158203125,
0.246826171875,
0.46728515625,
0.246826171875,
0.46728515625,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.08087158203125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.271484375,
0.08087158203125,
0.271484375
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.5009765625,
0.2734375,
0.67236328125,
0.2734375,
0.67236328125,
0.287353515625,
0.5009765625,
0.287353515625
],
[
0.08087158203125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.3017578125,
0.08087158203125,
0.3017578125
],
[
0.521484375,
0.28857421875,
0.89208984375,
0.28857421875,
0.89208984375,
0.3017578125,
0.521484375,
0.3017578125
],
[
0.080078125,
0.3037109375,
0.46728515625,
0.3037109375,
0.46728515625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.5009765625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.5009765625,
0.317626953125
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50244140625,
0.33203125
],
[
0.080078125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50244140625,
0.333984375,
0.64892578125,
0.333984375,
0.64892578125,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.08087158203125,
0.3515625,
0.46728515625,
0.3515625,
0.46728515625,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.52294921875,
0.348388671875,
0.8935546875,
0.348388671875,
0.8935546875,
0.3623046875,
0.52294921875,
0.3623046875
],
[
0.080078125,
0.365478515625,
0.4697265625,
0.365478515625,
0.4697265625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.50244140625,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.07843017578125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.07843017578125,
0.39453125
],
[
0.5009765625,
0.37890625,
0.89208984375,
0.37744140625,
0.89208984375,
0.392578125,
0.5009765625,
0.393310546875
],
[
0.07843017578125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.07843017578125,
0.4091796875
],
[
0.50341796875,
0.395751953125,
0.8896484375,
0.395751953125,
0.8896484375,
0.406005859375,
0.50341796875,
0.406005859375
],
[
0.08087158203125,
0.412353515625,
0.468017578125,
0.412353515625,
0.468017578125,
0.42236328125,
0.08087158203125,
0.42236328125
],
[
0.50341796875,
0.410888671875,
0.890625,
0.410888671875,
0.890625,
0.421630859375,
0.50341796875,
0.421630859375
],
[
0.080078125,
0.425537109375,
0.191162109375,
0.425537109375,
0.191162109375,
0.439453125,
0.080078125,
0.439453125
],
[
0.50244140625,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.50244140625,
0.438232421875
],
[
0.09722900390625,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.09722900390625,
0.455078125
],
[
0.50244140625,
0.43994140625,
0.6650390625,
0.43994140625,
0.6650390625,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.080078125,
0.45703125,
0.468017578125,
0.45703125,
0.468017578125,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50341796875,
0.46337890625,
0.76123046875,
0.46337890625,
0.76123046875,
0.477783203125,
0.50341796875,
0.477783203125
],
[
0.080078125,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.080078125,
0.4873046875,
0.4697265625,
0.4873046875,
0.4697265625,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.521484375,
0.48486328125,
0.890625,
0.486083984375,
0.890625,
0.50048828125,
0.521484375,
0.5
],
[
0.08087158203125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50341796875,
0.501953125,
0.89208984375,
0.501953125,
0.89208984375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.08087158203125,
0.51806640625,
0.468017578125,
0.51806640625,
0.468017578125,
0.52880859375,
0.08087158203125,
0.52880859375
],
[
0.50341796875,
0.51953125,
0.60107421875,
0.51953125,
0.60107421875,
0.52978515625,
0.50341796875,
0.52978515625
],
[
0.080078125,
0.5322265625,
0.4697265625,
0.5322265625,
0.4697265625,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.5,
0.53857421875,
0.81640625,
0.5361328125,
0.81640625,
0.56005859375,
0.5,
0.5625
],
[
0.080078125,
0.54736328125,
0.471435546875,
0.54736328125,
0.471435546875,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.080078125,
0.5625,
0.4697265625,
0.5625,
0.4697265625,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.53173828125,
0.56005859375,
0.75341796875,
0.56005859375,
0.75341796875,
0.58203125,
0.53173828125,
0.58203125
],
[
0.740234375,
0.5625,
0.79736328125,
0.5625,
0.79736328125,
0.57373046875,
0.740234375,
0.57373046875
],
[
0.080078125,
0.57763671875,
0.4697265625,
0.57763671875,
0.4697265625,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.53369140625,
0.580078125,
0.88818359375,
0.58056640625,
0.88818359375,
0.60302734375,
0.53369140625,
0.6015625
],
[
0.07843017578125,
0.59228515625,
0.34228515625,
0.59228515625,
0.34228515625,
0.60595703125,
0.07843017578125,
0.60595703125
],
[
0.09722900390625,
0.60791015625,
0.4697265625,
0.60791015625,
0.4697265625,
0.6220703125,
0.09722900390625,
0.6220703125
],
[
0.5185546875,
0.6123046875,
0.8935546875,
0.611328125,
0.8935546875,
0.62890625,
0.5185546875,
0.6298828125
],
[
0.07843017578125,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.07843017578125,
0.63720703125
],
[
0.50244140625,
0.6298828125,
0.890625,
0.6298828125,
0.890625,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.07843017578125,
0.6396484375,
0.471435546875,
0.6396484375,
0.471435546875,
0.65283203125,
0.07843017578125,
0.65283203125
],
[
0.5009765625,
0.642578125,
0.89208984375,
0.6416015625,
0.89208984375,
0.6591796875,
0.5009765625,
0.66015625
],
[
0.080078125,
0.65380859375,
0.4697265625,
0.65380859375,
0.4697265625,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.5048828125,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.5048828125,
0.673828125
],
[
0.080078125,
0.66845703125,
0.4697265625,
0.66845703125,
0.4697265625,
0.681640625,
0.080078125,
0.681640625
],
[
0.50244140625,
0.67529296875,
0.88818359375,
0.67529296875,
0.88818359375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.6845703125,
0.4697265625,
0.6845703125,
0.4697265625,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.69970703125,
0.40185546875,
0.69970703125,
0.40185546875,
0.71337890625,
0.07843017578125,
0.71337890625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.5009765625,
0.7197265625,
0.890625,
0.72021484375,
0.890625,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.72607421875,
0.27783203125,
0.72607421875,
0.27783203125,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.5009765625,
0.73486328125,
0.6337890625,
0.73486328125,
0.6337890625,
0.75,
0.5009765625,
0.75
],
[
0.098876953125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.098876953125,
0.7646484375
],
[
0.52294921875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.52294921875,
0.76513671875
],
[
0.080078125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.767578125,
0.890625,
0.767578125,
0.890625,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.583984375,
0.806640625,
0.75341796875,
0.806640625,
0.75341796875,
0.8212890625,
0.583984375,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5810546875,
0.82568359375,
0.81396484375,
0.82470703125,
0.81396484375,
0.84228515625,
0.5810546875,
0.84326171875
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.5,
0.85302734375,
0.8935546875,
0.8515625,
0.8935546875,
0.87060546875,
0.5,
0.8720703125
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.87060546875,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.8837890625,
0.8935546875,
0.8837890625,
0.8935546875,
0.9013671875,
0.5009765625,
0.9013671875
]
] |
[
"contributes to the applicability of the method, allowing it",
"tion, with twice the number of filters as the input channel",
"to be applied to live videos. In our experiments, we do not",
"size. Following an instance normalization and a LReLU [11]",
"use an input frame in order to show the generality of the",
"activation,the activations are re-ordered,sothat the width",
"approach. To encode the target image, we use a pre-trained",
"and height are doubled, while the channel size is halved",
"face classifier ResNet-50 network [12], trained over the VG-",
"Each residual block input is summed with the output of a",
"GFace2 dataset [4]]",
"Conv2D-LReLU-Conv2D chain.",
"The process during test time is similar to the steps taken in",
"A low-capacity U-net connection [38] is employed",
"the face swapping literature and involves the following steps:",
"(32x32x1), thus relieving the autoencoder's bottleneck, al-",
"(a) A square bounding box is extracted using the 'dlib' [2]]",
"lowing a stronger focus on the encoding of transfer-related",
"face detector. (b) 68 facial points are detected using [18]",
"information.Theconnectionsizedoesnotexceedthebottle",
"(c)A transformationmatrix is extracted,using anestimated",
"neck size (1024) and due to the distortion of the input image,",
"similaritytransformation(scale,rotationandtranslation)to",
"a collapse into a simple reconstructing autoencoder in early",
"an averaged face. (d) The estimated transformation is applied",
"training stages is averted.",
"totheinputface.(e)Thetransformedfaceispassedtoour",
"The discriminator consists of four strided convolutions",
"network, together with the representation of the target image",
"with LReLU activations, with instance normalization applied",
"obtaining both an output image and a mask. (f) The output",
"on all but the first one. A sigmoid activated convolution",
"image and mask are projected back, using the inverse of the",
"yields a single output.",
"similaritytransformation.(g)Wegenerateanoutputframe",
"The network has two versions: a lower resolution ver-",
"by linearly mixing, per pixel, the input and the network's",
"sion generating 128x128 images, and a higher resolution",
"transformed output image, according to the weights of the",
"version,generating256x256 images.The higher resolution",
"transformed mask. (h) The outcome is merged into the",
"decoderissimplifiedandenlargedandconsistsofalattice",
"originalframe,intheregiondefinedbytheconvexhull of",
"of6x(Upscaleblock->Residualblock).Unlessotherwise",
"the facial points.",
"specified, the results presented in the experiments are done",
"At training time, we perform the following steps: (a) The",
"with the high-res model.",
"face image is distorted and augmented. This is done by",
"3.2. Training and the Losses Used",
"applying random scaling, rotation and elastic deformation.",
"(b) The distorted image is fed into the network, together with",
"For training all networks, except for the discriminator",
"the representation of a target image. During training, we",
"D, we use a compound loss C, which is a weighted sum of",
"selectthe sameimage,undistorted.(c)A linearcombination",
"multiple parts:",
"of the masked output (computed as in step (g) above) and",
"C =aoLG + anCR\" +Q1CRasked + a2Craw",
"the undistorted input is fed to the discriminator. This is",
"the mixup technique [52] discussed below. (d) Losses are",
"+ a2Crau + a2Cmasked + a2C",
"cmasked",
"applied on the network's mask and image output, as well as",
"+a3Craw +a3Cmasked + a4Cm +asCm +asCm",
"to the masked output, as detailed below.",
"Note that there is a discrepancy between how the network",
"where LG is the generator's loss, Craw and Cmasked are",
"is trained and how it is applied. Not only do we not make",
"reconstruction losses for the output image of the decoder",
"any explicit effort to train on videos, the target images are",
"zraw and the version after applying the masking zmasked,",
"selected in a different way. During training, we extract the",
"L* and L* are reconstruction losses applied to the spatial",
"identity from the training image itself and not from an inde-",
"images derivatives, L, are the perceptual losses, and Cm",
"pendent target image. The method is still able to generalize",
"are regularization losses on the mask. The discriminator",
"to perform the real task on unconstrained videos.",
"network is trained using its own loss LD. Throughout our",
"experiments, we employ Qo = α1 = α2 = Q3 = 0.5, α4 =",
"3.1. Network architecture",
"3.10-3,Q5 =10-2",
"The architecture is illustrated in Fig. 2(a). The encoder is",
"To maintain realistic looking generator outputs, an adver-",
"composed of a convolutional layer, followed by five strided,",
"sarial loss is used with a convex combination of example",
"depth-wiseseparable[6|convolutionswithinstancenormal",
"pairs (known as mixup) [52] over a Least Square GAN [30]:",
"ization [48]. Subsequently, a single fully connected layer",
"is employed, and the target face representation is concate-",
"CD=D(8mz)-Ag12",
"nated. The decoder is composed of a fully connected layer,",
"LG = Q0D(8mz)-(1 -入)12",
"followed by a lattice of upscale and residual [12] blocks,",
"terminated with atanh activated convolutionforthe output",
"While, omr = 入g - + (1 -入g)zmasked and g is sampled",
"image, and a sigmoid activated convolution for the mask",
"out of a Beta distribution As ~ Beta(α, α), r is the undis-",
"output. Each upscale block is comprised of a 2D convolu-",
"torted input \"real\" sample and zmasked is the post masking"
] |
{
"arxiv_id": "1911.08348",
"language": "en",
"page_idx": 3
}
|
||
1911.08348-en-4
|
1911.08348-en-4.jpg
|
generated sample. A value of $\alpha=0.2$ is used throughout
the experiments.
Additional losses are exercised to both retain sourceto-output similarity, yet drive a perceptible transformation. Several losses are distributed equally between the raw and masked outputs, imposing constraints on both. An L1 reconstruction loss is used to enforce pixel-level similarity:
$$\mathcal{L}_{R}^{raw}=\alpha_{1}\|z^{raw}-x\|_{1}\quad\mathcal{L}_{R}^{masked}=\alpha_{1}\|z^{masked}-x\|_{1}$$
where $z^raw$ is the output image itself. This results in a nontrivial constraint, as the encoder input image is distorted. An edge-preserving loss is used to constrain pixel-level derivative differences in both the $x$ and $y$ image axes. Calculated as the absolute difference between the source and output derivatives in each axis direction for both the raw and masked outputs:
$$\mathcal{L}_{x}^{raw}=\alpha_{2}\|z_{x}^{raw}-x_{x}\|_{1}\quad\mathcal{L}_{x}^{masked}=\alpha_{2}\|z_{x}^{masked}-x_{x}\|_{1}\\\mathcal{L}_{y}^{raw}=\alpha_{2}\|z_{y}^{raw}-x_{y}\|_{1}\quad\mathcal{L}_{y}^{masked}=\alpha_{2}\|z_{y}^{masked}-x_{y}\|_{1}$$
where $x_x$ is the derivative of the undistorted input image $x$ along the $x$ axis, and similarly for outputs $z$ and the $y$ axis.
Additional losses are applied to the blending mask $m$, where 0 indicates that the value of this pixel would be taken from the input image $x,1$ indicates taking the value from $z^{raw}$, and intermediate values indicate linear mixing. We would like the mask to be both minimal and smooth and, therefore, employ the following losses:
$$\mathcal{L}^m=\|m\|_1\quad\mathcal{L}_x^m=\|m_x\|_1\quad\mathcal{L}_y^m=\|m_y\|_1$$
where $m_x$ and $m_y$ are the spatial derivatives of the mask.
## 3.2.1 A Multi-Image Perceptual Loss
A new variant of the perceptual loss[15] is employed to maintain source expression, pose and lighting conditions, while capturing the target identity essence. This is achieved by employing a perceptual loss between the undistorted source and generated output on several low-to-medium abstraction layers, while distancing the high abstraction layer perceptual loss between the target and generated output.
${\hat{\text{Let }a}_{n\times n}^{r}}$be the activations of an $n\times n$ spatial block within the face classifier network for image $r$,where in our case, $r$ can be either the input image $x$, the target image $t$, the raw output $z^{raw}$, or the masked output $z^{masked}.$
We consider the spatial activations maps of size 112× $112,56\times56,28\times28$ and $7\times7$, as well as the representation layer of size $1\times1.$ The lower layers (larger maps) are used to enforce similarity to the input image $x$,while the 7$\times7$ layer is used to enforce similarity to $t$, and the $1\times1$ feature vector is used to enforce dissimilarity to the target image.
Let us define $\ell_{n\times n}^{r_1,r_2}=c_n\|a_{r_1,n\times n}-a_{r_2,n\times n}\|_1$,where $c_n$ is a normalizing constant, corresponding to the size of the spatial activation map.
<FigureHere>
Figure 3. Sample results for video de-identification (zoom). Triplets
of source frame, converted frame and target are shown. The modi-
fied frame looks similar but the identity is completely different.
The perceptual loss is given by:
$$\mathcal{L}_{p}^{c}=\ell_{112\times112}^{x,z^{c}}+\ell_{56\times56}^{x,z^{c}}+\ell_{28\times28}^{x,z^{c}}+\ell_{7\times7}^{t,z^{c}}-\lambda\ell_{1\times1}^{t,z}$$
for $c$ that is either raw or $masked$, and where $\lambda>0$ is a hyperparameter, which determines the generated face's high level features distance from those of the target image.
The application of the multi-image perceptual loss during training is depicted in Fig. 2(b). During training, the target is the source, and there is only one input image. The resulting image has the texture, pose and expression of the source, but the face is modifed to distance the identity. Note that we refer to it as a multi-image perceptual loss, as its aim is to minimize the analog error term during inference (generalization error). However, as a training loss, it is only applied during train, where it receives a pair of images, similar to other perceptual losses.
Note that the perceptual loss parameters $c_n$ are normalizing constants obtained by counting the number of elements. In addition, $\alpha_0=\alpha_1=\alpha_2=\alpha_3$ are simply set to one, and $\alpha_4,\alpha_5$ were chosen arbitrarily. Therefore, there is effectively, only a single important hyperparameter: $\lambda$, which provides a direct control of the strength of the identity distance which requires tuning (see Fig. 9).
At inference time, the network is fed an input frame and a target image. The target image is transmitted through the face classifier, resulting in a target feature vector, which, in turn, is concatenated to the latent embedding space. Due to the way the network is trained, the decoder will drive the output image away from the target feature vector.
|
[
[
0.07843017578125,
0.0921630859375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.07843017578125,
0.106689453125
],
[
0.08087158203125,
0.10919189453125,
0.1871337890625,
0.10919189453125,
0.1871337890625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.098876953125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.098876953125,
0.1363525390625
],
[
0.080078125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.43310546875,
0.18310546875,
0.43310546875,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.1217041015625,
0.2083740234375,
0.2017822265625,
0.2083740234375,
0.2017822265625,
0.2197265625,
0.1217041015625,
0.2197265625
],
[
0.373291015625,
0.205810546875,
0.4248046875,
0.205810546875,
0.4248046875,
0.21337890625,
0.373291015625,
0.21337890625
],
[
0.4345703125,
0.210205078125,
0.459228515625,
0.210205078125,
0.459228515625,
0.2183837890625,
0.4345703125,
0.2183837890625
],
[
0.2197265625,
0.2108154296875,
0.24267578125,
0.2108154296875,
0.24267578125,
0.2183837890625,
0.2197265625,
0.2183837890625
],
[
0.0882568359375,
0.210205078125,
0.108642578125,
0.210205078125,
0.108642578125,
0.2183837890625,
0.0882568359375,
0.2183837890625
],
[
0.07843017578125,
0.2291259765625,
0.468017578125,
0.2310791015625,
0.468017578125,
0.24560546875,
0.07843017578125,
0.24365234375
],
[
0.080078125,
0.246826171875,
0.4697265625,
0.246826171875,
0.4697265625,
0.260009765625,
0.080078125,
0.260009765625
],
[
0.08087158203125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.274658203125,
0.08087158203125,
0.274658203125
],
[
0.080078125,
0.276611328125,
0.4697265625,
0.276611328125,
0.4697265625,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.080078125,
0.290283203125,
0.46728515625,
0.290283203125,
0.46728515625,
0.3037109375,
0.080078125,
0.3037109375
],
[
0.08087158203125,
0.30810546875,
0.46728515625,
0.30810546875,
0.46728515625,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.08087158203125,
0.323974609375,
0.1380615234375,
0.323974609375,
0.1380615234375,
0.335205078125,
0.08087158203125,
0.335205078125
],
[
0.07843017578125,
0.34033203125,
0.476318359375,
0.339111328125,
0.476318359375,
0.361083984375,
0.07843017578125,
0.3623046875
],
[
0.07843017578125,
0.359130859375,
0.48193359375,
0.359130859375,
0.48193359375,
0.3837890625,
0.07843017578125,
0.3837890625
],
[
0.50341796875,
0.35791015625,
0.890625,
0.35791015625,
0.890625,
0.371826171875,
0.50341796875,
0.371826171875
],
[
0.5009765625,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.385009765625,
0.5009765625,
0.385009765625
],
[
0.07843017578125,
0.38818359375,
0.4697265625,
0.389404296875,
0.4697265625,
0.404052734375,
0.07843017578125,
0.40283203125
],
[
0.50341796875,
0.38818359375,
0.8759765625,
0.38818359375,
0.8759765625,
0.3984375,
0.50341796875,
0.3984375
],
[
0.0767822265625,
0.404052734375,
0.46728515625,
0.40478515625,
0.46728515625,
0.419189453125,
0.0767822265625,
0.41845703125
],
[
0.1004638671875,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.434326171875,
0.1004638671875,
0.434326171875
],
[
0.52294921875,
0.41845703125,
0.73046875,
0.41845703125,
0.73046875,
0.431884765625,
0.52294921875,
0.431884765625
],
[
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
],
[
0.52783203125,
0.455078125,
0.65673828125,
0.458251953125,
0.65673828125,
0.47607421875,
0.52783203125,
0.472900390625
],
[
0.8037109375,
0.45263671875,
0.861328125,
0.458251953125,
0.85888671875,
0.4716796875,
0.80126953125,
0.466552734375
],
[
0.705078125,
0.462158203125,
0.71728515625,
0.462158203125,
0.71728515625,
0.46533203125,
0.705078125,
0.46533203125
],
[
0.7587890625,
0.4609375,
0.78173828125,
0.4609375,
0.78173828125,
0.46533203125,
0.7587890625,
0.46533203125
],
[
0.70361328125,
0.46337890625,
0.77685546875,
0.46337890625,
0.77685546875,
0.47021484375,
0.70361328125,
0.47021484375
],
[
0.08251953125,
0.46533203125,
0.468017578125,
0.46533203125,
0.468017578125,
0.479248046875,
0.08251953125,
0.479248046875
],
[
0.650390625,
0.46533203125,
0.69921875,
0.46533203125,
0.69921875,
0.4697265625,
0.650390625,
0.4697265625
],
[
0.080078125,
0.48095703125,
0.4697265625,
0.48095703125,
0.4697265625,
0.494384765625,
0.080078125,
0.494384765625
],
[
0.080078125,
0.496826171875,
0.340576171875,
0.496826171875,
0.340576171875,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.50341796875,
0.494384765625,
0.89208984375,
0.494384765625,
0.89208984375,
0.50830078125,
0.50341796875,
0.50830078125
],
[
0.50341796875,
0.51025390625,
0.890625,
0.51025390625,
0.890625,
0.5234375,
0.50341796875,
0.5234375
],
[
0.108642578125,
0.5166015625,
0.43701171875,
0.51806640625,
0.43701171875,
0.5361328125,
0.108642578125,
0.5341796875
],
[
0.50244140625,
0.5234375,
0.85302734375,
0.5244140625,
0.85302734375,
0.5390625,
0.50244140625,
0.53857421875
],
[
0.080078125,
0.54345703125,
0.453369140625,
0.54345703125,
0.453369140625,
0.556640625,
0.080078125,
0.556640625
],
[
0.52294921875,
0.54443359375,
0.890625,
0.54443359375,
0.890625,
0.5595703125,
0.52294921875,
0.5595703125
],
[
0.50341796875,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.080078125,
0.57568359375,
0.3505859375,
0.57568359375,
0.3505859375,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50341796875,
0.57470703125,
0.89208984375,
0.57470703125,
0.89208984375,
0.58837890625,
0.50341796875,
0.58837890625
],
[
0.50341796875,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.6015625,
0.50341796875,
0.6015625
],
[
0.080078125,
0.5986328125,
0.46728515625,
0.5986328125,
0.46728515625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50244140625,
0.60498046875,
0.8935546875,
0.60498046875,
0.8935546875,
0.61865234375,
0.50244140625,
0.61865234375
],
[
0.08087158203125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.08087158203125,
0.62890625
],
[
0.50244140625,
0.62060546875,
0.89208984375,
0.62060546875,
0.89208984375,
0.6337890625,
0.50244140625,
0.6337890625
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.62890625,
0.468017578125,
0.64306640625,
0.080078125,
0.64453125
],
[
0.5048828125,
0.63720703125,
0.8896484375,
0.63720703125,
0.8896484375,
0.64794921875,
0.5048828125,
0.64794921875
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.50244140625,
0.65087890625,
0.89208984375,
0.65087890625,
0.89208984375,
0.6640625,
0.50244140625,
0.6640625
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66650390625,
0.89208984375,
0.66650390625,
0.89208984375,
0.6796875,
0.50244140625,
0.6796875
],
[
0.08087158203125,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.6875,
0.08087158203125,
0.6875
],
[
0.5009765625,
0.6796875,
0.654296875,
0.6796875,
0.654296875,
0.693359375,
0.5009765625,
0.693359375
],
[
0.080078125,
0.68994140625,
0.449462890625,
0.68994140625,
0.449462890625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.09478759765625,
0.7041015625,
0.4697265625,
0.70263671875,
0.4697265625,
0.72021484375,
0.09478759765625,
0.7216796875
],
[
0.51953125,
0.69873046875,
0.89208984375,
0.69970703125,
0.89208984375,
0.7138671875,
0.51953125,
0.71337890625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.71630859375,
0.89208984375,
0.71630859375,
0.89208984375,
0.72998046875,
0.50244140625,
0.72998046875
],
[
0.50244140625,
0.73095703125,
0.611328125,
0.73095703125,
0.611328125,
0.744140625,
0.50244140625,
0.744140625
],
[
0.6953125,
0.73095703125,
0.89208984375,
0.73095703125,
0.89208984375,
0.744140625,
0.6953125,
0.744140625
],
[
0.62353515625,
0.7373046875,
0.6298828125,
0.7373046875,
0.6298828125,
0.74072265625,
0.62353515625,
0.74072265625
],
[
0.6611328125,
0.7373046875,
0.67333984375,
0.7373046875,
0.67333984375,
0.74072265625,
0.6611328125,
0.74072265625
],
[
0.50244140625,
0.74560546875,
0.89208984375,
0.74560546875,
0.89208984375,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.080078125,
0.736328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.080078125,
0.75
],
[
0.0767822265625,
0.74853515625,
0.418212890625,
0.7470703125,
0.418212890625,
0.7646484375,
0.0767822265625,
0.7666015625
],
[
0.098876953125,
0.76513671875,
0.468017578125,
0.7646484375,
0.468017578125,
0.77880859375,
0.098876953125,
0.77978515625
],
[
0.50244140625,
0.76123046875,
0.89208984375,
0.76123046875,
0.89208984375,
0.7744140625,
0.50244140625,
0.7744140625
],
[
0.08087158203125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.50244140625,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79248046875,
0.68115234375,
0.79248046875,
0.68115234375,
0.80322265625,
0.50244140625,
0.80322265625
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.52294921875,
0.8125,
0.890625,
0.8125,
0.890625,
0.82275390625,
0.52294921875,
0.82275390625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.455810546875,
0.84130859375,
0.455810546875,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.09722900390625,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.8720703125,
0.09722900390625,
0.8720703125
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.87060546875,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.2320556640625,
0.88720703125,
0.2320556640625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.82861328125,
0.88720703125,
0.82861328125,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"generated sample. A value of Q = 0.2 is used throughout",
"theexperiments",
"Additional losses are exercised to both retain source-",
"to-output similarity, yet drive a perceptible transformation",
"Several losses aredistributed equallybetweentherawand",
"masked outputs, imposing constraints on both. An L1 recon-",
"struction loss is used to enforce pixel-level similarity:",
"=aillzraw",
"masked",
"c",
"c",
"LD",
"where zraw is the output image itself. This results in a non",
"trivial constraint, as the encoder input image is distorted. An",
"edge-preserving loss is used to constrain pixel-level deriva-",
"tive differences in both the r and y image axes. Calculated as",
"the absolute difference between the source and output deriva",
"tives in each axis direction for both the raw and masked",
"outputs:",
"Craw = a2llzraw-alli Cmasked = a2llzmasked- aill",
"Craw =a2llzgaw -ayll Cmasked = a2llzmasked - rll",
"Figure 3. Sample results for video de-identification (zoom). Triplets",
"of source frame, converted frame and target are shown. The modi",
"where &is the derivative of the undistorted input image ",
"fied framelooks similar butthe identityis completelydifferent.",
"along the a axis, and similarly for outputs z and the y axis.",
"Additional losses are applied to the blending mask m.",
"The perceptual loss is given by:",
"where O indicates that the value of this pixel would be taken",
"from the input image &, 1 indicates taking the value from",
"Cg=lf12x112+l",
"-入et",
"",
"",
"628×28+",
"zraw, and intermediate values indicate linear mixing. We",
"56y56",
"would like the mask to be both minimal and smooth and,",
"therefore, employ the following losses:",
"for c that is either raw or masked, and where ^ > O is a",
"hyperparameter, which determines the generated face's high",
"cm=lmlcm=llmall1cm=llmyll1",
"level features distance from those of the target image.",
"where mr and my are the spatial derivatives of the mask.",
"The application of the multi-image perceptual loss during",
"training is depicted in Fig. 2(b). During training, the target is",
"3.2.1A Multi-Image Perceptual Loss",
"the source, and there is only one input image. The resulting",
"imagehasthetexture.poseand expression ofthe source,but",
"A new variant of the perceptual loss [15] is employed to",
"the face is modified to distance the identity. Note that we",
"maintain source expression, pose and lighting conditions.",
"refer to it as a multi-image perceptual loss, as its aim is to",
"while capturing the target identity essence. This is achieved",
"minimizetheanalogerrorterm duringinference(general",
"by employing a perceptual loss between the undistorted",
"ization error). However, as a training loss, it is only applied",
"source and generated output on several low-to-medium ab-",
"during train, where it receives a pair of images, similar to",
"straction layers,while distancing thehigh abstraction layer",
"other perceptual losses",
"perceptual loss between the target and generated output.",
"Let anxn be the activations of an n × n spatial block",
"Note that the perceptual loss parameters cn are normaliz-",
"within the face classifier network for image r, where in our",
"ing constants obtained by counting the number of elements.",
"In addition, Qo ",
" αg are simply set to one, and",
"",
"TP",
"Q4, Q5 were chosen arbitrarily. Therefore, there is effectively",
"case, r can be either the input image c, the target image t,",
"the raw output zraw, or the masked output zmasked.",
"We consider the spatial activations maps of size 112 x",
"only a single important hyperparameter: X, which provides",
"112, 56 × 56, 28 × 28 and 7 × 7, as well as the representation",
"a direct control of the strength of the identity distance which",
"layer of size 1 × 1. The lower layers (larger maps) are used",
"requires tuning (see Fig. 9)",
"to enforce similarity to the input image , while the 7 × 7",
"Atinferencetime,thenetworkisfed an inputframeand",
"layer is used to enforce similarity to t, and the 1 x 1 feature",
"a target image. The target image is transmitted through the",
"vector is used to enforce dissimilarity to the target image.",
"face classifier, resulting in a target feature vector, which, in",
"Let us define lnxf = Cn lari,nxn - ar2,nxnlli, where",
"turn,is concatenated tothe latentembedding space.Dueto",
"Cn is a normalizing constant, corresponding to the size of the",
"the way the network is trained, the decoder will drive the",
"spatial activation map.",
"output image away from the target feature vector."
] |
{
"arxiv_id": "1911.08348",
"language": "en",
"page_idx": 4
}
|
||
1912.09393-en-1
|
1912.09393-en-1.jpg
|
structed directly from WordNet[26], itself a hierarchy originally designed solely to represent semantic relationships between words. Shortly after ImageNet's introduction, works such as Deng et $al.[11]$,Zhao et al.[50], and Verma $et$ $al.[43]$ explicitly noted that the underpinning WordNet hierarchy suggested a way of quantifying the severity of mistakes, and experimented with hierarchical cost minimisation. Likewise, Deng $etal.[10]$ presented a straightforward method for using a hierarchy-derived similarity matrix to define a more semantically meaningful compatibility function for image retrieval. Despite this initial surge of interest and the promising results accompanying it, the community effectively discarded hierarchical measures after deciding that they were not communicating substantially different information about classifer performance than top-1 and top-5 accuracies$^{\mathrm{l}}$. When the celebrated results in Krizhevsky $et$ $al.[20]$ were reported in flat top-$k$ terms only, the precedent was firmly set for the work which followed in the deep learning era of image classification. Interest in optimising hierarchical performance measures waned accordingly.
We argue here that this problem is ripe for revisitation, and we begin by pointing to Fig. 1. Here, a mistake is defined as a top-1 prediction which differs from the groundtruth class, and the severity of such a mistake is the height of the lowest common ancestor of the predicted and groundtruth classes in the hierarchy. We see that while the flat top- 1 accuracies of state-of-the-art classifiers have improved to impressive levels over the years, the distributions of the severities of the errors that are made have changed very little over this time. We hypothesise that this is due, at least in part, to the scarcity of modern learning methods which attempt to exploit prior information and preferences about class relationships in the interest of “making better mistakes", whether this information is sourced from an offline taxonomy or otherwise. The few exceptions of which we are aware include Frome et al.[14], Wu et al.[44], Barz & Denzler [4], and a passing mention in Redmon & Farhadi[33]. In Sec. 2, we suggest a framework for thinking about these pieces of work, their predecessors, and some of their conceptual relatives.
The contributions of this work are as follows:
1. We review relevant literature within an explanatory
framework which unifies a fairly disjoint prior art.
2. Building on the perspective gained from the preced-
ing, we propose two methods that are both simple and effective at leveraging class hierarchies. Each uses a one-parameter drop-in generalisation of the standard cross-entropy loss. These loss variants can be tuned to produce different empirical tradeoffs between top-$k$
$^{1}$From Russakovsky et al.[37]:“[..] we found that all three measures oferror (top-5, top-1, and hierarchical) produced the same ordering of results. Thus, since ILSVRC2012 we have been exclusively using the top-5 metric which is the simplest and most suitable to the dataset."
and hierarchical performance measures, and reduce to
the standard setup in their respective limits.
3. We perform an extensive experimental evaluation to
both demonstrate the effectiveness of the said methods
compared to prior art and to encourage future work.
The PyTorch [29] code for all experiments is available at
github.com/fiveai/making-better-mistakes.
# 2. Framework and related work
We first suggest a simple framework for thinking about methods relevant to the problem of making better mistakes on image classification, beginning with the standard supervised setup. Consider a training set $\mathcal{S}=\{(x_i,C_i)\}_{i=1,\ldots,N}$ which pairs $N$ images $x_i\in\mathcal{I}$ with class labels $C_i\in\mathcal{C}.$ A network architecture implements the predictor function $\phi(x;\theta)$, whose parameters $\theta$ are learned by minimising
$$\begin{aligned}\frac{1}{N}\sum_{i=1}^{N}\mathcal{L}(\phi(x_i;\theta),y(C_i))+\mathcal{R}(\theta),\end{aligned}$$
(1)
where the loss function $\mathcal{L}$ compares the predictor's output $\phi(x_{i};\theta)$ to an embedded representation $y(C_{i})$ of each example's class, and $\mathcal{R}$ is a regulariser.
Under common choices such as cross-entropy for $\mathcal{L}$ and one-hot embedding for $y$,it is easy to see that the framework is agnostic of relationships between classes. The question is how such class relationships $\mathcal{H}$ can be incorporated into the loss in Eqn. 1. We identify the following three approaches:
1. Replacing class representation $y(C)$ with an alternate
embedding $y^{\mathcal{H}}(C).$ Such“label-embedding”methods, discussed in Sec. 2.1, can draw their embedding both from taxonomic hierarchies and alternative sources.
2. Altering the loss function $\mathcal{L}$ in terms of its arguments
to produce $\mathcal{L}^{\mathcal{H}}(\phi(x;\theta),y(C)),i.e.$ making the penalty assigned to a given output distribution and embedded label dependent on $\mathcal{H}.$ Methods using these“hierarchical losses" are covered in Sec. 2.2.
3. Altering the function $\phi(x;\theta)$ to $\phi^{\mathcal{H}}(x;\theta),i.e.$ mak-
ing hierarchically-informed architectural changes to the network, generally with the hope of introducing a favourable inductive bias. We cover these “hierarchical architectures" in Sec. 2.3 .
While a regulariser $\mathcal{R}^{\mathcal{H}}$ is certainly feasible, it is curiously
rare in practice: [50] is the only example we know of.
## $2. 1. \textbf{Label- embedding methods}$
These methods map class labels to vectors whose relative locations represent semantic relationships, and optimise a loss on these embedded vectors. The DeViSE method of Frome et al.[14] maps target classes onto a unit hypersphere, assigning terms with similar contexts to similar
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.53515625,
0.091552734375,
0.89208984375,
0.091552734375,
0.89208984375,
0.10479736328125,
0.53515625,
0.10479736328125
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.53515625,
0.10797119140625,
0.8203125,
0.10797119140625,
0.8203125,
0.121826171875,
0.53515625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.51318359375,
0.121826171875,
0.890625,
0.12249755859375,
0.890625,
0.136962890625,
0.51318359375,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.471435546875,
0.1383056640625,
0.471435546875,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.53515625,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.149658203125,
0.53515625,
0.149658203125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.53515625,
0.154052734375,
0.875,
0.154052734375,
0.875,
0.1673583984375,
0.53515625,
0.1673583984375
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50341796875,
0.1767578125,
0.89208984375,
0.1767578125,
0.89208984375,
0.1900634765625,
0.50341796875,
0.1900634765625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50341796875,
0.1925048828125,
0.78125,
0.1925048828125,
0.78125,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50341796875,
0.2203369140625,
0.771484375,
0.2203369140625,
0.771484375,
0.23486328125,
0.50341796875,
0.23486328125
],
[
0.08087158203125,
0.2298583984375,
0.46728515625,
0.2298583984375,
0.46728515625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.51953125,
0.242431640625,
0.8935546875,
0.2442626953125,
0.8935546875,
0.261962890625,
0.51953125,
0.260009765625
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.276611328125,
0.890625,
0.276611328125,
0.890625,
0.290283203125,
0.50244140625,
0.290283203125
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.5,
0.28857421875,
0.890625,
0.290283203125,
0.890625,
0.30810546875,
0.5,
0.30615234375
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.304931640625,
0.890625,
0.30615234375,
0.890625,
0.32080078125,
0.50244140625,
0.3193359375
],
[
0.08087158203125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50341796875,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50341796875,
0.335205078125
],
[
0.080078125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.337158203125,
0.865234375,
0.337158203125,
0.865234375,
0.35107421875,
0.50341796875,
0.35107421875
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.58642578125,
0.363037109375,
0.6259765625,
0.36865234375,
0.62158203125,
0.3837890625,
0.58251953125,
0.37890625
],
[
0.60888671875,
0.363037109375,
0.625,
0.363037109375,
0.625,
0.37255859375,
0.60888671875,
0.37255859375
],
[
0.576171875,
0.37255859375,
0.814453125,
0.363037109375,
0.81640625,
0.392578125,
0.57666015625,
0.401611328125
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.07843017578125,
0.37744140625,
0.441162109375,
0.380126953125,
0.441162109375,
0.393310546875,
0.07843017578125,
0.391357421875
],
[
0.869140625,
0.375,
0.89208984375,
0.375,
0.89208984375,
0.389404296875,
0.869140625,
0.389404296875
],
[
0.098876953125,
0.393310546875,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.098876953125,
0.4072265625
],
[
0.60546875,
0.393310546875,
0.6259765625,
0.393310546875,
0.6259765625,
0.401611328125,
0.60546875,
0.401611328125
],
[
0.07843017578125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.07843017578125,
0.423583984375
],
[
0.5048828125,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.4267578125,
0.5048828125,
0.4267578125
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.427978515625,
0.8896484375,
0.427978515625,
0.8896484375,
0.442626953125,
0.50244140625,
0.442626953125
],
[
0.08087158203125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.50244140625,
0.443115234375,
0.744140625,
0.443115234375,
0.744140625,
0.45703125,
0.50244140625,
0.45703125
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.521484375,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.472900390625,
0.521484375,
0.472900390625
],
[
0.080078125,
0.4697265625,
0.468017578125,
0.4697265625,
0.468017578125,
0.48291015625,
0.080078125,
0.48291015625
],
[
0.50244140625,
0.473388671875,
0.89208984375,
0.473388671875,
0.89208984375,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.08251953125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.08251953125,
0.498779296875
],
[
0.50244140625,
0.488037109375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.50244140625,
0.501953125
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.08087158203125,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.5009765625,
0.517578125,
0.8896484375,
0.51806640625,
0.8896484375,
0.53271484375,
0.5009765625,
0.5322265625
],
[
0.07843017578125,
0.52880859375,
0.4697265625,
0.52978515625,
0.4697265625,
0.54443359375,
0.07843017578125,
0.54345703125
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.5146484375,
0.54248046875,
0.89208984375,
0.54150390625,
0.89208984375,
0.55615234375,
0.5146484375,
0.556640625
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.53369140625,
0.55810546875,
0.89208984375,
0.556640625,
0.89208984375,
0.5712890625,
0.53369140625,
0.57275390625
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.53515625,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.53515625,
0.5859375
],
[
0.080078125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.53515625,
0.5888671875,
0.875,
0.5888671875,
0.875,
0.60302734375,
0.53515625,
0.60302734375
],
[
0.080078125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.080078125,
0.619140625
],
[
0.5146484375,
0.6015625,
0.89208984375,
0.60400390625,
0.89208984375,
0.61865234375,
0.5146484375,
0.6162109375
],
[
0.080078125,
0.62060546875,
0.468017578125,
0.62060546875,
0.468017578125,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.53515625,
0.6171875,
0.890625,
0.6171875,
0.890625,
0.6318359375,
0.53515625,
0.6318359375
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.5361328125,
0.6337890625,
0.89208984375,
0.6337890625,
0.89208984375,
0.64794921875,
0.5361328125,
0.64794921875
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.5361328125,
0.64794921875,
0.890625,
0.64794921875,
0.890625,
0.6611328125,
0.5361328125,
0.6611328125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.53515625,
0.662109375,
0.78125,
0.662109375,
0.78125,
0.67529296875,
0.53515625,
0.67529296875
],
[
0.08087158203125,
0.68310546875,
0.26708984375,
0.68310546875,
0.26708984375,
0.693359375,
0.08087158203125,
0.693359375
],
[
0.5146484375,
0.6787109375,
0.890625,
0.6787109375,
0.890625,
0.6923828125,
0.5146484375,
0.6923828125
],
[
0.098876953125,
0.6962890625,
0.398681640625,
0.6962890625,
0.398681640625,
0.71044921875,
0.098876953125,
0.71044921875
],
[
0.5361328125,
0.69580078125,
0.890625,
0.69580078125,
0.890625,
0.70556640625,
0.5361328125,
0.70556640625
],
[
0.53369140625,
0.708984375,
0.89208984375,
0.70849609375,
0.89208984375,
0.72265625,
0.53369140625,
0.7236328125
],
[
0.0906982421875,
0.72021484375,
0.468017578125,
0.7216796875,
0.468017578125,
0.736328125,
0.0906982421875,
0.73486328125
],
[
0.53515625,
0.7236328125,
0.890625,
0.7236328125,
0.890625,
0.7373046875,
0.53515625,
0.7373046875
],
[
0.11273193359375,
0.7373046875,
0.443603515625,
0.7373046875,
0.443603515625,
0.75048828125,
0.11273193359375,
0.75048828125
],
[
0.53369140625,
0.7392578125,
0.73388671875,
0.7392578125,
0.73388671875,
0.75244140625,
0.53369140625,
0.75244140625
],
[
0.0906982421875,
0.751953125,
0.465576171875,
0.751953125,
0.465576171875,
0.76513671875,
0.0906982421875,
0.76513671875
],
[
0.50244140625,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.111083984375,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.78076171875,
0.111083984375,
0.78076171875
],
[
0.11273193359375,
0.78271484375,
0.4697265625,
0.78271484375,
0.4697265625,
0.796875,
0.11273193359375,
0.796875
],
[
0.50244140625,
0.77880859375,
0.8544921875,
0.77880859375,
0.8544921875,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.11273193359375,
0.79736328125,
0.4697265625,
0.79736328125,
0.4697265625,
0.81103515625,
0.11273193359375,
0.81103515625
],
[
0.50341796875,
0.80322265625,
0.740234375,
0.80322265625,
0.740234375,
0.81640625,
0.50341796875,
0.81640625
],
[
0.11273193359375,
0.81298828125,
0.468017578125,
0.81298828125,
0.468017578125,
0.8271484375,
0.11273193359375,
0.8271484375
],
[
0.11273193359375,
0.82763671875,
0.4697265625,
0.82763671875,
0.4697265625,
0.84130859375,
0.11273193359375,
0.84130859375
],
[
0.521484375,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.521484375,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.1004638671875,
0.85107421875,
0.468017578125,
0.85107421875,
0.468017578125,
0.8642578125,
0.1004638671875,
0.8642578125
],
[
0.50244140625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50244140625,
0.869140625
],
[
0.080078125,
0.86279296875,
0.4697265625,
0.86279296875,
0.4697265625,
0.876953125,
0.080078125,
0.876953125
],
[
0.080078125,
0.87548828125,
0.468017578125,
0.87548828125,
0.468017578125,
0.8896484375,
0.080078125,
0.8896484375
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.8896484375,
0.40283203125,
0.8896484375,
0.40283203125,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.5009765625,
0.88720703125,
0.89208984375,
0.88623046875,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"structed directly from WordNet [26], itself a hierarchy origi-",
"and hierarchical performance measures, and reduce to",
"nally designed solely to represent semantic relationships be-",
"the standard setup in their respective limits.",
"tween words. Shortly after ImageNet's introduction, works",
"3. We perform an extensive experimental evaluation to",
"such as Deng et al. [11], Zhao et al. [50], and Verma et",
"bothdemonstratetheeffectivenessofthesaidmethods",
"al. [43] explicitly noted that the underpinning WordNet hi-",
"compared to prior art and to encourage future work.",
"erarchy suggestedaway ofquantifyingthe severity ofmis-",
"The PyTorch [29] code for all experiments is available at",
"takes, and experimented with hierarchical cost minimisa-",
"github.com/fiveai/making-better-mistakes.",
"tion.Likewise,Deng et al.Jlolpresented a straightforward",
"method for using a hierarchy-derived similarity matrix to",
"2. Framework and related work",
"defineamore semanticallymeaningfulcompatibilityfunc-",
"tion for image retrieval. Despite this initial surge of interest",
"We first suggest a simple framework for thinking about",
"and the promising results accompanying it, the community",
"methods relevant to the problem of making better mistakes",
"effectively discarded hierarchical measures after deciding",
"on image classification, beginning with the standard super-",
"that they were not communicating substantially different in-",
"vised setup. Consider a training set S = (ri, C,))i=],..N",
"formation about classifier performance than top-1 and top-5",
"which pairs N images r, e I with class labels C, e C.",
"accuracies.Whenthecelebratedresultsin Krizhevskyet",
"A network architecture implements the predictor function",
"al. [20] were reported in flat top-k terms only, the prece-",
"(r; 0), whose parameters are learned by minimising",
"dent was firmly set for the work which followed in the deep",
"17",
"N",
"ZC(o(r: 0), (C) + R(0),",
"learning era of image classification. Interest in optimising",
"hierarchical performancemeasures waned accordingly.",
"(1)",
"We argue here that this problem is ripe for revisitation,",
"i=1",
"and we begin by pointing to Fig. 1. Here, a mistake is de-",
"where the loss function C compares the predictor's output",
"fined as a top-1 prediction which differs from the ground-",
"(ri; 0) to an embedded representation y(C) of each ex-",
"truth class, and the severity of such a mistake is the height",
"ample's class, and R is a regulariser.",
"of the lowest common ancestor of the predicted and ground-",
"Under common choices such as cross-entropy for C and",
"truth classes in the hierarchy. We see that while the flat top-",
"one-hot embedding for y, it is easy to see that the framework",
"I accuracies of state-of-the-art classifiers have improved to",
"is agnostic of relationships between classes. The question is",
"impressive levels over the years, the distributions of the",
"how such class relationships H can be incorporated into the",
"severities of the errors that are madehavechanged very",
"loss in Eqn. 1. We identify the following three approaches:",
"little over this time.We hypothesise that this is due, at",
"least in part, to the scarcity of modern learning methods",
"1. Replacing class representation y(C) with an alternate",
"which attempt to exploit prior information and preferences",
"embedding y\"(C). Such “label-embedding\" methods,",
"about class relationships in the interest of “making better",
"discussed in Sec. 2.1, can draw their embedding both",
"mistakes\", whether this information is sourced from an of-",
"from taxonomic hierarchies and alternative sources.",
"fline taxonomy or otherwise.The few exceptions of which",
"2. Altering the loss function C in terms of its arguments",
"we are aware include Frome et al. [14], Wu et al. [44],",
"to produce CH((r; 0), y(C)), i.e. making the penalty",
"Barz & Denzler [4], and a passing mention in Redmon &",
"assigned to a given output distribution and embedded",
"Farhadi [33]. In Sec. 2, we suggest a framework for think-",
"label dependent on H. Methods using these \"hierar-",
"ing about these pieces of work, their predecessors, and some",
"chical losses\" are covered in Sec. 2.2.",
"oftheirconceptualrelatives.",
"3. Altering the function (c;0) to @H(c; 0), i.e. mak-",
"The contributions of this work are as follows:",
"ing hierarchically-informed architectural changes to",
"the network, generally with the hope of introducing a",
"1. We review relevant literature within an explanatory",
"favourable inductive bias. We cover these \"hierarchi-",
"framework which unifies a fairly disjoint prior art.",
"cal architectures\" in Sec. 2.3 .",
"2. Building on the perspective gained from the preced",
"While a regulariser RH is certainly feasible, it is curiously",
"ing, we propose two methods that are both simple and",
"effective at leveraging class hierarchies. Each uses a",
"rare in practice: [5O] is the only example we know of.",
"one-parameter drop-in generalisation of the standard",
"2.1. Label-embedding methods",
"cross-entropy loss.These loss variants can be tuned",
"to produce different empirical tradeoffs between top-k",
"These methods map class labels to vectors whose rela-",
"tive locations represent semantic relationships, and optimise",
"'From Russakovsky et al. [37]: \"[..] we found that all three measures",
"a loss on these embedded vectors. The DeViSE method",
"of error (top-5, top-l, and hierarchical) produced the same ordering of",
"results.Thus, since ILSVRC2012 we have been exclusively using the top-5",
"of Frome et al. [14] maps target classes onto a unit hy-",
"metricwhich isthe simplestand most suitable tothe dataset.",
"persphere, assigning terms with similar contexts to similar"
] |
{
"arxiv_id": "1912.09393",
"language": "en",
"page_idx": 1
}
|
||
1912.09393-en-2
|
1912.09393-en-2.jpg
|
representations through analysis of unannotated Wikipedia text[24]. The loss function is a ranking loss which penalises the extent to which the output is more cosine-similar to false label embeddings than to the correct one. They learn a linear mapping from a pre-trained visual feature pipeline to the embedded labels, then fine-tune the visual pipeline.
Romera-Paredes& Torr [35] note that their solution for learning an analogous linear mapping for zero-shot classification should easily extend to accommodating these sorts of embeddings. In Hinton et al.[18, Sec. 2], the role of the label embedding function is played by a temperaturescaled pre-existing classifier ensemble. This ensemble is “distilled" into a smaller DNN through cross-entropy minimisation against the ensemble's output. For zero-shot classification, Xian et $al.[45]$ experiment with various independent embedding methods, as is also done in Akata $et$ $al.[2]{:}$ annotated attributes, word2vec[25], glove[30], and the WordNet hierarchy. Their ranking loss function is functionally equivalent to that in Frome $etal.[14]$, and they learn a choice of linear mappings to these representations from the output of a fixed CNN. Barz &Denzler [4] present an embedding algorithm which maps examples onto a hypersphere such that all distances represent similarities derived from lowest common ancestor (LCA) height in a given hierarchy tree. They proceed by minimising the sum of two rather different losses: (1) a linear loss based on cosine distance to the embedded class vectors and (2) the standard cross-entropy loss on the output of a fully-connected layer added after the embedding layer.
## 2.2. Hierarchical losses
In these methods, the loss function itself is parametrised by the class hierarchy such that a higher penalty is assigned to the prediction of a more distant relative of the true label. Deng et al.[11] simply train kNN- and SVM-based classifiers to minimise the expected WordNet LCA height directly. Zhao et al.[50] modify standard multi-class logistic regression by replacing the output class probabilities with normalised class-similarity-weighted sums. They also regularise feature selection using an“overlapping-grouplasso penalty" which encourages the use of similar features for closely related classes, a rare example of a hierarchical regulariser. Verma $etal.[43]$ incorporate normalised LCA height into a“context-sensitive loss function"while learning a separate metric at each node in a taxonomy tree for nearest-neighbour classification. Wu et al.[44] implement granular classification of food images by sharing a standard deep network backbone between multiple fully-connected layers, each one outputting class probabilities at its respective hierarchy level. A separate label propagation step is used to smooth inconsistencies in the resulting marginal probabilities. Alsallakh et al.[5] likewise use a standard deep architecture as their starting point,but instead add
branches strategically to intermediate pipeline stages. They thereby force the net to classify into offline-determined superclasses at the respective levels, backpropagating error in $\hat{\text{these intermediate predictions accordingly. }}$ At test time, these additions are simply discarded.
## 2.3. Hierarchical architectures
These methods attempt to incorporate class hierarchy into the classifier architecture without necessarily changing the loss function otherwise. The core idea is to“divide and conquer" at the structural level, with the classifier assigning inputs to superclasses at earlier layers and making finegrained distinctions at later ones. In the context of language models, it was noted at least as early as Goodman [15] that classification with respect to an IS-A hierarchy tree could be formulated as a tree of classifiers outputting conditional probabilities, with the product of the conditionals along a given leaf's ancestry representing its posterior; motivated by efficiency, Morin & Bengio [27] applied this observation to a binary hierarchy derived from WordNet. Redmon & Farhadi [33] propose a modern deep-learning variant of this framework in the design of the YOLOv2 object detection and classification system. Using a version of WordNet pruned into a tree, they effectively train a conditional classifer at every parent node in the tree by using one softmax layer per sibling group and training under the usual crossentropy loss over leaf posteriors. While their main aim is to enable the integration of the COCO detection dataset with ImageNet, they suggest that graceful degradation on new or unknown object categories might be an incidental benefit. Brust & Denzler [7] propose an extension of conditional classifer chains to the more general case of DAGs.
The above approaches can be seen as a limiting case of hierarchical classification, in which every split in the hierarchy is cast as a separate classification problem. Many hierarchical classifers fall between this extreme and that of flat classification, working in terms of a coarser-grained conditionality in which a“generalist”makes assignments to groupings of the target classes before then distinguishing the group members from one another using “experts”. Xiao et al.[46], the quasi-ensemble section of Hinton $et$ $al.[18$, Sec.5], Yan $etal.[47]$, and Ahmed $etal.[1]$ all represent modern variations on this theme (which first appears no later than[19]). Additionally, the listed methods all use some form of low-level feature sharing either via architectural constraint or parameter cloning, and all infer the visual hierarchy dynamically through confusion clustering or latent parameter inference. Alsallakh et al.[5] make the one proposal of which we are aware which combines hierarchical architectural modifications (at train time) with a hierarchical loss, as described in Sec. 2.2. At test time, however, the architecture is that of an unmodified AlexNet, and all superclass“assignment”is purely implicit.
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
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,
0.471435546875,
0.10797119140625,
0.471435546875,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
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,
0.89208984375,
0.1383056640625,
0.89208984375,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.5009765625,
0.1514892578125,
0.74609375,
0.1514892578125,
0.74609375,
0.1654052734375,
0.5009765625,
0.1654052734375
],
[
0.080078125,
0.1685791015625,
0.416748046875,
0.1685791015625,
0.416748046875,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.09722900390625,
0.184326171875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1976318359375,
0.09722900390625,
0.1988525390625
],
[
0.50341796875,
0.18115234375,
0.73779296875,
0.18115234375,
0.73779296875,
0.1956787109375,
0.50341796875,
0.1956787109375
],
[
0.080078125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.21337890625,
0.080078125,
0.21337890625
],
[
0.51953125,
0.203857421875,
0.890625,
0.205810546875,
0.890625,
0.2203369140625,
0.51953125,
0.2183837890625
],
[
0.080078125,
0.2139892578125,
0.4697265625,
0.2139892578125,
0.4697265625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.5,
0.2197265625,
0.89208984375,
0.2215576171875,
0.89208984375,
0.23486328125,
0.5,
0.23291015625
],
[
0.080078125,
0.2298583984375,
0.471435546875,
0.2298583984375,
0.471435546875,
0.24365234375,
0.080078125,
0.24365234375
],
[
0.50244140625,
0.236083984375,
0.89208984375,
0.236083984375,
0.89208984375,
0.25,
0.50244140625,
0.25
],
[
0.080078125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.080078125,
0.25830078125
],
[
0.50244140625,
0.251953125,
0.890625,
0.251953125,
0.890625,
0.26513671875,
0.50244140625,
0.26513671875
],
[
0.080078125,
0.260009765625,
0.4697265625,
0.260009765625,
0.4697265625,
0.2734375,
0.080078125,
0.2734375
],
[
0.07843017578125,
0.2734375,
0.4697265625,
0.274658203125,
0.4697265625,
0.2890625,
0.07843017578125,
0.28857421875
],
[
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
],
[
0.07843017578125,
0.290283203125,
0.4697265625,
0.290283203125,
0.4697265625,
0.3037109375,
0.07843017578125,
0.3037109375
],
[
0.50244140625,
0.296630859375,
0.89208984375,
0.296630859375,
0.89208984375,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.07843017578125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.07843017578125,
0.3193359375
],
[
0.50341796875,
0.313232421875,
0.890625,
0.313232421875,
0.890625,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50244140625,
0.32568359375,
0.89208984375,
0.32568359375,
0.89208984375,
0.339111328125,
0.50244140625,
0.339111328125
],
[
0.080078125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50341796875,
0.343505859375,
0.890625,
0.343505859375,
0.890625,
0.354248046875,
0.50341796875,
0.354248046875
],
[
0.080078125,
0.35107421875,
0.4697265625,
0.35107421875,
0.4697265625,
0.3642578125,
0.080078125,
0.3642578125
],
[
0.50341796875,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.50341796875,
0.37060546875
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50244140625,
0.37255859375,
0.89208984375,
0.37255859375,
0.89208984375,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.080078125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.080078125,
0.39453125
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.50341796875,
0.40283203125,
0.89208984375,
0.40283203125,
0.89208984375,
0.416748046875,
0.50341796875,
0.416748046875
],
[
0.080078125,
0.410888671875,
0.46728515625,
0.410888671875,
0.46728515625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50244140625,
0.43310546875,
0.89208984375,
0.43310546875,
0.89208984375,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.080078125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.447509765625,
0.89208984375,
0.447509765625,
0.89208984375,
0.4609375,
0.50244140625,
0.4609375
],
[
0.080078125,
0.45703125,
0.4697265625,
0.45703125,
0.4697265625,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.5009765625,
0.462158203125,
0.89208984375,
0.4609375,
0.89208984375,
0.47607421875,
0.5009765625,
0.4765625
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.4716796875,
0.468017578125,
0.48486328125,
0.080078125,
0.48486328125
],
[
0.50244140625,
0.477783203125,
0.890625,
0.477783203125,
0.890625,
0.4912109375,
0.50244140625,
0.4912109375
],
[
0.07843017578125,
0.486083984375,
0.468017578125,
0.486083984375,
0.468017578125,
0.5,
0.07843017578125,
0.5
],
[
0.50244140625,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.50244140625,
0.505859375
],
[
0.080078125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.080078125,
0.51513671875
],
[
0.50244140625,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50244140625,
0.521484375
],
[
0.0767822265625,
0.51513671875,
0.29736328125,
0.5166015625,
0.29736328125,
0.53076171875,
0.0767822265625,
0.52978515625
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.080078125,
0.54345703125,
0.259033203125,
0.54443359375,
0.259033203125,
0.5595703125,
0.080078125,
0.55810546875
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.50244140625,
0.5517578125,
0.89208984375,
0.55322265625,
0.89208984375,
0.5673828125,
0.50244140625,
0.56640625
],
[
0.098876953125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.58349609375,
0.098876953125,
0.58349609375
],
[
0.50341796875,
0.568359375,
0.83642578125,
0.568359375,
0.83642578125,
0.58203125,
0.50341796875,
0.58203125
],
[
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
],
[
0.08087158203125,
0.6015625,
0.46728515625,
0.6015625,
0.46728515625,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.5009765625,
0.5986328125,
0.890625,
0.599609375,
0.890625,
0.6142578125,
0.5009765625,
0.61279296875
],
[
0.080078125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50244140625,
0.62890625
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.080078125,
0.64453125
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.08087158203125,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
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
],
[
0.5009765625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.5009765625,
0.673828125
],
[
0.08087158203125,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.6875,
0.08087158203125,
0.6875
],
[
0.5009765625,
0.6748046875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.07843017578125,
0.689453125,
0.468017578125,
0.68994140625,
0.468017578125,
0.705078125,
0.07843017578125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5009765625,
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,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85595703125,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5009765625,
0.869140625
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
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,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.830078125,
0.88720703125,
0.830078125,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"representations through analysis of unannotated Wikipedia",
"branches strategically to intermediate pipeline stages. They",
"text [24]. The loss function is a ranking loss which penalises",
"therebyforcethenetto classifyinto offline-determined su-",
"theextenttowhichtheoutputismorecosine-similartofalse",
"perclasses at the respective levels, backpropagating error in",
"label embeddings than to the correct one. They learn a lin-",
"these intermediate predictions accordingly. At test time,",
"earmappingfromapre-trainedvisualfeaturepipelinetothe",
"these additions are simply discarded.",
"embedded labels, then fine-tune the visual pipeline.",
"Romera-Paredes & Torr [35] note that their solution for",
"2.3.Hierarchical architectures",
"learning an analogous linear mapping for zero-shot classi-",
"These methods attempt to incorporate class hierarchy",
"fication should easily extend to accommodating these sorts",
"into the classifier architecture without necessarily changing",
"of embeddings. In Hinton et al. [18, Sec. 2], the role of",
"the loss function otherwise. The core idea is to “divide and",
"the label embedding function is played by a temperature-",
"conquer\" at the structural level, with the classifier assign-",
"scaled pre-existing classifier ensemble. This ensemble is",
"\"distilled\" into a smaller DNN through cross-entropy min-",
"ing inputs to superclasses at earlier layers and making fine-",
"grained distinctions at later ones. In the context of language",
"imisation against the ensemble's output. For zero-shot clas-",
"models, it was noted at least as early as Goodman [15] that",
"sification, Xian et al. [45] experiment with various inde-",
"classification with respect to an IS-A hierarchy tree could",
"pendent embedding methods, as is also done in Akata et",
"be formulated as a tree of classifiers outputting conditional",
"al. [2]: annotated attributes, word2vec [25], glove [30], and",
"probabilities, with the product of the conditionals along a",
"the WordNet hierarchy. Their ranking loss function is func-",
"given leaf's ancestry representing its posterior; motivated",
"tionally equivalent to that in Frome et al. [14], and they",
"by efficiency, Morin & Bengio [27] applied this observa-",
"learn a choice of linear mappings to these representations",
"tion to a binary hierarchy derived from WordNet. Redmon",
"from the output of a fixed CNN. Barz & Denzler [4] present",
"& Farhadi [33] propose a modern deep-learning variant of",
"an embedding algorithm which maps examples onto a hy-",
"this framework in the design of the YOLOv2 object detec-",
"persphere such that all distances represent similarities de-",
"tion and classification system. Using a version of WordNet",
"rived from lowest common ancestor (LCA) height in a given",
"pruned into a tree, they effectively train a conditional clas-",
"hierarchy tree. They proceed by minimising the sum of two",
"sifier at every parent node in the tree by using one softmax",
"rather different losses: (1) a linear loss based on cosine dis-",
"layer per sibling group and training under the usual cross-",
"tance to the embedded class vectors and (2) the standard",
"entropy loss over leaf posteriors. While their main aim is to",
"cross-entropy loss on the output of a fully-connected layer",
"enable the integration of the COCO detection dataset with",
"added after the embedding layer.",
"ImageNet, they suggest that graceful degradation on new",
"2.2.Hierarchical losses",
"or unknown object categories might be an incidental bene-",
"fit. Brust & Denzler [7] propose an extension of conditional",
"In these methods, the loss function itself is parametrised",
"classifier chains to the more general case of DAGs.",
"by the class hierarchy such that a higher penalty is assigned",
"The above approaches can be seen as a limiting case of",
"to the prediction of a more distant relative of the true la-",
"hierarchical classification, in which every split in the hier-",
"bel. Deng et al. [1i] simply train kNN- and SVM-based",
"archy is cast as a separate classification problem. Many",
"classifiers to minimise the expected WordNet LCA height",
"hierarchical classifiers fall between this extreme and that",
"directly. Zhao et al.[50l modify standard multi-class lo-",
"of flat classification, working in terms of a coarser-grained",
"gistic regression by replacing the output class probabilities",
"conditionality in which a “generalist\"makes assignments",
"with normalised class-similarity-weighted sums.Theyalso",
"to groupings of the target classes before then distinguish-",
"regularise feature selection using an “overlapping-group-",
"ing the group members from one another using “experts\".",
"lasso penalty\" which encourages the use of similar features",
"Xiao et al. [46], the quasi-ensemble section of Hinton et",
"for closely related classes, a rare example of a hierarchical",
"al. [18, Sec. 5], Yan et al. [47], and Ahmed et al. [1] all",
"regulariser. Verma et al. [43] incorporate normalised LCA",
"represent modern variations on this theme (which first ap-",
"height into a context-sensitive loss function\" while learn-",
"pears no later than [19]). Additionally, the listed methods",
"ing a separate metric at each node in a taxonomy tree for",
"all use some form of low-level feature sharing either via",
"nearest-neighbour classification.Wu et al.441implement",
"architectural constraint or parameter cloning,and all infer",
"granular classification of food images by sharing a standard",
"the visual hierarchy dynamically through confusion cluster-",
"deep network backbone between multiple fully-connected",
"ing or latent parameter inference. Alsallakh et al. [5] make",
"layers, each one outputting class probabilities at its respec-",
"the one proposal of which we are aware which combines",
"tive hierarchy level. A separate label propagation step is",
"hierarchical architecturalmodifications(attraintime)with",
"used to smooth inconsistencies in the resulting marginal",
"a hierarchical loss, as described in Sec. 2.2. At test time,",
"probabilities. Alsallakh et al. [5] likewise use a standard",
"however, the architecture is that of an unmodified AlexNet,",
"deep architecture as their starting point, but instead add",
"and all superclass “assignment\" is purely implicit."
] |
{
"arxiv_id": "1912.09393",
"language": "en",
"page_idx": 2
}
|
||
1912.09393-en-3
|
1912.09393-en-3.jpg
|
## $3. \textbf{Method}$
We now outline two simple methods that allow to leverage class hierarchies in order to make better mistakes on image classification. We concentrate on the case where the output of the network is a categorical distribution over classes for each input image and denote the corresponding distribution as $p(C)=\phi_{C}(x;\theta)$, where subscripts denote vector indices and $x$ and $\theta$ are omitted. In Sec. 3.1, we describe the hierarchical cross-entropy (HXE), a straightforward example of the hierarchical losses reviewed in Sec. 2.2. This approach expands each class probability into the chain of conditional probabilities defined by its lineage in a given hierarchy tree. It then reweights the corresponding terms in the loss so as to penalise classification mistakes in a way that is informed by the hierarchy. In Sec. 3.2, we suggest an easy choice of embedding function to implement the label-embedding framework of Sec. 2.1. The resulting soft labels are PMFs over $\mathcal{C}$ whose values decay exponentially w.r.t. an LCA-based distance to the ground truth.
## 3.1. Hierarchical cross-entropy
When the hierarchy $\mathcal{H}$ is a tree, it corresponds to a unique factorisation of the categorical distribution $p(C)$ over classes in terms of the conditional probabilities along the path connecting each class to the root of the tree. Denoting the path from a leaf node $C$ to the root $R$ as $C^(0)=$ $C, \ldots , C^{( \hat{h} ) }$ = $R$, the probability of class $C$ can be factorised as
$$\begin{aligned}p(C)=\prod_{l=0}^{h-1}p(C^{(l)}|C^{(l+1)}),\end{aligned}$$
(2)
where $h\equiv h(C)$ is the depth of node $C.$ Note that we have omitted the last term $p(\dot{C}^{(h)})=1.$ The conditionals can conversely be written in terms of the class probabilities as
$$p(C^{(l)}|C^{(l+1)})=\frac{\sum_{A\in\mathrm{Leaves}(C^{(l)})}p(A)}{\sum_{B\in\mathrm{Leaves}(C^{(l+1)})}p(B)},$$
(3)
where Leaves$(C)$ denotes the set of leaf nodes of the subtree
starting at node $C.$
A direct way to incorporate hierarchical information in the loss is to hierarchically factorise the output of the classifier according to Eqn. 2 and define the total loss as the reweighted sum of the cross-entropies of the conditional probabilities. This leads us to define the hierarchical crossentropy (HXE) as
$$\mathcal{L}_{\mathrm{HXE}}(p,C)=-\sum_{l=0}^{h-1}\lambda(C^{(l)})\log p(C^{(l)}|C^{(l+1)}),$$
(4)
where $\lambda(C^{(l)})$ is the weight associated with the edge node $C^{(l+1)}\to C^{(l)}$,see Fig. 2a. Although this loss is expressed in terms of conditional probabilities, it can easily be applied
to models that output class probabilities using Eqn. 3. Note that $\mathcal{L}_\mathrm{HXE}$ reduces to the standard cross-entropy when all weights are equal to 1. This limit case, which was briefly mentioned by Redmon &Farhadi in their YOLO-v2 paper [33], results only in architectural changes but does not incorporate hierarchical information in the loss directly.
Eqn. 4 has an interesting information-theoretical interpretation: since each term $\log p(C^{(l)}|C^{(l+1)})$ corresponds to the the information associated with the edge $C^{(l+1)}\to$ $C^{(l)}$ in the hierarchy, the HXE corresponds to discounting the information associated with each of these edges differently. Note that since the HXE is expressed in terms of conditional probabilities, the reweighting in Eqn.4 is not equivalent to reweighting the cross-entropy for each possible ground truth class independently (as done, for instance, in [21,9]). A sensible choice for the weights is to take
$$\lambda(C)=\exp(-\alpha h(C)),$$
(5)
where $h(C)$ is the depth of node $C$ and $\alpha>0$ is a hyperparameter that controls the extent to which information is discounted down the hierarchy. The higher the value of $\alpha$, the higher the preference for“generic”as opposed to “fine-grained” information, because classification errors related to nodes further away from the root receive a lower loss. While such a definition has the advantages of interpretability and simplicity, one could think of other meaningful weightings $(e.g.$ based on the branching factor of the hierarchy tree). We concentrate on Eqn. 5 here, while leaving the exploration of different strategies for future work.
### $3. 2. \textbf{Soft labels}$
Our second approach to incorporating hierarchical information, soft labels, is a label-embedding approach as described in Sec. 2.1. These methods use a mapping function $y(C)$ to associate classes with representations which encode class-relationship information that is absent in the trivial case of the one-hot representation. In the interest of simplicity, we choose a mapping function $y^\mathrm{soft}(C)$ which outputs a categorical distribution over the classes. This enables us to simply use the standard cross-entropy loss:
(6)
$$\mathcal{L}_{\mathrm{Soft}}(p,C)=-\sum_{A\in\mathcal{C}}y_{A}^{\mathrm{soft}}(C)\log p(A),$$
where the soft label embedding is given componentwise by
(7)
$$y_A^{\mathrm{soft}}(C)=\frac{\exp(-\beta d(A,C))}{\sum_{B\in\mathcal{C}}\exp(-\beta d(B,C))},$$
for class distance function $d$ and parameter $\beta.$ This loss is illustrated in Fig. 2b. For the distance function $d(C_i,C_j)$, we use the height of LCA$(C_i,C_j)$ divided by the height of the tree. To understand the role of the hyperparameter $\beta$,
|
[
[
0.080078125,
0.09027099609375,
0.1666259765625,
0.09027099609375,
0.1666259765625,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.12249755859375,
0.50244140625,
0.12249755859375
],
[
0.1004638671875,
0.11614990234375,
0.468017578125,
0.11614990234375,
0.468017578125,
0.1300048828125,
0.1004638671875,
0.1300048828125
],
[
0.50341796875,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.50341796875,
0.1363525390625
],
[
0.07843017578125,
0.1319580078125,
0.468017578125,
0.1307373046875,
0.468017578125,
0.14453125,
0.07843017578125,
0.1458740234375
],
[
0.5009765625,
0.1363525390625,
0.890625,
0.136962890625,
0.890625,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.146484375,
0.4697265625,
0.146484375,
0.4697265625,
0.160400390625,
0.080078125,
0.160400390625
],
[
0.5009765625,
0.15283203125,
0.89208984375,
0.1514892578125,
0.89208984375,
0.1666259765625,
0.5009765625,
0.1673583984375
],
[
0.080078125,
0.1622314453125,
0.4697265625,
0.1622314453125,
0.4697265625,
0.175537109375,
0.080078125,
0.175537109375
],
[
0.50341796875,
0.169189453125,
0.86669921875,
0.169189453125,
0.86669921875,
0.179931640625,
0.50341796875,
0.179931640625
],
[
0.080078125,
0.175537109375,
0.468017578125,
0.175537109375,
0.468017578125,
0.189453125,
0.080078125,
0.189453125
],
[
0.521484375,
0.18310546875,
0.8896484375,
0.18310546875,
0.8896484375,
0.1962890625,
0.521484375,
0.1962890625
],
[
0.080078125,
0.1912841796875,
0.468017578125,
0.1912841796875,
0.468017578125,
0.205810546875,
0.080078125,
0.205810546875
],
[
0.50244140625,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.080078125,
0.20703125,
0.4697265625,
0.20703125,
0.4697265625,
0.2203369140625,
0.080078125,
0.2203369140625
],
[
0.5009765625,
0.21337890625,
0.86767578125,
0.2108154296875,
0.86767578125,
0.2259521484375,
0.5009765625,
0.2279052734375
],
[
0.07843017578125,
0.2203369140625,
0.468017578125,
0.2215576171875,
0.468017578125,
0.236083984375,
0.07843017578125,
0.23486328125
],
[
0.87353515625,
0.2183837890625,
0.88623046875,
0.2183837890625,
0.88623046875,
0.2215576171875,
0.87353515625,
0.2215576171875
],
[
0.50244140625,
0.2266845703125,
0.89208984375,
0.2279052734375,
0.89208984375,
0.242431640625,
0.50244140625,
0.2412109375
],
[
0.080078125,
0.2374267578125,
0.4697265625,
0.2374267578125,
0.4697265625,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.50244140625,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.251953125,
0.468017578125,
0.251953125,
0.468017578125,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.08087158203125,
0.2841796875,
0.465576171875,
0.2841796875,
0.465576171875,
0.294189453125,
0.08087158203125,
0.294189453125
],
[
0.080078125,
0.29736328125,
0.4697265625,
0.29736328125,
0.4697265625,
0.311279296875,
0.080078125,
0.311279296875
],
[
0.50341796875,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.3017578125,
0.50341796875,
0.3017578125
],
[
0.5009765625,
0.302978515625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.5009765625,
0.31640625
],
[
0.08087158203125,
0.314453125,
0.46728515625,
0.314453125,
0.46728515625,
0.324462890625,
0.08087158203125,
0.324462890625
],
[
0.5009765625,
0.318115234375,
0.859375,
0.317626953125,
0.859375,
0.33203125,
0.5009765625,
0.333251953125
],
[
0.080078125,
0.32763671875,
0.468017578125,
0.32763671875,
0.468017578125,
0.341552734375,
0.080078125,
0.341552734375
],
[
0.07843017578125,
0.341552734375,
0.468017578125,
0.34228515625,
0.468017578125,
0.356689453125,
0.07843017578125,
0.35595703125
],
[
0.615234375,
0.347900390625,
0.77783203125,
0.347900390625,
0.77783203125,
0.3623046875,
0.615234375,
0.3623046875
],
[
0.865234375,
0.347900390625,
0.8935546875,
0.3447265625,
0.896484375,
0.361083984375,
0.86767578125,
0.3642578125
],
[
0.07843017578125,
0.356689453125,
0.468017578125,
0.35595703125,
0.468017578125,
0.37060546875,
0.07843017578125,
0.371826171875
],
[
0.080078125,
0.37255859375,
0.43701171875,
0.37255859375,
0.43701171875,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.50244140625,
0.375,
0.890625,
0.375732421875,
0.890625,
0.39013671875,
0.50244140625,
0.389404296875
],
[
0.07843017578125,
0.395751953125,
0.318603515625,
0.396484375,
0.318603515625,
0.410888671875,
0.07843017578125,
0.410400390625
],
[
0.5009765625,
0.391357421875,
0.89208984375,
0.39013671875,
0.89208984375,
0.40478515625,
0.5009765625,
0.406005859375
],
[
0.50244140625,
0.4072265625,
0.8935546875,
0.4072265625,
0.8935546875,
0.42041015625,
0.50244140625,
0.42041015625
],
[
0.098876953125,
0.419189453125,
0.468017578125,
0.419189453125,
0.468017578125,
0.43310546875,
0.098876953125,
0.43310546875
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.080078125,
0.43505859375,
0.46728515625,
0.43505859375,
0.46728515625,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.50341796875,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.50341796875,
0.45068359375
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.50244140625,
0.451904296875,
0.89208984375,
0.451904296875,
0.89208984375,
0.46533203125,
0.50244140625,
0.46533203125
],
[
0.08087158203125,
0.46728515625,
0.465576171875,
0.46728515625,
0.465576171875,
0.477783203125,
0.08087158203125,
0.477783203125
],
[
0.50244140625,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.07843017578125,
0.48095703125,
0.46728515625,
0.479248046875,
0.46728515625,
0.492431640625,
0.07843017578125,
0.494384765625
],
[
0.50244140625,
0.48291015625,
0.890625,
0.481689453125,
0.890625,
0.49560546875,
0.50244140625,
0.496826171875
],
[
0.0767822265625,
0.492431640625,
0.4697265625,
0.49365234375,
0.4697265625,
0.51123046875,
0.0767822265625,
0.51025390625
],
[
0.50341796875,
0.49755859375,
0.89208984375,
0.49755859375,
0.89208984375,
0.51123046875,
0.50341796875,
0.51123046875
],
[
0.080078125,
0.51220703125,
0.1446533203125,
0.51220703125,
0.1446533203125,
0.5234375,
0.080078125,
0.5234375
],
[
0.50341796875,
0.51220703125,
0.89208984375,
0.51220703125,
0.89208984375,
0.52587890625,
0.50341796875,
0.52587890625
],
[
0.238525390625,
0.5234375,
0.259033203125,
0.5234375,
0.259033203125,
0.5322265625,
0.238525390625,
0.5322265625
],
[
0.50244140625,
0.52783203125,
0.87744140625,
0.52783203125,
0.87744140625,
0.54150390625,
0.50244140625,
0.54150390625
],
[
0.1748046875,
0.5341796875,
0.366943359375,
0.52783203125,
0.36767578125,
0.5498046875,
0.176513671875,
0.55615234375
],
[
0.4453125,
0.5361328125,
0.4697265625,
0.5361328125,
0.4697265625,
0.5498046875,
0.4453125,
0.5498046875
],
[
0.238525390625,
0.5546875,
0.2607421875,
0.5546875,
0.2607421875,
0.56298828125,
0.238525390625,
0.56298828125
],
[
0.50341796875,
0.55322265625,
0.61669921875,
0.55322265625,
0.61669921875,
0.56640625,
0.50341796875,
0.56640625
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.521484375,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.521484375,
0.59033203125
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.5009765625,
0.59130859375,
0.890625,
0.59228515625,
0.890625,
0.60693359375,
0.5009765625,
0.60595703125
],
[
0.080078125,
0.599609375,
0.461669921875,
0.599609375,
0.461669921875,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50341796875,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50341796875,
0.6220703125
],
[
0.259033203125,
0.62255859375,
0.404296875,
0.62255859375,
0.404296875,
0.64453125,
0.259033203125,
0.64453125
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.1298828125,
0.63330078125,
0.238525390625,
0.62890625,
0.240234375,
0.646484375,
0.131591796875,
0.65087890625
],
[
0.44677734375,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.44677734375,
0.6494140625
],
[
0.250732421875,
0.64306640625,
0.4150390625,
0.64013671875,
0.4150390625,
0.662109375,
0.250732421875,
0.66552734375
],
[
0.50341796875,
0.6396484375,
0.8896484375,
0.6396484375,
0.8896484375,
0.6494140625,
0.50341796875,
0.6494140625
],
[
0.50341796875,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.50341796875,
0.66650390625
],
[
0.08087158203125,
0.671875,
0.4697265625,
0.671875,
0.4697265625,
0.68505859375,
0.08087158203125,
0.68505859375
],
[
0.5009765625,
0.66748046875,
0.89208984375,
0.66650390625,
0.89208984375,
0.68115234375,
0.5009765625,
0.681640625
],
[
0.080078125,
0.6875,
0.205078125,
0.6875,
0.205078125,
0.70068359375,
0.080078125,
0.70068359375
],
[
0.5009765625,
0.68310546875,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.5009765625,
0.6962890625
],
[
0.098876953125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71533203125,
0.098876953125,
0.71533203125
],
[
0.50244140625,
0.69873046875,
0.85888671875,
0.69873046875,
0.85888671875,
0.7119140625,
0.50244140625,
0.7119140625
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.56884765625,
0.72607421875,
0.82421875,
0.72607421875,
0.82421875,
0.74365234375,
0.56884765625,
0.74365234375
],
[
0.080078125,
0.73291015625,
0.468017578125,
0.73291015625,
0.468017578125,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.869140625,
0.72900390625,
0.8935546875,
0.72900390625,
0.8935546875,
0.7421875,
0.869140625,
0.7421875
],
[
0.080078125,
0.74755859375,
0.4697265625,
0.74755859375,
0.4697265625,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.67724609375,
0.74560546875,
0.70751953125,
0.74560546875,
0.70751953125,
0.75830078125,
0.67724609375,
0.75830078125
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.767578125,
0.8896484375,
0.76806640625,
0.8896484375,
0.78271484375,
0.50244140625,
0.7822265625
],
[
0.080078125,
0.7763671875,
0.19775390625,
0.7763671875,
0.19775390625,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.67236328125,
0.79541015625,
0.7900390625,
0.79541015625,
0.7900390625,
0.81005859375,
0.67236328125,
0.81005859375
],
[
0.57177734375,
0.80322265625,
0.63232421875,
0.7998046875,
0.6337890625,
0.818359375,
0.57275390625,
0.8212890625
],
[
0.86767578125,
0.80322265625,
0.89453125,
0.80322265625,
0.89453125,
0.8193359375,
0.86767578125,
0.8193359375
],
[
0.2303466796875,
0.81005859375,
0.471435546875,
0.81103515625,
0.471435546875,
0.83203125,
0.2303466796875,
0.83056640625
],
[
0.65283203125,
0.814453125,
0.81640625,
0.81103515625,
0.81640625,
0.82568359375,
0.65283203125,
0.8291015625
],
[
0.2181396484375,
0.833984375,
0.238525390625,
0.83203125,
0.240234375,
0.84228515625,
0.2197265625,
0.8447265625
],
[
0.10540771484375,
0.81494140625,
0.20751953125,
0.81298828125,
0.20751953125,
0.8291015625,
0.10540771484375,
0.83056640625
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.5048828125,
0.85302734375
],
[
0.08087158203125,
0.85498046875,
0.4697265625,
0.85498046875,
0.4697265625,
0.869140625,
0.08087158203125,
0.869140625
],
[
0.5009765625,
0.85498046875,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.5009765625,
0.869140625
],
[
0.0767822265625,
0.8671875,
0.47216796875,
0.869140625,
0.47216796875,
0.88720703125,
0.0767822265625,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"3. Method",
"to models that output class probabilities using Eqn. 3. Note",
"that CHxe reduces to the standard cross-entropy when all",
"We now outline two simple methods that allow to lever-",
"weights are equal to 1. This limit case, which was briefly",
"age class hierarchies in order to make better mistakes on",
"mentioned by Redmon & Farhadi in their YOLO-v2 pa-",
"image classification.We concentrate on the case where",
"per [33], results only in architectural changes but does not",
"the output of the network is a categorical distribution over",
"incorporatehierarchicalinformationinthelossdirectly.",
"classes for each input image and denote the corresponding",
"Eqn. 4 has an interesting information-theoretical inter-",
"distribution as p(C) = pc(c; 0), where subscripts denote",
"pretation: since each term log p(C()|C(t+1)) corresponds",
"vector indices and and are omitted. In Sec. 3., we",
"to the the information associated with the edge C(t+1)",
"describe the hierarchical cross-entropy (HXE), a straight-",
"",
"C(t) in the hierarchy, the HXE coresponds to discounting",
"forward example of the hierarchical losses reviewed in",
"the information associated with each of these edges differ-",
"Sec. 2.2. This approach expands each class probability into",
"ently. Note that since the HXE is expressed in terms of",
"the chain of conditional probabilities defined by its lineage",
"conditional probabilities, the reweighting in Eqn. 4 is not",
"inagivenhierarchytree.Itthenreweightsthecorrespond-",
"ing terms in the loss so as to penalise classification mistakes",
"equivalenttoreweightingthecross-entropyforeachpossi-",
"bleground truthclass independently (as done, forinstance,",
"inawaythatisinformedbythehierarchy.InSec.3.2,we",
"in [21, 9]. A sensible choice for the weights is to take",
"suggest an easy choice of embedding function to implement",
"the label-embedding framework of Sec. 2.1. The resulting",
"() = exp(-αh(C),",
"(5)",
"soft labels are PMFs over C whose values decay exponen-",
"tially w.r.t. an LCA-based distance to the ground truth.",
"where h(C) is the depth of node C and α > O is a hy-",
"3.1.Hierarchical cross-entropy",
"perparameter that controls the extent to which information",
"is discounted down the hierarchy. The higher the value of",
"When the hierarchy H is a tree, it corresponds to a",
"α, the higher the preference for \"generic\" as opposed to",
"unique factorisation of the categorical distribution p(C)",
"“\"fine-grained\" information, because classification errors re-",
"over classes in terms of the conditional probabilities along",
"lated to nodes further away from the root receive a lower",
"the path connecting each class to therootof the tree.De-",
"loss. While such a definition has the advantages of inter-",
"noting thepath from a leaf node C to the root R as C(o) =",
"pretability and simplicity, one could think of other mean-",
"C,..., C(h) = R, the probability of class C can be fac-",
"ingful weightings (e.g. based on the branching factor of the",
"torised as",
"hierarchy tree). We concentrate on Eqn. 5 here, while leav-",
"h-1",
"ing the exploration of different strategies for future work.",
"p(C) = II p(C()jc(+1),",
"(2)",
"1=0",
"3.2. Soft labels",
"where h = h(C) is the depth of node C. Note that we have",
"Our second approach to incorporating hierarchical infor-",
"omitted the last term p(C(h)) = 1. The conditionals can",
"mation, soft labels, is a label-embedding approach as de-",
"conversely be written in terms of the class probabilities as",
"scribed in Sec. 2.1. These methods use a mapping func-",
"ZAeLeaves(c() (A)",
"tion y(C) to associate classes with representations which",
"p(C()c(I+1) :",
"(3)",
"ZBeLeave(C(+1) P(B)",
"encodeclass-relationshipinformationthat is absentin the",
"trivial case of the one-hot representation. In the interest of",
"where Leaves(C) denotes the set of leaf nodes of the subtree",
"simplicity, we choose a mapping function ysoft(C)which",
"starting at node C.",
"outputs a categorical distribution over the classes. This en-",
"A direct way to incorporate hierarchical information in",
"ables us to simply use the standard cross-entropy loss:",
"the loss is to hierarchically factorise the output of the clas-",
"Cson(p, C) = - yof(C) log p(A),",
"sifier according to Eqn. 2 and define the total loss as the",
"(6)",
"reweighted sum of the cross-entropies of the conditional",
"AEC",
"probabilities. This leads us to define the hierarchical cross-",
"where the soft label embedding is given componentwise by",
"entropy (HXE) as",
"exp(-βd(A, C))",
"yo\"(C)",
"(7)",
"^(C() log p(c()|c(1+1),(4)",
"ZBec exp(-βd(B,C))",
"1=(",
"CHXE(P, C) = -",
"for class distancefunction d and parameter B.This loss is",
"where A(C(t)) is the weight associated with the edge node",
"illustrated in Fig. 2b. For the distance function d(Ci, C,),",
"C(t+1) = C(i), see Fig. 2a. Although this loss is expressed",
"we use the height of LCA(Ci, C,) divided by the height of",
"in terms of conditional probabilities, it can easily be applied",
"the tree. To understand the role of the hyperparameter β,"
] |
{
"arxiv_id": "1912.09393",
"language": "en",
"page_idx": 3
}
|
||
1912.09393-en-4
|
1912.09393-en-4.jpg
|
$$\begin{array}{ccccc}\mathrm{(a)}_{\lambda(D)}&\mathrm{R}\\\mathrm{A}&\mathrm{B}&\mathrm{C}\\\mathrm{1}&0&0&y_{A}^{\mathrm{soft}}(A)&y_{B}^{\mathrm{soft}}(A)\end{array}$$
Figure 2: Representations of the $HXE$ (Sec. 3.1) and soft $labels($Sec. 3.2) losses for a simple illustrative hierarchy are drawn in subfigures (a) and (b) respectively. The groundtruth class is underlined, and the edges contributing to the total value of the loss are drawn in bold.
note that values of $\beta$ that are much bigger than the typical inverse distance in the tree result in a label distribution that is nearly one-hot, $i.e.y_A(C)\simeq\delta_{AC}$, in which case the cross-entropy reduces to the familiar single-term log-loss expression. Conversely, for very small values of $\beta$ the label distribution is near-uniform. Between these extremes, greater probability mass is assigned to classes more closely related to the ground truth, with the magnitude of the difference controlled by $\beta.$
We offer two complementary interpretations that motivate this representation (besides its ease). For one, the distribution describing each target class can be considered to be a model of the actual uncertainty that a labeller would experience due to visual confusion between closely related classes$^2.$ It could also be thought of as encoding the extent to which a common response to different classes is required of the classifier, $i.e.$ the imposition of correlations between outputs, where higher correlations are expected for more closely related classes. This in turn suggests a connection to the superficially different but conceptually related $distil-$ lation method of Hinton et $al.[18$, Sec.2], in which correlations between a large network's responses to different classes are mimicked by a smaller network to desirable effect. Here, we simply supply these correlations directly, using widely available hierarchies.
Another important connection is the one to label smoothing [39], in which one-hot labels are combined with the uniform distribution. This technique has been used to regularise large neural networks $(e.g.[39,8,42,51])$, but has only recently [28] been studied more thoroughly.
## 4. Evaluation
In the following, we first describe the datasets (Sec. 4.1) and metrics (Sec. 4.2) of the setup common to all of our experiments. Then, in Sec. 4.3, we evaluate our two simple
$^{2}$In a recent work, Peterson $etal.[31]$ make use of soft labels expressing the distribution of human labellers for a subset of CIFAR-10, showing strong generalisation for classifiers trained on them.
proposals and compare them to the prior art. Finally, we experiment with random hierarchies to understand when information on class relatedness can help classification.
## 4.1. Datasets
In our experiments, we use tieredImageNet[34] (a large subset of ImageNet) and iNaturalist'19 [41], two datasets with hierarchies that are $a$) significantly different from one another and $b$)complex enough to cover a large number of visual concepts. ImageNet aims to populate the WordNet[26] hierarchy of nouns, with WordNet itself generated by inspecting IS-A lexical relationships. By contrast, iNaturalist'19[41] has a biological taxonomy [36] at its core.
tieredImageNet was originally introduced by Ren et $al.[34]$ for the problem of few-shot classification, in which the sets of classes between dataset splits are disjoint. The authors’ motivation in creating the dataset was to use the WordNet hierarchy to generate splits containing signifrcantly different classes, facilitating better assessment of few-shot classifiers by enforcing problem difficulty.
Although our task and motivations are different, we chose this dataset because of the large portion of the WordNet hierarchy spanned by its classes. To make it suitable for the problem of (standard) image classification, we resampled the dataset so as to represent all classes across the train, validation, and test splits. Moreover, since the methods proposed in Sec. 3.1 and YOLO-v2 [33] require that the graph representing the hierarchy be a tree, we modified the graph of the spanned WordNet hierarchy slightly to comply with this assumption (more details available in Appendix D). After this procedure, we obtained a tree of height 13 covering 608 classes. We refer to this modified version of tieredImageNet as tieredImageNet-H.
iNaturalist is a dataset of images of organisms that has mainly been used to evaluate fine-grained visual categorisation methods. The dataset construction protocol differs significantly from the one used for ImageNet in that it relies on passionate volunteers instead of workers paid per task[41]. Importantly, for the 2019 edition of the CVPR Fine-Grained Visual Categorization Workshop, metadata with hierarchical relationships between species have been released. In contrast to WordNet, this taxonomy is an 8-level complete
tree spanning 1010 classes that can readily be used in our experiments without modifications. Since the labels for the test set are not public, we randomly re-sampled three splits from the original train and validation splits into a new training, validation and test set (with respective probabilities of 0.7,0.15,and 0.15)
### $4. 2. \textbf{Metrics}$
We consider three measures of performance, covering
different interpretations of a classifier's mistakes.
|
[
[
0.2548828125,
0.091552734375,
0.281982421875,
0.091552734375,
0.281982421875,
0.10797119140625,
0.2548828125,
0.10797119140625
],
[
0.50341796875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.37744140625,
0.09661865234375,
0.384033203125,
0.09661865234375,
0.384033203125,
0.1029052734375,
0.37744140625,
0.1029052734375
],
[
0.080078125,
0.09344482421875,
0.10296630859375,
0.09344482421875,
0.10296630859375,
0.106689453125,
0.080078125,
0.106689453125
],
[
0.1650390625,
0.0985107421875,
0.17724609375,
0.0985107421875,
0.17724609375,
0.10986328125,
0.1650390625,
0.10986328125
],
[
0.5048828125,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.5048828125,
0.11932373046875
],
[
0.1168212890625,
0.11553955078125,
0.146240234375,
0.11553955078125,
0.146240234375,
0.1275634765625,
0.1168212890625,
0.1275634765625
],
[
0.50244140625,
0.12371826171875,
0.85302734375,
0.12371826171875,
0.85302734375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.125,
0.1319580078125,
0.1422119140625,
0.1319580078125,
0.1422119140625,
0.146484375,
0.125,
0.146484375
],
[
0.191162109375,
0.1319580078125,
0.223876953125,
0.1319580078125,
0.223876953125,
0.14453125,
0.191162109375,
0.14453125
],
[
0.08905029296875,
0.146484375,
0.1282958984375,
0.1483154296875,
0.12744140625,
0.1634521484375,
0.0882568359375,
0.1610107421875
],
[
0.1422119140625,
0.149658203125,
0.1805419921875,
0.149658203125,
0.1805419921875,
0.1634521484375,
0.1422119140625,
0.1634521484375
],
[
0.5048828125,
0.15087890625,
0.60107421875,
0.15087890625,
0.60107421875,
0.1622314453125,
0.5048828125,
0.1622314453125
],
[
0.285888671875,
0.1590576171875,
0.2998046875,
0.1641845703125,
0.293212890625,
0.175537109375,
0.279296875,
0.17041015625
],
[
0.35302734375,
0.1622314453125,
0.36767578125,
0.1590576171875,
0.371826171875,
0.1717529296875,
0.35693359375,
0.1749267578125
],
[
0.428955078125,
0.160400390625,
0.4453125,
0.160400390625,
0.4453125,
0.175537109375,
0.428955078125,
0.175537109375
],
[
0.1422119140625,
0.1654052734375,
0.158447265625,
0.1654052734375,
0.158447265625,
0.179931640625,
0.1422119140625,
0.179931640625
],
[
0.08905029296875,
0.1654052734375,
0.108642578125,
0.1654052734375,
0.108642578125,
0.179931640625,
0.08905029296875,
0.179931640625
],
[
0.1953125,
0.1654052734375,
0.2132568359375,
0.1654052734375,
0.2132568359375,
0.179931640625,
0.1953125,
0.179931640625
],
[
0.431396484375,
0.17236328125,
0.443603515625,
0.17236328125,
0.443603515625,
0.18310546875,
0.431396484375,
0.18310546875
],
[
0.521484375,
0.173583984375,
0.89208984375,
0.173583984375,
0.89208984375,
0.1875,
0.521484375,
0.1875
],
[
0.35302734375,
0.175537109375,
0.365234375,
0.17041015625,
0.371826171875,
0.179931640625,
0.359375,
0.1849365234375
],
[
0.1953125,
0.177978515625,
0.2115478515625,
0.177978515625,
0.2115478515625,
0.199462890625,
0.1953125,
0.199462890625
],
[
0.28515625,
0.175537109375,
0.293212890625,
0.175537109375,
0.293212890625,
0.18115234375,
0.28515625,
0.18115234375
],
[
0.1446533203125,
0.1787109375,
0.1568603515625,
0.1787109375,
0.1568603515625,
0.189453125,
0.1446533203125,
0.189453125
],
[
0.08905029296875,
0.179931640625,
0.108642578125,
0.179931640625,
0.108642578125,
0.2020263671875,
0.08905029296875,
0.2020263671875
],
[
0.2607421875,
0.18310546875,
0.32177734375,
0.18310546875,
0.32177734375,
0.2020263671875,
0.2607421875,
0.2020263671875
],
[
0.33251953125,
0.1817626953125,
0.39453125,
0.1817626953125,
0.39453125,
0.2008056640625,
0.33251953125,
0.2008056640625
],
[
0.143798828125,
0.1881103515625,
0.1568603515625,
0.1881103515625,
0.1568603515625,
0.199462890625,
0.143798828125,
0.199462890625
],
[
0.406005859375,
0.184326171875,
0.464111328125,
0.184326171875,
0.464111328125,
0.1988525390625,
0.406005859375,
0.1988525390625
],
[
0.50244140625,
0.189453125,
0.89208984375,
0.189453125,
0.89208984375,
0.20263671875,
0.50244140625,
0.20263671875
],
[
0.50244140625,
0.2052001953125,
0.89208984375,
0.2052001953125,
0.89208984375,
0.2183837890625,
0.50244140625,
0.2183837890625
],
[
0.080078125,
0.21533203125,
0.4697265625,
0.21533203125,
0.4697265625,
0.2298583984375,
0.080078125,
0.2298583984375
],
[
0.50244140625,
0.2197265625,
0.89208984375,
0.2197265625,
0.89208984375,
0.23291015625,
0.50244140625,
0.23291015625
],
[
0.08087158203125,
0.23291015625,
0.468017578125,
0.23291015625,
0.468017578125,
0.24365234375,
0.08087158203125,
0.24365234375
],
[
0.50244140625,
0.23486328125,
0.890625,
0.23486328125,
0.890625,
0.248779296875,
0.50244140625,
0.248779296875
],
[
0.080078125,
0.246826171875,
0.4697265625,
0.246826171875,
0.4697265625,
0.260009765625,
0.080078125,
0.260009765625
],
[
0.50244140625,
0.25,
0.89208984375,
0.25,
0.89208984375,
0.26318359375,
0.50244140625,
0.26318359375
],
[
0.080078125,
0.261962890625,
0.4697265625,
0.261962890625,
0.4697265625,
0.27587890625,
0.080078125,
0.27587890625
],
[
0.50341796875,
0.266357421875,
0.8896484375,
0.266357421875,
0.8896484375,
0.276611328125,
0.50341796875,
0.276611328125
],
[
0.080078125,
0.276611328125,
0.343017578125,
0.276611328125,
0.343017578125,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.50244140625,
0.27978515625,
0.8759765625,
0.27978515625,
0.8759765625,
0.29345703125,
0.50244140625,
0.29345703125
],
[
0.50244140625,
0.2998046875,
0.89208984375,
0.2998046875,
0.89208984375,
0.313232421875,
0.50244140625,
0.313232421875
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.50244140625,
0.31494140625,
0.89208984375,
0.31494140625,
0.89208984375,
0.328857421875,
0.50244140625,
0.328857421875
],
[
0.08087158203125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.34521484375,
0.08087158203125,
0.34521484375
],
[
0.50244140625,
0.330810546875,
0.89208984375,
0.330810546875,
0.89208984375,
0.3447265625,
0.50244140625,
0.3447265625
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50244140625,
0.34521484375,
0.89208984375,
0.34521484375,
0.89208984375,
0.359130859375,
0.50244140625,
0.359130859375
],
[
0.080078125,
0.3642578125,
0.471435546875,
0.3642578125,
0.471435546875,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50341796875,
0.361083984375,
0.88818359375,
0.361083984375,
0.88818359375,
0.375,
0.50341796875,
0.375
],
[
0.07843017578125,
0.37890625,
0.468017578125,
0.37744140625,
0.468017578125,
0.392578125,
0.07843017578125,
0.393310546875
],
[
0.50341796875,
0.37744140625,
0.890625,
0.37744140625,
0.890625,
0.38818359375,
0.50341796875,
0.38818359375
],
[
0.080078125,
0.393310546875,
0.468017578125,
0.393310546875,
0.468017578125,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.50244140625,
0.391357421875,
0.8408203125,
0.391357421875,
0.8408203125,
0.40478515625,
0.50244140625,
0.40478515625
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.521484375,
0.4072265625,
0.8935546875,
0.4072265625,
0.8935546875,
0.42041015625,
0.521484375,
0.42041015625
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.07843017578125,
0.439453125,
0.221435546875,
0.439453125,
0.221435546875,
0.45263671875,
0.07843017578125,
0.45263671875
],
[
0.50244140625,
0.4375,
0.890625,
0.4375,
0.890625,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.098876953125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.098876953125,
0.46728515625
],
[
0.50341796875,
0.453857421875,
0.8896484375,
0.453857421875,
0.8896484375,
0.464111328125,
0.50341796875,
0.464111328125
],
[
0.080078125,
0.4697265625,
0.4697265625,
0.4697265625,
0.4697265625,
0.48291015625,
0.080078125,
0.48291015625
],
[
0.50244140625,
0.46728515625,
0.89208984375,
0.46728515625,
0.89208984375,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.080078125,
0.48486328125,
0.4697265625,
0.48486328125,
0.4697265625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.48291015625,
0.890625,
0.48291015625,
0.890625,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.5009765625,
0.49755859375,
0.8935546875,
0.496826171875,
0.8935546875,
0.51123046875,
0.5009765625,
0.51220703125
],
[
0.07843017578125,
0.51513671875,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.07843017578125,
0.52880859375
],
[
0.50244140625,
0.51318359375,
0.89208984375,
0.51318359375,
0.89208984375,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.07843017578125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.07843017578125,
0.54345703125
],
[
0.50244140625,
0.52783203125,
0.89208984375,
0.52783203125,
0.89208984375,
0.54150390625,
0.50244140625,
0.54150390625
],
[
0.080078125,
0.54541015625,
0.4697265625,
0.54541015625,
0.4697265625,
0.5595703125,
0.080078125,
0.5595703125
],
[
0.50244140625,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.55810546875,
0.89208984375,
0.55810546875,
0.89208984375,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.080078125,
0.57470703125,
0.4697265625,
0.57470703125,
0.4697265625,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.50244140625,
0.5859375
],
[
0.080078125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50244140625,
0.58837890625,
0.76953125,
0.58837890625,
0.76953125,
0.6015625,
0.50244140625,
0.6015625
],
[
0.07843017578125,
0.60498046875,
0.468017578125,
0.60400390625,
0.468017578125,
0.61865234375,
0.07843017578125,
0.619140625
],
[
0.5009765625,
0.60693359375,
0.89208984375,
0.60791015625,
0.89208984375,
0.62255859375,
0.5009765625,
0.6220703125
],
[
0.080078125,
0.62060546875,
0.4697265625,
0.62060546875,
0.4697265625,
0.6337890625,
0.080078125,
0.6337890625
],
[
0.5009765625,
0.62353515625,
0.890625,
0.625,
0.890625,
0.63818359375,
0.5009765625,
0.63720703125
],
[
0.08087158203125,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.5009765625,
0.63818359375,
0.890625,
0.6396484375,
0.890625,
0.65380859375,
0.5009765625,
0.65283203125
],
[
0.080078125,
0.65087890625,
0.468017578125,
0.65087890625,
0.468017578125,
0.6640625,
0.080078125,
0.6640625
],
[
0.50341796875,
0.65576171875,
0.890625,
0.65576171875,
0.890625,
0.66650390625,
0.50341796875,
0.66650390625
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.50341796875,
0.66943359375,
0.890625,
0.66943359375,
0.890625,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.080078125,
0.68115234375,
0.291748046875,
0.68115234375,
0.291748046875,
0.6943359375,
0.080078125,
0.6943359375
],
[
0.50244140625,
0.6845703125,
0.89208984375,
0.6845703125,
0.89208984375,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.09722900390625,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.09722900390625,
0.708984375
],
[
0.50244140625,
0.69775390625,
0.890625,
0.69873046875,
0.890625,
0.71337890625,
0.50244140625,
0.7119140625
],
[
0.080078125,
0.7109375,
0.4697265625,
0.7109375,
0.4697265625,
0.724609375,
0.080078125,
0.724609375
],
[
0.50341796875,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.7265625,
0.50341796875,
0.7265625
],
[
0.0767822265625,
0.724609375,
0.468017578125,
0.72607421875,
0.468017578125,
0.74072265625,
0.0767822265625,
0.7392578125
],
[
0.5009765625,
0.72900390625,
0.890625,
0.72900390625,
0.890625,
0.7421875,
0.5009765625,
0.7421875
],
[
0.080078125,
0.7412109375,
0.4697265625,
0.7412109375,
0.4697265625,
0.7548828125,
0.080078125,
0.7548828125
],
[
0.50244140625,
0.744140625,
0.89208984375,
0.74365234375,
0.89208984375,
0.75830078125,
0.50244140625,
0.7587890625
],
[
0.0767822265625,
0.7548828125,
0.400390625,
0.755859375,
0.400390625,
0.77099609375,
0.0767822265625,
0.76953125
],
[
0.50244140625,
0.76025390625,
0.89208984375,
0.76025390625,
0.89208984375,
0.7734375,
0.50244140625,
0.7734375
],
[
0.50244140625,
0.7744140625,
0.89208984375,
0.7744140625,
0.89208984375,
0.78857421875,
0.50244140625,
0.78857421875
],
[
0.080078125,
0.78515625,
0.19287109375,
0.78515625,
0.19287109375,
0.798828125,
0.080078125,
0.798828125
],
[
0.50244140625,
0.79052734375,
0.890625,
0.79052734375,
0.890625,
0.8037109375,
0.50244140625,
0.8037109375
],
[
0.098876953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.098876953125,
0.82373046875
],
[
0.50244140625,
0.80615234375,
0.8935546875,
0.80615234375,
0.8935546875,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.07843017578125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.07843017578125,
0.8388671875
],
[
0.50244140625,
0.82080078125,
0.6337890625,
0.82080078125,
0.6337890625,
0.833984375,
0.50244140625,
0.833984375
],
[
0.080078125,
0.84033203125,
0.468017578125,
0.84033203125,
0.468017578125,
0.85400390625,
0.080078125,
0.85400390625
],
[
0.50341796875,
0.8486328125,
0.5947265625,
0.8486328125,
0.5947265625,
0.85986328125,
0.50341796875,
0.85986328125
],
[
0.096435546875,
0.86279296875,
0.471435546875,
0.86279296875,
0.471435546875,
0.876953125,
0.096435546875,
0.876953125
],
[
0.51953125,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.51953125,
0.8837890625
],
[
0.07843017578125,
0.87548828125,
0.471435546875,
0.87548828125,
0.471435546875,
0.8896484375,
0.07843017578125,
0.8896484375
],
[
0.07843017578125,
0.88818359375,
0.35693359375,
0.88818359375,
0.35693359375,
0.9013671875,
0.07843017578125,
0.9013671875
],
[
0.5009765625,
0.88525390625,
0.826171875,
0.88623046875,
0.826171875,
0.90087890625,
0.5009765625,
0.8994140625
]
] |
[
"(b)",
"proposals and compare them to the prior art. Finally,we",
"R",
"(a)",
"R",
"experimentwithrandomhierarchiestounderstandwhenin-",
"入(D)",
"formation on class relatedness can help classification.",
"D",
"A(C)",
"入(A) /",
"^(B)",
"4.1.Datasets",
"A",
"B",
"c",
"B",
"A",
"c",
"n",
"In our experiments, we use tieredImageNet [34] (a large",
"1",
"=O",
"1",
"Ⅱ1",
"1",
"uAf(A)",
"f(A)",
"0",
"yof(A)",
"subset of ImageNet) and iNaturalist'19 [41], two datasets",
"with hierarchies that are a) significantly different from one",
"Figure 2: Representations of the HXE (Sec. 3.1) and soft",
"another and b) complex enough to cover a large number",
"labels(Sec.3.2)lossesfora simpleillustrativehierarchyare",
"of visual concepts. ImageNet aims to populate the Word-",
"drawn in subfigures (a) and (b) respectively. The ground-",
"Net [26] hierarchy of nouns, with WordNet itself generated",
"truth class is underlined, and the edges contributing to the",
"byinspectingIS-Alexicalrelationships.Bycontrast.iNat-",
"total value of the loss are drawn in bold.",
"uralist'19 [4]] has a biological taxonomy [36] at its core.",
"tieredlmageNet was originally introduced by Ren et",
"note that values of β that are much bigger than the typi-",
"al. [34] for the problem of few-shot classification, in which",
"cal inversedistance in thetreeresultina label distribution",
"the sets of classes between dataset splits are disjoint. The",
"that is nearly one-hot, i.e. yA(C) AC, in which case the",
"authors' motivation in creating the dataset was to use the",
"cross-entropy reduces to the familiar single-term log-loss",
"WordNet hierarchy to generate splits containing signifi-",
"expression.Conversely,for very small values ofβ the la-",
"cantly different classes,facilitating better assessment of",
"bel distribution is near-uniform. Between these extremes,",
"few-shot classifiers by enforcing problem difficulty.",
"greater probability mass is assigned to classes more closely",
"Although our task and motivations are different, we",
"related to the ground truth, with the magnitude of the differ-",
"chose this dataset because of the large portion of the Word-",
"ence controlled by β.",
"Net hierarchy spanned by its classes. To make it suitable",
"We offer two complementary interpretations that moti-",
"for the problem of(standard)imageclassification,were-",
"vate this representation (besides its ease). For one, the dis-",
"sampled the dataset so as to represent all classes across the",
"tribution describing each target class can be considered to",
"train, validation, and test splits. Moreover, since the meth-",
"be a model of the actual uncertainty that a labeller would",
"ods proposed in Sec. 3.1 and YOLO-v2 [33] require that",
"experiencedueto visual confusion between closelyrelated",
"the graph representing the hierarchy be a tree, we modi-",
"classes2. It could also be thought of as encoding the extent",
"fied the graph of the spanned WordNet hierarchy slightly to",
"to which a common response to different classes is required",
"comply with this assumption (more details available in Ap-",
"of the classifier, i.e. the imposition of correlations between",
"pendix D). After this procedure, we obtained a tree of height",
"outputs, where higher correlations are expected for more",
"13 covering 608 classes. We refer to this modified version",
"closely related classes. This in turn suggests a connection",
"of tieredImageNet as tieredlmageNet-H.",
"to the superficially different but conceptually related distil-",
"iNaturalist is a dataset of images of organisms that has",
"lation method of Hinton et al. [18, Sec. 2], in which cor-",
"mainly been used to evaluate fine-grained visual categorisa-",
"relations betweena largenetwork'sresponsestodifferent",
"tion methods. The dataset construction protocol differs sig-",
"classes are mimicked by a smaller network to desirable ef-",
"nificantlyfromtheoneusedforImageNet inthatitrelies on",
"fect. Here, we simply supply these correlations directly, us-",
"passionate volunteers instead of workers paid per task [41]",
"ing widely available hierarchies.",
"Importantly, for the 2019 edition of the CVPR Fine-Grained",
"Another important connection is the one to label smooth-",
"Visual Categorization Workshop, metadata with hierarchi-",
"ing [39], in which one-hot labels are combined with the",
"calrelationshipsbetweenspecies havebeen released.In",
"uniform distribution. This technique has been used to reg-",
"contrast to WordNet, this taxonomy is an 8-level complete",
"ularise large neural networks (e.g. [39, 8, 42, 51), but has",
"tree spanning 1010 classes that can readily be used in our",
"only recently [28] been studied more thoroughly.",
"experiments without modifications. Since the labels for the",
"test set are not public, we randomly re-sampled three splits",
"4.Evaluation",
"from the original train and validation splits into a new train-",
"In the following, we first describe the datasets (Sec. 4.1)",
"ing, validation and test set (with respective probabilities of",
"and metrics (Sec. 4.2) of the setup common to all of our ex-",
"0.7, 0.15, and 0.15)",
"periments. Then, in Sec. 4.3, we evaluate our two simple",
"4.2. Metrics",
"2In a recent work, Peterson et al. [31] make use of soft labels express-",
"We consider three measures of performance, covering",
"ing the distribution of human labellers for a subset of CIFAR-10, showing",
" strong generalisation for classifiers trained on them.",
"different interpretations of a classifier's mistakes."
] |
{
"arxiv_id": "1912.09393",
"language": "en",
"page_idx": 4
}
|
||
2211.07955-en-0
|
2211.07955-en-0.jpg
|
## $\textbf{IntegratedPIFu: Integrated Pixel Aligned}$ Implicit Function for Single-view Human $\operatorname{Reconstruction}$
Kennard Yanting Chan$^1,2,3$, Guosheng Lin$^3$, Haiyu Zhao$^2$, and Weisi Lin$^1,3$
$^{1}$S-Lab, Nanyang Technological University
$^{2}{\mathrm{~SenseTime~Research}}$
$^3$ Nanyang Technological University
kenn0042@e.ntu.edu.sg, zhaohaiyu@sensetime.com, $\{$gslin,wslin$\}$@ntu.edu.sg
$\mathbf{Abstract.~We~propose~IntegratedPIFu,~a~new~pixel-aligned~implicit~model}$ that builds on the foundation set by PIFuHD. IntegratedPIFu shows how depth and human parsing information can be predicted and capitalized upon in a pixel-aligned implicit model. In addition, IntegratedPIFu introduces depth-oriented sampling, a novel training scheme that improve any pixel-aligned implicit model's ability to reconstruct important human features without noisy artefacts. Lastly, IntegratedPIFu presents a new architecture that, despite using less model parameters than PIFuHD, is able to improves the structural correctness of reconstructed meshes. Our results show that IntegratedPIFu significantly outperforms existing state-of-the-arts methods on single-view human reconstruction. We provide the code in our supplementary materials. Our code is available at https://github.com/kcyt/IntegratedPIFu.
Keywords: single-view human reconstruction, implicit function, depth
prediction, human parsing prediction
## 1 Introduction
Human digitization is an important topic with applications in areas like virtual reality, game production, and medical imaging. While high precision reconstruction of human body is already possible using high-end, multi-view capturing systems [5,6,17], it is largely inaccessible to general consumers. Increasingly, the research community has shown interest in developing deep learning models for human digitization from simple inputs such as a single image [1,20,28,27].
Many methods have been proposed for single-image clothed human reconstruction, but one class of approaches that have captured significant interest is pixel-aligned implicit models [26,8,27]. These methods model a human body as an implicit function, from which a mesh can be obtained using Marching Cubes algorithm [19]. One recent and notable pixel-aligned implicit model is PIFuHD [27] proposed by Saito $et\textit{al. Unlike previous pixel- aligned implicit models, PI- }$ FuHD is able to capture fine-level geometric details such as clothes wrinkles.
|
[
[
0.26318359375,
0.1458740234375,
0.74169921875,
0.1458740234375,
0.74169921875,
0.1634521484375,
0.26318359375,
0.1634521484375
],
[
0.265625,
0.1685791015625,
0.73876953125,
0.1685791015625,
0.73876953125,
0.18310546875,
0.265625,
0.18310546875
],
[
0.41259765625,
0.189453125,
0.59228515625,
0.1912841796875,
0.59228515625,
0.208984375,
0.41259765625,
0.20703125
],
[
0.233642578125,
0.236083984375,
0.771484375,
0.236083984375,
0.771484375,
0.250732421875,
0.233642578125,
0.250732421875
],
[
0.0228729248046875,
0.2568359375,
0.060455322265625,
0.2568359375,
0.060455322265625,
0.404052734375,
0.0228729248046875,
0.404052734375
],
[
0.3544921875,
0.261474609375,
0.646484375,
0.261962890625,
0.646484375,
0.27978515625,
0.3544921875,
0.279052734375
],
[
0.4306640625,
0.2822265625,
0.439453125,
0.2822265625,
0.439453125,
0.287353515625,
0.4306640625,
0.287353515625
],
[
0.435546875,
0.279052734375,
0.57666015625,
0.279052734375,
0.57666015625,
0.292236328125,
0.435546875,
0.292236328125
],
[
0.37744140625,
0.28857421875,
0.62158203125,
0.291748046875,
0.62158203125,
0.31005859375,
0.37744140625,
0.306884765625
],
[
0.223876953125,
0.306884765625,
0.78125,
0.306884765625,
0.78125,
0.32080078125,
0.223876953125,
0.32080078125
],
[
0.265625,
0.361083984375,
0.75341796875,
0.361083984375,
0.75341796875,
0.375,
0.265625,
0.375
],
[
0.2646484375,
0.375,
0.73876953125,
0.375,
0.73876953125,
0.38818359375,
0.2646484375,
0.38818359375
],
[
0.2646484375,
0.38818359375,
0.740234375,
0.38818359375,
0.740234375,
0.401611328125,
0.2646484375,
0.401611328125
],
[
0.021240234375,
0.396484375,
0.06451416015625,
0.396484375,
0.06451416015625,
0.488037109375,
0.021240234375,
0.488037109375
],
[
0.26318359375,
0.40283203125,
0.73876953125,
0.401611328125,
0.73876953125,
0.415283203125,
0.26318359375,
0.416748046875
],
[
0.2646484375,
0.417236328125,
0.740234375,
0.417236328125,
0.740234375,
0.43115234375,
0.2646484375,
0.43115234375
],
[
0.2646484375,
0.43115234375,
0.73876953125,
0.43115234375,
0.73876953125,
0.4443359375,
0.2646484375,
0.4443359375
],
[
0.2646484375,
0.4443359375,
0.740234375,
0.4443359375,
0.740234375,
0.458251953125,
0.2646484375,
0.458251953125
],
[
0.2646484375,
0.458251953125,
0.73876953125,
0.458251953125,
0.73876953125,
0.4716796875,
0.2646484375,
0.4716796875
],
[
0.26318359375,
0.4716796875,
0.73779296875,
0.47021484375,
0.73779296875,
0.48486328125,
0.26318359375,
0.486083984375
],
[
0.0228729248046875,
0.4892578125,
0.060455322265625,
0.4892578125,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.2646484375,
0.48486328125,
0.73876953125,
0.486083984375,
0.73876953125,
0.5,
0.2646484375,
0.498779296875
],
[
0.265625,
0.501953125,
0.736328125,
0.501953125,
0.736328125,
0.51220703125,
0.265625,
0.51220703125
],
[
0.2646484375,
0.51318359375,
0.73876953125,
0.51318359375,
0.73876953125,
0.52783203125,
0.2646484375,
0.52783203125
],
[
0.2646484375,
0.52783203125,
0.5400390625,
0.52783203125,
0.5400390625,
0.54150390625,
0.2646484375,
0.54150390625
],
[
0.2646484375,
0.5546875,
0.740234375,
0.5546875,
0.740234375,
0.568359375,
0.2646484375,
0.568359375
],
[
0.2646484375,
0.568359375,
0.51123046875,
0.568359375,
0.51123046875,
0.58203125,
0.2646484375,
0.58203125
],
[
0.2181396484375,
0.6123046875,
0.2467041015625,
0.6123046875,
0.2467041015625,
0.625,
0.2181396484375,
0.625
],
[
0.2484130859375,
0.611328125,
0.374267578125,
0.611328125,
0.374267578125,
0.62548828125,
0.2484130859375,
0.62548828125
],
[
0.2181396484375,
0.64453125,
0.78759765625,
0.64453125,
0.78759765625,
0.65771484375,
0.2181396484375,
0.65771484375
],
[
0.2181396484375,
0.6611328125,
0.78369140625,
0.6611328125,
0.78369140625,
0.6748046875,
0.2181396484375,
0.6748046875
],
[
0.2156982421875,
0.6748046875,
0.7861328125,
0.67529296875,
0.7861328125,
0.68994140625,
0.2156982421875,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.7861328125,
0.69140625,
0.7861328125,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.21728515625,
0.7197265625,
0.74609375,
0.7197265625,
0.74609375,
0.734375,
0.21728515625,
0.734375
],
[
0.24267578125,
0.736328125,
0.78369140625,
0.736328125,
0.78369140625,
0.75,
0.24267578125,
0.75
],
[
0.2181396484375,
0.751953125,
0.7861328125,
0.751953125,
0.7861328125,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.21728515625,
0.7666015625,
0.78759765625,
0.76513671875,
0.78759765625,
0.77978515625,
0.21728515625,
0.78076171875
],
[
0.21728515625,
0.77978515625,
0.78515625,
0.77978515625,
0.78515625,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.2197265625,
0.8125,
0.78369140625,
0.8125,
0.78369140625,
0.82568359375,
0.2197265625,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.7646484375,
0.8271484375,
0.7646484375,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"IntegratedPIFu:Integrated Pixel Aligned",
"Implicit Function for Single-view Human",
"Reconstruction",
"Kennard Yanting Chan1,2,3, Guosheng Lin3, Haiyu Zhao?, and Weisi Lin1,3",
"Z S",
"1 S-Lab, Nanyang Technological University",
"",
"SenseTime Research",
"3 Nanyang Technological University",
"kenn0042@e.ntu.edu.sg,zhaohaiyu@sensetime.com,(gslin,wslin)@ntu.edu.sg",
"Abstract. We propose IntegratedPIFu, a new pixel-aligned implicit model",
"that builds on the foundation set by PIFuHD. IntegratedPIFu shows how",
"depth and human parsing information can be predicted and capitalized",
"[A's]",
"upon in a pixel-aligned implicit model. In addition, IntegratedPIFu in-",
"troduces depth-oriented sampling, a novel training scheme that improve",
"any pixel-aligned implicit model's ability to reconstruct important hu-",
"man features without noisy artefacts. Lastly, IntegratedPIFu presents a",
"new architecture that, despite using less model parameters than PIFuHD,",
"is able to improves the structural correctness of reconstructed meshes.",
"1AS'1ZZ:AI",
"Our results show that IntegratedPIFu significantly outperforms existing",
"state-of-the-arts methods on single-viewhuman reconstruction.Wepro",
"vide the code in our supplementary materials. Our code is available at",
"https://github.com/kcyt/IntegratedPIFu.",
"Keywords: single-view human reconstruction, implicit function, depth",
"prediction, human parsing prediction",
"1",
"Introduction",
"Human digitization is an important topic with applications in areas like virtual",
"reality, game production, and medical imaging. While high precision reconstruc-",
"tion of human body is already possible using high-end, multi-view capturing",
"systems [5,6,17], it is largely inaccessible to general consumers. Increasingly, the",
"research community has shown interest in developing deep learning models for",
"human digitization from simple inputs such as a single image [1,20,28,27].",
"Many methods have been proposed for single-image clothed human recon-",
"struction, but one class of approaches that have captured significant interest is",
"pixel-aligned implicit models [26,8,27]. These methods model a human body as",
"an implicit function, from which a mesh can be obtained using Marching Cubes",
"algorithm [19]. One recent and notable pixel-aligned implicit model is PIFuHD",
"[27] proposed by Saito et al. Unlike previous pixel-aligned implicit models, PI-",
"FuHD is able to capture fine-level geometric details such as clothes wrinkles."
] |
{
"arxiv_id": "2211.07955",
"language": "en",
"page_idx": 0
}
|
||
2211.07955-en-1
|
2211.07955-en-1.jpg
|
2
K. Chan et al
However, pixel-aligned implicit models, including PIFuHD, are observed to be prone to problems such as depth ambiguity and generation of meshes with broken limbs [10,23]. In particular, depth ambiguity can lead to unnaturally elongated human features in the reconstructed meshes.
In order to overcome these issues, we propose IntegratedPIFu, which is a new and improved pixel-aligned implicit model that builds on the foundations set by PIFuHD. In order to mitigate the problem of depth ambiguity and broken limbs in reconstructed human meshes, IntegratedPIFu dedicates specialized networks for depth and human parsing predictions from a single RGB image. Predicted depth information mitigates depth ambiguity by serving as a proxy for actual depth. Human parsing predictions provide structural information critical for reconstructing human limbs without breakages or gaps.
In addition, IntegratedPIFu introduces depth-oriented sampling, a novel training scheme suitable for training any pixel-aligned implicit model. We show that this scheme is able to significantly improve the model's ability to capture important human features. Furthermore, IntegratedPIFu introduces High-resolution integrator, a new architecture that ensures coarse and fine-level components in a pixel-aligned implicit model can work together in closer tandem. We later show that our new architecture can improve the structural correctness of the reconstructed meshes.
The main contributions of our work consist of the following:
1. To the best of our knowledge, we are the first to introduce a method to
predict and incorporate relative depth and human parse maps into pixel-
aligned implicit models for single-view clothed human reconstruction.
2. Depth-oriented sampling, a novel training scheme that replaces the original
spatial sampling scheme used to train pixel-aligned implicit models. Our scheme is able to capture small but important human features, such as fingers and ears, that are often ignored by the original scheme. In addition, unlike the original scheme, depth-oriented sampling does not cause highly noticeable wavy artefacts on reconstructed meshes.
3. High-resolution integrator, a newly coined neural architecture that enables
much stronger interactions between the coarse and fne-level components in a pixel-aligned implicit model. It is also an efficient architecture that improves the structure of the reconstructed meshes despite using 39% less model parameters than its direct predecessor.
All three of our contributions can be applied to any existing pixel-aligned
${\mathrm{implicit~model.}}$
# 2 Related Work
2.1 Single-view Human Reconstruction
Several methods have been proposed to reconstruct human body shape from a single image. The methods can be classified into parametric and non-parametric
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.27294921875,
0.11614990234375,
0.36767578125,
0.11614990234375,
0.36767578125,
0.1275634765625,
0.27294921875,
0.1275634765625
],
[
0.24267578125,
0.147705078125,
0.7861328125,
0.147705078125,
0.7861328125,
0.1622314453125,
0.24267578125,
0.1622314453125
],
[
0.21728515625,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.177978515625,
0.21728515625,
0.177978515625
],
[
0.2181396484375,
0.1787109375,
0.78515625,
0.1787109375,
0.78515625,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.2156982421875,
0.1925048828125,
0.611328125,
0.1932373046875,
0.611328125,
0.2083740234375,
0.2156982421875,
0.20703125
],
[
0.24267578125,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.24267578125,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.78515625,
0.2235107421875,
0.78515625,
0.238037109375,
0.2181396484375,
0.238037109375
],
[
0.2181396484375,
0.2392578125,
0.7861328125,
0.2392578125,
0.7861328125,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.21728515625,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.268310546875,
0.21728515625,
0.268310546875
],
[
0.2181396484375,
0.26953125,
0.78759765625,
0.26953125,
0.78759765625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2181396484375,
0.285400390625,
0.7861328125,
0.285400390625,
0.7861328125,
0.298583984375,
0.2181396484375,
0.298583984375
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.314453125,
0.2181396484375,
0.314453125
],
[
0.2156982421875,
0.3125,
0.61669921875,
0.313232421875,
0.61669921875,
0.330810546875,
0.2156982421875,
0.330078125
],
[
0.24267578125,
0.330078125,
0.7890625,
0.330078125,
0.7890625,
0.343505859375,
0.24267578125,
0.343505859375
],
[
0.2181396484375,
0.34521484375,
0.78515625,
0.34521484375,
0.78515625,
0.359130859375,
0.2181396484375,
0.359130859375
],
[
0.21728515625,
0.359130859375,
0.78369140625,
0.35986328125,
0.78369140625,
0.375,
0.21728515625,
0.373779296875
],
[
0.2181396484375,
0.375732421875,
0.7861328125,
0.375732421875,
0.7861328125,
0.389404296875,
0.2181396484375,
0.389404296875
],
[
0.2156982421875,
0.39013671875,
0.7861328125,
0.389404296875,
0.7861328125,
0.404052734375,
0.2156982421875,
0.40478515625
],
[
0.21728515625,
0.406005859375,
0.7861328125,
0.406005859375,
0.7861328125,
0.419189453125,
0.21728515625,
0.419189453125
],
[
0.2181396484375,
0.42041015625,
0.7861328125,
0.42041015625,
0.7861328125,
0.434326171875,
0.2181396484375,
0.434326171875
],
[
0.2197265625,
0.4375,
0.374267578125,
0.4375,
0.374267578125,
0.447509765625,
0.2197265625,
0.447509765625
],
[
0.2442626953125,
0.449462890625,
0.67333984375,
0.449462890625,
0.67333984375,
0.464111328125,
0.2442626953125,
0.464111328125
],
[
0.223876953125,
0.47607421875,
0.7861328125,
0.47607421875,
0.7861328125,
0.490478515625,
0.223876953125,
0.490478515625
],
[
0.2467041015625,
0.4912109375,
0.78369140625,
0.4912109375,
0.78369140625,
0.505859375,
0.2467041015625,
0.505859375
],
[
0.2467041015625,
0.5068359375,
0.74169921875,
0.5068359375,
0.74169921875,
0.52099609375,
0.2467041015625,
0.52099609375
],
[
0.2254638671875,
0.521484375,
0.78759765625,
0.521484375,
0.78759765625,
0.5361328125,
0.2254638671875,
0.5361328125
],
[
0.2442626953125,
0.537109375,
0.7861328125,
0.5361328125,
0.7861328125,
0.55029296875,
0.2442626953125,
0.5517578125
],
[
0.2467041015625,
0.55322265625,
0.7861328125,
0.55322265625,
0.7861328125,
0.56640625,
0.2467041015625,
0.56640625
],
[
0.2459716796875,
0.5673828125,
0.7861328125,
0.56640625,
0.7861328125,
0.58056640625,
0.2459716796875,
0.58203125
],
[
0.2467041015625,
0.58349609375,
0.7861328125,
0.58349609375,
0.7861328125,
0.5966796875,
0.2467041015625,
0.5966796875
],
[
0.2467041015625,
0.59765625,
0.53515625,
0.59765625,
0.53515625,
0.6123046875,
0.2467041015625,
0.6123046875
],
[
0.2254638671875,
0.6123046875,
0.7861328125,
0.6123046875,
0.7861328125,
0.626953125,
0.2254638671875,
0.626953125
],
[
0.2459716796875,
0.626953125,
0.78515625,
0.626953125,
0.78515625,
0.64013671875,
0.2459716796875,
0.64013671875
],
[
0.2459716796875,
0.64306640625,
0.78515625,
0.64306640625,
0.78515625,
0.6572265625,
0.2459716796875,
0.6572265625
],
[
0.2467041015625,
0.65771484375,
0.7861328125,
0.65771484375,
0.7861328125,
0.67236328125,
0.2467041015625,
0.67236328125
],
[
0.2459716796875,
0.67236328125,
0.57275390625,
0.673828125,
0.57275390625,
0.68798828125,
0.2459716796875,
0.6875
],
[
0.2418212890625,
0.69580078125,
0.78759765625,
0.6962890625,
0.78759765625,
0.7138671875,
0.2418212890625,
0.71337890625
],
[
0.21728515625,
0.71337890625,
0.328369140625,
0.71337890625,
0.328369140625,
0.72802734375,
0.21728515625,
0.72802734375
],
[
0.2181396484375,
0.75390625,
0.236083984375,
0.75390625,
0.236083984375,
0.767578125,
0.2181396484375,
0.767578125
],
[
0.2484130859375,
0.75390625,
0.386474609375,
0.75390625,
0.386474609375,
0.767578125,
0.2484130859375,
0.767578125
],
[
0.2181396484375,
0.7861328125,
0.25244140625,
0.7861328125,
0.25244140625,
0.79736328125,
0.2181396484375,
0.79736328125
],
[
0.25732421875,
0.78515625,
0.55126953125,
0.78515625,
0.55126953125,
0.798828125,
0.25732421875,
0.798828125
],
[
0.2181396484375,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.82763671875,
0.7861328125,
0.82763671875,
0.7861328125,
0.84130859375,
0.2181396484375,
0.84130859375
]
] |
[
"2",
"K. Chan et al.",
"However, pixel-aligned implicit models, including PIFuHD, are observed to",
"be prone to problems such as depth ambiguity and generation of meshes with",
"broken limbs [10,23]. In particular, depth ambiguity can lead to unnaturally",
"elongated human features in the reconstructed meshes.",
"In order to overcome these issues, we propose IntegratedPIFu, which is a new",
"and improved pixel-aligned implicit model that builds on the foundations set by",
"PIFuHD. In order to mitigate the problem of depth ambiguity and broken limbs",
"in reconstructed human meshes, IntegratedPIFu dedicates specialized networks",
"for depth and human parsing predictions from a single RGB image. Predicted",
"depth information mitigates depth ambiguity by serving as a proxy for actual",
"depth. Human parsing predictions provide structural information critical for",
"reconstructing human limbs without breakages or gaps.",
"In addition, IntegratedPIFu introduces depth-oriented sampling, a novel train-",
"ing scheme suitable for training any pixel-aligned implicit model. We show that",
"this scheme is ableto significantly improve the model's ability to capture impor-",
"tant human features. Furthermore, IntegratedPIFu introduces High-resolution",
"integrator, a new architecture that ensures coarse and fine-level components in",
"a pixel-aligned implicit model can work together in closer tandem. We later",
"show that our new architecture can improve the structural correctness of the",
"reconstructedmeshes.",
"The main contributions of our work consist of the following:",
"1. To the best of our knowledge, we are the first to introduce a method to",
"predict and incorporate relative depth and human parse maps into pixel-",
"aligned implicit models for single-view clothed human reconstruction.",
"2. Depth-oriented sampling, a novel training scheme that replaces the original",
"spatial sampling scheme used to train pixel-aligned implicit models. Our",
"scheme is able to capture small but important human features, such as fingers",
"and ears, that are often ignored by the original scheme. In addition, unlike the",
"original scheme, depth-oriented sampling does not cause highly noticeable",
"wavy artefacts on reconstructed meshes.",
"3. High-resolution integrator, a newly coined neural architecture that enables",
"much stronger interactions between the coarse and fine-level components",
"in a pixel-aligned implicit model. It is also an efficient architecture that",
"improves the structure of the reconstructed meshes despite using 39% less",
"model parameters than its direct predecessor.",
"All three of our contributions can be applied to any existing pixel-aligned",
"implicit model.",
"2",
"Related Work",
"2.1",
"Single-view Human Reconstruction",
"Several methods have been proposed to reconstruct human body shape from a",
"single image. The methods can be classified into parametric and non-parametric"
] |
{
"arxiv_id": "2211.07955",
"language": "en",
"page_idx": 1
}
|
||
2211.07955-en-2
|
2211.07955-en-2.jpg
|
3
IntegratedPIFu
<FigureHere>
(a)
Fig.1. Compared with state-of-the-art methods, including (b) Geo-PIFu[8],(c) PIFu
[26], and (d) PIFuHD [27], our proposed model can precisely reconstruct various parts of a clothed human body without noisy artefacts. Subject's face censored as required
by dataset's owner
approaches. Parametric approaches, such as [13,15,14], recover human body shapes by estimating the parameters of a human parametric model (e.g. SMPL [18]) from an input image. However, these parametric methods are only able to reconstruct“naked”(without geometry of clothes) human meshes. Bhatnagar $et$ al.[3] and Jiang $et\textit{al. [ 12] additionally estimate the clothes on top of}$ a human parametric model. However, these methods are unable to reconstruct high-fidelity geometry shapes.
Non-parametric approaches, on the other hand, do not require a human parametric model. For example, [28] use a 3D CNN to directly estimate a volumetric body shape from a single image. Notably, a subclass of non-parametric approaches that have garner significant interest is pixel-aligned implicit models [26,10,8,27]. One of the first pixel-aligned implicit models is the Pixel-Aligned Implicit Function (PIFu) proposed by Saito $et\textit{al. [ 26] . PIFu is memory efficient and}$ capable of producing high-fidelity 3D reconstructions of clothed human meshes.
Building from PIFu, there have been other pixel-aligned implicit models being proposed. Examples include GeoPIFu [8] and StereoPIFu [10] that make use of latent voxel features and an additional view respectively to improve the quality of their reconstructed meshes. However, GeoPIFu is unable to reconstruct fine geometry details of a human body, and StereoPIFu requires an additional view and careful calibration of a binocular camera. On the other hand, PIFuHD [27], which is a multi-scale variant of PIFu and a follow-up work by Saito $et$ $al.$, is able to reconstruct very fine geometry details such as clothes wrinkles.
|
[
[
0.62841796875,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1287841796875,
0.62841796875,
0.1287841796875
],
[
0.77392578125,
0.11614990234375,
0.78759765625,
0.11614990234375,
0.78759765625,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.2197265625,
0.4375,
0.78515625,
0.4375,
0.78515625,
0.45068359375,
0.2197265625,
0.45068359375
],
[
0.2197265625,
0.451904296875,
0.78515625,
0.451904296875,
0.78515625,
0.46533203125,
0.2197265625,
0.46533203125
],
[
0.2197265625,
0.46533203125,
0.78759765625,
0.46533203125,
0.78759765625,
0.479248046875,
0.2197265625,
0.479248046875
],
[
0.2197265625,
0.48095703125,
0.3447265625,
0.48095703125,
0.3447265625,
0.4912109375,
0.2197265625,
0.4912109375
],
[
0.21728515625,
0.49755859375,
0.78515625,
0.496826171875,
0.78515625,
0.51123046875,
0.21728515625,
0.51220703125
],
[
0.2181396484375,
0.51318359375,
0.78369140625,
0.51318359375,
0.78369140625,
0.5263671875,
0.2181396484375,
0.5263671875
],
[
0.22216796875,
0.52978515625,
0.78369140625,
0.52978515625,
0.78369140625,
0.54052734375,
0.22216796875,
0.54052734375
],
[
0.2181396484375,
0.54248046875,
0.78173828125,
0.54248046875,
0.78173828125,
0.55615234375,
0.2181396484375,
0.55615234375
],
[
0.2181396484375,
0.55810546875,
0.7861328125,
0.55810546875,
0.7861328125,
0.5712890625,
0.2181396484375,
0.5712890625
],
[
0.2181396484375,
0.57275390625,
0.78515625,
0.57275390625,
0.78515625,
0.5859375,
0.2181396484375,
0.5859375
],
[
0.21728515625,
0.5869140625,
0.43310546875,
0.58837890625,
0.43310546875,
0.60302734375,
0.21728515625,
0.6015625
],
[
0.240234375,
0.60693359375,
0.78369140625,
0.609375,
0.78369140625,
0.62353515625,
0.240234375,
0.6220703125
],
[
0.2181396484375,
0.62353515625,
0.78369140625,
0.62353515625,
0.78369140625,
0.63818359375,
0.2181396484375,
0.63818359375
],
[
0.2181396484375,
0.6396484375,
0.7861328125,
0.6396484375,
0.7861328125,
0.65283203125,
0.2181396484375,
0.65283203125
],
[
0.2181396484375,
0.65380859375,
0.7861328125,
0.65283203125,
0.7861328125,
0.66748046875,
0.2181396484375,
0.66845703125
],
[
0.2197265625,
0.66943359375,
0.78173828125,
0.66943359375,
0.78173828125,
0.68310546875,
0.2197265625,
0.68310546875
],
[
0.2197265625,
0.68505859375,
0.78515625,
0.68505859375,
0.78515625,
0.69873046875,
0.2197265625,
0.69873046875
],
[
0.2197265625,
0.70068359375,
0.78369140625,
0.70068359375,
0.78369140625,
0.7138671875,
0.2197265625,
0.7138671875
],
[
0.24267578125,
0.7197265625,
0.78515625,
0.7197265625,
0.78515625,
0.73291015625,
0.24267578125,
0.73291015625
],
[
0.2181396484375,
0.73486328125,
0.78759765625,
0.73486328125,
0.78759765625,
0.74853515625,
0.2181396484375,
0.74853515625
],
[
0.2197265625,
0.75048828125,
0.78369140625,
0.75048828125,
0.78369140625,
0.7646484375,
0.2197265625,
0.7646484375
],
[
0.2197265625,
0.767578125,
0.78173828125,
0.767578125,
0.78173828125,
0.77783203125,
0.2197265625,
0.77783203125
],
[
0.2181396484375,
0.78076171875,
0.78369140625,
0.78076171875,
0.78369140625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.2181396484375,
0.79541015625,
0.78369140625,
0.79541015625,
0.78369140625,
0.8095703125,
0.2181396484375,
0.8095703125
],
[
0.21728515625,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.21728515625,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.72216796875,
0.8271484375,
0.72216796875,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"IntegratedPIFu",
"3",
"Fig.1. Compared with state-of-the-art methods, including (b) Geo-PIFu [8], (c) PIFu",
"[26], and (d) PIFuHD [27], our proposed model can precisely reconstruct various parts",
"of a clothed human body without noisy artefacts. Subject's face censored as required",
"bydataset's owner",
"approaches. Parametric approaches, such as [13,15,14], recover human body",
"shapes by estimating the parameters of a human parametric model (e.g. SMPL",
"18)from an input image.However,these parametric methods areonly able",
"to reconstruct “naked\" (without geometry of clothes) human meshes.Bhatna",
"gar et al. [3] and Jiang et al. [12] additionally estimate the clothes on top of",
"a human parametric model. However, these methods are unable to reconstruct",
"high-fidelitygeometry shapes.",
"Non-parametric approaches, on the other hand, do not require a human para-",
"metric model. For example, [28] use a 3D CNN to directly estimate a volu-",
"metric body shape from a single image. Notably, a subclass of non-parametric",
"approaches that have garner significant interest is pixel-aligned implicit models",
"[26,10,8,27]. One of the first pixel-aligned implicit models is the Pixel-Aligned Im-",
"plicit Function (PIFu) proposed by Saito et al. [26]. PIFu is memory efficient and",
"capable of producing high-fidelity 3D reconstructions of clothed human meshes.",
"Building from PIFu, there have been other pixel-aligned implicit models being",
"proposed. Examples include GeoPIFu [8] and StereoPIFu [10] that make use of",
"latent voxel features and an additional view respectively to improve the quality",
"of their reconstructed meshes.However,GeoPIFu is unable toreconstruct fine",
"geometry details of a human body, and StereoPIFu requires an additional view",
"and careful calibration of a binocular camera. On the other hand, PIFuHD [27]",
"which is a multi-scale variant of PIFu and a follow-up work by Saito et al., is",
"able to reconstruct very fine geometry details such as clothes wrinkles."
] |
{
"arxiv_id": "2211.07955",
"language": "en",
"page_idx": 2
}
|
||
2211.07955-en-3
|
2211.07955-en-3.jpg
|
4 K.Chan et al.
# 2.2 PIFu and PIFuHD
PIFu [26] can be interpreted as having two separate parts-1.Encoder and 2. Decoder. An illustration of the PIFU architecture is given in Fig. 2 (under) “Low-Resolution PIFu”). The encoder is a 2D convolutional network that extract feature maps from a RGB image. Points are sampled from the 3D camera space (of the RGB image), and their (x,y) coordinates are used to index the feature maps to retrieve a set of feature vectors. The z value of a point, together with the feature vector retrieved by the point, would then be used by the decoder (usually a multi-layer perceptron or MLP) to predict the binary occupancy value of that point. By evenly sampling the 3D camera space and employing the Marching Cube algorithm [19], a PIFu can reconstruct a 3D mesh.
PIFuHD [27] improves upon PIFu by using two PIFu modules that works at different resolutions. Additionally, PIFuHD also predicts frontal and rear normal maps from a single RGB image. In PIFuHD, the frst PIFu (i.e. low-res PIFu) is given a downsampled RGB image and a rear and frontal normal maps as input. The second PIFu (i.e. high-res PIFu) is given the same inputs but at the original resolution. For each point sampled from the 3D camera space, the high-res PIFu's decoder would obtain the point's corresponding feature vector from the encoder before combining that feature vector with additional features obtained from the low-res PIFu's decoder. The high-res PIFu's decoder would then predict the binary occupancy value of that point. A 3D mesh can then be generated using the same approach as PIFu.
## 2.3 Signed Distance Field
Our depth-oriented sampling technique drew inspiration from DeepSDF [22], which also employs a continuous signed distance function. But in our supplementary materials, we explain why DeepSDF, unlike DOS, is not a suitable solution for a pixel-aligned implicit model.
## 3 Method
Our method builds on PIFuHD, which we consider to be the current state-ofthe-art in this field. An overview of IntegratedPIFu is shown in Fig. 2. We train specialised networks to predict frontal normal, relative depth, and human parsing maps from a single 1024x1024 RGB image. Along with the image, these maps are downsampled (to 512x512 resolution), concatenated, and given to a PIFu The feature maps produced by the PIFu's encoder is then upsampled bilinearly before being fed into our High-Resolution Integrator. Our High-Resolution Integrator combines the upsampled feature maps with only the 1024x1024 RGB image and predicted frontal normal map before producing occupancy predictions using a MLP. We train IntegratedPIFu using our newly proposed depth-oriented $\operatorname{sampling}$scheme.
|
[
[
0.2156982421875,
0.11553955078125,
0.2320556640625,
0.11553955078125,
0.2320556640625,
0.1300048828125,
0.2156982421875,
0.1300048828125
],
[
0.27294921875,
0.11614990234375,
0.36767578125,
0.11614990234375,
0.36767578125,
0.1275634765625,
0.27294921875,
0.1275634765625
],
[
0.2181396484375,
0.149658203125,
0.26318359375,
0.149658203125,
0.26318359375,
0.1610107421875,
0.2181396484375,
0.1610107421875
],
[
0.256591796875,
0.147705078125,
0.414306640625,
0.147705078125,
0.414306640625,
0.1622314453125,
0.256591796875,
0.1622314453125
],
[
0.21728515625,
0.175537109375,
0.7861328125,
0.1767578125,
0.7861328125,
0.1912841796875,
0.21728515625,
0.1900634765625
],
[
0.21728515625,
0.1912841796875,
0.78515625,
0.1925048828125,
0.78515625,
0.20703125,
0.21728515625,
0.205810546875
],
[
0.221435546875,
0.2083740234375,
0.7861328125,
0.2083740234375,
0.7861328125,
0.2215576171875,
0.221435546875,
0.2215576171875
],
[
0.2181396484375,
0.222900390625,
0.7861328125,
0.222900390625,
0.7861328125,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.21728515625,
0.238037109375,
0.7861328125,
0.238037109375,
0.7861328125,
0.25244140625,
0.21728515625,
0.25244140625
],
[
0.21728515625,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.267578125,
0.21728515625,
0.267578125
],
[
0.2181396484375,
0.268310546875,
0.78369140625,
0.268310546875,
0.78369140625,
0.2822265625,
0.2181396484375,
0.2822265625
],
[
0.2181396484375,
0.28271484375,
0.78515625,
0.28271484375,
0.78515625,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.21728515625,
0.29736328125,
0.7861328125,
0.298583984375,
0.7861328125,
0.313232421875,
0.21728515625,
0.3125
],
[
0.2181396484375,
0.313232421875,
0.62158203125,
0.313232421875,
0.62158203125,
0.326904296875,
0.2181396484375,
0.326904296875
],
[
0.2442626953125,
0.328857421875,
0.78515625,
0.328857421875,
0.78515625,
0.343505859375,
0.2442626953125,
0.343505859375
],
[
0.2181396484375,
0.3447265625,
0.78759765625,
0.3447265625,
0.78759765625,
0.35791015625,
0.2181396484375,
0.35791015625
],
[
0.2181396484375,
0.359130859375,
0.78515625,
0.359130859375,
0.78515625,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.2156982421875,
0.373779296875,
0.7861328125,
0.373779296875,
0.7861328125,
0.38818359375,
0.2156982421875,
0.38818359375
],
[
0.21728515625,
0.39013671875,
0.7861328125,
0.39013671875,
0.7861328125,
0.404052734375,
0.21728515625,
0.404052734375
],
[
0.21728515625,
0.404052734375,
0.78759765625,
0.40478515625,
0.78759765625,
0.419189453125,
0.21728515625,
0.41845703125
],
[
0.2181396484375,
0.42041015625,
0.7861328125,
0.42041015625,
0.7861328125,
0.434326171875,
0.2181396484375,
0.434326171875
],
[
0.2181396484375,
0.43505859375,
0.7861328125,
0.43505859375,
0.7861328125,
0.448974609375,
0.2181396484375,
0.448974609375
],
[
0.2181396484375,
0.45068359375,
0.7861328125,
0.45068359375,
0.7861328125,
0.464111328125,
0.2181396484375,
0.464111328125
],
[
0.2181396484375,
0.46533203125,
0.7861328125,
0.46533203125,
0.7861328125,
0.479248046875,
0.2181396484375,
0.479248046875
],
[
0.21728515625,
0.48095703125,
0.53515625,
0.479736328125,
0.53515625,
0.494384765625,
0.21728515625,
0.49560546875
],
[
0.2197265625,
0.52294921875,
0.25244140625,
0.52294921875,
0.25244140625,
0.5341796875,
0.2197265625,
0.5341796875
],
[
0.259033203125,
0.521484375,
0.441162109375,
0.521484375,
0.441162109375,
0.5361328125,
0.259033203125,
0.5361328125
],
[
0.2181396484375,
0.55029296875,
0.78369140625,
0.55029296875,
0.78369140625,
0.56494140625,
0.2181396484375,
0.56494140625
],
[
0.2197265625,
0.56640625,
0.78369140625,
0.56640625,
0.78369140625,
0.58056640625,
0.2197265625,
0.58056640625
],
[
0.2181396484375,
0.58203125,
0.7861328125,
0.58203125,
0.7861328125,
0.59521484375,
0.2181396484375,
0.59521484375
],
[
0.21728515625,
0.5966796875,
0.52294921875,
0.59521484375,
0.52294921875,
0.60986328125,
0.21728515625,
0.611328125
],
[
0.2181396484375,
0.6396484375,
0.23779296875,
0.6396484375,
0.23779296875,
0.65283203125,
0.2181396484375,
0.65283203125
],
[
0.2467041015625,
0.63720703125,
0.330810546875,
0.63720703125,
0.330810546875,
0.65478515625,
0.2467041015625,
0.65478515625
],
[
0.2181396484375,
0.6748046875,
0.78369140625,
0.6748046875,
0.78369140625,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.78369140625,
0.70556640625,
0.78369140625,
0.72021484375,
0.2181396484375,
0.72021484375
],
[
0.21728515625,
0.7197265625,
0.78515625,
0.72021484375,
0.78515625,
0.73486328125,
0.21728515625,
0.734375
],
[
0.2197265625,
0.736328125,
0.78369140625,
0.736328125,
0.78369140625,
0.75,
0.2197265625,
0.75
],
[
0.2197265625,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.76513671875,
0.2197265625,
0.76513671875
],
[
0.21728515625,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.21728515625,
0.77978515625
],
[
0.21728515625,
0.77978515625,
0.78369140625,
0.77978515625,
0.78369140625,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.21728515625,
0.796875,
0.7861328125,
0.796875,
0.7861328125,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.78759765625,
0.81103515625,
0.78759765625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.21728515625,
0.82568359375,
0.343017578125,
0.8271484375,
0.343017578125,
0.84130859375,
0.21728515625,
0.84033203125
]
] |
[
"4",
"K. Chan et al.",
"2.2",
"PIFuand PIFuHD",
"PIFu [26] can be interpreted as having two separate parts - 1. Encoder and",
"2. Decoder. An illustration of the PIFU architecture is given in Fig. 2 (under",
"\"Low-Resolution PIFu\").The encoder is a 2D convolutional network that extract",
"feature maps from a RGB image. Points are sampled from the 3D camera space",
"(of the RGB image), and their (x,y) coordinates are used to index the feature",
"maps to retrieve a set of feature vectors. The z value of a point, together with the",
"feature vector retrieved by the point, would then be used by the decoder (usually",
"a multi-layer perceptron or MLP) to predict the binary occupancy value of that",
"point. By evenly sampling the 3D camera space and employing the Marching",
"Cube algorithm [19], a PIFu can reconstruct a 3D mesh.",
"PIFuHD [27] improves upon PIFu by using two PIFu modules that works at",
"different resolutions. Additionally, PIFuHD also predicts frontal and rear normal",
"maps from a single RGB image. In PIFuHD, the first PIFu (i.e. low-res PIFu)",
"is given a downsampled RGB image and a rear and frontal normal maps as",
"input. The second PIFu (i.e. high-res PIFu) is given the same inputs but at",
"the original resolution. For each point sampled from the 3D camera space, the",
"high-res PIFu's decoder would obtain the point's corresponding feature vector",
"from the encoder before combining that feature vector with additional features",
"obtained from the low-res PIFu's decoder. The high-res PIFu's decoder would",
"then predict the binary occupancy value of that point. A 3D mesh can then be",
"generated using the same approach as PIFu.",
"2.3",
"Signed Distance Field",
"Our depth-oriented sampling technique drew inspiration from DeepSDF [22]",
"which also employs a continuous signed distance function. But in our supple-",
"mentary materials, we explain why DeepSDF, unlike DOS, is not a suitable",
"solution for a pixel-aligned implicit model.",
"3 ",
" Method",
"Our method builds on PIFuHD, which we consider to be the current state-of-",
"the-art in this field. An overview of IntegratedPIFu is shown in Fig. 2. We train",
"specialised networks to predict frontal normal, relative depth, and human pars",
"ing maps from a single 1024x1024 RGB image. Along with the image, these maps",
"are downsampled (to 512x512 resolution), concatenated, and given to a PIFu.",
"The feature maps produced by the PIFu's encoder is then upsampled bilinearly",
"before being fed into our High-Resolution Integrator. Our High-Resolution In-",
"tegrator combines the upsampled feature maps with only the 1024x1024 RGB",
"image and predicted frontal normal map before producing occupancy predictions",
"using a MLP. We train IntegratedPIFu using our newly proposed depth-oriented",
"sampling scheme."
] |
{
"arxiv_id": "2211.07955",
"language": "en",
"page_idx": 3
}
|
||
2211.07955-en-4
|
2211.07955-en-4.jpg
|
5
IntegratedPIFu
<FigureHere>
Fig.2. Overview of our IntegratedPIFu framework. The Low-Resolution PIFu produces coarse feature maps that are further refined by High-Resolution Integrator. The HighResolution Integrator will then use a MLP to generate a reconstructed mesh. *We recommend the use of predicted depth and human parsing maps in the Low-Resolution PIFu, but the use of them in the High-Resolution Integrator is optional and depends heavily on the quality of the two maps.
We like to highlight that, unlike PIFuHD, IntegratedPIFu does not use a rear normal map as input as it tends to cause hallucinations of creases or wrinkles on the rear-side of a reconstructed mesh (More details in supplementary materials).
In short, IntegratedPIFu offers three new contributions: 1. Incorporating depth and parsing information 2. Depth-oriented sampling 3. High-resolution integrator. We will explain the motivations for and inner workings of each of them in the following sections.
### 3.1 Prediction and Incorporation of Relative Depth and Human Parse Information
Similar to PIFuHD [27], we use a pix2pixHD [29] network to predict a frontal normal map from a single RGB image. Additionally, we fnd that using this normal map as an additional input for relative depth and human parse prediction (carried out by specialised networks) helps to predict more accurate depth and human parse maps (shown in our supplementary materials).
Depth Prediction In our dataset, each input RGB image is rendered from a realistic human mesh. In order to reconstruct the 3D mesh shown in a single RGB image, a pixel-aligned implicit model is implicitly required to predict the depth of each mesh vertex in the 3D camera space. This is not an easy task due to depth ambiguity in a 2D RGB image. Also, 3D mesh vertices, when being projected onto an image, can have very similar RGB values, and yet have very different depth values in the 3D space.
Thus, we decided to partially shift the burden of depth prediction from a pixel-aligned implicit model to a separate, specialised network. This network (called depth predictor) would predict the relative depth values of mesh vertices
|
[
[
0.62841796875,
0.11614990234375,
0.73388671875,
0.11614990234375,
0.73388671875,
0.1287841796875,
0.62841796875,
0.1287841796875
],
[
0.77392578125,
0.11614990234375,
0.78759765625,
0.11614990234375,
0.78759765625,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.3203125,
0.1717529296875,
0.34228515625,
0.1717529296875,
0.34228515625,
0.2008056640625,
0.3203125,
0.2008056640625
],
[
0.328369140625,
0.2139892578125,
0.384033203125,
0.212158203125,
0.384033203125,
0.2235107421875,
0.328369140625,
0.2259521484375
],
[
0.334228515625,
0.212158203125,
0.343017578125,
0.212158203125,
0.343017578125,
0.216552734375,
0.334228515625,
0.216552734375
],
[
0.256591796875,
0.255615234375,
0.398681640625,
0.255615234375,
0.398681640625,
0.292236328125,
0.256591796875,
0.292236328125
],
[
0.36279296875,
0.2998046875,
0.371826171875,
0.2998046875,
0.371826171875,
0.3037109375,
0.36279296875,
0.3037109375
],
[
0.2181396484375,
0.31005859375,
0.78759765625,
0.31005859375,
0.78759765625,
0.323974609375,
0.2181396484375,
0.323974609375
],
[
0.2181396484375,
0.323974609375,
0.78515625,
0.323974609375,
0.78515625,
0.337158203125,
0.2181396484375,
0.337158203125
],
[
0.2181396484375,
0.33837890625,
0.7861328125,
0.33837890625,
0.7861328125,
0.3515625,
0.2181396484375,
0.3515625
],
[
0.2181396484375,
0.3515625,
0.7861328125,
0.3515625,
0.7861328125,
0.365478515625,
0.2181396484375,
0.365478515625
],
[
0.21728515625,
0.3642578125,
0.7861328125,
0.365478515625,
0.7861328125,
0.37890625,
0.21728515625,
0.37744140625
],
[
0.2156982421875,
0.37890625,
0.47802734375,
0.37890625,
0.47802734375,
0.392578125,
0.2156982421875,
0.392578125
],
[
0.24267578125,
0.4140625,
0.7861328125,
0.415283203125,
0.7861328125,
0.429931640625,
0.24267578125,
0.4287109375
],
[
0.21728515625,
0.43115234375,
0.7861328125,
0.43115234375,
0.7861328125,
0.4443359375,
0.21728515625,
0.4443359375
],
[
0.21728515625,
0.4462890625,
0.78369140625,
0.4462890625,
0.78369140625,
0.460205078125,
0.21728515625,
0.460205078125
],
[
0.2418212890625,
0.462158203125,
0.78759765625,
0.462158203125,
0.78759765625,
0.47607421875,
0.2418212890625,
0.47607421875
],
[
0.2197265625,
0.479248046875,
0.78515625,
0.479248046875,
0.78515625,
0.4892578125,
0.2197265625,
0.4892578125
],
[
0.2181396484375,
0.492431640625,
0.78759765625,
0.492431640625,
0.78759765625,
0.505859375,
0.2181396484375,
0.505859375
],
[
0.21728515625,
0.505859375,
0.439453125,
0.5068359375,
0.439453125,
0.521484375,
0.21728515625,
0.52099609375
],
[
0.2181396484375,
0.54736328125,
0.25244140625,
0.54736328125,
0.25244140625,
0.5595703125,
0.2181396484375,
0.5595703125
],
[
0.259033203125,
0.546875,
0.76123046875,
0.546875,
0.76123046875,
0.56103515625,
0.259033203125,
0.56103515625
],
[
0.261474609375,
0.5625,
0.41259765625,
0.5625,
0.41259765625,
0.57568359375,
0.261474609375,
0.57568359375
],
[
0.2156982421875,
0.58837890625,
0.7861328125,
0.58837890625,
0.7861328125,
0.6015625,
0.2156982421875,
0.6015625
],
[
0.2156982421875,
0.60400390625,
0.7861328125,
0.60302734375,
0.7861328125,
0.6171875,
0.2156982421875,
0.61865234375
],
[
0.21728515625,
0.61865234375,
0.78515625,
0.619140625,
0.78515625,
0.6337890625,
0.21728515625,
0.63330078125
],
[
0.2181396484375,
0.6337890625,
0.78515625,
0.6337890625,
0.78515625,
0.64794921875,
0.2181396484375,
0.64794921875
],
[
0.21728515625,
0.6494140625,
0.64794921875,
0.6484375,
0.64794921875,
0.66357421875,
0.21728515625,
0.6640625
],
[
0.21728515625,
0.68798828125,
0.7861328125,
0.689453125,
0.7861328125,
0.7041015625,
0.21728515625,
0.70263671875
],
[
0.2181396484375,
0.705078125,
0.7861328125,
0.705078125,
0.7861328125,
0.71826171875,
0.2181396484375,
0.71826171875
],
[
0.21728515625,
0.71826171875,
0.78759765625,
0.7197265625,
0.78759765625,
0.734375,
0.21728515625,
0.73291015625
],
[
0.2181396484375,
0.73486328125,
0.7861328125,
0.73486328125,
0.7861328125,
0.74853515625,
0.2181396484375,
0.74853515625
],
[
0.2181396484375,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.21728515625,
0.7666015625,
0.78515625,
0.76513671875,
0.78515625,
0.77880859375,
0.21728515625,
0.77978515625
],
[
0.2156982421875,
0.77880859375,
0.496826171875,
0.78076171875,
0.496826171875,
0.79443359375,
0.2156982421875,
0.79248046875
],
[
0.24267578125,
0.79541015625,
0.7861328125,
0.79541015625,
0.7861328125,
0.81005859375,
0.24267578125,
0.81005859375
],
[
0.2181396484375,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"IntegratedPIFu",
"5",
"一",
"Coecediedinnsborke",
"Py",
"大大贵",
"",
"Fig.2. Overview of our IntegratedPIFu framework. The Low-Resolution PIFu produces",
"coarse feature maps that are further refined by High-Resolution Integrator. The High-",
"Resolution Integrator will then use a MLP to generate a reconstructed mesh. *We",
"recommend the use of predicted depth and human parsing maps in the Low-Resolution",
"PIFu, but the use of them in the High-Resolution Integrator is optional and depends",
"heavily on the quality of the two maps.",
"We like to highlight that, unlike PIFuHD, IntegratedPIFu does not use a rear",
"normal map as input as it tends to cause hallucinations of creases or wrinkles on",
"the rear-side of a reconstructed mesh (More details in supplementary materials)",
"In short, IntegratedPIFu offers three new contributions: 1. Incorporating",
"depth andparsinginformation 2.Depth-oriented sampling 3.High-resolution",
"integrator. We will explain the motivations for and inner workings of each of",
"them in the following sections.",
"3.1",
"Prediction and Incorporation of Relative Depth and Human",
"Parse Information",
" Similar to PIFuHD [27l, we use a pix2pixHD [29] network to predict a frontal",
"normal map from a single RGB image. Additionally, we find that using this",
"normal map as an additional input for relative depth and human parse prediction",
"(carried out by specialised networks) helps to predict more accurate depth and",
"human parse maps (shown in our supplementary materials).",
"Depth Prediction In our dataset, each input RGB image is rendered from a",
"realistic human mesh. In order to reconstruct the 3D mesh shown in a single",
"RGB image, a pixel-aligned implicit model is implicitly required to predict the",
"depth of each mesh vertex in the 3D camera space. This is not an easy task due",
"to depth ambiguity in a 2D RGB image. Also, 3D mesh vertices, when being",
"projected onto an image, can have very similar RGB values, and yet have very",
"different depth values in the 3D space.",
"Thus, we decided to partially shift the burden of depth prediction from a",
"pixel-aligned implicit model to a separate, specialised network. This network",
"(called depth predictor) would predict the relative depth values of mesh vertices"
] |
{
"arxiv_id": "2211.07955",
"language": "en",
"page_idx": 4
}
|
||
1909.09256-en-0
|
1909.09256-en-0.jpg
|
## Triplet-Aware Scene Graph Embeddings
Brigit Schroeder$^*$
University of California, Santa Cruz
brschroe@ucsc.edu
Subarna Tripathi
Intel AI Lab
subarna.tripathi@intel.com
Hanlin Tang Intel AI Lab
hanlin.tangggintel.com
$ଶ୍ନି$
## $\mathbf{Abstract}$
Scene graphs have become an important form of struc-
$\begin{array}{cc}&\text{tured knowledge for tasks such as for image generation,}\\\textcircled{0}&\text{visual relation detection, visual question answering, and}\\\textcircled{0}&\text{image retrieval. While visualizing and interpreting word}\end{array}$
embeddings is well understood, scene graph embeddings have not been fully explored. In this work, we train scene graph embeddings in a layout generation task with different forms of supervision, specifically introducing triplet supervision and data augmentation. We see a significant performance increase in both metrics that measure the goodness $of\textit{layout prediction, mean intersection- over- union ( mIoU) }$ $(52.3\%$ vs. 49.2%) and relation score (61.7% vs. 54.1%), after the addition of triplet supervision and data augmentation. To understand how these different methods affect the scene graph representation, we apply several new visualization and evaluation methods to explore the evolution of the scene graph embedding. We find that triplet supervision significantly improves the embedding separability, which is highly correlated with the performance of the layout prediction model.
## $1. \textbf{Introduction}$
Scene graphs are a structured data format which encodes semantic relationships between objects [22]. Objects are represented as nodes in the graph and are connected by edges that expresses relationship, in the form of triplets. Each triplet is comprised of a <subject, predicate, object> which semantically describes the relationship between two objects such as <car, on, road> or <dog, left of, person>.
Scene graphs can be processed by graph convolutional neural networks (GCNN) which are able to pass information along graph edges[7]. Layout prediction models, such as those used image generation [7], are multi-stage networks (see Figure 1) which predict scene layout masks and object localization (bounding boxes), based upon the structure of the scene graph. The first stage of the layout predic-
$^{*}$Work done as a resident in Intel AI Lab.
tion model is a GCNN that is used to learn a scene graph embedding. Object embeddings are passed to the next stage of the network to predict a bounding box and mask for each object node in the scene graph. We are interested in layout prediction models in particular because a scene graph embedding is learned in the process of training a model. The performance metrics used for evaluating the quality of scene layout, namely mean intersection-over-union (mIoU) and relation score [15], can be used to correlate with the relative“quality” of the scene graph embedding. Relation score is particularly relevant as it measures the compliance of the generated layout to the scene graph, possibly giving insight into how well the embedding is structured, especially in terms of class separability.
In our work, we introduce several new supervisory signals that are conditioned upon triplet embeddings to train scene layout prediction models. We also apply data augmentation by using heuristic-based relationships [15] to maximize the number of triplets during training. The goal is to learn a triplet-aware scene graph embedding with the hypothesis additional supervision and data augmentation will enrichen the embedding representation. To this end, we also introduce several methods of embedding visualization and evaluation that help understand the degree of separability that is achieved from our improved layout prediction models.
### $2. \textbf{Related Work}$
Visual Genome [9], a human-annotated scene graph dataset, has known issues of incomplete and incorrect annotations. Additionally this only provides the bounding boxes for object instances but not their segmentation masks. On the other hand, the synthetic scene graphs generated from COCO stuff [2] are limited to simple geometric relationships (above, below, left, right, inside, surrounding) but are not hampered by incorrect annotations. COCO stuff also provides segmentation masks for instances. In this paper we use synthetic scene graphs from COCO stuff [2] for our experiments.
Generating scene graphs from visual features [22,24,
1
|
[
[
0.281005859375,
0.1319580078125,
0.68896484375,
0.1331787109375,
0.68896484375,
0.15087890625,
0.281005859375,
0.149658203125
],
[
0.1732177734375,
0.18310546875,
0.3095703125,
0.18310546875,
0.3095703125,
0.1976318359375,
0.1732177734375,
0.1976318359375
],
[
0.468017578125,
0.18310546875,
0.60107421875,
0.18310546875,
0.60107421875,
0.1976318359375,
0.468017578125,
0.1976318359375
],
[
0.73046875,
0.179931640625,
0.83251953125,
0.1817626953125,
0.83251953125,
0.199462890625,
0.73046875,
0.1976318359375
],
[
0.09722900390625,
0.2008056640625,
0.381591796875,
0.2008056640625,
0.381591796875,
0.2139892578125,
0.09722900390625,
0.2139892578125
],
[
0.484375,
0.2008056640625,
0.5849609375,
0.2008056640625,
0.5849609375,
0.21533203125,
0.484375,
0.21533203125
],
[
0.73388671875,
0.2020263671875,
0.830078125,
0.2020263671875,
0.830078125,
0.21337890625,
0.73388671875,
0.21337890625
],
[
0.16259765625,
0.2197265625,
0.314453125,
0.2197265625,
0.314453125,
0.23291015625,
0.16259765625,
0.23291015625
],
[
0.419189453125,
0.2197265625,
0.65185546875,
0.2197265625,
0.65185546875,
0.23291015625,
0.419189453125,
0.23291015625
],
[
0.68701171875,
0.2197265625,
0.8759765625,
0.2197265625,
0.8759765625,
0.23291015625,
0.68701171875,
0.23291015625
],
[
0.0236968994140625,
0.261962890625,
0.05963134765625,
0.261962890625,
0.05963134765625,
0.326904296875,
0.0236968994140625,
0.326904296875
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.5009765625,
0.26953125,
0.89208984375,
0.270751953125,
0.89208984375,
0.285400390625,
0.5009765625,
0.2841796875
],
[
0.50244140625,
0.285888671875,
0.89208984375,
0.285888671875,
0.89208984375,
0.2998046875,
0.50244140625,
0.2998046875
],
[
0.09722900390625,
0.300537109375,
0.46728515625,
0.3017578125,
0.46728515625,
0.31640625,
0.09722900390625,
0.31494140625
],
[
0.50244140625,
0.3017578125,
0.89208984375,
0.3017578125,
0.89208984375,
0.31494140625,
0.50244140625,
0.31494140625
],
[
0.02777099609375,
0.313232421875,
0.06451416015625,
0.31640625,
0.058013916015625,
0.36865234375,
0.021240234375,
0.365478515625
],
[
0.080078125,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.33203125,
0.080078125,
0.33203125
],
[
0.5009765625,
0.31640625,
0.890625,
0.31494140625,
0.890625,
0.330078125,
0.5009765625,
0.330810546875
],
[
0.080078125,
0.33203125,
0.4697265625,
0.333251953125,
0.4697265625,
0.347900390625,
0.080078125,
0.3466796875
],
[
0.50244140625,
0.33203125,
0.89208984375,
0.33203125,
0.89208984375,
0.34521484375,
0.50244140625,
0.34521484375
],
[
0.080078125,
0.348388671875,
0.4697265625,
0.348388671875,
0.4697265625,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.50244140625,
0.3466796875,
0.890625,
0.3466796875,
0.890625,
0.361083984375,
0.50244140625,
0.361083984375
],
[
0.08087158203125,
0.363037109375,
0.4697265625,
0.363037109375,
0.4697265625,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.50244140625,
0.3623046875,
0.89208984375,
0.3623046875,
0.89208984375,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.080078125,
0.37744140625,
0.4697265625,
0.37744140625,
0.4697265625,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.0269622802734375,
0.3642578125,
0.0555419921875,
0.3642578125,
0.0555419921875,
0.391357421875,
0.0269622802734375,
0.391357421875
],
[
0.50341796875,
0.37890625,
0.8896484375,
0.37890625,
0.8896484375,
0.389404296875,
0.50341796875,
0.389404296875
],
[
0.08087158203125,
0.393310546875,
0.4697265625,
0.393310546875,
0.4697265625,
0.4072265625,
0.08087158203125,
0.4072265625
],
[
0.50244140625,
0.392578125,
0.89208984375,
0.392578125,
0.89208984375,
0.406005859375,
0.50244140625,
0.406005859375
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.50244140625,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.08087158203125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.08087158203125,
0.4375
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50244140625,
0.436279296875
],
[
0.07843017578125,
0.4375,
0.468017578125,
0.438232421875,
0.468017578125,
0.45263671875,
0.07843017578125,
0.451904296875
],
[
0.5009765625,
0.438232421875,
0.890625,
0.438232421875,
0.890625,
0.451904296875,
0.5009765625,
0.451904296875
],
[
0.0318603515625,
0.448974609375,
0.058013916015625,
0.448974609375,
0.058013916015625,
0.48095703125,
0.0318603515625,
0.48095703125
],
[
0.03350830078125,
0.439453125,
0.045745849609375,
0.434326171875,
0.049835205078125,
0.43994140625,
0.037567138671875,
0.44580078125
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.50244140625,
0.45263671875,
0.890625,
0.45263671875,
0.890625,
0.466552734375,
0.50244140625,
0.466552734375
],
[
0.08087158203125,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.5009765625,
0.466552734375,
0.734375,
0.46728515625,
0.734375,
0.481689453125,
0.5009765625,
0.48095703125
],
[
0.08087158203125,
0.484130859375,
0.46728515625,
0.484130859375,
0.46728515625,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.51953125,
0.481689453125,
0.89208984375,
0.48291015625,
0.89208984375,
0.49755859375,
0.51953125,
0.496826171875
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.50341796875,
0.498779296875,
0.89208984375,
0.498779296875,
0.89208984375,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.50244140625,
0.51318359375,
0.89208984375,
0.51318359375,
0.89208984375,
0.52783203125,
0.50244140625,
0.52783203125
],
[
0.0236968994140625,
0.50048828125,
0.058013916015625,
0.50048828125,
0.058013916015625,
0.59765625,
0.0236968994140625,
0.59765625
],
[
0.07843017578125,
0.5146484375,
0.468017578125,
0.51318359375,
0.468017578125,
0.5263671875,
0.07843017578125,
0.52783203125
],
[
0.08087158203125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54150390625,
0.08087158203125,
0.54150390625
],
[
0.50244140625,
0.52880859375,
0.89208984375,
0.52880859375,
0.89208984375,
0.54248046875,
0.50244140625,
0.54248046875
],
[
0.080078125,
0.54443359375,
0.4697265625,
0.54443359375,
0.4697265625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50341796875,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.55615234375,
0.50341796875,
0.55615234375
],
[
0.080078125,
0.5595703125,
0.4697265625,
0.55810546875,
0.4697265625,
0.57275390625,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5595703125,
0.89208984375,
0.5595703125,
0.89208984375,
0.57275390625,
0.50244140625,
0.57275390625
],
[
0.080078125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.57373046875,
0.89208984375,
0.57373046875,
0.89208984375,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.0187835693359375,
0.58544921875,
0.056365966796875,
0.583984375,
0.060455322265625,
0.71044921875,
0.0228729248046875,
0.7109375
],
[
0.07843017578125,
0.5888671875,
0.156005859375,
0.5869140625,
0.156005859375,
0.6015625,
0.07843017578125,
0.60400390625
],
[
0.50341796875,
0.5888671875,
0.89208984375,
0.5888671875,
0.89208984375,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.50341796875,
0.60400390625,
0.89208984375,
0.60400390625,
0.89208984375,
0.6171875,
0.50341796875,
0.6171875
],
[
0.5048828125,
0.62060546875,
0.890625,
0.62060546875,
0.890625,
0.630859375,
0.5048828125,
0.630859375
],
[
0.035125732421875,
0.63525390625,
0.0400390625,
0.63525390625,
0.0400390625,
0.642578125,
0.035125732421875,
0.642578125
],
[
0.08087158203125,
0.63623046875,
0.20751953125,
0.63623046875,
0.20751953125,
0.65087890625,
0.08087158203125,
0.65087890625
],
[
0.50244140625,
0.6337890625,
0.89208984375,
0.6337890625,
0.89208984375,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.50244140625,
0.6494140625,
0.62841796875,
0.6494140625,
0.62841796875,
0.66357421875,
0.50244140625,
0.66357421875
],
[
0.09722900390625,
0.662109375,
0.4697265625,
0.662109375,
0.4697265625,
0.6767578125,
0.09722900390625,
0.6767578125
],
[
0.08087158203125,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.69140625,
0.08087158203125,
0.69140625
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50341796875,
0.6943359375,
0.640625,
0.6943359375,
0.640625,
0.708984375,
0.50341796875,
0.708984375
],
[
0.08087158203125,
0.70849609375,
0.46728515625,
0.70849609375,
0.46728515625,
0.7216796875,
0.08087158203125,
0.7216796875
],
[
0.080078125,
0.7236328125,
0.468017578125,
0.7236328125,
0.468017578125,
0.7373046875,
0.080078125,
0.7373046875
],
[
0.521484375,
0.71826171875,
0.89208984375,
0.7197265625,
0.89208984375,
0.734375,
0.521484375,
0.73291015625
],
[
0.08087158203125,
0.73779296875,
0.4697265625,
0.73779296875,
0.4697265625,
0.751953125,
0.08087158203125,
0.751953125
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.08087158203125,
0.75390625,
0.465576171875,
0.75390625,
0.465576171875,
0.767578125,
0.08087158203125,
0.767578125
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.098876953125,
0.76953125,
0.468017578125,
0.76953125,
0.468017578125,
0.78271484375,
0.098876953125,
0.78271484375
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.07843017578125,
0.78271484375,
0.468017578125,
0.7841796875,
0.468017578125,
0.798828125,
0.07843017578125,
0.79736328125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.0767822265625,
0.7998046875,
0.468017578125,
0.798828125,
0.468017578125,
0.81298828125,
0.0767822265625,
0.814453125
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.080078125,
0.81494140625,
0.468017578125,
0.81494140625,
0.468017578125,
0.8291015625,
0.080078125,
0.8291015625
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.830078125,
0.4697265625,
0.830078125,
0.4697265625,
0.84326171875,
0.08087158203125,
0.84326171875
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82470703125,
0.89208984375,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.08087158203125,
0.84521484375,
0.4697265625,
0.84521484375,
0.4697265625,
0.859375,
0.08087158203125,
0.859375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85986328125,
0.468017578125,
0.85986328125,
0.468017578125,
0.875,
0.080078125,
0.875
],
[
0.50244140625,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.50244140625,
0.869140625
],
[
0.50341796875,
0.87255859375,
0.5908203125,
0.87255859375,
0.5908203125,
0.8837890625,
0.50341796875,
0.8837890625
],
[
0.096435546875,
0.88720703125,
0.313720703125,
0.88720703125,
0.313720703125,
0.90087890625,
0.096435546875,
0.90087890625
],
[
0.52294921875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.52294921875,
0.90087890625
],
[
0.48193359375,
0.92822265625,
0.490234375,
0.92822265625,
0.490234375,
0.93310546875,
0.48193359375,
0.93310546875
]
] |
[
"Triplet-Aware Scene Graph Embeddings",
"Brigit Schroeder*",
"Subarna Tripathi",
"Hanlin Tang",
"University of California, Santa Cruz",
"Intel AI Lab",
"Intel AI Lab",
"brschroe@ucsc.edu",
"subarna.tripathi@intel.com",
"hanlin.tang@intel.com",
"6102",
"Abstract",
"tion model is a GCNN that is used to learn a scene graph",
"embedding. Object embeddings are passed to the next stage",
"Scene graphs have become an important form of struc-",
"of the network to predict a bounding box and mask for each",
"dos",
"tured knowledge for tasks such as for image generation,",
"object node in the scene graph. We are interested in lay-",
"visual relation detection, visual question answering, and",
"out prediction models in particular because a scene graph",
"image retrieval. While visualizing and interpreting word",
"embedding is learned in the process of training a model.",
"embeddings is well understood, scene graph embeddings",
"The performance metrics used for evaluating the quality of",
"have not been fully explored. In this work, we train scene",
"9",
"scenelayout,namelymeanintersection-over-union(mloU)",
"graph embeddings in a layout generation task with different",
"and relation score [15], can be used to correlate with the",
"LAD",
"forms of supervision, specifically introducing triplet super-",
"relative “quality\" of the scene graph embedding. Relation",
"vision and data augmentation. We see a significant perfor-",
"score is particularly relevant as it measures the compliance",
"mance increase in both metrics that measure the goodness",
"of the generated layout to the scene graph, possibly giv-",
"SO",
"",
"of layout prediction, mean intersection-over-union (mloU)",
"ing insight into how well the embedding is structured, espe-",
"(52.3% vs. 49.2%) and relation score (61.7% vs. 54.1%),",
"cially in terms of class separability.",
"after the addition of triplet supervision and data augmenta-",
"In our work, we introduce several new supervisory",
"tion. To understand how these different methods affect the",
"signals that are conditioned upon triplet embeddings to",
"train scene layout prediction models. We also apply data",
"1195260",
"scene graph representation, we apply several new visual-",
"ization and evaluation methodsto explore the evolution of",
"augmentation by using heuristic-based relationships [15] to",
"the scene graph embedding. We find that triplet supervision",
"maximize the number of triplets during training. The goal",
"significantly improves the embedding separability, which is",
"is to learn a triplet-aware scene graph embedding with the",
"highly correlated with the performance of the layout predic-",
"hypothesis additional supervision and data augmentation",
"6061:AXI",
"tion model.",
"will enrichen the embedding representation. To this end,",
"we also introduce several methods of embedding visu-",
"alization and evaluation that help understand the degree",
"",
"1. Introduction",
"of separability that is achieved from our improved layout",
"prediction models.",
" Scene graphs are a structured data format which encodes",
"semantic relationships between objects [22]. Objects are",
"represented as nodes in the graph and are connected by",
"2. Related Work",
"edges that expresses relationship, in the form of triplets.",
"Each triplet is comprised of a <subject, predicate, object>",
"Visual Genome [9], a human-annotated scene graph",
"which semantically describes the relationship between two",
"dataset, has known issues of incomplete and incorrect anno-",
"objects such as <car, on, road> or <dog, left of, person>.",
"tations. Additionally this only provides the bounding boxes",
"Scene graphs can be processed by graph convolutional",
"for object instances but not their segmentation masks. On",
"neural networks (GCNN) which are able to pass informa-",
"the other hand, the synthetic scene graphs generated from",
"tion along graph edges[7]. Layout prediction models, such",
"COCO stuff [2] are limited to simple geometric relation-",
"as those used image generation [7], are multi-stage net-",
"ships (above, below, left, right, inside, surrounding) but are",
"works (see Figure 1) which predict scene layout masks and",
"not hampered by incorrect annotations.COCo stuff also",
"object localization (bounding boxes), based upon the struc-",
"provides segmentation masks for instances. In this paper",
"ture of the scene graph. The first stage of the layout predic-",
"we use synthetic scene graphs from COCO stuff [2] for our",
"experiments.",
"*Work done as a resident in Intel AI Lab",
"Generating scene graphs from visual features [22, 24,",
"1"
] |
{
"arxiv_id": "1909.09256",
"language": "en",
"page_idx": 0
}
|
||
1909.09256-en-1
|
1909.09256-en-1.jpg
|
23,4,3,14,26,20,13,5] is a relatively explored task. Wan et al.[19] specifically predicts new triplets for scene graph completion using existing scene graphs and visual features. Another line of work that emerged recently [7,15,11,16,8,12,18] take scene graphs as input and produce final RGB images. All of these methods perform an intermediate layout prediction by learning embeddings of nodes. In closely related work, Belilovsky et al.[1] learn a joint visual-scene graph embedding for use in image retrieval. None of the above explored utilizing <subject, predicate, object> triplets as additional supervisory signal for more effective structured prediction.
Few papers that aim to detect visual relations from images also analyze the learned embeddings qualitatively. Dornadula $etal.[3]$ visualizes learned object category embeddings using 2-dimensional t-SNE plots. In GraphRCNN [23] Yang et al. showed how common sense of $<object-predicate>$ co-occurrence[6] emerges as the top scorers based upon the extracted weights that use visual features. They show similar common sense also appears for <object-object> co-occurrences. In the VTransE paper [25] the authors utilize t-SNE for visualizing the embeddings of predicate vectors and speculate when the model learns only co-occurrence vs. the actual meaning of the relationship by inspecting the neighbors in the t-SNE space. Note all these described models utilize visual features. In contrast, our model is trained from scene graphs without access to visual features. We provide quantitative and qualitative measures on why the embeddings trained by the proposed model is superior.
From an embedding introspection perspective, the closest work is [21]. In this concurrent work, Xu et al. introduced generic embeddings for common visual objects. They visualized embeddings of most common objects using t-SNE, and qualitatively studied results on vector decomposition and projection on specific axes anecdotally. We note that these embeddings are learned solely based on cooccurrence and thus they are not able to deal with relationships between objects.
## $3. \textbf{Method}$
Layout Prediction with Triplet Supervision. As part of our scene graph embedding introspection, we use a layout prediction network inspired by the image generation $\hat{\text{pipeline in}}[7].$ Figure 1 gives an overview of the network architecture. A GCNN processes an input scene graph to produce embeddings corresponding to object nodes in the graph. Singleton object embeddings are passed to the next stage of the layout prediction network per [7].
We utilize the object embeddings to form a set of triplet embeddings where each is composed of a <subject, pred- $icate,object>$ embedding. We pass these through a triplet mask prediction network. Figure 2 highlights the details of
the triplet mask prediction process. Rather than just learn individual class labels, the network learns to label objects as either subject or object, enforcing both an ordering and relationship between objects. We also pass triplet embeddings through a triplet superbox regression network, where we train the network to do joint localization over subject and object bounding boxes.
Ultimately, all of the outputs of the second stage of the layout prediction model are used to compose a scene layout mask with object localization.
Data Augmentation. We applied the method in [15] which uses heuristics to augment scene graphs with new spatial relations that induce a richer learned representation. The depth order between objects from an observer's viewpoint is quasi-exhaustively determined, and for 2D images, determining this order is non-trivial. Linear perspectivebased heuristics are utilized for augmenting spatial relationship vocabulary.
Training. We train the layout prediction network to minimize two additional triplet-based losses in addition to those used in [7]:
$\bullet$ Triplet mask loss, $L_triplet-mask$, penalizing differ-
ences between ground truth triplet masks and predicted
triplet masks with pixelwise cross-entropy loss.
$\bullet$ Triplet superbox loss, $L_superbox$, penalizing the $L_2$ dif-
ference between the ground truth and predicted triplet
superboxes.
$4. \textbf{Experiments}$
## 4.1. Dataset
COCO: We performed experiments on the 2017 COCOStuff dataset [2], which augments the COCO dataset [10] with additional stuff categories. The dataset annotates 40K train and 5K val images with bounding boxes and segmentation masks for 80 thing categories (people, cars, etc.) and 91 stuff categories (sky, grass, etc.). Similar to [7], we used thing and stuff annotations to construct synthetic scene graphs based on the 2D image coordinates of the objects, encoding six mutually exclusive geometric relationships: left of, right of, above, below, inside, surrounding. We ignored objects covering less than 2% of the image, and used images with 3 to 8 objects.
Table 1. Layout Prediction Model.
<table>
<tbody>
<tr>
<th>$\mathbf{Model}$</th>
<th>mloU</th>
<th>Relation $\operatorname{score}$</th>
</tr>
<tr>
<td>Baseline[7]</td>
<td>$49.2\%$</td>
<td>$54.1\%$</td>
</tr>
<tr>
<td>Triplet Supervision</td>
<td>$50.3\%$</td>
<td>$59.4\%$</td>
</tr>
<tr>
<td>Triplet Supervision+DA</td>
<td>$52.3\%$</td>
<td>$6\mathbf{1.7}\%$</td>
</tr>
</tbody>
</table>
|
[
[
0.07843017578125,
0.09027099609375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.07843017578125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.5009765625,
0.106689453125,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.080078125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.1405029296875,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.1405029296875,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.07843017578125,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.08087158203125,
0.1395263671875,
0.1422119140625,
0.1395263671875,
0.1422119140625,
0.149658203125,
0.08087158203125,
0.149658203125
],
[
0.50341796875,
0.1546630859375,
0.890625,
0.1546630859375,
0.890625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.080078125,
0.1673583984375,
0.4697265625,
0.1673583984375,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50341796875,
0.169189453125,
0.890625,
0.169189453125,
0.890625,
0.179931640625,
0.50341796875,
0.179931640625
],
[
0.080078125,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.6845703125,
0.18310546875,
0.6845703125,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1976318359375,
0.468017578125,
0.1976318359375,
0.468017578125,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.521484375,
0.1976318359375,
0.890625,
0.1976318359375,
0.890625,
0.2108154296875,
0.521484375,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.07843017578125,
0.2266845703125,
0.468017578125,
0.2279052734375,
0.468017578125,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.50341796875,
0.2298583984375,
0.69921875,
0.2298583984375,
0.69921875,
0.240478515625,
0.50341796875,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.521484375,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.521484375,
0.2568359375
],
[
0.08087158203125,
0.260009765625,
0.340576171875,
0.260009765625,
0.340576171875,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.2568359375,
0.890625,
0.25830078125,
0.890625,
0.272705078125,
0.50244140625,
0.271484375
],
[
0.09722900390625,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.09722900390625,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.302978515625,
0.468017578125,
0.302978515625,
0.468017578125,
0.31640625,
0.080078125,
0.31640625
],
[
0.5009765625,
0.302978515625,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.5009765625,
0.317626953125
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50341796875,
0.32080078125,
0.890625,
0.32080078125,
0.890625,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.5009765625,
0.333251953125,
0.89208984375,
0.333984375,
0.89208984375,
0.348388671875,
0.5009765625,
0.347900390625
],
[
0.08087158203125,
0.349853515625,
0.4697265625,
0.349853515625,
0.4697265625,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.50244140625,
0.349853515625,
0.609375,
0.349853515625,
0.609375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.52294921875,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.37890625,
0.52294921875,
0.37890625
],
[
0.07843017578125,
0.37744140625,
0.471435546875,
0.37890625,
0.471435546875,
0.393310546875,
0.07843017578125,
0.392578125
],
[
0.5,
0.37890625,
0.890625,
0.37890625,
0.890625,
0.392578125,
0.5,
0.392578125
],
[
0.07843017578125,
0.392578125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.07843017578125,
0.406005859375
],
[
0.50341796875,
0.396484375,
0.57861328125,
0.396484375,
0.57861328125,
0.40771484375,
0.50341796875,
0.40771484375
],
[
0.080078125,
0.40771484375,
0.46728515625,
0.4091796875,
0.46728515625,
0.423583984375,
0.080078125,
0.42236328125
],
[
0.5185546875,
0.417236328125,
0.8896484375,
0.41845703125,
0.8896484375,
0.436279296875,
0.5185546875,
0.43505859375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.53515625,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.53515625,
0.448974609375
],
[
0.08087158203125,
0.44140625,
0.465576171875,
0.44140625,
0.465576171875,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.080078125,
0.455078125,
0.46728515625,
0.455078125,
0.46728515625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.53515625,
0.45068359375,
0.84716796875,
0.45068359375,
0.84716796875,
0.464111328125,
0.53515625,
0.464111328125
],
[
0.08087158203125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.51953125,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.492431640625,
0.51953125,
0.492431640625
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.53515625,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.53515625,
0.505859375
],
[
0.08087158203125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51220703125,
0.08087158203125,
0.51220703125
],
[
0.5361328125,
0.50830078125,
0.615234375,
0.50830078125,
0.615234375,
0.51953125,
0.5361328125,
0.51953125
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.08087158203125,
0.5322265625,
0.2418212890625,
0.5322265625,
0.2418212890625,
0.54248046875,
0.08087158203125,
0.54248046875
],
[
0.50244140625,
0.53515625,
0.63134765625,
0.53515625,
0.63134765625,
0.5498046875,
0.50244140625,
0.5498046875
],
[
0.09722900390625,
0.54443359375,
0.468017578125,
0.54345703125,
0.468017578125,
0.55810546875,
0.09722900390625,
0.5595703125
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50341796875,
0.56103515625,
0.5947265625,
0.56103515625,
0.5947265625,
0.57275390625,
0.50341796875,
0.57275390625
],
[
0.080078125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.080078125,
0.5888671875,
0.468017578125,
0.59033203125,
0.468017578125,
0.60498046875,
0.080078125,
0.60400390625
],
[
0.521484375,
0.58349609375,
0.88818359375,
0.58349609375,
0.88818359375,
0.5966796875,
0.521484375,
0.5966796875
],
[
0.50244140625,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.080078125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.080078125,
0.619140625
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.6220703125,
0.4697265625,
0.6220703125,
0.4697265625,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50244140625,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.50244140625,
0.642578125
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.07843017578125,
0.6494140625,
0.46728515625,
0.65087890625,
0.46728515625,
0.66552734375,
0.07843017578125,
0.6640625
],
[
0.50341796875,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.50244140625,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.07843017578125,
0.66650390625,
0.2296142578125,
0.66650390625,
0.2296142578125,
0.6796875,
0.07843017578125,
0.6796875
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.080078125,
0.693359375,
0.1683349609375,
0.693359375,
0.1683349609375,
0.70849609375,
0.080078125,
0.70849609375
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.68798828125,
0.89208984375,
0.70263671875,
0.50244140625,
0.7041015625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.09722900390625,
0.72021484375,
0.471435546875,
0.72021484375,
0.471435546875,
0.734375,
0.09722900390625,
0.734375
],
[
0.50341796875,
0.7197265625,
0.8896484375,
0.7197265625,
0.8896484375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75,
0.6806640625,
0.75,
0.6806640625,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.07843017578125,
0.77978515625,
0.4697265625,
0.78076171875,
0.4697265625,
0.79541015625,
0.07843017578125,
0.79443359375
],
[
0.5888671875,
0.7822265625,
0.8056640625,
0.7822265625,
0.8056640625,
0.796875,
0.5888671875,
0.796875
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.52783203125,
0.806640625,
0.57861328125,
0.806640625,
0.57861328125,
0.8212890625,
0.52783203125,
0.8212890625
],
[
0.71142578125,
0.8095703125,
0.75244140625,
0.8095703125,
0.75244140625,
0.82080078125,
0.71142578125,
0.82080078125
],
[
0.765625,
0.8095703125,
0.86279296875,
0.8095703125,
0.86279296875,
0.82080078125,
0.765625,
0.82080078125
],
[
0.70751953125,
0.82275390625,
0.7509765625,
0.82275390625,
0.7509765625,
0.8369140625,
0.70751953125,
0.8369140625
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.53125,
0.82470703125,
0.60888671875,
0.82470703125,
0.60888671875,
0.8359375,
0.53125,
0.8359375
],
[
0.79150390625,
0.82275390625,
0.8349609375,
0.82275390625,
0.8349609375,
0.8369140625,
0.79150390625,
0.8369140625
],
[
0.70751953125,
0.8359375,
0.7509765625,
0.8359375,
0.7509765625,
0.85107421875,
0.70751953125,
0.85107421875
],
[
0.080078125,
0.8271484375,
0.3798828125,
0.8271484375,
0.3798828125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.53173828125,
0.8369140625,
0.65673828125,
0.8369140625,
0.65673828125,
0.85107421875,
0.53173828125,
0.85107421875
],
[
0.79150390625,
0.8359375,
0.8349609375,
0.8359375,
0.8349609375,
0.85107421875,
0.79150390625,
0.85107421875
],
[
0.70751953125,
0.849609375,
0.7509765625,
0.849609375,
0.7509765625,
0.8642578125,
0.70751953125,
0.8642578125
],
[
0.098876953125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.098876953125,
0.85498046875
],
[
0.53125,
0.85107421875,
0.69287109375,
0.85107421875,
0.69287109375,
0.8642578125,
0.53125,
0.8642578125
],
[
0.79150390625,
0.849609375,
0.83642578125,
0.849609375,
0.83642578125,
0.8642578125,
0.79150390625,
0.8642578125
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
]
] |
[
"23, 4, 3, 14, 26,20, 13, 5] is a relatively explored task.",
"the triplet mask prediction process. Rather than just learn",
"Wan et al. [19] specifically predicts new triplets for scene",
"individual class labels, the network learns to label objects",
"graph completion using existing scene graphs and visual",
"as either subject or object, enforcing both an ordering and",
"Another line of work that emerged recently",
"relationship between objects. We also pass triplet embed-",
"[7, 15, 11, 16, 8, 12, 18] take scene graphs as input and",
"features.",
"dings through a triplet superbox regression network, where",
"produce final RGB images. All of these methods perform",
"we train the network to do joint localization over subject",
"an intermediate layout prediction by learning embeddings",
"and object bounding boxes.",
"of nodes. In closely related work, Belilovsky et al. [1]",
"Ultimately, all of the outputs of the second stage of the",
"learn a joint visual-scene graph embedding for use in image",
"layout prediction model are used to compose a scene layout",
"retrieval.None of the above explored utilizing <subject,",
"maskwithobjectlocalization.",
"predicate, object> triplets as additional supervisory signal",
"Data Augmentation. We applied the method in [15]",
"formoreeffectivestructuredprediction.",
"which uses heuristics to augment scene graphs with new",
"Few papers that aim to detect visual relations from im-",
"spatial relations that induce a richer learned representation.",
"ages also analyze the learned embeddings qualitatively.",
"The depth order between objects from an observer's view-",
"Dornadula et al. [3] visualizes learned object category em-",
"point is quasi-exhaustively determined, and for 2D images,",
"beddingsusing 2-dimensional t-SNE plots.In Graph-",
"determining this order is non-trivial.Linear perspective-",
"RCNN [23] Yang et al. showed how common sense of",
"based heuristics are utilized for augmenting spatial relation-",
"<object-predicate> co-occurrence [6] emerges as the top",
"ship vocabulary.",
"scorers based upon the extracted weights that use visual",
"Training. We train the layout prediction network to min-",
"features.They show similar common sense also appears",
"imize two additional triplet-based losses in addition to those",
"for<object-object> co-occurrences.In theVTransEpaper",
"used in [7]:",
"[25] the authors utilize t-SNE for visualizing the embed-",
". Triplet mask loss, Ltriplet-mask, penalizing differ-",
"dings of predicate vectors and speculate when the model",
"ences between ground truth triplet masks and predicted",
"learnsonlyco-occurrencevs.theactualmeaning ofthere-",
"lationship by inspecting the neighbors in the t-SNE space.",
"triplet masks with pixelwise cross-entropy loss.",
"Noteall thesedescribedmodelsutilize visual features.In",
"• Triplet superbox loss, Lsuperboa, penalizing the L2 dif-",
"contrast, our model is trained from scene graphs without",
"ference between the ground truth and predicted triplet",
"access to visual features.Weprovidequantitativeand qual-",
"superboxes.",
"itative measures on why the embeddings trained by the pro-",
"posed model is superior.",
"4.Experiments",
"From an embedding introspection perspective, the clos-",
"est work is [21]. In this concurrent work, Xu et al. in-",
"4.1. Dataset",
"troduced generic embeddings for common visual objects.",
"They visualized embeddings of most common objects using",
"COCO: We performed experiments on the 2017 COCO",
"Stuff dataset [2], which augments the COCO dataset [10]",
"t-SNE, and qualitatively studied results on vector decom-",
"with additional stuff categories. The dataset annotates 40K",
"position and projection on specific axes anecdotally.We",
"train and 5K val images with bounding boxes and segmen-",
"note that these embeddings are learned solely based on co-",
"occurrence and thus they are not able to deal with relation-",
"tationmasksfor80thing categories(people.cars.etc.)and",
"91 stuff categories (sky, grass, etc.). Similar to [7], we",
"ships between objects.",
"used thing and stuff annotationsto construct synthetic scene",
"3. Method",
"graphs based on the 2D image coordinates of the objects,",
"encoding six mutually exclusive geometric relationships:",
"Layout Prediction with Triplet Supervision. As part",
"leff of, right of, above, below, inside, surrounding. We ig-",
"of our scene graph embedding introspection, we use a lay-",
"nored objects covering less than 2% of the image, and used",
"out prediction network inspired by the image generation",
"images with 3 to 8 objects.",
"pipeline in [7]. Figure 1 gives an overview of the network",
"architecture. A GCNN processes an input scene graph to",
"Table 1. Layout Prediction Model.",
"produce embeddings corresponding to object nodes in the",
"Model",
"mloU",
"Relation Score",
"49.2%",
"graph. Singleton object embeddings are passed to the next",
"Baseline [7]",
"54.1%",
"50.3%",
"stage of the layout prediction network per [7]",
"Triplet Supervision",
"59.4%",
"52.3%",
"We utilize the object embeddings to form a set of triplet",
"Triplet Supervision + DA",
"61.7%",
"embeddings where each is composed of a <subject, pred-",
"icate, object> embedding. We pass these through a triplet",
"mask prediction network. Figure 2 highlights the details of"
] |
{
"arxiv_id": "1909.09256",
"language": "en",
"page_idx": 1
}
|
||
1909.09256-en-2
|
1909.09256-en-2.jpg
|
<FigureHere>
Figure 1. Triplet-Supervised Layout Prediction Network. Here we show a multi-stage network which takes as input a scene graph and outputs a scene layout prediction. A graph convolutional neural network processes the scene graph to produce object embeddings, which can be used to form triplet embeddings composed of <subject, predicate, object>. Both single object and triplet embeddings are passed to the next stage to predict object masks, triplet masks and triplet superboxes. Ultimately, these are combined to predict a scene layout.
<FigureHere>
Figure 2. Triplet Mask Prediction.
Triplets containing a
$<subject,predicate,object>$found in a scene graph are used to predict corresponding triplet masks, labelling pixels either as subject and object. The mask prediction is used as supervisory signal during training.
With the triplet supervision and data augmentation, the individual classes become more separable, as seen in the progression of models. We quantified this observation by using an SVM with a linear kernel to classify the class categories (Table 2). The classification accuracy of the ten most frequent classes along with the mean classification accuracy across all classes is shown. The model trained with triplet supervision and data augmentation shows significantly better class separability over the baseline, as measured by classification accuracy. This is especially apparent in lower frequency classes such as wall, building and pavement.
### $4. 2. \textbf{Scene Layout Prediction}$
In order to evaluate the quality of scene layout prediction, we compared the predicted layout with the ground truth using intersection-over-union (mIoU) and relation score. The relation score measures the percentage of scene graph relations that are satisfied in the generated layout. The additional supervision and data augmentation introduced significantly improves over the baseline model [7] as can be seen in Table 1.
$4. 3. \textbf{Embedding Visualization and Separability}$
To visualize the results, we frst visualized the scene graph embedding using a t-SNE plot[17] (Figure 3) for the three models. For clarity, we only show the five most frequent classes in COCO-stuff dataset in the visualization.
In Figure 4, the mean scene embedding nodes for the fifty most frequent classes in COCO-stuff are visualized. We can see that several semantically similar objects clusters form, such as ones containing “animals”and“ground coverings" (indicated by colored boxes). Higher-level clusters such as“indoor/outdoor structural”(blue box) and“materials" (yellow box) which are also semantically similar at a higher level, are found closer together in embedding space. This kind of hierarchical clustering can also be seen in the 2d dendrogram in Figure 5 which visualizes a clustered heatmap of mean embedding distances between the top fifty most frequent classes. Distinct regions of blue (e.g. nearzero distance, other than along diagonal) indicate clusters of object classes that form correlated logical groupings in embedding space, such as outdoor objects (e.g. tree, bush, fence) and outdoor environments(e.g. sky, clouds, ground).
|
[
[
0.19775390625,
0.10797119140625,
0.223876953125,
0.10797119140625,
0.223876953125,
0.11614990234375,
0.19775390625,
0.11614990234375
],
[
0.310546875,
0.111083984375,
0.393798828125,
0.111083984375,
0.393798828125,
0.121826171875,
0.310546875,
0.121826171875
],
[
0.483642578125,
0.117431640625,
0.5185546875,
0.11553955078125,
0.5185546875,
0.12249755859375,
0.483642578125,
0.125
],
[
0.193603515625,
0.12249755859375,
0.2303466796875,
0.12249755859375,
0.2303466796875,
0.1331787109375,
0.193603515625,
0.1331787109375
],
[
0.48193359375,
0.12371826171875,
0.494384765625,
0.12371826171875,
0.494384765625,
0.1275634765625,
0.48193359375,
0.1275634765625
],
[
0.191162109375,
0.1331787109375,
0.3056640625,
0.1331787109375,
0.3056640625,
0.146484375,
0.191162109375,
0.146484375
],
[
0.1707763671875,
0.1351318359375,
0.1871337890625,
0.1351318359375,
0.1871337890625,
0.143310546875,
0.1707763671875,
0.143310546875
],
[
0.1846923828125,
0.11297607421875,
0.240234375,
0.11297607421875,
0.240234375,
0.125,
0.1846923828125,
0.125
],
[
0.1683349609375,
0.146484375,
0.2059326171875,
0.149658203125,
0.205078125,
0.1610107421875,
0.1666259765625,
0.1578369140625
],
[
0.56884765625,
0.141357421875,
0.62353515625,
0.141357421875,
0.62353515625,
0.15283203125,
0.56884765625,
0.15283203125
],
[
0.2017822265625,
0.147705078125,
0.2418212890625,
0.147705078125,
0.2418212890625,
0.1590576171875,
0.2017822265625,
0.1590576171875
],
[
0.25244140625,
0.149658203125,
0.281005859375,
0.149658203125,
0.281005859375,
0.1578369140625,
0.25244140625,
0.1578369140625
],
[
0.1953125,
0.1717529296875,
0.226318359375,
0.1717529296875,
0.226318359375,
0.189453125,
0.1953125,
0.189453125
],
[
0.2091064453125,
0.1666259765625,
0.2320556640625,
0.1641845703125,
0.233642578125,
0.17236328125,
0.2099609375,
0.1749267578125
],
[
0.1968994140625,
0.1944580078125,
0.2181396484375,
0.1944580078125,
0.2181396484375,
0.20263671875,
0.1968994140625,
0.20263671875
],
[
0.47802734375,
0.1988525390625,
0.51953125,
0.1988525390625,
0.51953125,
0.2020263671875,
0.47802734375,
0.2020263671875
],
[
0.20751953125,
0.2083740234375,
0.236083984375,
0.2083740234375,
0.236083984375,
0.216552734375,
0.20751953125,
0.216552734375
],
[
0.56884765625,
0.216552734375,
0.6259765625,
0.216552734375,
0.6259765625,
0.2279052734375,
0.56884765625,
0.2279052734375
],
[
0.1690673828125,
0.2291259765625,
0.253173828125,
0.2291259765625,
0.253173828125,
0.2392578125,
0.1690673828125,
0.2392578125
],
[
0.340576171875,
0.251953125,
0.34716796875,
0.251953125,
0.34716796875,
0.255126953125,
0.340576171875,
0.255126953125
],
[
0.400390625,
0.251953125,
0.428955078125,
0.251953125,
0.428955078125,
0.255615234375,
0.400390625,
0.255615234375
],
[
0.71337890625,
0.25,
0.7890625,
0.25,
0.7890625,
0.260009765625,
0.71337890625,
0.260009765625
],
[
0.348876953125,
0.25830078125,
0.419189453125,
0.25830078125,
0.419189453125,
0.268310546875,
0.348876953125,
0.268310546875
],
[
0.340576171875,
0.26953125,
0.361083984375,
0.26953125,
0.361083984375,
0.272705078125,
0.340576171875,
0.272705078125
],
[
0.39453125,
0.26953125,
0.426513671875,
0.26953125,
0.426513671875,
0.272705078125,
0.39453125,
0.272705078125
],
[
0.48193359375,
0.31494140625,
0.521484375,
0.31494140625,
0.521484375,
0.3193359375,
0.48193359375,
0.3193359375
],
[
0.56640625,
0.31005859375,
0.6396484375,
0.31005859375,
0.6396484375,
0.3212890625,
0.56640625,
0.3212890625
],
[
0.080078125,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.080078125,
0.348388671875,
0.89208984375,
0.348388671875,
0.89208984375,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.08087158203125,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.375,
0.08087158203125,
0.375
],
[
0.07843017578125,
0.376953125,
0.87353515625,
0.376953125,
0.87353515625,
0.39013671875,
0.07843017578125,
0.39013671875
],
[
0.28759765625,
0.427978515625,
0.361083984375,
0.4267578125,
0.361083984375,
0.4375,
0.28759765625,
0.438232421875
],
[
0.50244140625,
0.427978515625,
0.89208984375,
0.4287109375,
0.89208984375,
0.443115234375,
0.50244140625,
0.442626953125
],
[
0.29736328125,
0.439453125,
0.336669921875,
0.439453125,
0.336669921875,
0.443115234375,
0.29736328125,
0.443115234375
],
[
0.50244140625,
0.4443359375,
0.890625,
0.4443359375,
0.890625,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.08660888671875,
0.45263671875,
0.1405029296875,
0.45263671875,
0.1405029296875,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.41015625,
0.451904296875,
0.428955078125,
0.451904296875,
0.428955078125,
0.455810546875,
0.41015625,
0.455810546875
],
[
0.10455322265625,
0.464111328125,
0.125,
0.464111328125,
0.125,
0.472900390625,
0.10455322265625,
0.472900390625
],
[
0.50341796875,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.50244140625,
0.473388671875,
0.89208984375,
0.473388671875,
0.89208984375,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.50341796875,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.10540771484375,
0.5087890625,
0.11273193359375,
0.51123046875,
0.111083984375,
0.5146484375,
0.10296630859375,
0.51318359375
],
[
0.50244140625,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50244140625,
0.51806640625
],
[
0.09722900390625,
0.52294921875,
0.131591796875,
0.52294921875,
0.131591796875,
0.5322265625,
0.09722900390625,
0.5322265625
],
[
0.5009765625,
0.51953125,
0.89208984375,
0.51806640625,
0.89208984375,
0.5322265625,
0.5009765625,
0.53271484375
],
[
0.29736328125,
0.5390625,
0.318603515625,
0.5390625,
0.318603515625,
0.54248046875,
0.29736328125,
0.54248046875
],
[
0.50244140625,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50244140625,
0.548828125
],
[
0.330810546875,
0.55029296875,
0.4697265625,
0.55029296875,
0.4697265625,
0.564453125,
0.330810546875,
0.564453125
],
[
0.50244140625,
0.5498046875,
0.89208984375,
0.5498046875,
0.89208984375,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.50244140625,
0.56494140625,
0.890625,
0.56494140625,
0.890625,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.5517578125,
0.31787109375,
0.5517578125,
0.31787109375,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.080078125,
0.56640625,
0.468017578125,
0.56640625,
0.468017578125,
0.580078125,
0.080078125,
0.580078125
],
[
0.07843017578125,
0.5791015625,
0.4697265625,
0.5791015625,
0.4697265625,
0.59228515625,
0.07843017578125,
0.59228515625
],
[
0.50244140625,
0.580078125,
0.84619140625,
0.580078125,
0.84619140625,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.07843017578125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.07843017578125,
0.60693359375
],
[
0.0767822265625,
0.60595703125,
0.156005859375,
0.60791015625,
0.156005859375,
0.62255859375,
0.0767822265625,
0.62060546875
],
[
0.08087158203125,
0.66015625,
0.2998046875,
0.66015625,
0.2998046875,
0.6748046875,
0.08087158203125,
0.6748046875
],
[
0.52294921875,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.52294921875,
0.67236328125
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.1004638671875,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69873046875,
0.1004638671875,
0.69873046875
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.07843017578125,
0.70068359375,
0.4697265625,
0.70068359375,
0.4697265625,
0.7138671875,
0.07843017578125,
0.7138671875
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.5009765625,
0.72021484375,
0.8935546875,
0.7197265625,
0.8935546875,
0.734375,
0.5009765625,
0.73486328125
],
[
0.07843017578125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.07843017578125,
0.744140625
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.74560546875,
0.468017578125,
0.74560546875,
0.468017578125,
0.7587890625,
0.080078125,
0.7587890625
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76025390625,
0.468017578125,
0.76025390625,
0.468017578125,
0.7734375,
0.080078125,
0.7734375
],
[
0.50244140625,
0.7666015625,
0.8896484375,
0.7666015625,
0.8896484375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.7763671875,
0.468017578125,
0.7763671875,
0.468017578125,
0.787109375,
0.08087158203125,
0.787109375
],
[
0.50341796875,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.0767822265625,
0.79052734375,
0.2303466796875,
0.7890625,
0.2303466796875,
0.80322265625,
0.0767822265625,
0.8037109375
],
[
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
],
[
0.08087158203125,
0.8173828125,
0.435546875,
0.8173828125,
0.435546875,
0.83056640625,
0.08087158203125,
0.83056640625
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.101318359375,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.101318359375,
0.85302734375
],
[
0.50244140625,
0.84130859375,
0.8935546875,
0.84130859375,
0.8935546875,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.08087158203125,
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,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.88818359375,
0.88720703125,
0.88818359375,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"Graph",
"Object Embeddings",
"ObjectMast",
"Network",
"diet",
"wallconcrete wll-wood",
"dirt",
"Convolutional",
"below",
"Object Mask",
"rightof",
"above",
"lentor",
"giraff",
"giraff",
"rodietineNaty",
"rightof",
"Triplet Masks",
"Input: Scene Graph",
"0",
"objec",
"Layout Prediction",
"Triplet Embedding",
"",
"Bas",
"RegressionNe",
"Triplet Superbox",
"Figure 1. Triplet-Supervised Layout Prediction Network. Here we show a multi-stage network which takes as input a scene graph and",
"outputs a scene layout prediction. A graph convolutional neural network processes the scene graph to produce object embeddings, which",
"can beused toformtriplet embeddings composed of <subject,predicate,object>.Both single object and triplet embeddings arepassed",
"to the next stage to predict object masks, triplet masks and triplet superboxes. Ultimately, these are combined to predict a scene layout.",
"Triplet Mask Predictior",
"With the triplet supervision and data augmentation, the in-",
"ctmrayot",
"dividual classes become more separable, as seen in the pro-",
"ceneGranhTrinl",
"cene",
"wall",
"gression of models.We quantified this observation by using",
"an SVM with a linear kernel to classify the class categories",
"(Table 2). The classification accuracy of the ten most fre-",
"he",
"quent classes along with the mean classification accuracy",
"keyboard",
"across all classes is shown. The model trained with triplet",
"evbo",
"supervision and data augmentation shows significantly bet-",
"Triplets containing a",
"ter class separability over the baseline, as measured by clas-",
"sification accuracy. This is especially apparent in lower fre-",
"Figure 2.Triplet Mask Prediction.",
"<subject.predicate,object>found in a scene graph are used to pre-",
"dict corresponding triplet masks, labelling pixels either as subject",
"quency classes such as wall, building and pavement.",
"and object. The mask prediction is used as supervisory signal dur-",
"ing training.",
"4.2.Scene LayoutPrediction",
"In Figure 4, the mean scene embedding nodes for the",
"fifty most frequent classes in COCO-stuff are visualized.",
"In order to evaluate the quality of scene layout predic-",
"We can see that several semantically similar objects clusters",
"tion, we compared the predicted layout with the ground",
"form, such as ones containing \"animals\" and “ground cov-",
"truth using intersection-over-union (mloU) and relation",
"erings\" (indicated by colored boxes). Higher-level clusters",
"score.The relation scoremeasures the percentage of scene",
"such as “indoor/outdoor structural\" (blue box) and “mate-",
"graph relations that are satisfied in the generated layout.",
"rials(yellowbox)whicharealsosemanticallysimilarata",
"The additional supervision and data augmentation intro-",
"higher level, are found closer together in embedding space.",
"ducedsignificantlyimprovesoverthebaselinemodel[7las",
"This kind of hierarchical clustering can also be seen in the",
"can be seen in Table 1.",
"2d dendrogram in Figure 5 which visualizes a clustered",
"heatmap of mean embedding distances between the top fifty",
"4.3.Embedding Visualization and Separability",
"most frequent classes. Distinct regions of blue (e.g. near-",
"Tovisualizethe results.we firstvisualized the scene",
"zero distance, other than along diagonal) indicate clusters",
"graph embedding using a t-SNE plot [17] (Figure 3) for",
"of object classes that form correlated logical groupings in",
"the three models. For clarity, we only show the five most",
"embedding space, such as outdoor objects (e.g. tree, bush,",
"frequent classes in COCO-stuff dataset in the visualization.",
"fence) and outdoor environments(e.g. sky, clouds, ground)"
] |
{
"arxiv_id": "1909.09256",
"language": "en",
"page_idx": 2
}
|
||
1310.6257-en-0
|
1310.6257-en-0.jpg
|
<table>
<tbody>
<tr>
<td>$\mathbf{VLDBJ}$ manuscript No. will be inserted by the editor) </td>
</tr>
</tbody>
</table>
# Dissociation and Propagation for Approximate Lifted Inference with Standard Relational Database Management Systems
Wolfgang Gatterbauer $\cdot$ Dan Suciu
challenging data management problem since exact infer- $\Xi$ence is generally \# P- hard and is most often solved approxi- Probabilistic inference over large data sets is becoming a mately with sampling- based methods today. This paper pro- central data $\begin{array}{lll}&\text{poses an altermative approach for approximate evaluation of}\end{array}$ our approcach, every query is evaluated entirely in thes. Data sets with msisng values are base engineb yevaluating afzexd numberof querv, plans, often “completed”,using inference ing raphical models [8,each providing an upper boundo nthe true probabili
only the minimal necessary plans among all possible plans.
$\begin{array}{l\text{mptantulis is algcinthines arictc gentalzicinon ofdlll mecettinin.and prbbbhiististes altatabsencencenappliedo}\\&\text{thownTMTHis algefyrimprises aryulgtoritimpuerts.ostensin}&\text{bottopppengr,prosthinstecis}&\text{khowntobsterecis}&\text{kshowntobsterecis}&\text{kshownt$
corresponds to PTIME plans in probabilistic databases Keywords Probabilistic inference $\cdot$ Lifted inference $\cdot$
Probabilistic databases $\cdot$ Problem relaxation $\cdot$ Ranking $\cdot$
Query plans $\cdot$Query optimization
<table>
<tbody>
<tr>
<td>1</td>
<td>Gatterbauer ram</td>
<td>auer</td>
</tr>
<tr>
<td> </td>
<td>usiness</td>
<td>Business</td>
</tr>
<tr>
<td> </td>
<td>${\mathrm{niversity}}$</td>
<td>${\mathrm{niversity}}$</td>
</tr>
<tr>
<td> </td>
<td>E-mail: gatt@cmu.edu</td>
<td>$..edu$</td>
</tr>
<tr>
<td> </td>
<td>Dan Suciu</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>neerino</td>
<td>eering</td>
</tr>
<tr>
<td> </td>
<td>$\operatorname{ington}$</td>
<td>ington</td>
</tr>
</tbody>
</table>
### 1 Introduction
that are inferred from unstructured information [9] are also use query plans to guide a DPLL-based algorithm for Weighted Model Counting [35]. While both approaches deploy some advanced relational optimization techniques, at their core they are based on general purpose probabilistic inference techniques, which either run in exponential time (DPLL-based algorithms have been proven recently to take exponential time even for queries computable in polynomial time [4]), or require many iterations until convergence
In this paper, we propose a different approach to query evaluation with probabilistic databases (PDBs). In our approach, every query is evaluated entirely in the database en- $gine.$ Probability computation is done at query time, using
|
[
[
0.07135009765625,
0.05877685546875,
0.2476806640625,
0.0611572265625,
0.2476806640625,
0.07598876953125,
0.07135009765625,
0.0736083984375
],
[
0.07135009765625,
0.0736083984375,
0.279541015625,
0.0736083984375,
0.279541015625,
0.0872802734375,
0.07135009765625,
0.0872802734375
],
[
0.07135009765625,
0.163330078125,
0.85791015625,
0.163330078125,
0.85791015625,
0.18115234375,
0.07135009765625,
0.18115234375
],
[
0.0679931640625,
0.182861328125,
0.67236328125,
0.18408203125,
0.67236328125,
0.201904296875,
0.0679931640625,
0.20068359375
],
[
0.0697021484375,
0.22265625,
0.328369140625,
0.22265625,
0.328369140625,
0.2369384765625,
0.0697021484375,
0.2369384765625
],
[
0.026031494140625,
0.31640625,
0.056243896484375,
0.31640625,
0.056243896484375,
0.3349609375,
0.026031494140625,
0.3349609375
],
[
0.026031494140625,
0.327880859375,
0.05792236328125,
0.327880859375,
0.05792236328125,
0.358154296875,
0.026031494140625,
0.358154296875
],
[
0.0697021484375,
0.3486328125,
0.47265625,
0.3486328125,
0.47265625,
0.3623046875,
0.0697021484375,
0.3623046875
],
[
0.495361328125,
0.3466796875,
0.60693359375,
0.3486328125,
0.60693359375,
0.363525390625,
0.495361328125,
0.361083984375
],
[
0.0679931640625,
0.363525390625,
0.47265625,
0.363525390625,
0.47265625,
0.377685546875,
0.0679931640625,
0.377685546875
],
[
0.029388427734375,
0.3515625,
0.0545654296875,
0.3515625,
0.0545654296875,
0.366455078125,
0.029388427734375,
0.366455078125
],
[
0.033599853515625,
0.366455078125,
0.060455322265625,
0.3681640625,
0.056243896484375,
0.40087890625,
0.029388427734375,
0.3984375
],
[
0.06634521484375,
0.377197265625,
0.470947265625,
0.377685546875,
0.470947265625,
0.392578125,
0.06634521484375,
0.391357421875
],
[
0.495361328125,
0.377197265625,
0.89990234375,
0.37890625,
0.89990234375,
0.392578125,
0.495361328125,
0.39013671875
],
[
0.0679931640625,
0.393798828125,
0.470947265625,
0.393798828125,
0.470947265625,
0.40673828125,
0.0679931640625,
0.40673828125
],
[
0.496337890625,
0.393798828125,
0.8984375,
0.393798828125,
0.8984375,
0.40673828125,
0.496337890625,
0.40673828125
],
[
0.039459228515625,
0.402587890625,
0.05877685546875,
0.416259765625,
0.039459228515625,
0.429931640625,
0.0201568603515625,
0.416259765625
],
[
0.033599853515625,
0.3955078125,
0.04534912109375,
0.3955078125,
0.04534912109375,
0.399658203125,
0.033599853515625,
0.399658203125
],
[
0.06634521484375,
0.407958984375,
0.474365234375,
0.40673828125,
0.474365234375,
0.421630859375,
0.06634521484375,
0.4228515625
],
[
0.496337890625,
0.40673828125,
0.8984375,
0.40673828125,
0.8984375,
0.42041015625,
0.496337890625,
0.42041015625
],
[
0.0679931640625,
0.4228515625,
0.47265625,
0.4228515625,
0.47265625,
0.435791015625,
0.0679931640625,
0.435791015625
],
[
0.495361328125,
0.4228515625,
0.89990234375,
0.4228515625,
0.89990234375,
0.435791015625,
0.495361328125,
0.435791015625
],
[
0.495361328125,
0.43701171875,
0.8984375,
0.43701171875,
0.8984375,
0.45068359375,
0.495361328125,
0.45068359375
],
[
0.03106689453125,
0.42578125,
0.042816162109375,
0.424560546875,
0.04449462890625,
0.4287109375,
0.031890869140625,
0.429931640625
],
[
0.0679931640625,
0.438232421875,
0.47265625,
0.438232421875,
0.47265625,
0.451904296875,
0.0679931640625,
0.451904296875
],
[
0.0277099609375,
0.453125,
0.056243896484375,
0.453125,
0.056243896484375,
0.47265625,
0.0277099609375,
0.47265625
],
[
0.0679931640625,
0.453125,
0.470947265625,
0.453125,
0.470947265625,
0.466064453125,
0.0679931640625,
0.466064453125
],
[
0.496337890625,
0.451904296875,
0.8984375,
0.451904296875,
0.8984375,
0.465576171875,
0.496337890625,
0.465576171875
],
[
0.03106689453125,
0.468505859375,
0.05206298828125,
0.468505859375,
0.05206298828125,
0.485107421875,
0.03106689453125,
0.485107421875
],
[
0.0679931640625,
0.46728515625,
0.47265625,
0.46728515625,
0.47265625,
0.48095703125,
0.0679931640625,
0.48095703125
],
[
0.497802734375,
0.468505859375,
0.8984375,
0.468505859375,
0.8984375,
0.478515625,
0.497802734375,
0.478515625
],
[
0.0679931640625,
0.482177734375,
0.47265625,
0.482177734375,
0.47265625,
0.495361328125,
0.0679931640625,
0.495361328125
],
[
0.496337890625,
0.482177734375,
0.89990234375,
0.482177734375,
0.89990234375,
0.495361328125,
0.496337890625,
0.495361328125
],
[
0.0352783203125,
0.490478515625,
0.056243896484375,
0.490478515625,
0.056243896484375,
0.5029296875,
0.0352783203125,
0.5029296875
],
[
0.0679931640625,
0.496337890625,
0.47265625,
0.496337890625,
0.47265625,
0.51025390625,
0.0679931640625,
0.51025390625
],
[
0.496337890625,
0.496337890625,
0.89990234375,
0.496337890625,
0.89990234375,
0.51025390625,
0.496337890625,
0.51025390625
],
[
0.0679931640625,
0.51123046875,
0.470947265625,
0.51123046875,
0.470947265625,
0.52490234375,
0.0679931640625,
0.52490234375
],
[
0.0352783203125,
0.50048828125,
0.056243896484375,
0.50048828125,
0.056243896484375,
0.51416015625,
0.0352783203125,
0.51416015625
],
[
0.496337890625,
0.51123046875,
0.89990234375,
0.51123046875,
0.89990234375,
0.52490234375,
0.496337890625,
0.52490234375
],
[
0.0679931640625,
0.52685546875,
0.47265625,
0.52685546875,
0.47265625,
0.54052734375,
0.0679931640625,
0.54052734375
],
[
0.49365234375,
0.52490234375,
0.90185546875,
0.525390625,
0.90185546875,
0.54052734375,
0.49365234375,
0.5390625
],
[
0.06634521484375,
0.5390625,
0.470947265625,
0.54150390625,
0.470947265625,
0.5546875,
0.06634521484375,
0.552734375
],
[
0.031890869140625,
0.52978515625,
0.0545654296875,
0.53369140625,
0.050384521484375,
0.54736328125,
0.0277099609375,
0.54443359375
],
[
0.0352783203125,
0.54345703125,
0.0545654296875,
0.54345703125,
0.0545654296875,
0.55712890625,
0.0352783203125,
0.55712890625
],
[
0.496337890625,
0.54150390625,
0.76904296875,
0.54150390625,
0.76904296875,
0.5546875,
0.496337890625,
0.5546875
],
[
0.06634521484375,
0.5546875,
0.47021484375,
0.5546875,
0.47021484375,
0.568359375,
0.06634521484375,
0.568359375
],
[
0.5205078125,
0.55712890625,
0.89990234375,
0.55712890625,
0.89990234375,
0.57080078125,
0.5205078125,
0.57080078125
],
[
0.029388427734375,
0.56787109375,
0.056243896484375,
0.56787109375,
0.056243896484375,
0.58056640625,
0.029388427734375,
0.58056640625
],
[
0.031890869140625,
0.55859375,
0.042816162109375,
0.55859375,
0.042816162109375,
0.56640625,
0.031890869140625,
0.56640625
],
[
0.06634521484375,
0.5693359375,
0.470947265625,
0.57080078125,
0.470947265625,
0.5849609375,
0.06634521484375,
0.58447265625
],
[
0.495361328125,
0.57080078125,
0.89990234375,
0.57177734375,
0.89990234375,
0.5849609375,
0.495361328125,
0.58447265625
],
[
0.029388427734375,
0.5791015625,
0.0545654296875,
0.5791015625,
0.0545654296875,
0.5927734375,
0.029388427734375,
0.5927734375
],
[
0.06634521484375,
0.58447265625,
0.470947265625,
0.5849609375,
0.470947265625,
0.599609375,
0.06634521484375,
0.5986328125
],
[
0.029388427734375,
0.59130859375,
0.0545654296875,
0.59130859375,
0.0545654296875,
0.60498046875,
0.029388427734375,
0.60498046875
],
[
0.496337890625,
0.5859375,
0.89990234375,
0.5859375,
0.89990234375,
0.599609375,
0.496337890625,
0.599609375
],
[
0.0697021484375,
0.599609375,
0.470947265625,
0.599609375,
0.470947265625,
0.61328125,
0.0697021484375,
0.61328125
],
[
0.496337890625,
0.60107421875,
0.89990234375,
0.60107421875,
0.89990234375,
0.6142578125,
0.496337890625,
0.6142578125
],
[
0.03863525390625,
0.60693359375,
0.0545654296875,
0.6181640625,
0.039459228515625,
0.62744140625,
0.0235137939453125,
0.6162109375
],
[
0.0679931640625,
0.615234375,
0.47265625,
0.615234375,
0.47265625,
0.62890625,
0.0679931640625,
0.62890625
],
[
0.033599853515625,
0.62158203125,
0.0487060546875,
0.6259765625,
0.04449462890625,
0.63427734375,
0.029388427734375,
0.63134765625
],
[
0.497802734375,
0.61767578125,
0.8984375,
0.61767578125,
0.8984375,
0.62744140625,
0.497802734375,
0.62744140625
],
[
0.0679931640625,
0.6298828125,
0.47265625,
0.6298828125,
0.47265625,
0.6435546875,
0.0679931640625,
0.6435546875
],
[
0.496337890625,
0.63134765625,
0.89990234375,
0.63134765625,
0.89990234375,
0.64453125,
0.496337890625,
0.64453125
],
[
0.0679931640625,
0.64453125,
0.47265625,
0.64453125,
0.47265625,
0.65771484375,
0.0679931640625,
0.65771484375
],
[
0.029388427734375,
0.6328125,
0.056243896484375,
0.6328125,
0.056243896484375,
0.6455078125,
0.029388427734375,
0.6455078125
],
[
0.496337890625,
0.6455078125,
0.89990234375,
0.6455078125,
0.89990234375,
0.6591796875,
0.496337890625,
0.6591796875
],
[
0.0277099609375,
0.65771484375,
0.047027587890625,
0.65283203125,
0.056243896484375,
0.67041015625,
0.036956787109375,
0.67578125
],
[
0.06634521484375,
0.6591796875,
0.47265625,
0.6591796875,
0.47265625,
0.6728515625,
0.06634521484375,
0.6728515625
],
[
0.496337890625,
0.66015625,
0.8984375,
0.66015625,
0.8984375,
0.67333984375,
0.496337890625,
0.67333984375
],
[
0.06634521484375,
0.6748046875,
0.47265625,
0.6748046875,
0.47265625,
0.6884765625,
0.06634521484375,
0.6884765625
],
[
0.496337890625,
0.67578125,
0.89990234375,
0.67578125,
0.89990234375,
0.689453125,
0.496337890625,
0.689453125
],
[
0.0679931640625,
0.6884765625,
0.47021484375,
0.6884765625,
0.47021484375,
0.70166015625,
0.0679931640625,
0.70166015625
],
[
0.496337890625,
0.6904296875,
0.89990234375,
0.6904296875,
0.89990234375,
0.70361328125,
0.496337890625,
0.70361328125
],
[
0.06634521484375,
0.70361328125,
0.470947265625,
0.70361328125,
0.470947265625,
0.71728515625,
0.06634521484375,
0.71728515625
],
[
0.0352783203125,
0.69140625,
0.0411376953125,
0.69140625,
0.0411376953125,
0.69873046875,
0.0352783203125,
0.69873046875
],
[
0.03863525390625,
0.70703125,
0.05792236328125,
0.70703125,
0.05792236328125,
0.71630859375,
0.03863525390625,
0.71630859375
],
[
0.496337890625,
0.705078125,
0.89990234375,
0.705078125,
0.89990234375,
0.71875,
0.496337890625,
0.71875
],
[
0.0679931640625,
0.71875,
0.43994140625,
0.71875,
0.43994140625,
0.732421875,
0.0679931640625,
0.732421875
],
[
0.497802734375,
0.7197265625,
0.89990234375,
0.7197265625,
0.89990234375,
0.73291015625,
0.497802734375,
0.73291015625
],
[
0.496337890625,
0.73388671875,
0.89990234375,
0.73388671875,
0.89990234375,
0.74755859375,
0.496337890625,
0.74755859375
],
[
0.0697021484375,
0.74462890625,
0.438232421875,
0.74462890625,
0.438232421875,
0.75830078125,
0.0697021484375,
0.75830078125
],
[
0.496337890625,
0.7490234375,
0.89990234375,
0.7490234375,
0.89990234375,
0.7626953125,
0.496337890625,
0.7626953125
],
[
0.06634521484375,
0.75830078125,
0.44921875,
0.7587890625,
0.44921875,
0.77392578125,
0.06634521484375,
0.7724609375
],
[
0.495361328125,
0.76318359375,
0.8984375,
0.76318359375,
0.8984375,
0.77685546875,
0.495361328125,
0.77685546875
],
[
0.0697021484375,
0.77490234375,
0.298828125,
0.77490234375,
0.298828125,
0.7880859375,
0.0697021484375,
0.7880859375
],
[
0.495361328125,
0.77783203125,
0.8984375,
0.77783203125,
0.8984375,
0.79150390625,
0.495361328125,
0.79150390625
],
[
0.496337890625,
0.7919921875,
0.8984375,
0.7919921875,
0.8984375,
0.8056640625,
0.496337890625,
0.8056640625
],
[
0.0679931640625,
0.8017578125,
0.197265625,
0.80419921875,
0.197265625,
0.81787109375,
0.0679931640625,
0.8154296875
],
[
0.496337890625,
0.80810546875,
0.89990234375,
0.80810546875,
0.89990234375,
0.82177734375,
0.496337890625,
0.82177734375
],
[
0.0679931640625,
0.8154296875,
0.2225341796875,
0.81396484375,
0.2225341796875,
0.82763671875,
0.0679931640625,
0.8291015625
],
[
0.0697021484375,
0.8291015625,
0.2301025390625,
0.8291015625,
0.2301025390625,
0.8388671875,
0.0697021484375,
0.8388671875
],
[
0.496337890625,
0.82373046875,
0.86572265625,
0.82373046875,
0.86572265625,
0.83740234375,
0.496337890625,
0.83740234375
],
[
0.06634521484375,
0.8388671875,
0.204833984375,
0.8388671875,
0.204833984375,
0.8525390625,
0.06634521484375,
0.8525390625
],
[
0.5224609375,
0.8388671875,
0.8984375,
0.8388671875,
0.8984375,
0.8525390625,
0.5224609375,
0.8525390625
],
[
0.0679931640625,
0.85693359375,
0.133544921875,
0.85693359375,
0.133544921875,
0.8681640625,
0.0679931640625,
0.8681640625
],
[
0.496337890625,
0.85400390625,
0.89990234375,
0.85400390625,
0.89990234375,
0.86767578125,
0.496337890625,
0.86767578125
],
[
0.06634521484375,
0.8681640625,
0.35107421875,
0.8681640625,
0.35107421875,
0.8818359375,
0.06634521484375,
0.8818359375
],
[
0.49951171875,
0.87060546875,
0.896484375,
0.87060546875,
0.896484375,
0.88134765625,
0.49951171875,
0.88134765625
],
[
0.06634521484375,
0.88134765625,
0.219970703125,
0.88134765625,
0.219970703125,
0.89453125,
0.06634521484375,
0.89453125
],
[
0.496337890625,
0.88427734375,
0.89990234375,
0.88427734375,
0.89990234375,
0.89794921875,
0.496337890625,
0.89794921875
]
] |
[
"VLDBJ manuscript No.",
"(will be inserted by the editor)",
"Dissociation and Propagation for Approximate Lifted Inference with",
"Standard Relational Database Management Systems",
"Wolfgang Gatterbauer Dan Suciu",
"6",
"0",
"Abstract Probabilistic inference over large data sets is a",
"1Introduction",
"challenging data management problem since exact infer-",
"2",
"un",
"ence is generally #P-hard and is most often solved approxi-",
"Probabilistic inference over large data sets is becoming a",
"mately with sampling-based methods today. This paper pro-",
"central data management problem: Recent large knowl-",
"N",
"",
"poses an alternative approach for approximate evaluation of",
"edge bases, such as Yago [40], Nell [7], DeepDive [16],",
"conjunctive queries with standard relational databases: In",
"or Google's Knowledge Vault [19], have millions to bil-",
"lions of uncertain tuples. Data sets with missing values are",
"",
"our approach, every query is evaluated entirely in the data-",
"B",
"base engine by evaluating a fixed number of query plans,",
"often “completed\"using inference in graphical models [8,",
"D",
"each providing an upper bound on the true probability, then",
"6l,71lorsophisticated lowrankmatrixfactorizationtech-",
"taking their minimum. We provide an algorithm that takes",
"niques [21,70] that ultimately result in a large probabilistic",
"S",
"into account important schema information to enumerate",
"database. Data sets that result from crowdsourcing [1] or",
"only the minimal necessary plans among all possible plans.",
"C",
"that are inferred from unstructured information [9] are also",
"Importantly, this algorithm is a strict generalization of all",
"uncertain, and probabilistic databases have been applied to",
"known PTIME self-join-free conjunctive queries: A query",
"4",
"",
"bootstrapping over samples of data [80].",
"is in PTIME if and only if our algorithm returns one sin-",
"However, probabilistic inference is known to be #P-",
"5",
"贩",
"gle plan. Furthermore, our approach is a generalization of a",
"hard in the size of the database, even for some very sim-",
"2",
"family of efficient ranking methods from graphs to hyper-",
"6",
"ple queries [13]. Today's state of the art inference engines",
"graphs. We also adapt three relational query optimization",
"use either sampling-based methods or are based on some",
"0",
"techniques to evaluate all necessary plans very fast. We give",
"1",
"variantof theDPLLalgorithmforWeightedModelCount-",
"a detailed experimental evaluation of our approach and, in",
"ing [15]. For example, Tuffy [51], a popular implementation",
"the process, provide a new way of thinking about the value",
"3",
"of Markov Logic Networks (MLN) over relational data-",
"3",
" of probabilistic methods over non-probabilistic methods for",
"bases, uses Markov Chain Monte Carlo methods (MCMC).",
"ranking query answers. We also note that the techniques",
"Gibbs sampling can be significantly improved by adapting",
"developed in this paper apply immediately to lifted infer-",
"some classical relational optimization techniques [81]. For",
"ence from statistical relational models since lifted inference",
"",
"",
"another example, MayBMS [3] and its successor Sprout [55]",
"corresponds to PTIME plans in probabilistic databases.",
"use query plans to guide a DPLL-based algorithm for Weigh-",
"ted Model Counting [35]. While both approaches deploy",
"Keywords Probabilistic inference . Lifted inference .",
"some advanced relational optimization techniques, at their",
"Probabilistic databases -Problem relaxation Ranking ",
"core they are based on general purpose probabilistic in-",
"Queryplans·Queryoptimization",
"ference techniques, which either run in exponential time",
"(DPLL-based algorithms have been proven recently to take",
"Wolfgang Gatterbauer",
"exponential time even for queries computable in polynomial",
"Tepper School of Business",
"CarnegieMellon University",
"time [4), or require many iterations until convergence.",
"E-mail: gatt@cmu.edu",
"In this paper, we propose a different approach to query",
"Dan Suciu",
"evaluation with probabilistic databases (PDBs). In our ap-",
"Department of Computer Science & Engineering",
"proach,everyqueryisevaluatedentirelyinthedatabaseen-",
"University of Washington",
"gine. Probability computation is done at query time, using"
] |
{
"arxiv_id": "1310.6257",
"language": "en",
"page_idx": 0
}
|
||
1310.6257-en-3
|
1310.6257-en-3.jpg
|
Wolfgang Gatterbauer, Dan Suciu
all our optimizations enabled, computing hard queries over probabilistic databases incurs only a modest penalty over computing the same query on a deterministic database: For example, the 8-chain query (with 429 query plans) runs only a factor of<10 slower than on a deterministic database.
Prior publications. In recent work[30], we apply the idea of dissociation to both upper and lower bound the probability of Boolean functions, but discuss the connection to query evaluation only in passing. Parts of Section 3 and Section 4 are based on a workshop paper [28]. The remainder is based upon [31]. We added the connection to propagation on graphs, more detailed experiments, slightly changed the formalisms, and included extensive illustrating examples throughout. Due to space restrictions, some of our proofs had to be included in an online appendix on ArXiv [29].
4
On a query, a dissociation extends atoms with additional variables. We prove that a dissociation can only increase the probability of a query, and define the propagation score of a query as the minimum reliability of all dissociated queries that are“hierarchical”. This is justified by the fact that, in a $k$-partite graph, the propagation score is precisely the probability of one dissociated hierarchical query. Thus, in our definition, choosing a direction for the network in order to define the propagation score corresponds to choosing a particular dissociation that makes the query hierarchical.
(2) We show how the propagation score can be evaluated with the help of a query-dependent number of query plans (Section 4). We achieve this by establishing a one-toone correspondence between hierarchical dissociations and traditional query plans and showing that every query plan computes a probability that is an upper bound of query reliability. Moreover, we describe a natural partial order on the probabilities of query plans Thus every self-join-free conjunctive query can be approximated by a fixed number of query plans, and it suffices to iterate over all minimal plans, compute their probabilities, then take the minimum. We give an intuitive system R-style algorithm [67] that enumerates all minimal plans for a given query $q.$
(3) We generalize the algorithm to take into consideration schema knowledge on deterministic relations and functional dependencies (Section 5). In particular, we give a $uni-$ fred treatment and generalization of all previously known PTIME self-join free conjunctive queries, i.e. those that can be evaluated with a query plan in polynomial time in the size of the database, and show that our approach naturally generalizes all known PTIME queries: for every query that is PTIME (whether due to key constraints, or the presence of deterministic tables), reliability and propagation scores always coincide; for every query that is #P-hard, our approach still returns a unique, well-defined score in polynomial time.
(4) We give a set of targeted multi-query optimization techniques that considerably speed up the time needed to evaluate the propagation score (Section 6). Evaluating some queries may require a large number of plans (e.g., an 8- chain query requires 429 plans). Evaluating all plans sequentially would still be prohibitively expensive. Instead, we tailor three relational query optimization techniques to dissociation: $(i)$ combining all minimal plans into one single $query,(ii)$ reusing common subexpressions with views, and $(iii)$ performing deterministic semi-join reductions.
(5) We conduct a set of very extensive experiments in which we compare the quality of ranking and scalability of various alternative methods (probabilistic and not) against exact probabilistic inference on TPC-H data[72]. We devise a setup that measures the additional benefit of probabilistic inference for ranking over alternative methods, showing that our technique has high precision for ranking query answers based on their output probabilities. We also show that, with
## 2 Technical background
2.1 Probabilistic databases and self-join-free conjunctive
queries
A tuple-independent probabilistic database (TI-PDB) is a database $D$ plus a function $p(t)\in[0,1]$ associating an independent probability to each tuple $t\in D.$ We fix a relational vocabulary $\sigma=(R_1,\ldots,R_m)$ and denote with $D$ the database, i.e. the collection of tuples and their probabilities. A possible world is then a subset of $D$ generated by independently including each tuple $t$ in the world with probability $p(t).$ We use bold notation (e.g., x) to denote both sets or tuples. A self-join-free conjunctive query (sj-free CQ) is a fırst-order formula $q(\mathbf{z})=\exists x_1\ldots\exists x_k.(a_1\wedge\ldots\wedge a_m)$ where each atom $a_i$ represents a relation $R_i(\mathbf{x}_i)$,the variables $x_1,\ldots,x_k$ are called existential variables, and z are called the head variables (or free variables).$^2$
The term“self-join-free”means that the atoms refer to distinct relational symbols. We assume therefore w.l.o.g. that every relational symbol $R_1,\ldots,R_m$ occurs exactly once in the query. Unless otherwise stated, a“query”in this paper always denotes a sj-free CQ. As usual, we abbreviate a query by $q(\mathbf{z}):-a_1,\ldots,a_m$, and write HVar$(q)=\mathbf{z}$, EVar$(q)=\{x_1,\ldots,x_k\}$ and Var$(q)=$HVar$(q)\cup$EVar$(q)$ for the set of head variables, existential variables, and all variables of $q.$ If HVar$(q)=\emptyset$, then $q$ is called a Boolean query and EVar$(q)=$Var$(q).$ We also write Var$(a_i)$ for the variables in atom $a_i$ and at$(x_j)$ for the set of atoms that contain variable $x_j.$ The active domain of a variable $x_j$ is denoted $ADom_{x_j},^3$ and the active domain of the entire database is $ADom=\bigcup_jADom_{x_j}.$ The focus of probabilistic
${^{2}\text{W.l.o.g. we assume x}_{i}\text{ to be a tuple of only variables and don't}}$ write the constants. Selections can always be directly pushed into the database before executing the query.
$^{3}$ Defined formally as $ADom_{x_j}=\bigcup_{i:x_j\in\mathsf Var(R_i)}\pi_{x_j}(R_i).$
|
[
[
0.0679931640625,
0.0789794921875,
0.0814208984375,
0.0789794921875,
0.0814208984375,
0.08905029296875,
0.0679931640625,
0.08905029296875
],
[
0.7021484375,
0.07598876953125,
0.89990234375,
0.07659912109375,
0.89990234375,
0.0914306640625,
0.7021484375,
0.09027099609375
],
[
0.0697021484375,
0.10687255859375,
0.47265625,
0.10687255859375,
0.47265625,
0.12054443359375,
0.0697021484375,
0.12054443359375
],
[
0.496337890625,
0.10687255859375,
0.90185546875,
0.10687255859375,
0.90185546875,
0.12054443359375,
0.496337890625,
0.12054443359375
],
[
0.0679931640625,
0.122314453125,
0.47265625,
0.122314453125,
0.47265625,
0.135986328125,
0.0679931640625,
0.135986328125
],
[
0.495361328125,
0.122314453125,
0.90185546875,
0.122314453125,
0.90185546875,
0.135986328125,
0.495361328125,
0.135986328125
],
[
0.0679931640625,
0.13720703125,
0.474365234375,
0.13720703125,
0.474365234375,
0.15087890625,
0.0679931640625,
0.15087890625
],
[
0.495361328125,
0.1383056640625,
0.8984375,
0.13720703125,
0.8984375,
0.1484375,
0.495361328125,
0.149658203125
],
[
0.0679931640625,
0.1519775390625,
0.47265625,
0.1519775390625,
0.47265625,
0.1650390625,
0.0679931640625,
0.1650390625
],
[
0.495361328125,
0.1519775390625,
0.89990234375,
0.1519775390625,
0.89990234375,
0.1650390625,
0.495361328125,
0.1650390625
],
[
0.0679931640625,
0.166259765625,
0.47265625,
0.166259765625,
0.47265625,
0.179931640625,
0.0679931640625,
0.179931640625
],
[
0.49365234375,
0.1650390625,
0.8759765625,
0.1650390625,
0.8759765625,
0.1787109375,
0.49365234375,
0.1787109375
],
[
0.0679931640625,
0.1817626953125,
0.470947265625,
0.1817626953125,
0.470947265625,
0.1953125,
0.0679931640625,
0.1953125
],
[
0.5224609375,
0.1817626953125,
0.89990234375,
0.1817626953125,
0.89990234375,
0.1953125,
0.5224609375,
0.1953125
],
[
0.0697021484375,
0.1953125,
0.47265625,
0.1953125,
0.47265625,
0.208984375,
0.0697021484375,
0.208984375
],
[
0.495361328125,
0.196533203125,
0.89990234375,
0.196533203125,
0.89990234375,
0.210205078125,
0.495361328125,
0.210205078125
],
[
0.07135009765625,
0.2120361328125,
0.470947265625,
0.2120361328125,
0.470947265625,
0.22265625,
0.07135009765625,
0.22265625
],
[
0.497802734375,
0.213134765625,
0.8984375,
0.213134765625,
0.8984375,
0.223876953125,
0.497802734375,
0.223876953125
],
[
0.0679931640625,
0.2257080078125,
0.47265625,
0.2257080078125,
0.47265625,
0.2392578125,
0.0679931640625,
0.2392578125
],
[
0.495361328125,
0.226806640625,
0.89990234375,
0.2257080078125,
0.89990234375,
0.2392578125,
0.495361328125,
0.240478515625
],
[
0.0679931640625,
0.240478515625,
0.42822265625,
0.240478515625,
0.42822265625,
0.254150390625,
0.0679931640625,
0.254150390625
],
[
0.496337890625,
0.2410888671875,
0.89990234375,
0.2410888671875,
0.89990234375,
0.254638671875,
0.496337890625,
0.254638671875
],
[
0.09149169921875,
0.257080078125,
0.47265625,
0.257080078125,
0.47265625,
0.270751953125,
0.09149169921875,
0.270751953125
],
[
0.495361328125,
0.254638671875,
0.8984375,
0.255859375,
0.8984375,
0.270751953125,
0.495361328125,
0.26953125
],
[
0.0679931640625,
0.2724609375,
0.470947265625,
0.2724609375,
0.470947265625,
0.2861328125,
0.0679931640625,
0.2861328125
],
[
0.497802734375,
0.2724609375,
0.8984375,
0.2724609375,
0.8984375,
0.283203125,
0.497802734375,
0.283203125
],
[
0.06634521484375,
0.2861328125,
0.470947265625,
0.287353515625,
0.470947265625,
0.300537109375,
0.06634521484375,
0.2998046875
],
[
0.497802734375,
0.287353515625,
0.8984375,
0.287353515625,
0.8984375,
0.297607421875,
0.497802734375,
0.297607421875
],
[
0.0679931640625,
0.300537109375,
0.47265625,
0.300537109375,
0.47265625,
0.314208984375,
0.0679931640625,
0.314208984375
],
[
0.496337890625,
0.300537109375,
0.89990234375,
0.300537109375,
0.89990234375,
0.314208984375,
0.496337890625,
0.314208984375
],
[
0.0679931640625,
0.31640625,
0.47265625,
0.31640625,
0.47265625,
0.330078125,
0.0679931640625,
0.330078125
],
[
0.496337890625,
0.3154296875,
0.875,
0.3154296875,
0.875,
0.328857421875,
0.496337890625,
0.328857421875
],
[
0.0679931640625,
0.330078125,
0.47021484375,
0.330078125,
0.47021484375,
0.34326171875,
0.0679931640625,
0.34326171875
],
[
0.0679931640625,
0.344482421875,
0.470947265625,
0.345703125,
0.470947265625,
0.359375,
0.0679931640625,
0.358154296875
],
[
0.0697021484375,
0.3623046875,
0.47021484375,
0.3623046875,
0.47021484375,
0.372314453125,
0.0697021484375,
0.372314453125
],
[
0.496337890625,
0.361083984375,
0.67236328125,
0.361083984375,
0.67236328125,
0.374755859375,
0.496337890625,
0.374755859375
],
[
0.06634521484375,
0.3759765625,
0.474365234375,
0.3759765625,
0.474365234375,
0.3896484375,
0.06634521484375,
0.3896484375
],
[
0.0679931640625,
0.39013671875,
0.470947265625,
0.39013671875,
0.470947265625,
0.40380859375,
0.0679931640625,
0.40380859375
],
[
0.495361328125,
0.3896484375,
0.880859375,
0.3896484375,
0.880859375,
0.402587890625,
0.495361328125,
0.402587890625
],
[
0.0679931640625,
0.405029296875,
0.47265625,
0.405029296875,
0.47265625,
0.418701171875,
0.0679931640625,
0.418701171875
],
[
0.496337890625,
0.40673828125,
0.5498046875,
0.40673828125,
0.5498046875,
0.418701171875,
0.496337890625,
0.418701171875
],
[
0.0697021484375,
0.421630859375,
0.470947265625,
0.421630859375,
0.470947265625,
0.431640625,
0.0697021484375,
0.431640625
],
[
0.0679931640625,
0.435302734375,
0.322509765625,
0.435302734375,
0.322509765625,
0.448974609375,
0.0679931640625,
0.448974609375
],
[
0.496337890625,
0.43408203125,
0.89990234375,
0.43408203125,
0.89990234375,
0.44775390625,
0.496337890625,
0.44775390625
],
[
0.0948486328125,
0.45068359375,
0.470947265625,
0.45068359375,
0.470947265625,
0.46435546875,
0.0948486328125,
0.46435546875
],
[
0.497802734375,
0.45068359375,
0.8984375,
0.45068359375,
0.8984375,
0.46142578125,
0.497802734375,
0.46142578125
],
[
0.06634521484375,
0.46435546875,
0.47265625,
0.465576171875,
0.47265625,
0.478515625,
0.06634521484375,
0.47802734375
],
[
0.49365234375,
0.463134765625,
0.89990234375,
0.4619140625,
0.89990234375,
0.476806640625,
0.49365234375,
0.47802734375
],
[
0.0697021484375,
0.482177734375,
0.470947265625,
0.482177734375,
0.470947265625,
0.4921875,
0.0697021484375,
0.4921875
],
[
0.496337890625,
0.478515625,
0.89990234375,
0.478515625,
0.89990234375,
0.4921875,
0.496337890625,
0.4921875
],
[
0.0679931640625,
0.495361328125,
0.47265625,
0.495361328125,
0.47265625,
0.5087890625,
0.0679931640625,
0.5087890625
],
[
0.496337890625,
0.493408203125,
0.89990234375,
0.493408203125,
0.89990234375,
0.50732421875,
0.496337890625,
0.50732421875
],
[
0.0679931640625,
0.51025390625,
0.470947265625,
0.51025390625,
0.470947265625,
0.52392578125,
0.0679931640625,
0.52392578125
],
[
0.496337890625,
0.50830078125,
0.89990234375,
0.50830078125,
0.89990234375,
0.521484375,
0.496337890625,
0.521484375
],
[
0.0679931640625,
0.52490234375,
0.47265625,
0.52490234375,
0.47265625,
0.5380859375,
0.0679931640625,
0.5380859375
],
[
0.496337890625,
0.5224609375,
0.89990234375,
0.5224609375,
0.89990234375,
0.5361328125,
0.496337890625,
0.5361328125
],
[
0.0679931640625,
0.54052734375,
0.470947265625,
0.54052734375,
0.470947265625,
0.55419921875,
0.0679931640625,
0.55419921875
],
[
0.495361328125,
0.53759765625,
0.89990234375,
0.53515625,
0.89990234375,
0.5498046875,
0.495361328125,
0.5517578125
],
[
0.0679931640625,
0.5546875,
0.47265625,
0.5546875,
0.47265625,
0.568359375,
0.0679931640625,
0.568359375
],
[
0.497802734375,
0.55419921875,
0.8984375,
0.55419921875,
0.8984375,
0.56396484375,
0.497802734375,
0.56396484375
],
[
0.0679931640625,
0.5693359375,
0.474365234375,
0.5693359375,
0.474365234375,
0.5830078125,
0.0679931640625,
0.5830078125
],
[
0.495361328125,
0.56640625,
0.896484375,
0.56640625,
0.896484375,
0.580078125,
0.495361328125,
0.580078125
],
[
0.0679931640625,
0.58447265625,
0.47265625,
0.58447265625,
0.47265625,
0.59716796875,
0.0679931640625,
0.59716796875
],
[
0.496337890625,
0.58203125,
0.89990234375,
0.58203125,
0.89990234375,
0.595703125,
0.496337890625,
0.595703125
],
[
0.0679931640625,
0.599609375,
0.47265625,
0.599609375,
0.47265625,
0.61328125,
0.0679931640625,
0.61328125
],
[
0.49365234375,
0.5966796875,
0.90185546875,
0.59423828125,
0.90185546875,
0.609375,
0.49365234375,
0.61083984375
],
[
0.0679931640625,
0.61328125,
0.47265625,
0.61328125,
0.47265625,
0.626953125,
0.0679931640625,
0.626953125
],
[
0.496337890625,
0.61083984375,
0.73291015625,
0.61083984375,
0.73291015625,
0.62451171875,
0.496337890625,
0.62451171875
],
[
0.09320068359375,
0.6298828125,
0.47265625,
0.6298828125,
0.47265625,
0.6435546875,
0.09320068359375,
0.6435546875
],
[
0.5224609375,
0.626953125,
0.89990234375,
0.626953125,
0.89990234375,
0.64013671875,
0.5224609375,
0.64013671875
],
[
0.0697021484375,
0.646484375,
0.470947265625,
0.646484375,
0.470947265625,
0.65673828125,
0.0697021484375,
0.65673828125
],
[
0.496337890625,
0.642578125,
0.8984375,
0.642578125,
0.8984375,
0.65625,
0.496337890625,
0.65625
],
[
0.0679931640625,
0.66015625,
0.470947265625,
0.66015625,
0.470947265625,
0.67333984375,
0.0679931640625,
0.67333984375
],
[
0.496337890625,
0.65673828125,
0.89990234375,
0.65673828125,
0.89990234375,
0.67041015625,
0.496337890625,
0.67041015625
],
[
0.0679931640625,
0.6748046875,
0.470947265625,
0.6748046875,
0.470947265625,
0.6884765625,
0.0679931640625,
0.6884765625
],
[
0.495361328125,
0.67138671875,
0.89990234375,
0.67138671875,
0.89990234375,
0.68505859375,
0.495361328125,
0.68505859375
],
[
0.0679931640625,
0.689453125,
0.470947265625,
0.689453125,
0.470947265625,
0.703125,
0.0679931640625,
0.703125
],
[
0.496337890625,
0.68603515625,
0.8984375,
0.68603515625,
0.8984375,
0.69970703125,
0.496337890625,
0.69970703125
],
[
0.0697021484375,
0.7060546875,
0.47021484375,
0.7060546875,
0.47021484375,
0.71630859375,
0.0697021484375,
0.71630859375
],
[
0.495361328125,
0.70068359375,
0.89990234375,
0.69970703125,
0.89990234375,
0.71435546875,
0.495361328125,
0.71533203125
],
[
0.0679931640625,
0.7197265625,
0.47265625,
0.7197265625,
0.47265625,
0.73291015625,
0.0679931640625,
0.73291015625
],
[
0.496337890625,
0.71533203125,
0.89990234375,
0.71533203125,
0.89990234375,
0.72998046875,
0.496337890625,
0.72998046875
],
[
0.0679931640625,
0.73388671875,
0.470947265625,
0.73388671875,
0.470947265625,
0.74755859375,
0.0679931640625,
0.74755859375
],
[
0.496337890625,
0.72998046875,
0.89990234375,
0.72998046875,
0.89990234375,
0.74365234375,
0.496337890625,
0.74365234375
],
[
0.0697021484375,
0.74755859375,
0.470947265625,
0.74755859375,
0.470947265625,
0.76123046875,
0.0697021484375,
0.76123046875
],
[
0.496337890625,
0.7451171875,
0.89990234375,
0.7451171875,
0.89990234375,
0.7587890625,
0.496337890625,
0.7587890625
],
[
0.0697021484375,
0.76318359375,
0.414794921875,
0.76318359375,
0.414794921875,
0.77685546875,
0.0697021484375,
0.77685546875
],
[
0.497802734375,
0.76025390625,
0.89990234375,
0.76025390625,
0.89990234375,
0.77392578125,
0.497802734375,
0.77392578125
],
[
0.0948486328125,
0.77978515625,
0.470947265625,
0.77978515625,
0.470947265625,
0.79345703125,
0.0948486328125,
0.79345703125
],
[
0.496337890625,
0.77490234375,
0.89990234375,
0.77490234375,
0.89990234375,
0.7880859375,
0.496337890625,
0.7880859375
],
[
0.0697021484375,
0.79443359375,
0.474365234375,
0.79443359375,
0.474365234375,
0.80810546875,
0.0697021484375,
0.80810546875
],
[
0.497802734375,
0.79150390625,
0.8984375,
0.79150390625,
0.8984375,
0.8017578125,
0.497802734375,
0.8017578125
],
[
0.495361328125,
0.8046875,
0.90185546875,
0.8046875,
0.90185546875,
0.818359375,
0.495361328125,
0.818359375
],
[
0.0697021484375,
0.81103515625,
0.470947265625,
0.81103515625,
0.470947265625,
0.82177734375,
0.0697021484375,
0.82177734375
],
[
0.06463623046875,
0.82470703125,
0.470947265625,
0.82373046875,
0.470947265625,
0.83740234375,
0.06463623046875,
0.83837890625
],
[
0.496337890625,
0.8193359375,
0.89990234375,
0.8193359375,
0.89990234375,
0.83447265625,
0.496337890625,
0.83447265625
],
[
0.0679931640625,
0.8388671875,
0.470947265625,
0.8388671875,
0.470947265625,
0.8525390625,
0.0679931640625,
0.8525390625
],
[
0.50390625,
0.84423828125,
0.90185546875,
0.845703125,
0.90185546875,
0.85986328125,
0.50390625,
0.859375
],
[
0.06634521484375,
0.8525390625,
0.47265625,
0.85400390625,
0.47265625,
0.86767578125,
0.06634521484375,
0.8662109375
],
[
0.496337890625,
0.859375,
0.90185546875,
0.859375,
0.90185546875,
0.87255859375,
0.496337890625,
0.87255859375
],
[
0.0697021484375,
0.8681640625,
0.47265625,
0.8681640625,
0.47265625,
0.8818359375,
0.0697021484375,
0.8818359375
],
[
0.496337890625,
0.869140625,
0.70947265625,
0.87158203125,
0.70947265625,
0.88525390625,
0.496337890625,
0.8828125
],
[
0.06634521484375,
0.88427734375,
0.470947265625,
0.88427734375,
0.470947265625,
0.89794921875,
0.06634521484375,
0.89794921875
],
[
0.50390625,
0.88134765625,
0.82275390625,
0.8828125,
0.82275390625,
0.90087890625,
0.50390625,
0.8984375
]
] |
[
"4",
"Wolfgang Gatterbauer, Dan Suciu",
"On a query, a dissociation extends atoms with additional",
"all our optimizations enabled, computing hard queries over",
"variables. We prove that a dissociation can only increase the",
"probabilistic databases incurs only a modest penalty over",
"probability of a query, and define the propagation score of",
"computing the same query on a deterministic database:For",
"a query as the minimum reliability of all dissociated queries",
"example, the 8-chain query (with 429 query plans) runs only",
"that are \"hierarchical\" This is justified by the fact that, in a",
"a factor of < 10 slower than on a deterministic database.",
"k-partite graph, the propagation score is precisely the prob-",
"Prior publications. In recent work [30], we apply the",
"ability of one dissociated hierarchical query. Thus, in our",
"idea of dissociation to both upper and lower bound the prob-",
"definition, choosing a direction for the network in order to",
"ability of Boolean functions, but discuss the connection to",
"define the propagation score corresponds to choosing a par-",
"query evaluation only in passing. Parts of Section 3 and Sec-",
"ticular dissociation that makes the query hierarchical.",
"tion 4 are based on a workshop paper [28]. The remainder",
"(2) We show how the propagation score can be evalu-",
"is based upon [31]. We added the connection to propaga-",
"ated with the help of a query-dependent number of query",
"tion on graphs,more detailed experiments, slightly changed",
"plans (Section 4). We achieve this by establishing a one-to-",
"theformalisms,andincludedextensiveillustratingexamples",
"one correspondence between hierarchical dissociations and",
"throughout. Due to space restrictions, some of our proofs",
"traditional query plans and showing that every query plan",
"had to be included in an online appendix on ArXiv [29].",
"computes a probability that is an upper bound of query reli-",
"ability. Moreover, we describe a natural partial order on the",
"probabilitiesof queryplans Thus everyself-join-free con-",
"2 Technical background",
"junctive query can be approximated by a fixed number of",
"query plans, and it suffices to iterate over all minimal plans,",
"2.1 Probabilistic databases and self-join-free conjunctive",
"compute their probabilities, then take the minimum. We give",
"queries",
"an intuitive system R-style algorithm[67lthat enumerates",
"all minimal plans for a given query q.",
"A tuple-independent probabilistic database (TI-PDB) is a",
"(3) We generalize the algorithm to take into considera-",
"database D plus a function p(t) EO,1 associating an in-",
"tion schema knowledge on deterministic relations and func-",
"dependent probability to each tuple t e D. We fix a rela-",
"tional dependencies (Section 5).In particular, we give a uni-",
"tional vocabulary = (R1,...,Rm) and denote with D the",
"fied treatment and generalization of all previously known",
"database, i.e. the collection of tuples and their probabili-",
"PTIME self-join free conjunctive queries, i.e. those that can",
"ties. A possible world is then a subset of D generated by",
"be evaluated with a query plan in polynomial time in the size",
"independently including each tuple t in the world with prob-",
"of the database, and show that our approach naturally gen-",
"ability p(t). We use bold notation (e.g., x) to denote both",
"eralizes all known PTIME queries: for every query that is",
"sets or tuples. A self-join-free conjunctive query (sj-free",
"PTIME (whether due to key constraints, or the presence of",
"CQ) is a first-order formula g(z) = Ex1 ...Exk.(a1 A...Aam)",
"deterministic tables), reliability and propagation scores al-",
"where each atom a; represents a relation R;(x), the variables",
"ways coincide; for every query that is #P-hard, our approach",
"xi,..,Xk are called existential variables, and z are called the",
"still returns a unique, well-defined score in polynomial time.",
"head variables (or free variables).?",
"(4) We give a set of targeted multi-query optimization",
"The term “self-join-free\" means that the atoms refer to",
"technigues that considerably speed up the time needed to",
"distinct relational symbols. We assume therefore w.l.o.g.",
"evaluate the propagation score (Section 6). Evaluating some",
"that every relational symbol Ri,..., Rm occurs exactly once",
"queries may require a large number of plans (e.g., an 8-",
"in the query. Unless otherwise stated, a \"query\" in this",
"chain query requires 429 plans). Evaluating all plans se-",
"paper always denotes a sj-free CQ. As usual, we abbrevi-",
"quentiallywould still be prohibitively expensive.Instead,",
"ate a query by q(z):-a1,...,am, and write HVar(q) = z,",
"we tailor three relational query optimization techniques to",
"EVar(q) = [x1,...,X) and Var(q) = HVar(q) UEVar(q)",
"dissociation: (i) combining all minimal plans into one single",
"for the set of head variables, existential variables, and all",
"query, (ii) reusing common subexpressions with views, and",
"variables of q. If HVar(q) = O, then q is called a Boolean",
"(ii) performing deterministic semi-join reductions.",
"query and EVar(q) =- Var(q). We also write Var(ai) for",
"(5) We conduct a set of very extensive experiments in",
"the variables in atom a; and at(x;) for the set of atoms",
"which we compare the quality of ranking and scalability of",
"that contain variable x The active domain of a variable",
"x, is denoted ADomx,3 and the active domain of the entire",
"various alternative methods (probabilistic and not) against",
"exact probabilistic inference on TPC-H data [72]. We devise",
"database is ADom = U,ADomxj. The focus of probabilistic",
"a setup that measures the additional benefit of probabilistic",
"2 wl.o.g.we assume x, to be a tuple of only variables and don't",
"inference for ranking over alternative methods, showing that",
"write the constants. Selections can always be directly pushed into the",
"our technique has high precision for ranking query answers",
"database before executing the query.",
"based on their output probabilities. We also show that, with",
"3 Defined formally as ADomx, = Uixjevar(R) x,(R)."
] |
{
"arxiv_id": "1310.6257",
"language": "en",
"page_idx": 3
}
|
||
1310.6257-en-4
|
1310.6257-en-4.jpg
|
5
Dissociation and Propagation for Approximate Lifted Inference with Standard Relational Database Management Systems
a relational atom containing the variables $\mathbf{x},\pi_{\mathbf{x}}^p$ is the $proba-$ bilistic project operator with duplicate elimination (or short “projection"), and $\bowtie^p\begin{bmatrix}\ldots\end{bmatrix}$ is the probabilistic natural join (or short“join”) in prefix notation, which we allow to be $k-$ ary $(k\geq2).$ We require that joins and projections alternate in a plan and do not distinguish between join orders.
口
query evaluation is to compute $\mathbb{P}[q]$, i.e. the probability that the query is true in a randomly chosen world. We will refer to this probability as the “query reliability”r(q)[37].
It is known that the data complexity [77] of any query $q$ is either in PTIME or #P-hard [14]. The PTIME queries are also called“safe queries” and, for the case of sj-free CQs, are characterized precisely by a syntactic property called $hi-$ erarchical queries [13]. We briefly review these results:
Definition 3 (Hierarchical query) A query $q$ is called hierarchical iff for any two existential variables $x, y\in$EVar$( q) $, one of the following three conditions holds: at$(x)\subseteq$at$(y)$, at$(x)\supseteq$at$(y)$,or at$(x)\cap$at$(y)=\emptyset.$
For example, the query $q_1:-R(x,y),S(y,z),T(y,z,u)$ is hierarchical, while $q_2:-R(x,y),S(y,z),T(z,u)$ is not as neither of the three conditions holds for the variables y and z.
Theorem 4 (Hierarchy dichotomy [13]) Ifq is hierarchi- $cal,then\mathbb{P}[q]$ can be computed in PTIME in the size of D. Otherwise, computing $\mathbb{P}[q]$ is #P-hard in the size of D.
We next give an equivalent, recursive characterization of hierarchical queries, for which we need a few definitions. We write SepVar$(q)$ for the set of existential variables that appear in every atom (called“separator variables”). A connected component of $q$ (or short,“query component") is a subset of atoms that are connected via existential variables. A query $q$ is disconnected if its atoms can be partitioned into two non-empty sets that do not share any existential variables (e.g., $q:-R(x,y),S(z,u),T(u,\nu)$ is disconnected and has two query components:“R(x,y)” and “S(z,u),T(u,v)”). For every set of variables $\mathbf{x}$, denote $q-\mathbf{x}$ the query obtained by removing all variables x and decreasing the arities of the relational symbols that contain variables from x. Any query can become disconnected by removing a set of variables.
We write Var$(P)$ for all variables in a plan $P$ and HVar$(P)$ for its head variables, which are recursively defined as follows:(1) if $P=R_i(\mathbf{x})$, then HVar$(P)=\mathbf{x};(2)$ if $P=$ $\pi_{\mathbf{x}}^{p}(P^{\prime})$, then HVar$=\mathbf{x};$ and (3) if $P=\mathbf{\alpha}^{p}\left[P_{1},\ldots,P_{k}\right]$, then HVar$(P)=\bigcup_i=1^k$HVar$(P_i).$The existential variables EVar$(P)$ are then defined as Var$(P)-$HVar$(P).$
Every plan $P$ represents a query $q_P$ defined by taking all atoms mentioned in $P$ as the body and setting HVar$(q_P)=$ HVar$(P).$ A plan is called Boolean if HVar$(P)=\emptyset.$ We assume the usual sanity conditions on plans to be satisfied: for a projection $\pi_\mathbf{x}^pP$ we assume $\mathbf{x}\subseteq$HVar$(P)$, and each variable y is projected away at most once in a plan, i.e. there exists at most one operator $\pi_\mathbf{x}^pP$ s.t. $y\in$HVar$(P)-\mathbf{x}.$ For notational convenience, we also use the“project-away operator” $\pi_\mathrm{-y}^{p}P$ instead of $\pi_{\mathbf{x}}^pP$, where y are the variables being projected away, i.e. $\mathbf{x}=$HVar$(\pi_{-\mathbf{y}}^pP)=$HVar$(P)-\mathbf{y}.$
Each subplan $P$ returns an intermediate relation of arity |HVar$(P)|+1.$ The extra probability attribute stores a $score(t)$ for each output tuple $t\in P(D).$ Given a probabilistic database $D$ and a plan $P, \textit{score}( t)$ is defined inductively on the structure of $P$ as follows:(1) If $t\in R_i(\mathbf{x})$, then $score(t)=$ $\mathbb{P}[t]$,i.e. its probability in $D;(2)$ if $t\in\mathfrak{a}^p\left[P_1(D),\ldots,P_k(D)\right]$ where $t=\mathfrak{A}^{p}\left[t_{1},\ldots,t_{k}\right]$,then score$(t)=\prod_i=1^{k}score(t_{i});$and (3) if $t\in\pi_{\mathbf{x}}^pP(D)$, and $t_1,\ldots,t_n\in P(D)$ are all the tuples that project onto $t$, then score$(t)=\bigotimes_i=1^nscore(t_i)$, where “$\otimes”$ stands for the independent-or. In other words, score computes a probability by assuming that all tuples joined by $\mathfrak{\alpha}^p$ and all duplicates eliminated by $\pi^p$ are independent. Only if these conditions hold, then score is the correct query probability (also called “query reliability”[37]), but in general it is not. Therefore, score is also called an extensional semantics [26,57,62] and is, in general, not equal to the query probability, which is defined in terms of possible worlds: $score(P)\neq\mathbb{P}[q_p].^{5}$For a Boolean plan $P$,we get one single score, which we denote score(P).
The requirement that joins and projections alternate is w.l.o.g. because nested joins, such as $\alpha^p\left[\:\alpha^{p}\left[R_{1},R_{2}\right],R_{3}\right]$ or $\rtimes^{p}\left[R_{1},\rtimes^{p}\left[R_{2},R_{3}\right]\right]$,can be rewritten into $\rtimes^{\tilde{p}}\left[R_{1},R_{2},R_{3}\right]$ while keeping the same score, e.g., $(p_1p_2)p_3=p_1(p_2p_3).$ For the same reason we do not distinguish between different permutations in the joins, called join orders [50]. We do not focus on query optimization in this paper until Section 6.
$5Extensional$ approaches compute the probability of any formula
## Lemma 5 (Hierarchical queries) A query q is “hierarchi- $cal"iffeither:(I)q$ has a single atom; (2) q has k ≥ query components all of which are hierarchical; or (3) q has a separator variable x, and q-{ }x{ }is hierarchical.
Every hierarchical query can be computed in PTIME, but
non-hierarchical queries are #P-hard, in general.$^{\color{red}4}$
# 2.2 Probabilistic query plans
Unless otherwise stated, a“query plan”in this paper always
denotes a probabilistic query plan.
Definition 6(Query plans) A query plan P is given by the grammar $P\because=R_i(\mathbf{x})\mid\pi_{\mathbf{x}}^pP\mid\rtimes^p\left[P_1,\ldots,P_k\right]$ where $R_i(\mathbf{x})$ is
as a function of the probabilities of its subformulas according to syntactic rules, regardless of how those were derived. Intensional approaches reason in terms of possible worlds and keep track of dependencies [57].
$^{4}$ Non-hierarchical queries can be in PTIME when considering func-
tional dependencies or deterministic tables [13,54] (see Section 5).
|
[
[
0.0697021484375,
0.07781982421875,
0.76318359375,
0.07781982421875,
0.76318359375,
0.0885009765625,
0.0697021484375,
0.0885009765625
],
[
0.890625,
0.08197021484375,
0.896484375,
0.08197021484375,
0.896484375,
0.0872802734375,
0.890625,
0.0872802734375
],
[
0.0697021484375,
0.10687255859375,
0.47265625,
0.10687255859375,
0.47265625,
0.12054443359375,
0.0697021484375,
0.12054443359375
],
[
0.495361328125,
0.10687255859375,
0.8984375,
0.10687255859375,
0.8984375,
0.12054443359375,
0.495361328125,
0.12054443359375
],
[
0.0679931640625,
0.1217041015625,
0.474365234375,
0.12054443359375,
0.474365234375,
0.1353759765625,
0.0679931640625,
0.135986328125
],
[
0.495361328125,
0.122314453125,
0.89990234375,
0.122314453125,
0.89990234375,
0.135986328125,
0.495361328125,
0.135986328125
],
[
0.0697021484375,
0.13720703125,
0.43408203125,
0.13720703125,
0.43408203125,
0.15087890625,
0.0697021484375,
0.15087890625
],
[
0.496337890625,
0.135986328125,
0.8984375,
0.135986328125,
0.8984375,
0.149658203125,
0.496337890625,
0.149658203125
],
[
0.09149169921875,
0.15087890625,
0.47021484375,
0.15087890625,
0.47021484375,
0.1644287109375,
0.09149169921875,
0.1644287109375
],
[
0.496337890625,
0.1519775390625,
0.8984375,
0.1519775390625,
0.8984375,
0.1650390625,
0.496337890625,
0.1650390625
],
[
0.0679931640625,
0.166259765625,
0.470947265625,
0.166259765625,
0.470947265625,
0.179931640625,
0.0679931640625,
0.179931640625
],
[
0.497802734375,
0.166259765625,
0.89990234375,
0.166259765625,
0.89990234375,
0.179931640625,
0.497802734375,
0.179931640625
],
[
0.0697021484375,
0.1817626953125,
0.470947265625,
0.1817626953125,
0.470947265625,
0.1953125,
0.0697021484375,
0.1953125
],
[
0.496337890625,
0.18115234375,
0.85546875,
0.18115234375,
0.85546875,
0.19482421875,
0.496337890625,
0.19482421875
],
[
0.892578125,
0.1817626953125,
0.8984375,
0.188232421875,
0.892578125,
0.191162109375,
0.88671875,
0.185302734375
],
[
0.0679931640625,
0.196533203125,
0.470947265625,
0.196533203125,
0.470947265625,
0.210205078125,
0.0679931640625,
0.210205078125
],
[
0.0679931640625,
0.21142578125,
0.444091796875,
0.21142578125,
0.444091796875,
0.2244873046875,
0.0679931640625,
0.2244873046875
],
[
0.52294921875,
0.210205078125,
0.90283203125,
0.210205078125,
0.90283203125,
0.223876953125,
0.52294921875,
0.223876953125
],
[
0.496337890625,
0.2244873046875,
0.8984375,
0.2244873046875,
0.8984375,
0.2381591796875,
0.496337890625,
0.2381591796875
],
[
0.06634521484375,
0.235107421875,
0.47021484375,
0.236328125,
0.47021484375,
0.25048828125,
0.06634521484375,
0.2493896484375
],
[
0.497802734375,
0.2392578125,
0.8984375,
0.2392578125,
0.8984375,
0.254150390625,
0.497802734375,
0.254150390625
],
[
0.0679931640625,
0.251708984375,
0.470947265625,
0.251708984375,
0.470947265625,
0.265380859375,
0.0679931640625,
0.265380859375
],
[
0.496337890625,
0.254638671875,
0.89990234375,
0.254638671875,
0.89990234375,
0.268310546875,
0.496337890625,
0.268310546875
],
[
0.0679931640625,
0.26416015625,
0.470947265625,
0.2666015625,
0.470947265625,
0.28076171875,
0.0679931640625,
0.279052734375
],
[
0.496337890625,
0.270751953125,
0.90283203125,
0.270751953125,
0.90283203125,
0.283935546875,
0.496337890625,
0.283935546875
],
[
0.0679931640625,
0.281982421875,
0.31982421875,
0.281982421875,
0.31982421875,
0.295654296875,
0.0679931640625,
0.295654296875
],
[
0.496337890625,
0.28515625,
0.75732421875,
0.28515625,
0.75732421875,
0.298583984375,
0.496337890625,
0.298583984375
],
[
0.5224609375,
0.300537109375,
0.89990234375,
0.300537109375,
0.89990234375,
0.3154296875,
0.5224609375,
0.3154296875
],
[
0.0679931640625,
0.30712890625,
0.47021484375,
0.30712890625,
0.47021484375,
0.320556640625,
0.0679931640625,
0.320556640625
],
[
0.496337890625,
0.3154296875,
0.8984375,
0.31640625,
0.8984375,
0.330078125,
0.496337890625,
0.328857421875
],
[
0.0679931640625,
0.32177734375,
0.470947265625,
0.32177734375,
0.470947265625,
0.336181640625,
0.0679931640625,
0.336181640625
],
[
0.0697021484375,
0.336181640625,
0.4306640625,
0.336181640625,
0.4306640625,
0.349853515625,
0.0697021484375,
0.349853515625
],
[
0.496337890625,
0.330810546875,
0.8984375,
0.330810546875,
0.8984375,
0.344482421875,
0.496337890625,
0.344482421875
],
[
0.497802734375,
0.345703125,
0.89990234375,
0.345703125,
0.89990234375,
0.359375,
0.497802734375,
0.359375
],
[
0.0679931640625,
0.35986328125,
0.47021484375,
0.361083984375,
0.47021484375,
0.3759765625,
0.0679931640625,
0.374755859375
],
[
0.496337890625,
0.35986328125,
0.89990234375,
0.35986328125,
0.89990234375,
0.37353515625,
0.496337890625,
0.37353515625
],
[
0.0697021484375,
0.377197265625,
0.470947265625,
0.377197265625,
0.470947265625,
0.39013671875,
0.0697021484375,
0.39013671875
],
[
0.495361328125,
0.3759765625,
0.89990234375,
0.3759765625,
0.89990234375,
0.3896484375,
0.495361328125,
0.3896484375
],
[
0.0697021484375,
0.392578125,
0.432373046875,
0.392578125,
0.432373046875,
0.40625,
0.0697021484375,
0.40625
],
[
0.497802734375,
0.391357421875,
0.8984375,
0.391357421875,
0.8984375,
0.402099609375,
0.497802734375,
0.402099609375
],
[
0.495361328125,
0.405029296875,
0.8984375,
0.405029296875,
0.8984375,
0.418701171875,
0.495361328125,
0.418701171875
],
[
0.09320068359375,
0.41748046875,
0.470947265625,
0.41748046875,
0.470947265625,
0.43115234375,
0.09320068359375,
0.43115234375
],
[
0.495361328125,
0.419189453125,
0.89990234375,
0.419189453125,
0.89990234375,
0.432861328125,
0.495361328125,
0.432861328125
],
[
0.0697021484375,
0.432861328125,
0.468505859375,
0.432861328125,
0.468505859375,
0.446533203125,
0.0697021484375,
0.446533203125
],
[
0.496337890625,
0.43408203125,
0.7841796875,
0.43408203125,
0.7841796875,
0.44775390625,
0.496337890625,
0.44775390625
],
[
0.0697021484375,
0.44775390625,
0.47265625,
0.44775390625,
0.47265625,
0.46142578125,
0.0697021484375,
0.46142578125
],
[
0.52294921875,
0.449462890625,
0.89990234375,
0.449462890625,
0.89990234375,
0.463134765625,
0.52294921875,
0.463134765625
],
[
0.0697021484375,
0.4619140625,
0.47021484375,
0.46142578125,
0.47021484375,
0.474365234375,
0.0697021484375,
0.4755859375
],
[
0.497802734375,
0.466064453125,
0.8984375,
0.466064453125,
0.8984375,
0.476806640625,
0.497802734375,
0.476806640625
],
[
0.0697021484375,
0.476806640625,
0.470947265625,
0.476806640625,
0.470947265625,
0.490478515625,
0.0697021484375,
0.490478515625
],
[
0.497802734375,
0.48095703125,
0.8984375,
0.48095703125,
0.8984375,
0.4912109375,
0.497802734375,
0.4912109375
],
[
0.0697021484375,
0.4912109375,
0.470947265625,
0.4912109375,
0.470947265625,
0.5048828125,
0.0697021484375,
0.5048828125
],
[
0.497802734375,
0.493408203125,
0.89990234375,
0.493408203125,
0.89990234375,
0.50732421875,
0.497802734375,
0.50732421875
],
[
0.07135009765625,
0.50830078125,
0.47021484375,
0.50830078125,
0.47021484375,
0.5185546875,
0.07135009765625,
0.5185546875
],
[
0.497802734375,
0.5087890625,
0.8984375,
0.5087890625,
0.8984375,
0.5224609375,
0.497802734375,
0.5224609375
],
[
0.0697021484375,
0.521484375,
0.470947265625,
0.521484375,
0.470947265625,
0.53515625,
0.0697021484375,
0.53515625
],
[
0.496337890625,
0.52392578125,
0.896484375,
0.52392578125,
0.896484375,
0.53759765625,
0.496337890625,
0.53759765625
],
[
0.0697021484375,
0.5361328125,
0.470947265625,
0.5361328125,
0.470947265625,
0.5498046875,
0.0697021484375,
0.5498046875
],
[
0.496337890625,
0.5380859375,
0.89990234375,
0.5390625,
0.89990234375,
0.55419921875,
0.496337890625,
0.552734375
],
[
0.0697021484375,
0.55029296875,
0.47021484375,
0.55029296875,
0.47021484375,
0.5654296875,
0.0697021484375,
0.5654296875
],
[
0.497802734375,
0.5546875,
0.89990234375,
0.5546875,
0.89990234375,
0.568359375,
0.497802734375,
0.568359375
],
[
0.0697021484375,
0.56787109375,
0.47021484375,
0.56787109375,
0.47021484375,
0.57763671875,
0.0697021484375,
0.57763671875
],
[
0.496337890625,
0.5693359375,
0.8984375,
0.5693359375,
0.8984375,
0.5830078125,
0.496337890625,
0.5830078125
],
[
0.0679931640625,
0.58056640625,
0.470947265625,
0.58056640625,
0.470947265625,
0.59423828125,
0.0679931640625,
0.59423828125
],
[
0.496337890625,
0.58447265625,
0.896484375,
0.58447265625,
0.896484375,
0.59716796875,
0.496337890625,
0.59716796875
],
[
0.0679931640625,
0.59423828125,
0.47021484375,
0.59423828125,
0.47021484375,
0.60791015625,
0.0679931640625,
0.60791015625
],
[
0.495361328125,
0.599609375,
0.8984375,
0.59716796875,
0.8984375,
0.61083984375,
0.495361328125,
0.61328125
],
[
0.0679931640625,
0.60986328125,
0.453369140625,
0.60986328125,
0.453369140625,
0.62353515625,
0.0679931640625,
0.62353515625
],
[
0.497802734375,
0.6142578125,
0.90185546875,
0.6142578125,
0.90185546875,
0.62744140625,
0.497802734375,
0.62744140625
],
[
0.497802734375,
0.62744140625,
0.8984375,
0.62890625,
0.8984375,
0.642578125,
0.497802734375,
0.64111328125
],
[
0.0679931640625,
0.63525390625,
0.47021484375,
0.63525390625,
0.47021484375,
0.6484375,
0.0679931640625,
0.6484375
],
[
0.497802734375,
0.6435546875,
0.89990234375,
0.6435546875,
0.89990234375,
0.65673828125,
0.497802734375,
0.65673828125
],
[
0.0679931640625,
0.65087890625,
0.470947265625,
0.65087890625,
0.470947265625,
0.66455078125,
0.0679931640625,
0.66455078125
],
[
0.497802734375,
0.6591796875,
0.8984375,
0.6591796875,
0.8984375,
0.66943359375,
0.497802734375,
0.66943359375
],
[
0.07135009765625,
0.66748046875,
0.468505859375,
0.66748046875,
0.468505859375,
0.677734375,
0.07135009765625,
0.677734375
],
[
0.496337890625,
0.6728515625,
0.8984375,
0.6728515625,
0.8984375,
0.68603515625,
0.496337890625,
0.68603515625
],
[
0.0679931640625,
0.68115234375,
0.378662109375,
0.68115234375,
0.378662109375,
0.69482421875,
0.0679931640625,
0.69482421875
],
[
0.496337890625,
0.68701171875,
0.89990234375,
0.68701171875,
0.89990234375,
0.70068359375,
0.496337890625,
0.70068359375
],
[
0.0679931640625,
0.705078125,
0.470947265625,
0.70361328125,
0.470947265625,
0.71875,
0.0679931640625,
0.7197265625
],
[
0.497802734375,
0.703125,
0.89990234375,
0.703125,
0.89990234375,
0.71630859375,
0.497802734375,
0.71630859375
],
[
0.0697021484375,
0.72021484375,
0.399658203125,
0.72021484375,
0.399658203125,
0.7353515625,
0.0697021484375,
0.7353515625
],
[
0.496337890625,
0.71728515625,
0.7197265625,
0.71728515625,
0.7197265625,
0.73095703125,
0.496337890625,
0.73095703125
],
[
0.52294921875,
0.73291015625,
0.89990234375,
0.73291015625,
0.89990234375,
0.74658203125,
0.52294921875,
0.74658203125
],
[
0.496337890625,
0.74658203125,
0.892578125,
0.74658203125,
0.892578125,
0.76025390625,
0.496337890625,
0.76025390625
],
[
0.0679931640625,
0.76123046875,
0.26611328125,
0.76318359375,
0.26611328125,
0.77685546875,
0.0679931640625,
0.77490234375
],
[
0.497802734375,
0.7626953125,
0.89404296875,
0.7626953125,
0.89404296875,
0.77685546875,
0.497802734375,
0.77685546875
],
[
0.496337890625,
0.775390625,
0.89501953125,
0.77783203125,
0.89501953125,
0.79150390625,
0.496337890625,
0.7890625
],
[
0.0697021484375,
0.79150390625,
0.47265625,
0.79150390625,
0.47265625,
0.8046875,
0.0697021484375,
0.8046875
],
[
0.496337890625,
0.79345703125,
0.8984375,
0.79345703125,
0.8984375,
0.80419921875,
0.496337890625,
0.80419921875
],
[
0.0679931640625,
0.8046875,
0.30322265625,
0.8056640625,
0.30322265625,
0.82080078125,
0.0679931640625,
0.8193359375
],
[
0.496337890625,
0.80712890625,
0.89990234375,
0.80712890625,
0.89990234375,
0.82080078125,
0.496337890625,
0.82080078125
],
[
0.496337890625,
0.82177734375,
0.87890625,
0.82177734375,
0.87890625,
0.8359375,
0.496337890625,
0.8359375
],
[
0.06634521484375,
0.830078125,
0.47021484375,
0.83056640625,
0.47021484375,
0.845703125,
0.06634521484375,
0.84423828125
],
[
0.0697021484375,
0.8466796875,
0.470947265625,
0.8466796875,
0.470947265625,
0.86083984375,
0.0697021484375,
0.86083984375
],
[
0.497802734375,
0.84423828125,
0.90185546875,
0.8466796875,
0.90185546875,
0.8642578125,
0.497802734375,
0.8623046875
],
[
0.496337890625,
0.85986328125,
0.892578125,
0.85986328125,
0.892578125,
0.87353515625,
0.496337890625,
0.87353515625
],
[
0.0906982421875,
0.87255859375,
0.470947265625,
0.87255859375,
0.470947265625,
0.8828125,
0.0906982421875,
0.8828125
],
[
0.496337890625,
0.87353515625,
0.8984375,
0.87353515625,
0.8984375,
0.88427734375,
0.496337890625,
0.88427734375
],
[
0.06634521484375,
0.88427734375,
0.453369140625,
0.88427734375,
0.453369140625,
0.89794921875,
0.06634521484375,
0.89794921875
],
[
0.495361328125,
0.88427734375,
0.89990234375,
0.88427734375,
0.89990234375,
0.89794921875,
0.495361328125,
0.89794921875
]
] |
[
"Dissociation andPropagation forApproximateLifted Inferencewith Standard Relational DatabaseManagement Systems",
"4",
"query evaluation is to compute P[ql, i.e. the probability that",
"a relational atom containing the variables x, π is the proba-",
"the query is true in a randomly chosen world. We will refer",
"bilistic project operator with duplicate elimination (or short",
"to this probability as the “query reliability\" r(q) [37].",
"\"projection), and P...| is the probabilistic natural join",
"It is known that the data complexity [77] of any query g",
"(or short join\") in prefix notation, which we allow to be k-",
"is either in PTIME or #P-hard [14]. The PTIME queries are",
"ary (k ≥ 2). We require that joins and projections alternate",
"also called \"safe queries\" and, for the case of sj-free CQs,",
"in a plan and do not distinguish between join orders.",
"",
"are characterized precisely by a syntactic property called hi-",
"erarchical queries [13]. We briefly review these results:",
"We write Var(P) for all variables in a plan P and HVar(P)",
"for its head variables, which are recursively defined as",
"Definition 3 (Hierarchical query) A query q is called hier-",
"follows: (1) if P = R;(x), then HVar(P) = x; (2) if P =",
"archical iff for any two existential variables x,y E Evar(q),",
"元(P'), then HVar = x; and (3) if P =◆P [PI,..., Pk], then",
"one of the following three conditions holds: at(x) C at(y),",
"HVar(P) =U-,HVar(P).The existential variables EVar(P)",
"at(x) 2 at(y), or at(x)nat(y) = 0.",
"are then defined as Var(P) -HVar(P).",
"Every plan P represents a query qp defined by taking all",
"For example, the query q1 :- R(x, y), S(y,z), T(y,z, u) is hier-",
"atoms mentioned in P as the body and setting HVar(qp) =",
"archical, while q2 :-R(x,y), S(y,z),T(z,u) is not as neither",
"of the three conditions holds for the variables y and z.",
"HVar(P). A plan is called Boolean if HVar(P) = 0. We as-",
"sume the usual sanity conditions on plans to be satisfied: for",
"Theorem 4 (Hierarchy dichotomy [13]) If g is hierarchi-",
"a projection πP we assume x C HVar(P), and each variable",
"cal, then P[ql can be computed in PTIME in the size of D.",
"y is projected away at most once in a plan, i.e. there exists at",
"Otherwise, computing P[q] is #P-hard in the size of D.",
"mostoneoperator元pP s.t.yEHVar(P)-x.Fornotational",
"convenience, we also use the “project-away operator\"πP,P",
"We next give an equivalent, recursive characterization of",
"instead of p, where y are the variables being projected",
"hierarchical queries, for which we need a few definitions",
"away, i.e. x = HVar(元,P) = HVar(P) -y.",
"We write SepVar(q) for the set of existential variables that",
"Each subplan P returns an intermediate relation of ar-",
"appear in every atom (called\"separator variables\").A con-",
"ity[HVar(P)/+1.Theextraprobabilityattributestoresa",
"nected component of q (or short, \"query component\") is a",
"score(t)foreachoutputtupletEP(D).Givenaprobabilistic",
"subset of atoms that are connected via existential variables.",
"database D and a plan P, score(t) is defined inductively on",
"Aqueryqisdisconnected if its atomscanbepartitionedinto",
"the structure of P as follows: (1) If t E R;(x), then score(t) =",
"two non-empty sets that do not share any existential vari-",
"P[t], i.e. its probability in D; (2) if t E P[P(D),..., P (D)]",
"ables (e.g., q:-R(x,y),S(z,u),T(u, v) is disconnected and",
"where t =P [ti, ..,tl, then score(t) = II'- score(ti); and",
"has two query components: \"R(x,y)\" and \"S(z,u), T(u, v)\")",
"(3) if t E P(D), and ti,...,tn E P(D) are all the tuples that",
"Foreverysetofvariablesx,denoteq-xthequeryobtained",
"project onto t, then score(t) = @\"- score(t), where “\"",
"by removing all variables x and decreasing the arities of the",
"stands for the independent-or. In other words, score com",
"relational symbols that contain variables from x. Any query",
"putes a probability by assuming that all tuples joined by P",
"can become disconnected by removing a set of variables.",
"and all duplicates eliminated by P are independent. Only if",
"these conditions hold, then score is the correct query prob-",
"Lemma 5 (Hierarchical queries) A query q is “hierarchi-",
"ability (also called \"query reliability\" [37]), but in general",
"cal\" iff either: (1) q has a single atom; (2) q has k ≥ 2 query",
"it is not.Therefore.score is also called an extensional se-",
"componentsallofwhicharehierarchical;or(3)ghasasep-",
"mantics [26, 57,62] and is, in general, not equal to the query",
"arator variable x, and q -(x) is hierarchical.",
"probability, which is defined in terms of possible worlds:",
"Every hierarchical query can be computed in PTIME, but",
"score(P) P[qp]. For a Boolean plan P, we get one single",
"non-hierarchical queries are #P-hard, in general.4",
"score, which we denote score(P)",
"The requirement that joins and projections alternate is",
"w.l.o.g. because nested joins, such as p[P[Ri,R2],R3",
"2.2 Probabilistic query plans",
"or P[R1, P[R2,R3]], can be rewritten into P[R1, R2, R3",
"while keeping the same score, e.g., (pip2)p3 = pi(p2p3)",
"Unless otherwise stated, a “query plan\" in this paper always",
"Forthe samereason wedo not distinguishbetween different",
"denotes a probabilistic query plan.",
"permutations in the joins, called join orders [50]. We do not",
"focus on query optimization in this paper until Section 6.",
"Definition 6 (Query plans) A query plan Pis given by the",
"grammar P ::= Ri(x) | 元P P[Pi,,P| where Ri(x) is",
"5Extensional approaches compute the probability of anyformula",
"as a function of the probabilities of its subformulas according to syntac",
"Non-hierarchical queries can be in PTIME when considering func-",
"tic rules,regardless of howthose were derived.Intensional approaches",
"tional dependencies or deterministic tables [13,54] (see Section 5).",
"reason in terms of possible worlds and keep track of dependencies [57]."
] |
{
"arxiv_id": "1310.6257",
"language": "en",
"page_idx": 4
}
|
||
1502.03520-en-0
|
1502.03520-en-0.jpg
|
RAND-WALK: A latent variable model approach to word
# embeddings
Yingyu Liang
Tengyu Ma Andrej Risteski *
Sanjeev Arora Yuanzhi Li
Abstract
Semantic word embeddings represent the meaning of a word via a vector, and are created by diverse methods. Many use nonlinear operations on co-occurrence statistics, and have hand-tuned hyperparameters and reweighting methods.
This paper proposes a new generative model, a dynamic version of the log-linear topic model of Mnih and Hinton (2007). The methodological novelty is to use the prior to compute closed form expressions for word statistics. This provides a theoretical justification for nonlinear models like PMI, word2vec, and GloVe, as well as some hyperparameter choices. It also helps explain why low-dimensional semantic embeddings contain linear algebraic structure that allows solution of word analogies, as shown by Mikolov et al. (2013a) and many subsequent papers.
Experimental support is provided for the generative model assumptions, the most important of which
is that latent word vectors are fairly uniformly dispersed in space.
## 1 Introduction
Vector representations of words (word embeddings) try to capture relationships between words as distance or angle, and have many applications in computational linguistics and machine learning. They are constructed by various models whose unifying philosophy is that the meaning of a word is defined by“the company it keeps" (Firth, 1957), namely, co-occurrence statistics. The simplest methods use word vectors that explicitly represent co-occurrence statistics. Reweighting heuristics are known to improve these methods, as is dimension reduction (Deerwester et al., 1990). Some reweighting methods are nonlinear, which include taking the square root of co-occurrence counts (Rohde et al., 2006), or the logarithm, or the related Pointwise Mutual Information (PMI) (Church and Hanks, 1990). These are collectively referred to as Vector Space Models, surveyed in (Turney and Pantel, 2010).
Neural network language models (Rumelhart et al., 1986; 1988; Bengio et al., 2006; Collobert and Weston, 2008a) propose another way to construct embeddings: the word vector is simply the neural network's internal representation for the word. This method is nonlinear and nonconvex. It was popularized via word2vec, a family of energy-based models in (Mikolov et al., 2013b;c), followed by a matrix factorization approach called GloVe (Pennington et al., 2014). The first paper also showed how to solve analogies using linear algebra on word embeddings. Experiments and theory were used to suggest that these newer methods are related to the older PMI-based models, but with new hyperparameters and/or term reweighting methods (Levy and Goldberg, 2014b).
But note that even the old PMI method is a bit mysterious. The simplest version considers a symmetric $\begin{array}{c}\text{matrix with each row/column indexed by a word. The entry for }(w,w^{\prime})\mathrm{~is~PMI}(w,w^{\prime})=\log\frac{p(w,w^{\prime})}{p(w)p(w^{\prime})},\mathrm{~where}\end{array}$ $p(w,w^{\prime})$ is the empirical probability of words $w,w^\prime$ appearing within a window of certain size in the corpus, and $p(w)$ is the marginal probability of $w.$ (More complicated models could use asymmetric matrices with
$^*$Princeton University, Computer Science Department. $\{$arora,yuanzhil,yingyul,tengyu,risteski$\}$@cs.princeton.edu
This work was supported in part by NSF grants CCF-0832797, CCF-1117309, CCF-1302518, DMS-1317308, Simons Investigator Award, and Simons Collaboration Grant. Tengyu Ma was also supported by Simons Award for Graduate Students in Theoretical Award, and Simons Collaboration Grant. Tengyu Ma was also supported by Simons Award for Gradu ate Students in Theoretical Computer Science and IBM PhD Fellowship.
1
|
[
[
0.1707763671875,
0.136962890625,
0.82666015625,
0.1395263671875,
0.82666015625,
0.160400390625,
0.1707763671875,
0.1578369140625
],
[
0.43310546875,
0.1685791015625,
0.56787109375,
0.1685791015625,
0.56787109375,
0.1881103515625,
0.43310546875,
0.1881103515625
],
[
0.29736328125,
0.205810546875,
0.39794921875,
0.205810546875,
0.39794921875,
0.2235107421875,
0.29736328125,
0.2235107421875
],
[
0.426513671875,
0.2052001953125,
0.55224609375,
0.20703125,
0.55224609375,
0.2247314453125,
0.426513671875,
0.222900390625
],
[
0.583984375,
0.2083740234375,
0.68701171875,
0.2083740234375,
0.68701171875,
0.222900390625,
0.583984375,
0.222900390625
],
[
0.72119140625,
0.2083740234375,
0.865234375,
0.2083740234375,
0.865234375,
0.222900390625,
0.72119140625,
0.222900390625
],
[
0.021240234375,
0.26513671875,
0.058013916015625,
0.26513671875,
0.058013916015625,
0.401611328125,
0.021240234375,
0.401611328125
],
[
0.1396484375,
0.2083740234375,
0.261474609375,
0.2083740234375,
0.261474609375,
0.222900390625,
0.1396484375,
0.222900390625
],
[
0.468017578125,
0.268310546875,
0.53515625,
0.268310546875,
0.53515625,
0.27978515625,
0.468017578125,
0.27978515625
],
[
0.178955078125,
0.287353515625,
0.84326171875,
0.287353515625,
0.84326171875,
0.300537109375,
0.178955078125,
0.300537109375
],
[
0.156005859375,
0.3017578125,
0.84228515625,
0.3017578125,
0.84228515625,
0.31494140625,
0.156005859375,
0.31494140625
],
[
0.156005859375,
0.31494140625,
0.365234375,
0.31494140625,
0.365234375,
0.328857421875,
0.156005859375,
0.328857421875
],
[
0.178955078125,
0.328857421875,
0.84228515625,
0.328857421875,
0.84228515625,
0.34228515625,
0.178955078125,
0.34228515625
],
[
0.156005859375,
0.34228515625,
0.84228515625,
0.34228515625,
0.84228515625,
0.356689453125,
0.156005859375,
0.356689453125
],
[
0.156005859375,
0.356689453125,
0.84228515625,
0.356689453125,
0.84228515625,
0.37060546875,
0.156005859375,
0.37060546875
],
[
0.1568603515625,
0.37255859375,
0.84228515625,
0.37255859375,
0.84228515625,
0.38330078125,
0.1568603515625,
0.38330078125
],
[
0.1544189453125,
0.38330078125,
0.84326171875,
0.3837890625,
0.84326171875,
0.3984375,
0.1544189453125,
0.397705078125
],
[
0.021240234375,
0.40478515625,
0.056365966796875,
0.404052734375,
0.060455322265625,
0.4873046875,
0.02532958984375,
0.488037109375
],
[
0.1568603515625,
0.399658203125,
0.44287109375,
0.399658203125,
0.44287109375,
0.410400390625,
0.1568603515625,
0.410400390625
],
[
0.178955078125,
0.410888671875,
0.84326171875,
0.410888671875,
0.84326171875,
0.425537109375,
0.178955078125,
0.425537109375
],
[
0.156005859375,
0.425537109375,
0.59326171875,
0.425537109375,
0.59326171875,
0.439453125,
0.156005859375,
0.439453125
],
[
0.11767578125,
0.466552734375,
0.131591796875,
0.466552734375,
0.131591796875,
0.47607421875,
0.11767578125,
0.47607421875
],
[
0.15283203125,
0.464111328125,
0.301513671875,
0.464111328125,
0.301513671875,
0.479248046875,
0.15283203125,
0.479248046875
],
[
0.0196075439453125,
0.488037109375,
0.058013916015625,
0.488037109375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.1168212890625,
0.494384765625,
0.88232421875,
0.494384765625,
0.88232421875,
0.50830078125,
0.1168212890625,
0.50830078125
],
[
0.1168212890625,
0.51025390625,
0.88232421875,
0.51025390625,
0.88232421875,
0.5234375,
0.1168212890625,
0.5234375
],
[
0.1168212890625,
0.5244140625,
0.88232421875,
0.5244140625,
0.88232421875,
0.5390625,
0.1168212890625,
0.5390625
],
[
0.11517333984375,
0.54052734375,
0.88134765625,
0.54052734375,
0.88134765625,
0.5537109375,
0.11517333984375,
0.5537109375
],
[
0.11517333984375,
0.5546875,
0.88134765625,
0.5546875,
0.88134765625,
0.568359375,
0.11517333984375,
0.568359375
],
[
0.11517333984375,
0.5693359375,
0.88232421875,
0.5693359375,
0.88232421875,
0.58349609375,
0.11517333984375,
0.58349609375
],
[
0.1168212890625,
0.58544921875,
0.88232421875,
0.58544921875,
0.88232421875,
0.5986328125,
0.1168212890625,
0.5986328125
],
[
0.11517333984375,
0.599609375,
0.88232421875,
0.599609375,
0.88232421875,
0.6142578125,
0.11517333984375,
0.6142578125
],
[
0.11358642578125,
0.61572265625,
0.396240234375,
0.61572265625,
0.396240234375,
0.62890625,
0.11358642578125,
0.62890625
],
[
0.1396484375,
0.6298828125,
0.88134765625,
0.6298828125,
0.88134765625,
0.64453125,
0.1396484375,
0.64453125
],
[
0.1168212890625,
0.64599609375,
0.88134765625,
0.64599609375,
0.88134765625,
0.66015625,
0.1168212890625,
0.66015625
],
[
0.11517333984375,
0.6591796875,
0.88232421875,
0.66015625,
0.88232421875,
0.6748046875,
0.11517333984375,
0.673828125
],
[
0.1168212890625,
0.67529296875,
0.88134765625,
0.67529296875,
0.88134765625,
0.68994140625,
0.1168212890625,
0.68994140625
],
[
0.1168212890625,
0.69140625,
0.88134765625,
0.69140625,
0.88134765625,
0.705078125,
0.1168212890625,
0.705078125
],
[
0.11517333984375,
0.70556640625,
0.88232421875,
0.70556640625,
0.88232421875,
0.7197265625,
0.11517333984375,
0.7197265625
],
[
0.11517333984375,
0.7197265625,
0.88232421875,
0.72021484375,
0.88232421875,
0.73486328125,
0.11517333984375,
0.734375
],
[
0.11517333984375,
0.73486328125,
0.2459716796875,
0.73486328125,
0.2459716796875,
0.75,
0.11517333984375,
0.75
],
[
0.1405029296875,
0.75048828125,
0.88232421875,
0.75048828125,
0.88232421875,
0.76513671875,
0.1405029296875,
0.76513671875
],
[
0.1094970703125,
0.7587890625,
0.88623046875,
0.7646484375,
0.88623046875,
0.7890625,
0.1094970703125,
0.7841796875
],
[
0.779296875,
0.767578125,
0.81640625,
0.767578125,
0.81640625,
0.7744140625,
0.779296875,
0.7744140625
],
[
0.11517333984375,
0.78515625,
0.88134765625,
0.78515625,
0.88134765625,
0.7998046875,
0.11517333984375,
0.7998046875
],
[
0.11517333984375,
0.8017578125,
0.88232421875,
0.8017578125,
0.88232421875,
0.81494140625,
0.11517333984375,
0.81494140625
],
[
0.131591796875,
0.82275390625,
0.88134765625,
0.82373046875,
0.88134765625,
0.8369140625,
0.131591796875,
0.8359375
],
[
0.11358642578125,
0.83544921875,
0.88232421875,
0.8359375,
0.88232421875,
0.849609375,
0.11358642578125,
0.8486328125
],
[
0.1168212890625,
0.8486328125,
0.8837890625,
0.8486328125,
0.8837890625,
0.8623046875,
0.1168212890625,
0.8623046875
],
[
0.1168212890625,
0.8623046875,
0.390625,
0.8623046875,
0.390625,
0.87255859375,
0.1168212890625,
0.87255859375
],
[
0.494384765625,
0.8994140625,
0.5048828125,
0.8994140625,
0.5048828125,
0.9091796875,
0.494384765625,
0.9091796875
]
] |
[
"RAND-WALK: A latent variable model approach to word",
"embeddings",
"Yuanzhi Li",
"Yingyu Liang",
"Tengyu Ma",
"Andrej Risteski *",
"610z unf 61",
"Sanjeev Arora",
"Abstract",
"Semantic word embeddings represent the meaning of a word via a vector, and are created by diverse",
"methods. Many use nonlinear operations on co-occurrence statistics, and have hand-tuned hyperparam-",
"eters and reweighting methods.",
"This paper proposes a new generative model, a dynamic version of the log-linear topic model of Mnih",
"and Hinton (2007). The methodological novelty is to use the prior to compute closed form expressions",
"for word statistics.. This provides a theoretical justification for nonlinear models like PMI, word2vec,",
"and GloVe, as well as some hyperparameter choices.It also helps explain why low-dimensional semantic",
"embeddings contain linear algebraic structure that allows solution of word analogies, as shown by Mikolov",
"[DT'S]",
"et al.(2013a) and many subsequent papers",
"Experimental support is provided for the generative model assumptions, the most important of which",
"is that latent word vectors are fairly uniformly dispersed in space.",
"1",
"Introduction",
"00000",
"Vector representations of words (word embeddings) try to capture relationships between words as distance or",
"angle, and have many applications in computational linguistics and machine learning. They are constructed",
"by various models whose unifying philosophy is that the meaning of a word is defined by \"the company it",
"keeps\" (Firth, 1957), namely, co-occurrence statistics. The simplest methods use word vectors that explicitly",
"represent co-occurrence statistics. Reweighting heuristics are known to improve these methods, as is dimen-",
"sion reduction (Deerwester et al., 1990). Some reweighting methods are nonlinear, which include taking the",
"square root of co-occurrence counts (Rohde et al., 2006), or the logarithm, or the related Pointwise Mutual",
"Information (PMI) (Church and Hanks, 1990). These are collectively referred to as Vector Space Models,",
"surveyed in (Turney and Pantel, 2010).",
"Neural network language models (Rumelhart et al., 1986; 1988; Bengio et al., 2006; Collobert and Weston,",
"2008a) propose another way to construct embeddings: the word vector is simply the neural network's internal",
"representation for the word. This method is nonlinear and nonconvex. It was popularized via word2vec, a",
"family of energy-based models in (Mikolov et al., 2013b;c), followed by a matrix factorization approach called",
"GloVe (Pennington et al., 2014). The first paper also showed how to solve analogies using linear algebra on",
"word embeddings. Experiments and theory were used to suggest that these newer methods are related to",
"the older PMI-based models, but with new hyperparameters and/or term reweighting methods (Levy and",
"Goldberg, 2014b).",
"But note that even the old PMI method is a bit mysterious. The simplest version considers a symmetric",
"matrix with each row/column indexed by a word. The entry for (w, wl) is PMI(w, w') = log ), where",
"p(w.w",
"p(w, w') is the empirical probability of words w, w' appearing within a window of certain size in the corpus,",
"and p(w) is the marginal probability of w. (More complicated models could use asymmetric matrices with",
"*Princeton University,Computer Science Department..(arora,yuanzhil,yingyul,tengyu,risteski@cs.princeton.edu.",
"This workwas supported in part by NSFgrants CCF-0832797,CCF-1117309,CCF-1302518, DMS-1317308, Simons Investigator",
"Award, and Simons Collaboration Grant. Tengyu Ma was also supported by Simons Award for Graduate Students in Theoretical",
"ComputerScienceandIBMPhDFellowship.",
"1"
] |
{
"arxiv_id": "1502.03520",
"language": "en",
"page_idx": 0
}
|
||
1502.03520-en-1
|
1502.03520-en-1.jpg
|
columns corresponding to context words or phrases, and also involve tensorization.) Then word vectors are obtained by low-rank SVD on this matrix, or a related matrix with term reweightings. In particular, the PMI matrix is found to be closely approximated by a low rank matrix: there exist word vectors in say 300 dimensions, which is much smaller than the number of words in the dictionary, such that
$$\langle v_w,v_{w'}\rangle\approx\mathrm{PMI}(w,w')$$
(1.1)
where$\approx$should be interpreted loosely.
There appears to be no theoretical explanation for this empirical finding about the approximate low rank of the PMI matrix. The current paper addresses this. Specifically, we propose a probabilistic model of text generation that augments the log-linear topic model of Mnih and Hinton (2007) with dynamics, in the form of a random walk over a latent discourse space. The chief methodological contribution is using the model priors to analytically derive a closed-form expression that directly explains (1.1); see Theorem 2.2 in Section 2. Section 3 builds on this insight to give a rigorous justification for models such as word2vec and GloVe, including the hyperparameter choices for the latter. The insight also leads to a mathematical explanation for why these word embeddings allow analogies to be solved using linear algebra; see Section 4. Section 5 shows good empirical ft to this model's assumtions and predictions, including the surprising one that word vectors are pretty uniformly distributed (isotropic) in space.$^1$
## 1.1 Related work
Latent variable probabilistic models of language have been used for word embeddings before, including Latent Dirichlet Allocation (LDA) and its more complicated variants (see the survey (Blei, 2012)), and some neurally inspired nonlinear models (Mnih and Hinton, 2007; Maas et al., 2011). In fact, LDA evolved out of efforts in the 1990s to provide a generative model that “explains” the success of older vector space methods like Latent Semantic Indexing (Papadimitriou et al., 1998; Hofmann, 1999). However, none of these earlier generative models has been linked to PMI models.
Levy and Goldberg (2014b) tried to relate word2vec to PMI models. They showed that if there were no dimension constraint in word2vec, specifically, the “ skip- gram with negative sampling ( SGNS) ” version of the model, then its solutions would satisfy ( 1. 1) , provided the right hand side were replaced by PMI$( w, w^{\prime }) - \beta$ for some scalar $\beta.$ However, skip-gram is a discriminative model (due to the use of negative sampling), not generative. Furthermore, their argument only applies to very high-dimensional word embeddings, and thus does not address low-dimensional embeddings, which have superior quality in applications
Hashimoto et al. (2016) focuses on issues similar to our paper. They model text generation as a random walk on words, which are assumed to be embedded as vectors in a geometric space. Given that the last word produced was $w$, the probability that the next word is $w^\prime$ is assumed to be given by $h(|v_w-v_{w^{\prime}}|^2)$ for a suitable function $h$, and this model leads to an explanation of (1.1). By contrast our random walk involves a latent discourse vector, which has a clearer semantic interpretation and has proven useful in subsequent work, e.g. understanding structure of word embeddings for polysemous words Arora et al. (2016). Also our work clarifies some weighting and bias terms in the training objectives of previous methods (Section 3) and also the phenomenon discussed in the next paragraph.
Researchers have tried to understand why vectors obtained from the highly nonlinear word2vec models exhibit linear structures (Levy and Goldberg, 2014a; Pennington et al., 2014). Specifically, for analogies like “$man:woman::king:??,”queen$ happens to be the word whose vector $v_queen$ is the most similar to the vector $v_{king}-v_{man}+v_{woman}.$ This suggests that simple semantic relationships, such as masculine vs feminine tested in the above example, correspond approximately to a single direction in space, a phenomenon we will henceforth refer to as RELATIONS=LINES.
Section 4 surveys earlier attempts to explain this phenomenon and their shortcoming, namely, that they ignore the large approximation error in relationships like (1.1). This error appears larger than the difference between the best solution and the second best (incorrect) solution in analogy solving, so that this error could in principle lead to a complete failure in analogy solving. In our explanation, the low dimensionality of the
$^{1}$The code is available at https://github.com/PrincetonML/SemanticVector
2
|
[
[
0.11517333984375,
0.0921630859375,
0.88232421875,
0.0921630859375,
0.88232421875,
0.1060791015625,
0.11517333984375,
0.1060791015625
],
[
0.11517333984375,
0.1060791015625,
0.88232421875,
0.106689453125,
0.88232421875,
0.121826171875,
0.11517333984375,
0.12060546875
],
[
0.1168212890625,
0.12249755859375,
0.8837890625,
0.12249755859375,
0.8837890625,
0.136962890625,
0.1168212890625,
0.136962890625
],
[
0.11517333984375,
0.1383056640625,
0.7548828125,
0.1383056640625,
0.7548828125,
0.15283203125,
0.11517333984375,
0.15283203125
],
[
0.4150390625,
0.1622314453125,
0.58251953125,
0.1610107421875,
0.58251953125,
0.1787109375,
0.4150390625,
0.179931640625
],
[
0.84716796875,
0.1634521484375,
0.8837890625,
0.1634521484375,
0.8837890625,
0.179931640625,
0.84716796875,
0.179931640625
],
[
0.11517333984375,
0.1912841796875,
0.393798828125,
0.1912841796875,
0.393798828125,
0.2052001953125,
0.11517333984375,
0.2052001953125
],
[
0.1405029296875,
0.205810546875,
0.88134765625,
0.205810546875,
0.88134765625,
0.2203369140625,
0.1405029296875,
0.2203369140625
],
[
0.11517333984375,
0.2215576171875,
0.88232421875,
0.2215576171875,
0.88232421875,
0.236083984375,
0.11517333984375,
0.236083984375
],
[
0.11358642578125,
0.236083984375,
0.88232421875,
0.23486328125,
0.88232421875,
0.25,
0.11358642578125,
0.250732421875
],
[
0.11517333984375,
0.251953125,
0.88134765625,
0.251953125,
0.88134765625,
0.26513671875,
0.11517333984375,
0.26513671875
],
[
0.11517333984375,
0.267578125,
0.88134765625,
0.267578125,
0.88134765625,
0.281005859375,
0.11517333984375,
0.281005859375
],
[
0.11517333984375,
0.2822265625,
0.88134765625,
0.2822265625,
0.88134765625,
0.29541015625,
0.11517333984375,
0.29541015625
],
[
0.11517333984375,
0.296630859375,
0.88232421875,
0.296630859375,
0.88232421875,
0.31005859375,
0.11517333984375,
0.31005859375
],
[
0.11517333984375,
0.3125,
0.88232421875,
0.3125,
0.88232421875,
0.32568359375,
0.11517333984375,
0.32568359375
],
[
0.11358642578125,
0.32568359375,
0.8837890625,
0.326904296875,
0.8837890625,
0.341552734375,
0.11358642578125,
0.34033203125
],
[
0.11517333984375,
0.34228515625,
0.62841796875,
0.34228515625,
0.62841796875,
0.35595703125,
0.11517333984375,
0.35595703125
],
[
0.11517333984375,
0.375732421875,
0.295654296875,
0.375732421875,
0.295654296875,
0.39013671875,
0.11517333984375,
0.39013671875
],
[
0.11517333984375,
0.399658203125,
0.88134765625,
0.399658203125,
0.88134765625,
0.4140625,
0.11517333984375,
0.4140625
],
[
0.1168212890625,
0.417236328125,
0.8798828125,
0.417236328125,
0.8798828125,
0.427978515625,
0.1168212890625,
0.427978515625
],
[
0.11517333984375,
0.43115234375,
0.88232421875,
0.43115234375,
0.88232421875,
0.4443359375,
0.11517333984375,
0.4443359375
],
[
0.11517333984375,
0.44580078125,
0.88232421875,
0.44580078125,
0.88232421875,
0.458984375,
0.11517333984375,
0.458984375
],
[
0.1168212890625,
0.4609375,
0.88232421875,
0.4609375,
0.88232421875,
0.474853515625,
0.1168212890625,
0.474853515625
],
[
0.11517333984375,
0.4765625,
0.476318359375,
0.4765625,
0.476318359375,
0.490478515625,
0.11517333984375,
0.490478515625
],
[
0.1405029296875,
0.490478515625,
0.88134765625,
0.490478515625,
0.88134765625,
0.5048828125,
0.1405029296875,
0.5048828125
],
[
0.11517333984375,
0.5068359375,
0.88134765625,
0.5068359375,
0.88134765625,
0.52099609375,
0.11517333984375,
0.52099609375
],
[
0.11358642578125,
0.521484375,
0.8837890625,
0.521484375,
0.8837890625,
0.53515625,
0.11358642578125,
0.53515625
],
[
0.11517333984375,
0.5361328125,
0.88232421875,
0.5361328125,
0.88232421875,
0.55029296875,
0.11517333984375,
0.55029296875
],
[
0.1168212890625,
0.5517578125,
0.88232421875,
0.5517578125,
0.88232421875,
0.56494140625,
0.1168212890625,
0.56494140625
],
[
0.11517333984375,
0.56640625,
0.765625,
0.56640625,
0.765625,
0.58056640625,
0.11517333984375,
0.58056640625
],
[
0.1405029296875,
0.58203125,
0.88134765625,
0.58203125,
0.88134765625,
0.59521484375,
0.1405029296875,
0.59521484375
],
[
0.11517333984375,
0.59521484375,
0.88134765625,
0.59521484375,
0.88134765625,
0.60986328125,
0.11517333984375,
0.60986328125
],
[
0.11517333984375,
0.6123046875,
0.88232421875,
0.6123046875,
0.88232421875,
0.62548828125,
0.11517333984375,
0.62548828125
],
[
0.11358642578125,
0.626953125,
0.88232421875,
0.626953125,
0.88232421875,
0.64013671875,
0.11358642578125,
0.64013671875
],
[
0.11358642578125,
0.642578125,
0.88134765625,
0.642578125,
0.88134765625,
0.65576171875,
0.11358642578125,
0.65576171875
],
[
0.11517333984375,
0.6572265625,
0.88134765625,
0.6572265625,
0.88134765625,
0.67041015625,
0.11517333984375,
0.67041015625
],
[
0.11517333984375,
0.67236328125,
0.88232421875,
0.67236328125,
0.88232421875,
0.68603515625,
0.11517333984375,
0.68603515625
],
[
0.11358642578125,
0.68603515625,
0.5087890625,
0.6875,
0.5087890625,
0.7021484375,
0.11358642578125,
0.70068359375
],
[
0.1405029296875,
0.7021484375,
0.88232421875,
0.7021484375,
0.88232421875,
0.71533203125,
0.1405029296875,
0.71533203125
],
[
0.11273193359375,
0.71630859375,
0.88134765625,
0.71728515625,
0.88134765625,
0.73193359375,
0.11273193359375,
0.73095703125
],
[
0.11767578125,
0.73291015625,
0.88134765625,
0.73291015625,
0.88134765625,
0.74755859375,
0.11767578125,
0.74755859375
],
[
0.11767578125,
0.75,
0.1641845703125,
0.751953125,
0.16259765625,
0.76318359375,
0.1168212890625,
0.76123046875
],
[
0.1650390625,
0.7470703125,
0.88232421875,
0.744140625,
0.88232421875,
0.76220703125,
0.1650390625,
0.7646484375
],
[
0.1168212890625,
0.76318359375,
0.88232421875,
0.76318359375,
0.88232421875,
0.7763671875,
0.1168212890625,
0.7763671875
],
[
0.1168212890625,
0.77978515625,
0.410888671875,
0.77978515625,
0.410888671875,
0.79052734375,
0.1168212890625,
0.79052734375
],
[
0.1396484375,
0.79345703125,
0.88134765625,
0.79345703125,
0.88134765625,
0.806640625,
0.1396484375,
0.806640625
],
[
0.1168212890625,
0.8095703125,
0.8798828125,
0.8095703125,
0.8798828125,
0.82275390625,
0.1168212890625,
0.82275390625
],
[
0.1168212890625,
0.82373046875,
0.88232421875,
0.82373046875,
0.88232421875,
0.8369140625,
0.1168212890625,
0.8369140625
],
[
0.11517333984375,
0.8388671875,
0.88232421875,
0.8388671875,
0.88232421875,
0.85302734375,
0.11517333984375,
0.85302734375
],
[
0.134033203125,
0.85986328125,
0.58251953125,
0.85986328125,
0.58251953125,
0.875,
0.134033203125,
0.875
],
[
0.491943359375,
0.89794921875,
0.5087890625,
0.89794921875,
0.5087890625,
0.9111328125,
0.491943359375,
0.9111328125
]
] |
[
"columns corresponding to context words or phrases, and also involve tensorization.) Then word vectors are",
"obtained by low-rank SvD on this matrix, or a related matrix with term reweightings.In particular, the",
"PMI matrix is found to be closely approximated by a low rank matrix: there exist word vectors in say 300",
"dimensions, which is much smaller than the number of words in the dictionary, such that",
"(uaw,Uw\")~PMI(w,w\")",
"(1.1)",
"where ~ should be interpreted loosely.",
"There appears to be no theoretical explanation for this empirical finding about the approximate low",
"rank of the PMI matrix. The current paper addresses this. Specifically, we propose a probabilistic model",
"of text generation that augments the log-linear topic model of Mnih and Hinton (2007) with dynamics, in",
"the form of a random walk over a latent discourse space. The chief methodological contribution is using the",
"model priors to analytically derive a closed-form expression that directly explains (1.1): see Theorem 2.2",
"in Section 2. Section 3 builds on this insight to give a rigorous justification for models such as word2vec",
"and GloVe, including the hyperparameter choices for the latter. The insight also leads to a mathematical",
"explanation for why these word embeddings allow analogies to be solved using linear algebra; see Section 4.",
"Section 5 shows good empirical fit to this model's assumtions and predictions, including the surprising one",
"that word vectors are pretty uniformly distributed (isotropic) in space.1",
"1.1 Related work",
"Latent variable probabilistic models of language have been used for word embeddings before, including",
"Latent Dirichlet Allocation(LDA)and itsmore complicated variants (seethesurvey(Blei,2012)),and some",
"neurally inspired nonlinear models (Mnih and Hinton, 2007; Maas et al., 2011). In fact, LDA evolved out of",
"efforts in the 1990s to provide a generative model that “explains\" the success of older vector space methods",
"like Latent Semantic Indexing (Papadimitriou et al., 1998; Hofmann, 1999).However, none of these earlier",
"generative models has been linked to PMI models.",
"Levy and Goldberg (2014b) tried to relate word2vec to PMI models. They showed that if there were no",
"dimension constraint in word2vec, specifically, the “skip-gram with negative sampling (SGNS)\" version of the",
"model, then its solutions would satisfy (1.1), provided the right hand side were replaced by PMI(w, w')-β",
"for some scalar β. However, skip-gram is a discriminative model (due to the use of negative sampling), not",
"generative. Furthermore, their argument only applies to very high-dimensional word embeddings, and thus",
"does not address low-dimensional embeddings, which have superior quality in applications.",
"Hashimoto et al. (2016) focuses on issues similar to our paper. They model text generation as a random",
"walk on words, which are assumed to be embedded as vectors in a geometric space. Given that the last word",
"produced was w, the probability that the next word is w'is assumed to be given by h(lu-Uw/2) for a",
"suitable function h, and this model leads to an explanation of (1.1). By contrast our random walk involves",
"a latent discourse vector, which has a clearer semantic interpretation and has proven useful in subsequent",
"work, e.g. understanding structure of word embeddings for polysemous words Arora et al. (2016). Also our",
"work clarifies some weighting and bias terms in the training objectives of previous methods (Section 3) and",
"also the phenomenon discussed in the next paragraph.",
"Researchers have tried to understand why vectors obtained from the highly nonlinear word2vec models",
"exhibit linear structures (Levy and Goldberg, 2014a; Pennington et al., 2014). Specifically, for analogies like",
"\"man:woman:king:??,\" queen happens to be the word whose vector Uqueen is the most similar to the vector",
"Vking",
" Uman + Vwoman. This suggests that simple semantic relationships, such as masculine vs feminine",
"tested in the above example, correspond approximately to a single direction in space, a phenomenon we will",
"henceforthrefertoasRELATIONS=LINES.",
"Section 4 surveys earlier attempts to explain this phenomenon and their shortcoming, namely, that they",
"ignore the large approximation error in relationships like (1.1). This error appears larger than the difference",
"between the best solution and the second best (incorrect) solution in analogy solving, so that this error could",
"in principle lead to a complete failure in analogy solving. In our explanation, the low dimensionality of the",
"1The code is available at https://github.com/PrincetonML/SemanticVector",
"2"
] |
{
"arxiv_id": "1502.03520",
"language": "en",
"page_idx": 1
}
|
||
1502.03520-en-2
|
1502.03520-en-2.jpg
|
word vectors plays a key role. This can also be seen as a theoretical explanation of the old observation that dimension reduction improves the quality of word embeddings for various tasks. The intuitive explanation often given —that smaller models generalize better—turns out to be fallacious, since the training method for creating embeddings makes no reference to analogy solving. Thus there is no a priori reason why lowdimensional model parameters (i.e., lower model capacity) should lead to better performance in analogy solving, just as there is no reason they are better at some other unrelated task like predicting the weather.
1.2 Benefits of generative approaches
In addition to giving some form of “unification” of existing methods, our generative model also brings more intepretability to word embeddings beyond traditional cosine similarity and even analogy solving. For example, it led to an understanding of how the different senses of a polysemous word (e.g., bank) reside in linear superposition within the word embedding (Arora et al., 2016). Such insight into embeddings may prove useful in the numerous settings in NLP and neuroscience where they are used.
Another new explanatory feature of our model is that low dimensionality of word embeddings plays a key theoretical role -unlike in previous papers where the model is agnostic about the dimension of the embeddings, and the superiority of low-dimensional embeddings is an empirical fnding (starting with Deerwesten et al. (1990)). Specifically, our theoretical analysis makes the key assumption that the set of all word vectors (which are latent variables of the generative model) are spatially isotropic, which means that they have no preferred direction in space. Having $n$ vectors be isotropic in $d$ dimensions requires $d\ll n.$ This isotropy is needed in the calculations (i.e., multidimensional integral) that yield (1.1). It also holds empirically for our word vectors, as shown in Section 5.
The isotropy of low-dimensional word vectors also plays a key role in our explanation of the RELATIONS=LINES phenomenon (Section 4). The isotropy has a“purification” effect that mitigates the effect of the (rather large) approximation error in the PMI models.
2 Generative model and its properties
The model treats corpus generation as a dynamic process, where the $t$-th word is produced at step $t.$ The process is driven by the random walk of a discourse vector $c_t\in\Re^d.$ Its coordinates represent what is being $\bar{\text{talked about.}^{2}}$Each word has a (time-invariant) latent vector $v_w\in\Re^d$ that captures its correlations with the discourse vector. We model this bias with a log-linear word production model:
$$\Pr[w\text{ emitted at time }t\mid c_t]\propto\exp(\langle c_t,v_w\rangle).$$
(2.1)
The discourse vector $c_t$ does a slow random walk (meaning that $c_t+1$ is obtained from $c_t$ by adding a small random displacement vector), so that nearby words are generated under similar discourses. We are interested in the probabilities that word pairs co-occur near each other, so occasional big jumps in the random walk are allowed because they have negligible effect on these probabilities.
A similar log-linear model appears in Mnih and Hinton (2007) but without the random walk. The linean chain CRF of Collobert and Weston (2008b) is more general. The dynamic topic model of Blei and Lafferty (2006) utilizes topic dynamics, but with a linear word production model. Belanger and Kakade (2015) have proposed a dynamic model for text using Kalman Filters, where the sequence of words is generated from Gaussian linear dynamical systems, rather than the log-linear model in our case.
The novelty here over such past works is a theoretical analysis in the method-of-moments tradition (Hsu et al., 2012; Cohen et al., 2012). Assuming a prior on the random walk we analytically integrate out the hidden random variables and compute a simple closed form expression that approximately connects the model parameters to the observable joint probabilities (see Theorem 2.2). This is reminiscent of analysis of similar random walk models in fnance (Black and Scholes, 1973).
$^{2}$This is a different interpretation of the term“discourse”compared to some other settings in computational linguistics
3
|
[
[
0.1168212890625,
0.091552734375,
0.88232421875,
0.091552734375,
0.88232421875,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.1168212890625,
0.10797119140625,
0.88232421875,
0.10797119140625,
0.88232421875,
0.121826171875,
0.1168212890625,
0.121826171875
],
[
0.1168212890625,
0.125,
0.88134765625,
0.125,
0.88134765625,
0.1351318359375,
0.1168212890625,
0.1351318359375
],
[
0.11517333984375,
0.1383056640625,
0.88232421875,
0.1383056640625,
0.88232421875,
0.1514892578125,
0.11517333984375,
0.1514892578125
],
[
0.11358642578125,
0.1514892578125,
0.88232421875,
0.15283203125,
0.88232421875,
0.1673583984375,
0.11358642578125,
0.1666259765625
],
[
0.11358642578125,
0.1685791015625,
0.87841796875,
0.1685791015625,
0.87841796875,
0.1817626953125,
0.11358642578125,
0.1817626953125
],
[
0.11517333984375,
0.2020263671875,
0.488525390625,
0.2020263671875,
0.488525390625,
0.216552734375,
0.11517333984375,
0.216552734375
],
[
0.11358642578125,
0.2259521484375,
0.88232421875,
0.2259521484375,
0.88232421875,
0.240478515625,
0.11358642578125,
0.240478515625
],
[
0.11517333984375,
0.2412109375,
0.88232421875,
0.2412109375,
0.88232421875,
0.255126953125,
0.11517333984375,
0.255126953125
],
[
0.11358642578125,
0.2568359375,
0.88232421875,
0.2568359375,
0.88232421875,
0.270751953125,
0.11358642578125,
0.270751953125
],
[
0.11358642578125,
0.271484375,
0.88232421875,
0.271484375,
0.88232421875,
0.285888671875,
0.11358642578125,
0.285888671875
],
[
0.11517333984375,
0.287353515625,
0.72119140625,
0.287353515625,
0.72119140625,
0.300537109375,
0.11517333984375,
0.300537109375
],
[
0.1405029296875,
0.3017578125,
0.88134765625,
0.3017578125,
0.88134765625,
0.31494140625,
0.1405029296875,
0.31494140625
],
[
0.11517333984375,
0.317626953125,
0.88134765625,
0.317626953125,
0.88134765625,
0.330810546875,
0.11517333984375,
0.330810546875
],
[
0.11517333984375,
0.33203125,
0.88134765625,
0.33203125,
0.88134765625,
0.34521484375,
0.11517333984375,
0.34521484375
],
[
0.11517333984375,
0.3466796875,
0.88232421875,
0.3466796875,
0.88232421875,
0.35986328125,
0.11517333984375,
0.35986328125
],
[
0.11767578125,
0.3623046875,
0.88232421875,
0.3623046875,
0.88232421875,
0.375732421875,
0.11767578125,
0.375732421875
],
[
0.11517333984375,
0.376953125,
0.88232421875,
0.376953125,
0.88232421875,
0.39013671875,
0.11517333984375,
0.39013671875
],
[
0.11517333984375,
0.392578125,
0.88232421875,
0.392578125,
0.88232421875,
0.4072265625,
0.11517333984375,
0.4072265625
],
[
0.11517333984375,
0.40771484375,
0.378173828125,
0.40771484375,
0.378173828125,
0.421630859375,
0.11517333984375,
0.421630859375
],
[
0.1396484375,
0.42236328125,
0.8798828125,
0.42236328125,
0.8798828125,
0.436279296875,
0.1396484375,
0.436279296875
],
[
0.11517333984375,
0.438232421875,
0.8857421875,
0.438232421875,
0.8857421875,
0.451904296875,
0.11517333984375,
0.451904296875
],
[
0.11517333984375,
0.45263671875,
0.5400390625,
0.45263671875,
0.5400390625,
0.46728515625,
0.11517333984375,
0.46728515625
],
[
0.1168212890625,
0.4912109375,
0.1380615234375,
0.4912109375,
0.1380615234375,
0.505859375,
0.1168212890625,
0.505859375
],
[
0.15283203125,
0.4892578125,
0.57666015625,
0.490478515625,
0.57666015625,
0.50830078125,
0.15283203125,
0.5068359375
],
[
0.1168212890625,
0.521484375,
0.88232421875,
0.521484375,
0.88232421875,
0.53515625,
0.1168212890625,
0.53515625
],
[
0.11517333984375,
0.537109375,
0.88134765625,
0.537109375,
0.88134765625,
0.55029296875,
0.11517333984375,
0.55029296875
],
[
0.11517333984375,
0.5517578125,
0.88232421875,
0.5517578125,
0.88232421875,
0.56494140625,
0.11517333984375,
0.56494140625
],
[
0.11517333984375,
0.5673828125,
0.708984375,
0.5673828125,
0.708984375,
0.58056640625,
0.11517333984375,
0.58056640625
],
[
0.340576171875,
0.59326171875,
0.654296875,
0.5947265625,
0.654296875,
0.609375,
0.340576171875,
0.60791015625
],
[
0.84716796875,
0.59326171875,
0.8837890625,
0.59326171875,
0.8837890625,
0.60986328125,
0.84716796875,
0.60986328125
],
[
0.1396484375,
0.62060546875,
0.88232421875,
0.6220703125,
0.88232421875,
0.63623046875,
0.1396484375,
0.63525390625
],
[
0.1168212890625,
0.63818359375,
0.88232421875,
0.63818359375,
0.88232421875,
0.6513671875,
0.1168212890625,
0.6513671875
],
[
0.11358642578125,
0.6513671875,
0.88232421875,
0.65087890625,
0.88232421875,
0.66552734375,
0.11358642578125,
0.66650390625
],
[
0.11517333984375,
0.66748046875,
0.60888671875,
0.66748046875,
0.60888671875,
0.681640625,
0.11517333984375,
0.681640625
],
[
0.1405029296875,
0.68115234375,
0.88134765625,
0.68115234375,
0.88134765625,
0.69580078125,
0.1405029296875,
0.69580078125
],
[
0.1168212890625,
0.69775390625,
0.88134765625,
0.69775390625,
0.88134765625,
0.7109375,
0.1168212890625,
0.7109375
],
[
0.11767578125,
0.71337890625,
0.88232421875,
0.71337890625,
0.88232421875,
0.7265625,
0.11767578125,
0.7265625
],
[
0.1168212890625,
0.72802734375,
0.88232421875,
0.72802734375,
0.88232421875,
0.7421875,
0.1168212890625,
0.7421875
],
[
0.11517333984375,
0.74365234375,
0.69140625,
0.74365234375,
0.69140625,
0.7568359375,
0.11517333984375,
0.7568359375
],
[
0.1396484375,
0.7568359375,
0.88134765625,
0.75830078125,
0.88134765625,
0.77294921875,
0.1396484375,
0.771484375
],
[
0.11517333984375,
0.7734375,
0.88232421875,
0.7734375,
0.88232421875,
0.787109375,
0.11517333984375,
0.787109375
],
[
0.11358642578125,
0.787109375,
0.88232421875,
0.78857421875,
0.88232421875,
0.80322265625,
0.11358642578125,
0.8017578125
],
[
0.11358642578125,
0.80322265625,
0.8837890625,
0.8037109375,
0.8837890625,
0.818359375,
0.11358642578125,
0.8173828125
],
[
0.11517333984375,
0.818359375,
0.58837890625,
0.818359375,
0.58837890625,
0.83203125,
0.11517333984375,
0.83203125
],
[
0.13232421875,
0.84033203125,
0.861328125,
0.84228515625,
0.861328125,
0.85693359375,
0.13232421875,
0.85498046875
],
[
0.491943359375,
0.89794921875,
0.5087890625,
0.89794921875,
0.5087890625,
0.9111328125,
0.491943359375,
0.9111328125
]
] |
[
"word vectors plays a key role. This can also be seen as a theoretical explanation of the old observation that",
"dimension reduction improves the quality of word embeddings for various tasks. The intuitive explanation",
"often giventhatsmaller models generalize betterturns out to be fallacious,since the training method",
"for creating embeddings makes no reference to analogy solving. Thus there is no a priori reason why low-",
"dimensional model parameters (i.e., lower model capacity) should lead to better performance in analogy",
"solving, just as there is no reason they are better at some other unrelated task like predicting the weather.",
"1.2Benefits of generative approaches",
"In addition to giving some form of “unification\" of existing methods, our generative model also brings",
"more intepretability to word embeddings beyond traditional cosine similarity and even analogy solving. For",
"example, it led to an understanding of how the different senses of a polysemous word (e.g., bank) reside in",
"linear superposition within the word embedding (Arora et al., 2016). Such insight into embeddings may",
"prove useful in the numerous settings in NLP and neuroscience where they are used.",
"Another new explanatory feature of our model is that low dimensionality of word embeddings plays a key",
"theoretical role -unlike in previous papers where the model is agnostic about the dimension of the embed-",
"dings, and the superiority of low-dimensional embeddings is an empirical finding (starting with Deerwester",
"et al. (1990)). Specifically, our theoretical analysis makes the key assumption that the set of all word vectors",
"(which are latent variables of the generative model) are spatially isotropic, which means that they have no",
"preferred direction in space. Having n vectors be isotropic in d dimensions requires d < n. This isotropy is",
"needed in the calculations (i.e., multidimensional integral) that yield (1.1). It also holds empirically for our",
"word vectors, as shown in Section 5.",
"The isotropy of low-dimensional word vectors also plays a key role in our explanation of the RELA-",
"TIoNs-LINEs phenomenon (Section 4). The isotropy has a “purification\" effect that mitigates the effect of",
"the (rather large) approximation error in the PMI models.",
"2",
"Generative model and its properties",
"The model treats corpus generation as a dynamic process, where the t-th word is produced at step t. The",
"process is driven by the random walk of a discourse vector Ct E d. Its coordinates represent what is being",
"talked about.2 Each word has a (time-invariant) latent vector vw E d that captures its correlations with",
"the discourse vector. We model this bias with a log-linear word production model:",
"Pr[w emitted at time t / ct] α exp((ct, Uw)).",
"(2.1)",
"The discourse vector Ct does a slow random walk (meaning that ct+i is obtained from ct by adding a small",
"random displacement vector), so that nearby words are generated under similar discourses. We are interested",
"in the probabilities that word pairs co-occur near each other, so occasional big jumps in the random walk",
"are allowed because they have negligible effect on these probabilities.",
"A similar log-linear model appears in Mnih and Hinton (2007) but without the random walk. The linear",
"chain CRF of Collobert and Weston (2oo8b) is more general. The dynamic topic model of Blei and Lafferty",
"(2006) utilizes topic dynamics, but with a linear word production model. Belanger and Kakade (2015) have",
"proposed a dynamic model for text using Kalman Filters, where the sequence of words is generated from",
"Gaussian linear dynamical systems, rather than the log-linear model in our case.",
"The novelty here over such past works is a theoretical analysis in the method-of-moments tradition (Hsu",
"et al., 2012; Cohen et al., 2012). Assuming a prior on the random walk we analytically integrate out the",
"hidden random variables and compute a simple closed form expression that approximately connects the",
"model parameters to the observable joint probabilities (see Theorem 2.2). This is reminiscent of analysis of",
"similar random walk models in finance (Black and Scholes, 1973).",
"2This is a different interpretation of the term \"discourse\" compared to some other settings in computational linguistics.",
"3"
] |
{
"arxiv_id": "1502.03520",
"language": "en",
"page_idx": 2
}
|
||
1502.03520-en-3
|
1502.03520-en-3.jpg
|
Model details. Let $n$ denote the number of words and $d$ denote the dimension of the discourse space, where $1\leq d\leq n.$ Inspecting (2.1) suggests word vectors need to have varying lengths, to fit the empirical fnding that word probabilities satisfy a power law. Furthermore, we will assume that in the bulk, the word vectors are distributed uniformly in space, earlier referred to as isotropy. This can be quantified as a prior in the Bayesian tradition. More precisely, the ensemble of word vectors consists of i.i.d draws generated by $v=s\cdot\hat{v}$, where $\hat{v}$ is from the spherical Gaussian distribution, and $s$ is a scalar random variable. We assume $s$ is a random scalar with expectation $\tau=\Theta(1)$ and $s$ is always upper bounded by $\kappa$, which is another constant. Here $\tau$ governs the expected magnitude of $\langle v,c_t\rangle$, and it is particularly important to choose it to be $\Theta(1)$ so that the distribution Pr$[w|c_t]\propto\exp(\langle v_w,c_t\rangle)\text{ is interesting.}^3$ Moreover, the dynamic range of word probabilities will roughly equal $\exp(\kappa^2)$, so one should think of $\kappa$ as an absolute constant like 5. These details about $s$ are important for realistic modeling but not too important in our analysis. (Furthermore, readers uncomfortable with this simplistic Bayesian prior should look at Section 2.2 below.)
Finally, we clarify the nature of the random walk. We assume that the stationary distribution of the random walk is uniform over the unit sphere, denoted by $\mathcal{C}.$ The transition kernel of the random walk candom walk can be in any form so long as at each step the movement of the discourse vector is at most $\epsilon_2/\sqrt{d}$ in $\ell_2$ norm.$^4$ This is still fast enough to let the walk mix quickly in the space.
The following lemma (whose proof appears in the appendix) is central to the analysis. It says that under the Bayesian prior, the partition function $Z_c=\sum_w\exp(\langle v_w,c\rangle)$, which is the implied normalization in equation (2.1), is close to some constant $Z$ for most of the discourses $c.$ This can be seen as a plausible theoretical explanation of a phenomenon called $self-normalization$ in log-linear models: ignoring the partition function or treating it as a constant (which greatly simplifies training) is known to often give good results. This has also been studied in (Andreas and Klein, 2014).
$\textbf{Lemma 2. 1 ( Concentration of partition functions) . }If\textit{the word vectors satisfy the Bayesian prior described}$
$in\textit{ the model details, then}$
(2.2)
$$\Pr_{c\sim\mathcal{C}}\left[(1-\epsilon_z)Z\leq Z_c\leq(1+\epsilon_z)Z\right]\geq1-\delta,$$
for $\epsilon _{z}= \widetilde{O} ( 1/ \sqrt {n}) $, and $\delta = \exp ( - \Omega ( \log ^{2}n) ) .$
The concentration of the partition functions then leads to our main theorem (the proof is in the appendix). The theorem gives simple closed form approximations for $p(w)$, the probability of word $w$ in the corpus, and $p(w,w^{\prime})$, the probability that two words $w,w^\prime$ occur next to each other. The theorem states the result for the window size $q=2$, but the same analysis works for pairs that appear in a small window, say of size 10, as stated in Corollary 2.3. Recall that PMI$(w,w^{\prime})=\log[p(w,w^{\prime})/(p(w)p(w^{\prime}))].$
$\textbf{Theorem 2. 2. Suppose the word vectors satisfy the inequality ( 2. 2) , and window size q= 2. Then, }$
(2.3)
$$\begin{aligned}\log p(w,w^{\prime})&=\frac{\|v_{w}+v_{w^{\prime}}\|_{2}^{2}}{2d}-2\log Z\pm\epsilon,\\\log p(w)&=\frac{\|v_{w}\|_{2}^{2}}{2d}-\log Z\pm\epsilon.\end{aligned}$$
(2.4)
for $\epsilon = O( \epsilon _{z}) + \widetilde{O} ( 1/ d) + O( \epsilon _{2}) .$ Jointly these $imply:$
$$PMI\:(w,w')=\frac{\langle v_w,v_{w'}\rangle}{d}\pm O(\epsilon).$$
(2.5)
$\textbf{Remarks 1. Since the word vectors have }\ell _2$ norm of the order of $\sqrt d$, for two typical word vectors $v_w,v_w^{\prime}$, $\|v_w+v_{w^{\prime}}\|_2^2$ is of the order of $\Theta(d).$ Therefore the noise level $\epsilon$ is very small compared to the leading
$^{3}$A larger $\tau$ will make Pr$[w|c_t]$ too peaked and a smaller one will make it too uniform.
4 More precisely, the proof extends to any symmetric product stationary distribution $\mathcal{C}$ with sub-Gaussian coordinate
satisfying $\mathbb{E}_c\left[\|c\|^2\right]=1$, and the steps are such that for all $\dot c_t,\mathbb{E}_{p(c_{t+1}|c_{t})}[\exp(\kappa\sqrt{d}\|c_{t+1}-c_{t}\|)]\leq1+\epsilon_{2}$ for some small $\epsilon_2.$
4
|
[
[
0.1168212890625,
0.091552734375,
0.8798828125,
0.091552734375,
0.8798828125,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.1168212890625,
0.10797119140625,
0.88134765625,
0.10797119140625,
0.88134765625,
0.121826171875,
0.1168212890625,
0.121826171875
],
[
0.11517333984375,
0.12249755859375,
0.88232421875,
0.12249755859375,
0.88232421875,
0.1363525390625,
0.11517333984375,
0.1363525390625
],
[
0.11517333984375,
0.136962890625,
0.88232421875,
0.136962890625,
0.88232421875,
0.15087890625,
0.11517333984375,
0.15087890625
],
[
0.11517333984375,
0.15283203125,
0.88232421875,
0.15283203125,
0.88232421875,
0.1666259765625,
0.11517333984375,
0.1666259765625
],
[
0.11517333984375,
0.1673583984375,
0.88232421875,
0.1673583984375,
0.88232421875,
0.18115234375,
0.11517333984375,
0.18115234375
],
[
0.11358642578125,
0.18310546875,
0.88232421875,
0.18310546875,
0.88232421875,
0.1962890625,
0.11358642578125,
0.1962890625
],
[
0.11517333984375,
0.1976318359375,
0.88232421875,
0.1976318359375,
0.88232421875,
0.2108154296875,
0.11517333984375,
0.2108154296875
],
[
0.11517333984375,
0.21337890625,
0.8837890625,
0.21337890625,
0.8837890625,
0.2279052734375,
0.11517333984375,
0.2279052734375
],
[
0.1168212890625,
0.2279052734375,
0.8837890625,
0.2279052734375,
0.8837890625,
0.2412109375,
0.1168212890625,
0.2412109375
],
[
0.11517333984375,
0.24365234375,
0.88232421875,
0.24365234375,
0.88232421875,
0.2568359375,
0.11517333984375,
0.2568359375
],
[
0.11273193359375,
0.2568359375,
0.77294921875,
0.25830078125,
0.77294921875,
0.272705078125,
0.11273193359375,
0.271484375
],
[
0.1405029296875,
0.2734375,
0.88232421875,
0.2734375,
0.88232421875,
0.287353515625,
0.1405029296875,
0.287353515625
],
[
0.11358642578125,
0.28857421875,
0.88232421875,
0.287353515625,
0.88232421875,
0.3017578125,
0.11358642578125,
0.302978515625
],
[
0.11517333984375,
0.3037109375,
0.88134765625,
0.3037109375,
0.88134765625,
0.318115234375,
0.11517333984375,
0.318115234375
],
[
0.11358642578125,
0.317626953125,
0.576171875,
0.318115234375,
0.576171875,
0.333251953125,
0.11358642578125,
0.33203125
],
[
0.1405029296875,
0.33203125,
0.88134765625,
0.33203125,
0.88134765625,
0.3466796875,
0.1405029296875,
0.3466796875
],
[
0.1168212890625,
0.349853515625,
0.50244140625,
0.349853515625,
0.50244140625,
0.363037109375,
0.1168212890625,
0.363037109375
],
[
0.498291015625,
0.354248046875,
0.5087890625,
0.354248046875,
0.5087890625,
0.35791015625,
0.498291015625,
0.35791015625
],
[
0.5048828125,
0.349853515625,
0.88232421875,
0.349853515625,
0.88232421875,
0.363037109375,
0.5048828125,
0.363037109375
],
[
0.11358642578125,
0.3623046875,
0.88232421875,
0.363037109375,
0.88232421875,
0.37744140625,
0.11358642578125,
0.376953125
],
[
0.11517333984375,
0.380126953125,
0.88232421875,
0.380126953125,
0.88232421875,
0.393310546875,
0.11517333984375,
0.393310546875
],
[
0.1168212890625,
0.39453125,
0.88134765625,
0.39453125,
0.88134765625,
0.40771484375,
0.1168212890625,
0.40771484375
],
[
0.11517333984375,
0.4091796875,
0.52294921875,
0.4091796875,
0.52294921875,
0.423583984375,
0.11517333984375,
0.423583984375
],
[
0.11517333984375,
0.434326171875,
0.8837890625,
0.434326171875,
0.8837890625,
0.447509765625,
0.11517333984375,
0.447509765625
],
[
0.11517333984375,
0.449462890625,
0.301513671875,
0.449462890625,
0.301513671875,
0.46337890625,
0.11517333984375,
0.46337890625
],
[
0.3544921875,
0.46533203125,
0.64794921875,
0.46337890625,
0.64794921875,
0.477783203125,
0.3544921875,
0.479736328125
],
[
0.84619140625,
0.464111328125,
0.8837890625,
0.464111328125,
0.8837890625,
0.479736328125,
0.84619140625,
0.479736328125
],
[
0.11517333984375,
0.494384765625,
0.43701171875,
0.494384765625,
0.43701171875,
0.5087890625,
0.11517333984375,
0.5087890625
],
[
0.1405029296875,
0.52099609375,
0.8798828125,
0.52099609375,
0.8798828125,
0.5341796875,
0.1405029296875,
0.5341796875
],
[
0.11517333984375,
0.5341796875,
0.88134765625,
0.53515625,
0.88134765625,
0.5498046875,
0.11517333984375,
0.548828125
],
[
0.11358642578125,
0.55029296875,
0.88134765625,
0.5498046875,
0.88134765625,
0.564453125,
0.11358642578125,
0.56494140625
],
[
0.1168212890625,
0.56640625,
0.88134765625,
0.56640625,
0.88134765625,
0.580078125,
0.1168212890625,
0.580078125
],
[
0.11517333984375,
0.5791015625,
0.6806640625,
0.580078125,
0.6806640625,
0.5947265625,
0.11517333984375,
0.59326171875
],
[
0.1168212890625,
0.60498046875,
0.82861328125,
0.60498046875,
0.82861328125,
0.619140625,
0.1168212890625,
0.619140625
],
[
0.346435546875,
0.63720703125,
0.55126953125,
0.625,
0.5537109375,
0.6484375,
0.348876953125,
0.6611328125
],
[
0.54833984375,
0.642578125,
0.6435546875,
0.642578125,
0.6435546875,
0.65380859375,
0.54833984375,
0.65380859375
],
[
0.84619140625,
0.64013671875,
0.8857421875,
0.64013671875,
0.8857421875,
0.6572265625,
0.84619140625,
0.6572265625
],
[
0.490234375,
0.6494140625,
0.51318359375,
0.6494140625,
0.51318359375,
0.66357421875,
0.490234375,
0.66357421875
],
[
0.451904296875,
0.662109375,
0.5087890625,
0.662109375,
0.5087890625,
0.68505859375,
0.451904296875,
0.68505859375
],
[
0.375732421875,
0.67529296875,
0.45751953125,
0.673828125,
0.45751953125,
0.689453125,
0.375732421875,
0.69140625
],
[
0.50634765625,
0.6767578125,
0.5908203125,
0.6767578125,
0.5908203125,
0.68798828125,
0.50634765625,
0.68798828125
],
[
0.84716796875,
0.6748046875,
0.8837890625,
0.6748046875,
0.8837890625,
0.68994140625,
0.84716796875,
0.68994140625
],
[
0.468017578125,
0.68505859375,
0.494384765625,
0.68505859375,
0.494384765625,
0.69775390625,
0.468017578125,
0.69775390625
],
[
0.11517333984375,
0.708984375,
0.49267578125,
0.71044921875,
0.49267578125,
0.724609375,
0.11517333984375,
0.7236328125
],
[
0.378173828125,
0.72998046875,
0.62109375,
0.73193359375,
0.62109375,
0.76513671875,
0.378173828125,
0.76318359375
],
[
0.84716796875,
0.744140625,
0.8837890625,
0.744140625,
0.8837890625,
0.76123046875,
0.84716796875,
0.76123046875
],
[
0.5185546875,
0.755859375,
0.53173828125,
0.755859375,
0.53173828125,
0.7666015625,
0.5185546875,
0.7666015625
],
[
0.11517333984375,
0.78515625,
0.88134765625,
0.7861328125,
0.88134765625,
0.80029296875,
0.11517333984375,
0.7998046875
],
[
0.11767578125,
0.8017578125,
0.8837890625,
0.8017578125,
0.8837890625,
0.81640625,
0.11767578125,
0.81640625
],
[
0.134033203125,
0.82568359375,
0.65625,
0.82568359375,
0.65625,
0.8388671875,
0.134033203125,
0.8388671875
],
[
0.134033203125,
0.8369140625,
0.88232421875,
0.8369140625,
0.88232421875,
0.85107421875,
0.134033203125,
0.85107421875
],
[
0.1094970703125,
0.8466796875,
0.87109375,
0.84765625,
0.87109375,
0.86865234375,
0.1094970703125,
0.8671875
],
[
0.494384765625,
0.8984375,
0.50732421875,
0.8984375,
0.50732421875,
0.91015625,
0.494384765625,
0.91015625
]
] |
[
"Model details.Let n denote the number of words and d denote the dimension of the discourse space",
"where 1 ≤ d ≤ n. Inspecting (2.1) suggests word vectors need to have varying lengths, to fit the empirical",
"finding that word probabilities satisfy a power law. Furthermore, we will assume that in the bulk, the word",
"vectors are distributed uniformly in space, earlier referred to as isotropy. This can be quantified as a prior",
"in the Bayesian tradition. More precisely, the ensemble of word vectors consists of i.i.d draws generated by",
"u = s , where is from the spherical Gaussian distribution, and s is a scalar random variable. We assume",
"s is a random scalar with expectation T = (1) and s is always upper bounded by K, which is another",
"constant. Here T governs the expected magnitude of (u,ct), and it is particularly important to choose it to",
"be (1) so that the distribution Pr[w|ct] α exp((uw, ct)) is interesting.3 Moreover, the dynamic range of",
"word probabilities will roughly equal exp(k2), so one should think of k as an absolute constant like 5. These",
"details about s are important for realistic modeling but not too important in our analysis. (Furthermore,",
"readers uncomfortable with this simplistic Bayesian prior should look at Section 2.2below.)",
"Finally, we clarify the nature of the random walk. We assume that the stationary distribution of the",
"random walk is uniform over the unit sphere, denoted by C. The transition kernel of the random walk can",
"be in any form so long as at each step the movement of the discourse vector is at most e2/Vd in l2 norm.4",
"This is still fast enough to let the walk mix quickly in the space.",
"The following lemma (whose proof appears in the appendix) is central to the analysis. It says that",
"under the Bayesian prior, the partition function Ze=",
"=",
"exp((uw,c)),which istheimpliednormalization",
"in equation (2.1), is close to some constant z for most of the discourses c. This can be seen as a plausible",
"theoretical explanation of a phenomenon called self-normalization in log-linear models: ignoring the partition",
"function or treating it as a constant (which greatly simplifies training) is known to often give good results.",
"This has also been studied in (Andreas and Klein, 2014).",
"Lemma 2.1 (Concentration of partition functions). If the word vectors satisfy the Bayesian prior described",
"in themodel details, then",
"Pr [(1-)z≤ z ≤(1+z)zl≥ 1-8,",
"(2.2)",
"for Ez = O(1/ /n), and 8 = exp(-2(log² n).",
"The concentration of the partition functions then leads to our main theorem (the proof is in the appendix)",
"The theorem gives simple closed form approximations for p(w), the probability of word w in the corpus, and",
"p(w,w'), the probability that two words w, w' occur next to each other. The theorem states the result for",
"the window size q = 2, but the same analysis works for pairs that appear in a small window, say of size 10.",
"as stated in Corollary 2.3. Recall that PMI(w, w') = log[p(w, w')/(p(w)p(w'))).",
"Theorem 2.2. Suppose the word vectors satisfy the inequality (2.2), and window size q = 2. Then,",
"log p(w, wl) = Ilow + wll",
"-2logZ±E.",
"(2.3)",
"2d",
"Iowl/",
"log p(w) =",
"-log Z±e.",
"(2.4)",
"2d",
"for =O(e) +O(1/d)+O(2): Jointly these imply:",
"PMI (w ) = m) +0().",
"(2.5)",
"d",
"Remarks 1. Since the word vectors have l2 norm of the order of Va, for two typical word vectors Uw, Vu',",
"Iluu + uwll2 is of the order of O(d). Therefore the noise level e is very small compared to the leading",
"3A larger T will make Pr[wct] too peaked and a smaller one will make it too uniform.",
"4 More precisely, the proof extends to any symmetric product stationary distribution C with sub-Gaussian coordinate",
"satisfying[l], and thestsae suchthatallct,(e+)[ex(al+l]+2frmemall2",
"4"
] |
{
"arxiv_id": "1502.03520",
"language": "en",
"page_idx": 3
}
|
||
1502.03520-en-4
|
1502.03520-en-4.jpg
|
term$\frac1{2d}\|v_w+v_{w^{\prime}}\|_2^2.$ For PMI however, the noise level $O(\epsilon)$ could be comparable to the leading term, and
empirically we also find higher error here.
Remarks 2. Variants of the expression for joint probability in (2.3) had been hypothesized based upon
empirical evidence in Mikolov et al. (2013b) and also Globerson et al. (2007), and Maron et al. (2010) .
$\textbf{Remarks 3. Theorem 2. 2 directly leads to the extension to a general window size }q$ as follows:
$\textbf{Corollary 2. 3. }Let$ $p_{q}( w, w^{\prime })$ $be$ the $co$-occurrence probability $in$ windows $of$ size $q$, and $PMI_{q}( w, w^{\prime })$ $be$ the
$corresponding~PMI~value.~Then$
$$\begin{aligned}&\log p_{q}(w,w^{\prime})=\frac{\|v_{w}+v_{w^{\prime}}\|_{2}^{2}}{2d}-2\log Z+\gamma\pm\epsilon,\\&PMI_{q}\:(w,w^{\prime})=\frac{\langle v_{w},v_{w^{\prime}}\rangle}{d}+\gamma\pm O(\epsilon).\end{aligned}$$
where $\gamma = \log \left ( \frac {q( q- 1) }2\right ) .$
It is quite easy to see that Theorem 2.2 implies the Corollary 2.3, as when the window size is $q$ the pair $w,w^\prime$ could appear in any of $\binom q2$ positions within the window, and the joint probability of $w,w^\prime$ is roughly the same for any positions because the discourse vector changes slowly. (Of course, the error term gets worse as we consider larger window sizes, although for any constant size, the statement of the theorem is correct.) This is also consistent with the shift $\beta$ for fitting PMI in (Levy and Goldberg, 2014b), which showed that without dimension constraints, the solution to skip-gram with negative sampling satisfies PMI $( w, w^{\prime }) - \beta = \langle v_{w}, v_{w^{\prime }}\rangle$ for a constant $\beta$ that is related to the negative sampling in the optimization. Our result justifies via a generative model why this should be satisfied even for low dimensional word vectors.
# 2.1 Proof sketches
Here we provide the proof sketches, while the complete proof can be found in the appendix.
$\textbf{Proof sketch of Theorem 2. 2}$ Let $w$ and $w^\prime$ be two arbitrary words. Let $c$ and $c^\prime$ denote two consecutive
context vectors, where $c\sim\mathcal{C}$ and $c^\prime|c$ is defined by the Markov kernel $p(c^{\prime}\mid c).$
We start by using the law of total expectation, integrating out the hidden variables $c$ and $c^\prime:$
$$\begin{aligned}p(w,w^{\prime})&=\mathbb{E}_{c,c^{\prime}}\left[\Pr[w,w^{\prime}|c,c^{\prime}]\right]\\&=\mathbb{E}_{c,c^{\prime}}[p(w|c)p(w^{\prime}|c^{\prime})]\\&=\mathbb{E}_{c,c^{\prime}}\left[\frac{\exp(\langle v_{w},c\rangle)}{Z_{c}}\frac{\exp(\langle v_{w^{\prime}},c^{\prime}\rangle)}{Z_{c^{\prime}}}\right]\end{aligned}$$
(2.6)
An expectation like (2.6) would normally be difficult to analyze because of the partition functions. However, we can assume the inequality (2.2), that is, the partition function typically does not vary much for most of context vectors $c.$ Let $F$ be the event that both $c$ and $c^\prime$ are within $(1\pm\epsilon_z)Z.$ Then by (2.2) and the union bound, event $F$ happens with probability at least $1-2\exp(-\Omega(\log^2n)).$ We will split the right-hand side (RHS) of (2.6) into the parts according to whether $F$ happens or not.
$$\begin{aligned}\mathrm{RHS~of~}(2.6)&=\underbrace{\mathbb{E}_{c,c^{\prime}}\left[\frac{\exp(\langle v_{w},c\rangle)}{Z_{c}}\frac{\exp(\langle v_{w^{\prime}},c^{\prime}\rangle)}{Z_{c^{\prime}}}\mathbf{1}_{F}\right]}_{T_{1}}\\&+\underbrace{\mathbb{E}_{c,c^{\prime}}\left[\frac{\exp(\langle v_{w},c\rangle)}{Z_{c}}\frac{\exp(\langle v_{w^{\prime}},c^{\prime}\rangle)}{Z_{c^{\prime}}}\mathbf{1}_{F}\right]}_{T_{2}}\end{aligned}$$
(2.7)
5
|
[
[
0.11358642578125,
0.09027099609375,
0.8837890625,
0.0889892578125,
0.8837890625,
0.106689453125,
0.11358642578125,
0.10797119140625
],
[
0.1168212890625,
0.10797119140625,
0.418212890625,
0.10797119140625,
0.418212890625,
0.121826171875,
0.1168212890625,
0.121826171875
],
[
0.1168212890625,
0.1287841796875,
0.8798828125,
0.1287841796875,
0.8798828125,
0.143310546875,
0.1168212890625,
0.143310546875
],
[
0.11517333984375,
0.1458740234375,
0.85888671875,
0.1458740234375,
0.85888671875,
0.1590576171875,
0.11517333984375,
0.1590576171875
],
[
0.11358642578125,
0.1654052734375,
0.79833984375,
0.1666259765625,
0.79833984375,
0.18115234375,
0.11358642578125,
0.179931640625
],
[
0.1168212890625,
0.1900634765625,
0.88232421875,
0.1900634765625,
0.88232421875,
0.2052001953125,
0.1168212890625,
0.2052001953125
],
[
0.1168212890625,
0.205810546875,
0.346435546875,
0.205810546875,
0.346435546875,
0.2197265625,
0.1168212890625,
0.2197265625
],
[
0.330078125,
0.2235107421875,
0.67236328125,
0.2215576171875,
0.67236328125,
0.2568359375,
0.330078125,
0.2587890625
],
[
0.4794921875,
0.2474365234375,
0.5048828125,
0.2474365234375,
0.5048828125,
0.260009765625,
0.4794921875,
0.260009765625
],
[
0.330078125,
0.266357421875,
0.5146484375,
0.25830078125,
0.515625,
0.27978515625,
0.330810546875,
0.28857421875
],
[
0.52294921875,
0.26953125,
0.6005859375,
0.26953125,
0.6005859375,
0.2841796875,
0.52294921875,
0.2841796875
],
[
0.473876953125,
0.279052734375,
0.491943359375,
0.279052734375,
0.491943359375,
0.292236328125,
0.473876953125,
0.292236328125
],
[
0.1192626953125,
0.3037109375,
0.21728515625,
0.3037109375,
0.21728515625,
0.318115234375,
0.1192626953125,
0.318115234375
],
[
0.1396484375,
0.330078125,
0.88232421875,
0.330078125,
0.88232421875,
0.3447265625,
0.1396484375,
0.3447265625
],
[
0.11517333984375,
0.3466796875,
0.8837890625,
0.3466796875,
0.8837890625,
0.35986328125,
0.11517333984375,
0.35986328125
],
[
0.1168212890625,
0.3623046875,
0.88232421875,
0.3623046875,
0.88232421875,
0.375732421875,
0.1168212890625,
0.375732421875
],
[
0.1168212890625,
0.376953125,
0.88134765625,
0.376953125,
0.88134765625,
0.39013671875,
0.1168212890625,
0.39013671875
],
[
0.1168212890625,
0.392578125,
0.88134765625,
0.392578125,
0.88134765625,
0.406005859375,
0.1168212890625,
0.406005859375
],
[
0.11767578125,
0.4072265625,
0.88232421875,
0.4072265625,
0.88232421875,
0.42041015625,
0.11767578125,
0.42041015625
],
[
0.1168212890625,
0.42236328125,
0.88134765625,
0.42236328125,
0.88134765625,
0.436279296875,
0.1168212890625,
0.436279296875
],
[
0.11358642578125,
0.436279296875,
0.88232421875,
0.4375,
0.88232421875,
0.451904296875,
0.11358642578125,
0.45068359375
],
[
0.11517333984375,
0.47021484375,
0.306396484375,
0.47021484375,
0.306396484375,
0.48486328125,
0.11517333984375,
0.48486328125
],
[
0.1168212890625,
0.49560546875,
0.771484375,
0.49560546875,
0.771484375,
0.5087890625,
0.1168212890625,
0.5087890625
],
[
0.11517333984375,
0.5263671875,
0.88232421875,
0.5263671875,
0.88232421875,
0.54052734375,
0.11517333984375,
0.54052734375
],
[
0.11358642578125,
0.54150390625,
0.67333984375,
0.54248046875,
0.67333984375,
0.556640625,
0.11358642578125,
0.55615234375
],
[
0.1396484375,
0.556640625,
0.80224609375,
0.55615234375,
0.80224609375,
0.57080078125,
0.1396484375,
0.5712890625
],
[
0.34716796875,
0.58349609375,
0.55224609375,
0.58349609375,
0.55224609375,
0.59765625,
0.34716796875,
0.59765625
],
[
0.41015625,
0.60986328125,
0.55615234375,
0.60791015625,
0.55615234375,
0.62548828125,
0.41015625,
0.62744140625
],
[
0.455810546875,
0.63623046875,
0.640625,
0.63623046875,
0.640625,
0.65380859375,
0.455810546875,
0.65380859375
],
[
0.84716796875,
0.64599609375,
0.8837890625,
0.64599609375,
0.8837890625,
0.6611328125,
0.84716796875,
0.6611328125
],
[
0.490234375,
0.65576171875,
0.51318359375,
0.65576171875,
0.51318359375,
0.66943359375,
0.490234375,
0.66943359375
],
[
0.5810546875,
0.6572265625,
0.60546875,
0.6572265625,
0.60546875,
0.66845703125,
0.5810546875,
0.66845703125
],
[
0.1405029296875,
0.67822265625,
0.8798828125,
0.67822265625,
0.8798828125,
0.6923828125,
0.1405029296875,
0.6923828125
],
[
0.11767578125,
0.6962890625,
0.88134765625,
0.6962890625,
0.88134765625,
0.70703125,
0.11767578125,
0.70703125
],
[
0.1168212890625,
0.708984375,
0.88232421875,
0.708984375,
0.88232421875,
0.7236328125,
0.1168212890625,
0.7236328125
],
[
0.11517333984375,
0.7236328125,
0.8837890625,
0.724609375,
0.8837890625,
0.7392578125,
0.11517333984375,
0.73779296875
],
[
0.11517333984375,
0.7392578125,
0.64794921875,
0.7392578125,
0.64794921875,
0.75244140625,
0.11517333984375,
0.75244140625
],
[
0.463134765625,
0.76513671875,
0.65625,
0.76513671875,
0.65625,
0.77978515625,
0.463134765625,
0.77978515625
],
[
0.3203125,
0.77294921875,
0.453369140625,
0.77294921875,
0.453369140625,
0.7861328125,
0.3203125,
0.7861328125
],
[
0.498291015625,
0.7822265625,
0.51953125,
0.7822265625,
0.51953125,
0.796875,
0.498291015625,
0.796875
],
[
0.58837890625,
0.7822265625,
0.611328125,
0.7822265625,
0.611328125,
0.796875,
0.58837890625,
0.796875
],
[
0.438720703125,
0.78857421875,
0.44775390625,
0.78857421875,
0.44775390625,
0.79248046875,
0.438720703125,
0.79248046875
],
[
0.55224609375,
0.806640625,
0.5595703125,
0.806640625,
0.5595703125,
0.81298828125,
0.55224609375,
0.81298828125
],
[
0.4599609375,
0.8193359375,
0.64892578125,
0.82080078125,
0.64892578125,
0.83544921875,
0.4599609375,
0.833984375
],
[
0.418212890625,
0.830078125,
0.451904296875,
0.830078125,
0.451904296875,
0.84130859375,
0.418212890625,
0.84130859375
],
[
0.84716796875,
0.82763671875,
0.8837890625,
0.82763671875,
0.8837890625,
0.8447265625,
0.84716796875,
0.8447265625
],
[
0.495849609375,
0.8369140625,
0.51708984375,
0.8369140625,
0.51708984375,
0.8515625,
0.495849609375,
0.8515625
],
[
0.58642578125,
0.8369140625,
0.609375,
0.8369140625,
0.609375,
0.8515625,
0.58642578125,
0.8515625
],
[
0.54150390625,
0.85791015625,
0.564453125,
0.85791015625,
0.564453125,
0.87255859375,
0.54150390625,
0.87255859375
],
[
0.4306640625,
0.84130859375,
0.453369140625,
0.84130859375,
0.453369140625,
0.85498046875,
0.4306640625,
0.85498046875
],
[
0.49267578125,
0.8984375,
0.50634765625,
0.8984375,
0.50634765625,
0.91015625,
0.49267578125,
0.91015625
]
] |
[
"term 2alluw + Uw ll2. For PMI however, the noise level O(e) could be comparable to the leading term, and",
"empirically we also find higher error here.",
"Remarks 2. Variants of the expression for joint probability in (2.3) had been hypothesized based upon",
"empirical evidence in Mikolov et al. (2013b) and also Globerson et al. (2007), and Maron et al. (2010)",
"Remarks 3. Theorem 2.2 directly leads to the extension to a general window size q as follows:",
"Corollary 2.3. Let pg(w, w') be the co-occurrence probability in windows of size q, and PMl,(w, w') be the",
"corresponding PMI value. Then",
"log p(w, ) +/ -210g + ±6",
"2d",
"PMIg (w, w\") = (ow, w.)",
"+±0(e).",
"d",
"where =log",
"It is quite easy to see that Theorem 2.2 implies the Corollary 2.3, as when the window size is q the",
"pair w, w' could appear in any of () positions within the window, and the joint probability of w, w' is",
"roughly the same for any positions because the discourse vector changes slowly.(Of course, the error",
"term gets worse as we consider larger window sizes, although for any constant size, the statement of the",
"theorem is correct.) This is also consistent with the shift β for fitting PMI in (Levy and Goldberg, 2014b),",
"which showed that without dimension constraints, the solution to skip-gram with negative sampling satisfies",
"PMI (w, w') -β = (uw, Uw\") for a constant β that is related to the negative sampling in the optimization.",
"Our result justifies via a generative model why this should be satisfied even for low dimensional word vectors.",
"2.1Proof sketches",
"Here we provide the proof sketches, while the complete proof can be found in the appendix.",
"Proof sketch of Theorem 2.2 Let w and w' be two arbitrary words. Let c and c denote two consecutive",
"context vectors, where c ~ C and clc is defined by the Markov kernel p(c' I c).",
"We start by using the law of total expectation, integrating out the hidden variables c and c':",
"p(w,w') = E [Pr[w,w'lc,c]",
"= E [p(w|c)p(w'lc)]",
"exp(uw,c)) exp((uw),c\")",
"(2.6)",
"Ze",
"Zc'",
"An expectation like (2.6) would normally be difficult to analyze because of the partition functions",
"However,we can assume the inequality(2.2)that is, the partition function typically does not vary much for",
"most of context vectors c. Let F be the event that both c and c' are within (1 ±ez)Z. Then by (2.2) and the",
"union bound, event F happens with probability at least 1 - 2exp(-2(log n)). We will split the right-hand",
"side (RHS) of (2.6) into the parts according to whether F happens or not.",
"exp((uu,c)) exp((uwt,c)),",
"RHS of (2.6) = E",
"Zc",
"Ze",
".",
"T",
"[exp((uw,c))exp((Uu*,c\"))",
"+E",
"(2.7)",
"Zc",
"Ze",
"T2",
"c,c\"",
"5"
] |
{
"arxiv_id": "1502.03520",
"language": "en",
"page_idx": 4
}
|
||
2308.10280-en-1
|
2308.10280-en-1.jpg
|
2
IEEE ROBOTICS AND AUTOMATION LETTERS. PREPRINT VERSION. ACCEPTED AUGUST, 2023
<FigureHere>
Fig.2. (a) The system overview of MacFormer.(b) The detailed implementation of MacFormer and output size of each operation.
Former, an one-stage Map-Agent Coupled TransFormer, which directly couples map constraints with agent motions for real-time and robust trajectory prediction (Fig.1).In order backbones to encode map and fuse it with agent motions to enable direct utilization of the coupled relation between in high-level latent feature. By optimizing prediction outputs, the map and agent, two modules named coupled map and
reference extractor are carefully designed, which explicitly works[2],[21] used CNNs to encode the map and agents integrate map constraints into the system. Furthermore, to together into an image, then predicted trajectories using a assure that the map constraints are well learned during training, we propose a multi-task optimization strategy (MTOS). It ensures map-constrained prediction by forecasting future torNet[1],LaneGCN[3],and DSP[8] all vectorized the map
relative motions in coupled map, capturing motion priors, and
outputting final predictions with corresponding probabilities interactions of agents and map instead. Besides, TPCN[7] via map-conditioned regression. Lastly, to enhance computa- migrated point cloud learning backbone to implement the feational efficiency and promote real-time prediction, a concise ture fusion. To cover long-range prediction, mmTransformer bilateral query scheme for context fusion is devised. It allows [6], SceneTransformer[4] and MultiPath++[5] leveraged the parallel cross-domain information interaction between the map
and agent from their distinct perspectives and shared usage of fusion. Nevertheless, $implicit-fusion$ still leads to mode coltheir affinity matrix, significantly reducing the time and space lapse or infeasible forecasts that violate map constraints, which complexity of context fusion.
The proposed method was evaluated on three large- stability.
scale real-world benchmarks (Argoverse 1&2[18],[19] and
nuScenes [20]) and achieved state-of-the-art performance with fusion models, TNT[12] and DenseTNT[13] decomposed presignificantly lower inference latency and fewer parameters. diction objectives into two stage: target prediction and motion We also conducted extensive experiments to validate the effectiveness and robustness of our framework. Furthermore, we introduced our framework to classical prediction models, enhancing their performance while reducing parameters. In Similarly, HOME [14],GOHOME [15] and THOMAS [16]
summary, the contributions of this paper are:
1) A map-agent coupled framework for real-time and robust implement target prediction before generating full trajectories. trajectory prediction, which efficiently and effectively inte- However, despite the high intent correlation of targets, such grates map constraints via coupled map, reference extractor, models still cannot effectively capture motions of the entire and multi-task optimization strategy (MTOS).
2) An efficient and lightweight context fusion scheme, bilateral query, which allows parallel context fusion between map and agent from their individual views.
$\hat{3})$ Extensive evaluation in multiple real-world benchmarks.
It shows that MacFormer achieves state-of-the-art perfor- predictions via reference extractor, considers map constraints
mance, while it is faster and more lightweight than existing during optimization via MTOS.
models. Experiments also demonstrate the versatility and ro- Transformer. Transformer [22] has been successfully applied
bustness of the proposed framework, which enhances classical
prediction methods while maintaining satisfactory resilience to puter vision. Recently, some works[4]-[6] have utilized it
imperfect upstream results.
II. RELATED WORK
ferent directions: $implicit-fusion$ and target-driven.
1) Implicit-fusion. Implicit-fusion models use different they aim to capture the map structure in such fusion. Early fully connected layer. However, this approach cannot model relations between different scenario elements effectively. Vecand adopted graph-based architectures to extract features and global receptive field of Transformer to capture whole context illustrates their unsatisfactory map utilization and prediction
2) Target-driven. To overcome the limitations in $implicit-$ completion. They frst sampled several target candidates for the agent from map, selected qualified targets from candidates, then completed full trajectory conditioned on each target. forecasted a heatmap and greedily sampled target candidates to process, while also incurring redundant costs due to large quantities of unselected candidate.
In contrast, our method tackles these problems by directly and sufficiently leveraging map, which utilizes the coupled relation by coupled map, enables efficient map-constrained in various felds, $e.g.$, natural language processing and comand its variants for trajectory prediction due to its capability of global and flexible feature extraction that are highly applicable for context interaction. To leverage this architecture, our
Map utilization. Map utilization has received much attention proposed method devises a Transformer-based context encoder recently, where existing approaches mainly fall into two dif- and proposes an effective query module in reference extractor
|
[
[
0.080078125,
0.0309295654296875,
0.0882568359375,
0.0309295654296875,
0.0882568359375,
0.0404052734375,
0.080078125,
0.0404052734375
],
[
0.418212890625,
0.0296783447265625,
0.9189453125,
0.0296783447265625,
0.9189453125,
0.0404052734375,
0.418212890625,
0.0404052734375
],
[
0.09234619140625,
0.0738525390625,
0.1707763671875,
0.07574462890625,
0.1707763671875,
0.0889892578125,
0.09234619140625,
0.08709716796875
],
[
0.1968994140625,
0.07891845703125,
0.25244140625,
0.07891845703125,
0.25244140625,
0.09027099609375,
0.1968994140625,
0.09027099609375
],
[
0.2900390625,
0.07891845703125,
0.420654296875,
0.07891845703125,
0.420654296875,
0.0889892578125,
0.2900390625,
0.0889892578125
],
[
0.438720703125,
0.0802001953125,
0.53515625,
0.0802001953125,
0.53515625,
0.09027099609375,
0.438720703125,
0.09027099609375
],
[
0.56640625,
0.080810546875,
0.59326171875,
0.080810546875,
0.59326171875,
0.0889892578125,
0.56640625,
0.0889892578125
],
[
0.10540771484375,
0.09344482421875,
0.146240234375,
0.09344482421875,
0.146240234375,
0.10101318359375,
0.10540771484375,
0.10101318359375
],
[
0.373291015625,
0.0953369140625,
0.396240234375,
0.0953369140625,
0.396240234375,
0.0997314453125,
0.373291015625,
0.0997314453125
],
[
0.61767578125,
0.09027099609375,
0.62744140625,
0.08837890625,
0.6298828125,
0.0921630859375,
0.619140625,
0.0947265625
],
[
0.11767578125,
0.1029052734375,
0.1356201171875,
0.1029052734375,
0.1356201171875,
0.1060791015625,
0.11767578125,
0.1060791015625
],
[
0.2156982421875,
0.1060791015625,
0.2344970703125,
0.1060791015625,
0.2344970703125,
0.11297607421875,
0.2156982421875,
0.11297607421875
],
[
0.47802734375,
0.10162353515625,
0.5185546875,
0.10162353515625,
0.5185546875,
0.11236572265625,
0.47802734375,
0.11236572265625
],
[
0.1070556640625,
0.11297607421875,
0.11767578125,
0.10919189453125,
0.125,
0.12060546875,
0.11358642578125,
0.125
],
[
0.455078125,
0.117431640625,
0.471435546875,
0.117431640625,
0.471435546875,
0.12060546875,
0.455078125,
0.12060546875
],
[
0.459228515625,
0.177978515625,
0.498291015625,
0.177978515625,
0.498291015625,
0.1849365234375,
0.459228515625,
0.1849365234375
],
[
0.7998046875,
0.1767578125,
0.8095703125,
0.1767578125,
0.8095703125,
0.18115234375,
0.7998046875,
0.18115234375
],
[
0.11358642578125,
0.205810546875,
0.291748046875,
0.2052001953125,
0.291748046875,
0.216552734375,
0.11358642578125,
0.2171630859375
],
[
0.5009765625,
0.20703125,
0.57275390625,
0.20703125,
0.57275390625,
0.2171630859375,
0.5009765625,
0.2171630859375
],
[
0.80224609375,
0.205810546875,
0.810546875,
0.205810546875,
0.810546875,
0.2108154296875,
0.80224609375,
0.2108154296875
],
[
0.386474609375,
0.2083740234375,
0.439453125,
0.205810546875,
0.439453125,
0.21337890625,
0.386474609375,
0.21533203125
],
[
0.36279296875,
0.210205078125,
0.38232421875,
0.210205078125,
0.38232421875,
0.2139892578125,
0.36279296875,
0.2139892578125
],
[
0.0767822265625,
0.2298583984375,
0.80224609375,
0.2298583984375,
0.80224609375,
0.24365234375,
0.0767822265625,
0.24365234375
],
[
0.07843017578125,
0.248779296875,
0.5087890625,
0.248779296875,
0.5087890625,
0.261962890625,
0.07843017578125,
0.261962890625
],
[
0.5009765625,
0.248779296875,
0.8486328125,
0.248779296875,
0.8486328125,
0.261962890625,
0.5009765625,
0.261962890625
],
[
0.07843017578125,
0.26318359375,
0.49267578125,
0.26318359375,
0.49267578125,
0.276611328125,
0.07843017578125,
0.276611328125
],
[
0.52392578125,
0.264404296875,
0.92236328125,
0.264404296875,
0.92236328125,
0.27783203125,
0.52392578125,
0.27783203125
],
[
0.07843017578125,
0.27783203125,
0.498291015625,
0.27783203125,
0.498291015625,
0.291748046875,
0.07843017578125,
0.291748046875
],
[
0.50244140625,
0.27783203125,
0.92236328125,
0.279052734375,
0.92236328125,
0.29345703125,
0.50244140625,
0.292236328125
],
[
0.080078125,
0.29541015625,
0.49267578125,
0.29541015625,
0.49267578125,
0.30615234375,
0.080078125,
0.30615234375
],
[
0.50634765625,
0.294189453125,
0.9208984375,
0.294189453125,
0.9208984375,
0.30810546875,
0.50634765625,
0.30810546875
],
[
0.07843017578125,
0.30810546875,
0.494384765625,
0.30810546875,
0.494384765625,
0.3212890625,
0.07843017578125,
0.3212890625
],
[
0.50341796875,
0.31005859375,
0.9189453125,
0.31005859375,
0.9189453125,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.07843017578125,
0.323974609375,
0.49267578125,
0.323974609375,
0.49267578125,
0.337158203125,
0.07843017578125,
0.337158203125
],
[
0.50341796875,
0.324462890625,
0.9208984375,
0.324462890625,
0.9208984375,
0.33837890625,
0.50341796875,
0.33837890625
],
[
0.0767822265625,
0.339111328125,
0.494384765625,
0.339111328125,
0.494384765625,
0.352783203125,
0.0767822265625,
0.352783203125
],
[
0.50341796875,
0.34033203125,
0.92236328125,
0.34033203125,
0.92236328125,
0.354248046875,
0.50341796875,
0.354248046875
],
[
0.0767822265625,
0.354736328125,
0.5,
0.354736328125,
0.5,
0.36865234375,
0.0767822265625,
0.36865234375
],
[
0.50244140625,
0.354248046875,
0.9208984375,
0.354736328125,
0.9208984375,
0.369384765625,
0.50244140625,
0.36865234375
],
[
0.0767822265625,
0.369384765625,
0.495849609375,
0.36865234375,
0.495849609375,
0.38330078125,
0.0767822265625,
0.3837890625
],
[
0.491943359375,
0.373779296875,
0.5087890625,
0.373779296875,
0.5087890625,
0.380615234375,
0.491943359375,
0.380615234375
],
[
0.5048828125,
0.371826171875,
0.9189453125,
0.371826171875,
0.9189453125,
0.3818359375,
0.5048828125,
0.3818359375
],
[
0.07843017578125,
0.385009765625,
0.494384765625,
0.385009765625,
0.494384765625,
0.3984375,
0.07843017578125,
0.3984375
],
[
0.5048828125,
0.385009765625,
0.9189453125,
0.385009765625,
0.9189453125,
0.3984375,
0.5048828125,
0.3984375
],
[
0.0767822265625,
0.40087890625,
0.491943359375,
0.40087890625,
0.491943359375,
0.4140625,
0.0767822265625,
0.4140625
],
[
0.50341796875,
0.40087890625,
0.9208984375,
0.40087890625,
0.9208984375,
0.4140625,
0.50341796875,
0.4140625
],
[
0.0767822265625,
0.415283203125,
0.495849609375,
0.415283203125,
0.495849609375,
0.4287109375,
0.0767822265625,
0.4287109375
],
[
0.50244140625,
0.415283203125,
0.9208984375,
0.4140625,
0.9208984375,
0.4287109375,
0.50244140625,
0.429931640625
],
[
0.0767822265625,
0.43115234375,
0.495849609375,
0.43115234375,
0.495849609375,
0.4443359375,
0.0767822265625,
0.4443359375
],
[
0.50341796875,
0.43115234375,
0.9189453125,
0.43115234375,
0.9189453125,
0.4443359375,
0.50341796875,
0.4443359375
],
[
0.07843017578125,
0.44580078125,
0.494384765625,
0.44580078125,
0.494384765625,
0.458984375,
0.07843017578125,
0.458984375
],
[
0.50341796875,
0.44580078125,
0.9189453125,
0.44580078125,
0.9189453125,
0.458984375,
0.50341796875,
0.458984375
],
[
0.07843017578125,
0.460205078125,
0.494384765625,
0.460205078125,
0.494384765625,
0.473388671875,
0.07843017578125,
0.473388671875
],
[
0.50732421875,
0.460205078125,
0.9208984375,
0.460205078125,
0.9208984375,
0.473388671875,
0.50732421875,
0.473388671875
],
[
0.07843017578125,
0.474853515625,
0.494384765625,
0.474853515625,
0.494384765625,
0.488037109375,
0.07843017578125,
0.488037109375
],
[
0.50634765625,
0.47607421875,
0.9208984375,
0.47607421875,
0.9208984375,
0.4892578125,
0.50634765625,
0.4892578125
],
[
0.0767822265625,
0.490478515625,
0.495849609375,
0.490478515625,
0.495849609375,
0.50390625,
0.0767822265625,
0.50390625
],
[
0.50634765625,
0.4912109375,
0.9189453125,
0.4912109375,
0.9189453125,
0.5048828125,
0.50634765625,
0.5048828125
],
[
0.0767822265625,
0.505859375,
0.494384765625,
0.505859375,
0.494384765625,
0.51953125,
0.0767822265625,
0.51953125
],
[
0.50634765625,
0.505859375,
0.9208984375,
0.505859375,
0.9208984375,
0.51953125,
0.50634765625,
0.51953125
],
[
0.0767822265625,
0.52099609375,
0.279296875,
0.52099609375,
0.279296875,
0.5341796875,
0.0767822265625,
0.5341796875
],
[
0.50732421875,
0.521484375,
0.92236328125,
0.521484375,
0.92236328125,
0.53515625,
0.50732421875,
0.53515625
],
[
0.0931396484375,
0.53515625,
0.491943359375,
0.53515625,
0.491943359375,
0.548828125,
0.0931396484375,
0.548828125
],
[
0.50732421875,
0.537109375,
0.564453125,
0.537109375,
0.564453125,
0.548828125,
0.50732421875,
0.548828125
],
[
0.0767822265625,
0.5498046875,
0.490234375,
0.5498046875,
0.490234375,
0.56298828125,
0.0767822265625,
0.56298828125
],
[
0.52392578125,
0.5517578125,
0.91845703125,
0.5517578125,
0.91845703125,
0.56494140625,
0.52392578125,
0.56494140625
],
[
0.0767822265625,
0.56494140625,
0.494384765625,
0.56494140625,
0.494384765625,
0.5791015625,
0.0767822265625,
0.5791015625
],
[
0.50341796875,
0.5673828125,
0.9189453125,
0.5673828125,
0.9189453125,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.07843017578125,
0.58056640625,
0.494384765625,
0.58056640625,
0.494384765625,
0.5947265625,
0.07843017578125,
0.5947265625
],
[
0.50341796875,
0.58349609375,
0.9208984375,
0.58349609375,
0.9208984375,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.07843017578125,
0.59521484375,
0.494384765625,
0.59521484375,
0.494384765625,
0.609375,
0.07843017578125,
0.609375
],
[
0.50341796875,
0.5966796875,
0.9189453125,
0.5966796875,
0.9189453125,
0.60986328125,
0.50341796875,
0.60986328125
],
[
0.0767822265625,
0.60986328125,
0.494384765625,
0.611328125,
0.494384765625,
0.625,
0.0767822265625,
0.62353515625
],
[
0.50634765625,
0.61279296875,
0.9189453125,
0.61279296875,
0.9189453125,
0.626953125,
0.50634765625,
0.626953125
],
[
0.07843017578125,
0.62548828125,
0.49267578125,
0.62548828125,
0.49267578125,
0.6396484375,
0.07843017578125,
0.6396484375
],
[
0.50732421875,
0.62890625,
0.91845703125,
0.62890625,
0.91845703125,
0.6396484375,
0.50732421875,
0.6396484375
],
[
0.07843017578125,
0.6416015625,
0.495849609375,
0.6416015625,
0.495849609375,
0.65478515625,
0.07843017578125,
0.65478515625
],
[
0.50634765625,
0.642578125,
0.9208984375,
0.642578125,
0.9208984375,
0.65576171875,
0.50634765625,
0.65576171875
],
[
0.0767822265625,
0.6572265625,
0.384033203125,
0.6572265625,
0.384033203125,
0.67041015625,
0.0767822265625,
0.67041015625
],
[
0.5087890625,
0.6591796875,
0.9189453125,
0.6591796875,
0.9189453125,
0.66943359375,
0.5087890625,
0.66943359375
],
[
0.0931396484375,
0.67041015625,
0.495849609375,
0.67041015625,
0.495849609375,
0.6845703125,
0.0931396484375,
0.6845703125
],
[
0.50634765625,
0.67236328125,
0.9189453125,
0.67236328125,
0.9189453125,
0.68603515625,
0.50634765625,
0.68603515625
],
[
0.0767822265625,
0.6875,
0.495849609375,
0.6875,
0.495849609375,
0.70068359375,
0.0767822265625,
0.70068359375
],
[
0.50341796875,
0.68798828125,
0.9208984375,
0.68798828125,
0.9208984375,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.07843017578125,
0.7021484375,
0.495849609375,
0.7021484375,
0.495849609375,
0.71533203125,
0.07843017578125,
0.71533203125
],
[
0.50634765625,
0.70263671875,
0.92236328125,
0.70263671875,
0.92236328125,
0.71630859375,
0.50634765625,
0.71630859375
],
[
0.0767822265625,
0.71630859375,
0.3857421875,
0.71630859375,
0.3857421875,
0.72998046875,
0.0767822265625,
0.72998046875
],
[
0.50732421875,
0.71826171875,
0.9208984375,
0.71826171875,
0.9208984375,
0.73193359375,
0.50732421875,
0.73193359375
],
[
0.09234619140625,
0.73095703125,
0.494384765625,
0.73095703125,
0.494384765625,
0.744140625,
0.09234619140625,
0.744140625
],
[
0.50732421875,
0.73486328125,
0.740234375,
0.73486328125,
0.740234375,
0.74560546875,
0.50732421875,
0.74560546875
],
[
0.0767822265625,
0.7470703125,
0.49267578125,
0.7470703125,
0.49267578125,
0.76123046875,
0.0767822265625,
0.76123046875
],
[
0.525390625,
0.75048828125,
0.91845703125,
0.75048828125,
0.91845703125,
0.76123046875,
0.525390625,
0.76123046875
],
[
0.0767822265625,
0.76220703125,
0.371826171875,
0.76220703125,
0.371826171875,
0.77587890625,
0.0767822265625,
0.77587890625
],
[
0.50634765625,
0.7646484375,
0.9208984375,
0.7646484375,
0.9208984375,
0.77880859375,
0.50634765625,
0.77880859375
],
[
0.0906982421875,
0.77587890625,
0.494384765625,
0.7744140625,
0.494384765625,
0.7890625,
0.0906982421875,
0.79052734375
],
[
0.50341796875,
0.77978515625,
0.92236328125,
0.77978515625,
0.92236328125,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.07843017578125,
0.79150390625,
0.494384765625,
0.79150390625,
0.494384765625,
0.8046875,
0.07843017578125,
0.8046875
],
[
0.50732421875,
0.796875,
0.9189453125,
0.796875,
0.9189453125,
0.806640625,
0.50732421875,
0.806640625
],
[
0.0767822265625,
0.806640625,
0.495849609375,
0.806640625,
0.495849609375,
0.82080078125,
0.0767822265625,
0.82080078125
],
[
0.495849609375,
0.8095703125,
0.7197265625,
0.8095703125,
0.7197265625,
0.82275390625,
0.495849609375,
0.82275390625
],
[
0.0767822265625,
0.8212890625,
0.495849609375,
0.8212890625,
0.495849609375,
0.83544921875,
0.0767822265625,
0.83544921875
],
[
0.50634765625,
0.82470703125,
0.9208984375,
0.82470703125,
0.9208984375,
0.83837890625,
0.50634765625,
0.83837890625
],
[
0.080078125,
0.83837890625,
0.49267578125,
0.83837890625,
0.49267578125,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.5048828125,
0.84033203125,
0.91845703125,
0.84033203125,
0.91845703125,
0.85400390625,
0.5048828125,
0.85400390625
],
[
0.07843017578125,
0.8515625,
0.494384765625,
0.8515625,
0.494384765625,
0.86572265625,
0.07843017578125,
0.86572265625
],
[
0.5048828125,
0.85595703125,
0.9208984375,
0.85595703125,
0.9208984375,
0.869140625,
0.5048828125,
0.869140625
],
[
0.07843017578125,
0.8671875,
0.261474609375,
0.8671875,
0.261474609375,
0.880859375,
0.07843017578125,
0.880859375
],
[
0.50732421875,
0.87060546875,
0.92236328125,
0.87060546875,
0.92236328125,
0.8837890625,
0.50732421875,
0.8837890625
],
[
0.5107421875,
0.88720703125,
0.91845703125,
0.88720703125,
0.91845703125,
0.89794921875,
0.5107421875,
0.89794921875
],
[
0.2115478515625,
0.896484375,
0.358642578125,
0.896484375,
0.358642578125,
0.91015625,
0.2115478515625,
0.91015625
],
[
0.50634765625,
0.90087890625,
0.91845703125,
0.90087890625,
0.91845703125,
0.9140625,
0.50634765625,
0.9140625
],
[
0.07843017578125,
0.91552734375,
0.496826171875,
0.91552734375,
0.496826171875,
0.9287109375,
0.07843017578125,
0.9287109375
],
[
0.50244140625,
0.91650390625,
0.9208984375,
0.9140625,
0.9208984375,
0.92822265625,
0.50244140625,
0.9296875
],
[
0.07598876953125,
0.93115234375,
0.495849609375,
0.9296875,
0.495849609375,
0.9443359375,
0.07598876953125,
0.94580078125
],
[
0.50341796875,
0.93115234375,
0.9208984375,
0.93115234375,
0.9208984375,
0.9443359375,
0.50341796875,
0.9443359375
]
] |
[
"2",
"IEEEROBOTICSANDAUTOMATIONLETTERS.PREPRINTVERSION.ACCEPTEDAUGUST,2023",
"MacFormer",
"Coupled Layer",
"Transformer-based Context Encode",
"Mop-conditioned Decoder",
"MTOS",
"Agent Motion",
"entFe",
"t",
"Hietnr",
"MIP",
"Coptre Tatk",
"",
"",
"CoupledMoti",
"",
"Interested AgentOther AgentAgentFeature Coup",
"八ProbabilityD",
"",
"RelotiveMotions",
"alOe",
"Fig. 2.(a) The system overview of MacFormer. (b) The detailed implementation of MacFormer and output size of each operation.",
"Former, an one-stage Map-Agent Coupled TransFormer,",
"ferent directions: implicit-fusion and target-driven.",
"which directly couples map constraints with agent motions",
"1) Implicit-fusion. Implicit-fusion models use different",
"for real-time and robust trajectory prediction (Fig.1). In order",
"backbones to encode map and fuse it with agent motions",
"toenabledirect utilizationof thecoupled relationbetween",
"in high-level latent feature. By optimizing prediction outputs,",
"the map and agent, two modules named coupled map and",
"they aim to capture the map structure in such fusion. Early",
"reference extractor are carefully designed, which explicitly",
"works [2], [21] used CNNs to encode the map and agents",
"integrate map constraints into the system. Furthermore, to",
"together into an image, then predicted trajectories using a",
"assure that the map constraints are well learned during train-",
"fully connected layer. However, this approach cannot model",
"ing, we propose a multi-task optimization strategy (MTOS).",
"",
"relationsbetweendifferent scenarioelementseffectively.Vec-",
"It ensures map-constrained prediction by forecasting future",
"torNet [1], LaneGCN [3], and DSP [8] all vectorized the map",
"relative motions in coupled map, capturing motion priors, and",
"and adopted graph-based architectures to extract features and",
"outputting final predictions with corresponding probabilities",
"interactions of agents and map instead. Besides, TPCN [7]",
"via map-conditioned regression. Lastly, to enhance computa-",
"migrated point cloud learning backbone to implement the fea-",
"tional efficiency and promote real-time prediction, a concise",
"ture fusion. To cover long-range prediction, mmTransformer",
"bilateral query scheme for context fusion is devised. It allows",
"[6], SceneTransformer [4] and MultiPath++[5] leveraged the",
"parallel cross-domain information interaction between the map",
"global receptive field of Transformer to capture whole context",
"and agent from their distinct perspectives and shared usage of",
"fusion. Nevertheless, implicit-fusion still leads to mode col-",
"their affinity matrix, significantly reducing the time and space",
"lapse or infeasible forecasts that violate map constraints, which",
"complexity of context fusion.",
"illustrates their unsatisfactory map utilization and prediction",
"The proposed method was evaluated on three large-",
"stability.",
"scale real-world benchmarks (Argoverse 1&2 [18], J19] and",
"2) Target-driven. To overcome the limitations in implicit-",
"nuScenes [20]) and achieved state-of-the-art performance with",
"fusion models, TNT [12] and DenseTNT [13] decomposed pre-",
"significantly lower inference latency and fewer parameters.",
"diction objectives into two stage: target prediction and motion",
"We also conducted extensive experiments to validate the",
"completion. They first sampled several target candidates for",
"effectiveness and robustness of our framework. Furthermore,",
"the agent from map, selected qualified targets from candidates,",
"we introduced our framework to classical prediction models,",
"then completed full trajectory conditioned on each target",
"enhancing their performance while reducing parameters. In",
"Similarly, HOME [14], GOHOME [15] and THOMAS [16]",
"summary, the contributions of this paper are:",
"forecastedaheatmapandgreedilysampledtargetcandidatesto",
"1) A map-agent coupled framework for real-time and robust",
"implement target prediction before generating full trajectories.",
"trajectory prediction,which efficiently and effectively inte-",
"However, despite the high intent correlation of targets, such",
"grates map constraints via coupled map, reference extractor,",
"models still cannot effectively capture motions of the entire",
"and multi-task optimization strategy (MTOS).",
"process, while also incurring redundant costs due to large",
"2) An efficient and lightweight context fusion scheme,",
"quantities of unselected candidate.",
"bilateral query, which allows parallel context fusion between",
"In contrast,ourmethodtackles these problemsbydirectly",
"map and agent from their individual views.",
"and sufficiently leveraging map, which utilizes the coupled",
"3) Extensive evaluation in multiple real-world benchmarks.",
"relation by coupled map, enables efficient map-constrained",
"It shows that MacFormer achieves state-of-the-art perfor-",
"predictions via referenceextractor,considersmap constraints",
"mance, while it is faster and more lightweight than existing",
"during optimization via MTOS.",
"models. Experiments also demonstrate the versatility and ro-",
"Transformer. Transformer [22] has been successfully applied",
"bustnessoftheproposedframework.whichenhancesclassical",
"in various fields, e.g., natural language processing and com-",
"prediction methods while maintaining satisfactory resilience to",
"puter vision. Recently, some works [4]-[6] have utilized it",
"imperfect upstream results.",
"and its variants for trajectory prediction due to its capability of",
"global andflexiblefeatureextractionthatarehighlyapplicable",
"II. RELATED WORK",
"for context interaction. To leverage this architecture, our",
"Map utilization. Map utilization has received much attention",
"proposedmethod devises a Transformer-based context encoder",
"recently, where existing approaches mainly fall into two dif-",
"and proposes an effective query module in reference extractor"
] |
{
"arxiv_id": "2308.10280",
"language": "en",
"page_idx": 1
}
|
||
2402.14124-en-1
|
2402.14124-en-1.jpg
|
WWW'24, May 13-17, 2024, Singapore, Singapore
for adversaries to attack (while significant challenges for online job platforms to defend), yet our understanding on the attacks and potential defenses on online job platforms is rather limited.
To mitigate this gap in understanding, using the career prediction as target downstream task, we formulate three attack scenarios: (1) company promotion attack: amplifying the likelihood of target companies in the prediction model's result; (2) company demotion attack: diminishing the likelihood of target companies in the prediction model's result; (3) user promotion attack: amplifying the likelihood of target users being matched to certain companies, and propose a novel data poisoning attack, titled FRANCIS (Eake Resume-based dAta poisoNing attaCks on onlIne job platformS), which generates realistic fake resumes to mislead career prediction models. Figure 1 illustrates a concept of FRANCIS. Our contributions are as follows:
$\bullet$ To the best of our knowledge, FRANCIS is the first to demonstrate vulnerabilities by data poisoning attacks on online job platforms.
$\bullet$ We formulate novel attack scenarios and a data poisoning framework to generate fake resumes focusing on the weak nature of the current online job platforms.
$\bullet$ Extensive experiments show that even a small fraction of poisoning can alter the prediction results regardless of underlying career prediction and attack models.
$\bullet$ FRANCIS achieves improvement rates of up to 23.17 at 10%
injection, 4.98 at 1%, and 1.32 at 0.1% injection.
# 2 RELATED WORK
2.1 Data Poisoning Attack
Attacking online platforms is often possible [2,33],where the ultimate goal of an attacker is to exploit vulnerabilities in the platform's algorithms and generate malicious results that further their interests. [1,12]. Data poisoning attack is one of such harmful and practical attacks [1,41,42], where false information and malicious inputs are injected into the dataset to train a model, resulting in biased or incorrect predictions [26, 38]. Even though there are several works on data poisoning for web systems [19,41,42], attacking online professional job platforms $(e.$g., LinkedIn and Indeed) has not been well explored. The attack on these platforms can damage both companies and users, negatively affecting both business-to-consumer and business-to-business services [9]. As a practical attack, in this work, we propose fake resume attacks and show the pivotal vulnerability.
## 2.2 Career Prediction
“Career prediction" is an important downstream task in the HR domain, predicting the next potential job positions and/or companies from resumes. In a thorough survey of HR domain models [22], career prediction is identified as a pivotal process for comprehending and leveraging users' career activities, highlighting its application in recruitment. The growth of online professional networks has led to an unprecedented accumulation of online resumes, allowing unique opportunities for developing data-driven approaches to this task. Liu et al. [17] used multiple social media features such as Twitter for prediction with manually defined career patterns. NEMO [16], proposed by LinkedIn, is a model to predict an employee's next career move from contextual embedding. LinkedIn's patent on career move prediction [37] further substantiates the importance
Yamashita et al.
of this task, underscoring its practical application and significance. AHEAD [44] employs a heterogeneous company-position network to predict companies and positions simultaneously. TACTP [32] is a unified time-aware model to predict the next job with the estimated timing. NAOMI [35] is a long-term sequential model to predict the next k steps of pathways using multi-aspect embeddings and reasoning. In this work, we demonstrate the vulnerabilities of three state-of-the-art career prediction models [16,35,44].
2.3 HR-domain Downstream Tasks
There are various machine learning based downstream tasks that use resumes and career trajectory datasets in the HR domain [22]. For instance, skill extraction is a critical task for both companies and individuals, as companies want to assign their employees to the most effective department and individuals want to develop their skill sets [6, 28, 34]. Predicting employee turnover and job performance is another critical task, where models estimate the timing of employee turnover or how much they achieve based on multiple features [15,29,30]. Although our focus in this paper is data poisoning attack to the career prediction task, we believe that poisoned resumes could equally make other HR downstream tasks vulnerable. We leave this direction as future work.
## 3 PRELIMINARIES
3.1 Target Downstream Task
Online job platforms require users (i.e., job seekers) to create online profiles by submitting their career histories. While these user profiles are used for various HR functions, we specifically select career prediction, one of the essential real-world HR tasks, as our target downstream task [16]. In this task, the model predicts an individual's subsequent job based on past job histories. Then, a job platform uses such prediction results and provide both business-toconsumer$\left(i.\mathrm{e}.,\mathrm{B2C}\right)$and business-to-business$\left(i.\mathrm{e}.,\mathrm{B2B}\right)$services:$\left(1\right)$ For B2C side, the platform recommends a (ranked) list of companies that a job seeker matches well with, and (2) For B2B side, the platform recommends a (ranked) list of job seekers who matches well with a company so that recruiters can start recruiting actions. In other words, as the model results are used by both job seekers and companies, unique to online job platforms, it is particularly harmful if poisoned and manipulated. The details of this downstream task are explained in Section 4.5. To elucidate the repercussions of our fake resume attacks on online job platforms, we present the overview of the ecosystem in Figure 2.
# 3.2 Attack Settings
Gaining access to the specific parameters and model details of the downstream task is challenging due to their proprietary nature in commercial use. In response, we employ a black box approach by utilizing a surrogate model to generate fake resumes and then transfer it to career prediction models. For our target settings, we prefer a targeted attack approach, as it is potentially more detrimental than non-targeted attacks $(i.$e., decreasing overall model accuracy). Further details are provided in subsequent sections. Given these settings, the attacker's knowledge base is as follows:
$\bullet$ The specifics of the target prediction model, including parameters and architecture, remain unknown to the attackers (black box approach).
|
[
[
0.08416748046875,
0.0738525390625,
0.326904296875,
0.07452392578125,
0.326904296875,
0.0889892578125,
0.08416748046875,
0.08837890625
],
[
0.83642578125,
0.0770263671875,
0.9140625,
0.0770263671875,
0.9140625,
0.08709716796875,
0.83642578125,
0.08709716796875
],
[
0.08660888671875,
0.106689453125,
0.48046875,
0.106689453125,
0.48046875,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.51953125,
0.10919189453125,
0.91259765625,
0.10919189453125,
0.91259765625,
0.11932373046875,
0.51953125,
0.11932373046875
],
[
0.0849609375,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.44287109375,
0.1351318359375,
0.44287109375,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1337890625,
0.91259765625,
0.1337890625,
0.91259765625,
0.147705078125,
0.51708984375,
0.147705078125
],
[
0.10296630859375,
0.1483154296875,
0.48046875,
0.1483154296875,
0.48046875,
0.1622314453125,
0.10296630859375,
0.1622314453125
],
[
0.51708984375,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.0849609375,
0.1634521484375,
0.483642578125,
0.1634521484375,
0.483642578125,
0.1767578125,
0.0849609375,
0.1767578125
],
[
0.51708984375,
0.1634521484375,
0.91259765625,
0.1634521484375,
0.91259765625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.08660888671875,
0.1767578125,
0.483642578125,
0.1767578125,
0.483642578125,
0.1900634765625,
0.08660888671875,
0.1900634765625
],
[
0.5185546875,
0.1787109375,
0.9111328125,
0.1787109375,
0.9111328125,
0.189453125,
0.5185546875,
0.189453125
],
[
0.0849609375,
0.1912841796875,
0.483642578125,
0.1912841796875,
0.483642578125,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.51708984375,
0.1900634765625,
0.91259765625,
0.1900634765625,
0.91259765625,
0.203857421875,
0.51708984375,
0.203857421875
],
[
0.0849609375,
0.203857421875,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.0849609375,
0.2171630859375
],
[
0.515625,
0.203857421875,
0.8349609375,
0.2052001953125,
0.8349609375,
0.2197265625,
0.515625,
0.2183837890625
],
[
0.0849609375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.515625,
0.222900390625,
0.81396484375,
0.2235107421875,
0.81396484375,
0.238037109375,
0.515625,
0.2374267578125
],
[
0.0849609375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.0849609375,
0.246826171875
],
[
0.51708984375,
0.242431640625,
0.9140625,
0.242431640625,
0.9140625,
0.255615234375,
0.51708984375,
0.255615234375
],
[
0.0882568359375,
0.248779296875,
0.47802734375,
0.248779296875,
0.47802734375,
0.255615234375,
0.0882568359375,
0.255615234375
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.51708984375,
0.2568359375,
0.91259765625,
0.2568359375,
0.91259765625,
0.270751953125,
0.51708984375,
0.270751953125
],
[
0.08660888671875,
0.27587890625,
0.48046875,
0.27587890625,
0.48046875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.5185546875,
0.271484375,
0.9111328125,
0.271484375,
0.9111328125,
0.2822265625,
0.5185546875,
0.2822265625
],
[
0.0849609375,
0.287353515625,
0.483642578125,
0.287353515625,
0.483642578125,
0.300537109375,
0.0849609375,
0.300537109375
],
[
0.51708984375,
0.2841796875,
0.9140625,
0.2841796875,
0.9140625,
0.29736328125,
0.51708984375,
0.29736328125
],
[
0.5185546875,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.311279296875,
0.5185546875,
0.311279296875
],
[
0.11358642578125,
0.30615234375,
0.483642578125,
0.30615234375,
0.483642578125,
0.3193359375,
0.11358642578125,
0.3193359375
],
[
0.51953125,
0.313232421875,
0.91259765625,
0.313232421875,
0.91259765625,
0.323974609375,
0.51953125,
0.323974609375
],
[
0.125,
0.32080078125,
0.48193359375,
0.32080078125,
0.48193359375,
0.333984375,
0.125,
0.333984375
],
[
0.515625,
0.32568359375,
0.91259765625,
0.324462890625,
0.91259765625,
0.33837890625,
0.515625,
0.339111328125
],
[
0.12744140625,
0.3359375,
0.1871337890625,
0.3359375,
0.1871337890625,
0.347900390625,
0.12744140625,
0.347900390625
],
[
0.111083984375,
0.3466796875,
0.48046875,
0.347900390625,
0.48046875,
0.3623046875,
0.111083984375,
0.361083984375
],
[
0.5185546875,
0.341552734375,
0.91259765625,
0.341552734375,
0.91259765625,
0.3515625,
0.5185546875,
0.3515625
],
[
0.51708984375,
0.352783203125,
0.91259765625,
0.352783203125,
0.91259765625,
0.3662109375,
0.51708984375,
0.3662109375
],
[
0.12335205078125,
0.361083984375,
0.48046875,
0.361083984375,
0.48046875,
0.375,
0.12335205078125,
0.375
],
[
0.51708984375,
0.3662109375,
0.9140625,
0.3662109375,
0.9140625,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.1258544921875,
0.37744140625,
0.378173828125,
0.37744140625,
0.378173828125,
0.38818359375,
0.1258544921875,
0.38818359375
],
[
0.515625,
0.380615234375,
0.9140625,
0.380126953125,
0.9140625,
0.39453125,
0.515625,
0.395751953125
],
[
0.11358642578125,
0.389404296875,
0.483642578125,
0.389404296875,
0.483642578125,
0.40283203125,
0.11358642578125,
0.40283203125
],
[
0.5185546875,
0.396484375,
0.81884765625,
0.396484375,
0.81884765625,
0.4072265625,
0.5185546875,
0.4072265625
],
[
0.1258544921875,
0.404052734375,
0.48193359375,
0.404052734375,
0.48193359375,
0.417236328125,
0.1258544921875,
0.417236328125
],
[
0.125,
0.417236328125,
0.39794921875,
0.417236328125,
0.39794921875,
0.43115234375,
0.125,
0.43115234375
],
[
0.51953125,
0.419189453125,
0.5439453125,
0.419189453125,
0.5439453125,
0.427978515625,
0.51953125,
0.427978515625
],
[
0.5390625,
0.416748046875,
0.68701171875,
0.416748046875,
0.68701171875,
0.429931640625,
0.5390625,
0.429931640625
],
[
0.11273193359375,
0.43115234375,
0.483642578125,
0.43115234375,
0.483642578125,
0.4443359375,
0.11273193359375,
0.4443359375
],
[
0.515625,
0.43505859375,
0.76318359375,
0.436279296875,
0.76318359375,
0.45068359375,
0.515625,
0.449462890625
],
[
0.1258544921875,
0.4462890625,
0.406005859375,
0.4462890625,
0.406005859375,
0.45703125,
0.1258544921875,
0.45703125
],
[
0.515625,
0.453857421875,
0.9140625,
0.455078125,
0.9140625,
0.468505859375,
0.515625,
0.46728515625
],
[
0.1094970703125,
0.4716796875,
0.25732421875,
0.4716796875,
0.25732421875,
0.486083984375,
0.1094970703125,
0.486083984375
],
[
0.51708984375,
0.468505859375,
0.9140625,
0.468505859375,
0.9140625,
0.481689453125,
0.51708984375,
0.481689453125
],
[
0.51708984375,
0.48291015625,
0.9140625,
0.48291015625,
0.9140625,
0.496826171875,
0.51708984375,
0.496826171875
],
[
0.0849609375,
0.473388671875,
0.111083984375,
0.473388671875,
0.111083984375,
0.48486328125,
0.0849609375,
0.48486328125
],
[
0.0849609375,
0.4912109375,
0.310546875,
0.4912109375,
0.310546875,
0.505859375,
0.0849609375,
0.505859375
],
[
0.51708984375,
0.496826171875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.51708984375,
0.51025390625
],
[
0.0849609375,
0.51025390625,
0.483642578125,
0.51025390625,
0.483642578125,
0.5234375,
0.0849609375,
0.5234375
],
[
0.5185546875,
0.51220703125,
0.91259765625,
0.51220703125,
0.91259765625,
0.52294921875,
0.5185546875,
0.52294921875
],
[
0.0849609375,
0.5244140625,
0.48193359375,
0.5244140625,
0.48193359375,
0.53857421875,
0.0849609375,
0.53857421875
],
[
0.51708984375,
0.5244140625,
0.9140625,
0.5244140625,
0.9140625,
0.53857421875,
0.51708984375,
0.53857421875
],
[
0.08660888671875,
0.53857421875,
0.48046875,
0.53857421875,
0.48046875,
0.548828125,
0.08660888671875,
0.548828125
],
[
0.51708984375,
0.53857421875,
0.9140625,
0.53857421875,
0.9140625,
0.5517578125,
0.51708984375,
0.5517578125
],
[
0.0849609375,
0.5517578125,
0.483642578125,
0.5517578125,
0.483642578125,
0.56494140625,
0.0849609375,
0.56494140625
],
[
0.51708984375,
0.5517578125,
0.9140625,
0.5517578125,
0.9140625,
0.56494140625,
0.51708984375,
0.56494140625
],
[
0.08660888671875,
0.5673828125,
0.48046875,
0.5673828125,
0.48046875,
0.57763671875,
0.08660888671875,
0.57763671875
],
[
0.5185546875,
0.5673828125,
0.91259765625,
0.5673828125,
0.91259765625,
0.57763671875,
0.5185546875,
0.57763671875
],
[
0.08660888671875,
0.58056640625,
0.48046875,
0.58056640625,
0.48046875,
0.59130859375,
0.08660888671875,
0.59130859375
],
[
0.5185546875,
0.58056640625,
0.91259765625,
0.58056640625,
0.91259765625,
0.59130859375,
0.5185546875,
0.59130859375
],
[
0.0849609375,
0.59326171875,
0.48193359375,
0.59326171875,
0.48193359375,
0.60693359375,
0.0849609375,
0.60693359375
],
[
0.51708984375,
0.59326171875,
0.9140625,
0.59326171875,
0.9140625,
0.60693359375,
0.51708984375,
0.60693359375
],
[
0.08660888671875,
0.609375,
0.48046875,
0.609375,
0.48046875,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.5185546875,
0.60791015625,
0.91259765625,
0.60791015625,
0.91259765625,
0.61865234375,
0.5185546875,
0.61865234375
],
[
0.08660888671875,
0.62255859375,
0.48046875,
0.62255859375,
0.48046875,
0.63330078125,
0.08660888671875,
0.63330078125
],
[
0.5185546875,
0.6220703125,
0.91259765625,
0.6220703125,
0.91259765625,
0.6318359375,
0.5185546875,
0.6318359375
],
[
0.0849609375,
0.63525390625,
0.48193359375,
0.63525390625,
0.48193359375,
0.6484375,
0.0849609375,
0.6484375
],
[
0.51708984375,
0.63525390625,
0.91259765625,
0.63525390625,
0.91259765625,
0.6484375,
0.51708984375,
0.6484375
],
[
0.08660888671875,
0.65087890625,
0.48193359375,
0.65087890625,
0.48193359375,
0.6611328125,
0.08660888671875,
0.6611328125
],
[
0.51708984375,
0.6484375,
0.9140625,
0.6484375,
0.9140625,
0.662109375,
0.51708984375,
0.662109375
],
[
0.08660888671875,
0.6640625,
0.48193359375,
0.6640625,
0.48193359375,
0.6748046875,
0.08660888671875,
0.6748046875
],
[
0.51708984375,
0.6640625,
0.91259765625,
0.6640625,
0.91259765625,
0.6748046875,
0.51708984375,
0.6748046875
],
[
0.0849609375,
0.6767578125,
0.48193359375,
0.6767578125,
0.48193359375,
0.68994140625,
0.0849609375,
0.68994140625
],
[
0.51708984375,
0.6767578125,
0.91259765625,
0.6767578125,
0.91259765625,
0.68994140625,
0.51708984375,
0.68994140625
],
[
0.08416748046875,
0.6962890625,
0.26953125,
0.6962890625,
0.26953125,
0.7109375,
0.08416748046875,
0.7109375
],
[
0.5185546875,
0.6923828125,
0.7490234375,
0.6923828125,
0.7490234375,
0.70263671875,
0.5185546875,
0.70263671875
],
[
0.5146484375,
0.70703125,
0.6845703125,
0.70849609375,
0.6845703125,
0.72607421875,
0.5146484375,
0.724609375
],
[
0.0849609375,
0.71630859375,
0.483642578125,
0.71630859375,
0.483642578125,
0.72998046875,
0.0849609375,
0.72998046875
],
[
0.08660888671875,
0.73095703125,
0.483642578125,
0.73095703125,
0.483642578125,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.51708984375,
0.7265625,
0.91259765625,
0.7265625,
0.91259765625,
0.74072265625,
0.51708984375,
0.74072265625
],
[
0.08660888671875,
0.7470703125,
0.48193359375,
0.7470703125,
0.48193359375,
0.7568359375,
0.08660888671875,
0.7568359375
],
[
0.51708984375,
0.7421875,
0.91259765625,
0.7421875,
0.91259765625,
0.755859375,
0.51708984375,
0.755859375
],
[
0.08660888671875,
0.76025390625,
0.48046875,
0.76025390625,
0.48046875,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.51708984375,
0.755859375,
0.91259765625,
0.755859375,
0.91259765625,
0.76953125,
0.51708984375,
0.76953125
],
[
0.0849609375,
0.77294921875,
0.48193359375,
0.77294921875,
0.48193359375,
0.7861328125,
0.0849609375,
0.7861328125
],
[
0.515625,
0.76806640625,
0.9140625,
0.76953125,
0.9140625,
0.78271484375,
0.515625,
0.7822265625
],
[
0.08660888671875,
0.787109375,
0.48046875,
0.787109375,
0.48046875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.5185546875,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.79541015625,
0.5185546875,
0.79541015625
],
[
0.08660888671875,
0.8017578125,
0.48046875,
0.8017578125,
0.48046875,
0.8125,
0.08660888671875,
0.8125
],
[
0.51708984375,
0.79736328125,
0.9140625,
0.79736328125,
0.9140625,
0.81103515625,
0.51708984375,
0.81103515625
],
[
0.0849609375,
0.81298828125,
0.483642578125,
0.8125,
0.483642578125,
0.8271484375,
0.0849609375,
0.82763671875
],
[
0.51708984375,
0.81103515625,
0.91259765625,
0.81103515625,
0.91259765625,
0.82470703125,
0.51708984375,
0.82470703125
],
[
0.08660888671875,
0.8291015625,
0.48193359375,
0.8291015625,
0.48193359375,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.8369140625,
0.5185546875,
0.8369140625
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.8388671875,
0.8349609375,
0.8388671875,
0.8349609375,
0.85302734375,
0.51708984375,
0.85302734375
],
[
0.08660888671875,
0.85498046875,
0.48046875,
0.85498046875,
0.48046875,
0.869140625,
0.08660888671875,
0.869140625
],
[
0.54345703125,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.869140625,
0.54345703125,
0.869140625
],
[
0.08660888671875,
0.87060546875,
0.48046875,
0.87060546875,
0.48046875,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.55810546875,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.55810546875,
0.880859375
],
[
0.08660888671875,
0.88525390625,
0.48046875,
0.88525390625,
0.48046875,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.55615234375,
0.8818359375,
0.68896484375,
0.88330078125,
0.68896484375,
0.89794921875,
0.55615234375,
0.896484375
]
] |
[
"www24,May137,2024, Singapore,Singapore",
"Yamashita et al.",
"for adversaries to attack (while significant challenges for online",
"ofthistask,underscoringitspracticalapplicationandsignificance",
"job platforms to defend), yet our understanding on the attacks and",
"AHEAD [44] employs a heterogeneous company-position network",
"potential defenses on online job platforms is rather limited.",
"to predict companies and positions simultaneously. TACTP [32] is a",
"To mitigate this gap in understanding, using the career prediction",
"unified time-aware model to predict the next job with the estimated",
"as target downstream task, we formulate three attack scenarios: (1)",
"timing. NAOMI [35] is a long-term sequential model to predict",
"company promotion attack: amplifying the likelihood of target com-",
"thenextk steps ofpathwaysusingmulti-aspectembeddingsand",
"panies in the prediction model's result; (2) company demotion attack:",
"reasoning. In this work, we demonstrate the vulnerabilities of three",
"diminishing the likelihood of target companies in the prediction",
"state-of-the-art career prediction models [16, 35, 44].",
"model's result; (3) user promotion attack: amplifying the likelihood",
"2.3 HR-domain Downstream Tasks",
"of target users being matched to certain companies, and propose a",
"There are various machine learning based downstream tasks that",
"noveldatapoisoningattacktitledFRANCiS(FakeResume-basec",
"dAta poisoNing attaCks on onllne job platformS), which generates",
"use resumes and career trajectory datasets in the HR domain [22]",
"realisticfake resumes to mislead careerpredictionmodels.Figure1",
"For instance,skill extraction is a critical task for both companies",
"illustrates a concept of FRANCIS. Our contributions are as follows:",
"and individuals, as companies want to assign their employees to",
"the most effective department and individuals want to develop",
". To the best of our knowledge, FRANCIS is the first to demon-",
"their skill sets [6, 28, 34].Predicting employee turnover and job",
"strate vulnerabilities by data poisoning attacks on online job",
"performance is another critical task, where models estimate the",
"platforms",
". We formulate novel attack scenarios and a data poisoning",
"timing of employeeturnover orhowmuchtheyachievebased on",
"multiple features [15, 29, 30]. Although our focus in this paper is",
"framework to generate fake resumes focusing on the weak",
"data poisoning attack to the career prediction task, we believe that",
"nature of the current online jobplatforms",
"poisoned resumes could equally make other HR downstream tasks",
". Extensive experiments show that even a small fraction of",
"vulnerable.We leave this direction as futurework.",
"poisoning can alter the prediction results regardless of un",
"derlying career prediction and attack models.",
"3",
"PRELIMINARIES",
" FRANCIS achieves improvement rates of up to 23.17 at 10%",
"3.1 Target Downstream Task",
"injection,4.98 at 1%, and 1.32at 0.1% injection.",
"Online job platforms require users (ie, job seekers) to create on-",
"RELATED WORK",
"line profiles by submitting their career histories. While these user",
"profles are used for various HR functions, we specifically select",
"2",
"2.1 Data Poisoning Attack",
"career prediction, one of the essential real-world HR tasks, as our",
"Attacking online platforms is often possible [2, 33], where the ulti-",
"targetdownstream task[16].In thistask,themodel predicts an",
"mate goal of an attacker is to exploit vulnerabilities in the platform's",
"individual's subsequent job based on past job histories. Then, a job",
"algorithms and generatemalicious results that further their inter-",
"platform uses such prediction results and provide both business-to-",
"ests. [1, 12]. Data poisoning attack is one of such harmful and practi-",
"consumer (i.e., B2C) and business-to-business (ie., B2B) services: (1)",
"calattacks[1,41,42],wherefalseinformation andmaliciousinputs",
"ForB2Cside,theplatformrecommendsa (ranked)listof companies",
"are injected into the dataset to train a model, resulting in biased or",
"that a job seeker matches well with, and (2) For B2B side, the plat",
"incorrect predictions [26, 38]. Even though there are several works",
"form recommends a (ranked) list of job seekers who matches well",
"ondatapoisoningforwebsystems[19,41,42],attackingonlinepro-",
"witha company sothatrecruiterscan start recruitingactions.In",
"fessional jobplatforms (e.g.,LinkedInand Indeed)hasnotbeenwell",
"other words, as themodel results are used byboth job seekers and",
"explored. The attack on these platforms can damage both compa-",
"companies, unique to online job platforms, it is particularly harm",
"niesandusers,negativelyaffectingbothbusiness-to-consumerand",
"ful if poisoned and manipulated. The details of this downstream",
"business-to-business services[9].Asa practical attack,inthis work,",
"task areexplained in Section 4.5.To elucidate therepercussions of",
"we proposefake resume attacks and show the pivotal vulnerability",
"our fake resume attacks on online job platforms, we present the",
"2.2CareerPrediction",
"overviewoftheecosystem inFigure2.",
"3.2 Attack Settings",
"\"Career prediction\" is an important downstream task in the HR do-",
"main, predicting the next potential job positions and/or companies",
"Gaining access to the specific parameters and model details of the",
"fromresumes.InathoroughsurveyofHRdomainmodels[22],ca-",
"downstream task is challenging due to their proprietary nature in",
"reer predictionis identified asapivotal process forcomprehending",
"commercial use. In response, we employ a black box approach by",
"and leveraging users' career activities, highlighting its application",
"utilizing a surrogate model to generatefake resumes and then trans-",
"in recruitment.Thegrowth of onlineprofessional networks has",
"fer it to career predictionmodels.For our target settings, weprefer",
"ledto an unprecedentedaccumulationofonlineresumes,allowing",
"a targeted attack approach, as it is potentially more detrimental",
"unique opportunities for developing data-driven approaches to this",
"than non-targeted attacks (ie., decreasing overall model accuracy)",
"task.Liuetal.[17]used multiplesocialmediafeaturessuchasTwit-",
"Further details areprovided in subsequent sections.Given these",
"ter for prediction with manually defined career patterns. NEMO",
"settings, the attacker's knowledge base is as follows:",
"[16]. proposed by LinkedIn, is a model to predict an employee's",
". The specifics of the target prediction model, including pa-",
"nextcareermovefromcontextualembedding.Linkedn'spatent on",
"rametersand architecture,remainunknowntotheattackers",
"careermoveprediction[37]furthersubstantiatestheimportance",
"(black box approach)."
] |
{
"arxiv_id": "2402.14124",
"language": "en",
"page_idx": 1
}
|
||
2402.14124-en-2
|
2402.14124-en-2.jpg
|
Fake Resume Attacks: Data Poisoning on Online Job Platforms
WWW'24, May 13-17, 2024, Singapore, Singapore
<FigureHere>
Figure 2{:}Ecosystem of online job platforms and our attack scenarios. Users create their online accounts by registering their resumes, which are used for the career prediction model to predict their next career. Then, based on the predicted results, users receive the list of recommended companies as a B2C service while recruiters obtain the potential candidate lists as a B2B service. Our attack objects and scenarios are shown in red color. We propose (1) Company Promotion Attack, (2) Company Demotion Attack, and (3) User Promotion Attack. See more details in Section 3.
after the attack. Essentially, after the attack, a significantly larger portion of users would find these target companies among their top-N company recommendations.
$\bullet$ Attackers can only inject a limited number of fake resumes to evade the detection by the platform's security mechanism (e.g., fake resume filtering)
$\bullet$ It is relatively easy and cheap for attackers to create accounts
on a job platform.
$\bullet$ For credibility, attackers need to associate their fake resumes
with legitimate companies.
$\bullet$ All user profiles on the platform are accessible to the attackers, mirroring the visibility of professional profiles in real-world settings.
2. Company Demotion Attack: This approach is the inverse of the company promotion attack. Instead of increasing the likelihood, the aim is to decrease the likelihood and demote target companies. A plausible motivation is a corporate rivalry, where one company wishes to undermine the other company's presence on the platform.
3.3 Attack Scenarios and Objectives
$3. \textbf{User Promotion Attack: Some users, despite being keenly in- }$ terested in working for specific companies, say Google or Microsoft, may lack the necessary qualifications or experience. Consequently, these job seekers are unlikely to be recommended to the recruiters of Google or Microsoft. To promote such users for specific companies, therefore, this attack seeks to manipulate model outputs, ensuring target users to be featured in the shortlists provided to target companies. Shortlist systems consist of $K$ users for each company. The goal is to maximize the averaged display rate on the shortlist, which denotes the fraction of target companies whose top-K recommendations include target users.
We do not focus on the user demotion attack in our main discussion, as it is considerably less probable because demoting users does not provide direct benefits to attackers. However, for the sake of completeness, we present the user demotion case in Appendix F.
Our attack objects and scenarios are highlighted in red color within Figure 2. Our fake resume attack focuses on the delivery phase of a model's prediction result. Specifically, the aim is to alter the original predicted companies $X$ into target companies $Y$,thereby influencing Y's visibility to job seekers and the prominence of specific users in recruiters' shortlists. The foundation of these attacks lies in the unrestricted nature of resumes and account creation on online job platforms. We demonstrate three attack scenarios below:
## 3.4 Dataset
1. Company Promotion Attack: This approach targets specific companies $X$ and artificially increases the likelihood of $X$ to be recommended to job seekers. Imagine a small company that struggles to attract talents as job seekers are often gravitated toward larger and well-known companies. Then, an attacker may offer a promotion service to such a small company, claiming that“for some $, I can make your company to be twice more matched to job seekers than before? That is, the attacker's goal is to maximize the hit ratio of target companies. Suppose the career prediction model recommends N companies to each user. We denote the fraction of users whose top-N recommendations include the target company
We obtained our dataset from a popular career platform, FutureFit AI$^2.$ From this platform, we randomly sampled resumes of job seekers who have at least five legitimate work experiences within the United States. Given that job seekers tend to pursue positions
${\overline{{^2}\mathrm{https://www.fut}}}$turefit.ai
|
[
[
0.08660888671875,
0.0770263671875,
0.381591796875,
0.0770263671875,
0.381591796875,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.67236328125,
0.0738525390625,
0.91259765625,
0.07574462890625,
0.91259765625,
0.0889892578125,
0.67236328125,
0.08709716796875
],
[
0.1732177734375,
0.125,
0.2442626953125,
0.125,
0.2442626953125,
0.1351318359375,
0.1732177734375,
0.1351318359375
],
[
0.5009765625,
0.1300048828125,
0.53173828125,
0.1300048828125,
0.53173828125,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.55810546875,
0.1300048828125,
0.57275390625,
0.1300048828125,
0.57275390625,
0.1395263671875,
0.55810546875,
0.1395263671875
],
[
0.685546875,
0.1337890625,
0.744140625,
0.1337890625,
0.744140625,
0.14453125,
0.685546875,
0.14453125
],
[
0.214111328125,
0.143310546875,
0.23779296875,
0.143310546875,
0.23779296875,
0.1558837890625,
0.214111328125,
0.1558837890625
],
[
0.408447265625,
0.1337890625,
0.465576171875,
0.1337890625,
0.465576171875,
0.1458740234375,
0.408447265625,
0.1458740234375
],
[
0.253173828125,
0.1458740234375,
0.30810546875,
0.1458740234375,
0.30810546875,
0.1572265625,
0.253173828125,
0.1572265625
],
[
0.54345703125,
0.143310546875,
0.5888671875,
0.143310546875,
0.5888671875,
0.1546630859375,
0.54345703125,
0.1546630859375
],
[
0.61669921875,
0.146484375,
0.65625,
0.146484375,
0.65625,
0.1578369140625,
0.61669921875,
0.1578369140625
],
[
0.328369140625,
0.154052734375,
0.378173828125,
0.154052734375,
0.378173828125,
0.1641845703125,
0.328369140625,
0.1641845703125
],
[
0.268798828125,
0.1666259765625,
0.326904296875,
0.1666259765625,
0.326904296875,
0.1767578125,
0.268798828125,
0.1767578125
],
[
0.36279296875,
0.1717529296875,
0.38818359375,
0.1717529296875,
0.38818359375,
0.18310546875,
0.36279296875,
0.18310546875
],
[
0.56201171875,
0.1685791015625,
0.685546875,
0.1685791015625,
0.685546875,
0.175537109375,
0.56201171875,
0.175537109375
],
[
0.45751953125,
0.1767578125,
0.53125,
0.1767578125,
0.53125,
0.1881103515625,
0.45751953125,
0.1881103515625
],
[
0.5,
0.1912841796875,
0.61279296875,
0.1912841796875,
0.61279296875,
0.2020263671875,
0.5,
0.2020263671875
],
[
0.5908203125,
0.2108154296875,
0.611328125,
0.2108154296875,
0.611328125,
0.21533203125,
0.5908203125,
0.21533203125
],
[
0.64453125,
0.2108154296875,
0.6611328125,
0.2108154296875,
0.6611328125,
0.21533203125,
0.64453125,
0.21533203125
],
[
0.275390625,
0.2291259765625,
0.361083984375,
0.2291259765625,
0.361083984375,
0.240478515625,
0.275390625,
0.240478515625
],
[
0.438720703125,
0.2291259765625,
0.4453125,
0.2291259765625,
0.4453125,
0.2342529296875,
0.438720703125,
0.2342529296875
],
[
0.2998046875,
0.2392578125,
0.336669921875,
0.2392578125,
0.336669921875,
0.25,
0.2998046875,
0.25
],
[
0.373291015625,
0.24365234375,
0.40283203125,
0.24365234375,
0.40283203125,
0.251953125,
0.373291015625,
0.251953125
],
[
0.418212890625,
0.240478515625,
0.463134765625,
0.240478515625,
0.463134765625,
0.255126953125,
0.418212890625,
0.255126953125
],
[
0.72216796875,
0.248779296875,
0.74658203125,
0.246826171875,
0.74658203125,
0.250732421875,
0.72216796875,
0.25244140625
],
[
0.1854248046875,
0.255126953125,
0.223876953125,
0.2568359375,
0.22216796875,
0.26953125,
0.1846923828125,
0.267578125
],
[
0.75634765625,
0.2568359375,
0.81787109375,
0.2568359375,
0.81787109375,
0.268310546875,
0.75634765625,
0.268310546875
],
[
0.265625,
0.266357421875,
0.366943359375,
0.267578125,
0.366943359375,
0.279052734375,
0.265625,
0.27783203125
],
[
0.51953125,
0.266357421875,
0.56201171875,
0.268310546875,
0.560546875,
0.27978515625,
0.5185546875,
0.27783203125
],
[
0.2181396484375,
0.272705078125,
0.226318359375,
0.27783203125,
0.22216796875,
0.2822265625,
0.2132568359375,
0.27783203125
],
[
0.68115234375,
0.287353515625,
0.689453125,
0.287353515625,
0.689453125,
0.290283203125,
0.68115234375,
0.290283203125
],
[
0.4248046875,
0.2998046875,
0.58642578125,
0.2998046875,
0.58642578125,
0.313232421875,
0.4248046875,
0.313232421875
],
[
0.62353515625,
0.306884765625,
0.7587890625,
0.306884765625,
0.7587890625,
0.317626953125,
0.62353515625,
0.317626953125
],
[
0.426513671875,
0.313232421875,
0.5849609375,
0.313232421875,
0.5849609375,
0.323974609375,
0.426513671875,
0.323974609375
],
[
0.1895751953125,
0.349853515625,
0.2156982421875,
0.349853515625,
0.2156982421875,
0.35791015625,
0.1895751953125,
0.35791015625
],
[
0.33251953125,
0.359130859375,
0.41015625,
0.359130859375,
0.41015625,
0.369384765625,
0.33251953125,
0.369384765625
],
[
0.0849609375,
0.37744140625,
0.91259765625,
0.37744140625,
0.91259765625,
0.391357421875,
0.0849609375,
0.391357421875
],
[
0.0849609375,
0.391357421875,
0.9140625,
0.391357421875,
0.9140625,
0.40478515625,
0.0849609375,
0.40478515625
],
[
0.08416748046875,
0.40478515625,
0.9140625,
0.404052734375,
0.9140625,
0.41845703125,
0.08416748046875,
0.419189453125
],
[
0.0849609375,
0.419189453125,
0.91259765625,
0.419189453125,
0.91259765625,
0.43310546875,
0.0849609375,
0.43310546875
],
[
0.08660888671875,
0.43505859375,
0.611328125,
0.43505859375,
0.611328125,
0.44580078125,
0.08660888671875,
0.44580078125
],
[
0.11358642578125,
0.462158203125,
0.483642578125,
0.462158203125,
0.483642578125,
0.47607421875,
0.11358642578125,
0.47607421875
],
[
0.51708984375,
0.462158203125,
0.9140625,
0.462158203125,
0.9140625,
0.47607421875,
0.51708984375,
0.47607421875
],
[
0.125,
0.4765625,
0.483642578125,
0.4765625,
0.483642578125,
0.490478515625,
0.125,
0.490478515625
],
[
0.51708984375,
0.47607421875,
0.9140625,
0.47607421875,
0.9140625,
0.490478515625,
0.51708984375,
0.490478515625
],
[
0.125,
0.490478515625,
0.291748046875,
0.490478515625,
0.291748046875,
0.50390625,
0.125,
0.50390625
],
[
0.5185546875,
0.490478515625,
0.72802734375,
0.490478515625,
0.72802734375,
0.50390625,
0.5185546875,
0.50390625
],
[
0.11273193359375,
0.50390625,
0.48193359375,
0.50390625,
0.48193359375,
0.517578125,
0.11273193359375,
0.517578125
],
[
0.51708984375,
0.50830078125,
0.9140625,
0.50830078125,
0.9140625,
0.521484375,
0.51708984375,
0.521484375
],
[
0.12335205078125,
0.51806640625,
0.2344970703125,
0.51806640625,
0.2344970703125,
0.5322265625,
0.12335205078125,
0.5322265625
],
[
0.515625,
0.521484375,
0.9140625,
0.521484375,
0.9140625,
0.53515625,
0.515625,
0.53515625
],
[
0.11273193359375,
0.53076171875,
0.48193359375,
0.53076171875,
0.48193359375,
0.54443359375,
0.11273193359375,
0.54443359375
],
[
0.51708984375,
0.53515625,
0.9140625,
0.53515625,
0.9140625,
0.548828125,
0.51708984375,
0.548828125
],
[
0.1258544921875,
0.54736328125,
0.289306640625,
0.54736328125,
0.289306640625,
0.55810546875,
0.1258544921875,
0.55810546875
],
[
0.51708984375,
0.55029296875,
0.9111328125,
0.55029296875,
0.9111328125,
0.56103515625,
0.51708984375,
0.56103515625
],
[
0.1094970703125,
0.5595703125,
0.484375,
0.55810546875,
0.484375,
0.57275390625,
0.1094970703125,
0.57373046875
],
[
0.515625,
0.56298828125,
0.9140625,
0.56298828125,
0.9140625,
0.5771484375,
0.515625,
0.5771484375
],
[
0.125,
0.57373046875,
0.48193359375,
0.57373046875,
0.48193359375,
0.5869140625,
0.125,
0.5869140625
],
[
0.51708984375,
0.580078125,
0.9140625,
0.580078125,
0.9140625,
0.59326171875,
0.51708984375,
0.59326171875
],
[
0.12335205078125,
0.58837890625,
0.2484130859375,
0.58837890625,
0.2484130859375,
0.6015625,
0.12335205078125,
0.6015625
],
[
0.51708984375,
0.5947265625,
0.9140625,
0.5947265625,
0.9140625,
0.60791015625,
0.51708984375,
0.60791015625
],
[
0.12091064453125,
0.611328125,
0.386474609375,
0.611328125,
0.386474609375,
0.62548828125,
0.12091064453125,
0.62548828125
],
[
0.51708984375,
0.60693359375,
0.9140625,
0.60693359375,
0.9140625,
0.62060546875,
0.51708984375,
0.62060546875
],
[
0.515625,
0.62255859375,
0.9111328125,
0.62255859375,
0.9111328125,
0.63330078125,
0.515625,
0.63330078125
],
[
0.08660888671875,
0.6123046875,
0.1217041015625,
0.6123046875,
0.1217041015625,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.0849609375,
0.6298828125,
0.48193359375,
0.6298828125,
0.48193359375,
0.64306640625,
0.0849609375,
0.64306640625
],
[
0.515625,
0.63525390625,
0.9140625,
0.63623046875,
0.9140625,
0.6494140625,
0.515625,
0.6484375
],
[
0.0849609375,
0.64306640625,
0.48193359375,
0.64306640625,
0.48193359375,
0.6572265625,
0.0849609375,
0.6572265625
],
[
0.51708984375,
0.6484375,
0.9140625,
0.6484375,
0.9140625,
0.662109375,
0.51708984375,
0.662109375
],
[
0.08660888671875,
0.6591796875,
0.48193359375,
0.6591796875,
0.48193359375,
0.66943359375,
0.08660888671875,
0.66943359375
],
[
0.51708984375,
0.662109375,
0.9140625,
0.662109375,
0.9140625,
0.67529296875,
0.51708984375,
0.67529296875
],
[
0.0849609375,
0.671875,
0.48193359375,
0.671875,
0.48193359375,
0.68505859375,
0.0849609375,
0.68505859375
],
[
0.51708984375,
0.6767578125,
0.9140625,
0.6767578125,
0.9140625,
0.68994140625,
0.51708984375,
0.68994140625
],
[
0.08660888671875,
0.68505859375,
0.483642578125,
0.68505859375,
0.483642578125,
0.69873046875,
0.08660888671875,
0.69873046875
],
[
0.51708984375,
0.69140625,
0.9140625,
0.69140625,
0.9140625,
0.705078125,
0.51708984375,
0.705078125
],
[
0.08660888671875,
0.70068359375,
0.4794921875,
0.70068359375,
0.4794921875,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.515625,
0.705078125,
0.9140625,
0.705078125,
0.9140625,
0.71826171875,
0.515625,
0.71826171875
],
[
0.0849609375,
0.71337890625,
0.48193359375,
0.71337890625,
0.48193359375,
0.7265625,
0.0849609375,
0.7265625
],
[
0.5185546875,
0.72021484375,
0.7890625,
0.72021484375,
0.7890625,
0.73095703125,
0.5185546875,
0.73095703125
],
[
0.0849609375,
0.72802734375,
0.4248046875,
0.72802734375,
0.4248046875,
0.73779296875,
0.0849609375,
0.73779296875
],
[
0.53369140625,
0.73486328125,
0.91259765625,
0.736328125,
0.91259765625,
0.75,
0.53369140625,
0.74853515625
],
[
0.08660888671875,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.08660888671875,
0.75830078125
],
[
0.0849609375,
0.7568359375,
0.48046875,
0.7568359375,
0.48046875,
0.77099609375,
0.0849609375,
0.77099609375
],
[
0.51708984375,
0.751953125,
0.91259765625,
0.751953125,
0.91259765625,
0.76220703125,
0.51708984375,
0.76220703125
],
[
0.51708984375,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.77783203125,
0.51708984375,
0.77783203125
],
[
0.0849609375,
0.771484375,
0.483642578125,
0.771484375,
0.483642578125,
0.78515625,
0.0849609375,
0.78515625
],
[
0.51708984375,
0.77880859375,
0.9140625,
0.77880859375,
0.9140625,
0.79248046875,
0.51708984375,
0.79248046875
],
[
0.0882568359375,
0.7861328125,
0.4794921875,
0.7861328125,
0.4794921875,
0.796875,
0.0882568359375,
0.796875
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.515625,
0.796875,
0.62353515625,
0.79736328125,
0.62353515625,
0.8125,
0.515625,
0.81103515625
],
[
0.0849609375,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.51708984375,
0.81640625,
0.9140625,
0.81640625,
0.9140625,
0.830078125,
0.51708984375,
0.830078125
],
[
0.08660888671875,
0.8291015625,
0.4794921875,
0.8291015625,
0.4794921875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.51708984375,
0.830078125,
0.9140625,
0.830078125,
0.9140625,
0.84326171875,
0.51708984375,
0.84326171875
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.8447265625,
0.9140625,
0.8447265625,
0.9140625,
0.85791015625,
0.51708984375,
0.85791015625
],
[
0.08416748046875,
0.85400390625,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.08416748046875,
0.8671875
],
[
0.51708984375,
0.85791015625,
0.9140625,
0.85791015625,
0.9140625,
0.8720703125,
0.51708984375,
0.8720703125
],
[
0.08660888671875,
0.87060546875,
0.48193359375,
0.87060546875,
0.48193359375,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.0849609375,
0.88330078125,
0.48193359375,
0.88330078125,
0.48193359375,
0.896484375,
0.0849609375,
0.896484375
],
[
0.515625,
0.88330078125,
0.64208984375,
0.88330078125,
0.64208984375,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Fake Resume Attacks: Data Poisoning on Online Job Platforms",
"www*24,May 13-17,2024, singapore, Singapore",
"Register/Update",
"0=",
"1",
"amazon",
"8",
"amazon",
"UBER",
"UBER",
"airbnb",
"NIANTIC",
"CV/Resume",
"Input",
"OurFakeResumeModule",
"Data Poisoning",
"Obtain potential candidates",
"nDany",
"npan",
"Career Prediction",
".",
"Model",
"Predict",
"Meta",
"cnee",
"Users",
"Recruiters",
"NextJobPrediction",
"Shortlist",
"?",
"D",
"(1) Company Promotion Attack",
"(3)UserPromotionAttack",
"(2)CompanyDemotionAttack",
"Apply",
"JobApplication",
"Figure 2: Ecosystem of online job platforms and our attack scenarios. Users create their online accounts by registering their",
"resumes, which are used for the career prediction model to predict their next career. Then, based on the predicted results,",
"users receive the list of recommended companies as a B2C service while recruiters obtain the potential candidate lists as a B2B",
"service. Our attack objects and scenarios are shown in red color. We propose (1) Company Promotion Attack, (2) Company",
"DemotionAttack, and (3)UserPromotionAttack.Seemoredetailsin Section3.",
": Attackers can only inject a limited number of fake resumes",
"after the attack. Essentially, after the attack, a significantly larger",
"to evade the detection by the platform's security mechanism",
"portion of users would find these target companies among their",
"(e.g., fake resume filtering).",
"top-N company recommendations.",
". It is relatively easy and cheap for attackers to create accounts",
"2. Company Demotion Attack: This approach is the inverse of",
" on a job platform.",
"the company promotion attack. Instead of increasing the likelihood,",
"For credibility, attackers need to associate their fake resumes",
"the aim is to decrease the likelihood and demote target companies.",
"with legitimatecompanies.",
"Aplausiblemotivation is acorporaterivalry,whereone company",
". All user profiles on the platform are accessible to the at-",
"wishes to undermine the other company's presence on the platform.",
"tackers, mirroring the visibility of professional profiles in",
"3. User Promotion Attack: Some users, despite being keenly in-",
"real-world settings.",
"terested in working for specific companies, say Google or Microsoft,",
"Attack Scenarios and Objectives",
"may lack the necessary qualifications or experience.Consequently",
"thesejob seekersareunlikelytoberecommendedtotherecruiters",
"3.3",
"Our attack objects and scenarios are highlighted in red color within",
"of Google or Microsoft.Topromote suchusersfor specific com-",
"Figure 2. Our fake resume attack focuses on the delivery phase of a",
"panies, therefore, this attack seeks to manipulate model outputs,",
"model'spredictionresult.Specifically,theaim istoaltertheoriginal",
"ensuring target users to be featured in the shortlists provided to",
"predicted companies X into target companies Y,thereby influencing",
"target companies. Shortlist systems consist of K users for each",
"Y's visibility to job seekers and the prominence of specific users",
"company. The goal is to maximize the averaged display rate on the",
"inrecruiters shortlists.Thefoundation of these attacks lies inthe",
" shortlist, which denotes the fraction of target companies whose",
"unrestricted nature of resumes and account creation on online job",
"top-K recommendations include target users.",
"platforms.Wedemonstratethreeattackscenariosbelow",
"We do not focus on the user demotion attack in our main dis-",
"1. Company Promotion Attack: This approach targets specific",
"companies X and artificially increases the likelihood of X to be",
"cussion,asit is considerablyless probablebecausedemoting users",
"does not provide direct benefits to attackers. However, for the sake",
"recommended to job seekers. Imagine a small company that strug-",
"ofcompleteness, we present the user demotion case in AppendixF.",
"gles to attract talents as job seekers are often gravitated toward",
"larger and well-known companies. Then, an attacker may offer",
"3.4 Dataset",
"a promotion service to such a small company, claiming that \"for",
"We obtained our dataset from a popular career platform, Future-",
"somes,I can makeyourcompanyto betwicemorematched to job",
"Fit Al?. From this platform, we randomly sampled resumes of job",
"seekers than before\" That is, the attacker's goal is to maximize the",
"seekers who have at least five legitimate work experiences within",
"hit ratioof target companies.Suppose the career prediction model",
"the United States. Given that job seekers tend to pursue positions",
"recommendsN companiestoeachuser.Wedenotethefraction of",
"users whose top-N recommendations include the target company",
"2https:/www.futurefit.ai/"
] |
{
"arxiv_id": "2402.14124",
"language": "en",
"page_idx": 2
}
|
||
2402.14124-en-3
|
2402.14124-en-3.jpg
|
WWW'24,May 13-17, 2024, Singapore, Singapore
Table 1: Dataset Statistics
<table>
<tbody>
<tr>
<th> </th>
<th>Tech</th>
<th>Business</th>
</tr>
<tr>
<td>$\#$ of resumes</td>
<td>10,017</td>
<td>10,373</td>
</tr>
<tr>
<td>$\#$ of unique companies</td>
<td>11,679</td>
<td>12,144</td>
</tr>
</tbody>
</table>
within their current position types [43], and recruiters typically seek candidates for specific roles from ones having similar experiences, we tailored our dataset selection towards two domains-the technology (Tech) and business (Business) sectors.
To construct datasets encompassing positions within these two categories, we initiated two-step pre-processing. First, we standardized job titles in all resumes using a job title mapping model [36], which translates job titles into standardized ESCO-based position names [8]. Second, by leveraging ESCO skill definitions [14], we filtered out positions to retain only those pertinent to technology and business sectors. To further refine our data, we filtered out companies that only appeared once in our resume dataset. After this pre-processing, we obtained the datasets shown in Table 1. As our dataset also includes the company's general information $(e.$g.,# of employees), we label companies with less than 200 employees as Small" and companies with more than 10,000 employees as“Large” and use them as target companies in our attacks. For the statistics of companies per their sizes in our dataset, see Appendix B.
For ethical considerations, note that any personal identifiable information (PII) in the dataset has been anonymized, retaining only career trajectories for our experiments. While we cannot publicly release our dataset due to its commercial nature, we are open to sharing our dataset for research purposes upon valid requests (e.g., MOU signed). See Appendix A for a detailed description of our dataset collection and the ethics statement.
4 FAKE RESUME ATTACK FRAMEWORK
In this section, we propose an end-to-end fake resume generation framework FRANCIS that induces systematic prediction errors via data poisoning. In our scenario, the attacker develops an adversarial resume generator that produces a fake resume dataset $\mathcal{D}^*.$When a model is trained with $\mathcal{D}^*$ in addition to the original data, it assists the attacker in achieving the desired behavior.
Career prediction models aim to forecast the next career a person may hold based on their professional history. Let $\mathcal{D}=\{(\mathbf{x}_{i},y_{i})\}_{i=1}^{N}$ be a dataset containing N samples of career historv data $\mathbf{x}$ and be a dataset containing N samples of career history data $\mathbf{x}_i$ and the corresponding next companies $y_i$, where $y_i$ belongs to a set of $M$ possible companies $y.$ We denote $f:\mathbf{x}\to\mathcal{Y}$ as the career $\begin{array}{l}{\text{prediction model, parameterizedby}\theta_{f}.\text{To attack this model, our}}\\{\text{prediction model, parameterized by}\theta_{f}.\text{ To attack this model, our}}\end{array}$ fake resume attack approach comprises three unique modules.
4.1 $\textbf{Probabilistic Job Trajectory Generator}$
We design a conditional probabilistic job trajectory generator, denoted as $G(\mathbf{x}_{\mathrm{past}},z)$, tailored for career history data. The generator, $G$ takes the past career history data of a user as input and generates synthetic career data, $\mathbf{x}^*$, one token at a time. The generation procedure is contingent on two primary elements:
$\bullet$ The career history generated up to the current point, repre-
sented as $\mathbf{x}_\mathrm{past}.$
$\bullet$ A random latent variable, $z.$
Yamashita et al.
Each token within $\mathbf{x}^*$ is derived based on a conditional probability function at every time-step $t$ until it reaches the predetermined maximum sequence length $T.$ This process can be formally represented by:
(1)
$$\mathbf{x}^*=G(\mathbf{x}_{\mathrm{past}},z;\theta_G)$$
Here, $\theta_G$ denotes the learnable parameters intrinsic to the generator model $G.$ The training objective for $G$ can be modified to incorporate this conditional generation. Consequently, our initial objective function evolves to:
$$\min_{\theta_{G}}\frac{1}{N}\sum_{i=1}^{N}L(f(G(\mathbf{x}_{\mathrm{past},i},z;\theta_{G})),y_{i}),$$
(2)
where $\mathbf{x}_{\mathrm{past},i}$ symbolizes the previously generated career history
corresponding to the $i^{th}$ sample in the dataset.
# 4.2 Reality Regulation
We design a reality regulation function. To fabricate convincing synthetic career trajectories, our approach ensures fidelity to an underlying graph structure. Following state-of-the-art studies on formulating job transition graph [24,35,40,43],we create a graph consisting the user's job transitions, in which nodes represent companies and edges are company-company transitions as shown in Figure 3. For generating a career path, each job in the sequence should be adjacent or reachable within $n$ walking steps on the graph. This adjacency constraint can be mathematically represented as:
$$\forall c_{i},c_{j}\in\mathbf{x}^{*}:distance(c_{i},c_{j})\leq n$$
(3)
where distance$(c_i,c_j)$ computes the shortest path length between two company nodes $c_i$ and $c_j$ in the graph. Table 2 presents node degrees of large and small companies using our datasets. Average degree oflarge Tech companies is 42.89, of large Business companies is 36.10. Average degree of small Tech/Business companies remain at around 4 and the average degree of all Tech/Business companies keep at around 8. The average node degree in the graph varies between large and small companies.
## 4.3 Attack Module
We design an attack module to manipulate the adversarial generator $G$ to generate synthetic resumes that intentionally impact the results of the career prediction model. We follow a black box strategy rather than a white box approach as the black box strategy is more realistic ($i$.e, the victim model is untouchable) and does not require a transparent understanding about the victim model. As such, we design our surrogate model for career prediction to produce synthetic resumes that are then utilized by the actual and unseen victim model.
Our surrogate model $f$ predicts an individual's subsequent job,
aiming to optimize the following loss function:
$$L(f(\mathbf{x}_i;\theta_f))=-\sum_{c=1}^Cy_{ic}\log(f_c(\mathbf{x}_i;\theta_f))$$
(4)
where $C$ is the number of companies and $f_c$ is the predicted proba-
bility of company $c.$
For leveraging $f$ to guide $G$,we use backpropagation signals from $f.$ The aim is for G to generate a new resume, $x^*$, such that $f(x_i)$ (with a perturbation $y^*)$ results in a targeted prediction label
|
[
[
0.08416748046875,
0.0738525390625,
0.326904296875,
0.07452392578125,
0.326904296875,
0.0889892578125,
0.08416748046875,
0.08837890625
],
[
0.83642578125,
0.0770263671875,
0.9140625,
0.0770263671875,
0.9140625,
0.08709716796875,
0.83642578125,
0.08709716796875
],
[
0.19775390625,
0.10479736328125,
0.369384765625,
0.10479736328125,
0.369384765625,
0.11871337890625,
0.19775390625,
0.11871337890625
],
[
0.53173828125,
0.106689453125,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.53173828125,
0.12060546875
],
[
0.515625,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.515625,
0.1351318359375
],
[
0.30224609375,
0.1287841796875,
0.340576171875,
0.1287841796875,
0.340576171875,
0.14013671875,
0.30224609375,
0.14013671875
],
[
0.351318359375,
0.1287841796875,
0.416748046875,
0.1287841796875,
0.416748046875,
0.14013671875,
0.351318359375,
0.14013671875
],
[
0.51708984375,
0.1351318359375,
0.9150390625,
0.1351318359375,
0.9150390625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.146240234375,
0.143310546875,
0.226318359375,
0.143310546875,
0.226318359375,
0.1546630859375,
0.146240234375,
0.1546630859375
],
[
0.298095703125,
0.143310546875,
0.34228515625,
0.143310546875,
0.34228515625,
0.1546630859375,
0.298095703125,
0.1546630859375
],
[
0.365234375,
0.143310546875,
0.406982421875,
0.143310546875,
0.406982421875,
0.1546630859375,
0.365234375,
0.1546630859375
],
[
0.51318359375,
0.146484375,
0.583984375,
0.1483154296875,
0.583984375,
0.1634521484375,
0.51318359375,
0.1610107421875
],
[
0.143798828125,
0.1572265625,
0.291748046875,
0.1572265625,
0.291748046875,
0.17041015625,
0.143798828125,
0.17041015625
],
[
0.29736328125,
0.1578369140625,
0.34228515625,
0.1578369140625,
0.34228515625,
0.169189453125,
0.29736328125,
0.169189453125
],
[
0.365234375,
0.1578369140625,
0.406982421875,
0.1578369140625,
0.406982421875,
0.169189453125,
0.365234375,
0.169189453125
],
[
0.65283203125,
0.1622314453125,
0.77392578125,
0.1634521484375,
0.77392578125,
0.177978515625,
0.65283203125,
0.1767578125
],
[
0.89453125,
0.1641845703125,
0.9150390625,
0.1641845703125,
0.9150390625,
0.177978515625,
0.89453125,
0.177978515625
],
[
0.51708984375,
0.18115234375,
0.9150390625,
0.18115234375,
0.9150390625,
0.1944580078125,
0.51708984375,
0.1944580078125
],
[
0.0849609375,
0.189453125,
0.48193359375,
0.189453125,
0.48193359375,
0.20263671875,
0.0849609375,
0.20263671875
],
[
0.51708984375,
0.1944580078125,
0.9140625,
0.1944580078125,
0.9140625,
0.2083740234375,
0.51708984375,
0.2083740234375
],
[
0.08660888671875,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.21533203125,
0.08660888671875,
0.21533203125
],
[
0.51708984375,
0.2083740234375,
0.9140625,
0.2083740234375,
0.9140625,
0.2215576171875,
0.51708984375,
0.2215576171875
],
[
0.0849609375,
0.2171630859375,
0.48193359375,
0.2171630859375,
0.48193359375,
0.2310791015625,
0.0849609375,
0.2310791015625
],
[
0.515625,
0.2215576171875,
0.640625,
0.2215576171875,
0.640625,
0.23486328125,
0.515625,
0.23486328125
],
[
0.08660888671875,
0.23291015625,
0.38818359375,
0.23291015625,
0.38818359375,
0.24365234375,
0.08660888671875,
0.24365234375
],
[
0.63818359375,
0.24365234375,
0.65283203125,
0.24365234375,
0.65283203125,
0.25244140625,
0.63818359375,
0.25244140625
],
[
0.65625,
0.240478515625,
0.6640625,
0.240478515625,
0.6640625,
0.24560546875,
0.65625,
0.24560546875
],
[
0.66259765625,
0.248779296875,
0.826171875,
0.25,
0.826171875,
0.264404296875,
0.66259765625,
0.26318359375
],
[
0.89453125,
0.25,
0.9150390625,
0.25,
0.9150390625,
0.264404296875,
0.89453125,
0.264404296875
],
[
0.60546875,
0.250732421875,
0.6357421875,
0.250732421875,
0.6357421875,
0.261962890625,
0.60546875,
0.261962890625
],
[
0.10296630859375,
0.24560546875,
0.48046875,
0.24560546875,
0.48046875,
0.2587890625,
0.10296630859375,
0.2587890625
],
[
0.08416748046875,
0.2587890625,
0.48193359375,
0.2587890625,
0.48193359375,
0.272705078125,
0.08416748046875,
0.272705078125
],
[
0.6044921875,
0.255126953125,
0.6611328125,
0.25244140625,
0.66259765625,
0.272705078125,
0.60546875,
0.274658203125
],
[
0.650390625,
0.264404296875,
0.6708984375,
0.264404296875,
0.6708984375,
0.274658203125,
0.650390625,
0.274658203125
],
[
0.0849609375,
0.272705078125,
0.48193359375,
0.272705078125,
0.48193359375,
0.285888671875,
0.0849609375,
0.285888671875
],
[
0.51708984375,
0.279052734375,
0.9140625,
0.279052734375,
0.9140625,
0.29345703125,
0.51708984375,
0.29345703125
],
[
0.08660888671875,
0.28857421875,
0.48046875,
0.28857421875,
0.48046875,
0.298583984375,
0.08660888671875,
0.298583984375
],
[
0.51708984375,
0.29345703125,
0.7998046875,
0.29345703125,
0.7998046875,
0.30810546875,
0.51708984375,
0.30810546875
],
[
0.0849609375,
0.300537109375,
0.48193359375,
0.300537109375,
0.48193359375,
0.314453125,
0.0849609375,
0.314453125
],
[
0.08416748046875,
0.313232421875,
0.48193359375,
0.31494140625,
0.48193359375,
0.328857421875,
0.08416748046875,
0.326904296875
],
[
0.5185546875,
0.32080078125,
0.71337890625,
0.32080078125,
0.71337890625,
0.333984375,
0.5185546875,
0.333984375
],
[
0.0849609375,
0.32763671875,
0.48046875,
0.328857421875,
0.48046875,
0.34033203125,
0.0849609375,
0.339111328125
],
[
0.08660888671875,
0.34228515625,
0.4794921875,
0.34228515625,
0.4794921875,
0.352783203125,
0.08660888671875,
0.352783203125
],
[
0.515625,
0.337158203125,
0.9140625,
0.33837890625,
0.9140625,
0.352783203125,
0.515625,
0.3515625
],
[
0.0849609375,
0.35595703125,
0.483642578125,
0.35595703125,
0.483642578125,
0.369384765625,
0.0849609375,
0.369384765625
],
[
0.5185546875,
0.352783203125,
0.9140625,
0.352783203125,
0.9140625,
0.3662109375,
0.5185546875,
0.3662109375
],
[
0.08660888671875,
0.371826171875,
0.48046875,
0.371826171875,
0.48046875,
0.3818359375,
0.08660888671875,
0.3818359375
],
[
0.51708984375,
0.3662109375,
0.9140625,
0.3662109375,
0.9140625,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.0849609375,
0.38330078125,
0.483642578125,
0.38330078125,
0.483642578125,
0.396484375,
0.0849609375,
0.396484375
],
[
0.51708984375,
0.37890625,
0.9140625,
0.37890625,
0.9140625,
0.393310546875,
0.51708984375,
0.393310546875
],
[
0.0849609375,
0.397705078125,
0.48193359375,
0.397705078125,
0.48193359375,
0.410888671875,
0.0849609375,
0.410888671875
],
[
0.51708984375,
0.39453125,
0.9140625,
0.39453125,
0.9140625,
0.40771484375,
0.51708984375,
0.40771484375
],
[
0.0849609375,
0.41357421875,
0.48046875,
0.41357421875,
0.48046875,
0.423583984375,
0.0849609375,
0.423583984375
],
[
0.5185546875,
0.410400390625,
0.91259765625,
0.410400390625,
0.91259765625,
0.42041015625,
0.5185546875,
0.42041015625
],
[
0.0849609375,
0.4267578125,
0.441162109375,
0.4267578125,
0.441162109375,
0.4375,
0.0849609375,
0.4375
],
[
0.51708984375,
0.421630859375,
0.9140625,
0.421630859375,
0.9140625,
0.43505859375,
0.51708984375,
0.43505859375
],
[
0.1004638671875,
0.439453125,
0.48046875,
0.439453125,
0.48046875,
0.45263671875,
0.1004638671875,
0.45263671875
],
[
0.5185546875,
0.436279296875,
0.9140625,
0.436279296875,
0.9140625,
0.449462890625,
0.5185546875,
0.449462890625
],
[
0.08416748046875,
0.451904296875,
0.4794921875,
0.45263671875,
0.4794921875,
0.466552734375,
0.08416748046875,
0.46533203125
],
[
0.51708984375,
0.448974609375,
0.90625,
0.449462890625,
0.90625,
0.46337890625,
0.51708984375,
0.462158203125
],
[
0.08416748046875,
0.466552734375,
0.48046875,
0.466552734375,
0.48046875,
0.479736328125,
0.08416748046875,
0.479736328125
],
[
0.615234375,
0.46728515625,
0.81640625,
0.468505859375,
0.81640625,
0.48291015625,
0.615234375,
0.481689453125
],
[
0.89453125,
0.4697265625,
0.9150390625,
0.4697265625,
0.9150390625,
0.481689453125,
0.89453125,
0.481689453125
],
[
0.0849609375,
0.48095703125,
0.48193359375,
0.48095703125,
0.48193359375,
0.494384765625,
0.0849609375,
0.494384765625
],
[
0.08416748046875,
0.49365234375,
0.48193359375,
0.494384765625,
0.48193359375,
0.50830078125,
0.08416748046875,
0.5068359375
],
[
0.51708984375,
0.488037109375,
0.9140625,
0.488037109375,
0.9140625,
0.501953125,
0.51708984375,
0.501953125
],
[
0.08416748046875,
0.50830078125,
0.483642578125,
0.50830078125,
0.483642578125,
0.521484375,
0.08416748046875,
0.521484375
],
[
0.5185546875,
0.50390625,
0.9111328125,
0.50390625,
0.9111328125,
0.5146484375,
0.5185546875,
0.5146484375
],
[
0.08416748046875,
0.521484375,
0.346435546875,
0.521484375,
0.346435546875,
0.53515625,
0.08416748046875,
0.53515625
],
[
0.51708984375,
0.5166015625,
0.91259765625,
0.5166015625,
0.91259765625,
0.52978515625,
0.51708984375,
0.52978515625
],
[
0.51708984375,
0.52978515625,
0.9140625,
0.52978515625,
0.9140625,
0.54345703125,
0.51708984375,
0.54345703125
],
[
0.51708984375,
0.54345703125,
0.91259765625,
0.54345703125,
0.91259765625,
0.556640625,
0.51708984375,
0.556640625
],
[
0.0849609375,
0.5498046875,
0.439453125,
0.5498046875,
0.439453125,
0.56298828125,
0.0849609375,
0.56298828125
],
[
0.51708984375,
0.55615234375,
0.9140625,
0.55615234375,
0.9140625,
0.5693359375,
0.51708984375,
0.5693359375
],
[
0.0849609375,
0.568359375,
0.48193359375,
0.568359375,
0.48193359375,
0.58203125,
0.0849609375,
0.58203125
],
[
0.515625,
0.5712890625,
0.9140625,
0.5712890625,
0.9140625,
0.58544921875,
0.515625,
0.58544921875
],
[
0.0849609375,
0.58203125,
0.48193359375,
0.58203125,
0.48193359375,
0.59521484375,
0.0849609375,
0.59521484375
],
[
0.5146484375,
0.58544921875,
0.73779296875,
0.58544921875,
0.73779296875,
0.5986328125,
0.5146484375,
0.5986328125
],
[
0.0849609375,
0.5966796875,
0.48193359375,
0.5966796875,
0.48193359375,
0.60986328125,
0.0849609375,
0.60986328125
],
[
0.0849609375,
0.60986328125,
0.48193359375,
0.60986328125,
0.48193359375,
0.62353515625,
0.0849609375,
0.62353515625
],
[
0.515625,
0.611328125,
0.6806640625,
0.609375,
0.6806640625,
0.62353515625,
0.515625,
0.62548828125
],
[
0.0849609375,
0.62548828125,
0.48046875,
0.62548828125,
0.48046875,
0.63623046875,
0.0849609375,
0.63623046875
],
[
0.515625,
0.62744140625,
0.9140625,
0.62890625,
0.9140625,
0.64306640625,
0.515625,
0.642578125
],
[
0.0849609375,
0.63818359375,
0.365234375,
0.63720703125,
0.365234375,
0.6484375,
0.0849609375,
0.6494140625
],
[
0.5185546875,
0.64453125,
0.9111328125,
0.64453125,
0.9111328125,
0.65478515625,
0.5185546875,
0.65478515625
],
[
0.101318359375,
0.65087890625,
0.48193359375,
0.65087890625,
0.48193359375,
0.6640625,
0.101318359375,
0.6640625
],
[
0.51708984375,
0.6572265625,
0.9140625,
0.6572265625,
0.9140625,
0.67041015625,
0.51708984375,
0.67041015625
],
[
0.0849609375,
0.66552734375,
0.47802734375,
0.66552734375,
0.47802734375,
0.6796875,
0.0849609375,
0.6796875
],
[
0.5185546875,
0.67236328125,
0.9111328125,
0.67236328125,
0.9111328125,
0.68310546875,
0.5185546875,
0.68310546875
],
[
0.0849609375,
0.6787109375,
0.48046875,
0.6787109375,
0.48046875,
0.6923828125,
0.0849609375,
0.6923828125
],
[
0.5185546875,
0.68505859375,
0.91259765625,
0.68505859375,
0.91259765625,
0.69580078125,
0.5185546875,
0.69580078125
],
[
0.08416748046875,
0.6923828125,
0.48193359375,
0.69140625,
0.48193359375,
0.70556640625,
0.08416748046875,
0.70703125
],
[
0.51708984375,
0.69873046875,
0.9140625,
0.69873046875,
0.9140625,
0.7119140625,
0.51708984375,
0.7119140625
],
[
0.0849609375,
0.70703125,
0.48193359375,
0.70703125,
0.48193359375,
0.72021484375,
0.0849609375,
0.72021484375
],
[
0.5185546875,
0.7138671875,
0.91259765625,
0.7138671875,
0.91259765625,
0.724609375,
0.5185546875,
0.724609375
],
[
0.08660888671875,
0.72265625,
0.48046875,
0.72265625,
0.48046875,
0.73291015625,
0.08660888671875,
0.73291015625
],
[
0.515625,
0.7265625,
0.9140625,
0.72607421875,
0.9140625,
0.7392578125,
0.515625,
0.74072265625
],
[
0.08416748046875,
0.734375,
0.461669921875,
0.734375,
0.461669921875,
0.74755859375,
0.08416748046875,
0.74755859375
],
[
0.515625,
0.74072265625,
0.64892578125,
0.74072265625,
0.64892578125,
0.75390625,
0.515625,
0.75390625
],
[
0.53173828125,
0.75390625,
0.9140625,
0.75390625,
0.9140625,
0.767578125,
0.53173828125,
0.767578125
],
[
0.0849609375,
0.76123046875,
0.43701171875,
0.76220703125,
0.43701171875,
0.7763671875,
0.0849609375,
0.77587890625
],
[
0.515625,
0.76806640625,
0.80224609375,
0.76806640625,
0.80224609375,
0.7822265625,
0.515625,
0.7822265625
],
[
0.08416748046875,
0.78076171875,
0.483642578125,
0.78076171875,
0.483642578125,
0.79443359375,
0.08416748046875,
0.79443359375
],
[
0.7060546875,
0.787109375,
0.71826171875,
0.787109375,
0.71826171875,
0.79541015625,
0.7060546875,
0.79541015625
],
[
0.0849609375,
0.79541015625,
0.483642578125,
0.79541015625,
0.483642578125,
0.8095703125,
0.0849609375,
0.8095703125
],
[
0.5986328125,
0.79736328125,
0.68701171875,
0.79736328125,
0.68701171875,
0.8125,
0.5986328125,
0.8125
],
[
0.71142578125,
0.798828125,
0.8310546875,
0.798828125,
0.8310546875,
0.81298828125,
0.71142578125,
0.81298828125
],
[
0.89453125,
0.798828125,
0.9150390625,
0.798828125,
0.9150390625,
0.8125,
0.89453125,
0.8125
],
[
0.0849609375,
0.8095703125,
0.483642578125,
0.8095703125,
0.483642578125,
0.82275390625,
0.0849609375,
0.82275390625
],
[
0.705078125,
0.81298828125,
0.71826171875,
0.81298828125,
0.71826171875,
0.82275390625,
0.705078125,
0.82275390625
],
[
0.08660888671875,
0.82470703125,
0.48046875,
0.82470703125,
0.48046875,
0.83544921875,
0.08660888671875,
0.83544921875
],
[
0.51708984375,
0.8271484375,
0.9140625,
0.8271484375,
0.9140625,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.0849609375,
0.83837890625,
0.3896484375,
0.83837890625,
0.3896484375,
0.8486328125,
0.0849609375,
0.8486328125
],
[
0.515625,
0.84130859375,
0.6396484375,
0.84228515625,
0.6396484375,
0.85400390625,
0.515625,
0.85302734375
],
[
0.11273193359375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.11273193359375,
0.86865234375
],
[
0.53173828125,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.869140625,
0.53173828125,
0.869140625
],
[
0.12335205078125,
0.869140625,
0.2181396484375,
0.87060546875,
0.2181396484375,
0.8818359375,
0.12335205078125,
0.880859375
],
[
0.51708984375,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.11273193359375,
0.88330078125,
0.294189453125,
0.88330078125,
0.294189453125,
0.896484375,
0.11273193359375,
0.896484375
],
[
0.51708984375,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"www24,May137,224, Singapore,Singapor",
"Yamashita et al.",
"Table 1: Dataset Statistics",
"Each token within x* is derived based on a conditional probabil-",
"ity function at every time-step t until it reaches the predetermined",
"Tech",
"Business",
"maximum sequence length T. This process can be formally repre-",
"#ofresumes",
"10,017",
"10,373",
"sented by:",
"# ofunique companies",
"11,679",
"12,144",
"x=G(xpast,z; OG)",
"(1)",
"Here, O denotes the learnable parameters intrinsic to the generator",
"within their current position types [43], and recruiters typically",
"model G. The training objective for G can be modified to incorpo-",
"seek candidatesforspecificroles from ones havingsimilar experi",
"rate this conditional generation. Consequently, our initial objective",
"ences, we tailored our dataset selection towards two domains-the",
"function evolves to:",
"technology (Tech)and business (Business)sectors.",
"1",
"N",
"L(f(G(xpast,i, z; 0G)), yi),",
"(2)",
"min",
"To construct datasets encompassing positions within these two",
"categories,we initiated two-step pre-processing.First, we standard",
"",
"i=1",
"ized job titles in all resumes using a job title mapping model [36],",
"where Xpast,i symbolizes the previously generated career history",
"whichtranslatesjobtitlesintostandardizedESCO-based position",
"corresponding to the ith sample in the dataset.",
"names [8]. Second, by leveraging ESCO skill definitions [14], we",
"filtered out positions to retain only those pertinent to technology",
"4.2RealityRegulation",
"and business sectors.To further refine our data, we filtered out",
"companies that only appeared once in ourresume dataset.After",
"We design a reality regulation function. To fabricate convincing",
"this pre-processing, we obtained the datasets shown in Table 1. As",
"synthetic career trajectories, our approach ensures fidelity to an",
"ourdatasetalsoincludesthe companysgeneral information (e.g.,#",
"underlying graph structure. Following state-of-the-art studies on",
"of employees), we label companies with less than 200 employees as",
"formulating job transition graph [24, 35, 40, 43], we create a graph",
"\"Small\" and companies with more than 10,000 employees as\"Large\"",
"consisting the user's job transitions, in which nodes represent com-",
"and use them as target companies in our attacks.For the statistics",
"panies andedgesarecompany-companytransitionsasshownin",
"of companiespertheirsizesinourdataset,seeAppendixB",
"Figure 3. For generating a career path, each job in the sequence",
"For ethical considerations, note that any personal identifiable",
"should be adjacent or reachable within n walking steps on the graph.",
"information (PIl)in thedataset has beenanonymized,retaining only",
"This adjacency constraint can be mathematically represented as:",
"career trajectories for our experiments. While we cannot publicly",
"Vci,c, x*:distance(ci,cj)≤n",
"(3)",
"release our dataset due to its commercial nature, we are open to",
"sharing our dataset for research purposes upon valid requests (e.g.,",
"where distance(ci,c;) computes the shortest path length between",
"MOU signed). See Appendix A for a detailed description of our",
"two company nodes ciand cin the graph.Table 2presents node",
"dataset collection and the ethics statement.",
"degrees of large and small companies using our datasets. Average",
"degree oflarge Tech companies is 42.89, oflarge Business companies",
"is 36.10. Average degree of small Tech/Business companies remain",
"4FAKERESUMEATTACKFRAMEWORK",
"at around 4 and the average degree of all Tech/Business companies",
"In this section, we propose an end-to-end fake resume generation",
"keep at around 8. The average node degree in the graph varies",
"framework FRANCIS that induces systematic prediction errors via",
" between large and small companies.",
"data poisoning. In our scenario, the attacker develops an adversarial",
"resume generator that produces a fake resume dataset D*. When a",
"4.3AttackModule",
"modelis trained with D* in addition to the original data, it assists",
"We design an attack module to manipulate the adversarial gen-",
"the attacker in achieving thedesired behavior.",
"eratorGtogeneratesyntheticresumesthatintentionallyimpact",
"Career prediction models aim to forecast the next career a person",
"the results of the career prediction model. We follow a black box",
"may hold based on their professional history. Let D = ((xi, yi))",
"strategy rather than a white box approach as the black box strategy",
"be a dataset containing N samples of career history data xi and",
"is more realistic (ie.,the victim model is untouchable) and does",
"the corresponding next companies yi, where yi belongs to a set",
"not require a transparent understanding about the victim model.",
"of M possible companies y. We denote f : x - y as the career",
"Assuch,wedesignour surrogatemodelforcareerpredictionto",
"predictionmodel, parameterized by f.To attack thismodel, our",
"produce synthetic resumes that are then utilized by the actual and",
"fake resume attack approach comprises three unique modules.",
"unseen victim model.",
"Our surrogate model f predicts an individual's subsequent job,",
"4.1 Probabilistic JobTrajectory Generator",
"aiming to optimize the following loss function:",
"We design a conditional probabilistic job trajectory generator, de-",
"C",
"noted as G(xpast, z), tailored for career history data. The generator,",
"L(f(xi:0p)) =",
"yiclog(fe(xi:0g))",
"(4)",
"G takes the past career history data of a user as input and gener-",
"三",
"ates synthetic career data, x*,one token at a time.Thegeneration",
"where C is the number of companies and f is the predicted proba-",
"procedureis contingentontwoprimaryelements:",
"bilityof companyc.",
" The career history generated up to the current point, repre-",
"For leveraging f to guide G, we use backpropagation signals",
"sented asXpast.",
"from f. The aim is for G to generate a new resume, x*, such that",
". A random latent variable, z.",
"f(xi) (with a perturbation y*) results in a targeted prediction label"
] |
{
"arxiv_id": "2402.14124",
"language": "en",
"page_idx": 3
}
|
||
2402.14124-en-4
|
2402.14124-en-4.jpg
|
Fake Resume Attacks: Data Poisoning on Online Job Platforms
Table 2: Average degree of a job transition graph.
<table>
<tbody>
<tr>
<th>Company Category $(\#$ of employees)</th>
<th>Tech</th>
<th>Business</th>
</tr>
<tr>
<td>$\overline{{\mathrm{All~companies}}}$</td>
<td>9.37</td>
<td>8.49</td>
</tr>
<tr>
<td>Large companies ( > 10k employees) </td>
<td>42.89</td>
<td>36.10</td>
</tr>
<tr>
<td>Small companies ( < = 200 employees) </td>
<td>4.72</td>
<td>4.60</td>
</tr>
</tbody>
</table>
<FigureHere>
Figure 3: An example of a job transition graph.
$L^*$ from the set of companies for $x_i.$ Our optimization objective for
this function is:
(5)
$$\min_{\theta_G}L^*(f(G(\mathbf{x}_i;\theta_G)))$$
### 4.4 0bjective Function
We define objective functions in accordance with our three distinct attack scenarios. The attacker's goal is to craft realistic fake resumes that target the surrogate model by optimizing the objective function pertinent to each scenario.
Company Promotion Attack: In this scenario, the objective is to maximize the likelihood of target companies being predicted across as many users as possible.
$$L_{\text{promotion}}=-\frac{1}{N}\sum_{i=1}^{N}\sum_{j\in T}P_{ij}$$
(6)
Company Demotion Attack: The goal here is to minimize the likelihood of target companies in the surrogate model's predictions.
(7)
$$L_{\text{demotion}}=\frac{1}{N}\sum_{i=1}^{N}\sum_{j\in T}P_{ij}$$
User Promotion Attack: This attack aims to maximize the likelihood of specific users (or resumes) being associated with target companies, optimizing over a select group of users, denoted as $\mathcal{U}.$
(8)
$$L_\text{user-promotion}=-\frac{1}{U}\sum_{i\in U}\sum_{j\in T}P_{ij}$$
# 4.5 Surrogate Model
For the career prediction task, we adopt an RNN model with several state-of-the-art models. Following [16], we employ LSTM architecture to capture intricate patterns in job transitions that could indicate a user's future career shift
$\begin{array}{c}{\mathrm{Consider~the~dataset~}\mathcal{D}=\{(x_{i},y_{i})\}_{i=1}^{N},\mathrm{~where~each~}x_{i}\mathrm{~is~a~se-}}\\\end{array}$ quence of companies, and $y_i$ is the next potential company. We have a function $\mathcal{F}(x)$ representing a multiclass classification model. Trained with the categorical cross-entropy loss, it provides a prediction probability $\hat{y}$ as:
(9)
$$\hat{y}=\mathcal{F}(x;\theta_{f})$$
WWW'24, May 13-17, 2024, Singapore, Singapore
# where $\theta_f$ is the surrogate model's parameters. Following the prediction, the top-$k$ predicted companies arez
(10)
$$\mathbf{y}_{\mathrm{top-k}}=\mathrm{TopK}(\hat{y})$$
Our LSTM configuration consists of two layers with 128 units in the first and 64 in the second, a dropout layer (rate of 0.5), and the Adam optimizer for loss function optimization.
# 5 EVALUATION
In this section, we discuss the evaluation results of FRANCIS and the baseline models using a real-world dataset, as detailed in Section 3.4. Our evaluation seeks to address the following Research Questions (RQ):
(1) RQ1: Is it feasible to poison career prediction models?
(2) RQ2: How does our fake resume attack perform against
baseline approaches?
(3) RQ3: To what extent does injecting fake resumes affect the
performance of career prediction?
## 5.1 Evaluation Protocol
5.1.1 Attack Performance. To address RQ1 and RQ2, we evaluate the efficacy of our attack to various target models, as follows.
Degree of Attack Success: In the context of data poisoning attacks in career prediction for online job platforms, it is important to measure how well the attacks promote or demote the target in order to measure the success rate of the attack. For this, we use the Improvement Rate (IR) of the average target Hit Ratio (i.e., HR) in the original model as our measure. The improvement rate IR is defined as the increase in $HR$ after data injection over the $HR$ before data injection, as follows:
(11)
$$\mathrm{IR@}k=\frac{\mathrm{HR@k}_{\mathrm{after}}}{\mathrm{HR@k}_{\mathrm{before}}}$$
This gives us an indication of how much we are able to manipulate the visibility of the target through data poisoning. For a detailed explanation and intuition of this metric, see Appendix D. We vary the injection ratio in our experiments to discern its impact on the attack's success. Following previous studies [39], we set $k=10.$
Target Company and User Selection: In the company promotion and demotion attacks, we randomly sample 100 companies from $“$Small",“Large”, and random companies on our dataset(see Section 3.4 for the company definition), and measure the average $IR@10$ for the target companies. In the user promotion attack, we set “Large”companies as target companies assuming that some users want to get an interview or any recruitment opportunity for top companies competing with other job seekers, and extract users from those who never experienced “Large”companies (we name these as“Specific” users) or sample 20% users from all users as the target users (we name this as “Random”users). Afterward, we see the average HR@10 for the target companies in the target users. Target Victim Models: To attack the career prediction models, we set the three state-of-the-art models as target victim models: NEMO[16], AHEAD[44], and NAOMI [35]. All three models are designed and experienced in the data from online job platforms or real-world resumes.
|
[
[
0.08660888671875,
0.0770263671875,
0.38232421875,
0.0770263671875,
0.38232421875,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6708984375,
0.0738525390625,
0.91259765625,
0.07574462890625,
0.91259765625,
0.0889892578125,
0.6708984375,
0.08709716796875
],
[
0.11767578125,
0.1029052734375,
0.4453125,
0.10479736328125,
0.4453125,
0.11932373046875,
0.11767578125,
0.117431640625
],
[
0.515625,
0.1060791015625,
0.9140625,
0.106689453125,
0.9140625,
0.121826171875,
0.515625,
0.12060546875
],
[
0.51708984375,
0.121826171875,
0.779296875,
0.121826171875,
0.779296875,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.09234619140625,
0.1287841796875,
0.3203125,
0.1287841796875,
0.3203125,
0.1427001953125,
0.09234619140625,
0.1427001953125
],
[
0.346435546875,
0.1287841796875,
0.378173828125,
0.1287841796875,
0.378173828125,
0.141357421875,
0.346435546875,
0.141357421875
],
[
0.404296875,
0.1287841796875,
0.4697265625,
0.1287841796875,
0.4697265625,
0.14013671875,
0.404296875,
0.14013671875
],
[
0.0906982421875,
0.1458740234375,
0.1829833984375,
0.146484375,
0.1829833984375,
0.1578369140625,
0.0906982421875,
0.1572265625
],
[
0.34716796875,
0.1458740234375,
0.375732421875,
0.1458740234375,
0.375732421875,
0.1578369140625,
0.34716796875,
0.1578369140625
],
[
0.4248046875,
0.1458740234375,
0.453369140625,
0.1458740234375,
0.453369140625,
0.1578369140625,
0.4248046875,
0.1578369140625
],
[
0.66015625,
0.1458740234375,
0.7685546875,
0.1427001953125,
0.7685546875,
0.1572265625,
0.66015625,
0.160400390625
],
[
0.88818359375,
0.14453125,
0.9150390625,
0.14453125,
0.9150390625,
0.1572265625,
0.88818359375,
0.1572265625
],
[
0.08905029296875,
0.160400390625,
0.306396484375,
0.160400390625,
0.306396484375,
0.173583984375,
0.08905029296875,
0.173583984375
],
[
0.346435546875,
0.160400390625,
0.378173828125,
0.160400390625,
0.378173828125,
0.17236328125,
0.346435546875,
0.17236328125
],
[
0.42236328125,
0.160400390625,
0.455810546875,
0.160400390625,
0.455810546875,
0.17236328125,
0.42236328125,
0.17236328125
],
[
0.08905029296875,
0.17236328125,
0.31201171875,
0.173583984375,
0.31201171875,
0.1881103515625,
0.08905029296875,
0.1875
],
[
0.53369140625,
0.1666259765625,
0.9140625,
0.1666259765625,
0.9140625,
0.179931640625,
0.53369140625,
0.179931640625
],
[
0.34716796875,
0.1749267578125,
0.375732421875,
0.1749267578125,
0.375732421875,
0.1875,
0.34716796875,
0.1875
],
[
0.4248046875,
0.1749267578125,
0.451904296875,
0.1749267578125,
0.451904296875,
0.1875,
0.4248046875,
0.1875
],
[
0.51708984375,
0.1787109375,
0.9140625,
0.1787109375,
0.9140625,
0.1925048828125,
0.51708984375,
0.1925048828125
],
[
0.51708984375,
0.1944580078125,
0.7998046875,
0.1944580078125,
0.7998046875,
0.2083740234375,
0.51708984375,
0.2083740234375
],
[
0.51708984375,
0.2215576171875,
0.66015625,
0.2215576171875,
0.66015625,
0.23486328125,
0.51708984375,
0.23486328125
],
[
0.51708984375,
0.2412109375,
0.91259765625,
0.2412109375,
0.91259765625,
0.251953125,
0.51708984375,
0.251953125
],
[
0.51708984375,
0.25390625,
0.9140625,
0.25390625,
0.9140625,
0.267578125,
0.51708984375,
0.267578125
],
[
0.515625,
0.268310546875,
0.9111328125,
0.268310546875,
0.9111328125,
0.279052734375,
0.515625,
0.279052734375
],
[
0.5185546875,
0.28271484375,
0.61376953125,
0.28271484375,
0.61376953125,
0.294189453125,
0.5185546875,
0.294189453125
],
[
0.5361328125,
0.2998046875,
0.8857421875,
0.2998046875,
0.8857421875,
0.313232421875,
0.5361328125,
0.313232421875
],
[
0.12744140625,
0.313232421875,
0.439453125,
0.313232421875,
0.439453125,
0.326904296875,
0.12744140625,
0.326904296875
],
[
0.53515625,
0.313232421875,
0.9140625,
0.313232421875,
0.9140625,
0.326904296875,
0.53515625,
0.326904296875
],
[
0.55810546875,
0.328857421875,
0.68701171875,
0.328857421875,
0.68701171875,
0.339111328125,
0.55810546875,
0.339111328125
],
[
0.0849609375,
0.341552734375,
0.48193359375,
0.34228515625,
0.48193359375,
0.356689453125,
0.0849609375,
0.35595703125
],
[
0.53515625,
0.34033203125,
0.9140625,
0.34033203125,
0.9140625,
0.354248046875,
0.53515625,
0.354248046875
],
[
0.08660888671875,
0.35791015625,
0.1805419921875,
0.35791015625,
0.1805419921875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.55615234375,
0.354736328125,
0.765625,
0.354736328125,
0.765625,
0.36865234375,
0.55615234375,
0.36865234375
],
[
0.2156982421875,
0.369384765625,
0.348876953125,
0.37060546875,
0.348876953125,
0.385009765625,
0.2156982421875,
0.3837890625
],
[
0.463134765625,
0.371826171875,
0.484375,
0.371826171875,
0.484375,
0.385009765625,
0.463134765625,
0.385009765625
],
[
0.2156982421875,
0.380615234375,
0.238525390625,
0.37744140625,
0.24267578125,
0.391357421875,
0.2197265625,
0.39453125
],
[
0.51708984375,
0.38330078125,
0.72119140625,
0.38330078125,
0.72119140625,
0.396484375,
0.51708984375,
0.396484375
],
[
0.1094970703125,
0.404052734375,
0.281005859375,
0.40478515625,
0.281005859375,
0.419189453125,
0.1094970703125,
0.41845703125
],
[
0.51708984375,
0.40087890625,
0.9140625,
0.40087890625,
0.9140625,
0.4140625,
0.51708984375,
0.4140625
],
[
0.08905029296875,
0.40771484375,
0.1168212890625,
0.40771484375,
0.1168212890625,
0.416748046875,
0.08905029296875,
0.416748046875
],
[
0.51708984375,
0.415283203125,
0.90869140625,
0.415283203125,
0.90869140625,
0.4287109375,
0.51708984375,
0.4287109375
],
[
0.0849609375,
0.42236328125,
0.48193359375,
0.42236328125,
0.48193359375,
0.436279296875,
0.0849609375,
0.436279296875
],
[
0.0849609375,
0.4375,
0.483642578125,
0.4375,
0.483642578125,
0.45068359375,
0.0849609375,
0.45068359375
],
[
0.51708984375,
0.43310546875,
0.9140625,
0.43310546875,
0.9140625,
0.4462890625,
0.51708984375,
0.4462890625
],
[
0.0849609375,
0.45068359375,
0.483642578125,
0.45068359375,
0.483642578125,
0.464111328125,
0.0849609375,
0.464111328125
],
[
0.51708984375,
0.447509765625,
0.91259765625,
0.447509765625,
0.91259765625,
0.458251953125,
0.51708984375,
0.458251953125
],
[
0.08416748046875,
0.46533203125,
0.2467041015625,
0.464111328125,
0.2467041015625,
0.477783203125,
0.08416748046875,
0.479248046875
],
[
0.51708984375,
0.460205078125,
0.91259765625,
0.460205078125,
0.91259765625,
0.473388671875,
0.51708984375,
0.473388671875
],
[
0.51708984375,
0.47607421875,
0.91259765625,
0.47607421875,
0.91259765625,
0.486083984375,
0.51708984375,
0.486083984375
],
[
0.08660888671875,
0.484130859375,
0.48193359375,
0.484130859375,
0.48193359375,
0.494384765625,
0.08660888671875,
0.494384765625
],
[
0.5185546875,
0.4892578125,
0.91259765625,
0.4892578125,
0.91259765625,
0.5,
0.5185546875,
0.5
],
[
0.08416748046875,
0.496826171875,
0.483642578125,
0.496826171875,
0.483642578125,
0.51025390625,
0.08416748046875,
0.51025390625
],
[
0.51708984375,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.51708984375,
0.51513671875
],
[
0.08416748046875,
0.51025390625,
0.24267578125,
0.5087890625,
0.24267578125,
0.52294921875,
0.08416748046875,
0.5234375
],
[
0.515625,
0.51513671875,
0.9140625,
0.51513671875,
0.9140625,
0.52880859375,
0.515625,
0.52880859375
],
[
0.301513671875,
0.52783203125,
0.371826171875,
0.5390625,
0.366943359375,
0.55810546875,
0.295654296875,
0.546875
],
[
0.306396484375,
0.52880859375,
0.3203125,
0.52880859375,
0.3203125,
0.537109375,
0.306396484375,
0.537109375
],
[
0.51708984375,
0.53076171875,
0.7138671875,
0.53076171875,
0.7138671875,
0.54150390625,
0.51708984375,
0.54150390625
],
[
0.1968994140625,
0.54052734375,
0.273681640625,
0.54248046875,
0.273681640625,
0.556640625,
0.1968994140625,
0.5546875
],
[
0.2900390625,
0.53271484375,
0.3056640625,
0.53515625,
0.303955078125,
0.54541015625,
0.28759765625,
0.54248046875
],
[
0.463134765625,
0.54052734375,
0.483642578125,
0.54052734375,
0.483642578125,
0.5537109375,
0.463134765625,
0.5537109375
],
[
0.285888671875,
0.548828125,
0.310546875,
0.548828125,
0.310546875,
0.56005859375,
0.285888671875,
0.56005859375
],
[
0.7060546875,
0.54541015625,
0.77783203125,
0.5517578125,
0.775390625,
0.56640625,
0.70361328125,
0.56005859375
],
[
0.314453125,
0.556640625,
0.3447265625,
0.55615234375,
0.3447265625,
0.56298828125,
0.314453125,
0.564453125
],
[
0.64892578125,
0.55615234375,
0.70166015625,
0.55615234375,
0.70166015625,
0.57080078125,
0.64892578125,
0.57080078125
],
[
0.69775390625,
0.5625,
0.77783203125,
0.5673828125,
0.77685546875,
0.58203125,
0.69677734375,
0.5771484375
],
[
0.88818359375,
0.556640625,
0.91650390625,
0.556640625,
0.91650390625,
0.5712890625,
0.88818359375,
0.5712890625
],
[
0.08660888671875,
0.57080078125,
0.48193359375,
0.57080078125,
0.48193359375,
0.583984375,
0.08660888671875,
0.583984375
],
[
0.0849609375,
0.58544921875,
0.483642578125,
0.58544921875,
0.483642578125,
0.5986328125,
0.0849609375,
0.5986328125
],
[
0.515625,
0.58349609375,
0.91259765625,
0.58544921875,
0.91259765625,
0.5986328125,
0.515625,
0.5966796875
],
[
0.301513671875,
0.60400390625,
0.313720703125,
0.60400390625,
0.313720703125,
0.61279296875,
0.301513671875,
0.61279296875
],
[
0.51708984375,
0.5986328125,
0.9140625,
0.5986328125,
0.9140625,
0.6123046875,
0.51708984375,
0.6123046875
],
[
0.28515625,
0.611328125,
0.2998046875,
0.611328125,
0.2998046875,
0.61865234375,
0.28515625,
0.61865234375
],
[
0.31787109375,
0.60986328125,
0.361083984375,
0.6142578125,
0.358642578125,
0.6318359375,
0.314453125,
0.62744140625
],
[
0.51708984375,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.51708984375,
0.626953125
],
[
0.2998046875,
0.6142578125,
0.32177734375,
0.6142578125,
0.32177734375,
0.6220703125,
0.2998046875,
0.6220703125
],
[
0.2059326171875,
0.6142578125,
0.27783203125,
0.6171875,
0.27685546875,
0.63330078125,
0.205078125,
0.6298828125
],
[
0.463134765625,
0.61572265625,
0.484375,
0.61572265625,
0.484375,
0.62890625,
0.463134765625,
0.62890625
],
[
0.281005859375,
0.62255859375,
0.31201171875,
0.62255859375,
0.31201171875,
0.6337890625,
0.281005859375,
0.6337890625
],
[
0.2998046875,
0.6298828125,
0.339111328125,
0.6298828125,
0.339111328125,
0.64013671875,
0.2998046875,
0.64013671875
],
[
0.51708984375,
0.626953125,
0.9140625,
0.626953125,
0.9140625,
0.64013671875,
0.51708984375,
0.64013671875
],
[
0.51708984375,
0.6396484375,
0.8896484375,
0.6396484375,
0.8896484375,
0.65283203125,
0.51708984375,
0.65283203125
],
[
0.08660888671875,
0.64599609375,
0.483642578125,
0.64599609375,
0.483642578125,
0.6591796875,
0.08660888671875,
0.6591796875
],
[
0.0849609375,
0.66015625,
0.48193359375,
0.66015625,
0.48193359375,
0.673828125,
0.0849609375,
0.673828125
],
[
0.5146484375,
0.65576171875,
0.9140625,
0.6572265625,
0.9140625,
0.671875,
0.5146484375,
0.67041015625
],
[
0.08416748046875,
0.673828125,
0.48193359375,
0.67236328125,
0.48193359375,
0.68603515625,
0.08416748046875,
0.6875
],
[
0.515625,
0.671875,
0.9140625,
0.671875,
0.9140625,
0.68505859375,
0.515625,
0.68505859375
],
[
0.2998046875,
0.68310546875,
0.384033203125,
0.69775390625,
0.378173828125,
0.71630859375,
0.294189453125,
0.7021484375
],
[
0.51708984375,
0.6845703125,
0.91259765625,
0.6845703125,
0.91259765625,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.1846923828125,
0.69775390625,
0.281005859375,
0.70068359375,
0.281005859375,
0.71630859375,
0.1846923828125,
0.71337890625
],
[
0.463134765625,
0.69775390625,
0.483642578125,
0.69775390625,
0.483642578125,
0.7109375,
0.463134765625,
0.7109375
],
[
0.515625,
0.69775390625,
0.9140625,
0.69873046875,
0.9140625,
0.7119140625,
0.515625,
0.7109375
],
[
0.298095703125,
0.70556640625,
0.32275390625,
0.70556640625,
0.32275390625,
0.71728515625,
0.298095703125,
0.71728515625
],
[
0.3203125,
0.70263671875,
0.330078125,
0.70263671875,
0.330078125,
0.70849609375,
0.3203125,
0.70849609375
],
[
0.31201171875,
0.7119140625,
0.361083984375,
0.7119140625,
0.361083984375,
0.7236328125,
0.31201171875,
0.7236328125
],
[
0.51708984375,
0.7119140625,
0.9140625,
0.7119140625,
0.9140625,
0.72607421875,
0.51708984375,
0.72607421875
],
[
0.515625,
0.72607421875,
0.91259765625,
0.72802734375,
0.91259765625,
0.7412109375,
0.515625,
0.7392578125
],
[
0.0849609375,
0.734375,
0.261474609375,
0.73291015625,
0.261474609375,
0.74755859375,
0.0849609375,
0.74853515625
],
[
0.51708984375,
0.7412109375,
0.91259765625,
0.7412109375,
0.91259765625,
0.751953125,
0.51708984375,
0.751953125
],
[
0.0849609375,
0.751953125,
0.48046875,
0.751953125,
0.48046875,
0.76513671875,
0.0849609375,
0.76513671875
],
[
0.515625,
0.7548828125,
0.9140625,
0.7548828125,
0.9140625,
0.76806640625,
0.515625,
0.76806640625
],
[
0.08416748046875,
0.7666015625,
0.483642578125,
0.76513671875,
0.483642578125,
0.77880859375,
0.08416748046875,
0.77978515625
],
[
0.51708984375,
0.76806640625,
0.91259765625,
0.76806640625,
0.91259765625,
0.7822265625,
0.51708984375,
0.7822265625
],
[
0.0849609375,
0.78076171875,
0.483642578125,
0.78076171875,
0.483642578125,
0.79443359375,
0.0849609375,
0.79443359375
],
[
0.515625,
0.7822265625,
0.9140625,
0.7822265625,
0.9140625,
0.79541015625,
0.515625,
0.79541015625
],
[
0.08251953125,
0.79443359375,
0.298095703125,
0.79345703125,
0.298095703125,
0.806640625,
0.08251953125,
0.80810546875
],
[
0.515625,
0.796875,
0.9140625,
0.796875,
0.9140625,
0.81005859375,
0.515625,
0.81005859375
],
[
0.10296630859375,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.82275390625,
0.10296630859375,
0.82275390625
],
[
0.51708984375,
0.81005859375,
0.90283203125,
0.81005859375,
0.90283203125,
0.82373046875,
0.51708984375,
0.82373046875
],
[
0.0849609375,
0.82275390625,
0.48193359375,
0.82275390625,
0.48193359375,
0.8359375,
0.0849609375,
0.8359375
],
[
0.51708984375,
0.8271484375,
0.9140625,
0.8271484375,
0.9140625,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.0849609375,
0.8359375,
0.48046875,
0.8359375,
0.48046875,
0.849609375,
0.0849609375,
0.849609375
],
[
0.515625,
0.84130859375,
0.9150390625,
0.84130859375,
0.9150390625,
0.85498046875,
0.515625,
0.85498046875
],
[
0.08416748046875,
0.8486328125,
0.484375,
0.849609375,
0.484375,
0.8642578125,
0.08416748046875,
0.86279296875
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0849609375,
0.8642578125,
0.2320556640625,
0.8642578125,
0.2320556640625,
0.87744140625,
0.0849609375,
0.87744140625
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.240234375,
0.8798828125,
0.324462890625,
0.8818359375,
0.324462890625,
0.89794921875,
0.240234375,
0.89599609375
],
[
0.463134765625,
0.88330078125,
0.484375,
0.88330078125,
0.484375,
0.896484375,
0.463134765625,
0.896484375
],
[
0.515625,
0.88330078125,
0.640625,
0.88330078125,
0.640625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"Fake Resume Attacks: Data Poisoning on Online Job Platforms",
"www24,May1317,2024, singapore,Singapore",
"Table 2: Average degree of a job transition graph.",
"where , is the surrogate model's parameters.Following the pre-",
"diction, the top-k predicted companies are:",
"Company Category (# of employees)",
"Tech",
"Business",
"All companies",
"9.37",
"8.49",
"ytop-k = TopK(g)",
"(10)",
"Large companies (>10k employees)",
"42.89",
"36.10",
"Smallcompanies (<=200 employees)",
"Our LSTM configuration consists of two layers with 128 units in",
"4.72",
"4.60",
"the first and 64 in the second, a dropout layer (rate of 0.5), and the",
"Adam optimizer for loss function optimization.",
"5 EVALUATION",
"Inthis section,wediscusstheevaluationresults of FRANCIS and",
"the baseline models using a real-world dataset, as detailed in Sec-",
"tion3.4.Ourevaluation seeks toaddress thefollowing Research",
"Questions (RQ):",
"(1) RQ1: Is it feasible to poison career prediction models?",
"Figure 3: An example of a job transition graph.",
"(2) RQ2: How does our fake resume attack perform against",
"baselineapproaches?",
"L* from the set of companies for xi. Our optimization objective for",
"(3) RQ3: To what extent does injecting fake resumes affect the",
"this function is:",
"performance of career prediction?",
"min L*(f(G(xi; OG)",
"(5)",
"OG",
"5.1 EvaluationProtocol",
"ObjectiveFunction",
"5.1.1Attack Performance.To address RQ1 and RQ2, we evalu-",
"4.4",
"ate the efficacy of our attack to various target models, as follows.",
"We define objective functions in accordance with our three distinct",
"attack scenarios. The attacker's goal is to craft realistic fake resumes",
"Degree of Attack Success: In the context of data poisoning attacks",
"that target the surrogate model by optimizing the objective function",
"in careerpredictionforonlinejobplatforms,itis importantto",
"pertinent to each scenario.",
"measure how well the attacks promote or demote the target in",
"order to measure the success rate of the attack.For this, we use",
"CompanyPromotion Attack:In this scenario, the objective is to",
"the Improvement Rate (IR)of theaveragetarget Hit Ratio (i.e., HR)",
"maximize the likelihood of target companies being predicted across",
"in the original model as our measure. The improvement rate IR",
"as manyusers aspossible",
"is defined as the increase in HR after data injection over the HR",
"NEPij",
"N",
"before data injection, as follows:",
"Lpromotion=",
"1",
"(6)",
"N4",
"HR@kafter",
"lieT",
"IR@k =",
"HR@kbefore",
"(11)",
"Company Demotion Attack: The goal here is to minimize the",
"likelihood of target companies in the surrogate model's predictions.",
"This gives us an indication of how much we are able to manipulate",
"N",
"the visibility of the target through data poisoning. For a detailed",
"1",
"Pij",
"explanation and intuition of this metric, see Appendix D. We vary",
"",
"Ldemotion =",
"(7)",
"NL",
"i=l jeT",
"the injection ratio in our experiments to discern its impact on the",
"attack's success. Following previous studies [39], we set k=10.",
"User Promotion Attack: This attack aims to maximize the like-",
"lihood of specific users (or resumes) being associated with target",
"Target Company and User Selection: In the company promotion",
"companies, optimizing over a select group of users, denoted as u.",
"and demotion attacks, we randomly sample 1oo companies from",
"iEPuj",
"\"Small\",\"Large\", and random companies on our dataset (see Section",
"Luser-promotion",
"(8)",
"3.4for the company definition),and measure the average IR@10",
"U ",
"2",
"ieUje",
"for the target companies. In the user promotion attack, we set",
"\"Large\" companies as target companies assuming that some users",
"4.5Surrogate Model",
"want to get an interview or any recruitment opportunity for top",
"For the career prediction task, we adopt an RNN model with several",
"companies competing with other job seekers, and extract users",
"state-of-the-art models. Following [16], we employ LSTM archi-",
"from those who never experienced\"Large\"companies (we name",
"tecture to capture intricate patterns in job transitions that could",
"these as \"Specific\" users) or sample 20% users from all users as the",
"indicate a user's future career shift.",
"target users (we name this as \"Random\" users). Afterward, we see",
"Consider the dataset D = ((xi, yi))j, where each xi is a se-",
"the average HR@10 for the target companies in the target users.",
"quence of companies, and yi is the next potential company. We",
"Target Victim Models: To attack the career prediction models,",
"have a function F(x) representing a multiclass classification model",
"we set the three state-of-the-art models as target victim models:",
"Trained with the categorical cross-entropy loss, it provides a pre-",
"NEMO [16], AHEAD [44], and NAOMI [35]. All three models are",
"diction probability y as:",
"designedandexperiencedinthedatafromonlinejobplatformsor",
"9=F(x;0g)",
"(9)",
"real-world resumes."
] |
{
"arxiv_id": "2402.14124",
"language": "en",
"page_idx": 4
}
|
||
1803.11544-en-0
|
1803.11544-en-0.jpg
|
Guide Me: Interacting with D
Christian Rupprecht$^{1,2,*}$, Iro Laina$^{1,*}$, Nassir Navab$^1$, Gregory D.Hager$^{2,1}$, Federico Tombari$^{1}$
$^*$equal contribution
$^{1}$ Technische Universität München, Munich, Germany
$^{2}$ Johns Hopkins Univers
## $\mathbf{Abstract}$
Interaction and collaboratio ligent machines has become increasingly important as machine learning methods move that involve end users. While much prior work lies at the intersection of natural lang
trained convolutional neural improve its performance duri serting a layer that acts as a spatio-semantic guide into the network. This guide is train vations, either directly via
$\sum _{i= 1}^{\infty }$ $or indirectly through a recurrent model that translates hu- man language queries to interaction weights. Learning thevarbal intaraction is fullv automatic and doas not raauira{aligned}$
verbal interaction is fully automatic and does not require
$\textcircled {\infty }$ manual text annotatio
improve performance, and pro
the interaction between the guide and the CNN.
For example, in photo editin segment the foreground of an image from the background, the user might notice that the network has made a mistake. Instead of manually repairing the segmentation output or developing a post-processing algorithm based on some heuristics, a simpler and more effective way would be for the user to interact directly with the network through a directed hint, $e.g.$ pointing out that“the child on the bottom left of the image is in the foreground, not in the back- $ground’.$ The user that was previously presented with a fixed, black-box prediction is now able to influence and alter the outcome according to his needs. This property is particularly useful in high risk domains such as medical image analysis and computer-assisted diagnosis, where fully automated segmentation is not always robust in clinical applications and the experience of trained practitioners matters. Another relevant example is speeding up labor-intensive and repetitive labeling tasks, such as those needed to create datasets for semantic segmentation, especially those for which annotations are scarce and expensive to obtain.
We propose a novel idea to allow user-network feedbackbased interaction that aims at improving the performance of a pre-trained CNN at test time. The core idea is the definition of a spatio-semantic guiding mechanism that translates user feedback into changes in the internal activations of the network, thus acting as a means of re-thinking the inference process. The user input is modeled via a language-based approach, that enables interaction with a trained model in the form of a dialog. The user receives a first estimate, inputs a text query and the network replies with an updated prediction. Most previous interactive approaches place the user on the input/data side which means user input is required for the method to operate. In contrast, in our method, the user's input is optional and modifes the network, this means that the network does not depend on human interaction but can be adjusted by it.
We showcase this interactive module on the task of semantic image segmentation. One advantage of our method is that it does not depend on any explicit annotation for text-
# 1. Introduction
Convolutional neural networks (CNNs) continue to grow in their breadth of application and in their performance on challenging computer vision tasks, such as image classification, semantic segmentation, object detection, depth prediction and human pose estimation. To date, the majority of the techniques proposed for these applications train specific network architectures once and subsequently deploy them as static components inside an algorithm. However, it is unlikely that any static network will be perfect at the task it was designed for. If the deployed CNN were adaptable to feedback or specifications provided by a human user $on$- $line$, this interaction would hold the potential to improve the model's performance and benefit real-world applications
1
|
[
[
0.26953125,
0.1337890625,
0.701171875,
0.1337890625,
0.701171875,
0.1483154296875,
0.26953125,
0.1483154296875
],
[
0.098876953125,
0.18115234375,
0.87109375,
0.18115234375,
0.87109375,
0.1988525390625,
0.098876953125,
0.1988525390625
],
[
0.404296875,
0.21533203125,
0.56640625,
0.2171630859375,
0.56640625,
0.23486328125,
0.404296875,
0.23291015625
],
[
0.271240234375,
0.231689453125,
0.69677734375,
0.2342529296875,
0.69677734375,
0.251953125,
0.271240234375,
0.25
],
[
0.021240234375,
0.255615234375,
0.058013916015625,
0.255615234375,
0.058013916015625,
0.3984375,
0.021240234375,
0.3984375
],
[
0.285888671875,
0.25,
0.68310546875,
0.25244140625,
0.68310546875,
0.270751953125,
0.285888671875,
0.267578125
],
[
0.236083984375,
0.302978515625,
0.314453125,
0.302978515625,
0.314453125,
0.32080078125,
0.236083984375,
0.32080078125
],
[
0.521484375,
0.30615234375,
0.89208984375,
0.30615234375,
0.89208984375,
0.3193359375,
0.521484375,
0.3193359375
],
[
0.50341796875,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.50341796875,
0.335205078125
],
[
0.098876953125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.098876953125,
0.349853515625
],
[
0.5048828125,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.348388671875,
0.5048828125,
0.348388671875
],
[
0.08087158203125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.08087158203125,
0.365478515625
],
[
0.50244140625,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.08087158203125,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.380126953125,
0.08087158203125,
0.380126953125
],
[
0.5009765625,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.5009765625,
0.380126953125
],
[
0.08251953125,
0.3837890625,
0.46728515625,
0.3837890625,
0.46728515625,
0.39453125,
0.08251953125,
0.39453125
],
[
0.50341796875,
0.3837890625,
0.890625,
0.3837890625,
0.890625,
0.39453125,
0.50341796875,
0.39453125
],
[
0.080078125,
0.397705078125,
0.4697265625,
0.397705078125,
0.4697265625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.397705078125,
0.89208984375,
0.397705078125,
0.89208984375,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.08087158203125,
0.41357421875,
0.471435546875,
0.41357421875,
0.471435546875,
0.4267578125,
0.08087158203125,
0.4267578125
],
[
0.5009765625,
0.412353515625,
0.89208984375,
0.410888671875,
0.89208984375,
0.425537109375,
0.5009765625,
0.4267578125
],
[
0.080078125,
0.427978515625,
0.4697265625,
0.427978515625,
0.4697265625,
0.44140625,
0.080078125,
0.44140625
],
[
0.50244140625,
0.427978515625,
0.890625,
0.427978515625,
0.890625,
0.44140625,
0.50244140625,
0.44140625
],
[
0.080078125,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.443115234375,
0.8935546875,
0.443115234375,
0.8935546875,
0.45703125,
0.50244140625,
0.45703125
],
[
0.03350830078125,
0.449462890625,
0.0555419921875,
0.449462890625,
0.0555419921875,
0.479736328125,
0.03350830078125,
0.479736328125
],
[
0.080078125,
0.458251953125,
0.4697265625,
0.458251953125,
0.4697265625,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.50341796875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.080078125,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.080078125,
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
],
[
0.5009765625,
0.486083984375,
0.89208984375,
0.488037109375,
0.89208984375,
0.501953125,
0.5009765625,
0.5
],
[
0.02777099609375,
0.50244140625,
0.053924560546875,
0.50244140625,
0.053924560546875,
0.54150390625,
0.02777099609375,
0.54150390625
],
[
0.08087158203125,
0.50244140625,
0.468017578125,
0.50244140625,
0.468017578125,
0.5166015625,
0.08087158203125,
0.5166015625
],
[
0.50341796875,
0.50244140625,
0.890625,
0.50244140625,
0.890625,
0.5166015625,
0.50341796875,
0.5166015625
],
[
0.080078125,
0.51806640625,
0.468017578125,
0.51806640625,
0.468017578125,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50244140625,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.0236968994140625,
0.5322265625,
0.056365966796875,
0.5322265625,
0.056365966796875,
0.56494140625,
0.0236968994140625,
0.56494140625
],
[
0.080078125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50341796875,
0.53515625,
0.8896484375,
0.53515625,
0.8896484375,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.080078125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.080078125,
0.5625
],
[
0.50244140625,
0.548828125,
0.89208984375,
0.548828125,
0.89208984375,
0.5625,
0.50244140625,
0.5625
],
[
0.0318603515625,
0.564453125,
0.037567138671875,
0.564453125,
0.037567138671875,
0.5791015625,
0.0318603515625,
0.5791015625
],
[
0.07843017578125,
0.564453125,
0.4697265625,
0.564453125,
0.4697265625,
0.57763671875,
0.07843017578125,
0.57763671875
],
[
0.50244140625,
0.56298828125,
0.890625,
0.56298828125,
0.890625,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.08087158203125,
0.5791015625,
0.4697265625,
0.5791015625,
0.4697265625,
0.59228515625,
0.08087158203125,
0.59228515625
],
[
0.50244140625,
0.5791015625,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.02532958984375,
0.5888671875,
0.056365966796875,
0.5888671875,
0.056365966796875,
0.63818359375,
0.02532958984375,
0.63818359375
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5947265625,
0.85302734375,
0.5947265625,
0.85302734375,
0.60791015625,
0.50244140625,
0.60791015625
],
[
0.080078125,
0.609375,
0.4697265625,
0.609375,
0.4697265625,
0.62255859375,
0.080078125,
0.62255859375
],
[
0.52294921875,
0.6123046875,
0.890625,
0.6123046875,
0.890625,
0.62548828125,
0.52294921875,
0.62548828125
],
[
0.080078125,
0.625,
0.4697265625,
0.625,
0.4697265625,
0.63818359375,
0.080078125,
0.63818359375
],
[
0.021240234375,
0.6318359375,
0.056365966796875,
0.630859375,
0.05963134765625,
0.71044921875,
0.0236968994140625,
0.7109375
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.6396484375,
0.39453125,
0.6396484375,
0.39453125,
0.65283203125,
0.080078125,
0.65283203125
],
[
0.50341796875,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.50244140625,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.08251953125,
0.6796875,
0.20751953125,
0.6796875,
0.20751953125,
0.6943359375,
0.08251953125,
0.6943359375
],
[
0.50341796875,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.098876953125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.098876953125,
0.71826171875
],
[
0.50244140625,
0.7021484375,
0.890625,
0.70068359375,
0.890625,
0.71533203125,
0.50244140625,
0.71630859375
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.50341796875,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.50341796875,
0.73095703125
],
[
0.080078125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.080078125,
0.75
],
[
0.50341796875,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
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
],
[
0.08087158203125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50244140625,
0.76318359375,
0.89208984375,
0.76318359375,
0.89208984375,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.080078125,
0.7822265625,
0.4697265625,
0.7822265625,
0.4697265625,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50341796875,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50341796875,
0.79248046875
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50341796875,
0.806640625
],
[
0.07843017578125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.80810546875,
0.89208984375,
0.80810546875,
0.89208984375,
0.8212890625,
0.50244140625,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82373046875,
0.890625,
0.82373046875,
0.890625,
0.8369140625,
0.5009765625,
0.8369140625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.83837890625,
0.61767578125,
0.83837890625,
0.61767578125,
0.8515625,
0.50244140625,
0.8515625
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.521484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.521484375,
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,
0.455810546875,
0.88720703125,
0.455810546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Guide Me: Interacting with Deep Networks",
"Christian Rupprechtl,2*, Iro Lainal,*, Nassir Navab', Gregory D. Hager2,1, Federico Tombari",
"*equal contribution",
"1 Technische Universitat Munchen, Munich, Germany",
"810",
"2 Johns Hopkins University, Baltimore, MD, USA",
"Abstract",
"For example, in photo editing, when a CNN is used to",
"segment the foreground of an image from the background,",
"Interaction and collaboration between humans and intel-",
"the user might notice that the network has made a mis-",
"ligent machines has become increasingly important as ma-",
"take. Instead of manually repairing the segmentation output",
"chine learning methods move into real-world applications",
"or developing a post-processing algorithm based on some",
"that involve end users.While much prior work lies at the",
"heuristics,a simpler and more effective way wouldbefor",
"intersection of natural language and vision, such as image",
"the user to interact directly with the network through a di-",
"LAD",
"captioning or image generation from text descriptions, less",
"rected hint, e.g. pointing out that \"the child on the bot-",
"focus has been placed on the use of language to guide or im-",
"tom left of the image is in the foreground, not in the back-",
"prove the performance of a learned visual processing algo-",
"ground\"\"The user that was previously presented with a",
"'SO",
"rithm. In this paper, we explore methods to flexibly guide a",
"fixed,black-boxpredictionisnowabletoinfluenceandalter",
"trained convolutional neural network through user input to",
"the outcome according to his needs. This property is partic-",
"improve its performance during inference. We do so by in-",
"ularly useful in high risk domains such as medical image",
"",
"serting a layer that acts as a spatio-semantic guide into the",
"analysis and computer-assisted diagnosis, where fully auto-",
"network. This guide is trained to modify the network's acti-",
"mated segmentation is not always robust in clinical appli-",
"4",
"vations, either directly via an energy minimization scheme",
"cationsand the experience of trained practitioners matters.",
"or indirectly through a recurrent model that translates hu-",
"Another relevant example is speeding up labor-intensive",
"",
"man language queries to interaction weights. Learning the",
"and repetitive labeling tasks, such as those needed to cre-",
"verbal interaction is fully automatic and does not require",
"ate datasets for semantic segmentation, especially those for",
"6081",
"manual text annotations. We evaluate the method on two",
"which annotations are scarce and expensive to obtain.",
"datasets, showing that guiding a pre-trained network can",
"We propose a novel idea to allow user-network feedback-",
"improve performance, and provide extensive insights into",
"1:AIXIE",
"based interaction that aims at improving the performance of",
"the interaction between the guide and the CNN.",
"a pre-trained CNN at test time. The core idea is the defini-",
"tion of a spatio-semantic guiding mechanism that translates",
"user feedback into changes in the internal activations of the",
"1. Introduction",
"network, thus acting as a means of re-thinking the inference",
"Convolutional neural networks (CNNs) continue to grow",
"process. The user input is modeled via a language-based ap-",
"in their breadth of application and in their performance on",
"proach, that enables interaction with a trained model in the",
"challenging computer vision tasks, such as image classifi-",
"form of a dialog. The user receives a first estimate, inputs a",
"cation, semantic segmentation, object detection, depth pre-",
"text query and the network replies with an updated predic-",
"diction and human pose estimation. To date, the majority of",
"tion. Most previous interactive approaches place the user on",
"the techniques proposed for these applications train specific",
"the input/data side which means user input is required for",
"network architectures once and subsequently deploy them",
"the method to operate. In contrast, in our method, the user's",
"as static components inside an algorithm. However, it is",
"input is optional and modifies the network, this means that",
"unlikely that any static network will be perfect at the task",
"the network does not depend on human interaction but can",
"it was designed for. If the deployed CNN were adaptable",
"be adjusted by it.",
"to feedback or specifications provided by a human user on-",
"We showcase this interactive module on the task of se-",
"line, this interaction would hold the potential to improve the",
"mantic image segmentation. One advantage of our method",
"model's performance and benefit real-world applications.",
"is that it does not depend on any explicit annotation for text-",
"-"
] |
{
"arxiv_id": "1803.11544",
"language": "en",
"page_idx": 0
}
|
||
1803.11544-en-1
|
1803.11544-en-1.jpg
|
<FigureHere>
Figure 1. Overview. We introduce a system that is able to refine predictions of a CNN by injecting a guiding block into the network. The guiding can be performed using natural language through an RNN to process the text. In this example, the original network had difficulties to differentiate between the sky and the cloud classes. The user indicates that there is no sky and the prediction is updated, without any CNN weight updates and thus without additional training.
region correspondences. Yet, results indicate that the module can successfully transfer semantic information from the natural language domain to the visual domain, such that the network eventually produces a more accurate segmentation. As a side effect, this provides interesting insights into how CNNs structure their inference with respect to natural categories, providing an avenue for exploring the relationship between language and imagery.
map given by the user and [57] uses a second network operating both on the previous prediction and human feedback.
In our system, we integrate online interaction into the training by substituting the human input with an algorithm that dynamically generates hints from different modalities based on previous predictions. The CNN is already trained and only asks for the user's directions for the purpose of conditioned (on-demand) adjustments of an initial estimate.
# $2. \textbf{ Related Work}$
The intersection of vision and language To enable user interaction in a natural and intuitive way, we propose a novel idea that lies in the joint domain of natural language and vision. A relevant line of work in this feld is Visual Question Answering (VQA).A question is posed and the answer is based on the image context[1,45,64]. Specialized systems for VQA ground the question in the input image and focus on the relevant parts to answer complex queries with text responses [4,5,27,32,43,46]. Other examples include generation of referring expressions [36, 44, 47, 62], segmentation or object retrieval from referring expressions[19,29,28], image captioning [16,21,33,34,48,56] and visual dialog [18]. Most works focus on the combination of CNN and RNN models, often building attention mechanisms [3,42,59,60]. Most related to our approach is the recent method from [20,50] that proposes the use of a conditional batch normalization layer [30] and feature-wise adjustment for visual reasoning.
A key distinction between our approach and most of the summarized literature is that our system's output is visual and not textual, $i.e.$ it is neither an answer nor an image caption. The output of the interactive CNN is in the same domain as the initial one. Another major difference is that we do not rely on vision-text correspondences with paired questions-answers or captions; user interaction is simulated via textual expressions that we generate automatically
Interaction with neural networks Human-machine interaction is an extensively researched field. In [9] the user and an algorithm work together to solve fine grained recognition tasks, leveraging analytic strengths of the system and visual recognition power of the human. Prior to deep learning, several systems have been proposed for semi-automatic segmentation, that allowed the user to interfere with the result or to provide hints to the system via seed points [8,52], bounding boxes [24, 37, 54], contours or scribbles [7,23,58], eye gaze[26] or in the form of binary yes/no answers to a set of questions [11,55].
Most deep learning based segmentation methods, however, do not have an interface for human input during inference. The model and thus the attainable performance is fixed after the training phase. Directly integrating a human into a training loop with thousands of images is challenging. Nonetheless, some methods towards interactive deep learning have been proposed, such as weakly-supervised semantic segmentation from scribbles [38], user-provided clicks and Euclidean distance maps [61] or bounding boxes used as region initialization [17,53]. Additionally, a method for sparse, user-guided colorization of grayscale images is proposed in [63]. In the feld of medical imaging,[2] proposes to interactively improve segmentation by updating a seed-
|
[
[
0.77783203125,
0.10101318359375,
0.810546875,
0.10101318359375,
0.810546875,
0.11297607421875,
0.77783203125,
0.11297607421875
],
[
0.7509765625,
0.1142578125,
0.8388671875,
0.1142578125,
0.8388671875,
0.1256103515625,
0.7509765625,
0.1256103515625
],
[
0.3798828125,
0.126953125,
0.416748046875,
0.126953125,
0.416748046875,
0.1427001953125,
0.3798828125,
0.1427001953125
],
[
0.75732421875,
0.1287841796875,
0.83251953125,
0.1287841796875,
0.83251953125,
0.14013671875,
0.75732421875,
0.14013671875
],
[
0.771484375,
0.1427001953125,
0.81787109375,
0.1427001953125,
0.81787109375,
0.154052734375,
0.771484375,
0.154052734375
],
[
0.69921875,
0.1641845703125,
0.7236328125,
0.1641845703125,
0.7236328125,
0.1717529296875,
0.69921875,
0.1717529296875
],
[
0.17236328125,
0.1787109375,
0.1993408203125,
0.1787109375,
0.1993408203125,
0.186279296875,
0.17236328125,
0.186279296875
],
[
0.57666015625,
0.177978515625,
0.640625,
0.177978515625,
0.640625,
0.1881103515625,
0.57666015625,
0.1881103515625
],
[
0.381591796875,
0.184326171875,
0.4150390625,
0.184326171875,
0.4150390625,
0.1962890625,
0.381591796875,
0.1962890625
],
[
0.303955078125,
0.2215576171875,
0.336669921875,
0.2215576171875,
0.336669921875,
0.2342529296875,
0.303955078125,
0.2342529296875
],
[
0.459228515625,
0.2215576171875,
0.49267578125,
0.2215576171875,
0.49267578125,
0.2342529296875,
0.459228515625,
0.2342529296875
],
[
0.3798828125,
0.2291259765625,
0.4150390625,
0.2291259765625,
0.4150390625,
0.2412109375,
0.3798828125,
0.2412109375
],
[
0.301513671875,
0.2374267578125,
0.339111328125,
0.2374267578125,
0.339111328125,
0.250732421875,
0.301513671875,
0.250732421875
],
[
0.463134765625,
0.2374267578125,
0.488525390625,
0.2374267578125,
0.488525390625,
0.25,
0.463134765625,
0.25
],
[
0.54736328125,
0.27587890625,
0.6708984375,
0.27587890625,
0.6708984375,
0.285888671875,
0.54736328125,
0.285888671875
],
[
0.7294921875,
0.276611328125,
0.8095703125,
0.27587890625,
0.8095703125,
0.287353515625,
0.7294921875,
0.28857421875
],
[
0.07843017578125,
0.291748046875,
0.89208984375,
0.292236328125,
0.89208984375,
0.306884765625,
0.07843017578125,
0.30615234375
],
[
0.17236328125,
0.279052734375,
0.1993408203125,
0.279052734375,
0.1993408203125,
0.285888671875,
0.17236328125,
0.285888671875
],
[
0.080078125,
0.306884765625,
0.89208984375,
0.306884765625,
0.89208984375,
0.32080078125,
0.080078125,
0.32080078125
],
[
0.080078125,
0.3212890625,
0.89208984375,
0.3212890625,
0.89208984375,
0.335205078125,
0.080078125,
0.335205078125
],
[
0.07843017578125,
0.333984375,
0.420654296875,
0.335205078125,
0.420654296875,
0.349853515625,
0.07843017578125,
0.348388671875
],
[
0.07843017578125,
0.375,
0.468017578125,
0.373779296875,
0.468017578125,
0.38818359375,
0.07843017578125,
0.389404296875
],
[
0.5009765625,
0.375,
0.890625,
0.373779296875,
0.890625,
0.38818359375,
0.5009765625,
0.389404296875
],
[
0.080078125,
0.39013671875,
0.4697265625,
0.39013671875,
0.4697265625,
0.404052734375,
0.080078125,
0.404052734375
],
[
0.50244140625,
0.39013671875,
0.88623046875,
0.39013671875,
0.88623046875,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.080078125,
0.406005859375,
0.4697265625,
0.406005859375,
0.4697265625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.521484375,
0.4072265625,
0.89208984375,
0.4072265625,
0.89208984375,
0.42041015625,
0.521484375,
0.42041015625
],
[
0.080078125,
0.421630859375,
0.468017578125,
0.421630859375,
0.468017578125,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.5009765625,
0.421630859375,
0.890625,
0.421630859375,
0.890625,
0.43505859375,
0.5009765625,
0.43505859375
],
[
0.08251953125,
0.4375,
0.468017578125,
0.4375,
0.468017578125,
0.447509765625,
0.08251953125,
0.447509765625
],
[
0.50244140625,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.50244140625,
0.45068359375
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.50244140625,
0.451904296875,
0.89208984375,
0.451904296875,
0.89208984375,
0.46533203125,
0.50244140625,
0.46533203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50244140625,
0.46728515625,
0.89208984375,
0.46728515625,
0.89208984375,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.07843017578125,
0.48095703125,
0.28759765625,
0.48095703125,
0.28759765625,
0.494384765625,
0.07843017578125,
0.494384765625
],
[
0.50244140625,
0.481689453125,
0.890625,
0.481689453125,
0.890625,
0.49560546875,
0.50244140625,
0.49560546875
],
[
0.07843017578125,
0.5087890625,
0.2197265625,
0.51123046875,
0.2197265625,
0.52880859375,
0.07843017578125,
0.5263671875
],
[
0.50341796875,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50341796875,
0.521484375
],
[
0.50244140625,
0.5234375,
0.8935546875,
0.5234375,
0.8935546875,
0.537109375,
0.50244140625,
0.537109375
],
[
0.080078125,
0.53857421875,
0.468017578125,
0.53857421875,
0.468017578125,
0.5517578125,
0.080078125,
0.5517578125
],
[
0.5009765625,
0.537109375,
0.8935546875,
0.53857421875,
0.8935546875,
0.55322265625,
0.5009765625,
0.5517578125
],
[
0.080078125,
0.5537109375,
0.4697265625,
0.5537109375,
0.4697265625,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50341796875,
0.5546875,
0.890625,
0.5546875,
0.890625,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.080078125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.080078125,
0.58203125
],
[
0.50244140625,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.08087158203125,
0.60107421875,
0.46728515625,
0.60107421875,
0.46728515625,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.50244140625,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.0767822265625,
0.62890625,
0.4697265625,
0.62744140625,
0.4697265625,
0.642578125,
0.0767822265625,
0.64306640625
],
[
0.5009765625,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.642578125,
0.5009765625,
0.642578125
],
[
0.080078125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.08087158203125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.07843017578125,
0.68994140625,
0.328369140625,
0.68994140625,
0.328369140625,
0.7041015625,
0.07843017578125,
0.7041015625
],
[
0.5048828125,
0.69140625,
0.890625,
0.69140625,
0.890625,
0.7021484375,
0.5048828125,
0.7021484375
],
[
0.096435546875,
0.705078125,
0.468017578125,
0.7041015625,
0.468017578125,
0.71826171875,
0.096435546875,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73486328125,
0.8896484375,
0.73486328125,
0.8896484375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.07843017578125,
0.75,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.07843017578125,
0.7646484375
],
[
0.50244140625,
0.75,
0.8896484375,
0.75,
0.8896484375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76513671875,
0.826171875,
0.76513671875,
0.826171875,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.51953125,
0.77978515625,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.51953125,
0.79443359375
],
[
0.080078125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
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
],
[
0.5009765625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.5,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.465576171875,
0.88720703125,
0.465576171875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.859375,
0.88720703125,
0.859375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Hint:",
"\"Theskyisnot",
"CNN",
"visibleinthis",
"image\"",
"user",
"inpu",
"prediction",
"RNN",
"CNN",
"CNN",
"guide",
"head",
"tail",
"improvedprediction",
"groundtruth",
"Figure 1. Overview. We introduce a system that is able to refine predictions of a CNN by injecting a guiding block into the network. The",
"npu",
"guiding can be performed using natural language through an RNN to process the text. In this example, the original network had difficulties",
"to differentiate between the sky and the cloud classes. The user indicates that there is no sky and the prediction is updated, without any",
"CNN weight updates and thus without additional training.",
"region correspondences. Yet, results indicate that the mod-",
"map given by the user and [57] uses a second network oper-",
"ule can successfully transfer semantic information from the",
"ating both on the previous prediction and human feedback.",
"natural language domain to the visual domain, such that the",
"In our system, we integrate online interaction into the",
"network eventually produces a more accurate segmentation.",
"training by substituting the human input with an algorithm",
"Asasideeffect,thisprovides interestinginsightsintohow",
"that dynamically generates hints from different modalities",
"CNNs structure their inference with respect to natural cat-",
"based on previous predictions. The CNN is already trained",
"egories, providing an avenue for exploring the relationship",
"and only asks for the user's directions for the purpose of",
"between language and imagery.",
"conditioned (on-demand) adjustments of an initial estimate.",
"2. Related Work",
"The intersection of vision and languageTo enable user",
"interaction in a natural and intuitive way, we propose a",
"Interaction with neural networks Human-machine in-",
"novel idea that lies in the joint domain of natural language",
"teraction is an extensively researched field. In [9] the user",
"andvision.Arelevantline of workin thisfield is Visual",
"and an algorithm work together to solve fine grained recog-",
"Question Answering (VQA). A question is posed and the",
"nition tasks, leveraging analytic strengths of the system and",
"answer is based on the image context [1, 45, 64]. Spe-",
"visualrecognitionpowerofthehuman.Priortodeeplearn-",
"cialized systems for VQA ground the question in the in-",
"ing, several systems have been proposed for semi-automatic",
"put image and focus on the relevant parts to answer com-",
"segmentation, that allowed the user to interfere with the",
"plex queries with text responses [4, 5, 27, 32, 43, 46].",
"result or to provide hints to the system via seed points",
"Other examples include generation of referring expres-",
"[8, 52], bounding boxes [24, 37, 54], contours or scribbles",
"sions [36, 44, 47, 62], segmentation or object retrieval",
"[7, 23, 58], eye gaze [26] or in the form of binary yes/no",
"from referring expressions [19, 29, 28], image captioning",
"answers to a set of questions [11, 55].",
"[16,21,33.34,48,56]andvisual dialog[18].Mostworks",
"Most deep learning based segmentation methods, how-",
"focus on the combination of CNN and RNN models, often",
"ever, do not have an interface for human input during in-",
"building attention mechanisms [3, 42, 59, 60].Most related",
"ference.The model and thus the attainable performance is",
"to our approach is the recent method from [20, 50] that pro-",
"fixed after the training phase. Directly integrating a human",
"poses the use of a conditional batch normalization layer [30]",
"into a training loop with thousands of images is challenging.",
"and feature-wise adjustment for visual reasoning.",
"Nonetheless, some methods towards interactive deep learn-",
"A key distinction between our approach and most of the",
"ing have been proposed, such as weakly-supervised seman-",
"summarized literature is that our system's output is visual",
"tic segmentation from scribbles [38], user-provided clicks",
"and not textual, i.e. it is neither an answer nor an image",
"and Euclidean distance maps [61] or bounding boxes used",
"caption. The output of the interactive CNN is in the same",
"as region initialization [17, 53]. Additionally, a method for",
"domain as the initial one. Another major difference is that",
"sparse, user-guided colorization of grayscale images is pro-",
"we do not rely on vision-text correspondences with paired",
"posed in [63]. In the field of medical imaging, [2] proposes",
"questions-answers or captions; user interaction is simulated",
"to interactively improve segmentation by updating a seed-",
"via textual expressions that we generate automatically."
] |
{
"arxiv_id": "1803.11544",
"language": "en",
"page_idx": 1
}
|
||
1803.11544-en-2
|
1803.11544-en-2.jpg
|
Semantic segmentation In this paper, we focus on the application of semantic image segmentation, which is widely studied in the computer vision literature and significantly boosted by the success of deep learning methods [6,12,31, 39,41,51]. Our goal is to deploy out-of-the-box, state-ofthe-art models [12, 41] as estimators, that are then guided by our module to improve their former predictions with the help of a human user (or any given priors as hints).
## 3. Methods
In this section, we describe how our interaction module is integrated into a fuxed CNN following two different approaches: guiding with user clicks and back-propagation (Section 3.2) or natural language inputs (Section 3.3). A general overview for the latter case is shown in Figure 1.
We first define the main elements of our framework, which we refer to throughout this paper. The module we insert into the CNN is called the guide. The guide interacts with the guided CNN through a guiding block, which is built to adjust activation maps of the CNN to improve the final prediction for the given input. The guided CNN is thus split into two parts: the $head$, which processes the input until it reaches the guiding block, and the $tail$, that is the rest of the guided network up to the final prediction layer. More formally, by decomposing the network into a head $h$ and a tail function $t$, the output prediction $\tilde{y}$ given input $x$ can be written as $t(h(x))=\tilde{y}.$ We refer to the information that the guide uses to modify the guided network as the $hint.$
The split position is chosen manually. However, a reasonable choice is the (spatially) smallest encoding that the network produces, as this layer likely contains the most high-level information in a compact representation.We validate the choice of layers in the Section 4.2.
## 3.1. Guiding Block
The guiding block is the integral piece of our approach, it enables feedback to flow from the guide into the guided network. Essentially, the guide must be able to modify a set of activations in the guided network. Since activation maps usually consist of a large number of elements (e.g. $32\times32\times1024\approx1$ million), element-wise control is prone to over-fitting. The intuition behind our guiding block is that the network encodes specific features in each channel of a given layer. Thus, a guide that has the ability to strengthen or weaken the activations per channel, can emphasize or suppress aspects of the image representation and thus adapt the prediction in a semantically meaningful way.
The head predicts a feature representation $h(x)=A\in$ $\mathbb{R}^{H\times W\times C}$ with width $W$,height $\hat H$ and number of channels $C.$ Then, guiding can be expressed as a per feature map multiplication with a vector $\gamma^{(s)}\in\mathbb{R}^C$ and bias $\gamma^(b)\in\mathbb{R}^{\hat{C}}$,
(1)
$$A_c'=(1+\gamma_c^{(s)})A_c+\gamma_c^{(b)}\:,$$
where $c\in[1,\ldots,C]$ indexes the channels of the feature volume $A$ and the corresponding elements of the guiding vector $\gamma=(\gamma^{(s)},\gamma^{(b)}).$ Given this formulation, we are able to adjust a set of feature maps by emphasizing or hiding information per channel. Equation (1) can also be interpreted as the guide predicting a residual update (similar to ResNets [25]) for the activation map $A_c.\gamma$ plays the role of a filter on the feature maps. When $\gamma=0$, our guiding block reproduces the input feature map and thus has no effect in $\hat{\text{guiding the network. When }\gamma_{c}^{\hat{(}s)}}=-1$,channel $c$ would become suppressed as all its units would be set to 0. Conversely, for $\gamma_c^{(s)}=1$, the activation strength of that feature channel is doubled. Values smaller than $- 1\textit{invert afeature}$ map, emphasizing aspects that would have been otherwise cut-off by the ReLU unit that typically follows the weight layer (or vice versa).
While this approach, which is similar to the conditioning layer in [20, 50], supports per-channel updates and feature re-weighting via $\gamma$, it is not flexible enough to adjust features spatially since it modifies the whole feature map with the same weight. In other words, it is impossible for this module to encourage spatially localized changes in each feature map (“On the top left you missed ...”). To overcome this limitation, we extend the approach to the spatial dimensions $H$ and $W$ of the feature map, $i.e.$ we introduce two additional guiding vectors $\alpha\in\mathbb{R}^H$ and $\beta\in\mathbb{R}^W$ to modify the feature map $\bar{A}$ with spatial attention. In the following, we will index $A$ with $h,w$ and $c$ to uniquely identify a single element $A_h,w,c\in\mathbb{R}$ of $A$
$$A'_{h,w,c}=(1+\alpha_{h}+\beta_{w}+\gamma_{c}^{(s)})A_{h,w,c}+\gamma_{c}^{b}$$
(2)
The overall function that the guided network computes is
thus modified to
$$y^*=t\left[(1\oplus\alpha\oplus\beta\oplus\gamma^{(s)})\odot h(x)\oplus\gamma^{(b)}\right]\:,$$
(3)
where the tiling of the vectors $\alpha,\beta,\gamma$ along their appropriate dimensions is denoted with $\oplus$ and the Hadamard product with $\odot$. This way $\alpha$ and $\beta$ have spatial influence and $\gamma$ controls the semantic adjustment. Guiding with Equation 3 reduces the number of parameters from $H\times W\times C$ to $H+ W+ C$ = 1088 in the previous example, which is manageable to predict with a small guiding module.
Since fully convolutional architectures are a common choice for image prediction tasks, we employ linear interpolation of $\alpha$ and $\beta$ when the feature map spatial resolution varies. This choice reflects two properties of the guiding block. First, $\alpha$ and $\beta$ do not depend on fxed $H$ and $W.$ Second, one can select the granularity of the spatial resolution by changing the dimensionality of $\alpha$ and $\beta$ to match the spatial complexity of the hints that the guide follows.
We describe two fundamentally different ways to employ the guiding block. The first one- guiding by back-
|
[
[
0.08087158203125,
0.09027099609375,
0.46728515625,
0.09027099609375,
0.46728515625,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.50341796875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.07843017578125,
0.121826171875,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.121826171875,
0.89208984375,
0.121826171875,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.46728515625,
0.1383056640625,
0.46728515625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.08251953125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.08251953125,
0.1666259765625
],
[
0.50341796875,
0.1546630859375,
0.890625,
0.1546630859375,
0.890625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.08087158203125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.08087158203125,
0.18115234375
],
[
0.50341796875,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5048828125,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1976318359375,
0.5048828125,
0.1976318359375
],
[
0.080078125,
0.1988525390625,
0.41259765625,
0.1988525390625,
0.41259765625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.1962890625,
0.890625,
0.1976318359375,
0.890625,
0.212158203125,
0.50244140625,
0.2108154296875
],
[
0.50244140625,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2266845703125,
0.1748046875,
0.2266845703125,
0.1748046875,
0.2412109375,
0.080078125,
0.2412109375
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2291259765625,
0.89208984375,
0.24365234375,
0.50244140625,
0.2442626953125
],
[
0.50341796875,
0.24560546875,
0.89208984375,
0.24560546875,
0.89208984375,
0.2587890625,
0.50341796875,
0.2587890625
],
[
0.098876953125,
0.251953125,
0.465576171875,
0.251953125,
0.465576171875,
0.26513671875,
0.098876953125,
0.26513671875
],
[
0.50244140625,
0.261474609375,
0.61279296875,
0.260009765625,
0.61279296875,
0.274658203125,
0.50244140625,
0.27587890625
],
[
0.080078125,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.6044921875,
0.261962890625,
0.89208984375,
0.261962890625,
0.89208984375,
0.27587890625,
0.6044921875,
0.27587890625
],
[
0.50341796875,
0.276611328125,
0.89208984375,
0.276611328125,
0.89208984375,
0.290283203125,
0.50341796875,
0.290283203125
],
[
0.08087158203125,
0.28271484375,
0.468017578125,
0.28271484375,
0.468017578125,
0.296630859375,
0.08087158203125,
0.296630859375
],
[
0.08087158203125,
0.29736328125,
0.4697265625,
0.29736328125,
0.4697265625,
0.311279296875,
0.08087158203125,
0.311279296875
],
[
0.5009765625,
0.29345703125,
0.89208984375,
0.291748046875,
0.89208984375,
0.304931640625,
0.5009765625,
0.306884765625
],
[
0.080078125,
0.313232421875,
0.449462890625,
0.313232421875,
0.449462890625,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.50341796875,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.1004638671875,
0.32763671875,
0.46728515625,
0.32763671875,
0.46728515625,
0.341552734375,
0.1004638671875,
0.341552734375
],
[
0.50244140625,
0.323974609375,
0.6396484375,
0.323974609375,
0.6396484375,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.51953125,
0.337158203125,
0.89208984375,
0.3359375,
0.89208984375,
0.35107421875,
0.51953125,
0.3515625
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.080078125,
0.35791015625,
0.46728515625,
0.35791015625,
0.46728515625,
0.37255859375,
0.080078125,
0.37255859375
],
[
0.50341796875,
0.354248046875,
0.890625,
0.354248046875,
0.890625,
0.367431640625,
0.50341796875,
0.367431640625
],
[
0.07843017578125,
0.37255859375,
0.468017578125,
0.371826171875,
0.468017578125,
0.386474609375,
0.07843017578125,
0.386962890625
],
[
0.50341796875,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.080078125,
0.38818359375,
0.4697265625,
0.38818359375,
0.4697265625,
0.401611328125,
0.080078125,
0.401611328125
],
[
0.50244140625,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.396484375,
0.50244140625,
0.396484375
],
[
0.080078125,
0.40283203125,
0.4697265625,
0.40283203125,
0.4697265625,
0.416748046875,
0.080078125,
0.416748046875
],
[
0.50341796875,
0.3984375,
0.89208984375,
0.3984375,
0.89208984375,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.08251953125,
0.41845703125,
0.46728515625,
0.41845703125,
0.46728515625,
0.431884765625,
0.08251953125,
0.431884765625
],
[
0.50341796875,
0.41357421875,
0.89208984375,
0.41357421875,
0.89208984375,
0.427978515625,
0.50341796875,
0.427978515625
],
[
0.50244140625,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.442626953125,
0.50244140625,
0.442626953125
],
[
0.08087158203125,
0.43505859375,
0.46728515625,
0.43505859375,
0.46728515625,
0.44580078125,
0.08087158203125,
0.44580078125
],
[
0.08087158203125,
0.448974609375,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.50341796875,
0.443115234375,
0.890625,
0.443115234375,
0.890625,
0.45703125,
0.50341796875,
0.45703125
],
[
0.5009765625,
0.45703125,
0.89208984375,
0.458251953125,
0.89208984375,
0.472900390625,
0.5009765625,
0.4716796875
],
[
0.080078125,
0.46337890625,
0.4697265625,
0.46337890625,
0.4697265625,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50341796875,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.080078125,
0.479248046875,
0.4697265625,
0.479248046875,
0.4697265625,
0.492431640625,
0.080078125,
0.492431640625
],
[
0.5009765625,
0.488037109375,
0.89208984375,
0.4892578125,
0.89208984375,
0.50390625,
0.5009765625,
0.50244140625
],
[
0.08087158203125,
0.494384765625,
0.4697265625,
0.494384765625,
0.4697265625,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50390625,
0.89208984375,
0.51806640625,
0.50244140625,
0.517578125
],
[
0.080078125,
0.5087890625,
0.44775390625,
0.5087890625,
0.44775390625,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.098876953125,
0.5234375,
0.46728515625,
0.5244140625,
0.46728515625,
0.53857421875,
0.098876953125,
0.537109375
],
[
0.50341796875,
0.51953125,
0.67236328125,
0.51953125,
0.67236328125,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.08087158203125,
0.54150390625,
0.46728515625,
0.54150390625,
0.46728515625,
0.5517578125,
0.08087158203125,
0.5517578125
],
[
0.5498046875,
0.54443359375,
0.85107421875,
0.54248046875,
0.85107421875,
0.56298828125,
0.5498046875,
0.56494140625
],
[
0.86669921875,
0.548828125,
0.8896484375,
0.548828125,
0.8896484375,
0.56005859375,
0.86669921875,
0.56005859375
],
[
0.080078125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.080078125,
0.568359375
],
[
0.08087158203125,
0.57080078125,
0.465576171875,
0.57080078125,
0.465576171875,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50341796875,
0.57373046875,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50341796875,
0.5869140625
],
[
0.080078125,
0.583984375,
0.36767578125,
0.583984375,
0.36767578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50341796875,
0.5888671875,
0.615234375,
0.5888671875,
0.615234375,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.080078125,
0.60791015625,
0.2254638671875,
0.60791015625,
0.2254638671875,
0.62255859375,
0.080078125,
0.62255859375
],
[
0.55126953125,
0.6162109375,
0.58642578125,
0.6162109375,
0.58642578125,
0.62744140625,
0.55126953125,
0.62744140625
],
[
0.59716796875,
0.6142578125,
0.826171875,
0.6142578125,
0.826171875,
0.62890625,
0.59716796875,
0.62890625
],
[
0.83837890625,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.62744140625,
0.83837890625,
0.62744140625
],
[
0.09722900390625,
0.630859375,
0.46728515625,
0.6318359375,
0.46728515625,
0.646484375,
0.09722900390625,
0.64599609375
],
[
0.07843017578125,
0.64599609375,
0.471435546875,
0.646484375,
0.471435546875,
0.6611328125,
0.07843017578125,
0.66015625
],
[
0.50244140625,
0.64306640625,
0.890625,
0.64453125,
0.890625,
0.6591796875,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.662109375,
0.471435546875,
0.662109375,
0.471435546875,
0.67529296875,
0.080078125,
0.67529296875
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.69140625,
0.080078125,
0.69140625
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.6923828125,
0.468017578125,
0.6923828125,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.080078125,
0.70849609375,
0.468017578125,
0.70849609375,
0.468017578125,
0.7216796875,
0.080078125,
0.7216796875
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.705078125,
0.89208984375,
0.7197265625,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.736328125,
0.080078125,
0.736328125
],
[
0.50341796875,
0.72021484375,
0.8935546875,
0.72021484375,
0.8935546875,
0.734375,
0.50341796875,
0.734375
],
[
0.07843017578125,
0.73779296875,
0.4697265625,
0.73779296875,
0.4697265625,
0.751953125,
0.07843017578125,
0.751953125
],
[
0.50244140625,
0.73486328125,
0.84228515625,
0.73486328125,
0.84228515625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.75390625,
0.4697265625,
0.75390625,
0.4697265625,
0.767578125,
0.080078125,
0.767578125
],
[
0.52294921875,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.08087158203125,
0.77099609375,
0.465576171875,
0.77099609375,
0.465576171875,
0.78076171875,
0.08087158203125,
0.78076171875
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.7841796875,
0.468017578125,
0.7841796875,
0.468017578125,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.798828125,
0.46728515625,
0.798828125,
0.46728515625,
0.8125,
0.080078125,
0.8125
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.09722900390625,
0.8125,
0.468017578125,
0.81298828125,
0.468017578125,
0.8271484375,
0.09722900390625,
0.82568359375
],
[
0.50244140625,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.07843017578125,
0.82470703125,
0.4697265625,
0.82763671875,
0.4697265625,
0.8447265625,
0.07843017578125,
0.84130859375
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84326171875,
0.468017578125,
0.84326171875,
0.468017578125,
0.85693359375,
0.08087158203125,
0.85693359375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85791015625,
0.46728515625,
0.85693359375,
0.46728515625,
0.8720703125,
0.07843017578125,
0.87255859375
],
[
0.50341796875,
0.85693359375,
0.85107421875,
0.85693359375,
0.85107421875,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.1846923828125,
0.88623046875,
0.363525390625,
0.88623046875,
0.363525390625,
0.90087890625,
0.1846923828125,
0.90087890625
],
[
0.44677734375,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.44677734375,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Semantic segmentation In this paper, we focus on the ap-",
"where c E [1, ..., C] indexes the channels of the feature",
"plication of semantic image segmentation, which is widely",
"volume A and the corresponding elements of the guiding",
"studied in the computer vision literature and significantly",
"vector = ((s),(b). Given this formulation, we are able",
"boosted by the success of deep learning methods [6, 12, 31,",
"to adjust a set of feature maps by emphasizing or hiding in-",
"39, 41, 51l. Our goal is to deploy out-of-the-box, state-of-",
"formationperchannel.Equation(1)canalsobeinterpreted",
"the-art models [12, 41] as estimators, that are then guided",
"as the guide predicting a residual update (similar to ResNets",
"by our module to improve their former predictions with the",
"[25]) for the activation map Ac. plays the role of a filter",
"help of a human user (or any given priors as hints).",
"on the feature maps.When = O, our guiding block re-",
"produces the input feature map and thus has no effect in",
"3. Methods",
"guiding the network. When s) = -1, channel c would",
"become suppressed as all its units would be set to O. Con-",
"In this section, we describe how our interaction mod-",
"versely, for (s)",
"ule is integrated into a fixed CNN following two different",
" = 1, the activation strength of that feature",
"channel is doubled. Values smaller than -1 invert a feature",
"approaches: guiding with user clicks and back-propagation",
"(Section 3.2) or natural language inputs (Section 3.3). A",
"map, emphasizingaspects thatwould havebeen otherwise",
"general overview for the latter case is shown in Figure 1",
"cut-off by the ReLU unit that typically follows the weight",
"We first define the main elements of our framework,",
"layer (or vice versa).",
"While this approach, which is similar to the condition-",
"which we refer to throughout this paper.The module we",
"insert into the CNN is called the guide. The guide inter-",
"ing layer in [20, 50], supports per-channel updates and fea-",
"acts with the guided CNN through a guiding block, which",
"ture re-weighting via , it is not flexible enough to adjust",
"is built to adjust activation maps of the CNN to improve the",
"features spatially since it modifies the whole feature map",
"final prediction for the given input. The guided CNN is thus",
"with the same weight. In other words, it is impossible for",
"split into two parts: the head, which processes the input un-",
"this module to encourage spatially localized changes in each",
"feature map (\"On the top left you missed ..\"). To overcome",
"til it reaches the guiding block, and the tail, that is the rest",
"of the guided network up to the final prediction layer. More",
"this limitation, we extend the approach to the spatial dimen-",
"sions H and W of the feature map, i.e. we introduce two",
"formally, by decomposing the network into a head h and a",
"additional guiding vectors α e RH and β e RW to modify",
"tail function t, the output prediction y given input α can be",
"the feature map A with spatial attention. In the following,",
"written as t(h(r)) = y : We refer to the information that",
"we will index A with h, w and c to uniquely identify a single",
"the guide uses to modify the guided network as the hint.",
"The split position is chosen manually. However, a rea-",
"element Ah,w,c E R of A",
"sonablechoiceisthe(spatially)smallestencodingthatthe",
"An,w,e =(1+ ah + Bu +%)An,w,e+%",
"(2)",
"network produces, as this layer likely contains the most",
"high-level informationin acompactrepresentation.We val-",
"The overall function that the guided network computes is",
"idate the choice of layers in the Section 4.2.",
"thus modified to",
"3.1. Guiding Block",
"*=",
"[(1由α由β由(s))0h(r)(b)",
"(3)",
"The guiding block is the integral piece of our approach,",
"it enables feedback to flow from the guide into the guided",
"where the tiling of the vectors α, β, along their appropri-",
"network.Essentially,the guide must be able to modify a",
"ate dimensions is denoted with @ and the Hadamard prod-",
"set of activations in the guided network. Since activation",
"uct with o. This way α and β have spatial influence and",
"maps usually consist of a large number of elements (e.g.",
" controls the semantic adjustment. Guiding with Equation",
"32 × 32 × 1024 ~ 1 million), element-wise control is prone",
"3 reduces the number of parameters from H × W × C to",
"to over-fitting. The intuition behind our guiding block is",
"H + W + C = 1088 in the previous example, which is",
"that the network encodes specific features in each chan-",
"manageable to predict with a small guiding module.",
"nel of a given layer. Thus, a guide that has the ability to",
"Since fully convolutional architectures are a common",
"strengthen orweakentheactivationsper channel,canem-",
"choice for image prediction tasks, we employ linear inter-",
"phasize or suppress aspects of the image representation and",
"polation of α and β when the feature map spatial resolution",
"thus adapt the prediction in a semantically meaningful way",
"varies. This choice reflects two properties of the guiding",
"Thehead predicts afeature representation h(α)=A ",
"block. First, α and β do not depend on fixed H and W.",
"IRH ×W xC with width W, height H and number of channels",
"Second, one can select the granularity of the spatial resolu-",
"C. Then, guiding can be expressed as a per feature map",
"tion by changing the dimensionality of α and β to match the",
"multiplication with a vector (s) e RC and bias (b) e R",
"spatial complexity of the hints that the guide follows.",
"We describe two fundamentally different ways to em-",
"A=(1 +%))Ac +)",
"(1)",
"ploy the guiding block. The first one - guiding by back-"
] |
{
"arxiv_id": "1803.11544",
"language": "en",
"page_idx": 2
}
|
||
1803.11544-en-3
|
1803.11544-en-3.jpg
|
propagation (Section 3.2)-can be directly applied on a pretrained CNN that is kept constant. The second one aims at online interaction with neural networks via user feedback. The network should be able to deal with hints from different modalities, such as natural language-“the dog was mistaken for a horse". We describe how the guiding parameters $\alpha,\beta$ and $\gamma$ can be predicted with an appropriate module given a hint from a different input domain in Section 3.3, which also speeds up processing.
# 3.2. Guiding by Back-propagation
In this setup, our goal is to optimize the guiding parameters such that the network revises its decision making process and, without further learning, improves its initial prediction for the current input. The guiding block is placed between head and tail, and the guiding parameters are initialized to 0. For a given sample $x$, we formulate an energy minimization task to optimize $\alpha,\beta$ and $\gamma.$ The hint will be given as a sparse input $\hat{y}$ associated to a mask $\hat{m}.$ $\hat{y}$ and $\hat{m}$ have the same dimensionality as the prediction $\tilde{y}$ and the ground truth $y.$ $\hat{m}$ is a binary mask that indicates the locations where a hint $(i.e.$ prior knowledge) is given. Prior knowledge can be either directly given by the user or it could be a prior computed by another source.
In semantic segmentation, one can think of the hint as a single (or more) pixel(s) for which the user provides the true class-“this[pixel] is $a$ dog”as additional information. Prior to guiding, a certain loss $\mathcal{L}(t(h(x)),y)$ has been minimized during training of the network for a given task. We now optimize towards the same objective, e.g. per-pixel cross entropy for segmentation, but use the mask $\hat{m}$ to only consider the influence of the hint and minimize for the guiding parameters, as opposed to the network's parameters, $i.e.$
(4)
$$\alpha^*,\beta^*,\gamma^*=\underset{\alpha,\beta,\gamma}{\mathrm{argmin}}\left[\hat{m}\odot\mathcal{L}(y^*,\hat{y})\right]\:.$$
In this case, we only update the guiding variables for the current specific input $x$ and hint $\hat{y}$, whereas the network's weights are not trained further. The minimization finds the best parameters $\alpha^*,\beta^*,\gamma^*$ conditioned on the hint. The key insight is that this results in an overall adjusted prediction.
Since the guiding block and the network $t(h(x))$ are differentiable, we can minimize (4) using standard backpropagation and gradient descent with momentum. Intuitively, the tendency of gradient descent to fall into local optima is desirable here. We are looking for the smallest possible $\alpha,\beta$ and $\gamma$ that brings the guided prediction closer to the hint while avoiding degenerate solutions such as predicting the whole image as the hinted class.
## 3.3. Learning to Guide with Text
While the previous idea is straightforward and simple to apply to any network, it requires the hint to be given in the
<FigureHere>
Figure 2. Query Generator. We illustrate the process to automat-
ically generate queries to substitute the user during training
same domain as the network's output. We now explore a more natural way of human-machine interaction, in which the user can give hints to the network in natural language and the guiding mechanism is trained to update its parameters after parsing the user's inputs. To the best of our knowledge, this is a topic that has not been previously studied.
Training with queries Similar to prior work in related fields, we use a recurrent neural network (RNN) for processing natural language inputs. We first encode the input query using a word embedding matrix, pre-trained on a large corpus, to acquire a fixed-length embedding vector per word. The embedded words are fed as inputs to a Gated Recurrent Unit (GRU)[14, 15] at each time step. We freeze and do not further train the word embedding alongside our guiding module, to retain a meaningful word representation. The guiding parameters $\alpha,\beta$ and $\gamma$ are predicted as a linear mapping from the final hidden state of the GRU.
The language-guided module is trained as follows. We first generate an initial prediction with the fixed, taskspecific CNN, without influence from the guide. We then feed prediction and ground truth into a hint generator, which produces a query (e.g.“the sky is not visible in this image”, thus mimicking the user. The query is then encoded into a representation that becomes responsible for estimating $\alpha,\beta$ and $\gamma$ that will guide the feature map using (3) and subsequently update the prediction. The standard loss for the given task (pixel-wise cross entropy loss for semantic segmentation) is re-weighted giving positive weight(1) to the class(es) mentioned in the query to encourage changes in the prediction that coincide with the given hint. All wrongly predicted pixels are given a zero weight to prevent hints from being associated with other visual classes. Initially correct pixels are weighted 0.5 to discourage corrupting correctly classified regions.
Generating queries Previous work mostly relies on
human-annotated queries, which make them rich in variety;
|
[
[
0.0767822265625,
0.09027099609375,
0.4697265625,
0.0889892578125,
0.4697265625,
0.106689453125,
0.0767822265625,
0.10797119140625
],
[
0.7587890625,
0.09783935546875,
0.86767578125,
0.09783935546875,
0.86767578125,
0.111083984375,
0.7587890625,
0.111083984375
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.080078125,
0.12249755859375,
0.46728515625,
0.12249755859375,
0.46728515625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.69677734375,
0.12371826171875,
0.74169921875,
0.12371826171875,
0.74169921875,
0.1337890625,
0.69677734375,
0.1337890625
],
[
0.08087158203125,
0.1395263671875,
0.46728515625,
0.1395263671875,
0.46728515625,
0.149658203125,
0.08087158203125,
0.149658203125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.07843017578125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.08087158203125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.5498046875,
0.177978515625,
0.61767578125,
0.177978515625,
0.61767578125,
0.1881103515625,
0.5498046875,
0.1881103515625
],
[
0.080078125,
0.1962890625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.212158203125,
0.080078125,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.29736328125,
0.21337890625,
0.29736328125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.07843017578125,
0.23486328125,
0.3447265625,
0.2374267578125,
0.3447265625,
0.25390625,
0.07843017578125,
0.251953125
],
[
0.09722900390625,
0.261474609375,
0.46728515625,
0.261474609375,
0.46728515625,
0.274658203125,
0.09722900390625,
0.274658203125
],
[
0.71826171875,
0.25830078125,
0.869140625,
0.25830078125,
0.869140625,
0.268310546875,
0.71826171875,
0.268310546875
],
[
0.72119140625,
0.272705078125,
0.73876953125,
0.272705078125,
0.73876953125,
0.276611328125,
0.72119140625,
0.276611328125
],
[
0.7646484375,
0.272705078125,
0.7958984375,
0.271484375,
0.7958984375,
0.27587890625,
0.7646484375,
0.276611328125
],
[
0.8408203125,
0.272705078125,
0.865234375,
0.272705078125,
0.865234375,
0.276611328125,
0.8408203125,
0.276611328125
],
[
0.5595703125,
0.27783203125,
0.60693359375,
0.27783203125,
0.60693359375,
0.285400390625,
0.5595703125,
0.285400390625
],
[
0.080078125,
0.276611328125,
0.46728515625,
0.276611328125,
0.46728515625,
0.290283203125,
0.080078125,
0.290283203125
],
[
0.080078125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.304931640625,
0.080078125,
0.304931640625
],
[
0.50244140625,
0.290283203125,
0.8896484375,
0.290283203125,
0.8896484375,
0.3037109375,
0.50244140625,
0.3037109375
],
[
0.08087158203125,
0.306884765625,
0.4697265625,
0.306884765625,
0.4697265625,
0.32080078125,
0.08087158203125,
0.32080078125
],
[
0.5009765625,
0.304931640625,
0.85546875,
0.3037109375,
0.85546875,
0.318115234375,
0.5009765625,
0.3193359375
],
[
0.07843017578125,
0.3193359375,
0.4697265625,
0.3212890625,
0.4697265625,
0.3359375,
0.07843017578125,
0.333984375
],
[
0.08087158203125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50244140625,
0.3447265625,
0.89208984375,
0.3447265625,
0.89208984375,
0.35791015625,
0.50244140625,
0.35791015625
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.50341796875,
0.35986328125,
0.890625,
0.35986328125,
0.890625,
0.373779296875,
0.50341796875,
0.373779296875
],
[
0.08087158203125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.50244140625,
0.375732421875,
0.89208984375,
0.375732421875,
0.89208984375,
0.389404296875,
0.50244140625,
0.389404296875
],
[
0.08087158203125,
0.3837890625,
0.46728515625,
0.3837890625,
0.46728515625,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50244140625,
0.39013671875,
0.890625,
0.39013671875,
0.890625,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.080078125,
0.397705078125,
0.4697265625,
0.397705078125,
0.4697265625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50244140625,
0.40478515625,
0.890625,
0.40478515625,
0.890625,
0.41845703125,
0.50244140625,
0.41845703125
],
[
0.080078125,
0.412353515625,
0.46728515625,
0.412353515625,
0.46728515625,
0.425537109375,
0.080078125,
0.425537109375
],
[
0.50244140625,
0.42041015625,
0.87158203125,
0.42041015625,
0.87158203125,
0.434326171875,
0.50244140625,
0.434326171875
],
[
0.08087158203125,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.080078125,
0.442626953125,
0.38818359375,
0.442626953125,
0.38818359375,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.50341796875,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455078125,
0.50341796875,
0.455078125
],
[
0.098876953125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.4697265625,
0.098876953125,
0.4697265625
],
[
0.50341796875,
0.458984375,
0.8896484375,
0.458984375,
0.8896484375,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.080078125,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.5009765625,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.486083984375,
0.5009765625,
0.486083984375
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.080078125,
0.501953125
],
[
0.50341796875,
0.488037109375,
0.8935546875,
0.488037109375,
0.8935546875,
0.501953125,
0.50341796875,
0.501953125
],
[
0.080078125,
0.50244140625,
0.468017578125,
0.50244140625,
0.468017578125,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.5048828125,
0.50390625,
0.890625,
0.50390625,
0.890625,
0.5146484375,
0.5048828125,
0.5146484375
],
[
0.080078125,
0.51806640625,
0.465576171875,
0.51806640625,
0.465576171875,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50341796875,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.08087158203125,
0.53271484375,
0.468017578125,
0.53271484375,
0.468017578125,
0.54736328125,
0.08087158203125,
0.54736328125
],
[
0.5048828125,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54443359375,
0.5048828125,
0.54443359375
],
[
0.080078125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.080078125,
0.5625
],
[
0.50244140625,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.50341796875,
0.56298828125,
0.890625,
0.56298828125,
0.890625,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.080078125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.57763671875,
0.89208984375,
0.5771484375,
0.89208984375,
0.59130859375,
0.50244140625,
0.59228515625
],
[
0.50244140625,
0.59326171875,
0.8203125,
0.59326171875,
0.8203125,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.21728515625,
0.60595703125,
0.40185546875,
0.60595703125,
0.40185546875,
0.62060546875,
0.21728515625,
0.62060546875
],
[
0.44677734375,
0.60693359375,
0.4697265625,
0.60693359375,
0.4697265625,
0.619140625,
0.44677734375,
0.619140625
],
[
0.521484375,
0.60791015625,
0.89208984375,
0.60693359375,
0.89208984375,
0.6220703125,
0.521484375,
0.62255859375
],
[
0.2459716796875,
0.6220703125,
0.27783203125,
0.6220703125,
0.27783203125,
0.62890625,
0.2459716796875,
0.62890625
],
[
0.148681640625,
0.60791015625,
0.2115478515625,
0.60791015625,
0.2115478515625,
0.619140625,
0.148681640625,
0.619140625
],
[
0.50341796875,
0.62353515625,
0.890625,
0.62353515625,
0.890625,
0.63720703125,
0.50341796875,
0.63720703125
],
[
0.09722900390625,
0.64013671875,
0.468017578125,
0.6416015625,
0.468017578125,
0.65576171875,
0.09722900390625,
0.65478515625
],
[
0.50244140625,
0.63720703125,
0.890625,
0.63720703125,
0.890625,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.08087158203125,
0.65771484375,
0.4697265625,
0.65771484375,
0.4697265625,
0.671875,
0.08087158203125,
0.671875
],
[
0.50341796875,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50341796875,
0.66748046875
],
[
0.08087158203125,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.50244140625,
0.66845703125,
0.890625,
0.66748046875,
0.890625,
0.681640625,
0.50244140625,
0.68310546875
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.50341796875,
0.6845703125,
0.89208984375,
0.6845703125,
0.89208984375,
0.69775390625,
0.50341796875,
0.69775390625
],
[
0.080078125,
0.70263671875,
0.461669921875,
0.70263671875,
0.461669921875,
0.71630859375,
0.080078125,
0.71630859375
],
[
0.5048828125,
0.70068359375,
0.890625,
0.70068359375,
0.890625,
0.7109375,
0.5048828125,
0.7109375
],
[
0.1004638671875,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.1004638671875,
0.73193359375
],
[
0.50341796875,
0.7138671875,
0.89208984375,
0.7138671875,
0.89208984375,
0.72802734375,
0.50341796875,
0.72802734375
],
[
0.08087158203125,
0.73291015625,
0.46728515625,
0.73291015625,
0.46728515625,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.50244140625,
0.72802734375,
0.890625,
0.72802734375,
0.890625,
0.7412109375,
0.50244140625,
0.7412109375
],
[
0.07843017578125,
0.74853515625,
0.46728515625,
0.74755859375,
0.46728515625,
0.76123046875,
0.07843017578125,
0.76220703125
],
[
0.50244140625,
0.74365234375,
0.8896484375,
0.74365234375,
0.8896484375,
0.7568359375,
0.50244140625,
0.7568359375
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.7587890625,
0.89208984375,
0.7587890625,
0.89208984375,
0.7734375,
0.50244140625,
0.7734375
],
[
0.080078125,
0.77783203125,
0.4697265625,
0.77783203125,
0.4697265625,
0.79150390625,
0.080078125,
0.79150390625
],
[
0.50341796875,
0.7744140625,
0.89208984375,
0.7744140625,
0.89208984375,
0.78857421875,
0.50341796875,
0.78857421875
],
[
0.080078125,
0.79345703125,
0.4697265625,
0.79345703125,
0.4697265625,
0.806640625,
0.080078125,
0.806640625
],
[
0.5009765625,
0.78857421875,
0.890625,
0.7890625,
0.890625,
0.8037109375,
0.5009765625,
0.80322265625
],
[
0.07843017578125,
0.806640625,
0.468017578125,
0.80810546875,
0.468017578125,
0.82275390625,
0.07843017578125,
0.8212890625
],
[
0.50341796875,
0.8046875,
0.89208984375,
0.8046875,
0.89208984375,
0.818359375,
0.50341796875,
0.818359375
],
[
0.080078125,
0.82373046875,
0.366943359375,
0.82373046875,
0.366943359375,
0.8369140625,
0.080078125,
0.8369140625
],
[
0.5048828125,
0.8212890625,
0.8896484375,
0.8212890625,
0.8896484375,
0.83203125,
0.5048828125,
0.83203125
],
[
0.50244140625,
0.83544921875,
0.890625,
0.83544921875,
0.890625,
0.8486328125,
0.50244140625,
0.8486328125
],
[
0.080078125,
0.84765625,
0.330810546875,
0.84765625,
0.330810546875,
0.8623046875,
0.080078125,
0.8623046875
],
[
0.50244140625,
0.85107421875,
0.6650390625,
0.85107421875,
0.6650390625,
0.8642578125,
0.50244140625,
0.8642578125
],
[
0.098876953125,
0.869140625,
0.47216796875,
0.87060546875,
0.47216796875,
0.88525390625,
0.098876953125,
0.8837890625
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88525390625,
0.4697265625,
0.8994140625,
0.07843017578125,
0.9013671875
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"propagation (Section 3.2)- can be directly applied on a pre-",
"QueryGenerator",
"trained CNN that is kept constant.The second one aims at",
"online interaction with neural networks via user feedback.",
"error mask",
"Thenetworkshouldbeabletodealwithhintsfromdiffer-",
"entmodalities,such asnatural language-\"the dogwas mis-",
"taken for a horse\".We describe how the guiding parameters",
"Q, β and can be predicted with an appropriate module",
"initial prediction",
"given a hint from a different input domain in Section 3.3,",
"which also speeds up processing.",
"3.2. Guiding by Back-propagation",
"In this setup, our goal is to optimize the guiding param-",
"moteismissingontheright",
"e0",
"oeatin",
"ocatfo",
"groundtrut",
"eters such that the network revises its decision making pro-",
"cess and, without further learning, improves its initial pre-",
"Figure 2. Query Generator. We illustrate the process to automat-",
"diction for the current input. The guiding block is placed",
"ically generate queries to substitute the user during training.",
"between head and tail, and the guiding parameters are ini-",
"tialized to 0.For a given sample , we formulate an en-",
"same domain as the network's output.Wenow explorea",
"ergy minimization task to optimize Q, β and . The hint",
"more natural way of human-machine interaction, in which",
"will be given as a sparse input y associated to a mask m.",
"the user can give hints to the network in natural language",
"y and mhave the same dimensionality as the prediction y",
"and the guiding mechanism is trained to update its parame-",
"and the ground truth y. m is a binary mask that indicates",
"ters after parsing the user's inputs. To the best of our knowl-",
"the locations where a hint (i.e. prior knowledge) is given.",
"edge, this is a topic that has not been previously studied.",
"Prior knowledge can be either directly given by the user or",
"it could be a prior computed by another source.",
"Training with queries Similar to prior work in related",
"In semantic segmentation, one can think of the hint as",
"fields, we use a recurrent neural network (RNN)for pro-",
"a single (or more) pixel(s) for which the user provides the",
"cessing natural language inputs. We first encode the in-",
"true class -\"this [pixel] is a dog\" as additional informa-",
"put query using a word embedding matrix, pre-trained on",
"tion. Prior to guiding, a certain loss L(t(h()), y) has been",
"a large corpus,to acquire a fixed-length embedding vector",
"minimized during training of the network for a given task.",
"per word. The embedded words are fed as inputs to a Gated",
"We now optimize towards the same objective, e.g. per-pixel",
"RecurrentUnit(GRU)14.15ateachtimestep.Wefreeze",
"cross entropy for segmentation, but use the mask m to only",
"and do not further train the word embedding alongside our",
"consider the influence of the hint and minimize for the guid-",
"guiding module, to retain a meaningful word representation.",
"ing parameters, as opposed to the network's parameters, i.e.",
"The guiding parameters Q, β and are predicted as a linear",
"mapping from the final hidden state of the GRU.",
"= argmin [m o C(y*, 9)] :",
"(4)",
"The language-guided module is trained as follows. We",
"α.B.",
"α*,β*,",
"first generate an initial prediction with the fixed, task-",
"In this case, we only update the guiding variables for the",
"specific CNN, without influence from the guide. We then",
"current specific input and hint g, whereas the network's",
"feed prediction and ground truth into a hint generator, which",
"weights are not trained further. The minimization finds the",
"produces a query (e.g. \"the sky is not visible in this image\",",
"best parameters Q*, β*, * conditioned on the hint. The key",
"thus mimicking the user. The query is then encoded into a",
"insight is that this results in an overall adjusted prediction.",
"representationthatbecomesresponsibleforestimatinga,",
"Since the guiding block and the network t(h()) are",
"and that will guide the feature map using (3) and sub-",
"differentiable, we can minimize (4) using standard back-",
"sequently update the prediction. The standard loss for the",
"propagation and gradient descent with momentum.Intu-",
"given task (pixel-wise cross entropy loss for semantic seg-",
"itively, the tendency of gradient descent to fall into local",
"mentation) is re-weighted giving positive weight (1) to the",
"optima is desirable here. We are looking for the smallest",
"class(es) mentioned in the query to encourage changes in",
"possible a, β and that brings the guided prediction closer",
"the prediction that coincide with the given hint. All wrongly",
"to the hint while avoiding degenerate solutions such as pre-",
"predicted pixels are given a zero weight to prevent hints",
"dicting the whole image as the hinted class.",
"frombeing associated withother visual classes.Initially",
"correct pixels are weighted 0.5 to discourage corrupting cor-",
"3.3. Learning to Guide with Text",
"rectly classified regions.",
"While the previous idea is straightforward and simple to",
"Generating queries Previous work mostly relies on",
"apply to any network, it requires the hint to be given in the",
"human-annotated queries, which make them rich in variety;"
] |
{
"arxiv_id": "1803.11544",
"language": "en",
"page_idx": 3
}
|
||
1803.11544-en-4
|
1803.11544-en-4.jpg
|
however, in this case they would limit the model to a single interaction, since new annotations cannot be recorded adaptively during training. Instead, our approach uses visiononly datasets and does not require visual/textual annotations, such as captions [40], referring expressions [28, 35] and region-description correspondences [47]. Our method aims at aiding the network to correct predictions with various mistakes, rather than producing a segmentation result on demand given an input expression. Therefore, it requires textual expressions that are synthesized on-the-fly from visual categories, by comparing the initial prediction and the ground truth segmentation map.
For the generation of the queries we use a combination of functionality, semantic categories and spatial layouts (see Figure 2). Functional categories are defined by a set of operations that can be carried out on the output to improve it, such as discovering missing semantic classes, suppressing noise or replacing wrongly predicted pixels with another class. The set we used in our experiments consists of two operations, $i.e.$ find to handle classes missing in the initial prediction and remove to correct wrongly predicted labels.
Each query is built by its function and two placeholders, the entries of which are randomly selected at each training step from a set of plausible combinations based on prediction and the ground truth. We first divide the output of the network into a $N\times N$ grid. In each grid cell, we search for all erroneous classes, either missing or mistaken, while ignoring tiny spurious regions comprised of only a few pixels. Next, we randomly sample a class from the generated list of possible choices and use its semantic name for the textual expression $(e.g.“find$ the person"). The sampling probability is proportional to the potential improvement in the prediction. We then track the class position in the image based on the cells where it was found. Different combinations of cells define different spatial attention areas which can be then converted into text phrases such as“on the top left",“on the bottom”,“in the middle”.
Eventually, the proposed approach can generate textual phrases automatically and online. The guide is thus trained to understand language using vision-only annotations, $i.e.$ segmentation masks. The guiding block is able to discover semantic and spatial correspondences between the text input and the visual feature representation. During testing the guide can then interpret the commands of a real user.
## $4. \textbf{Experiments}$
We evaluate our guiding framework under multiple aspects. First, we guide semantic segmentation by backpropagation. This allows us to directly evaluate the performance of the guiding and show how it can be deployed into a model without any additional training. Second, we thoroughly investigate guiding with textual hints.
<table>
<tbody>
<tr>
<th>$\#$questions</th>
<th>0</th>
<th>1</th>
<th>5</th>
<th>10</th>
<th>15</th>
<th>20</th>
</tr>
<tr>
<td>$\mathbf{FCN-8s}$ mloll</td>
<td> </td>
<td>55 2 </td>
<td>73 1</td>
<td>$Q$ 76 1 T</td>
<td>777 2</td>
<td>81 1</td>
</tr>
<tr>
<td>p.accuracy</td>
<td>91.1</td>
<td>91.8</td>
<td>94.1</td>
<td>95.3</td>
<td>96.0</td>
<td>96.3</td>
</tr>
</tbody>
</table>
Table 1. Performance after a number of questions. We guide
a pre-trained FCN-8s [41] on PascalVOC 2012 val set[22] directly, using back-propagation. We report the mean intersection over union (mIoU) score and pixel accuracy. Every interaction with the user improves the result.
# 4.1. Guiding by Backpropagation
We investigate the performance gain by employing our guiding block directly on a fixed, pre-trained CNN. The task is semantic segmentation on the PascalVOC 2012 dataset [22]. We use a pre-trained FCN-8s network[41] and insert a guiding block in the smallest encoding layer
A user interaction scheme similar to the 20-question game of[55] is set up. After an inference step, the network is allowed to ask the user for the class of a single pixel and the guiding layer updates the feature representation using (4). The queried pixel is the one with the smallest posterior probability difference between the two most confident classes. This pixel has the highest interclass uncertainty, meaning that it is the most likely to flip. After each question the prediction is updated and the mean intersection over union (mIoU) is computed.
We have intentionally chosen a somewhat “outdated”architecture since we believe that user interaction is mostly necessary in tasks in which the performance is not close to optimal. We list the performance after 0,1,5,10,15 and 20 questions in Table 1, where 0 denotes the initial performance of FCN-8s without guiding. Over the course of 20 interactions with the user, a significant improvement of the performance from 62.6% to 81.0% is recorded. It is noteworthy that the top entry on the PascalVOC leaderboard (DeepLab-v3[13]) currently scores 86.9% mIoU, when trained with additional data. This demonstrates the benefit of guiding by back-propagation: it can be directly incorporated into a pre-trained CNN and, without any further training, it boosts a comparably low performance to reach the state of the art.
## 4.2. Guiding with Text Inputs
Due to the high accuracy of current methods on the PascalVOC semantic segmentation task, bringing a human into the loop to request improvements was not found to be meaningful with state-of-the-art models. We wish to evaluate our guiding module under a more challenging setting, in which even the performance of a state-of-the-art model is not satisfactory and interaction with a user can be beneficial.
For this purpose, we have chosen to use a dataset with
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.62158203125,
0.0947265625,
0.6396484375,
0.0947265625,
0.6396484375,
0.10919189453125,
0.62158203125,
0.10919189453125
],
[
0.71728515625,
0.0947265625,
0.734375,
0.0947265625,
0.734375,
0.10919189453125,
0.71728515625,
0.10919189453125
],
[
0.7587890625,
0.0947265625,
0.78173828125,
0.0947265625,
0.78173828125,
0.10797119140625,
0.7587890625,
0.10797119140625
],
[
0.80810546875,
0.0947265625,
0.82861328125,
0.0947265625,
0.82861328125,
0.10919189453125,
0.80810546875,
0.10919189453125
],
[
0.85546875,
0.0947265625,
0.87744140625,
0.0947265625,
0.87744140625,
0.10797119140625,
0.85546875,
0.10797119140625
],
[
0.6650390625,
0.09783935546875,
0.6806640625,
0.091552734375,
0.689453125,
0.10479736328125,
0.6748046875,
0.111083984375
],
[
0.5146484375,
0.09661865234375,
0.58642578125,
0.09661865234375,
0.58642578125,
0.10797119140625,
0.5146484375,
0.10797119140625
],
[
0.07843017578125,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.51318359375,
0.11553955078125,
0.57275390625,
0.11553955078125,
0.57275390625,
0.1300048828125,
0.51318359375,
0.1300048828125
],
[
0.080078125,
0.12371826171875,
0.4697265625,
0.12371826171875,
0.4697265625,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.60107421875,
0.1300048828125,
0.63818359375,
0.1300048828125,
0.63818359375,
0.1458740234375,
0.60107421875,
0.1458740234375
],
[
0.7939453125,
0.1300048828125,
0.8310546875,
0.1300048828125,
0.8310546875,
0.1458740234375,
0.7939453125,
0.1458740234375
],
[
0.650390625,
0.1319580078125,
0.6845703125,
0.1319580078125,
0.6845703125,
0.14453125,
0.650390625,
0.14453125
],
[
0.51318359375,
0.1307373046875,
0.55810546875,
0.1307373046875,
0.55810546875,
0.1458740234375,
0.51318359375,
0.1458740234375
],
[
0.69921875,
0.1319580078125,
0.73388671875,
0.1319580078125,
0.73388671875,
0.14453125,
0.69921875,
0.14453125
],
[
0.74658203125,
0.1319580078125,
0.78125,
0.1319580078125,
0.78125,
0.14453125,
0.74658203125,
0.14453125
],
[
0.84228515625,
0.1319580078125,
0.8759765625,
0.1319580078125,
0.8759765625,
0.14453125,
0.84228515625,
0.14453125
],
[
0.080078125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.60107421875,
0.14453125,
0.63818359375,
0.146484375,
0.63623046875,
0.1622314453125,
0.6005859375,
0.160400390625
],
[
0.74609375,
0.1458740234375,
0.78173828125,
0.1458740234375,
0.78173828125,
0.1610107421875,
0.74609375,
0.1610107421875
],
[
0.79345703125,
0.1458740234375,
0.8310546875,
0.1458740234375,
0.8310546875,
0.1610107421875,
0.79345703125,
0.1610107421875
],
[
0.8408203125,
0.14453125,
0.87841796875,
0.146484375,
0.87744140625,
0.1610107421875,
0.8388671875,
0.1590576171875
],
[
0.69775390625,
0.146484375,
0.73388671875,
0.146484375,
0.73388671875,
0.1590576171875,
0.69775390625,
0.1590576171875
],
[
0.5146484375,
0.1483154296875,
0.58642578125,
0.1483154296875,
0.58642578125,
0.160400390625,
0.5146484375,
0.160400390625
],
[
0.65185546875,
0.146484375,
0.6845703125,
0.146484375,
0.6845703125,
0.1590576171875,
0.65185546875,
0.1590576171875
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50244140625,
0.169189453125,
0.89208984375,
0.169189453125,
0.89208984375,
0.18310546875,
0.50244140625,
0.18310546875
],
[
0.07843017578125,
0.1817626953125,
0.468017578125,
0.1817626953125,
0.468017578125,
0.1956787109375,
0.07843017578125,
0.1956787109375
],
[
0.50244140625,
0.184326171875,
0.89208984375,
0.184326171875,
0.89208984375,
0.1976318359375,
0.50244140625,
0.1976318359375
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50244140625,
0.210205078125
],
[
0.080078125,
0.21337890625,
0.47216796875,
0.21337890625,
0.47216796875,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.212158203125,
0.89208984375,
0.212158203125,
0.89208984375,
0.2259521484375,
0.5009765625,
0.2259521484375
],
[
0.080078125,
0.2291259765625,
0.46728515625,
0.2291259765625,
0.46728515625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.5009765625,
0.2259521484375,
0.70166015625,
0.2259521484375,
0.70166015625,
0.2392578125,
0.5009765625,
0.2392578125
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.289306640625,
0.2587890625,
0.289306640625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.266357421875,
0.7587890625,
0.266357421875,
0.7587890625,
0.27978515625,
0.50341796875,
0.27978515625
],
[
0.09722900390625,
0.27587890625,
0.4697265625,
0.27587890625,
0.4697265625,
0.2890625,
0.09722900390625,
0.2890625
],
[
0.08087158203125,
0.292236328125,
0.468017578125,
0.292236328125,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.51953125,
0.28857421875,
0.89208984375,
0.2890625,
0.89208984375,
0.3037109375,
0.51953125,
0.302978515625
],
[
0.080078125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.080078125,
0.3193359375
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.50244140625,
0.3193359375
],
[
0.080078125,
0.3212890625,
0.4697265625,
0.3212890625,
0.4697265625,
0.335205078125,
0.080078125,
0.335205078125
],
[
0.50244140625,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.50244140625,
0.333984375
],
[
0.07843017578125,
0.335205078125,
0.4697265625,
0.3359375,
0.4697265625,
0.35107421875,
0.07843017578125,
0.349853515625
],
[
0.50341796875,
0.335205078125,
0.89208984375,
0.335205078125,
0.89208984375,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.080078125,
0.3515625,
0.4697265625,
0.3515625,
0.4697265625,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.50244140625,
0.35107421875,
0.80810546875,
0.35107421875,
0.80810546875,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.080078125,
0.3662109375,
0.4697265625,
0.3662109375,
0.4697265625,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.521484375,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.521484375,
0.380126953125
],
[
0.07843017578125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.39453125,
0.07843017578125,
0.39453125
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.07843017578125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.07843017578125,
0.410888671875
],
[
0.5009765625,
0.395751953125,
0.89208984375,
0.396484375,
0.89208984375,
0.410888671875,
0.5009765625,
0.410400390625
],
[
0.09722900390625,
0.4140625,
0.468017578125,
0.4140625,
0.468017578125,
0.427978515625,
0.09722900390625,
0.427978515625
],
[
0.5009765625,
0.410888671875,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.5009765625,
0.4248046875
],
[
0.07843017578125,
0.427978515625,
0.4697265625,
0.4287109375,
0.4697265625,
0.443115234375,
0.07843017578125,
0.442626953125
],
[
0.50341796875,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50341796875,
0.43994140625
],
[
0.08087158203125,
0.44580078125,
0.46728515625,
0.44580078125,
0.46728515625,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.50244140625,
0.44140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455078125,
0.50244140625,
0.455078125
],
[
0.07843017578125,
0.458984375,
0.4697265625,
0.458984375,
0.4697265625,
0.472900390625,
0.07843017578125,
0.472900390625
],
[
0.50244140625,
0.45703125,
0.89208984375,
0.45703125,
0.89208984375,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.080078125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.5009765625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.5009765625,
0.486083984375
],
[
0.080078125,
0.490478515625,
0.4697265625,
0.490478515625,
0.4697265625,
0.50390625,
0.080078125,
0.50390625
],
[
0.5009765625,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.5009765625,
0.50048828125
],
[
0.080078125,
0.5048828125,
0.4697265625,
0.5048828125,
0.4697265625,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.50390625,
0.6787109375,
0.50390625,
0.6787109375,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.080078125,
0.53271484375
],
[
0.521484375,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.521484375,
0.53076171875
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.50244140625,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.546875,
0.50244140625,
0.546875
],
[
0.080078125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.56298828125,
0.080078125,
0.56298828125
],
[
0.5009765625,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.5009765625,
0.56103515625
],
[
0.07843017578125,
0.56494140625,
0.4697265625,
0.564453125,
0.4697265625,
0.57763671875,
0.07843017578125,
0.5791015625
],
[
0.5009765625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.5009765625,
0.5771484375
],
[
0.080078125,
0.580078125,
0.471435546875,
0.580078125,
0.471435546875,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.5791015625,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.080078125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50244140625,
0.59326171875,
0.89208984375,
0.59326171875,
0.89208984375,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.609375,
0.89208984375,
0.609375,
0.89208984375,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.080078125,
0.62548828125,
0.4697265625,
0.62548828125,
0.4697265625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.50244140625,
0.62353515625,
0.8935546875,
0.62353515625,
0.8935546875,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.080078125,
0.64013671875,
0.3349609375,
0.64013671875,
0.3349609375,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.5009765625,
0.63818359375,
0.89208984375,
0.63720703125,
0.89208984375,
0.6513671875,
0.5009765625,
0.65283203125
],
[
0.09722900390625,
0.6572265625,
0.468017578125,
0.65576171875,
0.468017578125,
0.67041015625,
0.09722900390625,
0.671875
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.50341796875,
0.66845703125,
0.89208984375,
0.66845703125,
0.89208984375,
0.681640625,
0.50341796875,
0.681640625
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68505859375,
0.08087158203125,
0.68505859375
],
[
0.07843017578125,
0.68798828125,
0.4697265625,
0.68798828125,
0.4697265625,
0.7021484375,
0.07843017578125,
0.7021484375
],
[
0.50244140625,
0.6845703125,
0.890625,
0.6845703125,
0.890625,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50244140625,
0.69970703125,
0.890625,
0.69970703125,
0.890625,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.080078125,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.5009765625,
0.71337890625,
0.89208984375,
0.7138671875,
0.89208984375,
0.72900390625,
0.5009765625,
0.72802734375
],
[
0.080078125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.080078125,
0.74755859375
],
[
0.50244140625,
0.72998046875,
0.66650390625,
0.72998046875,
0.66650390625,
0.74365234375,
0.50244140625,
0.74365234375
],
[
0.080078125,
0.74755859375,
0.426513671875,
0.74755859375,
0.426513671875,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.50244140625,
0.7548828125,
0.7294921875,
0.7548828125,
0.7294921875,
0.76953125,
0.50244140625,
0.76953125
],
[
0.08087158203125,
0.7841796875,
0.2091064453125,
0.7841796875,
0.2091064453125,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.521484375,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.521484375,
0.79345703125
],
[
0.5009765625,
0.79345703125,
0.89208984375,
0.79443359375,
0.89208984375,
0.8095703125,
0.5009765625,
0.80810546875
],
[
0.098876953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.098876953125,
0.82373046875
],
[
0.50341796875,
0.8125,
0.8896484375,
0.8125,
0.8896484375,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.07843017578125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.5009765625,
0.869140625
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.85693359375,
0.8720703125,
0.85693359375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.404296875,
0.88720703125,
0.404296875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.52294921875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.52294921875,
0.90087890625
]
] |
[
"however, in this case they would limit the model to a single",
"0",
"5",
"10",
"15",
"20",
"1",
"#questions",
"interaction, since new annotations cannot be recorded adap-",
"FCN-8s",
"tively during training. Instead, our approach uses vision-",
"62.6",
"77.3",
"65.3",
"mloU",
"73.1",
"76.9",
"81.0",
"only datasets and does not require visual/textual annota-",
"91.1",
"95.3",
"96.0",
"96.3",
"94.1",
"p.accuracy",
"91.8",
"tions,such as captions[40],referring expressions[28,35]",
"andregion-description correspondences[47].Ourmethod",
"Table 1. Performance after a number of questions. We guide",
"aims at aiding the network to correct predictions with var-",
"a pre-trained FCN-8s [41] on PascalVOC 2012 val set [22] di-",
"ious mistakes, rather than producing a segmentation result",
"rectly,using back-propagation.We report the mean intersection",
"on demand given an input expression.Therefore, it requires",
"over union (mloU) score and pixel accuracy. Every interaction",
"textual expressions that are synthesized on-the-fly from vi-",
"with the user improves the result.",
"sual categories, by comparing the initial prediction and the",
"ground truth segmentation map.",
"4.1. Guiding by Backpropagation",
"For the generation of the queries we use a combination",
"offunctionality,semanticcategoriesand spatial layouts(see",
"We investigate the performance gain by employing our",
"Figure 2). Functional categories are defined by a set of op-",
"guiding block directly on a fixed, pre-trained CNN. The task",
"erations that can be carried out on the output to improve",
"is semantic segmentation on the PascalVOC 2012 dataset",
"it, such as discovering missing semantic classes, suppress-",
"[22]. We use a pre-trained FCN-8s network [41] and insert",
"ing noise or replacing wrongly predicted pixels with another",
"a guiding block in the smallest encoding layer.",
"class. The set we used in our experiments consists of two",
"A user interaction scheme similar to the 20-question",
"operations, i.e. f ind to handle classes missing in the initial",
"game of [55] is set up. After an inference step, the network",
"prediction and remove to correct wrongly predicted labels.",
"is allowed to ask the user for the class of a single pixel and",
"Each query is built by its function and two placeholders,",
"the guiding layer updates the feature representation using",
"the entries of which are randomly selected at each training",
"(4). The queried pixel is the one with the smallest poste-",
"stepfromasetofplausiblecombinationsbasedonpredic-",
"rior probability difference between the two most confident",
"tion and the ground truth. We first divide the output of the",
"classes. This pixel has the highest interclass uncertainty,",
"network into a N × N grid. In each grid cell, we search",
"meaning that it is the most likely to flip. After each ques-",
"for all erroneous classes, either missing or mistaken, while",
"tion the prediction is updated and the mean intersection over",
"ignoring tiny spurious regions comprised of only a few pix-",
"union(mloU)iscomputed",
"els. Next, we randomly sample a class from the generated",
"We have intentionally chosen a somewhat outdated\" ar-",
"list of possible choices and use its semantic name for the",
"chitecture since we believe that user interaction is mostly",
"textual expression (e.g. \"find the person\"). The sampling",
"necessary in tasks in which the performance is not close",
"probability is proportional to the potential improvement in",
"to optimal. We list the performance after 0, 1, 5, 10, 15",
"the prediction. We then track the class position in the image",
"and 20 questions in Table 1, where 0 denotes the initial",
"based on the cells where it was found. Different combina-",
"performance of FCN-8s without guiding. Over the course",
"tions of cells define different spatial attention areas which",
"of 20 interactions with the user, a significant improvement",
"can be then converted into text phrases such as “on the top",
"of the performance from 62.6% to 81.0% is recorded. It",
"leff\",\"on the bottom\", \"in the middle\"",
"is noteworthy that the top entry on the PascalvOC leader-",
"Eventually,the proposed approach can generate textual",
"board (DeepLab-v3 [13]) currently scores 86.9% mloU,",
"when trained with additional data. This demonstrates the",
"phrases automatically and online.Theguide is thustrained",
"to understand language using vision-only annotations, i.e.",
"benefit of guiding by back-propagation: it can be directly",
"segmentationmasks.Theguiding blockis abletodiscover",
"incorporated into a pre-trained CNN and, without any fur-",
"semantic and spatial correspondences between the text in-",
"ther training, it boosts a comparably low performance to",
"put and the visual feature representation. During testing the",
"reach the state of the art.",
"guide can then interpret the commands of a real user.",
"4.2. Guiding with Text Inputs",
"4.Experiments",
"Due to the high accuracy of current methods on the Pas-",
"calvOC semantic segmentation task, bringing a human into",
"We evaluate our guiding framework under multiple as-",
"thelooptorequestimprovementswasnotfoundtobemean-",
"pects.First, we guide semantic segmentation by back-",
"ingful with state-of-the-art models. We wish to evaluate our",
"propagation. This allows us to directly evaluate the per-",
"guiding module under a more challenging setting, in which",
"formance of the guiding and show how it can be deployed",
"even the performance of a state-of-the-art model is not sat-",
"into a model without any additional training. Second, we",
"isfactory and interaction with a user can be beneficial.",
"thoroughly investigate guiding with textual hints.",
"For this purpose, we have chosen to use a dataset with"
] |
{
"arxiv_id": "1803.11544",
"language": "en",
"page_idx": 4
}
|
||
1805.06549-en-0
|
1805.06549-en-0.jpg
|
Defoiling Foiled Image Captions
Pranava Madhyastha,Josiah Wang and Lucia Specia
Department of Computer Science
University of Sheffield, UK
$\{$p.madhyastha, j.k.wang, l.specia$\}$@sheffield.ac.uk
## Abstract
We address the task of detecting foiled image captions, i.e. identifying whether a caption contains a word that has been deliberately replaced by a semantically similar word, thus rendering it inaccurate with respect to the image being described. Solving this problem should in principle require a fine-grained understanding of images to detect linguistically valid perturbations in captions. In such contexts, encoding sufficiently descriptive image information becomes a key challenge. In this paper, we demonstrate that it is possible to solve this task using simple, interpretable yet powerful representations based on explicit object information. Our models achieve stateof-the-art performance on a standard dataset, with scores exceeding those achieved by humans on the task. We also measure the upperbound performance of our models using gold standard annotations. Our analysis reveals that the simpler model performs well even without image information, suggesting that the dataset contains strong linguistic bias.
son et al., 2017; Antol et al., 2015; Chen et al., 2015; Gao et al., 2015; Yu et al., 2015; Zhu et al., 2016).
FOIL (Shekhar et al.,2017b) is one such dataset. It was proposed to evaluate the ability of V2L models in understanding the interplay of objects and their attributes in the images and their relations in an image captioning framework. This is done by replacing a word in MSCOCO (Lin et al., 2014) captions with a‘foiled’word that is semantically similar or related to the original word (substituting dog with $cat)$, thus rendering the image caption unfaithful to the image content, while yet linguistically valid. Shekhar et al. (2017b) report poor performance for V2L models in classifying captions as foiled (or not). They suggested that their models (using image embeddings as input) are very poor at encoding structured visuallinguistic information to spot the mismatch between a foiled caption and the corresponding content depicted in the image.
In this paper, we focus on the foiled captions classification task (Section 2), and propose the use of explicit object detections as salient image cues for solving the task. In contrast to methods from previous work that make use of word based information extracted from captions (Heuer et al., 2016; Yao et al., 2016; Wu et al., 2018), we use explicit object category information directly extracted from the images. More specifically, we use an interpretable $bag\textit{of objects as image represen- }$ tation for the classifier. Our hypothesis is that, to truly `understand’the image, V2L models should exploit information about objects and their relations in the image and not just global, low-level image embeddings as used by most V2L models.
Our main contributions are:
## 1 Introduction
Models tackling vision-to-language (V2L) tasks, for example Image Captioning (IC) and Visual Question Answering (VQA), have demonstrated impressive results in recent years in terms of automatic metric scores. However, whether or not these models are actually learning to address the tasks they are designed for is questionable. For example, Hodosh and Hockenmaier (2016) showed that IC models do not understand images suffciently, as reflected by the generated captions. As a consequence, in the last few years many diagnostic tasks and datasets have been proposed aiming at investigating the capabilities of such models in more detail to determine whether and how these models are capable of exploiting visual and/or linguistic information (Shekhar et al., 2017b; John-
## 1. A model (Section 3) for foiled captions classification using a simple and interpretable
1
Please cite this paper as: Pranava Madhyastha, Josiah Wang, and Lucia Specia (2018)
Defoiling Foiled Image Captions.
In Proceedings of 2018 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies (NAACL-HLT 2018).
|
[
[
0.333251953125,
0.08074951171875,
0.66748046875,
0.08197021484375,
0.66748046875,
0.09979248046875,
0.333251953125,
0.09857177734375
],
[
0.273681640625,
0.135986328125,
0.73486328125,
0.135986328125,
0.73486328125,
0.149658203125,
0.273681640625,
0.149658203125
],
[
0.36865234375,
0.152587890625,
0.640625,
0.152587890625,
0.640625,
0.166259765625,
0.36865234375,
0.166259765625
],
[
0.391357421875,
0.169189453125,
0.615234375,
0.169189453125,
0.615234375,
0.182861328125,
0.391357421875,
0.182861328125
],
[
0.198974609375,
0.185302734375,
0.80029296875,
0.185302734375,
0.80029296875,
0.203125,
0.198974609375,
0.203125
],
[
0.264404296875,
0.262451171875,
0.343505859375,
0.26416015625,
0.343505859375,
0.27978515625,
0.264404296875,
0.27783203125
],
[
0.51123046875,
0.265380859375,
0.8818359375,
0.26416015625,
0.8818359375,
0.279052734375,
0.51123046875,
0.27978515625
],
[
0.5146484375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.019317626953125,
0.29443359375,
0.06634521484375,
0.29443359375,
0.0697021484375,
0.441162109375,
0.0218353271484375,
0.441162109375
],
[
0.14697265625,
0.290283203125,
0.459228515625,
0.290283203125,
0.459228515625,
0.303955078125,
0.14697265625,
0.303955078125
],
[
0.51318359375,
0.297607421875,
0.5634765625,
0.297607421875,
0.5634765625,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.145263671875,
0.3046875,
0.459228515625,
0.3046875,
0.459228515625,
0.319580078125,
0.145263671875,
0.319580078125
],
[
0.14697265625,
0.318359375,
0.459228515625,
0.318359375,
0.459228515625,
0.33203125,
0.14697265625,
0.33203125
],
[
0.52880859375,
0.3134765625,
0.884765625,
0.3154296875,
0.884765625,
0.330078125,
0.52880859375,
0.327880859375
],
[
0.14697265625,
0.333251953125,
0.459228515625,
0.333251953125,
0.459228515625,
0.3466796875,
0.14697265625,
0.3466796875
],
[
0.51318359375,
0.33203125,
0.884765625,
0.33203125,
0.884765625,
0.345703125,
0.51318359375,
0.345703125
],
[
0.14697265625,
0.3466796875,
0.459228515625,
0.3466796875,
0.459228515625,
0.35986328125,
0.14697265625,
0.35986328125
],
[
0.5146484375,
0.347412109375,
0.8818359375,
0.347412109375,
0.8818359375,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.14697265625,
0.3623046875,
0.4609375,
0.3623046875,
0.4609375,
0.3759765625,
0.14697265625,
0.3759765625
],
[
0.51318359375,
0.363525390625,
0.884765625,
0.363525390625,
0.884765625,
0.377685546875,
0.51318359375,
0.377685546875
],
[
0.14697265625,
0.3759765625,
0.459228515625,
0.3759765625,
0.459228515625,
0.3896484375,
0.14697265625,
0.3896484375
],
[
0.51318359375,
0.380126953125,
0.88330078125,
0.380126953125,
0.88330078125,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.14697265625,
0.3896484375,
0.45751953125,
0.3896484375,
0.45751953125,
0.402587890625,
0.14697265625,
0.402587890625
],
[
0.51318359375,
0.3955078125,
0.88330078125,
0.3955078125,
0.88330078125,
0.410400390625,
0.51318359375,
0.410400390625
],
[
0.1485595703125,
0.40625,
0.45751953125,
0.40625,
0.45751953125,
0.416259765625,
0.1485595703125,
0.416259765625
],
[
0.5146484375,
0.410888671875,
0.88330078125,
0.410888671875,
0.88330078125,
0.42578125,
0.5146484375,
0.42578125
],
[
0.14697265625,
0.418701171875,
0.459228515625,
0.418701171875,
0.459228515625,
0.431640625,
0.14697265625,
0.431640625
],
[
0.14697265625,
0.432861328125,
0.4609375,
0.432861328125,
0.4609375,
0.446533203125,
0.14697265625,
0.446533203125
],
[
0.51318359375,
0.427490234375,
0.88330078125,
0.427490234375,
0.88330078125,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.0218353271484375,
0.43701171875,
0.06378173828125,
0.43701171875,
0.06378173828125,
0.5185546875,
0.0218353271484375,
0.5185546875
],
[
0.14697265625,
0.44775390625,
0.4609375,
0.44775390625,
0.4609375,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.5146484375,
0.44482421875,
0.8818359375,
0.44482421875,
0.8818359375,
0.457763671875,
0.5146484375,
0.457763671875
],
[
0.1485595703125,
0.46142578125,
0.459228515625,
0.46142578125,
0.459228515625,
0.474365234375,
0.1485595703125,
0.474365234375
],
[
0.51318359375,
0.460205078125,
0.88330078125,
0.458984375,
0.88330078125,
0.473876953125,
0.51318359375,
0.474365234375
],
[
0.1485595703125,
0.476806640625,
0.45751953125,
0.476806640625,
0.45751953125,
0.48681640625,
0.1485595703125,
0.48681640625
],
[
0.51318359375,
0.4755859375,
0.8818359375,
0.4755859375,
0.8818359375,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.14697265625,
0.4892578125,
0.459228515625,
0.4892578125,
0.459228515625,
0.5029296875,
0.14697265625,
0.5029296875
],
[
0.51318359375,
0.4921875,
0.8818359375,
0.4912109375,
0.8818359375,
0.505859375,
0.51318359375,
0.50732421875
],
[
0.14697265625,
0.50439453125,
0.459228515625,
0.50439453125,
0.459228515625,
0.51708984375,
0.14697265625,
0.51708984375
],
[
0.5146484375,
0.50830078125,
0.884765625,
0.50830078125,
0.884765625,
0.5224609375,
0.5146484375,
0.5224609375
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.14697265625,
0.51708984375,
0.459228515625,
0.51708984375,
0.459228515625,
0.53076171875,
0.14697265625,
0.53076171875
],
[
0.5146484375,
0.52490234375,
0.88330078125,
0.52490234375,
0.88330078125,
0.5380859375,
0.5146484375,
0.5380859375
],
[
0.14697265625,
0.53076171875,
0.456787109375,
0.53076171875,
0.456787109375,
0.54443359375,
0.14697265625,
0.54443359375
],
[
0.51318359375,
0.54052734375,
0.88330078125,
0.5390625,
0.88330078125,
0.55419921875,
0.51318359375,
0.5546875
],
[
0.14697265625,
0.54638671875,
0.459228515625,
0.54638671875,
0.459228515625,
0.56005859375,
0.14697265625,
0.56005859375
],
[
0.1485595703125,
0.5625,
0.459228515625,
0.5625,
0.459228515625,
0.572265625,
0.1485595703125,
0.572265625
],
[
0.5146484375,
0.55712890625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57080078125,
0.5146484375,
0.57080078125
],
[
0.14697265625,
0.576171875,
0.45751953125,
0.57470703125,
0.45751953125,
0.5859375,
0.14697265625,
0.58740234375
],
[
0.51123046875,
0.57177734375,
0.8818359375,
0.572265625,
0.8818359375,
0.58740234375,
0.51123046875,
0.5859375
],
[
0.14697265625,
0.5888671875,
0.4609375,
0.5888671875,
0.4609375,
0.6025390625,
0.14697265625,
0.6025390625
],
[
0.51123046875,
0.58837890625,
0.7109375,
0.5888671875,
0.7109375,
0.60400390625,
0.51123046875,
0.6025390625
],
[
0.14697265625,
0.60400390625,
0.355224609375,
0.60400390625,
0.355224609375,
0.61767578125,
0.14697265625,
0.61767578125
],
[
0.53076171875,
0.60546875,
0.88330078125,
0.60546875,
0.88330078125,
0.619140625,
0.53076171875,
0.619140625
],
[
0.51123046875,
0.62060546875,
0.884765625,
0.62158203125,
0.884765625,
0.6357421875,
0.51123046875,
0.63525390625
],
[
0.124267578125,
0.6318359375,
0.130126953125,
0.6318359375,
0.130126953125,
0.63720703125,
0.124267578125,
0.63720703125
],
[
0.130126953125,
0.62744140625,
0.26025390625,
0.62744140625,
0.26025390625,
0.642578125,
0.130126953125,
0.642578125
],
[
0.51318359375,
0.63818359375,
0.884765625,
0.63818359375,
0.884765625,
0.65185546875,
0.51318359375,
0.65185546875
],
[
0.118408203125,
0.65283203125,
0.490234375,
0.65283203125,
0.490234375,
0.66748046875,
0.118408203125,
0.66748046875
],
[
0.51123046875,
0.65380859375,
0.884765625,
0.65283203125,
0.884765625,
0.66748046875,
0.51123046875,
0.66845703125
],
[
0.12005615234375,
0.66943359375,
0.489501953125,
0.66943359375,
0.489501953125,
0.68408203125,
0.12005615234375,
0.68408203125
],
[
0.51318359375,
0.67041015625,
0.88330078125,
0.67041015625,
0.88330078125,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.12005615234375,
0.68603515625,
0.489501953125,
0.68603515625,
0.489501953125,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.5146484375,
0.6884765625,
0.8818359375,
0.6884765625,
0.8818359375,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.118408203125,
0.703125,
0.486083984375,
0.703125,
0.486083984375,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.51123046875,
0.70068359375,
0.88330078125,
0.70166015625,
0.88330078125,
0.71630859375,
0.51123046875,
0.71533203125
],
[
0.118408203125,
0.71875,
0.490234375,
0.71875,
0.490234375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.51318359375,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.51318359375,
0.732421875
],
[
0.12005615234375,
0.73388671875,
0.490234375,
0.73388671875,
0.490234375,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.51123046875,
0.73388671875,
0.880859375,
0.73388671875,
0.880859375,
0.74755859375,
0.51123046875,
0.74755859375
],
[
0.118408203125,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.51318359375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.78076171875,
0.51318359375,
0.78076171875
],
[
0.12005615234375,
0.7822265625,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.783203125,
0.88330078125,
0.783203125,
0.88330078125,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.118408203125,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.118408203125,
0.8125
],
[
0.51123046875,
0.798828125,
0.8818359375,
0.79736328125,
0.8818359375,
0.8125,
0.51123046875,
0.8134765625
],
[
0.118408203125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.51123046875,
0.8154296875,
0.8818359375,
0.8154296875,
0.8818359375,
0.8291015625,
0.51123046875,
0.8291015625
],
[
0.1175537109375,
0.830078125,
0.489501953125,
0.83056640625,
0.489501953125,
0.845703125,
0.1175537109375,
0.84423828125
],
[
0.51318359375,
0.83203125,
0.8759765625,
0.83203125,
0.8759765625,
0.845703125,
0.51318359375,
0.845703125
],
[
0.118408203125,
0.84716796875,
0.490234375,
0.84716796875,
0.490234375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.53076171875,
0.8486328125,
0.73876953125,
0.8486328125,
0.73876953125,
0.8623046875,
0.53076171875,
0.8623046875
],
[
0.118408203125,
0.8642578125,
0.486083984375,
0.8642578125,
0.486083984375,
0.87451171875,
0.118408203125,
0.87451171875
],
[
0.1175537109375,
0.87744140625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.52880859375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.52880859375,
0.892578125
],
[
0.1175537109375,
0.89453125,
0.489501953125,
0.8935546875,
0.489501953125,
0.90771484375,
0.1175537109375,
0.9091796875
],
[
0.5498046875,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.5498046875,
0.9091796875
],
[
0.21240234375,
0.93115234375,
0.78857421875,
0.93115234375,
0.78857421875,
0.94482421875,
0.21240234375,
0.94482421875
],
[
0.8759765625,
0.931640625,
0.88330078125,
0.93701171875,
0.8759765625,
0.9423828125,
0.869140625,
0.93701171875
],
[
0.3837890625,
0.9453125,
0.615234375,
0.9453125,
0.615234375,
0.9560546875,
0.3837890625,
0.9560546875
],
[
0.20068359375,
0.95703125,
0.80029296875,
0.95703125,
0.80029296875,
0.97021484375,
0.20068359375,
0.97021484375
],
[
0.2275390625,
0.9716796875,
0.771484375,
0.9716796875,
0.771484375,
0.982421875,
0.2275390625,
0.982421875
]
] |
[
"Defoiling Foiled Image Captions",
"Pranava Madhyastha, Josiah Wang and Lucia Specia",
"Department of Computer Science",
"University of Sheffield, UK",
"[p.madhyastha,j.k.wang, l.specia]@sheffield,ac.uk",
"Abstract",
"son et al., 2017; Antol et al., 2015; Chen et al.",
"2015; Gao et al., 2015; Yu et al., 2015; Zhu et al.,",
"810291",
"We address the task of detecting foiled im-",
"2016).",
"age captions, i.e. identifying whether a cap-",
"tion contains a word that has been deliber-",
"FOIL (Shekhar et al.,2017b)is one such",
"ately replaced by a semantically similar word,",
"dataset. It was proposed to evaluate the ability of",
"thus rendering it inaccurate with respect to the",
"V2L models in understanding the interplay of ob-",
"image being described. Solving this problem",
"jects and their attributes in the images and their",
"should in principle require a fine-grained un-",
"relations in an image captioning framework. This",
"derstanding of images to detect linguistically",
"is done by replacing a word in MSCOCO (Lin",
"valid perturbationsin captions.In such con-",
"et al., 2014) captions with a foiled' word that is",
"texts, encoding sufficiently descriptive image",
"information becomes a key challenge. In this",
"semantically similar or related to the original word",
"[A's]",
"paper, we demonstrate that it is possible to",
"(substituting dog with cat), thus rendering the im-",
"solve this task using simple, interpretable yet",
"age caption unfaithful to the image content, while",
"powerfulrepresentationsbased on explicitob-",
"yet linguistically valid. Shekhar et al. (2017b) re-",
"ject information. Our models achieve state-",
"port poor performance for V2L models in classi-",
"of-the-art performance on a standard dataset,",
"fying captions as foiled (or not). They suggested",
"590010",
"with scores exceeding those achieved by hu-",
"that their models (using image embeddings as in-",
"mans on the task. We also measure the upper.",
"put) are very poor at encoding structured visual-",
"bound performance of our models using gold",
"standard annotations.Ouranalysisreveals that",
"linguistic information to spot the mismatch be-",
"thesimplermodelperformswellevenwithout",
"tween a foiled caption and the corresponding con-",
"image information, suggesting that the dataset",
"tent depicted in the image.",
"contains strong linguistic bias.",
"In this paper, we focus on the foiled captions",
"classification task (Section 2), and propose the use",
"",
"Introduction",
"of explicit object detections as salient image cues",
"Models tackling vision-to-language (V2L) tasks,",
"for solving the task. In contrast to methods from",
"for example Image Captioning (IC) and Visual",
"previous work that make use of word based in-",
"Question Answering (VQA), have demonstrated",
"formation extracted from captions (Heuer et al..",
"impressiveresultsinrecent vears in terms of au-",
"2016; Yao et al., 2016; Wu et al.,2018), we use",
"tomatic metric scores. However, whether or not",
"explicit object category information directly ex-",
"these models are actually learning to address the",
"tracted from the images. More specifically, we use",
"tasks they are designedforis guestionable.For ex-",
"an interpretable bag of objects as image represen-",
"ample, Hodosh and Hockenmaier (2016) showed",
"tation for the classifier. Our hypothesis is that, to",
"that IC models do not understand images suffi-",
"truly understand' the image, V2L models should",
"ciently, as reflected by the generated captions. As",
"exploit information about objects and their rela-",
"a consequence, in the last few years many diagnos-",
"tions in the image and not just global, low-level",
"tic tasks and datasets have been proposed aiming",
"image embeddings as used by most V2L models.",
"at investigating the capabilities of such models in",
"Our main contributions are:",
"more detail to determine whether and how these",
"models are capable of exploiting visual and/or lin-",
"1. A model (Section 3) for foiled captions clas-",
"guistic information (Shekhar et al., 2017b; John-",
"sification using a simple and interpretable",
"Please cite this paper as: Pranava Madhyastha, Josiah Wang, and Lucia Specia (2018)",
"1",
"DefoilingFoiled Image Captions",
"In Proceedings of 2018 Conference of the North American Chapter of the Association for",
"Computational Linguistics:HumanLanguageTechnologies(NAACL-HLT2O18)"
] |
{
"arxiv_id": "1805.06549",
"language": "en",
"page_idx": 0
}
|
||
1805.06549-en-2
|
1805.06549-en-2.jpg
|
<table>
<tbody>
<tr>
<th>$P_{0}S$ Type</th>
<th>$\mathbf{Train}^{foil}$</th>
<th>$\mathbf{Test}^{\int oil}$</th>
<th>Total Train</th>
<th>Total Test</th>
</tr>
<tr>
<td>Noun</td>
<td>153,229</td>
<td>75,278</td>
<td>306,458</td>
<td>150,556</td>
</tr>
<tr>
<td>Verb</td>
<td>6,314</td>
<td>2,788</td>
<td>60,262</td>
<td>33,616</td>
</tr>
<tr>
<td>Adjective</td>
<td>15,640</td>
<td>9,009</td>
<td>73,057</td>
<td>42,163</td>
</tr>
<tr>
<td>Adverb</td>
<td>1,011</td>
<td>451</td>
<td>53,381</td>
<td>30,738</td>
</tr>
<tr>
<td>Preposition</td>
<td>8,733</td>
<td>5,551</td>
<td>77,002</td>
<td>46,018</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<th>Feats</th>
<th>0verall</th>
<th>Real</th>
<th>Foil</th>
</tr>
<tr>
<td>Blind ( LSTM only$) ^{\dagger }$</td>
<td>55.62</td>
<td>86.20</td>
<td>25.04</td>
</tr>
<tr>
<td>$\mathrm{HieCoAtt}^{\dagger}$</td>
<td>64.14</td>
<td>91.89</td>
<td>36.38</td>
</tr>
<tr>
<td>$\mathbf{CNN+BOWMLP}$</td>
<td>88.42</td>
<td>86.89</td>
<td>89.97</td>
</tr>
<tr>
<td>Predict Mention+ BOW MLP</td>
<td>94.94</td>
<td>95.68</td>
<td>94.23</td>
</tr>
<tr>
<td>Predict Freq+ BOW MLP</td>
<td>95.14</td>
<td>95.82</td>
<td>94.48</td>
</tr>
<tr>
<td>Gold Mention+ BOW MLP</td>
<td>95.83</td>
<td>96.30</td>
<td>95.36</td>
</tr>
<tr>
<td>Gold Freq+ BOW MLP</td>
<td>96.45</td>
<td>96.04</td>
<td>96.85</td>
</tr>
<tr>
<td>CNN+ LSTM</td>
<td>87.45</td>
<td>86.78</td>
<td>88.14</td>
</tr>
<tr>
<td>Predict Freq+ LSTM</td>
<td>85.99</td>
<td>85.17</td>
<td>86.81</td>
</tr>
<tr>
<td>Gold Freq+ LSTM</td>
<td>87.38</td>
<td>86.62</td>
<td>88.18</td>
</tr>
<tr>
<td>Predict Freq+ MM- LSTM</td>
<td>87.90</td>
<td>86.73</td>
<td>88.95</td>
</tr>
<tr>
<td>Gold Freq+ MM- LSTM</td>
<td>89.02</td>
<td>88.35</td>
<td>89.72</td>
</tr>
<tr>
<td>Human (majority$) ^{\dagger }$</td>
<td>92.89</td>
<td>91.24</td>
<td>94.52</td>
</tr>
</tbody>
</table>
Table 1: Dataset statistics for different foiled parts of speech. The superscript foil indicates the number of foiled captions.
Our intuition is that an image description/caption is essentially a result of the interaction between important objects in the image (this includes spatial relations, co-occurrences, etc.). Thus, representations explicitly encoding objectlevel information are better suited for the foiled caption classification task.
Table 2: Accuracy on Nouns dataset. $^{\dagger}$ are taken directly from Shekhar et al.(2017b). HieCoAtt is the state of the art reported in the paper.
### 3.3 Classifiers
Three types of classifiers are explored: (a) Multilayer Perceptron (MLP): For BOW-based text representations, a two 100-dimensional hidden layer MLP with ReLU activation function is used with cross-entropy loss, and is optimized with Adam (learning rate 0.001); (b) LSTM Classifer: For LSTM-based text representations, a uni-directional LSTM classifer is used with 100-dimensional word embeddings and 200- dimensional hidden representations. We train it using cross-entropy loss and optimize it using Adam (learning rate 0.001). Image representations are appended to the final hidden state of the LSTM; (c) Multimodal LSTM (MM-LSTM) Classifier: As above, except that we initialize the LSTM with the image representation instead of appending it to its output. This can also be seen as am image grounded LSTM based classifier.
bag of objects information are the best performing models across classifiers. We also note that the performance is better than human performance. We hypothesize the following reasons for this: (a) human responses were crowd-sourced, which could have resulted in some noisy annotations; (b) our gold object-based features closely resembles the information used for data-generation as described in Shekhar et al. (2017b) for the foil noun dataset. The models using Predicted bag of objects from a detector are very close to the performance of Gold. The performance of models using simple bag of words (BOW) sentence representations and an MLP is better than that of models that use LSTMs. Also, the accuracy of the bag of objects model with Frequency counts is higher than with the binary Mention vector, which only encodes the presence of objects. The Multimodal LSTM (MM-LSTM) has a slightly better performance than LSTM classifiers. In all cases, we observe that the performance is on par with human-level accuracy. Our overall accuracy is substantially higher than that reported in Shekhar et al. (2017b). Interestingly, our implementation of CNN+LSTM produced better results than their equivalent model (they reported 61.07% vs. our 87.45%). We investigate this further in Section 5. Performance on other parts of speech: For
### 4 $\mathbf{Experiments}$
Data: We use the dataset for nouns from Shekhar et al. (2017b)$^{1}$ and the datasets for other parts of speech from Shekhar et al.(2017a) $^{2}.$ Statistics about the dataset are given in Table 1. The evaluation metric is accuracy per class and the average (overall) accuracy over the two classes.
Performance on nouns: The results of our experiments with foiled nouns are summarized in Table 2. First, we note that the models that use Gold
other parts of speech, we fix the image representation to Gold Frequency, and compare results using the BOW-based MLP and MM-LSTM. We also compare the scores to the state of the art reported in Shekhar et al.(2017a). Note that this
$^{\mathrm{l}}$https://foilunitn.github.io/
$^{2}$The authors have kindly provided us the datasets.
3
Please cite this paper as: Pranava Madhyastha, Josiah Wang, and Lucia Specia (2018)
Defoiling Foiled Image Captions.
In Proceedings of 2018 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies (NAACL-HLT 2018).
|
[
[
0.1292724609375,
0.0748291015625,
0.19482421875,
0.07659912109375,
0.19482421875,
0.09027099609375,
0.1292724609375,
0.0885009765625
],
[
0.2032470703125,
0.07598876953125,
0.264404296875,
0.0736083984375,
0.26611328125,
0.0885009765625,
0.204833984375,
0.09027099609375
],
[
0.273681640625,
0.07659912109375,
0.32080078125,
0.0748291015625,
0.322509765625,
0.0872802734375,
0.274658203125,
0.08905029296875
],
[
0.334228515625,
0.07659912109375,
0.405517578125,
0.07659912109375,
0.405517578125,
0.0885009765625,
0.334228515625,
0.0885009765625
],
[
0.4130859375,
0.07659912109375,
0.478515625,
0.07659912109375,
0.478515625,
0.0885009765625,
0.4130859375,
0.0885009765625
],
[
0.59619140625,
0.07598876953125,
0.63818359375,
0.07598876953125,
0.63818359375,
0.0914306640625,
0.59619140625,
0.0914306640625
],
[
0.72119140625,
0.07781982421875,
0.7734375,
0.07781982421875,
0.7734375,
0.08905029296875,
0.72119140625,
0.08905029296875
],
[
0.78662109375,
0.07781982421875,
0.8212890625,
0.07781982421875,
0.8212890625,
0.09027099609375,
0.78662109375,
0.09027099609375
],
[
0.8369140625,
0.07781982421875,
0.869140625,
0.07781982421875,
0.869140625,
0.09027099609375,
0.8369140625,
0.09027099609375
],
[
0.210693359375,
0.0926513671875,
0.2626953125,
0.0926513671875,
0.2626953125,
0.10687255859375,
0.210693359375,
0.10687255859375
],
[
0.276123046875,
0.0926513671875,
0.322509765625,
0.0926513671875,
0.322509765625,
0.10687255859375,
0.276123046875,
0.10687255859375
],
[
0.83447265625,
0.09320068359375,
0.875,
0.0955810546875,
0.873046875,
0.10809326171875,
0.8330078125,
0.105712890625
],
[
0.344970703125,
0.09442138671875,
0.393798828125,
0.09442138671875,
0.393798828125,
0.105712890625,
0.344970703125,
0.105712890625
],
[
0.145263671875,
0.09320068359375,
0.1822509765625,
0.09320068359375,
0.1822509765625,
0.105712890625,
0.145263671875,
0.105712890625
],
[
0.4248046875,
0.09442138671875,
0.47265625,
0.09442138671875,
0.47265625,
0.105712890625,
0.4248046875,
0.105712890625
],
[
0.55322265625,
0.0955810546875,
0.67822265625,
0.0955810546875,
0.67822265625,
0.1092529296875,
0.55322265625,
0.1092529296875
],
[
0.72705078125,
0.0955810546875,
0.767578125,
0.0955810546875,
0.767578125,
0.10687255859375,
0.72705078125,
0.10687255859375
],
[
0.78515625,
0.0955810546875,
0.82373046875,
0.0955810546875,
0.82373046875,
0.10809326171875,
0.78515625,
0.10809326171875
],
[
0.14697265625,
0.105712890625,
0.181396484375,
0.105712890625,
0.181396484375,
0.1175537109375,
0.14697265625,
0.1175537109375
],
[
0.2149658203125,
0.105712890625,
0.255126953125,
0.105712890625,
0.255126953125,
0.1175537109375,
0.2149658203125,
0.1175537109375
],
[
0.279541015625,
0.1051025390625,
0.318115234375,
0.10687255859375,
0.316650390625,
0.119384765625,
0.27783203125,
0.1175537109375
],
[
0.345947265625,
0.105712890625,
0.391357421875,
0.105712890625,
0.391357421875,
0.1175537109375,
0.345947265625,
0.1175537109375
],
[
0.4248046875,
0.105712890625,
0.47021484375,
0.105712890625,
0.47021484375,
0.1175537109375,
0.4248046875,
0.1175537109375
],
[
0.58251953125,
0.1092529296875,
0.6513671875,
0.1092529296875,
0.6513671875,
0.12054443359375,
0.58251953125,
0.12054443359375
],
[
0.72705078125,
0.1092529296875,
0.765625,
0.1092529296875,
0.765625,
0.12054443359375,
0.72705078125,
0.12054443359375
],
[
0.78515625,
0.1092529296875,
0.82373046875,
0.1092529296875,
0.82373046875,
0.1217041015625,
0.78515625,
0.1217041015625
],
[
0.83447265625,
0.1092529296875,
0.873046875,
0.1092529296875,
0.873046875,
0.1217041015625,
0.83447265625,
0.1217041015625
],
[
0.133544921875,
0.11517333984375,
0.19482421875,
0.1175537109375,
0.193115234375,
0.1318359375,
0.1318359375,
0.1300048828125
],
[
0.2149658203125,
0.1181640625,
0.258544921875,
0.1181640625,
0.258544921875,
0.1300048828125,
0.2149658203125,
0.1300048828125
],
[
0.279541015625,
0.1181640625,
0.318115234375,
0.1181640625,
0.318115234375,
0.1300048828125,
0.279541015625,
0.1300048828125
],
[
0.34765625,
0.1181640625,
0.391357421875,
0.1181640625,
0.391357421875,
0.1300048828125,
0.34765625,
0.1300048828125
],
[
0.4248046875,
0.1181640625,
0.468505859375,
0.1181640625,
0.468505859375,
0.1300048828125,
0.4248046875,
0.1300048828125
],
[
0.5556640625,
0.1258544921875,
0.6767578125,
0.1258544921875,
0.6767578125,
0.1395263671875,
0.5556640625,
0.1395263671875
],
[
0.72900390625,
0.1270751953125,
0.767578125,
0.1270751953125,
0.767578125,
0.1395263671875,
0.72900390625,
0.1395263671875
],
[
0.78515625,
0.1270751953125,
0.82373046875,
0.1270751953125,
0.82373046875,
0.1395263671875,
0.78515625,
0.1395263671875
],
[
0.83447265625,
0.1270751953125,
0.8759765625,
0.1270751953125,
0.8759765625,
0.14013671875,
0.83447265625,
0.14013671875
],
[
0.345947265625,
0.12890625,
0.39208984375,
0.12890625,
0.39208984375,
0.1436767578125,
0.345947265625,
0.1436767578125
],
[
0.14111328125,
0.1312255859375,
0.18896484375,
0.1312255859375,
0.18896484375,
0.142578125,
0.14111328125,
0.142578125
],
[
0.21826171875,
0.1312255859375,
0.25439453125,
0.1312255859375,
0.25439453125,
0.142578125,
0.21826171875,
0.142578125
],
[
0.28369140625,
0.1300048828125,
0.312255859375,
0.1300048828125,
0.312255859375,
0.142578125,
0.28369140625,
0.142578125
],
[
0.4248046875,
0.1312255859375,
0.468505859375,
0.1312255859375,
0.468505859375,
0.142578125,
0.4248046875,
0.142578125
],
[
0.52294921875,
0.1395263671875,
0.70947265625,
0.1395263671875,
0.70947265625,
0.152587890625,
0.52294921875,
0.152587890625
],
[
0.72705078125,
0.14013671875,
0.767578125,
0.14013671875,
0.767578125,
0.1519775390625,
0.72705078125,
0.1519775390625
],
[
0.7841796875,
0.1395263671875,
0.8251953125,
0.1395263671875,
0.8251953125,
0.152587890625,
0.7841796875,
0.152587890625
],
[
0.83447265625,
0.14013671875,
0.873046875,
0.14013671875,
0.873046875,
0.1519775390625,
0.83447265625,
0.1519775390625
],
[
0.4248046875,
0.141357421875,
0.470947265625,
0.141357421875,
0.470947265625,
0.1561279296875,
0.4248046875,
0.1561279296875
],
[
0.2166748046875,
0.142578125,
0.25439453125,
0.142578125,
0.25439453125,
0.155029296875,
0.2166748046875,
0.155029296875
],
[
0.280517578125,
0.142578125,
0.316650390625,
0.142578125,
0.316650390625,
0.155029296875,
0.280517578125,
0.155029296875
],
[
0.345947265625,
0.141357421875,
0.39208984375,
0.141357421875,
0.39208984375,
0.1561279296875,
0.345947265625,
0.1561279296875
],
[
0.72705078125,
0.1519775390625,
0.767578125,
0.15380859375,
0.765625,
0.166259765625,
0.7255859375,
0.1644287109375
],
[
0.1275634765625,
0.1436767578125,
0.197265625,
0.1436767578125,
0.197265625,
0.155029296875,
0.1275634765625,
0.155029296875
],
[
0.78515625,
0.152587890625,
0.82275390625,
0.152587890625,
0.82275390625,
0.1650390625,
0.78515625,
0.1650390625
],
[
0.8330078125,
0.152587890625,
0.875,
0.152587890625,
0.875,
0.166259765625,
0.8330078125,
0.166259765625
],
[
0.8330078125,
0.1644287109375,
0.873046875,
0.166259765625,
0.87158203125,
0.1787109375,
0.8310546875,
0.177001953125
],
[
0.53662109375,
0.15380859375,
0.6962890625,
0.152587890625,
0.6962890625,
0.1644287109375,
0.53662109375,
0.1650390625
],
[
0.52880859375,
0.1650390625,
0.70361328125,
0.1644287109375,
0.70361328125,
0.1787109375,
0.52880859375,
0.179931640625
],
[
0.72705078125,
0.166259765625,
0.76904296875,
0.166259765625,
0.76904296875,
0.179931640625,
0.72705078125,
0.179931640625
],
[
0.78515625,
0.166259765625,
0.82373046875,
0.166259765625,
0.82373046875,
0.1776123046875,
0.78515625,
0.1776123046875
],
[
0.54248046875,
0.179931640625,
0.68994140625,
0.1787109375,
0.68994140625,
0.190673828125,
0.54248046875,
0.191162109375
],
[
0.118408203125,
0.169189453125,
0.489501953125,
0.169189453125,
0.489501953125,
0.182861328125,
0.118408203125,
0.182861328125
],
[
0.72705078125,
0.179931640625,
0.76904296875,
0.179931640625,
0.76904296875,
0.193603515625,
0.72705078125,
0.193603515625
],
[
0.78515625,
0.179931640625,
0.82373046875,
0.179931640625,
0.82373046875,
0.191162109375,
0.78515625,
0.191162109375
],
[
0.8330078125,
0.179931640625,
0.875,
0.179931640625,
0.875,
0.193603515625,
0.8330078125,
0.193603515625
],
[
0.82861328125,
0.1953125,
0.8759765625,
0.1953125,
0.8759765625,
0.210205078125,
0.82861328125,
0.210205078125
],
[
0.118408203125,
0.18408203125,
0.490234375,
0.18408203125,
0.490234375,
0.19775390625,
0.118408203125,
0.19775390625
],
[
0.57275390625,
0.19775390625,
0.6630859375,
0.19775390625,
0.6630859375,
0.208984375,
0.57275390625,
0.208984375
],
[
0.72900390625,
0.19775390625,
0.767578125,
0.19775390625,
0.767578125,
0.208984375,
0.72900390625,
0.208984375
],
[
0.78515625,
0.196533203125,
0.82373046875,
0.196533203125,
0.82373046875,
0.208984375,
0.78515625,
0.208984375
],
[
0.54833984375,
0.2078857421875,
0.68408203125,
0.208984375,
0.68408203125,
0.223876953125,
0.54833984375,
0.22265625
],
[
0.118408203125,
0.1995849609375,
0.22412109375,
0.1995849609375,
0.22412109375,
0.21142578125,
0.118408203125,
0.21142578125
],
[
0.72900390625,
0.210205078125,
0.767578125,
0.210205078125,
0.767578125,
0.22265625,
0.72900390625,
0.22265625
],
[
0.78515625,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.22265625,
0.78515625,
0.22265625
],
[
0.83447265625,
0.210205078125,
0.87158203125,
0.210205078125,
0.87158203125,
0.22265625,
0.83447265625,
0.22265625
],
[
0.5576171875,
0.22265625,
0.67822265625,
0.22265625,
0.67822265625,
0.236328125,
0.5576171875,
0.236328125
],
[
0.72900390625,
0.22265625,
0.765625,
0.22265625,
0.765625,
0.235107421875,
0.72900390625,
0.235107421875
],
[
0.78515625,
0.223876953125,
0.82373046875,
0.223876953125,
0.82373046875,
0.235107421875,
0.78515625,
0.235107421875
],
[
0.83447265625,
0.223876953125,
0.873046875,
0.223876953125,
0.873046875,
0.235107421875,
0.83447265625,
0.235107421875
],
[
0.53564453125,
0.236328125,
0.69873046875,
0.236328125,
0.69873046875,
0.2493896484375,
0.53564453125,
0.2493896484375
],
[
0.72900390625,
0.236328125,
0.767578125,
0.236328125,
0.767578125,
0.248779296875,
0.72900390625,
0.248779296875
],
[
0.78515625,
0.236328125,
0.82373046875,
0.236328125,
0.82373046875,
0.248779296875,
0.78515625,
0.248779296875
],
[
0.83447265625,
0.2340087890625,
0.873046875,
0.236328125,
0.87158203125,
0.248779296875,
0.8330078125,
0.2464599609375
],
[
0.54150390625,
0.248779296875,
0.69287109375,
0.248779296875,
0.69287109375,
0.262451171875,
0.54150390625,
0.262451171875
],
[
0.1351318359375,
0.2381591796875,
0.48779296875,
0.2392578125,
0.48779296875,
0.254150390625,
0.1351318359375,
0.2529296875
],
[
0.72900390625,
0.2493896484375,
0.767578125,
0.2493896484375,
0.767578125,
0.26123046875,
0.72900390625,
0.26123046875
],
[
0.78515625,
0.2493896484375,
0.82373046875,
0.2493896484375,
0.82373046875,
0.262451171875,
0.78515625,
0.262451171875
],
[
0.8310546875,
0.248779296875,
0.8759765625,
0.248779296875,
0.8759765625,
0.26318359375,
0.8310546875,
0.26318359375
],
[
0.118408203125,
0.255859375,
0.48779296875,
0.255859375,
0.48779296875,
0.270751953125,
0.118408203125,
0.270751953125
],
[
0.724609375,
0.2666015625,
0.76904296875,
0.2666015625,
0.76904296875,
0.28076171875,
0.724609375,
0.28076171875
],
[
0.78271484375,
0.2666015625,
0.8759765625,
0.2666015625,
0.8759765625,
0.28076171875,
0.78271484375,
0.28076171875
],
[
0.118408203125,
0.2724609375,
0.489501953125,
0.2724609375,
0.489501953125,
0.2861328125,
0.118408203125,
0.2861328125
],
[
0.5576171875,
0.267333984375,
0.67529296875,
0.267333984375,
0.67529296875,
0.28076171875,
0.5576171875,
0.28076171875
],
[
0.1175537109375,
0.287353515625,
0.489501953125,
0.2880859375,
0.489501953125,
0.302734375,
0.1175537109375,
0.3017578125
],
[
0.51123046875,
0.29345703125,
0.78076171875,
0.29443359375,
0.78076171875,
0.308837890625,
0.51123046875,
0.30810546875
],
[
0.771484375,
0.292236328125,
0.88330078125,
0.295654296875,
0.88330078125,
0.31005859375,
0.771484375,
0.30712890625
],
[
0.1175537109375,
0.302734375,
0.489501953125,
0.303955078125,
0.489501953125,
0.318359375,
0.1175537109375,
0.317626953125
],
[
0.51318359375,
0.31005859375,
0.884765625,
0.31005859375,
0.884765625,
0.32373046875,
0.51318359375,
0.32373046875
],
[
0.118408203125,
0.320556640625,
0.489501953125,
0.320556640625,
0.489501953125,
0.334228515625,
0.118408203125,
0.334228515625
],
[
0.51318359375,
0.32373046875,
0.759765625,
0.32373046875,
0.759765625,
0.33740234375,
0.51318359375,
0.33740234375
],
[
0.118408203125,
0.33740234375,
0.31494140625,
0.33740234375,
0.31494140625,
0.350830078125,
0.118408203125,
0.350830078125
],
[
0.118408203125,
0.364013671875,
0.2451171875,
0.364013671875,
0.2451171875,
0.37890625,
0.118408203125,
0.37890625
],
[
0.51123046875,
0.366455078125,
0.880859375,
0.366455078125,
0.880859375,
0.380126953125,
0.51123046875,
0.380126953125
],
[
0.118408203125,
0.38720703125,
0.489501953125,
0.38720703125,
0.489501953125,
0.40087890625,
0.118408203125,
0.40087890625
],
[
0.51123046875,
0.3818359375,
0.8818359375,
0.3818359375,
0.8818359375,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.118408203125,
0.402587890625,
0.490234375,
0.402099609375,
0.490234375,
0.416259765625,
0.118408203125,
0.41748046875
],
[
0.51318359375,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.118408203125,
0.419189453125,
0.490234375,
0.419189453125,
0.490234375,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.5146484375,
0.4150390625,
0.884765625,
0.4150390625,
0.884765625,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.118408203125,
0.435791015625,
0.490234375,
0.435791015625,
0.490234375,
0.449462890625,
0.118408203125,
0.449462890625
],
[
0.51318359375,
0.431640625,
0.88330078125,
0.431640625,
0.88330078125,
0.4453125,
0.51318359375,
0.4453125
],
[
0.1175537109375,
0.45068359375,
0.489501953125,
0.451904296875,
0.489501953125,
0.466064453125,
0.1175537109375,
0.465576171875
],
[
0.51318359375,
0.44775390625,
0.88330078125,
0.44775390625,
0.88330078125,
0.46142578125,
0.51318359375,
0.46142578125
],
[
0.1175537109375,
0.46728515625,
0.489501953125,
0.466064453125,
0.489501953125,
0.48095703125,
0.1175537109375,
0.482177734375
],
[
0.51123046875,
0.46435546875,
0.884765625,
0.46435546875,
0.884765625,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.51123046875,
0.47802734375,
0.88330078125,
0.478515625,
0.88330078125,
0.493408203125,
0.51123046875,
0.4921875
],
[
0.12005615234375,
0.48388671875,
0.48779296875,
0.48388671875,
0.48779296875,
0.49755859375,
0.12005615234375,
0.49755859375
],
[
0.118408203125,
0.49951171875,
0.489501953125,
0.49951171875,
0.489501953125,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.51318359375,
0.495361328125,
0.884765625,
0.495361328125,
0.884765625,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.118408203125,
0.515625,
0.48779296875,
0.51416015625,
0.48779296875,
0.529296875,
0.118408203125,
0.52978515625
],
[
0.51318359375,
0.5126953125,
0.88330078125,
0.5126953125,
0.88330078125,
0.525390625,
0.51318359375,
0.525390625
],
[
0.39208984375,
0.5322265625,
0.48779296875,
0.5322265625,
0.48779296875,
0.54345703125,
0.39208984375,
0.54345703125
],
[
0.51318359375,
0.529296875,
0.88330078125,
0.529296875,
0.88330078125,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.5146484375,
0.5458984375,
0.880859375,
0.5458984375,
0.880859375,
0.5556640625,
0.5146484375,
0.5556640625
],
[
0.12176513671875,
0.53369140625,
0.397216796875,
0.53369140625,
0.397216796875,
0.54443359375,
0.12176513671875,
0.54443359375
],
[
0.1175537109375,
0.54736328125,
0.48779296875,
0.54736328125,
0.48779296875,
0.56103515625,
0.1175537109375,
0.56103515625
],
[
0.12005615234375,
0.56396484375,
0.349365234375,
0.56396484375,
0.349365234375,
0.57763671875,
0.12005615234375,
0.57763671875
],
[
0.51123046875,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.51123046875,
0.572265625
],
[
0.341796875,
0.56640625,
0.48779296875,
0.56640625,
0.48779296875,
0.57666015625,
0.341796875,
0.57666015625
],
[
0.1175537109375,
0.580078125,
0.490234375,
0.580078125,
0.490234375,
0.59326171875,
0.1175537109375,
0.59326171875
],
[
0.51123046875,
0.576171875,
0.8818359375,
0.57470703125,
0.8818359375,
0.5888671875,
0.51123046875,
0.59033203125
],
[
0.118408203125,
0.5966796875,
0.489501953125,
0.5966796875,
0.489501953125,
0.60986328125,
0.118408203125,
0.60986328125
],
[
0.51318359375,
0.5927734375,
0.884765625,
0.5927734375,
0.884765625,
0.60546875,
0.51318359375,
0.60546875
],
[
0.51123046875,
0.60693359375,
0.884765625,
0.60791015625,
0.884765625,
0.62255859375,
0.51123046875,
0.62158203125
],
[
0.12005615234375,
0.61328125,
0.489501953125,
0.61328125,
0.489501953125,
0.626953125,
0.12005615234375,
0.626953125
],
[
0.1175537109375,
0.62744140625,
0.490234375,
0.62890625,
0.490234375,
0.6435546875,
0.1175537109375,
0.642578125
],
[
0.51318359375,
0.62451171875,
0.88330078125,
0.62451171875,
0.88330078125,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.118408203125,
0.64453125,
0.490234375,
0.64453125,
0.490234375,
0.6591796875,
0.118408203125,
0.6591796875
],
[
0.51318359375,
0.64013671875,
0.8818359375,
0.64013671875,
0.8818359375,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.1175537109375,
0.6611328125,
0.4609375,
0.66015625,
0.4609375,
0.6748046875,
0.1175537109375,
0.67578125
],
[
0.51318359375,
0.65625,
0.88330078125,
0.65625,
0.88330078125,
0.66943359375,
0.51318359375,
0.66943359375
],
[
0.51318359375,
0.6728515625,
0.88330078125,
0.6728515625,
0.88330078125,
0.68701171875,
0.51318359375,
0.68701171875
],
[
0.145263671875,
0.69140625,
0.2626953125,
0.69140625,
0.2626953125,
0.705078125,
0.145263671875,
0.705078125
],
[
0.51318359375,
0.689453125,
0.884765625,
0.689453125,
0.884765625,
0.703125,
0.51318359375,
0.703125
],
[
0.1427001953125,
0.69482421875,
0.1485595703125,
0.69482421875,
0.1485595703125,
0.70068359375,
0.1427001953125,
0.70068359375
],
[
0.1234130859375,
0.6923828125,
0.145263671875,
0.6923828125,
0.145263671875,
0.703125,
0.1234130859375,
0.703125
],
[
0.51318359375,
0.705078125,
0.884765625,
0.705078125,
0.884765625,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.12005615234375,
0.71875,
0.181396484375,
0.71875,
0.181396484375,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.172119140625,
0.71630859375,
0.490234375,
0.71875,
0.490234375,
0.732421875,
0.172119140625,
0.72998046875
],
[
0.51318359375,
0.7216796875,
0.884765625,
0.7216796875,
0.884765625,
0.7353515625,
0.51318359375,
0.7353515625
],
[
0.12005615234375,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.51318359375,
0.73681640625,
0.884765625,
0.73681640625,
0.884765625,
0.75048828125,
0.51318359375,
0.75048828125
],
[
0.118408203125,
0.75,
0.484375,
0.7490234375,
0.484375,
0.76318359375,
0.118408203125,
0.76416015625
],
[
0.51318359375,
0.7529296875,
0.884765625,
0.7529296875,
0.884765625,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.12005615234375,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.76953125,
0.88330078125,
0.76953125,
0.88330078125,
0.783203125,
0.51318359375,
0.783203125
],
[
0.12005615234375,
0.78369140625,
0.489501953125,
0.78369140625,
0.489501953125,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.51318359375,
0.78515625,
0.8759765625,
0.78515625,
0.8759765625,
0.798828125,
0.51318359375,
0.798828125
],
[
0.118408203125,
0.798828125,
0.47021484375,
0.798828125,
0.47021484375,
0.8125,
0.118408203125,
0.8125
],
[
0.5146484375,
0.81396484375,
0.884765625,
0.81396484375,
0.884765625,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.12005615234375,
0.826171875,
0.489501953125,
0.826171875,
0.489501953125,
0.8388671875,
0.12005615234375,
0.8388671875
],
[
0.51318359375,
0.83203125,
0.8818359375,
0.83203125,
0.8818359375,
0.845703125,
0.51318359375,
0.845703125
],
[
0.12005615234375,
0.84326171875,
0.489501953125,
0.84326171875,
0.489501953125,
0.85400390625,
0.12005615234375,
0.85400390625
],
[
0.51318359375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1376953125,
0.88134765625,
0.4013671875,
0.8818359375,
0.4013671875,
0.896484375,
0.1376953125,
0.8955078125
],
[
0.51318359375,
0.8798828125,
0.88330078125,
0.8798828125,
0.88330078125,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.139404296875,
0.89453125,
0.44580078125,
0.8955078125,
0.44580078125,
0.9091796875,
0.139404296875,
0.90771484375
],
[
0.51123046875,
0.89453125,
0.884765625,
0.8935546875,
0.884765625,
0.90771484375,
0.51123046875,
0.9091796875
],
[
0.214111328125,
0.93115234375,
0.78857421875,
0.93115234375,
0.78857421875,
0.94482421875,
0.214111328125,
0.94482421875
],
[
0.87158203125,
0.93115234375,
0.884765625,
0.93115234375,
0.884765625,
0.9423828125,
0.87158203125,
0.9423828125
],
[
0.382080078125,
0.94482421875,
0.61865234375,
0.94482421875,
0.61865234375,
0.9580078125,
0.382080078125,
0.9580078125
],
[
0.20068359375,
0.95703125,
0.8017578125,
0.95703125,
0.8017578125,
0.97021484375,
0.20068359375,
0.97021484375
],
[
0.22412109375,
0.9697265625,
0.7734375,
0.96875,
0.7734375,
0.98291015625,
0.22412109375,
0.98388671875
]
] |
[
"Pos Type",
"Trainfoil",
"Testfoil",
"Total Train",
"Total Test",
"Feats",
"Overall",
"Real",
"Foil",
"153,229",
"75,278",
"25.04",
"306,458",
"Noun",
"150,556",
"Blind (LSTM only)t",
"55.62",
"86.20",
"Verb",
"6,314",
"2,788",
"60,262",
"33,616",
"HieCoAttt",
"64.14",
"91.89",
"36.38",
"Adjective",
"15,640",
"9.009",
"73.057",
"42,163",
"CNN + BOW MLP",
"88.42",
"86.89",
"89.97",
"53,381",
"Adverb",
"1,011",
"451",
"30,738",
"Predict Mention + BOW MLP",
"94.94",
"95.68",
"94.23",
"46,018",
"8,733",
"5,551",
"77,002",
"95.14",
"Preposition",
"95.82",
"94.48",
"95.36",
"Predict Freq+BOW MLP",
"Gold Mention + BOW MLP",
"95.83",
"96.30",
"GoldFreg+BOWMLP",
"Table 1: Dataset statistics for different foiled parts of",
"96.45",
"96.04",
"96.85",
"88.14",
"speech. The superscript foil indicates the number of",
"CNN +LSTM",
"87.45",
"86.78",
"Predict Freq + LSTM",
"foiled captions.",
"85.99",
"85.17",
"86.81",
"Gold Freq + LSTM",
"87.38",
"86.62",
"88.18",
"Predict Freq + MM-LSTM",
"87.90",
"86.73",
"88.95",
"Gold Freq + MM-LSTM",
"Our intuition is that an image descrip-",
"89.02",
"88.35",
"89.72",
"tion/caption is essentially a result of the interac-",
"92.89",
"91.2494.52",
"tion between important objects in the image (this",
"Human (majority)t",
"includes spatial relations, co-occurrences, etc.).",
"Table 2:Accuracy on Nouns dataset.",
"f are taken di-",
"Thus, representations explicitly encoding object-",
"rectly from Shekhar et al. (2017b). HieCoAtt is the",
"level information are better suited for the foiled",
"state of the art reported in the paper.",
"caption classification task.",
"3.3Classifiers",
"bag of objects information are the best performing",
"Three types of classifiers are explored: (a) Mul-",
"models across classifiers. We also note that the",
"tilayer Perceptron (MLP): For BOW-based text",
"performance is better than human performance",
"representations, a two 100-dimensional hidden",
"We hypothesize the following reasons for this:",
"layer MLP with ReLU activation function is",
"(a) human responses were crowd-sourced, which",
"used with cross-entropy loss, and is optimized",
"could have resulted in some noisy annotations; (b)",
"with Adam (learning rate 0.001); (b) LSTM",
"our gold object-based features closely resembles",
"the information used for data-generation as de-",
"Classifier: For LSTM-based text representa-",
"tions, a uni-directional LSTM classifier is used",
"scribed in Shekhar et al. (2017b) for the foil noun",
"with 100-dimensional word embeddings and 200-",
"dataset. The models using Predicted bag of ob-",
"We train it",
"jects from a detector are very close to the perfor-",
"manceof Gold.The performanceofmodels us-",
"dimensional hidden representations.",
"using cross-entropy loss and optimize it using",
"Adam (learning rate 0.001).",
"ing simple bag of words (BOW) sentence repre-",
"Image representa-",
"tions are appended to the final hidden state of",
"sentations and an MLP is better than that of mod-",
"the LSTM; (c) Multimodal LSTM (MM-LSTM)",
"els that use LSTMs. Also, the accuracy of the",
"bag of objects model with Frequency counts is",
"Classifier: As above, except that we initialize the",
"LSTM with the image representation instead of",
"higher than with the binary Mention vector, which",
"appending it to its output. This can also be seen",
"only encodes the presence of objects. The Multi-",
"as am image grounded LSTM based classifier.",
"modal LSTM (MM-LSTM) has a slightly better",
"performance than LSTM classifiers. In all cases,",
"Experiments",
"we observe that the performance is on par with",
"",
"4",
"human-level accuracy. Our overall accuracy is",
"Data:",
"We use thedataset for nouns from",
"substantially higher than that reported in Shekhar",
"Shekhar et al. (2017b) and the datasets for other",
"et al. (2017b).Interestingly, our implementation",
"parts of speech from Shekhar et al. (2017a) 2",
"of CNN+LSTM produced better results than their",
"Statistics about the dataset are given in Table 1.",
"equivalent model (they reported 61.07% vs. our",
"Theevaluationmetricis accuracyperclassandthe",
"87.45%). We investigate this further in Section 5.",
"average (overall) accuracy over the two classes.",
"Performance on other parts of speech: For",
"Performance on nouns: The results of our ex-",
"other parts of speech, we fix the image representa-",
"perimentswithfoiled nouns are summarizedinTa-",
"tion to Gold Frequency, and compare results us-",
"ble 2. First, we note that the models that use Gold",
"ing the BOW-based MLP and MM-LSTM. We",
"'https://foilunitn.github.io/",
"also compare the scores to the state of the art re-",
"2The authors have kindly provided us the datasets.",
"ported in Shekhar et al. (2017a). Note that this",
"Please cite this paper as: Pranava Madhyastha, Josiah Wang, and Lucia Specia (2018)",
"3",
"Defoiling Foiled Image Captions.",
"In Proceedings of 2018 Conference of the North American Chapter of the Association for",
"Computational Linguistics: Human Language Technologies (NAACL-HLT 2018)."
] |
{
"arxiv_id": "1805.06549",
"language": "en",
"page_idx": 2
}
|
||
2312.00296-en-0
|
2312.00296-en-0.jpg
|
Towards Aligned Canonical Correlation Analysis: Preliminary
Formulation and Proof-of-Concept Results
Biqian Cheng $\bar{\mathrm{UC}}$ Riverside
Riverside, CA, USA bchen158@ucr.edu
Evangelos E. Papalexakis
UC Riverside
Riverside, CA, USA epapalex@cs.ucr.edu
Jia Chen UC Riverside
Riverside, CA, USA
$\operatorname{jiac@ucr.edu}$
## $\mathbf{ABSTRACT}$
Canonical Correlation Analysis (CCA) has been a tried and tested analytical model which seeks to jointly embed two or more views of the data in a maximally correlated latent space. In addition to a powerful data mining tool, CCA has many implications for cutting-edge self-supervised representation learning approaches, as one can cast a number of recent approaches as variants of that model. In this work we explore what happens when a fundamental assumption of that model breaks: what if the alignment across different views of the data is unknown? Typically, we would first attempt to align the different views, and subsequently apply CCA or any other model in order to embed the data views. Can we do better if we align and embed at the same time? In this work, we seek to jointly solve the alignment and embedding by formulating and solving both problems under the same umbrella of Aligned Canonical Correlation Analysis (ACCA). We present a preliminary formulation and alternating optimization algorithm and proof-of-concept results.
many decades and has been around for quite a long time, it is still extremely relevant, not only as a standalone data mining tool, but also due to the fact that cutting-edge self-supervised representation learning techniques, such as Barlow Twins [Balestriero et al. 2023; Bielak et al. 2022; Zbontar et al. 2021; Zhang et al. 2021], can be essentially seen as variations of CCA, where the goal is to embed two views of the data (the original view and the augmentation) in a latent space where correlation is maximized.
Traditionally, in CCA-style analysis, we assume that entities across views have one-to-one correspondence across the two or more views of the data, and there is a wealth of algorithms that study different formulations for solving the problem of projecting those views in that desired maximally correlated space, both linearly and non-linearly [Andrew et al. 2013]. What if, however, this oneto-one correspondence is unknown? In this case, we are faced with two problems: (1) entity alignment and (2) CCA embedding. Motivated by recent results [Wu et al. 2022] in the related problem of misaligned joint tensor factorization, it turns out that formulating and solving the alignment and embedding problems jointly yields better results than solving each problem separately in multiple steps, as it appears that the two sub-problems work synergistically to produce better quality alignment and embeddings. In this work, we explore this type of formulation for CCA and we propose a new formulation, the Aligned Canonical Correlation Analysis (ACCA), where we seek to jointly compute the alignment and the embedding $\operatorname{space.}$
The closest formulation to our proposed model is found in [Sahbi 2018] where the author is considering linear transformation of the two views in CCA, however, is not seeking to recover the precise alignment matrix as our formulation does. In our on-going work we will consider scenarios where we can fairly compare the two formulations and understand pros and cons for either one.
The list of contributions in this preliminary work are:
# $\mathbf{KEYWORDS}$
Canonical Correlation Analysis, Alignment, Matching, Data Inte-
## gration
ACM Reference Format:
Biqian Cheng, Evangelos E. Papalexakis, and Jia Chen. 2023. Towards Aligned Canonical Correlation Analysis: Preliminary Formulation and Proofof-Concept Results. In Proceedings of KDD Conference (KDD'23). ACM, New
$\begin{array}{cc}{\mathrm{Ni}}&{\mathrm{of-Concept~Results.~In~Proceeedings~ofKDD~Conference~(KDD^{\prime}23).}}\\{\mathrm{N}}&{\mathrm{York,~NY,~USA,~4~pages.~https://doi.org/XXXXXXXX.XXXXXX}}\\\\{\mathrm{Cr}}&{\mathrm{1NTRODUCTION}}\end{array}$
# 1 INTRODUCTION
Canonical Correlation Analysis (CCA) [Harold 1936; Kettenring 1971] is a classical model which, given two different views of the same set of entities, e.g., two different bipartite graphs of (user, product) and (user, video) interactions or different feature representations for those entities in general, seeks to project those entities (users) in a low-dimensional space where the different projected views are maximally correlated. Essentially, CCA can jointly embed heterogeneous datasets in a common low-dimensional space, as it can be extended to more than two views [Chen et al.2019, 2022]. Even though CCA has been in and out of the spotlight for
$\bullet$ Novel Formulation: We propose the Aligned Canonical Correlation Analysis (ACCA) model, which seeks to jointly identify the best entity alingment and latent embedding for the dataset views.
$\bullet$ Proof of Concept: We derive an Alternating Optimization
algorithm and show preliminary results for solving the prob-
lem, demonstrating the feasibility of our effort.
# 2 васкєкоимр
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
$KDD^{\prime }23, August$ 07, 2023, Long $Beach$, $CA$, $USA$ $\textcircled{\circ}$ 2023 Association for Computing Machinery. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00 https://doi.org/XXXXXX.XXXXXX
Canonical correlation analysis (CCA) is a powerful tool to learn the shared latent components of two datasets by projecting them to the same space and enforcing the similarity of the projected $\dot{\text{data. Given two centered and aligned datasets X}}\in\mathbb{R}^{\hat{D}_{x}\times\tilde{N}}$and
|
[
[
0.10455322265625,
0.1029052734375,
0.89208984375,
0.10418701171875,
0.89208984375,
0.125,
0.10455322265625,
0.12371826171875
],
[
0.2279052734375,
0.1300048828125,
0.771484375,
0.1300048828125,
0.771484375,
0.147705078125,
0.2279052734375,
0.147705078125
],
[
0.1732177734375,
0.160400390625,
0.289306640625,
0.1622314453125,
0.289306640625,
0.18115234375,
0.1732177734375,
0.1787109375
],
[
0.39794921875,
0.1622314453125,
0.60302734375,
0.1622314453125,
0.60302734375,
0.179931640625,
0.39794921875,
0.179931640625
],
[
0.73046875,
0.1622314453125,
0.80615234375,
0.1622314453125,
0.80615234375,
0.179931640625,
0.73046875,
0.179931640625
],
[
0.1871337890625,
0.179931640625,
0.281005859375,
0.179931640625,
0.281005859375,
0.1944580078125,
0.1871337890625,
0.1944580078125
],
[
0.455078125,
0.179931640625,
0.5458984375,
0.179931640625,
0.5458984375,
0.1944580078125,
0.455078125,
0.1944580078125
],
[
0.72216796875,
0.179931640625,
0.81396484375,
0.179931640625,
0.81396484375,
0.1944580078125,
0.72216796875,
0.1944580078125
],
[
0.1666259765625,
0.1944580078125,
0.29736328125,
0.1944580078125,
0.29736328125,
0.208984375,
0.1666259765625,
0.208984375
],
[
0.43310546875,
0.1932373046875,
0.564453125,
0.1932373046875,
0.564453125,
0.2083740234375,
0.43310546875,
0.2083740234375
],
[
0.70361328125,
0.1944580078125,
0.833984375,
0.1944580078125,
0.833984375,
0.208984375,
0.70361328125,
0.208984375
],
[
0.1666259765625,
0.210205078125,
0.29736328125,
0.210205078125,
0.29736328125,
0.2235107421875,
0.1666259765625,
0.2235107421875
],
[
0.428955078125,
0.210205078125,
0.57177734375,
0.208984375,
0.57177734375,
0.2235107421875,
0.428955078125,
0.2247314453125
],
[
0.72216796875,
0.210205078125,
0.81396484375,
0.210205078125,
0.81396484375,
0.2247314453125,
0.72216796875,
0.2247314453125
],
[
0.08660888671875,
0.2342529296875,
0.1846923828125,
0.2342529296875,
0.1846923828125,
0.248779296875,
0.08660888671875,
0.248779296875
],
[
0.5185546875,
0.23486328125,
0.9140625,
0.23486328125,
0.9140625,
0.248779296875,
0.5185546875,
0.248779296875
],
[
0.0882568359375,
0.255126953125,
0.4794921875,
0.255126953125,
0.4794921875,
0.26513671875,
0.0882568359375,
0.26513671875
],
[
0.5185546875,
0.251953125,
0.91259765625,
0.251953125,
0.91259765625,
0.261962890625,
0.5185546875,
0.261962890625
],
[
0.0236968994140625,
0.267578125,
0.05963134765625,
0.267578125,
0.05963134765625,
0.326904296875,
0.0236968994140625,
0.326904296875
],
[
0.08660888671875,
0.267578125,
0.483642578125,
0.267578125,
0.483642578125,
0.281005859375,
0.08660888671875,
0.281005859375
],
[
0.51708984375,
0.264404296875,
0.9140625,
0.264404296875,
0.9140625,
0.27783203125,
0.51708984375,
0.27783203125
],
[
0.08416748046875,
0.27978515625,
0.483642578125,
0.281005859375,
0.483642578125,
0.294189453125,
0.08416748046875,
0.29345703125
],
[
0.5185546875,
0.279052734375,
0.91259765625,
0.279052734375,
0.91259765625,
0.2890625,
0.5185546875,
0.2890625
],
[
0.08660888671875,
0.296630859375,
0.48046875,
0.296630859375,
0.48046875,
0.306884765625,
0.08660888671875,
0.306884765625
],
[
0.51708984375,
0.291748046875,
0.9140625,
0.291748046875,
0.9140625,
0.304931640625,
0.51708984375,
0.304931640625
],
[
0.0849609375,
0.309326171875,
0.48193359375,
0.309326171875,
0.48193359375,
0.322509765625,
0.0849609375,
0.322509765625
],
[
0.515625,
0.3037109375,
0.9140625,
0.304931640625,
0.9140625,
0.3193359375,
0.515625,
0.318115234375
],
[
0.0310516357421875,
0.32080078125,
0.05963134765625,
0.322509765625,
0.053924560546875,
0.373779296875,
0.02532958984375,
0.371826171875
],
[
0.08660888671875,
0.3212890625,
0.48046875,
0.3212890625,
0.48046875,
0.335205078125,
0.08660888671875,
0.335205078125
],
[
0.51708984375,
0.3193359375,
0.9140625,
0.3193359375,
0.9140625,
0.333251953125,
0.51708984375,
0.333251953125
],
[
0.08660888671875,
0.33837890625,
0.48193359375,
0.33837890625,
0.48193359375,
0.348388671875,
0.08660888671875,
0.348388671875
],
[
0.515625,
0.333251953125,
0.79736328125,
0.333251953125,
0.79736328125,
0.3466796875,
0.515625,
0.3466796875
],
[
0.08660888671875,
0.3515625,
0.48193359375,
0.3515625,
0.48193359375,
0.3623046875,
0.08660888671875,
0.3623046875
],
[
0.53173828125,
0.3466796875,
0.9140625,
0.3466796875,
0.9140625,
0.35986328125,
0.53173828125,
0.35986328125
],
[
0.0849609375,
0.3642578125,
0.48046875,
0.3642578125,
0.48046875,
0.37744140625,
0.0849609375,
0.37744140625
],
[
0.5185546875,
0.3623046875,
0.91259765625,
0.3623046875,
0.91259765625,
0.37255859375,
0.5185546875,
0.37255859375
],
[
0.02777099609375,
0.37255859375,
0.0555419921875,
0.37255859375,
0.0555419921875,
0.389404296875,
0.02777099609375,
0.389404296875
],
[
0.08660888671875,
0.376953125,
0.48046875,
0.37744140625,
0.48046875,
0.391357421875,
0.08660888671875,
0.39013671875
],
[
0.5185546875,
0.375732421875,
0.91259765625,
0.375732421875,
0.91259765625,
0.386474609375,
0.5185546875,
0.386474609375
],
[
0.51708984375,
0.386962890625,
0.9140625,
0.386962890625,
0.9140625,
0.40087890625,
0.51708984375,
0.40087890625
],
[
0.08660888671875,
0.393310546875,
0.48046875,
0.393310546875,
0.48046875,
0.404052734375,
0.08660888671875,
0.404052734375
],
[
0.02532958984375,
0.401611328125,
0.058013916015625,
0.401611328125,
0.058013916015625,
0.443115234375,
0.02532958984375,
0.443115234375
],
[
0.08660888671875,
0.406005859375,
0.48193359375,
0.406005859375,
0.48193359375,
0.419189453125,
0.08660888671875,
0.419189453125
],
[
0.51708984375,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.51708984375,
0.415283203125
],
[
0.08660888671875,
0.419189453125,
0.48046875,
0.419189453125,
0.48046875,
0.43310546875,
0.08660888671875,
0.43310546875
],
[
0.51708984375,
0.416748046875,
0.91259765625,
0.416748046875,
0.91259765625,
0.429931640625,
0.51708984375,
0.429931640625
],
[
0.02777099609375,
0.43505859375,
0.05963134765625,
0.43505859375,
0.05963134765625,
0.479736328125,
0.02777099609375,
0.479736328125
],
[
0.08660888671875,
0.43505859375,
0.48046875,
0.43505859375,
0.48046875,
0.44580078125,
0.08660888671875,
0.44580078125
],
[
0.51708984375,
0.429931640625,
0.9140625,
0.429931640625,
0.9140625,
0.443115234375,
0.51708984375,
0.443115234375
],
[
0.0849609375,
0.447509765625,
0.483642578125,
0.447509765625,
0.483642578125,
0.4609375,
0.0849609375,
0.4609375
],
[
0.51708984375,
0.443115234375,
0.9140625,
0.443115234375,
0.9140625,
0.45703125,
0.51708984375,
0.45703125
],
[
0.08660888671875,
0.462158203125,
0.471435546875,
0.462158203125,
0.471435546875,
0.472900390625,
0.08660888671875,
0.472900390625
],
[
0.5185546875,
0.458984375,
0.91259765625,
0.458984375,
0.91259765625,
0.4697265625,
0.5185546875,
0.4697265625
],
[
0.51708984375,
0.4716796875,
0.91259765625,
0.4716796875,
0.91259765625,
0.48486328125,
0.51708984375,
0.48486328125
],
[
0.0228729248046875,
0.48486328125,
0.060455322265625,
0.48486328125,
0.060455322265625,
0.59033203125,
0.0228729248046875,
0.59033203125
],
[
0.0882568359375,
0.488037109375,
0.188720703125,
0.488037109375,
0.188720703125,
0.5,
0.0882568359375,
0.5
],
[
0.5185546875,
0.48486328125,
0.9140625,
0.48486328125,
0.9140625,
0.498779296875,
0.5185546875,
0.498779296875
],
[
0.08416748046875,
0.50390625,
0.48046875,
0.5048828125,
0.48046875,
0.51953125,
0.08416748046875,
0.51806640625
],
[
0.5185546875,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.51123046875,
0.5185546875,
0.51123046875
],
[
0.51708984375,
0.51318359375,
0.91259765625,
0.51318359375,
0.91259765625,
0.5263671875,
0.51708984375,
0.5263671875
],
[
0.08416748046875,
0.51953125,
0.13232421875,
0.517578125,
0.134033203125,
0.5322265625,
0.0849609375,
0.5341796875
],
[
0.51708984375,
0.5263671875,
0.9140625,
0.5263671875,
0.9140625,
0.54052734375,
0.51708984375,
0.54052734375
],
[
0.0849609375,
0.537109375,
0.2303466796875,
0.537109375,
0.2303466796875,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.5185546875,
0.54150390625,
0.91259765625,
0.54150390625,
0.91259765625,
0.5517578125,
0.5185546875,
0.5517578125
],
[
0.0849609375,
0.55029296875,
0.48193359375,
0.55029296875,
0.48193359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.5185546875,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.56640625,
0.5185546875,
0.56640625
],
[
0.0849609375,
0.56298828125,
0.484375,
0.56298828125,
0.484375,
0.5771484375,
0.0849609375,
0.5771484375
],
[
0.515625,
0.56640625,
0.9140625,
0.5673828125,
0.9140625,
0.58203125,
0.515625,
0.58056640625
],
[
0.0196075439453125,
0.57568359375,
0.062103271484375,
0.57568359375,
0.062103271484375,
0.70703125,
0.0196075439453125,
0.70703125
],
[
0.08660888671875,
0.57568359375,
0.483642578125,
0.57568359375,
0.483642578125,
0.5888671875,
0.08660888671875,
0.5888671875
],
[
0.0849609375,
0.58837890625,
0.42236328125,
0.58837890625,
0.42236328125,
0.6015625,
0.0849609375,
0.6015625
],
[
0.515625,
0.583984375,
0.5537109375,
0.58203125,
0.5556640625,
0.5947265625,
0.51708984375,
0.5966796875
],
[
0.53515625,
0.59521484375,
0.9140625,
0.59521484375,
0.9140625,
0.609375,
0.53515625,
0.609375
],
[
0.10455322265625,
0.61572265625,
0.25732421875,
0.61572265625,
0.25732421875,
0.62890625,
0.10455322265625,
0.62890625
],
[
0.5185546875,
0.611328125,
0.9111328125,
0.611328125,
0.9111328125,
0.6220703125,
0.5185546875,
0.6220703125
],
[
0.08905029296875,
0.61865234375,
0.1094970703125,
0.61865234375,
0.1094970703125,
0.626953125,
0.08905029296875,
0.626953125
],
[
0.5185546875,
0.62548828125,
0.9111328125,
0.62548828125,
0.9111328125,
0.63623046875,
0.5185546875,
0.63623046875
],
[
0.0849609375,
0.6318359375,
0.48193359375,
0.63330078125,
0.48193359375,
0.64794921875,
0.0849609375,
0.646484375
],
[
0.51953125,
0.6396484375,
0.91259765625,
0.6396484375,
0.91259765625,
0.6494140625,
0.51953125,
0.6494140625
],
[
0.08660888671875,
0.64794921875,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.08660888671875,
0.6611328125
],
[
0.51708984375,
0.6513671875,
0.9140625,
0.6513671875,
0.9140625,
0.66552734375,
0.51708984375,
0.66552734375
],
[
0.08660888671875,
0.66357421875,
0.48193359375,
0.66357421875,
0.48193359375,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.51708984375,
0.6640625,
0.869140625,
0.6640625,
0.869140625,
0.67822265625,
0.51708984375,
0.67822265625
],
[
0.08416748046875,
0.67529296875,
0.483642578125,
0.6748046875,
0.483642578125,
0.68798828125,
0.08416748046875,
0.689453125
],
[
0.53369140625,
0.6796875,
0.85888671875,
0.6796875,
0.85888671875,
0.68994140625,
0.53369140625,
0.68994140625
],
[
0.08660888671875,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.08660888671875,
0.70263671875
],
[
0.0882568359375,
0.7041015625,
0.48046875,
0.7041015625,
0.48046875,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.5439453125,
0.7041015625,
0.91259765625,
0.7041015625,
0.91259765625,
0.71728515625,
0.5439453125,
0.71728515625
],
[
0.0849609375,
0.71630859375,
0.48193359375,
0.71630859375,
0.48193359375,
0.72998046875,
0.0849609375,
0.72998046875
],
[
0.55810546875,
0.71826171875,
0.91259765625,
0.71826171875,
0.91259765625,
0.73193359375,
0.55810546875,
0.73193359375
],
[
0.0849609375,
0.72998046875,
0.48193359375,
0.72998046875,
0.48193359375,
0.74365234375,
0.0849609375,
0.74365234375
],
[
0.55810546875,
0.734375,
0.9111328125,
0.734375,
0.9111328125,
0.744140625,
0.55810546875,
0.744140625
],
[
0.08660888671875,
0.74560546875,
0.48046875,
0.74560546875,
0.48046875,
0.755859375,
0.08660888671875,
0.755859375
],
[
0.55810546875,
0.74755859375,
0.6650390625,
0.74755859375,
0.6650390625,
0.75830078125,
0.55810546875,
0.75830078125
],
[
0.0849609375,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.771484375,
0.0849609375,
0.771484375
],
[
0.5439453125,
0.7587890625,
0.91259765625,
0.7587890625,
0.91259765625,
0.7734375,
0.5439453125,
0.7734375
],
[
0.55615234375,
0.7744140625,
0.91259765625,
0.7744140625,
0.91259765625,
0.78857421875,
0.55615234375,
0.78857421875
],
[
0.08660888671875,
0.7822265625,
0.48046875,
0.7822265625,
0.48046875,
0.79248046875,
0.08660888671875,
0.79248046875
],
[
0.131591796875,
0.79248046875,
0.419189453125,
0.79248046875,
0.419189453125,
0.80322265625,
0.131591796875,
0.80322265625
],
[
0.414306640625,
0.79248046875,
0.48046875,
0.79248046875,
0.48046875,
0.80322265625,
0.414306640625,
0.80322265625
],
[
0.55615234375,
0.78857421875,
0.8408203125,
0.78857421875,
0.8408203125,
0.8017578125,
0.55615234375,
0.8017578125
],
[
0.08905029296875,
0.79541015625,
0.1282958984375,
0.79541015625,
0.1282958984375,
0.7998046875,
0.08905029296875,
0.7998046875
],
[
0.0849609375,
0.80322265625,
0.48193359375,
0.80322265625,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.0849609375,
0.81103515625,
0.48193359375,
0.81103515625,
0.48193359375,
0.82470703125,
0.0849609375,
0.82470703125
],
[
0.08660888671875,
0.82373046875,
0.48193359375,
0.82373046875,
0.48193359375,
0.833984375,
0.08660888671875,
0.833984375
],
[
0.51953125,
0.82568359375,
0.53173828125,
0.82568359375,
0.53173828125,
0.833984375,
0.51953125,
0.833984375
],
[
0.54150390625,
0.8212890625,
0.6748046875,
0.8212890625,
0.6748046875,
0.8359375,
0.54150390625,
0.8359375
],
[
0.0882568359375,
0.83349609375,
0.48193359375,
0.83349609375,
0.48193359375,
0.84326171875,
0.0882568359375,
0.84326171875
],
[
0.08660888671875,
0.84326171875,
0.336669921875,
0.84326171875,
0.336669921875,
0.85400390625,
0.08660888671875,
0.85400390625
],
[
0.51708984375,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0882568359375,
0.85498046875,
0.301513671875,
0.85498046875,
0.301513671875,
0.8623046875,
0.0882568359375,
0.8623046875
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.08660888671875,
0.86572265625,
0.30224609375,
0.86572265625,
0.30224609375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51708984375,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.51708984375,
0.8818359375
],
[
0.0882568359375,
0.87548828125,
0.1666259765625,
0.87548828125,
0.1666259765625,
0.88330078125,
0.0882568359375,
0.88330078125
],
[
0.2181396484375,
0.875,
0.3095703125,
0.875,
0.3095703125,
0.88525390625,
0.2181396484375,
0.88525390625
],
[
0.0849609375,
0.88525390625,
0.26708984375,
0.88525390625,
0.26708984375,
0.89599609375,
0.0849609375,
0.89599609375
],
[
0.515625,
0.8798828125,
0.9150390625,
0.8798828125,
0.9150390625,
0.89794921875,
0.515625,
0.89794921875
]
] |
[
"Towards Aligned Canonical Correlation Analysis: Preliminary",
"Formulation and Proof-of-Concept Results",
"Biqian Cheng",
"Evangelos E. Papalexakis",
"Jia Chen",
"UC Riverside",
"UC Riverside",
"UC Riverside",
"Riverside, CA, USA",
"Riverside, CA, USA",
"Riverside, CA, USA",
"bchen158@ucr.edu",
"epapalex@cs.ucr.edu",
"jiac@ucr.edu",
"ABSTRACT",
"many decades and has been around for quite a long time, it is still",
"Canonical Correlation Analysis (CCA)hasbeena tried and tested",
"extremely relevant, not only as a standalone data mining tool, but",
"",
"analytical model which seeks to jointly embed two or more views of",
"also due to the fact that cutting-edge self-supervised representation",
"the data in a maximally correlated latent space.In addition to a pow-",
"learningtechniques,such asBarlowTwins[Balestrieroetal.2023:",
"erful dataminingtool,CCAhasmanyimplicationsforcutting-edge",
"Bielak et al. 2022; Zbontar et al. 2021; Zhang et al. 2021], can be",
"self-supervised representation learning approaches, as one can cast",
"essentially seen as variations of CCA, where the goal is to embed",
"",
"a number of recent approaches as variants of that model. In this",
"two views of the data (the original view and the augmentation) in",
"workweexplorewhathappenswhenafundamentalassumptionof",
"a latent space where correlation is maximized.",
"that model breaks:what if the alignment across different views of",
"Traditionally, in CCA-style analysis, we assume that entities",
"the data is unknown? Typically, we would first attempt to align the",
"across views haveone-to-one correspondenceacrossthetwo or",
"8",
"different views, and subsequently apply CCA or any other model",
"moreviews of the data, and there is a wealth of algorithms that",
"study different formulations for solving the problem of projecting",
"in order to embed the data views. Can we do better if we align",
"[DT",
"and embed at the same time? In this work, we seek to jointly solve",
"those views in that desired maximally correlated space, both linearly",
"the alignment and embedding by formulating and solving both",
"and non-linearly [Andrew et al. 2013]. What if, however, this one",
"I'so]",
"problems underthe same umbrellaof Aligned Canonical Correla",
"to-one correspondence is unknown? In this case, we are faced",
"tion Analysis (ACCA). We present a preliminary formulation and",
"with two problems: (1) entity alignment and (2) CCA embedding.",
"alternating optimizationalgorithmandproof-of-conceptresults.",
"Motivated by recent results[Wu et al.2022] in therelated problem",
"of misaligned joint tensor factorization, it turns out thatformulating",
"2162002",
"KEYWORDS",
"and solving the alignment and embedding problems jointly yields",
"Canonical Correlation Analysis, Alignment, Matching, Data Inte",
"betterresults than solvingeachproblem separately inmultiple",
"steps, as it appears that the two sub-problems work synergistically",
"gration",
"to produce better quality alignment and embeddings. In this work,",
"ACM ReferenceFormat:",
"weexplorethistypeofformulationforCCAandweproposeanew",
"Biqian Cheng, Evangelos E. Papalexakis, and Jia Chen. 2023. Towards",
"formulation,theAligned CanonicalCorrelationAnalysis (ACCA)",
"Aligned Canonical Correlation Analysis: Preliminary Formulation and Proof-",
"where we seek to jointly compute the alignment and the embedding",
"'ZICAIXI",
"of-Concept Results. In Proceedings of KDD Conference (KDD'23). ACM, New",
"York, NY, USA, 4 pages. https://doi.org/XxXxxxX.xxxxXxx",
"space.",
"The closest formulation to our proposed model is found in [Sahbi",
"INTRODUCTION",
"2018] where the author is considering linear transformation of the",
"1",
"two views in CCA, however, is not seeking to recoverthe precise",
"Canonical Correlation Analysis (CCA) [Harold 1936; Kettenring",
"alignment matrix as ourformulation does.In our on-going work",
"1971j is a classical model which, given two different views of the",
"we will consider scenarios where we can fairly compare the two",
"same set of entities,e.g.,two different bipartite graphs of (user,",
"formulations and understand pros and cons for either one.",
"product) and (user, video) interactions or different feature represen-",
"Thelist of contributions in this preliminary work are:",
"tations for those entities in general, seeks to project those entities",
"(users)inalow-dimensionalspacewherethedifferentprojected",
".Novel Formulation: We propose the Aligned Canonical",
"views are maximally correlated. Essentially, CCA can jointly em-",
"Correlation Analysis (ACCA) model, which seeks to jointly",
"bed heterogeneous datasets in a common low-dimensional space,",
"identifythebestentityalingmentandlatentembeddingfor",
"as it can beextended tomore thantwoviews[Chen etal.2019",
"the dataset views.",
"2022]. Even though CCA has been in and out of the spotlight for",
"· Proof of Concept: We derive an Alternating Optimization",
"algorithm and show preliminary results for solving the prob-",
"Permission to make digital or hard copies of all or part of this work for personal or",
" use is granted without fee provided that copies are not made",
"or distributed",
"lem, demonstrating the feasibility of our effort.",
"sassroor",
"for profit or commercial advantage and that copies bear this notice and the full citation",
"on the first page. Copyrights for components of this work owned by others than ACM",
"must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,",
"2",
"BACKGROUND",
"to post on servers or to redistribute to lists, requires prior specific permission and/or a",
"fee. Request permissions from permissions@acm.org.",
"Canonical correlation analysis (CCA) is a powerful tool to learn",
"KDD'23.August07.2023,LongBeach,CA.USA",
"the shared latent components of two datasets by projecting them",
"2023Association for Computing Machinery",
"to the same space and enforcing the similarity of the projected",
"ACMISBN978-X",
"-x/YY/MM...$15.00",
"https://doi.org/xxxxxxx.xxxxxxx",
" data. Given two centered and aligned datasets X e RDxxN and"
] |
{
"arxiv_id": "2312.00296",
"language": "en",
"page_idx": 0
}
|
||
2105.04271-en-1
|
2105.04271-en-1.jpg
|
Syntactic Ambiguity. The second example sentence does not have an explicit clue about the relationship between“a device”and“a cellular base station".$^{\hat{2}}$ Thus, existing OpenIE systems often fail to split them, but incorrectly extract the first tuple (node-B; can be; a device a cellular base station). However, context sentence S4 includes an explicit cue to the relationship between the two terms and may thus help split them.
To minimize the aforementioned ambiguities, it is clear that we should leverage document-level context. However, all existing OpenIE datasets are generated or annotated at sentence-level. These datasets include standalone sentences but not their context sentences. Hence they are not suitable for evaluating context-aware tuple extraction.
We annotate the first Document-level contextaware Open Information Extraction (DocOIE) dataset. DocOIE consists of 800 expert annotated sentences from 80 documents; 10 sentences are randomly sampled for annotation from each of 80 documents. To the best of our knowledge, among all OpenIE datasets as of now, DocOIE contains the largest number of expert-annotated sentences. More importantly, DocOIE provides documentlevel contexts, enabling OpenIE models to take relevant contexts for accurate tuple extraction.
Furthermore, to show that document-level context is useful for OpenIE task, we develop the Document-level context-aware Open Information $\mathbf{E}$xtraction (DocIE)model. DocIE encodes a source sentence with its contextual information by using pre-trained BERT (Devlin et al., 2018). Because contextual sentences can be much longer than the source sentence, the syntactic/semantic information in source sentence might be dominated by that of the contexts. Our proposed DocIE model differentiates the source sentence and its contexts by segment tags, and adds additional transformer encoder layers only for the source sentence. In summary, our contributions are threefold:
$\cdot$ We propose a new task in OpenIE to extract relational tuples with document-level contexts. $\cdot$ We introduce DocOIE, an expert annotated
dataset for evaluating document-level OpenIE systems. DocOIE consists of 2,122 relational tuples in 800 expert-annotated sentences, with their document-level contexts.
· We present DocIE, a novel neural OpenIE sys- ${}^{2}$The missing of a comma could be a typo in the patent
document. Nevertheless, not all input sentences to OpenIE
system are typo free in real applications.
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>$\#$Sent.</th>
<th>Source</th>
<th>Annotation</th>
</tr>
<tr>
<td>OIE2016</td>
<td>3,200</td>
<td>QA- SRL</td>
<td>Automatic</td>
</tr>
<tr>
<td>$\mathbf{Wire57}$</td>
<td>57</td>
<td>Wikipedia $N_{f}$ $|ewswire$ T</td>
<td>Expert</td>
</tr>
<tr>
<td>CaRB</td>
<td>50</td>
<td>OIE2016</td>
<td>Expert</td>
</tr>
<tr>
<td>DocOIE</td>
<td>800</td>
<td>Patent</td>
<td>$\mathbf{Expert}$</td>
</tr>
</tbody>
</table>
Table 1: Existing OpenIE datasets with number of sen-
tences, sentence source and annotation type.
tem that can leverage on document-level con-
texts for relational tuple extraction.
## 2 $\textbf{Related Work}$
OpenIE Datasets. Since the OpenIE task was introduced by Yates et al. (2007), the earlier systems were mainly evaluated by using a small number of sentences, without a standardized evaluation procedure (Niklaus et al., 2018). OIE2016 (Stanovsky and Dagan, 2016) is the first large-scale dataset constructed for OpenIE tasks and comes with a standard scoring framework. In OIE2016, the gold tuples are automatically generated from a QA-SRL dataset (He et al., 2015) according to human crafted rules. Wire57 (Lechelle et al., 2019) improves the scorer and manually annotates 57 sentences as a benchmark dataset. Considering that OIE2016 dataset is noisy, Bhardwaj et al. (2019) provide a crowdsourcing dataset named CaRB. CaRB also has 50 expert annotated sentences and a sophisticated scoring framework
As summarized in Table 1, the number of expertannotated sentences in these datasets remains small. Furthermore, the sentences in these datasets do not come with contextual information. In contrast, our DocOIE dataset consists of 800 expert-annotated sentences, and comes with the source documents for accurate sentence interpretation.
$\mathbf{OpenIEModels. }$ TextRunner(Yates et al.,2007) is the frst OpenIE system, followed by Reverb (Fader et al., 2011), OLLIE (Schmitz et al., 2012), Clausie (Del Corro and Gemulla, 2013), CSD-IE (Bast and Haussmann, 2013), Stanford OpenIE (Angeli et al.,2015), Openie4 (Mausam, 2016), Openie5$^3$,NESTIE (Bhutani et al., 2016), MINIE (Gashteovski et al., 2017) and Graphene (Cetto et al., 2018). We consider them as traditional OpenIE models. These models ex-
$^{3}$github:dair-iitd/openie-standalone
|
[
[
0.118408203125,
0.0748291015625,
0.489501953125,
0.0748291015625,
0.489501953125,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.5224609375,
0.0789794921875,
0.58447265625,
0.0789794921875,
0.58447265625,
0.09320068359375,
0.5224609375,
0.09320068359375
],
[
0.60791015625,
0.0789794921875,
0.6650390625,
0.0789794921875,
0.6650390625,
0.09320068359375,
0.60791015625,
0.09320068359375
],
[
0.6767578125,
0.0789794921875,
0.73486328125,
0.0789794921875,
0.73486328125,
0.09320068359375,
0.6767578125,
0.09320068359375
],
[
0.76416015625,
0.0789794921875,
0.8505859375,
0.0789794921875,
0.8505859375,
0.09320068359375,
0.76416015625,
0.09320068359375
],
[
0.12005615234375,
0.09320068359375,
0.490234375,
0.09320068359375,
0.490234375,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.5224609375,
0.1009521484375,
0.59619140625,
0.1009521484375,
0.59619140625,
0.11517333984375,
0.5224609375,
0.11517333984375
],
[
0.61376953125,
0.1009521484375,
0.6650390625,
0.1009521484375,
0.6650390625,
0.11517333984375,
0.61376953125,
0.11517333984375
],
[
0.6767578125,
0.1009521484375,
0.748046875,
0.1009521484375,
0.748046875,
0.11517333984375,
0.6767578125,
0.11517333984375
],
[
0.118408203125,
0.10809326171875,
0.489501953125,
0.10809326171875,
0.489501953125,
0.1217041015625,
0.118408203125,
0.1217041015625
],
[
0.76416015625,
0.10272216796875,
0.8447265625,
0.10272216796875,
0.8447265625,
0.114013671875,
0.76416015625,
0.114013671875
],
[
0.5224609375,
0.1175537109375,
0.5810546875,
0.1175537109375,
0.5810546875,
0.1318359375,
0.5224609375,
0.1318359375
],
[
0.63818359375,
0.1175537109375,
0.66162109375,
0.11517333984375,
0.6630859375,
0.1300048828125,
0.6396484375,
0.1318359375
],
[
0.6767578125,
0.1175537109375,
0.8193359375,
0.1175537109375,
0.8193359375,
0.1318359375,
0.6767578125,
0.1318359375
],
[
0.118408203125,
0.12353515625,
0.489501953125,
0.12353515625,
0.489501953125,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.67529296875,
0.1318359375,
0.75390625,
0.1341552734375,
0.75390625,
0.1484375,
0.67529296875,
0.146728515625
],
[
0.1175537109375,
0.14013671875,
0.489501953125,
0.1395263671875,
0.489501953125,
0.15380859375,
0.1175537109375,
0.155029296875
],
[
0.5224609375,
0.146728515625,
0.57275390625,
0.1484375,
0.57080078125,
0.1644287109375,
0.5205078125,
0.162109375
],
[
0.6171875,
0.1484375,
0.6650390625,
0.1484375,
0.6650390625,
0.1644287109375,
0.6171875,
0.1644287109375
],
[
0.67822265625,
0.15087890625,
0.748046875,
0.15087890625,
0.748046875,
0.162109375,
0.67822265625,
0.162109375
],
[
0.76318359375,
0.1484375,
0.8818359375,
0.149658203125,
0.8818359375,
0.1644287109375,
0.76318359375,
0.163330078125
],
[
0.118408203125,
0.15673828125,
0.489501953125,
0.15673828125,
0.489501953125,
0.17041015625,
0.118408203125,
0.17041015625
],
[
0.5224609375,
0.1650390625,
0.57421875,
0.1650390625,
0.57421875,
0.179931640625,
0.5224609375,
0.179931640625
],
[
0.63818359375,
0.166259765625,
0.6630859375,
0.166259765625,
0.6630859375,
0.179931640625,
0.63818359375,
0.179931640625
],
[
0.6767578125,
0.1650390625,
0.75,
0.1650390625,
0.75,
0.179931640625,
0.6767578125,
0.179931640625
],
[
0.76171875,
0.166259765625,
0.8193359375,
0.166259765625,
0.8193359375,
0.18115234375,
0.76171875,
0.18115234375
],
[
0.118408203125,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.5205078125,
0.188232421875,
0.59033203125,
0.188232421875,
0.59033203125,
0.203125,
0.5205078125,
0.203125
],
[
0.6767578125,
0.188232421875,
0.72900390625,
0.188232421875,
0.72900390625,
0.203125,
0.6767578125,
0.203125
],
[
0.76318359375,
0.188232421875,
0.8193359375,
0.188232421875,
0.8193359375,
0.2037353515625,
0.76318359375,
0.2037353515625
],
[
0.63037109375,
0.189453125,
0.6630859375,
0.189453125,
0.6630859375,
0.201904296875,
0.63037109375,
0.201904296875
],
[
0.118408203125,
0.189453125,
0.489501953125,
0.189453125,
0.489501953125,
0.203125,
0.118408203125,
0.203125
],
[
0.1175537109375,
0.204833984375,
0.30908203125,
0.204833984375,
0.30908203125,
0.2197265625,
0.1175537109375,
0.2197265625
],
[
0.1368408203125,
0.2203369140625,
0.489501953125,
0.2203369140625,
0.489501953125,
0.2340087890625,
0.1368408203125,
0.2340087890625
],
[
0.51123046875,
0.218505859375,
0.8818359375,
0.218505859375,
0.8818359375,
0.232177734375,
0.51123046875,
0.232177734375
],
[
0.1175537109375,
0.2369384765625,
0.489501953125,
0.2369384765625,
0.489501953125,
0.25048828125,
0.1175537109375,
0.25048828125
],
[
0.51318359375,
0.2340087890625,
0.81005859375,
0.2340087890625,
0.81005859375,
0.2476806640625,
0.51318359375,
0.2476806640625
],
[
0.118408203125,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.5400390625,
0.26123046875,
0.884765625,
0.26123046875,
0.884765625,
0.275634765625,
0.5400390625,
0.275634765625
],
[
0.12005615234375,
0.26953125,
0.489501953125,
0.26953125,
0.489501953125,
0.283203125,
0.12005615234375,
0.283203125
],
[
0.5380859375,
0.27783203125,
0.80029296875,
0.27783203125,
0.80029296875,
0.29150390625,
0.5380859375,
0.29150390625
],
[
0.12005615234375,
0.287353515625,
0.48779296875,
0.287353515625,
0.48779296875,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.12005615234375,
0.3017578125,
0.490234375,
0.3017578125,
0.490234375,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.51318359375,
0.3046875,
0.66650390625,
0.3046875,
0.66650390625,
0.318359375,
0.51318359375,
0.318359375
],
[
0.1175537109375,
0.317626953125,
0.426513671875,
0.317626953125,
0.426513671875,
0.330810546875,
0.1175537109375,
0.330810546875
],
[
0.139404296875,
0.3349609375,
0.491943359375,
0.3349609375,
0.491943359375,
0.345703125,
0.139404296875,
0.345703125
],
[
0.51318359375,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.1175537109375,
0.3486328125,
0.489501953125,
0.3486328125,
0.489501953125,
0.3623046875,
0.1175537109375,
0.3623046875
],
[
0.51318359375,
0.347412109375,
0.884765625,
0.347412109375,
0.884765625,
0.361083984375,
0.51318359375,
0.361083984375
],
[
0.118408203125,
0.364013671875,
0.489501953125,
0.365234375,
0.489501953125,
0.380126953125,
0.118408203125,
0.37890625
],
[
0.51318359375,
0.364013671875,
0.88671875,
0.364013671875,
0.88671875,
0.377685546875,
0.51318359375,
0.377685546875
],
[
0.1175537109375,
0.3818359375,
0.490234375,
0.3818359375,
0.490234375,
0.3955078125,
0.1175537109375,
0.3955078125
],
[
0.51318359375,
0.380126953125,
0.88671875,
0.380126953125,
0.88671875,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.1175537109375,
0.3984375,
0.490234375,
0.3984375,
0.490234375,
0.412109375,
0.1175537109375,
0.412109375
],
[
0.51123046875,
0.394287109375,
0.88330078125,
0.3955078125,
0.88330078125,
0.410400390625,
0.51123046875,
0.4091796875
],
[
0.118408203125,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.118408203125,
0.4287109375
],
[
0.51318359375,
0.410888671875,
0.884765625,
0.410888671875,
0.884765625,
0.42578125,
0.51318359375,
0.42578125
],
[
0.1175537109375,
0.429931640625,
0.491943359375,
0.4287109375,
0.491943359375,
0.443603515625,
0.1175537109375,
0.44482421875
],
[
0.51318359375,
0.427490234375,
0.884765625,
0.427490234375,
0.884765625,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.118408203125,
0.446533203125,
0.49365234375,
0.446533203125,
0.49365234375,
0.46142578125,
0.118408203125,
0.46142578125
],
[
0.5146484375,
0.443603515625,
0.88330078125,
0.443603515625,
0.88330078125,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.118408203125,
0.463134765625,
0.491943359375,
0.463134765625,
0.491943359375,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.5146484375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.5146484375,
0.473876953125
],
[
0.12005615234375,
0.478515625,
0.490234375,
0.478515625,
0.490234375,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.51318359375,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.1175537109375,
0.495361328125,
0.459228515625,
0.495361328125,
0.459228515625,
0.5087890625,
0.1175537109375,
0.5087890625
],
[
0.51318359375,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.505859375,
0.51318359375,
0.505859375
],
[
0.1376953125,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.5224609375,
0.1376953125,
0.5224609375
],
[
0.51318359375,
0.50830078125,
0.88330078125,
0.50830078125,
0.88330078125,
0.521484375,
0.51318359375,
0.521484375
],
[
0.1175537109375,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.1175537109375,
0.54052734375
],
[
0.51318359375,
0.52392578125,
0.88330078125,
0.52392578125,
0.88330078125,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.118408203125,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5546875,
0.118408203125,
0.5546875
],
[
0.51318359375,
0.54052734375,
0.884765625,
0.54052734375,
0.884765625,
0.55419921875,
0.51318359375,
0.55419921875
],
[
0.12005615234375,
0.55859375,
0.490234375,
0.55859375,
0.490234375,
0.572265625,
0.12005615234375,
0.572265625
],
[
0.51318359375,
0.55712890625,
0.884765625,
0.55712890625,
0.884765625,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.12005615234375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.58740234375,
0.12005615234375,
0.58740234375
],
[
0.51318359375,
0.572265625,
0.884765625,
0.572265625,
0.884765625,
0.5859375,
0.51318359375,
0.5859375
],
[
0.1158447265625,
0.59130859375,
0.490234375,
0.59033203125,
0.490234375,
0.60498046875,
0.1158447265625,
0.60546875
],
[
0.51123046875,
0.5888671875,
0.705078125,
0.5888671875,
0.705078125,
0.6025390625,
0.51123046875,
0.6025390625
],
[
0.12005615234375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.53076171875,
0.60400390625,
0.88671875,
0.60400390625,
0.88671875,
0.6181640625,
0.53076171875,
0.6181640625
],
[
0.118408203125,
0.62353515625,
0.491943359375,
0.62353515625,
0.491943359375,
0.63720703125,
0.118408203125,
0.63720703125
],
[
0.51318359375,
0.62158203125,
0.88671875,
0.62158203125,
0.88671875,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.118408203125,
0.64013671875,
0.490234375,
0.64013671875,
0.490234375,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.51318359375,
0.63720703125,
0.88671875,
0.63720703125,
0.88671875,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.1175537109375,
0.65478515625,
0.489501953125,
0.65380859375,
0.489501953125,
0.66845703125,
0.1175537109375,
0.66943359375
],
[
0.51123046875,
0.65283203125,
0.884765625,
0.65380859375,
0.884765625,
0.66748046875,
0.51123046875,
0.66650390625
],
[
0.118408203125,
0.67138671875,
0.490234375,
0.67138671875,
0.490234375,
0.68505859375,
0.118408203125,
0.68505859375
],
[
0.51318359375,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.118408203125,
0.6884765625,
0.491943359375,
0.6884765625,
0.491943359375,
0.70166015625,
0.118408203125,
0.70166015625
],
[
0.5146484375,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.118408203125,
0.70361328125,
0.489501953125,
0.70361328125,
0.489501953125,
0.71728515625,
0.118408203125,
0.71728515625
],
[
0.51318359375,
0.70166015625,
0.78076171875,
0.70166015625,
0.78076171875,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.1175537109375,
0.72021484375,
0.426513671875,
0.71875,
0.426513671875,
0.732421875,
0.1175537109375,
0.73388671875
],
[
0.51318359375,
0.7255859375,
0.8818359375,
0.7255859375,
0.8818359375,
0.7392578125,
0.51318359375,
0.7392578125
],
[
0.1275634765625,
0.736328125,
0.489501953125,
0.736328125,
0.489501953125,
0.75048828125,
0.1275634765625,
0.75048828125
],
[
0.51123046875,
0.7421875,
0.884765625,
0.7412109375,
0.884765625,
0.755859375,
0.51123046875,
0.75732421875
],
[
0.1435546875,
0.751953125,
0.490234375,
0.751953125,
0.490234375,
0.765625,
0.1435546875,
0.765625
],
[
0.51318359375,
0.7587890625,
0.88330078125,
0.7587890625,
0.88330078125,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.1292724609375,
0.7685546875,
0.490234375,
0.7685546875,
0.490234375,
0.7822265625,
0.1292724609375,
0.7822265625
],
[
0.51318359375,
0.77490234375,
0.88330078125,
0.77490234375,
0.88330078125,
0.7880859375,
0.51318359375,
0.7880859375
],
[
0.1435546875,
0.78515625,
0.489501953125,
0.78515625,
0.489501953125,
0.798828125,
0.1435546875,
0.798828125
],
[
0.51318359375,
0.7890625,
0.88330078125,
0.79052734375,
0.88330078125,
0.8046875,
0.51318359375,
0.80419921875
],
[
0.1435546875,
0.80029296875,
0.489501953125,
0.80029296875,
0.489501953125,
0.81396484375,
0.1435546875,
0.81396484375
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.1435546875,
0.81640625,
0.489501953125,
0.81640625,
0.489501953125,
0.830078125,
0.1435546875,
0.830078125
],
[
0.51318359375,
0.822265625,
0.884765625,
0.822265625,
0.884765625,
0.8359375,
0.51318359375,
0.8359375
],
[
0.145263671875,
0.83447265625,
0.366943359375,
0.83447265625,
0.366943359375,
0.84423828125,
0.145263671875,
0.84423828125
],
[
0.51123046875,
0.8388671875,
0.88330078125,
0.8388671875,
0.88330078125,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.1275634765625,
0.84716796875,
0.490234375,
0.8486328125,
0.490234375,
0.86328125,
0.1275634765625,
0.8623046875
],
[
0.51318359375,
0.85498046875,
0.884765625,
0.85498046875,
0.884765625,
0.869140625,
0.51318359375,
0.869140625
],
[
0.1376953125,
0.869140625,
0.490234375,
0.87060546875,
0.490234375,
0.88525390625,
0.1376953125,
0.88427734375
],
[
0.51123046875,
0.87158203125,
0.88330078125,
0.87158203125,
0.88330078125,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.88427734375,
0.489501953125,
0.88427734375,
0.489501953125,
0.89794921875,
0.118408203125,
0.89794921875
],
[
0.118408203125,
0.896484375,
0.366943359375,
0.896484375,
0.366943359375,
0.91015625,
0.118408203125,
0.91015625
],
[
0.53564453125,
0.8955078125,
0.85205078125,
0.8955078125,
0.85205078125,
0.9091796875,
0.53564453125,
0.9091796875
]
] |
[
"Syntactic Ambiguity. The second example sen-",
"Dataset",
"#Sent.",
"Source",
"Annotation",
"tence does not have an explicit clue about the re-",
"OIE2016",
"3,200",
"QA-SRL",
"lationship between “a device\" and “\"a cellular base",
"Automatic",
"Wire57",
"57",
"Wikipedia Expert",
"station\"2 Thus, existing OpenIE systems often fail",
"Newswire",
"to split them, but incorrectly extract the first tuple",
"CaRB",
"1,282",
"OIE2016",
"Crowdsourcing",
"(node-B; can be; a device a cellular base station)",
"CaRB",
"50",
"OIE2016",
"Expert",
"However, context sentence S4 includes an explicit",
"DocOIE",
"Patent",
"Expert",
"800",
"cue to the relationship between the two terms and",
"may thus help split them.",
"To minimize the aforementioned ambiguities, it",
"Table l: Existing OpenIE datasets with number of sen",
"is clear that we should leverage document-level",
"tences, sentence source and annotation type.",
"context. However, all existing OpenlE datasets are",
"tem that can leverage on document-level con-",
"generated or annotated at sentence-level. These",
"texts for relational tuple extraction.",
"datasets include standalonesentencesbutnottheir",
"context sentences. Hence they are not suitable for",
"2Related Work",
"evaluating context-aware tuple extraction.",
"We annotate the first Document-level context-",
"OpenIE Datasets.Since the OpenIE task was in-",
"aware Open Information Extraction (DocOIE)",
"troduced by Yates et al. (2007), the earlier systems",
"dataset. DocOIE consists of 800 expert annotated",
"were mainly evaluated by using a small number of",
"sentences from 80 documents; 10 sentences are",
"sentences, without a standardized evaluation proce-",
"randomly sampled for annotation from each of 80",
"dure (Niklaus et al., 2018). OIE2016 (Stanovsky",
"documents. To the best of our knowledge, among",
"and Dagan, 2016) is the first large-scale dataset",
"all OpenIE datasets as of now, DocOIE contains",
"constructed for OpenIE tasks and comes with a",
"the largest number of expert-annotated sentences.",
"standard scoring framework. In OIE2016, the gold",
"More importantly, DocOIE provides document-",
"tuples are automatically generated from a QA-SRL",
"level contexts, enabling OpenIE models to take",
"dataset (He et al., 2015) according to human crafted",
"relevant contexts for accurate tuple extraction.",
"rules. Wire57 (Lechelle et al., 2019) improves",
"Furthermore,toshowthatdocument-level con-",
"the scorer and manually annotates 57 sentences",
"text is useful for OpenlE task, we develop the",
"as a benchmark dataset. Considering that OIE2016",
"Document-level context-aware Open Information",
"dataset is noisy, Bhardwaj et al. (2019) provide a",
"Extraction (DocIE) model. DoclE encodes a",
"crowdsourcing dataset named CaRB. CaRB also",
"sourcesentencewithitscontextualinformationby",
"has 50 expert annotated sentences and a sophisti-",
"using pre-trained BERT (Devlin et al., 2018). Be-",
"cated scoring framework.",
"causecontextualsentencescanbemuchlongerthan",
"As summarized in Table 1, the number of expert-",
"the source sentence, the syntactic/semantic infor-",
"annotated sentences in these datasets remains small.",
"mation in source sentence might be dominated by",
"Furthermore, the sentences in these datasets do not",
"that of the contexts. Our proposed DocIE model",
"come with contextual information. In contrast, our",
"differentiates the source sentence and its contexts",
"DocOIE dataset consists of 800 expert-annotated",
"by segment tags, and adds additional transformer",
"sentences, and comes with the source documents",
"encoder layers only for the source sentence.In",
"for accurate sentence interpretation.",
"summary, our contributions are threefold:",
"OpenIE Models. TextRunner (Yates et al., 2007)",
". We propose a new task in OpenlE to extract",
"is the first OpenlE system, followed by Re-",
"relational tuples with document-level contexts.",
"verb (Fader et al., 2011), OLLIE (Schmitz",
". We introduce DocOIE, an expert annotated",
"et al., 2012), Clausie (Del Corro and Gemulla",
"dataset for evaluating document-level OpenIE",
"2013), CSD-IE (Bast and Haussmann, 2013),",
"systems. DocOIE consists of 2,122 relational",
"Stanford OpenIE (Angeli et al., 2015), Ope",
"tuples in 800 expert-annotated sentences, with",
"nie4 (Mausam, 2016), Openie53, NESTIE (Bhutani",
"theirdocument-levelcontexts.",
"et al., 2016), MINIE (Gashteovski et al., 2017) and",
". We present DocIE, a novel neural OpenIE sys-",
"Graphene (Cetto et al., 2018). We consider them",
"2The missing of a comma could be a typo in the patent",
" as traditional OpenIE models. These models ex-",
"document. Nevertheless, not all input sentences to OpenlE",
"system are typo free in real applications.",
"3github:dair-iitd/openie-standalone"
] |
{
"arxiv_id": "2105.04271",
"language": "en",
"page_idx": 1
}
|
||
2105.04271-en-2
|
2105.04271-en-2.jpg
|
tract relational tuples based on handcrafted rules or statistical methods. They usually rely on prior syntactic or semantic analysis. Consequently, any error accumulated in the prior stages deteriorates model performance.
Recently, neural OpenIE systems have been developed and showed promising results (Cui et al., 2018; Zhan and Zhao, 2020; Kolluru et al., 2020a,b). Different from the traditional models, neural OpenIE models extract tuples in an end-toend manner, not requiring prior syntactic or semantic analysis. In principle, the traditional rule-based or statistical OpenIE models do not need training. However, neural OpenIE models need a large number of training samples to learn the extraction patterns. For instance, IMOJIE (Kolluru et al., 2020b) uses about 100,000 sentences for model training. It is unrealistic and expensive to manually annotate 100,000 sentences simply for training purpose. Therefore, a common practice in learning a neural OpenIE model is to use tuples automatically extracted by traditional systems as training data, $i.e.$, $a\textit{ bootstrapping strategy. We consider these im- }$ perfect training labels generated via bootstrapping as pseudo labels. The pseudo labels used in (Cui et al., 2018) are by Openie4, and those in (Kolluru et al., 2020b) are from multiple OpenIE systems.
To the best of our knowledge, no OpenIE models consider document-level contexts in the tuple extraction. Nonetheless, our neural model unavoidably requires extractions of pseudo labels bootstrapped from traditional models for training. To ensure reproducibility, as part of the DocOIE dataset, we also release the documents that are used for generating the pseudo labels.
## 3 $\textbf{DocOIE Dataset}$
We now present our Document-level context-aware Open Information Extraction (DocOIE) dataset. We first introduce the data selection and collection process, and then the annotation process by two experts. Moreover, we explain our annotation consistency measurement to indicate the high-level annotation consistency in DocOIE. In summary, DocOIE consists of two datasets: evaluation dataset and training dataset.
800 expert-
Evaluation dataset contains
annotated sentences, sampled from 80 documents in two domains (healthcare and transportation). Specifically, 10 sentences are sampled for annotation from each of the 40 documents in one domain.
<table>
<tbody>
<tr>
<th>Item</th>
<th>$\#$Item</th>
<th>Metric</th>
<th>Average</th>
<th>Min$\sim$Max</th>
</tr>
<tr>
<td>Doc</td>
<td>80</td>
<td>$N_{i}$ sent</td>
<td>101.78</td>
<td>$44{\sim}138$</td>
</tr>
<tr>
<td rowspan="1">Sent</td>
<td> </td>
<td>$L$ sent</td>
<td>22.69</td>
<td>$5{\sim}47$</td>
</tr>
<tr>
<td rowspan="3">Tuple</td>
<td> </td>
<td>$L$ sub</td>
<td>3.70</td>
<td>$1\sim17$</td>
</tr>
<tr>
<td>2,122</td>
<td>$L_{rel}$</td>
<td>3.39</td>
<td>$1\sim12$</td>
</tr>
<tr>
<td> </td>
<td>$L_{obj}$</td>
<td>3.94</td>
<td>$0{\sim}27$</td>
</tr>
</tbody>
</table>
Table 2: Statistics of DocOIE evaluation dataset. $N_{\{\cdot\}}$ denotes the number of units (i.e., sentence or tuple); $L_{\{\cdot\}}$ denotes the length (number of words) of the unit.
In total, 2,122 relational tuples are annotated in the 800 sampled sentences (refer to Table 2 for detailed statistics).
Training dataset contains 2,400 documents from the two domains (healthcare and transportation); 1,200 documents in each domain. All sentences from these documents are used to bootstrap pseudo labels for neural model training.$^{4}$
## 3.1 Dataset Collection
OpenIE, by definition, is to extract relational tuples in open domain. Ideally, sentences/documents in DocOIE dataset shall not be restricted to any particular document type or topical domain. However, it is challenging to include all types of documents and annotate them. In fact, all existing annotations are restricted to specific types of documents like news and Wikipedia articles (Niklaus et al., 2018).
Document Type Selection In building DocOIE,
we focus on formally written documents and leave it for future work to explore other kinds of documents. We select the type of formal documents with four criteria: (i) Adequacy: as a documentlevel context-aware dataset, each document shall have a reasonable number of sentences to provide sufficient context. (ii) Consistency: each document shall focus on a central topic. In such a way, sentences within the same document are correlated to and consistent with each other, which helps derive proper context. (iii) Informativeness: a document is considered informative if it contains informative entities like technical concepts, relations, and events. Intuitively, OpenIE models are more useful for extracting factual tuples in informative documents. (iv) Syntactic Variety: sentences in
$^{4}$Only document IDs are included in DocOIE, for docu-
ment collection at http://patft.uspto.gov/
|
[
[
0.118408203125,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.118408203125,
0.09027099609375
],
[
0.58251953125,
0.0789794921875,
0.63818359375,
0.0789794921875,
0.63818359375,
0.09320068359375,
0.58251953125,
0.09320068359375
],
[
0.646484375,
0.0789794921875,
0.705078125,
0.0789794921875,
0.705078125,
0.09320068359375,
0.646484375,
0.09320068359375
],
[
0.71533203125,
0.0789794921875,
0.7841796875,
0.0789794921875,
0.7841796875,
0.09320068359375,
0.71533203125,
0.09320068359375
],
[
0.7958984375,
0.07659912109375,
0.87890625,
0.0789794921875,
0.87890625,
0.09320068359375,
0.7958984375,
0.0914306640625
],
[
0.118408203125,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.118408203125,
0.10687255859375
],
[
0.5224609375,
0.08013916015625,
0.5634765625,
0.08013916015625,
0.5634765625,
0.09320068359375,
0.5224609375,
0.09320068359375
],
[
0.51904296875,
0.1009521484375,
0.56103515625,
0.1009521484375,
0.56103515625,
0.11639404296875,
0.51904296875,
0.11639404296875
],
[
0.646484375,
0.097412109375,
0.69873046875,
0.10272216796875,
0.6962890625,
0.119384765625,
0.64404296875,
0.11517333984375
],
[
0.72900390625,
0.1009521484375,
0.7841796875,
0.1009521484375,
0.7841796875,
0.11639404296875,
0.72900390625,
0.11639404296875
],
[
0.81689453125,
0.1009521484375,
0.880859375,
0.1009521484375,
0.880859375,
0.11517333984375,
0.81689453125,
0.11517333984375
],
[
0.118408203125,
0.1092529296875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.609375,
0.1015625,
0.63623046875,
0.1015625,
0.63623046875,
0.11517333984375,
0.609375,
0.11517333984375
],
[
0.6484375,
0.12353515625,
0.6962890625,
0.1258544921875,
0.6943359375,
0.141357421875,
0.646484375,
0.1395263671875
],
[
0.736328125,
0.12353515625,
0.7841796875,
0.12353515625,
0.7841796875,
0.1395263671875,
0.736328125,
0.1395263671875
],
[
0.8330078125,
0.12353515625,
0.87890625,
0.12353515625,
0.87890625,
0.1383056640625,
0.8330078125,
0.1383056640625
],
[
0.60107421875,
0.1312255859375,
0.634765625,
0.1312255859375,
0.634765625,
0.146728515625,
0.60107421875,
0.146728515625
],
[
0.52294921875,
0.133056640625,
0.56103515625,
0.133056640625,
0.56103515625,
0.1455078125,
0.52294921875,
0.1455078125
],
[
0.6484375,
0.1395263671875,
0.7001953125,
0.1436767578125,
0.69775390625,
0.157958984375,
0.6455078125,
0.15380859375
],
[
0.744140625,
0.1395263671875,
0.7841796875,
0.1395263671875,
0.7841796875,
0.155029296875,
0.744140625,
0.155029296875
],
[
0.8427734375,
0.1395263671875,
0.8818359375,
0.1395263671875,
0.8818359375,
0.155029296875,
0.8427734375,
0.155029296875
],
[
0.118408203125,
0.141357421875,
0.270263671875,
0.141357421875,
0.270263671875,
0.155029296875,
0.118408203125,
0.155029296875
],
[
0.118408203125,
0.1258544921875,
0.489501953125,
0.1258544921875,
0.489501953125,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.1368408203125,
0.15673828125,
0.489501953125,
0.15673828125,
0.489501953125,
0.17041015625,
0.1368408203125,
0.17041015625
],
[
0.6484375,
0.160888671875,
0.68994140625,
0.163330078125,
0.6884765625,
0.179931640625,
0.646484375,
0.1776123046875
],
[
0.74462890625,
0.162109375,
0.78515625,
0.162109375,
0.78515625,
0.1776123046875,
0.74462890625,
0.1776123046875
],
[
0.83447265625,
0.162109375,
0.880859375,
0.162109375,
0.880859375,
0.1776123046875,
0.83447265625,
0.1776123046875
],
[
0.12005615234375,
0.1728515625,
0.489501953125,
0.1728515625,
0.489501953125,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.58447265625,
0.1776123046875,
0.63623046875,
0.1776123046875,
0.63623046875,
0.1923828125,
0.58447265625,
0.1923828125
],
[
0.646484375,
0.177001953125,
0.68701171875,
0.1817626953125,
0.6826171875,
0.19775390625,
0.64208984375,
0.1923828125
],
[
0.83447265625,
0.1776123046875,
0.880859375,
0.1776123046875,
0.880859375,
0.193603515625,
0.83447265625,
0.193603515625
],
[
0.74658203125,
0.1787109375,
0.78271484375,
0.1787109375,
0.78271484375,
0.191162109375,
0.74658203125,
0.191162109375
],
[
0.52294921875,
0.1787109375,
0.57080078125,
0.1787109375,
0.57080078125,
0.193603515625,
0.52294921875,
0.193603515625
],
[
0.12005615234375,
0.188232421875,
0.490234375,
0.188232421875,
0.490234375,
0.201904296875,
0.12005615234375,
0.201904296875
],
[
0.646484375,
0.19482421875,
0.68701171875,
0.196533203125,
0.68603515625,
0.2120361328125,
0.6455078125,
0.210205078125
],
[
0.74462890625,
0.193603515625,
0.78515625,
0.193603515625,
0.78515625,
0.208984375,
0.74462890625,
0.208984375
],
[
0.8330078125,
0.19482421875,
0.87890625,
0.19482421875,
0.87890625,
0.210205078125,
0.8330078125,
0.210205078125
],
[
0.12005615234375,
0.204833984375,
0.490234375,
0.204833984375,
0.490234375,
0.218505859375,
0.12005615234375,
0.218505859375
],
[
0.118408203125,
0.221435546875,
0.49365234375,
0.221435546875,
0.49365234375,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.509765625,
0.221435546875,
0.88330078125,
0.223876953125,
0.88330078125,
0.2410888671875,
0.509765625,
0.2392578125
],
[
0.118408203125,
0.2381591796875,
0.495361328125,
0.2381591796875,
0.495361328125,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.51123046875,
0.240478515625,
0.88330078125,
0.240478515625,
0.88330078125,
0.254150390625,
0.51123046875,
0.254150390625
],
[
0.118408203125,
0.254150390625,
0.490234375,
0.254150390625,
0.490234375,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.51123046875,
0.254638671875,
0.87890625,
0.2529296875,
0.87890625,
0.267333984375,
0.51123046875,
0.26953125
],
[
0.118408203125,
0.268310546875,
0.489501953125,
0.26953125,
0.489501953125,
0.283935546875,
0.118408203125,
0.283203125
],
[
0.12005615234375,
0.287353515625,
0.48779296875,
0.287353515625,
0.48779296875,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.5146484375,
0.29443359375,
0.88330078125,
0.29443359375,
0.88330078125,
0.30810546875,
0.5146484375,
0.30810546875
],
[
0.118408203125,
0.3017578125,
0.489501953125,
0.3017578125,
0.489501953125,
0.3154296875,
0.118408203125,
0.3154296875
],
[
0.51318359375,
0.31005859375,
0.88330078125,
0.31005859375,
0.88330078125,
0.32373046875,
0.51318359375,
0.32373046875
],
[
0.118408203125,
0.31640625,
0.489501953125,
0.31640625,
0.489501953125,
0.330078125,
0.118408203125,
0.330078125
],
[
0.5146484375,
0.327880859375,
0.6533203125,
0.327880859375,
0.6533203125,
0.341552734375,
0.5146484375,
0.341552734375
],
[
0.12005615234375,
0.334228515625,
0.489501953125,
0.334228515625,
0.489501953125,
0.347412109375,
0.12005615234375,
0.347412109375
],
[
0.118408203125,
0.3486328125,
0.491943359375,
0.3486328125,
0.491943359375,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.51318359375,
0.3515625,
0.88330078125,
0.3515625,
0.88330078125,
0.365234375,
0.51318359375,
0.365234375
],
[
0.1175537109375,
0.364013671875,
0.490234375,
0.365234375,
0.490234375,
0.380126953125,
0.1175537109375,
0.37890625
],
[
0.51123046875,
0.36767578125,
0.88330078125,
0.36767578125,
0.88330078125,
0.38134765625,
0.51123046875,
0.38134765625
],
[
0.118408203125,
0.3818359375,
0.489501953125,
0.3818359375,
0.489501953125,
0.3955078125,
0.118408203125,
0.3955078125
],
[
0.51123046875,
0.383056640625,
0.88671875,
0.38427734375,
0.88671875,
0.3984375,
0.51123046875,
0.39794921875
],
[
0.12005615234375,
0.3984375,
0.49365234375,
0.3984375,
0.49365234375,
0.412109375,
0.12005615234375,
0.412109375
],
[
0.51318359375,
0.40087890625,
0.88330078125,
0.40087890625,
0.88330078125,
0.41455078125,
0.51318359375,
0.41455078125
],
[
0.1175537109375,
0.4150390625,
0.49365234375,
0.4150390625,
0.49365234375,
0.4287109375,
0.1175537109375,
0.4287109375
],
[
0.51318359375,
0.416259765625,
0.8154296875,
0.416259765625,
0.8154296875,
0.429931640625,
0.51318359375,
0.429931640625
],
[
0.118408203125,
0.43115234375,
0.490234375,
0.43115234375,
0.490234375,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.118408203125,
0.446533203125,
0.490234375,
0.446533203125,
0.490234375,
0.46142578125,
0.118408203125,
0.46142578125
],
[
0.51318359375,
0.443603515625,
0.7021484375,
0.443603515625,
0.7021484375,
0.457275390625,
0.51318359375,
0.457275390625
],
[
0.118408203125,
0.463134765625,
0.490234375,
0.463134765625,
0.490234375,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.51318359375,
0.46435546875,
0.884765625,
0.46435546875,
0.884765625,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.12005615234375,
0.478515625,
0.490234375,
0.478515625,
0.490234375,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.51318359375,
0.48095703125,
0.884765625,
0.48095703125,
0.884765625,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.118408203125,
0.49462890625,
0.480224609375,
0.49462890625,
0.480224609375,
0.50830078125,
0.118408203125,
0.50830078125
],
[
0.51123046875,
0.495361328125,
0.88671875,
0.496337890625,
0.88671875,
0.51123046875,
0.51123046875,
0.51025390625
],
[
0.1368408203125,
0.51123046875,
0.490234375,
0.51123046875,
0.490234375,
0.52490234375,
0.1368408203125,
0.52490234375
],
[
0.51318359375,
0.51318359375,
0.88671875,
0.51318359375,
0.88671875,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.118408203125,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.118408203125,
0.54052734375
],
[
0.51318359375,
0.52978515625,
0.88671875,
0.52978515625,
0.88671875,
0.54345703125,
0.51318359375,
0.54345703125
],
[
0.118408203125,
0.54345703125,
0.490234375,
0.54345703125,
0.490234375,
0.55712890625,
0.118408203125,
0.55712890625
],
[
0.51318359375,
0.5458984375,
0.884765625,
0.5458984375,
0.884765625,
0.55859375,
0.51318359375,
0.55859375
],
[
0.118408203125,
0.55859375,
0.48779296875,
0.55859375,
0.48779296875,
0.572265625,
0.118408203125,
0.572265625
],
[
0.51318359375,
0.5625,
0.884765625,
0.5625,
0.884765625,
0.576171875,
0.51318359375,
0.576171875
],
[
0.118408203125,
0.576171875,
0.489501953125,
0.576171875,
0.489501953125,
0.5888671875,
0.118408203125,
0.5888671875
],
[
0.51123046875,
0.57763671875,
0.88671875,
0.57666015625,
0.88671875,
0.59033203125,
0.51123046875,
0.59130859375
],
[
0.1158447265625,
0.59033203125,
0.48779296875,
0.59033203125,
0.48779296875,
0.60400390625,
0.1158447265625,
0.60400390625
],
[
0.118408203125,
0.60693359375,
0.48779296875,
0.60693359375,
0.48779296875,
0.62060546875,
0.118408203125,
0.62060546875
],
[
0.51318359375,
0.6015625,
0.88330078125,
0.6015625,
0.88330078125,
0.615234375,
0.51318359375,
0.615234375
],
[
0.118408203125,
0.62353515625,
0.359375,
0.62353515625,
0.359375,
0.63818359375,
0.118408203125,
0.63818359375
],
[
0.51318359375,
0.6181640625,
0.884765625,
0.6181640625,
0.884765625,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.51318359375,
0.63525390625,
0.884765625,
0.63525390625,
0.884765625,
0.6484375,
0.51318359375,
0.6484375
],
[
0.12005615234375,
0.65087890625,
0.29296875,
0.65087890625,
0.29296875,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.51318359375,
0.65087890625,
0.88330078125,
0.65087890625,
0.88330078125,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.51318359375,
0.66748046875,
0.884765625,
0.66748046875,
0.884765625,
0.68115234375,
0.51318359375,
0.68115234375
],
[
0.118408203125,
0.6767578125,
0.489501953125,
0.6767578125,
0.489501953125,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.5146484375,
0.68310546875,
0.88330078125,
0.68310546875,
0.88330078125,
0.69677734375,
0.5146484375,
0.69677734375
],
[
0.12005615234375,
0.693359375,
0.490234375,
0.693359375,
0.490234375,
0.70703125,
0.12005615234375,
0.70703125
],
[
0.51318359375,
0.69970703125,
0.88330078125,
0.69970703125,
0.88330078125,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.118408203125,
0.708984375,
0.490234375,
0.708984375,
0.490234375,
0.72265625,
0.118408203125,
0.72265625
],
[
0.51318359375,
0.71533203125,
0.88330078125,
0.71533203125,
0.88330078125,
0.728515625,
0.51318359375,
0.728515625
],
[
0.118408203125,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.118408203125,
0.73828125
],
[
0.51123046875,
0.73095703125,
0.88330078125,
0.73095703125,
0.88330078125,
0.74462890625,
0.51123046875,
0.74462890625
],
[
0.118408203125,
0.7412109375,
0.491943359375,
0.7412109375,
0.491943359375,
0.7548828125,
0.118408203125,
0.7548828125
],
[
0.51318359375,
0.74755859375,
0.884765625,
0.74755859375,
0.884765625,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.118408203125,
0.75830078125,
0.491943359375,
0.75830078125,
0.491943359375,
0.771484375,
0.118408203125,
0.771484375
],
[
0.5087890625,
0.7626953125,
0.884765625,
0.7626953125,
0.884765625,
0.775390625,
0.5087890625,
0.775390625
],
[
0.118408203125,
0.77392578125,
0.490234375,
0.77392578125,
0.490234375,
0.78759765625,
0.118408203125,
0.78759765625
],
[
0.51123046875,
0.77978515625,
0.88671875,
0.779296875,
0.88671875,
0.79345703125,
0.51123046875,
0.79443359375
],
[
0.1175537109375,
0.7880859375,
0.490234375,
0.7890625,
0.490234375,
0.80419921875,
0.1175537109375,
0.802734375
],
[
0.5146484375,
0.79736328125,
0.88671875,
0.79736328125,
0.88671875,
0.80810546875,
0.5146484375,
0.80810546875
],
[
0.118408203125,
0.80712890625,
0.274658203125,
0.80712890625,
0.274658203125,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51318359375,
0.8125,
0.88818359375,
0.8125,
0.88818359375,
0.826171875,
0.51318359375,
0.826171875
],
[
0.1175537109375,
0.830078125,
0.495361328125,
0.83056640625,
0.495361328125,
0.845703125,
0.1175537109375,
0.84423828125
],
[
0.51123046875,
0.82763671875,
0.88330078125,
0.82763671875,
0.88330078125,
0.84130859375,
0.51123046875,
0.84130859375
],
[
0.1175537109375,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.1175537109375,
0.86083984375
],
[
0.5146484375,
0.845703125,
0.88330078125,
0.845703125,
0.88330078125,
0.85546875,
0.5146484375,
0.85546875
],
[
0.118408203125,
0.86328125,
0.49365234375,
0.86328125,
0.49365234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.85986328125,
0.884765625,
0.85986328125,
0.884765625,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.1175537109375,
0.87744140625,
0.49365234375,
0.87890625,
0.49365234375,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.5322265625,
0.8818359375,
0.88671875,
0.8828125,
0.88671875,
0.89794921875,
0.5322265625,
0.896484375
],
[
0.118408203125,
0.8955078125,
0.49365234375,
0.8955078125,
0.49365234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.896484375,
0.83642578125,
0.896484375,
0.83642578125,
0.91015625,
0.51318359375,
0.91015625
]
] |
[
"tract relational tuples based on handcrafted rules",
"#Item",
"Metric",
"Average",
"Min~Max",
"or statistical methods. They usually rely on prior",
"Item",
"Doc",
"Nsent",
"101.78",
"44~138",
"syntactic or semantic analysis. Consequently, any",
"80",
"Lsent",
"22.69",
"5~47",
"800",
"Sent",
"Ntuple",
"2.65",
"1~8 ",
"model performance.",
"error accumulated in the prior stages deteriorates",
"Recently, neural OpenlE systems have been",
"Lsub",
"3.70",
"1~17",
"developed and showed promising results (Cui",
"2,122",
"Lrel",
"1~12",
"3.39",
"Tuple",
"et al., 2018; Zhan and Zhao, 2020; Kolluru et al.,",
"Lobj",
"3.94",
"0~27",
"2020a,b). Different from the traditional models,",
"neural OpenIE models extract tuples in an end-to-",
"Table 2: Statistics of DocOIE evaluation dataset. No",
"end manner, not requiring prior syntactic or seman-",
"denotes the number of units (i.e., sentence or tuple)",
"tic analysis. In principle, the traditional rule-based",
"L denotes the length (number of words) of the unit.",
"or statistical OpenIE models do not need training.",
"However.neural OpenlE modelsneed alargenum-",
"In total, 2, 122 relational tuples are annotated in",
"ber of training samples to learn the extraction pat-",
"the 800 sampled sentences (refer to Table 2 for",
"terns. For instance, IMOJIE (Kolluru et al., 2020b)",
"detailed statistics).",
"uses about 100,000 sentences for model training.",
"It is unrealistic and expensive to manually anno-",
"Training dataset contains 2,400 documents",
"tate 100,000 sentences simply for training purpose.",
"from the two domains (healthcare and transporta",
"Therefore, a common practice in learning a neural",
"tion); 1,200 documents in each domain. All sen-",
"OpenIE model is to use tuples automatically ex-",
"tences from these documents are used to bootstrap",
"tracted by traditional systems as training data, i.e.,",
"pseudo labels for neural model training.4",
"a bootstrapping strategy. We consider these im-",
"perfect training labels generated via bootstrapping",
"3.1 Dataset Collection",
"as pseudo labels. The pseudo labels used in (Cui",
"OpenIE, by definition, is to extract relational tuples",
"et al., 2018) are by Openie4, and those in (Kolluru",
"in open domain. Ideally, sentences/documents in",
"et al., 2020b) are from multiple OpenIE systems.",
"DocOIE dataset shall not be restricted to any par-",
"To the best of our knowledge, no OpenlE mod-",
"ticular document type or topical domain. However,",
"els consider document-level contexts in the tuple",
"it is challenging to include all types of documents",
"extraction. Nonetheless, our neural model un-",
"and annotate them. In fact, all existing annotations",
"avoidably requires extractions of pseudo labels",
"are restricted to specific types of documents like",
"bootstrapped from traditional models fortraining.",
"news and Wikipedia articles (Niklaus et al., 2018).",
"To ensure reproducibility, as part of the DocOIE",
"dataset, we also release the documents that are used",
"Document Type Selection In building DocOIE.",
"for generating the pseudo labels.",
"we focus on formally written documents and leave",
"it for future work to explore other kinds of doc-",
"3DocOIEDataset",
"uments. We select the type of formal documents",
"with four criteria: (i) Adequacy: as a document-",
"We now present our Document-level context-aware",
"level context-aware dataset, each document shall",
"Open Information Extraction (DocOIE) dataset.",
"have a reasonable number of sentences to provide",
"We first introduce the data selection and collection",
"sufficient context. (ii) Consistency: each docu-",
"process, and then the annotation process by two",
"ment shall focus on a central topic. In such a way",
"experts. Moreover, we explain our annotation con-",
"sentences within the same document are correlated",
"sistency measurement to indicate the high-level an-",
" to and consistent with each other, which helps de-",
"notation consistency in DocOIE. In summary, Do-",
"rive proper context. (ii) Informativeness: a doc-",
"cOIE consists of two datasets: evaluation dataset",
"ument is considered informativeif it contains in-",
"and training dataset.",
"formative entities like technical concepts, relations,",
"Evaluationdataset contains800expert-",
"and events. Intuitively, OpenIE models are more",
"annotated sentences, sampled from 80 documents",
"usefulforextractingfactual tuplesininformative",
"in two domains (healthcare and transportation).",
"documents. (iv) Syntactic Variety: sentences in",
"Specifically, 10 sentences are sampled for annota-",
"4Only document IDs are included in DocOIE, for docu-",
"tion from each of the 40 documents in one domain.",
"ment collection at http://pat ft.uspto.gov/"
] |
{
"arxiv_id": "2105.04271",
"language": "en",
"page_idx": 2
}
|
||
2105.04271-en-4
|
2105.04271-en-4.jpg
|
<table>
<tbody>
<tr>
<th>Consistency</th>
<th>Precision</th>
<th>$\mathbf{Recall}$</th>
<th>Fl</th>
</tr>
<tr>
<td>$A{\leftarrow}B$</td>
<td>90.7</td>
<td>92.4</td>
<td>91.6</td>
</tr>
<tr>
<td>$B\leftarrow A$</td>
<td>84.6</td>
<td>92.0</td>
<td>88.2</td>
</tr>
<tr>
<td>Average</td>
<td>87.7</td>
<td>92.2</td>
<td>89.9</td>
</tr>
</tbody>
</table>
Table 4: Annotation consistency estimated between annotators A and B. A< B indicates evaluation of A's annotations with B's annotations as ground truth.
truth. Among the tuple matching strategies in CaRB, we used the default binary lenient tuple matching, to estimate the consistency between the two annotators. Reported in Table 4, the two annotators reach high-level agreement in annotations with an average F1 of 89.9%.
Based on the high-level annotation consistency, in the third stage, each expert annotated 300 sentences from the remaining 600 sentences. The annotations are then validated by the other expert, and annotation disagreements are resolved through discussion.
## 3.4 DocOIE Training Dataset
Besides the 80 documents for expert annotations, we further sample 2,400 documents randomly (1,200 in each domain) from the documents collected in Section 3.1 to create DocOIE training dataset. The 1,200 documents in each domain contain around 120,000 sentences, which is sufficient for pseudo label generation, required by neural OpenIE models.
## 4 Pseudo Label by Bootstrapping
Following the common practice (Kolluru et al., 2020b; Cui et al., 2018), we generate pseudo labels by bootstrapping with traditional OpenIE models. Before we run these models on the DocOIE training dataset, we evaluate their performances on the DocOIE evaluation dataset, to select the models which can generate better quality pseudo labels.
We evaluate the models by using CaRB scorer (Bhardwaj et al., 2019). Table 5 reports the performance of five independent OpenIE models: Reverb (Fader et al., 2011), Clausie (Del Corro and Gemulla, 2013), Stanford OpenIE (Angeli et al., 2015), OpenIE4 (Mausam, 2016) and OpenIE5$^{6}.$ In addition to these five models, we also evaluated two combinations of Reverb and OpenIE4. With
$^{6}$github:dair-iitd/openie-standalone
<table>
<tbody>
<tr>
<th>OpenIEModel</th>
<th>$AUC$</th>
<th>Prec</th>
<th>Rec</th>
<th>Fl</th>
</tr>
<tr>
<td>Healthcare Dor</td>
<td>$la1n$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Reverb</td>
<td>35.4</td>
<td>79.9</td>
<td>42.8</td>
<td>55.8</td>
</tr>
<tr>
<td>Stanford</td>
<td>16.5</td>
<td>11.0</td>
<td>29.7</td>
<td>16.1 </td>
</tr>
<tr>
<td>Clausie</td>
<td>22.1</td>
<td>38.8</td>
<td>53.8</td>
<td>45.1 </td>
</tr>
<tr>
<td>OpenIE4</td>
<td>35.4</td>
<td>59.5</td>
<td>$\underline{55.1}$</td>
<td>57.2</td>
</tr>
<tr>
<td>OpenIE5</td>
<td>29.1</td>
<td>53.6</td>
<td>50.5</td>
<td>52.0</td>
</tr>
<tr>
<td>Rev+ Oie4</td>
<td>36.8</td>
<td>$\underline{75.8}$</td>
<td>47.7</td>
<td>58.6</td>
</tr>
<tr>
<td>Oie4+ Rev</td>
<td>$\underline{35.8}$</td>
<td>59.6</td>
<td>55.3</td>
<td>$\underline{57.4}$</td>
</tr>
<tr>
<td>Transportation</td>
<td>Oomain</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Reverb</td>
<td>29.3</td>
<td>79.1</td>
<td>36.3</td>
<td>49.7</td>
</tr>
<tr>
<td>Stanford</td>
<td>15.7</td>
<td>13.2</td>
<td>27.8</td>
<td>17.9</td>
</tr>
<tr>
<td>Clausie</td>
<td>18.0</td>
<td>36.2</td>
<td>48.4</td>
<td>41.4</td>
</tr>
<tr>
<td>OpenIE4</td>
<td>29.2</td>
<td>52.8</td>
<td>51.2</td>
<td>52.0</td>
</tr>
<tr>
<td>OpenIE5</td>
<td>25.0</td>
<td>50.9</td>
<td>43.8</td>
<td>47.1</td>
</tr>
<tr>
<td>Rev+ C -0ie4</td>
<td>31.0</td>
<td>$\underline{74.2}$</td>
<td>42.4</td>
<td>54.0</td>
</tr>
<tr>
<td>Oie4+ Rev</td>
<td>30.1</td>
<td>53.4</td>
<td>52.7</td>
<td>$\underline{53.0}$</td>
</tr>
</tbody>
</table>
Table 5: Performance of OpenIE models on DocOIE
evaluation dataset. The best scores are in boldface and
second best scores are underlined.
<table>
<tbody>
<tr>
<th>OpenIE model</th>
<th>Heal $\#Sent$</th>
<th>thcare $\#$Tuple</th>
<th>Transp $\#$Sent</th>
<th>oortation $\#$Tuple</th>
</tr>
<tr>
<td>OpenIE4</td>
<td>117k</td>
<td>$263k$</td>
<td>$1111k$</td>
<td>258k</td>
</tr>
<tr>
<td>Oie4+ Rev</td>
<td>$121k$</td>
<td>268k</td>
<td>114k</td>
<td>262k</td>
</tr>
<tr>
<td>Reverb</td>
<td>103k</td>
<td>$146k$</td>
<td>$97k$</td>
<td>$141k$</td>
</tr>
<tr>
<td>$\mathbf{Rev+0}$ $\cdot$Oie4 -</td>
<td>$121k$</td>
<td>$181k$</td>
<td>114k</td>
<td>173k</td>
</tr>
</tbody>
</table>
Table 6: Number of sentences and tuples extracted by Reverb, OpenIE4 and their combinations. The sentence is not included if it has no tuples extracted.
Rev+Oie4, Reverb is the main system and if Reverb fails to extract any tuples from a sentence, we complement the extraction by using Openie4. Similarly, Oie4+ Rev uses OpenIE4 as the main system, and the extractions are complemented by Reverb.
All the evaluated models show consistent performance in both domains. By F1 score, both Reverb and OpenIE4 are the best performing individual models and their combinations lead to the best and second best F1 scores in both domains. Accordingly, by applying OpenIE4, Reverb, and their combinations, the number of sentences and tuples extracted from the DocOIE training dataset are reported in Table 6. Note that, a sentence is not counted if it has no extracted tuples, which leads to the different sentence number.
|
[
[
0.15283203125,
0.0789794921875,
0.2476806640625,
0.0789794921875,
0.2476806640625,
0.09320068359375,
0.15283203125,
0.09320068359375
],
[
0.26025390625,
0.0789794921875,
0.337646484375,
0.0789794921875,
0.337646484375,
0.09320068359375,
0.26025390625,
0.09320068359375
],
[
0.349365234375,
0.0789794921875,
0.403076171875,
0.0789794921875,
0.403076171875,
0.09320068359375,
0.349365234375,
0.09320068359375
],
[
0.42236328125,
0.0789794921875,
0.447509765625,
0.0789794921875,
0.447509765625,
0.0926513671875,
0.42236328125,
0.0926513671875
],
[
0.5322265625,
0.0789794921875,
0.654296875,
0.0789794921875,
0.654296875,
0.0926513671875,
0.5322265625,
0.0926513671875
],
[
0.771484375,
0.07781982421875,
0.8095703125,
0.07781982421875,
0.8095703125,
0.09320068359375,
0.771484375,
0.09320068359375
],
[
0.82958984375,
0.0789794921875,
0.85400390625,
0.0789794921875,
0.85400390625,
0.0926513671875,
0.82958984375,
0.0926513671875
],
[
0.71875,
0.08013916015625,
0.75732421875,
0.08013916015625,
0.75732421875,
0.0926513671875,
0.71875,
0.0926513671875
],
[
0.6591796875,
0.08013916015625,
0.705078125,
0.08013916015625,
0.705078125,
0.0914306640625,
0.6591796875,
0.0914306640625
],
[
0.175537109375,
0.1009521484375,
0.22412109375,
0.1009521484375,
0.22412109375,
0.11517333984375,
0.175537109375,
0.11517333984375
],
[
0.27783203125,
0.1009521484375,
0.318115234375,
0.1009521484375,
0.318115234375,
0.11639404296875,
0.27783203125,
0.11639404296875
],
[
0.35693359375,
0.1015625,
0.39794921875,
0.1015625,
0.39794921875,
0.11517333984375,
0.35693359375,
0.11517333984375
],
[
0.41650390625,
0.1015625,
0.45751953125,
0.1015625,
0.45751953125,
0.11517333984375,
0.41650390625,
0.11517333984375
],
[
0.52880859375,
0.1015625,
0.67822265625,
0.1015625,
0.67822265625,
0.11517333984375,
0.52880859375,
0.11517333984375
],
[
0.175537109375,
0.1175537109375,
0.22412109375,
0.1175537109375,
0.22412109375,
0.133056640625,
0.175537109375,
0.133056640625
],
[
0.27783203125,
0.1175537109375,
0.31982421875,
0.1175537109375,
0.31982421875,
0.133056640625,
0.27783203125,
0.133056640625
],
[
0.355224609375,
0.11639404296875,
0.397216796875,
0.11639404296875,
0.397216796875,
0.1318359375,
0.355224609375,
0.1318359375
],
[
0.41650390625,
0.1175537109375,
0.456787109375,
0.1175537109375,
0.456787109375,
0.133056640625,
0.41650390625,
0.133056640625
],
[
0.6630859375,
0.12353515625,
0.70361328125,
0.12353515625,
0.70361328125,
0.1395263671875,
0.6630859375,
0.1395263671875
],
[
0.77099609375,
0.12353515625,
0.81005859375,
0.12353515625,
0.81005859375,
0.1395263671875,
0.77099609375,
0.1395263671875
],
[
0.82275390625,
0.12353515625,
0.8623046875,
0.12353515625,
0.8623046875,
0.1395263671875,
0.82275390625,
0.1395263671875
],
[
0.7197265625,
0.12469482421875,
0.75732421875,
0.12469482421875,
0.75732421875,
0.13720703125,
0.7197265625,
0.13720703125
],
[
0.55908203125,
0.12469482421875,
0.6171875,
0.12469482421875,
0.6171875,
0.1395263671875,
0.55908203125,
0.1395263671875
],
[
0.279541015625,
0.1395263671875,
0.318115234375,
0.1395263671875,
0.318115234375,
0.155029296875,
0.279541015625,
0.155029296875
],
[
0.7197265625,
0.1395263671875,
0.759765625,
0.1395263671875,
0.759765625,
0.155029296875,
0.7197265625,
0.155029296875
],
[
0.77099609375,
0.1395263671875,
0.81005859375,
0.1395263671875,
0.81005859375,
0.155029296875,
0.77099609375,
0.155029296875
],
[
0.82373046875,
0.1395263671875,
0.8623046875,
0.1395263671875,
0.8623046875,
0.155029296875,
0.82373046875,
0.155029296875
],
[
0.418212890625,
0.14013671875,
0.456787109375,
0.14013671875,
0.456787109375,
0.152587890625,
0.418212890625,
0.152587890625
],
[
0.358642578125,
0.14013671875,
0.3955078125,
0.14013671875,
0.3955078125,
0.152587890625,
0.358642578125,
0.152587890625
],
[
0.55517578125,
0.14013671875,
0.623046875,
0.14013671875,
0.623046875,
0.155029296875,
0.55517578125,
0.155029296875
],
[
0.66650390625,
0.141357421875,
0.7021484375,
0.141357421875,
0.7021484375,
0.15380859375,
0.66650390625,
0.15380859375
],
[
0.6630859375,
0.155029296875,
0.7021484375,
0.155029296875,
0.7021484375,
0.171630859375,
0.6630859375,
0.171630859375
],
[
0.1695556640625,
0.142578125,
0.231689453125,
0.142578125,
0.231689453125,
0.15380859375,
0.1695556640625,
0.15380859375
],
[
0.71728515625,
0.1561279296875,
0.75830078125,
0.1561279296875,
0.75830078125,
0.171630859375,
0.71728515625,
0.171630859375
],
[
0.77099609375,
0.1561279296875,
0.81005859375,
0.1561279296875,
0.81005859375,
0.171630859375,
0.77099609375,
0.171630859375
],
[
0.82275390625,
0.1561279296875,
0.861328125,
0.1561279296875,
0.861328125,
0.171630859375,
0.82275390625,
0.171630859375
],
[
0.118408203125,
0.17041015625,
0.489501953125,
0.17041015625,
0.489501953125,
0.18408203125,
0.118408203125,
0.18408203125
],
[
0.55908203125,
0.157958984375,
0.61865234375,
0.157958984375,
0.61865234375,
0.169189453125,
0.55908203125,
0.169189453125
],
[
0.6630859375,
0.171630859375,
0.70361328125,
0.171630859375,
0.70361328125,
0.18701171875,
0.6630859375,
0.18701171875
],
[
0.71875,
0.171630859375,
0.75830078125,
0.171630859375,
0.75830078125,
0.18701171875,
0.71875,
0.18701171875
],
[
0.77099609375,
0.171630859375,
0.8095703125,
0.171630859375,
0.8095703125,
0.18701171875,
0.77099609375,
0.18701171875
],
[
0.82373046875,
0.17333984375,
0.861328125,
0.17333984375,
0.861328125,
0.1859130859375,
0.82373046875,
0.1859130859375
],
[
0.1175537109375,
0.18408203125,
0.489501953125,
0.18408203125,
0.489501953125,
0.19775390625,
0.1175537109375,
0.19775390625
],
[
0.55322265625,
0.1728515625,
0.62451171875,
0.1728515625,
0.62451171875,
0.18701171875,
0.55322265625,
0.18701171875
],
[
0.6630859375,
0.18701171875,
0.7021484375,
0.18701171875,
0.7021484375,
0.203125,
0.6630859375,
0.203125
],
[
0.77099609375,
0.18701171875,
0.81005859375,
0.18701171875,
0.81005859375,
0.203125,
0.77099609375,
0.203125
],
[
0.82275390625,
0.18701171875,
0.8623046875,
0.18701171875,
0.8623046875,
0.203125,
0.82275390625,
0.203125
],
[
0.7197265625,
0.189453125,
0.75732421875,
0.189453125,
0.75732421875,
0.201904296875,
0.7197265625,
0.201904296875
],
[
0.55322265625,
0.188232421875,
0.62451171875,
0.188232421875,
0.62451171875,
0.203125,
0.55322265625,
0.203125
],
[
0.1175537109375,
0.1995849609375,
0.43408203125,
0.1995849609375,
0.43408203125,
0.213134765625,
0.1175537109375,
0.213134765625
],
[
0.71875,
0.203125,
0.75830078125,
0.203125,
0.75830078125,
0.2197265625,
0.71875,
0.2197265625
],
[
0.77099609375,
0.203125,
0.81005859375,
0.203125,
0.81005859375,
0.2197265625,
0.77099609375,
0.2197265625
],
[
0.6630859375,
0.2037353515625,
0.70361328125,
0.2037353515625,
0.70361328125,
0.2197265625,
0.6630859375,
0.2197265625
],
[
0.54833984375,
0.2037353515625,
0.62890625,
0.2037353515625,
0.62890625,
0.218505859375,
0.54833984375,
0.218505859375
],
[
0.82275390625,
0.2037353515625,
0.8623046875,
0.2037353515625,
0.8623046875,
0.2197265625,
0.82275390625,
0.2197265625
],
[
0.6630859375,
0.2197265625,
0.70361328125,
0.2197265625,
0.70361328125,
0.235107421875,
0.6630859375,
0.235107421875
],
[
0.71875,
0.2197265625,
0.759765625,
0.2197265625,
0.759765625,
0.235107421875,
0.71875,
0.235107421875
],
[
0.77099609375,
0.2197265625,
0.81005859375,
0.2197265625,
0.81005859375,
0.235107421875,
0.77099609375,
0.235107421875
],
[
0.82275390625,
0.2197265625,
0.86376953125,
0.2197265625,
0.86376953125,
0.235107421875,
0.82275390625,
0.235107421875
],
[
0.1175537109375,
0.2410888671875,
0.489501953125,
0.2410888671875,
0.489501953125,
0.254638671875,
0.1175537109375,
0.254638671875
],
[
0.5498046875,
0.221435546875,
0.62744140625,
0.221435546875,
0.62744140625,
0.2327880859375,
0.5498046875,
0.2327880859375
],
[
0.5322265625,
0.2410888671875,
0.7080078125,
0.2410888671875,
0.7080078125,
0.255859375,
0.5322265625,
0.255859375
],
[
0.118408203125,
0.25830078125,
0.489501953125,
0.25830078125,
0.489501953125,
0.271484375,
0.118408203125,
0.271484375
],
[
0.55908203125,
0.26318359375,
0.61865234375,
0.26318359375,
0.61865234375,
0.279052734375,
0.55908203125,
0.279052734375
],
[
0.6630859375,
0.26318359375,
0.70361328125,
0.26318359375,
0.70361328125,
0.279052734375,
0.6630859375,
0.279052734375
],
[
0.76904296875,
0.26318359375,
0.81005859375,
0.26318359375,
0.81005859375,
0.279052734375,
0.76904296875,
0.279052734375
],
[
0.82275390625,
0.26318359375,
0.8623046875,
0.26318359375,
0.8623046875,
0.279052734375,
0.82275390625,
0.279052734375
],
[
0.12005615234375,
0.275634765625,
0.48779296875,
0.275634765625,
0.48779296875,
0.2861328125,
0.12005615234375,
0.2861328125
],
[
0.7197265625,
0.26416015625,
0.755859375,
0.26416015625,
0.755859375,
0.276611328125,
0.7197265625,
0.276611328125
],
[
0.66650390625,
0.279052734375,
0.7021484375,
0.279052734375,
0.7021484375,
0.29443359375,
0.66650390625,
0.29443359375
],
[
0.7197265625,
0.279052734375,
0.759765625,
0.279052734375,
0.759765625,
0.29443359375,
0.7197265625,
0.29443359375
],
[
0.76904296875,
0.279052734375,
0.81005859375,
0.279052734375,
0.81005859375,
0.29443359375,
0.76904296875,
0.29443359375
],
[
0.82275390625,
0.279052734375,
0.8623046875,
0.279052734375,
0.8623046875,
0.29443359375,
0.82275390625,
0.29443359375
],
[
0.1158447265625,
0.289306640625,
0.489501953125,
0.2880859375,
0.489501953125,
0.302734375,
0.1158447265625,
0.303955078125
],
[
0.55517578125,
0.27978515625,
0.623046875,
0.27978515625,
0.623046875,
0.29443359375,
0.55517578125,
0.29443359375
],
[
0.5576171875,
0.295654296875,
0.62060546875,
0.295654296875,
0.62060546875,
0.31005859375,
0.5576171875,
0.31005859375
],
[
0.66650390625,
0.295654296875,
0.70361328125,
0.295654296875,
0.70361328125,
0.311279296875,
0.66650390625,
0.311279296875
],
[
0.7197265625,
0.29638671875,
0.75830078125,
0.29638671875,
0.75830078125,
0.308837890625,
0.7197265625,
0.308837890625
],
[
0.771484375,
0.29638671875,
0.81005859375,
0.29638671875,
0.81005859375,
0.308837890625,
0.771484375,
0.308837890625
],
[
0.82373046875,
0.29638671875,
0.861328125,
0.29638671875,
0.861328125,
0.308837890625,
0.82373046875,
0.308837890625
],
[
0.118408203125,
0.305908203125,
0.489501953125,
0.305908203125,
0.489501953125,
0.319580078125,
0.118408203125,
0.319580078125
],
[
0.6630859375,
0.311279296875,
0.7021484375,
0.311279296875,
0.7021484375,
0.32666015625,
0.6630859375,
0.32666015625
],
[
0.71875,
0.311279296875,
0.759765625,
0.311279296875,
0.759765625,
0.32666015625,
0.71875,
0.32666015625
],
[
0.77099609375,
0.311279296875,
0.81005859375,
0.311279296875,
0.81005859375,
0.32666015625,
0.77099609375,
0.32666015625
],
[
0.82275390625,
0.311279296875,
0.8623046875,
0.311279296875,
0.8623046875,
0.32666015625,
0.82275390625,
0.32666015625
],
[
0.1158447265625,
0.32177734375,
0.337646484375,
0.320556640625,
0.337646484375,
0.3349609375,
0.1158447265625,
0.336181640625
],
[
0.5517578125,
0.312255859375,
0.623046875,
0.31005859375,
0.62451171875,
0.325927734375,
0.55322265625,
0.327880859375
],
[
0.6630859375,
0.327880859375,
0.70361328125,
0.327880859375,
0.70361328125,
0.34326171875,
0.6630859375,
0.34326171875
],
[
0.71875,
0.327880859375,
0.75830078125,
0.327880859375,
0.75830078125,
0.34326171875,
0.71875,
0.34326171875
],
[
0.76904296875,
0.327880859375,
0.81201171875,
0.327880859375,
0.81201171875,
0.34326171875,
0.76904296875,
0.34326171875
],
[
0.82275390625,
0.327880859375,
0.8623046875,
0.327880859375,
0.8623046875,
0.34326171875,
0.82275390625,
0.34326171875
],
[
0.1368408203125,
0.339111328125,
0.489501953125,
0.339111328125,
0.489501953125,
0.352783203125,
0.1368408203125,
0.352783203125
],
[
0.5517578125,
0.328857421875,
0.623046875,
0.32666015625,
0.62451171875,
0.342529296875,
0.55322265625,
0.344482421875
],
[
0.6630859375,
0.34326171875,
0.7021484375,
0.34326171875,
0.7021484375,
0.359375,
0.6630859375,
0.359375
],
[
0.71875,
0.34326171875,
0.75830078125,
0.34326171875,
0.75830078125,
0.359375,
0.71875,
0.359375
],
[
0.82275390625,
0.34326171875,
0.8623046875,
0.34326171875,
0.8623046875,
0.359375,
0.82275390625,
0.359375
],
[
0.771484375,
0.344482421875,
0.8095703125,
0.344482421875,
0.8095703125,
0.35693359375,
0.771484375,
0.35693359375
],
[
0.54833984375,
0.344482421875,
0.62890625,
0.344482421875,
0.62890625,
0.359375,
0.54833984375,
0.359375
],
[
0.118408203125,
0.355224609375,
0.490234375,
0.355224609375,
0.490234375,
0.369384765625,
0.118408203125,
0.369384765625
],
[
0.6630859375,
0.359375,
0.7021484375,
0.359375,
0.7021484375,
0.3759765625,
0.6630859375,
0.3759765625
],
[
0.7197265625,
0.358154296875,
0.75830078125,
0.35986328125,
0.75732421875,
0.3759765625,
0.71728515625,
0.372314453125
],
[
0.77099609375,
0.359375,
0.8095703125,
0.359375,
0.8095703125,
0.374755859375,
0.77099609375,
0.374755859375
],
[
0.82373046875,
0.35986328125,
0.8623046875,
0.35986328125,
0.8623046875,
0.3759765625,
0.82373046875,
0.3759765625
],
[
0.1175537109375,
0.371826171875,
0.490234375,
0.371826171875,
0.490234375,
0.385498046875,
0.1175537109375,
0.385498046875
],
[
0.54833984375,
0.35986328125,
0.62744140625,
0.35986328125,
0.62744140625,
0.374755859375,
0.54833984375,
0.374755859375
],
[
0.1158447265625,
0.385986328125,
0.491943359375,
0.38720703125,
0.491943359375,
0.402099609375,
0.1158447265625,
0.40087890625
],
[
0.51123046875,
0.38427734375,
0.88330078125,
0.383056640625,
0.88330078125,
0.39794921875,
0.51123046875,
0.3984375
],
[
0.118408203125,
0.40380859375,
0.489501953125,
0.40380859375,
0.489501953125,
0.41748046875,
0.118408203125,
0.41748046875
],
[
0.5146484375,
0.40087890625,
0.88330078125,
0.40087890625,
0.88330078125,
0.410888671875,
0.5146484375,
0.410888671875
],
[
0.51318359375,
0.413330078125,
0.744140625,
0.413330078125,
0.744140625,
0.427001953125,
0.51318359375,
0.427001953125
],
[
0.118408203125,
0.42041015625,
0.202392578125,
0.42041015625,
0.202392578125,
0.431640625,
0.118408203125,
0.431640625
],
[
0.118408203125,
0.448974609375,
0.364501953125,
0.448974609375,
0.364501953125,
0.4619140625,
0.118408203125,
0.4619140625
],
[
0.53564453125,
0.4619140625,
0.59619140625,
0.4619140625,
0.59619140625,
0.476806640625,
0.53564453125,
0.476806640625
],
[
0.63818359375,
0.4619140625,
0.724609375,
0.4619140625,
0.724609375,
0.476806640625,
0.63818359375,
0.476806640625
],
[
0.75830078125,
0.4619140625,
0.873046875,
0.4619140625,
0.873046875,
0.476806640625,
0.75830078125,
0.476806640625
],
[
0.118408203125,
0.471435546875,
0.490234375,
0.471435546875,
0.490234375,
0.485107421875,
0.118408203125,
0.485107421875
],
[
0.5400390625,
0.47802734375,
0.591796875,
0.47802734375,
0.591796875,
0.4921875,
0.5400390625,
0.4921875
],
[
0.62109375,
0.47802734375,
0.6708984375,
0.47802734375,
0.6708984375,
0.4921875,
0.62109375,
0.4921875
],
[
0.755859375,
0.47802734375,
0.8037109375,
0.47802734375,
0.8037109375,
0.4921875,
0.755859375,
0.4921875
],
[
0.81689453125,
0.478515625,
0.875,
0.478515625,
0.875,
0.493408203125,
0.81689453125,
0.493408203125
],
[
0.1158447265625,
0.486328125,
0.489501953125,
0.48681640625,
0.489501953125,
0.501953125,
0.1158447265625,
0.50048828125
],
[
0.6826171875,
0.479736328125,
0.73779296875,
0.479736328125,
0.73779296875,
0.4912109375,
0.6826171875,
0.4912109375
],
[
0.62646484375,
0.49951171875,
0.6689453125,
0.49951171875,
0.6689453125,
0.515625,
0.62646484375,
0.515625
],
[
0.6962890625,
0.49951171875,
0.74072265625,
0.49951171875,
0.74072265625,
0.515625,
0.6962890625,
0.515625
],
[
0.76171875,
0.49951171875,
0.8037109375,
0.49951171875,
0.8037109375,
0.515625,
0.76171875,
0.515625
],
[
0.8310546875,
0.49951171875,
0.875,
0.49951171875,
0.875,
0.515625,
0.8310546875,
0.515625
],
[
0.52294921875,
0.50048828125,
0.59619140625,
0.50048828125,
0.59619140625,
0.515625,
0.52294921875,
0.515625
],
[
0.118408203125,
0.50439453125,
0.490234375,
0.50439453125,
0.490234375,
0.51708984375,
0.118408203125,
0.51708984375
],
[
0.5224609375,
0.5166015625,
0.60107421875,
0.5166015625,
0.60107421875,
0.53076171875,
0.5224609375,
0.53076171875
],
[
0.62646484375,
0.5166015625,
0.6689453125,
0.5166015625,
0.6689453125,
0.5322265625,
0.62646484375,
0.5322265625
],
[
0.6962890625,
0.5166015625,
0.73876953125,
0.5166015625,
0.73876953125,
0.5322265625,
0.6962890625,
0.5322265625
],
[
0.76171875,
0.5166015625,
0.8037109375,
0.5166015625,
0.8037109375,
0.5322265625,
0.76171875,
0.5322265625
],
[
0.82958984375,
0.5166015625,
0.875,
0.5166015625,
0.875,
0.5322265625,
0.82958984375,
0.5322265625
],
[
0.62646484375,
0.5322265625,
0.6689453125,
0.5322265625,
0.6689453125,
0.54736328125,
0.62646484375,
0.54736328125
],
[
0.118408203125,
0.5185546875,
0.489501953125,
0.51953125,
0.489501953125,
0.53369140625,
0.118408203125,
0.533203125
],
[
0.765625,
0.5322265625,
0.8037109375,
0.5322265625,
0.8037109375,
0.54736328125,
0.765625,
0.54736328125
],
[
0.8310546875,
0.5322265625,
0.875,
0.5322265625,
0.875,
0.54736328125,
0.8310546875,
0.54736328125
],
[
0.69873046875,
0.53369140625,
0.73779296875,
0.53369140625,
0.73779296875,
0.54638671875,
0.69873046875,
0.54638671875
],
[
0.118408203125,
0.53515625,
0.490234375,
0.53515625,
0.490234375,
0.548828125,
0.118408203125,
0.548828125
],
[
0.52294921875,
0.533203125,
0.58251953125,
0.533203125,
0.58251953125,
0.54736328125,
0.52294921875,
0.54736328125
],
[
0.52490234375,
0.5498046875,
0.603515625,
0.5498046875,
0.603515625,
0.56103515625,
0.52490234375,
0.56103515625
],
[
0.62744140625,
0.5498046875,
0.6689453125,
0.5498046875,
0.6689453125,
0.5634765625,
0.62744140625,
0.5634765625
],
[
0.69873046875,
0.5498046875,
0.74072265625,
0.5498046875,
0.74072265625,
0.5634765625,
0.69873046875,
0.5634765625
],
[
0.76171875,
0.54736328125,
0.8017578125,
0.54736328125,
0.8017578125,
0.5634765625,
0.76171875,
0.5634765625
],
[
0.8330078125,
0.5498046875,
0.875,
0.5498046875,
0.875,
0.5634765625,
0.8330078125,
0.5634765625
],
[
0.118408203125,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.118408203125,
0.5517578125,
0.489501953125,
0.5517578125,
0.489501953125,
0.5654296875,
0.118408203125,
0.5654296875
],
[
0.51123046875,
0.572265625,
0.88330078125,
0.57373046875,
0.88330078125,
0.58837890625,
0.51123046875,
0.58740234375
],
[
0.118408203125,
0.58447265625,
0.240966796875,
0.58447265625,
0.240966796875,
0.59716796875,
0.118408203125,
0.59716796875
],
[
0.51318359375,
0.5888671875,
0.884765625,
0.5888671875,
0.884765625,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.51318359375,
0.6025390625,
0.8037109375,
0.6025390625,
0.8037109375,
0.6162109375,
0.51318359375,
0.6162109375
],
[
0.1158447265625,
0.61328125,
0.420654296875,
0.615234375,
0.420654296875,
0.6298828125,
0.1158447265625,
0.62744140625
],
[
0.118408203125,
0.642578125,
0.491943359375,
0.642578125,
0.491943359375,
0.65673828125,
0.118408203125,
0.65673828125
],
[
0.5146484375,
0.6455078125,
0.884765625,
0.6455078125,
0.884765625,
0.66015625,
0.5146484375,
0.66015625
],
[
0.118408203125,
0.65673828125,
0.489501953125,
0.65771484375,
0.489501953125,
0.6728515625,
0.118408203125,
0.67138671875
],
[
0.51318359375,
0.662109375,
0.884765625,
0.662109375,
0.884765625,
0.67578125,
0.51318359375,
0.67578125
],
[
0.118408203125,
0.6748046875,
0.490234375,
0.6748046875,
0.490234375,
0.689453125,
0.118408203125,
0.689453125
],
[
0.51318359375,
0.6787109375,
0.884765625,
0.6787109375,
0.884765625,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.12005615234375,
0.6904296875,
0.491943359375,
0.6904296875,
0.491943359375,
0.70361328125,
0.12005615234375,
0.70361328125
],
[
0.51318359375,
0.69482421875,
0.884765625,
0.69482421875,
0.884765625,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.118408203125,
0.7060546875,
0.489501953125,
0.7060546875,
0.489501953125,
0.7197265625,
0.118408203125,
0.7197265625
],
[
0.5146484375,
0.7119140625,
0.8759765625,
0.7119140625,
0.8759765625,
0.72265625,
0.5146484375,
0.72265625
],
[
0.1175537109375,
0.7216796875,
0.489501953125,
0.72265625,
0.489501953125,
0.73681640625,
0.1175537109375,
0.736328125
],
[
0.118408203125,
0.7392578125,
0.470947265625,
0.7392578125,
0.470947265625,
0.7529296875,
0.118408203125,
0.7529296875
],
[
0.5341796875,
0.73388671875,
0.88671875,
0.73388671875,
0.88671875,
0.74755859375,
0.5341796875,
0.74755859375
],
[
0.1368408203125,
0.755859375,
0.489501953125,
0.755859375,
0.489501953125,
0.76953125,
0.1368408203125,
0.76953125
],
[
0.5146484375,
0.751953125,
0.884765625,
0.751953125,
0.884765625,
0.7626953125,
0.5146484375,
0.7626953125
],
[
0.1175537109375,
0.77099609375,
0.489501953125,
0.77099609375,
0.489501953125,
0.78515625,
0.1175537109375,
0.78515625
],
[
0.51318359375,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.51318359375,
0.7822265625,
0.884765625,
0.7822265625,
0.884765625,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.7880859375,
0.491943359375,
0.7880859375,
0.491943359375,
0.8017578125,
0.118408203125,
0.8017578125
],
[
0.118408203125,
0.80419921875,
0.489501953125,
0.80419921875,
0.489501953125,
0.81787109375,
0.118408203125,
0.81787109375
],
[
0.51318359375,
0.798828125,
0.884765625,
0.798828125,
0.884765625,
0.8125,
0.51318359375,
0.8125
],
[
0.51123046875,
0.81396484375,
0.884765625,
0.8134765625,
0.884765625,
0.82763671875,
0.51123046875,
0.8291015625
],
[
0.12005615234375,
0.82080078125,
0.490234375,
0.82080078125,
0.490234375,
0.83447265625,
0.12005615234375,
0.83447265625
],
[
0.118408203125,
0.8349609375,
0.490234375,
0.83447265625,
0.490234375,
0.8486328125,
0.118408203125,
0.849609375
],
[
0.51123046875,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.51123046875,
0.84423828125
],
[
0.51123046875,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.12005615234375,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.12005615234375,
0.8662109375
],
[
0.1175537109375,
0.8681640625,
0.489501953125,
0.8681640625,
0.489501953125,
0.8818359375,
0.1175537109375,
0.8818359375
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.14111328125,
0.8955078125,
0.459228515625,
0.8955078125,
0.459228515625,
0.9091796875,
0.14111328125,
0.9091796875
],
[
0.51318359375,
0.8955078125,
0.75732421875,
0.8955078125,
0.75732421875,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"Consistency",
"Precision",
"Recall",
"F1",
"OpenIE Model",
"Rec",
"F1",
"Prec",
"AUC",
"A←B",
"90.7",
"92.4",
"91.6",
"Healthcare Domain",
"B←A",
"84.6",
"92.0",
"88.2",
"35.4",
"42.8",
"55.8",
"79.9",
"Reverb",
"87.7",
"11.0",
"29.7",
"16.1",
"89.9",
"92.2",
"Stanford",
"16.5",
"22.1",
"Average",
"38.8",
"53.8",
"45.1",
"Table 4: Annotation consistency estimated between an-",
"Clausie",
"35.4",
"59.5",
"55.1",
"57.2",
"notators A and B. A+B indicates evaluation of A's an-",
"OpenlE4",
"29.1",
"50.5",
"52.0",
"53.6",
"OpenIE5",
"notations with B's annotations as ground truth.",
"75.8",
"47.7",
"36.8",
"Rev+Oie4",
"58.6",
"35.8",
"59.6",
"55.3",
"57.4",
"truth. Among the tuple matching strategies in",
"Oie4+Rev",
"Transportation Domain",
"CaRB, we used the default binary lenient tuple",
"Reverb",
"29.3",
"36.3",
"49.7",
"matching,toestimatethe consistencybetweenthe",
"79.1",
"15.7",
"13.2",
"27.8",
"17.9",
"two annotators. Reported in Table 4, the two an-",
"Stanford",
"Clausie",
"18.0",
"36.2",
"48.4",
"41.4",
"notators reach high-level agreement in annotations",
"29.2",
"52.8",
"51.2",
"52.0",
"with an average F1 of 89.9%.",
"OpenIE4",
"25.0",
"50.9",
"43.8",
"47.1",
"Based on the high-level annotation consistency,",
"OpenIE5",
"31.0",
"74.2",
"54.0",
"42.4",
"Rev+Oie4",
"in the third stage, each expert annotated 300 sen-",
"30.1",
"53.4",
"52.7",
"53.0",
"tences from the remaining 600 sentences. The an-",
"Oie4+Rev",
"notations are then validated by the other expert,",
"Table 5:Performance of OpenIE models on DocOIE",
"and annotation disagreements are resolved through",
"evaluation dataset.The best scores are in boldface and",
"second best scores are underlined.",
"discussion.",
"3.4DocOIE Training Dataset",
"OpenE",
"Healthcare",
"Transportation",
"Besides the 80 documents for expert annotations,",
"model",
"#Sent",
"#Sent",
"#Tuple",
"we further sample 2,400 documents randomly",
"#Tuple",
"117k",
"263k",
"111k",
"258k",
"OpenIE4",
"(1,200 in each domain) from the documents col-",
"Oie4+Rev",
"121k",
"268k",
"114k",
"262k",
"103k",
"lected in Section 3.1 to create DocOIE training",
"97k",
"141k",
"146k",
"dataset. The 1,200 documents in each domain con-",
"Reverb",
"Rev+Oie4",
"121k",
"181k",
"114k",
"173k",
"for pseudo label generation, required by neural",
"tain around 120,000 sentences, which is sufficient",
"Table 6: Number of sentences and tuples extracted by",
"OpenIE models.",
"Reverb, OpenIE4 and their combinations. The sentence",
"is not included if it has no tuples extracted.",
"4PseudoLabel byBootstrapping",
"Following the common practice (Kolluru et al.,",
"Rev+Oie4, Reverb is the main system and if Re-",
"2020b; Cui et al., 2018), we generate pseudo labels",
"verb fails to extract any tuples from a sentence, we",
"by bootstrapping with traditional OpenIE models.",
"complement the extraction by using Openie4. Simi-",
"Before we run these models on the DocOIE train-",
"larly, Oie4+ Rev uses OpenlE4 as the main system,",
"ing dataset, we evaluate their performances on the",
"andtheextractionsarecomplementedbyReverb.",
"DocOIE evaluation dataset, to select the models",
"which can generate better quality pseudo labels",
"All the evaluated models show consistent per-",
"We evaluate the models by using CaRB",
"formanceinbothdomains.ByFl score,bothRe-",
"scorer (Bhardwaj et al., 2019). Table 5 reports the",
"verb and OpenIE4 are the best performing indi-",
"vidual models and their combinations lead to the",
"performance of five independent OpenlE models:",
"Reverb (Fader et al., 2011), Clausie (Del Corro and",
"best and second best F1 scores in both domains",
"Accordingly, by applying OpenlE4, Reverb, and",
"Gemulla, 2013), Stanford OpenIE (Angeli et al.,",
"2015), OpenIE4 (Mausam, 2016) and OpenIE56.",
"their combinations, the number of sentences and",
"tuples extracted from the DocOIE training dataset",
"In addition to these five models, we also evaluated",
"two combinations of Reverb and OpenIE4. With",
"are reported in Table 6. Note that, a sentence is not",
"counted if it has no extracted tuples, which leads",
"6github:dair-iitd/openie-standalone",
"to the different sentence number."
] |
{
"arxiv_id": "2105.04271",
"language": "en",
"page_idx": 4
}
|
||
2106.04415-en-1
|
2106.04415-en-1.jpg
|
the time interval for the interest in daily necessities is about one month, while the time interval for the interest in Apple's products is longer, about one year. In summary, users' behavior for different types of items have different periodicities. (2) The interactivity between items is not explored effectively. These methods only model the correlation between adjacent items in the sequence, but do not consider the interactivity between items in the multi-interest extraction. In fact, multi-interest extraction can be viewed as the process of soft clustering between items, and the interactivity between items is effective for clustering tasks [Zhang et al.,2017], because items of the same type will learn similar representation through interaction. Thus, we argue that the time interval and interaction information between items in the user's sequence are more powerful to capture multi-interest representation.
To solve these problems, we propose a novel method, called PIMI, to explore Periodicity and Interactivity in MultiInterest framework for sequential recommendation. Firstly, we encode the time interval information between items in the sequence so that the periodicity information can be involved in the user's multi-interest representation, which can reflect the dynamic changes of the user behavior. Secondly, we design an ingenious graph structure. Previous GNN-based methods ignore the sequential information in the user behavior, our graph structure overcomes this shortcoming and captures the correlation between adjacent user behavior. What's more, the proposed graph structure can gather and scatter global and local item interactivity information with the virtual central node to improve the performance of multi-interest extraction. Finally, we obtain multi-interest representation for user based on the attention mechanism, which can be used to select candidate items and make recommendations. The main contributions of this work are summarized as follows:
$\cdot$ We incorporate the time interval information in the user
behavior sequence, which can model the periodicity of user's multiple interests and improve the quality of the user's representation.
· We design an innovative graph structure to capture the
global and local interactivity among items, and retain the sequential information at the same time, which can improve the quality of the item's representation.
·Our model PIMI achieves the state-of-the-art perfor-
mance on two real-world challenging datasets Amazon
and Taobao for the sequential recommendation.
## 2 $\textbf{Related Work}$
Sequential recommendation. Sequential recommendation systems are based on the user's behavior sequence to predict the next item that the user might be interested in. Many recent works about sequential recommendation focus on this problem. FPMC[Rendle et al., 2010] contains a common Markov chain and the normal matrix factorization model for sequential data. SDM[Lv et $al.,2019]$ combines user's longand short-term preferences to make recommendations, which models user's preferences based on LSTM and dense fullyconnected networks. Chorus[Wang et al., 2020] utilizes the information of the knowledge graph to model the relations
between items, and introduces temporal kernel functions for each item relation to better capture dynamic user demands. These methods give a single vector representation of the user based on behavior sequence, which is hard to reflect the realworld recommendation situation. Recently, MIND [Li et al., 2019] and ComiRec [ Cen $etal.,2020]$ attempt to use dynamic routing-based methods and attention-based methods to obtain multiple user's vectors to reflect multiple interests in the sequence. However, they do not explore the periodicity of multiple interests and the interactivity between items sufficiently, which are conducive to the extraction of multi-interest
Time information learning for recommendation. Time information is very important for recommendation. Most sequential recommendation methods sort items according to the timestamp of user's interactions, which implicitly uses time information. Few works attempt to model the time information in the sequence explicitly. For example, MTIN [Jiang $et$ $al.,2020]$ develops a parallel temporal mask network, which is able to learn multiple temporal information for recommendation. TiSASRec [Li et $\hat{al}.$, 2020] combines the advantages of absolute position and relative time interval encodings based on self-attention to predict future items.
Graph neural network. Graph embedding is to learn a mapping function which maps the nodes in a graph to lowdimensional latent ] representation [Zhou et al.,2018]. Some recent works utilize graph neural network [Scarselli et al., 2008] methods to obtain the representation of users and items, which can be used for recommendation tasks. For example, GATNE[Cen et al., 2019] supports both transductive and inductive embedding learning for attributed multiplex heterogeneous networks, which can learn the representation of users and items. However, GNNs essentially deal with the interactivity between nodes, they neglect the relationship between adjacent items in the user sequence.
Attention. The originality of attention mechanism can be traced back to decades ago in felds of computer vision [Xu et al., 2015]. It is also adapted to recommendation systems
and rather useful on real-world recommendation tasks. For instance, SASRec[Kang and McAuley, 2018] captures highorder dynamics in user behavior sequences based on the selfattention mechanism. GC-SAN [Xu et al., 2019] designs a multi-layer self-attention network to obtain contextualized non-local representation in sequence. CoSAN[Luo et al., 2020] proposes the collaborative self-attention network to learn the session representation by modeling the long-range dependencies between collaborative items
## 3 0ur Method
The existing sequential recommendation methods usually use a single vector to represent the user, it is hard to reflect user's multiple interests in real-world. Based on the above observation, we explore using multiple vectors to represent the user's multiple interests.
The recent multi-interest frameworks for sequential recommendation ignore two problems: the periodicity of user's interest and the interactivity between items in the sequence. We believe that the user's points of interest have different time
|
[
[
0.0882568359375,
0.0694580078125,
0.48779296875,
0.0694580078125,
0.48779296875,
0.08331298828125,
0.0882568359375,
0.08331298828125
],
[
0.51318359375,
0.0694580078125,
0.9111328125,
0.0694580078125,
0.9111328125,
0.08331298828125,
0.51318359375,
0.08331298828125
],
[
0.0849609375,
0.083984375,
0.486083984375,
0.083984375,
0.486083984375,
0.09783935546875,
0.0849609375,
0.09783935546875
],
[
0.5107421875,
0.08331298828125,
0.90869140625,
0.08331298828125,
0.90869140625,
0.09661865234375,
0.5107421875,
0.09661865234375
],
[
0.0849609375,
0.09783935546875,
0.484375,
0.09661865234375,
0.484375,
0.10986328125,
0.0849609375,
0.111083984375
],
[
0.51318359375,
0.09783935546875,
0.9111328125,
0.09783935546875,
0.9111328125,
0.111083984375,
0.51318359375,
0.111083984375
],
[
0.0882568359375,
0.11297607421875,
0.483642578125,
0.11297607421875,
0.483642578125,
0.12371826171875,
0.0882568359375,
0.12371826171875
],
[
0.5146484375,
0.11297607421875,
0.90869140625,
0.11297607421875,
0.90869140625,
0.12371826171875,
0.5146484375,
0.12371826171875
],
[
0.0849609375,
0.1256103515625,
0.486083984375,
0.1256103515625,
0.486083984375,
0.1395263671875,
0.0849609375,
0.1395263671875
],
[
0.51318359375,
0.1256103515625,
0.91015625,
0.1256103515625,
0.91015625,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.08660888671875,
0.1395263671875,
0.486083984375,
0.1395263671875,
0.486083984375,
0.15283203125,
0.08660888671875,
0.15283203125
],
[
0.51318359375,
0.1395263671875,
0.91259765625,
0.1395263671875,
0.91259765625,
0.15283203125,
0.51318359375,
0.15283203125
],
[
0.0849609375,
0.15283203125,
0.484375,
0.15283203125,
0.484375,
0.1666259765625,
0.0849609375,
0.1666259765625
],
[
0.51318359375,
0.1514892578125,
0.91259765625,
0.15283203125,
0.91259765625,
0.1666259765625,
0.51318359375,
0.1654052734375
],
[
0.08660888671875,
0.1685791015625,
0.486083984375,
0.1685791015625,
0.486083984375,
0.1787109375,
0.08660888671875,
0.1787109375
],
[
0.51318359375,
0.1685791015625,
0.91015625,
0.1685791015625,
0.91015625,
0.1787109375,
0.51318359375,
0.1787109375
],
[
0.0882568359375,
0.1817626953125,
0.486083984375,
0.1817626953125,
0.486083984375,
0.1925048828125,
0.0882568359375,
0.1925048828125
],
[
0.51123046875,
0.18115234375,
0.9111328125,
0.18115234375,
0.9111328125,
0.1944580078125,
0.51123046875,
0.1944580078125
],
[
0.0849609375,
0.1944580078125,
0.48779296875,
0.1944580078125,
0.48779296875,
0.2083740234375,
0.0849609375,
0.2083740234375
],
[
0.51318359375,
0.1944580078125,
0.9111328125,
0.1944580078125,
0.9111328125,
0.2083740234375,
0.51318359375,
0.2083740234375
],
[
0.0849609375,
0.2083740234375,
0.486083984375,
0.2083740234375,
0.486083984375,
0.2215576171875,
0.0849609375,
0.2215576171875
],
[
0.51318359375,
0.210205078125,
0.87158203125,
0.210205078125,
0.87158203125,
0.2203369140625,
0.51318359375,
0.2203369140625
],
[
0.08660888671875,
0.222900390625,
0.486083984375,
0.222900390625,
0.486083984375,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.5146484375,
0.2279052734375,
0.91259765625,
0.2279052734375,
0.91259765625,
0.2412109375,
0.5146484375,
0.2412109375
],
[
0.0849609375,
0.236083984375,
0.486083984375,
0.236083984375,
0.486083984375,
0.25,
0.0849609375,
0.25
],
[
0.51318359375,
0.242431640625,
0.91259765625,
0.242431640625,
0.91259765625,
0.255615234375,
0.51318359375,
0.255615234375
],
[
0.08660888671875,
0.251953125,
0.486083984375,
0.251953125,
0.486083984375,
0.261962890625,
0.08660888671875,
0.261962890625
],
[
0.51123046875,
0.255615234375,
0.91259765625,
0.255126953125,
0.91259765625,
0.26953125,
0.51123046875,
0.270751953125
],
[
0.0849609375,
0.264404296875,
0.4697265625,
0.264404296875,
0.4697265625,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.51318359375,
0.270751953125,
0.91015625,
0.270751953125,
0.91015625,
0.281005859375,
0.51318359375,
0.281005859375
],
[
0.10455322265625,
0.279052734375,
0.484375,
0.279052734375,
0.484375,
0.292236328125,
0.10455322265625,
0.292236328125
],
[
0.51123046875,
0.28271484375,
0.91015625,
0.28271484375,
0.91015625,
0.296630859375,
0.51123046875,
0.296630859375
],
[
0.0882568359375,
0.29345703125,
0.484375,
0.29345703125,
0.484375,
0.3037109375,
0.0882568359375,
0.3037109375
],
[
0.51318359375,
0.298583984375,
0.9140625,
0.298583984375,
0.9140625,
0.3125,
0.51318359375,
0.3125
],
[
0.08660888671875,
0.30615234375,
0.486083984375,
0.30615234375,
0.486083984375,
0.3193359375,
0.08660888671875,
0.3193359375
],
[
0.51123046875,
0.311279296875,
0.91259765625,
0.311279296875,
0.91259765625,
0.324462890625,
0.51123046875,
0.324462890625
],
[
0.08905029296875,
0.32080078125,
0.484375,
0.32080078125,
0.484375,
0.330810546875,
0.08905029296875,
0.330810546875
],
[
0.51123046875,
0.324462890625,
0.91259765625,
0.324462890625,
0.91259765625,
0.33837890625,
0.51123046875,
0.33837890625
],
[
0.0882568359375,
0.335205078125,
0.486083984375,
0.335205078125,
0.486083984375,
0.34521484375,
0.0882568359375,
0.34521484375
],
[
0.51123046875,
0.33837890625,
0.91015625,
0.33837890625,
0.91015625,
0.3515625,
0.51123046875,
0.3515625
],
[
0.0849609375,
0.347900390625,
0.486083984375,
0.347900390625,
0.486083984375,
0.361083984375,
0.0849609375,
0.361083984375
],
[
0.51318359375,
0.354248046875,
0.91259765625,
0.354248046875,
0.91259765625,
0.367431640625,
0.51318359375,
0.367431640625
],
[
0.0849609375,
0.361083984375,
0.484375,
0.361083984375,
0.484375,
0.375,
0.0849609375,
0.375
],
[
0.5146484375,
0.36865234375,
0.81787109375,
0.36865234375,
0.81787109375,
0.37890625,
0.5146484375,
0.37890625
],
[
0.08660888671875,
0.375,
0.486083984375,
0.375,
0.486083984375,
0.38818359375,
0.08660888671875,
0.38818359375
],
[
0.08660888671875,
0.389404296875,
0.486083984375,
0.389404296875,
0.486083984375,
0.40283203125,
0.08660888671875,
0.40283203125
],
[
0.5146484375,
0.386962890625,
0.91259765625,
0.386962890625,
0.91259765625,
0.40087890625,
0.5146484375,
0.40087890625
],
[
0.0849609375,
0.40283203125,
0.484375,
0.40283203125,
0.484375,
0.416748046875,
0.0849609375,
0.416748046875
],
[
0.51123046875,
0.40087890625,
0.9111328125,
0.399658203125,
0.9111328125,
0.4140625,
0.51123046875,
0.415283203125
],
[
0.0849609375,
0.416748046875,
0.486083984375,
0.416748046875,
0.486083984375,
0.429931640625,
0.0849609375,
0.429931640625
],
[
0.51318359375,
0.415283203125,
0.91259765625,
0.415283203125,
0.91259765625,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.08660888671875,
0.43115234375,
0.486083984375,
0.43115234375,
0.486083984375,
0.4443359375,
0.08660888671875,
0.4443359375
],
[
0.51123046875,
0.4287109375,
0.9111328125,
0.4287109375,
0.9111328125,
0.442626953125,
0.51123046875,
0.442626953125
],
[
0.0882568359375,
0.44580078125,
0.484375,
0.44580078125,
0.484375,
0.455810546875,
0.0882568359375,
0.455810546875
],
[
0.51318359375,
0.442626953125,
0.9111328125,
0.442626953125,
0.9111328125,
0.455810546875,
0.51318359375,
0.455810546875
],
[
0.0882568359375,
0.460205078125,
0.484375,
0.460205078125,
0.484375,
0.47021484375,
0.0882568359375,
0.47021484375
],
[
0.51318359375,
0.45703125,
0.91259765625,
0.45703125,
0.91259765625,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.0849609375,
0.4716796875,
0.486083984375,
0.4716796875,
0.486083984375,
0.48486328125,
0.0849609375,
0.48486328125
],
[
0.51318359375,
0.4697265625,
0.9111328125,
0.4697265625,
0.9111328125,
0.48291015625,
0.51318359375,
0.48291015625
],
[
0.08660888671875,
0.48486328125,
0.484375,
0.48486328125,
0.484375,
0.498779296875,
0.08660888671875,
0.498779296875
],
[
0.51123046875,
0.484130859375,
0.9111328125,
0.484130859375,
0.9111328125,
0.49755859375,
0.51123046875,
0.49755859375
],
[
0.0849609375,
0.5,
0.486083984375,
0.5,
0.486083984375,
0.51318359375,
0.0849609375,
0.51318359375
],
[
0.5146484375,
0.5,
0.9111328125,
0.5,
0.9111328125,
0.51025390625,
0.5146484375,
0.51025390625
],
[
0.0849609375,
0.51318359375,
0.441162109375,
0.51318359375,
0.441162109375,
0.5263671875,
0.0849609375,
0.5263671875
],
[
0.5146484375,
0.51318359375,
0.91015625,
0.51318359375,
0.91015625,
0.5234375,
0.5146484375,
0.5234375
],
[
0.51318359375,
0.52587890625,
0.9140625,
0.52587890625,
0.9140625,
0.5390625,
0.51318359375,
0.5390625
],
[
0.10540771484375,
0.53271484375,
0.484375,
0.53271484375,
0.484375,
0.546875,
0.10540771484375,
0.546875
],
[
0.51318359375,
0.54052734375,
0.74853515625,
0.54052734375,
0.74853515625,
0.5537109375,
0.51318359375,
0.5537109375
],
[
0.11767578125,
0.54736328125,
0.48779296875,
0.54736328125,
0.48779296875,
0.56103515625,
0.11767578125,
0.56103515625
],
[
0.515625,
0.56005859375,
0.59228515625,
0.56005859375,
0.59228515625,
0.5712890625,
0.515625,
0.5712890625
],
[
0.59326171875,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.59326171875,
0.5712890625
],
[
0.51123046875,
0.57275390625,
0.9111328125,
0.57275390625,
0.9111328125,
0.5859375,
0.51123046875,
0.5859375
],
[
0.11767578125,
0.56103515625,
0.48779296875,
0.56103515625,
0.48779296875,
0.57470703125,
0.11767578125,
0.57470703125
],
[
0.1168212890625,
0.57568359375,
0.261474609375,
0.57568359375,
0.261474609375,
0.5888671875,
0.1168212890625,
0.5888671875
],
[
0.10296630859375,
0.59228515625,
0.486083984375,
0.59326171875,
0.486083984375,
0.60791015625,
0.10296630859375,
0.60693359375
],
[
0.51318359375,
0.5869140625,
0.91259765625,
0.5869140625,
0.91259765625,
0.60107421875,
0.51318359375,
0.60107421875
],
[
0.51318359375,
0.6015625,
0.91259765625,
0.6015625,
0.91259765625,
0.61572265625,
0.51318359375,
0.61572265625
],
[
0.1192626953125,
0.609375,
0.486083984375,
0.609375,
0.486083984375,
0.62255859375,
0.1192626953125,
0.62255859375
],
[
0.51123046875,
0.61572265625,
0.91015625,
0.61572265625,
0.91015625,
0.62890625,
0.51123046875,
0.62890625
],
[
0.11767578125,
0.62255859375,
0.48779296875,
0.62255859375,
0.48779296875,
0.63623046875,
0.11767578125,
0.63623046875
],
[
0.5146484375,
0.6298828125,
0.90869140625,
0.6298828125,
0.90869140625,
0.64013671875,
0.5146484375,
0.64013671875
],
[
0.1192626953125,
0.63818359375,
0.43310546875,
0.63818359375,
0.43310546875,
0.6484375,
0.1192626953125,
0.6484375
],
[
0.51318359375,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.51318359375,
0.65576171875
],
[
0.10455322265625,
0.65576171875,
0.484375,
0.65576171875,
0.484375,
0.66943359375,
0.10455322265625,
0.66943359375
],
[
0.51123046875,
0.65576171875,
0.9111328125,
0.65576171875,
0.9111328125,
0.66943359375,
0.51123046875,
0.66943359375
],
[
0.11767578125,
0.67041015625,
0.48779296875,
0.67041015625,
0.48779296875,
0.6845703125,
0.11767578125,
0.6845703125
],
[
0.51318359375,
0.66943359375,
0.77392578125,
0.67041015625,
0.77392578125,
0.6845703125,
0.51318359375,
0.68310546875
],
[
0.7685546875,
0.671875,
0.91015625,
0.671875,
0.91015625,
0.681640625,
0.7685546875,
0.681640625
],
[
0.11767578125,
0.68310546875,
0.431396484375,
0.68310546875,
0.431396484375,
0.6962890625,
0.11767578125,
0.6962890625
],
[
0.51318359375,
0.6845703125,
0.9140625,
0.6845703125,
0.9140625,
0.69775390625,
0.51318359375,
0.69775390625
],
[
0.51123046875,
0.69775390625,
0.9111328125,
0.69775390625,
0.9111328125,
0.7109375,
0.51123046875,
0.7109375
],
[
0.1094970703125,
0.7138671875,
0.2344970703125,
0.7138671875,
0.2344970703125,
0.72900390625,
0.1094970703125,
0.72900390625
],
[
0.51318359375,
0.7119140625,
0.7958984375,
0.7119140625,
0.7958984375,
0.72607421875,
0.51318359375,
0.72607421875
],
[
0.108642578125,
0.7197265625,
0.11358642578125,
0.7197265625,
0.11358642578125,
0.724609375,
0.108642578125,
0.724609375
],
[
0.0882568359375,
0.73779296875,
0.301513671875,
0.73779296875,
0.301513671875,
0.74853515625,
0.0882568359375,
0.74853515625
],
[
0.08660888671875,
0.71533203125,
0.111083984375,
0.71533203125,
0.111083984375,
0.72802734375,
0.08660888671875,
0.72802734375
],
[
0.295654296875,
0.73779296875,
0.484375,
0.73779296875,
0.484375,
0.74853515625,
0.295654296875,
0.74853515625
],
[
0.51953125,
0.744140625,
0.52685546875,
0.75,
0.51953125,
0.7548828125,
0.51318359375,
0.75
],
[
0.53759765625,
0.7470703125,
0.54345703125,
0.7470703125,
0.54345703125,
0.75244140625,
0.53759765625,
0.75244140625
],
[
0.5390625,
0.7421875,
0.64892578125,
0.7421875,
0.64892578125,
0.7568359375,
0.5390625,
0.7568359375
],
[
0.08416748046875,
0.75,
0.48779296875,
0.75048828125,
0.48779296875,
0.76513671875,
0.08416748046875,
0.7646484375
],
[
0.0849609375,
0.76513671875,
0.486083984375,
0.76513671875,
0.486083984375,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.51318359375,
0.76318359375,
0.91259765625,
0.76318359375,
0.91259765625,
0.77783203125,
0.51318359375,
0.77783203125
],
[
0.08660888671875,
0.77978515625,
0.48779296875,
0.77978515625,
0.48779296875,
0.79052734375,
0.08660888671875,
0.79052734375
],
[
0.51318359375,
0.77978515625,
0.9111328125,
0.77978515625,
0.9111328125,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.0882568359375,
0.79248046875,
0.486083984375,
0.79248046875,
0.486083984375,
0.80615234375,
0.0882568359375,
0.80615234375
],
[
0.51318359375,
0.79248046875,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.51318359375,
0.80615234375
],
[
0.0849609375,
0.80615234375,
0.486083984375,
0.80615234375,
0.486083984375,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.51123046875,
0.80615234375,
0.91259765625,
0.8046875,
0.91259765625,
0.818359375,
0.51123046875,
0.8193359375
],
[
0.08660888671875,
0.8212890625,
0.484375,
0.8212890625,
0.484375,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.51318359375,
0.82080078125,
0.63232421875,
0.82080078125,
0.63232421875,
0.833984375,
0.51318359375,
0.833984375
],
[
0.0849609375,
0.833984375,
0.486083984375,
0.833984375,
0.486083984375,
0.84765625,
0.0849609375,
0.84765625
],
[
0.52783203125,
0.833984375,
0.91015625,
0.833984375,
0.91015625,
0.84765625,
0.52783203125,
0.84765625
],
[
0.08660888671875,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.5146484375,
0.849609375,
0.91015625,
0.849609375,
0.91015625,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.0849609375,
0.861328125,
0.48779296875,
0.861328125,
0.48779296875,
0.875,
0.0849609375,
0.875
],
[
0.51318359375,
0.861328125,
0.9140625,
0.861328125,
0.9140625,
0.875,
0.51318359375,
0.875
],
[
0.0849609375,
0.87548828125,
0.488525390625,
0.87548828125,
0.488525390625,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51123046875,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"the time interval for the interest in daily necessities is about",
"between items, and introduces temporal kernel functions for",
"one month, while the time interval for the interest in Apple's",
"each item relation to better capture dynamic user demands",
"products is longer, about one year.In summary,usersbe-",
"These methods give a single vector representation of the user",
"haviorfordifferenttypesof itemshavedifferentperiodici-",
"basedonbehaviorsequence,whichishardtoreflectthereal",
"ties. (2) The interactivity between items is not explored ef-",
"world recommendation situation. Recently, MIND [Li et al.",
"fectively. These methods only model the correlation between",
"2019] and ComiRec [Cen et al., 2020] attempt to use dynamic",
"adjacent items in the sequence, but do not consider the in-",
"routing-based methods and attention-based methods to obtain",
"teractivitybetween items in the multi-interest extraction.In",
"multipleuser's vectorstoreflectmultipleinterestsinthe se-",
"fact,multi-interestextraction canbe viewed as theprocess of",
"quence. However, they do not explore the periodicity of mul-",
"soft clustering between items, and the interactivity between",
"tiple interests and the interactivity between items sufficiently,",
"items is effective for clustering tasks [Zhang et al., 2017l, be-",
"whichareconducivetotheextraction ofmulti-interest.",
"cause items of the same type will learn similar representation",
"Time information learning for recommendation. Time",
"through interaction. Thus, we argue that the time interval and",
"information is very important for recommendation. Most se-",
"interactioninformationbetweenitemsintheuser's sequence",
"quential recommendation methods sort items according to the",
"are more powerful to capture multi-interest representation.",
"timestamp of user's interactions.which implicitly uses time",
"To solve these problems, we propose a novel method,",
"information. Few works attempt to model the time informa-",
"calledPIMI.toexplorePeriodicityandInteractivityinMulti-",
"tion in the sequence explicitly. For example, MTIN [Jiang et",
"Interest framework for sequential recommendation.Firstly,",
"al., 2020l develops a parallel temporal mask network, which",
"we encode the timeinterval informationbetween items in",
"is able to learn multiple temporal information for recommen-",
"the sequence sothat theperiodicityinformation can be in-",
"dation.TiSASRec [Li et al., 2020] combines the advan-",
"volved in the user's multi-interest representation, which can",
"tages of absolute position and relative time interval encodings",
"reflect the dynamic changes of the user behavior. Secondly",
"basedon self-attentiontopredictfutureitems.",
"we design an ingenious graph structure. Previous GNN-based",
"methods ignore the sequential information in the user behav-",
"Graph neural network. Graph embedding is to learn a",
"ior, our graph structure overcomes this shortcoming and cap-",
"mapping function which maps the nodes in a graph to low-",
"tures the correlation between adjacent user behavior. What's",
"dimensional latent J representation [Zhou et al., 2018]. Some",
"more, the proposed graph structure can gather and scatter",
"recent works utilize graph neural network [Scarselli et al.",
"globalandlocaliteminteractivityinformationwiththevirtual",
"2008l methods to obtain the representation of users and items,",
"centralnodetoimprovetheperformanceofmulti-interestex-",
"which can be used for recommendation tasks. For example,",
"traction.Finally, we obtain multi-interest representation for",
"GATNE [Cen et al., 2019] supports both transductive and in-",
"user based on the attention mechanism, which can be used to",
"ductive embedding learning for attributed multiplex heteroge-",
" select candidate items and make recommendations. The main",
"neous networks,whichcanlearntherepresentation of users",
"contributions of this work are summarized as follows:",
"anditems.However,GNNsessentiallydeal withtheinter-",
"activity between nodes, they neglect the relationship between",
": We incorporate the time interval information in the user",
"adjacent items in the user sequence.",
"behavior sequence, which can model the periodicity of",
"Attention.",
"The originality of attention mechanism can be",
"traced back to decades ago in fields of computer vision [Xu",
"user's multiple interests and improve the quality of the",
"user's representation.",
": We design an innovative graph structure to capture the",
"et al., 2015]. It is also adapted to recommendation systems",
"and rather useful on real-world recommendation tasks. For",
"global and local interactivity among items, and retain",
"instance, SASRec [Kang and McAuley, 2018] captures high-",
"the sequential information at the same time, which can",
"orderdynamicsinuserbehaviorseguencesbasedontheself",
"improvethequalityoftheitem'srepresentation.",
"attention mechanism. GC-SAN [Xu et al., 2019] designs",
". Our model PIMI achieves the state-of-the-art perfor-",
"a multi-layer self-attention network to obtain contextualized",
"mance on two real-world challenging datasets Amazon",
"non-local representation in sequence.",
"CoSAN[Luo et al..",
"and Taobao for the sequential recommendation.",
"2020l proposes the collaborative self-attention network to",
"learn the session representation by modeling the long-range",
"Related Work",
"dependencies between collaborative items.",
"",
"Sequentialrecommendation.",
"2",
"Sequentialrecommendation",
"",
"",
"Our Method",
" systems are based on the user's behavior sequence to predict",
"the next item that the user might be interested in. Many re-",
"The existing sequential recommendation methods usually use",
"centworks about seguentialrecommendationfocus on this",
"a single vectorto representtheuser,itis hard toreflectuser's",
"problem. FPMC [Rendle et al., 2010] contains a common",
"multiple interests in real-world. Based on the above observa-",
"Markov chain and the normal matrix factorization model for",
"tion, we explore using multiple vectors to represent the user's",
"sequential data.SDM[Lv et al.,2019] combines user'slong-",
"multiple interests.",
"and short-term preferences to make recommendations, which",
"The recent multi-interest frameworks for sequential recom-",
"models user's preferences based on LSTM and dense fully-",
"mendationignoretwoproblems:theperiodicityof user'sin-",
"connected networks. Chorus [Wang et al., 2020] utilizes the",
"terest and the interactivity between items in the sequence. We",
"information of the knowledge graph to model the relations",
"believe that the user's points of interest have different time"
] |
{
"arxiv_id": "2106.04415",
"language": "en",
"page_idx": 1
}
|
||
2106.04415-en-2
|
2106.04415-en-2.jpg
|
## 3.1 Problem Statement
Assume $U$ denote a set of users and $I$ denote all items. Each user $u\in U$ has a sequence of historical behavior in order. Given the user action history $S^u=(i_1^u,i_2^u,...,i_{|S^u|-1}^u),i_r^u\in I$ represents the $r$-th item in the sequence. The goal of sequential recommendation is to predict the next item that the user might be interested in.
<FigureHere>
Figure 2: The architecture of our proposed PIMI. The input of PIMI is the user's behavior sequence, which contains a series of item IDs. The item IDs are fed into the embedding layer to get the embedding representation of each item. Meanwhile, we construct a time interval matrix in the periodicity module and introduce the time information. We aggregate the embedding and time interval representation of each item and feed them into the interactivity module. The items' feature learned from the interactivity module are used to generate the user's multi-interest representation in the multi-interest extraction layer. Further more, the multi-interest vectors can be used to compute the sampled softmax loss through interest attention layer for training, or extract candidate items from the global items pool through interest clustering layer for testing.
## 3.2 Multi-Interest Framework
Embedding Layer
As shown in Figure 2, the input of PIMI is a user behavior sequence, which contains a series of item IDs representing the $\hat{\text{user's actions with items in time order. We convert the user}}$ behavior sequence $( i_{1}^{u}, i_{2}^{u}, . . . , i_{| S^{u}| - 1}^{u})$ into a fixed- length se- quence $s^{u}= ( i_{1}^{u}, i_{2}^{u}, . . . , i_{n}^{u}) $, where $n$ represents the maximum sequence length we consider. If the sequence length is greater than $n$,we truncate it and only take the most recent $n$ items, otherwise, we pad the sequence to a fixed length $n.$
We construct an embedding matrix $M^I\in\mathbb{R}^{|I|\times d}$ for all items, where $d$ is the dimension of embedding vector. The embedding look-up operation converts the IDs of items in the sequence into a unified low-dimension latent space. We can obtain its embedding:
$$E^I=[e_1,...,e_n]\in\mathbb{R}^{n\times d}$$
(1)
where $e_r\in\mathbb{R}^{1\times d}$ is the embedding of the $r$-th item.
Periodicity Module
Corresponding to the user's behavior sequence $s^u$, we can $\begin{array}{l}{\text{also obtain a time sequence }t^{u}=(t_{1}^{u},t_{2}^{u},...,t_{n}^{u}),\mathrm{which~con-}}\\{\text{tains the timestamp of each item in order. We only focus on}}\end{array}$ the relative length of the time interval in one user behavior sequence and model it as the relationship between any two items. Specifically, given a fixed-length time sequence $t^u=$ $(t_{1}^{u},t_{2}^{u},...,t_{n}^{u})$ of user $u$,the time interval $d_ab^{u}$ between item $a$ and item $b$ is defined as the number of days interacted by user $u$,where $d_ab^u\in\mathbb{N}.d_{ab}^u$ and $d_ba^u$ are equal according to this definition. We also set a threshold $p$ for the time interval to avoid sparse encoding, $d_{ab}^u=min\bar{(}p,d_{ab}^u).$ Hence, the time $\hat{\text{interval matrix }}M^{t}\in\mathbb{N}^{n\times n}$of a user sequence is:
$$M^t=\begin{bmatrix}d_{11}^u&d_{12}^u&...&d_{1n}^u\\d_{21}^u&d_{22}^u&...&d_{2n}^u\\...&...&...&...\\d_{n1}^u&d_{n2}^u&...&d_{nn}^u\end{bmatrix}$$
(2)
Similar to the items' embedding, time interval embedding matrix is $M^T\in\mathbb{R}^{n\times n\times d}.$ For each item in the sequence, we use the time-aware attention method to obtain the attention score matrix $A_1\in\mathbb{R}^{n\times n}$ of the time interval matrix:
$$A_1=softmax(M^TW_1)^\top $$
(3)
where $W_1\in\mathbb{R}^d$ is a trainable parameter. The superscript $T$ denotes the transpose of the matrix. The attention score matrix $A_1$ with size $n\times n$ represents the attention weight of each item for the time interval of other items in the sequence. When we sum up the embedding of time intervals according to the attention score, the broadcast mechanism in Python is used here, we can obtain a matrix representation $E^T\in\mathbb{R}^{n\times d}$ of items, which denotes the position of each item in the timeline of the overall sequence:
$$E^T=A_1M^T$$
(4)
Interactivity Module
After the embedding layer and the periodicity module, we
aggregate the embedding $E^I$ and time interval representation
periodicity, so adding time information into the multi-interest vectors improves the quality of user representation. Meanwhile, the interactivity between items in the sequence are effective in improving the quality of item representation.
In this section, we first formalize the sequential recommen-
dation problem. Then, we propose a novel method, to explore Periodicity and Interactivity in Multi-Interest framework for recommendation, called PIMI (as shown in Figure 2).
|
[
[
0.5107421875,
0.06756591796875,
0.75732421875,
0.06817626953125,
0.75732421875,
0.08331298828125,
0.5107421875,
0.08209228515625
],
[
0.16259765625,
0.0770263671875,
0.2132568359375,
0.0770263671875,
0.2132568359375,
0.091552734375,
0.16259765625,
0.091552734375
],
[
0.35302734375,
0.0738525390625,
0.410888671875,
0.0770263671875,
0.41015625,
0.0921630859375,
0.351318359375,
0.0889892578125
],
[
0.51318359375,
0.08709716796875,
0.64208984375,
0.08709716796875,
0.64208984375,
0.10101318359375,
0.51318359375,
0.10101318359375
],
[
0.08660888671875,
0.10162353515625,
0.19287109375,
0.0997314453125,
0.19287109375,
0.1142578125,
0.08660888671875,
0.11614990234375
],
[
0.371826171875,
0.0985107421875,
0.488525390625,
0.0985107421875,
0.488525390625,
0.11614990234375,
0.371826171875,
0.11614990234375
],
[
0.51318359375,
0.1029052734375,
0.91015625,
0.1029052734375,
0.91015625,
0.11614990234375,
0.51318359375,
0.11614990234375
],
[
0.12091064453125,
0.11614990234375,
0.250732421875,
0.11614990234375,
0.250732421875,
0.1300048828125,
0.12091064453125,
0.1300048828125
],
[
0.31787109375,
0.11871337890625,
0.439453125,
0.11871337890625,
0.439453125,
0.1287841796875,
0.31787109375,
0.1287841796875
],
[
0.51123046875,
0.117431640625,
0.91259765625,
0.117431640625,
0.91259765625,
0.1307373046875,
0.51123046875,
0.1307373046875
],
[
0.24267578125,
0.1331787109375,
0.250732421875,
0.1331787109375,
0.250732421875,
0.136962890625,
0.24267578125,
0.136962890625
],
[
0.51123046875,
0.1307373046875,
0.91259765625,
0.1307373046875,
0.91259765625,
0.14453125,
0.51123046875,
0.14453125
],
[
0.0906982421875,
0.1383056640625,
0.1298828125,
0.1363525390625,
0.131591796875,
0.1514892578125,
0.09234619140625,
0.154052734375
],
[
0.2034912109375,
0.141357421875,
0.2418212890625,
0.141357421875,
0.2418212890625,
0.15283203125,
0.2034912109375,
0.15283203125
],
[
0.324462890625,
0.141357421875,
0.363525390625,
0.1395263671875,
0.365234375,
0.154052734375,
0.325927734375,
0.1558837890625
],
[
0.5087890625,
0.1427001953125,
0.91259765625,
0.1427001953125,
0.91259765625,
0.1634521484375,
0.5087890625,
0.1634521484375
],
[
0.0931396484375,
0.15283203125,
0.134033203125,
0.15283203125,
0.134033203125,
0.1634521484375,
0.0931396484375,
0.1634521484375
],
[
0.1993408203125,
0.1558837890625,
0.363525390625,
0.1572265625,
0.363525390625,
0.1717529296875,
0.1993408203125,
0.17041015625
],
[
0.51318359375,
0.1610107421875,
0.91259765625,
0.1610107421875,
0.91259765625,
0.1749267578125,
0.51318359375,
0.1749267578125
],
[
0.191162109375,
0.1749267578125,
0.200927734375,
0.1749267578125,
0.200927734375,
0.1787109375,
0.191162109375,
0.1787109375
],
[
0.316162109375,
0.173583984375,
0.336669921875,
0.173583984375,
0.336669921875,
0.1817626953125,
0.316162109375,
0.1817626953125
],
[
0.38232421875,
0.1749267578125,
0.40283203125,
0.1749267578125,
0.40283203125,
0.1817626953125,
0.38232421875,
0.1817626953125
],
[
0.51123046875,
0.1749267578125,
0.91259765625,
0.1749267578125,
0.91259765625,
0.1881103515625,
0.51123046875,
0.1881103515625
],
[
0.4248046875,
0.175537109375,
0.455810546875,
0.175537109375,
0.455810546875,
0.179931640625,
0.4248046875,
0.179931640625
],
[
0.2548828125,
0.175537109375,
0.26708984375,
0.1767578125,
0.26708984375,
0.179931640625,
0.2548828125,
0.1787109375
],
[
0.111083984375,
0.1944580078125,
0.221435546875,
0.1944580078125,
0.221435546875,
0.2052001953125,
0.111083984375,
0.2052001953125
],
[
0.51123046875,
0.189453125,
0.9140625,
0.189453125,
0.9140625,
0.20263671875,
0.51123046875,
0.20263671875
],
[
0.295654296875,
0.20263671875,
0.396240234375,
0.20263671875,
0.396240234375,
0.2139892578125,
0.295654296875,
0.2139892578125
],
[
0.51123046875,
0.20263671875,
0.84716796875,
0.20263671875,
0.84716796875,
0.216552734375,
0.51123046875,
0.216552734375
],
[
0.1502685546875,
0.210205078125,
0.1666259765625,
0.210205078125,
0.1666259765625,
0.2139892578125,
0.1502685546875,
0.2139892578125
],
[
0.52783203125,
0.2139892578125,
0.9140625,
0.2139892578125,
0.9140625,
0.231689453125,
0.52783203125,
0.231689453125
],
[
0.11358642578125,
0.2247314453125,
0.125,
0.2247314453125,
0.125,
0.2291259765625,
0.11358642578125,
0.2291259765625
],
[
0.1953125,
0.2235107421875,
0.20751953125,
0.2235107421875,
0.20751953125,
0.2279052734375,
0.1953125,
0.2279052734375
],
[
0.51318359375,
0.23291015625,
0.9111328125,
0.23291015625,
0.9111328125,
0.24365234375,
0.51318359375,
0.24365234375
],
[
0.51123046875,
0.24560546875,
0.9140625,
0.24560546875,
0.9140625,
0.2587890625,
0.51123046875,
0.2587890625
],
[
0.51123046875,
0.2587890625,
0.9140625,
0.2587890625,
0.9140625,
0.2734375,
0.51123046875,
0.2734375
],
[
0.156005859375,
0.2734375,
0.16259765625,
0.276611328125,
0.16015625,
0.27978515625,
0.15283203125,
0.27783203125
],
[
0.5107421875,
0.271484375,
0.65673828125,
0.272705078125,
0.65673828125,
0.287353515625,
0.5107421875,
0.285888671875
],
[
0.625,
0.292236328125,
0.80126953125,
0.290283203125,
0.80126953125,
0.3125,
0.625,
0.314453125
],
[
0.8896484375,
0.29736328125,
0.9140625,
0.29736328125,
0.9140625,
0.311279296875,
0.8896484375,
0.311279296875
],
[
0.236083984375,
0.314453125,
0.330078125,
0.314453125,
0.330078125,
0.32568359375,
0.236083984375,
0.32568359375
],
[
0.51123046875,
0.3193359375,
0.8544921875,
0.3193359375,
0.8544921875,
0.337158203125,
0.51123046875,
0.337158203125
],
[
0.1732177734375,
0.333251953125,
0.1846923828125,
0.341552734375,
0.1732177734375,
0.349853515625,
0.16259765625,
0.341552734375
],
[
0.2115478515625,
0.32763671875,
0.2279052734375,
0.32763671875,
0.2279052734375,
0.347900390625,
0.2115478515625,
0.347900390625
],
[
0.26953125,
0.333984375,
0.28759765625,
0.333984375,
0.28759765625,
0.347900390625,
0.26953125,
0.347900390625
],
[
0.43701171875,
0.333251953125,
0.47802734375,
0.33837890625,
0.4755859375,
0.35107421875,
0.4345703125,
0.34521484375
],
[
0.330078125,
0.3359375,
0.343017578125,
0.3359375,
0.343017578125,
0.34521484375,
0.330078125,
0.34521484375
],
[
0.11767578125,
0.3359375,
0.1298828125,
0.3359375,
0.1298828125,
0.34521484375,
0.11767578125,
0.34521484375
],
[
0.39208984375,
0.3359375,
0.406005859375,
0.3359375,
0.406005859375,
0.34521484375,
0.39208984375,
0.34521484375
],
[
0.51318359375,
0.343505859375,
0.64892578125,
0.343505859375,
0.64892578125,
0.356689453125,
0.51318359375,
0.356689453125
],
[
0.2099609375,
0.349853515625,
0.22216796875,
0.349853515625,
0.22216796875,
0.363037109375,
0.2099609375,
0.363037109375
],
[
0.26318359375,
0.361083984375,
0.281005859375,
0.356689453125,
0.28515625,
0.367431640625,
0.26708984375,
0.371826171875
],
[
0.281982421875,
0.361083984375,
0.294189453125,
0.361083984375,
0.294189453125,
0.367431640625,
0.281982421875,
0.367431640625
],
[
0.51318359375,
0.35791015625,
0.91259765625,
0.35791015625,
0.91259765625,
0.37255859375,
0.51318359375,
0.37255859375
],
[
0.351318359375,
0.37744140625,
0.39453125,
0.38330078125,
0.39208984375,
0.393310546875,
0.348876953125,
0.38818359375
],
[
0.5107421875,
0.371826171875,
0.9111328125,
0.37255859375,
0.9111328125,
0.386962890625,
0.5107421875,
0.386474609375
],
[
0.1298828125,
0.37890625,
0.1707763671875,
0.37890625,
0.1707763671875,
0.389404296875,
0.1298828125,
0.389404296875
],
[
0.2296142578125,
0.37890625,
0.273681640625,
0.380615234375,
0.27294921875,
0.392578125,
0.2279052734375,
0.39013671875
],
[
0.289306640625,
0.37890625,
0.334228515625,
0.380615234375,
0.33251953125,
0.392578125,
0.28759765625,
0.39013671875
],
[
0.205078125,
0.386962890625,
0.2197265625,
0.389404296875,
0.2181396484375,
0.396484375,
0.2017822265625,
0.39453125
],
[
0.1829833984375,
0.380615234375,
0.20751953125,
0.380615234375,
0.20751953125,
0.38818359375,
0.1829833984375,
0.38818359375
],
[
0.265625,
0.38818359375,
0.289306640625,
0.38818359375,
0.289306640625,
0.396484375,
0.265625,
0.396484375
],
[
0.5146484375,
0.38818359375,
0.91259765625,
0.38818359375,
0.91259765625,
0.3984375,
0.5146484375,
0.3984375
],
[
0.330078125,
0.389404296875,
0.351318359375,
0.389404296875,
0.351318359375,
0.396484375,
0.330078125,
0.396484375
],
[
0.11358642578125,
0.389404296875,
0.1422119140625,
0.389404296875,
0.1422119140625,
0.397705078125,
0.11358642578125,
0.397705078125
],
[
0.1568603515625,
0.39013671875,
0.1968994140625,
0.39013671875,
0.1968994140625,
0.397705078125,
0.1568603515625,
0.397705078125
],
[
0.38818359375,
0.389404296875,
0.406982421875,
0.389404296875,
0.406982421875,
0.397705078125,
0.38818359375,
0.397705078125
],
[
0.51318359375,
0.40087890625,
0.91259765625,
0.40087890625,
0.91259765625,
0.4140625,
0.51318359375,
0.4140625
],
[
0.0849609375,
0.417236328125,
0.48779296875,
0.417236328125,
0.48779296875,
0.43115234375,
0.0849609375,
0.43115234375
],
[
0.5107421875,
0.415283203125,
0.91259765625,
0.4140625,
0.91259765625,
0.427978515625,
0.5107421875,
0.4287109375
],
[
0.0849609375,
0.429931640625,
0.486083984375,
0.429931640625,
0.486083984375,
0.443115234375,
0.0849609375,
0.443115234375
],
[
0.51123046875,
0.4287109375,
0.90625,
0.4287109375,
0.90625,
0.442626953125,
0.51123046875,
0.442626953125
],
[
0.0849609375,
0.443115234375,
0.48779296875,
0.443115234375,
0.48779296875,
0.45703125,
0.0849609375,
0.45703125
],
[
0.51318359375,
0.442626953125,
0.91259765625,
0.442626953125,
0.91259765625,
0.45703125,
0.51318359375,
0.45703125
],
[
0.08660888671875,
0.45703125,
0.484375,
0.45703125,
0.484375,
0.46728515625,
0.08660888671875,
0.46728515625
],
[
0.5146484375,
0.458251953125,
0.9111328125,
0.458251953125,
0.9111328125,
0.468505859375,
0.5146484375,
0.468505859375
],
[
0.0849609375,
0.468505859375,
0.486083984375,
0.468505859375,
0.486083984375,
0.481689453125,
0.0849609375,
0.481689453125
],
[
0.51318359375,
0.47021484375,
0.91259765625,
0.47021484375,
0.91259765625,
0.484130859375,
0.51318359375,
0.484130859375
],
[
0.0849609375,
0.48095703125,
0.48779296875,
0.48095703125,
0.48779296875,
0.494384765625,
0.0849609375,
0.494384765625
],
[
0.51123046875,
0.484130859375,
0.9140625,
0.484130859375,
0.9140625,
0.49755859375,
0.51123046875,
0.49755859375
],
[
0.0849609375,
0.49365234375,
0.484375,
0.49365234375,
0.484375,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.51123046875,
0.498779296875,
0.9140625,
0.498779296875,
0.9140625,
0.51220703125,
0.51123046875,
0.51220703125
],
[
0.0849609375,
0.505859375,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.0849609375,
0.51953125
],
[
0.5107421875,
0.50830078125,
0.8408203125,
0.5087890625,
0.8408203125,
0.5263671875,
0.5107421875,
0.52587890625
],
[
0.0849609375,
0.51806640625,
0.483642578125,
0.51806640625,
0.483642578125,
0.5322265625,
0.0849609375,
0.5322265625
],
[
0.08416748046875,
0.53076171875,
0.48779296875,
0.53076171875,
0.48779296875,
0.54443359375,
0.08416748046875,
0.54443359375
],
[
0.6611328125,
0.53271484375,
0.69384765625,
0.53271484375,
0.69384765625,
0.548828125,
0.6611328125,
0.548828125
],
[
0.70361328125,
0.53271484375,
0.771484375,
0.53515625,
0.771484375,
0.5498046875,
0.70361328125,
0.54736328125
],
[
0.765625,
0.52978515625,
0.80810546875,
0.5341796875,
0.8056640625,
0.55029296875,
0.76318359375,
0.546875
],
[
0.0849609375,
0.54345703125,
0.486083984375,
0.54345703125,
0.486083984375,
0.556640625,
0.0849609375,
0.556640625
],
[
0.66845703125,
0.546875,
0.69384765625,
0.5517578125,
0.689453125,
0.564453125,
0.6640625,
0.5595703125
],
[
0.70361328125,
0.54736328125,
0.76953125,
0.54736328125,
0.76953125,
0.5625,
0.70361328125,
0.5625
],
[
0.76708984375,
0.54345703125,
0.8037109375,
0.5498046875,
0.7998046875,
0.56494140625,
0.76318359375,
0.5595703125
],
[
0.609375,
0.55322265625,
0.640625,
0.55322265625,
0.640625,
0.568359375,
0.609375,
0.568359375
],
[
0.88818359375,
0.5537109375,
0.9150390625,
0.5537109375,
0.9150390625,
0.57080078125,
0.88818359375,
0.57080078125
],
[
0.08416748046875,
0.5673828125,
0.343017578125,
0.5693359375,
0.343017578125,
0.58349609375,
0.08416748046875,
0.58056640625
],
[
0.0849609375,
0.55615234375,
0.48779296875,
0.55615234375,
0.48779296875,
0.5693359375,
0.0849609375,
0.5693359375
],
[
0.66015625,
0.57373046875,
0.8095703125,
0.57373046875,
0.8095703125,
0.59130859375,
0.66015625,
0.59130859375
],
[
0.529296875,
0.5986328125,
0.91259765625,
0.5986328125,
0.91259765625,
0.6123046875,
0.529296875,
0.6123046875
],
[
0.5107421875,
0.609375,
0.9150390625,
0.609375,
0.9150390625,
0.626953125,
0.5107421875,
0.626953125
],
[
0.0849609375,
0.61865234375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.0849609375,
0.6318359375
],
[
0.08660888671875,
0.6318359375,
0.484375,
0.6318359375,
0.484375,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.51123046875,
0.626953125,
0.9140625,
0.626953125,
0.9140625,
0.64013671875,
0.51123046875,
0.64013671875
],
[
0.5107421875,
0.63818359375,
0.85693359375,
0.63818359375,
0.85693359375,
0.65283203125,
0.5107421875,
0.65283203125
],
[
0.0882568359375,
0.64794921875,
0.484375,
0.64794921875,
0.484375,
0.65771484375,
0.0882568359375,
0.65771484375
],
[
0.0849609375,
0.66015625,
0.443603515625,
0.66015625,
0.443603515625,
0.673828125,
0.0849609375,
0.673828125
],
[
0.62158203125,
0.6611328125,
0.7998046875,
0.6591796875,
0.7998046875,
0.6767578125,
0.62158203125,
0.6787109375
],
[
0.890625,
0.66357421875,
0.9140625,
0.66357421875,
0.9140625,
0.6767578125,
0.890625,
0.6767578125
],
[
0.10296630859375,
0.6796875,
0.484375,
0.6796875,
0.484375,
0.693359375,
0.10296630859375,
0.693359375
],
[
0.5107421875,
0.68505859375,
0.9140625,
0.6875,
0.9140625,
0.7041015625,
0.5107421875,
0.7021484375
],
[
0.08660888671875,
0.6923828125,
0.486083984375,
0.6923828125,
0.486083984375,
0.70556640625,
0.08660888671875,
0.70556640625
],
[
0.08660888671875,
0.70556640625,
0.484375,
0.70556640625,
0.484375,
0.7197265625,
0.08660888671875,
0.7197265625
],
[
0.5146484375,
0.70068359375,
0.91259765625,
0.70263671875,
0.91259765625,
0.71630859375,
0.5146484375,
0.7138671875
],
[
0.08416748046875,
0.72021484375,
0.435546875,
0.72021484375,
0.435546875,
0.734375,
0.08416748046875,
0.734375
],
[
0.51123046875,
0.71630859375,
0.9140625,
0.71630859375,
0.9140625,
0.72998046875,
0.51123046875,
0.72998046875
],
[
0.5107421875,
0.72900390625,
0.9111328125,
0.72998046875,
0.9111328125,
0.74365234375,
0.5107421875,
0.7421875
],
[
0.51318359375,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.51318359375,
0.7568359375
],
[
0.51123046875,
0.7568359375,
0.91259765625,
0.7568359375,
0.91259765625,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.0882568359375,
0.7666015625,
0.12744140625,
0.7666015625,
0.12744140625,
0.77783203125,
0.0882568359375,
0.77783203125
],
[
0.12091064453125,
0.76513671875,
0.27685546875,
0.76513671875,
0.27685546875,
0.77880859375,
0.12091064453125,
0.77880859375
],
[
0.5087890625,
0.76953125,
0.9140625,
0.767578125,
0.9140625,
0.78515625,
0.5087890625,
0.787109375
],
[
0.51123046875,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.51123046875,
0.798828125
],
[
0.0882568359375,
0.8037109375,
0.484375,
0.8037109375,
0.484375,
0.8173828125,
0.0882568359375,
0.8173828125
],
[
0.51318359375,
0.7998046875,
0.69921875,
0.7998046875,
0.69921875,
0.81298828125,
0.51318359375,
0.81298828125
],
[
0.08660888671875,
0.818359375,
0.484375,
0.818359375,
0.484375,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.6650390625,
0.8193359375,
0.7587890625,
0.8193359375,
0.7587890625,
0.8369140625,
0.6650390625,
0.8369140625
],
[
0.890625,
0.82275390625,
0.9140625,
0.82275390625,
0.9140625,
0.8369140625,
0.890625,
0.8369140625
],
[
0.08416748046875,
0.83056640625,
0.486083984375,
0.83056640625,
0.486083984375,
0.8486328125,
0.08416748046875,
0.8486328125
],
[
0.0849609375,
0.84765625,
0.486083984375,
0.84765625,
0.486083984375,
0.861328125,
0.0849609375,
0.861328125
],
[
0.51123046875,
0.84521484375,
0.6611328125,
0.84521484375,
0.6611328125,
0.859375,
0.51123046875,
0.859375
],
[
0.08660888671875,
0.8623046875,
0.488525390625,
0.8623046875,
0.488525390625,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51318359375,
0.861328125,
0.9140625,
0.861328125,
0.9140625,
0.875,
0.51318359375,
0.875
],
[
0.0849609375,
0.87548828125,
0.2344970703125,
0.87548828125,
0.2344970703125,
0.8896484375,
0.0849609375,
0.8896484375
],
[
0.51318359375,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51318359375,
0.8896484375
]
] |
[
"3.2Multi-InterestFramework",
"Testing",
"Training",
"Embedding Layer",
"Recommended Items",
"Isampled Sofimax Loss",
"As shown in Figure 2, the input of PIMI is a user behavior se-",
"Interest Clustering Layer",
"InterestAttentionLayer",
"quence, which contains a series of item IDs representing the",
"",
"user's actions with items in time order. We convert the user",
"Ttens",
"",
"",
"behavior sequence (i, i.., i's-) into a fixed-length se-",
"pool",
"Multi-Interest Extraction Layer",
"quence s\" - (i\", i2, .., iu), where n represents the maximum",
"",
"一",
"",
" sequence length we consider. If the sequence length is greater",
"",
"",
"InteractivityModule",
"than n, we truncate it and only take the most recent n items,",
"Periodicity Module",
"otherwise, we pad the sequence to a fixed length n.",
"",
"We construct an embedding matrix Ml e IRIlxd for all",
"1",
"1",
"items,where d is the dimension of embedding vector.The",
"embedding look-up operation converts the IDs of items in the",
"sequence into a unified low-dimension latent space. We can",
"1",
"obtain its embedding:",
"E' - [e,..en] e Rnxd",
"(1)",
"Embedding Layer",
"where er E RIxd is the embedding of the r-th item.",
"12",
"+.53",
"i4",
"itarget",
"is",
"ii",
"i6",
"Periodicity Module",
"百",
"Q",
"",
"Corresponding to the user's behavior sequence s\", we can",
"20days",
"also obtain a time sequence t = (t', t, .., tu), which con-",
"10days",
"20days",
"10days",
"T",
"Oda",
"T",
"tainsthetimestampof eachitem in order.Weonlyfocus on",
"T-",
"TK",
"下",
"HT6",
"the relative length of the time interval in one user behavior",
"Figure 2: The architecture of our proposed PIMI. The input of PIMI",
"sequence and model it as the relationship between any two",
"is the user's behavior sequence, which contains a series of item IDs.",
"items. Specifically, given a fixed-length time sequence tu",
"The item IDs are fed into the embedding layer to get the embedding",
"(t', t?, ., t) of user u, the time interval da, between item a",
"representation of each item.Meanwhile.weconstructatime inter-",
"anditembisdefinedasthenumberofdaysinteractedbyuser",
"val matrix in the periodicity module and introduce the time informa-",
"u, where dab E N. dab and doa are equal according to this",
"tion. We aggregate the embedding and time interval representation",
"definition. We also set a threshold p for the time interval to",
"of each item and feed them into the interactivity module. The items",
"avoid sparse encoding, dab = min(p, dab). Hence, the time",
"feature learned from the interactivity module are used to generate",
"interval matrix Mt e Nnxn of a user sequence is:",
"the user's multi-interest representation in the multi-interest extrac-",
"tion layer. Further more, the multi-interest vectors can be used to",
"[d",
"di2.",
"din",
"compute the sampled softmax loss through interest attention layer",
"d21",
"d22.",
"den",
"Mt",
"(2)",
"through interest clustering layer for testing.",
"for training, or extract candidate items from the global items pool",
"[dni d2 .. dnn]",
"Similar to the items' embedding, time interval embedding",
"matrix is MT e Rnxnxd, For each item in the sequence, we",
"periodicity,so adding time information into the multi-interest",
"vectors improves the quality of user representation. Mean-",
"use the time-aware attention method to obtain the attention",
" score matrix A, e IRnxn of the time interval matrix:",
"while,the interactivitybetweenitems inthe sequenceareef-",
"fective in improving the quality of item representation.",
"Ai = softmar(MTWi)T",
"(3)",
"In this section, we first formalize the sequential recommen-",
"where Wi e Rd is a trainable parameter. The superscript",
"dation problem. Then, we propose a novel method, to explore",
"Periodicity and Interactivity in Multi-Interest framework for",
"T denotes the transpose of the matrix.The attention score",
"recommendation, called PIMI (as shown in Figure 2)",
"matrix Ar with size n × n represents the attention weight of",
"each itemfor the time interval of other items in the sequence.",
"When we sum up the embedding of time intervals according",
"to the attention score, the broadcast mechanism in Python is",
"3.1",
" Problem Statement",
"used here, we can obtain a matrix representation ET e Rnxd",
"of items, which denotes the position of each item in the time-",
"Assume U denote a set of users and I denote all items. Each",
"line of the overall sequence:",
"user u e U has a sequence of historical behavior in order.",
"ET =AIMT",
"(4)",
"Given the user action history S\" = (, 2, ., isuj-1), u e I",
"represents the r-th item in the sequence. The goal of sequen-",
"Interactivity Module",
"tial recommendation is to predict the next item that the user",
"After the embedding layer and the periodicity module, we",
"might be interested in.",
"aggregate the embedding E and time interval representation"
] |
{
"arxiv_id": "2106.04415",
"language": "en",
"page_idx": 2
}
|
||
2106.04415-en-3
|
2106.04415-en-3.jpg
|
$E^T$ of the items and feed them into the interactivity module. In the interactivity module, we design an ingenious graph structure that regards each item in the sequence as a node. Our graph structure not only captures sequential information, but also allows items to interact via the graph neural network. Experimental results prove that the interaction between items can effectively improve multi-interest soft clustering.
Firstly, we construct a meaningful graph from the sequence. As shown in Figure 2, the graph structure contains one virtual central node and $n$ item nodes. The virtual central node is responsible for receiving and distributing feature among all item nodes. For each item node, the black edge represents the undirected connection with the virtual central node. Such a graph structure can make any two non-adjacent item nodes become two-hop neighbors, and can capture nonlocal information. Since the user's behavior is a sequence, we connect the item node in order, as shown by the red connection in the graph. Such graph structure can model the correlation between adjacent items, allow each item node to gather information from neighbors, and capture local information.
Next, we present how to obtain feature vectors of nodes via graph neural network. We use $c^l\in\mathbb{R}^{1\times d}$ and $H^l\in\mathbb{R}^{n\times d}$ to represent the virtual central node and all the item nodes at step $\hat{l}$ respectively. We initalize $H^0$ and $c^0$ as:
$$H^0=E^I+E^T$$
(5)
$$c^0=average(H^0)$$
(6)
The update of all nodes at step $l$ is divided into two stages: updating all item nodes and updating the virtual central node. In the first stage, each item node aggregates the following information: its adjacent node $h_r-1^{l-1}$ in sequence for local information, and the virtual central node $c^{l-1}$ for global information, in addition, its previous feature $h_{r}^{l-1}$, and its corresponding item embedding $e_r.$ After that, we update the feature of each item node $r$ at step $l$ based on the attention mechanism.
(7)
$$g_r^l=concat[h_{r-1}^{l-1};c^{l-1};h_r^{l-1};e_r]$$
$$h_{r}^{l}=MultiAtt(Q=h_{r}^{l-1},K=g_{r}^{l},V=g_{r}^{l})$$
(8)
where MultiAtt means Multi-Head Attention network. It
was proposed by Vaswani et al.[2017].
In the second stage, the virtual central node aggregates the information of all the item nodes $H^l$ and its previous feature $c^{l-1}.$ Similar to the item node, it also uses the attention mechanism to update the state.
(9)
$$q^l=concat[c^{l-1};H^l]$$
$$c^l=MultiAtt(Q=c^{l-1},K=q^l,V=q^l)$$
(10)
The overall update algorithm of the interactivity module is
shown in the Alg-1.
After $L$ rounds of update, the final feature matrix $H^l\in$ $\mathbb{R}^{n\times d}$ of item nodes can be used for multi-interest extraction of user interaction sequence.
Multi-Interest Extraction Layer
We use the self-attention method to extract multi-interest
from the user sequence. Given the hidden embedding representation $H\in\mathbb{R}^{n\times d}$ of all items from the interactivity module. We can obtain the attention weight $A_2\in\mathbb{R}^K\times n$ of multi-interest by the formula:
$$A_2=softmax(W_3tanh(W_2H^\top))$$
(11)
where $W_3$ and $W_2$ are trainable parameters of size $K\times4d$ and $4d\times d.\dot{K}$ denotes the number of user interests. The matrix $A_2$ with size $K\times n$ represents the $K$ perspectives of the user sequence, reflecting the $K$ interest of the user $u.$ Hence, the weighted sum of all item embedding with the attention weight can obtain the $K$ vector representation of the user to reflect the different interests.
$$M_u=A_2H$$
(12)
## 3.3 Training Phase
After computing the interest embedding from user behavior through the multi-interest extraction layer,based on a hard attention strategy in the interest attention layer, for the target item, we use the argmax operation to find the most relevant one among the $K$ vector representation:
$$m_u=M_u[:,argmax(M_u^\top e_o)]$$
(13)
where $M_u$ is user's multi-interest representation matrix, $e_o$
represents the embedding of the target item.
Given a training sample $(u,o)$ with the user embedding $m_u$ and the target item embedding $e_o$, we should maximize the probability of user $u$ interacting with item $o$ in the training phase. Due to the expensive computational cost, we utilize the sample softmax method to calculate the likelihood of the user $u$ interacting with the target item $o.$ Finally, we train our model by minimizing the following objective function:
(14)
$$\begin{aligned}\mathcal{L}(\theta)&=\sum_{u\in U}-\log\frac{exp(m_u^\top e_o)}{\sum_{v\in Sample(I)}exp(m_u^\top e_v)}\end{aligned}$$
## 3.4 Testing Phase
After the multi-interest extraction layer, we obtain multiple interests embedding for each user based on his/her past behavior, which can be used for recommendation prediction. In
|
[
[
0.08660888671875,
0.06756591796875,
0.484375,
0.06817626953125,
0.484375,
0.08331298828125,
0.08660888671875,
0.08209228515625
],
[
0.51318359375,
0.0694580078125,
0.72216796875,
0.0694580078125,
0.72216796875,
0.08331298828125,
0.51318359375,
0.08331298828125
],
[
0.0849609375,
0.083984375,
0.486083984375,
0.083984375,
0.486083984375,
0.09783935546875,
0.0849609375,
0.09783935546875
],
[
0.5107421875,
0.08709716796875,
0.91015625,
0.085205078125,
0.91015625,
0.0997314453125,
0.5107421875,
0.10162353515625
],
[
0.0849609375,
0.09783935546875,
0.484375,
0.09783935546875,
0.484375,
0.111083984375,
0.0849609375,
0.111083984375
],
[
0.5107421875,
0.10162353515625,
0.86376953125,
0.10162353515625,
0.86376953125,
0.11553955078125,
0.5107421875,
0.11553955078125
],
[
0.08660888671875,
0.11236572265625,
0.486083984375,
0.11236572265625,
0.486083984375,
0.1256103515625,
0.08660888671875,
0.1256103515625
],
[
0.5087890625,
0.11297607421875,
0.9150390625,
0.11236572265625,
0.9150390625,
0.1300048828125,
0.5087890625,
0.1307373046875
],
[
0.0849609375,
0.1256103515625,
0.484375,
0.1256103515625,
0.484375,
0.1395263671875,
0.0849609375,
0.1395263671875
],
[
0.51123046875,
0.1307373046875,
0.55126953125,
0.1307373046875,
0.55126953125,
0.1427001953125,
0.51123046875,
0.1427001953125
],
[
0.0849609375,
0.1383056640625,
0.486083984375,
0.1383056640625,
0.486083984375,
0.1514892578125,
0.0849609375,
0.1514892578125
],
[
0.51953125,
0.14453125,
0.654296875,
0.1427001953125,
0.654296875,
0.1572265625,
0.51953125,
0.1590576171875
],
[
0.08416748046875,
0.1514892578125,
0.431396484375,
0.1514892578125,
0.431396484375,
0.1654052734375,
0.08416748046875,
0.1654052734375
],
[
0.51953125,
0.1578369140625,
0.67333984375,
0.1578369140625,
0.67333984375,
0.17236328125,
0.51953125,
0.17236328125
],
[
0.101318359375,
0.1673583984375,
0.484375,
0.1673583984375,
0.484375,
0.18115234375,
0.101318359375,
0.18115234375
],
[
0.51953125,
0.173583984375,
0.685546875,
0.173583984375,
0.685546875,
0.1875,
0.51953125,
0.1875
],
[
0.0849609375,
0.18115234375,
0.48779296875,
0.18115234375,
0.48779296875,
0.1944580078125,
0.0849609375,
0.1944580078125
],
[
0.0849609375,
0.1944580078125,
0.486083984375,
0.1944580078125,
0.486083984375,
0.2083740234375,
0.0849609375,
0.2083740234375
],
[
0.51708984375,
0.1944580078125,
0.529296875,
0.186279296875,
0.5400390625,
0.1944580078125,
0.52783203125,
0.20263671875
],
[
0.5556640625,
0.189453125,
0.6806640625,
0.189453125,
0.6806640625,
0.199462890625,
0.5556640625,
0.199462890625
],
[
0.51953125,
0.2020263671875,
0.53759765625,
0.2020263671875,
0.53759765625,
0.21533203125,
0.51953125,
0.21533203125
],
[
0.5703125,
0.1962890625,
0.7998046875,
0.1962890625,
0.7998046875,
0.2171630859375,
0.5703125,
0.2171630859375
],
[
0.0849609375,
0.2083740234375,
0.486083984375,
0.2083740234375,
0.486083984375,
0.2215576171875,
0.0849609375,
0.2215576171875
],
[
0.51953125,
0.2171630859375,
0.53759765625,
0.2171630859375,
0.53759765625,
0.2310791015625,
0.51953125,
0.2310791015625
],
[
0.57177734375,
0.2139892578125,
0.875,
0.2139892578125,
0.875,
0.231689453125,
0.57177734375,
0.231689453125
],
[
0.0849609375,
0.222900390625,
0.486083984375,
0.222900390625,
0.486083984375,
0.236083984375,
0.0849609375,
0.236083984375
],
[
0.51953125,
0.231689453125,
0.5390625,
0.231689453125,
0.5390625,
0.2442626953125,
0.51953125,
0.2442626953125
],
[
0.5556640625,
0.231689453125,
0.609375,
0.231689453125,
0.609375,
0.24365234375,
0.5556640625,
0.24365234375
],
[
0.5556640625,
0.2442626953125,
0.705078125,
0.242431640625,
0.705078125,
0.2568359375,
0.5556640625,
0.2587890625
],
[
0.08416748046875,
0.236083984375,
0.486083984375,
0.23486328125,
0.486083984375,
0.248779296875,
0.08416748046875,
0.25
],
[
0.0849609375,
0.250732421875,
0.486083984375,
0.250732421875,
0.486083984375,
0.264404296875,
0.0849609375,
0.264404296875
],
[
0.521484375,
0.2474365234375,
0.5361328125,
0.2474365234375,
0.5361328125,
0.2568359375,
0.521484375,
0.2568359375
],
[
0.52783203125,
0.2568359375,
0.5390625,
0.26513671875,
0.52783203125,
0.2734375,
0.51708984375,
0.26513671875
],
[
0.56884765625,
0.255615234375,
0.84716796875,
0.255615234375,
0.84716796875,
0.2734375,
0.56884765625,
0.2734375
],
[
0.0849609375,
0.264404296875,
0.484375,
0.264404296875,
0.484375,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.5595703125,
0.26318359375,
0.56640625,
0.26318359375,
0.56640625,
0.266357421875,
0.5595703125,
0.266357421875
],
[
0.08416748046875,
0.276611328125,
0.48779296875,
0.27783203125,
0.48779296875,
0.291748046875,
0.08416748046875,
0.290283203125
],
[
0.51318359375,
0.272705078125,
0.59326171875,
0.272705078125,
0.59326171875,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.08660888671875,
0.29345703125,
0.484375,
0.29345703125,
0.484375,
0.3037109375,
0.08660888671875,
0.3037109375
],
[
0.0849609375,
0.30615234375,
0.486083984375,
0.30615234375,
0.486083984375,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.51318359375,
0.30810546875,
0.73779296875,
0.30810546875,
0.73779296875,
0.3212890625,
0.51318359375,
0.3212890625
],
[
0.08660888671875,
0.32080078125,
0.484375,
0.32080078125,
0.484375,
0.330810546875,
0.08660888671875,
0.330810546875
],
[
0.51123046875,
0.3212890625,
0.91259765625,
0.3212890625,
0.91259765625,
0.335205078125,
0.51123046875,
0.335205078125
],
[
0.08660888671875,
0.335205078125,
0.47216796875,
0.335205078125,
0.47216796875,
0.34521484375,
0.08660888671875,
0.34521484375
],
[
0.51318359375,
0.337158203125,
0.9111328125,
0.337158203125,
0.9111328125,
0.35107421875,
0.51318359375,
0.35107421875
],
[
0.101318359375,
0.3466796875,
0.484375,
0.3466796875,
0.484375,
0.35986328125,
0.101318359375,
0.35986328125
],
[
0.5107421875,
0.347900390625,
0.91259765625,
0.347900390625,
0.91259765625,
0.365478515625,
0.5107421875,
0.365478515625
],
[
0.08416748046875,
0.35791015625,
0.488525390625,
0.356689453125,
0.488525390625,
0.375,
0.08416748046875,
0.375732421875
],
[
0.5087890625,
0.363037109375,
0.9140625,
0.35986328125,
0.9140625,
0.37744140625,
0.5087890625,
0.380615234375
],
[
0.0849609375,
0.375,
0.488525390625,
0.375,
0.488525390625,
0.38818359375,
0.0849609375,
0.38818359375
],
[
0.5107421875,
0.37744140625,
0.7255859375,
0.37744140625,
0.7255859375,
0.391357421875,
0.5107421875,
0.391357421875
],
[
0.08416748046875,
0.38818359375,
0.3896484375,
0.38818359375,
0.3896484375,
0.40283203125,
0.08416748046875,
0.40283203125
],
[
0.59619140625,
0.397705078125,
0.8310546875,
0.397705078125,
0.8310546875,
0.412353515625,
0.59619140625,
0.412353515625
],
[
0.88134765625,
0.399658203125,
0.91259765625,
0.399658203125,
0.91259765625,
0.412353515625,
0.88134765625,
0.412353515625
],
[
0.2296142578125,
0.4091796875,
0.34228515625,
0.4072265625,
0.34228515625,
0.4248046875,
0.2296142578125,
0.4267578125
],
[
0.464111328125,
0.41357421875,
0.48779296875,
0.41357421875,
0.48779296875,
0.4267578125,
0.464111328125,
0.4267578125
],
[
0.51318359375,
0.417236328125,
0.91259765625,
0.417236328125,
0.91259765625,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.5146484375,
0.43310546875,
0.9111328125,
0.43310546875,
0.9111328125,
0.443115234375,
0.5146484375,
0.443115234375
],
[
0.2181396484375,
0.443115234375,
0.351318359375,
0.44580078125,
0.351318359375,
0.46337890625,
0.2181396484375,
0.4609375
],
[
0.464111328125,
0.447509765625,
0.48779296875,
0.447509765625,
0.48779296875,
0.460205078125,
0.464111328125,
0.460205078125
],
[
0.5146484375,
0.44580078125,
0.91259765625,
0.44580078125,
0.91259765625,
0.458984375,
0.5146484375,
0.458984375
],
[
0.5107421875,
0.458984375,
0.91259765625,
0.458251953125,
0.91259765625,
0.472900390625,
0.5107421875,
0.473388671875
],
[
0.10455322265625,
0.47021484375,
0.486083984375,
0.47021484375,
0.486083984375,
0.484130859375,
0.10455322265625,
0.484130859375
],
[
0.51318359375,
0.474853515625,
0.91259765625,
0.474853515625,
0.91259765625,
0.48486328125,
0.51318359375,
0.48486328125
],
[
0.08660888671875,
0.48486328125,
0.484375,
0.48486328125,
0.484375,
0.498779296875,
0.08660888671875,
0.498779296875
],
[
0.51123046875,
0.4873046875,
0.9140625,
0.4873046875,
0.9140625,
0.50048828125,
0.51123046875,
0.50048828125
],
[
0.101318359375,
0.498779296875,
0.486083984375,
0.498779296875,
0.486083984375,
0.51220703125,
0.101318359375,
0.51220703125
],
[
0.51123046875,
0.501953125,
0.65673828125,
0.501953125,
0.65673828125,
0.51513671875,
0.51123046875,
0.51513671875
],
[
0.08660888671875,
0.51318359375,
0.486083984375,
0.51318359375,
0.486083984375,
0.52783203125,
0.08660888671875,
0.52783203125
],
[
0.6708984375,
0.51953125,
0.75341796875,
0.51953125,
0.75341796875,
0.5341796875,
0.6708984375,
0.5341796875
],
[
0.87841796875,
0.51806640625,
0.9140625,
0.5166015625,
0.91650390625,
0.5322265625,
0.8798828125,
0.53515625
],
[
0.0849609375,
0.52978515625,
0.486083984375,
0.52978515625,
0.486083984375,
0.54443359375,
0.0849609375,
0.54443359375
],
[
0.08660888671875,
0.54345703125,
0.48779296875,
0.54345703125,
0.48779296875,
0.55810546875,
0.08660888671875,
0.55810546875
],
[
0.5146484375,
0.54443359375,
0.5390625,
0.54443359375,
0.5390625,
0.55615234375,
0.5146484375,
0.55615234375
],
[
0.5439453125,
0.54345703125,
0.67236328125,
0.54345703125,
0.67236328125,
0.556640625,
0.5439453125,
0.556640625
],
[
0.0849609375,
0.5595703125,
0.48779296875,
0.5595703125,
0.48779296875,
0.57275390625,
0.0849609375,
0.57275390625
],
[
0.5146484375,
0.56005859375,
0.91259765625,
0.56005859375,
0.91259765625,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.0849609375,
0.57275390625,
0.488525390625,
0.57275390625,
0.488525390625,
0.5859375,
0.0849609375,
0.5859375
],
[
0.5146484375,
0.57568359375,
0.9111328125,
0.57568359375,
0.9111328125,
0.5859375,
0.5146484375,
0.5859375
],
[
0.08660888671875,
0.5869140625,
0.1666259765625,
0.5869140625,
0.1666259765625,
0.5986328125,
0.08660888671875,
0.5986328125
],
[
0.51123046875,
0.58837890625,
0.9140625,
0.58837890625,
0.9140625,
0.6015625,
0.51123046875,
0.6015625
],
[
0.1732177734375,
0.60595703125,
0.398681640625,
0.60595703125,
0.398681640625,
0.62353515625,
0.1732177734375,
0.62353515625
],
[
0.51123046875,
0.6015625,
0.9140625,
0.6015625,
0.9140625,
0.61572265625,
0.51123046875,
0.61572265625
],
[
0.464111328125,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62255859375,
0.464111328125,
0.62255859375
],
[
0.51123046875,
0.6162109375,
0.775390625,
0.6162109375,
0.775390625,
0.6298828125,
0.51123046875,
0.6298828125
],
[
0.609375,
0.63525390625,
0.814453125,
0.63525390625,
0.814453125,
0.6494140625,
0.609375,
0.6494140625
],
[
0.88134765625,
0.63623046875,
0.91259765625,
0.63623046875,
0.91259765625,
0.6484375,
0.88134765625,
0.6484375
],
[
0.1356201171875,
0.64306640625,
0.435546875,
0.64306640625,
0.435546875,
0.65771484375,
0.1356201171875,
0.65771484375
],
[
0.463134765625,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.6572265625,
0.463134765625,
0.6572265625
],
[
0.51123046875,
0.65380859375,
0.91259765625,
0.65576171875,
0.91259765625,
0.67041015625,
0.51123046875,
0.66845703125
],
[
0.08660888671875,
0.66748046875,
0.48779296875,
0.66748046875,
0.48779296875,
0.68115234375,
0.08660888671875,
0.68115234375
],
[
0.51123046875,
0.66943359375,
0.8037109375,
0.66943359375,
0.8037109375,
0.68310546875,
0.51123046875,
0.68310546875
],
[
0.0849609375,
0.681640625,
0.348876953125,
0.68115234375,
0.348876953125,
0.6943359375,
0.0849609375,
0.69580078125
],
[
0.52685546875,
0.68115234375,
0.9111328125,
0.68310546875,
0.9111328125,
0.69775390625,
0.52685546875,
0.69580078125
],
[
0.10296630859375,
0.69580078125,
0.488525390625,
0.69580078125,
0.488525390625,
0.708984375,
0.10296630859375,
0.708984375
],
[
0.51123046875,
0.69775390625,
0.91259765625,
0.69775390625,
0.91259765625,
0.7109375,
0.51123046875,
0.7109375
],
[
0.08416748046875,
0.70849609375,
0.48779296875,
0.708984375,
0.48779296875,
0.7236328125,
0.08416748046875,
0.72265625
],
[
0.5146484375,
0.7119140625,
0.9111328125,
0.7119140625,
0.9111328125,
0.72265625,
0.5146484375,
0.72265625
],
[
0.08416748046875,
0.7216796875,
0.486083984375,
0.72265625,
0.486083984375,
0.7373046875,
0.08416748046875,
0.736328125
],
[
0.5146484375,
0.72607421875,
0.91259765625,
0.72607421875,
0.91259765625,
0.736328125,
0.5146484375,
0.736328125
],
[
0.08660888671875,
0.7392578125,
0.253173828125,
0.7392578125,
0.253173828125,
0.75,
0.08660888671875,
0.75
],
[
0.51318359375,
0.73779296875,
0.9140625,
0.73779296875,
0.9140625,
0.751953125,
0.51318359375,
0.751953125
],
[
0.2115478515625,
0.75830078125,
0.359375,
0.7568359375,
0.359375,
0.7744140625,
0.2115478515625,
0.77587890625
],
[
0.51318359375,
0.75390625,
0.9111328125,
0.75390625,
0.9111328125,
0.7646484375,
0.51318359375,
0.7646484375
],
[
0.464111328125,
0.76123046875,
0.48779296875,
0.76123046875,
0.48779296875,
0.7744140625,
0.464111328125,
0.7744140625
],
[
0.5107421875,
0.76513671875,
0.87158203125,
0.76513671875,
0.87158203125,
0.77880859375,
0.5107421875,
0.77880859375
],
[
0.71826171875,
0.78515625,
0.7998046875,
0.78515625,
0.7998046875,
0.7998046875,
0.71826171875,
0.7998046875
],
[
0.125,
0.79248046875,
0.419189453125,
0.79345703125,
0.419189453125,
0.81103515625,
0.125,
0.81005859375
],
[
0.54736328125,
0.79248046875,
0.62841796875,
0.79052734375,
0.62841796875,
0.80810546875,
0.54736328125,
0.81005859375
],
[
0.62158203125,
0.79150390625,
0.689453125,
0.79736328125,
0.68701171875,
0.81103515625,
0.619140625,
0.8046875
],
[
0.88134765625,
0.79345703125,
0.9150390625,
0.79345703125,
0.9150390625,
0.8095703125,
0.88134765625,
0.8095703125
],
[
0.60546875,
0.80029296875,
0.62158203125,
0.80615234375,
0.619140625,
0.81103515625,
0.60302734375,
0.80615234375
],
[
0.6689453125,
0.80615234375,
0.8486328125,
0.80029296875,
0.8486328125,
0.818359375,
0.6689453125,
0.82373046875
],
[
0.455810546875,
0.796875,
0.486083984375,
0.796875,
0.486083984375,
0.8095703125,
0.455810546875,
0.8095703125
],
[
0.6005859375,
0.81103515625,
0.6298828125,
0.81103515625,
0.6298828125,
0.82373046875,
0.6005859375,
0.82373046875
],
[
0.10455322265625,
0.8193359375,
0.486083984375,
0.8193359375,
0.486083984375,
0.83349609375,
0.10455322265625,
0.83349609375
],
[
0.5146484375,
0.83056640625,
0.55126953125,
0.83056640625,
0.55126953125,
0.84228515625,
0.5146484375,
0.84228515625
],
[
0.54345703125,
0.8291015625,
0.66015625,
0.830078125,
0.66015625,
0.8447265625,
0.54345703125,
0.84326171875
],
[
0.08660888671875,
0.83544921875,
0.2156982421875,
0.83544921875,
0.2156982421875,
0.84521484375,
0.08660888671875,
0.84521484375
],
[
0.101318359375,
0.8466796875,
0.486083984375,
0.84521484375,
0.486083984375,
0.85986328125,
0.101318359375,
0.861328125
],
[
0.51318359375,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.51318359375,
0.861328125
],
[
0.08416748046875,
0.859375,
0.488525390625,
0.8623046875,
0.488525390625,
0.876953125,
0.08416748046875,
0.87353515625
],
[
0.51318359375,
0.86279296875,
0.90673828125,
0.86279296875,
0.90673828125,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.08416748046875,
0.87548828125,
0.273681640625,
0.87548828125,
0.273681640625,
0.8896484375,
0.08416748046875,
0.8896484375
],
[
0.51123046875,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"ET of the items and feed them into the interactivity module.",
"Algorithm 1 Update Procedure",
"In the interactivity module, we design an ingenious graph",
"Input: Items embedding El and time interval embedding ET",
" structure that regards each item in the sequence as a node.",
"Parameter: Multi-head attention network parameters",
"Our graph structure not only captures sequential information,",
"Output: The feature representation H E Rnxd of all",
"but also allows items to interact via the graph neural network.",
"items",
"Experimental results prove that the interaction between items",
"1:HO←EI+ET",
"can effectively improve multi-interest soft clustering",
"2: co averaqe(Ho)",
"Firstly, we construct a meaningful graph from the se-",
"3: for step -1 to L do",
"quence. As shown in Figure 2, the graph structure contains",
"one virtual central node and n item nodes. The virtual cen-",
"X.",
"forr-ltondo",
"5:",
"gl - concat[h,-1;cl-1; hf-1;e,]",
"tral node is responsible for receiving and distributing feature",
"6:",
"h, = MultiAtt(Q= hi-i,K = gf,V = gf)",
"among all item nodes.For each item node, the black edge",
"7:",
"end for",
"q' = concat[cl-1: Hli",
"represents the undirected connection with the virtual central",
"node. Such a graph structure can make any two non-adjacent",
"8:",
"9",
"= MultiAtt(Q= cl-1,K=q,V= q)",
"item nodes become two-hop neighbors, and can capture non-",
"n",
"local information. Since the user's behavior is a sequence, we",
"10: end for",
"connecttheitemnode in order,as shown by the red connec-",
"tion in the graph. Such graph structure can model the corre-",
"Multi-Interest Extraction Layer",
"lationbetweenadjacentitems.alloweachitemnodetogather",
"We use the self-attention method to extract multi-interest",
"informationfromneighbors,andcapturelocalinformation.",
"from the user sequence. Given the hidden embedding rep-",
"Next, we present how to obtain feature vectors of nodes",
"resentation H e Rnxd of all items from the interactivity",
"via graph neural network. We use cl e RIxd and H' e Rnxd",
"module. We can obtain the attention weight A2 E IRKxn",
"to represent the virtual central node and all the item nodes at",
"of multi-interest by the formula:",
"step l respectively. We initalize H and c° as:",
"A2 = softmar(Wstanh(W2HT))",
"(11)",
"H°=EI +ET",
"(5)",
"where W: and W2 are trainable parameters of size K × 4d and",
"4d X d.K denotes the number of userinterests.The matrix",
"co = average(H°)",
"(6)",
"A2 with size K × n represents the K perspectives of the user",
"sequence, reflecting the K interest of the user u. Hence, the",
"The update of all nodes at step I is divided into two stages:",
"weighted sumof allitemembeddingwiththeattentionweight",
"updating all item nodes and updating the virtual central node.",
"can obtain the K vector representation of the user to reflect",
"In the first stage, each item node aggregates the follow-",
"the different interests.",
"ing information: its adjacent node h,-1 in sequence for local",
"Mu= A2H",
"(12)",
"information, and the virtual central node cl-1 for global in-",
"formation, in addition, its previous feature hf-1, and its cor-",
"3.3",
"Training Phase",
"responding item embedding er. After that, we update the",
"After computing the interest embedding from user behavior",
"feature of each item node r at step l based on the attention",
"through the multi-interest extractionlayer. based on a hard",
"mechanism.",
"attention strategy in the interest attention layer, for the target",
"g, = concat[h'-1;cl-1; h-1;er]",
"item, we use the argma operation to find the most relevant",
"(7)",
"one among the K vector representation:",
"mu = Mul:, argmar(MTeo)",
"(13)",
"ht = MultiAtt(Q = hf-1,K = gt, V = g)",
"(8)",
"where Mu is user's multi-interest representation matrix, eo",
"where MultiAtt means Multi-Head Attention network. It",
"represents the embedding of the target item.",
"wasproposedbyVaswanietal.[2017].",
"Given a training sample (u, o) with the user embedding mau",
"In the second stage, the virtual central node aggregates the",
"and the target item embedding eo, we should maximize the",
"information of all the item nodes H' and its previous feature",
"probability of user u interacting with item o in the training",
"cl-1. Similar to the item node, it also uses the attention mech-",
"phase.Due to the expensive computational cost,we utilize",
"anism to updatethe state.",
"the sample softmax method to calculate the likelihood of the",
"q' = concat[cl-1; H']",
"user uinteractingwiththetargetitem o.Finally,wetrain our",
"(9)",
"model by minimizing the following objective function:",
"erp(mTe)",
"cl = MultiAtt(Q=cl-1,K =q,V=q)",
"C(0) =",
"-log",
"(14)",
"",
"ZveSample(1) erp(mte,)",
"(10)",
"uEU",
"The overall update algorithm of the interactivity module is",
"3.4",
"Testing Phase",
"shownintheAlg-1.",
"After L rounds of update, the final feature matrix Hl E",
"After the multi-interest extraction layer, we obtain multiple",
"Rnxd of item nodes can be used for multi-interest extraction",
"interests embeddingfor eachuserbased onhis/her pastbe",
"of user interaction sequence.",
"havior, which can be used for recommendation prediction. In"
] |
{
"arxiv_id": "2106.04415",
"language": "en",
"page_idx": 3
}
|
||
2108.12139-en-0
|
2108.12139-en-0.jpg
|
Dealing with Typos for BERT-based Passage Retrieval and Ranking
Guido Zuccon
The University of Queensland Brisbane, QLD, Australia g.zuccon@uq.edu.au
Shengyao Zhuang
The University of Queensland Brisbane, QLD, Australia s.zhuang@uq.edu.au
### $\mathbf{Abstract}$
Passage retrieval and ranking is a key task in open-domain question answering and information retrieval. Current effective approaches mostly rely on pre-trained deep language model-based retrievers and rankers. These methods have been shown to effectively model the semantic matching between queries and passages, also in presence of keyword mismatch, i.e. passages that are relevant to a query but do not contain important query keywords. In this paper we consider the Dense Retriever (DR), a passage retrieval method, and the BERT re-ranker, a popular passage re-ranking method. In this context, we formally investigate how these models respond and adapt to a specific type of keyword mismatch - that caused by keyword typos occurring in queries. Through empirical investigation, we find that typos can lead to a significant drop in retrieval and ranking effectiveness. We then propose a simple typos-aware training framework for DR and BERT re-ranker to address this issue. Our experimental results on the MS MARCO passage ranking dataset show that, with our proposed typos-aware training, DR and BERT re-ranker can become robust to typos in queries, resulting in significantly improved effectiveness compared to models trained without appropriately accounting for typos.
## 1 Introduction
Passage ranking is a core task for many information retrieval related applications. In the context of conversational search and question answering, for example, passage ranking is often the first step in the system's pipeline: thus the quality of the ranking results will affect the effectivenesses of the downstream tasks. Traditional passage ranking models, such TF-IDF and BM25, use exact keyword matching signals, where a retrieved passage must contain at least one of the query's keywords. This mechanism however limits the capability of these models
to retrieve passages that are semantically relevant but use different keywords: this is the well-known vocabulary mismatch problem.
Recent advances in NLP have seen the introduction of deep language models (Devlin et al., 2018; Brown et al., 2020; Raffel et al., 2020); BERT (Devlin et al., 2018) in particular has shown generalised promise in language understanding tasks. BERT adopts the transformer encoder (Vaswani et al., 2017) as model architecture and uses WordPiece token embeddings (Wu et al., 2016) as model inputs. This design allows BERT to deal with the vocabulary mismatch problem. Hence practitioners have turned to design BERT-based passage ranking models (Lin et al., 2020).
Two main directions have been adopted to ex-
ploit BERT for effective passage ranking:
· Dense Retriever (DR) (Zhan et al., 2020; Xiong
et al., 2020; Gao et al., 2021b; Khattab and Zaharia, 2020; Karpukhin et al., 2020; Ding et al., 2020; Luan et al., 2020): queries and passages are separately encoded into low-dimensional dense representations with BERT. At indexing time, passage representations are computed and then stored in the index. At query time, a single query encoder inference is needed to obtain the query representation; then passage relevance scores are estimated by computing the similarity between the query and passages' representations.
BERT re-ranker, a.k.a. monoBERT (Nogueira and Cho, 2019; Dai and Callan, 2019; Gao et al., 2021a): the ranking task is modelled as a classification task that builds upon the BERT model. The input to BERT is a < $query,passage>$ pair and the relevance score can be computed by a linear layer on the $<CLS>$ token embedding, or the query likelihood estimated by the BERT model (Zhuang and Zuccon, 2021b). A key drawback of BERT re-ranker models is that multiple inferences are required at query time: this is a computationally expensive process
|
[
[
0.1502685546875,
0.0914306640625,
0.8486328125,
0.0926513671875,
0.8486328125,
0.10986328125,
0.1502685546875,
0.1092529296875
],
[
0.25439453125,
0.13720703125,
0.414794921875,
0.1383056640625,
0.414794921875,
0.15380859375,
0.25439453125,
0.152587890625
],
[
0.60546875,
0.1383056640625,
0.73046875,
0.1383056640625,
0.73046875,
0.152587890625,
0.60546875,
0.152587890625
],
[
0.21240234375,
0.155029296875,
0.45751953125,
0.155029296875,
0.45751953125,
0.169189453125,
0.21240234375,
0.169189453125
],
[
0.54248046875,
0.155029296875,
0.7900390625,
0.155029296875,
0.7900390625,
0.169189453125,
0.54248046875,
0.169189453125
],
[
0.2301025390625,
0.171630859375,
0.43994140625,
0.171630859375,
0.43994140625,
0.1859130859375,
0.2301025390625,
0.1859130859375
],
[
0.56103515625,
0.171630859375,
0.771484375,
0.171630859375,
0.771484375,
0.1859130859375,
0.56103515625,
0.1859130859375
],
[
0.225830078125,
0.189453125,
0.444091796875,
0.189453125,
0.444091796875,
0.203125,
0.225830078125,
0.203125
],
[
0.55517578125,
0.189453125,
0.77685546875,
0.188232421875,
0.77685546875,
0.203125,
0.55517578125,
0.2037353515625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.5087890625,
0.254150390625,
0.8818359375,
0.2529296875,
0.8818359375,
0.267333984375,
0.5087890625,
0.268310546875
],
[
0.509765625,
0.26953125,
0.880859375,
0.26953125,
0.880859375,
0.283203125,
0.509765625,
0.283203125
],
[
0.145263671875,
0.283203125,
0.459228515625,
0.283203125,
0.459228515625,
0.29638671875,
0.145263671875,
0.29638671875
],
[
0.51123046875,
0.2861328125,
0.744140625,
0.2861328125,
0.744140625,
0.2998046875,
0.51123046875,
0.2998046875
],
[
0.1435546875,
0.297607421875,
0.459228515625,
0.297607421875,
0.459228515625,
0.311279296875,
0.1435546875,
0.311279296875
],
[
0.52880859375,
0.303955078125,
0.8818359375,
0.303955078125,
0.8818359375,
0.318359375,
0.52880859375,
0.318359375
],
[
0.145263671875,
0.312255859375,
0.459228515625,
0.312255859375,
0.459228515625,
0.325927734375,
0.145263671875,
0.325927734375
],
[
0.0251922607421875,
0.322509765625,
0.05877685546875,
0.322509765625,
0.05877685546875,
0.3759765625,
0.0251922607421875,
0.3759765625
],
[
0.1427001953125,
0.32470703125,
0.459228515625,
0.325927734375,
0.459228515625,
0.34033203125,
0.1427001953125,
0.339111328125
],
[
0.51123046875,
0.320556640625,
0.884765625,
0.320556640625,
0.884765625,
0.3349609375,
0.51123046875,
0.3349609375
],
[
0.145263671875,
0.34033203125,
0.45751953125,
0.34033203125,
0.45751953125,
0.35400390625,
0.145263671875,
0.35400390625
],
[
0.51123046875,
0.3349609375,
0.88330078125,
0.336181640625,
0.88330078125,
0.350830078125,
0.51123046875,
0.349853515625
],
[
0.145263671875,
0.352783203125,
0.45751953125,
0.352783203125,
0.45751953125,
0.366455078125,
0.145263671875,
0.366455078125
],
[
0.51123046875,
0.352783203125,
0.8818359375,
0.352783203125,
0.8818359375,
0.366455078125,
0.51123046875,
0.366455078125
],
[
0.031890869140625,
0.3623046875,
0.0679931640625,
0.366455078125,
0.056243896484375,
0.41455078125,
0.0201568603515625,
0.410400390625
],
[
0.145263671875,
0.3681640625,
0.459228515625,
0.3681640625,
0.459228515625,
0.3818359375,
0.145263671875,
0.3818359375
],
[
0.51318359375,
0.37060546875,
0.8818359375,
0.37060546875,
0.8818359375,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.145263671875,
0.383056640625,
0.459228515625,
0.383056640625,
0.459228515625,
0.396728515625,
0.145263671875,
0.396728515625
],
[
0.51123046875,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.51123046875,
0.39794921875
],
[
0.145263671875,
0.394287109375,
0.45751953125,
0.396728515625,
0.45751953125,
0.410888671875,
0.145263671875,
0.4091796875
],
[
0.5146484375,
0.402587890625,
0.880859375,
0.402587890625,
0.880859375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.026031494140625,
0.410400390625,
0.05792236328125,
0.410400390625,
0.05792236328125,
0.439453125,
0.026031494140625,
0.439453125
],
[
0.145263671875,
0.410888671875,
0.453369140625,
0.410888671875,
0.453369140625,
0.424560546875,
0.145263671875,
0.424560546875
],
[
0.51318359375,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.145263671875,
0.43115234375,
0.459228515625,
0.43115234375,
0.459228515625,
0.4453125,
0.145263671875,
0.4453125
],
[
0.51123046875,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.446533203125,
0.51123046875,
0.446533203125
],
[
0.0218353271484375,
0.446533203125,
0.060455322265625,
0.4453125,
0.06378173828125,
0.5166015625,
0.0251922607421875,
0.51708984375
],
[
0.145263671875,
0.4453125,
0.459228515625,
0.4453125,
0.459228515625,
0.460205078125,
0.145263671875,
0.460205078125
],
[
0.51318359375,
0.449462890625,
0.8818359375,
0.449462890625,
0.8818359375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.1435546875,
0.458984375,
0.45751953125,
0.460205078125,
0.45751953125,
0.473876953125,
0.1435546875,
0.47265625
],
[
0.509765625,
0.465576171875,
0.8818359375,
0.46435546875,
0.8818359375,
0.478515625,
0.509765625,
0.479736328125
],
[
0.145263671875,
0.474365234375,
0.459228515625,
0.474365234375,
0.459228515625,
0.488037109375,
0.145263671875,
0.488037109375
],
[
0.51318359375,
0.482177734375,
0.8212890625,
0.482177734375,
0.8212890625,
0.495361328125,
0.51318359375,
0.495361328125
],
[
0.145263671875,
0.488037109375,
0.459228515625,
0.488037109375,
0.459228515625,
0.501953125,
0.145263671875,
0.501953125
],
[
0.145263671875,
0.5029296875,
0.459228515625,
0.5029296875,
0.459228515625,
0.5166015625,
0.145263671875,
0.5166015625
],
[
0.52880859375,
0.49951171875,
0.880859375,
0.49951171875,
0.880859375,
0.51318359375,
0.52880859375,
0.51318359375
],
[
0.145263671875,
0.5166015625,
0.45751953125,
0.5166015625,
0.45751953125,
0.52978515625,
0.145263671875,
0.52978515625
],
[
0.51123046875,
0.5166015625,
0.81787109375,
0.5166015625,
0.81787109375,
0.52978515625,
0.51123046875,
0.52978515625
],
[
0.145263671875,
0.53076171875,
0.459228515625,
0.53076171875,
0.459228515625,
0.54443359375,
0.145263671875,
0.54443359375
],
[
0.0218353271484375,
0.52099609375,
0.0621337890625,
0.52099609375,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.509765625,
0.533203125,
0.880859375,
0.53369140625,
0.880859375,
0.548828125,
0.509765625,
0.54736328125
],
[
0.145263671875,
0.5458984375,
0.459228515625,
0.5458984375,
0.459228515625,
0.55859375,
0.145263671875,
0.55859375
],
[
0.5263671875,
0.55029296875,
0.87890625,
0.55029296875,
0.87890625,
0.56103515625,
0.5263671875,
0.56103515625
],
[
0.145263671875,
0.55859375,
0.459228515625,
0.55859375,
0.459228515625,
0.572265625,
0.145263671875,
0.572265625
],
[
0.5263671875,
0.56640625,
0.8818359375,
0.56640625,
0.8818359375,
0.580078125,
0.5263671875,
0.580078125
],
[
0.145263671875,
0.572265625,
0.45751953125,
0.572265625,
0.45751953125,
0.5859375,
0.145263671875,
0.5859375
],
[
0.1435546875,
0.5859375,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.1435546875,
0.599609375
],
[
0.5263671875,
0.58056640625,
0.8818359375,
0.58203125,
0.8818359375,
0.5966796875,
0.5263671875,
0.595703125
],
[
0.145263671875,
0.6015625,
0.45751953125,
0.6015625,
0.45751953125,
0.615234375,
0.145263671875,
0.615234375
],
[
0.52490234375,
0.5986328125,
0.8818359375,
0.59716796875,
0.8818359375,
0.6123046875,
0.52490234375,
0.61328125
],
[
0.145263671875,
0.6162109375,
0.459228515625,
0.6162109375,
0.459228515625,
0.6298828125,
0.145263671875,
0.6298828125
],
[
0.5263671875,
0.6142578125,
0.8818359375,
0.6142578125,
0.8818359375,
0.62890625,
0.5263671875,
0.62890625
],
[
0.145263671875,
0.63134765625,
0.459228515625,
0.63134765625,
0.459228515625,
0.64453125,
0.145263671875,
0.64453125
],
[
0.5263671875,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.5263671875,
0.64453125
],
[
0.145263671875,
0.64453125,
0.459228515625,
0.64453125,
0.459228515625,
0.65771484375,
0.145263671875,
0.65771484375
],
[
0.52490234375,
0.6455078125,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.52490234375,
0.66015625
],
[
0.145263671875,
0.6591796875,
0.459228515625,
0.6591796875,
0.459228515625,
0.6728515625,
0.145263671875,
0.6728515625
],
[
0.52490234375,
0.662109375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.52490234375,
0.6767578125
],
[
0.145263671875,
0.6728515625,
0.459228515625,
0.6728515625,
0.459228515625,
0.68603515625,
0.145263671875,
0.68603515625
],
[
0.5263671875,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.5263671875,
0.693359375
],
[
0.145263671875,
0.68701171875,
0.3837890625,
0.68701171875,
0.3837890625,
0.70068359375,
0.145263671875,
0.70068359375
],
[
0.5263671875,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.5263671875,
0.708984375
],
[
0.12005615234375,
0.71630859375,
0.133544921875,
0.7197265625,
0.1292724609375,
0.728515625,
0.1158447265625,
0.7255859375
],
[
0.14111328125,
0.71728515625,
0.26025390625,
0.71728515625,
0.26025390625,
0.73095703125,
0.14111328125,
0.73095703125
],
[
0.5263671875,
0.7119140625,
0.82373046875,
0.7119140625,
0.82373046875,
0.7255859375,
0.5263671875,
0.7255859375
],
[
0.51318359375,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.7412109375,
0.51318359375,
0.7412109375
],
[
0.1175537109375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.52490234375,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.52490234375,
0.755859375
],
[
0.1158447265625,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.77392578125,
0.1158447265625,
0.77392578125
],
[
0.52490234375,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.771484375,
0.52490234375,
0.771484375
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.5263671875,
0.775390625,
0.88330078125,
0.775390625,
0.88330078125,
0.7890625,
0.5263671875,
0.7890625
],
[
0.1158447265625,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.1158447265625,
0.8056640625
],
[
0.52490234375,
0.79052734375,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.52490234375,
0.80419921875
],
[
0.1158447265625,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.5263671875,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.5263671875,
0.82177734375
],
[
0.11419677734375,
0.82373046875,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.11419677734375,
0.83740234375
],
[
0.5263671875,
0.82470703125,
0.88330078125,
0.82470703125,
0.88330078125,
0.83837890625,
0.5263671875,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.5263671875,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.5263671875,
0.85400390625
],
[
0.1175537109375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.52490234375,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.869140625,
0.52490234375,
0.869140625
],
[
0.1158447265625,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.5263671875,
0.87255859375,
0.88330078125,
0.87255859375,
0.88330078125,
0.8857421875,
0.5263671875,
0.8857421875
],
[
0.1158447265625,
0.88818359375,
0.489501953125,
0.88720703125,
0.489501953125,
0.90185546875,
0.1158447265625,
0.90283203125
],
[
0.5263671875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.5263671875,
0.90283203125
],
[
0.1158447265625,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.52490234375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.52490234375,
0.91845703125
]
] |
[
"Dealing with Typos for BERT-based Passage Retrieval and Ranking",
"Shengyao Zhuang",
"Guido Zuccon",
"The University of Queensland",
"The University of Queensland",
"Brisbane, QLD, Australia",
"Brisbane, QLD, Australia",
"s.zhuang@uq.edu.au",
"g.zuccon@uq.edu.au",
"Abstract",
"to retrieve passages that are semantically relevant",
"but use different keywords: this is the well-known",
"Passage retrieval and ranking is a key task in",
"vocabulary mismatch problem.",
"open-domain question answering and inform-",
"Recent advances in NLP have seen the intro-",
"ation retrieval. Current effective approaches",
"",
"mostly rely on pre-trained deep language",
"duction of deep language models (Devlin et al.,",
"model-based retrievers and rankers. These",
"2018; Brown et al., 2020; Raffel et al., 2020);",
"methods have been shown to effectively model",
"BERT (Devlin et al., 2018) in particular has",
"das",
"the semantic matching between queries and",
"shown generalised promise in language under-",
"passages, also in presence of keyword mis-",
"standing tasks. BERT adopts the transformer en-",
"match, i.e. passages that are relevant to a query",
"coder (Vaswani et al.,2017)asmodel architecture",
"3",
"but do not contain important query keywords.",
"and uses WordPiece token embeddings (Wu et al.,",
"In this paper we consider the Dense Retriever",
"2016) as model inputs. This design allows BERT to",
"['s]",
"(DR), a passage retrieval method, and the",
"deal with the vocabulary mismatch problem. Hence",
"BERT re-ranker, a popular passage re-ranking",
"practitioners have turned to design BERT-based",
"method. In this context, we formally invest-",
"passage ranking models (Lin et al., 2020).",
"igate how these models respond and adapt to",
"a specific type of keyword mismatch - that",
"Two main directions have been adopted to ex-",
"caused by keyword typos occurring in queries.",
"ploit BERT for effective passage ranking:",
"Through empirical investigation, we find that",
"Z651218012:A",
"· Dense Retriever (DR) (Zhan et al., 2020; Xiong",
"typos can lead to a significant drop in retrieval",
"et al.. 2020: Ga0 et al.. 2021b: Khattab and Za",
"and ranking effectiveness. We then propose",
"haria, 2020; Karpukhin et al., 2020; Ding et al.",
"a simple typos-aware training framework for",
"DR and BERT re-ranker to address this issue.",
"2020; Luan et al., 2020): queries and passages are",
"Our experimental results on the MS MARCO",
" separately encoded into low-dimensional dense",
"passage ranking dataset show that, with our",
"representations with BERT. At indexing time,",
"proposed typos-aware training, DR and BERT",
"passage representations are computed and then",
"re-ranker can become robust to typos in quer-",
" stored in the index. At query time, a single query",
"ies, resulting in significantly improved effect-",
"encoder inference is needed to obtain the query",
"iveness compared to models trained without",
"representation; then passage relevance scores are",
"appropriately accounting for typos.",
"estimated by computing the similarity between",
"1",
"Introduction",
"the query and passages' representations.",
": BERT re-ranker, a.k.a. monoBERT (Nogueira",
"Passage ranking is a core task for many information",
" and Cho, 2019; Dai and Callan, 2019: Gao et al..",
"retrieval related applications. In the context of con-",
"2021la): the ranking task is modelled as a classi",
"versational search and question answering, for ex-",
"fication task that builds upon the BERT model.",
"ample, passage ranking is often the first step in the",
"The input to BERT is a < query,passage >",
"system's pipeline: thus the quality of the ranking",
"pair and the relevance score can be computed",
"results will affect the effectivenesses of the down-",
"by a linear layer on the < CLS > token em-",
" stream tasks. Traditional passage ranking models.",
"bedding, or the query likelihood estimated by",
"such TF-IDF and BM25, use exact keyword match-",
"the BERT model (Zhuang and Zuccon, 2021b)",
"ing signals, where a retrieved passage must contain",
"A key drawback of BERT re-ranker models is",
"at least one of the query's keywords. This mechan-",
"that multiple inferences are required at query",
"ism however limits the capability of these models",
"time: this is a computationally expensive process"
] |
{
"arxiv_id": "2108.12139",
"language": "en",
"page_idx": 0
}
|
||
2108.12139-en-1
|
2108.12139-en-1.jpg
|
which results in high query latency (Zhuang and Zuccon, 2021a; MacAvaney et al., 2020; Hofstätter et al., 2020). Thus the use of these methods is confined to second stage re-ranking.
In principle, these methods are not affected by keyword mismatch because they use the latent embedding space to estimate the relevance of a query to a passage. This is supported by recent work that has shown the DR and BERT re-ranker provide better semantic matching (Zhan et al., 2020; MacAvaney et al., 2020; Formal et al., 2021).
In this paper we investigate the impact of a specific type of keyword mismatch: that caused by the presence of typos in the query. Traditional exact keyword matching methods perform badly on queries that contain typos. Extra query processing steps, such as spelling correction, are required for these methods to be tolerant to typos in queries (Martins and Silva, 2004). On the other hand, it is expected that BERT-based models can handle typos occurring in queries well. This is because BERT uses the WordPiece algorithm which splits a keyword that does not match an entry in the BERT vocabulary (typos are likely to not be present in this vocabulary) into character-level sub-tokens. This can be used to produce embeddings for out-of-vocabulary keywords, which are then passed as input to the BERT encoder. However, this intuition has never been tested before, and the capability of BERTbased passage ranking models to deal with typos in queries has not been quantiffed.
To address this gap, we frst formally investigate how the BERT-based DR and re-ranker respond and adapt to queries that contain typos. Specifically, we use different typo generators to produce typos for queries; we then compare the effectiveness of the rankers when using queries with typos vs. without typos. Interestingly, we find that these models fail to handle queries with typos-typos can lead to a significant drop in effectiveness for both DR and BERT re-ranker. In order to solve this issue and obtain typo-robust ranking models, we then propose a simple typos-aware training strategy, in which queries with typos are produced and used also for training. Our experimental results on the MS MARCO passage ranking dataset show that, with our typosaware training, DR and BERT re-ranker can become robust to typos in queries, without loss in effectiveness for queries without typos.
# 2 Methodology
With respect to BERT-based models for passage retrieval and typos in queries, we investigate the following research questions:
$\cdot \textbf{RQ1: What is the impact of typos in queries on}$
BERT-based DR and re-ranker effectiveness?
$\cdot\mathbb{RQ2:}$ Do different typo types affect the effect-
iveness of the BERT-based methods differently? $\cdot \textbf{RQ3: Does the proposed typos- aware training}$ improve the effectiveness of the BERT-based methods on queries with typos? Does it hurt their effectiveness on queries without typos?
## 2.1 Synthetic Typo Generation
To answer our research questions, a reasonably large set of queries with different types of typos is needed. As there is no available dataset for passage retrieval with labels that indicate the presence of typos in queries, we set off to create one such dataset. For this we augmented the MS MARCO passage retrieval dataset. (Manual inspection of this dataset did not reveal a considerable amount of queries with typos; the dataset curators likely did manually remove most typos). For augmentation, we synthetically generated typos from the original queries in the dataset, so that we could carefully control the number and types of typos.
For generating typos, we used the following operations that give rise to typos that often occur in real-world queries (Hagen et al., 2017)
$\cdot$Random character Insertion (RandInsert): Inserts a random letter into a random word, e.g., “search typo”->“search tyapo”,
$\cdot$Random character deletion (RandDelete): Deletes a random character of a random word, such as“search typo”->“search tpo”.
$\cdot$ Random character substitution (RandSub): Randomly replaces a character of a random word with a random letter, e.g.,“search typo” $->“$search type”.
$\cdot$ Swap neighbor character (SwapNeighbor): Randomly swaps a character with one of its neighbor characters, e.g.,“search typo”-> “search tyop”
$\cdot$ Swap adjacent keyboard character (SwapAdjacent): Randomly swaps a character with one of its adjacent letter on the keyboard$^{1}$,e.g.,“search typo"->“search typi”.
$^{\mathrm{l}}$e.g., on a $QWERTY$ keyboard, the list of adjacent charac-
ters for character 's'is['q',“w',`e',“a',“d',“z',“x',“c'].
|
[
[
0.5146484375,
0.08489990234375,
0.53076171875,
0.09027099609375,
0.5263671875,
0.09857177734375,
0.5087890625,
0.09320068359375
],
[
0.52880859375,
0.08489990234375,
0.65576171875,
0.08489990234375,
0.65576171875,
0.09979248046875,
0.52880859375,
0.09979248046875
],
[
0.130126953125,
0.1015625,
0.486083984375,
0.1015625,
0.486083984375,
0.11517333984375,
0.130126953125,
0.11517333984375
],
[
0.1318359375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1318359375,
0.09979248046875
],
[
0.1318359375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1318359375,
0.1318359375
],
[
0.51123046875,
0.1134033203125,
0.8818359375,
0.1134033203125,
0.8818359375,
0.127685546875,
0.51123046875,
0.127685546875
],
[
0.130126953125,
0.133056640625,
0.414794921875,
0.1341552734375,
0.414794921875,
0.1484375,
0.130126953125,
0.1478271484375
],
[
0.51123046875,
0.1312255859375,
0.8818359375,
0.1312255859375,
0.8818359375,
0.144287109375,
0.51123046875,
0.144287109375
],
[
0.51123046875,
0.146728515625,
0.7314453125,
0.146728515625,
0.7314453125,
0.1602783203125,
0.51123046875,
0.1602783203125
],
[
0.509765625,
0.162109375,
0.8818359375,
0.163330078125,
0.8818359375,
0.1776123046875,
0.509765625,
0.177001953125
],
[
0.1351318359375,
0.168701171875,
0.486083984375,
0.168701171875,
0.486083984375,
0.1787109375,
0.1351318359375,
0.1787109375
],
[
0.1175537109375,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.1175537109375,
0.196533203125
],
[
0.5263671875,
0.179931640625,
0.8623046875,
0.179931640625,
0.8623046875,
0.193603515625,
0.5263671875,
0.193603515625
],
[
0.1175537109375,
0.1995849609375,
0.489501953125,
0.1995849609375,
0.489501953125,
0.213134765625,
0.1175537109375,
0.213134765625
],
[
0.509765625,
0.19482421875,
0.88330078125,
0.1953125,
0.88330078125,
0.210205078125,
0.509765625,
0.208984375
],
[
0.1175537109375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.52490234375,
0.2120361328125,
0.8818359375,
0.2120361328125,
0.8818359375,
0.2257080078125,
0.52490234375,
0.2257080078125
],
[
0.1158447265625,
0.23095703125,
0.48779296875,
0.23095703125,
0.48779296875,
0.24462890625,
0.1158447265625,
0.24462890625
],
[
0.509765625,
0.226806640625,
0.8818359375,
0.22802734375,
0.8818359375,
0.2423095703125,
0.509765625,
0.2410888671875
],
[
0.1175537109375,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.52294921875,
0.243408203125,
0.8818359375,
0.2423095703125,
0.8818359375,
0.257080078125,
0.52294921875,
0.25830078125
],
[
0.1175537109375,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.52490234375,
0.260009765625,
0.8818359375,
0.2587890625,
0.8818359375,
0.273681640625,
0.52490234375,
0.27490234375
],
[
0.5263671875,
0.276611328125,
0.8564453125,
0.276611328125,
0.8564453125,
0.290283203125,
0.5263671875,
0.290283203125
],
[
0.1351318359375,
0.29345703125,
0.486083984375,
0.29345703125,
0.486083984375,
0.30810546875,
0.1351318359375,
0.30810546875
],
[
0.1175537109375,
0.31005859375,
0.48779296875,
0.31005859375,
0.48779296875,
0.32373046875,
0.1175537109375,
0.32373046875
],
[
0.51123046875,
0.305908203125,
0.767578125,
0.30712890625,
0.767578125,
0.32177734375,
0.51123046875,
0.320556640625
],
[
0.1175537109375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.1175537109375,
0.34033203125
],
[
0.50732421875,
0.330810546875,
0.880859375,
0.330810546875,
0.880859375,
0.344482421875,
0.50732421875,
0.344482421875
],
[
0.1175537109375,
0.342529296875,
0.490234375,
0.342529296875,
0.490234375,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.5087890625,
0.345703125,
0.880859375,
0.345703125,
0.880859375,
0.35986328125,
0.5087890625,
0.35986328125
],
[
0.1158447265625,
0.359375,
0.489501953125,
0.359375,
0.489501953125,
0.372314453125,
0.1158447265625,
0.372314453125
],
[
0.509765625,
0.363525390625,
0.88330078125,
0.363525390625,
0.88330078125,
0.377197265625,
0.509765625,
0.377197265625
],
[
0.1158447265625,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.1158447265625,
0.388427734375
],
[
0.5087890625,
0.37890625,
0.8818359375,
0.377685546875,
0.8818359375,
0.392578125,
0.5087890625,
0.393798828125
],
[
0.1175537109375,
0.391357421875,
0.489501953125,
0.391357421875,
0.489501953125,
0.405029296875,
0.1175537109375,
0.405029296875
],
[
0.509765625,
0.3955078125,
0.8818359375,
0.394287109375,
0.8818359375,
0.4091796875,
0.509765625,
0.410400390625
],
[
0.1158447265625,
0.40673828125,
0.489501953125,
0.40673828125,
0.489501953125,
0.42041015625,
0.1158447265625,
0.42041015625
],
[
0.51123046875,
0.410888671875,
0.880859375,
0.410400390625,
0.880859375,
0.424560546875,
0.51123046875,
0.42578125
],
[
0.1175537109375,
0.4228515625,
0.490234375,
0.4228515625,
0.490234375,
0.435791015625,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.427490234375,
0.8818359375,
0.427490234375,
0.8818359375,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.1158447265625,
0.438232421875,
0.489501953125,
0.43701171875,
0.489501953125,
0.451904296875,
0.1158447265625,
0.453125
],
[
0.509765625,
0.443603515625,
0.8818359375,
0.443603515625,
0.8818359375,
0.457275390625,
0.509765625,
0.457275390625
],
[
0.1175537109375,
0.454833984375,
0.48779296875,
0.454833984375,
0.48779296875,
0.468505859375,
0.1175537109375,
0.468505859375
],
[
0.51123046875,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.473876953125,
0.51123046875,
0.473876953125
],
[
0.1175537109375,
0.471435546875,
0.489501953125,
0.471435546875,
0.489501953125,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.509765625,
0.4755859375,
0.8818359375,
0.4755859375,
0.8818359375,
0.4892578125,
0.509765625,
0.4892578125
],
[
0.1158447265625,
0.48681640625,
0.491943359375,
0.486328125,
0.491943359375,
0.50048828125,
0.1158447265625,
0.501953125
],
[
0.509765625,
0.4921875,
0.8818359375,
0.4921875,
0.8818359375,
0.505859375,
0.509765625,
0.505859375
],
[
0.1175537109375,
0.50439453125,
0.489501953125,
0.50439453125,
0.489501953125,
0.51708984375,
0.1175537109375,
0.51708984375
],
[
0.509765625,
0.50830078125,
0.880859375,
0.50830078125,
0.880859375,
0.521484375,
0.509765625,
0.521484375
],
[
0.1175537109375,
0.51953125,
0.48779296875,
0.51953125,
0.48779296875,
0.533203125,
0.1175537109375,
0.533203125
],
[
0.509765625,
0.52490234375,
0.796875,
0.52490234375,
0.796875,
0.5380859375,
0.509765625,
0.5380859375
],
[
0.118408203125,
0.53759765625,
0.486083984375,
0.53759765625,
0.486083984375,
0.54736328125,
0.118408203125,
0.54736328125
],
[
0.5263671875,
0.54052734375,
0.880859375,
0.54052734375,
0.880859375,
0.5546875,
0.5263671875,
0.5546875
],
[
0.118408203125,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.118408203125,
0.5654296875
],
[
0.51318359375,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.1175537109375,
0.56787109375,
0.489501953125,
0.56787109375,
0.489501953125,
0.58056640625,
0.1175537109375,
0.58056640625
],
[
0.51318359375,
0.57373046875,
0.80419921875,
0.57373046875,
0.80419921875,
0.58740234375,
0.51318359375,
0.58740234375
],
[
0.1175537109375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.51318359375,
0.59033203125,
0.88330078125,
0.59033203125,
0.88330078125,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.1158447265625,
0.599609375,
0.370361328125,
0.599609375,
0.370361328125,
0.6142578125,
0.1158447265625,
0.6142578125
],
[
0.52490234375,
0.60498046875,
0.884765625,
0.60791015625,
0.884765625,
0.62255859375,
0.52490234375,
0.619140625
],
[
0.52294921875,
0.62353515625,
0.767578125,
0.62353515625,
0.767578125,
0.63720703125,
0.52294921875,
0.63720703125
],
[
0.1351318359375,
0.63134765625,
0.486083984375,
0.63134765625,
0.486083984375,
0.64453125,
0.1351318359375,
0.64453125
],
[
0.509765625,
0.63818359375,
0.88330078125,
0.63818359375,
0.88330078125,
0.65185546875,
0.509765625,
0.65185546875
],
[
0.1158447265625,
0.646484375,
0.486083984375,
0.646484375,
0.486083984375,
0.6611328125,
0.1158447265625,
0.6611328125
],
[
0.5263671875,
0.65625,
0.884765625,
0.65625,
0.884765625,
0.66943359375,
0.5263671875,
0.66943359375
],
[
0.1158447265625,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.1158447265625,
0.677734375
],
[
0.5263671875,
0.67138671875,
0.81005859375,
0.67138671875,
0.81005859375,
0.68505859375,
0.5263671875,
0.68505859375
],
[
0.1158447265625,
0.6796875,
0.48779296875,
0.6796875,
0.48779296875,
0.693359375,
0.1158447265625,
0.693359375
],
[
0.51123046875,
0.68701171875,
0.884765625,
0.68701171875,
0.884765625,
0.70166015625,
0.51123046875,
0.70166015625
],
[
0.1158447265625,
0.6953125,
0.48779296875,
0.6953125,
0.48779296875,
0.708984375,
0.1158447265625,
0.708984375
],
[
0.5263671875,
0.70361328125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71728515625,
0.5263671875,
0.71728515625
],
[
0.1175537109375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.5263671875,
0.7197265625,
0.88330078125,
0.7197265625,
0.88330078125,
0.73388671875,
0.5263671875,
0.73388671875
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.52490234375,
0.736328125,
0.6533203125,
0.736328125,
0.6533203125,
0.75,
0.52490234375,
0.75
],
[
0.1158447265625,
0.74365234375,
0.486083984375,
0.7421875,
0.486083984375,
0.75732421875,
0.1158447265625,
0.75830078125
],
[
0.51318359375,
0.751953125,
0.88330078125,
0.751953125,
0.88330078125,
0.765625,
0.51318359375,
0.765625
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.52490234375,
0.76708984375,
0.880859375,
0.76708984375,
0.880859375,
0.78076171875,
0.52490234375,
0.78076171875
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.52490234375,
0.783203125,
0.88330078125,
0.783203125,
0.88330078125,
0.79736328125,
0.52490234375,
0.79736328125
],
[
0.1175537109375,
0.79345703125,
0.486083984375,
0.79345703125,
0.486083984375,
0.80419921875,
0.1175537109375,
0.80419921875
],
[
0.5263671875,
0.8017578125,
0.62744140625,
0.8017578125,
0.62744140625,
0.8134765625,
0.5263671875,
0.8134765625
],
[
0.1158447265625,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.51318359375,
0.81640625,
0.8818359375,
0.81640625,
0.8818359375,
0.830078125,
0.51318359375,
0.830078125
],
[
0.1175537109375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.5263671875,
0.8330078125,
0.8818359375,
0.8330078125,
0.8818359375,
0.8466796875,
0.5263671875,
0.8466796875
],
[
0.1158447265625,
0.84033203125,
0.486083984375,
0.84033203125,
0.486083984375,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.5263671875,
0.8486328125,
0.8818359375,
0.8486328125,
0.8818359375,
0.86328125,
0.5263671875,
0.86328125
],
[
0.1175537109375,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5263671875,
0.8642578125,
0.69775390625,
0.8642578125,
0.69775390625,
0.87890625,
0.5263671875,
0.87890625
],
[
0.11419677734375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.1158447265625,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.5322265625,
0.892578125,
0.8818359375,
0.89111328125,
0.8818359375,
0.90625,
0.5322265625,
0.90673828125
],
[
0.1175537109375,
0.90478515625,
0.4072265625,
0.90478515625,
0.4072265625,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.5087890625,
0.90380859375,
0.8564453125,
0.90478515625,
0.8564453125,
0.91943359375,
0.5087890625,
0.91845703125
]
] |
[
"2",
"Methodology",
"Zuccon, 2021a; MacAvaney et al., 2020; Hofstat-",
"which results in high query latency (Zhuang and",
"ter et al., 2020). Thus the use of these methods",
"With respect to BERT-based models for passage",
"is confined to second stage re-ranking.",
"retrieval and typos in queries, we investigate the",
"following research questions:",
": RQl: What is the impact of typos in queries on",
"Inprinciple.thesemethodsarenotaffectedby",
"keyword mismatch because they use the latent",
"BERT-based DR and re-ranker effectiveness?",
"embedding space to estimate the relevance of a",
": RQ2: Do different typo types affect the effect-",
"query to a passage. This is supported by recent",
"iveness of the BERT-based methods differently?",
"work that has shown the DR and BERT re-ranker",
". RQ3: Does the proposed typos-aware training",
"provide better semantic matching (Zhan et al.,",
"improve the effectiveness of the BERT-based",
"2020; MacAvaney et al., 2020; Formal et al., 2021).",
"methods on queries with typos? Does it hurt",
"their effectiveness on queries without typos?",
"In this paper we investigate the impact of a spe-",
"cific type of keyword mismatch: that caused by the",
"2.1 Synthetic Typo Generation",
"presence of typos in the query. Traditional exact",
" To answer our research questions, a reasonably",
"keyword matching methods perform badly on quer-",
"large set of queries with different types of typos",
"ies that contain typos. Extra query processing steps,",
"is needed. As there is no available dataset for pas-",
"such as spelling correction, are required for these",
" sage retrieval with labels that indicate the presence",
"methods to be tolerant to typos in queries (Martins",
"of typos in queries, we set off to create one such",
"and Silva, 2004). On the other hand, it is expected",
"dataset. For this we augmented the MS MARCO",
"that BERT-based models can handle typos occur-",
"passage retrieval dataset. (Manual inspection of",
"ring in queries well. This is because BERT uses the",
"this dataset did not reveal a considerable amount of",
"WordPiece algorithm which splits a keyword that",
"queries with typos; the dataset curators likely did",
"does not match an entry in the BERT vocabulary",
"manually remove most typos). For augmentation",
"(typos are likely to not be present in this vocabu-",
"we synthetically generated typos from the original",
"lary) into character-level sub-tokens. This can be",
"queries in the dataset, so that we could carefully",
"used to produce embeddings for out-of-vocabulary",
"control the number and types of typos.",
"keywords.which are then passed as input to the",
"For generating typos, we used the following op",
"BERT encoder. However, this intuition has never",
"erations that give rise to typos that often occur in",
"been tested before, and the capability of BERT-",
"real-world queries (Hagen et al., 2017):",
"based passage ranking models to deal with typos",
": Random character Insertion (RandInsert):",
"in queries has not been quantified.",
"Inserts a random letter into a random word, e.g.,",
"\"search typo\"->“\"search tyapo\".",
"To address this gap, we first formally investigate",
": Random character deletion (RandDelete):",
"how the BERT-based DR and re-ranker respond and",
"Deletes a random character of a random word,",
"adapt to queries that contain typos. Specifically, we",
"such as “search typo\" -> \"search tpo\".",
"use different typo generators to produce typos for",
". Random character substitution (RandSub):",
"queries; we then compare the effectiveness of the",
"Randomly replaces a character of a random",
"rankers when using queries with typos vs. without",
"word with a random letter, e.g., “search typo\"",
"typos. Interestingly, we find that these models fail",
"->“\"search type\".",
"to handle queries with typos - typos can lead to a",
": Swap neighbor character (SwapNeighbor):",
"significant drop in effectiveness for both DR and",
"Randomly swaps a character with one of its",
"BERT re-ranker. In order to solve this issue and ob-",
"neighbor characters, e.g., “search typo\"->",
"tain typo-robust ranking models, wethenpropose a",
"“\"searchtyop\"",
"simple typos-aware training strategy, in which quer",
". Swap adjacent keyboard character (SwapAd-",
"ies with typos are produced and used also for train-",
"jacent): Randomly swaps a character with one of",
"ing. Our experimental results on the MS MARCO",
"its adjacent letter on the keyboard', e.g., “\"search",
"passage ranking dataset show that, with our typos-",
"typo\"->\"search typi\".",
" aware training, DR and BERT re-ranker can be-",
"come robust to typos in queries, without loss in",
"'e.g, on a QWERTY keyboard, the list of adjacent charac-",
"effectiveness for queries without typos.",
"ters for character's' is ['q',\"w','e', 'a', 'd', 'z, 'x', 'c']."
] |
{
"arxiv_id": "2108.12139",
"language": "en",
"page_idx": 1
}
|
||
2108.12139-en-2
|
2108.12139-en-2.jpg
|
Since queries in MS MARCO are relatively short $(\approx6$ keywords on average) (Nguyen et al., 2016), when generating typos for queries, we only consider keywords that have more than 3 characters and only randomly modify one keyword per query. We use open-source tool kits TextAttack (Morris et al., 2020) to implement these typo generators.
## 2.2 Typos-aware Training
To deal with queries with typos we propose to consider such queries also during the training phase of DR and BERT re-ranker: we call this typos-aware training. Specifically, for each original query that appears during the training phase, we draw an unbiased coin. If the result is head, we leave the query unchanged and use it for training. If it is tail (50% chances) we inject a typo in the query by uniformly sampling one of the considered typos generators (Section 2.1), and use the modified query for training. By doing so, at training time, the BERT-based methods will observe both the original, typos-free, queries and queries with different types of typos. Thus, in order to reduce the training loss, we force the methods to learn to be invariant to different types of typos.
Our typos-aware training can be considered a data augmentation approach, with small perturbations to some training queries: these do not change the underlying intent of the query or the relevance of the target passage. Data augmentation has been shown effective for a range of deep learning tasks, including computer vision (He et al., 2020; Chen et al., 2020; Grill et al., 2020) and NLP (Zhang et al., 2015; Wei and Zou, 2019; Xie et al., 2020; Jiao et al., 2020); however, the impact of data augmentation on ad-hoc retrieval remains to be studied.
## 3 Experimental Settings
3.1 Dataset and Evaluation Measures
For evaluation, we use the MS MARCO passage ranking dataset (Nguyen et al., 2016), which consists of 8.8M passages, $\approx503$K training queries and 6,980 dev queries. For typos-aware training, we modify training queries with a 50% chance. For dev queries, we experiment with both the original queries and with queries modified to contain typos. We produce typos for all queries in the dev set using the strategies in Section 2.1; we also consider the average effectiveness across typos queries.
We use the official metric MRR@ 10 to evaluate the ranking effectiveness of both DR and BERT
re-ranker. We use the BERT re-ranker as a second stage ranker, on top of the initial rankings provided by DR. This is unlike previous work (Lin et al., 2020), in which the BERT re-ranker is typically used on top of BM25. Our setting is motivated by the fact that BM25 would fail to retrieve the relevant target passages for queries that contain typos. Because of this, we also report Recall@ 1000 (labelled Recall) for DR, as this forms the basis of the first stage of retrieval and the number of retrieved relevant passages affects the effectiveness of the BERT re-ranker. Recall for BERT re-ranker is thus the same as that of DR, and is not reported.
## 3.2 DR Training Details
We follow Zhan et al. (2020) when training the DR. We adopt the BERT-Siamese architecture in which the query encoder and passage encoder share the BERT model parameters. This architecture has been used consistently in many recent approaches (Luan et al., 2020; Xiong et al., 2020). We use pairwise hinge loss with the "Train Triples" data provided in MS MARCO to fine-tune the “bertbase-uncased" model from the Huggingface library (Wolf et al., 2020). We use the ADAM optimizer, learning rate of 3e-6 with linear warm-up and decay scheduling. The model is trained on a single Tesla V100 GPU with a batch size of 26 and gradient accumulation step of 2 for 210K steps.
3.3 BERT re-ranker Training Details
To train the BERT re-ranker, we follow the training practice described by Nogueira and Cho (2019). We fine-tune a“bert-large-uncased" model with binary cross-entropy loss to perform binary classification on query-passage pairs. Negative pairs are randomly sampled from the top 1,000 passages retrieved by a trained DR model (without typosaware training). We set the ratio of positive pairs to negative pairs to 1:4. We use the same optimizer and learning rate scheduling used for DR; the model is trained on two Tesla V100 GPUs with a batch size of $2\times64$ for 70K steps.
For both DR and BERT re-ranker typos-aware training, we use exactly the same setting used for the standard training described above.
## 4 Results
Empirical results are reported in Table 1. We note that BM25 if outperformed by both DR and BERT re-ranker across all settings, confirming the superi-
|
[
[
0.1175537109375,
0.0843505859375,
0.490234375,
0.08489990234375,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09857177734375
],
[
0.50732421875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.50732421875,
0.09979248046875
],
[
0.1158447265625,
0.1015625,
0.489501953125,
0.1009521484375,
0.489501953125,
0.11517333984375,
0.1158447265625,
0.11639404296875
],
[
0.51123046875,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.51123046875,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.490234375,
0.1181640625,
0.490234375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.509765625,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.509765625,
0.1318359375
],
[
0.1158447265625,
0.133056640625,
0.48779296875,
0.133056640625,
0.48779296875,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.509765625,
0.1341552734375,
0.880859375,
0.1341552734375,
0.880859375,
0.1478271484375,
0.509765625,
0.1478271484375
],
[
0.1175537109375,
0.15087890625,
0.489501953125,
0.15087890625,
0.489501953125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.509765625,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.163330078125,
0.509765625,
0.163330078125
],
[
0.1158447265625,
0.1650390625,
0.48779296875,
0.1650390625,
0.48779296875,
0.1787109375,
0.1158447265625,
0.1787109375
],
[
0.509765625,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.509765625,
0.179931640625
],
[
0.1158447265625,
0.18115234375,
0.474365234375,
0.1817626953125,
0.474365234375,
0.196533203125,
0.1158447265625,
0.1953125
],
[
0.509765625,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.51123046875,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.1158447265625,
0.210205078125,
0.333251953125,
0.210205078125,
0.333251953125,
0.2244873046875,
0.1158447265625,
0.2244873046875
],
[
0.51123046875,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.11419677734375,
0.2298583984375,
0.491943359375,
0.23095703125,
0.491943359375,
0.2452392578125,
0.11419677734375,
0.24462890625
],
[
0.509765625,
0.2298583984375,
0.8818359375,
0.23095703125,
0.8818359375,
0.2452392578125,
0.509765625,
0.24462890625
],
[
0.1175537109375,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.1175537109375,
0.262451171875
],
[
0.50732421875,
0.2476806640625,
0.880859375,
0.2476806640625,
0.880859375,
0.26123046875,
0.50732421875,
0.26123046875
],
[
0.1158447265625,
0.262451171875,
0.489501953125,
0.26318359375,
0.489501953125,
0.27783203125,
0.1158447265625,
0.276611328125
],
[
0.509765625,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.509765625,
0.276611328125
],
[
0.1158447265625,
0.27978515625,
0.489501953125,
0.27978515625,
0.489501953125,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.509765625,
0.27978515625,
0.88330078125,
0.27978515625,
0.88330078125,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1158447265625,
0.29638671875,
0.489501953125,
0.295654296875,
0.489501953125,
0.31005859375,
0.1158447265625,
0.311279296875
],
[
0.1158447265625,
0.311279296875,
0.48779296875,
0.312255859375,
0.48779296875,
0.32666015625,
0.1158447265625,
0.325927734375
],
[
0.51318359375,
0.30712890625,
0.7138671875,
0.30712890625,
0.7138671875,
0.320556640625,
0.51318359375,
0.320556640625
],
[
0.1175537109375,
0.328857421875,
0.48779296875,
0.328857421875,
0.48779296875,
0.342529296875,
0.1175537109375,
0.342529296875
],
[
0.5087890625,
0.32666015625,
0.8818359375,
0.328857421875,
0.8818359375,
0.34326171875,
0.5087890625,
0.341552734375
],
[
0.1175537109375,
0.344482421875,
0.489501953125,
0.344482421875,
0.489501953125,
0.359375,
0.1175537109375,
0.359375
],
[
0.509765625,
0.345703125,
0.8818359375,
0.345703125,
0.8818359375,
0.359375,
0.509765625,
0.359375
],
[
0.1175537109375,
0.361083984375,
0.489501953125,
0.361083984375,
0.489501953125,
0.374755859375,
0.1175537109375,
0.374755859375
],
[
0.51123046875,
0.3623046875,
0.8818359375,
0.3623046875,
0.8818359375,
0.3759765625,
0.51123046875,
0.3759765625
],
[
0.1175537109375,
0.3759765625,
0.489501953125,
0.3759765625,
0.489501953125,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.51123046875,
0.377685546875,
0.8818359375,
0.377685546875,
0.8818359375,
0.392578125,
0.51123046875,
0.392578125
],
[
0.1175537109375,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.1175537109375,
0.40625
],
[
0.509765625,
0.394287109375,
0.8818359375,
0.394287109375,
0.8818359375,
0.407958984375,
0.509765625,
0.407958984375
],
[
0.1158447265625,
0.4091796875,
0.489501953125,
0.4091796875,
0.489501953125,
0.4228515625,
0.1158447265625,
0.4228515625
],
[
0.51123046875,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.1175537109375,
0.424560546875,
0.490234375,
0.424560546875,
0.490234375,
0.438232421875,
0.1175537109375,
0.438232421875
],
[
0.51123046875,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.51123046875,
0.440673828125
],
[
0.118408203125,
0.4423828125,
0.486083984375,
0.4423828125,
0.486083984375,
0.453125,
0.118408203125,
0.453125
],
[
0.51318359375,
0.4423828125,
0.8818359375,
0.4423828125,
0.8818359375,
0.4560546875,
0.51318359375,
0.4560546875
],
[
0.1175537109375,
0.457275390625,
0.489501953125,
0.457275390625,
0.489501953125,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.509765625,
0.458984375,
0.88330078125,
0.458984375,
0.88330078125,
0.47265625,
0.509765625,
0.47265625
],
[
0.118408203125,
0.474365234375,
0.2275390625,
0.474365234375,
0.2275390625,
0.486328125,
0.118408203125,
0.486328125
],
[
0.51123046875,
0.474365234375,
0.88330078125,
0.474365234375,
0.88330078125,
0.488037109375,
0.51123046875,
0.488037109375
],
[
0.1351318359375,
0.4892578125,
0.490234375,
0.4892578125,
0.490234375,
0.50439453125,
0.1351318359375,
0.50439453125
],
[
0.51123046875,
0.490478515625,
0.8818359375,
0.490478515625,
0.8818359375,
0.5048828125,
0.51123046875,
0.5048828125
],
[
0.1175537109375,
0.5048828125,
0.491943359375,
0.5048828125,
0.491943359375,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.51123046875,
0.50830078125,
0.880859375,
0.50830078125,
0.880859375,
0.5185546875,
0.51123046875,
0.5185546875
],
[
0.11419677734375,
0.5224609375,
0.48779296875,
0.5224609375,
0.48779296875,
0.5361328125,
0.11419677734375,
0.5361328125
],
[
0.509765625,
0.5224609375,
0.8818359375,
0.521484375,
0.8818359375,
0.5361328125,
0.509765625,
0.53759765625
],
[
0.11419677734375,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.55029296875,
0.11419677734375,
0.55029296875
],
[
0.51123046875,
0.5390625,
0.8623046875,
0.5390625,
0.8623046875,
0.55419921875,
0.51123046875,
0.55419921875
],
[
0.11419677734375,
0.55419921875,
0.486083984375,
0.552734375,
0.486083984375,
0.56787109375,
0.11419677734375,
0.568359375
],
[
0.1175537109375,
0.57080078125,
0.490234375,
0.57080078125,
0.490234375,
0.58447265625,
0.1175537109375,
0.58447265625
],
[
0.51318359375,
0.56787109375,
0.81689453125,
0.56787109375,
0.81689453125,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.1175537109375,
0.5849609375,
0.489501953125,
0.5849609375,
0.489501953125,
0.5986328125,
0.1175537109375,
0.5986328125
],
[
0.50537109375,
0.58837890625,
0.880859375,
0.5888671875,
0.880859375,
0.60400390625,
0.50537109375,
0.6025390625
],
[
0.1175537109375,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.615234375,
0.1175537109375,
0.615234375
],
[
0.5087890625,
0.60498046875,
0.880859375,
0.60498046875,
0.880859375,
0.6181640625,
0.5087890625,
0.6181640625
],
[
0.1158447265625,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.63134765625,
0.1158447265625,
0.63134765625
],
[
0.509765625,
0.62158203125,
0.8818359375,
0.62158203125,
0.8818359375,
0.6357421875,
0.509765625,
0.6357421875
],
[
0.1158447265625,
0.6328125,
0.489501953125,
0.63427734375,
0.489501953125,
0.6484375,
0.1158447265625,
0.64794921875
],
[
0.509765625,
0.63818359375,
0.88330078125,
0.63818359375,
0.88330078125,
0.65185546875,
0.509765625,
0.65185546875
],
[
0.1158447265625,
0.65087890625,
0.489501953125,
0.65087890625,
0.489501953125,
0.66455078125,
0.1158447265625,
0.66455078125
],
[
0.5087890625,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66748046875,
0.5087890625,
0.66748046875
],
[
0.51318359375,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.11248779296875,
0.67578125,
0.341796875,
0.677734375,
0.341796875,
0.6953125,
0.11248779296875,
0.693359375
],
[
0.5087890625,
0.68505859375,
0.880859375,
0.68505859375,
0.880859375,
0.69873046875,
0.5087890625,
0.69873046875
],
[
0.1175537109375,
0.705078125,
0.4248046875,
0.705078125,
0.4248046875,
0.71875,
0.1175537109375,
0.71875
],
[
0.51123046875,
0.703125,
0.8818359375,
0.703125,
0.8818359375,
0.71630859375,
0.51123046875,
0.71630859375
],
[
0.509765625,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.73095703125,
0.509765625,
0.73095703125
],
[
0.1175537109375,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.1175537109375,
0.74072265625
],
[
0.509765625,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.509765625,
0.7490234375
],
[
0.1158447265625,
0.74365234375,
0.489501953125,
0.74365234375,
0.489501953125,
0.75732421875,
0.1158447265625,
0.75732421875
],
[
0.51318359375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.765625,
0.76904296875,
0.76708984375,
0.76904296875,
0.78076171875,
0.51123046875,
0.779296875
],
[
0.1175537109375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.5283203125,
0.7822265625,
0.8818359375,
0.783203125,
0.8818359375,
0.79736328125,
0.5283203125,
0.79638671875
],
[
0.1158447265625,
0.79150390625,
0.48779296875,
0.79052734375,
0.48779296875,
0.8046875,
0.1158447265625,
0.8056640625
],
[
0.51318359375,
0.7998046875,
0.88330078125,
0.7998046875,
0.88330078125,
0.8134765625,
0.51318359375,
0.8134765625
],
[
0.1158447265625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.51318359375,
0.8154296875,
0.79443359375,
0.8154296875,
0.79443359375,
0.830078125,
0.51318359375,
0.830078125
],
[
0.1158447265625,
0.822265625,
0.490234375,
0.82373046875,
0.490234375,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.1158447265625,
0.83837890625,
0.489501953125,
0.8388671875,
0.489501953125,
0.85400390625,
0.1158447265625,
0.8525390625
],
[
0.51318359375,
0.84423828125,
0.609375,
0.84423828125,
0.609375,
0.859375,
0.51318359375,
0.859375
],
[
0.1158447265625,
0.85546875,
0.48779296875,
0.85546875,
0.48779296875,
0.869140625,
0.1158447265625,
0.869140625
],
[
0.11419677734375,
0.87060546875,
0.459228515625,
0.87158203125,
0.459228515625,
0.8857421875,
0.11419677734375,
0.88525390625
],
[
0.509765625,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1351318359375,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90185546875,
0.1351318359375,
0.90185546875
],
[
0.509765625,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90283203125,
0.48779296875,
0.90380859375,
0.48779296875,
0.91845703125,
0.1158447265625,
0.91748046875
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"Since queries in MS MARCO are relatively short",
"re-ranker. We use the BERT re-ranker as a second",
"(~ 6 keywords on average) (Nguyen et al., 2016),",
"stage ranker, on top of the initial rankings provided",
"when generating typos for queries, we only con-",
"by DR. This is unlike previous work (Lin et al.",
"sider keywords that have more than 3 characters",
"2020), in which the BERT re-ranker is typically",
"and only randomly modify one keyword per query.",
"used on top of BM25. Our setting is motivated",
"We use open-source tool kits TextAttack (Morris",
"by the fact that BM25 would fail to retrieve the",
"et al., 2020) to implement these typo generators.",
"relevant target passages for queries that contain",
"typos. Because of this, we also report Recall @ 1000",
"2.2Typos-aware Training",
"(labelled Recall) for DR, as this forms the basis",
"To deal with queries with typos we propose to con-",
"of the first stage of retrieval and the number of",
"sider such queries also during the training phase of",
" retrieved relevant passages affects the effectiveness",
"DR and BERT re-ranker: we call this typos-aware",
"of the BERT re-ranker. Recall for BERT re-ranker",
"training. Specifically, for each original query that",
"is thus the same as that of DR, and is not reported.",
"appears during the training phase, we draw an un-",
"biased coin. If the result is head, we leave the query",
"3.2 DR Training Details",
"unchanged and use it for training. If it is tail (50%",
"We follow Zhan et al. (2020) when training the",
"chances) we inject a typo in the query by uniformly",
"DR. We adopt the BERT-Siamese architecture in",
"sampling one of the considered typos generators",
"which the query encoder and passage encoder share",
"(Section 2.1), and use the modified query for train-",
"the BERT model parameters. This architecture",
"ing. By doing so, at training time, the BERT-based",
"has been used consistently in many recent ap",
"methods will observe both the original, typos-free,",
"proaches (Luan et al., 2020; Xiong et al., 2020)",
"queries and queries with different types of typos.",
"We use pairwise hinge loss with the \"Train Triples\"",
"Thus,in order toreduce the training loss, we force",
"data provided in MS MARCO to fine-tune the “bert-",
"the methods to learn to be invariant to different",
"base-uncased\" model from the Huggingface lib-",
"types of typos.",
"rary (Wolf et al., 2020). We use the ADAM op-",
"Our typos-aware training can be considered a",
"timizer, learning rate of 3e-6 with linear warm-up",
"data augmentation approach, with small perturba-",
"and decay scheduling.The model is trained on a",
"tions to some training queries: these do not change",
"single Tesla V100 GPU with a batch size of 26 and",
"the underlying intent of the query or the relevance",
"gradient accumulation step of 2 for 210K steps.",
"of the target passage. Data augmentation has been",
"shown effective for a range of deep learning tasks,",
"3.3 BERT re-ranker Training Details",
"including computer vision (He et al., 2020: Chen",
"Totrain the BERTre-ranker, we follow thetraining",
"et al., 2020; Grill et al., 2020) and NLP (Zhang",
"practice described by Nogueira and Cho (2019)",
"et al., 2015; Wei and Zou, 2019; Xie et al., 2020;",
"We fine-tune a “bert-large-uncased\" model with",
"Jiao et al., 2020); however, the impact of data aug-",
"binary cross-entropy loss to perform binary clas-",
"mentation on ad-hoc retrieval remains to be studied.",
" sification on query-passage pairs. Negative pairs",
"are randomly sampled from the top 1,o00 passages",
"3 Experimental Settings",
"retrieved by a trained DR model (without typos",
"3.1 Dataset and Evaluation Measures",
"aware training). We set the ratio of positive pairs",
"to negative pairs to l:4. We use the same optim",
"For evaluation, we use the MS MARCO passage",
"izer and learning rate scheduling used for DR; the",
"ranking dataset (Nguyen et al., 2016), which con-",
"model is trained on two Tesla V100 GPUs with a",
"sists of 8.8M passages, ~503K training queries and",
"batch size of 2 × 64 for 70K steps.",
"6,980 dev queries. For typos-aware training, we",
"For both DR and BERT re-ranker typos-aware",
"modify training queries with a 50% chance. For",
"training, we use exactly the same setting used for",
"dev queries, we experiment with both the original",
"the standard training described above.",
"queries and with queries modified to contain typos.",
"We produce typos for all queries in the dev set us-",
"4Results",
"ing the strategies in Section 2.1; we also consider",
"the average effectiveness across typos queries.",
"Empirical results are reported in Table 1. We note",
"We use the official metric MRR@ 10 to evaluate",
"that BM25 if outperformed by both DR and BERT",
"the ranking effectiveness of both DR and BERT",
"re-ranker across all settings, confirming the superi-"
] |
{
"arxiv_id": "2108.12139",
"language": "en",
"page_idx": 2
}
|
||
2108.12139-en-3
|
2108.12139-en-3.jpg
|
Table l: MS MARCO passage ranking results. Row 2 reports results averaged across all typos queries; rows 3-7 results for each typos type (for each type, typos are injected in all dev queries). Percentage reductions are computed w.r.t. the original queries; bold represents best performance across training methods for each of DR and BERT re-ranker. Statistical significant gains (two-tailed paired t-test with Bonferroni correction, $p<0.01)$ obtained by models with typos-aware training over the models with standard training (std.) are indicated by $\dagger.$
<table>
<tbody>
<tr>
<th rowspan="2">Typo type</th>
<th colspan="2">$BM25$</th>
<th colspan="2">$\mathbf{DR}\left(\mathbf{std.}\right)$</th>
<th colspan="2">$DR\left(\mathbf{typos-aware}\right)$</th>
<th rowspan="2">Re- ranker ( std. ) MRR@ 10</th>
<th rowspan="2">$\mathbf{Re-ranker}$ $(\mathbf{typos-aware})$ $\mathbf{MRR@}10$</th>
</tr>
<tr>
<th>MRR@ 10</th>
<th>Recall</th>
<th>$\mathbf{MRR@10}$</th>
<th>Recall</th>
<th>$\mathbf{MRR@10}$</th>
<th>Recall</th>
</tr>
<tr>
<td>original</td>
<td>.187</td>
<td>.857</td>
<td>.296</td>
<td>.940</td>
<td>.300</td>
<td>.940</td>
<td>.379</td>
<td>.374</td>
</tr>
<tr>
<td>$W.$ typos (avg)</td>
<td>$.120(-35.8\%)$</td>
<td>$.696(-18.6\%)$</td>
<td>$.141(-52.3\%)$</td>
<td>$.712(-24.3\%)$</td>
<td>$.219^{\dagger}(-27.0\%)$</td>
<td>$.857^{\dagger}(-8.8\%)$</td>
<td>$.250(-34.0\%)$</td>
<td>$.289^{\dagger}(-22.7\%)$</td>
</tr>
<tr>
<td>RandInsert</td>
<td>$.125(-33.1\%)$</td>
<td>$.693(-18.9\%)$</td>
<td>$.140(-52.7\%)$</td>
<td>$.711(-24.4\%)$</td>
<td>$.225^{\dagger}(-25.0\%)$</td>
<td>$.862^{\dagger}(-8.3\%)$</td>
<td>$.257(-32.2\%)$</td>
<td>$.297^{\dagger}(-20.6\%)$</td>
</tr>
<tr>
<td>${\mathrm{RandDelete}}$</td>
<td>$.118(-36.9\%)$</td>
<td>$.693(-18.9\%)$</td>
<td>$.154(-47.9\%)$</td>
<td>$.730(-22.3\%)$</td>
<td>$.217^{\dagger}(-27.6\%)$</td>
<td>$.853^{\dagger}(-9.3\%)$</td>
<td>$.257(-32.2\%)$</td>
<td>$.288^{\dagger}(-23.0\%$</td>
</tr>
<tr>
<td>RandSub</td>
<td>$.120(-35.8\%)$</td>
<td>$.702(-17.9\%)$</td>
<td>$.137(-53.7\%)$</td>
<td>$.714(-24.0\%)$</td>
<td>$.220^{\dagger}(-26.7\%)$</td>
<td>$.858^{\dagger}(-8.7\%)$</td>
<td>$.250(-34.0\%)$</td>
<td>$.291^{\dagger}(-22.2\%)$</td>
</tr>
<tr>
<td>SwapNeighbor</td>
<td>$.122(-34.7\%)$</td>
<td>$.702(-17.9\%)$</td>
<td>$.137(-53.7\%)$</td>
<td>$.705(-25.0\%)$</td>
<td>$.217^{\dagger}(-27.6\%)$</td>
<td>$.859^{\dagger}(-8.6\%)$</td>
<td>$.240(-36.7\%)$</td>
<td>$.284^{\dagger}(-24.1\%)$</td>
</tr>
<tr>
<td>SwapAdjacent</td>
<td>$.117(-37.4\%)$</td>
<td>$.691(-19.1\%)$</td>
<td>$.137(-53.7\%)$</td>
<td>$.702(-25.3\%)$</td>
<td>$.214^{\dagger}(-28.7\%)$</td>
<td>$.854^{\dagger}(-9.1\%)$</td>
<td>$.246(-35.1\%)$</td>
<td>$.286^{\dagger}(-23.5\%)$</td>
</tr>
</tbody>
</table>
<FigureHere>
Figure 1: Loss in terms of the rank position of the first relevant passage retrieved by DR when ranking for typos queries, compared to original queries. Each point on the x-axis refers to a query; x-axis ordered by decreasing loss when standard training is used.
ority of BERT-based methods. For RQ1, we compare the results obtained on the original queries with those on queries with typos, when models are trained using the standard procedure. We observe statistically significant losses in effectiveness for both DR (on average MRR@10 drops $52.3\%$ and Recall 24.3%) and BERT re-ranker (MRR@ 10 drops 34%). BERT re-ranker is performed on top of DR results, thus losses in Recall for DR are propagated to the BERT re-ranker. However, the effectiveness of DR on typos queries drops to about that of BM25, while BERT re-ranker stays superior.
In terms of the impact of different types of typos (RQ2), the results show that different typos have similar impact: they all hurt effectiveness heavily. DR appears however more tolerant to RandDelete typos (with a$\approx5\%$ smaller loss in MRR@10 than for other types of typos), while BERT re-ranker losses are generally uniform across typos types.
To answer RQ3, we compare the results of models produced with typos-aware training vs. with the standard training. Despite the typos-aware training, both DR and BERT re-ranker display significant losses in effectiveness when dealing with typos queries, compared to the original queries (Table 1). However, compared to the models with standard training, both methods are much more tolerant to all types of typos when typos-aware training is employed. In fact, losses in MRR@ 10 halve for DR (from 52.3% to 24.3%), and reduce by one third for BERT re-ranker (from 34% to 22.7%); all differences are statistically significant. Typosaware training seems to impact queries with typos produced by RandomInsert more than those with other types of typos. We also note that effectiveness obtained by models with typos-aware training is not different from that with standard training if
only queries without typos (original) are considered (minor differences are not statistically significant).
Figure 1 presents the rank loss obtained by DR when answering typos queries in place of the original queries (plot averaged across all types of typos; individual typos types show similar trends). A negative loss of $n$ means when using typos queries, the first relevant document is retrieved $n$ rank positions after that obtained when using the original queries. The figure shows that typos-aware training consistently provides smaller losses than the standard training. We also note there are few cases $(\approx300$ queries) in which typos queries provide gains compared to the original query. Queries with large losses often have typos for keywords that are essential to determine the intent of the query. Typos queries that exhibit gains generally display typos on non-essential keywords, e.g., stopwords
## 5 A Case Study
The results presented in the previous sections are conducted with synthetically generated typo queries. Accurate analysis of the MS MARCO dataset revels the presence of a very limit number of
|
[
[
0.1158447265625,
0.0831298828125,
0.8818359375,
0.0831298828125,
0.8818359375,
0.0968017578125,
0.1158447265625,
0.0968017578125
],
[
0.1175537109375,
0.097412109375,
0.8818359375,
0.097412109375,
0.8818359375,
0.11102294921875,
0.1175537109375,
0.11102294921875
],
[
0.1158447265625,
0.11224365234375,
0.8818359375,
0.11224365234375,
0.8818359375,
0.1258544921875,
0.1158447265625,
0.1258544921875
],
[
0.1158447265625,
0.1258544921875,
0.88330078125,
0.1258544921875,
0.88330078125,
0.1395263671875,
0.1158447265625,
0.1395263671875
],
[
0.1158447265625,
0.1395263671875,
0.85205078125,
0.1395263671875,
0.85205078125,
0.152587890625,
0.1158447265625,
0.152587890625
],
[
0.270263671875,
0.168701171875,
0.307373046875,
0.168701171875,
0.307373046875,
0.18115234375,
0.270263671875,
0.18115234375
],
[
0.426513671875,
0.169189453125,
0.480224609375,
0.169189453125,
0.480224609375,
0.18115234375,
0.426513671875,
0.18115234375
],
[
0.57080078125,
0.169189453125,
0.66650390625,
0.169189453125,
0.66650390625,
0.18115234375,
0.57080078125,
0.18115234375
],
[
0.705078125,
0.169189453125,
0.796875,
0.169189453125,
0.796875,
0.18115234375,
0.705078125,
0.18115234375
],
[
0.81005859375,
0.169189453125,
0.87158203125,
0.169189453125,
0.87158203125,
0.179931640625,
0.81005859375,
0.179931640625
],
[
0.8076171875,
0.18115234375,
0.880859375,
0.18115234375,
0.880859375,
0.188232421875,
0.8076171875,
0.188232421875
],
[
0.1234130859375,
0.1923828125,
0.177978515625,
0.1923828125,
0.177978515625,
0.2037353515625,
0.1234130859375,
0.2037353515625
],
[
0.208251953125,
0.1923828125,
0.264404296875,
0.1923828125,
0.264404296875,
0.203125,
0.208251953125,
0.203125
],
[
0.3720703125,
0.1923828125,
0.42822265625,
0.1923828125,
0.42822265625,
0.203125,
0.3720703125,
0.203125
],
[
0.456787109375,
0.1923828125,
0.491943359375,
0.1923828125,
0.491943359375,
0.2037353515625,
0.456787109375,
0.2037353515625
],
[
0.53564453125,
0.1923828125,
0.59375,
0.1923828125,
0.59375,
0.2037353515625,
0.53564453125,
0.2037353515625
],
[
0.62646484375,
0.1923828125,
0.66015625,
0.1923828125,
0.66015625,
0.2037353515625,
0.62646484375,
0.2037353515625
],
[
0.70361328125,
0.1923828125,
0.76171875,
0.1923828125,
0.76171875,
0.203125,
0.70361328125,
0.203125
],
[
0.79833984375,
0.188232421875,
0.85791015625,
0.190673828125,
0.8564453125,
0.204833984375,
0.796875,
0.203125
],
[
0.2091064453125,
0.204833984375,
0.2376708984375,
0.204833984375,
0.2376708984375,
0.21728515625,
0.2091064453125,
0.21728515625
],
[
0.2939453125,
0.193603515625,
0.32568359375,
0.193603515625,
0.32568359375,
0.201904296875,
0.2939453125,
0.201904296875
],
[
0.53662109375,
0.204833984375,
0.56689453125,
0.204833984375,
0.56689453125,
0.21728515625,
0.53662109375,
0.21728515625
],
[
0.62646484375,
0.204833984375,
0.654296875,
0.204833984375,
0.654296875,
0.21728515625,
0.62646484375,
0.21728515625
],
[
0.70361328125,
0.204833984375,
0.73486328125,
0.204833984375,
0.73486328125,
0.21728515625,
0.70361328125,
0.21728515625
],
[
0.29296875,
0.2060546875,
0.32080078125,
0.2060546875,
0.32080078125,
0.21728515625,
0.29296875,
0.21728515625
],
[
0.1234130859375,
0.2060546875,
0.1695556640625,
0.2060546875,
0.1695556640625,
0.21728515625,
0.1234130859375,
0.21728515625
],
[
0.372802734375,
0.2060546875,
0.40478515625,
0.2060546875,
0.40478515625,
0.21728515625,
0.372802734375,
0.21728515625
],
[
0.45751953125,
0.2060546875,
0.484375,
0.2060546875,
0.484375,
0.21728515625,
0.45751953125,
0.21728515625
],
[
0.8017578125,
0.2060546875,
0.82958984375,
0.2060546875,
0.82958984375,
0.21728515625,
0.8017578125,
0.21728515625
],
[
0.124267578125,
0.21728515625,
0.20068359375,
0.21728515625,
0.20068359375,
0.2286376953125,
0.124267578125,
0.2286376953125
],
[
0.210693359375,
0.21728515625,
0.28955078125,
0.21728515625,
0.28955078125,
0.22802734375,
0.210693359375,
0.22802734375
],
[
0.28955078125,
0.21728515625,
0.36865234375,
0.21728515625,
0.36865234375,
0.22802734375,
0.28955078125,
0.22802734375
],
[
0.37451171875,
0.21728515625,
0.450927734375,
0.21728515625,
0.450927734375,
0.22802734375,
0.37451171875,
0.22802734375
],
[
0.45751953125,
0.21728515625,
0.5341796875,
0.21728515625,
0.5341796875,
0.22802734375,
0.45751953125,
0.22802734375
],
[
0.5380859375,
0.21728515625,
0.62109375,
0.21728515625,
0.62109375,
0.22802734375,
0.5380859375,
0.22802734375
],
[
0.62646484375,
0.21728515625,
0.7021484375,
0.21728515625,
0.7021484375,
0.22802734375,
0.62646484375,
0.22802734375
],
[
0.7060546875,
0.21728515625,
0.78076171875,
0.21728515625,
0.78076171875,
0.22802734375,
0.7060546875,
0.22802734375
],
[
0.8037109375,
0.2161865234375,
0.8818359375,
0.2161865234375,
0.8818359375,
0.22802734375,
0.8037109375,
0.22802734375
],
[
0.1234130859375,
0.23095703125,
0.183837890625,
0.23095703125,
0.183837890625,
0.2423095703125,
0.1234130859375,
0.2423095703125
],
[
0.28955078125,
0.23095703125,
0.370361328125,
0.23095703125,
0.370361328125,
0.2423095703125,
0.28955078125,
0.2423095703125
],
[
0.53662109375,
0.23095703125,
0.62060546875,
0.23095703125,
0.62060546875,
0.2423095703125,
0.53662109375,
0.2423095703125
],
[
0.7060546875,
0.23095703125,
0.78076171875,
0.23095703125,
0.78076171875,
0.2423095703125,
0.7060546875,
0.2423095703125
],
[
0.8037109375,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.2423095703125,
0.8037109375,
0.2423095703125
],
[
0.3720703125,
0.232177734375,
0.450927734375,
0.232177734375,
0.450927734375,
0.2423095703125,
0.3720703125,
0.2423095703125
],
[
0.2091064453125,
0.232177734375,
0.287109375,
0.232177734375,
0.287109375,
0.2423095703125,
0.2091064453125,
0.2423095703125
],
[
0.455078125,
0.232177734375,
0.5341796875,
0.232177734375,
0.5341796875,
0.2423095703125,
0.455078125,
0.2423095703125
],
[
0.62451171875,
0.232177734375,
0.7021484375,
0.232177734375,
0.7021484375,
0.2423095703125,
0.62451171875,
0.2423095703125
],
[
0.376220703125,
0.2410888671875,
0.44921875,
0.2410888671875,
0.44921875,
0.2529296875,
0.376220703125,
0.2529296875
],
[
0.5400390625,
0.2410888671875,
0.62060546875,
0.2410888671875,
0.62060546875,
0.2529296875,
0.5400390625,
0.2529296875
],
[
0.62646484375,
0.2410888671875,
0.69873046875,
0.2410888671875,
0.69873046875,
0.2529296875,
0.62646484375,
0.2529296875
],
[
0.70947265625,
0.2410888671875,
0.78076171875,
0.2410888671875,
0.78076171875,
0.2529296875,
0.70947265625,
0.2529296875
],
[
0.8017578125,
0.2410888671875,
0.880859375,
0.2410888671875,
0.880859375,
0.2529296875,
0.8017578125,
0.2529296875
],
[
0.21240234375,
0.2423095703125,
0.285400390625,
0.2423095703125,
0.285400390625,
0.2529296875,
0.21240234375,
0.2529296875
],
[
0.1259765625,
0.2423095703125,
0.185546875,
0.2423095703125,
0.185546875,
0.2529296875,
0.1259765625,
0.2529296875
],
[
0.29296875,
0.2423095703125,
0.365234375,
0.2423095703125,
0.365234375,
0.2529296875,
0.29296875,
0.2529296875
],
[
0.459228515625,
0.2423095703125,
0.5322265625,
0.2423095703125,
0.5322265625,
0.2529296875,
0.459228515625,
0.2529296875
],
[
0.1259765625,
0.2529296875,
0.177978515625,
0.2529296875,
0.177978515625,
0.26318359375,
0.1259765625,
0.26318359375
],
[
0.210693359375,
0.2529296875,
0.285400390625,
0.2529296875,
0.285400390625,
0.26318359375,
0.210693359375,
0.26318359375
],
[
0.28955078125,
0.2529296875,
0.36865234375,
0.2529296875,
0.36865234375,
0.26318359375,
0.28955078125,
0.26318359375
],
[
0.37451171875,
0.2529296875,
0.450927734375,
0.2529296875,
0.450927734375,
0.26318359375,
0.37451171875,
0.26318359375
],
[
0.45751953125,
0.2529296875,
0.5322265625,
0.2529296875,
0.5322265625,
0.26318359375,
0.45751953125,
0.26318359375
],
[
0.5400390625,
0.2529296875,
0.62060546875,
0.2529296875,
0.62060546875,
0.26318359375,
0.5400390625,
0.26318359375
],
[
0.62451171875,
0.2529296875,
0.7001953125,
0.2529296875,
0.7001953125,
0.26318359375,
0.62451171875,
0.26318359375
],
[
0.70947265625,
0.2529296875,
0.78076171875,
0.2529296875,
0.78076171875,
0.26318359375,
0.70947265625,
0.26318359375
],
[
0.8017578125,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.26318359375,
0.8017578125,
0.26318359375
],
[
0.37451171875,
0.262451171875,
0.44921875,
0.26318359375,
0.44921875,
0.27490234375,
0.37451171875,
0.273681640625
],
[
0.5400390625,
0.262451171875,
0.61865234375,
0.262451171875,
0.61865234375,
0.273681640625,
0.5400390625,
0.273681640625
],
[
0.62451171875,
0.2587890625,
0.69873046875,
0.26123046875,
0.69873046875,
0.275634765625,
0.62451171875,
0.273681640625
],
[
0.8017578125,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.273681640625,
0.8017578125,
0.273681640625
],
[
0.2091064453125,
0.26318359375,
0.2822265625,
0.26318359375,
0.2822265625,
0.273681640625,
0.2091064453125,
0.273681640625
],
[
0.1259765625,
0.26318359375,
0.202392578125,
0.26318359375,
0.202392578125,
0.27490234375,
0.1259765625,
0.27490234375
],
[
0.29296875,
0.26318359375,
0.365234375,
0.26318359375,
0.365234375,
0.273681640625,
0.29296875,
0.273681640625
],
[
0.459228515625,
0.26318359375,
0.5322265625,
0.26318359375,
0.5322265625,
0.273681640625,
0.459228515625,
0.273681640625
],
[
0.70947265625,
0.26318359375,
0.78076171875,
0.26318359375,
0.78076171875,
0.27490234375,
0.70947265625,
0.27490234375
],
[
0.1259765625,
0.273681640625,
0.198974609375,
0.273681640625,
0.198974609375,
0.283935546875,
0.1259765625,
0.283935546875
],
[
0.2091064453125,
0.273681640625,
0.28369140625,
0.273681640625,
0.28369140625,
0.283935546875,
0.2091064453125,
0.283935546875
],
[
0.29296875,
0.273681640625,
0.365234375,
0.273681640625,
0.365234375,
0.283935546875,
0.29296875,
0.283935546875
],
[
0.376220703125,
0.2724609375,
0.44921875,
0.2724609375,
0.44921875,
0.283203125,
0.376220703125,
0.283203125
],
[
0.459228515625,
0.273681640625,
0.5322265625,
0.273681640625,
0.5322265625,
0.283935546875,
0.459228515625,
0.283935546875
],
[
0.5380859375,
0.273681640625,
0.61865234375,
0.273681640625,
0.61865234375,
0.283935546875,
0.5380859375,
0.283935546875
],
[
0.62646484375,
0.273681640625,
0.7001953125,
0.273681640625,
0.7001953125,
0.283935546875,
0.62646484375,
0.283935546875
],
[
0.70947265625,
0.273681640625,
0.78076171875,
0.273681640625,
0.78076171875,
0.283935546875,
0.70947265625,
0.283935546875
],
[
0.8017578125,
0.273681640625,
0.8818359375,
0.273681640625,
0.8818359375,
0.283935546875,
0.8017578125,
0.283935546875
],
[
0.1175537109375,
0.314208984375,
0.489501953125,
0.314208984375,
0.489501953125,
0.327880859375,
0.1175537109375,
0.327880859375
],
[
0.51708984375,
0.314208984375,
0.5322265625,
0.314208984375,
0.5322265625,
0.350830078125,
0.51708984375,
0.350830078125
],
[
0.5400390625,
0.3134765625,
0.568359375,
0.3134765625,
0.568359375,
0.325927734375,
0.5400390625,
0.325927734375
],
[
0.54248046875,
0.32470703125,
0.568359375,
0.32470703125,
0.568359375,
0.339111328125,
0.54248046875,
0.339111328125
],
[
0.603515625,
0.32177734375,
0.6943359375,
0.32177734375,
0.6943359375,
0.33203125,
0.603515625,
0.33203125
],
[
0.1175537109375,
0.330810546875,
0.48779296875,
0.330810546875,
0.48779296875,
0.344482421875,
0.1175537109375,
0.344482421875
],
[
0.5458984375,
0.33837890625,
0.568359375,
0.33837890625,
0.568359375,
0.35400390625,
0.5458984375,
0.35400390625
],
[
0.603515625,
0.333251953125,
0.7119140625,
0.333251953125,
0.7119140625,
0.34326171875,
0.603515625,
0.34326171875
],
[
0.1175537109375,
0.3466796875,
0.48779296875,
0.3466796875,
0.48779296875,
0.35986328125,
0.1175537109375,
0.35986328125
],
[
0.5205078125,
0.3486328125,
0.5263671875,
0.3486328125,
0.5263671875,
0.35693359375,
0.5205078125,
0.35693359375
],
[
0.515625,
0.355712890625,
0.5322265625,
0.355712890625,
0.5322265625,
0.396728515625,
0.515625,
0.396728515625
],
[
0.5458984375,
0.3515625,
0.5693359375,
0.3515625,
0.5693359375,
0.366455078125,
0.5458984375,
0.366455078125
],
[
0.1158447265625,
0.3623046875,
0.489501953125,
0.3623046875,
0.489501953125,
0.377197265625,
0.1158447265625,
0.377197265625
],
[
0.55517578125,
0.363525390625,
0.57080078125,
0.363525390625,
0.57080078125,
0.38134765625,
0.55517578125,
0.38134765625
],
[
0.1158447265625,
0.37890625,
0.489501953125,
0.37890625,
0.489501953125,
0.392578125,
0.1158447265625,
0.392578125
],
[
0.54150390625,
0.37890625,
0.568359375,
0.37890625,
0.568359375,
0.392578125,
0.54150390625,
0.392578125
],
[
0.5224609375,
0.393798828125,
0.53076171875,
0.399658203125,
0.52490234375,
0.405029296875,
0.515625,
0.3984375
],
[
0.54150390625,
0.392578125,
0.568359375,
0.392578125,
0.568359375,
0.40625,
0.54150390625,
0.40625
],
[
0.5400390625,
0.40625,
0.568359375,
0.40625,
0.568359375,
0.418701171875,
0.5400390625,
0.418701171875
],
[
0.1175537109375,
0.394287109375,
0.486083984375,
0.394287109375,
0.486083984375,
0.407958984375,
0.1175537109375,
0.407958984375
],
[
0.1158447265625,
0.410400390625,
0.484375,
0.410400390625,
0.484375,
0.42333984375,
0.1158447265625,
0.42333984375
],
[
0.5205078125,
0.42333984375,
0.52880859375,
0.42333984375,
0.52880859375,
0.427490234375,
0.5205078125,
0.427490234375
],
[
0.5380859375,
0.421630859375,
0.56689453125,
0.421630859375,
0.56689453125,
0.429931640625,
0.5380859375,
0.429931640625
],
[
0.1158447265625,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.1158447265625,
0.457275390625
],
[
0.1158447265625,
0.42578125,
0.486083984375,
0.427001953125,
0.486083984375,
0.441162109375,
0.1158447265625,
0.440673828125
],
[
0.509765625,
0.44482421875,
0.880859375,
0.44482421875,
0.880859375,
0.457763671875,
0.509765625,
0.457763671875
],
[
0.1158447265625,
0.458984375,
0.48779296875,
0.457763671875,
0.48779296875,
0.47265625,
0.1158447265625,
0.473876953125
],
[
0.509765625,
0.458984375,
0.8818359375,
0.458984375,
0.8818359375,
0.47265625,
0.509765625,
0.47265625
],
[
0.1175537109375,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.1175537109375,
0.4892578125
],
[
0.51318359375,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.1158447265625,
0.4912109375,
0.489501953125,
0.4912109375,
0.489501953125,
0.5048828125,
0.1158447265625,
0.5048828125
],
[
0.509765625,
0.48681640625,
0.88330078125,
0.48681640625,
0.88330078125,
0.50048828125,
0.509765625,
0.50048828125
],
[
0.51123046875,
0.501953125,
0.81005859375,
0.501953125,
0.81005859375,
0.515625,
0.51123046875,
0.515625
],
[
0.1351318359375,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.52685546875,
0.1351318359375,
0.52685546875
],
[
0.51123046875,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.53759765625,
0.51123046875,
0.53759765625
],
[
0.1175537109375,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.54345703125,
0.1175537109375,
0.54345703125
],
[
0.509765625,
0.5390625,
0.88330078125,
0.5380859375,
0.88330078125,
0.552734375,
0.509765625,
0.55419921875
],
[
0.1175537109375,
0.5458984375,
0.48779296875,
0.5458984375,
0.48779296875,
0.55859375,
0.1175537109375,
0.55859375
],
[
0.1175537109375,
0.56103515625,
0.48779296875,
0.56103515625,
0.48779296875,
0.576171875,
0.1175537109375,
0.576171875
],
[
0.52880859375,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.52880859375,
0.5693359375
],
[
0.1158447265625,
0.57666015625,
0.48779296875,
0.576171875,
0.48779296875,
0.59033203125,
0.1158447265625,
0.59130859375
],
[
0.51123046875,
0.572265625,
0.88330078125,
0.572265625,
0.88330078125,
0.5859375,
0.51123046875,
0.5859375
],
[
0.509765625,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.60107421875,
0.509765625,
0.60107421875
],
[
0.1175537109375,
0.59326171875,
0.48779296875,
0.59326171875,
0.48779296875,
0.60693359375,
0.1175537109375,
0.60693359375
],
[
0.1175537109375,
0.60986328125,
0.468505859375,
0.60986328125,
0.468505859375,
0.62353515625,
0.1175537109375,
0.62353515625
],
[
0.51123046875,
0.60498046875,
0.88330078125,
0.60498046875,
0.88330078125,
0.6181640625,
0.51123046875,
0.6181640625
],
[
0.509765625,
0.62060546875,
0.88330078125,
0.62060546875,
0.88330078125,
0.63427734375,
0.509765625,
0.63427734375
],
[
0.1351318359375,
0.6298828125,
0.48779296875,
0.6298828125,
0.48779296875,
0.6435546875,
0.1351318359375,
0.6435546875
],
[
0.509765625,
0.6357421875,
0.884765625,
0.63720703125,
0.884765625,
0.65087890625,
0.509765625,
0.6494140625
],
[
0.1175537109375,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.509765625,
0.65283203125,
0.8818359375,
0.65283203125,
0.8818359375,
0.66650390625,
0.509765625,
0.66650390625
],
[
0.1175537109375,
0.66455078125,
0.489501953125,
0.66455078125,
0.489501953125,
0.677734375,
0.1175537109375,
0.677734375
],
[
0.51318359375,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.1175537109375,
0.6796875,
0.48779296875,
0.6796875,
0.48779296875,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.51123046875,
0.68505859375,
0.8818359375,
0.68505859375,
0.8818359375,
0.69873046875,
0.51123046875,
0.69873046875
],
[
0.11419677734375,
0.69482421875,
0.486083984375,
0.69482421875,
0.486083984375,
0.7080078125,
0.11419677734375,
0.7080078125
],
[
0.509765625,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.509765625,
0.71337890625
],
[
0.1175537109375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.51123046875,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.73095703125,
0.51123046875,
0.73095703125
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.51318359375,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.1158447265625,
0.74462890625,
0.48779296875,
0.74462890625,
0.48779296875,
0.75830078125,
0.1158447265625,
0.75830078125
],
[
0.51123046875,
0.7490234375,
0.880859375,
0.7490234375,
0.880859375,
0.7626953125,
0.51123046875,
0.7626953125
],
[
0.1158447265625,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1158447265625,
0.7724609375
],
[
0.51123046875,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51123046875,
0.77978515625
],
[
0.1175537109375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.509765625,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.79443359375,
0.509765625,
0.79443359375
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.798828125,
0.8330078125,
0.798828125,
0.8330078125,
0.8125,
0.51123046875,
0.8125
],
[
0.1158447265625,
0.8056640625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82177734375,
0.1158447265625,
0.82080078125
],
[
0.1158447265625,
0.82470703125,
0.490234375,
0.82470703125,
0.490234375,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.5146484375,
0.82763671875,
0.66015625,
0.82763671875,
0.66015625,
0.8427734375,
0.5146484375,
0.8427734375
],
[
0.1158447265625,
0.84130859375,
0.489501953125,
0.84130859375,
0.489501953125,
0.85498046875,
0.1158447265625,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5087890625,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.87060546875,
0.5087890625,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1158447265625,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90283203125
],
[
0.5087890625,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.5087890625,
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,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"Table 1: MS MARCO passage ranking results. Row 2 reports results averaged across all typos queries; rows",
"3-7 results for each typos type (for each type, typos are injected in all dev queries). Percentage reductions are",
"computed w.r.t. the original queries; bold represents best performance across training methods for each of DR",
"and BERT re-ranker. Statistical significant gains (two-tailed paired t-test with Bonferroni correction, p < 0.01)",
"obtained by models with typos-aware training over the models with standard training (std.) are indicated by t",
"BM25",
"DR (std.)",
"DR (typos-aware)",
"Re-ranker (std.)",
"Re-ranker",
"(tvpos-aware)",
"Typo type",
"MRR@10",
"MRR@10",
"Recall",
"MRR@10",
"Recall",
"MRR@10",
"MRR@10",
".187",
"Recall",
".300",
".940",
".379",
".857",
"original",
".296",
".940",
".374",
"w. typos (avg)",
".120(-35.8%)",
".696(18.6%)",
".141(52.3%)",
".712(-24.3%)",
".219(-27.0%)",
".857t(-8.8%)",
".250(-34.0%)",
".289t(-22.7%)",
"RandInsert",
".693(-18.9%)",
".225(25.0%)",
"257(-32.2%)",
".297t(-20.6%)",
".140(-52.7%)",
".125(-33.1%)",
"-711(-24.4%)",
".862t(-8.3%)",
".154(-47.9%)",
".217(-27.6%)",
".853t(9.3%)",
".257(32.2%)",
".288t(-23.0%)",
".118(36.9%)",
"RandDelete",
".693(18.9%)",
".730(22.3%)",
"RandSub",
".120(-35.8%)",
".702(-17.9%)",
".137(-53.7%)",
".714(-24.0%)",
".220t(-26.7%)",
".858t(-8.7%)",
".250(-34.0%)",
".291t(-22.2%)",
".137(53.7%)",
".217(-27.6%)",
".859(8.6%)",
"284t(-24.1%)",
".122(34.7%)",
"SwapNeighbor",
".702(17.9%)",
".705(25.0%)",
".240(36.7%)",
"SwapAdjacent",
"117(-37.4%)",
".691(-19.1%)",
".137(-53.7%)",
".702(-25.3%)",
".214t(-28.7%)",
".854t(-9.1%)",
".246(35.1%)",
".286(-23.5%)",
"ority of BERT-based methods. For RQl, we com-",
"(b)od/a",
"1000",
"750",
"Standardtraining",
"pare the results obtained on the original queries",
"500",
"Typos-awaretraining",
"with those on queries with typos, when models",
"-",
"(b)eu",
"250",
"are trained using the standard procedure. We ob-",
"。",
" serve statistically significant losses in effectiveness",
"250",
"o",
"500",
"750",
"for both DR (on average MRR@10 drops 52.3%",
"and Recall 24.3%) and BERT re-ranker (MRR@10",
"D",
"1000",
"of DR results, thus losses in Recall for DR are",
"drops 34%). BERT re-ranker is performed on top",
"Figure I: Loss in terms of the rank position of the first",
"propagated to the BERT re-ranker. However, the",
"relevant passage retrieved by DR when ranking for ty-",
"effectiveness of DR on typos queries drops to about",
"pos queries, compared to original queries. Each point",
"that of BM25, while BERT re-ranker stays superior.",
" on the x-axis refers to a query; x-axis ordered by de-",
"creasing loss when standard training is used.",
"In terms of the impact of different types of typos",
"only queries without typos (original) are considered",
"(RQ2), the results show that different typos have",
"(minor differences are not statistically significant).",
"similar impact: they all hurt effectiveness heavily.",
"DR appears however more tolerant to RandDelete",
"Figure 1 presents the rank loss obtained by DR",
"typos (with a ~ 5% smaller loss in MRR@10 than",
"when answering typos queries in place of the ori-",
"ginal queries (plot averaged across all types of ty-",
"for other types of typos), while BERT re-ranker",
"losses are generally uniform across typos types.",
"pos; individual typos types show similar trends). A",
" negative loss of n means when using typos queries,",
"To answer RQ3, we compare the results of mod-",
"the first relevant document is retrieved n rank po-",
"els produced with typos-aware training vs. with the",
" sitions after that obtained when using the original",
"standard training. Despite the typos-aware training,",
"queries. The figure shows that typos-aware train",
"both DR and BERT re-ranker display significant",
"ing consistently provides smaller losses than the",
"lossesineffectivenesswhendealing withtypos",
" standard training. We also note there are few cases",
"queries, compared to the original queries (Table 1).",
"(~ 300 queries) in which typos queries provide",
"However, compared to the models with standard",
"gains compared to the original query. Queries with",
"training, both methods are much more tolerant to",
"large losses often have typos for keywords that are",
"all types of typos when typos-aware training is",
"essential to determine the intent of the query. Typos",
"employed. In fact, losses in MRR@10 halve for",
"queries that exhibit gains generally display typos",
"DR (from 52.3% to 24.3%), and reduce by one",
"on non-essential keywords, e.g., stopwords.",
"third for BERT re-ranker (from 34% to 22.7%);",
"all differences are statistically significant. Typos-",
"5A CaseStudy",
"aware training seems to impact queries with typos",
"produced by RandomInsert more than those with",
"The results presented in the previous sections are",
"other types of typos. We also note that effective-",
"conducted with synthetically generated typo quer",
"ness obtained bymodels with typos-aware training",
"ies. Accurate analysis of the MS MARCO data-",
"is not different from that with standard training if",
"set revels the presence of a very limit number of"
] |
{
"arxiv_id": "2108.12139",
"language": "en",
"page_idx": 3
}
|
||
2108.12139-en-4
|
2108.12139-en-4.jpg
|
queries containing typos - these typos are legitimate errors made by the user issuing the query. For instance, the MS MARCO dev set contains the mistyped query-“sydeny climate”2(qid: 506025). Without typos-aware training, the considered DR cannot retrieve the relevant passage in the top 1,000 results. However, with typos-aware training, the DR is able to rank the relevant passage at rank 127 for this typo query. This suggests that DRs trained with our typos-aware training with synthetic typo generation may be able to generalize to real-world typo queries, aside from those synthetic (though realistic) typo queries we considered in our extensive empirical evaluation. In future work, we want to further test our proposed typos-aware training with more real-world typo queries by acquiring a real query log with typos and perform relevance annotations on the MS MARCO passage collection.
# 6 Conclusion
In this paper we studied the impact of typos in queries on popular BERT-based passage retrieval methods. We reported significant drops in effectiveness across different types of typos for both DR and BERT re-ranker: these methods are not tolerant to typos in queries when solely relying on the BERT encoder. We then proposed a typos-aware training strategy for DR and BERT re-ranker, which controls the exposure of the models to queries with typos during training. With our typos-aware training, both DR and BERT re-ranker showed to be much more tolerant to typos in queries. We believe our typos-aware training can be used (more extensively than in this paper) as a standard data augmentation step in the DR and BERT re-ranker's training loop since the computations for typos generation are very light and can provide extra gains on typos queries, without hurting effectiveness on queries without typos. Code, typos queries and results files at https://github.com/ielab/ $t$ypos-aware-BERT.
## Acknowledgements
Dr Guido Zuccon is the recipient of an Australian Research Council DECRA Research Fellowship (DE180101579). This research is partially funded by the Grain Research and Development Corporation project AgAsk (UOQ2003-009RTX).
$^{2}$The correct spelling is “sydney climate”.
# References
Tom B Brown, Benjamin Mann, Nick Ryder, Melanie
Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. $arXivpreprintarXiv:2005.14165.$
Ting Chen, Simon Kornblith, Mohammad Norouzi,
and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In $In-$ ternational conference on machine learning, pages 1597-1607. PMLR.
Zhuyun Dai and Jamie Callan. 2019. Deeper text un-
derstanding for ir with contextual neural language modeling. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 985-988.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv: I810.04805.
Yingqi Qu Yuchen Ding, Jing Liu, Kai Liu, Ruiy-
ang Ren, Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2020. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. arXivpreprint $\hat{ar}Xiv:2010.08\hat{I}9l.$
Thibault Formal, Benjamin Piwowarski, and Stéphane
Clinchant. 2021. A white box analysis of colbert. In The 43rd European Conference On Information Retrieval $(ECIR).$
Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021a. Re-
think training of bert rerankers in multi-stage retrieval pipeline. In The 43rd European Conference On Information Retrieval (ECIR).
Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Ben-
$\dot{\text{jamin Van Durme, and Jamie Callan. 2021b. Com-}}$ plementing lexical retrieval with semantic residual embedding. In The 43rd European Conference On Information Retrieval (ECIR).
Jean-Bastien Grill, Florian Strub, Florent Altché
Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. 2020. Bootstrap your own latent: A new approach to self-supervised learning. $arXi\nu preprint$ $arXiv:2006.07733.$
Matthias Hagen, Martin Potthast, Marcel Gohsen, Anja
Rathgeber, and Benno Stein. 2017. A large-scale query spelling correction corpus. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1261-1264.
|
[
[
0.1175537109375,
0.08612060546875,
0.490234375,
0.08612060546875,
0.490234375,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51318359375,
0.0872802734375,
0.60791015625,
0.0872802734375,
0.60791015625,
0.09857177734375,
0.51318359375,
0.09857177734375
],
[
0.1175537109375,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.1175537109375,
0.11639404296875
],
[
0.509765625,
0.10986328125,
0.8818359375,
0.10986328125,
0.8818359375,
0.12353515625,
0.509765625,
0.12353515625
],
[
0.1175537109375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.52880859375,
0.12353515625,
0.8818359375,
0.12353515625,
0.8818359375,
0.13720703125,
0.52880859375,
0.13720703125
],
[
0.1158447265625,
0.1341552734375,
0.48779296875,
0.133056640625,
0.48779296875,
0.1478271484375,
0.1158447265625,
0.1484375
],
[
0.53076171875,
0.1383056640625,
0.880859375,
0.1383056640625,
0.880859375,
0.1484375,
0.53076171875,
0.1484375
],
[
0.1175537109375,
0.15087890625,
0.486083984375,
0.15087890625,
0.486083984375,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.53076171875,
0.149658203125,
0.88330078125,
0.149658203125,
0.88330078125,
0.163330078125,
0.53076171875,
0.163330078125
],
[
0.11419677734375,
0.166259765625,
0.486083984375,
0.1650390625,
0.486083984375,
0.179931640625,
0.11419677734375,
0.18115234375
],
[
0.52880859375,
0.163330078125,
0.82275390625,
0.163330078125,
0.82275390625,
0.177001953125,
0.52880859375,
0.177001953125
],
[
0.1158447265625,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.1158447265625,
0.196533203125
],
[
0.509765625,
0.190673828125,
0.880859375,
0.190673828125,
0.880859375,
0.2037353515625,
0.509765625,
0.2037353515625
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.52880859375,
0.204833984375,
0.88330078125,
0.204833984375,
0.88330078125,
0.218505859375,
0.52880859375,
0.218505859375
],
[
0.1175537109375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.5283203125,
0.218505859375,
0.880859375,
0.218505859375,
0.880859375,
0.232177734375,
0.5283203125,
0.232177734375
],
[
0.1158447265625,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1158447265625,
0.2452392578125
],
[
0.5283203125,
0.2286376953125,
0.8818359375,
0.23095703125,
0.8818359375,
0.2452392578125,
0.5283203125,
0.243408203125
],
[
0.1175537109375,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.52880859375,
0.243408203125,
0.6650390625,
0.243408203125,
0.6650390625,
0.257080078125,
0.52880859375,
0.257080078125
],
[
0.1158447265625,
0.26318359375,
0.48779296875,
0.262451171875,
0.48779296875,
0.276611328125,
0.1158447265625,
0.27783203125
],
[
0.509765625,
0.271484375,
0.880859375,
0.271484375,
0.880859375,
0.28515625,
0.509765625,
0.28515625
],
[
0.1158447265625,
0.27978515625,
0.489501953125,
0.27978515625,
0.489501953125,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.5283203125,
0.28515625,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.5283203125,
0.298583984375
],
[
0.1158447265625,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1158447265625,
0.308837890625
],
[
0.5283203125,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.5283203125,
0.312255859375
],
[
0.1158447265625,
0.312255859375,
0.486083984375,
0.312255859375,
0.486083984375,
0.325927734375,
0.1158447265625,
0.325927734375
],
[
0.5283203125,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.325927734375,
0.5283203125,
0.325927734375
],
[
0.1175537109375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.1175537109375,
0.341552734375
],
[
0.5263671875,
0.32470703125,
0.84619140625,
0.32373046875,
0.84619140625,
0.33837890625,
0.5263671875,
0.339111328125
],
[
0.1175537109375,
0.345703125,
0.486083984375,
0.345703125,
0.486083984375,
0.355712890625,
0.1175537109375,
0.355712890625
],
[
0.1158447265625,
0.359375,
0.480224609375,
0.359375,
0.480224609375,
0.37353515625,
0.1158447265625,
0.37353515625
],
[
0.51318359375,
0.35400390625,
0.8818359375,
0.35400390625,
0.8818359375,
0.36767578125,
0.51318359375,
0.36767578125
],
[
0.52880859375,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.52880859375,
0.38134765625
],
[
0.5283203125,
0.37890625,
0.88330078125,
0.38134765625,
0.88330078125,
0.394287109375,
0.5283203125,
0.392578125
],
[
0.118408203125,
0.392578125,
0.133544921875,
0.392578125,
0.133544921875,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.14697265625,
0.392578125,
0.2451171875,
0.392578125,
0.2451171875,
0.40380859375,
0.14697265625,
0.40380859375
],
[
0.5263671875,
0.393798828125,
0.79248046875,
0.391357421875,
0.79248046875,
0.405029296875,
0.5263671875,
0.40673828125
],
[
0.11419677734375,
0.418701171875,
0.486083984375,
0.418701171875,
0.486083984375,
0.432861328125,
0.11419677734375,
0.432861328125
],
[
0.133544921875,
0.3955078125,
0.1427001953125,
0.3955078125,
0.1427001953125,
0.3984375,
0.133544921875,
0.3984375
],
[
0.51123046875,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435302734375,
0.51123046875,
0.435302734375
],
[
0.1175537109375,
0.435302734375,
0.486083984375,
0.435302734375,
0.486083984375,
0.448974609375,
0.1175537109375,
0.448974609375
],
[
0.52880859375,
0.435302734375,
0.8818359375,
0.435302734375,
0.8818359375,
0.448974609375,
0.52880859375,
0.448974609375
],
[
0.11419677734375,
0.449462890625,
0.48779296875,
0.45068359375,
0.48779296875,
0.465576171875,
0.11419677734375,
0.46435546875
],
[
0.52880859375,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.52880859375,
0.46142578125
],
[
0.52880859375,
0.46142578125,
0.8818359375,
0.46142578125,
0.8818359375,
0.474365234375,
0.52880859375,
0.474365234375
],
[
0.1158447265625,
0.46728515625,
0.486083984375,
0.46728515625,
0.486083984375,
0.48095703125,
0.1158447265625,
0.48095703125
],
[
0.52880859375,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.488037109375,
0.52880859375,
0.488037109375
],
[
0.1175537109375,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.495361328125,
0.1175537109375,
0.495361328125
],
[
0.52880859375,
0.488037109375,
0.6591796875,
0.48681640625,
0.6591796875,
0.498779296875,
0.52880859375,
0.49951171875
],
[
0.1158447265625,
0.49951171875,
0.48779296875,
0.49755859375,
0.48779296875,
0.5126953125,
0.1158447265625,
0.51416015625
],
[
0.1175537109375,
0.5166015625,
0.48779296875,
0.5166015625,
0.48779296875,
0.52978515625,
0.1175537109375,
0.52978515625
],
[
0.51318359375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.1175537109375,
0.5322265625,
0.490234375,
0.5322265625,
0.490234375,
0.5458984375,
0.1175537109375,
0.5458984375
],
[
0.52880859375,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.52880859375,
0.5419921875
],
[
0.5283203125,
0.54052734375,
0.8818359375,
0.54150390625,
0.8818359375,
0.5556640625,
0.5283203125,
0.5546875
],
[
0.1175537109375,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5625,
0.1175537109375,
0.5625
],
[
0.52880859375,
0.5546875,
0.6533203125,
0.5546875,
0.6533203125,
0.568359375,
0.52880859375,
0.568359375
],
[
0.1175537109375,
0.56396484375,
0.489501953125,
0.56396484375,
0.489501953125,
0.57763671875,
0.1175537109375,
0.57763671875
],
[
0.1158447265625,
0.580078125,
0.489501953125,
0.580078125,
0.489501953125,
0.59326171875,
0.1158447265625,
0.59326171875
],
[
0.5087890625,
0.5830078125,
0.8818359375,
0.58203125,
0.8818359375,
0.5966796875,
0.5087890625,
0.59716796875
],
[
0.1158447265625,
0.595703125,
0.490234375,
0.59423828125,
0.490234375,
0.609375,
0.1158447265625,
0.60986328125
],
[
0.5263671875,
0.595703125,
0.8818359375,
0.595703125,
0.8818359375,
0.609375,
0.5263671875,
0.609375
],
[
0.1158447265625,
0.61083984375,
0.486083984375,
0.61083984375,
0.486083984375,
0.62451171875,
0.1158447265625,
0.62451171875
],
[
0.52880859375,
0.60986328125,
0.8818359375,
0.60986328125,
0.8818359375,
0.62353515625,
0.52880859375,
0.62353515625
],
[
0.1175537109375,
0.62744140625,
0.48779296875,
0.62744140625,
0.48779296875,
0.64111328125,
0.1175537109375,
0.64111328125
],
[
0.53076171875,
0.62451171875,
0.75830078125,
0.62451171875,
0.75830078125,
0.63525390625,
0.53076171875,
0.63525390625
],
[
0.11419677734375,
0.6435546875,
0.48779296875,
0.642578125,
0.48779296875,
0.65673828125,
0.11419677734375,
0.65771484375
],
[
0.51123046875,
0.65185546875,
0.88330078125,
0.65185546875,
0.88330078125,
0.6650390625,
0.51123046875,
0.6650390625
],
[
0.1175537109375,
0.66015625,
0.48779296875,
0.66015625,
0.48779296875,
0.67333984375,
0.1175537109375,
0.67333984375
],
[
0.5283203125,
0.6650390625,
0.88330078125,
0.6650390625,
0.88330078125,
0.6787109375,
0.5283203125,
0.6787109375
],
[
0.1175537109375,
0.6767578125,
0.48779296875,
0.6767578125,
0.48779296875,
0.6904296875,
0.1175537109375,
0.6904296875
],
[
0.5283203125,
0.6787109375,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.5283203125,
0.6923828125
],
[
0.1158447265625,
0.693359375,
0.48779296875,
0.693359375,
0.48779296875,
0.70703125,
0.1158447265625,
0.70703125
],
[
0.52880859375,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.52880859375,
0.705078125
],
[
0.1158447265625,
0.708984375,
0.489501953125,
0.708984375,
0.489501953125,
0.72265625,
0.1158447265625,
0.72265625
],
[
0.53076171875,
0.7060546875,
0.73486328125,
0.7060546875,
0.73486328125,
0.71630859375,
0.53076171875,
0.71630859375
],
[
0.1158447265625,
0.724609375,
0.486083984375,
0.724609375,
0.486083984375,
0.73828125,
0.1158447265625,
0.73828125
],
[
0.51123046875,
0.732421875,
0.88330078125,
0.732421875,
0.88330078125,
0.7451171875,
0.51123046875,
0.7451171875
],
[
0.1158447265625,
0.7412109375,
0.301513671875,
0.7412109375,
0.301513671875,
0.7548828125,
0.1158447265625,
0.7548828125
],
[
0.5283203125,
0.74462890625,
0.88330078125,
0.74462890625,
0.88330078125,
0.75830078125,
0.5283203125,
0.75830078125
],
[
0.52880859375,
0.7587890625,
0.88330078125,
0.7587890625,
0.88330078125,
0.7724609375,
0.52880859375,
0.7724609375
],
[
0.118408203125,
0.771484375,
0.285400390625,
0.771484375,
0.285400390625,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.52880859375,
0.771484375,
0.884765625,
0.771484375,
0.884765625,
0.78515625,
0.52880859375,
0.78515625
],
[
0.52880859375,
0.78515625,
0.79833984375,
0.78515625,
0.79833984375,
0.798828125,
0.52880859375,
0.798828125
],
[
0.79248046875,
0.78759765625,
0.880859375,
0.78759765625,
0.880859375,
0.79736328125,
0.79248046875,
0.79736328125
],
[
0.1175537109375,
0.80029296875,
0.48779296875,
0.80029296875,
0.48779296875,
0.81396484375,
0.1175537109375,
0.81396484375
],
[
0.52880859375,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.52880859375,
0.8125
],
[
0.1175537109375,
0.81640625,
0.486083984375,
0.81640625,
0.486083984375,
0.830078125,
0.1175537109375,
0.830078125
],
[
0.52880859375,
0.81103515625,
0.66015625,
0.81103515625,
0.66015625,
0.82470703125,
0.52880859375,
0.82470703125
],
[
0.1175537109375,
0.83203125,
0.486083984375,
0.83203125,
0.486083984375,
0.8466796875,
0.1175537109375,
0.8466796875
],
[
0.51318359375,
0.8388671875,
0.8818359375,
0.8388671875,
0.8818359375,
0.8525390625,
0.51318359375,
0.8525390625
],
[
0.1158447265625,
0.8486328125,
0.48779296875,
0.8486328125,
0.48779296875,
0.8623046875,
0.1158447265625,
0.8623046875
],
[
0.52880859375,
0.8525390625,
0.8818359375,
0.8525390625,
0.8818359375,
0.8662109375,
0.52880859375,
0.8662109375
],
[
0.1158447265625,
0.8642578125,
0.4248046875,
0.8642578125,
0.4248046875,
0.87744140625,
0.1158447265625,
0.87744140625
],
[
0.52880859375,
0.8662109375,
0.88330078125,
0.8662109375,
0.88330078125,
0.8798828125,
0.52880859375,
0.8798828125
],
[
0.52880859375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.52880859375,
0.892578125
],
[
0.52880859375,
0.892578125,
0.884765625,
0.892578125,
0.884765625,
0.90625,
0.52880859375,
0.90625
],
[
0.1368408203125,
0.90283203125,
0.391357421875,
0.90380859375,
0.391357421875,
0.91845703125,
0.1368408203125,
0.91748046875
],
[
0.5283203125,
0.90625,
0.654296875,
0.90283203125,
0.654296875,
0.91650390625,
0.5283203125,
0.91943359375
]
] |
[
"queries containing typos -- these typos are legit-",
"References",
"imate errors made by the user issuing the query.",
"Tom B Brown, Benjamin Mann, Nick Ryder, Melanie",
"For instance, the MS MARCO dev set contains the",
"Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind",
"mistyped query-“sydenyclimate\"2(qid:506025).",
"Neelakantan,PranavShyam,Girish Sastry,Amanda",
"Without typos-aware training, the considered DR",
"Askell, et al. 2020. Language models are few-shot",
"cannot retrieve the relevant passage in the top 1,o00",
"learners. arXiv preprint arXiv:2005.14165.",
"results. However, with typos-aware training, the",
"Ting Chen, Simon Kornblith, Mohammad Norouzi.",
"DR is able to rank the relevant passage at rank 127",
"and Geoffrey Hinton.2020. A simple framework for",
"for this typo query. This suggests that DRs trained",
"contrastive learning of visual representations. In In-",
"with our typos-aware training with synthetic typo",
"ternational conference on machine learning,pages",
"generation may be able to generalize to real-world",
"1597-1607.PMLR.",
"typo queries, aside from those synthetic (though",
"Zhuyun Dai and Jamie Callan. 2019. Deeper text un",
"realistic) typo queries we considered in our extens-",
"derstanding for ir with contextual neural language",
"ive empirical evaluation. In future work, we want",
"modeling. In Proceedings of the 42nd International",
"to further test our proposed typos-aware training",
"ACM SIGIR Conference on Research and Develop-",
"with more real-world typo queries by acquiring a",
"ment in Information Retrieval, pages 985-988.",
"realquerylogwithtypos andperformrelevancean-",
"notations on the MS MARCO passage collection.",
"Jacob Devlin, Ming-Wei Chang, Kenton Lee, and",
"Kristina Toutanova. 2018. Bert: Pre-training of deep",
"bidirectional transformers for language understand-",
"6.",
"Conclusion",
"ing.arXiv preprint arXiv:1810.04805.",
"In this paper we studied the impact of typos in",
"",
"Yingqi Qu Yuchen Ding, Jing Liu, Kai Liu, Ruiy-",
"queries on popular BERT-based passage retrieval",
"ang Ren, Xin Zhao, Daxiang Dong, Hua Wu, and",
"methods. We reported significant drops in effective-",
"Haifeng Wang.2020.Rocketqa:An optimized",
"training approach to dense passage retrieval for",
"ness across different types of typos for both DR and",
"open-domain question answering.arXiv preprint",
"BERT re-ranker:these methods are not tolerantto",
"arXiv:2010.08191.",
"typos in queries when solely relying on the BERT",
"encoder. We then proposed a typos-aware training",
"ThibaultFormal,Benjamin Piwowarski,and Stephane",
"strategy for DR and BERT re-ranker, which con-",
"Clinchant. 2021. A white box analysis of colbert.",
"In The 43rd European Conference On Information",
"trols the exposure of the models to queries with",
"Retrieval (ECIR).",
"typos during training. With our typos-aware train-",
"ing, both DR and BERT re-ranker showed to be",
"Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021a. Re-",
"much more tolerant to typos in queries. We be-",
" think training of bert rerankers in multi-stage re-",
"lieve our typos-aware training can be used (more",
"trieval pipeline. In The 43rd European Conference",
"extensively than in this paper) as a standard data",
"OnInformationRetrieval (ECIR).",
"augmentation step in the DR and BERT re-ranker's",
"Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Ben-",
"training loop since the computations for typos gen-",
"jamin Van Durme, and Jamie Callan. 2021b. Com-",
"eration are very light and can provide extra gains",
"plementing lexical retrieval with semantic residual",
"on typos queries, without hurting effectiveness on",
"embedding. In The 43rd European Conference On",
"queries without typos. Code, typos queries and res-",
"InformationRetrieval (ECIR).",
"ults files at https: //github.com/ielab/",
"Jean-Bastien Grill,Florian Strub, Florent Altche,",
"typos-aware-BERT.",
"Corentin Tallec, Pierre H Richemond, Elena Buchat-",
"skaya, Carl Doersch, Bernardo Avila Pires, Zhao-",
"Acknowledgements",
"han Daniel Guo, Mohammad Gheshlaghi Azar, et al.",
"2020.Bootstrap your own latent:",
"Anewap-",
"Dr Guido Zuccon is the recipient of an Australian",
"proach to self-supervised learning. arXiv preprint",
"Research Council DECRA Research Fellowship",
"arXiv:2006.07733.",
"(DE180101579). This research is partially funded",
"Matthias Hagen, Martin Potthast, Marcel Gohsen, Anja",
"by the Grain Research and Development Corpora-",
"Rathgeber, and Benno Stein. 2017. A large-scale",
"tion project AgAsk (UOQ2003-009RTX).",
"query spelling correction corpus. In Proceedings of",
"the 4Oth International ACM SIGIR Conference on",
"Research and Development in Information Retrieval,",
"2The correct spelling is “\"sydney climate\".",
"pages1261-1264."
] |
{
"arxiv_id": "2108.12139",
"language": "en",
"page_idx": 4
}
|
||
2311.00523-en-0
|
2311.00523-en-0.jpg
|
## Learning impartial policies for sequential counterfactual explanations using Deep $\operatorname{Reinforcement}$ Learning
Emmanouil Panagiotou$^{1,2[0000-0001-9134-9387]}$ and Eirini
$\begin{array}{l}{{\mathrm{Moutsi}^{2}[0000-0001-5729-1003]}}\\{{\mathrm{Ntoutsi}^{2}[0000-0001-5729-1003]}}\end{array}$
$^{1}$ Freie Universität Berlin, Department of Mathematics &Computer Science, Berlin,
Germany
emmanouil.panagiotou@fu-berlin.de
$^2\text{ Universit\"{a} t der Bundeswehr M\"{u} nchen, Faculty for Informatik, Munich, Germany}$
Abstract. In the field of explainable Artificial Intelligence (XAI), sequential counterfactual (SCF) examples are often used to alter the decision of a trained classifier by implementing a sequence of modifications to the input instance. Although certain test-time algorithms aim to optimize for each new instance individually, recently Reinforcement Learning (RL) methods have been proposed that seek to learn policies for discovering SCFs, thereby enhancing scalability. As is typical in RL, the formulation of the RL problem, including the specification of state space actions, and rewards, can often be ambiguous. In this work, we identify shortcomings in existing methods that can result in policies with undesired properties, such as a bias towards specific actions. We propose to use the output probabilities of the classifier to create a more informative reward, to mitigate this effect.
Keywords: Sequential counterfactuals$\cdot$Reinforcement Learning$\cdot$Model-
agnostic
## 1 Introduction
Predictive Machine Learning (ML) models have been used in various fields to make predictions by generalizing knowledge learned from data. Yet, in recent years, generating explanations for model decisions has become increasingly important, shifting interest to the feld of explainable AI (XAI)[7]. A popular category of XAI methods are the so-called“counterfactual explanations”that offer a"what-if" analysis of model decisions by identifying the minimal set of changes or interventions needed to alter the outcome of the model for a given instance, allowing users to understand the factors that influenced the model's decision. Common desiderata for counterfactual explanations, as highlighted by [4], include proximity (distance) to the original instance, sparsity in terms of changes made, and the plausibility of the counterfactual instances w.r.t. the particular problem domain.
|
[
[
0.26708984375,
0.146484375,
0.73779296875,
0.146484375,
0.73779296875,
0.1610107421875,
0.26708984375,
0.1610107421875
],
[
0.27685546875,
0.169189453125,
0.72802734375,
0.169189453125,
0.72802734375,
0.184326171875,
0.27685546875,
0.184326171875
],
[
0.36279296875,
0.189453125,
0.64208984375,
0.1912841796875,
0.64208984375,
0.208984375,
0.36279296875,
0.20703125
],
[
0.29736328125,
0.231689453125,
0.708984375,
0.231689453125,
0.708984375,
0.25244140625,
0.29736328125,
0.25244140625
],
[
0.396240234375,
0.25,
0.6044921875,
0.246826171875,
0.6044921875,
0.264404296875,
0.396240234375,
0.267578125
],
[
0.0228729248046875,
0.26513671875,
0.060455322265625,
0.26513671875,
0.060455322265625,
0.3984375,
0.0228729248046875,
0.3984375
],
[
0.2181396484375,
0.27587890625,
0.78369140625,
0.27783203125,
0.78369140625,
0.29541015625,
0.2181396484375,
0.29345703125
],
[
0.47216796875,
0.29541015625,
0.53173828125,
0.29541015625,
0.53173828125,
0.306884765625,
0.47216796875,
0.306884765625
],
[
0.374267578125,
0.30810546875,
0.6298828125,
0.30810546875,
0.6298828125,
0.3212890625,
0.374267578125,
0.3212890625
],
[
0.2254638671875,
0.318115234375,
0.77685546875,
0.32080078125,
0.77685546875,
0.3359375,
0.2254638671875,
0.333984375
],
[
0.0318603515625,
0.375,
0.0400390625,
0.37060546875,
0.04412841796875,
0.375,
0.03594970703125,
0.37890625
],
[
0.2646484375,
0.375,
0.73876953125,
0.375732421875,
0.73876953125,
0.39013671875,
0.2646484375,
0.389404296875
],
[
0.0228729248046875,
0.395751953125,
0.062103271484375,
0.395751953125,
0.062103271484375,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.265625,
0.39013671875,
0.740234375,
0.39013671875,
0.740234375,
0.404052734375,
0.265625,
0.404052734375
],
[
0.2646484375,
0.40478515625,
0.740234375,
0.40478515625,
0.740234375,
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
],
[
0.265625,
0.44580078125,
0.73779296875,
0.44580078125,
0.73779296875,
0.458984375,
0.265625,
0.458984375
],
[
0.26318359375,
0.458251953125,
0.73876953125,
0.458984375,
0.73876953125,
0.473388671875,
0.26318359375,
0.472900390625
],
[
0.2646484375,
0.473388671875,
0.73876953125,
0.473388671875,
0.73876953125,
0.4873046875,
0.2646484375,
0.4873046875
],
[
0.0228729248046875,
0.4873046875,
0.060455322265625,
0.4873046875,
0.060455322265625,
0.70556640625,
0.0228729248046875,
0.70556640625
],
[
0.2646484375,
0.4873046875,
0.73779296875,
0.4873046875,
0.73779296875,
0.50048828125,
0.2646484375,
0.50048828125
],
[
0.2646484375,
0.501953125,
0.73876953125,
0.501953125,
0.73876953125,
0.51513671875,
0.2646484375,
0.51513671875
],
[
0.26318359375,
0.51513671875,
0.740234375,
0.51513671875,
0.740234375,
0.52880859375,
0.26318359375,
0.52880859375
],
[
0.2646484375,
0.52880859375,
0.740234375,
0.52880859375,
0.740234375,
0.54248046875,
0.2646484375,
0.54248046875
],
[
0.26318359375,
0.54345703125,
0.46728515625,
0.54345703125,
0.46728515625,
0.556640625,
0.26318359375,
0.556640625
],
[
0.2646484375,
0.57080078125,
0.74609375,
0.57080078125,
0.74609375,
0.583984375,
0.2646484375,
0.583984375
],
[
0.2646484375,
0.5859375,
0.3203125,
0.5859375,
0.3203125,
0.59765625,
0.2646484375,
0.59765625
],
[
0.2181396484375,
0.626953125,
0.236083984375,
0.626953125,
0.236083984375,
0.64013671875,
0.2181396484375,
0.64013671875
],
[
0.2484130859375,
0.626953125,
0.373291015625,
0.626953125,
0.373291015625,
0.6416015625,
0.2484130859375,
0.6416015625
],
[
0.2181396484375,
0.66015625,
0.7861328125,
0.66015625,
0.7861328125,
0.6748046875,
0.2181396484375,
0.6748046875
],
[
0.21728515625,
0.67529296875,
0.7861328125,
0.67529296875,
0.7861328125,
0.68994140625,
0.21728515625,
0.68994140625
],
[
0.21728515625,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.70556640625,
0.21728515625,
0.70556640625
],
[
0.2181396484375,
0.70556640625,
0.7861328125,
0.70556640625,
0.7861328125,
0.72021484375,
0.2181396484375,
0.72021484375
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.2181396484375,
0.736328125,
0.78759765625,
0.736328125,
0.78759765625,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.751953125,
0.7861328125,
0.751953125,
0.7861328125,
0.76513671875,
0.2181396484375,
0.76513671875
],
[
0.2156982421875,
0.7666015625,
0.7861328125,
0.76513671875,
0.7861328125,
0.77978515625,
0.2156982421875,
0.78076171875
],
[
0.21728515625,
0.77978515625,
0.78515625,
0.78076171875,
0.78515625,
0.79541015625,
0.21728515625,
0.79443359375
],
[
0.2197265625,
0.796875,
0.78759765625,
0.796875,
0.78759765625,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78759765625,
0.8125,
0.78759765625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2156982421875,
0.8271484375,
0.4150390625,
0.82568359375,
0.4150390625,
0.84033203125,
0.2156982421875,
0.84130859375
]
] |
[
"Learning impartial policies for sequential",
"counterfactual explanations using Deep",
"Reinforcement Learning",
"Emmanouil Panagiotoul,2000-0001-9134-9387] and Eirini",
"Ntoutsi2[0000-0001-5729-1003]",
"",
"1 Freie Universitat Berlin, Department of Mathematics & Computer Science, Berlin,",
"Germany",
"emmanouil.panagiotou@fu-berlin.de",
"2 Universitat der Bundeswehr Minchen, Faculty for Informatik, Munich, Germany",
"",
"Abstract. In the field of explainable Artificial Intelligence (XAI), se-",
"[T'S]",
"quential counterfactual (SCF) examples are often used to alter the deci-",
"sion of a trained classifier by implementing a sequence of modifications",
"to the input instance. Although certain test-time algorithms aim to opti-",
"mize for each new instance individually, recently Reinforcement Learning",
"(RL) methods have been proposed that seek to learn policies for discov-",
"ering SCFs, thereby enhancing scalability. As is typical in RL, the for-",
"mulation of the RL problem, including the specification of state space,",
"AE00",
"actions, and rewards, can often be ambiguous. In this work, we identify",
"shortcomings in existing methods that can result in policies with unde-",
"sired properties, such as a bias towards specific actions. We propose to",
"use the output probabilities of the classifier to create a more informative",
"reward, to mitigate this effect.",
"Keywords: Sequential counterfactuals -Reinforcement Learning ·Model-",
"agnostic",
"1",
"Introduction",
"Predictive Machine Learning (ML) models have been used in various fields to",
"make predictions by generalizing knowledge learned from data. Yet, in recent",
"years, generating explanations for model decisions has become increasingly im-",
"portant, shifting interest to the field of explainable AI (XAI) [7]. A popular",
"category of XAI methods are the so-called “counterfactual explanations\" that",
"offer a \"what-if\" analysis of model decisions by identifying the minimal set of",
"changes or interventions needed to alter the outcome of the model for a given",
"instance, allowing users to understand the factors that influenced the model's",
"decision. Common desiderata for counterfactual explanations, as highlighted by",
"[4], include prorimity (distance) to the original instance, sparsity in terms of",
"changes made, and the plausibility of the counterfactual instances w.r.t. the",
"particular problem domain."
] |
{
"arxiv_id": "2311.00523",
"language": "en",
"page_idx": 0
}
|
||
2311.00523-en-1
|
2311.00523-en-1.jpg
|
2 E. Panagiotou et al.
Recently, a specific type of counterfactual explanations, called sequential counterfactuals (SCFs) has been proposed [14,16,17,13,15,5] that, rather than making instantaneous or simultaneous modifications to the input, propose a sequence of alterations, each building upon the previous one, until the classifier's decision changes. Such approaches allow for considering the consequences of cer- $\tan$changes to other attributes, for instance, increasing the age of an individual might require an increase in their level of education to maintain consistency or plausibility. SCF generation algorithms look for the optimal order of changes to an input instance, to alter the decision of a model.
In the generation of SCFs, certain methods employ search algorithms that need to initiate optimization for each new instance [14,8], leading to efficiency problems. On the other hand, feed-forward methods [18,3] leverage Reinforcement Learning (RL) to learn a scalable policy for generating SCFs. The problem is formulated as a Markov Decision Process (MDP), where states present the original input instance and its alternations, and actions correspond to all allowed changes to input features (discrete or continuous). The immediate $re$- ward for taking an action depends on the output of the black-box model and on the distance of the altered instance to its original state (proximity), to ensure minimal feature changes (action sparsity).
As is typical in RL, the problem formulation plays a crucial role in shaping the policy that the agent learns. In this study, we focus on analyzing certain modeling choices, specifically those related to the action space and the shaping of the reward function, which is influenced by the output of the classifier. We observe that rewarding the RL agent based solely on the sparse binary decision of the classifier can result in problematic policies that only modify features highly correlated with the classifier's output. While these methods demonstrate satisfactory performance in terms of action sparsity, satisfability, and distance, they prove to be ineffective in practice as they learn to take identical actions for any given input, leading to what we call $feature\textit{ over- }utilization.$ In extreme cases, this phenomenon leads to altering a single feature for any given input, thereby reducing the practicality and diversity of generated counterfactual explanations. We present an action entropy metric as a means to quantify the extent of feature over-utilization. Moreover, we propose to overcome this issue by taking continuous actions and implementing a denser reward function when the classifier's output probability is available instead of solely relying on the class labels.
The rest of this paper is organized as follows: related work is discussed in Section 2. In Section 3 we introduce our proposed method. Experimental results are provided in Section 4. Conclusions and future work are discussed in Section 5.
## 2 Related Work
In recent years, many XAI techniques have been proposed to solve the problem of finding SCFs for a pre-trained model. Model-agnostic methods are more generalizable, as they do not require differentiability or information on the internal
|
[
[
0.2181396484375,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.2181396484375,
0.1275634765625
],
[
0.26953125,
0.1142578125,
0.406982421875,
0.11553955078125,
0.406982421875,
0.1300048828125,
0.26953125,
0.1287841796875
],
[
0.24267578125,
0.147705078125,
0.78759765625,
0.147705078125,
0.78759765625,
0.1622314453125,
0.24267578125,
0.1622314453125
],
[
0.21728515625,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.177978515625,
0.21728515625,
0.177978515625
],
[
0.2181396484375,
0.1787109375,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.21728515625,
0.1932373046875,
0.7861328125,
0.1925048828125,
0.7861328125,
0.20703125,
0.21728515625,
0.2083740234375
],
[
0.2181396484375,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.78759765625,
0.2235107421875,
0.78759765625,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.21728515625,
0.238037109375,
0.7861328125,
0.2392578125,
0.7861328125,
0.25390625,
0.21728515625,
0.25244140625
],
[
0.21728515625,
0.25244140625,
0.78759765625,
0.25390625,
0.78759765625,
0.268310546875,
0.21728515625,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.5810546875,
0.26953125,
0.5810546875,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2418212890625,
0.2841796875,
0.78515625,
0.2841796875,
0.78515625,
0.29736328125,
0.2418212890625,
0.29736328125
],
[
0.2156982421875,
0.298583984375,
0.78515625,
0.2998046875,
0.78515625,
0.314453125,
0.2156982421875,
0.313232421875
],
[
0.21728515625,
0.313232421875,
0.78369140625,
0.314453125,
0.78369140625,
0.328857421875,
0.21728515625,
0.32763671875
],
[
0.2181396484375,
0.330078125,
0.78369140625,
0.330078125,
0.78369140625,
0.343505859375,
0.2181396484375,
0.343505859375
],
[
0.21728515625,
0.343505859375,
0.78515625,
0.3447265625,
0.78515625,
0.359130859375,
0.21728515625,
0.35791015625
],
[
0.2181396484375,
0.35986328125,
0.7861328125,
0.35986328125,
0.7861328125,
0.373779296875,
0.2181396484375,
0.373779296875
],
[
0.2181396484375,
0.375732421875,
0.7861328125,
0.375732421875,
0.7861328125,
0.389404296875,
0.2181396484375,
0.389404296875
],
[
0.2197265625,
0.39013671875,
0.7861328125,
0.39013671875,
0.7861328125,
0.404052734375,
0.2197265625,
0.404052734375
],
[
0.21728515625,
0.404052734375,
0.7861328125,
0.406005859375,
0.7861328125,
0.42041015625,
0.21728515625,
0.41845703125
],
[
0.2181396484375,
0.42041015625,
0.51953125,
0.42041015625,
0.51953125,
0.434326171875,
0.2181396484375,
0.434326171875
],
[
0.2442626953125,
0.436279296875,
0.7861328125,
0.436279296875,
0.7861328125,
0.449462890625,
0.2442626953125,
0.449462890625
],
[
0.21728515625,
0.45068359375,
0.7861328125,
0.45068359375,
0.7861328125,
0.46533203125,
0.21728515625,
0.46533203125
],
[
0.21728515625,
0.46533203125,
0.7861328125,
0.466552734375,
0.7861328125,
0.48095703125,
0.21728515625,
0.479736328125
],
[
0.2181396484375,
0.48095703125,
0.78369140625,
0.48095703125,
0.78369140625,
0.494384765625,
0.2181396484375,
0.494384765625
],
[
0.2181396484375,
0.49560546875,
0.78759765625,
0.49560546875,
0.78759765625,
0.5087890625,
0.2181396484375,
0.5087890625
],
[
0.2181396484375,
0.51123046875,
0.78515625,
0.51123046875,
0.78515625,
0.5244140625,
0.2181396484375,
0.5244140625
],
[
0.2181396484375,
0.52587890625,
0.78369140625,
0.52587890625,
0.78369140625,
0.54052734375,
0.2181396484375,
0.54052734375
],
[
0.21728515625,
0.54052734375,
0.78515625,
0.54150390625,
0.78515625,
0.55615234375,
0.21728515625,
0.5546875
],
[
0.2181396484375,
0.556640625,
0.78515625,
0.556640625,
0.78515625,
0.57080078125,
0.2181396484375,
0.57080078125
],
[
0.2181396484375,
0.5712890625,
0.78515625,
0.5712890625,
0.78515625,
0.5859375,
0.2181396484375,
0.5859375
],
[
0.21728515625,
0.5859375,
0.78369140625,
0.5859375,
0.78369140625,
0.599609375,
0.21728515625,
0.599609375
],
[
0.21728515625,
0.6015625,
0.78515625,
0.60302734375,
0.78515625,
0.6162109375,
0.21728515625,
0.61572265625
],
[
0.2197265625,
0.6171875,
0.7861328125,
0.6171875,
0.7861328125,
0.630859375,
0.2197265625,
0.630859375
],
[
0.21728515625,
0.630859375,
0.78369140625,
0.6318359375,
0.78369140625,
0.646484375,
0.21728515625,
0.64599609375
],
[
0.2181396484375,
0.64794921875,
0.78759765625,
0.64794921875,
0.78759765625,
0.6611328125,
0.2181396484375,
0.6611328125
],
[
0.2181396484375,
0.66357421875,
0.74658203125,
0.66357421875,
0.74658203125,
0.6767578125,
0.2181396484375,
0.6767578125
],
[
0.24267578125,
0.67822265625,
0.7861328125,
0.67822265625,
0.7861328125,
0.69140625,
0.24267578125,
0.69140625
],
[
0.21728515625,
0.69140625,
0.7861328125,
0.6923828125,
0.7861328125,
0.70703125,
0.21728515625,
0.70556640625
],
[
0.2181396484375,
0.70849609375,
0.7861328125,
0.70849609375,
0.7861328125,
0.7216796875,
0.2181396484375,
0.7216796875
],
[
0.21728515625,
0.7236328125,
0.2344970703125,
0.7236328125,
0.2344970703125,
0.7373046875,
0.21728515625,
0.7373046875
],
[
0.2181396484375,
0.76318359375,
0.236083984375,
0.76318359375,
0.236083984375,
0.7763671875,
0.2181396484375,
0.7763671875
],
[
0.2484130859375,
0.76318359375,
0.386474609375,
0.76318359375,
0.386474609375,
0.77783203125,
0.2484130859375,
0.77783203125
],
[
0.21728515625,
0.79541015625,
0.78515625,
0.79541015625,
0.78515625,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.21728515625,
0.81005859375,
0.78369140625,
0.8125,
0.78369140625,
0.8271484375,
0.21728515625,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"2",
"E. Panagiotou et al.",
"Recently, a specific type of counterfactual explanations, called sequential",
"counterfactuals (SCFs) has been proposed [14,16,17,13,15,5] that, rather than",
"making instantaneous or simultaneous modifications to the input, propose a se-",
"quence of alterations, each building upon the previous one, until the classifier's",
"decision changes. Such approaches allow for considering the consequences of cer-",
"tain changes to other attributes, for instance, increasing the age of an individual",
"might require an increase in their level of education to maintain consistency or",
"plausibility. SCF generation algorithms look for the optimal order of changes to",
"an input instance, to alter the decision of a model.",
"In the generation of SCFs, certain methods employ search algorithms that",
"need to initiate optimization for each new instance [14,8], leading to efficiency",
"problems. On the other hand, feed-forward methods [18,3] leverage Reinforce-",
"ment Learning (RL) to learn a scalable policy for generating SCFs. The prob-",
"lem is formulated as a Markov Decision Process (MDP), where states present",
"the original input instance and its alternations, and actions correspond to all",
"allowed changes to input features (discrete or continuous). The immediate re-",
"ward for taking an action depends on the output of the black-box model and on",
"the distance of the altered instance to its original state (proximity), to ensure",
"minimal feature changes (action sparsity).",
"As is typical in RL, the problem formulation plays a crucial role in shaping",
"the policy that the agent learns. In this study, we focus on analyzing certain",
"modeling choices, specifically those related to the action space and the shaping",
"of the reward function, which is influenced by the output of the classifier. We ob-",
"serve that rewarding the RL agent based solely on the sparse binary decision of",
"the classifier can result in problematic policies that only modify features highly",
"correlated with the classifier's output. While these methods demonstrate satis",
"factory performance in terms of action sparsity, satisfiability, and distance, they",
"prove to be ineffective in practice as they learn to take identical actions for any",
"given input, leading to what we call feature over-utilization. In extreme cases,",
"this phenomenon leads to altering a single feature for any given input, thereby",
"reducing thepracticality and diversity of generated counterfactual explanations.",
"We present an action entropy metric as a means to quantify the extent of feature",
"over-utilization. Moreover, we propose to overcome this issue by taking contin-",
"uous actions and implementing a denser reward function when the classifier's",
"output probability is available instead of solely relying on the class labels.",
"The rest of this paper is organized as follows: related work is discussed in",
"Section 2. In Section 3 we introduce our proposed method. Experimental results",
"are provided in Section 4. Conclusions and future work are discussed in Section",
"5.",
"2",
"Related Work",
"In recent years, many XAI techniques have been proposed to solve the problem",
"of finding SCFs for a pre-trained model. Model-agnostic methods are more gen-",
"eralizable, as they do not require differentiability or information on the internal"
] |
{
"arxiv_id": "2311.00523",
"language": "en",
"page_idx": 1
}
|
||
2311.00523-en-2
|
2311.00523-en-2.jpg
|
Sequential counterfactual policies using RL 3
architecture of the models, but only use the output for decision-making. The majority are search algorithms that re-optimize for each new given instance, for example by searching in the neighborhood [2,1],or by population-based approaches [14,6]. These algorithms start from a given instance and navigate through trial and error to find a CF, often optimizing for multiple objectives. In our setting, a RL policy has to be learned that can be applied as a feed-forward method after training, on any given instance.
In [18] a Deep Q-Network (DQN) is used to predict discrete feature changes, which are defined as constant steps. Although the reward function incorporates classifier probabilities whenever feasible, its effect on the SCF generation is not discussed. Besides, the use of fxed increments presents issues by assuming all discrete features to be ordinal and restricting changes in continuous features to discrete steps only. This is further detailed in Section 3.
To overcome this problem, [3] proposes to use a parametrized DQN (P-DQN) [19] network, which combines a DQN with a policy gradient method (DDPG) [11], that can be described as the continuous counterpart of a DQN. Doing that, P-DQN can choose among all features using the DQN, and predict a more precise continuous feature change using the DDPG network. Nonetheless, the reward depends only on the class labels, which leads to over-utilization of some features as presented in our experiments.
## 3 Overcoming feature over-utilization through informative reward design
In this section, we address the issue of action over-utilization that arises when employing sparse rewards in RL techniques for generating SCFs. To overcome this issue, we propose the adoption of a denser reward function that employs the probabilities obtained from the black-box classifier.
3.1 Limitations of existing approaches
$\mathbf{DQN:As~in~[18],~a~DQN~takes~discrete~steps~of~(\pm0.05)~for~continuous~features}$ that are in the range of [-1,1],and $(\pm1)$ for discrete features. This formulation is problematic because i) some discrete features can be nominal, e.g. Occupations don't necessarily have a specific ordering, ii) for continuous features, the step size is discrete and constant, e.g. working hours $\pm30min$, and iii) the constant step size of $\pm0.05$ does not take feature interdependencies into account, for example, it can translate to $\pm30min$ for working hours, but $\pm2500EUR$ for capital gain, depending on the normalization method. The paper mentions that a dense reward using the black-box probabilities is applied when available. Furthermore, results report a very low action sparsity (close to 1) and distance (0.04), which is desired when generating SCFs, as a lesser number of changes, closer to the original instance, are more interpretable. However, this also means that for all input instances, CFs are found after changing a single feature just once, as the distance is almost the step size of 0.05, leading to feature over-utilization and
|
[
[
0.4453125,
0.11553955078125,
0.734375,
0.11553955078125,
0.734375,
0.1287841796875,
0.4453125,
0.1287841796875
],
[
0.77392578125,
0.117431640625,
0.78759765625,
0.117431640625,
0.78759765625,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.2181396484375,
0.147705078125,
0.78369140625,
0.147705078125,
0.78369140625,
0.1610107421875,
0.2181396484375,
0.1610107421875
],
[
0.21728515625,
0.1634521484375,
0.78369140625,
0.1634521484375,
0.78369140625,
0.1767578125,
0.21728515625,
0.1767578125
],
[
0.21728515625,
0.1767578125,
0.78515625,
0.177978515625,
0.78515625,
0.1925048828125,
0.21728515625,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.7861328125,
0.208984375,
0.7861328125,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.21728515625,
0.2235107421875,
0.78369140625,
0.2235107421875,
0.78369140625,
0.2374267578125,
0.21728515625,
0.2374267578125
],
[
0.2181396484375,
0.2392578125,
0.443603515625,
0.2392578125,
0.443603515625,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.24267578125,
0.25244140625,
0.78369140625,
0.25244140625,
0.78369140625,
0.267578125,
0.24267578125,
0.267578125
],
[
0.2181396484375,
0.268310546875,
0.7861328125,
0.268310546875,
0.7861328125,
0.2822265625,
0.2181396484375,
0.2822265625
],
[
0.2181396484375,
0.2841796875,
0.78515625,
0.2841796875,
0.78515625,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.2181396484375,
0.2998046875,
0.7861328125,
0.2998046875,
0.7861328125,
0.313232421875,
0.2181396484375,
0.313232421875
],
[
0.2181396484375,
0.314453125,
0.7861328125,
0.314453125,
0.7861328125,
0.32763671875,
0.2181396484375,
0.32763671875
],
[
0.21728515625,
0.32763671875,
0.62109375,
0.328857421875,
0.62109375,
0.343505859375,
0.21728515625,
0.34228515625
],
[
0.24267578125,
0.343505859375,
0.78369140625,
0.34228515625,
0.78369140625,
0.356689453125,
0.24267578125,
0.35791015625
],
[
0.2197265625,
0.359130859375,
0.78369140625,
0.359130859375,
0.78369140625,
0.37255859375,
0.2197265625,
0.37255859375
],
[
0.221435546875,
0.375,
0.7861328125,
0.375,
0.7861328125,
0.38818359375,
0.221435546875,
0.38818359375
],
[
0.2181396484375,
0.389404296875,
0.7861328125,
0.389404296875,
0.7861328125,
0.404052734375,
0.2181396484375,
0.404052734375
],
[
0.2181396484375,
0.40478515625,
0.7861328125,
0.40478515625,
0.7861328125,
0.41845703125,
0.2181396484375,
0.41845703125
],
[
0.2197265625,
0.421630859375,
0.78369140625,
0.421630859375,
0.78369140625,
0.431884765625,
0.2197265625,
0.431884765625
],
[
0.2181396484375,
0.43505859375,
0.51318359375,
0.43505859375,
0.51318359375,
0.448974609375,
0.2181396484375,
0.448974609375
],
[
0.2197265625,
0.47607421875,
0.233642578125,
0.47607421875,
0.233642578125,
0.48486328125,
0.2197265625,
0.48486328125
],
[
0.25244140625,
0.473388671875,
0.68310546875,
0.473388671875,
0.68310546875,
0.488037109375,
0.25244140625,
0.488037109375
],
[
0.250732421875,
0.4912109375,
0.5048828125,
0.4912109375,
0.5048828125,
0.505859375,
0.250732421875,
0.505859375
],
[
0.2181396484375,
0.521484375,
0.7861328125,
0.521484375,
0.7861328125,
0.5361328125,
0.2181396484375,
0.5361328125
],
[
0.2181396484375,
0.537109375,
0.7861328125,
0.537109375,
0.7861328125,
0.55029296875,
0.2181396484375,
0.55029296875
],
[
0.2181396484375,
0.55322265625,
0.7861328125,
0.55322265625,
0.7861328125,
0.56640625,
0.2181396484375,
0.56640625
],
[
0.2181396484375,
0.5673828125,
0.5849609375,
0.5673828125,
0.5849609375,
0.58056640625,
0.2181396484375,
0.58056640625
],
[
0.21728515625,
0.60498046875,
0.5439453125,
0.60498046875,
0.5439453125,
0.619140625,
0.21728515625,
0.619140625
],
[
0.2181396484375,
0.62890625,
0.7861328125,
0.62890625,
0.7861328125,
0.64306640625,
0.2181396484375,
0.64306640625
],
[
0.2181396484375,
0.64599609375,
0.7861328125,
0.64599609375,
0.7861328125,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.2181396484375,
0.66015625,
0.78515625,
0.66015625,
0.78515625,
0.673828125,
0.2181396484375,
0.673828125
],
[
0.21728515625,
0.67529296875,
0.78515625,
0.67529296875,
0.78515625,
0.689453125,
0.21728515625,
0.689453125
],
[
0.2181396484375,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.2181396484375,
0.705078125
],
[
0.2181396484375,
0.70556640625,
0.78369140625,
0.70556640625,
0.78369140625,
0.7197265625,
0.2181396484375,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.2181396484375,
0.734375
],
[
0.2181396484375,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75,
0.2181396484375,
0.75
],
[
0.2181396484375,
0.75048828125,
0.78369140625,
0.75048828125,
0.78369140625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2181396484375,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.78076171875,
0.7861328125,
0.78076171875,
0.7861328125,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.2197265625,
0.79736328125,
0.78515625,
0.79736328125,
0.78515625,
0.80810546875,
0.2197265625,
0.80810546875
],
[
0.2181396484375,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2181396484375,
0.82470703125
],
[
0.21728515625,
0.8271484375,
0.78759765625,
0.8271484375,
0.78759765625,
0.84033203125,
0.21728515625,
0.84033203125
]
] |
[
"Sequential counterfactual policies using RL",
"3",
"architecture of the models, but only use the output for decision-making. The ma-",
"jority are search algorithms that re-optimize for each new given instance, for ex-",
"ample by searching in the neighborhood [2,1], or by population-based approaches",
"[14,6]. These algorithms start from a given instance and navigate through trial",
"and error to find a CF, often optimizing for multiple objectives. In our setting, a",
"RL policy has to be learned that can be applied as a feed-forward method after",
"training, on any given instance.",
"In [18] a Deep Q-Network (DQN) is used to predict discrete feature changes,",
"which are defined as constant steps. Although the reward function incorporates",
"classifier probabilities whenever feasible, its effect on the SCF generation is not",
"discussed. Besides, the use of fixed increments presents issues by assuming all",
"discrete features to be ordinal and restricting changes in continuous features to",
"discrete steps only.This is further detailed in Section 3.",
"To overcome this problem, [3] proposes to use a parametrized DQN (P-DQN)",
"[19] network, which combines a DQN with a policy gradient method (DDPG)",
"[1i], that can be described as the continuous counterpart of a DQN. Doing",
"that, P-DQN can choose among all features using the DQN, and predict a more",
"precise continuous feature change using the DDPG network. Nonetheless, the",
"reward depends onlyon theclasslabels.whichleadstoover-utilizationof some",
"features as presented in our experiments.",
"3",
"Overcoming feature over-utilization through",
"informative reward design",
"In this section, we address the issue of action over-utilization that arises when",
"employing sparse rewards in RL techniques for generating SCFs. To overcome",
"this issue, we propose the adoption of a denser reward function that employs the",
"probabilities obtained from the black-box classifier.",
"3.1Limitations of existing approaches",
"DQN: As in [18], a DQN takes discrete steps of (±0.05) for continuous features",
"that are in the range of [-1, 1l, and (±1) for discrete features. This formulation is",
"problematic because i) some discrete features can be nominal, e.g. Occupations",
"don't necessarily have a specific ordering, ii) for continuous features, the step",
"size is discrete and constant, e.g. working hours ±30min, and ii) the constant",
"step size of ±0.05 does not take feature interdependencies into account, for ex-",
"ample, it can translate to ±30min for working hours, but ±2500EUR for capital",
"gain, depending on the normalization method. The paper mentions that a dense",
"reward using the black-box probabilities is applied when available. Furthermore.",
"results report a very low action sparsity (close to 1) and distance (0.04), which",
"is desired when generating SCFs, as a lesser number of changes, closer to the",
"original instance, are more interpretable.However,this also means that for all",
"input instances, CFs are found after changing a single feature just once, as the",
"distance is almost the step size of 0.05, leading to feature over-utilization and"
] |
{
"arxiv_id": "2311.00523",
"language": "en",
"page_idx": 2
}
|
||
2311.00523-en-3
|
2311.00523-en-3.jpg
|
4 E. Panagiotou et al.
scarcity of diverse options. Such an unvarying single-step policy does not provide useful SCF explanations, and potentially other XAI approaches such as global feature importance methods [10,12] are more ftting in this case. Due to these many disadvantages, we do not follow this approach.
P-DQN: As proposed in [3] the P-DQN [19] model is used to perform continuous changes sequentially. Taking an action consists of a high-level feature choice $k$ (discrete) and the new value of that feature $u_k$ (continuous).This assumption considers all feature changes $u_k$ to be continuous, which necessitates discretization when handling discrete features. However, P-DQN remains more accurate and adaptable compared to the DQN method, as it enables continuous changes for continuous features without the restriction of discrete steps. Analytically, the action space for P-DQN is defined as:
$$A=\left\{(k,u_k)\mid u_k\in[-1,1]\quad k\in[K]\right\}$$
(1)
Where $[K]=\{1,2,...,K\}$ are the different features. Each time a feature is changed using an action, the timestep $t$ is incremented. A state $s_t=(x_t,b_t)$ is defned as a tuple of some instance $x_t$ concatenated with a binary indicator vector of previously used actions $b_t.$ When taking an action $a_t=(k,u_k)$ the new state $s_t+1=(x_{t+1},b_{t+1})$ is created by updating the value of the chosen feature $k$ as $u_{k}$, i.e. $x_t+1[k]=u_{k}$, and incrementing the indicator vector $b_t+1[k]+=1.$ Finally, the reward function used in this method [3] is defined as:
$$R_t^{bin}=\begin{Bmatrix}Pen&\text{if}\:s_{t+1}=\text{failure,}\\\\Pos&\text{if}\:s_{t+1}=\text{success,}\\\\0&else.\end{Bmatrix}-\beta*\delta(s_{t+1},s_0)$$
(2)
Where Pen is a negative penalty if a failure terminal state is reached, i.e. violating a feature constraint or re-using the same action, Pos is a large positive reward if a success terminal state is reached (counterfactual found), $\delta()$ is a distance measure (the $l_2$ norm in our case), and $\beta$ is a parameter controlling the weight of the distance measure. It is evident that this reward function is sparse since it only has access to the binary class labels and not the probabilities. Therefore we refer to this reward function as $R^bin.$ In Section 4, we demonstrate that the learned policy using this reward is suboptimal in terms of action variety.
3.2 P-DQN &classifier probability rewards
To construct a more dense reward, we assume that the classifier $f(\mathcal{X})=\mathcal{Y}$ makes a binary decision $\mathcal{Y}$ given an input $\mathcal{X}$ based on a probability measure. We denote the probability of an instance belonging to the target class as $P(\mathcal{X})=P(f(\mathcal{X})=$ target). This way we can redefine the reward function for a non-terminal state as follows:
(3)
$$R_t^{prob}=\alpha*\left(P(s_{t+1})-P(s_0)\right)-\beta*\delta(s_{t+1},s_0)$$
|
[
[
0.21728515625,
0.11614990234375,
0.2296142578125,
0.11614990234375,
0.2296142578125,
0.1275634765625,
0.21728515625,
0.1275634765625
],
[
0.26953125,
0.1142578125,
0.406005859375,
0.11553955078125,
0.406005859375,
0.1300048828125,
0.26953125,
0.1287841796875
],
[
0.2181396484375,
0.147705078125,
0.78515625,
0.147705078125,
0.78515625,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2197265625,
0.1634521484375,
0.7861328125,
0.1634521484375,
0.7861328125,
0.177978515625,
0.2197265625,
0.177978515625
],
[
0.2181396484375,
0.1787109375,
0.78515625,
0.1787109375,
0.78515625,
0.1925048828125,
0.2181396484375,
0.1925048828125
],
[
0.2197265625,
0.1932373046875,
0.5947265625,
0.1932373046875,
0.5947265625,
0.2083740234375,
0.2197265625,
0.2083740234375
],
[
0.2197265625,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.222900390625,
0.2197265625,
0.222900390625
],
[
0.2181396484375,
0.2247314453125,
0.78515625,
0.2247314453125,
0.78515625,
0.238037109375,
0.2181396484375,
0.238037109375
],
[
0.21728515625,
0.238037109375,
0.7861328125,
0.2392578125,
0.7861328125,
0.25390625,
0.21728515625,
0.25244140625
],
[
0.2181396484375,
0.255126953125,
0.78515625,
0.255126953125,
0.78515625,
0.268310546875,
0.2181396484375,
0.268310546875
],
[
0.2181396484375,
0.26953125,
0.78515625,
0.26953125,
0.78515625,
0.2841796875,
0.2181396484375,
0.2841796875
],
[
0.2181396484375,
0.285400390625,
0.7861328125,
0.285400390625,
0.7861328125,
0.298583984375,
0.2181396484375,
0.298583984375
],
[
0.21728515625,
0.298583984375,
0.7861328125,
0.2998046875,
0.7861328125,
0.314453125,
0.21728515625,
0.313232421875
],
[
0.21728515625,
0.314453125,
0.490234375,
0.31494140625,
0.490234375,
0.330078125,
0.21728515625,
0.328857421875
],
[
0.369384765625,
0.3447265625,
0.63623046875,
0.3447265625,
0.63623046875,
0.365478515625,
0.369384765625,
0.365478515625
],
[
0.76123046875,
0.34521484375,
0.7890625,
0.34521484375,
0.7890625,
0.3642578125,
0.76123046875,
0.3642578125
],
[
0.2418212890625,
0.375,
0.7861328125,
0.373779296875,
0.7861328125,
0.391357421875,
0.2418212890625,
0.392578125
],
[
0.21728515625,
0.391357421875,
0.78515625,
0.392578125,
0.78515625,
0.4072265625,
0.21728515625,
0.406005859375
],
[
0.2181396484375,
0.40771484375,
0.78369140625,
0.40771484375,
0.78369140625,
0.421630859375,
0.2181396484375,
0.421630859375
],
[
0.2181396484375,
0.42236328125,
0.78515625,
0.42236328125,
0.78515625,
0.4375,
0.2181396484375,
0.4375
],
[
0.2181396484375,
0.4375,
0.7861328125,
0.4375,
0.7861328125,
0.451904296875,
0.2181396484375,
0.451904296875
],
[
0.2181396484375,
0.45263671875,
0.78515625,
0.45263671875,
0.78515625,
0.46728515625,
0.2181396484375,
0.46728515625
],
[
0.2181396484375,
0.468505859375,
0.685546875,
0.468505859375,
0.685546875,
0.481689453125,
0.2181396484375,
0.481689453125
],
[
0.375732421875,
0.496826171875,
0.42236328125,
0.496826171875,
0.42236328125,
0.51220703125,
0.375732421875,
0.51220703125
],
[
0.431396484375,
0.49560546875,
0.564453125,
0.49560546875,
0.564453125,
0.51318359375,
0.431396484375,
0.51318359375
],
[
0.314453125,
0.51953125,
0.36279296875,
0.5166015625,
0.365234375,
0.5322265625,
0.316162109375,
0.5361328125
],
[
0.5849609375,
0.51953125,
0.68896484375,
0.521484375,
0.68896484375,
0.5390625,
0.5849609375,
0.537109375
],
[
0.76123046875,
0.52099609375,
0.78759765625,
0.52099609375,
0.78759765625,
0.5390625,
0.76123046875,
0.5390625
],
[
0.42724609375,
0.521484375,
0.55126953125,
0.521484375,
0.55126953125,
0.5361328125,
0.42724609375,
0.5361328125
],
[
0.38232421875,
0.52294921875,
0.4150390625,
0.52294921875,
0.4150390625,
0.53515625,
0.38232421875,
0.53515625
],
[
0.375732421875,
0.54541015625,
0.396240234375,
0.54541015625,
0.396240234375,
0.5595703125,
0.375732421875,
0.5595703125
],
[
0.5185546875,
0.54345703125,
0.56640625,
0.54345703125,
0.56640625,
0.5625,
0.5185546875,
0.5625
],
[
0.2442626953125,
0.57080078125,
0.78369140625,
0.57080078125,
0.78369140625,
0.58544921875,
0.2442626953125,
0.58544921875
],
[
0.2181396484375,
0.5859375,
0.78515625,
0.5859375,
0.78515625,
0.599609375,
0.2181396484375,
0.599609375
],
[
0.2197265625,
0.6015625,
0.7861328125,
0.6015625,
0.7861328125,
0.61572265625,
0.2197265625,
0.61572265625
],
[
0.2197265625,
0.6171875,
0.78515625,
0.6171875,
0.78515625,
0.630859375,
0.2197265625,
0.630859375
],
[
0.2197265625,
0.6318359375,
0.78515625,
0.6318359375,
0.78515625,
0.64599609375,
0.2197265625,
0.64599609375
],
[
0.2181396484375,
0.646484375,
0.78369140625,
0.646484375,
0.78369140625,
0.66015625,
0.2181396484375,
0.66015625
],
[
0.2197265625,
0.662109375,
0.78515625,
0.662109375,
0.78515625,
0.67529296875,
0.2197265625,
0.67529296875
],
[
0.2181396484375,
0.67822265625,
0.78515625,
0.67822265625,
0.78515625,
0.69140625,
0.2181396484375,
0.69140625
],
[
0.21728515625,
0.71630859375,
0.5849609375,
0.71630859375,
0.5849609375,
0.73095703125,
0.21728515625,
0.73095703125
],
[
0.2181396484375,
0.7470703125,
0.78515625,
0.7470703125,
0.78515625,
0.76123046875,
0.2181396484375,
0.76123046875
],
[
0.2181396484375,
0.76220703125,
0.7861328125,
0.76220703125,
0.7861328125,
0.77587890625,
0.2181396484375,
0.77587890625
],
[
0.2181396484375,
0.77783203125,
0.78515625,
0.77783203125,
0.78515625,
0.79150390625,
0.2181396484375,
0.79150390625
],
[
0.21728515625,
0.79345703125,
0.7861328125,
0.79345703125,
0.7861328125,
0.806640625,
0.21728515625,
0.806640625
],
[
0.2181396484375,
0.8095703125,
0.293212890625,
0.8095703125,
0.293212890625,
0.82080078125,
0.2181396484375,
0.82080078125
],
[
0.33251953125,
0.8212890625,
0.41259765625,
0.8212890625,
0.41259765625,
0.8388671875,
0.33251953125,
0.8388671875
],
[
0.420654296875,
0.8212890625,
0.67333984375,
0.82275390625,
0.67333984375,
0.84130859375,
0.420654296875,
0.84033203125
],
[
0.76318359375,
0.82470703125,
0.7861328125,
0.82470703125,
0.7861328125,
0.8388671875,
0.76318359375,
0.8388671875
]
] |
[
"4",
"E. Panagiotou et al.",
"scarcity of diverse options. Such an unvarying single-step policy does not provide",
"useful SCF explanations, and potentially other XAI approaches such as global",
"feature importance methods [1o,12] are more fitting in this case. Due to these",
"many disadvantages, we do not follow this approach.",
"P-DQN: As proposed in [3] the P-DQN [19] model is used to perform continu-",
"ous changes sequentially. Taking an action consists of a high-level feature choice",
"k (discrete) and the new value of that feature uk (continuous). This assumption",
"considers all feature changes uk to be continuous, which necessitates discretiza-",
"tion when handling discrete features. However, P-DQN remains more accurate",
"and adaptable compared to the DQN method, as it enables continuous changes",
"for continuous features without the restriction of discrete steps. Analytically, the",
"action space for P-DQN is defined as:",
"A= [(k, uk) [ uk E[-1,1] ke[K]]",
"(1) ",
"Where [K] = (1, 2, ., K} are the different features. Each time a feature is",
"changed using an action, the timestep t is incremented. A state st = (rt,bt)",
"is defined as a tuple of some instance rt concatenated with a binary indicator",
"vector of previously used actions bt. When taking an action at = (k, uk) the new",
"state st+1 = (ct+1, bt+1) is created by updating the value of the chosen feature",
"k as uk, i.e. at+i[k] = uk, and incrementing the indicator vector bt+i[k] += 1.",
"Finally, the reward function used in this method [3] is defined as:",
"(Pen",
"if st+i =failure,)",
"Rbin",
"β * 8(st+1, 80)",
"(2)",
"if st+i = success",
"Pos",
"(0",
"else.J",
"Where Pen is a negative penalty if a failure terminal state is reached, i.e.",
"violating a feature constraint or re-using the same action, Pos is a large positive",
"reward if a success terminal state is reached (counterfactual found), o() is a",
"distance measure (the l2 norm in our case), and β is a parameter controlling the",
"weight of the distance measure. It is evident that this reward function is sparse",
"since it only has access to the binary class labels and not the probabilities.",
"Therefore we refer to this reward function as Rbin. In Section 4, we demonstrate",
"that the learned policy using this reward is suboptimal in terms of action variety.",
"3.2P-DQN &classifierprobabilityrewards",
"To construct a more dense reward, we assume that the classifier f(X) = makes",
"a binary decision given an input X based on a probability measure. We denote",
"the probability of an instance belonging to the target class as P(X) = P(f(X) =",
"target). This way we can redefine the reward function for a non-terminal state",
"as follows:",
"Rprob = Q*",
"(P(s$+1) - P(so)) - β * (s++1, 8o)",
"(3)"
] |
{
"arxiv_id": "2311.00523",
"language": "en",
"page_idx": 3
}
|
||
2311.00523-en-4
|
2311.00523-en-4.jpg
|
Sequential counterfactual policies using RL 5
Where $P(s_{t+1})-P(s_{0})$ is the difference in target class probability between the next state and the initial state and $\alpha,\beta$ are parameters controlling the balance between the difference of probability and distance respectively. This is beneficial because the RL agent has immediate feedback on each step regarding the direction taken in relation to the target class. In contrast, when using the aforementioned binary reward $(R^{bin})$, the agent must explore the space without receiving any positive feedback until a successful state is reached (found CF). This additional information on every step results in a more dense reward function, that not only discourages large changes but also guides towards the target class. We denote this reward function with $R^prob$ since it incorporates the utilization of target probabilities.
### 4 Experiments &Results
In this section, we compare the P-DQN model's performance when using the binary reward $R^bin$ and when employing the target class probabilities reward $R^{prob}.$
## 4.1 Datasets
Motivated by previous approaches we use four tabular datasets in our comparative study. Specifically, the Adult Income, Credit Approval, German Credit, and German Risk[9]. Each dataset has a binary output label and is comprised of mixed features. Some datasets also have immutable features (e.g. Race) that are considered when defining the states but are not taken into account in the action $\operatorname{space}.$
Table 1. Characteristics of the four public datasets used.
<table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{D}$ataset</th>
<th rowspan="2">$\mathbf{N.}$ instances</th>
<th colspan="3">$\mathbf{Features}$</th>
</tr>
<tr>
<th>N. of Continuous</th>
<th>N. of Discrete</th>
<th>N. of Immutable $\begin{array}{c}{-1}\\{-1}\end{array}$</th>
</tr>
<tr>
<td>Adult $\operatorname{Income}$ 1</td>
<td>48842</td>
<td>4</td>
<td>3</td>
<td>4</td>
</tr>
<tr>
<td>Credit Approval</td>
<td>690</td>
<td>5</td>
<td>10</td>
<td>0</td>
</tr>
<tr>
<td>German Credit</td>
<td>1000</td>
<td>5</td>
<td>111</td>
<td>4</td>
</tr>
<tr>
<td>German Risk</td>
<td>1000</td>
<td>3</td>
<td>4</td>
<td>2</td>
</tr>
</tbody>
</table>
## 4.2 Evaluation measures
Based on previous works on SCF generation [3,18,14], we use the metrics of satisfability, distance (i.e. proximity), and action sparsity to evaluate the quality of the generated CF examples. Additionally, we measure the target probability of the found CF and define a novel entropy metric to evaluate the diversity of the actions taken by the policy. When an agent has learned to modify only a
|
[
[
0.4453125,
0.11553955078125,
0.73388671875,
0.11553955078125,
0.73388671875,
0.1287841796875,
0.4453125,
0.1287841796875
],
[
0.77392578125,
0.11871337890625,
0.7861328125,
0.11871337890625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.2181396484375,
0.147705078125,
0.7861328125,
0.147705078125,
0.7861328125,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2181396484375,
0.1622314453125,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2181396484375,
0.1767578125
],
[
0.21728515625,
0.1767578125,
0.7861328125,
0.177978515625,
0.7861328125,
0.1925048828125,
0.21728515625,
0.1912841796875
],
[
0.2181396484375,
0.1932373046875,
0.78369140625,
0.1932373046875,
0.78369140625,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2181396484375,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2156982421875,
0.2203369140625,
0.78759765625,
0.2215576171875,
0.78759765625,
0.2392578125,
0.2156982421875,
0.238037109375
],
[
0.2181396484375,
0.238037109375,
0.78515625,
0.238037109375,
0.78515625,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.267578125,
0.2181396484375,
0.267578125
],
[
0.2181396484375,
0.26953125,
0.78515625,
0.26953125,
0.78515625,
0.28271484375,
0.2181396484375,
0.28271484375
],
[
0.2181396484375,
0.2841796875,
0.78759765625,
0.2841796875,
0.78759765625,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.2156982421875,
0.2998046875,
0.36279296875,
0.298583984375,
0.36279296875,
0.313232421875,
0.2156982421875,
0.314453125
],
[
0.2197265625,
0.339111328125,
0.2344970703125,
0.339111328125,
0.2344970703125,
0.349853515625,
0.2197265625,
0.349853515625
],
[
0.25,
0.33837890625,
0.476318359375,
0.33837890625,
0.476318359375,
0.3515625,
0.25,
0.3515625
],
[
0.21728515625,
0.36865234375,
0.78369140625,
0.36865234375,
0.78369140625,
0.3818359375,
0.21728515625,
0.3818359375
],
[
0.21728515625,
0.3837890625,
0.7861328125,
0.3837890625,
0.7861328125,
0.3984375,
0.21728515625,
0.3984375
],
[
0.2156982421875,
0.3984375,
0.2607421875,
0.396484375,
0.261474609375,
0.410888671875,
0.21728515625,
0.41357421875
],
[
0.2197265625,
0.4375,
0.25244140625,
0.4375,
0.25244140625,
0.449462890625,
0.2197265625,
0.449462890625
],
[
0.2607421875,
0.438232421875,
0.334228515625,
0.438232421875,
0.334228515625,
0.449462890625,
0.2607421875,
0.449462890625
],
[
0.2181396484375,
0.462158203125,
0.78369140625,
0.462158203125,
0.78369140625,
0.4765625,
0.2181396484375,
0.4765625
],
[
0.2181396484375,
0.477783203125,
0.7861328125,
0.477783203125,
0.7861328125,
0.4912109375,
0.2181396484375,
0.4912109375
],
[
0.2181396484375,
0.492431640625,
0.7861328125,
0.492431640625,
0.7861328125,
0.505859375,
0.2181396484375,
0.505859375
],
[
0.21728515625,
0.505859375,
0.78515625,
0.5068359375,
0.78515625,
0.521484375,
0.21728515625,
0.52099609375
],
[
0.2181396484375,
0.52294921875,
0.7861328125,
0.52294921875,
0.7861328125,
0.5361328125,
0.2181396484375,
0.5361328125
],
[
0.2181396484375,
0.5390625,
0.26318359375,
0.5390625,
0.26318359375,
0.55029296875,
0.2181396484375,
0.55029296875
],
[
0.31201171875,
0.5771484375,
0.69140625,
0.5771484375,
0.69140625,
0.59033203125,
0.31201171875,
0.59033203125
],
[
0.259033203125,
0.60400390625,
0.31787109375,
0.60400390625,
0.31787109375,
0.61572265625,
0.259033203125,
0.61572265625
],
[
0.343017578125,
0.60302734375,
0.438720703125,
0.60302734375,
0.438720703125,
0.6142578125,
0.343017578125,
0.6142578125
],
[
0.57275390625,
0.60400390625,
0.6357421875,
0.60400390625,
0.6357421875,
0.61572265625,
0.57275390625,
0.61572265625
],
[
0.44287109375,
0.6162109375,
0.765625,
0.6162109375,
0.765625,
0.630859375,
0.44287109375,
0.630859375
],
[
0.240234375,
0.630859375,
0.336669921875,
0.630859375,
0.336669921875,
0.64599609375,
0.240234375,
0.64599609375
],
[
0.371826171875,
0.630859375,
0.4150390625,
0.630859375,
0.4150390625,
0.646484375,
0.371826171875,
0.646484375
],
[
0.490234375,
0.630859375,
0.5087890625,
0.630859375,
0.5087890625,
0.662109375,
0.490234375,
0.662109375
],
[
0.59716796875,
0.6318359375,
0.61376953125,
0.6318359375,
0.61376953125,
0.6484375,
0.59716796875,
0.6484375
],
[
0.701171875,
0.630859375,
0.7197265625,
0.630859375,
0.7197265625,
0.6640625,
0.701171875,
0.6640625
],
[
0.37744140625,
0.64453125,
0.408447265625,
0.64453125,
0.408447265625,
0.66015625,
0.37744140625,
0.66015625
],
[
0.5947265625,
0.642578125,
0.61767578125,
0.642578125,
0.61767578125,
0.662109375,
0.5947265625,
0.662109375
],
[
0.59326171875,
0.65771484375,
0.61669921875,
0.65771484375,
0.61669921875,
0.6767578125,
0.59326171875,
0.6767578125
],
[
0.233642578125,
0.64599609375,
0.343017578125,
0.64453125,
0.343017578125,
0.6591796875,
0.233642578125,
0.66015625
],
[
0.23779296875,
0.6591796875,
0.34228515625,
0.6591796875,
0.34228515625,
0.673828125,
0.23779296875,
0.673828125
],
[
0.374267578125,
0.6591796875,
0.41015625,
0.6591796875,
0.41015625,
0.6748046875,
0.374267578125,
0.6748046875
],
[
0.491943359375,
0.6591796875,
0.5087890625,
0.6591796875,
0.5087890625,
0.68798828125,
0.491943359375,
0.68798828125
],
[
0.70166015625,
0.6591796875,
0.7197265625,
0.6591796875,
0.7197265625,
0.68798828125,
0.70166015625,
0.68798828125
],
[
0.375732421875,
0.671875,
0.41015625,
0.671875,
0.41015625,
0.6875,
0.375732421875,
0.6875
],
[
0.59716796875,
0.671875,
0.61376953125,
0.671875,
0.61376953125,
0.68603515625,
0.59716796875,
0.68603515625
],
[
0.259033203125,
0.7412109375,
0.4306640625,
0.7412109375,
0.4306640625,
0.7548828125,
0.259033203125,
0.7548828125
],
[
0.24267578125,
0.673828125,
0.334228515625,
0.673828125,
0.334228515625,
0.68505859375,
0.24267578125,
0.68505859375
],
[
0.2197265625,
0.7421875,
0.25244140625,
0.7421875,
0.25244140625,
0.75390625,
0.2197265625,
0.75390625
],
[
0.2181396484375,
0.76513671875,
0.78759765625,
0.76513671875,
0.78759765625,
0.77978515625,
0.2181396484375,
0.77978515625
],
[
0.21728515625,
0.77978515625,
0.78369140625,
0.77978515625,
0.78369140625,
0.79443359375,
0.21728515625,
0.79443359375
],
[
0.2181396484375,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.2181396484375,
0.81005859375
],
[
0.21728515625,
0.81005859375,
0.78759765625,
0.81103515625,
0.78759765625,
0.82568359375,
0.21728515625,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Sequential counterfactual policies using RL",
"5",
"Where P(st+i)-P(so) is the difference in target class probability between the",
"next state and the initial state and Q, β are parameters controlling the balance",
"between the difference of probability and distance respectively. This is beneficial",
"because the RL agent has immediate feedback on each step regarding the direc-",
"tion taken in relation to the target class. In contrast, when using the aforemen-",
"tioned binary reward (Rbin), the agent must explore the space without receiving",
"any positive feedback until a successful state is reached (found CF). This addi-",
"tional information on every step results in a more dense reward function, that",
"not only discourages large changes but also guides towards the target class. We",
"denote this reward function with Rprob since it incorporates the utilization of",
"targetprobabilities.",
"4",
"Experiments&Results",
"In this section, we compare the P-DQN model's performance when using the",
"binary reward Rbin and when employing the target class probabilities reward",
"Rprob",
"4.1",
"Datasets",
"Motivated by previous approaches we use four tabular datasets in our compara-",
"tive study. Specifically, the Adult Income, Credit Approval, German Credit, and",
"German Risk [9]. Each dataset has a binary output label and is comprised of",
"mixed features. Some datasets also have immutable features (e.g. Race) that are",
"considered when defining the states but are not taken into account in the action",
"space.",
"Table 1. Characteristics of the four public datasets used.",
"Dataset",
"N.instances",
"Features",
"N.of ContinuousN. of DiscreteN.of Immutable",
"Adult Income",
"48842",
"45",
"3",
"40",
"690",
"10",
"11",
"Credit Approval",
"German Credit",
"1000",
"53",
"42",
"1000",
"4",
"Evaluation measures",
"GermanRisk",
"4.2",
"Based on previous works on SCF generation [3,18,14], we use the metrics of",
"satisfiability, distance (i.e. proximity), and action sparsity to evaluate the quality",
"of the generated CF examples. Additionally, we measure the target probability",
"of the found CF and define a novel entropy metric to evaluate the diversity of",
"the actions taken by the policy. When an agent has learned to modify only a"
] |
{
"arxiv_id": "2311.00523",
"language": "en",
"page_idx": 4
}
|
||
2204.11122-en-0
|
2204.11122-en-0.jpg
|
Can domain adaptation make object recognition work for everyone?
Nikhil Naik$^{3}$
$\textbf{Viraj Prabhu}^{* , 1}$ $\textbf{Ramprasaath R. Selvaraju}^{\dagger , 2}$ $\textbf{Judy Hoffman}^{1}$
$\{$virajp, judy}@gatech.edu
ram@artera.ai
nnaik@salesforce.com
# Abstract
Despite the rapid progress in deep visual recognition, modern computer vision datasets significantly overrepresent the developed world and models trained on such datasets underperform on images from unseen geographies. We investigate the effectiveness of unsupervised domain adaptation $( UDA) \textit{of such models across geographies at closing this}$ performance gap. To do so, we first curate two shifts from existing datasets to study the Geographical DA problem, and discover new challenges beyond data distribution shift: context $shift, wherein$ object surroundings may change $signif-$ icantlyacross geographies, and subpopulation shift, wherein the intra-category distributions may shift. We demonstrate the inefficacy of standard DA methods at Geographical DA, highlighting the need for specialized geographical adaptation solutions to address the challenge of making object recognition work for everyone.
## $1. \textbf{Introduction}$
As deep-learning based computer vision systems gain widespread adoption, it is crucial that they perform equitably across diverse geographical deployments. However, prior work[1] has found that in practice modern computer vision datasets significantly overrepresent the developed world and models trained on such datasets systematically underperform on images from the rest of the world[2](see Fig. 1). Labeling images from target geographies is a natural solution but may be expensive and difficult to scale. Unsupervised domain adaptation[3-5](UDA) has extensively studied the problem of adapting models trained on a labeled source to an unlabeled target domain. However, UDA typically considers specific kinds of shifts in data generating distributions (e.g. synthetic to real data[6], or clipart to sketch images[7]), rather than distribution shifts across space and time in the real world. In this work, we investigate the effectiveness of UDA techniques at the practical application of adapting trained object recognition models to novel geographies.
$^*$Work done partially as intern at Salesforce Research.
$^{\dagger}$Work done at Salesforce Research.
Test (Rest of the world)
Train (North America)
$^{\prime\prime}$statue$^\prime\prime$
<FigureHere>
Figure 1. Modern computer vision datasets overrepresent the developed world[1]. This leads object recognition models trained on them (left) to underperform on images from novel geographies [2] (right-we show country of origin and model prediction above each image). In this work we investigate the effectiveness of domain adaptation [3] methods in bridging this performance gap.
Geographical domain adaptation presents two novel challenges beyond shifting data distributions: context shift and subpopulation shift. Context shift arises from a change in visual context for a given category across geographies $(e.g.$ predominantly indoor v/s outdoor basketball courts). Subpopulation shift arises from a change in within-category data distributions (e.g. for a‘toothbrush’category, the relative proportion of electric v/s mechanical varieties may change across geographies). In our experiments, we demonstrate the inefficacy of conventional adaptation strategies in addressing these additional challenges.
Some prior work has studied the problem of transferring deep visual models to new geographies. De Vries et al.[2] benchmark the drop in performance of publicly available vision API's on images from diverse geographies from the Dollar Street dataset[8], but do not propose a mitigation strategy. Recently, Dubey et al. [9] formulate this as a domain generalization problem and propose a solution that makes use of auxiliary target domain embeddings. We instead pose the problem as one of domain adaptation so as to leverage the full potential of unlabeled target data by allowing model updates on it. We make the following contributions:
1. To study the Geographical DA problem, we propose
|
[
[
0.148681640625,
0.1337890625,
0.826171875,
0.1337890625,
0.826171875,
0.1483154296875,
0.148681640625,
0.1483154296875
],
[
0.1168212890625,
0.1817626953125,
0.2459716796875,
0.18115234375,
0.2459716796875,
0.1962890625,
0.1168212890625,
0.1976318359375
],
[
0.283447265625,
0.179931640625,
0.53173828125,
0.177978515625,
0.53173828125,
0.1988525390625,
0.283447265625,
0.2008056640625
],
[
0.5703125,
0.18310546875,
0.69921875,
0.18310546875,
0.69921875,
0.1976318359375,
0.5703125,
0.1976318359375
],
[
0.74169921875,
0.1817626953125,
0.8486328125,
0.18115234375,
0.8486328125,
0.1962890625,
0.74169921875,
0.1976318359375
],
[
0.259033203125,
0.1976318359375,
0.37744140625,
0.199462890625,
0.37744140625,
0.2171630859375,
0.259033203125,
0.21533203125
],
[
0.4150390625,
0.199462890625,
0.50341796875,
0.199462890625,
0.50341796875,
0.2139892578125,
0.4150390625,
0.2139892578125
],
[
0.5400390625,
0.1962890625,
0.71337890625,
0.199462890625,
0.71337890625,
0.2171630859375,
0.5400390625,
0.2139892578125
],
[
0.178955078125,
0.2197265625,
0.39208984375,
0.2197265625,
0.39208984375,
0.2342529296875,
0.178955078125,
0.2342529296875
],
[
0.4306640625,
0.2215576171875,
0.5458984375,
0.2215576171875,
0.5458984375,
0.231689453125,
0.4306640625,
0.231689453125
],
[
0.57861328125,
0.2215576171875,
0.7548828125,
0.2215576171875,
0.7548828125,
0.231689453125,
0.57861328125,
0.231689453125
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.51708984375,
0.272705078125,
0.63818359375,
0.272705078125,
0.63818359375,
0.28271484375,
0.51708984375,
0.28271484375
],
[
0.75244140625,
0.272705078125,
0.8798828125,
0.272705078125,
0.8798828125,
0.28271484375,
0.75244140625,
0.28271484375
],
[
0.5390625,
0.2841796875,
0.61376953125,
0.2841796875,
0.61376953125,
0.294189453125,
0.5390625,
0.294189453125
],
[
0.1004638671875,
0.302978515625,
0.468017578125,
0.302978515625,
0.468017578125,
0.31640625,
0.1004638671875,
0.31640625
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.08087158203125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.6953125,
0.356689453125,
0.701171875,
0.356689453125,
0.701171875,
0.3642578125,
0.6953125,
0.3642578125
],
[
0.08087158203125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.6953125,
0.365478515625,
0.69921875,
0.365478515625,
0.69921875,
0.37060546875,
0.6953125,
0.37060546875
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.080078125,
0.392578125
],
[
0.07843017578125,
0.393310546875,
0.468017578125,
0.392578125,
0.468017578125,
0.4072265625,
0.07843017578125,
0.40771484375
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.080078125,
0.42236328125
],
[
0.080078125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.080078125,
0.4375
],
[
0.080078125,
0.439453125,
0.4697265625,
0.439453125,
0.4697265625,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.50244140625,
0.448974609375
],
[
0.080078125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.50341796875,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.4609375,
0.50341796875,
0.4609375
],
[
0.08087158203125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.50341796875,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.4765625,
0.50341796875,
0.4765625
],
[
0.50244140625,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.08087158203125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.50341796875,
0.492431640625,
0.890625,
0.492431640625,
0.890625,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.080078125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.5009765625,
0.50390625,
0.8408203125,
0.5048828125,
0.8408203125,
0.51953125,
0.5009765625,
0.51806640625
],
[
0.080078125,
0.5146484375,
0.4697265625,
0.5146484375,
0.4697265625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.080078125,
0.52978515625,
0.281982421875,
0.52978515625,
0.281982421875,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.52294921875,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.52294921875,
0.55029296875
],
[
0.50341796875,
0.5537109375,
0.890625,
0.5537109375,
0.890625,
0.564453125,
0.50341796875,
0.564453125
],
[
0.08087158203125,
0.56298828125,
0.20751953125,
0.56298828125,
0.20751953125,
0.57763671875,
0.08087158203125,
0.57763671875
],
[
0.50244140625,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.50341796875,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50341796875,
0.59765625
],
[
0.098876953125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.098876953125,
0.60400390625
],
[
0.5009765625,
0.5986328125,
0.890625,
0.59765625,
0.890625,
0.611328125,
0.5009765625,
0.6123046875
],
[
0.080078125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.080078125,
0.619140625
],
[
0.50341796875,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.080078125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50244140625,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.080078125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64306640625,
0.89208984375,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.6572265625,
0.89208984375,
0.671875,
0.50244140625,
0.67236328125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.08087158203125,
0.68310546875,
0.46728515625,
0.68310546875,
0.46728515625,
0.693359375,
0.08087158203125,
0.693359375
],
[
0.50244140625,
0.689453125,
0.68701171875,
0.689453125,
0.68701171875,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.07843017578125,
0.69580078125,
0.468017578125,
0.6943359375,
0.468017578125,
0.708984375,
0.07843017578125,
0.71044921875
],
[
0.521484375,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.521484375,
0.71728515625
],
[
0.080078125,
0.71044921875,
0.46728515625,
0.71044921875,
0.46728515625,
0.7236328125,
0.080078125,
0.7236328125
],
[
0.50244140625,
0.71826171875,
0.8896484375,
0.71826171875,
0.8896484375,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.080078125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.080078125,
0.74072265625
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.7412109375,
0.468017578125,
0.7412109375,
0.468017578125,
0.7548828125,
0.080078125,
0.7548828125
],
[
0.50341796875,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.080078125,
0.7568359375,
0.4697265625,
0.7568359375,
0.4697265625,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.50341796875,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.771484375,
0.4697265625,
0.771484375,
0.4697265625,
0.7861328125,
0.080078125,
0.7861328125
],
[
0.5009765625,
0.77978515625,
0.89208984375,
0.77880859375,
0.89208984375,
0.79345703125,
0.5009765625,
0.79443359375
],
[
0.080078125,
0.787109375,
0.468017578125,
0.787109375,
0.468017578125,
0.80029296875,
0.080078125,
0.80029296875
],
[
0.50341796875,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.080078125,
0.8017578125,
0.468017578125,
0.8017578125,
0.468017578125,
0.81494140625,
0.080078125,
0.81494140625
],
[
0.50341796875,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.080078125,
0.8173828125,
0.46728515625,
0.8173828125,
0.46728515625,
0.83056640625,
0.080078125,
0.83056640625
],
[
0.080078125,
0.83056640625,
0.468017578125,
0.83203125,
0.468017578125,
0.8466796875,
0.080078125,
0.84521484375
],
[
0.50244140625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.50341796875,
0.84033203125,
0.89208984375,
0.84033203125,
0.89208984375,
0.85400390625,
0.50341796875,
0.85400390625
],
[
0.080078125,
0.84765625,
0.44287109375,
0.84765625,
0.44287109375,
0.861328125,
0.080078125,
0.861328125
],
[
0.5009765625,
0.85595703125,
0.84326171875,
0.85595703125,
0.84326171875,
0.869140625,
0.5009765625,
0.869140625
],
[
0.098876953125,
0.87548828125,
0.38232421875,
0.87548828125,
0.38232421875,
0.8896484375,
0.098876953125,
0.8896484375
],
[
0.098876953125,
0.88623046875,
0.28759765625,
0.88720703125,
0.28759765625,
0.90087890625,
0.098876953125,
0.8994140625
],
[
0.5146484375,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5146484375,
0.90087890625
]
] |
[
"Can domain adaptation make object recognition work for evervone?",
"Viraj Prabhu*1",
"Ramprasaath R. Selvaraju t.2",
"Judy Hoffman'",
"Nikhil Naik3",
"1Georgia Tech",
"2Artera AI",
"3Salesforce Research",
"fvirajp,judyl@gatech.edu",
"ram@artera.ai",
"nnaik@salesforce.com",
"Abstract",
"Train(NorthAmerica)",
"Test (Rest oftheworld)",
"label =\"statue\"",
"Despite the rapid progress in deep visual recognition",
"modern computer vision datasets significantly overrepresent",
"the developed world and models trained on such datasets",
"underperform on images from unseen geographies. We inves",
"O",
"tigate the effectiveness of unsupervised domain adaptation",
"A",
"(UDA) of such models across geographies at closing this",
"performance gap. To do so, we first curate two shifts from",
"existing datasets to study the Geographical DA problem",
"and discover new challenges beyond data distribution shift",
"context shift, wherein object surroundings may change signif-",
"Figure 1. Modern computer vision datasets overrepresent the de",
"icantly across geographies, and subpopulation shift, wherein",
"velopedworld[l.Thisleadsobjectrecognitionmodelstrained on",
"the intra-category distributions may shift. We demonstrate",
"them (left) to underperform on images from novel geographies [2]",
"(right- we show country of origin and model prediction above each",
"the inefficacy of standard DA methods at Geographical DA,",
"image).Inthisworkweinvestigatetheeffectivenessof domain",
"highlighting the need for specialized geographical adap-",
"adaptation [3] methods in bridging this performance gap.",
"tation solutions to address the challenge of making object",
"recognition work for everyone.",
"Geographical domain adaptation presents two novel chal-",
"lengesbeyond shiftingdatadistributions:contextshiftand",
"1. Introduction",
"subpopulation shift. Context shift arises from a change in",
"visual context for a given category across geographies (e.g.",
"As deep-learning based computer vision systems gain",
"predominantly indoor v/s outdoorbasketball courts).Sub",
"widespread adoption, it is crucial that they perform equitably",
"population shift arises from a change in within-category data",
"across diverse geographical deployments.However, prior",
"distributions (e.g. for a ‘toothbrush' category, the relative",
"work [I] has found that in practice modern computer vision",
"proportion of electric v/s mechanical varieties may change",
"datasets significantly overrepresent the developed world and",
"across geographies). In our experiments, we demonstrate the",
"models trained on such datasets systematically underperform",
"inefficacy of conventional adaptation strategies in addressing",
"onimagesfromtherestoftheworld[2l(seeFig.1).La",
"these additional challenges.",
"beling images from target geographies is a natural solution",
"Some prior work has studied the problem of transferring",
"but may be expensive and difficult to scale. Unsupervised",
"deep visual models to new geographies. De Vries et al. [2]",
"domain adaptation [3-5] (UDA) has extensively studied the",
"benchmark the drop in performance of publicly available",
"problem of adapting models trained on a labeled source to an",
"vision API's on images from diverse geographies from the",
"unlabeled target domain. However, UDA typically considers",
"Dollar Street dataset [8], but do not propose a mitigation strat-",
"specific kinds of shifts in data generating distributions (e.g",
"egy. Recently, Dubey et al. [9] formulate this as a domain",
"synthetic to real data [6], or clipart to sketch images [7].",
"generalization problem and propose a solution that makes",
"rather than distribution shifts across space and time in the",
"use of auxiliary target domain embeddings. We instead pose",
"real world. In this work, we investigate the effectiveness",
"of UDA techniques at the practical application of adapting",
"the problem as one of domain adaptation so as to leverage",
"the full potential of unlabeled target data by allowing model",
"trained object recognition models to novel geographies.",
"updates on it. We make the following contributions:",
"* Work done partially as intern at Salesforce Research.",
"Work done at Salesforce Research.",
"1. To study the Geographical DA problem, we propose"
] |
{
"arxiv_id": "2204.11122",
"language": "en",
"page_idx": 0
}
|
||
2204.11122-en-3
|
2204.11122-en-3.jpg
|
<FigureHere>
Figure 4. Subpopulation shift for select categories across domains. We plot normalized cluster assignments per-domain as approximate subpopulation distributions-blue denotes within-category subpopulation distribution on source and orange denotes target. As seen, subpopulation distributions shift significantly across domains. On the right we visualize random images from some of the discovered clusters, and verify that they generally correspond to distinct subpopulations.
In Fig. 4, we show examples of subpopulation shift in the Dollar Street-DA and GeoYFCC-DA benchmarks. In the absence of subpopulation-level annotations, we use a simple strategy to obtain approximate annotations: we use a pretrained model (ResNet50[22] trained on ImageNet) to extract features for source and target images of a given category, and perform agglomerative clustering on the combined set of embeddings. We then use the inferred cluster assignments as subpopulation annotations. We also plot the normalized within-class distribution of cluster assignments on the source and target domains, and measure the Wasserstein distance between the two as a measure of the degree of subpopulation shift
As seen, this simple strategy discovers distinct clusters corresponding to semantically distinct subpopulations: $e.g.$ for “cleaning equipment” on Dollar Street-DA we discover separate clusters roughly corresponding to brooms, vaccuum cleaners, mops, and miscellaneous cleaning items. Crucially, we find that the intra-class distribution of many categories changes significantly across geographies $(e.g.$ brooms make up a significantly larger proportion of cleaning equipment in the target domain than in the source).
To quantitatively validate
Data and label distribution shift. As in conventional domain adaptation, geographical DA also presents data distribution shift$\left(\mathcal{P}_{\mathcal S}(\mathbf{x})\neq\mathcal{P}_{\mathcal{T}}(\mathbf{x})\right)$as object appearances change across geographies (see Fig. 1), and label distribution shift $(\mathcal{P}_{\mathcal S}(y)\neq\mathcal{P}_{\mathcal T}(y))$, as task label distributions change across domains (see Fig. 2).
In addition, geographical adaptation presents two new
challenges: context and subpopulation shift.
Context shift. We define context $c(\mathbf{x})$ for image x with label $y$ as the $task-irrelevant$ information in the image—this loosely corresponds to the background or surroundings of the object of interest. We define context shift as $\mathcal{P}_S(c(\mathbf{x})|y)\neq$ $\mathcal{P}_{\mathcal{T}}(c(\mathbf{x})|y)$,representing a change in object context across geographical domains.
In Fig. 3 we show qualitative examples of context shift within our proposed Dollar Street-DA and GeoYFCC-DA shifts for a few categories. For example, we find that in Dollar Street-DA, most“toothbrush”images in the source domain tend to be photographed inside bathrooms, whereas the surroundings in the target domains are significantly more diverse $(e.g.$ walls and roofs). We see similar trends in the GeoYFCC-DA shift (e.g. indoor v/s outdoor basketball games). As deep neural networks are known to often employ “shortcut learning"[23] of potentially spurious features (e.g. object backgrounds) to make predictions, we hypothesize (and experimentally verify in Sec. 4.4) that such a context shift will present a challenge to visual recognition models deployed in new geographies.
Subpopulation shift. We define subpopulation shift as $\mathcal{P} _{\mathcal{S} }( \mathbf{x} | y)$ $\neq$ $\mathcal{P} _{\mathcal{T} }( \mathbf{x} | y) $, representing a change in withincategory distribution across domains.
<FigureHere>
the subpopulation shift, we compute the degree of perclass subpopulation shift: for each class, we compute the normalized subpopulation distribution per-domain (as visualized in Fig. $\color{red}{4,\mathrm{left}),\mathrm{measure}}$ the cross-domain Wasserstein
Figure 5. Verifying sub-
population shift.
|
[
[
0.459228515625,
0.09661865234375,
0.5009765625,
0.09661865234375,
0.5009765625,
0.10418701171875,
0.459228515625,
0.10418701171875
],
[
0.52783203125,
0.09661865234375,
0.57177734375,
0.09661865234375,
0.57177734375,
0.10418701171875,
0.52783203125,
0.10418701171875
],
[
0.60107421875,
0.0985107421875,
0.640625,
0.0985107421875,
0.640625,
0.1029052734375,
0.60107421875,
0.1029052734375
],
[
0.67333984375,
0.0985107421875,
0.708984375,
0.0985107421875,
0.708984375,
0.1029052734375,
0.67333984375,
0.1029052734375
],
[
0.293212890625,
0.11297607421875,
0.3857421875,
0.11297607421875,
0.3857421875,
0.12371826171875,
0.293212890625,
0.12371826171875
],
[
0.289306640625,
0.1029052734375,
0.390625,
0.1029052734375,
0.390625,
0.11297607421875,
0.289306640625,
0.11297607421875
],
[
0.334228515625,
0.1337890625,
0.369384765625,
0.1337890625,
0.369384765625,
0.141357421875,
0.334228515625,
0.141357421875
],
[
0.2467041015625,
0.1427001953125,
0.261474609375,
0.1427001953125,
0.261474609375,
0.175537109375,
0.2467041015625,
0.175537109375
],
[
0.338134765625,
0.14453125,
0.365234375,
0.14453125,
0.365234375,
0.1514892578125,
0.338134765625,
0.1514892578125
],
[
0.314453125,
0.2020263671875,
0.36767578125,
0.2020263671875,
0.36767578125,
0.212158203125,
0.314453125,
0.212158203125
],
[
0.293212890625,
0.2197265625,
0.386474609375,
0.2197265625,
0.386474609375,
0.2298583984375,
0.293212890625,
0.2298583984375
],
[
0.298095703125,
0.2298583984375,
0.378173828125,
0.2298583984375,
0.378173828125,
0.240478515625,
0.298095703125,
0.240478515625
],
[
0.38232421875,
0.251953125,
0.406982421875,
0.251953125,
0.406982421875,
0.255615234375,
0.38232421875,
0.255615234375
],
[
0.378173828125,
0.2587890625,
0.40185546875,
0.2587890625,
0.40185546875,
0.266357421875,
0.378173828125,
0.266357421875
],
[
0.08087158203125,
0.34521484375,
0.89208984375,
0.34521484375,
0.89208984375,
0.35986328125,
0.08087158203125,
0.35986328125
],
[
0.08087158203125,
0.3623046875,
0.890625,
0.3623046875,
0.890625,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.080078125,
0.375,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.080078125,
0.38818359375
],
[
0.080078125,
0.38818359375,
0.48046875,
0.38818359375,
0.48046875,
0.401611328125,
0.080078125,
0.401611328125
],
[
0.08087158203125,
0.4287109375,
0.4697265625,
0.4287109375,
0.4697265625,
0.442626953125,
0.08087158203125,
0.442626953125
],
[
0.521484375,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.442626953125,
0.521484375,
0.442626953125
],
[
0.08087158203125,
0.4443359375,
0.4697265625,
0.4443359375,
0.4697265625,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.50244140625,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50244140625,
0.458251953125
],
[
0.080078125,
0.460205078125,
0.4697265625,
0.460205078125,
0.4697265625,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.5009765625,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.07843017578125,
0.474853515625,
0.4697265625,
0.473388671875,
0.4697265625,
0.488037109375,
0.07843017578125,
0.4892578125
],
[
0.50244140625,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.4892578125,
0.50244140625,
0.4892578125
],
[
0.08087158203125,
0.490478515625,
0.4697265625,
0.490478515625,
0.4697265625,
0.50390625,
0.08087158203125,
0.50390625
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.080078125,
0.505859375,
0.221435546875,
0.505859375,
0.221435546875,
0.51953125,
0.080078125,
0.51953125
],
[
0.50244140625,
0.505859375,
0.8935546875,
0.505859375,
0.8935546875,
0.51953125,
0.50244140625,
0.51953125
],
[
0.096435546875,
0.52099609375,
0.468017578125,
0.51953125,
0.468017578125,
0.5341796875,
0.096435546875,
0.53515625
],
[
0.50244140625,
0.52099609375,
0.8935546875,
0.52099609375,
0.8935546875,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.080078125,
0.5361328125,
0.370849609375,
0.5361328125,
0.370849609375,
0.5498046875,
0.080078125,
0.5498046875
],
[
0.50244140625,
0.53515625,
0.89208984375,
0.53515625,
0.89208984375,
0.548828125,
0.50244140625,
0.548828125
],
[
0.080078125,
0.5517578125,
0.468017578125,
0.55029296875,
0.468017578125,
0.56494140625,
0.080078125,
0.56640625
],
[
0.50244140625,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50244140625,
0.564453125
],
[
0.08087158203125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.580078125,
0.08087158203125,
0.580078125
],
[
0.50244140625,
0.56494140625,
0.89208984375,
0.56494140625,
0.89208984375,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.08251953125,
0.583984375,
0.46728515625,
0.583984375,
0.46728515625,
0.5947265625,
0.08251953125,
0.5947265625
],
[
0.50244140625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.08087158203125,
0.59765625,
0.468017578125,
0.59765625,
0.468017578125,
0.611328125,
0.08087158203125,
0.611328125
],
[
0.5009765625,
0.59521484375,
0.89208984375,
0.5947265625,
0.89208984375,
0.609375,
0.5009765625,
0.60986328125
],
[
0.08087158203125,
0.61279296875,
0.4697265625,
0.61279296875,
0.4697265625,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.5009765625,
0.60986328125,
0.6337890625,
0.609375,
0.6337890625,
0.62353515625,
0.5009765625,
0.625
],
[
0.080078125,
0.62890625,
0.2303466796875,
0.62744140625,
0.2303466796875,
0.6416015625,
0.080078125,
0.642578125
],
[
0.521484375,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.6416015625,
0.521484375,
0.6416015625
],
[
0.09722900390625,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.09722900390625,
0.65771484375
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64306640625,
0.89208984375,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.07843017578125,
0.66015625,
0.4697265625,
0.65771484375,
0.4697265625,
0.671875,
0.07843017578125,
0.673828125
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.08087158203125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.50244140625,
0.68798828125,
0.8935546875,
0.68798828125,
0.8935546875,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.07843017578125,
0.7041015625,
0.4697265625,
0.70263671875,
0.4697265625,
0.71728515625,
0.07843017578125,
0.71826171875
],
[
0.5009765625,
0.70263671875,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.5009765625,
0.71630859375
],
[
0.080078125,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.73095703125,
0.5048828125,
0.73095703125
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.74853515625,
0.46728515625,
0.74853515625,
0.46728515625,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.50244140625,
0.74853515625,
0.74853515625,
0.74853515625,
0.74853515625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.080078125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.51953125,
0.7646484375,
0.701171875,
0.76513671875,
0.701171875,
0.77978515625,
0.51953125,
0.77880859375
],
[
0.7685546875,
0.7666015625,
0.8857421875,
0.7666015625,
0.8857421875,
0.7734375,
0.7685546875,
0.7734375
],
[
0.74169921875,
0.77587890625,
0.76708984375,
0.77783203125,
0.765625,
0.7890625,
0.740234375,
0.787109375
],
[
0.81787109375,
0.77783203125,
0.869140625,
0.77783203125,
0.869140625,
0.78857421875,
0.81787109375,
0.78857421875
],
[
0.5009765625,
0.78076171875,
0.701171875,
0.77978515625,
0.701171875,
0.79443359375,
0.5009765625,
0.79541015625
],
[
0.07843017578125,
0.77978515625,
0.4697265625,
0.77978515625,
0.4697265625,
0.79443359375,
0.07843017578125,
0.79443359375
],
[
0.74169921875,
0.787109375,
0.76708984375,
0.787109375,
0.76708984375,
0.79541015625,
0.74169921875,
0.79541015625
],
[
0.8408203125,
0.78857421875,
0.869140625,
0.78857421875,
0.869140625,
0.79248046875,
0.8408203125,
0.79248046875
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.08087158203125,
0.8095703125
],
[
0.50244140625,
0.796875,
0.70166015625,
0.796875,
0.70166015625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.08087158203125,
0.82373046875
],
[
0.50244140625,
0.81103515625,
0.701171875,
0.81103515625,
0.701171875,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.701171875,
0.8271484375,
0.701171875,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.27685546875,
0.84130859375,
0.27685546875,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.70166015625,
0.84130859375,
0.70166015625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.78125,
0.84326171875,
0.81787109375,
0.84326171875,
0.81787109375,
0.8466796875,
0.78125,
0.8466796875
],
[
0.8408203125,
0.84228515625,
0.88134765625,
0.84228515625,
0.88134765625,
0.8466796875,
0.8408203125,
0.8466796875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.70166015625,
0.85693359375,
0.70166015625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.734375,
0.859375,
0.8935546875,
0.859375,
0.8935546875,
0.87255859375,
0.734375,
0.87255859375
],
[
0.080078125,
0.8720703125,
0.47216796875,
0.87060546875,
0.47216796875,
0.88525390625,
0.080078125,
0.88623046875
],
[
0.50341796875,
0.8720703125,
0.70166015625,
0.8720703125,
0.70166015625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.734375,
0.875,
0.8310546875,
0.875,
0.8310546875,
0.88623046875,
0.734375,
0.88623046875
],
[
0.080078125,
0.88720703125,
0.328369140625,
0.88720703125,
0.328369140625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.70166015625,
0.88720703125,
0.70166015625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"clustero",
"cluster 1",
"cluster",
"custer",
"(Dollar Street-DA)",
"cleaning equipment",
"Source",
"paurisse",
"Targe",
"Cluster label",
"vessel/watercraft",
"(GeoYFCC-DA)",
"aource",
"Targe",
"Figure 4. Subpopulation shift for select categories across domains. We plot normalized cluster assignments per-domain as approximate",
"subpopulation distributions-blue denotes within-category subpopulation distribution on sourceand orange denotes target.As seen.",
"subpopulation distributions shift significantly across domains.On theright we visualize random images from some of the discovered clusters.",
"and verify that they generally correspond to distinct subpopulations.",
"Data and label distribution shift. As in conventional do-",
"In Fig. 4, we show examples of subpopulation shift in",
"main adaptation, geographical DA also presents data distri-",
"the Dollar Street-DA and GeoYFCC-DA benchmarks. In",
"bution shift (Ps(x) P(x)) as object appearances change",
"the absence of subpopulation-level annotations, we use a",
"across geographies (see Fig. 1), and label distribution shift",
"simple strategy to obtain approximate annotations: we use",
"(Ps(y) ± Pr(y)), as task label distributions change across",
"a pretrained model (ResNet50 [22] trained on ImageNet)",
"domains (see Fig. 2).",
"to extract features for source and target images of a given",
"In addition, geographical adaptation presents two new",
"category, and perform agglomerative clustering on the com-",
"challenges: context and subpopulation shift.",
"bined set of embeddings. We then use the inferred cluster",
"Context shift. We define context c(x) for image x with",
"assignments as subpopulation annotations. We also plot the",
"labelyasthetask-irrelevantinformationintheimage--this",
"normalized within-class distribution of cluster assignments",
"looselycorrespondstothebackgroundorsurroundingsofthe",
"on the source and target domains, and measure the Wasser-",
"object of interest. We define context shift as Ps(c(x)ly)",
"stein distance between the two as a measure of the degree of",
"P(c(x)ly), representing a change in object context across",
"subpopulation shift.",
"geographical domains.",
"As seen, this simple strategy discovers distinct clusters",
"In Fig. 3 we show qualitative examples of context shift",
"corresponding to semantically distinct subpopulations: e.g.",
"withinourproposedDollarStreet-DAandGeoYFCC-DA",
"for \"cleaning equipment\" on Dollar Street-DA we discover",
"shifts for a few categories. For example, we find that in",
"separate clusters roughly corresponding to brooms, vaccuum",
"Dollar Street-DA, most “toothbrush\"images in the source",
"cleaners, mops, and miscellaneous cleaning items. Crucially,",
"domain tend to be photographed inside bathrooms, whereas",
"we find that the intra-class distribution of many categories",
"the surroundings in the target domains are significantly more",
"changessignificantlyacrossgeographies(e.g.broomsmake",
"diverse (e.g. walls and roofs). We see similar trends in",
"up a significantly larger proportion of cleaning equipment in",
"the GeoYFCC-DA shift (e.g. indoor v/s outdoor basketball",
"the target domain than in the source).",
"games). As deep neural networks are known to often employ",
"To quantitatively validate",
"Verifying subpopulation shift",
"0.175",
"Random",
"the subpopulation shift, we",
"\"shortcut learning\" [23] of potentially spurious features (e.g.",
"0.150",
"iengran",
"object backgrounds) to make predictions, we hypothesize",
"compute the degree of per-",
"(and experimentally verify in Sec. 4.4) that such a context",
"class subpopulation shift: for",
"shift will present a challenge to visual recognition models",
"each class, we compute the",
"deployed in new geographies.",
"normalized subpopulation dis-",
"arstreet-",
"GeYFcc-D",
"Subpopulation shift. We define subpopulation shift as",
"tribution per-domain (as visu-",
"Figure 5. Verifying sub-",
"Ps(xly) ≠ P(xy), representing a change in within-",
"alized in Fig. 4, left), measure",
"population shift.",
"category distribution across domains.",
"the cross-domain Wasserstein"
] |
{
"arxiv_id": "2204.11122",
"language": "en",
"page_idx": 3
}
|
||
2204.11122-en-4
|
2204.11122-en-4.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Method}$</th>
<th>$\mathbf{D}_{0}\mathbf{l}\mathbf{ar}\mathbf{S}\mathbf{t}\mathbf{r}\mathbf{e}\mathbf{t}\mathbf{-}\mathbf{D}\mathbf{A}$</th>
<th>$\mathbf{GeoYFCC-DA}$</th>
</tr>
<tr>
<td>SOurce</td>
<td>$54.66\pm0.62$</td>
<td>42.88</td>
</tr>
<tr>
<td>$\operatorname{target}$oracle*</td>
<td>$67.73\pm0.30$</td>
<td>56.78</td>
</tr>
<tr>
<td>$\mathbf{MMD}$ 10</td>
<td>$55.77\pm0.75$</td>
<td>43.53</td>
</tr>
<tr>
<td>$\mathbf{DANN}$ [4]</td>
<td>$54.80_{\pm0.38}$</td>
<td>42.64</td>
</tr>
<tr>
<td>SENTRY [5]</td>
<td>$55.73\pm0.34$</td>
<td>42.58</td>
</tr>
<tr>
<td>$SST$</td>
<td>$58.71\pm0.53$</td>
<td>45.22</td>
</tr>
</tbody>
</table>
Table 1. Average accuracy on the target test set (20%) for the Dollar Street-DA and GeoYFCC-DA shifts. * denotes that the target oracle was trained on target data non-overlapping with the test set (80%) whereas DA methods were adapted without labels on the entire target dataset.
distance, and average across classes. Fig. 5 presents results for this measure for our proposed geographical shifts versus randomly constructed source and target domains of the same size, for the Dollar Street and GeoYFCC datasets. As seen, geographical shifts lead to significantly higher subpopulation shift.
## $4. \textbf{Experiments}$
# 4.1. Setup
To account for label imbalance, we report per-class average accuracy as our metric. As described in Sec. 3, we consider adaptation on two shifts:
Dollar Street-DA: We consider images from North America and Europe as the source domain (2.9k) and images from Asia, Africa, and South America as the target domain (8.8k). GeoYFCC-DA: We consider images from North America as the source domain (24.1k) and images from Asia, Australia, and South America as the target domain (59.4k).
On both shifts we create a 90%-10% train-test split on the source domain and report transfer performance on 20% heldout target data, but use the entire target dataset for unsupervised adaptation. On Dollar Street-DA, due to the relatively small size of the dataset, we report performance mean and standard deviation over three experimental runs.
# 4.2. Domain Adaptation Baselines
We benchmark 4 representative DA methods from the literature on our Geographical DA shifts: one domain discrepancy-based method, one domain adversarial method, and two self-training based methods.
$\mathbf{1} ) \textbf{MMD[ 10] : Aligns domains by computing mean source}$ and target embeddings in a reproducing kernel hilbert space and minimizing their distance as a maximum mean discrepancy measure.
2) DANN [4]: Domain-adversarial neural networks adversarially learn a domain discriminator to distinguish source
and target features against a feature encoder that is trained
to fool the discriminator.
3) SENTRY [5]: SENTRY measures model predictive consistency across randomly augmented versions of each target image and selectively increases predictive entropy on highly consistent instances, while decreasing it on highly inconsistent ones. SENTRY also uses pseudolabel-based approximating class balancing on the target domain, and employs a slightly modified ResNet50 architecture[24].
4) Selective Self-training (SST): We implement a simplified self-training baseline that self-trains against predicted labels on target instances on which the model is atleast 90% confident, and also employs pseudolabel-based target class balancing and a modified ResNet50 architecture[24].
Finally, we also report performance for a target oracle that is trained in a supervised fashion on the target domain. The target oracle is meant to represent a performance upper bound in the absence of domain shift.
Implementation details. We use a ResNet50[22] as our CNN architecture, and for SENTRY and SST use a modified few-shot variant[24] that replaces the last linear layer with a K-way (where K is the number of classes) fully-connected layer without bias. Input activations to this layer are $L_2$- normalized and its output is passed into a softmax layer with a temperature of 0.05. For optimization, we use Adam [25] with a learning rate of 0.001 and weight decay of $5e^{-4}.$ We use a batch size of 128 across experiments. All DA methods are applied to a source model that is first trained using supervised training on the source domain for 50 epochs We employ 100 epochs of adaptation on Dollar Street-DA and 40 epochs on GeoYFCC-DA. All methods additionally minimize a supervised cross-entropy loss on source labels during adaptation (with a loss weight of 1.0 and 0.1 on Dollar Street-DA and GeoYFCC-DA respectively). To combat label distribution shift, we follow Tan et al.[26] and use classbalanced sampling on the source domain across experiments.
### $4. 3. \textbf{Results}$
Table 1 presents average accuracy on the target test set for the Dollar Street-DA and GeoYFCC-DA shifts. We observe: $\triangleright \textbf{Geographical shifts lead to significant performance}$ drops (Row 1 v/s 2). As seen, the target oracle achieves 67.7% and 56.78% whereas the source model achieves a significantly lower performance of 54.66%(-13.1%) and $42.88\%(-13.9\%).$ Clearly, geographical variations pose a significant challenge to transfer.
$\triangleright \textbf{DA methods offer limited improvements( Rows 3- 6) . }$ All the benchmarked methods achieve limited success at geographical adaptation, sometimes performing no better than the source model. Surprisingly, we find the simple SST method to achieve the best (albeit small) improvement over the source model $(+4.1\%/+2.3\%)$, but still well short of the target oracle (-9%/-11.6%). Altogether, these results
|
[
[
0.09234619140625,
0.09344482421875,
0.156005859375,
0.0953369140625,
0.1544189453125,
0.111083984375,
0.0906982421875,
0.10919189453125
],
[
0.2059326171875,
0.09661865234375,
0.33251953125,
0.09661865234375,
0.33251953125,
0.10986328125,
0.2059326171875,
0.10986328125
],
[
0.3505859375,
0.09661865234375,
0.459228515625,
0.09661865234375,
0.459228515625,
0.10986328125,
0.3505859375,
0.10986328125
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.50244140625,
0.10797119140625,
0.66650390625,
0.10797119140625,
0.66650390625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.233642578125,
0.11871337890625,
0.30810546875,
0.11871337890625,
0.30810546875,
0.1331787109375,
0.233642578125,
0.1331787109375
],
[
0.38232421875,
0.11871337890625,
0.426513671875,
0.11871337890625,
0.426513671875,
0.1331787109375,
0.38232421875,
0.1331787109375
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.2303466796875,
0.1331787109375,
0.3095703125,
0.1351318359375,
0.3095703125,
0.149658203125,
0.2303466796875,
0.147705078125
],
[
0.09234619140625,
0.11932373046875,
0.143798828125,
0.11932373046875,
0.143798828125,
0.1337890625,
0.09234619140625,
0.1337890625
],
[
0.09234619140625,
0.1363525390625,
0.1895751953125,
0.1331787109375,
0.1895751953125,
0.147705078125,
0.09234619140625,
0.15087890625
],
[
0.38232421875,
0.1337890625,
0.426513671875,
0.1337890625,
0.426513671875,
0.149658203125,
0.38232421875,
0.149658203125
],
[
0.50244140625,
0.1395263671875,
0.8935546875,
0.1395263671875,
0.8935546875,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.0906982421875,
0.154052734375,
0.17236328125,
0.1558837890625,
0.17236328125,
0.1717529296875,
0.0906982421875,
0.169189453125
],
[
0.38232421875,
0.1558837890625,
0.426513671875,
0.1558837890625,
0.426513671875,
0.1717529296875,
0.38232421875,
0.1717529296875
],
[
0.50244140625,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.50341796875,
0.17041015625,
0.8896484375,
0.17041015625,
0.8896484375,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.233642578125,
0.1572265625,
0.30810546875,
0.1572265625,
0.30810546875,
0.1717529296875,
0.233642578125,
0.1717529296875
],
[
0.0931396484375,
0.17236328125,
0.1707763671875,
0.17236328125,
0.1707763671875,
0.1875,
0.0931396484375,
0.1875
],
[
0.233642578125,
0.17236328125,
0.30810546875,
0.17236328125,
0.30810546875,
0.1875,
0.233642578125,
0.1875
],
[
0.381591796875,
0.17236328125,
0.426513671875,
0.17236328125,
0.426513671875,
0.1875,
0.381591796875,
0.1875
],
[
0.2320556640625,
0.1875,
0.306396484375,
0.1875,
0.306396484375,
0.2020263671875,
0.2320556640625,
0.2020263671875
],
[
0.5009765625,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.5009765625,
0.1962890625
],
[
0.38232421875,
0.1881103515625,
0.42724609375,
0.1881103515625,
0.42724609375,
0.20263671875,
0.38232421875,
0.20263671875
],
[
0.0931396484375,
0.1881103515625,
0.1854248046875,
0.1881103515625,
0.1854248046875,
0.20263671875,
0.0931396484375,
0.20263671875
],
[
0.38232421875,
0.20263671875,
0.426513671875,
0.20263671875,
0.426513671875,
0.2183837890625,
0.38232421875,
0.2183837890625
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.0931396484375,
0.2052001953125,
0.12744140625,
0.2052001953125,
0.12744140625,
0.2171630859375,
0.0931396484375,
0.2171630859375
],
[
0.233642578125,
0.203857421875,
0.30810546875,
0.203857421875,
0.30810546875,
0.2183837890625,
0.233642578125,
0.2183837890625
],
[
0.50244140625,
0.2139892578125,
0.80126953125,
0.2139892578125,
0.80126953125,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.50244140625,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.080078125,
0.236083984375,
0.468017578125,
0.236083984375,
0.468017578125,
0.25,
0.080078125,
0.25
],
[
0.50244140625,
0.2442626953125,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.50244140625,
0.25830078125
],
[
0.080078125,
0.250732421875,
0.468017578125,
0.250732421875,
0.468017578125,
0.264404296875,
0.080078125,
0.264404296875
],
[
0.080078125,
0.26513671875,
0.471435546875,
0.26513671875,
0.471435546875,
0.279052734375,
0.080078125,
0.279052734375
],
[
0.50244140625,
0.260009765625,
0.89208984375,
0.260009765625,
0.89208984375,
0.2734375,
0.50244140625,
0.2734375
],
[
0.080078125,
0.279052734375,
0.4697265625,
0.279052734375,
0.4697265625,
0.292236328125,
0.080078125,
0.292236328125
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.080078125,
0.29345703125,
0.160888671875,
0.29345703125,
0.160888671875,
0.304931640625,
0.080078125,
0.304931640625
],
[
0.5009765625,
0.2890625,
0.85107421875,
0.2890625,
0.85107421875,
0.302978515625,
0.5009765625,
0.302978515625
],
[
0.521484375,
0.30615234375,
0.89208984375,
0.30615234375,
0.89208984375,
0.3193359375,
0.521484375,
0.3193359375
],
[
0.080078125,
0.32568359375,
0.468017578125,
0.32568359375,
0.468017578125,
0.339111328125,
0.080078125,
0.339111328125
],
[
0.5009765625,
0.32080078125,
0.89208984375,
0.32080078125,
0.89208984375,
0.333984375,
0.5009765625,
0.333984375
],
[
0.5009765625,
0.335205078125,
0.8935546875,
0.3359375,
0.8935546875,
0.35107421875,
0.5009765625,
0.349853515625
],
[
0.080078125,
0.341552734375,
0.4697265625,
0.341552734375,
0.4697265625,
0.354736328125,
0.080078125,
0.354736328125
],
[
0.080078125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.50341796875,
0.352783203125,
0.74853515625,
0.352783203125,
0.74853515625,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.080078125,
0.37060546875,
0.4697265625,
0.37060546875,
0.4697265625,
0.3837890625,
0.080078125,
0.3837890625
],
[
0.50244140625,
0.3662109375,
0.89208984375,
0.3662109375,
0.89208984375,
0.380126953125,
0.50244140625,
0.380126953125
],
[
0.080078125,
0.386474609375,
0.468017578125,
0.386474609375,
0.468017578125,
0.40087890625,
0.080078125,
0.40087890625
],
[
0.50244140625,
0.3818359375,
0.89208984375,
0.3818359375,
0.89208984375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.07843017578125,
0.401611328125,
0.1168212890625,
0.401611328125,
0.1168212890625,
0.415283203125,
0.07843017578125,
0.415283203125
],
[
0.5009765625,
0.395751953125,
0.89208984375,
0.396484375,
0.89208984375,
0.410888671875,
0.5009765625,
0.410400390625
],
[
0.50341796875,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.50341796875,
0.425537109375
],
[
0.080078125,
0.431884765625,
0.2091064453125,
0.431884765625,
0.2091064453125,
0.4462890625,
0.080078125,
0.4462890625
],
[
0.50341796875,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.439453125,
0.50341796875,
0.439453125
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.07843017578125,
0.455810546875,
0.158447265625,
0.458251953125,
0.158447265625,
0.473388671875,
0.07843017578125,
0.4716796875
],
[
0.50341796875,
0.458984375,
0.8896484375,
0.458984375,
0.8896484375,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.50244140625,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50244140625,
0.486083984375
],
[
0.098876953125,
0.481689453125,
0.4697265625,
0.481689453125,
0.4697265625,
0.49560546875,
0.098876953125,
0.49560546875
],
[
0.50244140625,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.0767822265625,
0.49755859375,
0.468017578125,
0.496826171875,
0.468017578125,
0.51123046875,
0.0767822265625,
0.51220703125
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.07843017578125,
0.51318359375,
0.30224609375,
0.51318359375,
0.30224609375,
0.5263671875,
0.07843017578125,
0.5263671875
],
[
0.50244140625,
0.51806640625,
0.89208984375,
0.51806640625,
0.89208984375,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.080078125,
0.52880859375,
0.4697265625,
0.52880859375,
0.4697265625,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.50244140625,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.08087158203125,
0.54541015625,
0.46728515625,
0.54541015625,
0.46728515625,
0.55615234375,
0.08087158203125,
0.55615234375
],
[
0.50244140625,
0.548828125,
0.890625,
0.548828125,
0.890625,
0.5625,
0.50244140625,
0.5625
],
[
0.080078125,
0.5595703125,
0.4697265625,
0.5595703125,
0.4697265625,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.08087158203125,
0.57470703125,
0.4697265625,
0.57470703125,
0.4697265625,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.50341796875,
0.57763671875,
0.89208984375,
0.57763671875,
0.89208984375,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.08087158203125,
0.59228515625,
0.468017578125,
0.59228515625,
0.468017578125,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.5009765625,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.5009765625,
0.60595703125
],
[
0.07843017578125,
0.60595703125,
0.398681640625,
0.60595703125,
0.398681640625,
0.619140625,
0.07843017578125,
0.619140625
],
[
0.50244140625,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.09722900390625,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.09722900390625,
0.63525390625
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.07843017578125,
0.63623046875,
0.468017578125,
0.63525390625,
0.468017578125,
0.6494140625,
0.07843017578125,
0.65087890625
],
[
0.08087158203125,
0.65380859375,
0.46728515625,
0.65380859375,
0.46728515625,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.5048828125,
0.65087890625,
0.59326171875,
0.65087890625,
0.59326171875,
0.662109375,
0.5048828125,
0.662109375
],
[
0.080078125,
0.66748046875,
0.468017578125,
0.66748046875,
0.468017578125,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.51953125,
0.671875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.51953125,
0.68603515625
],
[
0.080078125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.080078125,
0.69580078125
],
[
0.5009765625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.07843017578125,
0.69775390625,
0.461669921875,
0.69775390625,
0.461669921875,
0.7109375,
0.07843017578125,
0.7109375
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.72265625,
0.34228515625,
0.7236328125,
0.34228515625,
0.73779296875,
0.080078125,
0.7373046875
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.50244140625,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.1004638671875,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.1004638671875,
0.76220703125
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.080078125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.080078125,
0.77783203125
],
[
0.5009765625,
0.7646484375,
0.89208984375,
0.76513671875,
0.89208984375,
0.77978515625,
0.5009765625,
0.77880859375
],
[
0.080078125,
0.77880859375,
0.4697265625,
0.77880859375,
0.4697265625,
0.79248046875,
0.080078125,
0.79248046875
],
[
0.5009765625,
0.77978515625,
0.7138671875,
0.77978515625,
0.7138671875,
0.79345703125,
0.5009765625,
0.79345703125
],
[
0.080078125,
0.79443359375,
0.32275390625,
0.79443359375,
0.32275390625,
0.80810546875,
0.080078125,
0.80810546875
],
[
0.5009765625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.080078125,
0.81005859375,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.080078125,
0.82373046875
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.07843017578125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.8388671875,
0.07843017578125,
0.8388671875
],
[
0.50341796875,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.1748046875,
0.85791015625,
0.1748046875,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.47216796875,
0.8720703125,
0.47216796875,
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,
0.4697265625,
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
]
] |
[
"Method",
"Dollar Street-DA",
"GeoYFCC-DA",
"and target features against a feature encoder that is trained",
"to fool the discriminator.",
"54.66±0.62",
"42.88",
"3) SENTRY [5]: SENTRY measures model predictive con-",
"67.73± 0.30",
"source",
"target oracle*",
"56.78",
"sistency across randomly augmented versions of each target",
"MMD [10]",
"43.53",
"image and selectively increases predictive entropy on highly",
"consistent instances,while decreasing it on highlyinconsis.",
"55.77±0.75",
"DANN [4]",
"54.80±0.38",
"42.64",
"55.73±0.34",
"tent ones. SENTRY also uses pseudolabel-based approxi",
"42.58",
"SENTRY [5]",
"45.22",
"mating class balancing on the target domain, and employs a",
"SST",
"58.71±0.53",
"slightly modified ResNet50 architecture [24].",
"4) Selective Self-training (SST): We implement a simpli-",
"Table 1. Average accuracy on the target test set (20%) for the Dollar",
"fied self-training baseline that self-trains against predicted",
"Street-DA and GeoYFCC-DA shifts. * denotes that the target oracle",
"was trained on target data non-overlapping with the test set (80%)",
"labels on target instances on which the model is atleast 90%",
"whereas DA methods were adapted without labels on the entire",
"confident, and also employs pseudolabel-based target class",
"target dataset",
"balancing and a modified ResNet50 architecture [24]",
"Finally, we also report performance for a target oracle",
"distance, and average across classes. Fig. 5 presents results",
"that is trained in a supervised fashion on the target domain.",
"The target oracle is meant to represent a performance upper",
"for this measure for our proposed geographical shifts versus",
"randomly constructed source and target domains of the same",
"bound intheabsence of domain shift.",
"size, for the Dollar Street and GeoYFCC datasets. As seen,",
"Implementation details. We use a ResNet50 [22] as our",
"geographical shifts lead to significantly higher subpopulation",
"CNN architecture, and for SENTRY and SST use a modified",
"shift.",
"few-shot variant [24] that replaces the last linear layer with",
"a K-way (where K is the number of classes) fully-connected",
"4. Experiments",
"layer without bias.Input activations to this layer are L2",
"normalized and its output is passed into a softmax layer with",
"4.1. Setup",
"a temperature of 0.05.For optimization, we use Adam [25]",
"with a learning rate of 0.001 and weight decay of 5e-4",
"To account for label imbalance, we report per-class av-",
"We use a batch size of 128 across experiments. All DA",
"erage accuracy as our metric. As described in Sec. 3, we",
"methods are applied to a source model that is first trained",
"consider adaptation on two shifts:",
"using supervised training on the source domain for 50 epochs.",
"Dollar Street-DA: We consider images from North America",
"We employ 100 epochs of adaptation on Dollar Street-DA",
"andEuropeas the sourcedomain(2.9k)andimagesfrom",
"and 40 epochs on GeoYFCC-DA. All methods additionally",
"Asia, Africa, and South America as the target domain (8.8k)",
"minimize a supervised cross-entropy loss on source labels",
"GeoYFCC-DA: We consider images from North America as",
"during adaptation (with a loss weight of 1.0 and 0.1 on Dollar",
"thesourcedomain(24.1k)andimagesfromAsia,Australia",
"Street-DA and GeoYFCC-DA respectively). To combat label",
"and South America as the target domain (59.4k).",
"distribution shift, we follow Tan et al. [26] and use class-",
"On both shifts we create a 90%-10% train-test split on",
"balanced sampling on the source domain across experiments.",
"the source domain and report transfer performance on 20%",
"heldout target data, but use theentire targetdatasetfor un",
"4.3. Results",
"supervised adaptation. On Dollar Street-DA, due to the",
"Table I presents average accuracy on the target test set for",
"relatively small size of the dataset, we report performance",
"the Dollar Street-DA and GeoYFCC-DA shifts. We observe:",
"mean and standard deviation over three experimental runs.",
" Geographical shifts lead to significant performance",
"4.2.DomainAdaptationBaselines",
"drops (Row 1 v/s 2). As seen, the target oracle achieves",
"67.7% and 56.78% whereas the source model achieves a",
"We benchmark 4 representative DA methods from the",
"significantly lower performance of 54.66% (-13.1%) and",
"literature on our Geographical DA shifts: one domain",
"42.88% (-13.9%). Clearly, geographical variations pose a",
"discrepancy-based method, one domain adversarial method,",
"significant challenge to transfer.",
"and two self-training based methods.",
" DA methods offer limited improvements (Rows 3-6)",
"1 MMD [10]: Aligns domains by computing mean source",
"All the benchmarked methods achieve limited success at",
"and target embeddings in a reproducing kernel hilbert space",
"geographical adaptation, sometimes performing no better",
"and minimizing their distance as a maximum mean discrep",
"than the source model. Surprisingly, we find the simple SST",
"ancymeasure.",
"method toachievethebest (albeit small)improvement over",
"2) DANN [4]: Domain-adversarial neural networks adver-",
"the source model (+4.1% / +2.3%), but still well short of",
"sarially learn a domain discriminator to distinguish source",
"the target oracle (-9% / -11.6%). Altogether, these results"
] |
{
"arxiv_id": "2204.11122",
"language": "en",
"page_idx": 4
}
|
||
2212.08408-en-0
|
2212.08408-en-0.jpg
|
$\textbf{Decoder Tuning: Efficient Language Understanding as Decoding}$
Ganqu Cui$^1,\textbf{Wentao Li}^1,\textbf{Ning Ding}^1,\textbf{Longtao Huang}^2,\textbf{Zhiyuan Liu}^{1,3*}$,Maosong Sun$^1,3*$
$^{1}$ NLP Group, DCST, IAI, BNRIST, Tsinghua University, Beijing
$^{2}$ Alibaba Group $^3$ IICTUS, Shanghai cgq22@mails.tsinghua.edu.cn
## $\mathbf{Abstract}$
With the evergrowing sizes of pre-trained models (PTMs), it has been an emerging practice to only provide the inference APIs for users, namely model-as-a-service (MaaS) setting. To adapt PTMs with model parameters frozen, most current approaches focus on the input side, seeking for powerful prompts to stimulate models for correct answers. However, we argue that input-side adaptation could be arduous due to the lack of gradient signals and they usually require thousands of API queries, resulting in high computation and time costs. In light of this, we present Decoder Tuning (DecT), which in contrast optimizes task-specific decoder networks on the output side. Specifically, DecT fırst extracts prompt-stimulated output scores for initial predictions. On top of that, we train an additional decoder network on the output representations to incorporate posterior data knowledge. By gradientbased optimization, DecT can be trained within several seconds and requires only one PTM query per sample. Empirically, we conduct extensive natural language understanding experiments and show that DecT significantly outperforms state-of-the-art algorithms with a 200× speed-up. Our codes are available at https://github.com/thunlp/DecT.
## 1 Introduction
Recent advances in pre-trained models (PTMs) demonstrate the power of the “pre-trainingfine-tuning" paradigm, which empowers broad downstream NLP tasks with a single backbone model (Devlin et al., 2019; Raffel et al., 2020; Radford et al., 2019). Given the million even billionscale models, model-as-a-service (MaaS) has become an emerging practice in deploying massive PTMs, where users can only get access to model inference APIs (Brown et al., 2020; Sun et al., 2022b). Under such a scenario, PTMs' parameters
${^*\text{Correspon}}$ding Author.
<FigureHere>
Figure 1: Accuracy v.s. training time for MaaS adaptation methods under different training shots. We plot $\{1,4,16,64,256\}$-shot DecT,0-shot Prompt, 1-shot ICL and $\{1,4,16\}$-shot for other algorithms. DecT outperforms all baselines by a large margin with a $20\bar{0\times}$speed-up.
are frozen, and users cannot fine-tune the model on downstream tasks for adaptation. To find an alternative way, researchers have studied MaaS PTM adaptation methods extensively.
Most existing approaches in this line are based on $prompts$, which modify inputs with specific patterns. By wrapping inputs into cloze-style questions or prepending inputs with a few demonstrative examples, PTMs could produce the right outputs directly and show strong “in-context”learning abilities (Petroni et al., 2019; Brown et al., 2020) without any parameter update. Besides heuristic prompt design, some recent works try to optimize the input prompts without gradients. Among them, Black-box Tuning (BBT) (Sun et al., 2022b) and BBTv2 (Sun et al., 2022a) apply evolutionary algorithm (Hansen and Ostermeier, 2001) on continuous prompt tokens, while RLPrompt (Deng et al., 2022) adopts reinforcement learning to find discrete prompt tokens. Nevertheless, gradientfree optimization is rather difficult and these input-
|
[
[
0.16455078125,
0.0914306640625,
0.8330078125,
0.0926513671875,
0.8330078125,
0.10986328125,
0.16455078125,
0.1092529296875
],
[
0.12176513671875,
0.135986328125,
0.884765625,
0.133056640625,
0.884765625,
0.15380859375,
0.12176513671875,
0.15673828125
],
[
0.2359619140625,
0.155029296875,
0.765625,
0.155029296875,
0.765625,
0.169189453125,
0.2359619140625,
0.169189453125
],
[
0.345947265625,
0.168701171875,
0.65576171875,
0.169189453125,
0.65576171875,
0.18701171875,
0.345947265625,
0.1859130859375
],
[
0.339111328125,
0.189453125,
0.6650390625,
0.189453125,
0.6650390625,
0.203125,
0.339111328125,
0.203125
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.56103515625,
0.25830078125,
0.61279296875,
0.25830078125,
0.61279296875,
0.2724609375,
0.56103515625,
0.2724609375
],
[
0.67822265625,
0.2587890625,
0.771484375,
0.26123046875,
0.771484375,
0.2724609375,
0.67822265625,
0.270751953125
],
[
0.8017578125,
0.26318359375,
0.8095703125,
0.26318359375,
0.8095703125,
0.26953125,
0.8017578125,
0.26953125
],
[
0.8193359375,
0.2587890625,
0.8447265625,
0.2587890625,
0.8447265625,
0.271484375,
0.8193359375,
0.271484375
],
[
0.8017578125,
0.270751953125,
0.8134765625,
0.273681640625,
0.8076171875,
0.283935546875,
0.7958984375,
0.28076171875
],
[
0.81787109375,
0.268310546875,
0.86376953125,
0.270751953125,
0.8623046875,
0.28515625,
0.81689453125,
0.283203125
],
[
0.67919921875,
0.2724609375,
0.7197265625,
0.2724609375,
0.7197265625,
0.283203125,
0.67919921875,
0.283203125
],
[
0.5693359375,
0.27490234375,
0.580078125,
0.27490234375,
0.580078125,
0.27978515625,
0.5693359375,
0.27978515625
],
[
0.580078125,
0.271484375,
0.623046875,
0.271484375,
0.623046875,
0.283203125,
0.580078125,
0.283203125
],
[
0.66650390625,
0.27490234375,
0.68115234375,
0.27490234375,
0.68115234375,
0.279052734375,
0.66650390625,
0.279052734375
],
[
0.1435546875,
0.28076171875,
0.459228515625,
0.281982421875,
0.459228515625,
0.29638671875,
0.1435546875,
0.295654296875
],
[
0.5810546875,
0.28076171875,
0.6396484375,
0.283203125,
0.6396484375,
0.29443359375,
0.5810546875,
0.292236328125
],
[
0.145263671875,
0.29638671875,
0.459228515625,
0.29638671875,
0.459228515625,
0.31005859375,
0.145263671875,
0.31005859375
],
[
0.5380859375,
0.29638671875,
0.5576171875,
0.29638671875,
0.5576171875,
0.31005859375,
0.5380859375,
0.31005859375
],
[
0.14697265625,
0.312255859375,
0.45751953125,
0.312255859375,
0.45751953125,
0.322509765625,
0.14697265625,
0.322509765625
],
[
0.0218353271484375,
0.2998046875,
0.06378173828125,
0.2998046875,
0.06378173828125,
0.438232421875,
0.0218353271484375,
0.438232421875
],
[
0.509765625,
0.318359375,
0.54833984375,
0.3046875,
0.56494140625,
0.32666015625,
0.5263671875,
0.34033203125
],
[
0.145263671875,
0.32470703125,
0.459228515625,
0.32470703125,
0.459228515625,
0.33837890625,
0.145263671875,
0.33837890625
],
[
0.515625,
0.32470703125,
0.56494140625,
0.32470703125,
0.56494140625,
0.4091796875,
0.515625,
0.4091796875
],
[
0.145263671875,
0.339111328125,
0.459228515625,
0.339111328125,
0.459228515625,
0.352783203125,
0.145263671875,
0.352783203125
],
[
0.64990234375,
0.336181640625,
0.72119140625,
0.336181640625,
0.72119140625,
0.3466796875,
0.64990234375,
0.3466796875
],
[
0.145263671875,
0.35400390625,
0.459228515625,
0.35400390625,
0.459228515625,
0.36767578125,
0.145263671875,
0.36767578125
],
[
0.145263671875,
0.36767578125,
0.459228515625,
0.36767578125,
0.459228515625,
0.38134765625,
0.145263671875,
0.38134765625
],
[
0.145263671875,
0.38134765625,
0.45751953125,
0.38134765625,
0.45751953125,
0.394287109375,
0.145263671875,
0.394287109375
],
[
0.1485595703125,
0.39794921875,
0.45751953125,
0.39794921875,
0.45751953125,
0.407958984375,
0.1485595703125,
0.407958984375
],
[
0.145263671875,
0.410400390625,
0.4609375,
0.410400390625,
0.4609375,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.5400390625,
0.412109375,
0.55322265625,
0.410400390625,
0.5556640625,
0.42041015625,
0.54248046875,
0.4228515625
],
[
0.57080078125,
0.42041015625,
0.611328125,
0.4228515625,
0.609375,
0.43701171875,
0.5693359375,
0.435302734375
],
[
0.7421875,
0.42333984375,
0.76416015625,
0.42333984375,
0.76416015625,
0.435791015625,
0.7421875,
0.435791015625
],
[
0.6513671875,
0.42578125,
0.6689453125,
0.42578125,
0.6689453125,
0.435302734375,
0.6513671875,
0.435302734375
],
[
0.14697265625,
0.424560546875,
0.459228515625,
0.424560546875,
0.459228515625,
0.438232421875,
0.14697265625,
0.438232421875
],
[
0.69775390625,
0.42578125,
0.71728515625,
0.42578125,
0.71728515625,
0.435302734375,
0.69775390625,
0.435302734375
],
[
0.7900390625,
0.42578125,
0.81201171875,
0.42578125,
0.81201171875,
0.435791015625,
0.7900390625,
0.435791015625
],
[
0.8369140625,
0.42578125,
0.85546875,
0.42578125,
0.85546875,
0.435302734375,
0.8369140625,
0.435302734375
],
[
0.0201568603515625,
0.435791015625,
0.06378173828125,
0.435302734375,
0.0679931640625,
0.51708984375,
0.0251922607421875,
0.5185546875
],
[
0.145263671875,
0.439453125,
0.459228515625,
0.439453125,
0.459228515625,
0.453125,
0.145263671875,
0.453125
],
[
0.66015625,
0.43701171875,
0.755859375,
0.43701171875,
0.755859375,
0.448974609375,
0.66015625,
0.448974609375
],
[
0.145263671875,
0.453125,
0.459228515625,
0.453125,
0.459228515625,
0.466064453125,
0.145263671875,
0.466064453125
],
[
0.145263671875,
0.468505859375,
0.45751953125,
0.468505859375,
0.45751953125,
0.478515625,
0.145263671875,
0.478515625
],
[
0.51318359375,
0.47021484375,
0.8818359375,
0.47021484375,
0.8818359375,
0.48388671875,
0.51318359375,
0.48388671875
],
[
0.145263671875,
0.48095703125,
0.4609375,
0.48095703125,
0.4609375,
0.49462890625,
0.145263671875,
0.49462890625
],
[
0.509765625,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.498779296875,
0.509765625,
0.498779296875
],
[
0.145263671875,
0.495361328125,
0.459228515625,
0.495361328125,
0.459228515625,
0.5087890625,
0.145263671875,
0.5087890625
],
[
0.5146484375,
0.49951171875,
0.88330078125,
0.49951171875,
0.88330078125,
0.51318359375,
0.5146484375,
0.51318359375
],
[
0.145263671875,
0.51025390625,
0.459228515625,
0.51025390625,
0.459228515625,
0.52392578125,
0.145263671875,
0.52392578125
],
[
0.51123046875,
0.51318359375,
0.88330078125,
0.51318359375,
0.88330078125,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.0218353271484375,
0.51953125,
0.0621337890625,
0.51953125,
0.0621337890625,
0.72998046875,
0.0218353271484375,
0.72998046875
],
[
0.1435546875,
0.52392578125,
0.45751953125,
0.52392578125,
0.45751953125,
0.53759765625,
0.1435546875,
0.53759765625
],
[
0.509765625,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.509765625,
0.54150390625
],
[
0.1435546875,
0.5380859375,
0.459228515625,
0.5380859375,
0.459228515625,
0.5517578125,
0.1435546875,
0.5517578125
],
[
0.51318359375,
0.54345703125,
0.615234375,
0.54345703125,
0.615234375,
0.5546875,
0.51318359375,
0.5546875
],
[
0.1435546875,
0.552734375,
0.459228515625,
0.5517578125,
0.459228515625,
0.5654296875,
0.1435546875,
0.56640625
],
[
0.14697265625,
0.56640625,
0.459228515625,
0.56640625,
0.459228515625,
0.580078125,
0.14697265625,
0.580078125
],
[
0.145263671875,
0.58056640625,
0.45751953125,
0.58056640625,
0.45751953125,
0.59423828125,
0.145263671875,
0.59423828125
],
[
0.51318359375,
0.5830078125,
0.8818359375,
0.5830078125,
0.8818359375,
0.5966796875,
0.51318359375,
0.5966796875
],
[
0.145263671875,
0.595703125,
0.459228515625,
0.595703125,
0.459228515625,
0.609375,
0.145263671875,
0.609375
],
[
0.51123046875,
0.5986328125,
0.88330078125,
0.5986328125,
0.88330078125,
0.6123046875,
0.51123046875,
0.6123046875
],
[
0.145263671875,
0.609375,
0.459228515625,
0.609375,
0.459228515625,
0.62255859375,
0.145263671875,
0.62255859375
],
[
0.509765625,
0.6142578125,
0.8818359375,
0.61328125,
0.8818359375,
0.62744140625,
0.509765625,
0.62890625
],
[
0.145263671875,
0.62255859375,
0.45751953125,
0.62255859375,
0.45751953125,
0.6357421875,
0.145263671875,
0.6357421875
],
[
0.51123046875,
0.63134765625,
0.74658203125,
0.63134765625,
0.74658203125,
0.64453125,
0.51123046875,
0.64453125
],
[
0.145263671875,
0.63720703125,
0.45751953125,
0.63720703125,
0.45751953125,
0.65087890625,
0.145263671875,
0.65087890625
],
[
0.145263671875,
0.65185546875,
0.459228515625,
0.65185546875,
0.459228515625,
0.6650390625,
0.145263671875,
0.6650390625
],
[
0.52880859375,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.52880859375,
0.6611328125
],
[
0.145263671875,
0.6650390625,
0.453369140625,
0.6650390625,
0.453369140625,
0.6787109375,
0.145263671875,
0.6787109375
],
[
0.509765625,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.51123046875,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.51123046875,
0.693359375
],
[
0.1158447265625,
0.69482421875,
0.26025390625,
0.69482421875,
0.26025390625,
0.708984375,
0.1158447265625,
0.708984375
],
[
0.51123046875,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.71044921875,
0.51123046875,
0.71044921875
],
[
0.51123046875,
0.71142578125,
0.8818359375,
0.71044921875,
0.8818359375,
0.724609375,
0.51123046875,
0.7255859375
],
[
0.1175537109375,
0.72021484375,
0.48779296875,
0.72021484375,
0.48779296875,
0.7353515625,
0.1175537109375,
0.7353515625
],
[
0.51318359375,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51318359375,
0.7412109375
],
[
0.1158447265625,
0.736328125,
0.486083984375,
0.736328125,
0.486083984375,
0.75,
0.1158447265625,
0.75
],
[
0.509765625,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.509765625,
0.755859375
],
[
0.1175537109375,
0.7529296875,
0.48779296875,
0.7529296875,
0.48779296875,
0.76708984375,
0.1175537109375,
0.76708984375
],
[
0.51123046875,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.1175537109375,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.1175537109375,
0.7822265625
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.1175537109375,
0.78515625,
0.489501953125,
0.78515625,
0.489501953125,
0.798828125,
0.1175537109375,
0.798828125
],
[
0.51318359375,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.1175537109375,
0.80029296875,
0.490234375,
0.80029296875,
0.490234375,
0.81396484375,
0.1175537109375,
0.81396484375
],
[
0.51123046875,
0.80712890625,
0.8818359375,
0.8056640625,
0.8818359375,
0.82080078125,
0.51123046875,
0.82177734375
],
[
0.1175537109375,
0.81787109375,
0.490234375,
0.81787109375,
0.490234375,
0.83056640625,
0.1175537109375,
0.83056640625
],
[
0.51123046875,
0.822265625,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.51123046875,
0.83740234375
],
[
0.1175537109375,
0.83447265625,
0.48779296875,
0.83447265625,
0.48779296875,
0.84716796875,
0.1175537109375,
0.84716796875
],
[
0.509765625,
0.84033203125,
0.88330078125,
0.8388671875,
0.88330078125,
0.85400390625,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.849609375,
0.48779296875,
0.849609375,
0.48779296875,
0.8642578125,
0.1175537109375,
0.8642578125
],
[
0.5087890625,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.5087890625,
0.87060546875
],
[
0.1175537109375,
0.865234375,
0.489501953125,
0.865234375,
0.489501953125,
0.87890625,
0.1175537109375,
0.87890625
],
[
0.51318359375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1175537109375,
0.8818359375,
0.489501953125,
0.8818359375,
0.489501953125,
0.8955078125,
0.1175537109375,
0.8955078125
],
[
0.51318359375,
0.88818359375,
0.88330078125,
0.88818359375,
0.88330078125,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.145263671875,
0.90478515625,
0.2939453125,
0.90478515625,
0.2939453125,
0.91845703125,
0.145263671875,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.88330078125,
0.90478515625,
0.88330078125,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Decoder Tuning: Efficient Language Understanding as Decoding",
"Ganqu Cui', Wentao Li, Ning Ding', Longtao Huang, Zhiyuan Liul.3, Maosong Sunl.3*",
"1 NLP Group, DCST, IAI, BNRIST, Tsinghua University, Beijing",
"2 Alibaba Group 3 IICTUS, Shanghai",
"cgq22@mails.tsinghua.edu.cn",
"Abstract",
"BBT",
"PromptBoosting",
"O",
"ICL",
"★",
"Prompt",
"DecT",
"-",
"BBTv2",
"",
"With the evergrowing sizes of pre-trained mod-",
"RLPrompt",
"els (PTMs), it has been an emerging practice",
"80",
"to onlyprovidethe inference APIs for users",
" ",
"S",
"namely model-as-a-service (MaaS) setting. To",
".",
"adapt PTMs with model parameters frozen,",
"200×faster",
"most current approaches focus on the input side,",
"seeking for powerful prompts to stimulate mod-",
"els for correct answers. However, we argue",
"thatinput-sideadaptation could bearduousdue",
"to the lack of gradient signals and they usu-",
"50",
"010°",
"103",
"10",
"ally require thousands of API queries, resulting",
"102",
"104",
"10°",
"[T's]",
"in high computation and time costs. In light",
"Training Time (s)",
"of this, we present Decoder Tuning (DecT),",
"which incontrast optimizes task-specificde-",
"Figure l: Accuracy v.s. training time for MaaS adapta",
"coder networks on the output side. Specif-",
"tion methods under different training shots. We plot",
"ically, DecT first extracts prompt-stimulated",
"[1, 4, 16, 64, 256]-shot DecT, 0-shot Prompt, 1-shot",
"output scores for initial predictions. On top",
"ICL and [1, 4, 16]-shot for other algorithms. DecT",
"808021211",
"of that, we train an additional decoder net",
" outperforms all baselines by a large margin with a",
"work on the output representations to incor-",
"200xspeed-up.",
"porate posterior data knowledge. By gradient-",
"based optimization, DecT can be trained within",
"several seconds and requires only one PTM",
"are frozen, and users cannot fine-tune the model on",
"query per sample. Empirically, we conduct",
"downstream tasks for adaptation. To find an alter-",
"extensive natural language understanding ex-",
"native way, researchers have studied Maas PTM",
"periments and show that DecT significantly",
"adaptation methods extensively.",
"outperforms state-of-the-art algorithms with",
"a 200x speed-up. Our codes are available at",
"Most existing approaches in this line are based",
"https://github.com/thunlp/DecT.",
" on prompts, which modify inputs with specific pat",
"terns. By wrapping inputs into cloze-style ques-",
"1Introduction",
"tions or prepending inputs with a few demonstra-",
"tive examples, PTMs could produce the right out-",
"Recent advances in pre-trained models (PTMs)",
"puts directly and show strong“in-context\"learning",
"demonstrate the power of the “pre-training-",
"abilities (Petroni et al., 2019; Brown et al., 2020)",
"fine-tuning\" paradigm, which empowers broad",
"without any parameter update. Besides heuristic",
"downstream NLP tasks with a single backbone",
"prompt design, some recent works try to optimize",
"model (Devlin et al., 2019; Raffel et al., 2020; Rad-",
"the input prompts without gradients. Among them,",
"ford et al., 2019). Given the million even billion-",
"Black-box Tuning (BBT) (Sun et al., 2022b) and",
"scale models, model-as-a-service (MaaS) has be-",
"BBTv2 (Sun et al., 2022a) apply evolutionary al-",
"come an emerging practice in deploying massive",
"gorithm (Hansen and Ostermeier, 2001) on con-",
"PTMs, where users can only get access to model",
"tinuous prompt tokens, while RLPrompt (Deng",
"inference APIs (Brown et al., 2020; Sun et al.,",
"et al., 2022) adopts reinforcement learning to find",
"2022b). Under such a scenario, PTMs' parameters",
"discrete prompt tokens.Nevertheless, gradient-",
"*Corresponding Author.",
"free optimization is rather difficult and these input-"
] |
{
"arxiv_id": "2212.08408",
"language": "en",
"page_idx": 0
}
|
||
2212.08408-en-1
|
2212.08408-en-1.jpg
|
side methods need to query the PTMs thousands of times for optimization, which leads to huge inference costs in terms of time and computation resources. Moreover, their final performance is not satisfying as well.
Given the flaws of input-side adaptation, we turn to output-side adaptation, which builds tunable decoder networks on model outputs. Comparatively, output-side adaptation enjoys two major advantages: (1) We can directly tune decoder networks on top of model outputs with back-propagation rather than arduous alternatives. (2) We can reduce thousands of model queries to only once per sample. However, designing decoder networks is not straightforward. Past studies have shown that merely tuning an MLP or LSTM (Hochreiter and Schmidhuber, 1997) over output features cannot provide satisfying results (Sun et al.,2022a,b), leaving this path underexplored.
In this work, we aim to solve the performance issue for output-side adaptation, and we argue that there are two critical reasons behind it: (1) Simply utilizing PTMs as feature extractors ignores the infilling ability of PTMs, which is a strong prior for adaptation.(2) MLP and LSTM are not proper networks especially when training data is not sufficient.
Based on these findings, we present Decoder Tuning (DecT), an enhanced output-side adaptation method. Specifically, DecT has two crucial design choices to address the above issues. First, DecT queries the PTM with prompts and adopts model output scores as the initial predictions, which takes advantage of internal model knowledge. Second, on top of the output representations, we select a Prototypical Network (ProtoNet) (Snell et al.,2017) as the decoder network and train it to fit the training data, which is more suitable for few-shot learning. In this way, DecT modifies the initial model scores with subsequent training data, thus achieving better performance.
Through few-shot learning experiments on ten language understanding datasets, we highlight three advantages of DecT (see Figure 1). (1) DecT achieves over $3\%$ absolute accuracy improvement on average, greatly outperforming previous works. (2) DecT is highly efficient. Compared with major prompt engineering baselines, DecT dramatically reduces the average adaptation time from over 9,800 seconds (BBTv2) to 3 seconds. (3) DecT only requires one PTM query for each example,
while other input-side optimization methods need about $10^4$ calls. This advantage is vital when PTM calls are not for free. In addition, we conduct extensive ablation studies and validate the impact of each component of DecT.
### 2 Preliminaries
Given a set of training data $\mathcal{D}_\mathrm{train}=\{(x_i,y_i)\}_{i=1}^N$ and PTM $\mathcal{M}$,we need to predict the label $y\in$ $\{1,\ldots,K\}$ for sample $x$,where $K$ is the number of classes. We assume that each class has the same amount of $n$ training samples.
In the MaaS setting, $\mathcal{M}$ is a black-box inference API with fixed parameters. Therefore, we can only query the model with input $x$ and get corresponding outputs. To better utilize the PTMs, it has been a common practice to wrap input samples into prompts. Specifically, we enclose each input $x$ into a template $\mathcal{T}$ with a [MASK] token (here we assume using a masked language model). Then, we query $\mathcal{M}$ with $\mathcal{T}(x)$ and get the final layer hidden states h at the [MASK] position and scores $\mathbf{s}=S_{\mathcal{M}}(\mathcal{T}(x))\in\mathbb{R}^{K}$ over label words $\mathcal{V}.$ Take sentiment analysis as an example, we can use
$\mathcal{T}(x)=x$ In summary, it was [MASK].
as the template with $\mathcal{V}=\{$bad,great$\}$ as label words for negative and positive sentiment respectively. The output scores on these label words further correspond to the classes.
# 3 Methodology
In this section, we elaborate on our proposed Decoder Tuning (DecT) method for the classification task. We start with reviewing current input-side adaptation methods, then give an overview of DecT and finally detail it step-by-step.
## 3.1 Input-side Adaptation
Previous MaaS adaptation methods seek for optimal prompts that stimulate PTMs to output correct answers$^1.$ Without loss of generality, we formulate these methods with a transformation function $f(\cdot)$ which pre-processes the input $x.f(\cdot)$ can be specialized by adding demonstrations (Brown et al., 2020), discrete prompt tokens (Deng et al., 2022) or soft ones (Sun et al., 2022a,b). Denote the final score as $q(x)$ and probability as
$^{1}$BBTv2(Sun et al., 2022a) further optimizes prompt to-
kens in the intermediate layers, but we omit this here.
|
[
[
0.1175537109375,
0.08489990234375,
0.490234375,
0.0843505859375,
0.490234375,
0.09857177734375,
0.1175537109375,
0.09979248046875
],
[
0.509765625,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.1158447265625,
0.1015625,
0.490234375,
0.1015625,
0.490234375,
0.11517333984375,
0.1158447265625,
0.11517333984375
],
[
0.51318359375,
0.1015625,
0.8818359375,
0.1015625,
0.8818359375,
0.11517333984375,
0.51318359375,
0.11517333984375
],
[
0.1175537109375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.509765625,
0.133056640625,
0.880859375,
0.133056640625,
0.880859375,
0.146728515625,
0.509765625,
0.146728515625
],
[
0.118408203125,
0.1519775390625,
0.2509765625,
0.1519775390625,
0.2509765625,
0.162109375,
0.118408203125,
0.162109375
],
[
0.51123046875,
0.149658203125,
0.705078125,
0.149658203125,
0.705078125,
0.163330078125,
0.51123046875,
0.163330078125
],
[
0.1368408203125,
0.169189453125,
0.486083984375,
0.169189453125,
0.486083984375,
0.182861328125,
0.1368408203125,
0.182861328125
],
[
0.51318359375,
0.1776123046875,
0.6630859375,
0.1776123046875,
0.6630859375,
0.1923828125,
0.51318359375,
0.1923828125
],
[
0.1158447265625,
0.1859130859375,
0.489501953125,
0.1859130859375,
0.489501953125,
0.1995849609375,
0.1158447265625,
0.1995849609375
],
[
0.1175537109375,
0.201904296875,
0.490234375,
0.201904296875,
0.490234375,
0.215576171875,
0.1175537109375,
0.215576171875
],
[
0.51123046875,
0.204833984375,
0.87158203125,
0.204833984375,
0.87158203125,
0.2197265625,
0.51123046875,
0.2197265625
],
[
0.1175537109375,
0.218505859375,
0.490234375,
0.218505859375,
0.490234375,
0.232177734375,
0.1175537109375,
0.232177734375
],
[
0.51123046875,
0.221435546875,
0.880859375,
0.221435546875,
0.880859375,
0.235107421875,
0.51123046875,
0.235107421875
],
[
0.1175537109375,
0.2340087890625,
0.489501953125,
0.2340087890625,
0.489501953125,
0.2476806640625,
0.1175537109375,
0.2476806640625
],
[
0.51318359375,
0.2369384765625,
0.8818359375,
0.2369384765625,
0.8818359375,
0.251708984375,
0.51318359375,
0.251708984375
],
[
0.1175537109375,
0.25048828125,
0.48779296875,
0.25048828125,
0.48779296875,
0.26416015625,
0.1175537109375,
0.26416015625
],
[
0.51318359375,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.265380859375,
0.51318359375,
0.265380859375
],
[
0.1175537109375,
0.2666015625,
0.490234375,
0.2666015625,
0.490234375,
0.27978515625,
0.1175537109375,
0.27978515625
],
[
0.51318359375,
0.26953125,
0.736328125,
0.26953125,
0.736328125,
0.283203125,
0.51318359375,
0.283203125
],
[
0.118408203125,
0.283935546875,
0.486083984375,
0.283935546875,
0.486083984375,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.52880859375,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.2998046875,
0.52880859375,
0.2998046875
],
[
0.1175537109375,
0.2998046875,
0.48779296875,
0.2998046875,
0.48779296875,
0.3134765625,
0.1175537109375,
0.3134765625
],
[
0.51318359375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.1175537109375,
0.3154296875,
0.48779296875,
0.3154296875,
0.48779296875,
0.328857421875,
0.1175537109375,
0.328857421875
],
[
0.5087890625,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.330810546875,
0.5087890625,
0.330810546875
],
[
0.1175537109375,
0.330810546875,
0.489501953125,
0.330810546875,
0.489501953125,
0.344482421875,
0.1175537109375,
0.344482421875
],
[
0.51123046875,
0.334228515625,
0.8818359375,
0.333251953125,
0.8818359375,
0.347412109375,
0.51123046875,
0.3486328125
],
[
0.1175537109375,
0.3466796875,
0.490234375,
0.3466796875,
0.490234375,
0.35986328125,
0.1175537109375,
0.35986328125
],
[
0.509765625,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.509765625,
0.364013671875
],
[
0.1175537109375,
0.363525390625,
0.49365234375,
0.363525390625,
0.49365234375,
0.377197265625,
0.1175537109375,
0.377197265625
],
[
0.51123046875,
0.366455078125,
0.880859375,
0.366455078125,
0.880859375,
0.380126953125,
0.51123046875,
0.380126953125
],
[
0.1175537109375,
0.37890625,
0.32568359375,
0.37890625,
0.32568359375,
0.393798828125,
0.1175537109375,
0.393798828125
],
[
0.51123046875,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.133544921875,
0.3984375,
0.48779296875,
0.3984375,
0.48779296875,
0.412109375,
0.133544921875,
0.412109375
],
[
0.509765625,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.412109375,
0.509765625,
0.412109375
],
[
0.1158447265625,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.1158447265625,
0.4287109375
],
[
0.509765625,
0.41455078125,
0.8818359375,
0.41455078125,
0.8818359375,
0.4287109375,
0.509765625,
0.4287109375
],
[
0.1158447265625,
0.43115234375,
0.48779296875,
0.43115234375,
0.48779296875,
0.44482421875,
0.1158447265625,
0.44482421875
],
[
0.51123046875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.51123046875,
0.44482421875
],
[
0.1175537109375,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.46142578125,
0.1175537109375,
0.46142578125
],
[
0.5087890625,
0.446533203125,
0.8818359375,
0.4453125,
0.8818359375,
0.460205078125,
0.5087890625,
0.46142578125
],
[
0.1158447265625,
0.4619140625,
0.489501953125,
0.463134765625,
0.489501953125,
0.47802734375,
0.1158447265625,
0.476806640625
],
[
0.51123046875,
0.463134765625,
0.8486328125,
0.463134765625,
0.8486328125,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.1158447265625,
0.47802734375,
0.489501953125,
0.478515625,
0.489501953125,
0.493408203125,
0.1158447265625,
0.4921875
],
[
0.11419677734375,
0.495361328125,
0.48779296875,
0.495361328125,
0.48779296875,
0.5087890625,
0.11419677734375,
0.5087890625
],
[
0.5498046875,
0.4921875,
0.8447265625,
0.4921875,
0.8447265625,
0.50732421875,
0.5498046875,
0.50732421875
],
[
0.1158447265625,
0.51123046875,
0.160400390625,
0.51123046875,
0.160400390625,
0.525390625,
0.1158447265625,
0.525390625
],
[
0.51318359375,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.1351318359375,
0.53076171875,
0.489501953125,
0.53076171875,
0.489501953125,
0.5458984375,
0.1351318359375,
0.5458984375
],
[
0.51123046875,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.55419921875,
0.51123046875,
0.55419921875
],
[
0.11419677734375,
0.5458984375,
0.48779296875,
0.54638671875,
0.48779296875,
0.56103515625,
0.11419677734375,
0.56005859375
],
[
0.509765625,
0.5546875,
0.8818359375,
0.5546875,
0.8818359375,
0.568359375,
0.509765625,
0.568359375
],
[
0.1158447265625,
0.5634765625,
0.48779296875,
0.5634765625,
0.48779296875,
0.57666015625,
0.1158447265625,
0.57666015625
],
[
0.509765625,
0.57177734375,
0.736328125,
0.57177734375,
0.736328125,
0.5849609375,
0.509765625,
0.5849609375
],
[
0.1175537109375,
0.5791015625,
0.486083984375,
0.5791015625,
0.486083984375,
0.5927734375,
0.1175537109375,
0.5927734375
],
[
0.1175537109375,
0.595703125,
0.48779296875,
0.595703125,
0.48779296875,
0.609375,
0.1175537109375,
0.609375
],
[
0.5087890625,
0.59716796875,
0.65576171875,
0.5986328125,
0.65576171875,
0.6142578125,
0.5087890625,
0.61328125
],
[
0.1158447265625,
0.61083984375,
0.48779296875,
0.61083984375,
0.48779296875,
0.62451171875,
0.1158447265625,
0.62451171875
],
[
0.1175537109375,
0.62744140625,
0.490234375,
0.62744140625,
0.490234375,
0.64111328125,
0.1175537109375,
0.64111328125
],
[
0.509765625,
0.6259765625,
0.8818359375,
0.6259765625,
0.8818359375,
0.64013671875,
0.509765625,
0.64013671875
],
[
0.118408203125,
0.6455078125,
0.48779296875,
0.6455078125,
0.48779296875,
0.65625,
0.118408203125,
0.65625
],
[
0.509765625,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65625,
0.509765625,
0.65625
],
[
0.1175537109375,
0.66015625,
0.48779296875,
0.66015625,
0.48779296875,
0.67333984375,
0.1175537109375,
0.67333984375
],
[
0.509765625,
0.65771484375,
0.8818359375,
0.65771484375,
0.8818359375,
0.67138671875,
0.509765625,
0.67138671875
],
[
0.1175537109375,
0.67578125,
0.48779296875,
0.67578125,
0.48779296875,
0.689453125,
0.1175537109375,
0.689453125
],
[
0.509765625,
0.6748046875,
0.88330078125,
0.6748046875,
0.88330078125,
0.6884765625,
0.509765625,
0.6884765625
],
[
0.1158447265625,
0.6904296875,
0.489501953125,
0.69140625,
0.489501953125,
0.7060546875,
0.1158447265625,
0.705078125
],
[
0.51123046875,
0.69140625,
0.75146484375,
0.69140625,
0.75146484375,
0.705078125,
0.51123046875,
0.705078125
],
[
0.1158447265625,
0.70703125,
0.486083984375,
0.70703125,
0.486083984375,
0.72021484375,
0.1158447265625,
0.72021484375
],
[
0.51318359375,
0.71728515625,
0.73046875,
0.71728515625,
0.73046875,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.1158447265625,
0.7236328125,
0.48779296875,
0.7236328125,
0.48779296875,
0.73681640625,
0.1158447265625,
0.73681640625
],
[
0.11419677734375,
0.74072265625,
0.21826171875,
0.73828125,
0.21826171875,
0.7529296875,
0.11419677734375,
0.7548828125
],
[
0.51318359375,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.1351318359375,
0.75830078125,
0.48779296875,
0.7587890625,
0.48779296875,
0.77392578125,
0.1351318359375,
0.7724609375
],
[
0.509765625,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.7685546875,
0.509765625,
0.7685546875
],
[
0.1175537109375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.771484375,
0.88330078125,
0.771484375,
0.88330078125,
0.78515625,
0.51123046875,
0.78515625
],
[
0.1175537109375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51318359375,
0.78759765625,
0.88330078125,
0.78759765625,
0.88330078125,
0.80029296875,
0.51318359375,
0.80029296875
],
[
0.1158447265625,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1158447265625,
0.82177734375
],
[
0.51318359375,
0.80419921875,
0.88330078125,
0.80419921875,
0.88330078125,
0.818359375,
0.51318359375,
0.818359375
],
[
0.1175537109375,
0.82470703125,
0.491943359375,
0.82470703125,
0.491943359375,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.82080078125,
0.8818359375,
0.82080078125,
0.8818359375,
0.83447265625,
0.51123046875,
0.83447265625
],
[
0.118408203125,
0.84033203125,
0.491943359375,
0.84033203125,
0.491943359375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51318359375,
0.8359375,
0.8818359375,
0.8359375,
0.8818359375,
0.849609375,
0.51318359375,
0.849609375
],
[
0.1158447265625,
0.85546875,
0.49365234375,
0.85498046875,
0.49365234375,
0.869140625,
0.1158447265625,
0.87060546875
],
[
0.51123046875,
0.8515625,
0.880859375,
0.8515625,
0.880859375,
0.865234375,
0.51123046875,
0.865234375
],
[
0.1175537109375,
0.87255859375,
0.491943359375,
0.87255859375,
0.491943359375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.8681640625,
0.8818359375,
0.8681640625,
0.8818359375,
0.8828125,
0.509765625,
0.8828125
],
[
0.1175537109375,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.1175537109375,
0.90185546875
],
[
0.53076171875,
0.89013671875,
0.88330078125,
0.892578125,
0.88330078125,
0.90673828125,
0.53076171875,
0.90478515625
],
[
0.1175537109375,
0.90478515625,
0.49365234375,
0.90478515625,
0.49365234375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.50537109375,
0.90478515625,
0.8369140625,
0.90478515625,
0.8369140625,
0.91845703125,
0.50537109375,
0.91845703125
]
] |
[
"side methods need to query the PTMs thousands",
"while other input-side optimization methods need",
"of times for optimization, which leads to huge in-",
"about 104 calls. This advantage is vital when PTM",
"ference costs in terms of time and computation",
"calls are not for free. In addition, we conduct ex-",
"resources. Moreover, their final performance is not",
"tensive ablation studies and validate the impact of",
"satisfying as well.",
"each component of DecT.",
"Given the flaws of input-side adaptation, we turn",
"2Preliminaries",
"to output-side adaptation, which builds tunable de-",
"coder networks on model outputs. Comparatively,",
"Given a set of training data Drain = [(ci, yi)]N",
"output-side adaptation enjoys two major advan-",
"and PTM M, we need to predict the label y E",
"tages: (1) We can directly tune decoder networks",
"[1, ..., K] for sample &, where K is the number",
"on top of model outputs with back-propagation",
"of classes. We assume that each class has the same",
"rather than arduous alternatives. (2) We can re-",
"amount of n training samples.",
"duce thousands of model queries to only once per",
"In the MaaS setting, M is a black-box infer-",
"sample. However, designing decoder networks is",
"ence API with fixed parameters. Therefore, we can",
"not straightforward. Past studies have shown that",
" only query the model with input and get corre",
"merely tuning an MLP or LSTM (Hochreiter and",
"sponding outputs. To better utilize the PTMs, it",
"Schmidhuber, 1997) over output features cannot",
"has been a common practice to wrap input samples",
"provide satisfying results (Sun et al., 2022a,b), leav-",
"into prompts. Specifically, we enclose each input",
"ing this path underexplored.",
"α into a template T with a [MASK] token (here",
"In this work, we aim to solve the performance",
"we assume using a masked language model). Then",
"issue for output-side adaptation, and we argue that",
"we query M with T(α) and get the final layer hid-",
"there are two critical reasons behind it: (1) Simply",
"den states h at the [MASK] position and scores",
"utilizing PTMs as feature extractors ignores the",
"S = SM(T(α)) RK over label words V. Take",
"infilling ability of PTMs, which is a strong prior",
"sentiment analysis as an example, we can use",
"for adaptation. (2) MLP and LSTM are not proper",
"networks especially when training data is not suffi-",
"T(c) = In summary, it was [MASK].",
"cient.",
"as the template with V = [bad, great) as label",
"Based on these findings, we present Decoder",
"words for negative and positive sentiment respec-",
"Tuning (DecT), an enhanced output-side adaptation",
"tively. The output scores on these label words fur",
"method. Specifically, DecT has two crucial design",
"ther correspond to the classes.",
"choices to address the above issues. First, DecT",
"queries the PTM with prompts and adopts model",
"3Methodology",
"output scores as the initial predictions, which takes",
"advantage of internal model knowledge. Second,",
"In this section, we elaborate on our proposed De-",
"on top of the output representations, we select a",
"coder Tuning (DecT) method for the classification",
"Prototypical Network (ProtoNet) (Snell et al., 2017)",
"task. We start with reviewing current input-side",
"as the decoder network and train it to fit the training",
"adaptation methods, then give an overview of DecT",
"data, which is more suitable for few-shot learning.",
"and finally detail it step-by-step.",
"In this way, DecT modifies the initial model scores",
"3.1Input-sideAdaptation",
"with subsequent training data, thus achieving better",
"performance.",
"Previous MaaS adaptation methods seek for opti",
"Through few-shot learning experiments on ten",
"mal prompts that stimulate PTMs to output correct",
"language understanding datasets, we highlight",
"answers'. Without loss of generality, we formu-",
"three advantages of DecT (see Figure 1). (1) DecT",
"late these methods with a transformation function",
"achieves over 3% absolute accuracy improvement",
"f() which pre-processes the input a. f() can",
"on average, greatly outperforming previous works.",
"be specialized by adding demonstrations (Brown",
"(2) DecT is highly efficient. Compared with ma-",
"et al., 2020), discrete prompt tokens (Deng et al.",
"jor prompt engineering baselines, DecT dramati-",
"2022) or soft ones (Sun et al., 2022a,b). De-",
"cally reduces the average adaptation time from over",
"note the final score as q(αc) and probability as",
"9,800 seconds (BBTv2) to 3 seconds. (3) DecT",
"'BBTv2 (Sun et al., 2022a) further optimizes prompt to-",
"only requires one PTM query for each example,",
"kens in the intermediate layers, but we omit this here."
] |
{
"arxiv_id": "2212.08408",
"language": "en",
"page_idx": 1
}
|
||
2212.08408-en-3
|
2212.08408-en-3.jpg
|
data on top of PTM outputs, which combine both
knowledge effectively.
The design choice of Dec(·) is fairly flexible. In practice, we select Prototypical Networks (ProtoNet) (Snell et al., 2017) due to their simplicity and remarkable performance in few-shot learning and prompt-based tuning (Cui et al., 2022). For this, we project the hidden states with a linear layer parameterized by W and get sample representation
$$\mathbf{v}=\mathbf{Wh}.$$
(3)
On prototypes, classical approaches model them as points in the embedding space, which overlook the different class characteristics. Inspired by Ding et al. (2022a), we model prototypes as hyperspheres with an additional radius parameter. Concretely, the prototype for class $k$ contains two parameters, center position vector z$_k$ and radius scalar $r_k.$ We randomly initialize z$_k$ and initialize $r_k$ as the average distance between z$_k$ and instances in class $k:$
$$r_k=\dfrac{1}{N_k}\sum_i^{y_i=k}\|\mathbf{v}_i-\mathbf{z}_k\|_2.$$
(4)
As for the score function, we calculate the Euclidean distances between instances and prototypes.
(5)
$$\mathrm{Dec}(\mathbf{h},k)=-\|\mathbf{W}\mathbf{h}-\mathbf{z}_k\|_2+r_k.$$
According to Eq. 2, the final logit is
$$q(x,k)=-\|\mathbf{Wh}-\mathbf{z}_k\|_2+r_k+\lambda\mathbf{\hat{s}}_k.$$
(6)
From a geometric view, the score function calculates the distance from instance $x$ to the“surface” of the prototype, where $r_k+\lambda\mathbf{\hat{s}}_k$ is the whole radius acting like the bias term. With the scores, we can calculate the predicted probability by the Softmax function:
(7)
$$P(y=k|x)=\frac{\exp(q(x,k))}{\sum_{k'=1}^{K}\exp(q(x,k'))},$$
and we can optimize W and $r_k$ by the cross-
entropy loss
(8)
$$\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N}\log P(y_i|x_i).$$
## 4 Experiments
In this section, we first introduce the experimental settings (Section 4.1), then discuss the results for few-shot experiments (Section 4.2), efficiency comparison (Section 4.3), and experiment results for more training data (Section 4.4).
## 4.1 Experimental Settings
Datasets. We conduct experiments on four typical natural language understanding tasks. For sentiment analysis, we select SST2 (Socher et al., 2013), Yelp P. (Zhang et al., 2015) and IMDB (Maas et al., 2011). For text classification, we use AG's News, Yahoo (Zhang et al., 2015) and DBPedia (Lehmann et al., 2015). For natural language inference (NLI), we adopt RTE (Dagan et al., 2005; Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), SNLI (Bowman et al., 2015) and MNLI (Williams et al., 2018). For entity typing, we experiment on FewNERD (Ding et al., 2021b). We report dataset statistics in Appendix A.1
Splits. We randomly sample $n=1,4,16$ data instances for each class from the training set for few-shot learning, and sample same amount data for validation. For datasets in GLUE (Wang et al., 2019)(SST2, RTE, MNLI) and SNLI, we use the original validation sets as test sets following Zhang et al. (2021). For other datasets, we evaluate on their original test sets.
Baselines. We compare with representative MaaS PTM adaptation methods. Prompt refers to directly performing zero-shot classification with template-wrapped examples. In-context learning $(\mathbf{ICL})$ (Brown et al., 2020) further concatenates some exemplars before the test samples. BBT (Sun et al., 2022b) optimizes soft prompt tokens with an evolutionary algorithm, and BBTv2 (Sun et al., 2022a) further inserts deep prompts to intermediate layers for better performance. RLPrompt (Deng et al., 2022) is another recent algorithm that optimizes discrete prompts with reinforcement learning. PromptBoosting (Hou et al., 2022) is a concurrent work that applies boosting algorithm for prompt ensembling. We report the details of base- ${\tilde{l}}$ines in Appendix A.2.
Environments. For all experiments, we use NVIDIA A100 and RTX 2080 Ti GPUs. We implement DecT with PyTorch (Paszke et al., 2019), HuggingFace Tansformers (Wolf et al., 2020), and OpenPrompt (Ding et al., 2022b).
Implementation Details. For all methods, we use the same RoBERTa$_\mathrm{LARGE}$ (Liu et al., 2019) as the backbone model. For DecT, we set the representation dimension to 128 and optimize the parameters for 30 epochs with Adam optimizer (Kingma and Ba, 2015). The learning rate is 0.01. On the
|
[
[
0.1175537109375,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.72900390625,
0.08612060546875,
0.72900390625,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.1158447265625,
0.10272216796875,
0.28369140625,
0.10272216796875,
0.28369140625,
0.11639404296875,
0.1158447265625,
0.11639404296875
],
[
0.51318359375,
0.10687255859375,
0.8818359375,
0.10687255859375,
0.8818359375,
0.12054443359375,
0.51318359375,
0.12054443359375
],
[
0.1368408203125,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1368408203125,
0.1318359375
],
[
0.51123046875,
0.12353515625,
0.88330078125,
0.12353515625,
0.88330078125,
0.13720703125,
0.51123046875,
0.13720703125
],
[
0.1158447265625,
0.133056640625,
0.48779296875,
0.133056640625,
0.48779296875,
0.146728515625,
0.1158447265625,
0.146728515625
],
[
0.509765625,
0.1395263671875,
0.88330078125,
0.1383056640625,
0.88330078125,
0.152587890625,
0.509765625,
0.15380859375
],
[
0.1175537109375,
0.1519775390625,
0.486083984375,
0.1519775390625,
0.486083984375,
0.162109375,
0.1175537109375,
0.162109375
],
[
0.5087890625,
0.15380859375,
0.88330078125,
0.155029296875,
0.88330078125,
0.169189453125,
0.5087890625,
0.168701171875
],
[
0.1158447265625,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.1158447265625,
0.179931640625
],
[
0.51123046875,
0.17041015625,
0.884765625,
0.171630859375,
0.884765625,
0.1859130859375,
0.51123046875,
0.185302734375
],
[
0.1158447265625,
0.1817626953125,
0.489501953125,
0.18115234375,
0.489501953125,
0.1953125,
0.1158447265625,
0.196533203125
],
[
0.51123046875,
0.18701171875,
0.8818359375,
0.18701171875,
0.8818359375,
0.20068359375,
0.51123046875,
0.20068359375
],
[
0.1158447265625,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1158447265625,
0.2120361328125
],
[
0.51123046875,
0.2037353515625,
0.8818359375,
0.2037353515625,
0.8818359375,
0.21728515625,
0.51123046875,
0.21728515625
],
[
0.1158447265625,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.51123046875,
0.2203369140625,
0.8818359375,
0.2203369140625,
0.8818359375,
0.2340087890625,
0.51123046875,
0.2340087890625
],
[
0.2626953125,
0.2369384765625,
0.340087890625,
0.2369384765625,
0.340087890625,
0.251708984375,
0.2626953125,
0.251708984375
],
[
0.46337890625,
0.2392578125,
0.489501953125,
0.2392578125,
0.489501953125,
0.251708984375,
0.46337890625,
0.251708984375
],
[
0.51123046875,
0.236328125,
0.880859375,
0.236328125,
0.880859375,
0.2493896484375,
0.51123046875,
0.2493896484375
],
[
0.51318359375,
0.251708984375,
0.8818359375,
0.251708984375,
0.8818359375,
0.265380859375,
0.51318359375,
0.265380859375
],
[
0.1368408203125,
0.26123046875,
0.486083984375,
0.26123046875,
0.486083984375,
0.275634765625,
0.1368408203125,
0.275634765625
],
[
0.51123046875,
0.267333984375,
0.880859375,
0.268310546875,
0.880859375,
0.283203125,
0.51123046875,
0.281982421875
],
[
0.1158447265625,
0.27783203125,
0.48779296875,
0.27783203125,
0.48779296875,
0.29150390625,
0.1158447265625,
0.29150390625
],
[
0.51123046875,
0.283935546875,
0.8818359375,
0.283203125,
0.8818359375,
0.297607421875,
0.51123046875,
0.298583984375
],
[
0.1158447265625,
0.292236328125,
0.486083984375,
0.29345703125,
0.486083984375,
0.30810546875,
0.1158447265625,
0.30712890625
],
[
0.509765625,
0.300537109375,
0.84033203125,
0.300537109375,
0.84033203125,
0.314208984375,
0.509765625,
0.314208984375
],
[
0.1158447265625,
0.308837890625,
0.48779296875,
0.31005859375,
0.48779296875,
0.32470703125,
0.1158447265625,
0.32373046875
],
[
0.1158447265625,
0.32470703125,
0.489501953125,
0.325927734375,
0.489501953125,
0.34033203125,
0.1158447265625,
0.339111328125
],
[
0.509765625,
0.325927734375,
0.880859375,
0.325927734375,
0.880859375,
0.339111328125,
0.509765625,
0.339111328125
],
[
0.1175537109375,
0.342529296875,
0.489501953125,
0.342529296875,
0.489501953125,
0.355712890625,
0.1175537109375,
0.355712890625
],
[
0.509765625,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.355712890625,
0.509765625,
0.355712890625
],
[
0.1158447265625,
0.35693359375,
0.486083984375,
0.35693359375,
0.486083984375,
0.37060546875,
0.1158447265625,
0.37060546875
],
[
0.509765625,
0.358154296875,
0.880859375,
0.358154296875,
0.880859375,
0.371826171875,
0.509765625,
0.371826171875
],
[
0.11419677734375,
0.372314453125,
0.490234375,
0.37353515625,
0.490234375,
0.388427734375,
0.11419677734375,
0.38720703125
],
[
0.509765625,
0.37353515625,
0.880859375,
0.37353515625,
0.880859375,
0.38720703125,
0.509765625,
0.38720703125
],
[
0.1175537109375,
0.39013671875,
0.489501953125,
0.39013671875,
0.489501953125,
0.40380859375,
0.1175537109375,
0.40380859375
],
[
0.51318359375,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.51318359375,
0.405029296875
],
[
0.1175537109375,
0.407958984375,
0.1351318359375,
0.407958984375,
0.1351318359375,
0.41748046875,
0.1175537109375,
0.41748046875
],
[
0.51123046875,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.51123046875,
0.421630859375
],
[
0.246826171875,
0.4228515625,
0.307373046875,
0.413330078125,
0.312255859375,
0.427490234375,
0.2509765625,
0.43701171875
],
[
0.274658203125,
0.427001953125,
0.40478515625,
0.427001953125,
0.40478515625,
0.453125,
0.274658203125,
0.453125
],
[
0.51123046875,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.51123046875,
0.43701171875
],
[
0.2032470703125,
0.435302734375,
0.235107421875,
0.435302734375,
0.235107421875,
0.446533203125,
0.2032470703125,
0.446533203125
],
[
0.46337890625,
0.432861328125,
0.489501953125,
0.432861328125,
0.489501953125,
0.446533203125,
0.46337890625,
0.446533203125
],
[
0.2435302734375,
0.440673828125,
0.287109375,
0.440673828125,
0.287109375,
0.4560546875,
0.2435302734375,
0.4560546875
],
[
0.509765625,
0.438232421875,
0.6767578125,
0.439453125,
0.6767578125,
0.45361328125,
0.509765625,
0.453125
],
[
0.1175537109375,
0.463134765625,
0.490234375,
0.463134765625,
0.490234375,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.51123046875,
0.465576171875,
0.8818359375,
0.465576171875,
0.8818359375,
0.479736328125,
0.51123046875,
0.479736328125
],
[
0.1158447265625,
0.478515625,
0.495361328125,
0.479736328125,
0.495361328125,
0.49462890625,
0.1158447265625,
0.493408203125
],
[
0.509765625,
0.482177734375,
0.8818359375,
0.482177734375,
0.8818359375,
0.495361328125,
0.509765625,
0.495361328125
],
[
0.1158447265625,
0.498779296875,
0.162109375,
0.498779296875,
0.162109375,
0.51025390625,
0.1158447265625,
0.51025390625
],
[
0.51123046875,
0.49755859375,
0.8818359375,
0.49755859375,
0.8818359375,
0.5126953125,
0.51123046875,
0.5126953125
],
[
0.17041015625,
0.5185546875,
0.429931640625,
0.51953125,
0.429931640625,
0.53369140625,
0.17041015625,
0.533203125
],
[
0.51123046875,
0.51416015625,
0.8818359375,
0.51416015625,
0.8818359375,
0.529296875,
0.51123046875,
0.529296875
],
[
0.466796875,
0.5224609375,
0.486083984375,
0.5224609375,
0.486083984375,
0.5322265625,
0.466796875,
0.5322265625
],
[
0.51123046875,
0.529296875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.51123046875,
0.54345703125
],
[
0.1351318359375,
0.5419921875,
0.4013671875,
0.5419921875,
0.4013671875,
0.55712890625,
0.1351318359375,
0.55712890625
],
[
0.51123046875,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56005859375,
0.51123046875,
0.56005859375
],
[
0.15869140625,
0.56640625,
0.44921875,
0.56640625,
0.44921875,
0.58056640625,
0.15869140625,
0.58056640625
],
[
0.509765625,
0.56103515625,
0.8818359375,
0.5625,
0.8818359375,
0.57666015625,
0.509765625,
0.576171875
],
[
0.455078125,
0.5693359375,
0.486083984375,
0.5693359375,
0.486083984375,
0.57763671875,
0.455078125,
0.57763671875
],
[
0.51123046875,
0.57763671875,
0.8818359375,
0.57666015625,
0.8818359375,
0.59130859375,
0.51123046875,
0.5927734375
],
[
0.1158447265625,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6015625,
0.1158447265625,
0.6015625
],
[
0.51123046875,
0.59326171875,
0.8818359375,
0.59423828125,
0.8818359375,
0.609375,
0.51123046875,
0.60791015625
],
[
0.1158447265625,
0.60546875,
0.489501953125,
0.60546875,
0.489501953125,
0.619140625,
0.1158447265625,
0.619140625
],
[
0.51123046875,
0.609375,
0.880859375,
0.60986328125,
0.880859375,
0.62451171875,
0.51123046875,
0.62353515625
],
[
0.1158447265625,
0.62158203125,
0.48779296875,
0.62060546875,
0.48779296875,
0.63525390625,
0.1158447265625,
0.6357421875
],
[
0.51318359375,
0.626953125,
0.8818359375,
0.626953125,
0.8818359375,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.1158447265625,
0.63818359375,
0.48779296875,
0.63818359375,
0.48779296875,
0.65185546875,
0.1158447265625,
0.65185546875
],
[
0.51123046875,
0.642578125,
0.8818359375,
0.642578125,
0.8818359375,
0.65673828125,
0.51123046875,
0.65673828125
],
[
0.1158447265625,
0.65283203125,
0.489501953125,
0.65380859375,
0.489501953125,
0.66845703125,
0.1158447265625,
0.66748046875
],
[
0.509765625,
0.65771484375,
0.8818359375,
0.65673828125,
0.8818359375,
0.67138671875,
0.509765625,
0.6728515625
],
[
0.1175537109375,
0.67138671875,
0.187255859375,
0.67138671875,
0.187255859375,
0.68310546875,
0.1175537109375,
0.68310546875
],
[
0.51318359375,
0.6748046875,
0.8818359375,
0.6748046875,
0.8818359375,
0.6884765625,
0.51318359375,
0.6884765625
],
[
0.306396484375,
0.6884765625,
0.403076171875,
0.6884765625,
0.403076171875,
0.703125,
0.306396484375,
0.703125
],
[
0.16455078125,
0.6953125,
0.258544921875,
0.6953125,
0.258544921875,
0.71044921875,
0.16455078125,
0.71044921875
],
[
0.509765625,
0.69140625,
0.8818359375,
0.6904296875,
0.8818359375,
0.705078125,
0.509765625,
0.7060546875
],
[
0.25439453125,
0.70166015625,
0.258544921875,
0.70166015625,
0.258544921875,
0.70703125,
0.25439453125,
0.70703125
],
[
0.466796875,
0.69970703125,
0.486083984375,
0.69970703125,
0.486083984375,
0.708984375,
0.466796875,
0.708984375
],
[
0.276123046875,
0.7080078125,
0.43408203125,
0.7060546875,
0.43408203125,
0.7236328125,
0.276123046875,
0.7255859375
],
[
0.51123046875,
0.7060546875,
0.67919921875,
0.70703125,
0.67919921875,
0.7216796875,
0.51123046875,
0.72021484375
],
[
0.1158447265625,
0.72998046875,
0.49365234375,
0.72998046875,
0.49365234375,
0.74365234375,
0.1158447265625,
0.74365234375
],
[
0.51123046875,
0.732421875,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.51123046875,
0.7451171875
],
[
0.11419677734375,
0.74658203125,
0.2091064453125,
0.74462890625,
0.2091064453125,
0.7587890625,
0.11419677734375,
0.76123046875
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.2822265625,
0.7685546875,
0.297119140625,
0.7685546875,
0.297119140625,
0.77783203125,
0.2822265625,
0.77783203125
],
[
0.5146484375,
0.765625,
0.880859375,
0.765625,
0.880859375,
0.779296875,
0.5146484375,
0.779296875
],
[
0.208251953125,
0.77783203125,
0.2283935546875,
0.7861328125,
0.219970703125,
0.79736328125,
0.198974609375,
0.7890625
],
[
0.26123046875,
0.77490234375,
0.267822265625,
0.77490234375,
0.267822265625,
0.7822265625,
0.26123046875,
0.7822265625
],
[
0.297119140625,
0.78076171875,
0.4013671875,
0.78076171875,
0.4013671875,
0.7958984375,
0.297119140625,
0.7958984375
],
[
0.46337890625,
0.78076171875,
0.489501953125,
0.78076171875,
0.489501953125,
0.79345703125,
0.46337890625,
0.79345703125
],
[
0.51123046875,
0.78076171875,
0.8818359375,
0.78076171875,
0.8818359375,
0.7958984375,
0.51123046875,
0.7958984375
],
[
0.26025390625,
0.7880859375,
0.273681640625,
0.79345703125,
0.264404296875,
0.802734375,
0.252685546875,
0.79736328125
],
[
0.28369140625,
0.80029296875,
0.295654296875,
0.80029296875,
0.295654296875,
0.8046875,
0.28369140625,
0.8046875
],
[
0.51123046875,
0.79736328125,
0.76171875,
0.79736328125,
0.76171875,
0.81103515625,
0.51123046875,
0.81103515625
],
[
0.118408203125,
0.81787109375,
0.1318359375,
0.81787109375,
0.1318359375,
0.826171875,
0.118408203125,
0.826171875
],
[
0.1351318359375,
0.8193359375,
0.1427001953125,
0.8193359375,
0.1427001953125,
0.82373046875,
0.1351318359375,
0.82373046875
],
[
0.14697265625,
0.81640625,
0.2568359375,
0.81640625,
0.2568359375,
0.82763671875,
0.14697265625,
0.82763671875
],
[
0.51318359375,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83740234375,
0.51318359375,
0.83740234375
],
[
0.1175537109375,
0.84033203125,
0.49365234375,
0.84033203125,
0.49365234375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51123046875,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.509765625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.509765625,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.509765625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1158447265625,
0.90283203125,
0.3837890625,
0.90380859375,
0.3837890625,
0.91845703125,
0.1158447265625,
0.91748046875
],
[
0.509765625,
0.90283203125,
0.8818359375,
0.90380859375,
0.8818359375,
0.91845703125,
0.509765625,
0.91748046875
]
] |
[
"data on top of PTM outputs, which combine both",
"4.1 Experimental Settings",
"knowledge effectively.",
"Datasets.We conduct experiments on four typi-",
"The design choice of Dec() is fairly flexible.",
"cal natural language understanding tasks. For senti-",
"In practice, we select Prototypical Networks (Pro-",
"ment analysis, we select SST2 (Socher et al., 2013),",
"toNet)(Snell et al..2017)due to their simplicity",
"Yelp P. (Zhang et al., 2015) and IMDB (Maas et al.,",
"and remarkable performance in few-shot learning",
"2011). For text classification, we use AG's News,",
"and prompt-based tuning (Cui et al., 2022). For",
"Yahoo (Zhang et al., 2015) and DBPedia (Lehmann",
"this, we project the hidden states with a linear layer",
"et al., 2015).For natural language inference",
"parameterized by W and get sample representation",
"(NLD), we adopt RTE (Dagan et al., 2005; Haim",
"v = Wh.",
"(3)",
"et al., 2006; Giampiccolo et al., 2007; Bentivogli",
"et al., 2009), SNLI (Bowman et al., 2015) and",
"On prototypes, classical approaches model them",
"MNLI (Williams et al., 2018). For entity typing",
"as points in the embedding space, which overlook",
"we experiment on FewNERD (Ding et al., 2021b)",
"the different class characteristics. Inspired by Ding",
"We report dataset statistics in Appendix A.1.",
"et al. (2022a), we model prototypes as hyperspheres",
"with an additional radius parameter. Concretely,",
"Splits. We randomly sample n = 1, 4. 16 data",
"the prototype for class k contains two parameters,",
"instances for each class from the training set for",
"center position vector zk and radius scalar rk. We",
"few-shot learning, and sample same amount data",
"randomly initialize zk and initialize rk as the av-",
"for validation. For datasets in GLUE (Wang et al.",
"erage distance between zZk and instances in class",
"2019) (SST2, RTE, MNLI) and SNLI, we use the",
"k:",
"original validation sets as test sets following Zhang",
"1 yi=h",
"Z IIvi - z/ll2.",
"et al. (2021). For other datasets, we evaluate on",
"Tk :",
"(4)",
"Nk",
"their original test sets.",
"As for the score function, we calculate the Eu-",
"Baselines.We compare with representative",
"clidean distances between instances and proto-",
"MaaS PTM adaptation methods. Prompt refers",
"types.",
"to directly performing zero-shot classification with",
"Dec(h, k) = -Wh - zkll2 + rk:",
"template-wrapped examples. In-context learning",
"(5)",
"(ICL) (Brown et al., 2020) further concatenates",
"According to Eq. 2, the final logit is",
"some exemplars before the test samples. BBT (Sun",
"q(α, k) = -l/Wh -zkll2 +rk + 入sk.",
"et al., 2022b) optimizes soft prompt tokens with",
"（6)",
"an evolutionary algorithm, and BBTv2 (Sun et al..",
"From a geometric view, the score function calcu-",
"2022a) further inserts deep prompts to intermediate",
"lates the distance from instance α to the “surface\"",
"layers for better performance. RLPrompt (Deng",
"of the prototype,whererk+s,is the wholeradius",
"et al., 2022) is another recent algorithm that opti-",
"acting like the bias term. With the scores, we can",
"mizes discrete prompts with reinforcement learn-",
"calculate the predicted probability by the Softmax",
"ing. PromptBoosting (Hou et al., 2022) is a con-",
"function:",
"current work that applies boosting algorithm for",
"exp(q(c, k))",
"P(y = klr)",
"prompt ensembling. We report the details of base-",
"",
"(7)",
"Ek=1 exp(q(, k))",
"lines in Appendix A.2.",
"and we can optimize W and rk by the cross-",
"Environments.For all experiments,we use",
"entropy loss",
"NVIDIAA100andRTX2080TiGPUs.Weim-",
"N",
"plement DecT with PyTorch (Paszke et al., 2019)",
"C",
"",
"log P(yi|ci).",
"(8)",
"HuggingFace Tansformers (Wolf et al., 2020), and",
"N",
"",
"OpenPrompt (Ding et al., 2022b).",
"4",
"",
"Experiments",
"Implementation Details. For all methods, we",
"In this section, we first introduce the experimen-",
"use the same RoBERTaLARGE (Liu et al., 2019) as",
"tal settings (Section 4.1),then discusstheresults",
"the backbone model. For DecT, we set the repre",
"for few-shot experiments (Section 4.2), efficiency",
" sentation dimension to 128 and optimize the param",
"comparison (Section 4.3), and experiment results",
"eters for 30 epochs withAdamoptimizer (Kingma",
"for more training data (Section 4.4).",
"and Ba, 2015). The learning rate is 0.01. On the"
] |
{
"arxiv_id": "2212.08408",
"language": "en",
"page_idx": 3
}
|
||
2212.08408-en-4
|
2212.08408-en-4.jpg
|
<table>
<tbody>
<tr>
<th>$m$</th>
<th>$\mathbf{Method}$</th>
<th>$SST2$</th>
<th>$IMDB$</th>
<th>Yelp</th>
<th>$AG$</th>
<th>$DB$</th>
<th>Yahoo</th>
<th>$RTE$</th>
<th>$SNLI$</th>
<th>$\mathbf{MNLI-m/mm}$</th>
<th>$\mathbf{NERD}$</th>
<th>$Avg.$</th>
</tr>
<tr>
<td>0</td>
<td>$\mathbf{Prompt}$</td>
<td>83.3</td>
<td>89.4</td>
<td>87.1</td>
<td>80.9</td>
<td>68.4</td>
<td>49.9</td>
<td>52.4</td>
<td>40.7</td>
<td>50.8/51.7</td>
<td>21.1</td>
<td>61.4</td>
</tr>
<tr>
<td rowspan="5">0 1</td>
<td>$ICL$</td>
<td>81.5 7 1 T</td>
<td>$65.6_{1}.$</td>
<td>81.1</td>
<td>66.7</td>
<td>71.7</td>
<td>53.2,</td>
<td>45.0.</td>
<td>46.1</td>
<td>$53.6_{0.5}/53.9_{0.8}$</td>
<td>34.7</td>
<td>59.4</td>
</tr>
<tr>
<td>$BBT$</td>
<td>$83.4_{1.3}$</td>
<td>$89.0_{0.1}$</td>
<td>$89.7_{0.1}$</td>
<td>$75.4_{0.8}$</td>
<td>$59.1_{1.7}$</td>
<td>$31.2_{2.7}$</td>
<td>$52.3_{1.4}$</td>
<td>$38.5_{0.8}$</td>
<td>$43.4_{2.5}/42.9_{3.3}$</td>
<td>$14.1_{2.3}$</td>
<td>$56.3_{1.5}$</td>
</tr>
<tr>
<td>$BBTv2$</td>
<td>$83.3_{2.5}$</td>
<td>$89.0_{0.2}$</td>
<td>$89.9_{0.2}$</td>
<td>$74.3_{3.2}$</td>
<td>$74.2_{5.2}$</td>
<td>$34.0_{3.5}$</td>
<td>$48.2_{5.7}$</td>
<td>$38.6_{4.0}$</td>
<td>$44.2_{3.2}/44.3_{4.5}$</td>
<td>$29.0_{0.8}$</td>
<td>$59.0_{3.0}$</td>
</tr>
<tr>
<td>$\mathbf{RLPrompt}$</td>
<td>$63.5_{6.3}$</td>
<td>$65.0_{6.5}$</td>
<td>$66.3_{6.9}$</td>
<td>$72.5_{4.5}$</td>
<td>65.65.5</td>
<td>$38.1_{5.8}$</td>
<td>$53.8_{5.3}$</td>
<td>$36.5_{3.0}$</td>
<td>$40.3_{2.0}/41.0_{2.1}$</td>
<td>$14.5_{1.8}$</td>
<td>$50.6_{4.7}$</td>
</tr>
<tr>
<td>${\mathbf{PromptBoosting}}$</td>
<td>$86.7_{2.6}$</td>
<td>$82.4_{6.1}$</td>
<td>$88.7_{2.5}$</td>
<td>$58.7_{11.8}$</td>
<td>$73.0_{4.8}$</td>
<td>$23.7_{7.0}$</td>
<td>50.05.9</td>
<td>$43.5_{6.1}$</td>
<td>$36.8_{1.6}/36.3_{2.3}$</td>
<td>$22.0_{0.8}$</td>
<td>$54.7_{4.6}$</td>
</tr>
<tr>
<td> </td>
<td>DecT</td>
<td>$90.8_{0.2}$</td>
<td>$91.2_{0.3}$</td>
<td>$94.8_{0.1}$</td>
<td>$79.9_{1.1}$</td>
<td>$78.8_{0.9}$</td>
<td>$55.2_{0.8}$</td>
<td>$56.0_{2.7}$</td>
<td>$47.7_{4.1}$</td>
<td>$52.2_{2.7}/53.3_{3.0}$</td>
<td>$35.7_{1.5}$</td>
<td>$66.9_{1.6}$</td>
</tr>
<tr>
<td rowspan="6">4</td>
<td>$ICL$</td>
<td>60.39.8</td>
<td>$80.4_{6.6}$</td>
<td>$77.4_{14.6}$</td>
<td>$65.1_{5.4}$</td>
<td>71.76.5</td>
<td>49.99.9</td>
<td>$42.7_{3.9}$</td>
<td>$42.1_{3.2}$</td>
<td>$44.7_{5.9}/45.2_{6.0}$</td>
<td>$31.7_{4.8}$</td>
<td>$55.6_{7.0}$</td>
</tr>
<tr>
<td>$BBT$</td>
<td>$84.5_{1.2}$</td>
<td>$89.8_{0.9}$</td>
<td>$90.2_{0.6}$</td>
<td>$79.0_{2.1}$</td>
<td>$67.7_{3.5}$</td>
<td>$42.9_{0.6}$</td>
<td>$48.4_{4.0}$</td>
<td>$40.5_{1.3}$</td>
<td>$41.2_{1.7}/40.7_{2.0}$</td>
<td>$19.4_{1.5}$</td>
<td>$58.6_{1.8}$</td>
</tr>
<tr>
<td>$BBTv2$</td>
<td>$86.6_{2.2}$</td>
<td>$89.4_{0.6}$</td>
<td>$90.3_{0.5}$</td>
<td>$79.1_{2.1}$</td>
<td>$89.0_{1.7}$</td>
<td>$46.0_{1.4}$</td>
<td>$46.2_{2.3}$</td>
<td>$40.8_{4.3}$</td>
<td>$44.0_{0.9}/44.8_{1.6}$</td>
<td>$31.9_{1.4}$</td>
<td>$62.6_{1.7}$</td>
</tr>
<tr>
<td>$\mathbf{RLPrompt}$</td>
<td>$80.7_{7.5}$</td>
<td>$75.8_{10.1}$</td>
<td>$78.8_{7.3}$</td>
<td>$76.1_{4.8}$</td>
<td>76.35.9</td>
<td>$45.0_{3.1}$</td>
<td>$53.5_{2.9}$</td>
<td>$36.3_{2.6}$</td>
<td>$44.4_{2.9}/45.5_{3.8}$</td>
<td>$16.7_{2.4}$</td>
<td>$57.4_{4.8}$</td>
</tr>
<tr>
<td>${\mathbf{PromptBoosting}}$</td>
<td>$88.9_{2.3}$</td>
<td>$83.0_{5.2}$</td>
<td>$92.3_{2.1}$</td>
<td>$78.2_{6.8}$</td>
<td>$90.1_{0.7}$</td>
<td>$36.4_{5.1}$</td>
<td>53.55.9</td>
<td>$53.4_{3.4}$</td>
<td>$39.8_{4.5}/40.3_{5.7}$</td>
<td>$40.9_{2.5}$</td>
<td>$63.4_{4.0}$</td>
</tr>
<tr>
<td>DecT</td>
<td>$87.6_{1.6}$</td>
<td>$89.6_{0.9}$</td>
<td>$94.8_{0.7}$</td>
<td>$81.9_{2.6}$</td>
<td>$89.1_{0.6}$</td>
<td>$59.9_{2.1}$</td>
<td>$56.7_{2.7}$</td>
<td>$53.2_{2.9}$</td>
<td>$52.2_{2.3}/53.4_{2.4}$</td>
<td>$46.7_{1.7}$</td>
<td>$69.5_{1.9}$</td>
</tr>
<tr>
<td rowspan="6">16</td>
<td>$ICL$</td>
<td>$71.5_{15.8}$</td>
<td>$80.6_{6.0}$</td>
<td>$73.7_{14.5}$</td>
<td>$64.4_{6.0}$</td>
<td>71.89.1</td>
<td>52.65.7</td>
<td>$43.8_{7.0}$</td>
<td>$42.0_{6.3}$</td>
<td>$51.4_{3.0}/52.1_{3.3}$</td>
<td>$35.1_{2.6}$</td>
<td>$58.1_{7.2}$</td>
</tr>
<tr>
<td>$BBT$</td>
<td>$89.6_{0.3}^b$</td>
<td>$89.3_{0.4}$</td>
<td>$91.5_{4}^{\mathrm{t}}$ 1.2 1</td>
<td>$81.5_{1}^{\prime}$ $'0.8$ 1</td>
<td>87.8! 1</td>
<td>$48.3_{1.4}$</td>
<td>$52.6_{j}^{b}$ 2.2</td>
<td>$46.6^{t}$ 1.3</td>
<td>$40.0_{2.6}/39.9_{2.9}$</td>
<td>$17.8_{1.4}$</td>
<td>$62.3_{1.5}$</td>
</tr>
<tr>
<td>$\mathbf{BBTv2}$</td>
<td>$90.3_1^a$</td>
<td>$88.6_{2.1}$</td>
<td>$92.9_{0}^{a}$</td>
<td>$85.3_{0}^{a}$</td>
<td>$93.6_{0}^{a}$</td>
<td>$52.0_{1.4}$</td>
<td>$56.7_{3}^a$</td>
<td>$57.3_{2}^a$</td>
<td>$50.1_{2.4}/51.7_{3.2}$</td>
<td>$33.3_{1.0}$</td>
<td>$68.3_{1.7}$</td>
</tr>
<tr>
<td>$\mathbf{RLPrompt}$</td>
<td>$87.0_{2.6}$</td>
<td>$87.6_{2.4}$</td>
<td>$95.1_{1.0}^c$</td>
<td>$80.2_{0.7}^c$</td>
<td>$80.8_{3.3}$</td>
<td>$48.1_{2.2}$</td>
<td>$54.3_{2.8}$</td>
<td>$41.1_{5.0}$</td>
<td>$43.3_{3.9}/44.3_{4.5}$</td>
<td>$17.5_{1.4}$</td>
<td>$61.8_{2.7}$</td>
</tr>
<tr>
<td>${\mathbf{PromptBoosting}}$</td>
<td>$87.6_{i}^{c}$</td>
<td>$86.2_{3.1}$</td>
<td>$94.7_{1.0}$</td>
<td>$85.2_{0}^{d}$</td>
<td>$95.0_{0.5}$</td>
<td>$46.6_{2.4}$</td>
<td>$60.0_{5}^\mathrm{d}$</td>
<td>$61.3_{2}^{\mathrm{d}}$</td>
<td>$/50.4_{5.1}$ $52. 5_{1}^{d}$ .</td>
<td>$52.1_{2.6}$</td>
<td>70.1. $\cdot2.6$</td>
</tr>
<tr>
<td>DecT</td>
<td>91.00.5</td>
<td>$91.0_{0.9}$</td>
<td>$95.4_{0.3}$</td>
<td>$86.4_{0.4}$</td>
<td>94.60.5</td>
<td>$64.2_{0.7}$</td>
<td>$59.7_{1.8}$</td>
<td>$60.5_{0.8}$</td>
<td>$55.3_{1.3}/56.8_{1.5}$</td>
<td>$53.5_{1.8}$</td>
<td>$73.5_{1.\mathrm{c}}$</td>
</tr>
</tbody>
</table>
Table 1: Experiment results for MaaS adaptation methods. Some baseline results are taken from corresponding papers $(^a$Sun et al. (2022a), $^b$Sun et al. (2022b), cDeng et al. (2022), $^d$Hou et al. (2022)). We run other experiments over 5 random seeds and report average accuracy and standard deviation $(\%).$ Best results are in bold
<table>
<tbody>
<tr>
<th>$\mathbf{Method}$</th>
<th>$\mathbf{Tr.Time}$ $(\mathbf{s})$</th>
<th>$\#Query$</th>
<th>$\#$Param.$(\mathbf{K})$</th>
</tr>
<tr>
<td>$ICL$</td>
<td>0</td>
<td>0</td>
<td>0</td>
</tr>
<tr>
<td>$BBT$</td>
<td>10,512</td>
<td>8,000</td>
<td>0.5</td>
</tr>
<tr>
<td>$\mathbf{BBTv2}$</td>
<td>9,856</td>
<td>8,000</td>
<td>12</td>
</tr>
<tr>
<td>$\mathbf{RLPrompt}$</td>
<td>65,579</td>
<td>12,000</td>
<td>3,100</td>
</tr>
<tr>
<td>$\mathbf{PromptBoosting}$</td>
<td>644</td>
<td>10</td>
<td>0.4</td>
</tr>
<tr>
<td>DecT</td>
<td>3</td>
<td>1</td>
<td>130</td>
</tr>
</tbody>
</table>
selection of $\lambda$,we directly set $\lambda=1/n$ for most datasets based on the intuition that $\lambda$ should decrease as the amount of training data increases. On MNLI and FewNERD, we tune $\lambda$ on the validation set and select $\lambda=1$ and $\lambda=1/16$ respectively. We give the templates and label words in Appendix A.3.
### 4.2 Main Results
Table 2: Efficiency comparison of MaaS adaptation methods. Training time is the average wall clock time measured in the 16-shot setting.“Tr.”stands for Training and “Param.” stands for Parameter.
Table 1 presents the main few-shot learning results.
From the results, we have these observations:
zero-shot prompt and ICL are strong baselines. However, as shown in the table, ICL gives the best results in the 1-shot setting, and it can hardly improve with more training data due to the input length restriction. To compare, merely optimizing the input prompts (BBT and RLPrompt) can hardly outperform them, showing the limitation of input-side prompt optimization. In contrast, two other baselines, BBTv2 and PromptBoosting, are more powerful because they either inserts additional learnable prompt tokens inside the PTM or ensembles the outputs of different prompts. With the superior results of DecT, we argue that outputside optimization is a promising way for MaaS PTM adaptation.
Overall, DecT outperforms the state-of-theart baseline methods by a large margin (more than $3\%$ on average), especially under extreme data scarcity, showing its superior performance. Across different tasks, DecT and baselines obtain similar results on some easy sentiment analysis and topic classification tasks, but we highlight that DecT is much more favorable on difficult datasets, such as Yahoo and FewNERD. While other baseline methods struggle to optimize well, DecT surpasses them significantly (about 10% on Yahoo and $20\%$ on FewNERD under 16-shot setting compared with BBTv2 and ICL).
On stability, DecT also has consistently low variance and some baselines (ICL, RLPrompt and PromptBoosting) are unstable. Given the difficulty of few-shot PTM adaptation, it is of great significance that the adaptation method is robust to random seeds.
On baselines, optimization-free methods, i.e.
## 4.3 Efficiency Comparison
Despite the superior performance, another major advantage of DecT is its high efficiency. In Fig-
|
[
[
0.47607421875,
0.08489990234375,
0.50146484375,
0.08489990234375,
0.50146484375,
0.09857177734375,
0.47607421875,
0.09857177734375
],
[
0.5205078125,
0.08489990234375,
0.56494140625,
0.08489990234375,
0.56494140625,
0.09979248046875,
0.5205078125,
0.09979248046875
],
[
0.84228515625,
0.08489990234375,
0.87158203125,
0.0885009765625,
0.86962890625,
0.09979248046875,
0.8388671875,
0.097412109375
],
[
0.2509765625,
0.08612060546875,
0.287109375,
0.08612060546875,
0.287109375,
0.097412109375,
0.2509765625,
0.097412109375
],
[
0.1175537109375,
0.08612060546875,
0.19140625,
0.08612060546875,
0.19140625,
0.097412109375,
0.1175537109375,
0.097412109375
],
[
0.3046875,
0.08612060546875,
0.343505859375,
0.08612060546875,
0.343505859375,
0.09857177734375,
0.3046875,
0.09857177734375
],
[
0.364501953125,
0.08612060546875,
0.3955078125,
0.08612060546875,
0.3955078125,
0.1009521484375,
0.364501953125,
0.1009521484375
],
[
0.420654296875,
0.08612060546875,
0.44580078125,
0.08612060546875,
0.44580078125,
0.09857177734375,
0.420654296875,
0.09857177734375
],
[
0.5810546875,
0.08612060546875,
0.61279296875,
0.08612060546875,
0.61279296875,
0.09857177734375,
0.5810546875,
0.09857177734375
],
[
0.6337890625,
0.08612060546875,
0.6689453125,
0.08612060546875,
0.6689453125,
0.09857177734375,
0.6337890625,
0.09857177734375
],
[
0.6884765625,
0.08612060546875,
0.767578125,
0.08612060546875,
0.767578125,
0.097412109375,
0.6884765625,
0.097412109375
],
[
0.78515625,
0.08612060546875,
0.8271484375,
0.08612060546875,
0.8271484375,
0.097412109375,
0.78515625,
0.097412109375
],
[
0.474365234375,
0.097412109375,
0.50732421875,
0.1015625,
0.5029296875,
0.1175537109375,
0.47021484375,
0.1134033203125
],
[
0.63818359375,
0.097412109375,
0.6689453125,
0.1009521484375,
0.66650390625,
0.1175537109375,
0.634765625,
0.114013671875
],
[
0.8427734375,
0.09857177734375,
0.8818359375,
0.1015625,
0.87890625,
0.11639404296875,
0.84033203125,
0.1134033203125
],
[
0.580078125,
0.09979248046875,
0.61376953125,
0.09979248046875,
0.61376953125,
0.11517333984375,
0.580078125,
0.11517333984375
],
[
0.11419677734375,
0.09857177734375,
0.19140625,
0.1009521484375,
0.19140625,
0.11517333984375,
0.11419677734375,
0.1134033203125
],
[
0.25439453125,
0.1015625,
0.285400390625,
0.1015625,
0.285400390625,
0.114013671875,
0.25439453125,
0.114013671875
],
[
0.30908203125,
0.1015625,
0.339111328125,
0.1015625,
0.339111328125,
0.114013671875,
0.30908203125,
0.114013671875
],
[
0.364501953125,
0.1015625,
0.39208984375,
0.1015625,
0.39208984375,
0.114013671875,
0.364501953125,
0.114013671875
],
[
0.4189453125,
0.1015625,
0.44921875,
0.1015625,
0.44921875,
0.114013671875,
0.4189453125,
0.114013671875
],
[
0.5283203125,
0.1015625,
0.55908203125,
0.1015625,
0.55908203125,
0.114013671875,
0.5283203125,
0.114013671875
],
[
0.69873046875,
0.1009521484375,
0.75732421875,
0.1009521484375,
0.75732421875,
0.11517333984375,
0.69873046875,
0.11517333984375
],
[
0.7900390625,
0.1015625,
0.8212890625,
0.1015625,
0.8212890625,
0.114013671875,
0.7900390625,
0.114013671875
],
[
0.4130859375,
0.114013671875,
0.453369140625,
0.11639404296875,
0.45166015625,
0.1318359375,
0.411376953125,
0.1300048828125
],
[
0.466796875,
0.114013671875,
0.509765625,
0.11639404296875,
0.5087890625,
0.1318359375,
0.465087890625,
0.1300048828125
],
[
0.67919921875,
0.11517333984375,
0.7734375,
0.1175537109375,
0.7734375,
0.1318359375,
0.67919921875,
0.1300048828125
],
[
0.8388671875,
0.11517333984375,
0.88330078125,
0.1175537109375,
0.8818359375,
0.1318359375,
0.8369140625,
0.1300048828125
],
[
0.353515625,
0.11517333984375,
0.403076171875,
0.1175537109375,
0.4013671875,
0.1318359375,
0.3525390625,
0.1300048828125
],
[
0.246826171875,
0.11639404296875,
0.291259765625,
0.11639404296875,
0.291259765625,
0.1312255859375,
0.246826171875,
0.1312255859375
],
[
0.300537109375,
0.11639404296875,
0.345947265625,
0.11639404296875,
0.345947265625,
0.1312255859375,
0.300537109375,
0.1312255859375
],
[
0.5205078125,
0.11639404296875,
0.56494140625,
0.11639404296875,
0.56494140625,
0.1312255859375,
0.5205078125,
0.1312255859375
],
[
0.57666015625,
0.1175537109375,
0.61865234375,
0.1175537109375,
0.61865234375,
0.1312255859375,
0.57666015625,
0.1312255859375
],
[
0.63232421875,
0.1175537109375,
0.6708984375,
0.1175537109375,
0.6708984375,
0.1300048828125,
0.63232421875,
0.1300048828125
],
[
0.78515625,
0.1175537109375,
0.8271484375,
0.1175537109375,
0.8271484375,
0.1312255859375,
0.78515625,
0.1312255859375
],
[
0.466796875,
0.1258544921875,
0.509765625,
0.127685546875,
0.5087890625,
0.142578125,
0.465087890625,
0.14013671875
],
[
0.14111328125,
0.1175537109375,
0.172119140625,
0.1175537109375,
0.172119140625,
0.1300048828125,
0.14111328125,
0.1300048828125
],
[
0.2476806640625,
0.1270751953125,
0.291259765625,
0.12890625,
0.28955078125,
0.141357421875,
0.246826171875,
0.1395263671875
],
[
0.30322265625,
0.1270751953125,
0.344970703125,
0.12890625,
0.343505859375,
0.141357421875,
0.301513671875,
0.1395263671875
],
[
0.4130859375,
0.1270751953125,
0.453369140625,
0.1300048828125,
0.45166015625,
0.142578125,
0.41064453125,
0.1383056640625
],
[
0.5205078125,
0.1270751953125,
0.56494140625,
0.12890625,
0.5634765625,
0.141357421875,
0.51904296875,
0.1395263671875
],
[
0.62890625,
0.1270751953125,
0.67236328125,
0.1300048828125,
0.6708984375,
0.142578125,
0.62744140625,
0.1395263671875
],
[
0.68408203125,
0.1258544921875,
0.77099609375,
0.127685546875,
0.77099609375,
0.142578125,
0.68408203125,
0.14013671875
],
[
0.8369140625,
0.1270751953125,
0.8818359375,
0.1300048828125,
0.880859375,
0.142578125,
0.83642578125,
0.1395263671875
],
[
0.35693359375,
0.127685546875,
0.39794921875,
0.1300048828125,
0.397216796875,
0.141357421875,
0.355224609375,
0.1395263671875
],
[
0.14111328125,
0.127685546875,
0.173828125,
0.127685546875,
0.173828125,
0.14013671875,
0.14111328125,
0.14013671875
],
[
0.5751953125,
0.127685546875,
0.61865234375,
0.1300048828125,
0.6171875,
0.141357421875,
0.57421875,
0.1395263671875
],
[
0.78515625,
0.127685546875,
0.82861328125,
0.1300048828125,
0.8271484375,
0.141357421875,
0.7841796875,
0.1395263671875
],
[
0.30322265625,
0.1383056640625,
0.343505859375,
0.14013671875,
0.341796875,
0.152587890625,
0.301513671875,
0.15087890625
],
[
0.41064453125,
0.1383056640625,
0.456787109375,
0.1383056640625,
0.456787109375,
0.152587890625,
0.41064453125,
0.152587890625
],
[
0.466796875,
0.1383056640625,
0.509765625,
0.14013671875,
0.5087890625,
0.152587890625,
0.465087890625,
0.15087890625
],
[
0.51904296875,
0.13720703125,
0.56494140625,
0.1395263671875,
0.5634765625,
0.15380859375,
0.51708984375,
0.1519775390625
],
[
0.62890625,
0.1383056640625,
0.67236328125,
0.141357421875,
0.6708984375,
0.152587890625,
0.62744140625,
0.149658203125
],
[
0.8369140625,
0.1383056640625,
0.8818359375,
0.141357421875,
0.880859375,
0.152587890625,
0.83642578125,
0.149658203125
],
[
0.2476806640625,
0.14013671875,
0.291259765625,
0.14013671875,
0.291259765625,
0.1519775390625,
0.2476806640625,
0.1519775390625
],
[
0.14111328125,
0.14013671875,
0.187255859375,
0.14013671875,
0.187255859375,
0.1519775390625,
0.14111328125,
0.1519775390625
],
[
0.358642578125,
0.14013671875,
0.399658203125,
0.14013671875,
0.399658203125,
0.1519775390625,
0.358642578125,
0.1519775390625
],
[
0.5751953125,
0.14013671875,
0.6171875,
0.14013671875,
0.6171875,
0.1519775390625,
0.5751953125,
0.1519775390625
],
[
0.68701171875,
0.14013671875,
0.76904296875,
0.14013671875,
0.76904296875,
0.1519775390625,
0.68701171875,
0.1519775390625
],
[
0.7841796875,
0.14013671875,
0.8271484375,
0.14013671875,
0.8271484375,
0.1519775390625,
0.7841796875,
0.1519775390625
],
[
0.139404296875,
0.1484375,
0.208251953125,
0.15087890625,
0.208251953125,
0.1650390625,
0.139404296875,
0.163330078125
],
[
0.2451171875,
0.1484375,
0.29296875,
0.15087890625,
0.291259765625,
0.1650390625,
0.2435302734375,
0.163330078125
],
[
0.30322265625,
0.149658203125,
0.343505859375,
0.1519775390625,
0.341796875,
0.1644287109375,
0.301513671875,
0.162109375
],
[
0.35693359375,
0.149658203125,
0.399658203125,
0.1519775390625,
0.39794921875,
0.1644287109375,
0.355224609375,
0.162109375
],
[
0.4130859375,
0.149658203125,
0.455078125,
0.1519775390625,
0.453369140625,
0.1644287109375,
0.411376953125,
0.162109375
],
[
0.468505859375,
0.149658203125,
0.509765625,
0.1519775390625,
0.5087890625,
0.1644287109375,
0.466796875,
0.162109375
],
[
0.51904296875,
0.1484375,
0.56494140625,
0.15087890625,
0.5634765625,
0.1650390625,
0.51708984375,
0.163330078125
],
[
0.62890625,
0.149658203125,
0.67236328125,
0.152587890625,
0.6708984375,
0.1650390625,
0.62744140625,
0.162109375
],
[
0.8369140625,
0.149658203125,
0.8818359375,
0.152587890625,
0.880859375,
0.1644287109375,
0.83642578125,
0.160888671875
],
[
0.68701171875,
0.1519775390625,
0.76904296875,
0.1519775390625,
0.76904296875,
0.163330078125,
0.68701171875,
0.163330078125
],
[
0.5751953125,
0.1519775390625,
0.61865234375,
0.1519775390625,
0.61865234375,
0.163330078125,
0.5751953125,
0.163330078125
],
[
0.78515625,
0.1519775390625,
0.8271484375,
0.1519775390625,
0.8271484375,
0.163330078125,
0.78515625,
0.163330078125
],
[
0.139404296875,
0.1602783203125,
0.2392578125,
0.162109375,
0.2392578125,
0.177001953125,
0.139404296875,
0.174560546875
],
[
0.5205078125,
0.1602783203125,
0.56689453125,
0.163330078125,
0.5634765625,
0.1776123046875,
0.51708984375,
0.174560546875
],
[
0.62890625,
0.1602783203125,
0.67333984375,
0.162109375,
0.67236328125,
0.177001953125,
0.62744140625,
0.174560546875
],
[
0.30322265625,
0.160888671875,
0.343505859375,
0.163330078125,
0.341796875,
0.17578125,
0.301513671875,
0.17333984375
],
[
0.2476806640625,
0.160888671875,
0.291259765625,
0.1644287109375,
0.28955078125,
0.177001953125,
0.246826171875,
0.17333984375
],
[
0.358642578125,
0.160888671875,
0.4013671875,
0.1644287109375,
0.399658203125,
0.17578125,
0.35693359375,
0.1728515625
],
[
0.466796875,
0.160888671875,
0.509765625,
0.163330078125,
0.5087890625,
0.17578125,
0.465087890625,
0.17333984375
],
[
0.5751953125,
0.160888671875,
0.61865234375,
0.1644287109375,
0.6171875,
0.17578125,
0.57421875,
0.1728515625
],
[
0.78515625,
0.160888671875,
0.8271484375,
0.163330078125,
0.8251953125,
0.174560546875,
0.7841796875,
0.1728515625
],
[
0.8369140625,
0.160888671875,
0.8818359375,
0.1644287109375,
0.880859375,
0.177001953125,
0.83642578125,
0.17333984375
],
[
0.68701171875,
0.163330078125,
0.767578125,
0.163330078125,
0.767578125,
0.174560546875,
0.68701171875,
0.174560546875
],
[
0.41064453125,
0.162109375,
0.45751953125,
0.1644287109375,
0.456787109375,
0.17578125,
0.408935546875,
0.17333984375
],
[
0.411376953125,
0.17041015625,
0.45751953125,
0.17333984375,
0.455078125,
0.188232421875,
0.408935546875,
0.185302734375
],
[
0.8369140625,
0.169189453125,
0.88671875,
0.174560546875,
0.88330078125,
0.190673828125,
0.8330078125,
0.185302734375
],
[
0.353515625,
0.171630859375,
0.4013671875,
0.17333984375,
0.399658203125,
0.188232421875,
0.3525390625,
0.1859130859375
],
[
0.301513671875,
0.171630859375,
0.344970703125,
0.17333984375,
0.343505859375,
0.188232421875,
0.300537109375,
0.1859130859375
],
[
0.46337890625,
0.171630859375,
0.51123046875,
0.17333984375,
0.509765625,
0.188232421875,
0.462646484375,
0.1859130859375
],
[
0.57275390625,
0.171630859375,
0.62060546875,
0.17333984375,
0.61865234375,
0.188232421875,
0.57080078125,
0.1859130859375
],
[
0.62744140625,
0.171630859375,
0.67333984375,
0.17333984375,
0.67236328125,
0.188232421875,
0.62646484375,
0.1859130859375
],
[
0.5205078125,
0.1728515625,
0.56494140625,
0.17578125,
0.5634765625,
0.188232421875,
0.51904296875,
0.185302734375
],
[
0.14111328125,
0.17333984375,
0.177978515625,
0.17333984375,
0.177978515625,
0.1859130859375,
0.14111328125,
0.1859130859375
],
[
0.246826171875,
0.1728515625,
0.291259765625,
0.17578125,
0.28955078125,
0.18701171875,
0.2451171875,
0.18408203125
],
[
0.6884765625,
0.174560546875,
0.767578125,
0.174560546875,
0.767578125,
0.1859130859375,
0.6884765625,
0.1859130859375
],
[
0.7841796875,
0.174560546875,
0.8271484375,
0.174560546875,
0.8271484375,
0.1859130859375,
0.7841796875,
0.1859130859375
],
[
0.14111328125,
0.1859130859375,
0.172119140625,
0.188232421875,
0.17041015625,
0.201904296875,
0.139404296875,
0.1995849609375
],
[
0.3046875,
0.18701171875,
0.343505859375,
0.189453125,
0.341796875,
0.201904296875,
0.30322265625,
0.1995849609375
],
[
0.5224609375,
0.18701171875,
0.5634765625,
0.189453125,
0.5615234375,
0.201904296875,
0.5205078125,
0.1995849609375
],
[
0.355224609375,
0.188232421875,
0.4013671875,
0.190673828125,
0.399658203125,
0.201904296875,
0.353515625,
0.1995849609375
],
[
0.246826171875,
0.188232421875,
0.291259765625,
0.188232421875,
0.291259765625,
0.203125,
0.246826171875,
0.203125
],
[
0.5751953125,
0.188232421875,
0.61865234375,
0.190673828125,
0.6171875,
0.201904296875,
0.57421875,
0.1995849609375
],
[
0.7841796875,
0.188232421875,
0.8271484375,
0.190673828125,
0.8251953125,
0.201904296875,
0.78271484375,
0.1995849609375
],
[
0.8388671875,
0.188232421875,
0.8818359375,
0.190673828125,
0.880859375,
0.201904296875,
0.8369140625,
0.1995849609375
],
[
0.468505859375,
0.189453125,
0.509765625,
0.189453125,
0.509765625,
0.203125,
0.468505859375,
0.203125
],
[
0.4130859375,
0.189453125,
0.455078125,
0.189453125,
0.455078125,
0.203125,
0.4130859375,
0.203125
],
[
0.63232421875,
0.189453125,
0.67333984375,
0.189453125,
0.67333984375,
0.203125,
0.63232421875,
0.203125
],
[
0.6884765625,
0.189453125,
0.767578125,
0.189453125,
0.767578125,
0.20068359375,
0.6884765625,
0.20068359375
],
[
0.465087890625,
0.19775390625,
0.51318359375,
0.201904296875,
0.509765625,
0.215576171875,
0.462646484375,
0.21142578125
],
[
0.5224609375,
0.198974609375,
0.5634765625,
0.201904296875,
0.5615234375,
0.21435546875,
0.5205078125,
0.21142578125
],
[
0.63037109375,
0.198974609375,
0.67236328125,
0.201904296875,
0.6708984375,
0.21435546875,
0.62890625,
0.21142578125
],
[
0.68408203125,
0.19775390625,
0.77099609375,
0.1995849609375,
0.77099609375,
0.21435546875,
0.68408203125,
0.2120361328125
],
[
0.78662109375,
0.198974609375,
0.8271484375,
0.201904296875,
0.8251953125,
0.21435546875,
0.7841796875,
0.210205078125
],
[
0.246826171875,
0.1995849609375,
0.29296875,
0.1995849609375,
0.29296875,
0.21435546875,
0.246826171875,
0.21435546875
],
[
0.14111328125,
0.1995849609375,
0.173828125,
0.1995849609375,
0.173828125,
0.2120361328125,
0.14111328125,
0.2120361328125
],
[
0.300537109375,
0.1995849609375,
0.34765625,
0.1995849609375,
0.34765625,
0.21435546875,
0.300537109375,
0.21435546875
],
[
0.35693359375,
0.1995849609375,
0.399658203125,
0.201904296875,
0.39794921875,
0.213134765625,
0.355224609375,
0.21142578125
],
[
0.5751953125,
0.1995849609375,
0.61865234375,
0.201904296875,
0.6171875,
0.21435546875,
0.57421875,
0.2120361328125
],
[
0.8388671875,
0.1995849609375,
0.8818359375,
0.201904296875,
0.880859375,
0.21435546875,
0.8369140625,
0.2120361328125
],
[
0.246826171875,
0.208984375,
0.29296875,
0.21142578125,
0.291259765625,
0.2257080078125,
0.2451171875,
0.223876953125
],
[
0.4130859375,
0.20068359375,
0.455078125,
0.20068359375,
0.455078125,
0.21435546875,
0.4130859375,
0.21435546875
],
[
0.358642578125,
0.208984375,
0.399658203125,
0.2120361328125,
0.39794921875,
0.2244873046875,
0.35693359375,
0.221435546875
],
[
0.8369140625,
0.208984375,
0.88330078125,
0.2120361328125,
0.880859375,
0.226806640625,
0.83447265625,
0.223876953125
],
[
0.4130859375,
0.210205078125,
0.453369140625,
0.2120361328125,
0.45166015625,
0.2244873046875,
0.411376953125,
0.22265625
],
[
0.30322265625,
0.210205078125,
0.344970703125,
0.2120361328125,
0.343505859375,
0.2244873046875,
0.301513671875,
0.22265625
],
[
0.466796875,
0.210205078125,
0.5087890625,
0.2120361328125,
0.50732421875,
0.2244873046875,
0.465087890625,
0.22265625
],
[
0.5224609375,
0.210205078125,
0.5615234375,
0.2120361328125,
0.56103515625,
0.2244873046875,
0.5205078125,
0.22265625
],
[
0.62890625,
0.210205078125,
0.67236328125,
0.213134765625,
0.6708984375,
0.2257080078125,
0.62744140625,
0.22265625
],
[
0.7841796875,
0.210205078125,
0.8271484375,
0.2120361328125,
0.8251953125,
0.223876953125,
0.78271484375,
0.221435546875
],
[
0.5751953125,
0.21142578125,
0.61865234375,
0.213134765625,
0.6171875,
0.2244873046875,
0.57421875,
0.22265625
],
[
0.14111328125,
0.21142578125,
0.187255859375,
0.21142578125,
0.187255859375,
0.22265625,
0.14111328125,
0.22265625
],
[
0.68701171875,
0.2120361328125,
0.76904296875,
0.2120361328125,
0.76904296875,
0.223876953125,
0.68701171875,
0.223876953125
],
[
0.301513671875,
0.2203369140625,
0.344970703125,
0.223876953125,
0.343505859375,
0.236328125,
0.300537109375,
0.2327880859375
],
[
0.5751953125,
0.2203369140625,
0.6171875,
0.223876953125,
0.615234375,
0.236328125,
0.57421875,
0.2327880859375
],
[
0.62890625,
0.2203369140625,
0.6708984375,
0.223876953125,
0.6689453125,
0.236328125,
0.62744140625,
0.2327880859375
],
[
0.246826171875,
0.221435546875,
0.291259765625,
0.221435546875,
0.291259765625,
0.236328125,
0.246826171875,
0.236328125
],
[
0.14111328125,
0.221435546875,
0.20654296875,
0.223876953125,
0.20654296875,
0.236328125,
0.14111328125,
0.2340087890625
],
[
0.358642578125,
0.221435546875,
0.399658203125,
0.223876953125,
0.39794921875,
0.236328125,
0.35693359375,
0.2340087890625
],
[
0.411376953125,
0.221435546875,
0.453369140625,
0.223876953125,
0.45166015625,
0.236328125,
0.41064453125,
0.2340087890625
],
[
0.466796875,
0.221435546875,
0.5087890625,
0.223876953125,
0.50732421875,
0.236328125,
0.465087890625,
0.2340087890625
],
[
0.5224609375,
0.221435546875,
0.5615234375,
0.223876953125,
0.56103515625,
0.236328125,
0.5205078125,
0.2340087890625
],
[
0.68701171875,
0.223876953125,
0.76904296875,
0.223876953125,
0.76904296875,
0.235107421875,
0.68701171875,
0.235107421875
],
[
0.78515625,
0.223876953125,
0.8271484375,
0.223876953125,
0.8271484375,
0.235107421875,
0.78515625,
0.235107421875
],
[
0.8369140625,
0.221435546875,
0.880859375,
0.2244873046875,
0.87890625,
0.2369384765625,
0.83642578125,
0.2340087890625
],
[
0.411376953125,
0.232177734375,
0.456787109375,
0.2340087890625,
0.455078125,
0.2493896484375,
0.41064453125,
0.2476806640625
],
[
0.46337890625,
0.232177734375,
0.51123046875,
0.2340087890625,
0.509765625,
0.248779296875,
0.462646484375,
0.2464599609375
],
[
0.62646484375,
0.232177734375,
0.67333984375,
0.2340087890625,
0.67236328125,
0.248779296875,
0.62451171875,
0.2464599609375
],
[
0.68408203125,
0.232177734375,
0.771484375,
0.2340087890625,
0.771484375,
0.248779296875,
0.68408203125,
0.2464599609375
],
[
0.35693359375,
0.2327880859375,
0.39794921875,
0.235107421875,
0.397216796875,
0.2476806640625,
0.355224609375,
0.2452392578125
],
[
0.246826171875,
0.2327880859375,
0.291259765625,
0.235107421875,
0.28955078125,
0.2476806640625,
0.2451171875,
0.2452392578125
],
[
0.5751953125,
0.2327880859375,
0.61865234375,
0.236328125,
0.6171875,
0.248779296875,
0.57421875,
0.2452392578125
],
[
0.78662109375,
0.2327880859375,
0.8271484375,
0.235107421875,
0.8251953125,
0.2476806640625,
0.78515625,
0.2452392578125
],
[
0.8388671875,
0.2327880859375,
0.8818359375,
0.236328125,
0.880859375,
0.248779296875,
0.8369140625,
0.2452392578125
],
[
0.3046875,
0.235107421875,
0.344970703125,
0.235107421875,
0.344970703125,
0.2464599609375,
0.3046875,
0.2464599609375
],
[
0.1427001953125,
0.235107421875,
0.2376708984375,
0.235107421875,
0.2376708984375,
0.2464599609375,
0.1427001953125,
0.2464599609375
],
[
0.5205078125,
0.235107421875,
0.5615234375,
0.235107421875,
0.5615234375,
0.2464599609375,
0.5205078125,
0.2464599609375
],
[
0.2476806640625,
0.2423095703125,
0.291259765625,
0.2476806640625,
0.2880859375,
0.26123046875,
0.2451171875,
0.255859375
],
[
0.353515625,
0.2423095703125,
0.403076171875,
0.2464599609375,
0.399658203125,
0.260009765625,
0.35107421875,
0.255859375
],
[
0.465087890625,
0.2423095703125,
0.51318359375,
0.248779296875,
0.509765625,
0.26123046875,
0.462646484375,
0.254638671875
],
[
0.51904296875,
0.2423095703125,
0.56494140625,
0.24462890625,
0.5634765625,
0.2587890625,
0.51708984375,
0.257080078125
],
[
0.8369140625,
0.2410888671875,
0.88671875,
0.2476806640625,
0.8818359375,
0.262451171875,
0.8330078125,
0.255859375
],
[
0.41064453125,
0.243408203125,
0.45751953125,
0.2452392578125,
0.456787109375,
0.260009765625,
0.408935546875,
0.25830078125
],
[
0.14111328125,
0.243408203125,
0.17626953125,
0.2452392578125,
0.175537109375,
0.25830078125,
0.139404296875,
0.255859375
],
[
0.57275390625,
0.243408203125,
0.62060546875,
0.2452392578125,
0.61865234375,
0.260009765625,
0.57080078125,
0.25830078125
],
[
0.63037109375,
0.243408203125,
0.6708984375,
0.2452392578125,
0.6689453125,
0.260009765625,
0.62890625,
0.25830078125
],
[
0.67919921875,
0.243408203125,
0.77490234375,
0.2452392578125,
0.77490234375,
0.260009765625,
0.67919921875,
0.25830078125
],
[
0.7841796875,
0.243408203125,
0.82861328125,
0.2464599609375,
0.8271484375,
0.2587890625,
0.78271484375,
0.255859375
],
[
0.139404296875,
0.2587890625,
0.17041015625,
0.26123046875,
0.1695556640625,
0.273681640625,
0.1376953125,
0.271484375
],
[
0.3046875,
0.24462890625,
0.343505859375,
0.2464599609375,
0.341796875,
0.25830078125,
0.30322265625,
0.255859375
],
[
0.2451171875,
0.260009765625,
0.29296875,
0.260009765625,
0.29296875,
0.27490234375,
0.2451171875,
0.27490234375
],
[
0.30322265625,
0.260009765625,
0.343505859375,
0.262451171875,
0.341796875,
0.273681640625,
0.301513671875,
0.271484375
],
[
0.355224609375,
0.260009765625,
0.4013671875,
0.262451171875,
0.399658203125,
0.273681640625,
0.353515625,
0.271484375
],
[
0.5224609375,
0.260009765625,
0.5634765625,
0.262451171875,
0.5615234375,
0.273681640625,
0.5205078125,
0.271484375
],
[
0.57666015625,
0.260009765625,
0.6171875,
0.262451171875,
0.615234375,
0.273681640625,
0.5751953125,
0.271484375
],
[
0.68603515625,
0.260009765625,
0.76904296875,
0.262451171875,
0.76904296875,
0.273681640625,
0.68603515625,
0.271484375
],
[
0.78515625,
0.2587890625,
0.8271484375,
0.262451171875,
0.8251953125,
0.27490234375,
0.7841796875,
0.271484375
],
[
0.84033203125,
0.260009765625,
0.8818359375,
0.262451171875,
0.880859375,
0.273681640625,
0.8388671875,
0.271484375
],
[
0.468505859375,
0.26123046875,
0.50732421875,
0.26123046875,
0.50732421875,
0.273681640625,
0.468505859375,
0.273681640625
],
[
0.4130859375,
0.26123046875,
0.455078125,
0.26123046875,
0.455078125,
0.27490234375,
0.4130859375,
0.27490234375
],
[
0.63232421875,
0.26123046875,
0.6708984375,
0.26123046875,
0.6708984375,
0.273681640625,
0.63232421875,
0.273681640625
],
[
0.57421875,
0.268310546875,
0.62060546875,
0.2724609375,
0.6171875,
0.2880859375,
0.57080078125,
0.283935546875
],
[
0.62890625,
0.268310546875,
0.67529296875,
0.2724609375,
0.67236328125,
0.287353515625,
0.62646484375,
0.283203125
],
[
0.411376953125,
0.26953125,
0.455078125,
0.271484375,
0.453369140625,
0.287353515625,
0.41064453125,
0.28515625
],
[
0.14111328125,
0.271484375,
0.175537109375,
0.271484375,
0.175537109375,
0.283935546875,
0.14111328125,
0.283935546875
],
[
0.2451171875,
0.271484375,
0.291259765625,
0.271484375,
0.291259765625,
0.2861328125,
0.2451171875,
0.2861328125
],
[
0.30322265625,
0.271484375,
0.344970703125,
0.273681640625,
0.343505859375,
0.2861328125,
0.301513671875,
0.283935546875
],
[
0.355224609375,
0.271484375,
0.4013671875,
0.271484375,
0.4013671875,
0.2861328125,
0.355224609375,
0.2861328125
],
[
0.465087890625,
0.271484375,
0.509765625,
0.271484375,
0.509765625,
0.2861328125,
0.465087890625,
0.2861328125
],
[
0.5224609375,
0.271484375,
0.5634765625,
0.273681640625,
0.5615234375,
0.28515625,
0.5205078125,
0.283203125
],
[
0.68603515625,
0.271484375,
0.767578125,
0.2724609375,
0.767578125,
0.283935546875,
0.68603515625,
0.283203125
],
[
0.78515625,
0.271484375,
0.8271484375,
0.273681640625,
0.8251953125,
0.28515625,
0.7841796875,
0.283203125
],
[
0.8388671875,
0.271484375,
0.880859375,
0.273681640625,
0.87890625,
0.28515625,
0.8369140625,
0.283203125
],
[
0.35693359375,
0.279052734375,
0.4013671875,
0.28515625,
0.397216796875,
0.2998046875,
0.3525390625,
0.29345703125
],
[
0.4130859375,
0.279052734375,
0.456787109375,
0.283935546875,
0.453369140625,
0.2998046875,
0.408935546875,
0.29443359375
],
[
0.465087890625,
0.28076171875,
0.509765625,
0.283203125,
0.5087890625,
0.297607421875,
0.46337890625,
0.295654296875
],
[
0.5205078125,
0.28076171875,
0.56494140625,
0.283935546875,
0.5634765625,
0.29638671875,
0.51904296875,
0.29345703125
],
[
0.57421875,
0.27978515625,
0.62060546875,
0.28515625,
0.6171875,
0.2998046875,
0.57080078125,
0.29443359375
],
[
0.62890625,
0.28076171875,
0.67333984375,
0.283203125,
0.67236328125,
0.298583984375,
0.62744140625,
0.29638671875
],
[
0.3046875,
0.281982421875,
0.343505859375,
0.283935546875,
0.341796875,
0.29638671875,
0.30322265625,
0.29443359375
],
[
0.2476806640625,
0.281982421875,
0.291259765625,
0.283935546875,
0.28955078125,
0.29638671875,
0.246826171875,
0.29443359375
],
[
0.8388671875,
0.281982421875,
0.8818359375,
0.28515625,
0.880859375,
0.29638671875,
0.8369140625,
0.29345703125
],
[
0.68603515625,
0.283203125,
0.76904296875,
0.283935546875,
0.76904296875,
0.295654296875,
0.68603515625,
0.29443359375
],
[
0.139404296875,
0.283203125,
0.187255859375,
0.283203125,
0.187255859375,
0.29443359375,
0.139404296875,
0.29443359375
],
[
0.78515625,
0.283203125,
0.8271484375,
0.28515625,
0.8251953125,
0.29638671875,
0.7841796875,
0.29443359375
],
[
0.35693359375,
0.289306640625,
0.403076171875,
0.295654296875,
0.39794921875,
0.311279296875,
0.3525390625,
0.3046875
],
[
0.411376953125,
0.290283203125,
0.456787109375,
0.295654296875,
0.453369140625,
0.31005859375,
0.408935546875,
0.3046875
],
[
0.1175537109375,
0.283935546875,
0.145263671875,
0.283935546875,
0.145263671875,
0.295654296875,
0.1175537109375,
0.295654296875
],
[
0.2476806640625,
0.292236328125,
0.28955078125,
0.29443359375,
0.2880859375,
0.30712890625,
0.246826171875,
0.3046875
],
[
0.62744140625,
0.290283203125,
0.67333984375,
0.295654296875,
0.6708984375,
0.31005859375,
0.62451171875,
0.3046875
],
[
0.8388671875,
0.292236328125,
0.880859375,
0.295654296875,
0.87890625,
0.30810546875,
0.8369140625,
0.3046875
],
[
0.5224609375,
0.29345703125,
0.5634765625,
0.295654296875,
0.5615234375,
0.30712890625,
0.5205078125,
0.3046875
],
[
0.466796875,
0.29345703125,
0.5087890625,
0.295654296875,
0.50732421875,
0.30712890625,
0.465087890625,
0.3046875
],
[
0.5751953125,
0.29345703125,
0.6171875,
0.295654296875,
0.615234375,
0.30712890625,
0.57421875,
0.3046875
],
[
0.68603515625,
0.29345703125,
0.76904296875,
0.295654296875,
0.76904296875,
0.30712890625,
0.68603515625,
0.3046875
],
[
0.78515625,
0.29345703125,
0.8271484375,
0.295654296875,
0.8251953125,
0.30712890625,
0.7841796875,
0.3046875
],
[
0.30322265625,
0.29443359375,
0.344970703125,
0.29443359375,
0.344970703125,
0.30810546875,
0.30322265625,
0.30810546875
],
[
0.1427001953125,
0.29443359375,
0.20654296875,
0.29443359375,
0.20654296875,
0.305908203125,
0.1427001953125,
0.305908203125
],
[
0.465087890625,
0.302734375,
0.51318359375,
0.30712890625,
0.509765625,
0.320556640625,
0.462646484375,
0.317626953125
],
[
0.57275390625,
0.302734375,
0.62109375,
0.30712890625,
0.61865234375,
0.32177734375,
0.5693359375,
0.317626953125
],
[
0.62646484375,
0.302734375,
0.67333984375,
0.3046875,
0.67236328125,
0.319580078125,
0.62451171875,
0.317626953125
],
[
0.68408203125,
0.302734375,
0.76904296875,
0.3046875,
0.76904296875,
0.320556640625,
0.68408203125,
0.318359375
],
[
0.4130859375,
0.303955078125,
0.453369140625,
0.30712890625,
0.45166015625,
0.319580078125,
0.41064453125,
0.3154296875
],
[
0.355224609375,
0.303955078125,
0.399658203125,
0.30712890625,
0.39794921875,
0.319580078125,
0.353515625,
0.31640625
],
[
0.5205078125,
0.303955078125,
0.56494140625,
0.30712890625,
0.5634765625,
0.319580078125,
0.51904296875,
0.31640625
],
[
0.8388671875,
0.303955078125,
0.880859375,
0.30712890625,
0.87890625,
0.319580078125,
0.8369140625,
0.31640625
],
[
0.30322265625,
0.3046875,
0.343505859375,
0.30712890625,
0.341796875,
0.319580078125,
0.301513671875,
0.317626953125
],
[
0.246826171875,
0.3046875,
0.28955078125,
0.30712890625,
0.2880859375,
0.319580078125,
0.2451171875,
0.317626953125
],
[
0.7841796875,
0.3046875,
0.8271484375,
0.30712890625,
0.8251953125,
0.318359375,
0.78271484375,
0.31640625
],
[
0.14111328125,
0.3154296875,
0.17626953125,
0.317626953125,
0.175537109375,
0.330078125,
0.139404296875,
0.327880859375
],
[
0.1427001953125,
0.305908203125,
0.2376708984375,
0.305908203125,
0.2376708984375,
0.317626953125,
0.1427001953125,
0.317626953125
],
[
0.301513671875,
0.3154296875,
0.345947265625,
0.317626953125,
0.344970703125,
0.330078125,
0.300537109375,
0.327880859375
],
[
0.355224609375,
0.3154296875,
0.4013671875,
0.317626953125,
0.399658203125,
0.330078125,
0.353515625,
0.327880859375
],
[
0.411376953125,
0.3154296875,
0.455078125,
0.317626953125,
0.453369140625,
0.330078125,
0.41064453125,
0.327880859375
],
[
0.466796875,
0.3154296875,
0.5087890625,
0.317626953125,
0.50732421875,
0.330078125,
0.465087890625,
0.327880859375
],
[
0.51904296875,
0.3154296875,
0.5634765625,
0.317626953125,
0.5615234375,
0.330078125,
0.51708984375,
0.327880859375
],
[
0.57421875,
0.3154296875,
0.62060546875,
0.3154296875,
0.62060546875,
0.330078125,
0.57421875,
0.330078125
],
[
0.62890625,
0.3154296875,
0.67333984375,
0.3154296875,
0.67333984375,
0.330078125,
0.62890625,
0.330078125
],
[
0.8369140625,
0.3154296875,
0.88330078125,
0.318359375,
0.8818359375,
0.330810546875,
0.83642578125,
0.327880859375
],
[
0.68408203125,
0.317626953125,
0.771484375,
0.317626953125,
0.771484375,
0.328857421875,
0.68408203125,
0.328857421875
],
[
0.246826171875,
0.317626953125,
0.291259765625,
0.317626953125,
0.291259765625,
0.328857421875,
0.246826171875,
0.328857421875
],
[
0.7841796875,
0.317626953125,
0.8271484375,
0.317626953125,
0.8271484375,
0.328857421875,
0.7841796875,
0.328857421875
],
[
0.11419677734375,
0.341552734375,
0.8818359375,
0.342529296875,
0.8818359375,
0.35693359375,
0.11419677734375,
0.355712890625
],
[
0.1158447265625,
0.35693359375,
0.8818359375,
0.35693359375,
0.8818359375,
0.37060546875,
0.1158447265625,
0.37060546875
],
[
0.1158447265625,
0.37060546875,
0.80029296875,
0.37060546875,
0.80029296875,
0.38427734375,
0.1158447265625,
0.38427734375
],
[
0.1175537109375,
0.412109375,
0.48779296875,
0.412109375,
0.48779296875,
0.42578125,
0.1175537109375,
0.42578125
],
[
0.509765625,
0.412109375,
0.568359375,
0.412109375,
0.568359375,
0.427001953125,
0.509765625,
0.427001953125
],
[
0.63037109375,
0.410400390625,
0.71533203125,
0.412109375,
0.71533203125,
0.427001953125,
0.63037109375,
0.424560546875
],
[
0.72119140625,
0.412109375,
0.7841796875,
0.412109375,
0.7841796875,
0.427001953125,
0.72119140625,
0.427001953125
],
[
0.791015625,
0.412109375,
0.8818359375,
0.412109375,
0.8818359375,
0.427001953125,
0.791015625,
0.427001953125
],
[
0.1158447265625,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.441162109375,
0.1158447265625,
0.441162109375
],
[
0.5087890625,
0.429931640625,
0.5439453125,
0.429931640625,
0.5439453125,
0.4453125,
0.5087890625,
0.4453125
],
[
0.6630859375,
0.429931640625,
0.6826171875,
0.429931640625,
0.6826171875,
0.44482421875,
0.6630859375,
0.44482421875
],
[
0.744140625,
0.429931640625,
0.76171875,
0.429931640625,
0.76171875,
0.44482421875,
0.744140625,
0.44482421875
],
[
0.82861328125,
0.429931640625,
0.84814453125,
0.429931640625,
0.84814453125,
0.44482421875,
0.82861328125,
0.44482421875
],
[
0.1175537109375,
0.44482421875,
0.48779296875,
0.44482421875,
0.48779296875,
0.457763671875,
0.1175537109375,
0.457763671875
],
[
0.509765625,
0.44482421875,
0.5458984375,
0.44482421875,
0.5458984375,
0.457275390625,
0.509765625,
0.457275390625
],
[
0.6484375,
0.443603515625,
0.69775390625,
0.443603515625,
0.69775390625,
0.457763671875,
0.6484375,
0.457763671875
],
[
0.73291015625,
0.44482421875,
0.77490234375,
0.44482421875,
0.77490234375,
0.457763671875,
0.73291015625,
0.457763671875
],
[
0.82373046875,
0.44482421875,
0.8505859375,
0.44482421875,
0.8505859375,
0.457763671875,
0.82373046875,
0.457763671875
],
[
0.5087890625,
0.457275390625,
0.5634765625,
0.457275390625,
0.5634765625,
0.471435546875,
0.5087890625,
0.471435546875
],
[
0.64990234375,
0.457275390625,
0.6943359375,
0.457275390625,
0.6943359375,
0.471435546875,
0.64990234375,
0.471435546875
],
[
0.73291015625,
0.457763671875,
0.77490234375,
0.457763671875,
0.77490234375,
0.471435546875,
0.73291015625,
0.471435546875
],
[
0.82861328125,
0.457275390625,
0.85205078125,
0.457275390625,
0.85205078125,
0.47021484375,
0.82861328125,
0.47021484375
],
[
0.73046875,
0.468505859375,
0.77734375,
0.47021484375,
0.77685546875,
0.486328125,
0.72900390625,
0.48388671875
],
[
0.1158447265625,
0.460205078125,
0.489501953125,
0.460205078125,
0.489501953125,
0.473876953125,
0.1158447265625,
0.473876953125
],
[
0.51318359375,
0.471435546875,
0.58447265625,
0.471435546875,
0.58447265625,
0.482666015625,
0.51318359375,
0.482666015625
],
[
0.64990234375,
0.471435546875,
0.6962890625,
0.471435546875,
0.6962890625,
0.482666015625,
0.64990234375,
0.482666015625
],
[
0.81787109375,
0.471435546875,
0.85986328125,
0.471435546875,
0.85986328125,
0.485107421875,
0.81787109375,
0.485107421875
],
[
0.5087890625,
0.482177734375,
0.62109375,
0.48388671875,
0.62109375,
0.498779296875,
0.5087890625,
0.496337890625
],
[
0.1158447265625,
0.474365234375,
0.489501953125,
0.4755859375,
0.489501953125,
0.490478515625,
0.1158447265625,
0.4892578125
],
[
0.6572265625,
0.48388671875,
0.6884765625,
0.48388671875,
0.6884765625,
0.496337890625,
0.6572265625,
0.496337890625
],
[
0.7421875,
0.48388671875,
0.765625,
0.48388671875,
0.765625,
0.49755859375,
0.7421875,
0.49755859375
],
[
0.82373046875,
0.48388671875,
0.85205078125,
0.48388671875,
0.85205078125,
0.49755859375,
0.82373046875,
0.49755859375
],
[
0.1158447265625,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.505859375,
0.1158447265625,
0.505859375
],
[
0.5087890625,
0.496337890625,
0.55322265625,
0.496337890625,
0.55322265625,
0.51123046875,
0.5087890625,
0.51123046875
],
[
0.6630859375,
0.496337890625,
0.68408203125,
0.496337890625,
0.68408203125,
0.51123046875,
0.6630859375,
0.51123046875
],
[
0.74658203125,
0.49755859375,
0.759765625,
0.49755859375,
0.759765625,
0.51025390625,
0.74658203125,
0.51025390625
],
[
0.82373046875,
0.493408203125,
0.8564453125,
0.49755859375,
0.85205078125,
0.51318359375,
0.8193359375,
0.51025390625
],
[
0.1175537109375,
0.51025390625,
0.22412109375,
0.51025390625,
0.22412109375,
0.521484375,
0.1175537109375,
0.521484375
],
[
0.509765625,
0.52490234375,
0.8818359375,
0.52490234375,
0.8818359375,
0.5380859375,
0.509765625,
0.5380859375
],
[
0.118408203125,
0.54052734375,
0.156982421875,
0.54052734375,
0.156982421875,
0.552734375,
0.118408203125,
0.552734375
],
[
0.154541015625,
0.54052734375,
0.264404296875,
0.54052734375,
0.264404296875,
0.5517578125,
0.154541015625,
0.5517578125
],
[
0.51123046875,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.552734375,
0.51123046875,
0.552734375
],
[
0.51123046875,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.51123046875,
0.56787109375
],
[
0.1175537109375,
0.5634765625,
0.489501953125,
0.5634765625,
0.489501953125,
0.57666015625,
0.1175537109375,
0.57666015625
],
[
0.509765625,
0.56640625,
0.77490234375,
0.56640625,
0.77490234375,
0.580078125,
0.509765625,
0.580078125
],
[
0.1175537109375,
0.5791015625,
0.453369140625,
0.5791015625,
0.453369140625,
0.5927734375,
0.1175537109375,
0.5927734375
],
[
0.1351318359375,
0.5966796875,
0.490234375,
0.5966796875,
0.490234375,
0.60986328125,
0.1351318359375,
0.60986328125
],
[
0.1158447265625,
0.61328125,
0.489501953125,
0.61328125,
0.489501953125,
0.626953125,
0.1158447265625,
0.626953125
],
[
0.51123046875,
0.61083984375,
0.88330078125,
0.61083984375,
0.88330078125,
0.62451171875,
0.51123046875,
0.62451171875
],
[
0.1175537109375,
0.62890625,
0.489501953125,
0.62890625,
0.489501953125,
0.6435546875,
0.1175537109375,
0.6435546875
],
[
0.509765625,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.509765625,
0.6396484375
],
[
0.1158447265625,
0.6435546875,
0.489501953125,
0.64453125,
0.489501953125,
0.6591796875,
0.1158447265625,
0.65771484375
],
[
0.51123046875,
0.6435546875,
0.880859375,
0.6435546875,
0.880859375,
0.65673828125,
0.51123046875,
0.65673828125
],
[
0.1175537109375,
0.6611328125,
0.48779296875,
0.6611328125,
0.48779296875,
0.6748046875,
0.1175537109375,
0.6748046875
],
[
0.509765625,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.509765625,
0.6728515625
],
[
0.1175537109375,
0.6767578125,
0.48779296875,
0.6767578125,
0.48779296875,
0.6904296875,
0.1175537109375,
0.6904296875
],
[
0.51123046875,
0.67578125,
0.88330078125,
0.67578125,
0.88330078125,
0.689453125,
0.51123046875,
0.689453125
],
[
0.1158447265625,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.1158447265625,
0.70703125
],
[
0.5087890625,
0.6904296875,
0.8818359375,
0.6904296875,
0.8818359375,
0.70361328125,
0.5087890625,
0.70361328125
],
[
0.1175537109375,
0.708984375,
0.489501953125,
0.708984375,
0.489501953125,
0.72265625,
0.1175537109375,
0.72265625
],
[
0.5087890625,
0.7060546875,
0.8818359375,
0.7060546875,
0.8818359375,
0.7197265625,
0.5087890625,
0.7197265625
],
[
0.1175537109375,
0.72705078125,
0.490234375,
0.72705078125,
0.490234375,
0.73681640625,
0.1175537109375,
0.73681640625
],
[
0.509765625,
0.7236328125,
0.8818359375,
0.7236328125,
0.8818359375,
0.73681640625,
0.509765625,
0.73681640625
],
[
0.1175537109375,
0.7421875,
0.490234375,
0.7421875,
0.490234375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.509765625,
0.73828125,
0.8818359375,
0.7392578125,
0.8818359375,
0.75439453125,
0.509765625,
0.7529296875
],
[
0.1175537109375,
0.75830078125,
0.489501953125,
0.75830078125,
0.489501953125,
0.771484375,
0.1175537109375,
0.771484375
],
[
0.51123046875,
0.75732421875,
0.880859375,
0.75732421875,
0.880859375,
0.76708984375,
0.51123046875,
0.76708984375
],
[
0.1175537109375,
0.77392578125,
0.48779296875,
0.77392578125,
0.48779296875,
0.78759765625,
0.1175537109375,
0.78759765625
],
[
0.509765625,
0.771484375,
0.88330078125,
0.771484375,
0.88330078125,
0.78515625,
0.509765625,
0.78515625
],
[
0.1158447265625,
0.7880859375,
0.287109375,
0.7890625,
0.287109375,
0.80419921875,
0.1158447265625,
0.802734375
],
[
0.51123046875,
0.7880859375,
0.8818359375,
0.7880859375,
0.8818359375,
0.8017578125,
0.51123046875,
0.8017578125
],
[
0.133544921875,
0.80712890625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82080078125,
0.133544921875,
0.82080078125
],
[
0.51123046875,
0.80419921875,
0.88330078125,
0.80419921875,
0.88330078125,
0.81787109375,
0.51123046875,
0.81787109375
],
[
0.1175537109375,
0.822265625,
0.48779296875,
0.822265625,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83740234375
],
[
0.5087890625,
0.8193359375,
0.8818359375,
0.818359375,
0.8818359375,
0.8330078125,
0.5087890625,
0.83447265625
],
[
0.1175537109375,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.8525390625,
0.1175537109375,
0.8525390625
],
[
0.51123046875,
0.83447265625,
0.63818359375,
0.8349609375,
0.63818359375,
0.849609375,
0.51123046875,
0.8486328125
],
[
0.11419677734375,
0.85400390625,
0.489501953125,
0.85498046875,
0.489501953125,
0.869140625,
0.11419677734375,
0.8681640625
],
[
0.51123046875,
0.8642578125,
0.736328125,
0.865234375,
0.736328125,
0.8798828125,
0.51123046875,
0.87890625
],
[
0.1158447265625,
0.87158203125,
0.491943359375,
0.87158203125,
0.491943359375,
0.88525390625,
0.1158447265625,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.20068359375,
0.8896484375,
0.20068359375,
0.90087890625,
0.118408203125,
0.90087890625
],
[
0.51123046875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1351318359375,
0.90478515625,
0.490234375,
0.90478515625,
0.490234375,
0.91845703125,
0.1351318359375,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"DB",
"Yahoo",
"Avg.",
"SST2",
"nMethod",
"IMDB",
"Yelp",
"AG",
"RTE",
"SNLI",
"MNLI-m/mm",
"NERD",
"68.4",
"40.7",
"61.4",
"52.4",
"Prompt",
"83.3",
"89.4",
"87.1",
"80.9",
"49.9",
"50.8/51.7",
"21.1",
"66.74.8",
"71.72.6",
"53.60.5/53.90.8",
"59.44.9",
"81.110.6",
"81.53.7",
"65.611.4",
"53.26.2",
"45.04.7",
"46.15.3",
"34.73.3",
"59.11.7",
"ICL",
"83.41.3",
"89.00.1",
"75.40.8",
"31.22.7",
"38.50.8",
"43.42.5/42.93.3",
"56.31.5",
"89.70.1",
"BBT",
"52.31.4",
"14.12.3",
"89.00.2",
"74.33.2",
"74.25.2",
"34.03.5",
"38.64.0",
"59.03.0",
"83.32.5",
"BBTv2",
"89.90.2",
"48.25.7",
"44.23.2/44.34.5",
"29.00.8",
"RLPrompt",
"63.56.3",
"65.06.5",
"66.36.9",
"72.54.5",
"65.65.5",
"38.15.8",
"36.53.0",
"50.64.7",
"40.32.0/41.02.1",
"53.85.3",
"14.51.8",
"PromptBoosting",
"23.77.0",
"43.56.1",
"82.46.1",
"86.72.6",
"88.72.5",
"73.04.8",
"50.05.9",
"22.00.8",
"54.74.6",
"36.81.6/36.32.3",
"58.711.8",
"79.91.1",
"66.91.6",
"94.80.1",
"91.20.3",
"78.80.9",
"56.02.7",
"47.74.1",
"55.20.8",
"DecT",
"90.80.2",
"52.22.7/53.33.0",
"35.71.5",
"ICL",
"80.46.6",
"49.99.9",
"77.414.6",
"60.39.8",
"42.73.9",
"31.74.8",
"55.67.0",
"71.76.5",
"65.15.4",
"42.13.2",
"44.75.9/45.26.0",
"67.73.5",
"42.90.6",
"40.51.3",
"41.21.7/40.72.0",
"19.41.5",
"84.51.2",
"BBT",
"89.80.9",
"90.20.6",
"48.44.0",
"58.61.8",
"86.62.2",
"79.02.1",
"90.30.5",
"62.61.7",
"79.12.1",
"89.40.6",
"89.01.7",
"46.01.4",
"40.84.3",
"31.91.4",
"46.22.3",
"BBTv2",
"44.00.9/44.81.6",
"75.810.1",
"53.52.9",
"36.32.6",
"80.775",
"RLPrompt",
"78.87.3",
"76.14.8",
"76.35.9",
"45.03.1",
"44.42.9/45.53.8",
"16.72.4",
"57.44.8",
"78.26.8",
"90.10.7",
"53.43.4",
"39.84.5/40.35.7",
"92.32.1",
"88.92.3",
"53.55.9",
"40.92.5",
"63.44.0",
"83.05.2",
"PromptBoosting",
"36.45.1",
"87.61.6",
"94.80.7",
"89.10.6",
"59.92.1",
"69.51.9",
"81.92.6",
"DecT",
"56.72.7",
"53.22.9",
"52.22.3/53.42.4",
"46.71.7",
"ICL",
"89.60.9",
"71.515.8",
"80.66.0",
"73.714.5",
"52.65.7",
"43.87.0",
"51.43.0/52.13.3",
"35.12.6",
"58.17.2",
"71.89.1",
"64.46.0",
"42.06.3",
"52.62.2",
"46.64.3",
"81.50.8",
"BBT",
"89.60.3",
"89.30.4",
"91.50.2",
"87.85.0",
"48.31.4",
"40.02.6/39.92.9",
"17.81.4",
"62.31.5",
"92.90.6",
"85.30.5",
"93.68.7",
"52.01.4",
"56.73.3",
"57.39.3",
"88.62.1",
"90.32.7",
"68.31.7",
"50.12.4/51.73.2",
"BBTv2",
"33.31.0",
"95.11.0",
"80.26.7",
"16",
"87.02.6",
"41.15.0",
"61.82.7",
"48.12.2",
"80.83.3",
"54.32.8",
"43.33.9/44.34.5",
"17.51.4",
"87.62.4",
"RLPrompt",
"95.00.5",
"60.0g.5",
"61.39.5",
"52.54.s/50.45.1",
"85.2d.9",
"94.71.0",
"46.62.4",
"70.12.6",
"86.23.1",
"87.69.0",
"52.12.6",
"DecT",
"PromptBoosting",
"91.00.9",
"95.40.3",
"86.40.4",
"94.60.5",
"64.20.7",
"59.71.8",
"60.50.8",
"73.51.0",
"55.31.3/56.81.5",
"91.00.5",
"53.51.8",
"Table 1: Experiment results for MaaS adaptation methods. Some baseline results are taken from corresponding",
"papers (aSun et al. (2022a), bSun et al. (2022b), cDeng et al. (2022), dHou et al. (2022). We run other experiments",
"over 5 random seeds and report average accuracy and standard deviation (%). Best results are in bold.",
"selection of 入, we directly set > = 1/n for most",
"Method",
"Tr. Time (s)",
"#Query",
"# Param. (K)",
"datasets based on the intuition that ^ should de-",
"ICL",
"0",
"0",
"0",
"crease as the amount of training data increases. On",
"BBT",
"10,512",
"8,000",
"0.5",
"BBTv2",
"9,856",
"8,000",
"12",
"12,000",
"MNLI and FewNERD, we tune ^ on the valida-",
"RLPrompt",
"65,579",
"3,100",
"PromptBoosting",
"tion set and select 入 = 1 and 入 = 1/16 respec-",
"644",
"10",
"0.4",
"tively. We give the templates and label words in",
"DecT",
"3",
"1",
"130",
"Appendix A.3",
"Table 2: Efficiency comparison of MaaS adaptation",
"4.2",
"MainResults",
"methods. Training time is the average wall clock time",
"measured in the 16-shot setting. \"Tr\" stands for Train-",
"Table I presents the main few-shot learning results.",
"ing and Param.\" stands for Parameter.",
"From the results, we have these observations:",
"Overall, DecT outperforms the state-of-the-",
"art baseline methods by a large margin (more",
"zero-shot prompt and ICL are strong baselines.",
"than 3% on average), especially under extreme",
"However, as shown in the table, ICL gives the",
"data scarcity, showing its superior performance.",
"best results in the 1-shot setting, and it can hardly",
"Across different tasks, DecT and baselines obtain",
" improve with more training data due to the input",
"similar results on some easy sentiment analysis",
"length restriction. To compare, merely optimiz-",
"and topic classification tasks, but we highlight that",
"ing the input prompts (BBT and RLPrompt) can",
"DecT is much more favorable on difficult datasets,",
" hardly outperform them, showing the limitation of",
"suchasYahooandFewNERD.Whileotherbase-",
"input-sideprompt optimization. In contrast, two",
"line methods struggle to optimize well, DecT sur-",
"other baselines, BBTv2 and PromptBoosting, are",
"passes them significantly (about 10% on Yahoo and",
"morepowerfulbecausetheyeitherinsertsaddi",
"20% on FewNERD under 16-shot setting compared",
"tional learnable prompt tokens inside the PTM or",
"with BBTv2 and ICL).",
"ensembles the outputs of different prompts. With",
"On stability, DecT also has consistently low",
"the superior results of DecT, we argue that output-",
"variance and some baselines (ICL, RLPrompt and",
"side optimization is a promising way for Maas",
"PromptBoosting) are unstable. Given the difficulty",
"PTM adaptation.",
"of few-shot PTM adaptation, it is of great signifi-",
"4.3 Efficiency Comparison",
"cance that the adaptation method is robust to ran-",
"dom seeds.",
"Despite the superior performance, another major",
"On baselines, optimization-free methods, i.e.",
"advantage of DecT is its high efficiency. In Fig"
] |
{
"arxiv_id": "2212.08408",
"language": "en",
"page_idx": 4
}
|
||
2108.06417-en-1
|
2108.06417-en-1.jpg
|
Our method sets a new state of the art on the task of monocular 3D detection on KITTI-3D[14] and nuScenes [5] with significant improvements compared to previous state-of-the-art methods. The simplicity of its training procedure and its end-to-end optimization allows for effective use of large-scale depth data, leading to impressive multi-class detection accuracy.
## $2. \textbf{Related work}$
Monocular 3D detection. A large body of work in imagebased 3D detection builds upon 2D detectors, and aims to lift them to 3D using various cues from object shapes and scene geometry. These priors are often injected by aligning 2D keypoints with their projective 3D counterparts [6,1,2,21,52], or by learning a low-dimensional shape representation [28, 43]. Other methods try to leverage geometric consistency between 2D and 3D structures, formulating the inference as a constrained optimization problem [45,38,46,12,44,24,27,34]. These methods commonly use additional data (e.g. 3D CAD models, instance segmentation), or assume rigidity of objects. Inspired by Lidarbased methods [8,70, 30], another line of work employs view transformation (i.e. birds-eye-view) to overcome the limitations of perspective range view,e.g. occlusion or variability in size [25, 56, 60]. These methods often require precise camera extrinsics, or are accurate only in close vicinity. End-to-end 3D detectors. Alternatively, researchers have attempted to directly regress 3D bounding boxes from CNN features [59, 39, 76, 23, 3, 9]. Typically, these approaches extend standard 2D detectors (single-stage [54,76] or twostage [55]) by adding heads that predict various 3D cuboid parameterizations . The use of depth-aware convolution or dense 3D anchors [10, 53] enabled higher accuracy. Incorporating uncertainty in estimating the 3D box (or its depth) has also been shown to greatly improve detection accuracy [59,9,58]. In [59], the authors proposed a disentangled loss for 3D box regression that helps stabilize training. DD3D also falls in the end-to-end 3D detector category, however, our emphasis is on learning a good depth representation via large-scale self-supervised pre-training on raw data, which leads to robust 3D detection.
Pseudo-Lidar methods. Starting with the pioneering work of[67], these methods leverage advances in monocular depth estimation and train Lidar-based detectors on the resulting pseudo pointcloud, producing impressive results [69,51,68,42]. Recent methods have improved upon [67] by correcting the monocular depth with sparse Lidar readings [71], decorating the pseudo pointcloud with colors [42], using 2D detection to segment foreground pointcloud regions [49,69,41], and structured sparsification of the monocular pointcloud[63]. Recently,[58] has shown a bias in the PL results on the representative KITTI- 3D[14] benchmark, while this paradigm remains the state-
of-the-art. Multiple researchers showed that inaccurate $depth\textit{ estimation is a major source of error in PL methods}$ [58, 66]. In this work, we build our reference PL method based on [49] and [41] to investigate the benefits of using large-scale depth data and compare it with DD3D.
Monocular depth estimation. Estimating per-pixel depth is a key task for both DD3D (as a pre-training task) and PL (as the first of a two-stage pipeline). It is itself a thriving research area: the community has pushed toward accurate dense depth prediction via supervised [35,29,19,50,26, 18,32,13,31] as well as self-supervised[48,75,62,16, 15,64] methods. We note that supervised monocular depth training used in this work require no annotations from human, allowing us to scale our methods to a large amount of raw data.
$3. \textbf{ Dense depth pre- training for 3D detection}$
Given a single image and its camera intrinsics matrix as input, the goal of monocular 3D detection is to generate a set of multi-class 3D bounding boxes relative to camera coordinates. During inference, DD3D does not require any additional data, such as per-pixel depth estimates, 2D bounding boxes, segmentations, or 3D CAD models. DD3D is also camera-aware: it scales the depth of putative 3D boxes according to the camera intrinsics.
## 3.1. Architecture
DD3D is a fully convolutional single-stage network that extends FCOS [61] to perform 3D detection and dense depth prediction. The architecture (see Figure 2) is composed of a backbone network and three sub-networks (or heads) that are shared among all multi-scale features. The backbone takes an RGB image as input, and computes convolutional features at different scales. As in [61], we adopt a feature pyramid network (FPN) [36] as the backbone
Three head networks are applied to each feature map produced by the backbone, and perform independent prediction tasks. The classification module predicts object category. It produces $C$ real values, where $C$ is the number of object categories. The 2D box module produces classagnostic bounding boxes and center-ness by predicting 4 offsets from the feature location to the sides of each bounding box and a scalar associated with center-ness. We refer the readers to [61] for more details regarding the 2D detection architecture.
3D detection head. This head predicts 3D bounding boxes and per-pixel depth. It takes FPN features as input, and applies four 2D convolutions with $3\times3$ kernels that generate 12 real values for each feature location. These are decoded into 3D bounding box, per-pixel depth map, and 3D prediction confidence, as described below:
$\bullet\mathbf{q}=(q_w,q_x,q_y,q_z)$ is the quaternion representation of
|
[
[
0.09722900390625,
0.0921630859375,
0.471435546875,
0.0921630859375,
0.471435546875,
0.1060791015625,
0.09722900390625,
0.1060791015625
],
[
0.5009765625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.5009765625,
0.1060791015625
],
[
0.07843017578125,
0.1060791015625,
0.4697265625,
0.106689453125,
0.4697265625,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.5048828125,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.080078125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.5009765625,
0.1363525390625,
0.8935546875,
0.1383056640625,
0.8935546875,
0.15283203125,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.5009765625,
0.15283203125,
0.8310546875,
0.1514892578125,
0.8310546875,
0.1654052734375,
0.5009765625,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.0767822265625,
0.1817626953125,
0.339111328125,
0.18310546875,
0.339111328125,
0.1976318359375,
0.0767822265625,
0.1962890625
],
[
0.5009765625,
0.1817626953125,
0.8896484375,
0.1817626953125,
0.8896484375,
0.1956787109375,
0.5009765625,
0.1956787109375
],
[
0.50341796875,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.080078125,
0.212158203125,
0.2132568359375,
0.212158203125,
0.2132568359375,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.5048828125,
0.2298583984375,
0.8896484375,
0.2298583984375,
0.8896484375,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.080078125,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.50341796875,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.080078125,
0.25244140625,
0.4697265625,
0.25244140625,
0.4697265625,
0.266357421875,
0.080078125,
0.266357421875
],
[
0.50341796875,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.271484375,
0.50341796875,
0.271484375
],
[
0.080078125,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.28271484375,
0.46728515625,
0.28271484375,
0.46728515625,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.50244140625,
0.2890625,
0.8935546875,
0.2890625,
0.8935546875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.07843017578125,
0.296630859375,
0.4697265625,
0.29736328125,
0.4697265625,
0.3125,
0.07843017578125,
0.311279296875
],
[
0.50341796875,
0.30615234375,
0.564453125,
0.30615234375,
0.564453125,
0.31640625,
0.50341796875,
0.31640625
],
[
0.080078125,
0.313232421875,
0.468017578125,
0.313232421875,
0.468017578125,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.07843017578125,
0.326904296875,
0.468017578125,
0.32763671875,
0.468017578125,
0.34228515625,
0.07843017578125,
0.341552734375
],
[
0.50341796875,
0.333251953125,
0.87158203125,
0.333251953125,
0.87158203125,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.08087158203125,
0.35986328125,
0.468017578125,
0.35986328125,
0.468017578125,
0.37060546875,
0.08087158203125,
0.37060546875
],
[
0.521484375,
0.35791015625,
0.89208984375,
0.35791015625,
0.89208984375,
0.371826171875,
0.521484375,
0.371826171875
],
[
0.080078125,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.50244140625,
0.373779296875,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.50244140625,
0.386962890625
],
[
0.080078125,
0.38818359375,
0.468017578125,
0.38818359375,
0.468017578125,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.50244140625,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.401611328125,
0.50244140625,
0.401611328125
],
[
0.080078125,
0.404052734375,
0.4697265625,
0.404052734375,
0.4697265625,
0.417236328125,
0.080078125,
0.417236328125
],
[
0.50244140625,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.50244140625,
0.416748046875
],
[
0.07843017578125,
0.417236328125,
0.4697265625,
0.41845703125,
0.4697265625,
0.431884765625,
0.07843017578125,
0.43115234375
],
[
0.50244140625,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.08087158203125,
0.43505859375,
0.46728515625,
0.43505859375,
0.46728515625,
0.44580078125,
0.08087158203125,
0.44580078125
],
[
0.5009765625,
0.43310546875,
0.89208984375,
0.431884765625,
0.89208984375,
0.4462890625,
0.5009765625,
0.447509765625
],
[
0.080078125,
0.448974609375,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.080078125,
0.462158203125
],
[
0.50244140625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.50244140625,
0.462158203125
],
[
0.080078125,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.080078125,
0.477783203125
],
[
0.50244140625,
0.464111328125,
0.7294921875,
0.464111328125,
0.7294921875,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.080078125,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.080078125,
0.492431640625
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50341796875,
0.488037109375,
0.6337890625,
0.488037109375,
0.6337890625,
0.501953125,
0.50341796875,
0.501953125
],
[
0.080078125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.521484375,
0.5087890625,
0.89208984375,
0.51123046875,
0.89208984375,
0.52587890625,
0.521484375,
0.5234375
],
[
0.080078125,
0.5244140625,
0.4697265625,
0.5244140625,
0.4697265625,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.50244140625,
0.5263671875,
0.89208984375,
0.5263671875,
0.89208984375,
0.54052734375,
0.50244140625,
0.54052734375
],
[
0.07843017578125,
0.53857421875,
0.468017578125,
0.5390625,
0.468017578125,
0.5537109375,
0.07843017578125,
0.55322265625
],
[
0.50244140625,
0.54150390625,
0.89208984375,
0.54150390625,
0.89208984375,
0.5546875,
0.50244140625,
0.5546875
],
[
0.080078125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.080078125,
0.568359375
],
[
0.50341796875,
0.556640625,
0.89208984375,
0.556640625,
0.89208984375,
0.57080078125,
0.50341796875,
0.57080078125
],
[
0.080078125,
0.5693359375,
0.4697265625,
0.5693359375,
0.4697265625,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50244140625,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.50244140625,
0.5859375
],
[
0.080078125,
0.58544921875,
0.468017578125,
0.58544921875,
0.468017578125,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.50244140625,
0.5869140625,
0.89208984375,
0.5869140625,
0.89208984375,
0.60107421875,
0.50244140625,
0.60107421875
],
[
0.08087158203125,
0.6015625,
0.46728515625,
0.6015625,
0.46728515625,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.50341796875,
0.6015625,
0.89208984375,
0.6015625,
0.89208984375,
0.61572265625,
0.50341796875,
0.61572265625
],
[
0.080078125,
0.6162109375,
0.465576171875,
0.6162109375,
0.465576171875,
0.626953125,
0.080078125,
0.626953125
],
[
0.50244140625,
0.6171875,
0.86376953125,
0.6171875,
0.86376953125,
0.630859375,
0.50244140625,
0.630859375
],
[
0.07843017578125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.07843017578125,
0.64306640625
],
[
0.521484375,
0.630859375,
0.89208984375,
0.6318359375,
0.89208984375,
0.646484375,
0.521484375,
0.64599609375
],
[
0.07843017578125,
0.64306640625,
0.468017578125,
0.64453125,
0.468017578125,
0.6591796875,
0.07843017578125,
0.65771484375
],
[
0.50244140625,
0.64794921875,
0.89208984375,
0.64794921875,
0.89208984375,
0.6611328125,
0.50244140625,
0.6611328125
],
[
0.07843017578125,
0.65771484375,
0.468017578125,
0.66015625,
0.468017578125,
0.6748046875,
0.07843017578125,
0.67236328125
],
[
0.50244140625,
0.662109375,
0.89208984375,
0.662109375,
0.89208984375,
0.67529296875,
0.50244140625,
0.67529296875
],
[
0.07843017578125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.07843017578125,
0.689453125
],
[
0.50244140625,
0.67822265625,
0.89208984375,
0.67822265625,
0.89208984375,
0.69140625,
0.50244140625,
0.69140625
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5009765625,
0.69140625,
0.8896484375,
0.69140625,
0.8896484375,
0.70556640625,
0.5009765625,
0.70556640625
],
[
0.080078125,
0.70556640625,
0.3505859375,
0.70556640625,
0.3505859375,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70849609375,
0.89208984375,
0.70849609375,
0.89208984375,
0.7216796875,
0.50244140625,
0.7216796875
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.50341796875,
0.724609375,
0.890625,
0.724609375,
0.890625,
0.73486328125,
0.50341796875,
0.73486328125
],
[
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,
0.751953125,
0.50244140625,
0.751953125
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.5009765625,
0.75244140625,
0.89208984375,
0.751953125,
0.89208984375,
0.7666015625,
0.5009765625,
0.767578125
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76806640625,
0.61376953125,
0.76806640625,
0.61376953125,
0.7822265625,
0.50244140625,
0.7822265625
],
[
0.080078125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.78271484375,
0.89208984375,
0.78271484375,
0.89208984375,
0.796875,
0.50244140625,
0.796875
],
[
0.07843017578125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.50244140625,
0.798828125,
0.890625,
0.798828125,
0.890625,
0.8125,
0.50244140625,
0.8125
],
[
0.07843017578125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.5009765625,
0.8125,
0.89208984375,
0.81298828125,
0.89208984375,
0.82763671875,
0.5009765625,
0.8271484375
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.5048828125,
0.830078125,
0.890625,
0.830078125,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84228515625,
0.88818359375,
0.84326171875,
0.88818359375,
0.85791015625,
0.5009765625,
0.85693359375
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.859375,
0.74169921875,
0.859375,
0.74169921875,
0.87255859375,
0.5009765625,
0.87255859375
],
[
0.080078125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.07843017578125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.52685546875,
0.88720703125,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.52685546875,
0.9013671875
]
] |
[
"Our method sets a new state of the art on the",
"of-the-art. Multiple researchers showed that inaccurate",
"taskof monocular 3D detection on KITTI-3D[14] and",
"depth estimation is a major source of error in PL methods",
"nuScenes[5]with significantimprovementscompared to",
"[58,66l.Inthiswork,webuildourreferencePLmethod",
"previous state-of-the-art methods.The simplicity of its",
"based on [49] and [41] to investigate the benefits of using",
"training procedure and its end-to-end optimization allows",
"large-scale depth data and compare it with DD3D.",
"for effective use of large-scale depth data, leading to im-",
"Monocular depth estimation. Estimating per-pixel depth",
"pressive multi-class detection accuracy.",
"is a key task for both DD3D (as a pre-training task) and PL",
"(as the first of a two-stage pipeline). It is itself a thriving",
"2. Related work",
"research area: the community has pushed toward accurate",
"densedepthpredictionviasupervised[35,29,19,50,26.",
"Monocular 3D detection. A large body of work in image-",
"18, 32, 13, 31] as well as self-supervised [48, 75, 62, 16,",
"based 3D detection builds upon 2D detectors, and aims",
"15, 64] methods. We note that supervised monocular depth",
"to lift them to 3D using various cues from object shapes",
"training used in this work require no annotations from hu-",
"and scene geometry.These priors are often injected by",
"man, allowing us to scale our methods to a large amount of",
"aligning 2Dkeypoints with their projective 3Dcounterparts",
"rawdata.",
"[6, 1, 2, 21, 52], or by learning a low-dimensional shape",
"representation [28, 43]. Other methods try to leverage ge-",
"3. Dense depth pre-training for 3D detection",
"ometric consistency between 2D and 3D structures, formu-",
"latingtheinferenceasaconstrainedoptimizationproblem",
"Given a single image and its camera intrinsics matrix as",
"[45, 38, 46, 12, 44, 24, 27, 34]. These methods commonly",
"input, the goal of monocular 3D detection is to generate a",
"use additional data (e.g. 3D CAD models, instance segmen-",
"set of multi-class 3D bounding boxes relative to camera co-",
"tation), or assume rigidity of objects. Inspired by Lidar-",
"ordinates. During inference, DD3D does not require any ad-",
"based methods[8,70, 30],another line of work employs",
"ditional data, such as per-pixel depth estimates, 2D bound-",
"viewtransformation (i.e.birds-eye-view)to overcome the",
"ing boxes, segmentations, or 3D CAD models. DD3D is",
"limitations of perspective range view, e.g. occlusion or vari-",
"also camera-aware: it scales the depth of putative 3D boxes",
"ability in size [25, 56, 60]. These methods often require pre-",
"according to the camera intrinsics.",
"cise camera extrinsics, or are accurate only in close vicinity.",
"End-to-end 3D detectors. Alternatively, researchers have",
"3.1. Architecture",
"attempted to directly regress 3D bounding boxes from CNN",
"DD3D is a fully convolutional single-stage network that",
"features [59, 39, 76, 23, 3, 9]. Typically, these approaches",
"extends FCOS [61] to perform 3D detection and dense",
"extend standard 2D detectors (single-stage [54, 76] or two-",
"depth prediction. The architecture (see Figure 2) is com-",
"stage [55]) by adding heads that predict various 3D cuboid",
"posed of a backbone network and three sub-networks (or",
"parameterizations . The use of depth-aware convolution or",
"heads) that are shared among all multi-scale features. The",
"dense 3D anchors [10, 53] enabled higher accuracy. Incor-",
"backbone takes an RGB image as input, and computes con-",
"poratinguncertaintyin estimatingthe3Dbox(orits depth)",
"volutional features at different scales. As in [61], we adopt",
"has also been shown to greatly improve detection accu-",
"a feature pyramid network (FPN) [36] as the backbone.",
"racy [59, 9, 58]. In [59], the authors proposed a disentan-",
"Three head networks are applied to each feature map",
"gled loss for 3D box regression that helps stabilize training.",
"produced by the backbone, and perform independent pre-",
"DD3D also falls in the end-to-end 3D detector category,",
"diction tasks. The classification module predicts object cat-",
"however, our emphasis is on learning a good depth repre-",
"egory. It produces C real values, where C is the number",
"sentation via large-scale self-supervised pre-training on raw",
"of object categories. The 2D box module produces class-",
"data, which leads to robust 3D detection.",
"agnostic bounding boxes and center-ness by predicting 4",
"Pseudo-Lidar methods. Starting with the pioneering",
"offsetsfromthefeaturelocationtothesidesofeachbound-",
"work of[67l,these methodsleverage advances inmonoc-",
"ing box and a scalar associated with center-ness. We refer",
"ular depth estimation and train Lidar-based detectors on",
"the readers to [61]for more details regarding the 2D detec-",
"the resulting pseudo pointcloud, producing impressive re-",
"tion architecture",
"sults[69,51,68,421.Recent methods have improved",
"3D detection head. This head predicts 3D bounding boxes",
"upon [67] by correcting the monocular depth with sparse",
"and per-pixel depth. It takes FPN features as input, and ap-",
"Lidar readings [71], decorating the pseudo pointcloud with",
"plies four 2D convolutions with 3 × 3 kernels that generate",
"colors [42], using 2D detection to segment foreground",
"12real valuesforeachfeaturelocation.Thesearedecoded",
"pointcloud regions [49, 69, 41], and structured sparsifica-",
"into3D bounding box,per-pixel depth map,and 3D predic-",
"tion of the monocular pointcloud [63]. Recently, [58] has",
"tion confidence, as described below:",
"shown a bias in the PL results on the representative KITTI-",
"3D [14] benchmark, while this paradigm remains the state-",
"q=(qw,qr,qy,qz)is thequaternion representation of"
] |
{
"arxiv_id": "2108.06417",
"language": "en",
"page_idx": 1
}
|
||
2108.06417-en-2
|
2108.06417-en-2.jpg
|
<FigureHere>
Figure 2: DD3D is a fully convolutional single-stage architecture that performs monocular 3D detection and dense depth prediction. To maximize reuse of pre-trained features, the inference of dense depth and 3D bounding box share most of the parameters. The heads are shared among all FPN layers.
allocentric orientation [43,59,39] of the 3D bounding box. It is normalized and transformed to an egocentric orientation [43]. Note that we predict orientations with the full 3 degrees of freedom.
$\bullet z_{\{c,p\}}$ represent the depth predictions. $z_c$ is decoded to the z-component of 3D bounding box centers and therefore only associated with foreground features, while $z_p$ is decoded to the monocular depth to the closest surface and is associated with every pixel. To decode them to metric depth, we unnormalized them using per-level parameters as follows:
(1)
$$\begin{aligned}&d=\frac{c}{p}\cdot(\sigma_{l}\cdot z+\mu_{l}),\\&p=\sqrt{\frac{1}{f_{x}^{2}}+\frac{1}{f_{y}^{2}}},\end{aligned}$$
(2)
where $z\in\{z_c,z_p\}$ are network output, $d\in\{d_c,d_p\}$ are predicted depths, $(\sigma_l,\mu_l)$ are learnable scaling factors and offsets defined for each FPN level, $p$ is the pixel size computed from the focal lengths, $f_x$ and $f_y$, and $c$ is a constant.
We note that this design of using camera focal lengths endows DD3D with camera-awareness, by allowing us to infer the depth not only from the input image, but also from the pixel size. We found that this is particularly useful for stable training. Specifically, when the input image is resized during training, we keep the ground-truth 3D bounding box unchanged, but modify the camera resolution as follows:
(3)
$$\mathbf{K}=\begin{bmatrix}r_x&&r_y&&1\end{bmatrix}\begin{bmatrix}f_x&&0&&p_x\\0&&f_y&&p_y\\0&&0&&1\end{bmatrix},$$
where $r_x$ and $r_y$ are resize rates, and K is the new camera intrinsic matrix that is used in Eqs. 2 and 4. Finally, $\{z_p\}$ collectively represent low-resolution versions of the dense depth maps computed from each FPN features. To recover the full resolution of the dense depth maps, we apply bilinear interpolation to match the size of the input image.
$\bullet\mathbf{o}=(\Delta_{u},\Delta_{v})$ represent offsets from the feature location to the 3D bounding box center projected onto the camera plane. This is decoded to the 3D center via unprojection:
(4)
$$\mathbf{C}=\mathbf{K}^{-1}\begin{bmatrix}u_b+\alpha_l\Delta_u\\v_b+\alpha_l\Delta_v\\1\end{bmatrix}d_c,$$
where $(u_b,v_b)$ is the feature location in image space, $\alpha_l$ is a learnable scaling factor assigned to each FPN level.
$\bullet$ $\delta =$ $( \delta _W, \delta _H, \delta _H)$ represents the deviation in the size
of the 3D bounding box from the class-specific canonical size, i. e. s= ( $W_{0}e^{\delta w}$, $H_{0}e^{\delta _{\mathrm{н } }}$, $D_{0}e^{\delta _{\mathrm{D} }}$) . As in [ 57] , $(W_{0},H_{0},D_{0})$ is the canonical box size for each class, and is pre-computed from the training data as its average size.
$\bullet\beta_{3\mathrm{D}}$ represents the confidence of the 3D bounding box
prediction [59]. It is transformed into a probability: $p_{3\mathrm{D} }$ = (1 + $e^{- \beta _{3\mathrm{D} }}) ^{- 1}$, and multiplied by the class probability computed from the classification head[61] to account for the relative confidence to the 2D confidence. The adjusted probability is used as the final score for the candidate.
## 3.2. Losses
We adopt the classification loss and 2D bounding box
regresssion loss from FCOS [61]:
$$\mathcal{L}_{2\mathrm{D}}=\mathcal{L}_{reg}+\mathcal{L}_{cls}+\mathcal{L}_{ctr},$$
(5)
where the 2D box regression loss $L_{reg}$ is the IOU loss [73], the classification loss $L_{cls}$ is the binary focal loss (i.e. onevs-all), and the center-ness loss $L_{ctr}$ is the binary cross entropy loss. For 3D bounding box regression, we use the disentangled L1 loss described in [59], i.e
(6)
$$\mathcal{L}_{\mathrm{3D}}(\mathbf{B}^*,\hat{\mathbf{B}})=\frac{1}{8}||\mathbf{B}^*-\hat{\mathbf{B}}||_1,$$
|
[
[
0.441162109375,
0.0953369140625,
0.455078125,
0.0953369140625,
0.455078125,
0.1300048828125,
0.441162109375,
0.1300048828125
],
[
0.72607421875,
0.117431640625,
0.74658203125,
0.117431640625,
0.74658203125,
0.1319580078125,
0.72607421875,
0.1319580078125
],
[
0.76708984375,
0.11871337890625,
0.77783203125,
0.11871337890625,
0.77783203125,
0.1287841796875,
0.76708984375,
0.1287841796875
],
[
0.7939453125,
0.11932373046875,
0.81396484375,
0.11932373046875,
0.81396484375,
0.1319580078125,
0.7939453125,
0.1319580078125
],
[
0.441162109375,
0.1287841796875,
0.455078125,
0.1287841796875,
0.455078125,
0.1749267578125,
0.441162109375,
0.1749267578125
],
[
0.7294921875,
0.1319580078125,
0.74658203125,
0.1319580078125,
0.74658203125,
0.1458740234375,
0.7294921875,
0.1458740234375
],
[
0.76123046875,
0.1331787109375,
0.779296875,
0.1331787109375,
0.779296875,
0.147705078125,
0.76123046875,
0.147705078125
],
[
0.87109375,
0.1319580078125,
0.89208984375,
0.1319580078125,
0.89208984375,
0.146484375,
0.87109375,
0.146484375
],
[
0.7939453125,
0.1337890625,
0.81396484375,
0.1337890625,
0.81396484375,
0.147705078125,
0.7939453125,
0.147705078125
],
[
0.5986328125,
0.1337890625,
0.62158203125,
0.1337890625,
0.62158203125,
0.1427001953125,
0.5986328125,
0.1427001953125
],
[
0.334228515625,
0.14453125,
0.34716796875,
0.14453125,
0.34716796875,
0.147705078125,
0.334228515625,
0.147705078125
],
[
0.6337890625,
0.1395263671875,
0.650390625,
0.1395263671875,
0.650390625,
0.143310546875,
0.6337890625,
0.143310546875
],
[
0.72802734375,
0.146484375,
0.74658203125,
0.146484375,
0.74658203125,
0.1610107421875,
0.72802734375,
0.1610107421875
],
[
0.7646484375,
0.1483154296875,
0.77783203125,
0.1483154296875,
0.77783203125,
0.1590576171875,
0.7646484375,
0.1590576171875
],
[
0.79736328125,
0.149658203125,
0.81201171875,
0.149658203125,
0.81201171875,
0.1578369140625,
0.79736328125,
0.1578369140625
],
[
0.5361328125,
0.1749267578125,
0.890625,
0.1749267578125,
0.890625,
0.1881103515625,
0.5361328125,
0.1881103515625
],
[
0.3349609375,
0.1900634765625,
0.34716796875,
0.1900634765625,
0.34716796875,
0.1944580078125,
0.3349609375,
0.1944580078125
],
[
0.53515625,
0.189453125,
0.89208984375,
0.189453125,
0.89208984375,
0.20263671875,
0.53515625,
0.20263671875
],
[
0.443603515625,
0.210205078125,
0.449462890625,
0.210205078125,
0.44775390625,
0.2235107421875,
0.44287109375,
0.2235107421875
],
[
0.53515625,
0.2052001953125,
0.89208984375,
0.2052001953125,
0.89208984375,
0.2183837890625,
0.53515625,
0.2183837890625
],
[
0.5361328125,
0.2197265625,
0.89208984375,
0.2197265625,
0.89208984375,
0.23291015625,
0.5361328125,
0.23291015625
],
[
0.5361328125,
0.23486328125,
0.89208984375,
0.23486328125,
0.89208984375,
0.248779296875,
0.5361328125,
0.248779296875
],
[
0.5361328125,
0.25,
0.89208984375,
0.25,
0.89208984375,
0.264404296875,
0.5361328125,
0.264404296875
],
[
0.334228515625,
0.264404296875,
0.406005859375,
0.264404296875,
0.406005859375,
0.271484375,
0.334228515625,
0.271484375
],
[
0.53515625,
0.26513671875,
0.75634765625,
0.26513671875,
0.75634765625,
0.279052734375,
0.53515625,
0.279052734375
],
[
0.529296875,
0.290283203125,
0.890625,
0.290283203125,
0.890625,
0.3037109375,
0.529296875,
0.3037109375
],
[
0.080078125,
0.2998046875,
0.46728515625,
0.2998046875,
0.46728515625,
0.313232421875,
0.080078125,
0.313232421875
],
[
0.53515625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.53515625,
0.318115234375
],
[
0.080078125,
0.31494140625,
0.468017578125,
0.31494140625,
0.468017578125,
0.328857421875,
0.080078125,
0.328857421875
],
[
0.53369140625,
0.32080078125,
0.8935546875,
0.3193359375,
0.8935546875,
0.333984375,
0.53369140625,
0.335205078125
],
[
0.080078125,
0.330810546875,
0.468017578125,
0.330810546875,
0.468017578125,
0.3447265625,
0.080078125,
0.3447265625
],
[
0.5361328125,
0.33837890625,
0.6259765625,
0.33837890625,
0.6259765625,
0.349853515625,
0.5361328125,
0.349853515625
],
[
0.080078125,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.359130859375,
0.080078125,
0.359130859375
],
[
0.080078125,
0.35986328125,
0.468017578125,
0.35986328125,
0.468017578125,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.69677734375,
0.3623046875,
0.77783203125,
0.3623046875,
0.77783203125,
0.376953125,
0.69677734375,
0.376953125
],
[
0.07843017578125,
0.373779296875,
0.158447265625,
0.375732421875,
0.158447265625,
0.39013671875,
0.07843017578125,
0.38818359375
],
[
0.62109375,
0.37890625,
0.66845703125,
0.37890625,
0.66845703125,
0.389404296875,
0.62109375,
0.389404296875
],
[
0.708984375,
0.380126953125,
0.771484375,
0.380126953125,
0.771484375,
0.39013671875,
0.708984375,
0.39013671875
],
[
0.7861328125,
0.38330078125,
0.79736328125,
0.37890625,
0.80126953125,
0.385009765625,
0.7900390625,
0.389404296875
],
[
0.869140625,
0.37890625,
0.8935546875,
0.37890625,
0.8935546875,
0.392578125,
0.869140625,
0.392578125
],
[
0.73046875,
0.396484375,
0.74267578125,
0.396484375,
0.74267578125,
0.40478515625,
0.73046875,
0.40478515625
],
[
0.11273193359375,
0.410400390625,
0.468017578125,
0.410888671875,
0.468017578125,
0.425537109375,
0.11273193359375,
0.4248046875
],
[
0.5361328125,
0.42041015625,
0.8896484375,
0.42041015625,
0.8896484375,
0.43505859375,
0.5361328125,
0.43505859375
],
[
0.11273193359375,
0.4267578125,
0.4697265625,
0.4267578125,
0.4697265625,
0.43994140625,
0.11273193359375,
0.43994140625
],
[
0.11273193359375,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.11273193359375,
0.455078125
],
[
0.53369140625,
0.436279296875,
0.890625,
0.43505859375,
0.890625,
0.449462890625,
0.53369140625,
0.45068359375
],
[
0.11273193359375,
0.455810546875,
0.30810546875,
0.455810546875,
0.30810546875,
0.4697265625,
0.11273193359375,
0.4697265625
],
[
0.53369140625,
0.45068359375,
0.57421875,
0.45068359375,
0.57421875,
0.46533203125,
0.53369140625,
0.46533203125
],
[
0.52294921875,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.52294921875,
0.490478515625
],
[
0.108642578125,
0.4892578125,
0.4697265625,
0.486083984375,
0.4697265625,
0.50390625,
0.108642578125,
0.5068359375
],
[
0.53515625,
0.492431640625,
0.890625,
0.492431640625,
0.890625,
0.505859375,
0.53515625,
0.505859375
],
[
0.11273193359375,
0.5048828125,
0.4697265625,
0.5048828125,
0.4697265625,
0.51806640625,
0.11273193359375,
0.51806640625
],
[
0.5361328125,
0.5068359375,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.5361328125,
0.52099609375
],
[
0.11358642578125,
0.521484375,
0.46728515625,
0.521484375,
0.46728515625,
0.5322265625,
0.11358642578125,
0.5322265625
],
[
0.53759765625,
0.52294921875,
0.890625,
0.52294921875,
0.890625,
0.5361328125,
0.53759765625,
0.5361328125
],
[
0.11358642578125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.11358642578125,
0.548828125
],
[
0.53515625,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.5517578125,
0.53515625,
0.5517578125
],
[
0.11273193359375,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.11273193359375,
0.564453125
],
[
0.53369140625,
0.55322265625,
0.59619140625,
0.55029296875,
0.59619140625,
0.56494140625,
0.53369140625,
0.5673828125
],
[
0.11273193359375,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.11273193359375,
0.5791015625
],
[
0.11273193359375,
0.58056640625,
0.348876953125,
0.58056640625,
0.348876953125,
0.5947265625,
0.11273193359375,
0.5947265625
],
[
0.52294921875,
0.57763671875,
0.890625,
0.57763671875,
0.890625,
0.59228515625,
0.52294921875,
0.59228515625
],
[
0.53515625,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.53515625,
0.60595703125
],
[
0.53173828125,
0.60595703125,
0.8935546875,
0.60498046875,
0.8935546875,
0.62255859375,
0.53173828125,
0.62353515625
],
[
0.253173828125,
0.6123046875,
0.2607421875,
0.6123046875,
0.2607421875,
0.6162109375,
0.253173828125,
0.6162109375
],
[
0.265625,
0.6142578125,
0.361083984375,
0.6142578125,
0.361083984375,
0.62890625,
0.265625,
0.62890625
],
[
0.443603515625,
0.61279296875,
0.471435546875,
0.61279296875,
0.471435546875,
0.6298828125,
0.443603515625,
0.6298828125
],
[
0.5361328125,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.5361328125,
0.63623046875
],
[
0.2181396484375,
0.61572265625,
0.2467041015625,
0.61572265625,
0.2467041015625,
0.626953125,
0.2181396484375,
0.626953125
],
[
0.253173828125,
0.62890625,
0.2607421875,
0.62890625,
0.2607421875,
0.63525390625,
0.253173828125,
0.63525390625
],
[
0.53369140625,
0.63818359375,
0.890625,
0.63818359375,
0.890625,
0.6513671875,
0.53369140625,
0.6513671875
],
[
0.313720703125,
0.64794921875,
0.32275390625,
0.64794921875,
0.32275390625,
0.65478515625,
0.313720703125,
0.65478515625
],
[
0.443603515625,
0.6513671875,
0.471435546875,
0.6513671875,
0.471435546875,
0.66845703125,
0.443603515625,
0.66845703125
],
[
0.5361328125,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.5361328125,
0.66650390625
],
[
0.26953125,
0.66552734375,
0.285888671875,
0.66552734375,
0.285888671875,
0.673828125,
0.26953125,
0.673828125
],
[
0.29736328125,
0.66357421875,
0.328369140625,
0.65771484375,
0.334228515625,
0.67529296875,
0.30224609375,
0.68115234375
],
[
0.5361328125,
0.67041015625,
0.68896484375,
0.67041015625,
0.68896484375,
0.68115234375,
0.5361328125,
0.68115234375
],
[
0.11358642578125,
0.6962890625,
0.46728515625,
0.6962890625,
0.46728515625,
0.7109375,
0.11358642578125,
0.7109375
],
[
0.5009765625,
0.6943359375,
0.58837890625,
0.6962890625,
0.58837890625,
0.7109375,
0.5009765625,
0.708984375
],
[
0.11273193359375,
0.7119140625,
0.46728515625,
0.7119140625,
0.46728515625,
0.72607421875,
0.11273193359375,
0.72607421875
],
[
0.521484375,
0.71728515625,
0.890625,
0.71826171875,
0.890625,
0.73291015625,
0.521484375,
0.73193359375
],
[
0.11273193359375,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.11273193359375,
0.74072265625
],
[
0.50244140625,
0.734375,
0.7236328125,
0.734375,
0.7236328125,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.111083984375,
0.7421875,
0.46728515625,
0.7421875,
0.46728515625,
0.755859375,
0.111083984375,
0.755859375
],
[
0.11273193359375,
0.75830078125,
0.23779296875,
0.75830078125,
0.23779296875,
0.76806640625,
0.11273193359375,
0.76806640625
],
[
0.60546875,
0.76123046875,
0.78759765625,
0.76123046875,
0.78759765625,
0.77587890625,
0.60546875,
0.77587890625
],
[
0.86767578125,
0.76123046875,
0.89453125,
0.76123046875,
0.89453125,
0.77783203125,
0.86767578125,
0.77783203125
],
[
0.11273193359375,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.11273193359375,
0.79443359375
],
[
0.1094970703125,
0.79443359375,
0.4697265625,
0.796875,
0.4697265625,
0.81103515625,
0.1094970703125,
0.8095703125
],
[
0.50341796875,
0.79052734375,
0.89208984375,
0.79052734375,
0.89208984375,
0.8046875,
0.50341796875,
0.8046875
],
[
0.111083984375,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.111083984375,
0.82470703125
],
[
0.50244140625,
0.80615234375,
0.89208984375,
0.80615234375,
0.89208984375,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.11273193359375,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.11273193359375,
0.84033203125
],
[
0.50244140625,
0.8212890625,
0.89208984375,
0.8212890625,
0.89208984375,
0.83544921875,
0.50244140625,
0.83544921875
],
[
0.5009765625,
0.83544921875,
0.8935546875,
0.833984375,
0.8935546875,
0.8486328125,
0.5009765625,
0.849609375
],
[
0.111083984375,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.111083984375,
0.85498046875
],
[
0.11273193359375,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.11273193359375,
0.87060546875
],
[
0.50244140625,
0.85107421875,
0.78173828125,
0.85107421875,
0.78173828125,
0.8642578125,
0.50244140625,
0.8642578125
],
[
0.11273193359375,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88623046875,
0.11273193359375,
0.88623046875
],
[
0.59619140625,
0.8818359375,
0.69775390625,
0.8798828125,
0.69775390625,
0.89794921875,
0.59619140625,
0.8994140625
],
[
0.705078125,
0.880859375,
0.79736328125,
0.880859375,
0.79736328125,
0.8984375,
0.705078125,
0.8984375
],
[
0.11273193359375,
0.88720703125,
0.33251953125,
0.88720703125,
0.33251953125,
0.90087890625,
0.11273193359375,
0.90087890625
],
[
0.869140625,
0.8837890625,
0.8935546875,
0.8837890625,
0.8935546875,
0.89794921875,
0.869140625,
0.89794921875
]
] |
[
"cls. head",
"f.",
"0",
"Pr",
"2D det. head",
"0",
"fy",
"(3)",
"Py",
"K:",
"",
"",
"0",
"0",
"1",
"where rr and ru are resize rates, and K is the new",
"3D",
"camera intrinsic matrix that is used in Eqs. 2 and 4.",
"oele",
"Finally, {zp collectively represent low-resolution ver-",
"sions of the dense depth maps computed from each",
"FPN features.To recover the full resolution of the",
"dense depth maps, we apply bilinear interpolation to",
"3Dconfidence8.",
"match the size of the input image.",
"o = (△u,△,) represent offsets from the feature lo-",
"Figure 2: DD3D is a fully convolutional single-stage ar-",
"cation to the 3D bounding box center projected onto",
"chitecture that performs monocular 3D detection and dense",
"the camera plane. This is decoded to the 3D center via",
"depth prediction. To maximize reuse of pre-trained features,",
"unprojection:",
"the inference of dense depth and 3D bounding box share",
"most of the parameters. The heads are shared among all",
"ub+Qdu",
"FPN layers.",
"C=K",
"h+A，",
"d",
"(4)",
"1",
"allocentric orientation [43, 59, 39] of the 3D bounding",
"where (ub, Vb) is the feature location in image space,",
"box. It is normalized and transformed to an egocentric",
"orientation [43]. Note that we predict orientations with",
"Qi is a learnable scaling factor assigned to each FPN",
"the full 3 degrees of freedom.",
"level.",
". = (ow,oH,oH) represents the deviation in the size",
"Z(c,p) represent the depth predictions: ze is decoded",
"of the 3D bounding box from the class-specific canon-",
"to the z-component of 3D bounding box centers and",
"ical size, i.e. s = (Woedw, Hoed, Doedp). As in [57],",
"therefore only associated with foreground features.",
"(Wo, Ho, Do) is the canonical box size for each class,",
"while zp is decoded to the monocular depth to the clos-",
"and is pre-computed from the training data as its aver-",
"est surface and is associated with every pixel. To de-",
"age size.",
"code them to metric depth, we unnormalized them us-",
"ing per-level parameters as follows:",
": β3D represents the confidence of the 3D bounding box",
"prediction [59]. It is transformed into a probability:",
"p3D = (1 + e-βaD)-1, and multiplied by the class",
"7",
": (oi -z+μ),",
"(1)",
"probability computed from the classification head [6]]",
"d=",
"D",
"to account for the relative confidence to the 2D con-",
"",
"(2)",
"fidence. The adjusted probability is used as the final",
"1",
"2",
"scorefor the candidate.",
"where z E [zc, zp] are network output, d E {de, dp]",
"3.2.Losses",
"arepredicted depths, (ot,μ)are learnable scaling fac-",
"We adopt the classification loss and 2D bounding box",
"tors and offsets defined for each FPN level, p is the",
"regresssion loss from FCOS [6]]:",
"pixel size computed from the focal lengths, f and fy",
"and cis a constant.",
"C2D=Lreg+Ccls+Cctr:",
"(5)",
"We note that this design of using camera focal lengths",
"endows DD3D with camera-awareness, by allowing",
"where the 2D box regression loss Lreg is the IOU loss [73],",
"us to infer the depth not only from the input image,",
"the classification loss Lcls is the binary focal loss (i.e. one-",
"but also from the pixel size. We found that this is par-",
"vs-all), and the center-ness loss Lctr is the binary cross en-",
"tropy loss. For 3D bounding box regression, we use the",
"ticularly useful for stable training. Specifically, when",
"the input image is resized during training, we keep the",
"disentangled L1 loss described in [59], i.e.",
"ground-truth 3D bounding box unchanged, but modify",
"C3D(B*,B) =",
"B*-B1,",
"the camera resolution as follows:",
"(6)"
] |
{
"arxiv_id": "2108.06417",
"language": "en",
"page_idx": 2
}
|
||
2108.06417-en-3
|
2108.06417-en-3.jpg
|
where $\mathbf{B}^*$ and Ê are the 8 vertices of ground-truth and candidate 3D boxes. This loss is replicated 4 times by using only one of the predicted 3D box components (orientation, projected center, depth, and size), while replacing other three with their ground-truth values.
Also as in [59], we adopt the self-supervised loss for 3D confidence which uses the error in 3D box prediction to compute a surrogate target for 3D confidence (relative to the 2D confidence):
$$p_{\mathrm{3D}}^*=e^{-\frac{1}{T}}\mathcal{L}_{\mathrm{3D}}(\mathbf{B}^*,\hat{\mathbf{B}})\:,$$
(7)
where $T$ is the temperature parameter. The confidence loss $\mathcal{L}_{conf}$ is the binary cross entropy between $p_3$D and $p_{3\mathrm{D}}^*.$ In summary, the total loss of DD3D is defined as follows:
(8)
$$\mathcal{L}_{\mathrm{DD}}=\mathcal{L}_{\mathrm{2D}}+\mathcal{L}_{\mathrm{3D}}+\mathcal{L}_{conf}.$$
# 3.3. Depth pre-training
During pre-training, we use per-pixel depth predictions from all FPN levels, i.e. $\{z_p\}.$ We consider pixels that have valid ground-truth depth from the sparse Lidar pointclouds projected onto the camera plane, and compute L1 distance from the predicted values:
(9)
(10)
$$\begin{aligned}&\mathcal{L}_{depth}^{l}=||\mathbf{D}^{*}-\hat{\mathbf{D}}_{l}||_{1}\odot\mathbf{M},\\&\mathcal{L}_{depth}=\sum_{l}\mathcal{L}_{depth}^{l},\end{aligned}$$
where $\mathbf{D}^*$ is the ground-truth depth map, $\hat{\mathbf{D}}_l$ is the predicted depth map from the $l$-th level in FPN (i.e. interpolated $z_p)$, and M is the binary indicator for valid pixels. We observed using all FPN levels in the objective, rather than e.g. using only the highest resolution features, enables stable training, especially when training from scratch. We also observed that the L1 loss yields stable training with large batch sizes and high-resolution input, compared to SILog loss[11,31] that is popular in monocular depth estimation literature.
We note that the two paths in DD3D from the input image to the 3D bounding box and to the dense depth prediction differ only in the last $3\times3$ convolutional layer, and thus share nearly all parameters. This allows for effective transfer from the pre-trained representation to the target task.
While pre-training, the camera-awareness of DD3D allows us to use camera intrinsics that are substantially different from the ones of the target domain, while still enjoying effective transfer. Specifically, from Eqs. 1 and 2, the error in depth prediction, $z$, caused by the difference in resolution between the two domains is corrected by the difference in pixel size, p.
# 4. Pseudo-Lidar 3D detection
An alternative way to utilize a large set of image-Lidar frames is to adopt the Pseudo-Lidar (PL) paradigm and aim
to improve the depth network component using the large scale data. PL is a two-stage method: first, given an image it applies a monocular depth network to predict per-pixel depth. The dense depth map is transformed into a 3D point cloud, and then a Lidar-based 3D detector is used to predict 3D bounding boxes. The modularity of PL methods enables us to quantify the role of improved depth predictors brought by a large-scale image-LiDAR dataset (see the supplementary material for additional details.)
Monocular depth estimation. The aim of monocular depth estimation is to compute the depth $\hat{D}=f_D\left(I\left(p\right)\right)$for each pixel $p\in I.$ Similarly to Eq. 9, given the ground truth depth measurement $\mathbf{D}^*$ acquired from Lidar pointclouds, we define a loss by the error between the predicted and the ground-truth depth. Here we instead use the SILog loss[11, 31], which yields better performance than L1 for PackNet. Network architecture. As the depth network, we use PackNet[17], a state-of-the-art monocular depth prediction architecture which uses packing and unpacking blocks with 3D convolutions. By avoiding feature sub-sampling, PackNet recovers fine structures in the depth map with high precision; moreover, PackNet has been shown to generalize better thanks to its increased capacity [17].
3D detection. To predict 3D bounding boxes from the input image and the estimated depth map, we follow the method proposed by [49, 41]. We first convert the estimated depth map into a 3D pointcloud similarly to Eq. 4, and concatenate each 3D point with the corresponding pixel values. This results in a 6D tensor encompassing colors along with 3D coordinates. We use an off-the-shelf 2D detector to identify proposal regions in input images, and apply a 3D detection network to each RoI region of the 6-channel image to produce 3D bounding boxes.
Backbone, detection head and 3D confidence. We follow[41] and process each RoI with a ResNet-18[20] backbone that uses Squeeze-and-Excitation layers [22]. As the RoI contains both objects as well as background pixels, the resulting features are filtered via foreground masks computed based on the associated RoI depth map[42]. The de- $\hat{\text{tection head follows[41]and operates in 3 distance ranges,}}$ producing one bounding box for each range. The final output is then selected based on the mean depth of the input RoI. Following [59, 58], we modify the detection head to also output a 3D confidence value $\gamma$ per detection, which is linked to the 3D detection loss.
Loss function. The 3D regression loss [49] is defined as:
$$\mathcal{L}_{\mathrm{3D}}^{\mathrm{PL}}=\mathcal{L}_{center}+\mathcal{L}_{size}+\mathcal{L}_{heading}+\mathcal{L}_{corners}.$$
(11)
In addition, we define a loss that links the predicted 3D confidence $\gamma$ with the 3D bounding box coordinates loss[59] using a Binary Cross Entropy (BCE) formulation with target $\tilde{\hat{\gamma}}=e^{-\dot{\mathcal{L}}_{corners}}.$ The final PL 3D detection loss is: $\tilde{\mathcal L}_{\mathrm{PL}}=\mathcal{L}_{3\mathrm{D}}^{\mathrm{PL}}+\mathcal{L}_{conf}^{\mathrm{PL}}.$
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.106689453125,
0.4697265625,
0.106689453125,
0.4697265625,
0.12060546875,
0.080078125,
0.12060546875
],
[
0.50244140625,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
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,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.358642578125,
0.15283203125,
0.358642578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.1004638671875,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.1004638671875,
0.18115234375
],
[
0.50244140625,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50244140625,
0.18115234375
],
[
0.07843017578125,
0.18115234375,
0.4697265625,
0.1817626953125,
0.4697265625,
0.1962890625,
0.07843017578125,
0.1956787109375
],
[
0.50341796875,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.080078125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.1962890625,
0.890625,
0.1976318359375,
0.890625,
0.212158203125,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.2115478515625,
0.21337890625,
0.2115478515625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.73779296875,
0.21337890625,
0.73779296875,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.50341796875,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.50341796875,
0.240478515625
],
[
0.1968994140625,
0.2392578125,
0.343017578125,
0.2310791015625,
0.3447265625,
0.248779296875,
0.19775390625,
0.2568359375
],
[
0.44677734375,
0.240478515625,
0.4697265625,
0.240478515625,
0.4697265625,
0.25390625,
0.44677734375,
0.25390625
],
[
0.50244140625,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.50244140625,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.26513671875,
0.4697265625,
0.26513671875,
0.4697265625,
0.279052734375,
0.080078125,
0.279052734375
],
[
0.50244140625,
0.2734375,
0.89208984375,
0.2734375,
0.89208984375,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.08251953125,
0.281005859375,
0.468017578125,
0.281005859375,
0.468017578125,
0.29541015625,
0.08251953125,
0.29541015625
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.07843017578125,
0.29541015625,
0.439453125,
0.294189453125,
0.439453125,
0.309326171875,
0.07843017578125,
0.31005859375
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.176513671875,
0.318115234375,
0.369384765625,
0.3193359375,
0.369384765625,
0.337158203125,
0.176513671875,
0.3359375
],
[
0.44677734375,
0.3212890625,
0.468017578125,
0.3212890625,
0.468017578125,
0.333984375,
0.44677734375,
0.333984375
],
[
0.50341796875,
0.318115234375,
0.8857421875,
0.318115234375,
0.8857421875,
0.33203125,
0.50341796875,
0.33203125
],
[
0.50341796875,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50341796875,
0.347900390625
],
[
0.080078125,
0.343505859375,
0.2607421875,
0.34521484375,
0.2607421875,
0.35986328125,
0.080078125,
0.35791015625
],
[
0.50341796875,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.1004638671875,
0.36865234375,
0.468017578125,
0.36865234375,
0.468017578125,
0.3818359375,
0.1004638671875,
0.3818359375
],
[
0.50341796875,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.50341796875,
0.37744140625
],
[
0.07843017578125,
0.3818359375,
0.4697265625,
0.38330078125,
0.4697265625,
0.397705078125,
0.07843017578125,
0.396484375
],
[
0.50341796875,
0.380126953125,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.50341796875,
0.393310546875
],
[
0.08087158203125,
0.3984375,
0.4697265625,
0.3984375,
0.4697265625,
0.412353515625,
0.08087158203125,
0.412353515625
],
[
0.5009765625,
0.393310546875,
0.890625,
0.39453125,
0.890625,
0.4091796875,
0.5009765625,
0.40771484375
],
[
0.080078125,
0.4140625,
0.4697265625,
0.4140625,
0.4697265625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.50244140625,
0.410400390625,
0.89208984375,
0.410400390625,
0.89208984375,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.4287109375,
0.253173828125,
0.4287109375,
0.253173828125,
0.442626953125,
0.080078125,
0.442626953125
],
[
0.50244140625,
0.425537109375,
0.78369140625,
0.425537109375,
0.78369140625,
0.439453125,
0.50244140625,
0.439453125
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.176513671875,
0.451904296875,
0.371826171875,
0.45068359375,
0.371826171875,
0.4697265625,
0.176513671875,
0.47021484375
],
[
0.443603515625,
0.45263671875,
0.471435546875,
0.45263671875,
0.471435546875,
0.4697265625,
0.443603515625,
0.4697265625
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.25732421875,
0.47021484375,
0.32275390625,
0.479736328125,
0.31787109375,
0.5,
0.25244140625,
0.490478515625
],
[
0.5009765625,
0.47021484375,
0.89208984375,
0.4697265625,
0.89208984375,
0.484130859375,
0.5009765625,
0.48486328125
],
[
0.43701171875,
0.47607421875,
0.471435546875,
0.47607421875,
0.471435546875,
0.4912109375,
0.43701171875,
0.4912109375
],
[
0.176513671875,
0.47607421875,
0.256591796875,
0.47607421875,
0.256591796875,
0.49365234375,
0.176513671875,
0.49365234375
],
[
0.50244140625,
0.48486328125,
0.890625,
0.484130859375,
0.890625,
0.498779296875,
0.50244140625,
0.5
],
[
0.50244140625,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.52978515625,
0.080078125,
0.52978515625
],
[
0.50341796875,
0.51513671875,
0.89208984375,
0.51513671875,
0.89208984375,
0.52880859375,
0.50341796875,
0.52880859375
],
[
0.07843017578125,
0.53076171875,
0.468017578125,
0.5322265625,
0.468017578125,
0.546875,
0.07843017578125,
0.54541015625
],
[
0.50244140625,
0.52880859375,
0.89208984375,
0.52978515625,
0.89208984375,
0.54443359375,
0.50244140625,
0.54345703125
],
[
0.080078125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.50341796875,
0.54541015625,
0.89208984375,
0.54541015625,
0.89208984375,
0.56005859375,
0.50341796875,
0.56005859375
],
[
0.07843017578125,
0.56103515625,
0.4697265625,
0.5625,
0.4697265625,
0.5771484375,
0.07843017578125,
0.57568359375
],
[
0.50341796875,
0.5625,
0.8896484375,
0.5625,
0.8896484375,
0.57275390625,
0.50341796875,
0.57275390625
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.50244140625,
0.57568359375,
0.736328125,
0.57568359375,
0.736328125,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.080078125,
0.59130859375,
0.4697265625,
0.59130859375,
0.4697265625,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50341796875,
0.59033203125,
0.89208984375,
0.59033203125,
0.89208984375,
0.60400390625,
0.50341796875,
0.60400390625
],
[
0.080078125,
0.60693359375,
0.4697265625,
0.60693359375,
0.4697265625,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.50341796875,
0.60791015625,
0.890625,
0.60791015625,
0.890625,
0.61865234375,
0.50341796875,
0.61865234375
],
[
0.080078125,
0.6220703125,
0.46728515625,
0.6220703125,
0.46728515625,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50244140625,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63525390625,
0.50244140625,
0.63525390625
],
[
0.080078125,
0.63720703125,
0.44677734375,
0.63720703125,
0.44677734375,
0.65087890625,
0.080078125,
0.65087890625
],
[
0.50244140625,
0.63720703125,
0.89208984375,
0.63720703125,
0.89208984375,
0.65087890625,
0.50244140625,
0.65087890625
],
[
0.1004638671875,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.1004638671875,
0.66552734375
],
[
0.50244140625,
0.6513671875,
0.890625,
0.6513671875,
0.890625,
0.66552734375,
0.50244140625,
0.66552734375
],
[
0.080078125,
0.66748046875,
0.4697265625,
0.66748046875,
0.4697265625,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.50341796875,
0.66748046875,
0.890625,
0.66748046875,
0.890625,
0.68115234375,
0.50341796875,
0.68115234375
],
[
0.08087158203125,
0.6845703125,
0.4697265625,
0.6845703125,
0.4697265625,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.5009765625,
0.68115234375,
0.89208984375,
0.681640625,
0.89208984375,
0.6962890625,
0.5009765625,
0.69580078125
],
[
0.08087158203125,
0.69970703125,
0.46728515625,
0.69970703125,
0.46728515625,
0.71044921875,
0.08087158203125,
0.71044921875
],
[
0.5009765625,
0.6962890625,
0.8896484375,
0.6962890625,
0.8896484375,
0.71044921875,
0.5009765625,
0.71044921875
],
[
0.080078125,
0.71337890625,
0.44775390625,
0.71337890625,
0.44775390625,
0.7265625,
0.080078125,
0.7265625
],
[
0.50341796875,
0.7119140625,
0.89208984375,
0.7119140625,
0.89208984375,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.1004638671875,
0.72802734375,
0.468017578125,
0.72802734375,
0.468017578125,
0.7421875,
0.1004638671875,
0.7421875
],
[
0.50244140625,
0.7265625,
0.89208984375,
0.7265625,
0.89208984375,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.07843017578125,
0.7412109375,
0.468017578125,
0.7421875,
0.468017578125,
0.7568359375,
0.07843017578125,
0.755859375
],
[
0.50244140625,
0.7421875,
0.89208984375,
0.7421875,
0.89208984375,
0.755859375,
0.50244140625,
0.755859375
],
[
0.080078125,
0.75830078125,
0.4697265625,
0.75830078125,
0.4697265625,
0.771484375,
0.080078125,
0.771484375
],
[
0.50244140625,
0.7568359375,
0.71142578125,
0.7568359375,
0.71142578125,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.080078125,
0.7734375,
0.4697265625,
0.7734375,
0.4697265625,
0.787109375,
0.080078125,
0.787109375
],
[
0.50244140625,
0.77294921875,
0.87744140625,
0.77294921875,
0.87744140625,
0.7861328125,
0.50244140625,
0.7861328125
],
[
0.080078125,
0.78857421875,
0.4697265625,
0.78857421875,
0.4697265625,
0.80322265625,
0.080078125,
0.80322265625
],
[
0.53515625,
0.79443359375,
0.83642578125,
0.796875,
0.83642578125,
0.8173828125,
0.53515625,
0.81494140625
],
[
0.08087158203125,
0.8046875,
0.468017578125,
0.8046875,
0.468017578125,
0.81494140625,
0.08087158203125,
0.81494140625
],
[
0.86376953125,
0.8017578125,
0.89208984375,
0.8017578125,
0.89208984375,
0.81298828125,
0.86376953125,
0.81298828125
],
[
0.080078125,
0.82080078125,
0.16259765625,
0.82080078125,
0.16259765625,
0.83203125,
0.080078125,
0.83203125
],
[
0.50244140625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.08087158203125,
0.8466796875,
0.325927734375,
0.8466796875,
0.325927734375,
0.85986328125,
0.08087158203125,
0.85986328125
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50244140625,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.50244140625,
0.869140625
],
[
0.098876953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.560546875,
0.8720703125,
0.611328125,
0.869140625,
0.611328125,
0.8798828125,
0.560546875,
0.8818359375
],
[
0.62109375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.62109375,
0.88525390625
],
[
0.5,
0.8818359375,
0.64794921875,
0.8818359375,
0.64794921875,
0.9072265625,
0.5,
0.9072265625
],
[
0.5048828125,
0.87255859375,
0.56396484375,
0.87255859375,
0.56396484375,
0.8837890625,
0.5048828125,
0.8837890625
],
[
0.07843017578125,
0.88623046875,
0.468017578125,
0.88525390625,
0.468017578125,
0.8994140625,
0.07843017578125,
0.90087890625
]
] |
[
"where B* and B are the 8 vertices of ground-truth and can-",
"to improve the depth network component using the large",
"didate 3D boxes. This loss is replicated 4 times by us-",
"scale data. PL is a two-stage method: first, given an image",
"ing only one of the predicted 3D box components (orien-",
"it applies a monocular depth network to predict per-pixel",
"tation, projected center, depth, and size), while replacing",
"depth. The dense depth map is transformed into a 3D point",
"other three with their ground-truth values.",
"cloud, and then a Lidar-based 3D detector is used to predict",
"Also as in [59], we adopt the self-supervised loss for",
"3D bounding boxes. The modularity of PL methods enables",
"3D confidencewhich uses the error in 3D box prediction",
"us to quantify the role of improved depth predictors brought",
"to compute a surrogate target for 3D confidence (relative to",
"by a large-scale image-LiDAR dataset (see the supplemen-",
"the 2D confidence):",
"tary material for additional details.)",
"Monoculardepthestimation.Theaim ofmonoculardepth",
"p3D =e-CaD(B*,B)",
"(7)",
"estimation is to compute the depth D = fD (I (p)) for each",
"pixel p E I. Similarly to Eq. 9, given the ground truth",
"where T is the temperature parameter. The confidence loss",
"depth measurement D* acquired from Lidar pointclouds,",
"Ceonf is the binary cross entropy between psD and psp. In",
"we define a loss by the error between the predicted and the",
"summary, the total loss of DD3D is defined as follows:",
"ground-truth depth. Here we instead use the SILog loss [11,",
"CDD = C2D + L3D + Cconf.",
"(8)",
"31], which yields better performance than L1 for PackNet.",
"Network architecture. As the depth network, we use Pack-",
"3.3.Depth pre-training",
"Net [17], a state-of-the-art monocular depth prediction ar-",
"During pre-training, we use per-pixel depth predictions",
"chitecture which uses packing and unpacking blocks with",
"from all FPN levels, i.e. (zp. We consider pixels that have",
"3D convolutions. By avoiding feature sub-sampling,Pack-",
"valid ground-truth depth from the sparse Lidar pointclouds",
"Net recovers fine structures in the depth map with high pre-",
"projected onto the camera plane, and compute Ll distance",
"cision; moreover, PackNet has been shown to generalize",
"from the predicted values:",
"better thanks to its increased capacity [17].",
"3D detection. To predict 3D bounding boxes from the input",
"Cdepth = I|D* - Dilli M,",
"(9)",
"image and the estimated depth map, we follow the method",
"C'depth,",
"proposed by [49, 41]. We first convert the estimated depth",
"(10)",
"Cdepth =>",
"map into a 3D pointcloud similarly to Eq. 4, and concate-",
"nate each 3D point with the corresponding pixel values.",
"where D* is the ground-truth depth map, D, is the predicted",
"This results in a 6D tensor encompassing colors along with",
"depth map from the l-th level in FPN (i.e. interpolated zp),",
"3D coordinates.We use an off-the-shelf 2D detector to",
"and M is the binary indicator for valid pixels. We observed",
"identify proposal regions in input images, and apply a 3D",
"using all FPN levels in the objective, rather than e.g. using",
"detectionnetworktoeachRoIregionofthe6-channelim-",
"only the highest resolution features, enables stable training,",
"age to produce 3D bounding boxes.",
"especially whentrainingfrom scratch.Wealso observed",
"Backbone, detection head and 3D confidence. We fol-",
"that the L1 loss yields stable training with large batch sizes",
"low[41]andprocesseachRowithaResNet-18[20]back-",
"and high-resolution input, compared to SILog loss [11, 31]",
"bone that uses Squeeze-and-Excitation layers [22]. As the",
"that is popular in monocular depth estimation literature.",
"Rol contains both objects as well as background pixels, the",
"We note that the two paths in DD3D from the input im-",
"resulting features are filtered via foreground masks com-",
"age to the 3D bounding box and to the dense depth predic-",
"puted based on the associated Rol depth map [42]. The de-",
"tiondifferonlyinthelast3x3convolutional layer,andthus",
"tection head follows [41] and operates in 3 distance ranges,",
"share nearly all parameters.This allows for effectivetrans-",
"producing one bounding box for each range. The final out-",
"fer from the pre-trained representation to the target task.",
"put is then selected based on the mean depth of the input",
"While pre-training, the camera-awareness of DD3D al-",
"Rol.Following [59, 58], we modify the detection head to",
"lows us to use camera intrinsics that are substantially differ-",
"also output a 3D confidence value per detection, which is",
"ent from the ones of the target domain, while still enjoying",
"linked to the 3D detection loss.",
"effective transfer. Specifically, from Eqs. 1 and 2, the error",
"Loss function. The 3D regression loss [49] is defined as:",
"in depth prediction, z, caused by the difference in resolution",
"C = Leenter + Csize + Cheading + Corners.",
"betweenthetwodomainsiscorrectedbythedifferencein",
"(11)",
"pixel size, p.",
"In addition, we define a loss that links the predicted 3D con-",
"4.Pseudo-Lidar 3Ddetection",
"fidence with the 3D bounding box coordinates loss [59]",
"using a Binary Cross Entropy (BCE) formulation with tar-",
"An alternative way to utilize a large set of image-Lidar",
"e-eo",
"ersThefinal PL 3D detection loss is:",
"CpL = CSb + CPon.",
"get =",
"frames is to adopt the Pseudo-Lidar (PL) paradigm and aim"
] |
{
"arxiv_id": "2108.06417",
"language": "en",
"page_idx": 3
}
|
||
2108.06417-en-4
|
2108.06417-en-4.jpg
|
## 5. Experimental Setup
$5.1.Datasets$
KITTI-3D. The KITTI-3D detection benchmark[14] consists of urban driving scenes with 8 object classes. The benchmark evaluates 3D detection accuracy on three classes (Car, Pedestrian, and Cyclist) using two average precision (AP) metrics computed with class-specific thresholds on intersection-over-union (IoU) of 3D bounding boxes or Bird-Eye-View (2D) bounding boxes. We refer to these metrics as 3D AP and $BEVAP.$ We use the revised AP$|_R_{40}$ metrics [57]. The training set consists of 7481 images, the test set of 7518 images. The objects in the test set are organized into three partitions according to their difficulty level (easy, moderate, hard), and are evaluated separately. For the analysis in Section 6.2, we follow the common practice of splitting the training set into 3712 and 3769 images [7], and report validation results on the latter. We refer to these splits as KITTI-3D train and KITTI-3D val.
nuScenes. The nuScenes 3D detection benchmark[5] consists of 1000 multi-modal videos with 6 cameras covering the full 360-degree field of view. The videos are split into 700 for training, 150 for validation, and 150 for testing. The benchmark requires to report 3D bounding boxes of 10 object classes over 2Hz-sampled video frames. The evaluation metric, nuScenes detection score (NDS), is computed by combining the detection accuracy $(mAP)$ computed over four different thresholds on center distance with five $true-$ positive metrics. We report NDS and mAP, along with the three true-positive metrics that concern 3D detection, i.e. $ATE,ASE$, and $AOE.$
KITTI-Depth. We use the KITTI-Depth dataset[14] to fine-tune the depth networks of our PL models. It contains
over 93 thousands depth maps associated with the images in the KITTI raw dataset. The standard monocular depth protocol[74,15,16]is to use the Eigen splits[11]. However, as described in [58], up to a third of its training images overlap with KITTI-3D images, leading to biased results for PL models. To avoid this bias, we generate a new split by removing training images that are geographically close (i.e. within 50m) to any of the KITTI-3D images. We denote this split by Eigen-clean and use it to fine-tune the depth networks of our PL models.
DDAD15M. To pre-train DD3D and our PL models, we use an in-house dataset that consists of 25000 multi-camera videos of urban driving scenes. DDAD15M is a larger version of DDAD[16]: it contains high-resolution Lidar sensors to generate pointclouds and 6 cameras synchronized with 10 Hz scans. Most videos are 10-second long, which amounts to approximately 15M image frames in total. Unless noted differently, we use the entire dataset for pretraining.
## 5.2. Implementation detail
DD3D. We use V2-99 [33] extended to an FPN as the backbone network. When pre-training DD3D, we first initialize the backbone with parameters pre-trained on the 2D detection task using the COCO dataset[37], and perform the pre-training on dense depth prediction using the DDAD15M dataset.
We use a test-time augmentation by resizing and flipping the input images. We observed $2.3\%$ gain in “Car”BEV AP on KITTI $val$, but no improvement on the nuScenes validation set. All metrics on DD3D in Section 6.2 are averages over 4 training runs. We observed the variance over the runs to be small, i.e. $0.5\sim1.2\%$ BEV AP.
$\mathbf{PL.~When~training~PackNet~[16],~we~use~only~the~front}$ camera images of DDAD15M to pre-train PackNet, and train until convergence. We then fine-tune the network on KITTI Eigen-clean split for 5 epochs. For more details on training DD3D and PL, please refer to the supplementary material.
### $6. \textbf{Results}$
## $6. 1. \textbf{Benchmark evaluation}$
In this section, we evaluate DD3D on the KITTI- 3D and nuScenes benchmarks. DD3D is pre-trained on DDAD15M, and then fine-tuned on the training set of each dataset. We also evaluate PL on KITTI-3D. Its depth network is pre-trained on DDAD15M and fine-tuned on KITTI $Eigen-clean$, and its detection network is trained on the KITTI-3D train.
KITTI-3D. In Table 1 we compare the accuracy of DD3D to state-of-the-art methods on the KITTI-3D benchmark. DD3D achieves a significant improvement over all methods with 16.34% 3D AP on Moderate Cars, which amounts to a 23% improvement from the previous best method $(13.25\%).$ Qualitative visualization is presented in Figure 3. In Table 4 we evaluate DD3D also on the Pedestrian and Cyclist classes. DD3D outperforms all other approaches on the Pedestrian category, with an 80.5% improvement from the previous best method (9.30% vs 5.14% 3D AP). On Cyclist DD3D achieves the second best result, reducing the gap to[27] that uses ground-truth pointclouds to train a per-instance pointcloud reconstruction module and a two-stage regression network to refine 3D object proposals.
We next report the accuracy of our PL detector in Table 1. The accuracy of our PL detector is on par with stateof-the-art methods, however it performs significantly worse than DD3D (13.05% vs. 16.34%). We will discuss this result in the context of generalizability of PL methods in Section 6.2.
nuScenes. In Table 2 we compare DD3D with other monocular methods reported on the nuScenes detection bench-
|
[
[
0.080078125,
0.0889892578125,
0.26708984375,
0.091552734375,
0.26708984375,
0.1060791015625,
0.080078125,
0.10418701171875
],
[
0.50341796875,
0.091552734375,
0.7060546875,
0.091552734375,
0.7060546875,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.080078125,
0.11932373046875,
0.1805419921875,
0.11932373046875,
0.1805419921875,
0.1337890625,
0.080078125,
0.1337890625
],
[
0.521484375,
0.1142578125,
0.89208984375,
0.1142578125,
0.89208984375,
0.1275634765625,
0.521484375,
0.1275634765625
],
[
0.50244140625,
0.1300048828125,
0.89208984375,
0.1300048828125,
0.89208984375,
0.14453125,
0.50244140625,
0.14453125
],
[
0.080078125,
0.147705078125,
0.468017578125,
0.147705078125,
0.468017578125,
0.1610107421875,
0.080078125,
0.1610107421875
],
[
0.50244140625,
0.1458740234375,
0.8935546875,
0.1458740234375,
0.8935546875,
0.1590576171875,
0.50244140625,
0.1590576171875
],
[
0.07843017578125,
0.1634521484375,
0.47216796875,
0.1634521484375,
0.47216796875,
0.1767578125,
0.07843017578125,
0.1767578125
],
[
0.5009765625,
0.1590576171875,
0.890625,
0.160400390625,
0.890625,
0.1749267578125,
0.5009765625,
0.173583984375
],
[
0.080078125,
0.177978515625,
0.4697265625,
0.177978515625,
0.4697265625,
0.1912841796875,
0.080078125,
0.1912841796875
],
[
0.50244140625,
0.175537109375,
0.89208984375,
0.175537109375,
0.89208984375,
0.1900634765625,
0.50244140625,
0.1900634765625
],
[
0.08087158203125,
0.1932373046875,
0.468017578125,
0.1932373046875,
0.468017578125,
0.20703125,
0.08087158203125,
0.20703125
],
[
0.50244140625,
0.1900634765625,
0.6396484375,
0.1900634765625,
0.6396484375,
0.203857421875,
0.50244140625,
0.203857421875
],
[
0.07843017578125,
0.208984375,
0.468017578125,
0.208984375,
0.468017578125,
0.222900390625,
0.07843017578125,
0.222900390625
],
[
0.521484375,
0.205810546875,
0.89208984375,
0.205810546875,
0.89208984375,
0.2197265625,
0.521484375,
0.2197265625
],
[
0.080078125,
0.2235107421875,
0.4697265625,
0.2235107421875,
0.4697265625,
0.2374267578125,
0.080078125,
0.2374267578125
],
[
0.50244140625,
0.2203369140625,
0.89208984375,
0.2203369140625,
0.89208984375,
0.2342529296875,
0.50244140625,
0.2342529296875
],
[
0.080078125,
0.2392578125,
0.4697265625,
0.2392578125,
0.4697265625,
0.25244140625,
0.080078125,
0.25244140625
],
[
0.50341796875,
0.2374267578125,
0.8896484375,
0.2374267578125,
0.8896484375,
0.2474365234375,
0.50341796875,
0.2474365234375
],
[
0.0767822265625,
0.250732421875,
0.4697265625,
0.25244140625,
0.4697265625,
0.270751953125,
0.0767822265625,
0.268310546875
],
[
0.50244140625,
0.250732421875,
0.89208984375,
0.250732421875,
0.89208984375,
0.264404296875,
0.50244140625,
0.264404296875
],
[
0.07843017578125,
0.268310546875,
0.468017578125,
0.26953125,
0.468017578125,
0.28271484375,
0.07843017578125,
0.2822265625
],
[
0.50244140625,
0.266357421875,
0.89208984375,
0.266357421875,
0.89208984375,
0.27978515625,
0.50244140625,
0.27978515625
],
[
0.0767822265625,
0.28271484375,
0.46728515625,
0.2841796875,
0.46728515625,
0.298583984375,
0.0767822265625,
0.29736328125
],
[
0.5009765625,
0.281005859375,
0.75244140625,
0.281005859375,
0.75244140625,
0.294189453125,
0.5009765625,
0.294189453125
],
[
0.080078125,
0.298583984375,
0.468017578125,
0.298583984375,
0.468017578125,
0.3125,
0.080078125,
0.3125
],
[
0.521484375,
0.296630859375,
0.89208984375,
0.296630859375,
0.89208984375,
0.31005859375,
0.521484375,
0.31005859375
],
[
0.080078125,
0.313232421875,
0.468017578125,
0.313232421875,
0.468017578125,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.5009765625,
0.31005859375,
0.89208984375,
0.311279296875,
0.89208984375,
0.32568359375,
0.5009765625,
0.324462890625
],
[
0.080078125,
0.330078125,
0.471435546875,
0.330078125,
0.471435546875,
0.343505859375,
0.080078125,
0.343505859375
],
[
0.50244140625,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50244140625,
0.34033203125
],
[
0.080078125,
0.3447265625,
0.4697265625,
0.3447265625,
0.4697265625,
0.35791015625,
0.080078125,
0.35791015625
],
[
0.5009765625,
0.34033203125,
0.89208984375,
0.341552734375,
0.89208984375,
0.35595703125,
0.5009765625,
0.354736328125
],
[
0.080078125,
0.35986328125,
0.4697265625,
0.35986328125,
0.4697265625,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.5,
0.35595703125,
0.890625,
0.356689453125,
0.890625,
0.371826171875,
0.5,
0.37060546875
],
[
0.07843017578125,
0.375,
0.33251953125,
0.375,
0.33251953125,
0.38818359375,
0.07843017578125,
0.38818359375
],
[
0.5009765625,
0.371826171875,
0.56396484375,
0.371826171875,
0.56396484375,
0.386474609375,
0.5009765625,
0.386474609375
],
[
0.07843017578125,
0.392578125,
0.46728515625,
0.393310546875,
0.46728515625,
0.40771484375,
0.07843017578125,
0.4072265625
],
[
0.50244140625,
0.40087890625,
0.58837890625,
0.40087890625,
0.58837890625,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.07843017578125,
0.40771484375,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.07843017578125,
0.421630859375
],
[
0.080078125,
0.4248046875,
0.4697265625,
0.4248046875,
0.4697265625,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.5009765625,
0.423583984375,
0.71142578125,
0.4248046875,
0.71142578125,
0.439453125,
0.5009765625,
0.438232421875
],
[
0.07843017578125,
0.438232421875,
0.4697265625,
0.439453125,
0.4697265625,
0.453857421875,
0.07843017578125,
0.45263671875
],
[
0.521484375,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.462158203125,
0.521484375,
0.462158203125
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.08087158203125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50244140625,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.080078125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.07843017578125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54443359375,
0.07843017578125,
0.54443359375
],
[
0.50341796875,
0.52587890625,
0.890625,
0.52587890625,
0.890625,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.080078125,
0.54443359375,
0.46728515625,
0.54443359375,
0.46728515625,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.50244140625,
0.5390625,
0.615234375,
0.5390625,
0.615234375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.51953125,
0.55322265625,
0.8935546875,
0.5537109375,
0.8935546875,
0.568359375,
0.51953125,
0.5673828125
],
[
0.080078125,
0.56005859375,
0.2181396484375,
0.56005859375,
0.2181396484375,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5693359375,
0.890625,
0.5693359375,
0.890625,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.080078125,
0.57763671875,
0.4697265625,
0.5791015625,
0.4697265625,
0.59326171875,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.5947265625,
0.4697265625,
0.5947265625,
0.4697265625,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.5009765625,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.5009765625,
0.61279296875
],
[
0.080078125,
0.60986328125,
0.4697265625,
0.60986328125,
0.4697265625,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.5009765625,
0.6142578125,
0.8935546875,
0.6142578125,
0.8935546875,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.080078125,
0.626953125,
0.46728515625,
0.626953125,
0.46728515625,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.07843017578125,
0.64013671875,
0.468017578125,
0.6396484375,
0.468017578125,
0.65380859375,
0.07843017578125,
0.65478515625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.080078125,
0.66845703125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.50244140625,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.50244140625,
0.689453125
],
[
0.080078125,
0.68603515625,
0.46728515625,
0.68603515625,
0.46728515625,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.07843017578125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71533203125,
0.07843017578125,
0.71533203125
],
[
0.5009765625,
0.7041015625,
0.89208984375,
0.705078125,
0.89208984375,
0.7197265625,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.080078125,
0.72998046875
],
[
0.50244140625,
0.7216796875,
0.89208984375,
0.7216796875,
0.89208984375,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.080078125,
0.73193359375,
0.468017578125,
0.73193359375,
0.468017578125,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50244140625,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.74755859375,
0.2646484375,
0.74755859375,
0.2646484375,
0.75830078125,
0.080078125,
0.75830078125
],
[
0.5009765625,
0.75048828125,
0.8896484375,
0.75048828125,
0.8896484375,
0.7646484375,
0.5009765625,
0.7646484375
],
[
0.08087158203125,
0.7646484375,
0.46728515625,
0.7646484375,
0.46728515625,
0.77880859375,
0.08087158203125,
0.77880859375
],
[
0.50244140625,
0.7666015625,
0.52783203125,
0.7666015625,
0.52783203125,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.07843017578125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.07843017578125,
0.79443359375
],
[
0.521484375,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79541015625,
0.89208984375,
0.79541015625,
0.89208984375,
0.8095703125,
0.50244140625,
0.8095703125
],
[
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
],
[
0.07843017578125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.84033203125
],
[
0.5009765625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.5009765625,
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
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.583984375,
0.85791015625,
0.583984375,
0.869140625,
0.50341796875,
0.869140625
],
[
0.0767822265625,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.0767822265625,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.0767822265625,
0.88525390625,
0.1396484375,
0.88720703125,
0.1396484375,
0.9013671875,
0.0767822265625,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"5.Experimental Setup",
"5.2.Implementation detail",
"5.1. Datasets",
"DD3D. We use V2-99 [33] extended to an FPN as the",
"backbone network. When pre-training DD3D, we first ini-",
"KITTI-3D. The KITTI-3D detection benchmark [14] con-",
"tialize the backbone with parameters pre-trained on the",
"sists of urban driving scenes with 8 object classes.The",
"2D detection task using the COCO dataset [37], and per-",
"benchmark evaluates 3D detection accuracy on three classes",
"form the pre-training on dense depth prediction using the",
"(Car, Pedestrian, and Cyclist) using two average preci-",
"DDAD15M dataset.",
"sion (AP) metrics computed with class-specific thresholds",
"We use a test-time augmentation by resizing and flipping",
"on intersection-over-union (loU) of 3D bounding boxes or",
"the input images. We observed 2.3% gain in \"Car\" BEV AP",
"Bird-Eye-View (2D) bounding boxes.We refer to these",
"onKITTI val.but noimprovement onthe nuScenes valida-",
"metrics as 3D AP and BEV AP. We use the revised AP|Rao",
"tion set. All metrics on DD3D in Section 6.2 are averages",
"metrics [57].The training set consists of 7481 images, the",
"over 4 training runs. We observed the variance over the runs",
"test set of 7518 images. The objects in the test set are orga-",
"to be small, i.e. 0.5 ~ 1.2% BEV AP.",
"nized into three partitions according to their difficulty level",
"PL. When training PackNet [16], we use only the front",
"(easy,moderate, hard), and are evaluated separately. For the",
"camera images of DDAD15M to pre-train PackNet, and",
"analysis in Section 6.2, we follow the common practice of",
"train until convergence. We then fine-tune the network on",
"splitting the training set into 3712 and 3769 images [7], and",
"KITTI Eigen-clean split for 5 epochs. For more details on",
"report validation results on the latter. We refer to these splits",
"training DD3D and PL, please refer to the supplementary",
"as KITTI-3D train and KITTI-3D val.",
"material.",
"nuScenes. The nuScenes 3D detection benchmark [5] con-",
"6. Results",
"sists of 1000 multi-modal videos with 6 cameras covering",
"the full 360-degree field of view. The videos are split into",
"6.1. Benchmark evaluation",
"700 for training, 150 for validation, and 150 for testing.The",
"In this section, we evaluate DD3D on the KITTI-",
"benchmark requires to report 3D bounding boxes of 10 ob-",
"3D and nuScenes benchmarks. DD3D is pre-trained on",
"jectclassesover2Hz-sampled video frames.Theevalua-",
"tion metric, nuScenes detection score (NDS), is computed",
"DDAD15M, and then fine-tuned on the training set of each",
"dataset. We also evaluate PL on KITTI-3D. Its depth net-",
"by combining the detection accuracy (mAP) computed over",
"work is pre-trained on DDAD15M and fine-tuned on KITTI",
"four different thresholds on center distance with five true-",
"positive metrics. We report NDS and mAP, along with the",
"Eigen-clean. and its detection network is trained on the",
"three true-positive metrics that concern 3D detection, i.e.",
"KITTI-3D train.",
"KITTI-3D. In Table 1 we compare the accuracy of",
"ATE, ASE, and AOE.",
"DD3D to state-of-the-art methods on the KITTI-3D bench-",
"KITTI-Depth. We use the KITTI-Depth dataset [14] to",
"mark. DD3D achieves a significant improvement over all",
"fine-tune the depth networks of our PL models. It contains",
"methods with 16.34% 3D AP on Moderate Cars, which",
"over 93 thousands depth maps associated with the images",
"amounts to a 23% improvement from the previous best",
"in the KITTI raw dataset.The standard monocular depth",
"method (13.25%). Qualitative visualization is presented in",
"protocol [74, 15, 16] is to use the Eigen splits [11]. How-",
"Figure 3. In Table 4 we evaluate DD3D also on the Pedes-",
"ever, as described in [58], up to a third of its training images",
"trian and Cyclist classes. DD3D outperforms all other ap-",
"overlap with KITTI-3D images, leading to biased results for",
"proaches on the Pedestrian category, with an 80.5% im-",
"PL models. To avoid this bias, we generate a new split by",
"provement from the previous best method (9.30% vs 5.14%",
"removing training images that are geographically close (i.e.",
"3D AP). On Cyclist DD3D achieves the second best result,",
"within 50m) to any of the KITTI-3D images. We denote",
"reducing the gap to [27] that uses ground-truth pointclouds",
"this split by Eigen-clean and use it to fine-tune the depth",
"to train a per-instance pointcloud reconstruction module and",
"networks ofourPLmodels.",
"a two-stage regression network to refine 3D object propos-",
"DDAD15M. To pre-train DD3D and our PL models, we",
"als.",
"use an in-house dataset that consists of 25000 multi-camera",
"We next report the accuracy of our PL detector in Ta-",
"videos of urban driving scenes. DDAD15M is a larger ver-",
"ble 1. The accuracy of our PL detector is on par with state-",
"sion of DDAD [16]: it contains high-resolution Lidar sen-",
"of-the-art methods, however it performs significantly worse",
"sors to generate pointclouds and 6 cameras synchronized",
"than DD3D (13.05% vs. 16.34%). We will discuss this",
"with 10 Hz scans. Most videos are 10-second long, which",
"result in the context of generalizability of PL methods in",
"amounts to approximately 15M image frames in total. Un-",
"Section 6.2.",
"less noted differently, we use the entire dataset for pre-",
"nuScenes. In Table 2 we compare DD3D with other monoc-",
"training.",
"ular methods reported on the nuScenes detection bench-"
] |
{
"arxiv_id": "2108.06417",
"language": "en",
"page_idx": 4
}
|
||
2003.11883-en-0
|
2003.11883-en-0.jpg
|
# $\textbf{DCNAS: Densely Connected Neural Architecture Search for Semantic Image}$ Segmentation
Xiong Zhang$^1*$,Hongmin Xu$^2$,Hong Mo$^3$,Jianchao Tan$^2$,Cheng Yang$^1$,Lei Wang$^1$,Wenqi Ren$^4$
$^{1}$Joyy Inc.,$^{2}$ AI Platform, Kwai Inc., $^{3}$State Key Lab of VR, Beihang University, $^{4}$SKLOIS, IIE,CAS
## Abstract
ExistingNAS methods for dense image prediction tasks usually compromise on restricted search space or search on proxy task to meet the achievable computational demands. To allow as wide as possible network architectures and avoid the gap between realistic and proxy setting, we propose a novel Densely Connected NAS (DCNAS) framework, which directly searches the optimal network structures for the multi-scale representations of visual information, over a large- scale target dataset without proxy. Specifically, by connecting cells with each other using learnable weights, we introduce a densely connected search space to cover an abundance of mainstream network designs. Moreover, by combining both path-level and channel-level sampling strategies, we design a fusion module and mixture layer to reduce the memory consumption of ample search space, hence favoring the proxyless searching. Compared with contemporary works, experiments reveal that the proxyless searching scheme is capable of bridging the gap between searching and training environments. Further, DCNAS achieves new state-of-the-art performances on public semantic image segmentation benchmarks, including 84.3% $on$ Cityscapes, and 86.9% on PASCAL VOC 2012. We also retain leading performances when evaluating the architecture on the more challenging ADE20K and PASCALContext dataset.
# 1. Introduction
The heavy computation overheads of early Neural Architecture Search (NAS) methods [81] hinder their applications in real-world problems. Recently, limited search space strategies [82, 37, 38] significantly shorten the searching time of NAS algorithms, which make NAS approaches achieve superhuman performance in image classification task. However, to meet the low consumption of searching time, NAS methods with a constrained search space throw down the multi-scale representation of high-resolution im-
$^{*}$corresponding author, zhangxiong@yy.com
<table>
<tbody>
<tr>
<th>Method</th>
<th>Proxyless</th>
<th>$GPUI$ Days</th>
<th>$\mathbf{FLOPs}(G)$</th>
<th>个 $\rho$</th>
<th>$\tau\uparrow$</th>
<th>${\mathbf{m}}$IoU$(\%)$1</th>
</tr>
<tr>
<td>$DPC$ :[3]</td>
<td>$X$</td>
<td>2600</td>
<td>684.0</td>
<td>0.46</td>
<td>0.37</td>
<td>82.7</td>
</tr>
<tr>
<td>Auto-DeepLab [36]</td>
<td>$X$</td>
<td>3</td>
<td>695.0</td>
<td>0.31</td>
<td>0.21</td>
<td>82.1</td>
</tr>
<tr>
<td>CAS[ 73] </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>72.3</td>
</tr>
<tr>
<td>GAS[35]</td>
<td> </td>
<td>6.7</td>
<td> </td>
<td>-</td>
<td> </td>
<td>73.5</td>
</tr>
<tr>
<td>FasterSeg $\because[8]$</td>
<td>$x$</td>
<td>2</td>
<td>28.2</td>
<td>0.35</td>
<td>0.25</td>
<td>71.5</td>
</tr>
<tr>
<td>$\mathbf{Fast-NAS}[44]$</td>
<td>$x$</td>
<td>8</td>
<td>435.7</td>
<td>0.42</td>
<td>0.33</td>
<td>78.9</td>
</tr>
<tr>
<td>SparseMask [59]</td>
<td>$X$</td>
<td>4.2</td>
<td>36.4</td>
<td>0.49</td>
<td>0.38</td>
<td>68.6</td>
</tr>
<tr>
<td>Ours ( DCNAS) </td>
<td> </td>
<td>$\bar{5}.\bar{6}$</td>
<td>$\overline{29}4.6$</td>
<td>$\bar{0}.\bar{7}\bar{3}$</td>
<td>$\bar{0}.\bar{5}5$</td>
<td>$\overline{84}.\overline{3}$</td>
</tr>
</tbody>
</table>
Table 1. Comparisons with state of the arts. The table presents
the comparison results of [3, 36, 73, 35, 8, 66, 59] and our DCNAS comprehensively. In which, $\rho$ and $\tau$ represent the Pearson Correlation Coefficient and the Kendall Rank Correlation Coefficient, respectively, both of which are measured according to the performances of the super-net and the fine-tuned stand-alone model, Proxyless applies a proxyless searching paradigm, GPU Days presents the searching cost, mIoU(%) and FLOPs(G) report the accuracy and the flops of the best model.
age. As a result, those methods are not suitable for dense image prediction tasks $(e.g.$ semantic image segmentation, object detection, and monocular depth estimation).
To efficiently search appropriate network structure that can combine both the local and global clues of the semantic features, researches recently focus on improving NAS frameworks by designing new search spaces to handle multi-scale features. For instance, DPC [3] introduces a recursive search space, and Auto-DeepLab[36] proposes a hierarchical search space. However, designing new search space for dense image prediction tasks proves challenging: one has to delicately compose a flexible search space that covers as much as possible optimal network architectures. Meanwhile, efficiently address memory consumption and the heavy computation problems accompanied with ample search space for the high-resolution imagery.
In this work, we propose an efficient and proxyless NAS framework to search the optimal model structure for semantic image segmentation. Our approach is based on two principal considerations. Firstly, the search space should be comprehensive enough to handle most of the mainstream architecture designs, even some undiscovered highquality model structures. As shown in Figure l, we design a reticular-like and fully densely connected search space, which contains various paths in the search space. Con-
|
[
[
0.1004638671875,
0.1319580078125,
0.87109375,
0.1331787109375,
0.87109375,
0.15087890625,
0.1004638671875,
0.149658203125
],
[
0.418212890625,
0.1578369140625,
0.5556640625,
0.1578369140625,
0.5556640625,
0.17236328125,
0.418212890625,
0.17236328125
],
[
0.108642578125,
0.2052001953125,
0.86279296875,
0.2052001953125,
0.86279296875,
0.2197265625,
0.108642578125,
0.2197265625
],
[
0.15283203125,
0.2235107421875,
0.81787109375,
0.2235107421875,
0.81787109375,
0.238037109375,
0.15283203125,
0.238037109375
],
[
0.0228729248046875,
0.261962890625,
0.060455322265625,
0.261962890625,
0.060455322265625,
0.36865234375,
0.0228729248046875,
0.36865234375
],
[
0.035125732421875,
0.26953125,
0.043304443359375,
0.2734375,
0.039215087890625,
0.27783203125,
0.0310516357421875,
0.2734375
],
[
0.236083984375,
0.272705078125,
0.314453125,
0.272705078125,
0.314453125,
0.290283203125,
0.236083984375,
0.290283203125
],
[
0.5048828125,
0.27587890625,
0.5458984375,
0.27587890625,
0.5458984375,
0.285888671875,
0.5048828125,
0.285888671875
],
[
0.650390625,
0.2734375,
0.701171875,
0.27587890625,
0.701171875,
0.287353515625,
0.650390625,
0.285400390625
],
[
0.70751953125,
0.27587890625,
0.7587890625,
0.27587890625,
0.7587890625,
0.287353515625,
0.70751953125,
0.287353515625
],
[
0.82666015625,
0.27587890625,
0.8837890625,
0.27587890625,
0.8837890625,
0.287353515625,
0.82666015625,
0.287353515625
],
[
0.7685546875,
0.27783203125,
0.78515625,
0.27783203125,
0.78515625,
0.285888671875,
0.7685546875,
0.285888671875
],
[
0.6005859375,
0.276611328125,
0.6435546875,
0.276611328125,
0.6435546875,
0.287353515625,
0.6005859375,
0.287353515625
],
[
0.7998046875,
0.27783203125,
0.81640625,
0.27783203125,
0.81640625,
0.285888671875,
0.7998046875,
0.285888671875
],
[
0.50634765625,
0.285888671875,
0.5458984375,
0.28857421875,
0.5439453125,
0.2998046875,
0.5048828125,
0.29736328125
],
[
0.7939453125,
0.285888671875,
0.8203125,
0.285888671875,
0.8203125,
0.311279296875,
0.7939453125,
0.311279296875
],
[
0.83837890625,
0.285400390625,
0.869140625,
0.285400390625,
0.869140625,
0.330078125,
0.83837890625,
0.330078125
],
[
0.72119140625,
0.28857421875,
0.74658203125,
0.28857421875,
0.74658203125,
0.2998046875,
0.72119140625,
0.2998046875
],
[
0.6640625,
0.28857421875,
0.689453125,
0.28857421875,
0.689453125,
0.300537109375,
0.6640625,
0.300537109375
],
[
0.7685546875,
0.2890625,
0.7890625,
0.2890625,
0.7890625,
0.2998046875,
0.7685546875,
0.2998046875
],
[
0.765625,
0.29541015625,
0.7900390625,
0.29541015625,
0.7900390625,
0.309326171875,
0.765625,
0.309326171875
],
[
0.66650390625,
0.296630859375,
0.685546875,
0.296630859375,
0.685546875,
0.309326171875,
0.66650390625,
0.309326171875
],
[
0.72119140625,
0.296630859375,
0.7490234375,
0.296630859375,
0.7490234375,
0.309326171875,
0.72119140625,
0.309326171875
],
[
0.5087890625,
0.29736328125,
0.59228515625,
0.296630859375,
0.59228515625,
0.30810546875,
0.5087890625,
0.309326171875
],
[
0.615234375,
0.290283203125,
0.6298828125,
0.290283203125,
0.6298828125,
0.309326171875,
0.615234375,
0.309326171875
],
[
0.1004638671875,
0.306884765625,
0.468017578125,
0.306884765625,
0.468017578125,
0.32080078125,
0.1004638671875,
0.32080078125
],
[
0.5087890625,
0.30810546875,
0.55126953125,
0.30810546875,
0.55126953125,
0.318115234375,
0.5087890625,
0.318115234375
],
[
0.62109375,
0.30810546875,
0.6259765625,
0.3125,
0.62109375,
0.31640625,
0.615234375,
0.3125
],
[
0.6640625,
0.304931640625,
0.689453125,
0.304931640625,
0.689453125,
0.349853515625,
0.6640625,
0.349853515625
],
[
0.5087890625,
0.31494140625,
0.55126953125,
0.31494140625,
0.55126953125,
0.326904296875,
0.5087890625,
0.326904296875
],
[
0.8037109375,
0.311279296875,
0.8095703125,
0.311279296875,
0.8095703125,
0.3193359375,
0.8037109375,
0.3193359375
],
[
0.080078125,
0.322509765625,
0.4697265625,
0.322509765625,
0.4697265625,
0.3359375,
0.080078125,
0.3359375
],
[
0.615234375,
0.323974609375,
0.6298828125,
0.323974609375,
0.6298828125,
0.348388671875,
0.615234375,
0.348388671875
],
[
0.71826171875,
0.3212890625,
0.7490234375,
0.3212890625,
0.7490234375,
0.348388671875,
0.71826171875,
0.348388671875
],
[
0.765625,
0.323974609375,
0.7900390625,
0.323974609375,
0.7900390625,
0.3359375,
0.765625,
0.3359375
],
[
0.7958984375,
0.323974609375,
0.81884765625,
0.323974609375,
0.81884765625,
0.3359375,
0.7958984375,
0.3359375
],
[
0.84228515625,
0.323974609375,
0.865234375,
0.323974609375,
0.865234375,
0.337158203125,
0.84228515625,
0.337158203125
],
[
0.7958984375,
0.333984375,
0.81884765625,
0.333984375,
0.81884765625,
0.3466796875,
0.7958984375,
0.3466796875
],
[
0.5087890625,
0.32568359375,
0.56396484375,
0.326904296875,
0.56396484375,
0.335205078125,
0.5087890625,
0.333984375
],
[
0.765625,
0.335205078125,
0.7900390625,
0.335205078125,
0.7900390625,
0.3466796875,
0.765625,
0.3466796875
],
[
0.84228515625,
0.333984375,
0.86669921875,
0.333984375,
0.86669921875,
0.3466796875,
0.84228515625,
0.3466796875
],
[
0.5087890625,
0.3359375,
0.5703125,
0.3359375,
0.5703125,
0.3466796875,
0.5087890625,
0.3466796875
],
[
0.080078125,
0.33837890625,
0.4697265625,
0.33837890625,
0.4697265625,
0.3515625,
0.080078125,
0.3515625
],
[
0.72216796875,
0.3447265625,
0.74853515625,
0.3447265625,
0.74853515625,
0.35595703125,
0.72216796875,
0.35595703125
],
[
0.79345703125,
0.3447265625,
0.82275390625,
0.3447265625,
0.82275390625,
0.35595703125,
0.79345703125,
0.35595703125
],
[
0.8408203125,
0.3447265625,
0.86669921875,
0.3447265625,
0.86669921875,
0.35595703125,
0.8408203125,
0.35595703125
],
[
0.6708984375,
0.34521484375,
0.685546875,
0.34521484375,
0.685546875,
0.354736328125,
0.6708984375,
0.354736328125
],
[
0.5087890625,
0.34521484375,
0.583984375,
0.34521484375,
0.583984375,
0.35595703125,
0.5087890625,
0.35595703125
],
[
0.61767578125,
0.3466796875,
0.625,
0.3466796875,
0.625,
0.354248046875,
0.61767578125,
0.354248046875
],
[
0.7685546875,
0.3466796875,
0.7900390625,
0.3466796875,
0.7900390625,
0.354736328125,
0.7685546875,
0.354736328125
],
[
0.76318359375,
0.352783203125,
0.81884765625,
0.352783203125,
0.81884765625,
0.363037109375,
0.76318359375,
0.363037109375
],
[
0.5048828125,
0.354248046875,
0.57275390625,
0.354248046875,
0.57275390625,
0.3642578125,
0.5048828125,
0.3642578125
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.66845703125,
0.354248046875,
0.685546875,
0.354248046875,
0.685546875,
0.3623046875,
0.66845703125,
0.3623046875
],
[
0.72119140625,
0.354248046875,
0.74658203125,
0.354248046875,
0.74658203125,
0.3623046875,
0.72119140625,
0.3623046875
],
[
0.8408203125,
0.354248046875,
0.86669921875,
0.354248046875,
0.86669921875,
0.3623046875,
0.8408203125,
0.3623046875
],
[
0.02532958984375,
0.3623046875,
0.058013916015625,
0.3623046875,
0.058013916015625,
0.396484375,
0.02532958984375,
0.396484375
],
[
0.080078125,
0.36865234375,
0.4697265625,
0.36865234375,
0.4697265625,
0.3818359375,
0.080078125,
0.3818359375
],
[
0.50341796875,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.080078125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.396484375,
0.080078125,
0.396484375
],
[
0.50341796875,
0.38330078125,
0.68115234375,
0.38330078125,
0.68115234375,
0.396484375,
0.50341796875,
0.396484375
],
[
0.67333984375,
0.38330078125,
0.88818359375,
0.38330078125,
0.88818359375,
0.393310546875,
0.67333984375,
0.393310546875
],
[
0.08087158203125,
0.399658203125,
0.465576171875,
0.399658203125,
0.465576171875,
0.410400390625,
0.08087158203125,
0.410400390625
],
[
0.50244140625,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.50244140625,
0.410400390625
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.080078125,
0.41357421875,
0.468017578125,
0.41357421875,
0.468017578125,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.080078125,
0.427978515625,
0.46728515625,
0.427978515625,
0.46728515625,
0.44140625,
0.080078125,
0.44140625
],
[
0.5009765625,
0.4248046875,
0.890625,
0.4248046875,
0.890625,
0.438232421875,
0.5009765625,
0.438232421875
],
[
0.080078125,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.451904296875,
0.50244140625,
0.451904296875
],
[
0.0310516357421875,
0.448974609375,
0.058013916015625,
0.448974609375,
0.058013916015625,
0.481689453125,
0.0310516357421875,
0.481689453125
],
[
0.50244140625,
0.451904296875,
0.89208984375,
0.451904296875,
0.89208984375,
0.46533203125,
0.50244140625,
0.46533203125
],
[
0.080078125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.50244140625,
0.466552734375,
0.89208984375,
0.466552734375,
0.89208984375,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.080078125,
0.473388671875,
0.46728515625,
0.473388671875,
0.46728515625,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.5,
0.479736328125,
0.79150390625,
0.479736328125,
0.79150390625,
0.49365234375,
0.5,
0.49365234375
],
[
0.0236968994140625,
0.492431640625,
0.058013916015625,
0.492431640625,
0.058013916015625,
0.5673828125,
0.0236968994140625,
0.5673828125
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.080078125,
0.501953125
],
[
0.080078125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.517578125,
0.080078125,
0.517578125
],
[
0.08087158203125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.53076171875,
0.08087158203125,
0.53076171875
],
[
0.5009765625,
0.521484375,
0.8935546875,
0.52099609375,
0.8935546875,
0.53515625,
0.5009765625,
0.5361328125
],
[
0.080078125,
0.5341796875,
0.4697265625,
0.5341796875,
0.4697265625,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50244140625,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.50244140625,
0.55029296875
],
[
0.080078125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.56298828125,
0.080078125,
0.56298828125
],
[
0.50341796875,
0.5537109375,
0.833984375,
0.5537109375,
0.833984375,
0.564453125,
0.50341796875,
0.564453125
],
[
0.03350830078125,
0.56005859375,
0.041656494140625,
0.5625,
0.03594970703125,
0.58056640625,
0.02777099609375,
0.5791015625
],
[
0.041656494140625,
0.56298828125,
0.049835205078125,
0.56298828125,
0.049835205078125,
0.56640625,
0.041656494140625,
0.56640625
],
[
0.080078125,
0.564453125,
0.468017578125,
0.564453125,
0.468017578125,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.52294921875,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.52294921875,
0.58203125
],
[
0.080078125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.02532958984375,
0.5888671875,
0.056365966796875,
0.5888671875,
0.056365966796875,
0.6337890625,
0.02532958984375,
0.6337890625
],
[
0.50244140625,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.5009765625,
0.5986328125,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.5009765625,
0.6123046875
],
[
0.080078125,
0.60791015625,
0.46728515625,
0.60791015625,
0.46728515625,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.0228729248046875,
0.6220703125,
0.060455322265625,
0.6220703125,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.07843017578125,
0.62353515625,
0.43701171875,
0.62255859375,
0.43701171875,
0.63720703125,
0.07843017578125,
0.63818359375
],
[
0.4306640625,
0.625,
0.468017578125,
0.625,
0.468017578125,
0.63623046875,
0.4306640625,
0.63623046875
],
[
0.5048828125,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.6416015625,
0.5048828125,
0.6416015625
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.65380859375,
0.46728515625,
0.65380859375,
0.46728515625,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.50244140625,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.08087158203125,
0.66943359375,
0.188720703125,
0.66943359375,
0.188720703125,
0.68310546875,
0.08087158203125,
0.68310546875
],
[
0.50244140625,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.50244140625,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.08087158203125,
0.71533203125,
0.2091064453125,
0.71533203125,
0.2091064453125,
0.72998046875,
0.08087158203125,
0.72998046875
],
[
0.50341796875,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.098876953125,
0.74072265625,
0.468017578125,
0.74072265625,
0.468017578125,
0.75390625,
0.098876953125,
0.75390625
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.755859375,
0.468017578125,
0.755859375,
0.468017578125,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.50244140625,
0.75048828125,
0.79833984375,
0.75048828125,
0.79833984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.521484375,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.521484375,
0.77978515625
],
[
0.08087158203125,
0.77294921875,
0.468017578125,
0.77294921875,
0.468017578125,
0.78271484375,
0.08087158203125,
0.78271484375
],
[
0.080078125,
0.7861328125,
0.46728515625,
0.7861328125,
0.46728515625,
0.7998046875,
0.080078125,
0.7998046875
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.8017578125,
0.4697265625,
0.8017578125,
0.4697265625,
0.81494140625,
0.080078125,
0.81494140625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.8173828125,
0.468017578125,
0.8173828125,
0.468017578125,
0.83056640625,
0.080078125,
0.83056640625
],
[
0.080078125,
0.83203125,
0.468017578125,
0.83203125,
0.468017578125,
0.84521484375,
0.080078125,
0.84521484375
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.84765625,
0.4697265625,
0.84765625,
0.4697265625,
0.861328125,
0.080078125,
0.861328125
],
[
0.080078125,
0.8623046875,
0.468017578125,
0.8623046875,
0.468017578125,
0.87548828125,
0.080078125,
0.87548828125
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09722900390625,
0.88818359375,
0.3349609375,
0.88818359375,
0.3349609375,
0.9013671875,
0.09722900390625,
0.9013671875
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"DCNAS: Densely Connected Neural Architecture Search for Semantic Image",
"Segmentation",
"Xiong Zhangl* Hongmin Xu?, Hong Mo3, Jianchao Tan?, Cheng Yang', Lei Wangl, Wenqi Ren4",
"IJoyy Inc.,?AIPlatform, Kwai Inc.,3StateKeyLab of VR,Beihang University,4SKLOIS,IIE,CAS",
" ",
"",
"Abstract",
"Method",
"GPU Days",
"FLOPs(G)",
"mloU(%)",
"et",
"Proxyless",
"T个",
"DPC[3]",
"0.7",
"3",
"684.0",
"2600",
"0.46",
"0.31",
"3",
"695.0",
"Auto-DeepLab [36]",
"××",
"Existing NAS methods for dense image prediction tasks",
"CAS [73]",
".",
"-0283",
"GAS[35]",
"1",
"usually compromise on restricted search space or search",
"××",
"23",
"0.35",
"0.25",
"71.5",
"0.33",
"FasterSeg[8]",
"0.42 ",
"78.9",
"Fast-NAS []",
"on proxy task to meet the achievable computational de-",
"36.4",
"0.38",
"68.6",
"4.2",
"SparseMask [59]",
"x",
"0.49",
"0.73-0.55",
"Ours(DCNAS)",
"mands. To allow as wide as possible network architectures",
"5.6",
"294.6",
"84.3",
"2",
"and avoid the gap between realistic and proxy setting, we",
"Table 1. Comparisons with state of the arts. The table presents",
"propose a novel Densely Connected NAS (DCNAS) frame-",
"the comparison results of [3,",
"36.73.35.8.66.591andourDC",
"work, which directly searches the optimal network struc-",
"NAS comprehensively. In which, p and represent the Pear-",
"LAD",
"tures for the multi-scale representations of visual informa-",
"son Correlation Coefficient and the Kendall Rank Correlation Co-",
"tion, over a large-scale target dataset without proxy. Specif-",
"efficient, respectively, both of which are measured according to",
"ically, by connecting cells with each other using learnable",
"the performances of the super-net and the fine-tuned stand-alone",
"S",
"model, Proxyless applies a proxyless searching paradigm, GPU",
"weights, we introduce a densely connected search space to",
"Days presents the searching cost, mloU(%) and FLOPs(G) re-",
"cover an abundance of mainstream network designs. More-",
"port the accuracy and the flops of the best model.",
"21088",
"over, by combining both path-level and channel-level sam-",
"pling strategies, we design a fusion module and mixture",
"layerto reduce thememoryconsumption ofamplesearch",
"age. As a result, those methods are not suitable for dense",
"space, hence favoring the proxyless searching. Compared",
"image prediction tasks (e.g. semantic image segmentation,",
"with contemporary works, experiments reveal that the prox-",
"objectdetection,andmonoculardepthestimation).",
"",
"",
"yless searching scheme is capable of bridging the gap be-",
"To efficiently search appropriate network structure that",
"tween searching and training environments. Further, DC-",
"6002",
"can combine both the local and global clues of the se-",
"NAS achieves new state-of-the-art performances on public",
"mantic features, researches recently focus on improving",
"semantic image segmentation benchmarks, including 84.3%",
"NAS frameworks by designing new search spaces to han-",
"TAX",
"on Cityscapes, and 86.9% on PASCAL VOC 2012.",
"We",
"dlemulti-scalefeatures.Forinstance,DPC[3introducesa",
"also retain leading performances when evaluating the ar-",
"recursive search space, and Auto-DeepLab [36] proposes a",
"chitecture on the more challenging ADE2OK and PASCAL-",
"hierarchical search space. However, designing new search",
"Context dataset.",
"space for dense image prediction tasks proves challenging:",
"one has to delicately compose a flexible search space that",
"covers as much as possible optimal network architectures.",
"1. Introduction",
"Meanwhile, efficiently address memory consumption and",
"The heavy computation overheads of early Neural Ar-",
"the heavy computation problems accompanied with ample",
"chitecture Search (NAS) methods [81] hinder their appli-",
"search space for the high-resolution imagery.",
"In this work, we propose an efficient and proxyless NAS",
"cationsinreal-worldproblems.Recently,limitedsearch",
"space strategies [82, 37, 38] significantly shorten the search-",
"framework to search the optimal model structure for se-",
"ing time of NAS algorithms, which make NAS approaches",
"mantic image segmentation. Our approach is based on two",
"principal considerations. Firstly, the search space should",
"achieve superhuman performance in image classification",
"task. However, to meet the low consumption of searching",
"be comprehensive enough to handle most of the main-",
"stream architecture designs, even some undiscovered high-",
"time, NAS methods with a constrained search space throw",
"down the multi-scale representation of high-resolution im-",
"quality model structures. As shown in Figure I, we design",
"a reticular-like and fully densely connected search space,",
"\"corresponding author, zhangxiong@yy.com",
"which contains various paths in the search space. Con-"
] |
{
"arxiv_id": "2003.11883",
"language": "en",
"page_idx": 0
}
|
||
2003.11883-en-1
|
2003.11883-en-1.jpg
|
<FigureHere>
Figure 1. Framework of DCNAS. $Top:$ The densely connected search space (DCSS) with layer $L$ and max downsampling rate 32. To stable the searching procedure, we keep the beginning STEM and final Upsample block unchanged. Dashed lines represent candidate connections in DCSS, to keep clarity, we only demonstrate several connections among all the candidates. $Bottom\textit{Left: Fusion module}$ targets at aggregating feature-maps derived by previous layers, and solving the intensive GPU memory requirement problem by sampling a portion of connections from all possible ones. $Botton\textit{Right: Mixture layer may further save GPU memory consumption and accelerating}$ the searching process by sampling and operating on a portion of features while bypassing the others.
[26,17,74,68,69,27,70,21] and advances the state-ofthe-art performance. Transferring the model to ADE20K [78] and PASCAL-Context[43] datasets, our model obtains the best result compared with state-of-the-art approaches [12,17,22,24,71,69,31], according to the corresponding evaluation metrics.
Further, considering NAS methods depend on agent metrics to explore promising networks efficiently. If the agent metric provides a strong relative ranking of models, this enables the discovery of high performing models when trained to convergence. However, studies about the relative ranking of models in the searching period are less explored in literature [44,59,8,3,36]. In this work, we enrich this field and investigate most related methods on how well the accuracy in searching phase for specific architectures correlates with that of the fine-tuned stand-alone model.
To summarize, our main contributions are as follows:
sequently, DCNAS may derive whichever model architecture designs by selecting appropriate paths among the whole set of these connections. Additionally, DCNAS aggregates contextual semantics encoded in multi-scale imageries to derive long-range context information, which has been recently found to be vital in the state-of-the-art manually designed image segmentation models [40,74,5,7,70]. Secondly, we observe that previous NAS approaches [3,36] heavily rely on the proxy task or proxy dataset to reduce the cost of GPU hours and to alleviate the high GPU memory consumption problem. However, architectures optimized over the proxy are not guaranteed to be suitable in realistic setting [2], because of the gap between the proxy and the target configuration. Regarding this concern, we relax the discrete architectures into continuous representation to save GPU hours and design a fusion module which applies both path-level and channel-level sampling strategies during the searching procedure to reduce memory demand. Based on that, one may employ the stochastic gradient descent (SGD) to perform the proxyless searching procedure to select the optimal architecture from all the candidate models without the help of proxy datasets or proxy tasks. The searching procedure takes about 5.6 GPU days on Cityscapes [11].
We apply our approach to semantic image segmentation tasks on several public benchmarks, our model achieves the best performance compared with state-of-the-art hand-craft $\hat{\text{models[51,74,80,7,24,12,17] and other contempo-}}$ rary NAS approaches [36, 3, 59, 44, 8]. We also evaluate the optimal model identified by DCNAS on PASCAL VOC 2012[15],the model outperforms other leading approaches
$\bullet$ We design a novel entirely densely connected search
space, allowing to explore various existing designs and
to cover arbitrary model architecture patterns.
$\bullet$ A novel proxyless searching paradigm is implemented to
efficiently and directly explore the most promising model among all the candidates encoded in DCNAS, on largescale segmentation datasets $(e.g.$, Cityscapes [11]).
·The DCNAS outperforms contemporary NAS methods
[44,59,8,3,36] and demonstrates new state-of-the-art performance on Cityscapes [11], PASCAL-VOC 2012 [15],PASCAL-Context[43] and ADE20K[78] datasets.
|
[
[
0.12744140625,
0.09344482421875,
0.200927734375,
0.09344482421875,
0.200927734375,
0.10479736328125,
0.12744140625,
0.10479736328125
],
[
0.334228515625,
0.091552734375,
0.525390625,
0.091552734375,
0.525390625,
0.10479736328125,
0.334228515625,
0.10479736328125
],
[
0.68896484375,
0.0947265625,
0.73193359375,
0.0947265625,
0.73193359375,
0.10479736328125,
0.68896484375,
0.10479736328125
],
[
0.1641845703125,
0.106689453125,
0.2034912109375,
0.106689453125,
0.2034912109375,
0.11871337890625,
0.1641845703125,
0.11871337890625
],
[
0.390625,
0.10919189453125,
0.398681640625,
0.117431640625,
0.39208984375,
0.121826171875,
0.384033203125,
0.11297607421875
],
[
0.68701171875,
0.106689453125,
0.73193359375,
0.106689453125,
0.73193359375,
0.117431640625,
0.68701171875,
0.117431640625
],
[
0.351318359375,
0.11236572265625,
0.358642578125,
0.11236572265625,
0.358642578125,
0.11871337890625,
0.351318359375,
0.11871337890625
],
[
0.2099609375,
0.11236572265625,
0.2197265625,
0.11236572265625,
0.2197265625,
0.117431640625,
0.2099609375,
0.117431640625
],
[
0.420654296875,
0.11614990234375,
0.42724609375,
0.111083984375,
0.43310546875,
0.11553955078125,
0.426513671875,
0.12060546875
],
[
0.5986328125,
0.1142578125,
0.60888671875,
0.1142578125,
0.60888671875,
0.117431640625,
0.5986328125,
0.117431640625
],
[
0.136474609375,
0.125,
0.1502685546875,
0.125,
0.1502685546875,
0.1351318359375,
0.136474609375,
0.1351318359375
],
[
0.348876953125,
0.1256103515625,
0.35693359375,
0.1256103515625,
0.35693359375,
0.1287841796875,
0.348876953125,
0.1287841796875
],
[
0.3857421875,
0.1256103515625,
0.39794921875,
0.1256103515625,
0.39794921875,
0.1319580078125,
0.3857421875,
0.1319580078125
],
[
0.134033203125,
0.1514892578125,
0.1519775390625,
0.1514892578125,
0.1519775390625,
0.1654052734375,
0.134033203125,
0.1654052734375
],
[
0.134033203125,
0.179931640625,
0.15283203125,
0.179931640625,
0.15283203125,
0.1925048828125,
0.134033203125,
0.1925048828125
],
[
0.1396484375,
0.2108154296875,
0.1478271484375,
0.2108154296875,
0.1478271484375,
0.216552734375,
0.1396484375,
0.216552734375
],
[
0.365234375,
0.2266845703125,
0.431396484375,
0.2266845703125,
0.431396484375,
0.2374267578125,
0.365234375,
0.2374267578125
],
[
0.7587890625,
0.2266845703125,
0.78173828125,
0.2266845703125,
0.78173828125,
0.238037109375,
0.7587890625,
0.238037109375
],
[
0.81884765625,
0.2310791015625,
0.8349609375,
0.2310791015625,
0.8349609375,
0.2342529296875,
0.81884765625,
0.2342529296875
],
[
0.160888671875,
0.2374267578125,
0.1683349609375,
0.231689453125,
0.1732177734375,
0.236083984375,
0.1666259765625,
0.2412109375
],
[
0.2132568359375,
0.2310791015625,
0.281982421875,
0.2310791015625,
0.281982421875,
0.238037109375,
0.2132568359375,
0.238037109375
],
[
0.78515625,
0.2587890625,
0.83837890625,
0.2587890625,
0.83837890625,
0.266357421875,
0.78515625,
0.266357421875
],
[
0.53125,
0.26513671875,
0.5458984375,
0.26513671875,
0.5458984375,
0.268310546875,
0.53125,
0.268310546875
],
[
0.53125,
0.27587890625,
0.54833984375,
0.27587890625,
0.54833984375,
0.279052734375,
0.53125,
0.279052734375
],
[
0.78515625,
0.2890625,
0.83642578125,
0.2890625,
0.83642578125,
0.296630859375,
0.78515625,
0.296630859375
],
[
0.60302734375,
0.2998046875,
0.61376953125,
0.2998046875,
0.61376953125,
0.3037109375,
0.60302734375,
0.3037109375
],
[
0.62353515625,
0.3017578125,
0.646484375,
0.3017578125,
0.646484375,
0.309326171875,
0.62353515625,
0.309326171875
],
[
0.080078125,
0.33203125,
0.89208984375,
0.33203125,
0.89208984375,
0.34521484375,
0.080078125,
0.34521484375
],
[
0.5390625,
0.302978515625,
0.54833984375,
0.302978515625,
0.54833984375,
0.30615234375,
0.5390625,
0.30615234375
],
[
0.07843017578125,
0.3466796875,
0.89208984375,
0.3466796875,
0.89208984375,
0.35986328125,
0.07843017578125,
0.35986328125
],
[
0.08087158203125,
0.3623046875,
0.8896484375,
0.3623046875,
0.8896484375,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.080078125,
0.375,
0.89208984375,
0.375,
0.89208984375,
0.38818359375,
0.080078125,
0.38818359375
],
[
0.07843017578125,
0.386962890625,
0.89208984375,
0.38818359375,
0.89208984375,
0.40283203125,
0.07843017578125,
0.401611328125
],
[
0.07843017578125,
0.40283203125,
0.6708984375,
0.40283203125,
0.6708984375,
0.416748046875,
0.07843017578125,
0.416748046875
],
[
0.07843017578125,
0.44140625,
0.4697265625,
0.44140625,
0.4697265625,
0.455078125,
0.07843017578125,
0.455078125
],
[
0.50341796875,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50341796875,
0.455078125
],
[
0.080078125,
0.45703125,
0.4697265625,
0.45703125,
0.4697265625,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50244140625,
0.45703125,
0.890625,
0.45703125,
0.890625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.07843017578125,
0.4716796875,
0.4697265625,
0.4716796875,
0.4697265625,
0.486083984375,
0.07843017578125,
0.486083984375
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.07843017578125,
0.4873046875,
0.4697265625,
0.4873046875,
0.4697265625,
0.50048828125,
0.07843017578125,
0.50048828125
],
[
0.5009765625,
0.48486328125,
0.89208984375,
0.4873046875,
0.89208984375,
0.501953125,
0.5009765625,
0.5
],
[
0.08087158203125,
0.50390625,
0.465576171875,
0.50390625,
0.465576171875,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.5048828125,
0.50390625,
0.8896484375,
0.50390625,
0.8896484375,
0.5146484375,
0.5048828125,
0.5146484375
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.50244140625,
0.517578125,
0.65185546875,
0.517578125,
0.65185546875,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.07843017578125,
0.53271484375,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.07843017578125,
0.546875
],
[
0.521484375,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.546875,
0.521484375,
0.546875
],
[
0.080078125,
0.54736328125,
0.468017578125,
0.54736328125,
0.468017578125,
0.56103515625,
0.080078125,
0.56103515625
],
[
0.50244140625,
0.548828125,
0.89208984375,
0.548828125,
0.89208984375,
0.5625,
0.50244140625,
0.5625
],
[
0.080078125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.080078125,
0.57763671875,
0.468017578125,
0.57763671875,
0.468017578125,
0.59130859375,
0.080078125,
0.59130859375
],
[
0.50244140625,
0.5791015625,
0.89208984375,
0.5791015625,
0.89208984375,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.080078125,
0.59228515625,
0.4697265625,
0.59228515625,
0.4697265625,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.5009765625,
0.59228515625,
0.8896484375,
0.59228515625,
0.8896484375,
0.60595703125,
0.5009765625,
0.60595703125
],
[
0.080078125,
0.60791015625,
0.4697265625,
0.60791015625,
0.4697265625,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.50341796875,
0.60986328125,
0.8896484375,
0.60986328125,
0.8896484375,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.080078125,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.080078125,
0.63720703125
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.080078125,
0.63818359375,
0.4697265625,
0.63818359375,
0.4697265625,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.50244140625,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.08087158203125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.5009765625,
0.65283203125,
0.890625,
0.65283203125,
0.890625,
0.66650390625,
0.5009765625,
0.66650390625
],
[
0.08087158203125,
0.66845703125,
0.468017578125,
0.66845703125,
0.468017578125,
0.681640625,
0.08087158203125,
0.681640625
],
[
0.50244140625,
0.66845703125,
0.7685546875,
0.66845703125,
0.7685546875,
0.681640625,
0.50244140625,
0.681640625
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.521484375,
0.68505859375,
0.87158203125,
0.68505859375,
0.87158203125,
0.69873046875,
0.521484375,
0.69873046875
],
[
0.080078125,
0.69873046875,
0.4697265625,
0.69873046875,
0.4697265625,
0.7119140625,
0.080078125,
0.7119140625
],
[
0.07843017578125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.07843017578125,
0.72802734375
],
[
0.50341796875,
0.71044921875,
0.890625,
0.71044921875,
0.890625,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.0767822265625,
0.72900390625,
0.4697265625,
0.72802734375,
0.4697265625,
0.7421875,
0.0767822265625,
0.74365234375
],
[
0.5185546875,
0.7265625,
0.89208984375,
0.7265625,
0.89208984375,
0.74072265625,
0.5185546875,
0.74072265625
],
[
0.07843017578125,
0.744140625,
0.4697265625,
0.74365234375,
0.4697265625,
0.7568359375,
0.07843017578125,
0.75830078125
],
[
0.5185546875,
0.7412109375,
0.8203125,
0.7412109375,
0.8203125,
0.7548828125,
0.5185546875,
0.7548828125
],
[
0.080078125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.77294921875,
0.080078125,
0.77294921875
],
[
0.07843017578125,
0.7734375,
0.449462890625,
0.77294921875,
0.449462890625,
0.787109375,
0.07843017578125,
0.78857421875
],
[
0.5048828125,
0.76806640625,
0.890625,
0.76806640625,
0.890625,
0.7822265625,
0.5048828125,
0.7822265625
],
[
0.5185546875,
0.7841796875,
0.89208984375,
0.7841796875,
0.89208984375,
0.79736328125,
0.5185546875,
0.79736328125
],
[
0.1004638671875,
0.79443359375,
0.468017578125,
0.79443359375,
0.468017578125,
0.8095703125,
0.1004638671875,
0.8095703125
],
[
0.51708984375,
0.798828125,
0.8896484375,
0.7998046875,
0.8896484375,
0.81298828125,
0.51708984375,
0.8125
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5185546875,
0.814453125,
0.85107421875,
0.814453125,
0.85107421875,
0.8291015625,
0.5185546875,
0.8291015625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.5185546875,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5185546875,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5185546875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5185546875,
0.88525390625
],
[
0.07843017578125,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.51953125,
0.88623046875,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.51953125,
0.90087890625
]
] |
[
"Scale Input",
"DenselyConnectedSearchSpace",
"Output",
"STEM",
"6",
"Upsample",
"4",
"0",
"6",
"-",
"1/4",
"",
"",
"/g",
"/16",
"/o",
"Mixture Layer",
"④",
"aad",
"",
"FusionModule",
"FusionModule",
"33",
"RA",
"Fixedoperator",
"k",
"Em,e",
"Figure 1. Framework of DCNAS. Top: The densely connected search space (DCSS) with layer L and max downsampling rate 32. To",
"",
" stable the searching procedure, we keep the beginning STEM and final Upsample block unchanged. Dashed lines represent candidate",
"connections inDCsS,tokeepclarity,we onlydemonstrateseveral connections among all thecandidates.Bottom Left:Fusionmodule",
"targets at aggregating feature-maps derived by previous layers, and solving the intensive GPU memory requirement problem by sampling a",
"portion of connections from all possible ones. Bottom Right: Mixture layer may further save GPU memory consumption and accelerating",
"the searching process by sampling and operating on a portion of features while bypassing the others.",
"sequently, DCNAS may derive whichever model architec-",
"[26, 17, 74, 68, 69, 27, 70, 21] and advances the state-of-",
"turedesigns by selecting appropriate paths among the whole",
"the-art performance. Transferring the model to ADE2OK",
" set of these connections. Additionally, DCNAS aggregates",
"[78] and PASCAL-Context [43] datasets, our model obtains",
"contextual semantics encoded in multi-scale imageries to",
"the best result compared with state-of-the-art approaches",
"derive long-range context information,which has beenre-",
"[12,17,22,24,71,69,31l.according to thecorrespond-",
"cently found to be vital in the state-of-the-art manually de-",
"ing evaluation metrics.",
"signed image segmentation models [40, 74, 5.7.701. Sec-",
"Further, considering NAS methods depend on agent met-",
"ondly, we observe that previous NAS approaches [3, 36]",
"rics to explore promising networks efficiently. If the agent",
"heavily rely on the proxy task or proxy dataset to reduce the",
"metric provides a strong relative ranking of models, this en-",
"cost of GPU hours and to alleviate the high GPU memory",
"ables the discovery of high performing models when trained",
"consumption problem.However, architectures optimized",
"to convergence. However, studies about the relative ranking",
"over the proxy are not guaranteed to be suitable in realistic",
"ofmodelsinthesearchingperiodarelessexploredinlitera-",
"setting [2], because of the gap between the proxy and the",
"ture [44, 59, 8, 3, 36]. In this work, we enrich this field and",
"target configuration. Regarding this concern, we relax the",
"investigate most related methods on how well the accuracy",
"discretearchitecturesintocontinuousrepresentationtosave",
"in searching phase for specific architectures correlates with",
"GPU hours and design a fusion module which applies both",
"that of the fine-tuned stand-alone model.",
"path-level and channel-level sampling strategies during the",
"To summarize, our main contributions are as follows:",
"searching procedure to reduce memory demand. Based on",
"that, one may employ the stochastic gradient descent (SGD)",
". We design a novel entirely densely connected search",
"to perform the proxyless searching procedure to select the",
"space, allowing to explore various existing designs and",
"optimal architecture from all the candidatemodels without",
"to cover arbitrary model architecture patterns.",
"the help of proxy datasets or proxy tasks. The searching",
"procedure takes about 5.6 GPU days on Cityscapes [11].",
"·A novel proxyless searching paradigm is implemented to",
"efficiently and directly explore the most promising model",
"We apply our approach to semantic image segmentation",
"among all the candidates encoded in DCNAS, on large-",
"tasks on several public benchmarks, our model achieves the",
"scale segmentation datasets (e.g., Cityscapes [11]).",
"best performance compared with state-of-the-art hand-craft",
"models [51, 74, 80, 7, 24,12,17] and other contempo-",
"· The DCNAS outperforms contemporary NAS methods",
"rary NAS approaches [36, 3, 59, 44, 8]. We also evaluate",
"[44, 59, 8, 3, 36] and demonstrates new state-of-the-art",
"the optimal model identified by DCNAS on PASCAL VOC",
"performance on Cityscapes [11], PASCAL-VOC 2012",
"2012 [15], the model outperforms other leading approaches",
"[15], PASCAL-Context [43] and ADE20K [78] datasets."
] |
{
"arxiv_id": "2003.11883",
"language": "en",
"page_idx": 1
}
|
||
2003.11883-en-2
|
2003.11883-en-2.jpg
|
$\bullet$ We conduct a large scale experiment to understand the
correlation of model performance between the searching and training periods for most contemporary works [36, 3, 59, 44, 8] focusing on semantic segmentation task.
## $2. \textbf{ Related Work}$
### $2. 1. \textbf{Neural Architecture Search Space}$
Early NAS works[81,49] directly search the whole network architecture. Although those works achieve impressive results, their expensive computation overheads (e.g. thousands of GPU days) hinder their applications in realworld scenarios. To alleviate this situation, researchers proposed restricted search spaces. NASNet[82] first proposed a cell-based search space. Concretely, NASet searches the fundamental cell structure, which is easier than searches the whole network architectures on a small proxy dataset. After that, many works [38,47,2,56,58] adopt the cell-based search space design and make improvements in several ways. One line of works [2,56,58] attempt to search highefficiency network architectures for resource-constrained platforms, such as mobile phones. Another line of works [9, 16] try to improve previous cell-based NAS methods. PDRATS [9] first bridge the depth gap between search and evaluation phases. Fang $et$ al.[16] proposes the routing blocks to determine the number of blocks automatically.
Despite the success of repeated cell-based methods for image classification tasks, other dense image prediction tasks, such as semantic image segmentation and object detection, demand more delicate and complicated network architectures to capture multi-scale information of the image. Therefore, researchers begin to explore more flexible search space for dense image prediction tasks. For instance, DPC [3] proposes a recursive search space to build a multi-scale representation of a high-resolution image. Auto-DeepLab [36] introduces a hierarchical search space to exhibit multilevel representations of the image.
## 2.2. Semantic Image Segmentation
The pioneer work FCN[42] makes great progress in semantic segmentation with a fully convolutional network. Afterwards, researchers devote enormous efforts to explore the tremendous potentials of convolutional neural networks (CNNs) on dense image prediction tasks. Several kinds of research works [74,34,6,14,4,50,46,33,48,7,18] focus on how to efficiently utilize multi-scale context information to improve the performance of semantic segmentation systems. Some works [34,6,14] takes an image pyramid as input to capture larger objects in down-sampled input image. Some other works try to make use of multi-scale context information. PSPNet[74] uses spatial pyramid pooling at different manually set scales. Deeplab[4] performs multiple rates atrous convolution operations. Also, there are some
works [50,46,33,48,7,18] that employ encoder-decoder
structure to captures long-range context information.
Recent works [3, 36, 44, 52, 59, 8, 30, 73, 72, 60, 41] shift to automatically design network architectures. Research works, such as [44,52,59,8,30,73], target to search network structures that are friendly to resource-constrained devices for semantic segmentation task. Other works, such as DPC[3] and Auto-DeepLab[36], intent to directly search the highest-performance network architecture for semantic segmentation without considering computation overheads. Most of those works [36,44,52,59,8,30,73] suffer from limited search space and search on proxy tasks, because of high GPU memory consumption and expensive computational burden. Yet the optimal network structures for multiscale feature representations require ample search space.
The most similar works to ours are DPC[3], AutoDeepLab[36], and DenseNAS[16]. [3] proposes an ample search space but requires thousands of GPU days for searching, whereas we can efficiently search over at least the same search space using only 5.6 GPU days. [36] achieves comparable searching time like ours, yet contains a much smaller search space than DCNAS. [16] places candidate connections between adjacent $M$ cells only, while DCNAS comprises a wholly densely connected search space. Besides, both [3, 36, 16] search on a proxy dataset, whereas DCNAS directly search on the target dataset.
# 3. Methodology
In this section, we introduce several vital components that constitute DCNAS in detail, including a densely connected search space that is general enough to cover enormous network design, the derived continuously differentiable representation of the search space, and the approach that can save the cost of GPU hours and reduce excessive memory footprint, which favors proxyless searching.
## 3.1. Densely Connected Search Space
The densely connected search space( DCSS) involves two primitives, the mixture layer, and the fusion module. The mixture layer is defined as a mixture of the candidate operators, and the fusion module aims at aggregating semantic features from preceding fusion modules. A range of reticular-like and densely connected fusion modules constitute the architecture of DCSS, as shown in Figure l.
Mixture Layer. The mixture layer is the elementary structure in search space represents a collection of available operations. Similar to[2], we construct the operator space $\hat{\mathcal{O}}$ with various configurations of mobileNet-v3[23], i.e., kernel sizes $k\in\{3,5,7\}$, expansion ratios $r\in\{3,6\}.$ Due to the dense connections contained in the search space, the DCSS inherently supports the zero operation and identity mapping. Hence $\mathcal{O}$ does not contain the above operators
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.0931396484375,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.0931396484375,
0.12060546875
],
[
0.50244140625,
0.10797119140625,
0.84716796875,
0.10797119140625,
0.84716796875,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.09478759765625,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.09478759765625,
0.136962890625
],
[
0.521484375,
0.12371826171875,
0.8896484375,
0.12371826171875,
0.8896484375,
0.136962890625,
0.521484375,
0.136962890625
],
[
0.09478759765625,
0.136962890625,
0.42724609375,
0.136962890625,
0.42724609375,
0.15087890625,
0.09478759765625,
0.15087890625
],
[
0.50341796875,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.15283203125,
0.50341796875,
0.15283203125
],
[
0.50244140625,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.08087158203125,
0.1673583984375,
0.2181396484375,
0.1673583984375,
0.2181396484375,
0.1817626953125,
0.08087158203125,
0.1817626953125
],
[
0.50341796875,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.50244140625,
0.184326171875,
0.89208984375,
0.184326171875,
0.89208984375,
0.1976318359375,
0.50244140625,
0.1976318359375
],
[
0.080078125,
0.1932373046875,
0.373291015625,
0.1932373046875,
0.373291015625,
0.20703125,
0.080078125,
0.20703125
],
[
0.5048828125,
0.2008056640625,
0.890625,
0.2008056640625,
0.890625,
0.2108154296875,
0.5048828125,
0.2108154296875
],
[
0.098876953125,
0.2171630859375,
0.468017578125,
0.2171630859375,
0.468017578125,
0.2310791015625,
0.098876953125,
0.2310791015625
],
[
0.50341796875,
0.2139892578125,
0.89208984375,
0.2139892578125,
0.89208984375,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.080078125,
0.2342529296875,
0.46728515625,
0.2342529296875,
0.46728515625,
0.2442626953125,
0.080078125,
0.2442626953125
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.50244140625,
0.24365234375
],
[
0.080078125,
0.2474365234375,
0.468017578125,
0.2474365234375,
0.468017578125,
0.261474609375,
0.080078125,
0.261474609375
],
[
0.50244140625,
0.2442626953125,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.50244140625,
0.25830078125
],
[
0.07843017578125,
0.261474609375,
0.468017578125,
0.261962890625,
0.468017578125,
0.276611328125,
0.07843017578125,
0.27587890625
],
[
0.50244140625,
0.260009765625,
0.89208984375,
0.260009765625,
0.89208984375,
0.2734375,
0.50244140625,
0.2734375
],
[
0.07843017578125,
0.276611328125,
0.468017578125,
0.27783203125,
0.468017578125,
0.291748046875,
0.07843017578125,
0.290283203125
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.080078125,
0.292236328125,
0.4697265625,
0.292236328125,
0.4697265625,
0.306884765625,
0.080078125,
0.306884765625
],
[
0.50244140625,
0.290283203125,
0.890625,
0.290283203125,
0.890625,
0.3037109375,
0.50244140625,
0.3037109375
],
[
0.08087158203125,
0.309326171875,
0.46728515625,
0.309326171875,
0.46728515625,
0.3193359375,
0.08087158203125,
0.3193359375
],
[
0.5009765625,
0.30615234375,
0.875,
0.30615234375,
0.875,
0.3193359375,
0.5009765625,
0.3193359375
],
[
0.08087158203125,
0.324462890625,
0.468017578125,
0.324462890625,
0.468017578125,
0.335205078125,
0.08087158203125,
0.335205078125
],
[
0.51953125,
0.3193359375,
0.89208984375,
0.32080078125,
0.89208984375,
0.335205078125,
0.51953125,
0.333984375
],
[
0.080078125,
0.33837890625,
0.468017578125,
0.33837890625,
0.468017578125,
0.3515625,
0.080078125,
0.3515625
],
[
0.50341796875,
0.3359375,
0.890625,
0.3359375,
0.890625,
0.349853515625,
0.50341796875,
0.349853515625
],
[
0.07843017578125,
0.354248046875,
0.4697265625,
0.354248046875,
0.4697265625,
0.367431640625,
0.07843017578125,
0.367431640625
],
[
0.50244140625,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.080078125,
0.36865234375,
0.468017578125,
0.36865234375,
0.468017578125,
0.3818359375,
0.080078125,
0.3818359375
],
[
0.50341796875,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.37890625,
0.50341796875,
0.37890625
],
[
0.50244140625,
0.3818359375,
0.85302734375,
0.3818359375,
0.85302734375,
0.395751953125,
0.50244140625,
0.395751953125
],
[
0.8544921875,
0.38330078125,
0.8896484375,
0.38330078125,
0.8896484375,
0.39453125,
0.8544921875,
0.39453125
],
[
0.5009765625,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.410888671875,
0.5009765625,
0.410888671875
],
[
0.08087158203125,
0.385009765625,
0.46728515625,
0.385009765625,
0.46728515625,
0.395751953125,
0.08087158203125,
0.395751953125
],
[
0.080078125,
0.3984375,
0.4697265625,
0.3984375,
0.4697265625,
0.412353515625,
0.080078125,
0.412353515625
],
[
0.080078125,
0.41357421875,
0.4697265625,
0.41357421875,
0.4697265625,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50244140625,
0.412353515625,
0.890625,
0.412353515625,
0.890625,
0.425537109375,
0.50244140625,
0.425537109375
],
[
0.08087158203125,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.50341796875,
0.4287109375,
0.8896484375,
0.4287109375,
0.8896484375,
0.439453125,
0.50341796875,
0.439453125
],
[
0.080078125,
0.443115234375,
0.4697265625,
0.443115234375,
0.4697265625,
0.45703125,
0.080078125,
0.45703125
],
[
0.50244140625,
0.442626953125,
0.8896484375,
0.442626953125,
0.8896484375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.080078125,
0.458984375,
0.4697265625,
0.458984375,
0.4697265625,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.50341796875,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.50341796875,
0.4716796875
],
[
0.080078125,
0.473388671875,
0.44775390625,
0.473388671875,
0.44775390625,
0.4873046875,
0.080078125,
0.4873046875
],
[
0.5009765625,
0.47021484375,
0.79833984375,
0.4716796875,
0.79833984375,
0.486083984375,
0.5009765625,
0.48486328125
],
[
0.1004638671875,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.501953125,
0.1004638671875,
0.501953125
],
[
0.07843017578125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.07843017578125,
0.51806640625
],
[
0.5009765625,
0.50244140625,
0.63232421875,
0.5048828125,
0.63232421875,
0.51953125,
0.5009765625,
0.517578125
],
[
0.080078125,
0.521484375,
0.465576171875,
0.521484375,
0.465576171875,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.080078125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.080078125,
0.548828125
],
[
0.521484375,
0.53076171875,
0.89208984375,
0.53076171875,
0.89208984375,
0.54443359375,
0.521484375,
0.54443359375
],
[
0.080078125,
0.5498046875,
0.46728515625,
0.5498046875,
0.46728515625,
0.56298828125,
0.080078125,
0.56298828125
],
[
0.50244140625,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.08251953125,
0.56640625,
0.468017578125,
0.56640625,
0.468017578125,
0.5771484375,
0.08251953125,
0.5771484375
],
[
0.50244140625,
0.56103515625,
0.890625,
0.56103515625,
0.890625,
0.57470703125,
0.50244140625,
0.57470703125
],
[
0.080078125,
0.580078125,
0.4697265625,
0.580078125,
0.4697265625,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.5009765625,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.5009765625,
0.59033203125
],
[
0.080078125,
0.5947265625,
0.468017578125,
0.59326171875,
0.468017578125,
0.60791015625,
0.080078125,
0.609375
],
[
0.50244140625,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.60498046875,
0.50244140625,
0.60498046875
],
[
0.080078125,
0.60986328125,
0.4697265625,
0.60986328125,
0.4697265625,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.62060546875,
0.50244140625,
0.62060546875
],
[
0.080078125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.5009765625,
0.62060546875,
0.84716796875,
0.62060546875,
0.84716796875,
0.63525390625,
0.5009765625,
0.63525390625
],
[
0.080078125,
0.6416015625,
0.3056640625,
0.6416015625,
0.3056640625,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.50341796875,
0.6494140625,
0.7890625,
0.6494140625,
0.7890625,
0.66357421875,
0.50341796875,
0.66357421875
],
[
0.07843017578125,
0.6640625,
0.346435546875,
0.66552734375,
0.346435546875,
0.6796875,
0.07843017578125,
0.6787109375
],
[
0.521484375,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.521484375,
0.6875
],
[
0.09722900390625,
0.689453125,
0.468017578125,
0.689453125,
0.468017578125,
0.7041015625,
0.09722900390625,
0.7041015625
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.71630859375,
0.50341796875,
0.71630859375
],
[
0.08087158203125,
0.72021484375,
0.471435546875,
0.72021484375,
0.471435546875,
0.734375,
0.08087158203125,
0.734375
],
[
0.50244140625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.75,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.08087158203125,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50244140625,
0.76513671875,
0.8447265625,
0.76513671875,
0.8447265625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.08087158203125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.521484375,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.080078125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
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,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50244140625,
0.8388671875
],
[
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
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.07843017578125,
0.869140625
],
[
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
],
[
0.5009765625,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
" We conduct a large scale experiment to understand the",
"works [50, 46, 33, 48, 7, 18] that employ encoder-decoder",
"correlation of model performance between the searching",
"structure to captures long-range context information.",
"and training periods for most contemporary works [36, 3,",
"Recent works [3, 36, 44, 52,59, 8, 30, 73, 72, 60, 41]",
"59, 44, 8] focusing on semantic segmentation task.",
"shift to automatically design network architectures. Re-",
"search works, such as [44, 52, 59, 8, 30, 73], target to search",
"2.Related Work",
"network structures that are friendly to resource-constrained",
"devices for semantic segmentation task. Other works, such",
"2.1. Neural Architecture Search Space",
"asDPC[3landAuto-DeepLab[36l.intenttodirectlysearch",
"Early NAS works [81, 49] directly search the whole net-",
"the highest-performance network architecture for semantic",
"workarchitecture.Althoughthoseworksachieveimpres-",
"segmentation without considering computation overheads.",
"sive results, their expensive computation overheads (e.g.",
"Most of those works [36, 44, 52, 59, 8, 30, 73] suffer from",
"thousands of GPU days) hinder their applications in real-",
"limited search space and search on proxy tasks, because of",
"world scenarios. To alleviate this situation, researchers pro-",
"high GPU memory consumption and expensive computa-",
"posed restricted search spaces. NASNet [82] first proposed",
"tional burden. Yet the optimal network structures for multi-",
"acell-based searchspace.Concretely.NASetsearchesthe",
"scale feature representations require ample search space.",
"fundamental cell structure,whichiseasierthan searchesthe",
"The most similar works to ours are DPC [3], Auto-",
"whole network architectures on a small proxy dataset. Af-",
"DeepLab [36], and DenseNAS [16]. [3] proposes an am-",
"ter that, many works [38, 47, 2, 56, 58] adopt the cell-based",
"ple search space but requires thousands of GPU days for",
"search space design and make improvements in several",
"searching,whereas we can efficiently search over at least",
"the same search space using only 5.6 GPU days.",
"[36]",
"achieves comparable searching time like ours, yet contains a",
"ways.Onelineofworks[2,56,58]attempttosearchhigh-",
"efficiency network architectures for resource-constrained",
"platforms, such as mobile phones. Another line of works",
"much smaller search space than DCNAS. [16] places candi-",
"[9, 16] try to improve previous cell-based NAS methods. P-",
"date connections between adjacent M cells only,whileDC-",
"DRATS [9] first bridge the depth gap between search and",
"NAS comprises a wholly densely connected search space.",
"evaluation phases. Fang et al.[16] proposes the routing",
"Besides, both [3, 36, 16] search on a proxy dataset, whereas",
"blocks to determine the number of blocks automatically.",
"DCNAS directly search on the target dataset.",
"Despitethe successof repeated cell-basedmethodsfor",
"image classification tasks, other dense image prediction",
"3.Methodology",
"tasks,such as semantic image segmentation and object de",
"tection, demand more delicate and complicated network ar-",
"In this section, we introduce several vital components",
"chitectures to capture multi-scale information of the image.",
"that constitute DCNAS in detail, including a densely con-",
"Therefore,researchersbegintoexploremoreflexiblesearch",
"nected search space that is general enough to cover enor-",
"space for dense image prediction tasks. For instance, DPC",
"mous network design, the derived continuously differen-",
"[3] proposes a recursive search space to build a multi-scale",
"tiable representation of the search space, and the approach",
"representation of a high-resolution image. Auto-DeepLab",
"that can save the cost of GPU hours and reduce excessive",
"[36] introduces a hierarchical search space to exhibit multi-",
"memory footprint, which favors proxyless searching.",
"level representations of the image.",
"3.1. Densely Connected Search Space",
"2.2. Semantic Image Segmentation",
"The densely connected search space （DCSS) involves",
"The pioneer work FCN [42] makes great progress in se-",
"two primitives, the mixture layer, and the fusion module.",
"mantic segmentation with a fully convolutional network.",
"The mixture layeris defined as a mixture of the candidate",
"Afterwards, researchers devote enormous efforts to explore",
"operators, and the fusion module aims at aggregating se-",
"thetremendouspotentials of convolutionalneural networks",
"mantic features from preceding fusion modules. A range of",
"(CNNs) on dense image prediction tasks. Several kinds of",
"reticular-like and densely connected fusion modules consti-",
"research works[74.34.6.14.4.50.46.33.48.7.181focus",
"tute the architecture of DCSS, as shown in Figure 1.",
"onhowtoefficientlyutilizemulti-scalecontextinformation",
"Mixture Layer. The mixture layer is the elementary",
"to improve the performance of semantic segmentation sys-",
"structure in search space represents a collection of available",
"tems. Some works [34, 6, 14] takes an image pyramid as in-",
"operations. Similar to [2], we construct the operator space",
"put to capture larger objects in down-sampled input image.",
"O with various configurations of mobileNet-v3 [23], i.e.,",
"Some other works try to make use of multi-scale context in-",
"kernel sizes k E (3, 5, 7), expansion ratios r E f3, 6]. Due",
"formation. PSPNet [74] uses spatial pyramid pooling at dif-",
"to the dense connections contained in the search space, the",
"ferent manually set scales. Deeplab [4] performs multiple",
"DCSS inherently supports the zero operation and identity",
"rates atrous convolution operations. Also, there are some",
"mapping. Hence O does not contain the above operators"
] |
{
"arxiv_id": "2003.11883",
"language": "en",
"page_idx": 2
}
|
||
2003.11883-en-3
|
2003.11883-en-3.jpg
|
<FigureHere>
Figure 2. Mainstream Network Designs in DCNAS. The figure demonstrates several mainstream network design patterns [42,46, 5,54] that embedded in our Densely Connected Search Space.
explicitly. Similarly, considering the search space inherently favors the multi-scale semantic features aggregation, which increases the receptive field, therefore, the mixture layer does not support atrous convolution explicitly.
Fusion Module. To explore various paths in DCSS, we introduce the fusion module with the ability of aggregating semantic features from preceding fusion modules besides the feature pyramids and attaching transformed semantic features to succeeding ones. As illustrated by the bottomleft part in Figure 1, the fusion module consists of the shape-alignment layer besides the mixture layer. The shapealignment layer is in a multi-branch parallel form. Given an array of feature-maps with different shapes (e.g., spatial resolutions and channel widths), the shape-alignment layer dispatches each feature-map to the corresponding branches to align them to the target shape. Semantic features are well-aligned and fully aggregated, then feed into the mixture layer to perform efficient multi-scale features fusion.
Search Space. Benefiting from the sophisticated design of the fusion module, we introduce a set of reticulararranged fusion modules and place dense connections between them to form a search space that supports more flexible architecture configurations than similar works [36,8]. The search algorithm is admitted to select a subset of the fusion models together with the appropriate connections be-
tween them to derive whatever model architecture (see Figure 2 for instance), hence enables a general architecture search. Specifically, we construct the super network with an array of fusion modules $\{\mathcal{M}_{(s,l)}\}$, where $s$ indicates the spatial confguration and $l$ refers to the layer index, which is $\hat{\text{demonstrated in Figure }1}$,in which, each $\mathcal{M}_(s,l)$ aggregates the semantic features coming from $\{\mathcal{M}_{(s^{\prime},l^{\prime})}\}_{l^{\prime}<l}.$
## $3. 2. \textbf{Differentiable Representation}$
Reinforcement learning representations in [1,81] and evolutionary representations in [53,49] both tend to be computationally intensive, hence probably not suitable for semantic image segmentation task. Regarding this concern, we draw on the experience of continuous relaxation in [38, 62] and relax the discrete architectures into continuous representation. Notably, we derive a continuous representation for both the mixture layer and the fusion module, hence leading to a fully differentiable search space, based on which one can apply the stochastic gradient descent method to search promising architectures.
Mixture Layer. We assign an architecture parameter $\alpha_{(s,l)}^{o}$ to each operator $o$ in mixture layer $\ell_(s,l)$ contained ${\mathrm{in~fusion~module~}\mathcal{M}}_{(s,l)},{\mathrm{and~derive~the~continuously~rep-}}$ resentation of the mixture layer by defining it as a weighted sum of outputs from all candidate operations. The architecture weight of the operation is computed as a soft-max over all architecture parameters in mixture layer $\ell_{(s,l)}$,
$$w_{(s,l)}^{o}=\frac{\exp\{\alpha_{(s,l)}^{o}\}}{\sum_{o'\in\mathcal{O}}\exp\{\alpha_{(s,l)}^{o'}\}},$$
(1)
and the output of mixture layer $\ell_{(s,l)}$ can be estimated as,
(2)
$$O_{(s,l)}=\sum_{o\in\mathcal{O}}w_{(s,l)}^{o}o\left(\mathcal{I}_{(s,l)}\right),$$
where $\mathcal{I}_(s,l)$ refers to the input feature-maps of $\ell_(s,l).$
Fusion Module. Similar to the relaxation paradigm adopted in the mixture layer, to relax the connections as a continuous representation, we assign an architecture param- $\operatorname*{eter}\beta_{(s^{\prime},l^{\prime})\to(s,l)}\operatorname*{for~the~path~from}\mathcal{M}_{(s^{\prime},l^{\prime})}$to$\mathcal{M}_{(s,l)}.$Since $\begin{array}{l}{\mathrm{the~}\mathcal{M}_{(s,l)}}&{\mathrm{aggregates~all~semantic~features~}\{\mathcal{I}_{(s^{\prime},l^{\prime})}\}_{l^{\prime}<l},}\\{\mathrm{we~estimate~the~transmission~probability~of~each~path~using}}\end{array}$ a soft-max function over all available connections,
$$p_{(s',l')\to(s,l)}=\frac{\exp\left\{\beta_{(s',l')\to(s,l)}\right\}}{\sum_{s''\in\mathbb{S},l''<l}\exp\left\{\beta_{(s'',l'')\to(s,l)}\right\}},$$
(3)
and the aggregating process can be formalized as,
$$\mathcal{I}_{(s,l)}=\sum_{s'\in\mathbb{S},l'<l}p_{(s',l')\to(s,l)}H_{(s',l')\to(s,l)}(O_{(s',l')}),$$
where $H_{(s^{\prime},l^{\prime}\to(s,l)}$ denotes the corresponding transformation branch in shape-alignment layer of fusion module $\begin{array}{l}\mathcal{M}_{(s,l)},\text{ which transforms the semantic features produced}\\\text{by mixture layer }\ell_{(s^{\prime},l^{\prime})}\text{ to the shape specified by }\mathcal{M}_{(s,l)}.\end{array}$
|
[
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.16015625,
0.1558837890625,
0.38818359375,
0.1558837890625,
0.38818359375,
0.169189453125,
0.16015625,
0.169189453125
],
[
0.5009765625,
0.15283203125,
0.89208984375,
0.1514892578125,
0.89208984375,
0.1666259765625,
0.5009765625,
0.1673583984375
],
[
0.50244140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.18310546875,
0.50244140625,
0.18310546875
],
[
0.10455322265625,
0.18310546875,
0.1446533203125,
0.189453125,
0.1405029296875,
0.2008056640625,
0.101318359375,
0.1944580078125
],
[
0.5,
0.1787109375,
0.8349609375,
0.1817626953125,
0.8349609375,
0.199462890625,
0.5,
0.1962890625
],
[
0.181396484375,
0.1900634765625,
0.2320556640625,
0.1900634765625,
0.2320556640625,
0.2008056640625,
0.181396484375,
0.2008056640625
],
[
0.426513671875,
0.1849365234375,
0.443603515625,
0.1881103515625,
0.441162109375,
0.1976318359375,
0.423095703125,
0.1944580078125
],
[
0.50244140625,
0.203857421875,
0.7646484375,
0.2052001953125,
0.7646484375,
0.2197265625,
0.50244140625,
0.2183837890625
],
[
0.52294921875,
0.2291259765625,
0.89208984375,
0.2291259765625,
0.89208984375,
0.242431640625,
0.52294921875,
0.242431640625
],
[
0.1568603515625,
0.2474365234375,
0.3896484375,
0.2474365234375,
0.3896484375,
0.25830078125,
0.1568603515625,
0.25830078125
],
[
0.50244140625,
0.2442626953125,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.50244140625,
0.25830078125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.108642578125,
0.276611328125,
0.11358642578125,
0.276611328125,
0.11358642578125,
0.28271484375,
0.108642578125,
0.28271484375
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.50341796875,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.3017578125,
0.50341796875,
0.3017578125
],
[
0.50244140625,
0.3037109375,
0.8935546875,
0.3037109375,
0.8935546875,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.50244140625,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.400390625,
0.32763671875,
0.416748046875,
0.32763671875,
0.416748046875,
0.330810546875,
0.400390625,
0.330810546875
],
[
0.1544189453125,
0.335205078125,
0.393798828125,
0.3359375,
0.393798828125,
0.35107421875,
0.1544189453125,
0.349853515625
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.50341796875,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.108642578125,
0.369384765625,
0.11358642578125,
0.369384765625,
0.11358642578125,
0.375,
0.108642578125,
0.375
],
[
0.50341796875,
0.365478515625,
0.890625,
0.365478515625,
0.890625,
0.37890625,
0.50341796875,
0.37890625
],
[
0.50244140625,
0.380126953125,
0.81787109375,
0.380126953125,
0.81787109375,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.52294921875,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.4091796875,
0.52294921875,
0.4091796875
],
[
0.5009765625,
0.4091796875,
0.8935546875,
0.4072265625,
0.8935546875,
0.4248046875,
0.5009765625,
0.4267578125
],
[
0.1666259765625,
0.425537109375,
0.38232421875,
0.425537109375,
0.38232421875,
0.439453125,
0.1666259765625,
0.439453125
],
[
0.50244140625,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.442626953125,
0.4697265625,
0.442626953125,
0.4697265625,
0.455810546875,
0.080078125,
0.455810546875
],
[
0.5009765625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.5009765625,
0.453857421875
],
[
0.080078125,
0.45703125,
0.468017578125,
0.45703125,
0.468017578125,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50244140625,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.08087158203125,
0.472900390625,
0.44775390625,
0.472900390625,
0.44775390625,
0.48291015625,
0.08087158203125,
0.48291015625
],
[
0.50244140625,
0.47021484375,
0.8935546875,
0.47021484375,
0.8935546875,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.5009765625,
0.481689453125,
0.82861328125,
0.48486328125,
0.82861328125,
0.50244140625,
0.5009765625,
0.5
],
[
0.6806640625,
0.50244140625,
0.765625,
0.505859375,
0.7646484375,
0.5263671875,
0.6787109375,
0.5234375
],
[
0.080078125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.080078125,
0.53076171875
],
[
0.5986328125,
0.51513671875,
0.65673828125,
0.521484375,
0.65283203125,
0.54052734375,
0.5947265625,
0.5341796875
],
[
0.86767578125,
0.51953125,
0.89208984375,
0.51953125,
0.89208984375,
0.53271484375,
0.86767578125,
0.53271484375
],
[
0.6689453125,
0.52587890625,
0.7890625,
0.52783203125,
0.7890625,
0.546875,
0.6689453125,
0.54443359375
],
[
0.080078125,
0.53271484375,
0.468017578125,
0.53271484375,
0.468017578125,
0.546875,
0.080078125,
0.546875
],
[
0.08087158203125,
0.54736328125,
0.4697265625,
0.54736328125,
0.4697265625,
0.56103515625,
0.08087158203125,
0.56103515625
],
[
0.5009765625,
0.55029296875,
0.87841796875,
0.55029296875,
0.87841796875,
0.568359375,
0.5009765625,
0.568359375
],
[
0.080078125,
0.56298828125,
0.419189453125,
0.56298828125,
0.419189453125,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.6748046875,
0.5771484375,
0.7939453125,
0.5771484375,
0.7939453125,
0.59521484375,
0.6748046875,
0.59521484375
],
[
0.86767578125,
0.57763671875,
0.89453125,
0.57763671875,
0.89453125,
0.5947265625,
0.86767578125,
0.5947265625
],
[
0.6005859375,
0.5791015625,
0.6650390625,
0.5791015625,
0.6650390625,
0.59326171875,
0.6005859375,
0.59326171875
],
[
0.098876953125,
0.58056640625,
0.468017578125,
0.58056640625,
0.468017578125,
0.5947265625,
0.098876953125,
0.5947265625
],
[
0.080078125,
0.5966796875,
0.4697265625,
0.5966796875,
0.4697265625,
0.60986328125,
0.080078125,
0.60986328125
],
[
0.65869140625,
0.59521484375,
0.68701171875,
0.59521484375,
0.68701171875,
0.60693359375,
0.65869140625,
0.60693359375
],
[
0.08087158203125,
0.61279296875,
0.4697265625,
0.61279296875,
0.4697265625,
0.62353515625,
0.08087158203125,
0.62353515625
],
[
0.498291015625,
0.609375,
0.85107421875,
0.60986328125,
0.85107421875,
0.630859375,
0.498291015625,
0.6298828125
],
[
0.080078125,
0.626953125,
0.4697265625,
0.626953125,
0.4697265625,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.52294921875,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.6416015625,
0.52294921875,
0.6416015625
],
[
0.08087158203125,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.65380859375,
0.08087158203125,
0.65380859375
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64306640625,
0.89208984375,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.65576171875,
0.468017578125,
0.65576171875,
0.468017578125,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.67041015625,
0.50341796875,
0.67041015625
],
[
0.080078125,
0.67236328125,
0.468017578125,
0.67236328125,
0.468017578125,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.5009765625,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68994140625,
0.5009765625,
0.68994140625
],
[
0.07843017578125,
0.6875,
0.468017578125,
0.6875,
0.468017578125,
0.70068359375,
0.07843017578125,
0.70068359375
],
[
0.5009765625,
0.6875,
0.89208984375,
0.6875,
0.89208984375,
0.705078125,
0.5009765625,
0.705078125
],
[
0.080078125,
0.7021484375,
0.468017578125,
0.7021484375,
0.468017578125,
0.71630859375,
0.080078125,
0.71630859375
],
[
0.50244140625,
0.7021484375,
0.89208984375,
0.7041015625,
0.89208984375,
0.71826171875,
0.50244140625,
0.71630859375
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.5009765625,
0.71728515625,
0.83251953125,
0.71728515625,
0.83251953125,
0.73095703125,
0.5009765625,
0.73095703125
],
[
0.080078125,
0.73193359375,
0.4697265625,
0.73193359375,
0.4697265625,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.66650390625,
0.73779296875,
0.7998046875,
0.73779296875,
0.7998046875,
0.755859375,
0.66650390625,
0.755859375
],
[
0.080078125,
0.74755859375,
0.4697265625,
0.74755859375,
0.4697265625,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.5185546875,
0.75048828125,
0.61669921875,
0.75244140625,
0.61669921875,
0.7666015625,
0.5185546875,
0.7646484375
],
[
0.62158203125,
0.7548828125,
0.8486328125,
0.755859375,
0.8486328125,
0.7763671875,
0.62158203125,
0.77587890625
],
[
0.869140625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.869140625,
0.76318359375
],
[
0.080078125,
0.76220703125,
0.46728515625,
0.76220703125,
0.46728515625,
0.77587890625,
0.080078125,
0.77587890625
],
[
0.07843017578125,
0.7763671875,
0.455078125,
0.7763671875,
0.455078125,
0.79052734375,
0.07843017578125,
0.79052734375
],
[
0.50244140625,
0.77978515625,
0.82666015625,
0.77978515625,
0.82666015625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.09722900390625,
0.79443359375,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.09722900390625,
0.8095703125
],
[
0.58642578125,
0.8046875,
0.61669921875,
0.8046875,
0.61669921875,
0.81640625,
0.58642578125,
0.81640625
],
[
0.61279296875,
0.8046875,
0.89208984375,
0.8037109375,
0.89208984375,
0.8212890625,
0.61279296875,
0.82275390625
],
[
0.080078125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.51708984375,
0.80615234375,
0.57666015625,
0.80615234375,
0.57666015625,
0.8212890625,
0.51708984375,
0.8212890625
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.57177734375,
0.82275390625,
0.62841796875,
0.82275390625,
0.62841796875,
0.833984375,
0.57177734375,
0.833984375
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84033203125,
0.8935546875,
0.84033203125,
0.8935546875,
0.85791015625,
0.50244140625,
0.85791015625
],
[
0.07843017578125,
0.85498046875,
0.46728515625,
0.85595703125,
0.46728515625,
0.87060546875,
0.07843017578125,
0.869140625
],
[
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
],
[
0.5009765625,
0.87060546875,
0.8935546875,
0.86865234375,
0.8935546875,
0.88623046875,
0.5009765625,
0.88818359375
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.875,
0.88623046875,
0.875,
0.90380859375,
0.5009765625,
0.90380859375
]
] |
[
"tween them to derive whatever model architecture (see Fig-",
"ure 2 for instance), hence enables a general architecture",
"search. Specifically, we construct the super network with",
"an array of fusion modules [M(s,)], where s indicates the",
"(a) Network Architecture of FCN-32s [42]",
"spatial configuration and I refers to the layer index,which is",
"demonstrated in Figure 1, in which, each M(s,1) aggregates",
"",
"the semantic features coming from (M(s',1\")lu\"<l.",
"8",
"",
"3.2.Differentiable Representation",
"Reinforcement learning representations in [1, 81] and",
"(b)NetworkArchitecture ofHourGlass[46]",
"evolutionary representations in [53, 49] both tend to be",
"computationally intensive, hence probably not suitable for",
"",
"semantic image segmentation task. Regarding this con-",
"cern, we draw on the experience of continuous relaxation",
"in [38, 62] and relax the discrete architectures into continu-",
"ous representation. Notably, we derive a continuous repre-",
"",
"(c) Network Architecture of DeepLabV3 [5]",
"sentation for both the mixture layer and the fusion mod-",
"ule, hence leading to a fully differentiable search space,",
"E",
"based on which one can apply the stochastic gradient de-",
"scent method to search promising architectures.",
"Mixture Layer. We assign an architecture parameter",
"(s,) to each operator o in mixture layer l(s,1) contained",
"(d) Network Architecture of HRNet [54]",
"in fusion module M(s,l), and derive the continuously rep-",
"Figure 2. Mainstream Network Designs in DCNAS. The figure",
"resentation of the mixture layer by defining it as a weighted",
"demonstrates several mainstream network design patterns [42, 46,",
"sum of outputs from all candidate operations. The architec-",
"5,54lthatembeddedinourDenselyConnectedSearchSpace.",
"ture weight of the operation is computed as a soft-max over",
"all architecture parameters in mixture layer l(s,1),",
"exp((s,)]",
"explicitly. Similarly, considering the search space inher-",
"wis,n) =",
"(1)",
"Co\"eo exp(as,)]",
"ently favors the multi-scale semantic features aggregation,",
"which increases the receptive field, therefore, the mixture",
"and the output of mixture layer l(s,t) can be estimated as,",
"layer does not support atrous convolution explicitly.",
"Cws,1)o (I(s,1) ,",
"(2)",
"O(s,1) =",
"Fusion Module. To explore various paths in DCSS, we",
"introduce the fusion module with the ability of aggregating",
"OEO",
"semanticfeaturesfromprecedingfusionmodulesbesides",
"where I(s,) refers to the input feature-maps of l(s,1).",
"the feature pyramids and attaching transformed semantic",
"Fusion Module.Similar to the relaxation paradigm",
"features to succeedingones. As illustrated bythebottom-",
"adopted in the mixture layer, to relax the connections as a",
"left part in Figure 1, the fusion module consists of the",
"continuousrepresentation,weassignanarchitectureparam-",
"shape-alignment layer besides the mixture layer. The shape-",
"eter β(s,\")→(s,1) for thepath from M(s,\") to M(s,t). Since",
"alignment layer is in a multi-branch parallel form. Given",
"the M(s,1) aggregates all semantic features [(s',u)lv<l,",
"an array of feature-maps with different shapes (e.g, spatial",
"we estimate the transmission probability of each path using",
"resolutions and channel widths), the shape-alignment layer",
"a soft-max function over all available connections.",
"dispatches each feature-map to the corresponding branches",
"exp [β(s*,1)→(s,1)",
"to align them to the target shape. Semantic features are",
"P(s1\")→(s,l) =",
"Es\"es, \"<1 exp (β(s\",\")→(s,)]",
"(3)",
"well-aligned and fully aggregated, then feed into the mix-",
"ture layer to perform efficient multi-scale features fusion.",
"and the aggregating process can be formalized as.",
"Search Space. Benefiting from the sophisticated de-",
"1",
"P(s'1)-→(s,t)H(s,u)-(s,)(O(s'1)), (4)",
"sign of the fusion module, we introduce a set of reticular-",
"L(s,1) =",
"arranged fusion modules and place dense connections be-",
"s'es,U'<l",
"tween them to form a search space that supports more flex-",
"where H(s',l'-→(s,t) denotes the corresponding transforma-",
"ible architecture configurations than similar works [36, 8].",
"tion branch in shape-alignment layer of fusion module",
"The search algorithm is admitted to select a subset of the",
"M(s,1), which transforms the semantic features produced",
"fusion models together with the appropriate connections be-",
"by mixture layer l(s',\") to the shape specified by M(s,)."
] |
{
"arxiv_id": "2003.11883",
"language": "en",
"page_idx": 3
}
|
||
2003.11883-en-4
|
2003.11883-en-4.jpg
|
# 3.3. Reducing Memory Footprint
Continuous representation of network architectures can largely reduce the cost of GPU hours, while the memory footprint grows linearly w.r.t. the size of candidate operation set and the number of connections, hence suffering from the high GPU memory consumption problem. While image segmentation intrinsically requires high-resolution semantic features, which may lead to even more GPU memory consumption. To resolve the excessive memory consumption problem, we apply the sampling strategy both in the fusion module and the mixture layer to solve the excessive memory footprint problem and further reduce the cost of GPU hours. Taking the above tactics, one can proxylessly search the model over large-scale dataset, hence avoiding the gap between target and proxy dataset.
Fusion Module. The DCSS comprises $8L(L+1)$ connections in total, where $L$ is layer number, thus making it impossible to optimize the whole set of candidate connections in each iteration because of the extravagant GPU memory demand. Therefore, in each iteration during the search, for each fusion module, we sample several connections among the corresponding potential connections. Concretely, taking fusion module $\mathcal{M}_{(s,l)}$ for instance, all the preceding fusion modules $\{\mathcal{M}_{(s^{\prime},l^{\prime})}\}_{l^{\prime}<l}^{\prime}$ are associated with it, in each search iteration, we perform sampling without replacement to activate several transmission paths, and the probability distribution is defined as:
(5)
$$p_{(s',l')\to(s,l)}=\frac{\exp\left\{\beta_{(s',l')\to(s,l)}/\tau\right\}}{\sum_{s''\in\mathbb{S},l''<l}\exp\left\{\beta_{(s'',l'')\to(s,l)}/\tau\right\}},$$
where $\beta_(s^{\prime\prime},l^{\prime\prime})\to(s,l)$ share the same definition in Equation 3, and the temperature variable $\tau$ starts from a high temperature then anneal to a small but non-zero value. Assume there are $n$ fusion modules $\{\mathcal{M}_s_i,l_i\}_{i<n}$ selected from $\{\mathcal{M}_{(s^{\prime},l^{\prime})}\}_{l^{\prime}<l}$,then we may approximate Equation 4 with:
(6)
$$\mathcal{I}_{(s,l)}=\sum_{k<n}w_{(s_{k},l_{k})\to(s,l)}H_{(s_{k},l_{k})\to(s,l)}(O_{(s_{k},l_{k})}),$$
$\begin{array}{l}{\mathrm{where~}w_{(s_{k},l_{k})\to(s,l)}}&{\mathrm{refers~to~the~normalized~blending}}\\{\mathrm{weight,~}H\mathrm{~and~}O\mathrm{~shares~the~same~definition~in~Equation~4.}}\\{\mathrm{weight,~}H\mathrm{~and~}O\mathrm{~shares~the~same~definition~in~Equation~4.}}\end{array}$
Mixture Layer. A typical way to apply the sampling trick is to sample the operator $o$ from the candidate operator space $\mathcal{O}$ in each mixture layer. While in practice, we find that the naive sampling strategy makes the searching process unstable and sometime does not convergence. Inspired by[64], a portion of channels are sampled among the input features, then the lucky ones are sent into the mixed transformation of $|\mathcal{O}|$ operators while bypasses the others. Specifically, for each $\ell_{(s,l)}$,we assign a random variable $S_{(s,l)}$ that control the ratio of sampled channels, which trades off the search accuracy and efficiency, and we may
estimate $O_{(s,l)}$ in Equation 2 as:
(7)
$$\sum_{o\in\mathcal{O}}w_{(s,l)}^{o}o\left(S_{(s,l)}\mathcal{I}_{(s,l)}\right)+(1-S_{(s,l)})\mathcal{I}_{(s,l)},$$
$\begin{array}{l}\mathrm{where~}w_{(s,l)}^{o}&\mathrm{indicates~the~normalized~blending~weight~de-}\\\mathrm{fined~by~Equation~1.~While~different~to~[64],~the~edge}\end{array}$ normalization layers are removed and the feature masks $S_{(s,l)}$ remain unchanged in the searching procedure, and experimental result demonstrates that this simplified variant proves more suitable for semantic image segmentation. Regarding the sampling ratio, we do not tune the sampling ratio and fix it to be $^1/4$, empirically, our experiments demonstrate that it works well. With this scheme, one shall gain $4\times$ acceleration of the searching process, and save about $75\%$ GPU memory consumption.
### 3.4. Search Procedure
Benefiting from the fully differentiable representation of the search space and the sampling strategies mentioned above, one may apply the stochastic gradient descent algorithm to search for appropriate models by optimizing architecture parameters $\{\alpha,\beta\}$ on the target dataset without proxy. Similar to previous works [38,3,36,8],we partition the training data into two parts trainA and $trainB$, which are used for updating the convolutional weights $w$ and architecture parameters $\{\alpha,\beta\}$ respectively. We solve this optimization problem alternatively:
$\bullet$ update weights $w$ by $\nabla_w\mathcal{L}_{trainA}(w,\alpha,\beta)$,
$\bullet$ update parameters $\{\alpha,\beta\}$ by $\nabla_\alpha,\beta\mathcal{L}_{trainB}(w,\alpha,\beta)$,
in which the loss function $\mathcal{L}$ mainly consists of the classical
cross entropy calculated on each mini-batch.
We shall point out that, in practice, $\mathcal{L}$ comprises of several practical regularization terms in addition to crossentropy, experiments reveal the regularization terms yield a faster convergence rate of the searching procedure and lead to a better searching accuracy (see supplementary material).
## 3.5. Decoding Network Structure
Once the search procedure terminates, one may derive the suitable operator for each mixture layer and the optimal architecture based on the architecture parameters $\alpha$ and $\beta.$ For mixture layer $\ell_(s,l)$, we select the candidate operation that has maximum operation weight, $i.e.$, $\arg\max_{o\in\mathcal{O}}\alpha_{(s,l)}^{o}.$ In terms of the network architecture, we use Breadth-First Search algorithm (see supplemental material) to derive the network architecture.
## $4. \textbf{Experiments}$
To evaluate DCNAS comprehensively, we first verify the effect of the proxyless searching paradigm by measuring the
|
[
[
0.08087158203125,
0.091552734375,
0.334228515625,
0.091552734375,
0.334228515625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.5009765625,
0.091552734375,
0.71337890625,
0.091552734375,
0.71337890625,
0.10479736328125,
0.5009765625,
0.10479736328125
],
[
0.09722900390625,
0.11553955078125,
0.4697265625,
0.11614990234375,
0.4697265625,
0.1300048828125,
0.09722900390625,
0.1287841796875
],
[
0.54736328125,
0.11932373046875,
0.859375,
0.121826171875,
0.859375,
0.1427001953125,
0.54736328125,
0.14013671875
],
[
0.08087158203125,
0.1307373046875,
0.468017578125,
0.1307373046875,
0.468017578125,
0.14453125,
0.08087158203125,
0.14453125
],
[
0.859375,
0.1256103515625,
0.8896484375,
0.1256103515625,
0.8896484375,
0.136962890625,
0.859375,
0.136962890625
],
[
0.080078125,
0.1458740234375,
0.46728515625,
0.1458740234375,
0.46728515625,
0.1590576171875,
0.080078125,
0.1590576171875
],
[
0.54833984375,
0.14013671875,
0.57861328125,
0.14013671875,
0.57861328125,
0.1514892578125,
0.54833984375,
0.1514892578125
],
[
0.07843017578125,
0.160400390625,
0.4697265625,
0.1610107421875,
0.4697265625,
0.175537109375,
0.07843017578125,
0.1749267578125
],
[
0.5,
0.160400390625,
0.8935546875,
0.1572265625,
0.8935546875,
0.177978515625,
0.5,
0.18115234375
],
[
0.080078125,
0.1767578125,
0.4697265625,
0.1767578125,
0.4697265625,
0.1900634765625,
0.080078125,
0.1900634765625
],
[
0.50341796875,
0.177978515625,
0.89208984375,
0.177978515625,
0.89208984375,
0.1925048828125,
0.50341796875,
0.1925048828125
],
[
0.080078125,
0.1912841796875,
0.468017578125,
0.1912841796875,
0.468017578125,
0.2052001953125,
0.080078125,
0.2052001953125
],
[
0.50341796875,
0.1932373046875,
0.890625,
0.1932373046875,
0.890625,
0.20703125,
0.50341796875,
0.20703125
],
[
0.080078125,
0.205810546875,
0.468017578125,
0.205810546875,
0.468017578125,
0.2197265625,
0.080078125,
0.2197265625
],
[
0.50244140625,
0.208984375,
0.89208984375,
0.20703125,
0.89208984375,
0.2215576171875,
0.50244140625,
0.2235107421875
],
[
0.080078125,
0.2215576171875,
0.46728515625,
0.2215576171875,
0.46728515625,
0.23486328125,
0.080078125,
0.23486328125
],
[
0.50244140625,
0.2235107421875,
0.89208984375,
0.2235107421875,
0.89208984375,
0.2374267578125,
0.50244140625,
0.2374267578125
],
[
0.080078125,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.50244140625,
0.2392578125,
0.89208984375,
0.2392578125,
0.89208984375,
0.25244140625,
0.50244140625,
0.25244140625
],
[
0.080078125,
0.251953125,
0.46728515625,
0.251953125,
0.46728515625,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.5009765625,
0.25244140625,
0.8935546875,
0.25390625,
0.8935546875,
0.268310546875,
0.5009765625,
0.267578125
],
[
0.080078125,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.27978515625,
0.080078125,
0.27978515625
],
[
0.50244140625,
0.26953125,
0.89208984375,
0.26953125,
0.89208984375,
0.28271484375,
0.50244140625,
0.28271484375
],
[
0.08087158203125,
0.2822265625,
0.46728515625,
0.28271484375,
0.46728515625,
0.294189453125,
0.08087158203125,
0.29345703125
],
[
0.50244140625,
0.2841796875,
0.89208984375,
0.2841796875,
0.89208984375,
0.29736328125,
0.50244140625,
0.29736328125
],
[
0.07843017578125,
0.29541015625,
0.4697265625,
0.296630859375,
0.4697265625,
0.311279296875,
0.07843017578125,
0.31005859375
],
[
0.50341796875,
0.2998046875,
0.89208984375,
0.2998046875,
0.89208984375,
0.313232421875,
0.50341796875,
0.313232421875
],
[
0.080078125,
0.3125,
0.35546875,
0.3125,
0.35546875,
0.32568359375,
0.080078125,
0.32568359375
],
[
0.50341796875,
0.314453125,
0.72216796875,
0.314453125,
0.72216796875,
0.32763671875,
0.50341796875,
0.32763671875
],
[
0.098876953125,
0.32763671875,
0.46728515625,
0.32763671875,
0.46728515625,
0.341552734375,
0.098876953125,
0.341552734375
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.080078125,
0.356689453125
],
[
0.50341796875,
0.33837890625,
0.67236328125,
0.33837890625,
0.67236328125,
0.352783203125,
0.50341796875,
0.352783203125
],
[
0.080078125,
0.35791015625,
0.468017578125,
0.35791015625,
0.468017578125,
0.371826171875,
0.080078125,
0.371826171875
],
[
0.52294921875,
0.361083984375,
0.890625,
0.361083984375,
0.890625,
0.375,
0.52294921875,
0.375
],
[
0.080078125,
0.37255859375,
0.468017578125,
0.37255859375,
0.468017578125,
0.386474609375,
0.080078125,
0.386474609375
],
[
0.50244140625,
0.376953125,
0.89208984375,
0.376953125,
0.89208984375,
0.39013671875,
0.50244140625,
0.39013671875
],
[
0.08087158203125,
0.389404296875,
0.46728515625,
0.389404296875,
0.46728515625,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.50244140625,
0.392578125,
0.890625,
0.392578125,
0.890625,
0.4072265625,
0.50244140625,
0.4072265625
],
[
0.080078125,
0.40283203125,
0.4697265625,
0.40283203125,
0.4697265625,
0.416748046875,
0.080078125,
0.416748046875
],
[
0.50244140625,
0.40771484375,
0.8896484375,
0.40771484375,
0.8896484375,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.080078125,
0.41845703125,
0.46728515625,
0.41845703125,
0.46728515625,
0.431884765625,
0.080078125,
0.431884765625
],
[
0.50244140625,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.4375,
0.50244140625,
0.4375
],
[
0.07843017578125,
0.429931640625,
0.4697265625,
0.431884765625,
0.4697265625,
0.449462890625,
0.07843017578125,
0.447509765625
],
[
0.50244140625,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.451904296875,
0.50244140625,
0.451904296875
],
[
0.080078125,
0.447509765625,
0.471435546875,
0.447509765625,
0.471435546875,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.50244140625,
0.45263671875,
0.89208984375,
0.45263671875,
0.89208984375,
0.466552734375,
0.50244140625,
0.466552734375
],
[
0.08087158203125,
0.464111328125,
0.46728515625,
0.464111328125,
0.46728515625,
0.477783203125,
0.08087158203125,
0.477783203125
],
[
0.5009765625,
0.46728515625,
0.890625,
0.466552734375,
0.890625,
0.48095703125,
0.5009765625,
0.481689453125
],
[
0.080078125,
0.477783203125,
0.4697265625,
0.479248046875,
0.4697265625,
0.49365234375,
0.080078125,
0.492431640625
],
[
0.50244140625,
0.48291015625,
0.890625,
0.48291015625,
0.890625,
0.496826171875,
0.50244140625,
0.496826171875
],
[
0.080078125,
0.49365234375,
0.346435546875,
0.49365234375,
0.346435546875,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.5009765625,
0.49755859375,
0.7099609375,
0.49755859375,
0.7099609375,
0.51220703125,
0.5009765625,
0.51220703125
],
[
0.238525390625,
0.521484375,
0.386474609375,
0.521484375,
0.386474609375,
0.5390625,
0.238525390625,
0.5390625
],
[
0.50244140625,
0.52099609375,
0.8056640625,
0.521484375,
0.8056640625,
0.5390625,
0.50244140625,
0.53857421875
],
[
0.08905029296875,
0.53271484375,
0.193603515625,
0.53515625,
0.193603515625,
0.5498046875,
0.08905029296875,
0.54736328125
],
[
0.193603515625,
0.5390625,
0.431396484375,
0.5390625,
0.431396484375,
0.56005859375,
0.193603515625,
0.56005859375
],
[
0.43701171875,
0.5341796875,
0.465576171875,
0.5341796875,
0.465576171875,
0.54541015625,
0.43701171875,
0.54541015625
],
[
0.5009765625,
0.54541015625,
0.86767578125,
0.546875,
0.86767578125,
0.564453125,
0.5009765625,
0.56298828125
],
[
0.07843017578125,
0.5673828125,
0.471435546875,
0.5673828125,
0.471435546875,
0.58837890625,
0.07843017578125,
0.58837890625
],
[
0.50244140625,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.50244140625,
0.5859375
],
[
0.080078125,
0.58544921875,
0.468017578125,
0.58544921875,
0.468017578125,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.5,
0.58837890625,
0.79736328125,
0.5869140625,
0.79736328125,
0.6015625,
0.5,
0.60302734375
],
[
0.08087158203125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.08087158203125,
0.6142578125
],
[
0.521484375,
0.60400390625,
0.89208984375,
0.60400390625,
0.89208984375,
0.6171875,
0.521484375,
0.6171875
],
[
0.08087158203125,
0.6162109375,
0.468017578125,
0.6162109375,
0.468017578125,
0.630859375,
0.08087158203125,
0.630859375
],
[
0.50341796875,
0.62060546875,
0.88623046875,
0.62060546875,
0.88623046875,
0.630859375,
0.50341796875,
0.630859375
],
[
0.08087158203125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64794921875,
0.08087158203125,
0.64794921875
],
[
0.50341796875,
0.6337890625,
0.89208984375,
0.6337890625,
0.89208984375,
0.64794921875,
0.50341796875,
0.64794921875
],
[
0.50244140625,
0.6484375,
0.89208984375,
0.6484375,
0.89208984375,
0.662109375,
0.50244140625,
0.662109375
],
[
0.1650390625,
0.66357421875,
0.468017578125,
0.662109375,
0.468017578125,
0.6796875,
0.1650390625,
0.68115234375
],
[
0.50244140625,
0.6640625,
0.890625,
0.6640625,
0.890625,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.156005859375,
0.681640625,
0.1805419921875,
0.681640625,
0.1805419921875,
0.68994140625,
0.156005859375,
0.68994140625
],
[
0.098876953125,
0.66650390625,
0.1641845703125,
0.66650390625,
0.1641845703125,
0.67822265625,
0.098876953125,
0.67822265625
],
[
0.50244140625,
0.6875,
0.75634765625,
0.6875,
0.75634765625,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.07843017578125,
0.7041015625,
0.4697265625,
0.7021484375,
0.4697265625,
0.7197265625,
0.07843017578125,
0.7216796875
],
[
0.521484375,
0.7109375,
0.890625,
0.7109375,
0.890625,
0.72607421875,
0.521484375,
0.72607421875
],
[
0.080078125,
0.72021484375,
0.464111328125,
0.72021484375,
0.464111328125,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.7265625,
0.890625,
0.7265625,
0.890625,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.098876953125,
0.734375,
0.468017578125,
0.73486328125,
0.468017578125,
0.75,
0.098876953125,
0.74853515625
],
[
0.50244140625,
0.7421875,
0.890625,
0.7421875,
0.890625,
0.755859375,
0.50244140625,
0.755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75830078125,
0.890625,
0.75830078125,
0.890625,
0.771484375,
0.50244140625,
0.771484375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.77294921875,
0.890625,
0.77294921875,
0.890625,
0.7861328125,
0.50244140625,
0.7861328125
],
[
0.07843017578125,
0.77978515625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.07843017578125,
0.79345703125
],
[
0.5,
0.7861328125,
0.89208984375,
0.7841796875,
0.89208984375,
0.8017578125,
0.5,
0.8037109375
],
[
0.08087158203125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50244140625,
0.8017578125,
0.890625,
0.8017578125,
0.890625,
0.81640625,
0.50244140625,
0.81640625
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.8173828125,
0.79150390625,
0.8173828125,
0.79150390625,
0.83056640625,
0.50244140625,
0.83056640625
],
[
0.080078125,
0.82568359375,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.8466796875,
0.63232421875,
0.8466796875,
0.63232421875,
0.861328125,
0.50341796875,
0.861328125
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.521484375,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.521484375,
0.8837890625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.8994140625
]
] |
[
"3.3. Reducing Memory Footprint",
"estimate O() in Equation 2 as:",
"Continuous representation of network architectures can",
"Zws,t)0(S(s,1)I(s,D)+ (1 - S(s,1)I(s,1),",
"largely reduce the cost of GPU hours,while the memory",
"(7)",
"footprint grows linearly w.r.t. the size of candidate oper-",
"OEO",
"ation set and the number of connections, hence suffering",
"where ws,) indicates the normalized blending weight de-",
"from the high GPU memory consumption problem. While",
"fined by Equation 1. While different to [64], the edge",
"image segmentation intrinsically requires high-resolution",
"normalization layers are removed and the feature masks",
"semantic features, which may lead to even more GPU mem-",
"S(s,) remain unchanged in the searching procedure, and",
"ory consumption.To resolve the excessive memory con-",
"experimental result demonstrates that this simplified variant",
"sumption problem, we apply the sampling strategy both in",
"proves more suitable for semantic image segmentation. Re-",
"the fusion module and the mixture layer to solve the exces-",
"garding the sampling ratio, we do not tune the sampling ra-",
"sive memory footprint problem and further reduce the cost",
"tio and fix it to be 1/4, empirically, our experiments demon-",
"ofGPUhours.Takingtheabovetactics,onecanproxylessly",
"strate that it works well. With this scheme, one shall gain",
"search the model over large-scale dataset, hence avoiding",
"4 × acceleration of the searching process, and save about",
"the gap between target and proxy dataset.",
"75% GPU memory consumption.",
"Fusion Module. The DCSS comprises 8L(L + 1) con-",
"nections in total, where L is layer number, thus making",
"3.4. Search Procedure",
"it impossible to optimize the whole set of candidate con-",
"Benefiting from the fully differentiable representation",
"nections in each iteration because of the extravagant GPU",
"of the search space and the sampling strategies mentioned",
"memory demand.Therefore,in eachiteration during the",
"above, one may apply the stochastic gradient descent al-",
"search, for each fusion module, we sample several con-",
"gorithm to search for appropriate models by optimizing ar-",
"nections among the corresponding potential connections.",
"chitecture parameters [α, β) on the target dataset without",
"Concretely, taking fusion module M(s,) for instance, all",
"proxy. Similar to previous works [38, 3, 36, 8], we partition",
"the preceding fusion modules [M(s',1)lu<1 are associated",
"the training data into two parts trainA and trainB, which are",
"with it, in each search iteration, we perform sampling with-",
"used for updating the convolutional weights w and archi-",
"out replacement to activate several transmission paths, and",
"tecture parameters [α, βB] respectively. We solve this opti-",
"the probability distribution is defined as:",
"mization problem alternatively:",
"exp [β(s*,1\")→(s,1) /T)",
" update weights w by V LtrainA(w, a, β),",
"P(st,1)-→(s,) =",
"Ds\"es,\"<Iexp [B(s\",u)-(s,)/T]",
"(5)",
". update parameters [a, β) by Va,β CtrainB(w, Q, β),",
"where β(s\",\")→(s,t) share the same definition in Equation",
"in which the loss function C mainly consists of the classical",
"3, and the temperature variable T starts from a high tem-",
"cross entropy calculated on each mini-batch.",
"perature then anneal to a small but non-zero value. As-",
"We shall point out that, in practice, L comprises of",
"sume there are n fusion modules [Mst,l Ji<n selected from",
"several practical regularization terms in addition to cross",
"[M(s',1)lu<t, then we may approximate Equation 4 with:",
"entropy, experiments reveal the regularization terms yield a",
"faster convergence rate of the searching procedure and lead",
"W(sk,ln)-(s,)H(snelk)→(s,)(O(sx,x), (6)",
"to a better searching accuracy (see supplementary material).",
"k<n",
"L(s.1) = ",
"3.5. Decoding Network Structure",
"where w(s,lx)-(s,l) refers to the normalized blending",
"Once the search procedure terminates, one may de-",
"weight, H and O shares the same definition in Equation 4.",
"rive the suitable operator for each mixture layer and the",
"Mixture Layer.A typical way to apply the sampling",
"optimal architecture based on the architecture parameters",
"trick is to sample the operator o from the candidate oper-",
"α and β. For mixture layer l(s.l), we select the candi-",
"ator space O in each mixture layer. While in practice, we",
"date operation that has maximum operation weight, i.e.,",
"find that the naive sampling strategy makes the searching",
"arg maxoe oo(s.1). In terms of the network architecture,",
"process unstable and sometime does not convergence. In-",
"we use Breadth-First Search algorithm (see supplemental",
"spired by [64], a portion of channels are sampled among",
"material) to derive the network architecture.",
"the input features, then the lucky ones are sent into the",
"mixed transformation of O| operators while bypasses the",
"4.Experiments",
"others. Specifically, for each E(s.l), we assign a random vari-",
"able S(s,1) that control the ratio of sampled channels, which",
"To evaluate DCNAS comprehensively, we first verify the",
"trades off the search accuracy and efficiency, and we may",
"effect of the proxyless searching paradigm by measuring the"
] |
{
"arxiv_id": "2003.11883",
"language": "en",
"page_idx": 4
}
|
||
1912.13471-en-0
|
1912.13471-en-0.jpg
|
$\begin{array}{ll}\mathrm{OneGAN:~Simultaneous~Unsupervised~Learning}\\\mathrm{C}&\mathrm{of~Conditional~Image~Generation,~Foreground}\\\mathrm{C}&\mathrm{Segmentation,~and~Fine-Grained~Clustering}\end{array}$
三
Yaniv Benny$^1$ and Lior Wolf$^1,2$
$^{1}$ Tel-Aviv University, Israel $^{2}$ Facebook AI Research
$\textbf{Abstract. We present a method for simultaneously learning, in an un- }$ supervised manner, (i) a conditional image generator, (ii) foreground extraction and segmentation, (iii) clustering into a two-level class hierarchy, and (iv) object removal and background completion, all done without any use of annotation. The method combines a Generative Adversarial Network and a Variational Auto-Encoder, with multiple encoders, generators and discriminators, and benefits from solving all tasks at once. The input to the training scheme is a varied collection of unlabeled images from the same domain, as well as a set of background images without a foreground object. In addition, the image generator can mix the background from one image, with a foreground that is conditioned either on that of a second image or on the index of a desired cluster. The method obtains state of the art results in comparison to the literature methods, when compared to the current state of the art in each of the tasks.
# $\xi _{1}$ Introduction
We hypothesize that solving multiple unsupervised tasks together, enables one to improve on the performance of the best methods that solve each individually. The underlying motivation is that in unsupervised learning, the structure of the data is a key source of knowledge and each task exposes a different aspect of it. We advocate for solving the various tasks in phases, where easier tasks are addressed frst, and the other tasks are introduced gradually, while constantly updating the solutions of the previous sets of tasks. The method consists of multiple networks that are trained end-to-end and side-by-side to solve multiple tasks. The method starts from learning background image synthesis and image generation of objects from a particular domain. It then advances to more complex tasks, such as clustering, semantic segmentation and object removal. Finally we show the model's ability to perform image-to-image translation. The entire learning process is unsupervised, meaning that no annotated information is used. In particular, the method does not employ class labels, segmentation masks, bounding boxes, etc. However, it does require a separate set of clean background images, which are easy to obtain in many cases.
|
[
[
0.093017578125,
0.052001953125,
0.9072265625,
0.053619384765625,
0.9072265625,
0.07470703125,
0.093017578125,
0.07232666015625
],
[
0.032623291015625,
0.06414794921875,
0.0821533203125,
0.06414794921875,
0.08697509765625,
0.245361328125,
0.037445068359375,
0.245361328125
],
[
0.10870361328125,
0.0828857421875,
0.89013671875,
0.0828857421875,
0.89013671875,
0.10400390625,
0.10870361328125,
0.10400390625
],
[
0.1243896484375,
0.10882568359375,
0.87451171875,
0.11126708984375,
0.87451171875,
0.1356201171875,
0.1243896484375,
0.134033203125
],
[
0.33203125,
0.17138671875,
0.66552734375,
0.1673583984375,
0.66552734375,
0.1893310546875,
0.33203125,
0.193359375
],
[
0.358642578125,
0.2047119140625,
0.64013671875,
0.20556640625,
0.64013671875,
0.2242431640625,
0.358642578125,
0.223388671875
],
[
0.0338134765625,
0.2291259765625,
0.08331298828125,
0.2274169921875,
0.08697509765625,
0.33642578125,
0.037445068359375,
0.337158203125
],
[
0.37451171875,
0.2225341796875,
0.61962890625,
0.2242431640625,
0.61962890625,
0.2420654296875,
0.37451171875,
0.240478515625
],
[
0.1497802734375,
0.300537109375,
0.84912109375,
0.302978515625,
0.84912109375,
0.320068359375,
0.1497802734375,
0.318359375
],
[
0.1497802734375,
0.322509765625,
0.8466796875,
0.322509765625,
0.8466796875,
0.337158203125,
0.1497802734375,
0.337158203125
],
[
0.043487548828125,
0.34521484375,
0.0784912109375,
0.34521484375,
0.0784912109375,
0.396484375,
0.043487548828125,
0.396484375
],
[
0.1473388671875,
0.339599609375,
0.8466796875,
0.34033203125,
0.8466796875,
0.35498046875,
0.1473388671875,
0.353271484375
],
[
0.1461181640625,
0.356689453125,
0.84765625,
0.355712890625,
0.84765625,
0.370361328125,
0.1461181640625,
0.371337890625
],
[
0.1473388671875,
0.37451171875,
0.84912109375,
0.373779296875,
0.84912109375,
0.38818359375,
0.1473388671875,
0.38916015625
],
[
0.04107666015625,
0.385986328125,
0.0809326171875,
0.385986328125,
0.0809326171875,
0.459716796875,
0.04107666015625,
0.459716796875
],
[
0.1497802734375,
0.392333984375,
0.8466796875,
0.392333984375,
0.8466796875,
0.406982421875,
0.1497802734375,
0.406982421875
],
[
0.1497802734375,
0.41015625,
0.84912109375,
0.41015625,
0.84912109375,
0.4248046875,
0.1497802734375,
0.4248046875
],
[
0.1473388671875,
0.42822265625,
0.84912109375,
0.428955078125,
0.84912109375,
0.443603515625,
0.1473388671875,
0.44189453125
],
[
0.1497802734375,
0.44677734375,
0.8515625,
0.44677734375,
0.8515625,
0.46142578125,
0.1497802734375,
0.46142578125
],
[
0.043487548828125,
0.455810546875,
0.0772705078125,
0.455810546875,
0.0772705078125,
0.50048828125,
0.043487548828125,
0.50048828125
],
[
0.1497802734375,
0.4638671875,
0.8466796875,
0.4638671875,
0.8466796875,
0.478515625,
0.1497802734375,
0.478515625
],
[
0.1473388671875,
0.481689453125,
0.8466796875,
0.48095703125,
0.8466796875,
0.49560546875,
0.1473388671875,
0.496337890625
],
[
0.04107666015625,
0.49462890625,
0.07244873046875,
0.4931640625,
0.07611083984375,
0.5322265625,
0.044677734375,
0.53271484375
],
[
0.1497802734375,
0.49951171875,
0.84765625,
0.49951171875,
0.84765625,
0.51416015625,
0.1497802734375,
0.51416015625
],
[
0.02777099609375,
0.5224609375,
0.09661865234375,
0.51806640625,
0.10992431640625,
0.62890625,
0.04107666015625,
0.6328125
],
[
0.1473388671875,
0.5166015625,
0.8466796875,
0.517578125,
0.8466796875,
0.5322265625,
0.1473388671875,
0.5302734375
],
[
0.1497802734375,
0.5361328125,
0.80078125,
0.5361328125,
0.80078125,
0.5498046875,
0.1497802734375,
0.5498046875
],
[
0.1292724609375,
0.59228515625,
0.310302734375,
0.59228515625,
0.310302734375,
0.60986328125,
0.1292724609375,
0.60986328125
],
[
0.0821533203125,
0.64013671875,
0.91650390625,
0.64013671875,
0.91650390625,
0.65478515625,
0.0821533203125,
0.65478515625
],
[
0.0809326171875,
0.65869140625,
0.91552734375,
0.65869140625,
0.91552734375,
0.67333984375,
0.0809326171875,
0.67333984375
],
[
0.0821533203125,
0.67919921875,
0.91650390625,
0.67919921875,
0.91650390625,
0.69384765625,
0.0821533203125,
0.69384765625
],
[
0.0809326171875,
0.69873046875,
0.92041015625,
0.69873046875,
0.92041015625,
0.71337890625,
0.0809326171875,
0.71337890625
],
[
0.0809326171875,
0.71728515625,
0.91650390625,
0.71728515625,
0.91650390625,
0.73193359375,
0.0809326171875,
0.73193359375
],
[
0.0821533203125,
0.73681640625,
0.91552734375,
0.73681640625,
0.91552734375,
0.75146484375,
0.0821533203125,
0.75146484375
],
[
0.0821533203125,
0.75732421875,
0.92041015625,
0.75732421875,
0.92041015625,
0.77197265625,
0.0821533203125,
0.77197265625
],
[
0.0809326171875,
0.7763671875,
0.91552734375,
0.7763671875,
0.91552734375,
0.79052734375,
0.0809326171875,
0.79052734375
],
[
0.0772705078125,
0.79345703125,
0.91796875,
0.79443359375,
0.91796875,
0.8125,
0.0772705078125,
0.810546875
],
[
0.0809326171875,
0.81640625,
0.91650390625,
0.8154296875,
0.91650390625,
0.82958984375,
0.0809326171875,
0.8310546875
],
[
0.0809326171875,
0.83447265625,
0.91552734375,
0.8349609375,
0.91552734375,
0.849609375,
0.0809326171875,
0.84912109375
],
[
0.0821533203125,
0.8544921875,
0.91552734375,
0.8544921875,
0.91552734375,
0.869140625,
0.0821533203125,
0.869140625
],
[
0.0809326171875,
0.873046875,
0.9130859375,
0.873046875,
0.9130859375,
0.8876953125,
0.0809326171875,
0.8876953125
],
[
0.0772705078125,
0.89013671875,
0.91650390625,
0.89111328125,
0.91650390625,
0.9091796875,
0.0772705078125,
0.908203125
],
[
0.0784912109375,
0.91162109375,
0.92041015625,
0.91162109375,
0.92041015625,
0.9287109375,
0.0784912109375,
0.9287109375
],
[
0.0784912109375,
0.931640625,
0.5859375,
0.931640625,
0.5859375,
0.9462890625,
0.0784912109375,
0.9462890625
]
] |
[
"OneGAN: Simultaneous Unsupervised Learning",
"[",
"of Conditional Image Generation, Foreground",
"Segmentation, and Fine-Grained Clustering",
"Yaniv Bennyl and Lior Wolfl,2",
"1 Tel-Aviv University, Israel",
"['S]",
"2FacebookAIResearch",
"Abstract. We present a method for simultaneouslylearning, in an un-",
"supervised manner, (i)a conditional image generator, (ii) foreground ex-",
"ZAI",
"traction andsegmentation,(i)clusteringintoatwo-levelclasshierarchy",
"and (iv)object removal and background completion, all done without",
"any use of annotation.The method combines a Generative Adversarial",
"",
"Network and a Variational Auto-Encoder, with multiple encoders,gener-",
"ators and discriminators, and benefits from solving all tasks at once.The",
"input tothe training schemeis a varied collection of unlabeled images",
"from the same domain, as well as a set of background images without a",
"[Z",
"foreground object. In addition, the image generator can mix the back-",
"ground from one image,with a foreground that is conditioned either on",
"[6]",
"that of a second image or on the index of a desired cluster. The method",
"T:AIXie",
"obtainsstate of theartresults incomparisontotheliteraturemethods",
"when compared to the current state of the art in each of the tasks.",
"Introduction",
"We hypothesize that solving multiple unsupervised tasks together,enables one",
"to improve ontheperformance of thebestmethodsthat solveeach individually.",
"The underlying motivation isthat in unsupervised learning, the structure of the",
"data is akey source of knowledge and eachtask exposes a different aspect of",
"it.We advocate for solving the various tasks in phases,where easier tasks are",
"addressed first, and the other tasks are introduced gradually. while constantly",
"updating the solutions of the previous sets of tasks. The method consists of",
"multiple networksthat are trained end-to-end and side-by-sideto solve multiple",
"tasks.The method starts from learning background image synthesis and image",
"generationofobjectsfromaparticulardomain.Itthenadvancestomorecomplex",
"tasks, such as clustering,semantic segmentation and object removal.Finally",
"we show the model's ability to perform image-to-image translation.The entire",
"learning processis unsupervised,meaningthatnoannotatedinformation is used",
"In particular, the method does not employ class labels, segmentation masks,",
"bounding boxes, etc. However, it does require a separate set of clean background",
"images,whichareeasyto obtaininmanycases."
] |
{
"arxiv_id": "1912.13471",
"language": "en",
"page_idx": 0
}
|
||
1912.13471-en-1
|
1912.13471-en-1.jpg
|
2
Y. Benny and L. Wolf
Contributions Beyond the conceptual novelty of a method that treats singlehandedly multiple unsupervised tasks, which were previously solved by individual methods, the method displays a host of technical novelties, including: (i) a novel architecture that supports multiple paths addressing multiple tasks, (ii) employing bypass paths that allow a smooth transition between autoencoding and generation based on a random seed, (iii) backpropagation through three paths in each iteration, (iv) mixup module, which applies interpolation between latent representations of the generation and reconstruction paths, and more. Due to each of these novelties, backed by the ablation studies, we obtain state of the art results compared to the literature methods in each of the individual tasks.
## $\begin{array}{cc}2&\text{Related work}\end{array}$
Since our work touches on many tasks, we focus the literature review on general concepts and on the most relevant work. Generative models are typically based on Generative Adversarial Networks [11] or Variational Auto-Encoders [18]. In addition, these two can be combined [21]. Conditional image generation conditions the output on an initial variable, most commonly, the target class. CGAN [23] and InfoGAN [7] proposed different methods to apply the condition on the discriminator. Our work is more similar to InfoGAN, since we do not use labeled data and the label is not linked to any real image and no conditional discriminator can be applied. The condition is maintained by a classifier that tries to predict the conditioned label and, as a result, forces the generator to condition the result on that label. Semantic segmentation deals with the classification of the image pixels based on their class labels. For the supervised setting Unet [24], DeepLab [4], DeepLabV3 [5], HRNet [30] have shown great performance leaps using a regression loss. For the unsupervised case, more creative solutions are considered. In [6,16,32,31,8,27,14,3] a variety of methods have been used including inpainting, learning feature representation, clustering or video frames comparison. In Clustering, deep learning methods are the current state of the art. JULE [33] and DEPICT [10], cluster based on a learned feature representation. IIC [14] trains a classifier directly.
The most similar approach to ours is FineGAN [26], which our generators and discriminators are based upon. However, there are many significant differences and additions: (i) We added a set of encoders, which are trained to support new tasks. (ii) While FineGAN employs one-hot input, our generators use coded input, which is important for our autoencoding path. (iii) We added a skip connection, followed by a mixup module that combines the bypass tensor with the pre-image tensor. The mixup also allows passing only one of the tensors, making either the bypass or the pre-image optional in each flow. (iv) We employ single foreground generator instead of FineGAN's double hierarchical design, where we have found one generator to be dominant and the second one redundant. (v) Our model uses layer normalization [1] instead of batch normalization, which better performs for large number of classes, small batch size, and alternating paths. (vi) We define a new normalization method for the generators, where GLU[9]
|
[
[
0.07489013671875,
0.0138092041015625,
0.1026611328125,
0.0138092041015625,
0.1026611328125,
0.03167724609375,
0.07489013671875,
0.03167724609375
],
[
0.158203125,
0.01543426513671875,
0.38037109375,
0.01543426513671875,
0.38037109375,
0.030059814453125,
0.158203125,
0.030059814453125
],
[
0.08331298828125,
0.05767822265625,
0.9130859375,
0.05767822265625,
0.9130859375,
0.07232666015625,
0.08331298828125,
0.07232666015625
],
[
0.0809326171875,
0.07879638671875,
0.9130859375,
0.07879638671875,
0.9130859375,
0.09344482421875,
0.0809326171875,
0.09344482421875
],
[
0.0784912109375,
0.09503173828125,
0.92041015625,
0.0966796875,
0.92041015625,
0.11456298828125,
0.0784912109375,
0.1129150390625
],
[
0.0821533203125,
0.11700439453125,
0.9130859375,
0.11700439453125,
0.9130859375,
0.131591796875,
0.0821533203125,
0.131591796875
],
[
0.0784912109375,
0.1348876953125,
0.91650390625,
0.1356201171875,
0.91650390625,
0.153564453125,
0.0784912109375,
0.1527099609375
],
[
0.0821533203125,
0.156005859375,
0.9130859375,
0.156005859375,
0.9130859375,
0.1705322265625,
0.0821533203125,
0.1705322265625
],
[
0.0809326171875,
0.1754150390625,
0.91552734375,
0.1754150390625,
0.91552734375,
0.1900634765625,
0.0809326171875,
0.1900634765625
],
[
0.0784912109375,
0.194091796875,
0.91650390625,
0.193359375,
0.91650390625,
0.2080078125,
0.0784912109375,
0.208740234375
],
[
0.0809326171875,
0.2144775390625,
0.91552734375,
0.2144775390625,
0.91552734375,
0.2291259765625,
0.0809326171875,
0.2291259765625
],
[
0.0809326171875,
0.2340087890625,
0.90087890625,
0.2340087890625,
0.90087890625,
0.24853515625,
0.0809326171875,
0.24853515625
],
[
0.0784912109375,
0.28515625,
0.10992431640625,
0.28515625,
0.10992431640625,
0.30126953125,
0.0784912109375,
0.30126953125
],
[
0.1243896484375,
0.28515625,
0.320068359375,
0.28515625,
0.320068359375,
0.302978515625,
0.1243896484375,
0.302978515625
],
[
0.0784912109375,
0.32666015625,
0.91650390625,
0.32666015625,
0.91650390625,
0.344482421875,
0.0784912109375,
0.344482421875
],
[
0.0809326171875,
0.348388671875,
0.91552734375,
0.348388671875,
0.91552734375,
0.363037109375,
0.0809326171875,
0.363037109375
],
[
0.0821533203125,
0.3671875,
0.91796875,
0.3671875,
0.91796875,
0.3818359375,
0.0821533203125,
0.3818359375
],
[
0.0772705078125,
0.38427734375,
0.91796875,
0.385986328125,
0.91796875,
0.40380859375,
0.0772705078125,
0.4013671875
],
[
0.0809326171875,
0.406982421875,
0.91552734375,
0.406982421875,
0.91552734375,
0.421630859375,
0.0809326171875,
0.421630859375
],
[
0.0809326171875,
0.4248046875,
0.91650390625,
0.4248046875,
0.91650390625,
0.44189453125,
0.0809326171875,
0.44189453125
],
[
0.0809326171875,
0.445068359375,
0.91650390625,
0.445068359375,
0.91650390625,
0.459716796875,
0.0809326171875,
0.459716796875
],
[
0.0809326171875,
0.465576171875,
0.9130859375,
0.465576171875,
0.9130859375,
0.47998046875,
0.0809326171875,
0.47998046875
],
[
0.0821533203125,
0.484130859375,
0.91552734375,
0.484130859375,
0.91552734375,
0.498779296875,
0.0821533203125,
0.498779296875
],
[
0.0772705078125,
0.501953125,
0.91552734375,
0.5029296875,
0.91552734375,
0.5205078125,
0.0772705078125,
0.51904296875
],
[
0.0809326171875,
0.52294921875,
0.91552734375,
0.52294921875,
0.91552734375,
0.53759765625,
0.0809326171875,
0.53759765625
],
[
0.0821533203125,
0.54248046875,
0.9130859375,
0.54248046875,
0.9130859375,
0.55712890625,
0.0821533203125,
0.55712890625
],
[
0.0809326171875,
0.5595703125,
0.92041015625,
0.5595703125,
0.92041015625,
0.57763671875,
0.0809326171875,
0.57763671875
],
[
0.0809326171875,
0.58154296875,
0.91650390625,
0.58154296875,
0.91650390625,
0.59619140625,
0.0809326171875,
0.59619140625
],
[
0.0821533203125,
0.60009765625,
0.91552734375,
0.60009765625,
0.91552734375,
0.61474609375,
0.0821533203125,
0.61474609375
],
[
0.0784912109375,
0.619140625,
0.91796875,
0.62158203125,
0.91796875,
0.63623046875,
0.0784912109375,
0.6337890625
],
[
0.0809326171875,
0.64013671875,
0.9130859375,
0.64013671875,
0.9130859375,
0.65478515625,
0.0809326171875,
0.65478515625
],
[
0.0809326171875,
0.65869140625,
0.91796875,
0.65869140625,
0.91796875,
0.67333984375,
0.0809326171875,
0.67333984375
],
[
0.0772705078125,
0.67578125,
0.68701171875,
0.677734375,
0.68701171875,
0.6953125,
0.0772705078125,
0.69384765625
],
[
0.11712646484375,
0.6962890625,
0.92041015625,
0.6962890625,
0.92041015625,
0.7138671875,
0.11712646484375,
0.7138671875
],
[
0.0809326171875,
0.71728515625,
0.9130859375,
0.71728515625,
0.9130859375,
0.73193359375,
0.0809326171875,
0.73193359375
],
[
0.0772705078125,
0.7353515625,
0.92041015625,
0.73583984375,
0.92041015625,
0.75390625,
0.0772705078125,
0.7529296875
],
[
0.0809326171875,
0.75732421875,
0.91650390625,
0.75732421875,
0.91650390625,
0.77197265625,
0.0809326171875,
0.77197265625
],
[
0.0809326171875,
0.7763671875,
0.9130859375,
0.7763671875,
0.9130859375,
0.79052734375,
0.0809326171875,
0.79052734375
],
[
0.0784912109375,
0.79443359375,
0.91650390625,
0.79443359375,
0.91650390625,
0.8125,
0.0784912109375,
0.8125
],
[
0.0772705078125,
0.81298828125,
0.91650390625,
0.81396484375,
0.91650390625,
0.83203125,
0.0772705078125,
0.8310546875
],
[
0.0821533203125,
0.8349609375,
0.91552734375,
0.8349609375,
0.91552734375,
0.849609375,
0.0821533203125,
0.849609375
],
[
0.0784912109375,
0.85302734375,
0.91796875,
0.85302734375,
0.91796875,
0.8701171875,
0.0784912109375,
0.8701171875
],
[
0.0784912109375,
0.873046875,
0.91552734375,
0.873046875,
0.91552734375,
0.8876953125,
0.0784912109375,
0.8876953125
],
[
0.0784912109375,
0.89208984375,
0.91650390625,
0.89208984375,
0.91650390625,
0.90673828125,
0.0784912109375,
0.90673828125
],
[
0.0821533203125,
0.912109375,
0.9130859375,
0.912109375,
0.9130859375,
0.9267578125,
0.0821533203125,
0.9267578125
],
[
0.08331298828125,
0.931640625,
0.91552734375,
0.931640625,
0.91552734375,
0.9462890625,
0.08331298828125,
0.9462890625
]
] |
[
"2",
"Y.BennvandL.Wolf",
"ContributionsBeyond the conceptual novelty of a method that treatssingle",
"handedly multipleunsupervised tasks, which were previously solved by individ-",
"ual methods, the method displays a host of technical novelties, including: (i) a",
"novel architecture that supports multiple paths addressing multiple tasks, (ii)",
"employing bypass paths that allow a smooth transition between autoencoding",
"and generation based on a random seed, (iii) backpropagation through three",
"paths in each iteration,(iv)mixupmodule, which applies interpolation between",
"latent representations of thegenerationand reconstruction paths.and more.Due",
"to each of these novelties, backed by the ablation studies, we obtain state of the",
"art results compared to the literature methods in each of the individual tasks",
"2",
"Related work",
"Since our work touches on many tasks, we focus the literature review on gen-",
"eral concepts and on the most relevant work. Generative models aretypically",
"basedonGenerativeAdversarialNetworks[11]orVariational Auto-Encoders[18]",
"In addition,thesetwocan be combined[21i].Conditional imagegeneration",
"conditions the output on an initial variable, most commonly, the target class.",
"CGAN [23] and InfoGAN [7] proposed different methods to apply the condition",
"on the discriminator. Our work is more similar to InfoGAN, since we do not",
"use labeled data and the label is not linked to any real image and no condi-",
"tional discriminator can be applied. The condition is maintained by a classifier",
"that triesto predictthe conditioned label and, as a result, forces the genera-",
"tor to condition the result on that label. Semantic segmentation deals with",
"the classification of the image pixels based on their class labels.For the super",
"vised setting Unet [24], DeepLab [4], DeepLabV3 [5], HRNet [30] have shown",
"great performance leaps using a regression loss.Forthe unsupervised case, more",
"creative solutions are considered.In [6,16.32,31.8,27,14.3]a variety of methods",
"havebeen used including inpainting,learningfeaturerepresentation,clustering",
"or video frames comparison.In Clustering,deep learning methods arethe cur-",
"rent stateoftheart.JULE[33] and DEPICT[iol.clusterbasedonalearned",
"feature representation.IIC [14] trains a classifier directly.",
"The most similar approach to ours is FineGAN [26], which our generators",
"and discriminators are based upon. However, there are many significant differ-",
"ences and additions: (i) We added a set of encoders, which are trained to support",
"newtasks.(ii)WhileFineGANemploysone-hot input,ourgeneratorsuse coded",
"input,whichisimportantforourautoencodingpath.(iii)Weaddedaskip con",
"nection, followed by a mixup module that combines the bypass tensor with the",
"pre-image tensor: The mixup also allows passing only one of the tensors, making",
"either thebypassorthepre-image optional ineachflow.(iv)Weemploysingle",
"foreground generator instead of FineGAN's double hierarchical design, where we",
"havefound onegeneratortobe dominant andthesecond oneredundant.(v)Our",
"model useslayernormalization[i]instead of batchnormalization,whichbetter",
"performsforlarge number of classes,small batch size,and alternating paths",
"(vi)Wedefineanewnormalizationmethodforthegenerators,whereGLUJ9"
] |
{
"arxiv_id": "1912.13471",
"language": "en",
"page_idx": 1
}
|
||
1912.13471-en-2
|
1912.13471-en-2.jpg
|
OneGAN
3
activation layers were used as non-linear activations. (vii) We add many losses, regularization terms, and training techniques that were not used in FineGAN, many of which are completely novel, as far as we can ascertain. As a result, our work outperforms FineGAN in all tasks and is capable of performing new tasks that its predecessor could not handle.
Mixup [35] is a technique for applying a weighted sum between two latent variables in order to synthesize a new latent variable. We use it to merge different paths in the model by mixing latent variables that are part coded by the encoder and part produced by the lower levels of the generation. As far as we can ascertain, this is the first usage of mixup to merge information from different paths. Our mixup is applied to image reconstruction in four different locations.
# 3 Method
To solve the tasks of clustering, foreground segmentation, and conditional generation with minimal supervision, our method trains multiple neural networks side-by-side. Each task is solved by applying the networks in a task-dependent order. Similarly, the method is trained by applying the networks in two different paths in each iteration, each path is optimized with a specific set of losses.
$\textbf{Architecture}$ The architecture of the compound network consists of two generators, three encoders and two discriminators. Fig.1 and Fig.2 illustrate the two training paths. In the generation path, the generators produce a synthetic image conditioned on selected code, the encoders then retrieve the latent code from the generated images. In the reconstruction path, the encoders code an input image into latent code which is used by the generators to reconstruct the image. The reconstruction path is applied twice in each iteration, once on real images and once on fake images from the generation path. The reconstruction of fake images adds multiple capabilities of self-supervision such as reconstructing the background and mask, which is not applicable with real images without additional supervision. The sub-networks are fully detailed in the supplementary. Generators The generation is performed by merging the results of two separate generators that run in parallel to produce the output image. One generator is dedicated for generating the background and the other for the foreground. The generators are conditioned on a two-level hierarchical categorization. Each category has a unique child class $\phi_c$ and a parent class $\phi_p$ shared by multiple child classes. These classes are represented by the one-hot vectors $(e_c,e_p).$ An additional background one-hot vector $e_bg$ affects the generation of the background images. Since there is a tight coupling between the class of the object (water bird, tropical bird, etc.) and the expected background, the typology of the background follows the coarse hierarchy, i.e. the parent class. The generator architecture is influenced by.
(1)
$$e_{c}[i]=\delta_{i,\phi_{c}},\quad e_{p}[i]=\delta_{i,\phi_{p}},\quad e_{bg}=e_{p}$$
The generation starts by converting the one-hot vectors into code vectors using learned embeddings. Such an embedding is often used when working with
|
[
[
0.74267578125,
0.01543426513671875,
0.8408203125,
0.01543426513671875,
0.8408203125,
0.033294677734375,
0.74267578125,
0.033294677734375
],
[
0.8974609375,
0.0138092041015625,
0.923828125,
0.0138092041015625,
0.923828125,
0.033294677734375,
0.8974609375,
0.033294677734375
],
[
0.0809326171875,
0.056854248046875,
0.91552734375,
0.056854248046875,
0.91552734375,
0.07470703125,
0.0809326171875,
0.07470703125
],
[
0.0821533203125,
0.07879638671875,
0.9130859375,
0.07879638671875,
0.9130859375,
0.09344482421875,
0.0821533203125,
0.09344482421875
],
[
0.0809326171875,
0.0982666015625,
0.91552734375,
0.0982666015625,
0.91552734375,
0.11212158203125,
0.0809326171875,
0.11212158203125
],
[
0.0821533203125,
0.11700439453125,
0.91552734375,
0.11700439453125,
0.91552734375,
0.131591796875,
0.0821533203125,
0.131591796875
],
[
0.0821533203125,
0.1373291015625,
0.475830078125,
0.1373291015625,
0.475830078125,
0.15185546875,
0.0821533203125,
0.15185546875
],
[
0.113525390625,
0.1551513671875,
0.91650390625,
0.153564453125,
0.91650390625,
0.17138671875,
0.113525390625,
0.1722412109375
],
[
0.0821533203125,
0.1754150390625,
0.91552734375,
0.1754150390625,
0.91552734375,
0.1900634765625,
0.0821533203125,
0.1900634765625
],
[
0.0821533203125,
0.194091796875,
0.91650390625,
0.194091796875,
0.91650390625,
0.208740234375,
0.0821533203125,
0.208740234375
],
[
0.0772705078125,
0.212890625,
0.92041015625,
0.2120361328125,
0.92041015625,
0.2298583984375,
0.0772705078125,
0.230712890625
],
[
0.0809326171875,
0.2340087890625,
0.9130859375,
0.2340087890625,
0.9130859375,
0.24853515625,
0.0809326171875,
0.24853515625
],
[
0.0809326171875,
0.25341796875,
0.9130859375,
0.25341796875,
0.9130859375,
0.26806640625,
0.0809326171875,
0.26806640625
],
[
0.0784912109375,
0.300537109375,
0.10992431640625,
0.300537109375,
0.10992431640625,
0.317626953125,
0.0784912109375,
0.317626953125
],
[
0.10748291015625,
0.305419921875,
0.11834716796875,
0.305419921875,
0.11834716796875,
0.31103515625,
0.10748291015625,
0.31103515625
],
[
0.11474609375,
0.2998046875,
0.2427978515625,
0.2998046875,
0.2427978515625,
0.318359375,
0.11474609375,
0.318359375
],
[
0.0821533203125,
0.339599609375,
0.91162109375,
0.339599609375,
0.91162109375,
0.353271484375,
0.0821533203125,
0.353271484375
],
[
0.0809326171875,
0.35986328125,
0.91552734375,
0.35986328125,
0.91552734375,
0.373779296875,
0.0809326171875,
0.373779296875
],
[
0.0784912109375,
0.377685546875,
0.91650390625,
0.377685546875,
0.91650390625,
0.3955078125,
0.0784912109375,
0.3955078125
],
[
0.0784912109375,
0.397216796875,
0.91650390625,
0.397216796875,
0.91650390625,
0.4150390625,
0.0784912109375,
0.4150390625
],
[
0.0821533203125,
0.41845703125,
0.8671875,
0.41845703125,
0.8671875,
0.432861328125,
0.0821533203125,
0.432861328125
],
[
0.0821533203125,
0.437744140625,
0.260986328125,
0.437744140625,
0.260986328125,
0.45166015625,
0.0821533203125,
0.45166015625
],
[
0.264404296875,
0.43701171875,
0.91552734375,
0.43701171875,
0.91552734375,
0.45166015625,
0.264404296875,
0.45166015625
],
[
0.0784912109375,
0.455810546875,
0.91650390625,
0.454833984375,
0.91650390625,
0.472900390625,
0.0784912109375,
0.4736328125
],
[
0.0821533203125,
0.476806640625,
0.91552734375,
0.476806640625,
0.91552734375,
0.491455078125,
0.0821533203125,
0.491455078125
],
[
0.0772705078125,
0.49462890625,
0.91650390625,
0.4931640625,
0.91650390625,
0.5107421875,
0.0772705078125,
0.51171875
],
[
0.0809326171875,
0.5166015625,
0.91650390625,
0.5166015625,
0.91650390625,
0.5302734375,
0.0809326171875,
0.5302734375
],
[
0.0809326171875,
0.53369140625,
0.91552734375,
0.53369140625,
0.91552734375,
0.54833984375,
0.0809326171875,
0.54833984375
],
[
0.0772705078125,
0.55224609375,
0.92041015625,
0.5517578125,
0.92041015625,
0.5693359375,
0.0772705078125,
0.5703125
],
[
0.0809326171875,
0.57373046875,
0.91796875,
0.57373046875,
0.91796875,
0.58837890625,
0.0809326171875,
0.58837890625
],
[
0.0772705078125,
0.5908203125,
0.91796875,
0.59130859375,
0.91796875,
0.609375,
0.0772705078125,
0.607421875
],
[
0.0809326171875,
0.61083984375,
0.92041015625,
0.61083984375,
0.92041015625,
0.62890625,
0.0809326171875,
0.62890625
],
[
0.0809326171875,
0.6318359375,
0.91552734375,
0.6318359375,
0.91552734375,
0.646484375,
0.0809326171875,
0.646484375
],
[
0.0784912109375,
0.6484375,
0.91552734375,
0.64990234375,
0.91552734375,
0.6669921875,
0.0784912109375,
0.666015625
],
[
0.0821533203125,
0.671875,
0.91796875,
0.671875,
0.91796875,
0.6865234375,
0.0821533203125,
0.6865234375
],
[
0.0809326171875,
0.6904296875,
0.91552734375,
0.6904296875,
0.91552734375,
0.705078125,
0.0809326171875,
0.705078125
],
[
0.0821533203125,
0.7099609375,
0.91552734375,
0.7099609375,
0.91552734375,
0.724609375,
0.0821533203125,
0.724609375
],
[
0.0784912109375,
0.72705078125,
0.91552734375,
0.72705078125,
0.91552734375,
0.744140625,
0.0784912109375,
0.744140625
],
[
0.0784912109375,
0.74658203125,
0.91650390625,
0.74755859375,
0.91650390625,
0.76513671875,
0.0784912109375,
0.7646484375
],
[
0.0821533203125,
0.767578125,
0.9130859375,
0.767578125,
0.9130859375,
0.7822265625,
0.0821533203125,
0.7822265625
],
[
0.0809326171875,
0.7861328125,
0.91796875,
0.7861328125,
0.91796875,
0.80322265625,
0.0809326171875,
0.80322265625
],
[
0.0821533203125,
0.80517578125,
0.92041015625,
0.80517578125,
0.92041015625,
0.82275390625,
0.0821533203125,
0.82275390625
],
[
0.0784912109375,
0.82373046875,
0.91796875,
0.82470703125,
0.91796875,
0.84228515625,
0.0784912109375,
0.841796875
],
[
0.0772705078125,
0.84326171875,
0.38525390625,
0.84423828125,
0.38525390625,
0.86181640625,
0.0772705078125,
0.861328125
],
[
0.29345703125,
0.87548828125,
0.703125,
0.87744140625,
0.703125,
0.8984375,
0.29345703125,
0.89697265625
],
[
0.88427734375,
0.87548828125,
0.92138671875,
0.87548828125,
0.92138671875,
0.8974609375,
0.88427734375,
0.8974609375
],
[
0.0809326171875,
0.9091796875,
0.91552734375,
0.91162109375,
0.91552734375,
0.9287109375,
0.0809326171875,
0.9267578125
],
[
0.0809326171875,
0.931640625,
0.91552734375,
0.931640625,
0.91552734375,
0.9462890625,
0.0809326171875,
0.9462890625
]
] |
[
"OneGAN",
"3",
"activation layers were used as non-linear activations. (vi) We add many losses,",
"regularization terms, and training techniques that were not used in FineGAN",
"many of which are completely novel, as far as we can ascertain. As a result, our",
"work outperforms FineGAN in all tasks and is capable of performing new tasks",
"thatitspredecessorcouldnothandle",
"Mixup [35] is a technique for applying aweighted sum between two latent",
"variables in order to synthesize a new latent variable. We use it to merge dif-",
"ferent paths in the model by mixing latent variables that are part coded by the",
"encoder and part produced by the lower levels of the generation. As far as we",
"can ascertain, this is the first usage of mixup to merge information from different",
"paths. Our mixup is applied to image reconstruction infour different locations.",
"3 ",
"",
" Method",
"To solve thetasks of clustering.foreground segmentation.and conditional gen",
"eration with minimal supervision, our method trains multiple neural networks",
"side-by-side. Each task is solved by applying the networks in a task-dependent",
"order. Similarly, the method is trained by applying the networks in two different",
"paths in eachiteration, each path is optimized with aspecific set of losses.",
"Architecture",
"Thearchitecture of the compound network consists of two",
"generators, three encoders and two discriminators: Fig. 1 and Fig. 2 illustrate",
"thetwotrainingpaths.Inthegenerationpath,thegeneratorsproduceasynthetic",
"image conditioned on selected code,the encodersthenretrieve the latent code",
"from thegenerated images.In thereconstructionpath,theencoders code an",
"input image into latent codewhichis used bythegenerators to reconstruct the",
"image. The reconstruction path is applied twice in each iteration, once on real",
"imagesand onceonfakeimagesfromthegenerationpath.Thereconstruction of",
"fake images adds multiple capabilities of self-supervision such as reconstructing",
"the background and mask, which is not applicable with real images without",
"additional supervision.The sub-networksarefullydetailed in the supplementary",
"GeneratorsThe generation is performed by merging theresults of two sepa-",
"rategeneratorsthatruninparallel toproducetheoutputimage.Onegenerator",
"is dedicated for generatingthebackground and the other for the foreground.",
"The generators are conditioned on a two-level hierarchical categorization.Each",
"category has a unique child class @e and a parent class @p shared by multi-",
"ple child classes. These classes are represented by the one-hot vectors (ec,ep).",
"An additional background one-hot vector ebg affects thegeneration of theback",
"ground images. Since there is a tight coupling between the class of the object",
"(water bird, tropical bird, etc.) and the expected background, the typology of",
"the background follows the coarse hierarchy, i.e. the parent class. The generator",
"architectureis influenced by.",
"ec[i]=i中e,ep[i] =di,Φp，ebg=ep",
"(1)",
"The generation starts by converting the one-hot vectors into code vectors us-",
"ing learned embeddings.Such an embeddingis often used when working with"
] |
{
"arxiv_id": "1912.13471",
"language": "en",
"page_idx": 2
}
|
||
1912.13471-en-3
|
1912.13471-en-3.jpg
|
4
Y. Benny and L. Wolf
<FigureHere>
$\mathbf{Fig.1.Flow~of~the~generation~path.~The~generators~decode~the~four~priors~}(e_{bg},e_{p},e_{c},z)$ and produce three separate images (foreground, background, mask) that are combined into the final image. The generated image is then coded by three encoders to retrieve the latent variables and priors.
<FigureHere>
Fig. 2. Flow of the reconstruction path. The same sub-networks are rearranged to perform image reconstruction. The image is coded with the shape and style encoders and then decoded by the foreground generator to produce the foreground image and mask. Then the background encoder and generator code the masked image and produces a background image. The output image combines the foreground and background images. The mixup modules, placed in four different locations, merge the encoders' predicted codes with intermediate stages of the generation, acting as a robust skip connection.
categorical values. A fourth vector $z$ is sampled from a multi-variate gaussian
distribution to represent non-categorical features.
$$v_{bg}=V_{bg}(e_{bg}),\quad v_{p}=V_{p}(e_{p}),\quad v_{c}=V_{c}(e_{c}),\quad z\sim\mathcal{N}(0,1)^{d_{z}}$$
(2)
$\begin{array}{c}\text{The background generator }G_{bg}\text{ receives the background vector }v_{bg}\text{ and noise}\\z\text{ and produces a background image }I_{bg}.\text{ The foreground generator }G_{fg}\text{ receives}\end{array}$ the parent vector $e_p$, child vector $e_c$ and the same $z$ used in the background generation and produces a foreground image $I_{fg}$ and a foreground mask $I_m.$ The generator is optimized such that all foreground images with the same $e_p$ will
|
[
[
0.0772705078125,
0.0162506103515625,
0.09783935546875,
0.0162506103515625,
0.09783935546875,
0.0308685302734375,
0.0772705078125,
0.0308685302734375
],
[
0.16064453125,
0.0162506103515625,
0.379150390625,
0.0162506103515625,
0.379150390625,
0.0308685302734375,
0.16064453125,
0.0308685302734375
],
[
0.38525390625,
0.0755615234375,
0.422607421875,
0.08367919921875,
0.416748046875,
0.09503173828125,
0.379150390625,
0.0869140625
],
[
0.755859375,
0.07958984375,
0.775390625,
0.0828857421875,
0.77294921875,
0.09100341796875,
0.75244140625,
0.08770751953125
],
[
0.136474609375,
0.1226806640625,
0.1436767578125,
0.1226806640625,
0.1436767578125,
0.129150390625,
0.136474609375,
0.129150390625
],
[
0.755859375,
0.138916015625,
0.77783203125,
0.138916015625,
0.77783203125,
0.1527099609375,
0.755859375,
0.1527099609375
],
[
0.8212890625,
0.1405029296875,
0.8466796875,
0.1422119140625,
0.84423828125,
0.1494140625,
0.81982421875,
0.1478271484375
],
[
0.85400390625,
0.1422119140625,
0.86474609375,
0.1422119140625,
0.86474609375,
0.1478271484375,
0.85400390625,
0.1478271484375
],
[
0.75732421875,
0.1820068359375,
0.78271484375,
0.1820068359375,
0.78271484375,
0.1900634765625,
0.75732421875,
0.1900634765625
],
[
0.75830078125,
0.203125,
0.775390625,
0.203125,
0.775390625,
0.2120361328125,
0.75830078125,
0.2120361328125
],
[
0.60888671875,
0.218505859375,
0.64599609375,
0.218505859375,
0.64599609375,
0.2225341796875,
0.60888671875,
0.2225341796875
],
[
0.0784912109375,
0.281982421875,
0.91552734375,
0.284423828125,
0.91552734375,
0.30126953125,
0.0784912109375,
0.2998046875
],
[
0.0821533203125,
0.3037109375,
0.91552734375,
0.3037109375,
0.91552734375,
0.317626953125,
0.0821533203125,
0.317626953125
],
[
0.0809326171875,
0.320068359375,
0.91650390625,
0.320068359375,
0.91650390625,
0.334716796875,
0.0809326171875,
0.334716796875
],
[
0.0809326171875,
0.337890625,
0.38037109375,
0.337890625,
0.38037109375,
0.3525390625,
0.0809326171875,
0.3525390625
],
[
0.388916015625,
0.37451171875,
0.439697265625,
0.37451171875,
0.439697265625,
0.385009765625,
0.388916015625,
0.385009765625
],
[
0.57470703125,
0.375244140625,
0.61376953125,
0.375244140625,
0.61376953125,
0.382568359375,
0.57470703125,
0.382568359375
],
[
0.64013671875,
0.3916015625,
0.67138671875,
0.3955078125,
0.66650390625,
0.40771484375,
0.63525390625,
0.40380859375
],
[
0.8466796875,
0.38427734375,
0.88525390625,
0.38427734375,
0.88525390625,
0.3916015625,
0.8466796875,
0.3916015625
],
[
0.2451171875,
0.39306640625,
0.271728515625,
0.39306640625,
0.271728515625,
0.406982421875,
0.2451171875,
0.406982421875
],
[
0.7509765625,
0.426513671875,
0.75830078125,
0.42333984375,
0.76318359375,
0.428955078125,
0.755859375,
0.43212890625
],
[
0.25,
0.452392578125,
0.26318359375,
0.452392578125,
0.26318359375,
0.4638671875,
0.25,
0.4638671875
],
[
0.0821533203125,
0.6025390625,
0.9130859375,
0.6025390625,
0.9130859375,
0.6171875,
0.0821533203125,
0.6171875
],
[
0.0809326171875,
0.62158203125,
0.91650390625,
0.62158203125,
0.91650390625,
0.63623046875,
0.0809326171875,
0.63623046875
],
[
0.0809326171875,
0.63916015625,
0.9130859375,
0.63916015625,
0.9130859375,
0.65380859375,
0.0809326171875,
0.65380859375
],
[
0.0784912109375,
0.65576171875,
0.91796875,
0.6572265625,
0.91796875,
0.671875,
0.0784912109375,
0.67041015625
],
[
0.0809326171875,
0.67578125,
0.9130859375,
0.67578125,
0.9130859375,
0.6904296875,
0.0809326171875,
0.6904296875
],
[
0.0809326171875,
0.69384765625,
0.91552734375,
0.69384765625,
0.91552734375,
0.70654296875,
0.0809326171875,
0.70654296875
],
[
0.0809326171875,
0.7109375,
0.90478515625,
0.7109375,
0.90478515625,
0.7255859375,
0.0809326171875,
0.7255859375
],
[
0.0809326171875,
0.7626953125,
0.91796875,
0.7626953125,
0.91796875,
0.78076171875,
0.0809326171875,
0.78076171875
],
[
0.0821533203125,
0.78369140625,
0.6025390625,
0.78369140625,
0.6025390625,
0.79833984375,
0.0821533203125,
0.79833984375
],
[
0.18115234375,
0.81787109375,
0.81396484375,
0.8154296875,
0.81396484375,
0.83447265625,
0.18115234375,
0.8369140625
],
[
0.88525390625,
0.8173828125,
0.91796875,
0.8173828125,
0.91796875,
0.8359375,
0.88525390625,
0.8359375
],
[
0.11474609375,
0.8505859375,
0.91796875,
0.8515625,
0.91796875,
0.869140625,
0.11474609375,
0.8681640625
],
[
0.0784912109375,
0.8681640625,
0.91796875,
0.87158203125,
0.91796875,
0.89208984375,
0.0784912109375,
0.8896484375
],
[
0.0809326171875,
0.89111328125,
0.91650390625,
0.89111328125,
0.91650390625,
0.9091796875,
0.0809326171875,
0.9091796875
],
[
0.0809326171875,
0.91162109375,
0.9130859375,
0.91162109375,
0.9130859375,
0.9287109375,
0.0809326171875,
0.9287109375
],
[
0.0784912109375,
0.9287109375,
0.91796875,
0.93017578125,
0.91796875,
0.947265625,
0.0784912109375,
0.9462890625
]
] |
[
"4",
"Y.Bennyand L.Wolf",
"Ggi",
"Eh",
"n",
"Ep",
"Thape",
"cto",
"Style",
"E.",
"ux:chd",
"Fig.1.Flow of the generationpath.The generatorsdecode thefourpriors (ebg,ep,ec,z)",
"and producethree separateimages(foreground,background,mask)that are combined",
"into the final image.The generated image is then coded by three encoders to retrieve",
"the latent variables and priors.",
"Embedding",
"Generato",
"Gbe",
"real/fake",
"Ebg",
"A",
"E",
"Fig.2. Flow of the reconstruction path. The same sub-networks are rearranged to per",
"form image reconstruction. The image is coded with the shape and style encoders and",
"then decoded bythe foreground generator toproducethe foreground imageand mask",
"Then the background encoder and generator codethe masked image and produces a",
"background image.The output image combines theforeground and background images",
"The mixup modules,placed in four different locations,mergethe encoders'predicted",
"codes with intermediate stages of the generation, acting as a robust skip connection.",
"categorical values. A fourth vector z is sampled from a multi-variate gaussian",
"distribution to representnon-categorical features.",
"Vbg = Vbg(ebg),Up =Vp(ep), Ve= Ve(ec), z~ N(O, 1)ds",
"(2)",
"The background generator Gbg receives the background vector Vbg and noise",
"z and produces a background image Ibg. The foreground generator Gfg receives",
"the parent vector ep, child vector ee and the same z used in the background",
"generation and produces a foreground image Ifg and a foreground mask Im",
"Thegenerator is optimized such that allforeground imageswith the same epwill"
] |
{
"arxiv_id": "1912.13471",
"language": "en",
"page_idx": 3
}
|
||
1912.13471-en-4
|
1912.13471-en-4.jpg
|
OneGAN5
have the same object shape and all images with the same $e_c$ will have a similar object appearance. The latent vector $z$ is implicitly conditioned to represent all non-categorical information, such as pose, orientation, size, etc. It is used in both the background and foreground generation, so that the images produced by both networks will merge into a coherent image. Each generator is a composition of sub-modules applied back to back, with intermediate pre-images $(A_{bg},A_{fg})\colon$
$$\begin{aligned}I_{bg}&=G_{bg_1}(A_{bg}),\quad&A_{bg}&=G_{bg_0}(v_{bg},z)\\(I_{fg},I_{m})&=G_{fg_2}(G_{fg_1}(A_{fg},v_p),v_c),\quad&A_{fg}&=G_{fg_0}(v_p,z)\end{aligned}$$
(3)
(4)
The final generated image is: (where o denotes element-wise multiplication)
$$I=I_{bg}\circ(1-I_{m})+I_{fg}\circ I_{m}$$
(5)
Encoders Unlike FineGAN, which performs only the generation task, ou
method requires the use of encoders. We introduce three encoders: background encoder $E_bg$,shape encoder $E_p$, and style encoder $E_c.$ They run in semi-parallel to predict both the latent codes $(v_{bg},v_{p},v_{c},z)$ of an input image and the underlying one-hot vectors $(e_bg,e_p,e_c).$ All encoders are fed with image $I$ as input. The background encoder is also fed with the mask $I_m.$ During image reconstruction, there is no initial image mask, therefore it first has to be generated by encoding the shape and style features and applying the foreground generator. The lack of ground-truth mask is why the encoders do not run fully in parallel. In addition, the background and shape encoders also produce bypass tensors $(B_bg,B_{fg})$ to be used as skip connections between the encoders and the generators.
(6)
(7)
$$\begin{aligned}(B_{bg})&=E_{bg}(I,I_{m})\\(\hat{e}_{p},\mu_{p},\sigma_{p},B_{fg},\mu_{z},\sigma_{z})&=E_{p}(I)\\(\hat{e}_{c},\mu_{c},\sigma_{c})&=E_{c}(I)\end{aligned}$$
(8)
Following Variational Auto-Encoder literature, $(\mu,\sigma)$ are three paired vectors of sizes $(d_z,d_p,d_c)$ defning the mean and variance to sample each element of $(\hat{z},\hat{v}_{p},\hat{v}_{c})$ from a gaussian distribution .
$\mathbf{Mixup}$ At the intersection between the encoders and generators, we introduce a novel method to merge information coded by the encoders and information produced by the embeddings and lower levels of the generators. The mixup module [35],mixes two input variables with a weight parameter $\beta.$ The rationale behind this application is that during generation there is no data coming from the encoders, so the mixup is turned off and only information from the embeddings and lower levels of the generators are passed forward. During reconstruction, we want our method to utilize the skip connections to improve performance and also use the predicted embeddings $(v_p,v_c)$ to represent the object's shape and style. The contrast between the two paths leads to a difficulty in optimizing them simultaneously. The introduction of the mixup simplifies this by having both paths active during forward path and back-propagation. In contrast to regular residual connections, the ever changing $\beta$ used in the mixup forces both inputs to be independent representations and not complement each other.
|
[
[
0.74267578125,
0.01543426513671875,
0.841796875,
0.01543426513671875,
0.841796875,
0.033294677734375,
0.74267578125,
0.033294677734375
],
[
0.89990234375,
0.0162506103515625,
0.9228515625,
0.0162506103515625,
0.9228515625,
0.03167724609375,
0.89990234375,
0.03167724609375
],
[
0.0821533203125,
0.05767822265625,
0.91552734375,
0.05767822265625,
0.91552734375,
0.07232666015625,
0.0821533203125,
0.07232666015625
],
[
0.0821533203125,
0.07879638671875,
0.91650390625,
0.07879638671875,
0.91650390625,
0.09344482421875,
0.0821533203125,
0.09344482421875
],
[
0.0809326171875,
0.0982666015625,
0.91650390625,
0.0982666015625,
0.91650390625,
0.1129150390625,
0.0809326171875,
0.1129150390625
],
[
0.0821533203125,
0.11700439453125,
0.91552734375,
0.11700439453125,
0.91552734375,
0.131591796875,
0.0821533203125,
0.131591796875
],
[
0.0809326171875,
0.1373291015625,
0.91796875,
0.1373291015625,
0.91796875,
0.15185546875,
0.0809326171875,
0.15185546875
],
[
0.07611083984375,
0.153564453125,
0.88427734375,
0.156005859375,
0.88427734375,
0.173828125,
0.07611083984375,
0.17138671875
],
[
0.2210693359375,
0.1851806640625,
0.3974609375,
0.1868896484375,
0.3974609375,
0.2080078125,
0.2210693359375,
0.20556640625
],
[
0.6435546875,
0.18603515625,
0.83837890625,
0.18603515625,
0.83837890625,
0.2071533203125,
0.6435546875,
0.2071533203125
],
[
0.88525390625,
0.1868896484375,
0.91796875,
0.1868896484375,
0.91796875,
0.20556640625,
0.88525390625,
0.20556640625
],
[
0.64013671875,
0.2120361328125,
0.8359375,
0.2120361328125,
0.8359375,
0.233154296875,
0.64013671875,
0.233154296875
],
[
0.88525390625,
0.2120361328125,
0.92041015625,
0.2120361328125,
0.92041015625,
0.230712890625,
0.88525390625,
0.230712890625
],
[
0.11712646484375,
0.24365234375,
0.9130859375,
0.2445068359375,
0.9130859375,
0.262451171875,
0.11712646484375,
0.2607421875
],
[
0.1654052734375,
0.212890625,
0.53857421875,
0.212890625,
0.53857421875,
0.230712890625,
0.1654052734375,
0.230712890625
],
[
0.343017578125,
0.275390625,
0.65234375,
0.27783203125,
0.65234375,
0.29736328125,
0.343017578125,
0.295654296875
],
[
0.88427734375,
0.275390625,
0.92138671875,
0.275390625,
0.92138671875,
0.298828125,
0.88427734375,
0.298828125
],
[
0.0821533203125,
0.31103515625,
0.2041015625,
0.31103515625,
0.2041015625,
0.32568359375,
0.0821533203125,
0.32568359375
],
[
0.210205078125,
0.31591796875,
0.2186279296875,
0.31591796875,
0.2186279296875,
0.320068359375,
0.210205078125,
0.320068359375
],
[
0.2198486328125,
0.30859375,
0.91650390625,
0.31103515625,
0.91650390625,
0.3291015625,
0.2198486328125,
0.32666015625
],
[
0.0809326171875,
0.332275390625,
0.91650390625,
0.332275390625,
0.91650390625,
0.346923828125,
0.0809326171875,
0.346923828125
],
[
0.0809326171875,
0.350830078125,
0.91650390625,
0.350830078125,
0.91650390625,
0.365478515625,
0.0809326171875,
0.365478515625
],
[
0.0809326171875,
0.36962890625,
0.91552734375,
0.36962890625,
0.91552734375,
0.387451171875,
0.0809326171875,
0.387451171875
],
[
0.0784912109375,
0.38818359375,
0.91552734375,
0.38818359375,
0.91552734375,
0.40625,
0.0784912109375,
0.40625
],
[
0.0809326171875,
0.40869140625,
0.91552734375,
0.40869140625,
0.91552734375,
0.42333984375,
0.0809326171875,
0.42333984375
],
[
0.0784912109375,
0.42578125,
0.91796875,
0.42822265625,
0.91796875,
0.445068359375,
0.0784912109375,
0.443603515625
],
[
0.0772705078125,
0.44677734375,
0.92041015625,
0.445068359375,
0.92041015625,
0.463134765625,
0.0772705078125,
0.4638671875
],
[
0.0821533203125,
0.467041015625,
0.91552734375,
0.467041015625,
0.91552734375,
0.481689453125,
0.0821533203125,
0.481689453125
],
[
0.0809326171875,
0.48486328125,
0.91796875,
0.4873046875,
0.91796875,
0.50439453125,
0.0809326171875,
0.5029296875
],
[
0.0809326171875,
0.5068359375,
0.81640625,
0.5068359375,
0.81640625,
0.521484375,
0.0809326171875,
0.521484375
],
[
0.490234375,
0.53759765625,
0.6845703125,
0.53759765625,
0.6845703125,
0.5556640625,
0.490234375,
0.5556640625
],
[
0.88427734375,
0.5361328125,
0.92138671875,
0.5361328125,
0.92138671875,
0.55810546875,
0.88427734375,
0.55810546875
],
[
0.30908203125,
0.56201171875,
0.638671875,
0.5595703125,
0.638671875,
0.5810546875,
0.30908203125,
0.58349609375
],
[
0.88427734375,
0.55908203125,
0.92041015625,
0.55908203125,
0.92041015625,
0.58349609375,
0.88427734375,
0.58349609375
],
[
0.437255859375,
0.58740234375,
0.63623046875,
0.58740234375,
0.63623046875,
0.60498046875,
0.437255859375,
0.60498046875
],
[
0.88427734375,
0.583984375,
0.92138671875,
0.583984375,
0.92138671875,
0.60595703125,
0.88427734375,
0.60595703125
],
[
0.0809326171875,
0.619140625,
0.91796875,
0.619140625,
0.91796875,
0.63671875,
0.0809326171875,
0.63671875
],
[
0.0821533203125,
0.64013671875,
0.91650390625,
0.64013671875,
0.91650390625,
0.65478515625,
0.0821533203125,
0.65478515625
],
[
0.0809326171875,
0.658203125,
0.49755859375,
0.6572265625,
0.49755859375,
0.67431640625,
0.0809326171875,
0.67578125
],
[
0.0784912109375,
0.677734375,
0.91796875,
0.677734375,
0.91796875,
0.6953125,
0.0784912109375,
0.6953125
],
[
0.0809326171875,
0.69873046875,
0.91650390625,
0.69873046875,
0.91650390625,
0.71337890625,
0.0809326171875,
0.71337890625
],
[
0.0821533203125,
0.71728515625,
0.91552734375,
0.71728515625,
0.91552734375,
0.73193359375,
0.0821533203125,
0.73193359375
],
[
0.0809326171875,
0.73681640625,
0.91796875,
0.73681640625,
0.91796875,
0.75146484375,
0.0809326171875,
0.75146484375
],
[
0.0809326171875,
0.75537109375,
0.9130859375,
0.75537109375,
0.9130859375,
0.77001953125,
0.0809326171875,
0.77001953125
],
[
0.0772705078125,
0.7734375,
0.91796875,
0.77490234375,
0.91796875,
0.7919921875,
0.0772705078125,
0.791015625
],
[
0.0784912109375,
0.79443359375,
0.92041015625,
0.79443359375,
0.92041015625,
0.8125,
0.0784912109375,
0.8125
],
[
0.0809326171875,
0.8154296875,
0.91650390625,
0.8154296875,
0.91650390625,
0.82958984375,
0.0809326171875,
0.82958984375
],
[
0.0809326171875,
0.83447265625,
0.91552734375,
0.83447265625,
0.91552734375,
0.8515625,
0.0809326171875,
0.8515625
],
[
0.0809326171875,
0.85302734375,
0.91650390625,
0.85302734375,
0.91650390625,
0.8701171875,
0.0809326171875,
0.8701171875
],
[
0.0809326171875,
0.873046875,
0.91650390625,
0.873046875,
0.91650390625,
0.8876953125,
0.0809326171875,
0.8876953125
],
[
0.0772705078125,
0.89111328125,
0.91796875,
0.89013671875,
0.91796875,
0.908203125,
0.0772705078125,
0.9091796875
],
[
0.0809326171875,
0.912109375,
0.91796875,
0.912109375,
0.91796875,
0.9267578125,
0.0809326171875,
0.9267578125
],
[
0.0784912109375,
0.931640625,
0.78759765625,
0.931640625,
0.78759765625,
0.9462890625,
0.0784912109375,
0.9462890625
]
] |
[
"OneGAN",
"5",
"have the same object shape and all images with the same ec will have a similan",
"object appearance.The latent vector z is implicitly conditioned to represent all",
"non-categorical information, such as pose, orientation, size, etc.It is used in both",
"the background and foreground generation, so that the images produced by both",
"networks will merge into a coherent image.Each generator is a composition of",
"sub-modules applied back to back, with intermediate pre-images (Abg, Afg):",
"Ibg = Gbgi (Abg),",
"Abg = Gbgo (Ubg, 2)",
"(3)",
"Afg=Gfgo(up,2)",
"(4)",
"The final generated image is: (where o denotes element-wise multiplication)",
"(Ifg, Im) = Gfg2(Gfg1(Afg, Up), Ue),",
"I=Ibgo(1-Im)+IfgoIm",
"(5)",
"Encoders",
"",
"Unlike FineGAN, which performs only the generation task, our",
"method requires the use of encoders.We introduce three encoders: background",
"encoder Ebg,shape encoder Ep,and style encoder Ee.They run in semi-parallel",
"to predict both the latent codes (vbg, Up, Vc, z) of an input image and the underly-",
"ing one-hot vectors (ebg, ep, ec). All encoders are fed with image I as input. The",
"background encoder is also fed with the mask Im.During image reconstruction.",
"there is no initial image mask, therefore it first has to be generated by encoding",
"the shape and style features and applying the foreground generator. The lack of",
"ground-truth mask is why the encoders do not run fully in parallel. In addition.",
"the background and shape encoders also produce bypass tensors (Bbg,Bfg) to",
"be used as skip connections between the encoders and the generators.",
"(Bbg) = Ebg(I, Im)",
"(6)",
"(ep, Mμp, Op, Bfg, Hz,0z) = Ep(I)",
"(7)",
"(éc, μc,c) = Ec(I)",
"(8)",
"Following Variational Auto-Encoder literature, (μ,) are three paired vectors",
"of sizes (dz,d,d.) defining the mean and variance to sample each element of",
"(2,ip,o)fromagaussiandistribution.",
"MixupAt the intersection between the encoders and generators, we introduce",
"a novel method to merge information coded by the encoders and information",
"produced by the embeddings and lower levels of the generators.The mixup",
"module[35l,mixestwoinput variableswitha weightparameter β.Therationale",
"behind this application is that duringgeneration thereisno data comingfrom the",
"encoders,so themixup is turned off and only information from theembeddings",
"and lower levels of the generators are passed forward. During reconstruction, we",
"wantourmethodtoutilizetheskip connectionstoimproveperformanceandalso",
"use the predicted embeddings (up,uc) to represent the object's shape and style.",
"The contrast between the two paths leads to a difficulty in optimizing them",
"simultaneously.The introduction of the mixup simplifies this byhaving both",
"paths active during forward path and back-propagation. In contrast to regular",
"residual connections,the ever changingβ usedinthemixupforcesboth inputs",
"to be independent representations and not complement each other."
] |
{
"arxiv_id": "1912.13471",
"language": "en",
"page_idx": 4
}
|
||
2305.12961-en-0
|
2305.12961-en-0.jpg
|
$\textbf{Enhanced Meta Label Correction for Coping with Label Corruption}$
Chaim Baskin
Technion- Israel Institute of Technology
chaimbaskingtechnion.ac.il
### Mitchell Keren Taraday Technion- Israel Institute of Technology
mitchell@campus.technion.ac.il
$ଝୁ$
## $\mathbf{Abstract}$
Traditional methods for learning with the presence of
$\begin{array}{lll}&noisy&labels&have&successfully&handled&datasets&with&ar-\\\text{ರ್ರ}&&iffcially&injected&noise&but&still&fall&short&of&adequately\\&&handling&real-world&noise.&With&the&increasing&use&of\end{array}$
meta-learning in the diverse fhelds of machine learning, researchers leveraged auxiliary small clean datasets to metacorrect the training labels. Nonetheless, existing meta- label
$\sum_{\begin{array}{ccc}&\text{correction approaches are not fully exploiting their poten-}\\&\text{tial. In this study, we propose an Enhanced Meta Label}\\&\textbf{Correction approach abbreviated as EMLCfor the learning}\end{array}}$ $\begin{array}{lll}{&Correction~approach~abbreviated~as~EMLC~for~the~learning}\\{\textcircled{0}}&{with~noisy~labels~(LNL)~problem.~We~re~-examine~the~meta-2~learning}\\{\textcircled{0}}&{learning~process~and~introduce~faster~and~more~accurate}\end{array}$
meta-gradient derivations. We propose a novel teacher architecture tailored explicitly to the LNL problem, equipped with novel training objectives. EMLC outperforms prior approaches and achieves state-of-the-art results in all standard benchmarks. Notably, EMLC enhances the previous art on the noisy real-world dataset ClothinglM by 1.52% while requiring ×0.5 the time per epoch and with much faster convergence of the meta- objective when compared to the baseline approach.
### $1. \textbf{Introduction}$
The remarkable success of Deep Neural Networks (DNNs) for visual classification tasks is predominantly due to the availability of massive labeled datasets. In many practical applications, obtaining the required amounts of reliable labeled data is intractable. As a result, numerous works over the past decade have sought ways to reduce the amount of labeled data required for classification tasks. Notably, semi-supervised learning exploits unlabeled data [33,24], transfer learning exploits prior knowledge obtained from different tasks [7,39], and self-supervised learning exploits data augmentations for label agnostic representation learning [3,11]. Nonetheless, for many applications, the quality of the labeled data can be sacrificed for the sake of its quan-
$^{1}$Project page:
https://sites.google.com/view/
emlc-paper.
<FigureHere>
Figure 1: Comparison of the training labels recovery (solid $li$ne) and the wrong labels recovery (dashed line) over training epochs on the CIFAR-10 dataset with 90% symmetric noise. EMLC (ours) is comapred to MLC [41] and to MLC [41] with self-supervised pretraining.
tity. A prominent example is the process of crawling search engines and online websites as demonstrated by [17,32]. The crawling process is often easy to automate however results in significant amounts of label noise with complicated patterns. A fundamental problem with such data, however, is that classical learning methodologies tend to fail when significant label noise is present [36]. Therefore, designing learning frameworks that are able to cope with label corruption is a task of great importance.
While traditional methods for learning with noisy labels (LNL) are capable of handling data with immense artificial, injected noise, their ability to handle real-world noise remains highly limited. Thus, noisy labeled datasets alone limit the capability of learning from real-world noisy labeled datasets. Luckily, in many real-world applications, while obtaining large amounts of labeled data is infeasible, obtaining small amounts of labeled data is usually attainable. Thus, the paramount objective is to find methods that can incorporate both large amounts of noisy labeled data and small amounts of clean data. A natural choice is to adopt a meta-learning framework, a popular design regularly used for solving various tasks. Consequently, recent
|
[
[
0.146240234375,
0.1337890625,
0.826171875,
0.1337890625,
0.826171875,
0.1483154296875,
0.146240234375,
0.1483154296875
],
[
0.20751953125,
0.18310546875,
0.393798828125,
0.18310546875,
0.393798828125,
0.1976318359375,
0.20751953125,
0.1976318359375
],
[
0.61376953125,
0.18310546875,
0.72802734375,
0.18310546875,
0.72802734375,
0.1976318359375,
0.61376953125,
0.1976318359375
],
[
0.136474609375,
0.1988525390625,
0.4599609375,
0.2008056640625,
0.4599609375,
0.21533203125,
0.136474609375,
0.21337890625
],
[
0.50732421875,
0.199462890625,
0.8310546875,
0.2008056640625,
0.8310546875,
0.21533203125,
0.50732421875,
0.2139892578125
],
[
0.1666259765625,
0.2197265625,
0.43310546875,
0.2197265625,
0.43310546875,
0.23291015625,
0.1666259765625,
0.23291015625
],
[
0.5556640625,
0.2197265625,
0.78515625,
0.2197265625,
0.78515625,
0.23291015625,
0.5556640625,
0.23291015625
],
[
0.0236968994140625,
0.264404296875,
0.05963134765625,
0.264404296875,
0.05963134765625,
0.324462890625,
0.0236968994140625,
0.324462890625
],
[
0.23779296875,
0.26953125,
0.313720703125,
0.26953125,
0.313720703125,
0.2841796875,
0.23779296875,
0.2841796875
],
[
0.52392578125,
0.2734375,
0.54345703125,
0.2734375,
0.54345703125,
0.285888671875,
0.52392578125,
0.285888671875
],
[
0.52294921875,
0.285400390625,
0.5439453125,
0.285400390625,
0.5439453125,
0.2998046875,
0.52294921875,
0.2998046875
],
[
0.09722900390625,
0.2998046875,
0.4697265625,
0.300537109375,
0.4697265625,
0.31494140625,
0.09722900390625,
0.314453125
],
[
0.52294921875,
0.296630859375,
0.54345703125,
0.296630859375,
0.54345703125,
0.311279296875,
0.52294921875,
0.311279296875
],
[
0.52294921875,
0.30810546875,
0.5439453125,
0.30810546875,
0.5439453125,
0.3212890625,
0.52294921875,
0.3212890625
],
[
0.0269622802734375,
0.31494140625,
0.056365966796875,
0.31640625,
0.053924560546875,
0.365478515625,
0.0236968994140625,
0.3642578125
],
[
0.08087158203125,
0.318115234375,
0.465576171875,
0.318115234375,
0.465576171875,
0.328857421875,
0.08087158203125,
0.328857421875
],
[
0.51123046875,
0.318115234375,
0.5439453125,
0.318115234375,
0.5439453125,
0.335205078125,
0.51123046875,
0.335205078125
],
[
0.080078125,
0.33203125,
0.468017578125,
0.33203125,
0.468017578125,
0.34521484375,
0.080078125,
0.34521484375
],
[
0.5146484375,
0.32763671875,
0.5458984375,
0.330810546875,
0.54150390625,
0.348388671875,
0.5107421875,
0.34521484375
],
[
0.07843017578125,
0.34521484375,
0.27685546875,
0.3466796875,
0.27685546875,
0.361083984375,
0.07843017578125,
0.35986328125
],
[
0.26953125,
0.34521484375,
0.4697265625,
0.3466796875,
0.4697265625,
0.361083984375,
0.26953125,
0.35986328125
],
[
0.51123046875,
0.343505859375,
0.5439453125,
0.343505859375,
0.5439453125,
0.359130859375,
0.51123046875,
0.359130859375
],
[
0.52294921875,
0.35595703125,
0.5400390625,
0.354736328125,
0.54150390625,
0.3662109375,
0.52392578125,
0.36865234375
],
[
0.0269622802734375,
0.3642578125,
0.053924560546875,
0.3642578125,
0.053924560546875,
0.392578125,
0.0269622802734375,
0.392578125
],
[
0.080078125,
0.3623046875,
0.468017578125,
0.3623046875,
0.468017578125,
0.375732421875,
0.080078125,
0.375732421875
],
[
0.52392578125,
0.367431640625,
0.54345703125,
0.367431640625,
0.54345703125,
0.380615234375,
0.52392578125,
0.380615234375
],
[
0.08087158203125,
0.37890625,
0.465576171875,
0.37890625,
0.465576171875,
0.389404296875,
0.08087158203125,
0.389404296875
],
[
0.521484375,
0.386474609375,
0.53369140625,
0.37744140625,
0.5439453125,
0.386474609375,
0.53173828125,
0.39453125
],
[
0.080078125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.525390625,
0.393310546875,
0.54150390625,
0.393310546875,
0.54150390625,
0.40283203125,
0.525390625,
0.40283203125
],
[
0.75244140625,
0.399658203125,
0.765625,
0.399658203125,
0.765625,
0.4091796875,
0.75244140625,
0.4091796875
],
[
0.7958984375,
0.399658203125,
0.8095703125,
0.399658203125,
0.8095703125,
0.40771484375,
0.7958984375,
0.40771484375
],
[
0.02532958984375,
0.406005859375,
0.056365966796875,
0.406005859375,
0.056365966796875,
0.45068359375,
0.02532958984375,
0.45068359375
],
[
0.84228515625,
0.401611328125,
0.85107421875,
0.401611328125,
0.85107421875,
0.4072265625,
0.84228515625,
0.4072265625
],
[
0.080078125,
0.40771484375,
0.468017578125,
0.40771484375,
0.468017578125,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.69921875,
0.40771484375,
0.7294921875,
0.40771484375,
0.7294921875,
0.416748046875,
0.69921875,
0.416748046875
],
[
0.080078125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.54150390625,
0.423583984375,
0.6298828125,
0.423583984375,
0.6298828125,
0.434326171875,
0.54150390625,
0.434326171875
],
[
0.6689453125,
0.425537109375,
0.74853515625,
0.425537109375,
0.74853515625,
0.43310546875,
0.6689453125,
0.43310546875
],
[
0.78515625,
0.423583984375,
0.8896484375,
0.423583984375,
0.8896484375,
0.434326171875,
0.78515625,
0.434326171875
],
[
0.54150390625,
0.43505859375,
0.62744140625,
0.43505859375,
0.62744140625,
0.442626953125,
0.54150390625,
0.442626953125
],
[
0.66259765625,
0.43505859375,
0.74609375,
0.43505859375,
0.74609375,
0.442626953125,
0.66259765625,
0.442626953125
],
[
0.78173828125,
0.43505859375,
0.88818359375,
0.43505859375,
0.88818359375,
0.442626953125,
0.78173828125,
0.442626953125
],
[
0.03350830078125,
0.449462890625,
0.0555419921875,
0.449462890625,
0.0555419921875,
0.479736328125,
0.03350830078125,
0.479736328125
],
[
0.080078125,
0.436279296875,
0.4697265625,
0.4375,
0.4697265625,
0.451904296875,
0.080078125,
0.45068359375
],
[
0.08087158203125,
0.453857421875,
0.465576171875,
0.453857421875,
0.465576171875,
0.464111328125,
0.08087158203125,
0.464111328125
],
[
0.50244140625,
0.458984375,
0.89208984375,
0.458984375,
0.89208984375,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.07843017578125,
0.466552734375,
0.4697265625,
0.46728515625,
0.4697265625,
0.481689453125,
0.07843017578125,
0.48095703125
],
[
0.50244140625,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.080078125,
0.48291015625,
0.468017578125,
0.48291015625,
0.468017578125,
0.496826171875,
0.080078125,
0.496826171875
],
[
0.02532958984375,
0.49365234375,
0.056365966796875,
0.49365234375,
0.056365966796875,
0.52978515625,
0.02532958984375,
0.52978515625
],
[
0.50244140625,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.07843017578125,
0.496826171875,
0.471435546875,
0.49755859375,
0.471435546875,
0.51220703125,
0.07843017578125,
0.51123046875
],
[
0.50244140625,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.50244140625,
0.517578125
],
[
0.080078125,
0.51220703125,
0.4697265625,
0.51220703125,
0.4697265625,
0.5263671875,
0.080078125,
0.5263671875
],
[
0.50341796875,
0.51953125,
0.74609375,
0.51953125,
0.74609375,
0.53271484375,
0.50341796875,
0.53271484375
],
[
0.02777099609375,
0.53076171875,
0.053924560546875,
0.53076171875,
0.053924560546875,
0.5498046875,
0.02777099609375,
0.5498046875
],
[
0.080078125,
0.52880859375,
0.468017578125,
0.52880859375,
0.468017578125,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.02532958984375,
0.54150390625,
0.056365966796875,
0.54150390625,
0.056365966796875,
0.580078125,
0.02532958984375,
0.580078125
],
[
0.07843017578125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.07843017578125,
0.556640625
],
[
0.07843017578125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.07843017578125,
0.57275390625
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.080078125,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.5869140625,
0.080078125,
0.5869140625
],
[
0.0196075439453125,
0.583984375,
0.058013916015625,
0.583984375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.039215087890625,
0.5859375,
0.0482177734375,
0.5859375,
0.0482177734375,
0.59130859375,
0.039215087890625,
0.59130859375
],
[
0.07843017578125,
0.5888671875,
0.4697265625,
0.5888671875,
0.4697265625,
0.60302734375,
0.07843017578125,
0.60302734375
],
[
0.50244140625,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.50341796875,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.611328125,
0.50341796875,
0.611328125
],
[
0.080078125,
0.60400390625,
0.238525390625,
0.60400390625,
0.238525390625,
0.6171875,
0.080078125,
0.6171875
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.50244140625,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.08087158203125,
0.646484375,
0.20751953125,
0.646484375,
0.20751953125,
0.6611328125,
0.08087158203125,
0.6611328125
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.50244140625,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50244140625,
0.671875
],
[
0.098876953125,
0.671875,
0.4697265625,
0.671875,
0.4697265625,
0.68505859375,
0.098876953125,
0.68505859375
],
[
0.5009765625,
0.671875,
0.890625,
0.67236328125,
0.890625,
0.6875,
0.5009765625,
0.68603515625
],
[
0.08087158203125,
0.6875,
0.4697265625,
0.6875,
0.4697265625,
0.70068359375,
0.08087158203125,
0.70068359375
],
[
0.50244140625,
0.68798828125,
0.7255859375,
0.68798828125,
0.7255859375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.7021484375,
0.46728515625,
0.7021484375,
0.46728515625,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.51953125,
0.7041015625,
0.89208984375,
0.705078125,
0.89208984375,
0.7197265625,
0.51953125,
0.71826171875
],
[
0.080078125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.080078125,
0.73193359375,
0.4697265625,
0.73193359375,
0.4697265625,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.74755859375,
0.471435546875,
0.74755859375,
0.471435546875,
0.76123046875,
0.080078125,
0.76123046875
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
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,
0.7763671875,
0.46728515625,
0.7763671875,
0.46728515625,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.79248046875,
0.468017578125,
0.79248046875,
0.468017578125,
0.80615234375,
0.080078125,
0.80615234375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.080078125,
0.8212890625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.82373046875,
0.468017578125,
0.82373046875,
0.468017578125,
0.8369140625,
0.080078125,
0.8369140625
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.83837890625,
0.468017578125,
0.83837890625,
0.468017578125,
0.8515625,
0.080078125,
0.8515625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85302734375,
0.468017578125,
0.85400390625,
0.468017578125,
0.8671875,
0.080078125,
0.8662109375
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09478759765625,
0.87353515625,
0.1953125,
0.876953125,
0.1953125,
0.8916015625,
0.09478759765625,
0.88818359375
],
[
0.2279052734375,
0.87548828125,
0.4697265625,
0.87548828125,
0.4697265625,
0.8896484375,
0.2279052734375,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.07843017578125,
0.88818359375,
0.1650390625,
0.8896484375,
0.1650390625,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Enhanced Meta Label Correction for Coping with Label Corruption",
"Mitchell Keren Taraday",
"Chaim Baskin",
"Technion-Israel Institute of Technology",
"Technion -Israel Institute of Technology",
"mitchell@campus.technion.ac.il",
"chaimbaskin@technion.ac.il",
"",
"Abstract",
"1.0",
"0.9",
"Traditional methods for learning with the presence of",
"0.8",
"0.7",
"10",
"noisy labels have successfullyhandled datasets with ar-",
"00.6",
"tificially injected noise but still fall short of adequately",
"00.5",
"handling real-world noise.",
"With the increasing use of",
"0.4",
"0.3",
"0",
"meta-learning in the diverse fields of machine learning, re-",
"0.2 ",
"searchers leveraged auxiliary small clean datasetsto meta-",
"Q",
"correct the training labels. Nonetheless, existing meta-label",
"0.0",
"10",
"12",
"LAD",
"14",
"correction approaches are not fully exploiting their poten-",
"Epoch",
"tial. In this study, we propose an Enhanced Meta Label",
"EMLC label recovery",
"MLClabelrecovery",
"MLC+SSL label recovery",
"EMLCcleanednoise",
"MLCcleanednoise",
"MLC+SSLcleanednoise",
"'SO",
"Correction approach abbreviated as EMLC for the learning",
"withnoisylabels(LNL)problem.Were-examinethemeta-",
"Figure l: Comparison of the training labels recovery (solid",
"learning process and introduce faster and more accurate",
"line) and the wrong labels recovery (dashed line) over train-",
"meta-gradient derivations. We propose a novel teacher ar-",
"ZA",
"ing epochs on the CIFAR-10 dataset with 90% symmetric",
"chitecture tailored explicitly to the LNL problem, equipped",
"noise. EMLC (ours) is comapred to MLC[41] and to MLC",
"with novel training objectives. EMLC outperforms prior",
"[41] with self-supervised pretraining.",
"6",
"approaches and achieves state-of-the-art results in all stan-",
"160",
"dard benchmarks. Notably, EMLC enhances the previous",
"art on the noisy real-world dataset ClothinglM by 1.52%",
"tity. A prominent example is the process of crawling search",
"while requiring x0.5 the time per epoch and with much",
"'SSZAIX",
"",
"faster convergence of the meta-objective when compared to",
"engines and online websites as demonstrated by [17, 32].",
"The crawling process is often easy to automate however re-",
"the baseline approach.",
"sults in significant amounts of label noise with complicated",
"patterns. A fundamental problem with such data, however",
"1. Introduction",
"is that classical learning methodologies tend to fail when",
"significant label noise is present [36]. Therefore, designing",
"The remarkable success of Deep Neural Networks",
"learning frameworks that are able to cope with label corrup-",
"(DNNs) for visual classification tasks is predominantly due",
"tion is a task of great importance.",
"to the availability of massive labeled datasets. In many prac-",
"While traditional methods for learning with noisy labels",
"tical applications, obtaining the required amounts of reli-",
"(LNL) are capable of handling data with immense artificial,",
"able labeled data is intractable. As a result, numerous works",
"injected noise, their ability to handle real-world noise re-",
"over the past decade have sought ways to reduce the amount",
"mains highly limited.Thus, noisylabeled datasets alone",
"of labeled data required for classification tasks. Notably,",
"limit the capability of learning from real-world noisy la-",
"semi-supervised learning exploits unlabeled data [33, 24]",
"beled datasets. Luckily, in many real-world applications.",
"transfer learning exploits prior knowledge obtained from",
"while obtaining large amounts of labeled data is infeasible,",
"different tasks [7, 39], and self-supervised learning exploits",
"obtaining small amounts of labeled data is usually attain-",
"data augmentations for label agnostic representation learn-",
"able. Thus, the paramount objective is to find methods that",
"ing [3, 11]. Nonetheless, for many applications, the quality",
"can incorporate both large amounts of noisy labeled data",
"of the labeled data can be sacrificed for the sake of its quan-",
"and small amounts of clean data. A natural choice is to",
"Iprojectpage:",
"https://sites.google.com/view/",
"adopt a meta-learning framework, a popular design regu-",
"emlc-paper.",
"larly used for solving various tasks. Consequently, recent"
] |
{
"arxiv_id": "2305.12961",
"language": "en",
"page_idx": 0
}
|
||
2305.12961-en-1
|
2305.12961-en-1.jpg
|
trends in LNL leverage meta-learning using auxiliary small clean datasets. Prominent examples include meta-sample weighting [26,38], meta-robustification to artificially injected noise[16], meta-label correction [41], and meta-soft label correction [31].
In this work, we propose EMLC- an enhanced metalabel correction approach for learning from label-corrupted data. We first revise the bi-level optimization process by deriving a more accurate meta gradient used to optimize the teacher. In particular, we derive an exact form of the meta-gradient for the one-step look-ahead approximation and suggest an improved meta-gradient approximation for the multi-step look-ahead approximation. We further provide an algorithm for efficiently computing our derivations using a modern hardware accelerator. We empirically validate our derivations, demonstrating a significant improvement in convergence speed and training time. We further propose a dedicated teacher architecture that employs a feature extraction to generate initial predictions and incorporates them with the noisy label signal to generate refined soft labels for the student. Our teacher architecture is completely independent of the student, avoiding the confirmation bias problem. In addition, we propose a novel auxiliary adversarial training objective for enhancing the effectiveness of the teacher's label correction mechanism. As demonstrated in fig. 1, our teacher proves to have a superior ability of purifying the training labels.
Our contributions can be summarized as follows:
## $\cdot$ We derive fast and more accurate procedures for computing the meta-gradient used to optimize the teacher.
· We design a unique teacher architecture in conjunction
with a novel training objective toward an improved la-
bel correction process.
$\cdot$ We combine these two components into a single yet effective framework termed EMLC. We demonstrate the effectiveness of EMLC on both synthetic and realworld label-corrupted standard benchmarks.
## $2. \textbf{ Related work}$
The LNL problem was previously addressed by various approaches. For instance, multiple works iteratively modified the labels to better align with the model's predictions [25,27], estimated the noise transition matrix [8], or applied a regularization [21,37]. The common pitfall of these methods is the phenomenon that DNNs tend to develop a confirmation bias, which causes the model to the confirm the corrupted labels. To this end, Han et al.[9] proposed a two-model framework that employs a cross-model sample selection for averting the confırmation bias problem. Being remarkably successful, many follow-up works tried to build on the two-model framework by either improving the
sample selection process or using semi-supervised tools to exploit the high-loss samples and leveraging label-agnostic pretraining [2,15,40].
Nevertheless, there is still a noticeable gap between the effectiveness of LNL methods and fully supervised training, especially on datasets with real-world noise. In numerous practical scenarios, acquiring substantial quantities of labeled data may be impractical, but obtaining a small amount of labeled data is typically achievable. This observation motivates the use of meta-learning for tackling the LNL problem. Consequently, multiple researchers have tried to exploit auxiliary small labeled datasets to apply meta-learning for LNL research.
Outside of LNL research, meta-learning has begun to appear in diverse felds in machine learning, including neural architecture search [18], hyperparameter tuning [20], fewshot learning [6] and semi-supervised learning [24]. Metalearning often involves two types of objectives: an inner (lower-level) objective and a meta (upper-level) objective. Usually, meta-learning tasks are equipped with a large training set and a small validation set. The validation data (in many cases) is necessary for updating the meta-objective.
In the context of LNL, researchers leverage metalearning to mitigate the effect of the label noise. In particular, [26, 38] try to weight samples to degrade the effect of noisy samples on the loss. MLNT [16] use a studentteacher paradigm and try to artificially inject random labels into the data used to train the student model. The student is encouraged to be close to the teacher model that does not observe the noisy data. Meta-label correction [41] and MSLC[31] follow the student-teacher paradigm but, unlike MLNT[16], use the teacher to produce soft training targets for the student. The soft targets are produced in such a way that when the student is trained with respect to these targets, it performs well on the clean data.
EMLC builds upon the meta-label correction framework, which, as presented in prior work [41,31], does not appear to have fully exploited its potential. We claim that the main pitfall of current label correction frameworks is that their teacher architecture is strongly entangled to the student, resulting in a severe confırmation bias. On the contrary, EMLC is composed of a completely independent teacher, averting this problem. In addition, EMLC leverages artificial noise injection for robust training. Opposed to[16] however, we propose adversarial noise injection that we discover to be more effective for robustifying models against label noise. Another major difference is that we use the artificial noise injection to better train the teacher whereas [16] use it to train the student.
## $3. \textbf{Methods}$
In this section we discuss the distinct components of the proposed method. We start by introducing the reader to the
|
[
[
0.080078125,
0.091552734375,
0.468017578125,
0.091552734375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
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
],
[
0.08087158203125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.5009765625,
0.12249755859375,
0.654296875,
0.121826171875,
0.654296875,
0.1363525390625,
0.5009765625,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.5185546875,
0.1363525390625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.5185546875,
0.15087890625
],
[
0.080078125,
0.15283203125,
0.22216796875,
0.15283203125,
0.22216796875,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.1514892578125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1666259765625
],
[
0.098876953125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.098876953125,
0.1817626953125
],
[
0.50341796875,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
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,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.50244140625,
0.1976318359375,
0.8935546875,
0.1976318359375,
0.8935546875,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.08087158203125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.5048828125,
0.2298583984375,
0.890625,
0.2298583984375,
0.890625,
0.240478515625,
0.5048828125,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50341796875,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50341796875,
0.272705078125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.2734375,
0.62158203125,
0.2734375,
0.62158203125,
0.287353515625,
0.50244140625,
0.287353515625
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.51953125,
0.287353515625,
0.88818359375,
0.28857421875,
0.88818359375,
0.302978515625,
0.51953125,
0.3017578125
],
[
0.080078125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.08087158203125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50244140625,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.07843017578125,
0.349853515625,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.07843017578125,
0.363037109375
],
[
0.50244140625,
0.349853515625,
0.89208984375,
0.349853515625,
0.89208984375,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50341796875,
0.365478515625,
0.890625,
0.365478515625,
0.890625,
0.37890625,
0.50341796875,
0.37890625
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.5048828125,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.391357421875,
0.5048828125,
0.391357421875
],
[
0.08087158203125,
0.395751953125,
0.46728515625,
0.395751953125,
0.46728515625,
0.406005859375,
0.08087158203125,
0.406005859375
],
[
0.50244140625,
0.39453125,
0.89208984375,
0.39453125,
0.89208984375,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.08087158203125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.5009765625,
0.410400390625,
0.87841796875,
0.410400390625,
0.87841796875,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.521484375,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.521484375,
0.438232421875
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.44140625,
0.8896484375,
0.44140625,
0.8896484375,
0.451904296875,
0.50341796875,
0.451904296875
],
[
0.080078125,
0.455078125,
0.4697265625,
0.455078125,
0.4697265625,
0.468505859375,
0.080078125,
0.468505859375
],
[
0.50244140625,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.08087158203125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.50244140625,
0.468505859375,
0.88818359375,
0.4697265625,
0.88818359375,
0.484130859375,
0.50244140625,
0.48291015625
],
[
0.080078125,
0.48486328125,
0.328369140625,
0.48486328125,
0.328369140625,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50341796875,
0.48486328125,
0.89208984375,
0.48486328125,
0.89208984375,
0.498779296875,
0.50341796875,
0.498779296875
],
[
0.09722900390625,
0.5,
0.419189453125,
0.5,
0.419189453125,
0.5146484375,
0.09722900390625,
0.5146484375
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.098876953125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.098876953125,
0.54052734375
],
[
0.50341796875,
0.53076171875,
0.89208984375,
0.53076171875,
0.89208984375,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.111083984375,
0.54150390625,
0.464111328125,
0.54150390625,
0.464111328125,
0.5546875,
0.111083984375,
0.5546875
],
[
0.50244140625,
0.54443359375,
0.890625,
0.54443359375,
0.890625,
0.55810546875,
0.50244140625,
0.55810546875
],
[
0.098876953125,
0.56494140625,
0.468017578125,
0.56640625,
0.468017578125,
0.58056640625,
0.098876953125,
0.580078125
],
[
0.50341796875,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.50341796875,
0.57568359375,
0.890625,
0.57568359375,
0.890625,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.11273193359375,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.59521484375,
0.11273193359375,
0.59521484375
],
[
0.1094970703125,
0.59521484375,
0.2646484375,
0.5966796875,
0.2646484375,
0.611328125,
0.1094970703125,
0.60986328125
],
[
0.50341796875,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.60498046875,
0.50341796875,
0.60498046875
],
[
0.50244140625,
0.60595703125,
0.72802734375,
0.60595703125,
0.72802734375,
0.619140625,
0.50244140625,
0.619140625
],
[
0.098876953125,
0.62060546875,
0.4697265625,
0.6220703125,
0.4697265625,
0.63623046875,
0.098876953125,
0.63525390625
],
[
0.52294921875,
0.62060546875,
0.890625,
0.62060546875,
0.890625,
0.6337890625,
0.52294921875,
0.6337890625
],
[
0.111083984375,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.65087890625,
0.111083984375,
0.65087890625
],
[
0.50341796875,
0.63525390625,
0.89208984375,
0.63525390625,
0.89208984375,
0.6484375,
0.50341796875,
0.6484375
],
[
0.11273193359375,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.11273193359375,
0.66552734375
],
[
0.5009765625,
0.6494140625,
0.89208984375,
0.65087890625,
0.89208984375,
0.66552734375,
0.5009765625,
0.6640625
],
[
0.11273193359375,
0.66748046875,
0.40283203125,
0.66748046875,
0.40283203125,
0.68115234375,
0.11273193359375,
0.68115234375
],
[
0.50341796875,
0.66650390625,
0.890625,
0.66650390625,
0.890625,
0.6796875,
0.50341796875,
0.6796875
],
[
0.50244140625,
0.681640625,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.080078125,
0.693359375,
0.214111328125,
0.6943359375,
0.214111328125,
0.708984375,
0.080078125,
0.70849609375
],
[
0.50244140625,
0.69580078125,
0.89208984375,
0.6962890625,
0.89208984375,
0.7109375,
0.50244140625,
0.71044921875
],
[
0.50341796875,
0.7109375,
0.89208984375,
0.7109375,
0.89208984375,
0.724609375,
0.50341796875,
0.724609375
],
[
0.09722900390625,
0.71826171875,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.09722900390625,
0.73291015625
],
[
0.50341796875,
0.7265625,
0.890625,
0.7265625,
0.890625,
0.74072265625,
0.50341796875,
0.74072265625
],
[
0.07843017578125,
0.73486328125,
0.46728515625,
0.734375,
0.46728515625,
0.74853515625,
0.07843017578125,
0.75
],
[
0.5009765625,
0.74072265625,
0.890625,
0.7412109375,
0.890625,
0.7548828125,
0.5009765625,
0.75390625
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.7568359375,
0.890625,
0.7568359375,
0.890625,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.08087158203125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.5009765625,
0.77099609375,
0.8935546875,
0.771484375,
0.8935546875,
0.7861328125,
0.5009765625,
0.78515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.07843017578125,
0.79443359375,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.07843017578125,
0.8095703125
],
[
0.50341796875,
0.78857421875,
0.8896484375,
0.78857421875,
0.8896484375,
0.798828125,
0.50341796875,
0.798828125
],
[
0.50244140625,
0.8017578125,
0.890625,
0.8017578125,
0.890625,
0.81494140625,
0.50244140625,
0.81494140625
],
[
0.08087158203125,
0.8125,
0.46728515625,
0.8125,
0.46728515625,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.50244140625,
0.8173828125,
0.67236328125,
0.8173828125,
0.67236328125,
0.83056640625,
0.50244140625,
0.83056640625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84521484375,
0.5986328125,
0.84521484375,
0.5986328125,
0.85986328125,
0.50244140625,
0.85986328125
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.51953125,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.51953125,
0.8837890625
],
[
0.07843017578125,
0.88623046875,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.5009765625,
0.88720703125,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.9013671875
]
] |
[
"trends in LNL leverage meta-learning using auxiliary small",
"sample selection process or using semi-supervised tools to",
"clean datasets. Prominent examples include meta-sample",
"exploit the high-loss samples and leveraging label-agnostic",
"weighting [26, 38], meta-robustification to artificially in-",
"pretraining [2, 15, 40].",
"jected noise [16], meta-label correction [41], and meta-soft",
"Nevertheless, there is still a noticeable gap between the",
"label correction [31].",
"effectiveness of LNL methods and fully supervised training,",
"In this work, we propose EMLC - an enhanced meta-",
"especially on datasets with real-world noise.In numerous",
"label correction approach for learning from label-corrupted",
"practical scenarios, acquiring substantial quantities of la-",
"data.We first revise the bi-level optimization process by",
"beled data may be impractical, but obtaining a small amount",
"deriving a more accurate meta gradient used to optimize",
"of labeled data is typically achievable. This observation mo-",
"the teacher. In particular, we derive an exact form of the",
"tivatestheuseof meta-learningfortacklingtheLNLprob-",
"meta-gradient for the one-step look-ahead approximation",
"lem. Consequently, multiple researchers have tried to ex-",
"and suggest an improved meta-gradient approximation for",
"ploit auxiliary small labeled datasets to apply meta-learning",
"the multi-step look-ahead approximation. We further pro-",
"for LNL research.",
"vide an algorithm for efficiently computing our derivations",
"Outside of LNL research, meta-learning has begun to ap",
"using a modern hardware accelerator. We empirically val-",
"pear in diverse fields in machine learning, including neural",
"idateour derivations.demonstrating a significant improve-",
"architecture search [18], hyperparameter tuning [20], few-",
"ment in convergence speed and training time. We further",
"shot learning [6] and semi-supervised learning [24]. Meta-",
"propose a dedicated teacher architecture that employs a fea-",
"learning often involves two types of objectives: an inner",
"ture extraction to generate initial predictions and incorpo-",
"(lower-level) objective and a meta (upper-level) objective.",
"ratesthemwiththenoisylabel signal togenerate refined",
"Usually,meta-learningtasks areequipped witha largetrain-",
"softlabelsforthestudent.Ourteacherarchitectureis com-",
"ing set and a small validation set. The validation data (in",
"pletely independent of the student, avoiding the confirma-",
"many cases) is necessary for updating the meta-objective.",
"tion bias problem. In addition, we propose a novel auxil-",
"In the context of LNL, researchers leverage meta-",
"iary adversarial training objective for enhancing the effec-",
"learningtomitigatetheeffectof the label noise.Inpar-",
"tiveness of the teacher's label correction mechanism. As",
"ticular, [26, 38] try to weight samples to degrade the effect",
"demonstrated in fig. 1, our teacher proves to have a superior",
"of noisy samples on the loss. MLNT [16] use a student-",
"ability of purifying the training labels",
"teacher paradigm and try to artificially inject random labels",
" Our contributions can be summarized as follows",
"into the data used to train the student model. The student",
"isencouraged tobe close to the teachermodel that does",
".We derive fast and more accurate procedures for com-",
"not observe the noisy data. Meta-label correction [41] and",
"puting the meta-gradient used to optimize the teacher.",
"MSLC [31] follow the student-teacher paradigm but, unlike",
".We design a unique teacher architecture in conjunction",
"MLNT [16], use the teacher to produce soft training targets",
"for the student. The soft targets are produced in such a way",
"with a novel training objective toward an improved la-",
"bel correction process.",
"that when the student is trained with respect to these targets,",
"it performs well on the clean data.",
".We combine these two components into a single yet",
"EMLC builds upon the meta-label correction framework,",
"effective framework termed EMLC.We demonstrate",
"which, as presented in prior work [41, 31], does not appear",
"the effectiveness of EMLC on both synthetic and real-",
"to have fully exploited its potential. We claim that the main",
"world label-corrupted standard benchmarks.",
"pitfall of current label correction frameworks is that their",
"teacher architecture is strongly entangled to the student,",
"2. Related work",
"resulting in a severe confirmation bias. On the contrary,",
"EMLC is composed of a completely independent teacher,",
"The LNL problem was previously addressed by various",
"averting this problem. In addition, EMLC leverages arti-",
"approaches. For instance, multiple works iteratively mod-",
"ficial noise injection for robust training. Opposed to [16]",
"ified the labels to better align with the model's predictions",
"however, we propose adversarial noise injection that we dis-",
"[25, 27], estimated the noise transition matrix [8], or ap-",
"cover to be more effective for robustifying models against",
"plied a regularization [21, 37]. The common pitfall of these",
"methods is the phenomenon that DNNs tend to develop a",
"labelnoise.Anothermajordifferenceisthatweusethearti-",
"ficial noise injection to better train the teacher whereas [16]",
"confirmationbias,which causes the modelto theconfirm",
"use it to train the student.",
"the corrupted labels. To this end, Han et al. [9] proposed a",
"two-model framework that employs a cross-model sample",
"3. Methods",
"selection for averting the confirmation bias problem. Be-",
"ing remarkably successful, many follow-up works tried to",
"In this section we discuss the distinct components of the",
"build on the two-model framework by either improving the",
"proposed method. We start by introducing the reader to the"
] |
{
"arxiv_id": "2305.12961",
"language": "en",
"page_idx": 1
}
|
||
2305.12961-en-3
|
2305.12961-en-3.jpg
|
## $3. 3. \textbf{Bi- Level Optimization}$
Bi-level optimization problems appear in diverse subfields in machine learning and in meta-learning in particular [18,20,6,24]. Such problems are composed of an outer level optimization problem that is usually constrained by another inner level problem. We relate to the upper level problem parameters as the meta-parameters and to the parameters in the lower level as the main parameters. The fact that each meta-parameter value defines a new lower level problem makes applying gradient-based optimization methods to such situations particularly difficult due to the iteration complexity. A solution to this problem is to approximate $w^*$ with $k$-step look-ahead SGD, optimizing both the main and meta- parameters simultaneously. The joint optimization process can be summarized in algorithm l.
<table>
<tbody>
<tr>
<td>$^{\prime}$</td>
<td>$\mathbf{N}(\mathbf{i})$ $|.00k$</td>
</tr>
<tr>
<td>a</td>
<td>proximation</td>
</tr>
</tbody>
</table>
$\|$ Sample clean and noisy batches from the datasets
2: 3:
$\mathcal{B},\widetilde{\mathcal{B}}\leftarrow Sample(\mathcal{D}),Sample(\widetilde{\mathcal{D}})$
$\|$ Update $w$ by descending $\widetilde{\mathcal{L}}(w,\alpha)$ w.r.t $w$
4:
$w^{(t+1)}\leftarrow w^{(t)}-\eta_w\nabla_w\widetilde{\mathcal{L}}(w,\alpha^{(t)})\bigg|$
5:
$\left|_{w=w^{(t)}}\right|_{w=w^{(t)}}$
$\textbf{if}t\mathrm{mod}k= k- 1\textbf{ then}$
6:
Input: Clean and noisy datasets $\mathcal{D},\widetilde{\mathcal{D}}$, number of train-
ing steps $T$, initial parameters $w^(0),\alpha^{(0)}$, learning rates
$\eta_{w},\eta_{\alpha}.$
Output: Optimized parameters $w^{(T)},\alpha^{(T)}.$
$1: \textbf{for }t= 0, . . . , T- 1\textbf{do}$
$\|$ Unroll $w^{(t+1)}(\alpha)$ to compute the meta gradient
7: 8:
$\begin{array}{l}{g_{\alpha}=\mathrm{META\_GRAD}(\mathcal{L}(w^{(t+1)}),\alpha)}\\{\|\mathrm{Update~}\alpha\text{ by descending }\mathcal{L}(w^{(t+1)}(\alpha))\mathrm{~w.r.t~}\alpha}\end{array}$
9:
$\alpha^{(t+1)}\leftarrow\alpha^{(t)}-\eta_{\alpha}g_{\alpha}$
10:
else
11:
${\dot{\alpha}^{(t+1)}}\leftarrow\alpha^{(t)}$
12:
end if
13:
$14: \textbf{end for}$
During training, the higher order dependencies of $w$ on $\alpha$ are neglected, thereby allowing the meta-gradient to be computed, as we discuss in the next section.
$3. 4. \textbf{ Revisiting the Meta- Gradient Approximation}$
As opposed to Meta-Label Correction [41], we propose better approximations for both the single and multi-step meta-gradients when updating the teacher. In the next section, we establish an efficient method for carrying out the meta-gradient computation.
We explicitly differentiate between the cases $k=1$ and $k>1.$ We begin by deriving the $one-step$ approximation meta-gradient.
Proposition 1. (One-Step Meta-Gradient) Let $g_w:=$ $\nabla _{w}\bar{\mathcal{L} } ( w^{( t+ 1) })$ where $w^{( t+ 1) }$ are the $student^{\prime }s$ new $pa-$ rameters obtained when last updated. Let H' := $\nabla _{w\alpha }\widetilde{\mathcal{L} } ( w^{( t) }, \alpha ^{( t) })$ where $w^{( t) }$ are the $student^{\prime }s$ original $pa-$ rameters (before'its update). Let $\alpha^{(t)}$ be the teacher's parameters before its update. Then, the one-step metagradient at time t denoted by $\nabla_\alpha^{(t)}$ is given by:
$$[\nabla_\alpha^{(t)}]^T=-\eta_wg_w^TH_{w\alpha}^{(t)}$$
(7)
The proof of proposition l is given in appendix A.1. We proceed by considering the case of $k>1.$ In this case, we do not compute the exact meta-gradient. Instead, we approximate it.
Proposition 2. (k-Step Meta-Gradient Approximation with $Exponential\textit{ Moving Average of Mixed Hessians) Assume}$ $\begin{array}{l}{that~t~\mathrm{mod}\:k=k-1~and~let~g_{w}:=\nabla_{w}{\mathcal{L}}(w^{(t+1)})~where}\\{w^{(t+1)}~is~the~student's~new~parameters~obtained~when~last}\end{array}$ updated. Consider the mixed Hessians in the last k steps: $\begin{array}{l}{{\dot{H}_{w\alpha}^{(\tau)}:=\:\nabla_{w\alpha}\widetilde{\mathcal{L}}(w^{(\tau)},\alpha^{(\tau)})\:where\:t-k+1\leq\tau\leq t.}}\\{{\dot{H}_{w\alpha}^{(\tau)}:=\:\nabla_{w\alpha}\widetilde{\mathcal{L}}(w^{(\tau)},\alpha^{(\tau)})\:where\:t-k+1\leq\tau\leq t.}}\\{{\dot{H}_{w\alpha}^{(\tau)}(t)}}\end{array}$ Let $\alpha ^{( t) }$ = $\alpha ^{( t- 1) }$ = ... = $\alpha ^{( t- k+ 1) }$ $be$ the $teacher^{\prime }s$ $pa-$ rameters before its update. Then, the k-step meta-gradient denotedby $\nabla _{\alpha }$ can be approximated by:
(8)
$$[\nabla_\alpha^{(t)}]^T\approx-\eta_wg_w^T\sum_{\tau=t-k+1}^t\gamma_w^{t-\tau}H_{w\alpha}^{(\tau)}$$
where $\gamma _{w}= 1- \eta _{w}.$
The proof of proposition 2 is given in appendix A.2.
3.5.Computational Efficiency of the Meta-Gradient
Calculating the one-step meta-gradient as in proposition 1 or the multi-step meta-gradient as in proposition 2 in a straightforward manner is inefficient, as it requires computing higher-order derivatives of large DNNs. In this section, we provide a few workarounds that allow us to carry out the computations efficiently. Our first observation is that the sample-wise mixed Hessian can be expressed as a product of two frst-order Jacobians. Formally:
Proposition 3. Let us consider the sample-wise lower $loss\nobreak {: } $ $\widetilde{\ell } _{i}( w, \alpha )$ := $CE( q_{\alpha }( y| x_{i}, \widetilde{y} _{i}) , p_{w}( \dot{y} | x_{i}) ) .$ Recall $from~eq.~(6)~that~\widetilde{\mathcal{L}}(w,\alpha)~=~\frac{1}{n}\sum_{i=1}^{n}\widetilde{\ell}_{i}(w,\alpha).~De-$ note $by$ $J_{w}( i)$ := $J_{w}( \log ( p_{w}( y| x_{i}) ) )$ and $by$ $J_{\alpha }( i)$ := $J_{\alpha }( q_{\alpha }( y| x_{i}, \widetilde{y} _{i}) ) .$ $We$ claim $that:$
$$\nabla_{w\alpha}\widetilde{\ell_i}(w,\alpha)=-[J_w(i)]^T[J_\alpha(i)]$$
(9)
The proof of proposition 3 is given in appendix A.3.
Corollary 3.1. Using proposition l, the one-step meta-
gradient can be expressed as:
(10)
$$[\nabla_\alpha^{(t)}]^T=\dfrac{\eta_w}{n}\sum_{i=1}^ng_w^T[J_{w^{(t)}}(i)]^T[J_{\alpha^{(t)}}(i)]$$
|
[
[
0.08087158203125,
0.091552734375,
0.28515625,
0.091552734375,
0.28515625,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.50244140625,
0.09027099609375,
0.859375,
0.09027099609375,
0.859375,
0.10479736328125,
0.50244140625,
0.10479736328125
],
[
0.87353515625,
0.0985107421875,
0.88232421875,
0.0985107421875,
0.88232421875,
0.10162353515625,
0.87353515625,
0.10162353515625
],
[
0.50244140625,
0.10479736328125,
0.8935546875,
0.10479736328125,
0.8935546875,
0.12249755859375,
0.50244140625,
0.12249755859375
],
[
0.09722900390625,
0.11297607421875,
0.468017578125,
0.1142578125,
0.468017578125,
0.1287841796875,
0.09722900390625,
0.1275634765625
],
[
0.07843017578125,
0.1287841796875,
0.468017578125,
0.1300048828125,
0.468017578125,
0.14453125,
0.07843017578125,
0.143310546875
],
[
0.50244140625,
0.125,
0.859375,
0.12249755859375,
0.859375,
0.136962890625,
0.50244140625,
0.1395263671875
],
[
0.86767578125,
0.1300048828125,
0.8857421875,
0.1300048828125,
0.8857421875,
0.1337890625,
0.86767578125,
0.1337890625
],
[
0.50341796875,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.154052734375,
0.50341796875,
0.154052734375
],
[
0.080078125,
0.1458740234375,
0.468017578125,
0.1458740234375,
0.468017578125,
0.1590576171875,
0.080078125,
0.1590576171875
],
[
0.50244140625,
0.1546630859375,
0.89208984375,
0.1546630859375,
0.89208984375,
0.1685791015625,
0.50244140625,
0.1685791015625
],
[
0.08251953125,
0.1622314453125,
0.46728515625,
0.1622314453125,
0.46728515625,
0.17236328125,
0.08251953125,
0.17236328125
],
[
0.5009765625,
0.169189453125,
0.8896484375,
0.169189453125,
0.8896484375,
0.18310546875,
0.5009765625,
0.18310546875
],
[
0.080078125,
0.175537109375,
0.468017578125,
0.175537109375,
0.468017578125,
0.189453125,
0.080078125,
0.189453125
],
[
0.5,
0.1849365234375,
0.80810546875,
0.184326171875,
0.80810546875,
0.2020263671875,
0.5,
0.20263671875
],
[
0.07843017578125,
0.1912841796875,
0.468017578125,
0.1900634765625,
0.468017578125,
0.203857421875,
0.07843017578125,
0.2052001953125
],
[
0.08087158203125,
0.2083740234375,
0.46728515625,
0.2083740234375,
0.46728515625,
0.2183837890625,
0.08087158203125,
0.2183837890625
],
[
0.61669921875,
0.203857421875,
0.775390625,
0.205810546875,
0.775390625,
0.2279052734375,
0.61669921875,
0.2259521484375
],
[
0.869140625,
0.2108154296875,
0.89208984375,
0.2108154296875,
0.89208984375,
0.2247314453125,
0.869140625,
0.2247314453125
],
[
0.080078125,
0.2215576171875,
0.468017578125,
0.2215576171875,
0.468017578125,
0.23486328125,
0.080078125,
0.23486328125
],
[
0.080078125,
0.2374267578125,
0.46728515625,
0.2374267578125,
0.46728515625,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.521484375,
0.23291015625,
0.89208984375,
0.2342529296875,
0.89208984375,
0.248779296875,
0.521484375,
0.2474365234375
],
[
0.080078125,
0.251953125,
0.46728515625,
0.251953125,
0.46728515625,
0.26513671875,
0.080078125,
0.26513671875
],
[
0.50244140625,
0.25,
0.8896484375,
0.25,
0.8896484375,
0.26318359375,
0.50244140625,
0.26318359375
],
[
0.080078125,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.27978515625,
0.080078125,
0.27978515625
],
[
0.50244140625,
0.26513671875,
0.89208984375,
0.26513671875,
0.89208984375,
0.279052734375,
0.50244140625,
0.279052734375
],
[
0.080078125,
0.281005859375,
0.468017578125,
0.281005859375,
0.468017578125,
0.294189453125,
0.080078125,
0.294189453125
],
[
0.50244140625,
0.27978515625,
0.60546875,
0.276611328125,
0.60546875,
0.291748046875,
0.50244140625,
0.294189453125
],
[
0.080078125,
0.29541015625,
0.46728515625,
0.29541015625,
0.46728515625,
0.309326171875,
0.080078125,
0.309326171875
],
[
0.50341796875,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.314453125,
0.50341796875,
0.314453125
],
[
0.080078125,
0.3125,
0.4248046875,
0.3125,
0.4248046875,
0.32568359375,
0.080078125,
0.32568359375
],
[
0.50341796875,
0.31640625,
0.890625,
0.31640625,
0.890625,
0.330078125,
0.50341796875,
0.330078125
],
[
0.5009765625,
0.330810546875,
0.890625,
0.330078125,
0.890625,
0.3447265625,
0.5009765625,
0.34521484375
],
[
0.08251953125,
0.34228515625,
0.468017578125,
0.34228515625,
0.468017578125,
0.35595703125,
0.08251953125,
0.35595703125
],
[
0.5,
0.343505859375,
0.8935546875,
0.3447265625,
0.8935546875,
0.3623046875,
0.5,
0.361083984375
],
[
0.07843017578125,
0.356689453125,
0.2254638671875,
0.35595703125,
0.2254638671875,
0.37060546875,
0.07843017578125,
0.371826171875
],
[
0.50244140625,
0.3623046875,
0.890625,
0.3623046875,
0.890625,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.07843017578125,
0.376953125,
0.468017578125,
0.376953125,
0.468017578125,
0.39453125,
0.07843017578125,
0.39453125
],
[
0.50244140625,
0.375732421875,
0.890625,
0.375732421875,
0.890625,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.10540771484375,
0.393310546875,
0.471435546875,
0.393310546875,
0.471435546875,
0.410888671875,
0.10540771484375,
0.410888671875
],
[
0.5009765625,
0.39013671875,
0.89208984375,
0.39013671875,
0.89208984375,
0.40771484375,
0.5009765625,
0.40771484375
],
[
0.108642578125,
0.41357421875,
0.156005859375,
0.41357421875,
0.156005859375,
0.4248046875,
0.108642578125,
0.4248046875
],
[
0.50341796875,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.42236328125,
0.50341796875,
0.42236328125
],
[
0.07843017578125,
0.423583984375,
0.36767578125,
0.421630859375,
0.36767578125,
0.439453125,
0.07843017578125,
0.44140625
],
[
0.5009765625,
0.42236328125,
0.7685546875,
0.423583984375,
0.7685546875,
0.438232421875,
0.5009765625,
0.4375
],
[
0.0882568359375,
0.43994140625,
0.259033203125,
0.43994140625,
0.259033203125,
0.453857421875,
0.0882568359375,
0.453857421875
],
[
0.57177734375,
0.45263671875,
0.70166015625,
0.455078125,
0.70166015625,
0.473388671875,
0.57177734375,
0.4716796875
],
[
0.7060546875,
0.451904296875,
0.73388671875,
0.460205078125,
0.72607421875,
0.474853515625,
0.69775390625,
0.46533203125
],
[
0.744140625,
0.455810546875,
0.8203125,
0.453857421875,
0.82177734375,
0.472900390625,
0.74609375,
0.474853515625
],
[
0.1298828125,
0.455810546875,
0.471435546875,
0.455810546875,
0.471435546875,
0.4697265625,
0.1298828125,
0.4697265625
],
[
0.08905029296875,
0.458984375,
0.10455322265625,
0.458984375,
0.10455322265625,
0.468505859375,
0.08905029296875,
0.468505859375
],
[
0.869140625,
0.458984375,
0.89208984375,
0.458984375,
0.89208984375,
0.472900390625,
0.869140625,
0.472900390625
],
[
0.08905029296875,
0.472900390625,
0.10296630859375,
0.472900390625,
0.10296630859375,
0.484130859375,
0.08905029296875,
0.484130859375
],
[
0.13232421875,
0.47021484375,
0.35693359375,
0.47021484375,
0.35693359375,
0.48486328125,
0.13232421875,
0.48486328125
],
[
0.708984375,
0.477783203125,
0.74853515625,
0.477783203125,
0.74853515625,
0.48486328125,
0.708984375,
0.48486328125
],
[
0.0882568359375,
0.488037109375,
0.10540771484375,
0.488037109375,
0.10540771484375,
0.501953125,
0.0882568359375,
0.501953125
],
[
0.131591796875,
0.488037109375,
0.414306640625,
0.488037109375,
0.414306640625,
0.501953125,
0.131591796875,
0.501953125
],
[
0.50244140625,
0.49365234375,
0.63818359375,
0.49560546875,
0.63818359375,
0.51025390625,
0.50244140625,
0.50830078125
],
[
0.08660888671875,
0.51025390625,
0.10540771484375,
0.51025390625,
0.10540771484375,
0.5244140625,
0.08660888671875,
0.5244140625
],
[
0.1282958984375,
0.505859375,
0.35302734375,
0.5068359375,
0.35302734375,
0.52587890625,
0.1282958984375,
0.5244140625
],
[
0.521484375,
0.5146484375,
0.86279296875,
0.51513671875,
0.86279296875,
0.52978515625,
0.521484375,
0.52880859375
],
[
0.08660888671875,
0.5341796875,
0.10455322265625,
0.5341796875,
0.10455322265625,
0.54736328125,
0.08660888671875,
0.54736328125
],
[
0.131591796875,
0.53271484375,
0.29736328125,
0.53271484375,
0.29736328125,
0.546875,
0.131591796875,
0.546875
],
[
0.15283203125,
0.54443359375,
0.48046875,
0.54541015625,
0.48046875,
0.56298828125,
0.15283203125,
0.5625
],
[
0.50341796875,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.08660888671875,
0.54736328125,
0.10540771484375,
0.54736328125,
0.10540771484375,
0.5625,
0.08660888671875,
0.5625
],
[
0.15283203125,
0.56103515625,
0.381591796875,
0.56005859375,
0.381591796875,
0.5791015625,
0.15283203125,
0.580078125
],
[
0.52294921875,
0.5625,
0.890625,
0.5625,
0.890625,
0.57568359375,
0.52294921875,
0.57568359375
],
[
0.1544189453125,
0.57568359375,
0.473876953125,
0.57568359375,
0.473876953125,
0.59326171875,
0.1544189453125,
0.59326171875
],
[
0.08660888671875,
0.56298828125,
0.10540771484375,
0.56298828125,
0.10540771484375,
0.57763671875,
0.08660888671875,
0.57763671875
],
[
0.08660888671875,
0.57763671875,
0.10455322265625,
0.57763671875,
0.10455322265625,
0.59228515625,
0.08660888671875,
0.59228515625
],
[
0.5009765625,
0.57568359375,
0.89208984375,
0.5771484375,
0.89208984375,
0.59130859375,
0.5009765625,
0.59033203125
],
[
0.15283203125,
0.58837890625,
0.3095703125,
0.59228515625,
0.3095703125,
0.60986328125,
0.15283203125,
0.60595703125
],
[
0.08251953125,
0.5947265625,
0.10455322265625,
0.5947265625,
0.10455322265625,
0.60693359375,
0.08251953125,
0.60693359375
],
[
0.50341796875,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.50341796875,
0.60595703125
],
[
0.08251953125,
0.609375,
0.10455322265625,
0.609375,
0.10455322265625,
0.6220703125,
0.08251953125,
0.6220703125
],
[
0.131591796875,
0.609375,
0.16015625,
0.609375,
0.16015625,
0.62060546875,
0.131591796875,
0.62060546875
],
[
0.50244140625,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50244140625,
0.6220703125
],
[
0.08251953125,
0.62353515625,
0.10455322265625,
0.62353515625,
0.10455322265625,
0.63720703125,
0.08251953125,
0.63720703125
],
[
0.156005859375,
0.62060546875,
0.25244140625,
0.62060546875,
0.25244140625,
0.63525390625,
0.156005859375,
0.63525390625
],
[
0.5048828125,
0.625,
0.890625,
0.625,
0.890625,
0.63525390625,
0.5048828125,
0.63525390625
],
[
0.1298828125,
0.63720703125,
0.17724609375,
0.63720703125,
0.17724609375,
0.6513671875,
0.1298828125,
0.6513671875
],
[
0.50244140625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.5009765625,
0.6513671875,
0.8935546875,
0.65283203125,
0.8935546875,
0.66748046875,
0.5009765625,
0.66650390625
],
[
0.08251953125,
0.6396484375,
0.10455322265625,
0.6396484375,
0.10455322265625,
0.6513671875,
0.08251953125,
0.6513671875
],
[
0.08251953125,
0.65283203125,
0.160888671875,
0.65283203125,
0.160888671875,
0.66748046875,
0.08251953125,
0.66748046875
],
[
0.50244140625,
0.66845703125,
0.75732421875,
0.66845703125,
0.75732421875,
0.681640625,
0.50244140625,
0.681640625
],
[
0.1004638671875,
0.6875,
0.468017578125,
0.6875,
0.468017578125,
0.7021484375,
0.1004638671875,
0.7021484375
],
[
0.50341796875,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50341796875,
0.70263671875
],
[
0.08087158203125,
0.70263671875,
0.468017578125,
0.70263671875,
0.468017578125,
0.71630859375,
0.08087158203125,
0.71630859375
],
[
0.5009765625,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.71728515625,
0.36767578125,
0.71728515625,
0.36767578125,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.496826171875,
0.71728515625,
0.89453125,
0.71826171875,
0.89453125,
0.7392578125,
0.496826171875,
0.73779296875
],
[
0.50244140625,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.751953125,
0.50244140625,
0.751953125
],
[
0.08087158203125,
0.74365234375,
0.459228515625,
0.74365234375,
0.459228515625,
0.7568359375,
0.08087158203125,
0.7568359375
],
[
0.50244140625,
0.751953125,
0.7197265625,
0.75048828125,
0.7197265625,
0.76513671875,
0.50244140625,
0.7666015625
],
[
0.1004638671875,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.78076171875,
0.1004638671875,
0.78076171875
],
[
0.5849609375,
0.7763671875,
0.810546875,
0.7763671875,
0.810546875,
0.79150390625,
0.5849609375,
0.79150390625
],
[
0.87109375,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79150390625,
0.87109375,
0.79150390625
],
[
0.080078125,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.796875,
0.080078125,
0.796875
],
[
0.08087158203125,
0.7998046875,
0.465576171875,
0.7998046875,
0.465576171875,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.52294921875,
0.80029296875,
0.86376953125,
0.80029296875,
0.86376953125,
0.814453125,
0.52294921875,
0.814453125
],
[
0.07843017578125,
0.81298828125,
0.468017578125,
0.81298828125,
0.468017578125,
0.8271484375,
0.07843017578125,
0.8271484375
],
[
0.50244140625,
0.82080078125,
0.890625,
0.82275390625,
0.890625,
0.8369140625,
0.50244140625,
0.83544921875
],
[
0.080078125,
0.82763671875,
0.26318359375,
0.82763671875,
0.26318359375,
0.84130859375,
0.080078125,
0.84130859375
],
[
0.50244140625,
0.83837890625,
0.69921875,
0.83837890625,
0.69921875,
0.8515625,
0.50244140625,
0.8515625
],
[
0.101318359375,
0.8447265625,
0.468017578125,
0.8447265625,
0.468017578125,
0.85498046875,
0.101318359375,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.4697265625,
0.85791015625,
0.4697265625,
0.87255859375,
0.07843017578125,
0.8720703125
],
[
0.55615234375,
0.85986328125,
0.8388671875,
0.85986328125,
0.8388671875,
0.8896484375,
0.55615234375,
0.8896484375
],
[
0.08087158203125,
0.875,
0.178955078125,
0.875,
0.178955078125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.859375,
0.87060546875,
0.890625,
0.87060546875,
0.890625,
0.88330078125,
0.859375,
0.88330078125
],
[
0.6357421875,
0.880859375,
0.6611328125,
0.880859375,
0.6611328125,
0.892578125,
0.6357421875,
0.892578125
],
[
0.66015625,
0.8896484375,
0.6787109375,
0.8896484375,
0.6787109375,
0.89794921875,
0.66015625,
0.89794921875
]
] |
[
"3.3.Bi-Level Optimization",
"Proposition 1. (One-Step Meta-Gradient) Let gu",
"--",
"VuC(w(t+1)) where w(t+1) are the student's new pa-",
"Bi-level optimization problems appear in diverse sub-",
"fields in machine learning and in meta-learning in partic-",
"rameters obtained when last updated. Let H(t)",
"",
"VwaC(w(t), α(t)) where w(t) are the student's original pa-",
"ular [18, 20, 6, 24]. Such problems are composed of an",
"rameters (before its update). Let α(t) be the teacher's",
"outerleveloptimizationproblemthatisusuallyconstrained",
"parameters before its update. Then, the one-step meta-",
"by another inner level problem. We relate to the upper level",
"gradient at time t denoted by t) is given by:",
"problem parameters as the meta-parameters and to the pa-",
"rametersinthelowerlevelasthemainparameters.Thefact",
"[(]T = g",
"(7)",
"that each meta-parameter value defines a new lower level",
"problem makes applying gradient-based optimization meth-",
"The proof of proposition 1 is given in appendix A.1. We",
"ods to such situations particularly difficult due to the itera-",
"proceed by considering the case of k > 1. In this case,",
"tion complexity. A solution to this problem is to approxi-",
"we do not compute the exact meta-gradient. Instead, we",
"mate w* with k-step look-ahead SGD, optimizing both the",
"approximate it.",
"main and meta- parameters simultaneously. The joint opti-",
"Proposition 2. (k-Step Meta-Gradient Approximation with",
"mization process can be summarized in algorithm 1.",
"Exponential Moving Average of Mixed Hessians) Assume",
"that t mod k = k - 1 and let gw := wC(w(t+1)) where",
"Algorithm 1 Bi-Level Optimization via k-step SGD Look-",
"w(t+1) is the student's new parameters obtained when last",
"ahead Approximation",
"updated. Consider the mixed Hessians in the last k steps:",
"Input: Clean and noisy datasets D, D, number of train-",
"H := VwaL(w(t),a(t) where t - k +1 ≤ T ≤ t.",
"ing steps T, initial parameters w(0), α(0), learning rates",
"Let α(t) = α(t-i) --... - Q(t-h+1) be the teacher's pa-",
"Nw,Na.",
"rameters before its update. Then, the k-step meta-gradient",
"Output: Optimized parameters w(T), Q(T)",
"denoted by & can be approximated by:",
"1:fort=0..T-1do",
"[V(t)]T ~-ng ",
"N",
"-TH(T)",
"Il Sample clean and noisy batches from the datasets",
"2:",
"(8)",
"3:",
"B, B Sample(D), Sample(D)",
"t-k十",
"4:",
"//UpdatewbydescendingC(w,a)w.r.tw",
"where w =1-nw.",
"5:",
"w(t+1) w(t) -nwVwL(w,Q(t",
"The proof of proposition 2 is given in appendix A.2",
"6:",
"if t mod k = k - 1 then",
"// Unroll w(t+1)(a) to compute the meta gradient",
"3.5.Computational Efficiency of the Meta-Gradient",
"7:",
"ga = META-GRAD(C(w(++1),α)",
"Calculating the one-step meta-gradient as in proposi-",
"// Update αbydescending C(w(t+1)(a)w.r.t a",
"8:",
"9:",
"tion 1 or the multi-step meta-gradient as in proposition 2 in a",
"Q(t+1) α(t) _naga",
"10:",
"straightforward manner is inefficient, as it requires comput-",
"11:",
"else",
"ing higher-order derivatives of large DNNs. In this section,",
"12: ",
"Q(t+1) Q(t)",
"weprovideafewworkaroundsthatallowustocarryoutthe",
"end if",
"computations efficiently. Our first observation is that the",
"sample-wise mixed Hessian can be expressed as a product",
"13:",
"14: end for",
"of two first-order Jacobians. Formally:",
"During training, the higher order dependencies of w on",
"Proposition 3. Let us consider the sample-wise lower",
"Q are neglected, thereby allowing the meta-gradient to be",
"loss: l;(w,a) := CE(qa(ylxi,yi),pw(yri)). Recall",
"computed, as we discuss in the next section.",
"from eq. (6) that C(w,a) = Er-lei(w,a). De-",
"note by Jw(i) := Ju(log(puw(ylci))) and by Ja(i) :=",
"3.4.Revisiting the Meta-Gradient Approximation",
"Je(qa (ylci, yi)). We claim that:",
"As opposed to Meta-Label Correction [41], we propose",
"Vwal;(w, α) = -[Jw()]T [Ja()]",
"(9)",
"better approximations for both the single and multi-step",
"meta-gradientswhenupdatingtheteacher.Inthenext sec-",
"The proof of proposition 3 is given in appendix A.3.",
"tion, we establish an efficient method for carrying out the",
"Corollary 3.1. Using proposition I, the one-step meta-",
"meta-gradient computation.",
"gradient can be expressed as:",
"Weexplicitlydifferentiatebetweenthecasesk=land",
"k > 1. We begin by deriving the one-step approximation",
"[] =[()][a(0)]",
"meta-gradient.",
"(10)",
"n",
"=1"
] |
{
"arxiv_id": "2305.12961",
"language": "en",
"page_idx": 3
}
|
||
2302.14434-en-0
|
2302.14434-en-0.jpg
|
## A Hierarchical Representation Network for Accurate and Detailed Face $\textbf{Reconstruction from In- The- Wild Images}$
Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, Xuansong Xie
DAMO Academy, Alibaba Group
$\{$biwen.lbw, jianqiang.rjq, mengyang.fmy, miaomiao.cmm}galibaba-inc.com,
xingtong.xxs@taobao.com
<FigureHere>
Figure 1. Example of low-frequency geometry, MF, HF facial details and the hierarchical representation.
## Abstract
Limited by the nature of the low-dimensional represen-
tational capacity of 3DMM, most of the 3DMM- based face $\geq$reconstruction ( FR) methods fail to recover high- frequency facial details, such as wrinkles, dimples, etc. Some attempt
to solve the problem by introducing detail maps or non- $linearoperations,however,theresultsarestillnotvivid.To$ this end, we in this paper present a novel hierarchical representation network (HRN) to achieve accurate and detailed face reconstruction from a single image. Specifically, we implement the geometry disentanglement and introduce the hierarchical representation to fulfill detailed face modeling. Meanwhile, 3D priors of facial details are incorporated to enhance the accuracy and authenticity of the reconstruction results. We also propose a de-retouching module to achieve better decoupling of the geometry and appearance. It is noteworthy that our framework can be extended to a multiview fashion by considering detail consistency of different views. Extensive experiments on two single-view and two multi-view FR benchmarks demonstrate that our method outperforms the existing methods in both reconstruction accuracy and visual effects. Finally, we introduce a highquality 3D face dataset FaceHD-100 to boost the research of high-fidelity face reconstruction. The project homepage isat https: / hyoung lbw. github. io/ HRN- homepage/.
treatment, film production, etc. While extensive works already achieved excellent reconstruction performance using specialized hardware like LightStage [2,11,35], estimating highly detailed face models from single or sparse-view images is still a challenging problem. Based on 3DMM[8], a statistical model learned from a collection of face scans, many works [16,22,23,32] attempt to reconstruct the 3D face from a single image and achieve impressive results. However, limited by the nature of the low dimensional representational ability of the 3DMM, these methods can not recover the detailed facial geometry.
Recently, some methods [13,24,38] devote to capturing high-frequency facial details such as wrinkles by predicting a displacement map. They achieve realistic results, however, fail to model the mid-frequency details, such as the detailed contour of the jaw, cheeck, etc. To this end, some works try to capture the overall details by introducing latent encoding of details [19] or non-linear operations [20,44]. Nevertheless, it is hard to make a trade-off when handling the mid- and high-frequency details simultaneously. Another challenge is how to obtain accurate shapes and detailed 3D facial priors considering multifarious lightings and skins for different images.[10,13] resort to the wrinkle statistics computed from 3D face scans to fulfill realistic high-frequency details, but still fail to model the midfrequency details.
Based on the observations above, we introduce a hierarchical representation network (HRN) for accurate and detailed face reconstruction from single image, as shown in
## $1. \textbf{Introduction}$
High-fidelity 3D face reconstruction finds a wide range of applications in many scenarios, such as AR/VR, medical
|
[
[
0.1282958984375,
0.1337890625,
0.84326171875,
0.1337890625,
0.84326171875,
0.1483154296875,
0.1282958984375,
0.1483154296875
],
[
0.275390625,
0.1546630859375,
0.69677734375,
0.1558837890625,
0.69677734375,
0.173583984375,
0.275390625,
0.17236328125
],
[
0.19287109375,
0.2052001953125,
0.77685546875,
0.2052001953125,
0.77685546875,
0.2197265625,
0.19287109375,
0.2197265625
],
[
0.351318359375,
0.2215576171875,
0.619140625,
0.2235107421875,
0.619140625,
0.238037109375,
0.351318359375,
0.236083984375
],
[
0.1748046875,
0.242431640625,
0.7958984375,
0.242431640625,
0.7958984375,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.021240234375,
0.2568359375,
0.058013916015625,
0.2568359375,
0.060455322265625,
0.410888671875,
0.0236968994140625,
0.410888671875
],
[
0.3798828125,
0.260009765625,
0.58837890625,
0.260009765625,
0.58837890625,
0.2734375,
0.3798828125,
0.2734375
],
[
0.328369140625,
0.37890625,
0.406982421875,
0.37890625,
0.406982421875,
0.397705078125,
0.328369140625,
0.397705078125
],
[
0.384033203125,
0.397705078125,
0.486083984375,
0.397705078125,
0.486083984375,
0.4091796875,
0.384033203125,
0.4091796875
],
[
0.55615234375,
0.397705078125,
0.61376953125,
0.397705078125,
0.61376953125,
0.4091796875,
0.55615234375,
0.4091796875
],
[
0.71142578125,
0.397705078125,
0.775390625,
0.397705078125,
0.775390625,
0.4091796875,
0.71142578125,
0.4091796875
],
[
0.36767578125,
0.40283203125,
0.374267578125,
0.40283203125,
0.374267578125,
0.406005859375,
0.36767578125,
0.406005859375
],
[
0.0236968994140625,
0.3984375,
0.058013916015625,
0.3984375,
0.058013916015625,
0.48486328125,
0.0236968994140625,
0.48486328125
],
[
0.408447265625,
0.4072265625,
0.455810546875,
0.406005859375,
0.455810546875,
0.4140625,
0.408447265625,
0.415283203125
],
[
0.57275390625,
0.40771484375,
0.5986328125,
0.40771484375,
0.5986328125,
0.415283203125,
0.57275390625,
0.415283203125
],
[
0.73046875,
0.40771484375,
0.7548828125,
0.40771484375,
0.7548828125,
0.415283203125,
0.73046875,
0.415283203125
],
[
0.178955078125,
0.434326171875,
0.79345703125,
0.434326171875,
0.79345703125,
0.4443359375,
0.178955078125,
0.4443359375
],
[
0.238525390625,
0.460205078125,
0.313720703125,
0.460205078125,
0.313720703125,
0.474853515625,
0.238525390625,
0.474853515625
],
[
0.50244140625,
0.4609375,
0.890625,
0.4609375,
0.890625,
0.474853515625,
0.50244140625,
0.474853515625
],
[
0.1004638671875,
0.479248046875,
0.468017578125,
0.479248046875,
0.468017578125,
0.492431640625,
0.1004638671875,
0.492431640625
],
[
0.5009765625,
0.47607421875,
0.890625,
0.4765625,
0.890625,
0.4912109375,
0.5009765625,
0.490478515625
],
[
0.0269622802734375,
0.49365234375,
0.056365966796875,
0.49365234375,
0.056365966796875,
0.5546875,
0.0269622802734375,
0.5546875
],
[
0.08087158203125,
0.494384765625,
0.471435546875,
0.494384765625,
0.471435546875,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.50341796875,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.50341796875,
0.505859375
],
[
0.08087158203125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.08087158203125,
0.52294921875
],
[
0.50341796875,
0.5087890625,
0.8896484375,
0.5087890625,
0.8896484375,
0.51953125,
0.50341796875,
0.51953125
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.5009765625,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.5361328125,
0.5009765625,
0.5361328125
],
[
0.080078125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.50341796875,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.548828125,
0.50341796875,
0.548828125
],
[
0.0318603515625,
0.55029296875,
0.049835205078125,
0.55029296875,
0.049835205078125,
0.564453125,
0.0318603515625,
0.564453125
],
[
0.080078125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.080078125,
0.568359375
],
[
0.50244140625,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.0310516357421875,
0.56298828125,
0.049835205078125,
0.56298828125,
0.049835205078125,
0.5771484375,
0.0310516357421875,
0.5771484375
],
[
0.080078125,
0.5693359375,
0.46728515625,
0.5693359375,
0.46728515625,
0.58349609375,
0.080078125,
0.58349609375
],
[
0.50341796875,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.080078125,
0.58544921875,
0.4697265625,
0.58544921875,
0.4697265625,
0.5986328125,
0.080078125,
0.5986328125
],
[
0.5009765625,
0.58056640625,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.5009765625,
0.5947265625
],
[
0.0294189453125,
0.59228515625,
0.052276611328125,
0.59228515625,
0.052276611328125,
0.60595703125,
0.0294189453125,
0.60595703125
],
[
0.07843017578125,
0.60107421875,
0.468017578125,
0.60107421875,
0.468017578125,
0.6142578125,
0.07843017578125,
0.6142578125
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.0318603515625,
0.611328125,
0.056365966796875,
0.61865234375,
0.0482177734375,
0.63525390625,
0.0236968994140625,
0.62744140625
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.5009765625,
0.6123046875,
0.74169921875,
0.6123046875,
0.74169921875,
0.62548828125,
0.5009765625,
0.62548828125
],
[
0.07843017578125,
0.62890625,
0.468017578125,
0.6298828125,
0.468017578125,
0.64453125,
0.07843017578125,
0.64306640625
],
[
0.521484375,
0.62744140625,
0.89208984375,
0.62890625,
0.89208984375,
0.64306640625,
0.521484375,
0.642578125
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64306640625,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.0294189453125,
0.6298828125,
0.052276611328125,
0.6298828125,
0.052276611328125,
0.642578125,
0.0294189453125,
0.642578125
],
[
0.03594970703125,
0.646484375,
0.051483154296875,
0.646484375,
0.051483154296875,
0.6640625,
0.03594970703125,
0.6640625
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.08087158203125,
0.662109375,
0.468017578125,
0.662109375,
0.468017578125,
0.67236328125,
0.08087158203125,
0.67236328125
],
[
0.50244140625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.080078125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.080078125,
0.689453125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.03594970703125,
0.6845703125,
0.052276611328125,
0.6845703125,
0.052276611328125,
0.70556640625,
0.03594970703125,
0.70556640625
],
[
0.080078125,
0.689453125,
0.468017578125,
0.689453125,
0.468017578125,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.471435546875,
0.72021484375,
0.471435546875,
0.734375,
0.080078125,
0.734375
],
[
0.50244140625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.5009765625,
0.73291015625,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5009765625,
0.7470703125
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.7666015625,
0.88818359375,
0.7666015625,
0.88818359375,
0.7763671875,
0.50341796875,
0.7763671875
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.08087158203125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.5048828125,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
0.806640625,
0.5048828125,
0.806640625
],
[
0.08087158203125,
0.8125,
0.4697265625,
0.8125,
0.4697265625,
0.82568359375,
0.08087158203125,
0.82568359375
],
[
0.50244140625,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.07843017578125,
0.82568359375,
0.40185546875,
0.8271484375,
0.40185546875,
0.84130859375,
0.07843017578125,
0.84033203125
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82470703125,
0.890625,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.50244140625,
0.84033203125,
0.62353515625,
0.84033203125,
0.62353515625,
0.85400390625,
0.50244140625,
0.85400390625
],
[
0.08087158203125,
0.84765625,
0.20751953125,
0.84765625,
0.20751953125,
0.861328125,
0.08087158203125,
0.861328125
],
[
0.521484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.521484375,
0.869140625
],
[
0.098876953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.098876953125,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"A Hierarchical Representation Network for Accurate and Detailed Face",
"Reconstruction from In-The-Wild Images",
"Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, Xuansong Xie",
"DAMO Academy, Alibaba Group",
"[biwen.lbw,jianqiang.rjq,mengyang.fmy,miaomiao.cmm]@alibaba-inc.com,",
"",
"xingtong.xxs@taobao.com",
"..",
"identity&expression",
"deformation",
"displacement",
"",
"[']",
"coefficient",
"map",
"map",
"Figure1.Example oflow-frequency geometry,MF,HFfacial details and thehierarchical representation.",
"Abstract",
"treatment, film production, etc. While extensive works al-",
"Limited by the nature of the low-dimensional represen-",
"ready achieved excellent reconstruction performance using",
"",
"tational capacityof 3DMM, most of the 3DMM-based face",
"specialized hardware like LightStage [2, 11,35], estimating",
"reconstruction (FR) methods fail to recover high-frequency",
"highlydetailedfacemodelsfromsingleor sparse-viewim-",
"facial details, such as wrinkles, dimples, etc. Some attempt",
"ages is still a challenging problem. Based on 3DMM [8],",
"to solve the problem by introducing detail maps or non-",
"a statistical model learned froma collection offace scans.",
"4",
"linear operations, however, the results are still not vivid. To",
"many works [16,22, 23, 321 attempt to reconstruct the 3D",
"4",
"this end, we in this paper present a novel hierarchical repre-",
"face from a single image and achieve impressive results.",
"sentation network (HRN) to achieve accurate and detailed",
"However,limited bythenatureofthelow dimensional rep-",
"2",
"face reconstruction from a single image.Specifically, we",
"resentational ability of the 3DMM, these methods can not",
"3",
"implement the geometry disentanglement and introduce the",
"recover the detailed facial geometry.",
"hierarchical representation to fulfill detailed face modeling.",
"Recently, some methods [13, 24, 38] devote to capturing",
"high-frequency facial details such as wrinkles by predicting",
"2",
"V",
"Meanwhile, 3D priors of facial details are incorporated to",
"enhancetheaccuracyandauthenticityofthereconstruction",
"a displacement map. They achieve realistic results, how-",
"results. We also propose a de-retouching module to achieve",
"ever, fail to model the mid-frequency details, such as the",
"",
"better decoupling of the geometry and appearance. It is",
"detailed contour of the jaw, cheeck, etc. To this end, some",
"noteworthy that our framework can be extended to a multi-",
"works try to capture the overall details by introducing latent",
"view fashion by considering detail consistency of different",
"encoding of details [19] or non-linear operations [20, 44].",
"views.Extensive experiments on two single-view and two",
"Nevertheless, it is hard to make a trade-off when handling",
"multi-view FR benchmarks demonstrate that our method",
"the mid- and high-frequency details simultaneously. An-",
"outperforms the existing methods in both reconstruction ac-",
"other challenge is how to obtain accurate shapes and de-",
"curacy and visual effects. Finally, we introduce a high-",
"tailed 3D facial priors considering multifarious lightings",
"quality 3D face dataset FaceHD-100 to boost the research",
"and skins for differentimages.lo,l3lresorttothe wrin-",
"of high-fidelity face reconstruction. The project homepage",
"kle statistics computed from 3D face scans to fulfill realis-",
"is at https:/younglbw.github.io/HRN-homepagel.",
"tic high-frequency details, but still fail to model the mid-",
"frequencydetails.",
"1. Introduction",
"Based on the observations above, we introduce a hierar-",
"High-fidelity 3D face reconstruction finds a wide range",
"chical representation network (HRN) for accurate and de-",
"of applications in many scenarios, such as AR/VR, medical",
"tailed face reconstruction from single image, as shown in"
] |
{
"arxiv_id": "2302.14434",
"language": "en",
"page_idx": 0
}
|
||
2302.14434-en-1
|
2302.14434-en-1.jpg
|
Fig. 2. Firstly, we decouple the facial geometry into lowfrequency geometry, mid-frequency (MF) details, and highfrequency (HF) details. Then, in a hierarchical fashion, we model these parts with face-wise blendshape coefficients, vertex-wise deformation map, and pixel-wise displacement map, respectively (shown in Fig. 1). Concretely, we employ two image translation networks [27] to estimate the corresponding detail maps (deformation and displacement map), and further employ them to generate the detailed face model in a coarse-to-fine manner. Moreover, we introduce the 3D priors of MF and HF details by fitting face scans with our hierarchical representation to facilitate accurate and faithful modeling. Inspired by [33], we propose a de-retouching module to adaptively refine the base texture to overcome the ambiguities between skin blemishes and illuminations. Extensive experiments show that our method outperforms the existing methods on two large-scale benchmarks, exhibiting excellent performance in terms of detail capturing and accurate shape modeling. Thanks to the detail disentanglement strategy and the guidance of detail priors, we extend HRN to a multi-view fashion and achieve accurate FR from only a few views. Finally, to boost the research of sparse-view and high-fidelity FR, we introduce a high-quality 3D face dataset named FaceHD-100.
Our main contributions in this work are as follows:
(A) We present a hierarchical modeling strategy and propose a novel framework HRN for single-view FR task. Our HRN produces accurate and highly detailed FR results and outperforms the existing state-of-the-art methods on two large-scale single-view FR benchmarks.
(B) We introduce detail priors to guide the faithful modeling of hierarchical details and design a de-retouching module to facilitate the decoupling of geometry and appearance.
(C) We extend HRN to a multi-view fashion to form MVHRN, which enables accurate face modeling from sparseview images and outperforms the existing methods on two large-scale multi-view FR benchmarks.
$(\mathbf{D})$ To boost the research on sparse-view and high-fidelity FR tasks, we introduce a high-quality 3D face dataset FaceHD-100, containing 2,000 detailed 3D face models and corresponding high-definition multi-view images.
$2. \textbf{Related Work}$
Single-View Face Reconstruction. Recovering 3D face from a single image is an ill-posed problem, but the advent of the 3D morphable model (3DMM) [8,39] has made it possible. The 3DMM provides strong prior knowledge and can represent complicated face geometry with lowdimensional coefficients. In this formulation, current methods can be categorized into either optimization-based [7,8, 25,31,51] or learning-based [15,40,43,52]. Optimizationbased approaches usually need costly analysis-by-synthesis processes and are sensitive to initialization, while learning-
based methods directly train neural networks to regress the low-dimensional coefficients of 3DMM and recover 3D face through efficient forward inference.
However, the original 3DMM models inherently lie in low-dimensional linear space and lack fine details. Many works [14,19,42,45,48] are proposed to overcome this limitation. Tran $etal.[45]$ present a nonlinear 3DMM model and achieve more powerful representational abilities. Sela et al.[42] employ image-to-image network to generate pixel-based geometric representation for high quality reconstructions. In addition to static face geometry details, Feng $etal.[19]$ present an animatable displacement model to generate dynamic expression-depended wrinkles. Yang et al.[48] predict displacement maps via pix2pixHD network and combine them according to blendshape weights for dynamic details synthesis. Compared with these approaches, our method further introduces facial detail priors and can recover high fidelity facial details with hierarchical geometry representations.
Multi-View Face Reconstruction. Traditional multi-view stereo (MVS) methods [6,9,21] are designed for 3D reconstruction given a set of multi-view images, but they heavily rely on the precision of camera calibration, and can hardly recover intact geometry in the sparse-view situation. To address these problems, many face-specialized multi-view reconstruction methods [5,17,26,36,37,46,47] are proposed. Ramon et $al.[37]$ introduce siamese neural networks to extract relevant features from each view, and learn the 3D shape and the individual camera poses simultaneously. Wu et al.[46] exploit both 3DMM and multi-view geometric constraints by estimating the alignment loss between multi-view inputs. Bai $etal.[5]$ leverage non-rigid multiview stereo optimization to explicitly enforce multi-view appearance consistency, which is able to capture mediumlevel details. With the emergence of implicit 3D representation, Xiao et $al.[47]$ propose to learn an implicit function to recover detailed geometry from calibrated multi-view images, but the implicit function learning is time-consuming which needs dozen of seconds and is sensitive to camera count and pose estimation.
Rather than a specific design for multi-view inputs, our single-view model can be easily transferred to sparse-view face reconstruction task by adding hierarchical detail consistency between different views. Our method is robust to the calibration error of cameras thanks to the coarse-to-fine learning scheme.
## 3. Methods
## 3.1.0verview
In this paper, we propose a novel hierarchical representation network for accurate and detailed face reconstruction from single image. Fig. 2 illustrates the overview of our
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50244140625,
0.106689453125,
0.89208984375,
0.106689453125,
0.89208984375,
0.12060546875,
0.50244140625,
0.12060546875
],
[
0.07843017578125,
0.121826171875,
0.4697265625,
0.12249755859375,
0.4697265625,
0.136962890625,
0.07843017578125,
0.1363525390625
],
[
0.50341796875,
0.125,
0.7685546875,
0.125,
0.7685546875,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.07843017578125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.51953125,
0.1383056640625,
0.8935546875,
0.136962890625,
0.8935546875,
0.1514892578125,
0.51953125,
0.15283203125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.50244140625,
0.154052734375,
0.89208984375,
0.154052734375,
0.89208984375,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.50341796875,
0.17041015625,
0.890625,
0.17041015625,
0.890625,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.080078125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.5009765625,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.21337890625,
0.5009765625,
0.21337890625
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.5009765625,
0.21337890625,
0.89208984375,
0.2139892578125,
0.89208984375,
0.2291259765625,
0.5009765625,
0.2279052734375
],
[
0.08087158203125,
0.2298583984375,
0.46728515625,
0.2298583984375,
0.46728515625,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50244140625,
0.2298583984375,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.50244140625,
0.24365234375
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.2442626953125,
0.890625,
0.24365234375,
0.890625,
0.25830078125,
0.5009765625,
0.2587890625
],
[
0.08087158203125,
0.260009765625,
0.46728515625,
0.260009765625,
0.46728515625,
0.270751953125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.2734375,
0.50244140625,
0.2734375
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.07843017578125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.07843017578125,
0.302978515625
],
[
0.50244140625,
0.290283203125,
0.89208984375,
0.290283203125,
0.89208984375,
0.3037109375,
0.50244140625,
0.3037109375
],
[
0.07843017578125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.07843017578125,
0.317626953125
],
[
0.50244140625,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.08087158203125,
0.32080078125,
0.468017578125,
0.32080078125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50341796875,
0.3212890625,
0.890625,
0.3212890625,
0.890625,
0.33203125,
0.50341796875,
0.33203125
],
[
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
],
[
0.08087158203125,
0.35107421875,
0.46728515625,
0.35107421875,
0.46728515625,
0.361083984375,
0.08087158203125,
0.361083984375
],
[
0.50341796875,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.080078125,
0.3642578125,
0.4697265625,
0.3642578125,
0.4697265625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.50244140625,
0.365478515625,
0.7548828125,
0.365478515625,
0.7548828125,
0.37890625,
0.50244140625,
0.37890625
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50244140625,
0.380615234375,
0.89208984375,
0.380615234375,
0.89208984375,
0.39453125,
0.50244140625,
0.39453125
],
[
0.080078125,
0.39453125,
0.46728515625,
0.39453125,
0.46728515625,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50341796875,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.40771484375,
0.50341796875,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.50341796875,
0.41357421875,
0.8896484375,
0.41357421875,
0.8896484375,
0.423583984375,
0.50341796875,
0.423583984375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.080078125,
0.438232421875
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.080078125,
0.439453125,
0.26953125,
0.439453125,
0.26953125,
0.45263671875,
0.080078125,
0.45263671875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.09722900390625,
0.455078125,
0.4345703125,
0.455078125,
0.4345703125,
0.468505859375,
0.09722900390625,
0.468505859375
],
[
0.50244140625,
0.45703125,
0.890625,
0.45703125,
0.890625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.08087158203125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.5,
0.4716796875,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.5,
0.48486328125
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.5009765625,
0.4873046875,
0.89208984375,
0.486083984375,
0.89208984375,
0.5,
0.5009765625,
0.50048828125
],
[
0.080078125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.080078125,
0.5146484375
],
[
0.50341796875,
0.50390625,
0.890625,
0.50390625,
0.890625,
0.5146484375,
0.50341796875,
0.5146484375
],
[
0.08087158203125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.52783203125,
0.08087158203125,
0.52783203125
],
[
0.50244140625,
0.517578125,
0.88818359375,
0.517578125,
0.88818359375,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.080078125,
0.52978515625,
0.3447265625,
0.52978515625,
0.3447265625,
0.54345703125,
0.080078125,
0.54345703125
],
[
0.5009765625,
0.53076171875,
0.8896484375,
0.53076171875,
0.8896484375,
0.54541015625,
0.5009765625,
0.54541015625
],
[
0.08087158203125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.08087158203125,
0.5595703125
],
[
0.50244140625,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.56005859375,
0.4697265625,
0.56005859375,
0.4697265625,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5625,
0.89208984375,
0.5625,
0.89208984375,
0.57568359375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.57568359375,
0.4306640625,
0.57568359375,
0.4306640625,
0.5888671875,
0.080078125,
0.5888671875
],
[
0.50341796875,
0.5791015625,
0.8896484375,
0.5791015625,
0.8896484375,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.08087158203125,
0.59033203125,
0.468017578125,
0.59033203125,
0.468017578125,
0.60400390625,
0.08087158203125,
0.60400390625
],
[
0.50244140625,
0.59326171875,
0.890625,
0.59326171875,
0.890625,
0.60693359375,
0.50244140625,
0.60693359375
],
[
0.080078125,
0.60595703125,
0.4697265625,
0.60595703125,
0.4697265625,
0.619140625,
0.080078125,
0.619140625
],
[
0.5009765625,
0.60595703125,
0.89208984375,
0.60693359375,
0.89208984375,
0.6220703125,
0.5009765625,
0.62060546875
],
[
0.080078125,
0.6220703125,
0.4697265625,
0.6220703125,
0.4697265625,
0.63525390625,
0.080078125,
0.63525390625
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.080078125,
0.63623046875,
0.340576171875,
0.63623046875,
0.340576171875,
0.6494140625,
0.080078125,
0.6494140625
],
[
0.50244140625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.08087158203125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.080078125,
0.66650390625,
0.4697265625,
0.66650390625,
0.4697265625,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.5009765625,
0.66748046875,
0.89208984375,
0.66748046875,
0.89208984375,
0.68115234375,
0.5009765625,
0.68115234375
],
[
0.08087158203125,
0.68310546875,
0.468017578125,
0.68310546875,
0.468017578125,
0.693359375,
0.08087158203125,
0.693359375
],
[
0.50341796875,
0.68505859375,
0.6787109375,
0.68505859375,
0.6787109375,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.07843017578125,
0.6962890625,
0.406005859375,
0.6962890625,
0.406005859375,
0.71044921875,
0.07843017578125,
0.71044921875
],
[
0.521484375,
0.69775390625,
0.89208984375,
0.69775390625,
0.89208984375,
0.7119140625,
0.521484375,
0.7119140625
],
[
0.50341796875,
0.7138671875,
0.890625,
0.7138671875,
0.890625,
0.72802734375,
0.50341796875,
0.72802734375
],
[
0.080078125,
0.72607421875,
0.2181396484375,
0.72607421875,
0.2181396484375,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.50244140625,
0.72998046875,
0.89208984375,
0.72998046875,
0.89208984375,
0.74365234375,
0.50244140625,
0.74365234375
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.7470703125,
0.890625,
0.7470703125,
0.890625,
0.7568359375,
0.50341796875,
0.7568359375
],
[
0.5009765625,
0.7587890625,
0.890625,
0.7587890625,
0.890625,
0.77294921875,
0.5009765625,
0.77294921875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.07843017578125,
0.77978515625,
0.4697265625,
0.78076171875,
0.4697265625,
0.79541015625,
0.07843017578125,
0.79443359375
],
[
0.50244140625,
0.77587890625,
0.61767578125,
0.77587890625,
0.61767578125,
0.7890625,
0.50244140625,
0.7890625
],
[
0.07843017578125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81103515625,
0.07843017578125,
0.81103515625
],
[
0.07843017578125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.80615234375,
0.5986328125,
0.80615234375,
0.5986328125,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.83203125,
0.611328125,
0.83203125,
0.611328125,
0.8466796875,
0.50244140625,
0.8466796875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.521484375,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.521484375,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
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,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"Fig. 2. Firstly, we decouple the facial geometry into low-",
"based methods directly train neural networks to regress",
"frequency geometry, mid-frequency (MF) details, and high-",
"the low-dimensional coefficients of 3DMM and recover 3D",
"frequency (HF) details. Then, in a hierarchical fashion, we",
"facethroughefficientforward inference.",
"model these parts with face-wise blendshape coefficients,",
"However, the original 3DMM models inherently lie in",
"vertex-wise deformation map,and pixel-wise displacement",
"low-dimensional linear space and lack fine details. Many",
"map, respectively (shown in Fig.1). Concretely, we employ",
"works [14, 19,42,45, 48 are proposed to overcome this",
"two image translation networks [27] to estimate the corre-",
"limitation.Tran et al.[45] present a nonlinear 3DMM",
"sponding detail maps (deformation and displacement map),",
"model and achieve more powerful representational abil-",
"and further employ them to generate the detailed face model",
"ities.Sela et al.[42] employ image-to-image network",
"inacoarse-to-finemanner.Moreover,weintroducethe3D",
"to generate pixel-based geometric representation for high",
"priors of MF and HF details by fitting face scans with our",
"quality reconstructions.In addition to static face geom-",
"hierarchical representation tofacilitate accurate and faith-",
"etry details, Feng et al. [19] present an animatable dis-",
"ful modeling. Inspired by [33], we propose a de-retouching",
"placement model to generate dynamic expression-depended",
"module to adaptively refine the base texture to overcome the",
"wrinkles. Yang et al. [48] predict displacement maps via",
"ambiguities between skin blemishes and illuminations. Ex-",
"pix2pixHD network and combine them according to blend-",
"tensiveexperimentsshowthat ourmethod outperforms the",
"shape weightsfor dynamic details synthesis.Compared",
"existing methods on two large-scale benchmarks, exhibiting",
"with these approaches, our method further introduces facial",
"excellentperformanceintermsofdetailcapturingandaccu-",
"detail priors and canrecoverhighfidelityfacial details with",
"rate shape modeling. Thanks to the detail disentanglement",
"hierarchical geometry representations.",
"strategy and the guidance of detail priors, we extend HRN",
"Multi-View Face Reconstruction. Traditional multi-view",
"to a multi-view fashion and achieve accurate FR from only",
"stereo(MVS)methods[6.9.21laredesignedfor3Drecon-",
"a few views. Finally, to boost the research of sparse-view",
"struction given a set of multi-viewimages,but theyheavily",
"and high-fidelity FR, we introduce a high-quality 3D face",
"rely on the precision of camera calibration, and can hardly",
"dataset named FaceHD-100.",
"recover intact geometry in the sparse-view situation. To",
"Our main contributions in this work are as follows:",
"address these problems, many face-specialized multi-view",
"(A) We present a hierarchical modeling strategy and pro-",
"reconstruction methods [5, 17,26, 36, 37, 46, 47] are pro-",
"pose a novel framework HRN for single-view FR task. Our",
"posed.Ramonetal.[37]introducesiameseneuralnetworks",
"HRN produces accurate and highly detailed FR results and",
"to extract relevant features from each view,and learn the",
"outperforms theexistingstate-of-the-art methodsontwo",
"3D shape and the individual camera poses simultaneously",
"large-scale single-view FR benchmarks.",
"Wu et al. [46] exploit both 3DMM and multi-view geo-",
"(B) We introduce detail priors to guide the faithful modeling",
"metric constraints by estimating the alignment loss between",
"of hierarchical details and design a de-retouching module to",
"multi-view inputs. Bai et al. [5] leverage non-rigid multi-",
"facilitate the decoupling of geometry and appearance.",
"view stereo optimization to explicitly enforce multi-view",
"(C) We extend HRN to a multi-view fashion to form MV-",
"appearance consistency, which is able to capture medium-",
"HRN, which enables accurate face modeling from sparse-",
"level details.With the emergence of implicit 3D representa-",
"view images and outperforms the existing methods on two",
"tion, Xiao et al. [47] propose to learn an implicit function to",
"large-scale multi-view FR benchmarks.",
"recover detailed geometry from calibrated multi-view im-",
"(D) To boost the research on sparse-view and high-fidelity",
"ages, but the implicit function learning is time-consuming",
"FR tasks, we introduce a high-quality 3D face dataset",
"which needs dozen of seconds and is sensitive to camera",
"FaceHD-100.containing2.000detailed3Dfacemodelsand",
"count and pose estimation.",
"corresponding high-definition multi-view images.",
"Rather than a specific design for multi-view inputs, our",
"single-view model can be easily transferred to sparse-view",
"2. Related Work",
"face reconstruction task by adding hierarchical detail con-",
"Single-View Face Reconstruction. Recovering 3D face",
"sistency between different views.Ourmethod is robustto",
"the calibration error of cameras thanks to the coarse-to-fine",
"from a single image is an ill-posed problem, but the ad-",
"vent of the 3D morphable model (3DMM) [8, 39] has made",
"learning scheme.",
"it possible. The 3DMM provides strong prior knowledge",
"and can represent complicated face geometry with low-",
"3.Methods",
"dimensional coefficients. In this formulation, current meth-",
"3.1.Overview",
"ods can be categorized into either optimization-based [7, 8,",
"25,31,51] or learning-based [15, 40, 43,52]. Optimization-",
"In this paper, we propose a novel hierarchical represen-",
"based approaches usually need costly analysis-by-synthesis",
"tation network for accurate and detailed face reconstruction",
"processes and are sensitive to initialization, while learning-",
"from single image. Fig. 2 illustrates the overview of our"
] |
{
"arxiv_id": "2302.14434",
"language": "en",
"page_idx": 1
}
|
||
2302.14434-en-2
|
2302.14434-en-2.jpg
|
<FigureHere>
Figure 2. Overview of the proposed hierarchical representation network (HRN).
framework. We first employ 3DMM to predict a coarse mesh and albedo (blue area in Fig. 2). Then we develop a hierarchical modeling (Sec. 3.2) strategy to handle the complex facial details in a coarse-to-fine manner (green and purple area). To facilitate the accurate and faithful modeling of the hierarchical details, the 3D priors are incorporated through adversarial and semi-supervised learning (Sec. 3.3). Besides, we propose a de-retouching module (Sec. 3.4) to fulfill better decoupling of the geometry and appearance, alleviating the ambiguities between the various skin texture and illuminations. Moreover, we extend our framework to a multi-view fashion (Sec. 3.5) and introduce a high quality 3D face dataset (Sec. 3.6) to boost the research on sparseview face reconstruction. For simplification, we specify the related loss functions and training strategy in each section.
# 3.2. Hierarchical Modeling
frequency detail, while the fineness increases in turn.
We design the hierarchical representation to model the above three components respectively. For the lowfrequency part, we adopt the BFM as our base model and output the low-dimensional coefficients to fulfill a coarse reconstruction of the input face. Then we introduce a threechannel deformation map, which lies in the UV space and indicates the offset of each vertex relative to the coarse result. Worked as the representation of the mid-frequency details, the deformation map provides a flexible way to manipulate the geometries. We use the size of 64×64 to represent the deformation map to balance the fineness and smoothness of the mid-frequency details. For the high-frequency details, we employ the displacement map following [19], which is a one-channel map $(256\times256)$ denoting the geometry deformation along the direction of the normals. The displacement map is converted to detailed normals used in the rendering process in a pixel-wise manner to exhibit all the tiny details, breaking the limitation of the vertex density of the base model. Accordingly, we are enabled to describe an arbitrary complex face with these representations.
As shown in Fig. 2, given a portrait image $I$, we firstly utilize a regression network as the face analyzer to predict the BFM coefficients and obtain the coarse aligned mesh $M_0$ and albedo $A_0$ using the corresponding basis from the 3DMM database. Combined $I$ and $M_0$, we are able to acquire the inpainted texture $T$ in UV space by applying the $\hat{\text{differentiable rendering with a coarse-to-fine strategy. And}}$ we concatenate the position map $P$ and $T$ as the input of the following modules for hierarchical details learning. We adopt two pix2pix [27] networks to synthesize the deformation map and displacement map in sequence. Note that, considering the deformation map will change the facial geometry and lead to the misalignment between $T$ and the
3DMM exhibit great performance in expressing various shape of the face, while the low-dimensional representation itself severely stem its learning on the details, leading to the imperfect alignment to the face or the over-smoothed results. Some methods extend 3DMM by introducing displacement maps to reconstruct some details such as wrinkles and bumps. However, focused on the high-frequency part, a simple displacement map still fails to handle some larger-scale details, such as the contours of the jaw and cheek. Based on such observation, we decouple the facial geometry into three components:(1) low-frequency part, which provides the coarse shape roughly aligned to the input face; (2) mid-frequency details, which describe the details of the contour and local shape relative to the low-frequency part; (3) and high-frequency details, such as wrinkles, micro bumps, etc. As shown in the Fig. 1, the scales decrease from the low-frequency part to the high-
|
[
[
0.330078125,
0.0985107421875,
0.358642578125,
0.0985107421875,
0.358642578125,
0.106689453125,
0.330078125,
0.106689453125
],
[
0.4794921875,
0.0985107421875,
0.525390625,
0.0985107421875,
0.525390625,
0.10919189453125,
0.4794921875,
0.10919189453125
],
[
0.56201171875,
0.09783935546875,
0.61376953125,
0.09783935546875,
0.61376953125,
0.10919189453125,
0.56201171875,
0.10919189453125
],
[
0.416748046875,
0.10479736328125,
0.431396484375,
0.10479736328125,
0.431396484375,
0.125,
0.416748046875,
0.125
],
[
0.316162109375,
0.106689453125,
0.371826171875,
0.10919189453125,
0.371826171875,
0.12060546875,
0.316162109375,
0.11871337890625
],
[
0.1446533203125,
0.10101318359375,
0.2091064453125,
0.10101318359375,
0.2091064453125,
0.11236572265625,
0.1446533203125,
0.11236572265625
],
[
0.473876953125,
0.10797119140625,
0.52783203125,
0.10797119140625,
0.52783203125,
0.11932373046875,
0.473876953125,
0.11932373046875
],
[
0.560546875,
0.10986328125,
0.61767578125,
0.10986328125,
0.61767578125,
0.12060546875,
0.560546875,
0.12060546875
],
[
0.6787109375,
0.121826171875,
0.69677734375,
0.121826171875,
0.69677734375,
0.1300048828125,
0.6787109375,
0.1300048828125
],
[
0.16015625,
0.11236572265625,
0.191162109375,
0.11236572265625,
0.191162109375,
0.11932373046875,
0.16015625,
0.11932373046875
],
[
0.160888671875,
0.12371826171875,
0.1846923828125,
0.12371826171875,
0.1846923828125,
0.1275634765625,
0.160888671875,
0.1275634765625
],
[
0.390625,
0.1300048828125,
0.459228515625,
0.1300048828125,
0.459228515625,
0.14013671875,
0.390625,
0.14013671875
],
[
0.491943359375,
0.1300048828125,
0.51318359375,
0.1300048828125,
0.51318359375,
0.1383056640625,
0.491943359375,
0.1383056640625
],
[
0.59228515625,
0.1287841796875,
0.60693359375,
0.1256103515625,
0.611328125,
0.1363525390625,
0.59619140625,
0.1395263671875
],
[
0.6806640625,
0.1307373046875,
0.708984375,
0.1307373046875,
0.708984375,
0.1427001953125,
0.6806640625,
0.1427001953125
],
[
0.343017578125,
0.136962890625,
0.35693359375,
0.147705078125,
0.34716796875,
0.1546630859375,
0.334228515625,
0.14453125
],
[
0.404296875,
0.1395263671875,
0.44287109375,
0.1395263671875,
0.44287109375,
0.149658203125,
0.404296875,
0.149658203125
],
[
0.5009765625,
0.1337890625,
0.5146484375,
0.14453125,
0.50244140625,
0.154052734375,
0.488525390625,
0.143310546875
],
[
0.1519775390625,
0.1514892578125,
0.2254638671875,
0.1514892578125,
0.2254638671875,
0.1622314453125,
0.1519775390625,
0.1622314453125
],
[
0.580078125,
0.149658203125,
0.6005859375,
0.154052734375,
0.59619140625,
0.1666259765625,
0.576171875,
0.1622314453125
],
[
0.6611328125,
0.1514892578125,
0.6650390625,
0.1514892578125,
0.6650390625,
0.1578369140625,
0.6611328125,
0.1578369140625
],
[
0.705078125,
0.15283203125,
0.765625,
0.15283203125,
0.765625,
0.1641845703125,
0.705078125,
0.1641845703125
],
[
0.5009765625,
0.1578369140625,
0.5087890625,
0.1641845703125,
0.5009765625,
0.17041015625,
0.49267578125,
0.1641845703125
],
[
0.336669921875,
0.1572265625,
0.358642578125,
0.1610107421875,
0.3544921875,
0.173583984375,
0.33251953125,
0.169189453125
],
[
0.62158203125,
0.1590576171875,
0.6689453125,
0.1622314453125,
0.66845703125,
0.1749267578125,
0.62109375,
0.1717529296875
],
[
0.80126953125,
0.1622314453125,
0.8505859375,
0.1622314453125,
0.8505859375,
0.173583984375,
0.80126953125,
0.173583984375
],
[
0.58251953125,
0.1654052734375,
0.5986328125,
0.1654052734375,
0.5986328125,
0.173583984375,
0.58251953125,
0.173583984375
],
[
0.2646484375,
0.160400390625,
0.26953125,
0.160400390625,
0.271240234375,
0.169189453125,
0.265625,
0.169189453125
],
[
0.250732421875,
0.1717529296875,
0.289306640625,
0.1717529296875,
0.289306640625,
0.1849365234375,
0.250732421875,
0.1849365234375
],
[
0.86767578125,
0.1673583984375,
0.8798828125,
0.1673583984375,
0.8798828125,
0.173583984375,
0.86767578125,
0.173583984375
],
[
0.326904296875,
0.1787109375,
0.366943359375,
0.1787109375,
0.366943359375,
0.21533203125,
0.326904296875,
0.21533203125
],
[
0.64208984375,
0.1787109375,
0.65673828125,
0.1787109375,
0.65673828125,
0.199462890625,
0.64208984375,
0.199462890625
],
[
0.2646484375,
0.1849365234375,
0.27294921875,
0.1849365234375,
0.27294921875,
0.1944580078125,
0.2646484375,
0.1944580078125
],
[
0.82421875,
0.179931640625,
0.8388671875,
0.179931640625,
0.8388671875,
0.199462890625,
0.82421875,
0.199462890625
],
[
0.1707763671875,
0.1875,
0.1953125,
0.1875,
0.1953125,
0.1956787109375,
0.1707763671875,
0.1956787109375
],
[
0.1446533203125,
0.1944580078125,
0.148681640625,
0.1944580078125,
0.148681640625,
0.199462890625,
0.1446533203125,
0.199462890625
],
[
0.16259765625,
0.1944580078125,
0.2034912109375,
0.1962890625,
0.2017822265625,
0.2083740234375,
0.160888671875,
0.205810546875
],
[
0.44677734375,
0.2008056640625,
0.496826171875,
0.2008056640625,
0.496826171875,
0.212158203125,
0.44677734375,
0.212158203125
],
[
0.6357421875,
0.2008056640625,
0.67724609375,
0.2008056640625,
0.67724609375,
0.2392578125,
0.6357421875,
0.2392578125
],
[
0.81201171875,
0.20263671875,
0.85107421875,
0.1988525390625,
0.85888671875,
0.2374267578125,
0.81884765625,
0.2412109375
],
[
0.734375,
0.203857421875,
0.78515625,
0.2052001953125,
0.78515625,
0.21337890625,
0.734375,
0.212158203125
],
[
0.271240234375,
0.20263671875,
0.279296875,
0.20263671875,
0.279296875,
0.2083740234375,
0.271240234375,
0.2083740234375
],
[
0.52392578125,
0.2052001953125,
0.576171875,
0.2052001953125,
0.576171875,
0.212158203125,
0.52392578125,
0.212158203125
],
[
0.0743408203125,
0.21533203125,
0.136474609375,
0.21533203125,
0.136474609375,
0.2259521484375,
0.0743408203125,
0.2259521484375
],
[
0.191162109375,
0.216552734375,
0.273681640625,
0.216552734375,
0.273681640625,
0.2279052734375,
0.191162109375,
0.2279052734375
],
[
0.3349609375,
0.212158203125,
0.358642578125,
0.216552734375,
0.35546875,
0.2298583984375,
0.330810546875,
0.2266845703125
],
[
0.44775390625,
0.2247314453125,
0.459228515625,
0.2247314453125,
0.459228515625,
0.2291259765625,
0.44775390625,
0.2291259765625
],
[
0.4599609375,
0.2247314453125,
0.498291015625,
0.2247314453125,
0.498291015625,
0.2291259765625,
0.4599609375,
0.2291259765625
],
[
0.51123046875,
0.2266845703125,
0.52685546875,
0.2266845703125,
0.52685546875,
0.2342529296875,
0.51123046875,
0.2342529296875
],
[
0.6787109375,
0.2259521484375,
0.70751953125,
0.2259521484375,
0.70751953125,
0.2374267578125,
0.6787109375,
0.2374267578125
],
[
0.7158203125,
0.2266845703125,
0.73046875,
0.2266845703125,
0.73046875,
0.23486328125,
0.7158203125,
0.23486328125
],
[
0.6435546875,
0.240478515625,
0.6650390625,
0.240478515625,
0.6650390625,
0.25390625,
0.6435546875,
0.25390625
],
[
0.2607421875,
0.2298583984375,
0.283447265625,
0.2298583984375,
0.283447265625,
0.248779296875,
0.2607421875,
0.248779296875
],
[
0.826171875,
0.240478515625,
0.8486328125,
0.240478515625,
0.8486328125,
0.25244140625,
0.826171875,
0.25244140625
],
[
0.148681640625,
0.25390625,
0.293212890625,
0.255126953125,
0.293212890625,
0.268310546875,
0.148681640625,
0.267578125
],
[
0.51123046875,
0.25830078125,
0.58837890625,
0.25830078125,
0.58837890625,
0.268310546875,
0.51123046875,
0.268310546875
],
[
0.7197265625,
0.2587890625,
0.7998046875,
0.2587890625,
0.7998046875,
0.26953125,
0.7197265625,
0.26953125
],
[
0.371826171875,
0.274658203125,
0.38232421875,
0.274658203125,
0.38232421875,
0.2841796875,
0.371826171875,
0.2841796875
],
[
0.7060546875,
0.274658203125,
0.71142578125,
0.274658203125,
0.71142578125,
0.27978515625,
0.7060546875,
0.27978515625
],
[
0.5810546875,
0.2822265625,
0.61669921875,
0.2822265625,
0.61669921875,
0.29345703125,
0.5810546875,
0.29345703125
],
[
0.6787109375,
0.2822265625,
0.72119140625,
0.2841796875,
0.7197265625,
0.29541015625,
0.67724609375,
0.29345703125
],
[
0.79345703125,
0.276611328125,
0.830078125,
0.281005859375,
0.82666015625,
0.29345703125,
0.7900390625,
0.2890625
],
[
0.363525390625,
0.29541015625,
0.38818359375,
0.292236328125,
0.39208984375,
0.317626953125,
0.36767578125,
0.32080078125
],
[
0.583984375,
0.296630859375,
0.64453125,
0.296630859375,
0.64453125,
0.30810546875,
0.583984375,
0.30810546875
],
[
0.7958984375,
0.296630859375,
0.85546875,
0.296630859375,
0.85546875,
0.30810546875,
0.7958984375,
0.30810546875
],
[
0.268798828125,
0.298583984375,
0.289306640625,
0.298583984375,
0.289306640625,
0.3125,
0.268798828125,
0.3125
],
[
0.233642578125,
0.298583984375,
0.2442626953125,
0.298583984375,
0.2442626953125,
0.302978515625,
0.233642578125,
0.302978515625
],
[
0.233642578125,
0.304931640625,
0.24267578125,
0.304931640625,
0.24267578125,
0.3125,
0.233642578125,
0.3125
],
[
0.5849609375,
0.306884765625,
0.6396484375,
0.311279296875,
0.63818359375,
0.322509765625,
0.583984375,
0.318115234375
],
[
0.7939453125,
0.30615234375,
0.85546875,
0.30810546875,
0.85546875,
0.3193359375,
0.7939453125,
0.317626953125
],
[
0.54736328125,
0.313232421875,
0.5537109375,
0.313232421875,
0.5537109375,
0.31640625,
0.54736328125,
0.31640625
],
[
0.2998046875,
0.3193359375,
0.339111328125,
0.3212890625,
0.338134765625,
0.333251953125,
0.298095703125,
0.330810546875
],
[
0.361083984375,
0.32080078125,
0.396240234375,
0.322509765625,
0.39453125,
0.335205078125,
0.359375,
0.333251953125
],
[
0.7060546875,
0.317626953125,
0.71337890625,
0.317626953125,
0.71337890625,
0.32080078125,
0.7060546875,
0.32080078125
],
[
0.1422119140625,
0.324462890625,
0.2034912109375,
0.324462890625,
0.2034912109375,
0.335205078125,
0.1422119140625,
0.335205078125
],
[
0.2303466796875,
0.324462890625,
0.2484130859375,
0.324462890625,
0.2484130859375,
0.333984375,
0.2303466796875,
0.333984375
],
[
0.53369140625,
0.32568359375,
0.5703125,
0.32568359375,
0.5703125,
0.337158203125,
0.53369140625,
0.337158203125
],
[
0.6611328125,
0.32568359375,
0.7255859375,
0.32568359375,
0.7255859375,
0.3359375,
0.6611328125,
0.3359375
],
[
0.74267578125,
0.32568359375,
0.779296875,
0.32568359375,
0.779296875,
0.3359375,
0.74267578125,
0.3359375
],
[
0.471435546875,
0.32763671875,
0.515625,
0.32763671875,
0.515625,
0.33837890625,
0.471435546875,
0.33837890625
],
[
0.3056640625,
0.330078125,
0.334228515625,
0.330078125,
0.334228515625,
0.341552734375,
0.3056640625,
0.341552734375
],
[
0.365234375,
0.330078125,
0.390625,
0.330078125,
0.390625,
0.341552734375,
0.365234375,
0.341552734375
],
[
0.2484130859375,
0.35595703125,
0.72119140625,
0.354736328125,
0.72119140625,
0.369384765625,
0.2484130859375,
0.37060546875
],
[
0.080078125,
0.386962890625,
0.471435546875,
0.386962890625,
0.471435546875,
0.40087890625,
0.080078125,
0.40087890625
],
[
0.5009765625,
0.386474609375,
0.8486328125,
0.386474609375,
0.8486328125,
0.399658203125,
0.5009765625,
0.399658203125
],
[
0.08087158203125,
0.404052734375,
0.46728515625,
0.404052734375,
0.46728515625,
0.4140625,
0.08087158203125,
0.4140625
],
[
0.51953125,
0.401611328125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.51953125,
0.415283203125
],
[
0.07843017578125,
0.416748046875,
0.4697265625,
0.417236328125,
0.4697265625,
0.431884765625,
0.07843017578125,
0.43115234375
],
[
0.50244140625,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.080078125,
0.43310546875,
0.4697265625,
0.43310546875,
0.4697265625,
0.4462890625,
0.080078125,
0.4462890625
],
[
0.5009765625,
0.43310546875,
0.89208984375,
0.431884765625,
0.89208984375,
0.4462890625,
0.5009765625,
0.447509765625
],
[
0.07843017578125,
0.448974609375,
0.468017578125,
0.447509765625,
0.468017578125,
0.4609375,
0.07843017578125,
0.462158203125
],
[
0.5009765625,
0.448974609375,
0.89208984375,
0.448974609375,
0.89208984375,
0.462158203125,
0.5009765625,
0.462158203125
],
[
0.08087158203125,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.474853515625,
0.08087158203125,
0.474853515625
],
[
0.5009765625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.5009765625,
0.4765625
],
[
0.080078125,
0.477783203125,
0.46728515625,
0.477783203125,
0.46728515625,
0.4912109375,
0.080078125,
0.4912109375
],
[
0.50244140625,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50244140625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.080078125,
0.50830078125,
0.468017578125,
0.50830078125,
0.468017578125,
0.521484375,
0.080078125,
0.521484375
],
[
0.50244140625,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50244140625,
0.521484375
],
[
0.08087158203125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53515625,
0.08087158203125,
0.53515625
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.07843017578125,
0.53857421875,
0.471435546875,
0.53857421875,
0.471435546875,
0.5517578125,
0.07843017578125,
0.5517578125
],
[
0.50244140625,
0.5390625,
0.8935546875,
0.5390625,
0.8935546875,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.07843017578125,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.5673828125,
0.07843017578125,
0.5673828125
],
[
0.50244140625,
0.5537109375,
0.8935546875,
0.5537109375,
0.8935546875,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.080078125,
0.58203125
],
[
0.50341796875,
0.568359375,
0.890625,
0.5693359375,
0.890625,
0.58349609375,
0.50341796875,
0.58203125
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.5009765625,
0.58349609375,
0.890625,
0.583984375,
0.890625,
0.5986328125,
0.5009765625,
0.59765625
],
[
0.080078125,
0.599609375,
0.463134765625,
0.599609375,
0.463134765625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.50341796875,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50341796875,
0.61279296875
],
[
0.07843017578125,
0.62060546875,
0.28759765625,
0.62255859375,
0.28759765625,
0.63720703125,
0.07843017578125,
0.63525390625
],
[
0.50341796875,
0.6162109375,
0.890625,
0.6162109375,
0.890625,
0.626953125,
0.50341796875,
0.626953125
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.09722900390625,
0.64306640625,
0.471435546875,
0.64453125,
0.471435546875,
0.6591796875,
0.09722900390625,
0.65771484375
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.08087158203125,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.6875,
0.08087158203125,
0.6875
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.8486328125,
0.68994140625,
0.8486328125,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.521484375,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.521484375,
0.7197265625
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73486328125
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.07843017578125,
0.75048828125,
0.4697265625,
0.75,
0.4697265625,
0.7646484375,
0.07843017578125,
0.76513671875
],
[
0.50341796875,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.5009765625,
0.7646484375,
0.890625,
0.76513671875,
0.890625,
0.77978515625,
0.5009765625,
0.77880859375
],
[
0.08087158203125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.07843017578125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.07843017578125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.080078125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
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,
0.4697265625,
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
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
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,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88623046875,
0.89208984375,
0.88525390625,
0.89208984375,
0.8994140625,
0.5009765625,
0.90087890625
]
] |
[
"Initial",
"Improved",
"Training",
"→",
"Albedomap",
"Pretraining",
"Albedo map",
"Reconstruction",
"loss",
"",
"",
"De-Retouching",
"",
"1-",
"￥",
"1",
"Module",
"",
"3DMM database",
"",
"",
"Input portrait",
"A",
"Ao",
"Rendering",
"Rendering",
"A",
"",
"Modeler",
"A",
"?",
"",
"",
"←",
"Face",
"",
"Analyzer",
"Training",
"?",
"®",
"Deformatior",
"",
"Deformatior",
"Input portrait",
"3DMM coefficients",
"Mo",
"O",
"ase-to-fine",
"Mo",
"lossco",
"",
"M1",
"",
"M2",
"Training Data Generation",
"Deformationmap",
"Displacement map",
"T",
"1",
"losssmo",
"Re-align",
"losSyeg",
"A",
"Discriminator",
"Discriminator",
">",
"",
"P",
"losSadv.mid",
"losSadv_high",
"",
"Position",
"Texture",
"",
"Input portrait",
"Mo",
"pix2pix",
"p-Concat",
"pix2pix",
"Concat",
"map",
"map",
"Figure 2. Overview of the proposed hierarchical representation network (HRN).",
"framework. We first employ 3DMM to predict a coarse",
"frequency detail, while the fineness increases in turn.",
"mesh and albedo (blue area in Fig.2).Then we develop",
"We design the hierarchical representation to model",
"a hierarchical modeling (Sec. 3.2) strategy to handle the",
"the above three components respectively.For the low-",
"complex facial details in a coarse-to-fine manner (green and",
"frequency part, we adopt the BFM as our base model and",
"purple area). To facilitate the accurate and faithful model-",
"output the low-dimensional coefficients to fulfill a coarse",
"ingof thehierarchicaldetails,the3Dpriorsareincorporated",
"reconstruction of the input face. Then we introduce a three-",
"through adversarial and semi-supervised learning (Sec. 3.3).",
"channel deformation map, which lies in the UV space and",
"Besides, we propose a de-retouching module (Sec. 3.4) to",
"indicates the offset of each vertex relative to the coarse re-",
"fulfill better decoupling of the geometry and appearance,",
"sult. Worked as the representation of the mid-frequency de-",
"alleviating the ambiguities between the various skintexture",
"tails, the deformation map provides a flexible way to manip-",
"and illuminations. Moreover, we extend our framework to",
"ulate the geometries. We use the size of 64 × 64 to represent",
"a multi-view fashion (Sec. 3.5) and introduce a high quality",
"the deformation map to balance the fineness and smooth-",
"3D face dataset (Sec. 3.6) to boost the research on sparse-",
"ness of the mid-frequency details. For the high-frequency",
"view face reconstruction. For simplification, we specify the",
"details, we employ the displacement map following [19]",
"related loss functions and training strategy in each section.",
"which is a one-channel map (256 × 256) denoting the ge-",
"3.2.Hierarchical Modeling",
"ometry deformation along thedirection ofthenormals.The",
"displacement map is converted to detailed normals used in",
"3DMM exhibit great performance in expressing various",
"the rendering process in a pixel-wise manner to exhibit all",
"shape of the face, while the low-dimensional representation",
"the tiny details, breaking the limitation of the vertex density",
"itself severely stem its learning on the details, leading to",
"of the base model. Accordingly, we are enabled to describe",
"the imperfect alignment to the face or the over-smoothed",
"an arbitrary complex face with these representations.",
"results. Some methods extend 3DMM by introducing dis-",
"As shown in Fig. 2, given a portrait image I, we firstly",
"placement maps to reconstruct some details such as wrin-",
"utilize a regression network as the face analyzer to predict",
"kles and bumps. However, focused on the high-frequency",
"the BFM coefficients and obtain the coarse aligned mesh",
"part, a simple displacement map still fails to handle some",
"Mo and albedo Ao using the corresponding basis from the",
"larger-scale details, such as the contours of the jaw and",
"3DMM database. Combined I and Mo, we are able to ac-",
"cheek. Based on such observation, we decouple the fa-",
"quire the inpainted texture T in UV space by applying the",
"cial geometry into three components: (1) low-frequency",
"differentiable rendering with a coarse-to-fine strategy. And",
"part, which provides the coarse shape roughly aligned to",
"we concatenate the position map P and T as the input of",
"the input face; (2) mid-frequency details, which describe",
"the following modules for hierarchical details learning. We",
"the details of the contour and local shape relative to the",
"adopt two pix2pix [27] networks to synthesize the defor-",
"low-frequency part; (3) and high-frequency details, such as",
"mation map and displacement map in sequence. Note that,",
"wrinkles, micro bumps, etc. As shown in the Fig. 1, the",
"considering the deformation map will change the facial ge-",
"scales decrease from the low-frequency part to the high-",
"ometry and lead to the misalignment between T and the"
] |
{
"arxiv_id": "2302.14434",
"language": "en",
"page_idx": 2
}
|
||
2302.14434-en-3
|
2302.14434-en-3.jpg
|
deformed mesh, we generate the realigned texture $T^{\prime}$ as the input of second pix2pix network by projecting the threechannel deformation map to the 2D space and transform it to a reversal flow $F$ to re-align $T.$ Taking advantage of the abundant details from $T$ and the pixel-wise learning strategy, we manage to obtain the accurate detail maps which are further employed in a coarse-to-fine manner to generate the detailed face mesh $M_1$ and $M_{2}.$ Finally, combined with the lighting $L$ and the refined albedo generated from the deretouching module (Sec. 3.4), we accomplish detailed face reconstruction from the single image.
Overall, the framework is trained in a self-supervised manner guided with 3D detail priors learned from face scans (Sec. 3.3). To reduce the training complexity, we adopt the pre-trained encoder and MLP from the[15] as the face analyzer to predict the coefficients and generate the corresponding $P$ and $T$ for the following details learning. The two image translation networks are trained jointly and the related loss functions are composed of three components:
Reconstruction Loss. The reconstruction loss is calculated between the rendered face and the input face and consists of the photometric loss $L_photo$ , perception-level loss $L_per$ and landmark loss $L_{lan}$ following [15]. Thanks to the delighted albedo and the illumination system of 3DMM, the photometric loss will enforce the deformation of the facial geometry to fit the various shadows and highlight areas of the input face. It is crucial that we apply the reconstruction loss on both images rendered from the $M_1$ and $M_2$,which benefit the disentanglement of the MF and HF details.
Details Loss. We apply the total variation loss $L_{tv}$ [29] to encourage the smoothness of the deformation map, and use the L1 regularization loss $L_reg$ to limit the scale of the displacement map.
Contour-aware Loss. We propose a novel contour-aware loss $L_{con}$ to fulfill accurate modeling of the face contour. The $L_{con}$ works on $M_1$ and aims to pull the vertices of edge to align the face contour. As shown in Fig. 3, we firstly project vertices of $M_1$ into the image space. Then we predict the face mask $M_{face}$ using the pre-trained face matting network[34] and implement post process to obtain the left side and right side points for each row. Given a vertex $p$ and the corresponding projected point $p^\prime$ on $M_{face}$, we obtain the vector $l_p$ and $r_p$ (from $p^\prime$ to the edge points in the horizontal direction). Then $L_con$ can be describe as:
(1)
$$\begin{aligned}L_{con}&=\frac{1}{N_p}\sum_{p\in M_1}(f(p)\mathbb{1}[y(p')>\delta]),\end{aligned}$$
(2)
$$f(p)=|h(\frac{\boldsymbol{l_p}\cdot\boldsymbol{r_p}}{max(||\boldsymbol{l_p}||,||\boldsymbol{r_p}||)}+\lambda)-\lambda|,$$
where $h$ is the ReLU function, and $\lambda$ denotes a soft margin relative to the face contour $(\lambda=0.01$ as default), $1[y(p^{\prime})>$ $\delta]$ indicates whether $p^\prime$ is on the lower part of the image $(\delta=$
<FigureHere>
Figure 3. The details of the proposed contour loss. (a) input image,
(b) the projected vertices, (c) the predicted face mask.
<FigureHere>
Figure 4. An example of the 3D priors of facial details. (a) the raw image and scan, (b) the hierarchical representation and the corresponding fitted mesh.
100 as default). As we can see, $L_con$ punish the vertices outside the soft margin (such as the blue and gray points in Fig. 3) of the face and pull them to the face contour, while keeping the vertices inside the face intact. Combined with $L_{tv}$ of the deformation map, $L_con$ will avoid the unsmooth effect near the face contour. Note that we only focus on the lower part of the face contour to avoid the distraction of the hair. Compared to the common segmentation loss, $L_{con}$ gives a more straightforward direction for optimizing the face contour and is easier for training. We conduct an ablation study to reveal the effectiveness of $L_{con}$ in Sec.4.4.
## 3.3. 3D Priors of Facial Details
Although facial details can be roughly learned from single image using the reconstruction loss (Sec.3.2), it suffers from unreality and ambiguousness due to its ill-posed essence. Adding additional regularization may help to narrow the solution space, but also lead to severe degradation in detail accuracy and fidelity
To address this problem, we exploit the 3D priors of facial details derived from face scans and corresponding multi-view images in our framework. Firstly, given an raw image and its corresponding scan, we transform the raw scan to align to the image in BFM space (the details can be found in the supplemental files). Then we can obtain the ground-truth deformation map and displacement map for each image by fitting the image and scan using the loss functions mentioned in Sec. 3.2 with additional supervision on vertices distance following [3]. Thanks to the powerful hierarchical representation, the details of scans can be accurately captured. See Fig. 4 for example.
We take advantage of the 3D priors of details on two
|
[
[
0.080078125,
0.09027099609375,
0.46728515625,
0.09027099609375,
0.46728515625,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
0.080078125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.07843017578125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.07843017578125,
0.1514892578125,
0.46728515625,
0.15283203125,
0.46728515625,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.07843017578125,
0.1685791015625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.18115234375,
0.07843017578125,
0.1817626953125
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.56201171875,
0.1912841796875,
0.57861328125,
0.1912841796875,
0.57861328125,
0.2008056640625,
0.56201171875,
0.2008056640625
],
[
0.68701171875,
0.1925048828125,
0.70166015625,
0.1925048828125,
0.70166015625,
0.2008056640625,
0.68701171875,
0.2008056640625
],
[
0.81787109375,
0.1932373046875,
0.8310546875,
0.1932373046875,
0.8310546875,
0.2020263671875,
0.81787109375,
0.2020263671875
],
[
0.5009765625,
0.203857421875,
0.89208984375,
0.2052001953125,
0.89208984375,
0.2197265625,
0.5009765625,
0.2183837890625
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50341796875,
0.2197265625,
0.8203125,
0.2197265625,
0.8203125,
0.23291015625,
0.50341796875,
0.23291015625
],
[
0.08087158203125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.080078125,
0.24365234375,
0.324462890625,
0.24365234375,
0.324462890625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.09722900390625,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.09722900390625,
0.271484375
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.08087158203125,
0.28857421875,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.08087158203125,
0.3017578125
],
[
0.66650390625,
0.285400390625,
0.69140625,
0.285400390625,
0.69140625,
0.292236328125,
0.66650390625,
0.292236328125
],
[
0.080078125,
0.3037109375,
0.4697265625,
0.3037109375,
0.4697265625,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.73193359375,
0.3017578125,
0.744140625,
0.3017578125,
0.744140625,
0.31005859375,
0.73193359375,
0.31005859375
],
[
0.080078125,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.33203125,
0.080078125,
0.33203125
],
[
0.5849609375,
0.33203125,
0.59619140625,
0.33203125,
0.59619140625,
0.3359375,
0.5849609375,
0.3359375
],
[
0.78515625,
0.328857421875,
0.79833984375,
0.328857421875,
0.79833984375,
0.33837890625,
0.78515625,
0.33837890625
],
[
0.07843017578125,
0.349853515625,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.07843017578125,
0.363037109375
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50341796875,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.35986328125,
0.50341796875,
0.35986328125
],
[
0.08087158203125,
0.3662109375,
0.408447265625,
0.3662109375,
0.408447265625,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.50244140625,
0.3623046875,
0.89208984375,
0.3623046875,
0.89208984375,
0.375732421875,
0.50244140625,
0.375732421875
],
[
0.08087158203125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.5009765625,
0.376953125,
0.6611328125,
0.376953125,
0.6611328125,
0.39013671875,
0.5009765625,
0.39013671875
],
[
0.080078125,
0.395751953125,
0.4697265625,
0.395751953125,
0.4697265625,
0.4091796875,
0.080078125,
0.4091796875
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.412353515625,
0.8935546875,
0.412353515625,
0.8935546875,
0.4267578125,
0.50244140625,
0.4267578125
],
[
0.07843017578125,
0.4248046875,
0.468017578125,
0.423583984375,
0.468017578125,
0.438232421875,
0.07843017578125,
0.439453125
],
[
0.50244140625,
0.427978515625,
0.89208984375,
0.427978515625,
0.89208984375,
0.44140625,
0.50244140625,
0.44140625
],
[
0.08087158203125,
0.43994140625,
0.4697265625,
0.43994140625,
0.4697265625,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.50244140625,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.45703125,
0.50244140625,
0.45703125
],
[
0.080078125,
0.455810546875,
0.4697265625,
0.455810546875,
0.4697265625,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50244140625,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.07843017578125,
0.47021484375,
0.468017578125,
0.4697265625,
0.468017578125,
0.48291015625,
0.07843017578125,
0.484130859375
],
[
0.50341796875,
0.473388671875,
0.89208984375,
0.473388671875,
0.89208984375,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.488037109375,
0.8935546875,
0.488037109375,
0.8935546875,
0.501953125,
0.50244140625,
0.501953125
],
[
0.08087158203125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51220703125,
0.08087158203125,
0.51220703125
],
[
0.5048828125,
0.5048828125,
0.890625,
0.5048828125,
0.890625,
0.51513671875,
0.5048828125,
0.51513671875
],
[
0.080078125,
0.51513671875,
0.431396484375,
0.51513671875,
0.431396484375,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.50244140625,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.07843017578125,
0.52880859375,
0.468017578125,
0.52978515625,
0.468017578125,
0.54443359375,
0.07843017578125,
0.54345703125
],
[
0.50244140625,
0.5341796875,
0.89208984375,
0.53271484375,
0.89208984375,
0.54736328125,
0.50244140625,
0.548828125
],
[
0.0767822265625,
0.54443359375,
0.468017578125,
0.54541015625,
0.468017578125,
0.56005859375,
0.0767822265625,
0.5595703125
],
[
0.50244140625,
0.548828125,
0.89208984375,
0.548828125,
0.89208984375,
0.5625,
0.50244140625,
0.5625
],
[
0.0767822265625,
0.55810546875,
0.471435546875,
0.56005859375,
0.471435546875,
0.5771484375,
0.0767822265625,
0.57470703125
],
[
0.50244140625,
0.564453125,
0.89208984375,
0.564453125,
0.89208984375,
0.5791015625,
0.50244140625,
0.5791015625
],
[
0.080078125,
0.5771484375,
0.2034912109375,
0.5771484375,
0.2034912109375,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.08087158203125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.08087158203125,
0.60498046875
],
[
0.50244140625,
0.5888671875,
0.73876953125,
0.5888671875,
0.73876953125,
0.60400390625,
0.50244140625,
0.60400390625
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.62060546875,
0.080078125,
0.62060546875
],
[
0.521484375,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.521484375,
0.626953125
],
[
0.080078125,
0.6220703125,
0.471435546875,
0.6220703125,
0.471435546875,
0.63623046875,
0.080078125,
0.63623046875
],
[
0.5009765625,
0.62890625,
0.890625,
0.62744140625,
0.890625,
0.642578125,
0.5009765625,
0.64306640625
],
[
0.08087158203125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6484375,
0.08087158203125,
0.6484375
],
[
0.080078125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50341796875,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66748046875,
0.4697265625,
0.66748046875,
0.4697265625,
0.681640625,
0.080078125,
0.681640625
],
[
0.50341796875,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.08087158203125,
0.6845703125,
0.46728515625,
0.6845703125,
0.46728515625,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.50244140625,
0.68798828125,
0.701171875,
0.68798828125,
0.701171875,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.69775390625,
0.468017578125,
0.69775390625,
0.468017578125,
0.7109375,
0.080078125,
0.7109375
],
[
0.0767822265625,
0.708984375,
0.4697265625,
0.7109375,
0.4697265625,
0.72900390625,
0.0767822265625,
0.7265625
],
[
0.521484375,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.521484375,
0.71728515625
],
[
0.5009765625,
0.71826171875,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.5009765625,
0.73193359375
],
[
0.080078125,
0.72802734375,
0.468017578125,
0.72802734375,
0.468017578125,
0.7412109375,
0.080078125,
0.7412109375
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.7421875,
0.463134765625,
0.7421875,
0.463134765625,
0.7568359375,
0.080078125,
0.7568359375
],
[
0.50244140625,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.236083984375,
0.7744140625,
0.396240234375,
0.7744140625,
0.396240234375,
0.7890625,
0.236083984375,
0.7890625
],
[
0.44677734375,
0.77587890625,
0.4697265625,
0.77587890625,
0.4697265625,
0.7890625,
0.44677734375,
0.7890625
],
[
0.5009765625,
0.77880859375,
0.89208984375,
0.77978515625,
0.89208984375,
0.79443359375,
0.5009765625,
0.79345703125
],
[
0.2115478515625,
0.7841796875,
0.226318359375,
0.78857421875,
0.22216796875,
0.79736328125,
0.20751953125,
0.79345703125
],
[
0.22216796875,
0.79150390625,
0.268798828125,
0.79150390625,
0.268798828125,
0.80615234375,
0.22216796875,
0.80615234375
],
[
0.50244140625,
0.79443359375,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.80810546875
],
[
0.5048828125,
0.8125,
0.8896484375,
0.8125,
0.8896484375,
0.82275390625,
0.5048828125,
0.82275390625
],
[
0.2484130859375,
0.8173828125,
0.29736328125,
0.8173828125,
0.29736328125,
0.83203125,
0.2484130859375,
0.83203125
],
[
0.1544189453125,
0.7763671875,
0.191162109375,
0.7763671875,
0.191162109375,
0.78857421875,
0.1544189453125,
0.78857421875
],
[
0.1380615234375,
0.82568359375,
0.214111328125,
0.82568359375,
0.214111328125,
0.84033203125,
0.1380615234375,
0.84033203125
],
[
0.338134765625,
0.82568359375,
0.41259765625,
0.82568359375,
0.41259765625,
0.84033203125,
0.338134765625,
0.84033203125
],
[
0.4453125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.8388671875,
0.4453125,
0.8388671875
],
[
0.50244140625,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.2091064453125,
0.83203125,
0.338134765625,
0.833984375,
0.338134765625,
0.8515625,
0.2091064453125,
0.849609375
],
[
0.50341796875,
0.84033203125,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.50341796875,
0.85400390625
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.080078125,
0.869140625
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50341796875,
0.869140625
],
[
0.07843017578125,
0.869140625,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.8837890625
],
[
0.5009765625,
0.869140625,
0.7958984375,
0.87060546875,
0.7958984375,
0.88525390625,
0.5009765625,
0.8837890625
],
[
0.080078125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.52294921875,
0.88623046875,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.52294921875,
0.90087890625
]
] |
[
"deformed mesh, we generate the realigned texture T as the",
"input of second pix2pix network by projecting the three-",
"channel deformation map to the 2D space and transform it",
"to a reversal flow F to re-align T. Taking advantage of the",
"abundant details from T and the pixel-wise learning strat-",
"egy,wemanageto obtaintheaccuratedetail mapswhich",
"are further employed in a coarse-to-fine manner to generate",
"(a)",
"(b)",
"(c)",
"Figure 3. The details of the proposed contour loss. (a) input image,",
"the detailed face mesh Mi and M2. Finally, combined with",
"the lighting L and the refined albedo generated from the de-",
"(b) the projected vertices, (c) the predicted face mask.",
"retouchingmodule(Sec.3.4),weaccomplishdetailedface",
"reconstruction from the single image.",
"Overall, the framework is trained in a self-supervised",
"manner guided with 3D detail priors learned from face scans",
"(Sec. 3.3). To reduce the training complexity, we adopt the",
"fittin",
"pre-trained encoder and MLP from the [15] as the face ana-",
"",
"lyzerto predict the coefficients and generate the correspond-",
"al",
"(b)",
"age translation networks are trained jointly and the related",
"ing P and T for the following details learning. The two im-",
"Figure 4.An example of the3D priors of facial details.(a)the",
"lossfunctionsarecomposedofthreecomponents:",
"raw image and scan, (b) the hierarchical representation and the",
"ReconstructionLoss.Thereconstructionloss iscalculated",
"corresponding fitted mesh.",
"between the rendered face and the input face and consists",
"of the photometric loss Lphoto, perception-level loss Lper",
"100 as default). As we can see, Lcon punish the vertices",
"and landmark loss Lian following [15]. Thanks to the de-",
"outside the soft margin (such as the blue and gray points in",
"lighted albedo and the illumination system of 3DMM, the",
"Fig. 3) of the face and pull them to the face contour, while",
"photometric loss will enforce the deformation of the facial",
"keeping the vertices inside the face intact. Combined with",
"geometry to fit the various shadows and highlight areas of",
"Lt, of the deformation map, Lcon will avoid the unsmooth",
"the input face. It is crucial that we apply the reconstruction",
"effect near the face contour. Note that we only focus on",
"lossonbothimagesrenderedfromtheM,andM2,which",
"thelowerpart ofthe face contourto avoid thedistraction",
"benefit the disentanglement of the MF and HF details.",
"of the hair. Compared to the common segmentation loss",
"Details Loss. We apply the total variation loss Ltu [29]",
"Lcon gives a more straightforward direction for optimizing",
"to encourage the smoothness of the deformation map, and",
"the face contour and is easier for training. We conduct an",
"use the LI regularization loss Lreg to limit the scale of the",
"ablation study to reveal the effectiveness of Lcon in Sec. 4.4.",
"displacement map.",
"Contour-aware Loss. We propose a novel contour-aware",
"3.3. 3D Priors of Facial Details",
"loss Lcon to fulfill accurate modeling of the face contour.",
"Although facial details can be roughly learned from sin-",
"The Lcon works on Mi and aims to pull the vertices of",
"gle image using the reconstruction loss (Sec. 3.2), it suf-",
"edgetoaligntheface contour.As shown in Fig.3.we",
"firstly project vertices of Mi into the image space. Then",
"fers from unreality and ambiguousness dueto its ill-posed",
"essence. Adding additional regularization may help to nar-",
"we predict the face mask Mface using the pre-trained face",
"row the solution space,but alsolead to severe degradation",
"mattingnetwork[34landimplementpostprocesstoobtain",
"in detail accuracy and fidelity.",
"the left side and right side points for each row. Given a ver-",
"tex p and the corresponding projected point p' on Mface,",
"To address this problem, we exploit the 3D priors of",
"facial details derived from face scans and corresponding",
"we obtain the vector Ip and rp (from p' to the edge points",
"multi-view images in our framework. Firstly, given an raw",
"in the horizontal direction). Then Lcon can be describe as:",
"image and its corresponding scan, we transform the raw",
"scan to align to the image in BFM space (the details can",
"(f(p)1[y(p) >),",
"(1)",
"be found in the supplemental files). Then we can obtain",
"N",
"PpeMi",
"the ground-truth deformation map and displacement map",
"for eachimagebyfittingthe imageand scanusingthe loss",
"lp-rp",
"Lcon",
"f(p) =[h(",
"+>)一入,",
"(2)",
"functions mentioned in Sec. 3.2 with additional supervision",
"mar(/pll, /rpl)",
"on vertices distance following [3]. Thanks to the power-",
"where h is the ReLU function, and denotes a soft margin",
"ful hierarchical representation, the details of scans can be",
"relative to the face contour ( = 0.01 as default), 1I [y(p') >",
"accurately captured. See Fig. 4 for example.",
") indicates whether p' is on the lower part of the image (d =",
"We take advantage of the 3D priors of details on two"
] |
{
"arxiv_id": "2302.14434",
"language": "en",
"page_idx": 3
}
|
||
2302.14434-en-4
|
2302.14434-en-4.jpg
|
<FigureHere>
Figure 6. The pipeline of the proposed MV-HRN.
aspects. On the one hand, we develop two discriminators and use the adversarial loss [28] $L_{adv\_mid}$ and $L_{adv\_high}$ to supervise the domain distribution of the deformation map and displacement map. On the other hand, we acquire the paired data as mentioned above from 3D scans to conduct supervised learning to guide the self-supervised learning in Sec. 3.2. Specifically, we supplement the L1 loss $L_{mid}$ and $L_{high}$ for the predicted deformation map and displacement map respectively. Note that a mask is used in training to remove the distraction of eyes and hair area from scans.
## $3. 4. \textbf{ De- Retouching Module}$
A face image is the result of a combination of geometry, lighting, and face albedo. Prior works assume that the face albedo is smooth and model it with the low-frequency albedo from 3DMM. However, the actual skin texture is full of high-frequency details such as moles, scars, freckles, and other blemishes, which bring ambiguities to the geometry details learning especially in the single view FR task. Inspired by the[33], we propose a de-retouching module (DRM), which aims to generate the face albedo with highfrequency details and facilitate more precise decoupling of $\hat{\text{geometry and appearance.}}$
We collect 10, 000 face images from FFHQ [30], and hire a team of professional image editors to process the images, with the goal of removing the skin blemishes and other texture details while maintaining the shape-related content such as wrinkles, bumps, etc. Then we transform the paired
images into the texture maps in UV space by applying the process specified in Sec. 3.2 and train an image translation network $G$ to fulfill skin retouching. Given the re-aligned texture $T^{\prime}$ , we firstly employ $G$ to remove its texture details and get $T_0$, as shown in Fig. 5. We aim to bake the texture details into the coarse albedo $A_0$ to obtain the improved albedo $A^{\prime}$ for rendering. We make an assumption that the shading from $A_0$ to $T_0$ should be consistent with the one $\operatorname{from}\tilde{A}^{\prime}$to$T^{\prime},\operatorname{as:}$
(3)
$$\begin{aligned}&T_0=A_0\odot S,\\&T^{'}=A^{'}\odot S,\end{aligned}$$
(4)
where $S$ denotes the shading map, $\odot$ denotes element-wise matrix multiplication. Then we can solve the equations and obtain $A^{\prime}$ as:
(5)
$$A^{'}=A_0\odot B\approx A_0\odot\frac{T^{'}+\phi(T_0)}{T_0+\phi(T_0)},$$
$$\phi(\boldsymbol{T_0})=\frac{1}{\frac{\boldsymbol{T_0}\odot\boldsymbol{T_0}\odot\boldsymbol{T_0}}{\varepsilon}+\varepsilon},$$
(6)
where $\phi(T_{0})$ avoids the value explosion near 0 and $\varepsilon=$ $1e-6$ as default. Compared to $A_0$,the de-retouched albedo $A^{\prime}$ contains more high-frequency texture details, which alleviate the ambiguities between geometry and appearance, especially in single view FR task.
### $3.5.MV-HRN$
Thanks to the hierarchical modeling and the 3D priors guidance, we can easily adapt the HRN to a multi-view fashion to fulfill precise modeling of the global facial geometry with only a few views by adding the geometry consistency between different views. Fig. 6 shows the pipeline of MV-HRN. We assume that the low-frequency identity part and the mid-frequency details are consistent between different views, while the lighting, expression, and HF details should be view-dependent to overcome the disturbance. Therefore, we develop a canonical space, which contains the shared identity coefficient and deformation map that are initialized by averaging all the single-view results, to represent the shared intrinsic face shape. Then other BFM coeffcients and the displacement map of each view are utilized to dependently model the pose, lighting, expression and highfrequency details. Then we apply the loss functions mentioned in Sec. 3.2 and Sec. 3.3 to iteratively optimize all the coefficients and detail maps. Through the fitting process, the face shape is gradually restricted to a smaller and more accurate space under the supervision of different views. Extensive experiments show that MV-HRN achieves accurate reconstruction given only a few $(2\sim5)$ views of images in a short time (less than one minute).
|
[
[
0.50341796875,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.106689453125,
0.890625,
0.12060546875,
0.50244140625,
0.121826171875
],
[
0.50341796875,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50341796875,
0.1363525390625
],
[
0.5009765625,
0.1363525390625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.5009765625,
0.15087890625
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.30810546875,
0.160400390625,
0.31201171875,
0.160400390625,
0.31201171875,
0.1654052734375,
0.30810546875,
0.1654052734375
],
[
0.50341796875,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.5048828125,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1944580078125,
0.5048828125,
0.1944580078125
],
[
0.50341796875,
0.1976318359375,
0.89208984375,
0.1976318359375,
0.89208984375,
0.2108154296875,
0.50341796875,
0.2108154296875
],
[
0.373291015625,
0.2139892578125,
0.39208984375,
0.2139892578125,
0.39208984375,
0.2215576171875,
0.373291015625,
0.2215576171875
],
[
0.5009765625,
0.2108154296875,
0.62841796875,
0.212158203125,
0.62841796875,
0.2266845703125,
0.5009765625,
0.2259521484375
],
[
0.64453125,
0.24365234375,
0.74853515625,
0.24560546875,
0.74853515625,
0.261474609375,
0.64453125,
0.2587890625
],
[
0.86767578125,
0.2442626953125,
0.89453125,
0.2442626953125,
0.89453125,
0.261474609375,
0.86767578125,
0.261474609375
],
[
0.6435546875,
0.267578125,
0.74853515625,
0.26953125,
0.74853515625,
0.287353515625,
0.6435546875,
0.285400390625
],
[
0.12335205078125,
0.246826171875,
0.42236328125,
0.24560546875,
0.42236328125,
0.260009765625,
0.12335205078125,
0.261474609375
],
[
0.86767578125,
0.271484375,
0.89453125,
0.271484375,
0.89453125,
0.28857421875,
0.86767578125,
0.28857421875
],
[
0.50341796875,
0.29736328125,
0.89208984375,
0.29736328125,
0.89208984375,
0.311279296875,
0.50341796875,
0.311279296875
],
[
0.50244140625,
0.313232421875,
0.89208984375,
0.313232421875,
0.89208984375,
0.326904296875,
0.50244140625,
0.326904296875
],
[
0.50244140625,
0.32763671875,
0.58837890625,
0.32763671875,
0.58837890625,
0.339111328125,
0.50244140625,
0.339111328125
],
[
0.73388671875,
0.352783203125,
0.81787109375,
0.352783203125,
0.81787109375,
0.37060546875,
0.73388671875,
0.37060546875
],
[
0.56884765625,
0.359130859375,
0.73779296875,
0.361083984375,
0.73779296875,
0.37890625,
0.56884765625,
0.376953125
],
[
0.86767578125,
0.361083984375,
0.89453125,
0.361083984375,
0.89453125,
0.37744140625,
0.86767578125,
0.37744140625
],
[
0.73193359375,
0.37060546875,
0.81884765625,
0.37255859375,
0.81884765625,
0.386962890625,
0.73193359375,
0.385009765625
],
[
0.7197265625,
0.401611328125,
0.73193359375,
0.401611328125,
0.73193359375,
0.410888671875,
0.7197265625,
0.410888671875
],
[
0.61279296875,
0.4091796875,
0.67333984375,
0.4091796875,
0.67333984375,
0.423583984375,
0.61279296875,
0.423583984375
],
[
0.66845703125,
0.410888671875,
0.78125,
0.416748046875,
0.779296875,
0.431884765625,
0.66650390625,
0.4267578125
],
[
0.869140625,
0.4091796875,
0.8935546875,
0.4091796875,
0.8935546875,
0.42236328125,
0.869140625,
0.42236328125
],
[
0.318603515625,
0.416748046875,
0.334228515625,
0.419189453125,
0.326904296875,
0.436279296875,
0.31201171875,
0.43310546875
],
[
0.1168212890625,
0.4248046875,
0.131591796875,
0.4248046875,
0.131591796875,
0.43310546875,
0.1168212890625,
0.43310546875
],
[
0.22216796875,
0.427978515625,
0.233642578125,
0.427978515625,
0.233642578125,
0.431884765625,
0.22216796875,
0.431884765625
],
[
0.4150390625,
0.427978515625,
0.423095703125,
0.427978515625,
0.423095703125,
0.43115234375,
0.4150390625,
0.43115234375
],
[
0.1258544921875,
0.442626953125,
0.420654296875,
0.44140625,
0.420654296875,
0.455810546875,
0.1258544921875,
0.45703125
],
[
0.50341796875,
0.443115234375,
0.875,
0.443115234375,
0.875,
0.45703125,
0.50341796875,
0.45703125
],
[
0.5048828125,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.47021484375,
0.5048828125,
0.47021484375
],
[
0.080078125,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.5048828125,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.5048828125,
0.4873046875
],
[
0.07843017578125,
0.488037109375,
0.471435546875,
0.488037109375,
0.471435546875,
0.50244140625,
0.07843017578125,
0.50244140625
],
[
0.50244140625,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.07843017578125,
0.501953125,
0.468017578125,
0.50244140625,
0.468017578125,
0.517578125,
0.07843017578125,
0.5166015625
],
[
0.5009765625,
0.50390625,
0.7255859375,
0.50244140625,
0.7255859375,
0.517578125,
0.5009765625,
0.51806640625
],
[
0.080078125,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.07843017578125,
0.5341796875,
0.4697265625,
0.53271484375,
0.4697265625,
0.546875,
0.07843017578125,
0.54736328125
],
[
0.50341796875,
0.52978515625,
0.61279296875,
0.52978515625,
0.61279296875,
0.54345703125,
0.50341796875,
0.54345703125
],
[
0.08087158203125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.56103515625,
0.08087158203125,
0.56103515625
],
[
0.521484375,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.521484375,
0.5673828125
],
[
0.080078125,
0.564453125,
0.4697265625,
0.564453125,
0.4697265625,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50341796875,
0.5693359375,
0.89208984375,
0.5693359375,
0.89208984375,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.07843017578125,
0.57763671875,
0.471435546875,
0.57568359375,
0.471435546875,
0.59326171875,
0.07843017578125,
0.59521484375
],
[
0.5009765625,
0.58349609375,
0.890625,
0.583984375,
0.890625,
0.5986328125,
0.5009765625,
0.59765625
],
[
0.080078125,
0.5947265625,
0.4697265625,
0.5947265625,
0.4697265625,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50244140625,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50244140625,
0.61279296875
],
[
0.07843017578125,
0.60986328125,
0.4453125,
0.60986328125,
0.4453125,
0.62353515625,
0.07843017578125,
0.62353515625
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.50244140625,
0.62890625,
0.89208984375,
0.630859375,
0.89208984375,
0.64453125,
0.50244140625,
0.642578125
],
[
0.080078125,
0.63525390625,
0.2900390625,
0.63525390625,
0.2900390625,
0.6484375,
0.080078125,
0.6484375
],
[
0.5048828125,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.098876953125,
0.6591796875,
0.46728515625,
0.6591796875,
0.46728515625,
0.67236328125,
0.098876953125,
0.67236328125
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.080078125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.08087158203125,
0.69140625,
0.46728515625,
0.69140625,
0.46728515625,
0.7021484375,
0.08087158203125,
0.7021484375
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.5009765625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.08087158203125,
0.7216796875,
0.468017578125,
0.7216796875,
0.468017578125,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.77880859375,
0.468017578125,
0.77978515625,
0.468017578125,
0.79443359375,
0.080078125,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.080078125,
0.79541015625,
0.471435546875,
0.79541015625,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.07843017578125,
0.81103515625,
0.250732421875,
0.81005859375,
0.250732421875,
0.82373046875,
0.07843017578125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.098876953125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
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,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.5009765625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50244140625,
0.86865234375
],
[
0.07843017578125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.5009765625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.734375,
0.88720703125,
0.734375,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"images into the texture maps in UV space by applying the",
"process specified in Sec. 3.2 and train an image translation",
"network G to fulfill skin retouching. Given the re-aligned",
"texture T', we firstly employ G to remove its texture details",
"and get To, as shown in Fig. 5. We aim to bake the texture",
"",
"details into the coarse albedo Ao to obtain the improved",
"albedo Aforrendering.Wemakeanassumptionthatthe",
"shading from Ao to To should be consistent with the one",
"Blen",
"from A'to T',as:",
"To = Ao O S,",
"(3)",
"T=A'OS,",
"Figure 5. The details of the de-retouching module.",
"(4)",
"where S denotes the shading map, denotes element-wise",
"matrix multiplication. Then we can solve the equations and",
"obtain A as:",
"T' +Φ(To)",
"A'=AoOB~ Ao0",
"(5)",
"To + Φ(To)",
"1",
"Φ(To) =",
"TooTooTo+e",
"(6)",
"(0",
"(a)",
"TD",
"",
"Figure 6. The pipeline of the proposed MV-HRN.",
"where p(To) avoids the value explosion near 0 and ",
"1e-6asdefault.ComparedtoAo,thede-retouchedalbedo",
"aspects. On the one hand, we develop two discriminators",
"A' contains more high-frequency texture details, which al-",
"and use the adversarial loss [28] Ladv-mid and Ladu.high to",
"leviate the ambiguities between geometry and appearance,",
"supervise the domain distribution of the deformation map",
"especially in single view FR task.",
"and displacement map. On the other hand, we acquire the",
"paired data as mentioned above from 3D scans to conduct",
"3.5. MV-HRN",
"supervised learning to guidethe self-supervised learning in",
"Thanks to the hierarchical modeling and the 3D priors",
"Sec. 3.2. Specifically, we supplement the L1 loss Lmid and",
"guidance, we can easily adapt the HRN to a multi-view",
"Lhigh for the predicted deformation map and displacement",
"fashion to fulfill precise modeling of the global facial geom-",
"map respectively. Note that a mask is used in training to",
"etry with only a few views by adding the geometry consis-",
"remove the distraction of eyes and hair area from scans.",
"tency between different views. Fig. 6 shows the pipeline of",
"MV-HRN. We assume that the low-frequency identity part",
"3.4. De-Retouching Module",
"and the mid-frequency details are consistent between dif-",
"A face image is the result of a combination of geome-",
"ferent views, while the lighting, expression, and HF details",
"try, lighting, and face albedo. Prior works assume that the",
"should be view-dependent to overcome the disturbance.",
"facealbedo is smoothand model it withthe low-frequency",
"Therefore, we develop a canonical space, which contains",
"albedo from 3DMM. However, the actual skin texture is",
"the shared identity coefficient and deformation map that are",
"full of high-freguency details such as moles.scars,freckles.",
"initialized by averaging all the single-view results, to repre-",
"and other blemishes, which bring ambiguities to the geom-",
"sent the shared intrinsic face shape. Then other BFM coeffi-",
"etry details learning especially in the single view FR task.",
"cients and the displacement map of each view are utilizedto",
"Inspired by the [33], we propose a de-retouching module",
"dependently model the pose, lighting, expression and high-",
"(DRM), which aims to generate the face albedo with high-",
"frequency details. Then we apply the loss functions men-",
"frequency details and facilitate more precise decoupling of",
"tioned in Sec. 3.2 and Sec. 3.3 to iteratively optimize all the",
"geometryandappearance",
"coefficients and detail maps.Through the fitting process,",
"We collect 10, 000 face images from FFHQ [30], and",
"the face shape is gradually restricted to a smaller and more",
"hire a team of professional image editors to process the im-",
"accurate space under the supervision of different views.Ex-",
"ages, with the goal of removing the skin blemishes and other",
"tensive experiments show that MV-HRN achieves accurate",
"texture details while maintaining the shape-related content",
"reconstruction given only a few (2 ~ 5) views of images in",
"such as wrinkles, bumps, etc. Then we transform the paired",
"a short time (less than one minute)."
] |
{
"arxiv_id": "2302.14434",
"language": "en",
"page_idx": 4
}
|
||
1809.01649-en-0
|
1809.01649-en-0.jpg
|
00
DF-Net: Unsupervised Joint Learning of
$\begin{array}{c}\text{三 Depth and Flow using Cross-Task Consistency}\\\text{C}\\\text{C}\end{array}$
$ప్సి$
Yuliang Zou$^1$,Zelun Luo$^2$,and Jia-Bin Huang$^{\mathrm{I}}$
$^{1}$Virginia Tech$\:^2$Stanford University
$w$
<FigureHere>
$\sum$Fig.1: Joint learning v.s. separate learning. Single-view depth prediction
$\bigcirc$and optical flow estimation are two highly correlated tasks. Existing work,
$\underbrace{寸}$ however, often addresses these two tasks in isolation. In this paper, we propose $\dot{\text{\"{O}a novel cross-task consistency loss to couple the training of these two problems}}$ $\sum_{\text{constraints, we show substantially improved results fored both tasks.}}^{\text{using unlabeled monocular videos. Through enforcing the underlying geometric}}$ $ଚ୍ରି$
$\infty0$
一
$\therefore$
汉
$స్రాని$
$\mathbf{Abstract.~We~present~an~unsupervised~learning~framework~for~simulta-}$ neously training single-view depth prediction and optical flow estimation models using unlabeled video sequences. Existing unsupervised methods often exploit brightness constancy and spatial smoothness priors to train depth or flow models. In this paper, we propose to leverage geometric consistency as additional supervisory signals. Our core idea is that for rigid regions we can use the predicted scene depth and camera motion to synthesize 2D optical flow by backprojecting the induced 3D scene flow. The discrepancy between the rigid flow (from depth prediction and camera motion) and the estimated flow (from optical flow model) allows us to impose a cross-task consistency loss. While all the networks are jointly optimized during training, they can be applied independently at test time. Extensive experiments demonstrate that our depth and flow models compare favorably with state-of-the-art unsupervised methods.
## 1 Introduction
Single-view depth prediction and optical flow estimation are two fundamental problems in computer vision. While the two tasks aim to recover highly correlated information from the scene (i.e., the scene structure and the dense motion field between consecutive frames), existing efforts typically study each problem $\begin{array}{c}\text{in isolation. In this paper, we demonstrate the benefits of exploring the geometric}\\\end{array}$ relationship between depth, camera motion, and flow for unsupervised learning of depth and flow estimation models.
|
[
[
0.1510009765625,
0.052001953125,
0.84765625,
0.05279541015625,
0.84765625,
0.07391357421875,
0.1510009765625,
0.07232666015625
],
[
0.043487548828125,
0.061737060546875,
0.0809326171875,
0.061737060546875,
0.0809326171875,
0.0869140625,
0.043487548828125,
0.0869140625
],
[
0.032623291015625,
0.07391357421875,
0.09417724609375,
0.07391357421875,
0.09417724609375,
0.194091796875,
0.032623291015625,
0.194091796875
],
[
0.1014404296875,
0.08203125,
0.8935546875,
0.0828857421875,
0.8935546875,
0.10479736328125,
0.1014404296875,
0.10400390625
],
[
0.251220703125,
0.1422119140625,
0.74267578125,
0.1422119140625,
0.74267578125,
0.1600341796875,
0.251220703125,
0.1600341796875
],
[
0.306884765625,
0.1746826171875,
0.4638671875,
0.17626953125,
0.4638671875,
0.19580078125,
0.306884765625,
0.193359375
],
[
0.46142578125,
0.18115234375,
0.48193359375,
0.18115234375,
0.48193359375,
0.1893310546875,
0.46142578125,
0.1893310546875
],
[
0.475830078125,
0.1746826171875,
0.689453125,
0.1778564453125,
0.689453125,
0.19580078125,
0.475830078125,
0.1925048828125
],
[
0.03985595703125,
0.1893310546875,
0.0784912109375,
0.1893310546875,
0.0784912109375,
0.2120361328125,
0.03985595703125,
0.2120361328125
],
[
0.03985595703125,
0.2298583984375,
0.0809326171875,
0.2298583984375,
0.0809326171875,
0.291748046875,
0.03985595703125,
0.291748046875
],
[
0.0809326171875,
0.2242431640625,
0.1026611328125,
0.2242431640625,
0.1026611328125,
0.2705078125,
0.0809326171875,
0.2705078125
],
[
0.044677734375,
0.267333984375,
0.10992431640625,
0.267333984375,
0.10992431640625,
0.322509765625,
0.044677734375,
0.322509765625
],
[
0.0809326171875,
0.2998046875,
0.08697509765625,
0.305419921875,
0.0809326171875,
0.30859375,
0.07489013671875,
0.302978515625
],
[
0.207763671875,
0.318359375,
0.271728515625,
0.318359375,
0.271728515625,
0.337158203125,
0.207763671875,
0.337158203125
],
[
0.423828125,
0.32177734375,
0.60888671875,
0.32177734375,
0.60888671875,
0.33642578125,
0.423828125,
0.33642578125
],
[
0.6787109375,
0.32177734375,
0.90234375,
0.32177734375,
0.90234375,
0.33642578125,
0.6787109375,
0.33642578125
],
[
0.059173583984375,
0.349365234375,
0.90576171875,
0.350830078125,
0.90576171875,
0.371337890625,
0.059173583984375,
0.370361328125
],
[
0.043487548828125,
0.371337890625,
0.8671875,
0.36962890625,
0.8671875,
0.389892578125,
0.043487548828125,
0.392333984375
],
[
0.060394287109375,
0.3916015625,
0.9130859375,
0.39306640625,
0.9130859375,
0.40771484375,
0.060394287109375,
0.40625
],
[
0.04107666015625,
0.406982421875,
0.91650390625,
0.41015625,
0.91650390625,
0.4306640625,
0.04107666015625,
0.42822265625
],
[
0.0772705078125,
0.4306640625,
0.92041015625,
0.4306640625,
0.92041015625,
0.448486328125,
0.0772705078125,
0.448486328125
],
[
0.0821533203125,
0.45166015625,
0.7802734375,
0.45166015625,
0.7802734375,
0.46630859375,
0.0821533203125,
0.46630859375
],
[
0.03985595703125,
0.4638671875,
0.0809326171875,
0.4638671875,
0.0809326171875,
0.5263671875,
0.03985595703125,
0.5263671875
],
[
0.1522216796875,
0.496337890625,
0.8466796875,
0.496337890625,
0.8466796875,
0.5107421875,
0.1522216796875,
0.5107421875
],
[
0.037445068359375,
0.51513671875,
0.0809326171875,
0.51513671875,
0.08331298828125,
0.6181640625,
0.03985595703125,
0.6181640625
],
[
0.1497802734375,
0.51513671875,
0.8466796875,
0.51513671875,
0.8466796875,
0.52978515625,
0.1497802734375,
0.52978515625
],
[
0.1497802734375,
0.53271484375,
0.84912109375,
0.53271484375,
0.84912109375,
0.54736328125,
0.1497802734375,
0.54736328125
],
[
0.1497802734375,
0.55078125,
0.84765625,
0.55078125,
0.84765625,
0.5654296875,
0.1497802734375,
0.5654296875
],
[
0.1497802734375,
0.5693359375,
0.84912109375,
0.5693359375,
0.84912109375,
0.58349609375,
0.1497802734375,
0.58349609375
],
[
0.1497802734375,
0.58740234375,
0.84912109375,
0.58740234375,
0.84912109375,
0.60205078125,
0.1497802734375,
0.60205078125
],
[
0.1497802734375,
0.603515625,
0.84765625,
0.603515625,
0.84765625,
0.6181640625,
0.1497802734375,
0.6181640625
],
[
0.1473388671875,
0.62158203125,
0.84912109375,
0.62158203125,
0.84912109375,
0.63623046875,
0.1473388671875,
0.63623046875
],
[
0.1473388671875,
0.63916015625,
0.84912109375,
0.63916015625,
0.84912109375,
0.65380859375,
0.1473388671875,
0.65380859375
],
[
0.1473388671875,
0.6572265625,
0.84765625,
0.6572265625,
0.84765625,
0.67041015625,
0.1473388671875,
0.67041015625
],
[
0.1461181640625,
0.67578125,
0.84912109375,
0.67431640625,
0.84912109375,
0.68896484375,
0.1461181640625,
0.6904296875
],
[
0.1473388671875,
0.69189453125,
0.84765625,
0.69189453125,
0.84765625,
0.70654296875,
0.1473388671875,
0.70654296875
],
[
0.1473388671875,
0.7109375,
0.84765625,
0.7109375,
0.84765625,
0.7255859375,
0.1473388671875,
0.7255859375
],
[
0.1448974609375,
0.72802734375,
0.8408203125,
0.72705078125,
0.8408203125,
0.744140625,
0.1448974609375,
0.74560546875
],
[
0.08331298828125,
0.77978515625,
0.1014404296875,
0.77978515625,
0.1014404296875,
0.79052734375,
0.08331298828125,
0.79052734375
],
[
0.12078857421875,
0.77490234375,
0.30908203125,
0.7763671875,
0.30908203125,
0.79541015625,
0.12078857421875,
0.79345703125
],
[
0.0784912109375,
0.8125,
0.91796875,
0.81298828125,
0.91796875,
0.8310546875,
0.0784912109375,
0.82958984375
],
[
0.0809326171875,
0.8349609375,
0.91162109375,
0.83447265625,
0.91162109375,
0.84912109375,
0.0809326171875,
0.849609375
],
[
0.0821533203125,
0.85400390625,
0.91552734375,
0.85400390625,
0.91552734375,
0.8681640625,
0.0821533203125,
0.8681640625
],
[
0.0772705078125,
0.8701171875,
0.91650390625,
0.87158203125,
0.91650390625,
0.8896484375,
0.0772705078125,
0.8876953125
],
[
0.0809326171875,
0.89208984375,
0.91552734375,
0.89208984375,
0.91552734375,
0.90673828125,
0.0809326171875,
0.90673828125
],
[
0.0772705078125,
0.9091796875,
0.92041015625,
0.91162109375,
0.92041015625,
0.9287109375,
0.0772705078125,
0.9267578125
],
[
0.0809326171875,
0.931640625,
0.4697265625,
0.931640625,
0.4697265625,
0.9462890625,
0.0809326171875,
0.9462890625
]
] |
[
"DF-Net: Unsupervised Joint Learning of",
"8",
"310des",
"Depth and Flow using Cross-Task Consistency",
"Yuliang Zoul, Zelun Luo?, and Jia-Bin Huang",
"1Virginia Tech",
"",
"2StanfordUniversity",
"5",
"LA'",
"a",
"",
"",
"Input",
"Separatelearning",
"Joint learning (Ours)",
"Fig.1:Joint learning v.s.separate learning.Single-view depth prediction",
"and optical flow estimation are two highly correlated tasks. Existing work,",
"+however,oftenaddressesthesetwotasks in isolation.Inthispaper,wepropose",
"Oa novel cross-task consistency loss to couple the training of these two problems",
"using unlabeled monocular videos. Through enforcing the underlying geometric",
"constraints,we show substantiallyimprovedresults for bothtasks.",
"6081",
"Abstract. We present an unsupervised learning framework for simulta-",
"I:AIXJE",
"neouslytrainingsingle-viewdepthprediction and optical flowestimation",
"models using unlabeled video sequences. Existing unsupervised methods",
"often exploit brightness constancy and spatial smoothness priors to train",
"depth or flow models.In this paper,we propose to leverage geometric",
"consistency as additional supervisory signals. Our core idea is that for",
"rigid regions we can use the predicted scene depth and camera motion",
"tosynthesize 2Doptical flowby backprojectingthe induced3D scene",
"fow.Thediscrepancybetweentherigid flow (fromdepthpredictionand",
"cameramotion)andtheestimatedflow(fromopticalflowmodel)allows",
"us to impose a cross-task consistency loss. While all the networks are",
"jointly optimized during training,they can be applied independently at",
"test time. Extensive experiments demonstrate that our depth and fow",
"models compare favorably with state-of-the-art unsupervised methods.",
"1",
"Introduction",
"Single-view depth prediction and optical flow estimationaretwo fundamental",
"problemsincomputervision.Whilethetwotasks aim to recoverhighlycorre",
"lated information fromthe scene (i.e.,the scenestructure and the dense motion",
"field between consecutive frames),existing efforts typically study each problem",
"inisolation.Inthispaper,wedemonstratethebenefitsofexploringthegeometric",
"relationship between depth,camera motion, and flow for unsupervised learning",
"ofdepthand flow estimationmodels."
] |
{
"arxiv_id": "1809.01649",
"language": "en",
"page_idx": 0
}
|
||
1809.01649-en-1
|
1809.01649-en-1.jpg
|
2
Y. Zou, Z. Luo, and J.-B. Huang
<FigureHere>
<FigureHere>
sequences
Pixelwise ground truth
Fig.2: Supervised v.s. unsupervised learning. Supervised learning of
depth or flow networks requires large amount of training data with pixelwise
ground truth annotations, which are difficult to acquire in real scenes. In
contrast, our work leverages the readily available unlabeled video sequences to
jointly train the depth and flow models.
With the rapid development of deep convolutional neural networks (CNNs), numerous approaches have been proposed to tackle dense prediction problems in an end-to-end manner. However, supervised training CNN for such tasks often involves in constructing large-scale, diverse datasets with dense pixelwise ground truth labels. Collecting such densely labeled datasets in real-world requires significant amounts of human efforts and is prone to error. Existing efforts of RGB-D dataset construction [18,45,53,54] often have limited scope (e.g., in terms of locations, scenes, and objects), and hence are lack of diversity. For optical flow, dense motion annotations are even more difficult to acquire [37]. Consequently, existing CNN-based methods rely on synthetic datasets for training the models [5,12,16,24]. These synthetic datasets, however, do not capture the complexity of motion blur, occlusion, and natural image statistics from real scenes. The trained models usually do not generalize well to unseen scenes without fine-tuning on sufficient ground truth data in a new visual domain.
Several work [17,21,28] have been proposed to capitalize on large-scale realworld videos to train the CNNs in the unsupervised setting. The main idea lies to exploit the brightness constancy and spatial smoothness assumptions of flow fields or disparity maps as supervisory signals. These assumptions, however, often do not hold at motion boundaries and hence makes the training unstable.
Many recent efforts [59,60,65,73] explore the geometric relationship between the two problems. With the estimated depth and camera pose, these methods can produce dense optical flow by backprojecting the 3D scene flow induced from camera ego-motion. However, these methods implicitly assume perfect depth and camera pose estimation when “synthesizing”the optical flow. The errors in either depth or camera pose estimation inevitably produce inaccurate flow predictions.
In this paper, we present a technique for jointly learning a single-view depth estimation model and a flow prediction model using unlabeled videos as shown in Figure 2. Our key observation is that the predictions from depth, pose, and optical flow should be consistent with each other. By exploiting this geometry cue, we present a novel cross-task consistency loss that provides additional supervisory signals for training both networks. We validate the effectiveness of
|
[
[
0.07489013671875,
0.0138092041015625,
0.1014404296875,
0.0138092041015625,
0.1014404296875,
0.033294677734375,
0.07489013671875,
0.033294677734375
],
[
0.158203125,
0.01543426513671875,
0.48193359375,
0.017059326171875,
0.48193359375,
0.03167724609375,
0.158203125,
0.030059814453125
],
[
0.1461181640625,
0.1705322265625,
0.39013671875,
0.1705322265625,
0.39013671875,
0.1884765625,
0.1461181640625,
0.1884765625
],
[
0.54345703125,
0.1697998046875,
0.8935546875,
0.17138671875,
0.8935546875,
0.1893310546875,
0.54345703125,
0.1868896484375
],
[
0.08331298828125,
0.20556640625,
0.85986328125,
0.20556640625,
0.85986328125,
0.2200927734375,
0.08331298828125,
0.2200927734375
],
[
0.0821533203125,
0.225830078125,
0.89013671875,
0.225830078125,
0.89013671875,
0.240478515625,
0.0821533203125,
0.240478515625
],
[
0.0772705078125,
0.2445068359375,
0.8525390625,
0.2420654296875,
0.8525390625,
0.260009765625,
0.0772705078125,
0.262451171875
],
[
0.0784912109375,
0.26318359375,
0.90576171875,
0.26318359375,
0.90576171875,
0.281005859375,
0.0784912109375,
0.281005859375
],
[
0.0772705078125,
0.28271484375,
0.50244140625,
0.281982421875,
0.50244140625,
0.2998046875,
0.0772705078125,
0.300537109375
],
[
0.11956787109375,
0.344482421875,
0.91162109375,
0.344482421875,
0.91162109375,
0.359130859375,
0.11956787109375,
0.359130859375
],
[
0.0772705078125,
0.363037109375,
0.91796875,
0.3623046875,
0.91796875,
0.380126953125,
0.0772705078125,
0.381103515625
],
[
0.0809326171875,
0.382568359375,
0.91162109375,
0.382568359375,
0.91162109375,
0.397216796875,
0.0809326171875,
0.397216796875
],
[
0.0809326171875,
0.40380859375,
0.91650390625,
0.40380859375,
0.91650390625,
0.41845703125,
0.0809326171875,
0.41845703125
],
[
0.0784912109375,
0.421630859375,
0.91552734375,
0.4208984375,
0.91552734375,
0.437744140625,
0.0784912109375,
0.439453125
],
[
0.0821533203125,
0.44189453125,
0.91650390625,
0.44189453125,
0.91650390625,
0.45654296875,
0.0821533203125,
0.45654296875
],
[
0.0772705078125,
0.458984375,
0.91796875,
0.459716796875,
0.91796875,
0.4775390625,
0.0772705078125,
0.476806640625
],
[
0.0809326171875,
0.481689453125,
0.91552734375,
0.481689453125,
0.91552734375,
0.496337890625,
0.0809326171875,
0.496337890625
],
[
0.0821533203125,
0.50048828125,
0.9130859375,
0.50048828125,
0.9130859375,
0.51513671875,
0.0821533203125,
0.51513671875
],
[
0.0821533203125,
0.5205078125,
0.9130859375,
0.5205078125,
0.9130859375,
0.53515625,
0.0821533203125,
0.53515625
],
[
0.0784912109375,
0.53955078125,
0.91552734375,
0.5400390625,
0.91552734375,
0.5546875,
0.0784912109375,
0.55419921875
],
[
0.0809326171875,
0.55908203125,
0.91650390625,
0.55908203125,
0.91650390625,
0.57373046875,
0.0809326171875,
0.57373046875
],
[
0.0784912109375,
0.57763671875,
0.91552734375,
0.57763671875,
0.91552734375,
0.59228515625,
0.0784912109375,
0.59228515625
],
[
0.0809326171875,
0.59765625,
0.8310546875,
0.59765625,
0.8310546875,
0.61181640625,
0.0809326171875,
0.61181640625
],
[
0.11834716796875,
0.6181640625,
0.91162109375,
0.6181640625,
0.91162109375,
0.6328125,
0.11834716796875,
0.6328125
],
[
0.0821533203125,
0.638671875,
0.91650390625,
0.638671875,
0.91650390625,
0.65234375,
0.0821533203125,
0.65234375
],
[
0.0809326171875,
0.6572265625,
0.91796875,
0.6572265625,
0.91796875,
0.671875,
0.0809326171875,
0.671875
],
[
0.0784912109375,
0.67333984375,
0.91552734375,
0.67431640625,
0.91552734375,
0.69189453125,
0.0784912109375,
0.69140625
],
[
0.0809326171875,
0.6953125,
0.91552734375,
0.6953125,
0.91552734375,
0.7099609375,
0.0809326171875,
0.7099609375
],
[
0.11474609375,
0.7138671875,
0.91796875,
0.71484375,
0.91796875,
0.73291015625,
0.11474609375,
0.73193359375
],
[
0.0809326171875,
0.7353515625,
0.91552734375,
0.73583984375,
0.91552734375,
0.75048828125,
0.0809326171875,
0.75
],
[
0.0809326171875,
0.75537109375,
0.91552734375,
0.75537109375,
0.91552734375,
0.77001953125,
0.0809326171875,
0.77001953125
],
[
0.0809326171875,
0.77490234375,
0.91552734375,
0.77490234375,
0.91552734375,
0.78857421875,
0.0809326171875,
0.78857421875
],
[
0.0821533203125,
0.79443359375,
0.91650390625,
0.79443359375,
0.91650390625,
0.80908203125,
0.0821533203125,
0.80908203125
],
[
0.0809326171875,
0.81396484375,
0.91552734375,
0.81396484375,
0.91552734375,
0.82861328125,
0.0809326171875,
0.82861328125
],
[
0.11712646484375,
0.8349609375,
0.91552734375,
0.8349609375,
0.91552734375,
0.849609375,
0.11712646484375,
0.849609375
],
[
0.0809326171875,
0.85302734375,
0.92041015625,
0.85302734375,
0.92041015625,
0.8701171875,
0.0809326171875,
0.8701171875
],
[
0.0772705078125,
0.8701171875,
0.91796875,
0.87158203125,
0.91796875,
0.8896484375,
0.0772705078125,
0.8876953125
],
[
0.0821533203125,
0.89208984375,
0.91552734375,
0.89208984375,
0.91552734375,
0.90673828125,
0.0821533203125,
0.90673828125
],
[
0.0809326171875,
0.912109375,
0.9130859375,
0.912109375,
0.9130859375,
0.9267578125,
0.0809326171875,
0.9267578125
],
[
0.0809326171875,
0.931640625,
0.92041015625,
0.931640625,
0.92041015625,
0.9462890625,
0.0809326171875,
0.9462890625
]
] |
[
"2",
"Y.Zou,Z.Luo,and J.-B.Huang",
"Pixelwiseground truth",
"Unlabeledvideosequences (ours)",
"Fig.2: Supervised v.s.unsupervised learning.Supervised learning of",
"depth or flow networks requires large amount of training data withpixelwise",
"ground truth annotations, which are difficult to acquire in real scenes. In",
"contrast, our work leverages the readily available unlabeled video sequences to",
"jointlytrain thedepth and flowmodels.",
"With therapid developmentof deepconvolutional neural networks (CNNs)",
"numerous approaches have been proposed to tackle dense prediction problems",
"in an end-to-end manner. However, supervised training CNN for such tasks of",
"ten involves in constructing large-scale, diverse datasets with dense pixelwise",
"ground truth labels. Collecting such densely labeled datasets in real-world re-",
"quires significant amounts of human efforts and is prone to error.Existing efforts",
"of RGB-D dataset construction [18,45,53,54] often have limited scope (e.g., in",
"terms of locations, scenes, and objects),and hence are lack of diversity. For",
"optical flow, dense motion annotations are even more difficult to acguire [37l",
"Consequently, existing CNN-based methods rely on synthetic datasets for train",
"ingthemodels[5,12,16,24].These syntheticdatasets,however,do not capture",
"the complexity of motion blur, occlusion, and natural image statistics from real",
"scenes. The trained models usually do not generalize well to unseen scenes with-",
"out fine-tuning on sufficient ground truth data in a new visual domain.",
"Several work[17.21,28lhavebeenproposed to capitalizeonlarge-scalereal",
"world videos to train the CNNs in the unsupervised setting.The main idea",
"lies to exploit the brightness constancy and spatial smoothness assumptions of",
"flow fields or disparity maps as supervisory signals. These assumptions, however,",
"often do not hold at motion boundaries and hence makes the training unstable.",
"Many recent efforts[59,60,65,73] explore the geometric relationship between",
"thetwoproblems.Withtheestimated depth and camera pose,thesemethods",
"canproducedenseopticalflowbybackprojectingthe3Dscenefowinducedfrom",
"cameraego-motion.However,thesemethodsimplicitlyassumeperfectdepthand",
"cameraposeestimationwhen“\"synthesizing\"theopticalflow.Theerrorsineither",
"depthorcameraposeestimationinevitablyproduceinaccurateflowpredictions",
"In thispaper,wepresenta techniquefor jointlylearning asingle-view depth",
"estimation model and a flow prediction model using unlabeled videos as shown",
"in Figure2.Our key observation is that thepredictions from depth,pose, and",
"opticalflowshould be consistent with eachother.By exploiting this geometry",
"cue,we present a novel cross-task consistency loss thatprovides additional su-",
"pervisory signals for training both networks.We validate the effectiveness of"
] |
{
"arxiv_id": "1809.01649",
"language": "en",
"page_idx": 1
}
|
||
1809.01649-en-2
|
1809.01649-en-2.jpg
|
3
Unsupervised Joint Learning using Cross-Task Consistency
the proposed approach through extensive experiments on several benchmark datasets. Experimental results show that our joint training method significantly $\begin{array}{c}\text{improves the performance of both models (Figure 1). The proposed depth and}\\\end{array}$ flow models compare favorably with state-of-the-art unsupervised methods
We make the following contributions. (1) We propose an unsupervised learning framework to simultaneously train a depth prediction network and an optical flow network. We achieve this by introducing a cross-task consistency loss that enforces geometric consistency. (2) We show that through the proposed unsupervised training our depth and flow models compare favorably with existing unsupervised algorithms and achieve competitive performance with supervised methods on several benchmark datasets. (3) We release the source code and pretrained models to facilitate future research: http://yuliang.vision/DF-Net/
### 2 $\textbf{Related Work}$
Supervised learning of depth and flow. Supervised learning using CNNs has emerged to be an effective approach for depth and flow estimation to avoid hand-crafted objective functions and computationally expensive optimization at test time. The availability of RGB-D datasets and deep learning leads to a line of work on single-view depth estimation [13,14,35,38,62,72]. While promising results have been shown, these methods rely on the absolute ground truth depth maps. These depth maps, however, are expensive and difficult to collect. Some efforts [8,74] have been made to relax the difficulty of collecting absolute depth by exploring learning from $relative/ordinal$ depth annotations. Recent work also explores gathering training datasets from web videos [7] or Internet photos [36] using structure-from-motion and multi-view stereo algorithms.
Compared to ground truth depth datasets, constructing optical flow datasets of diverse scenes in real-world is even more challenging. Consequently, existing approaches [12,26,47] typically rely on synthetic datasets [5,12] for training. Due to the limited scalability of constructing diverse, high-quality training data, fully supervised approaches often require fine-tuning on sufficient ground truth labels in new visual domains to perform well. In contrast, our approach leverages the readily available real-world videos to jointly train the depth and flow models. The ability to learn from unlabeled data enables unsupervised pre-training for domains with limited amounts of ground truth data.
Self-supervised learning of depth and flow. To alleviate the dependency on large-scale annotated datasets, several works have been proposed to exploit the classical assumptions of brightness constancy and spatial smoothness on the disparity map or the flow field [17,21,28,43,71]. The core idea is to treat the estimated depth and flow as latent layers and use them to differentiably warp the source frame to the target frame, where the source and target frames can either be the stereo pair or two consecutive frames in a video sequence. A photometric loss between the synthesized frame and the target frame can then serve as an unsupervised proxy loss to train the network. Using photometric loss
|
[
[
0.260986328125,
0.01543426513671875,
0.8359375,
0.01543426513671875,
0.8359375,
0.030059814453125,
0.260986328125,
0.030059814453125
],
[
0.8974609375,
0.0138092041015625,
0.923828125,
0.0138092041015625,
0.923828125,
0.033294677734375,
0.8974609375,
0.033294677734375
],
[
0.0809326171875,
0.056854248046875,
0.91650390625,
0.056854248046875,
0.91650390625,
0.07470703125,
0.0809326171875,
0.07470703125
],
[
0.0821533203125,
0.07879638671875,
0.9130859375,
0.07879638671875,
0.9130859375,
0.09344482421875,
0.0821533203125,
0.09344482421875
],
[
0.0809326171875,
0.0982666015625,
0.91650390625,
0.0982666015625,
0.91650390625,
0.1129150390625,
0.0809326171875,
0.1129150390625
],
[
0.0809326171875,
0.11700439453125,
0.87548828125,
0.11700439453125,
0.87548828125,
0.131591796875,
0.0809326171875,
0.131591796875
],
[
0.11956787109375,
0.1373291015625,
0.91162109375,
0.1373291015625,
0.91162109375,
0.15185546875,
0.11956787109375,
0.15185546875
],
[
0.0821533203125,
0.15673828125,
0.91650390625,
0.15673828125,
0.91650390625,
0.17138671875,
0.0821533203125,
0.17138671875
],
[
0.0821533203125,
0.1754150390625,
0.91650390625,
0.1754150390625,
0.91650390625,
0.1900634765625,
0.0821533203125,
0.1900634765625
],
[
0.0784912109375,
0.194091796875,
0.91552734375,
0.194091796875,
0.91552734375,
0.2120361328125,
0.0784912109375,
0.2120361328125
],
[
0.0772705078125,
0.2144775390625,
0.91650390625,
0.2144775390625,
0.91650390625,
0.2315673828125,
0.0772705078125,
0.2315673828125
],
[
0.0809326171875,
0.2340087890625,
0.91552734375,
0.2340087890625,
0.91552734375,
0.24853515625,
0.0809326171875,
0.24853515625
],
[
0.0809326171875,
0.255126953125,
0.9130859375,
0.255126953125,
0.9130859375,
0.26806640625,
0.0809326171875,
0.26806640625
],
[
0.0809326171875,
0.273681640625,
0.91162109375,
0.273681640625,
0.91162109375,
0.288330078125,
0.0809326171875,
0.288330078125
],
[
0.0809326171875,
0.324951171875,
0.1038818359375,
0.324951171875,
0.1038818359375,
0.34033203125,
0.0809326171875,
0.34033203125
],
[
0.1243896484375,
0.3232421875,
0.32861328125,
0.3232421875,
0.32861328125,
0.341064453125,
0.1243896484375,
0.341064453125
],
[
0.0809326171875,
0.375244140625,
0.91650390625,
0.375244140625,
0.91650390625,
0.39306640625,
0.0809326171875,
0.39306640625
],
[
0.0772705078125,
0.39404296875,
0.91796875,
0.3955078125,
0.91796875,
0.41357421875,
0.0772705078125,
0.411865234375
],
[
0.0809326171875,
0.416015625,
0.91796875,
0.416015625,
0.91796875,
0.4306640625,
0.0809326171875,
0.4306640625
],
[
0.0784912109375,
0.4345703125,
0.91650390625,
0.436279296875,
0.91650390625,
0.450927734375,
0.0784912109375,
0.44921875
],
[
0.0809326171875,
0.454833984375,
0.91552734375,
0.455810546875,
0.91552734375,
0.470458984375,
0.0809326171875,
0.469482421875
],
[
0.0809326171875,
0.474365234375,
0.91552734375,
0.474365234375,
0.91552734375,
0.489013671875,
0.0809326171875,
0.489013671875
],
[
0.0809326171875,
0.49462890625,
0.91650390625,
0.49462890625,
0.91650390625,
0.50927734375,
0.0809326171875,
0.50927734375
],
[
0.0821533203125,
0.51416015625,
0.91650390625,
0.51416015625,
0.91650390625,
0.52880859375,
0.0821533203125,
0.52880859375
],
[
0.0821533203125,
0.53369140625,
0.91552734375,
0.53369140625,
0.91552734375,
0.54833984375,
0.0821533203125,
0.54833984375
],
[
0.0784912109375,
0.5498046875,
0.91552734375,
0.5517578125,
0.91552734375,
0.5693359375,
0.0784912109375,
0.56689453125
],
[
0.0821533203125,
0.57275390625,
0.7353515625,
0.57275390625,
0.7353515625,
0.5859375,
0.0821533203125,
0.5859375
],
[
0.11712646484375,
0.5908203125,
0.92041015625,
0.5908203125,
0.92041015625,
0.607421875,
0.11712646484375,
0.607421875
],
[
0.0772705078125,
0.609375,
0.91796875,
0.60986328125,
0.91796875,
0.6279296875,
0.0772705078125,
0.62646484375
],
[
0.0784912109375,
0.62939453125,
0.91796875,
0.62939453125,
0.91796875,
0.6474609375,
0.0784912109375,
0.6474609375
],
[
0.0809326171875,
0.65087890625,
0.9130859375,
0.65087890625,
0.9130859375,
0.66552734375,
0.0809326171875,
0.66552734375
],
[
0.0809326171875,
0.66943359375,
0.91650390625,
0.66943359375,
0.91650390625,
0.68408203125,
0.0809326171875,
0.68408203125
],
[
0.0809326171875,
0.68896484375,
0.91650390625,
0.68896484375,
0.91650390625,
0.70361328125,
0.0809326171875,
0.70361328125
],
[
0.0809326171875,
0.70751953125,
0.9130859375,
0.70751953125,
0.9130859375,
0.72216796875,
0.0809326171875,
0.72216796875
],
[
0.0809326171875,
0.7255859375,
0.91650390625,
0.7255859375,
0.91650390625,
0.7431640625,
0.0809326171875,
0.7431640625
],
[
0.0809326171875,
0.74755859375,
0.63623046875,
0.74755859375,
0.63623046875,
0.76220703125,
0.0809326171875,
0.76220703125
],
[
0.11712646484375,
0.77587890625,
0.91552734375,
0.77587890625,
0.91552734375,
0.79052734375,
0.11712646484375,
0.79052734375
],
[
0.0784912109375,
0.79443359375,
0.91650390625,
0.79443359375,
0.91650390625,
0.8125,
0.0784912109375,
0.8125
],
[
0.0821533203125,
0.8154296875,
0.91552734375,
0.8154296875,
0.91552734375,
0.82958984375,
0.0821533203125,
0.82958984375
],
[
0.0821533203125,
0.8349609375,
0.91552734375,
0.8349609375,
0.91552734375,
0.849609375,
0.0821533203125,
0.849609375
],
[
0.0821533203125,
0.85400390625,
0.9130859375,
0.85400390625,
0.9130859375,
0.8681640625,
0.0821533203125,
0.8681640625
],
[
0.0809326171875,
0.873046875,
0.91650390625,
0.873046875,
0.91650390625,
0.8876953125,
0.0809326171875,
0.8876953125
],
[
0.0772705078125,
0.89208984375,
0.91650390625,
0.89013671875,
0.91650390625,
0.90673828125,
0.0772705078125,
0.9091796875
],
[
0.0809326171875,
0.912109375,
0.91650390625,
0.912109375,
0.91650390625,
0.9267578125,
0.0809326171875,
0.9267578125
],
[
0.07611083984375,
0.93115234375,
0.91796875,
0.93017578125,
0.91796875,
0.947265625,
0.07611083984375,
0.94873046875
]
] |
[
"Unsupervised JointLearning using Cross-Task Consistency",
"3",
"the proposed approach through extensive experiments on several benchmark",
"datasets.Experimental results show that our joint training method significantly",
"improves theperformance of bothmodels (Figure 1).The proposed depth and",
"flow models comparefavorably with state-of-the-art unsupervised methods.",
"Wemakethefollowingcontributions.(1)Weproposeanunsupervisedlearn",
"ingframework to simultaneouslytrainadepthpredictionnetwork and anoptical",
"flow network. We achieve this by introducing a cross-task consistency loss that",
"enforces geometric consistency. (2) We show that through the proposed unsu-",
"pervised training our depth and flow models compare favorably with existing",
"unsupervised algorithms and achieve competitive performance with supervised",
"methodsonseveralbenchmarkdatasets.(3)Wereleasethesourcecodeandpre",
"trained models to facilitatefuture research:http://yuliang.vision/DF-Net/",
"2",
"Related Work",
"Supervised learning of depth and flow. Supervised learning using CNNs",
"has emerged to be an effective approach for depth and flow estimation to avoid",
"hand-crafted objectivefunctions and computationallyexpensive optimization at",
"test time.The availability of RGB-D datasets and deeplearning leads to a line",
"of work on single-view depth estimation[13,14,35,38,62,72].Whilepromising",
"results have been shown, these methods rely on the absolute ground truth depth",
"maps.These depth maps,however, are expensive and difficult to collect. Some",
"efforts8,74 have been made torelaxthedifficulty of collecting absolutedepth",
"byexploring learningfrom relative/ordinal depthannotations.Recent work also",
"explores gathering training datasets from web videos [7] or Internet photos [36]",
"usingstructure-from-motion andmulti-viewstereo algorithms",
"Compared to ground truth depth datasets, constructing optical flow datasets",
"of diverse scenes in real-world is even more challenging.Consequently, existing",
"approaches [12,26,47] typically rely on synthetic datasets [5,12] for training. Due",
"to the limited scalability of constructing diverse, high-quality training data,fully",
"supervised approaches often require fine-tuning on suficient ground truth labels",
"in new visual domains to perform well. In contrast, our approach leverages the",
"readily available real-world videos to jointly train the depth and flow models",
"The ability to learn from unlabeled data enables unsupervised pre-training for",
"domains with limited amounts of ground truth data.",
"Self-supervised learningof depthandflow.Toalleviatethedependency",
"on large-scale annotated datasets, several works have been proposed to exploit",
"the classical assumptions of brightness constancy and spatial smoothness on",
"the disparity map or the flow field17,21,28,43.71l.The core idea is to treat",
"the estimated depth and flow as latent layers and use them to differentiably",
"warp the sourceframetothe targetframe,wherethe sourceandtarget frames",
"can either bethe stereopair or two consecutive frames ina video sequence.A",
"photometric loss between thesynthesized frame and thetarget frame can then",
"serve as an unsupervised proxy loss to train the network.Using photometric loss"
] |
{
"arxiv_id": "1809.01649",
"language": "en",
"page_idx": 2
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.