id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
1707.05165-en-0
|
1707.05165-en-0.jpg
|
# A Comprehensive Implementation of Conceptual Space$s^{\star }$
Lucas Bechberger$^{\star\star}$ and Kai-Uwe Kühnberger
Institute of Cognitive Science, Osnabrück University, Osnabrück, Germany
lucas. bechberger@ uni- osnabrueck. de, kai- uwe. kuehnberger@ uni- osnabrueck. de
$\mathbf{Abstract.~The~highly~influential~framework~of~conceptual~spaces~pro-}$ vides a geometric way of representing knowledge. Instances are represented by points and concepts are represented by regions in a (poten-1) tially) high-dimensional space. Based on our recent formalization, we present a comprehensive implementation of the conceptual spaces framework that is not only capable of representing concepts with inter-domain correlations, but that also offers a variety of operations on these concepts.
## 1 Introduction
One common criticism of symbolic AI approaches is that the symbols they operate on do not contain any meaning: For the system, they are just arbitrary tokens that can be manipulated in some way. This lack of inherent meaning in abstract symbols is called the symbol grounding problem [18].One approach towards solving this problem is to devise a grounding mechanism that connects abstract symbols to the real world, i.e., to perception and action.
The cognitive framework of conceptual spaces [16,17] attempts to bridge this gap between symbolic and subsymbolic AI by proposing an intermediate conceptual layer based on geometric representations. A conceptual space is spanned by a number of quality dimensions that are based on perception and/or subsymbolic processing. Regions in this space correspond to concepts and can be referred to as abstract symbols.
The framework of conceptual spaces has been highly influential in the last 15 years within cognitive science and cognitive linguistics [14,15,27]. It has also sparked considerable research in various subfelds of artificial intelligence, ranging from robotics and computer vision [11,12] over the semantic web [2] to plausible reasoning [13,24].
Most practical implementations of the conceptual spaces framework are rather ad-hoc and tailored towards a specific application. They tend to ignore important aspects of the framework and should thus be regarded as only partial implementations of the framework. Furthermore, these implementations are usually not publicly accessible, which greatly reduces their value for the research community.
$^{\star}$ The final paper is published at http://ceur-ws.org/Vol-2090/
$^{\star\star}$ Corresponding author, ORCID: 0000-0002-1962-1777
|
[
[
0.2197265625,
0.143310546875,
0.78515625,
0.14453125,
0.78515625,
0.1622314453125,
0.2197265625,
0.1610107421875
],
[
0.45751953125,
0.1673583984375,
0.5458984375,
0.1673583984375,
0.5458984375,
0.1849365234375,
0.45751953125,
0.1849365234375
],
[
0.338134765625,
0.2139892578125,
0.66650390625,
0.2139892578125,
0.66650390625,
0.2291259765625,
0.338134765625,
0.2291259765625
],
[
0.253173828125,
0.2412109375,
0.7490234375,
0.2412109375,
0.7490234375,
0.255126953125,
0.253173828125,
0.255126953125
],
[
0.223876953125,
0.255615234375,
0.78125,
0.255615234375,
0.78125,
0.26953125,
0.223876953125,
0.26953125
],
[
0.021240234375,
0.268310546875,
0.06451416015625,
0.268310546875,
0.06451416015625,
0.416748046875,
0.021240234375,
0.416748046875
],
[
0.2646484375,
0.302978515625,
0.73779296875,
0.3037109375,
0.73779296875,
0.318115234375,
0.2646484375,
0.317626953125
],
[
0.265625,
0.317626953125,
0.736328125,
0.317626953125,
0.736328125,
0.330810546875,
0.265625,
0.330810546875
],
[
0.2646484375,
0.333251953125,
0.73779296875,
0.333251953125,
0.73779296875,
0.3466796875,
0.2646484375,
0.3466796875
],
[
0.265625,
0.347900390625,
0.73779296875,
0.347900390625,
0.73779296875,
0.35791015625,
0.265625,
0.35791015625
],
[
0.2646484375,
0.35986328125,
0.736328125,
0.35986328125,
0.736328125,
0.373779296875,
0.2646484375,
0.373779296875
],
[
0.265625,
0.373779296875,
0.73876953125,
0.373779296875,
0.73876953125,
0.386962890625,
0.265625,
0.386962890625
],
[
0.2646484375,
0.386962890625,
0.73779296875,
0.386962890625,
0.73779296875,
0.40087890625,
0.2646484375,
0.40087890625
],
[
0.0228729248046875,
0.40771484375,
0.062103271484375,
0.40771484375,
0.062103271484375,
0.49365234375,
0.0228729248046875,
0.49365234375
],
[
0.2254638671875,
0.425537109375,
0.2344970703125,
0.434326171875,
0.2254638671875,
0.43994140625,
0.2156982421875,
0.431884765625
],
[
0.25,
0.4267578125,
0.373291015625,
0.4267578125,
0.373291015625,
0.43994140625,
0.25,
0.43994140625
],
[
0.21728515625,
0.455810546875,
0.78369140625,
0.45703125,
0.78369140625,
0.4716796875,
0.21728515625,
0.47021484375
],
[
0.2181396484375,
0.472900390625,
0.78515625,
0.472900390625,
0.78515625,
0.4873046875,
0.2181396484375,
0.4873046875
],
[
0.0228729248046875,
0.486083984375,
0.05963134765625,
0.486083984375,
0.05963134765625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.21728515625,
0.486083984375,
0.78515625,
0.4873046875,
0.78515625,
0.501953125,
0.21728515625,
0.50048828125
],
[
0.2181396484375,
0.50244140625,
0.7861328125,
0.50244140625,
0.7861328125,
0.5166015625,
0.2181396484375,
0.5166015625
],
[
0.2181396484375,
0.51806640625,
0.7861328125,
0.51806640625,
0.7861328125,
0.5322265625,
0.2181396484375,
0.5322265625
],
[
0.21728515625,
0.5322265625,
0.685546875,
0.53271484375,
0.685546875,
0.54736328125,
0.21728515625,
0.546875
],
[
0.2442626953125,
0.548828125,
0.78515625,
0.548828125,
0.78515625,
0.5625,
0.2442626953125,
0.5625
],
[
0.2181396484375,
0.56298828125,
0.78173828125,
0.56298828125,
0.78173828125,
0.57763671875,
0.2181396484375,
0.57763671875
],
[
0.2197265625,
0.5791015625,
0.7861328125,
0.5791015625,
0.7861328125,
0.59228515625,
0.2197265625,
0.59228515625
],
[
0.2181396484375,
0.59326171875,
0.78369140625,
0.59326171875,
0.78369140625,
0.60693359375,
0.2181396484375,
0.60693359375
],
[
0.2181396484375,
0.609375,
0.7861328125,
0.609375,
0.7861328125,
0.62255859375,
0.2181396484375,
0.62255859375
],
[
0.2181396484375,
0.625,
0.4453125,
0.625,
0.4453125,
0.63818359375,
0.2181396484375,
0.63818359375
],
[
0.2418212890625,
0.63818359375,
0.78515625,
0.63720703125,
0.78515625,
0.6513671875,
0.2418212890625,
0.65283203125
],
[
0.2197265625,
0.65380859375,
0.7861328125,
0.65380859375,
0.7861328125,
0.66748046875,
0.2197265625,
0.66748046875
],
[
0.2181396484375,
0.66943359375,
0.78369140625,
0.66943359375,
0.78369140625,
0.68310546875,
0.2181396484375,
0.68310546875
],
[
0.21728515625,
0.6845703125,
0.78369140625,
0.6845703125,
0.78369140625,
0.69775390625,
0.21728515625,
0.69775390625
],
[
0.2181396484375,
0.69970703125,
0.3798828125,
0.69970703125,
0.3798828125,
0.71337890625,
0.2181396484375,
0.71337890625
],
[
0.2442626953125,
0.72900390625,
0.78759765625,
0.72900390625,
0.78759765625,
0.74365234375,
0.2442626953125,
0.74365234375
],
[
0.2181396484375,
0.744140625,
0.78515625,
0.744140625,
0.78515625,
0.75830078125,
0.2181396484375,
0.75830078125
],
[
0.2181396484375,
0.76025390625,
0.78369140625,
0.76025390625,
0.78369140625,
0.7734375,
0.2181396484375,
0.7734375
],
[
0.2181396484375,
0.77587890625,
0.78515625,
0.77587890625,
0.78515625,
0.7890625,
0.2181396484375,
0.7890625
],
[
0.21728515625,
0.7890625,
0.78369140625,
0.79052734375,
0.78369140625,
0.8046875,
0.21728515625,
0.8037109375
],
[
0.221435546875,
0.8125,
0.6611328125,
0.8125,
0.6611328125,
0.8271484375,
0.221435546875,
0.8271484375
],
[
0.214111328125,
0.82373046875,
0.58837890625,
0.82470703125,
0.58837890625,
0.84228515625,
0.214111328125,
0.84130859375
]
] |
[
"A Comprehensive Implementation of Conceptual",
"Spaces*",
"Lucas Bechberger** and Kai-Uwe Kihnberger",
"Institute of Cognitive Science, Osnabrick University, Osnabrick, Germany",
"lucas.bechberger@uni-osnabrueck.de,kai-uwe.kuehnberger@uni-osnabrueck.de",
"8",
"Abstract. The highly influential framework of conceptual spaces pro-",
"vides a geometric way of representing knowledge. Instances are repre",
"sented by points and concepts are represented by regions in a (poten-",
"tially) high-dimensional space.Based on our recent formalization,we",
"present a comprehensive implementation of the conceptual spaces frame",
"work that is not only capable of representing concepts with inter-domain",
"correlations, but that also offers a variety of operations on these concepts.",
"[I's]",
"7",
"Introduction",
"One common criticism of symbolic AI approaches is that the symbols they op-",
"erate on do not contain any meaning: For the system, they are just arbitrary",
"CAS9ISO'LOLI:AIXIE",
"tokens that can be manipulated in some way. This lack of inherent meaning in",
"abstract symbols is called the symbol grounding problem [18]. One approach",
"towards solving this problem is to devise a grounding mechanism that connects",
"abstract symbols to the real world, i.e., to perception and action.",
"The cognitive framework of conceptual spaces [16,17] attempts to bridge this",
"gap between symbolic and subsymbolic AI by proposing an intermediate con",
"ceptual layer based on geometric representations. A conceptual space is spanned",
"by a number of quality dimensions that are based on perception and/or sub-",
"symbolic processing. Regions in this space correspond to concepts and can be",
"referred to as abstract symbols.",
"The framework of conceptual spaces has been highly influential in the last",
"15 years within cognitive science and cognitive linguistics [14,15,27]. It has also",
"sparked considerable research in various subfields of artificial intelligence, rang-",
"ing from robotics and computer vision [11,12] over the semantic web [2] to plau-",
"sible reasoning [13,24].",
"Most practical implementations of the conceptual spaces framework are rather",
"ad-hoc and tailored towards a specific application. They tend to ignore important",
"aspects of the framework and should thus be regarded as only partial implemen-",
"tations of the framework. Furthermore, these implementations are usually not",
"publicly accessible, which greatly reduces their value for the research community",
"* The final paper is published at http://ceur-ws.org/Vo1-2090/",
"* Corresponding author, ORCID: 0000-0002-1962-1777"
] |
{
"arxiv_id": "1707.05165",
"language": "en",
"page_idx": 0
}
|
||
1707.05165-en-1
|
1707.05165-en-1.jpg
|
In this paper, we present a thorough and comprehensive implementation of the conceptual spaces framework which is based on our formalization reported in [7] and [8]. Its source code is publicly available on GitHub to researchers anywhere in the world. Instead of investing a considerable amount of time into developing their own implementation, researchers can use our implementation off-the-shelf and focus on their specific application scenario
The remainder of this paper is structured as follows: Section 2 introduces the framework of conceptual spaces and our formalization. In Section 3, we give an overview of our implementation and in Section 4 we illustrate its usage. Section 5 summarizes related work and Section 6 concludes the paper.
# 2 Conceptual Spaces
This section presents the cognitive framework of conceptual spaces as described
in [16] and summarizes our formalization as reported in [7] and [8]
2.1 Dimensions, Domains, and Distance
A conceptual space is spanned by a set $D$ of so-called“quality dimensions". Each of these dimensions $d\in D$ represents a way in which two stimuli can be judged to be similar or different. Examples for quality dimensions include temperature, weight, time, pitch, and hue. The distance between two points $x$ and $y$ with respect to a dimension $d$ is denoted as $|x_d-y_d|.$
A domain $\delta\subseteq D$ is a set of dimensions that inherently belong together Different perceptual modalities (like color, shape, or taste) are represented by different domains. The color domain for instance consists of the three dimensions hue, saturation, and brightness. Distance within a domain $\delta$ is measured by the weighted Euclidean metric $d_E.$
The overall conceptual space $CS$ is defned as the product space of all dimensions. Distance within the overall conceptual space is measured by the weighted Manhattan metric $d_{M}$ of the intra-domain distances. This is supported by both psychological evidence [5,25] and mathematical considerations [3]. Let $\Delta$be the set of all domains in $CS.$ The combined distance $d_C^{\Delta}$ within $CS$ is defned as follows:
$$d_C^\Delta(x,y,W)=\sum\limits_{\delta\in\Delta}w_\delta\cdot\sqrt{\sum\limits_{d\in\delta}w_d\cdot|x_d-y_d|^2}$$
$\begin{array}{l}{\mathrm{The~parameter~}W=\langle W_{\Delta},\{W_{\delta}\}_{\delta\in\Delta}\rangle~contains~two~parts:~}W_{\Delta}&{\mathrm{is~the~set~of~pos}}\\{\mathrm{itive~domain~weights~}w_{\delta}~\mathrm{with~}\sum_{\delta\in\Delta}w_{\delta}=|\Delta|.~\mathrm{Moreover,~}W~\mathrm{contains~for~each}}\\{\mathrm{domain~}\delta\in\Delta~\mathrm{a~set~}W_{\delta}~\mathrm{of~positive~dimension~weights~}w_{d}~\mathrm{wi}}\end{array}$
The similarity of two points in a conceptual space is inversely related to their
distance. This can be written as follows:
$Sim( x, y) = e^{- c\cdot d( x, y) }$ with a constant $c>0$ and a given metric $d$
|
[
[
0.24267578125,
0.147705078125,
0.78759765625,
0.147705078125,
0.78759765625,
0.1622314453125,
0.24267578125,
0.1622314453125
],
[
0.2181396484375,
0.1622314453125,
0.78759765625,
0.1634521484375,
0.78759765625,
0.177978515625,
0.2181396484375,
0.1767578125
],
[
0.21728515625,
0.1787109375,
0.7861328125,
0.1787109375,
0.7861328125,
0.1925048828125,
0.21728515625,
0.1925048828125
],
[
0.2181396484375,
0.1932373046875,
0.7861328125,
0.1932373046875,
0.7861328125,
0.2083740234375,
0.2181396484375,
0.2083740234375
],
[
0.21728515625,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.222900390625,
0.21728515625,
0.222900390625
],
[
0.2181396484375,
0.2235107421875,
0.646484375,
0.2235107421875,
0.646484375,
0.2374267578125,
0.2181396484375,
0.2374267578125
],
[
0.24267578125,
0.238037109375,
0.7861328125,
0.238037109375,
0.7861328125,
0.25244140625,
0.24267578125,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.7861328125,
0.25390625,
0.7861328125,
0.268310546875,
0.2181396484375,
0.268310546875
],
[
0.2181396484375,
0.268310546875,
0.78369140625,
0.268310546875,
0.78369140625,
0.2822265625,
0.2181396484375,
0.2822265625
],
[
0.21728515625,
0.28271484375,
0.6640625,
0.2841796875,
0.6640625,
0.298583984375,
0.21728515625,
0.29736328125
],
[
0.221435546875,
0.32568359375,
0.233642578125,
0.32568359375,
0.233642578125,
0.337158203125,
0.221435546875,
0.337158203125
],
[
0.250732421875,
0.324462890625,
0.43310546875,
0.324462890625,
0.43310546875,
0.339111328125,
0.250732421875,
0.339111328125
],
[
0.2181396484375,
0.35595703125,
0.7861328125,
0.35595703125,
0.7861328125,
0.37060546875,
0.2181396484375,
0.37060546875
],
[
0.21728515625,
0.371826171875,
0.69677734375,
0.371826171875,
0.69677734375,
0.386474609375,
0.21728515625,
0.386474609375
],
[
0.21728515625,
0.410400390625,
0.55810546875,
0.410400390625,
0.55810546875,
0.4248046875,
0.21728515625,
0.4248046875
],
[
0.2181396484375,
0.436279296875,
0.78515625,
0.436279296875,
0.78515625,
0.45068359375,
0.2181396484375,
0.45068359375
],
[
0.21728515625,
0.45263671875,
0.7861328125,
0.45263671875,
0.7861328125,
0.466552734375,
0.21728515625,
0.466552734375
],
[
0.21728515625,
0.466552734375,
0.78369140625,
0.46728515625,
0.78369140625,
0.481689453125,
0.21728515625,
0.48095703125
],
[
0.2197265625,
0.481689453125,
0.78515625,
0.481689453125,
0.78515625,
0.49560546875,
0.2197265625,
0.49560546875
],
[
0.21728515625,
0.494384765625,
0.564453125,
0.49755859375,
0.564453125,
0.51220703125,
0.21728515625,
0.5087890625
],
[
0.2442626953125,
0.51220703125,
0.78369140625,
0.51220703125,
0.78369140625,
0.5263671875,
0.2442626953125,
0.5263671875
],
[
0.2181396484375,
0.52783203125,
0.78369140625,
0.52783203125,
0.78369140625,
0.54150390625,
0.2181396484375,
0.54150390625
],
[
0.2197265625,
0.54248046875,
0.7861328125,
0.54248046875,
0.7861328125,
0.55615234375,
0.2197265625,
0.55615234375
],
[
0.2181396484375,
0.55810546875,
0.78515625,
0.55810546875,
0.78515625,
0.5712890625,
0.2181396484375,
0.5712890625
],
[
0.21728515625,
0.5712890625,
0.438720703125,
0.57275390625,
0.438720703125,
0.5869140625,
0.21728515625,
0.5859375
],
[
0.24267578125,
0.58837890625,
0.78173828125,
0.58837890625,
0.78173828125,
0.60302734375,
0.24267578125,
0.60302734375
],
[
0.2181396484375,
0.60302734375,
0.7861328125,
0.60302734375,
0.7861328125,
0.6162109375,
0.2181396484375,
0.6162109375
],
[
0.2181396484375,
0.61865234375,
0.78515625,
0.61865234375,
0.78515625,
0.6318359375,
0.2181396484375,
0.6318359375
],
[
0.2181396484375,
0.63330078125,
0.78369140625,
0.63330078125,
0.78369140625,
0.64794921875,
0.2181396484375,
0.64794921875
],
[
0.2181396484375,
0.6484375,
0.78515625,
0.6484375,
0.78515625,
0.662109375,
0.2181396484375,
0.662109375
],
[
0.21728515625,
0.66357421875,
0.275390625,
0.66357421875,
0.275390625,
0.67822265625,
0.21728515625,
0.67822265625
],
[
0.34716796875,
0.68115234375,
0.52392578125,
0.68115234375,
0.52392578125,
0.69873046875,
0.34716796875,
0.69873046875
],
[
0.51318359375,
0.67822265625,
0.65869140625,
0.681640625,
0.65673828125,
0.70263671875,
0.51123046875,
0.69873046875
],
[
0.453369140625,
0.7021484375,
0.486083984375,
0.7021484375,
0.486083984375,
0.71337890625,
0.453369140625,
0.71337890625
],
[
0.52294921875,
0.7021484375,
0.55126953125,
0.7021484375,
0.55126953125,
0.71337890625,
0.52294921875,
0.71337890625
],
[
0.2181396484375,
0.72021484375,
0.78369140625,
0.72021484375,
0.78369140625,
0.73779296875,
0.2181396484375,
0.73779296875
],
[
0.21728515625,
0.736328125,
0.7861328125,
0.736328125,
0.7861328125,
0.75390625,
0.21728515625,
0.75390625
],
[
0.2156982421875,
0.75,
0.7646484375,
0.751953125,
0.7646484375,
0.76953125,
0.2156982421875,
0.767578125
],
[
0.2442626953125,
0.7822265625,
0.78515625,
0.78271484375,
0.78515625,
0.79736328125,
0.2442626953125,
0.796875
],
[
0.2181396484375,
0.798828125,
0.515625,
0.798828125,
0.515625,
0.8125,
0.2181396484375,
0.8125
],
[
0.261474609375,
0.8212890625,
0.74267578125,
0.82080078125,
0.74267578125,
0.84228515625,
0.261474609375,
0.84326171875
]
] |
[
"In this paper, we present a thorough and comprehensive implementation of",
"the conceptual spaces framework which is based on our formalization reported",
"in [7] and [8]. Its source code is publicly available on GitHub to researchers",
"anywhere in the world. Instead of investing a considerable amount of time into",
"developing their own implementation, researchers can use our implementation",
"off-the-shelf and focus on their specific application scenario.",
"The remainder of this paper is structured as follows: Section 2 introduces the",
"framework of conceptual spaces and our formalization. In Section 3, we give an",
"overview of our implementation and in Section 4 we illustrate its usage. Section",
"5 summarizes related work and Section 6 concludes the paper.",
"2",
"Conceptual Spaces",
"This section presents the cognitive framework of conceptual spaces as described",
"in [16] and summarizes our formalization as reported in [7] and [8]",
"2.1 Dimensions, Domains, and Distance",
"A conceptual space is spanned by a set D of so-called “quality dimensions\". Each",
"of these dimensions d E D represents a way in which two stimuli can be judged",
"to be similar or different. Examples for quality dimensions include temperature",
"weight, time, pitch, and hue. The distance between two points and y with",
"respect to a dimension d is denoted as rd - yal.",
"A domain D is a set of dimensions that inherently belong together.",
"Different perceptual modalities (like color, shape, or taste) are represented by",
"different domains. The color domain for instance consists of the three dimensions",
"hue, saturation, and brightness. Distance within a domain d is measured by the",
"weighted Euclidean metric de.",
"The overall conceptual space CS is defined as the product space of all dimen.",
"sions. Distance within the overall conceptual space is measured by the weighted",
"Manhattan metric dM of the intra-domain distances. This is supported by both",
"psychological evidence [5,25] and mathematical considerations [3]. Let △ be the",
"set of all domains in CS. The combined distance d within CS is defined as",
"follows:",
"d(r,y,W) = ws:",
" wd · [rd - yal2",
"SE",
"ded",
"The parameter W = (W, [Ws)se) contains two parts: W is the set of pos-",
"itive domain weights ws with se ws = [A/. Moreover, W contains for each",
"domain e △ a set Ws of positive dimension weights wd with des wd = 1.",
"The similarity of two points in a conceptual space is inversely related to their",
"distance. This can be written as follows :",
"Sim(a, ) =e-cd(r,y) with a constant c >0 and a given metric d"
] |
{
"arxiv_id": "1707.05165",
"language": "en",
"page_idx": 1
}
|
||
1707.05165-en-2
|
1707.05165-en-2.jpg
|
<FigureHere>
Fig.1. Left: Intuitive way to define regions for the concepts of “adult" and “child” (solid) as well as representation by using convex sets (dashed). Right: Representation by using star-shaped sets with central points marked by crosses.
Betweenness is a logical predicate $B(x,y,z)$ that is true if and only if $y$ is considered to be between $x$ and $z.$ It can be defned based on a given metric $d:$
$$B_d(x,y,z):\Longleftrightarrow\:d(x,y)+d(y,z)=d(x,z)$$
The betweenness relation based on $d_E$ results in the line segment connecting the points $x$ and $z$,whereas the betweenness relation based on $d_M$ results in an axis-parallel cuboid between the points $x$ and $z.$ One can define convexity and star-shapedness based on the notion of betweenness:
$\textbf{Definition 1. }( \textit{Convexity})$
$A\textit{ set C CS is convex under a metric d : }$
$$\forall x\in C,z\in C,y\in CS:(B_{d}(x,y,z)\to y\in C)$$
${\mathrm{Definition~2.~}}(Star-shapedness)$
$A$ set $S\subseteq CS$ $is$ star- shaped under $a$ metric $d$ with respect $to$ $a$ set $P\subseteq S$ $: \Longleftrightarrow$
$\forall p\in P,z\in S,y\in CS:(B_d(p,y,z)\to y\in S)$
# 2.2 Properties and Concepts
Gärdenfors [16] distinguishes properties like“red”,“round”, and“sweet”from full-fleshed concepts like “apple” or “dog” by observing that properties can be defined on individual domains (e.g., color, shape, taste), whereas full-fleshed concepts involve multiple domains. Each domain involved in representing a concept has a certain importance, which is reflected by so-called “salience weights”, Another important aspect of concepts are the correlations between the different domains, which are important for both learning [ 10] and reasoning [ 22, Ch 8] .
Based on the principle of cognitive economy, Gärdenfors argues that both properties and concepts should be represented as convex sets. However, this convexity assumption has recently been criticized [19] and we demonstrated in [7] that one cannot geometrically encode correlations between domains when using convex sets: The left part of Figure 1 shows two domains, age and height, and the concepts of child and adult. The solid ellipses illustrate the intuitive way of defining these concepts. As domains are combined with the Manhattan metric, a convex set corresponds in this case to an axis-parallel cuboid. One can easily see
|
[
[
0.428955078125,
0.1572265625,
0.461669921875,
0.1572265625,
0.461669921875,
0.1685791015625,
0.428955078125,
0.1685791015625
],
[
0.521484375,
0.15283203125,
0.525390625,
0.15283203125,
0.525390625,
0.1578369140625,
0.521484375,
0.1578369140625
],
[
0.64208984375,
0.1610107421875,
0.646484375,
0.1610107421875,
0.646484375,
0.1666259765625,
0.64208984375,
0.1666259765625
],
[
0.50634765625,
0.1685791015625,
0.521484375,
0.1685791015625,
0.51953125,
0.1976318359375,
0.5048828125,
0.1976318359375
],
[
0.3095703125,
0.177978515625,
0.314453125,
0.177978515625,
0.314453125,
0.1875,
0.3095703125,
0.1875
],
[
0.398681640625,
0.2279052734375,
0.4306640625,
0.2279052734375,
0.4306640625,
0.24365234375,
0.398681640625,
0.24365234375
],
[
0.60107421875,
0.2291259765625,
0.6298828125,
0.2291259765625,
0.6298828125,
0.24365234375,
0.60107421875,
0.24365234375
],
[
0.2181396484375,
0.2568359375,
0.78369140625,
0.255615234375,
0.78369140625,
0.270751953125,
0.2181396484375,
0.271484375
],
[
0.2181396484375,
0.270751953125,
0.78515625,
0.271484375,
0.78515625,
0.285888671875,
0.2181396484375,
0.285400390625
],
[
0.2181396484375,
0.285400390625,
0.6435546875,
0.285400390625,
0.6435546875,
0.2998046875,
0.2181396484375,
0.2998046875
],
[
0.2442626953125,
0.32763671875,
0.78515625,
0.32763671875,
0.78515625,
0.34228515625,
0.2442626953125,
0.34228515625
],
[
0.2181396484375,
0.343505859375,
0.78369140625,
0.343505859375,
0.78369140625,
0.356689453125,
0.2181396484375,
0.356689453125
],
[
0.351318359375,
0.37060546875,
0.654296875,
0.37060546875,
0.654296875,
0.385009765625,
0.351318359375,
0.385009765625
],
[
0.2442626953125,
0.396484375,
0.7861328125,
0.396484375,
0.7861328125,
0.410400390625,
0.2442626953125,
0.410400390625
],
[
0.2181396484375,
0.412353515625,
0.78515625,
0.412353515625,
0.78515625,
0.425537109375,
0.2181396484375,
0.425537109375
],
[
0.2181396484375,
0.4267578125,
0.7861328125,
0.4267578125,
0.7861328125,
0.43994140625,
0.2181396484375,
0.43994140625
],
[
0.2181396484375,
0.442626953125,
0.59326171875,
0.442626953125,
0.59326171875,
0.455810546875,
0.2181396484375,
0.455810546875
],
[
0.2181396484375,
0.46728515625,
0.414306640625,
0.46728515625,
0.414306640625,
0.481689453125,
0.2181396484375,
0.481689453125
],
[
0.21728515625,
0.481689453125,
0.564453125,
0.48095703125,
0.564453125,
0.49560546875,
0.21728515625,
0.496826171875
],
[
0.28759765625,
0.49755859375,
0.61376953125,
0.49755859375,
0.61376953125,
0.51220703125,
0.28759765625,
0.51220703125
],
[
0.2181396484375,
0.52294921875,
0.455810546875,
0.52294921875,
0.455810546875,
0.537109375,
0.2181396484375,
0.537109375
],
[
0.2181396484375,
0.53857421875,
0.779296875,
0.53857421875,
0.779296875,
0.5517578125,
0.2181396484375,
0.5517578125
],
[
0.2900390625,
0.5537109375,
0.60693359375,
0.5537109375,
0.60693359375,
0.568359375,
0.2900390625,
0.568359375
],
[
0.2197265625,
0.59033203125,
0.25244140625,
0.59033203125,
0.25244140625,
0.60302734375,
0.2197265625,
0.60302734375
],
[
0.259033203125,
0.5888671875,
0.465576171875,
0.5888671875,
0.465576171875,
0.60400390625,
0.259033203125,
0.60400390625
],
[
0.2181396484375,
0.6142578125,
0.7861328125,
0.6142578125,
0.7861328125,
0.62890625,
0.2181396484375,
0.62890625
],
[
0.2181396484375,
0.6298828125,
0.7861328125,
0.6298828125,
0.7861328125,
0.64453125,
0.2181396484375,
0.64453125
],
[
0.2181396484375,
0.64599609375,
0.7861328125,
0.64599609375,
0.7861328125,
0.6591796875,
0.2181396484375,
0.6591796875
],
[
0.21728515625,
0.66015625,
0.78369140625,
0.6611328125,
0.78369140625,
0.6748046875,
0.21728515625,
0.673828125
],
[
0.2181396484375,
0.67529296875,
0.78515625,
0.67529296875,
0.78515625,
0.689453125,
0.2181396484375,
0.689453125
],
[
0.2197265625,
0.69140625,
0.78369140625,
0.69140625,
0.78369140625,
0.705078125,
0.2197265625,
0.705078125
],
[
0.21728515625,
0.70556640625,
0.77392578125,
0.70556640625,
0.77392578125,
0.7197265625,
0.21728515625,
0.7197265625
],
[
0.24267578125,
0.72021484375,
0.7861328125,
0.72021484375,
0.7861328125,
0.734375,
0.24267578125,
0.734375
],
[
0.2181396484375,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2181396484375,
0.75
],
[
0.2197265625,
0.75048828125,
0.7861328125,
0.75048828125,
0.7861328125,
0.7646484375,
0.2197265625,
0.7646484375
],
[
0.221435546875,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.221435546875,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.7861328125,
0.78076171875,
0.7861328125,
0.79541015625,
0.2181396484375,
0.79541015625
],
[
0.21728515625,
0.796875,
0.78759765625,
0.796875,
0.78759765625,
0.81005859375,
0.21728515625,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.7861328125,
0.8125,
0.7861328125,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"adult",
"",
"",
"",
"I",
"age",
"age",
"Fig.1. Left: Intuitive way to define regions for the concepts of \"adult\" and “child\"",
"(solid) as well as representation by using convex sets (dashed). Right: Representation",
"by using star-shaped sets with central points marked by crosses.",
"Betweenness is a logical predicate B(c, y, z) that is true if and only if y is",
"considered to be between and z. It can be defined based on a given metric d:",
"Ba(c, y,z) : → d(,y) + d(y,z) =d(a, z)",
"The betweenness relation based on de results in the line segment connecting",
"the points and z, whereas the betweenness relation based on dm results in an",
"axis-parallel cuboid between the points r and z. One can define convexity and",
"star-shapedness based on the notion of betweenness:",
"Definition 1. (Converity)",
"A set CCCs is convexunderametricd:←",
"Vr EC,zEC,y E CS : (Ba(r,y,z) -→y EC)",
"Definition 2. (Star-shapedness)",
"A set S C CS is star-shaped under a metric d with respect to a set P C S : ",
"Vp E P,z E S, y ECS : (Ba(p, y,z) -→y E S)",
"2.2",
"Properties and Concepts",
"Gardenfors [16] distinguishes properties like “red\",“round\", and “sweet\" from",
"full-fleshed concepts like “apple\" or “dog\" by observing that properties can be",
"defined on individual domains (e.g, color, shape, taste), whereas full-fleshed",
"concepts involve multiple domains. Each domain involved in representing a con-",
"cept has a certain importance, which is reflected by so-called “salience weights\".",
"Another important aspect of concepts are the correlations between the different",
"domains, which are important for both learning [10] and reasoning [22, Ch 8].",
"Based on the principle of cognitive economy, Gardenfors argues that both",
"properties and concepts should be represented as convex sets. However, this",
"convexity assumption has recently been criticized [19] and we demonstrated in",
"[7] that one cannot geometrically encode correlations between domains when us-",
"ing convex sets: The left part of Figure 1 shows two domains, age and height, and",
"the concepts of child and adult. The solid ellipses illustrate the intuitive way of",
"defining these concepts. As domains are combined with the Manhattan metric, a",
"convex set corresponds in this case to an axis-parallel cuboid. One can easily see"
] |
{
"arxiv_id": "1707.05165",
"language": "en",
"page_idx": 2
}
|
||
1707.05165-en-3
|
1707.05165-en-3.jpg
|
that this convex representation (dashed rectangles) is not satisfactory, because the correlation of the two domains is not encoded. We therefore proposed in [7] to relax the convexity criterion and to use star-shaped sets, which is illustrated in the right part of Figure 1. This enables a geometric representation of correlations while still being only a minimal departure from the original framework
We have based our formalization on axis-parallel cuboids that can be de scribed by a triple $\langle\Delta_{C},p^{-},p^{+}\rangle$ consisting of a set of domains $\Delta_{C}$ on which this cuboid $C$ is defined and two points $p^-$ and $p^+$,such that
$$x\in C\iff\forall\delta\in\Delta_{C}:\forall d\in\delta:p_{d}^{-}\leq x_{d}\leq p_{d}^{+}$$
$\begin{array}{l}{\mathrm{These~cuboids~are~convex~under~}d_{C}^{\Delta}.~It~is~also~easy~to~see~that~any~union~of}\\{\mathrm{convex~sets~that~have~a~non-empty~intersection~is~star-shaped~[26].~We~define}}\end{array}$
the core of a concept as follows:
$\textbf{Definition 3. }( Simple\textit{star- shaped set})$
$A$ simple star-shaped set $S\textit{ is described as a tuple }\langle \Delta _S, \{ C_1, \ldots , C_m\} \rangle . \Delta _S\subseteq \Delta$ $isasetofdomainsonwhichthecuboids\left\{C_{1},\ldots,C_{m}\right\}(andthusalsoS)are$ $defned. Moreover, we\textit{ require that the central region P}: = \bigcap _{i= 1}^{m}C_{i}\neq \emptyset . Then$ $the\textit{ simple star- shaped set S is defined as}$
$$S:=\bigcup\limits_{i=1}^mC_i$$
In order to represent imprecise concept boundaries, we use fuzzy sets [9,28,29]. A fuzzy set is characterized by its membership function $\mu:CS\to[0,1]$ which assigns a degree of membership to each point in the conceptual space. The membership of a point to a fuzzy concept is based on its maximal similarity to any of the points in the concept's core:
$\textbf{Definition 4. }( Fuzzy\textit{ simple star- }shaped\textit{ set})$
$A$ fuzzy simple star-shaped set $\widetilde{S} \textit{is described by a quadruple }\langle S, \mu _0, c, W\rangle \textit{where}$ $S= \langle \Delta _{S}, \{ C_{1}, \ldots , C_{m}\} \rangle$ $is$ $a$ non-empty simple star-shaped $set.$ The parameter $\mu_{0}\in(0,1]\textit{ controls the highest possible membership to \widetilde{S} and is usually set to 1.}$ The sensitivity parameter $c> 0$ controls the rate $of$ the exponential decay $in$ the similarity $function.$ $Finally$, $W=$ $\langle W_{\Delta _S}, \{ W_\delta \} _{\delta \in \Delta _S}\rangle$ contains positive weights for all domains $in$ $\Delta _{S}$ and all dimensions within these $domains$, reflecting their respective $importance.$ $We$ require that $\sum _{\delta \in \Delta _{S}}w_{\delta }=$ $| \Delta _{S}|$ and that $\forall \delta \in \Delta _{S}$ : $\sum _{d\in \delta }w_{d}= 1. The\textit{ membership function of }\tilde{S} \textit{ is then defned as follows: }$
$$\mu_{\widetilde{S}}(x)=\mu_0\cdot\max_{y\in S}(e^{-c\cdot d_C^\Delta(x,y,W)})$$
The sensitivity parameter $c$ controls the overall degree of fuzziness of $\bar{S}$ by determining how fast the membership drops to zero. The weights $W$ represent not only the relative importance of the respective domain or dimension for the
|
[
[
0.2181396484375,
0.1483154296875,
0.78369140625,
0.1483154296875,
0.78369140625,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2181396484375,
0.1622314453125,
0.78515625,
0.1634521484375,
0.78515625,
0.177978515625,
0.2181396484375,
0.1767578125
],
[
0.2197265625,
0.1787109375,
0.78515625,
0.1787109375,
0.78515625,
0.1925048828125,
0.2197265625,
0.1925048828125
],
[
0.2181396484375,
0.1956787109375,
0.78173828125,
0.1956787109375,
0.78173828125,
0.205810546875,
0.2181396484375,
0.205810546875
],
[
0.2181396484375,
0.208984375,
0.7685546875,
0.208984375,
0.7685546875,
0.222900390625,
0.2181396484375,
0.222900390625
],
[
0.2442626953125,
0.238037109375,
0.78369140625,
0.238037109375,
0.78369140625,
0.25244140625,
0.2442626953125,
0.25244140625
],
[
0.2181396484375,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.268310546875,
0.2181396484375,
0.268310546875
],
[
0.21728515625,
0.268310546875,
0.62744140625,
0.268310546875,
0.62744140625,
0.2822265625,
0.21728515625,
0.2822265625
],
[
0.336669921875,
0.29345703125,
0.66259765625,
0.29541015625,
0.66259765625,
0.3125,
0.336669921875,
0.31005859375
],
[
0.21728515625,
0.3212890625,
0.7861328125,
0.3212890625,
0.7861328125,
0.3359375,
0.21728515625,
0.3359375
],
[
0.2181396484375,
0.337158203125,
0.78369140625,
0.337158203125,
0.78369140625,
0.3515625,
0.2181396484375,
0.3515625
],
[
0.2181396484375,
0.354248046875,
0.450927734375,
0.354248046875,
0.450927734375,
0.367431640625,
0.2181396484375,
0.367431640625
],
[
0.2197265625,
0.37890625,
0.50244140625,
0.37890625,
0.50244140625,
0.393310546875,
0.2197265625,
0.393310546875
],
[
0.2197265625,
0.39453125,
0.78515625,
0.39453125,
0.78515625,
0.4091796875,
0.2197265625,
0.4091796875
],
[
0.21728515625,
0.40771484375,
0.78759765625,
0.4072265625,
0.78759765625,
0.4248046875,
0.21728515625,
0.425537109375
],
[
0.2197265625,
0.4248046875,
0.7861328125,
0.4248046875,
0.7861328125,
0.438232421875,
0.2197265625,
0.438232421875
],
[
0.2197265625,
0.43994140625,
0.51318359375,
0.43994140625,
0.51318359375,
0.453857421875,
0.2197265625,
0.453857421875
],
[
0.465576171875,
0.4765625,
0.494384765625,
0.4765625,
0.494384765625,
0.4892578125,
0.465576171875,
0.4892578125
],
[
0.5146484375,
0.47607421875,
0.53759765625,
0.481689453125,
0.53369140625,
0.4912109375,
0.5107421875,
0.48486328125
],
[
0.24267578125,
0.51513671875,
0.78173828125,
0.51513671875,
0.78173828125,
0.52880859375,
0.24267578125,
0.52880859375
],
[
0.2181396484375,
0.52978515625,
0.78515625,
0.52978515625,
0.78515625,
0.54345703125,
0.2181396484375,
0.54345703125
],
[
0.2181396484375,
0.54541015625,
0.78369140625,
0.54541015625,
0.78369140625,
0.5595703125,
0.2181396484375,
0.5595703125
],
[
0.2181396484375,
0.56103515625,
0.78369140625,
0.56103515625,
0.78369140625,
0.57470703125,
0.2181396484375,
0.57470703125
],
[
0.2181396484375,
0.57568359375,
0.46728515625,
0.57568359375,
0.46728515625,
0.5888671875,
0.2181396484375,
0.5888671875
],
[
0.2181396484375,
0.6015625,
0.5439453125,
0.6015625,
0.5439453125,
0.61572265625,
0.2181396484375,
0.61572265625
],
[
0.21728515625,
0.61572265625,
0.78759765625,
0.6162109375,
0.78759765625,
0.6337890625,
0.21728515625,
0.63330078125
],
[
0.2181396484375,
0.63330078125,
0.78515625,
0.6318359375,
0.78515625,
0.646484375,
0.2181396484375,
0.64794921875
],
[
0.2181396484375,
0.6494140625,
0.78515625,
0.6484375,
0.78515625,
0.66357421875,
0.2181396484375,
0.6640625
],
[
0.221435546875,
0.66552734375,
0.7861328125,
0.66552734375,
0.7861328125,
0.6787109375,
0.221435546875,
0.6787109375
],
[
0.2197265625,
0.6787109375,
0.7890625,
0.6787109375,
0.7890625,
0.6962890625,
0.2197265625,
0.6962890625
],
[
0.2197265625,
0.69580078125,
0.7861328125,
0.69580078125,
0.7861328125,
0.708984375,
0.2197265625,
0.708984375
],
[
0.21728515625,
0.70849609375,
0.7890625,
0.70703125,
0.7890625,
0.72802734375,
0.21728515625,
0.72900390625
],
[
0.21728515625,
0.72802734375,
0.73046875,
0.72607421875,
0.73046875,
0.74365234375,
0.21728515625,
0.74560546875
],
[
0.38232421875,
0.7587890625,
0.61376953125,
0.751953125,
0.615234375,
0.77294921875,
0.384033203125,
0.77978515625
],
[
0.486083984375,
0.77294921875,
0.50634765625,
0.77294921875,
0.50634765625,
0.78076171875,
0.486083984375,
0.78076171875
],
[
0.2442626953125,
0.79541015625,
0.78369140625,
0.79541015625,
0.78369140625,
0.81005859375,
0.2442626953125,
0.81005859375
],
[
0.2181396484375,
0.8125,
0.78369140625,
0.8125,
0.78369140625,
0.82568359375,
0.2181396484375,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"that this convex representation (dashed rectangles) is not satisfactory, because",
"the correlation of the two domains is not encoded. We therefore proposed in [7]",
"to relax the convexity criterion and to use star-shaped sets, which is illustrated",
"in therightpart of Figurel.This enables ageometricrepresentation of correla",
"tions while still being only a minimal departure from the original framework.",
"We have based our formalization on axis-parallel cuboids that can be de-",
"scribed by a triple (△c,p=,p+) consisting of a set of domains c on which this",
"cuboid is defined and two points p- and pt, such that",
"EEC:dEd:a",
"These cuboids are convex under d. It is also easy to see that any union of",
"convex sets that have a non-empty intersection is star-shaped [26]. We define",
"the core of a concept as follows:",
"Definition 3. (Simple star-shaped set)",
"A simple star-shaped set S is described as a tuple (△s, (Ci,:..,Cm)). As ",
"is a set of domains on which the cuboids (Ci,...,Cm) (and thus also S) are",
"defined. Moreover, we require that the central region P := -, Ci + O. Then",
"the simple star-shaped set S is defined as",
"S:=",
"10",
"In order to represent imprecise concept boundaries, we use fuzzy sets [9,28,29]",
"A fuzzy set is characterized by its membership function μ : CS -→ [0, 1] which",
"assigns a degree of membership to each point in the conceptual space. The mem-",
"bership of a point to a fuzzy concept is based on its maximal similarity to any",
"of the points in the concept's core:",
"Definition 4. (Fuzzy simple star-shaped set)",
"A fuzzy simple star-shaped set S is described by a quadruple (S, μo,c, W) where",
"S = <s, [Ci,...,Cm]) is a non-empty simple star-shaped set. The parameter",
"Ho E (O,1) controls the highest possible membership to S and is usually set to 1.",
"The sensitivity parameter c > O controls the rate of the erponential decay in the",
"similarity function. Finally, W = (Ws, {Ws]ses) contains positive weights",
"for all domains in s and all dimensions within these domains, reflecting their",
"respective importance. We require that EseAs ws = IAsl and that Vs e As :",
"des Wa = 1. The membership function of S is then defined as follows:",
"μg(z) = μo max(e-cd(r,y,W)",
"yEs",
"The sensitivity parameter c controls the overall degree of fuzziness of S by",
"determining how fast the membership drops to zero. The weights W represent",
"not only the relative importance of the respective domain or dimension for the"
] |
{
"arxiv_id": "1707.05165",
"language": "en",
"page_idx": 3
}
|
||
1707.05165-en-4
|
1707.05165-en-4.jpg
|
<FigureHere>
$\mathbf{Fig.2.~Left:~Three~cuboids~}C_{1},C_{2},C_{3}$ with nonempty intersection. Middle: Resulting simple star-shaped set $S$ based on these cuboids. Right: Fuzzy simple star-shaped set $\tilde{S}$ based on $S$ with three $\alpha$-cuts for $\alpha\in\{1.0,0.5,0.25\}.$
represented concept, but they also influence the relative fuzziness with respect to this domain or dimension. Note that if $|\Delta_S|=1$, then $\widetilde{S}$ represents a property, and if$|\Delta_S|>1$,then $\widetilde{S}$ represents a concept. Figure 2 illustrates these definitions (the $x$ and $y$ axes are assumed to belong to different domains and are combined with $d_{M}$ using equal weights).
# 2.3 Operations on Concepts
Our formalization provides a number of operations, which can be used to create new concepts from existing ones and to describe relations between concepts. We summarize them here only briefly, more details can be found in [7] and [8].
Intersection. The intersection of two concepts can be interpreted as the logical “and”-e.g., intersecting the property “green”with the concept“banana" results in the set of all objects that are both green and bananas. The intersection of two concepts is defined as follows: The core of the resulting concept is the intersection of the highest intersecting $\alpha\text{-cuts}^1$ of both original concepts, i.e., $S^\prime=\tilde{S}_1^{\alpha^{\prime}}\cap$ $\widetilde{S}_{2}^{\alpha^{\prime}}$with$\alpha^\prime=\max\{\alpha\in[0,1]:\widetilde{S}_{1}^{\alpha}\cap\widetilde{S}_{2}^{\alpha}\neq\emptyset\}.$As this intersection is not guaranteed to be a valid core, we approximate it by cuboids. If these cuboids do not intersect, we compute the arithmetic mean of the cuboids' centers and extend each cuboid, such that it contains this central point. We use $\alpha^\prime$ as the resulting concept's $\mu_0$ parameter, set $c^\prime:=\min(c^{(1)},c^{(2)})$, and derive a new set of weights by interpolating between the two original sets of weights
Union. The union of two concepts can be used to construct more abstract categories (e.g., defning“fruit”as the union of“apple”,“banana",“coconut",etc.). We define the union of two concepts as follows: The core of the new concept is
${^{1}\text{The }\alpha \mathrm{- cut~of}}$ $\widetilde{S}$ is defined as $\widetilde{S}^\alpha:=\{x\in CS\mid\mu_{\tilde{S}}(x)\geq\alpha\}.$
|
[
[
0.443603515625,
0.1787109375,
0.455078125,
0.1817626953125,
0.450927734375,
0.1900634765625,
0.439453125,
0.1875
],
[
0.62841796875,
0.177978515625,
0.69140625,
0.177978515625,
0.69140625,
0.1925048828125,
0.62841796875,
0.1925048828125
],
[
0.44287109375,
0.208984375,
0.463134765625,
0.208984375,
0.463134765625,
0.2197265625,
0.44287109375,
0.2197265625
],
[
0.62841796875,
0.203857421875,
0.65625,
0.203857421875,
0.65625,
0.222900390625,
0.62841796875,
0.222900390625
],
[
0.2181396484375,
0.304931640625,
0.7861328125,
0.30615234375,
0.7861328125,
0.32080078125,
0.2181396484375,
0.3193359375
],
[
0.2181396484375,
0.32080078125,
0.7861328125,
0.32080078125,
0.7861328125,
0.333984375,
0.2181396484375,
0.333984375
],
[
0.21728515625,
0.333251953125,
0.5810546875,
0.333984375,
0.5810546875,
0.348388671875,
0.21728515625,
0.347900390625
],
[
0.21728515625,
0.37890625,
0.78515625,
0.37890625,
0.78515625,
0.393310546875,
0.21728515625,
0.393310546875
],
[
0.21728515625,
0.392578125,
0.78515625,
0.39453125,
0.78515625,
0.4091796875,
0.21728515625,
0.4072265625
],
[
0.2181396484375,
0.410400390625,
0.78369140625,
0.410400390625,
0.78369140625,
0.423583984375,
0.2181396484375,
0.423583984375
],
[
0.2181396484375,
0.4248046875,
0.78759765625,
0.425537109375,
0.78759765625,
0.43994140625,
0.2181396484375,
0.439453125
],
[
0.2181396484375,
0.44140625,
0.4345703125,
0.44140625,
0.4345703125,
0.455078125,
0.2181396484375,
0.455078125
],
[
0.2181396484375,
0.479736328125,
0.461669921875,
0.479736328125,
0.461669921875,
0.494384765625,
0.2181396484375,
0.494384765625
],
[
0.2181396484375,
0.50830078125,
0.78515625,
0.50830078125,
0.78515625,
0.521484375,
0.2181396484375,
0.521484375
],
[
0.21728515625,
0.52294921875,
0.7861328125,
0.52294921875,
0.7861328125,
0.5361328125,
0.21728515625,
0.5361328125
],
[
0.2181396484375,
0.53857421875,
0.7509765625,
0.53857421875,
0.7509765625,
0.5517578125,
0.2181396484375,
0.5517578125
],
[
0.2181396484375,
0.5771484375,
0.7861328125,
0.5771484375,
0.7861328125,
0.59033203125,
0.2181396484375,
0.59033203125
],
[
0.2197265625,
0.59228515625,
0.78515625,
0.59228515625,
0.78515625,
0.60693359375,
0.2197265625,
0.60693359375
],
[
0.21728515625,
0.60595703125,
0.7861328125,
0.60693359375,
0.7861328125,
0.6220703125,
0.21728515625,
0.62060546875
],
[
0.2181396484375,
0.62255859375,
0.78515625,
0.62255859375,
0.78515625,
0.63623046875,
0.2181396484375,
0.63623046875
],
[
0.2181396484375,
0.63818359375,
0.78759765625,
0.63818359375,
0.78759765625,
0.6513671875,
0.2181396484375,
0.6513671875
],
[
0.2197265625,
0.65380859375,
0.78515625,
0.65380859375,
0.78515625,
0.66845703125,
0.2197265625,
0.66845703125
],
[
0.2197265625,
0.66943359375,
0.78515625,
0.66943359375,
0.78515625,
0.68310546875,
0.2197265625,
0.68310546875
],
[
0.2181396484375,
0.68505859375,
0.7861328125,
0.68505859375,
0.7861328125,
0.69873046875,
0.2181396484375,
0.69873046875
],
[
0.21728515625,
0.69970703125,
0.78515625,
0.69970703125,
0.78515625,
0.71337890625,
0.21728515625,
0.71337890625
],
[
0.21728515625,
0.7138671875,
0.7861328125,
0.71337890625,
0.7861328125,
0.72802734375,
0.21728515625,
0.72900390625
],
[
0.21728515625,
0.73095703125,
0.705078125,
0.73095703125,
0.705078125,
0.744140625,
0.21728515625,
0.744140625
],
[
0.2181396484375,
0.767578125,
0.78173828125,
0.76806640625,
0.78173828125,
0.78271484375,
0.2181396484375,
0.7822265625
],
[
0.2197265625,
0.78515625,
0.78173828125,
0.78515625,
0.78173828125,
0.798828125,
0.2197265625,
0.798828125
],
[
0.2197265625,
0.7998046875,
0.78515625,
0.7998046875,
0.78515625,
0.81298828125,
0.2197265625,
0.81298828125
],
[
0.2197265625,
0.82275390625,
0.61767578125,
0.82568359375,
0.61767578125,
0.84228515625,
0.2197265625,
0.8388671875
]
] |
[
"S",
"s=S1.0",
"P",
"P",
"Fig.2. Left: Three cuboids Ci, C2, Cs with nonempty intersection. Middle: Resulting",
"simple star-shaped set S based on these cuboids. Right: Fuzzy simple star-shaped set",
"S based on S with three α-cuts for α E [1.0, 0.5,0.25].",
"represented concept, but they also influence the relative fuzziness with respect",
"to this domain or dimension. Note that if s = 1, then S represents a property,",
"and if Asl > 1, then S represents a concept. Figure 2 illustrates these definitions",
"(the r and y axes are assumed to belong to different domains and are combined",
"with dm using equal weights).",
"2.3 Operations on Concepts",
"Our formalization provides a number of operations, which can be used to create",
"new concepts from existing ones and to describe relations between concepts. We",
"summarize them here only briefly, more details can be found in [7] and [8]]",
"Intersection. The intersection of two concepts can be interpreted as the logical",
"\"and\"-e.g., intersecting the property\"green\"with the concept “banana\" results",
"in the set of all objects that are both green and bananas.The intersection of two",
"concepts is defined as follows: The core of the resulting concept is the intersection",
"of the highest intersecting Q-cuts' of both original concepts, i.e., S' = So n",
"Sg' with Q' = maxα e [0,1] : S n Sg + 0]. As this intersection is not",
"guaranteed to be a valid core, we approximate it by cuboids. If these cuboids",
"do not intersect, we compute the arithmetic mean of the cuboids' centers and",
"extend each cuboid, such that it contains this central point. We use α' as the",
"resulting concept's μo parameter, set c' := min(c(1),c(2)), and derive a new set",
"of weights by interpolating between the two original sets of weights.",
"Union. The union of two concepts can be used to construct more abstract cate",
"gories (e.g., defining“\"fruit\" as the union of \"apple\",\"banana\",“coconut\", etc.)",
"We define the union of two concepts as follows: The core of the new concept is",
"1 The a-cut of S is defined as sa := (α E CS I μg(a) ≥ a)."
] |
{
"arxiv_id": "1707.05165",
"language": "en",
"page_idx": 4
}
|
||
2310.15664-en-0
|
2310.15664-en-0.jpg
|
# Expression Syntax Information Bottleneck for Math Word Problems
Min Yan$g^*$
Chengming Li$^*$
lichengming@mail.sysu.edu.cn
Sun Yat-sen University
Shenzhen, Guangdong, China
$\operatorname{Jing}$Xiong
xiongj69@mail2.sysu.edu.cn
Sun Yat-sen University
Shenzhen, Guangdong, China
$\operatorname*{min.yang@siat.ac.cn}$
SIAT, Chinese Academy of Sciences
Shenzhen, Guangdong, China
Xiping Hu
$\bar{\text{huxiping@mail.sysu.edu.cn}}$
Sun Yat-sen University
Shenzhen, Guangdong, China
$\operatorname{Bin}$Hu
bh@lzu.edu.cn
Lanzhou University Lanzhou, Gansu, China
# $\mathbf{ABSTRACT}$
# 1 INTRODUCTION
Math Word Problems (MWP) aims to automatically solve mathematical questions given in texts. Previous studies tend to design complex models to capture additional information in the original text so as to enable the model to gain more comprehensive features. In this paper, we turn our attention in the opposite direction, and work on how to discard redundant features containing spurious correlations for MWP. To this end, we design an Expression Syntax
correlations for MWP. To this end, we design an Expression Syntax $\textbf{Information Bottleneck method for MWP ( called ESIB) based on}$ variational information bottleneck, which extracts essential features of expression syntax tree while filtering la
containing syntax-irrelevant features. The key idea of $ESIB$ is to encourage multiple models to predict the same expression syntax tree for different problem representations of the same problem by mutual learning so as to capture consistent information of expression syntax tree and discard latent-specific redundancy. To improve the generalization ability of the model and generate more diverse expressions, we design a self-distillation loss to encourage the model to rely more on the expression syntax information in the latent space. Experimental results on two large-scale benchmarks show that our model not only achieves state-of-the-art results but also generates more diverse solutions. The code is available.$^1$
Math Word Problems (MWP) is challenging and draws much attention from researchers in the field of natural language processing [3,25] and information retrieval (e.g., mathematical understanding) [8,26]. MWP aims to automatically answer mathematical questions given in a natural language, which requires the model not only understand what facts are presented in a text, but also possess the reasoning capability to answer the mathematical question. Table 1 shows three examples of MWP with three mathematical problems and their solution expressions with answer.
Inspired by the success of deep learning [14,21], the encoderdecoder framework with attention mechanisms [2] have been dominated in MWP [18-20], which bring the state-of-the-art to a new level. The key idea is to use an encoder to learn representations of problem text and employ a decoder to generate the corresponding solution expression and answer. Subsequently, several studies propose sequence-to-tree models, which explore the tree structure information presented in the text and improve the generation of solution expressions [23,24,28].
However, the previous MWP methods appear to rely on spurious correlations between the shallow heuristics in problem and solution expression [15]. For example, as shown in Table 1, previous models may associate Problem 1 and Problem 2 with the mathematical formula“x×y÷z”, since these two problems have similar semantic patterns, e.g., calculating the speed. Based on this association, the models could generate wrong solution expression for Problem 3 which has similar semantic problem expression like the text segment"place A to place B" in Problems 1-2. In particular, the models that learn spurious correlations are more likely to generate wrong solution expression“220×25%÷30%”, rather than $“220\div(25\%+30\%)”$for Problem 3. We define such a false association as spurious correlation. Some recent studies have revealed that MWP solvers relying on spurious correlations could achieve high accuracy [9,15]. These models can even compute correct answers without paying attention to the question part in the problem such as the text segment“how many kilometers is the total length of the two places?” in Problem 3 calculating the distance. In addition, the solution expression is sensitive to the perturbed latent representations [10], since the semantically similar mathematical problems, even with totally different solution expressions and questions, can be encoded closely in the latent space. We believe it is an evidence that redundant information containing spurious correlations is encoded
# $\mathbf{KEYWORDS}$
Math Word Problems, Mutual learning, Spurious correlations, Vari-
ational information bottleneck
ACM Reference Format:
Jing Xiong, Chengming Li, Min Yang, Xiping Hu, and Bin Hu. 2022. Expres-
sion Syntax Information Bottleneck for Math Word Problems. In Proceedings ofthe 45th International ACM SIGIR Conference on Research and Development inInformation Retrieval (SIGIR '22), fuly 11-15, 2022, Madrid, Spain. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3477495.3531824
$^{*}$Corresponding authors.
$^{1}$https://github.com/menik1126/math\_ESIB
Permission to make digital or hard copies of part or all 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 third-party components of this work must be honored For all other uses, contact the owner/author(s).
SIGIR'22, fuly 11-15, 2022, Madrid, Spain
$\textcircled{\circ}$ 2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8732-3/22/07...$15.00
https://doi.org/10.1145/3477495.3531824
|
[
[
0.2303466796875,
0.10479736328125,
0.7685546875,
0.10479736328125,
0.7685546875,
0.12249755859375,
0.2303466796875,
0.12249755859375
],
[
0.338134765625,
0.1287841796875,
0.66015625,
0.1300048828125,
0.66015625,
0.147705078125,
0.338134765625,
0.146484375
],
[
0.1846923828125,
0.1558837890625,
0.27783203125,
0.1578369140625,
0.27783203125,
0.175537109375,
0.1846923828125,
0.173583984375
],
[
0.439453125,
0.1590576171875,
0.55810546875,
0.1590576171875,
0.55810546875,
0.173583984375,
0.439453125,
0.173583984375
],
[
0.72216796875,
0.1558837890625,
0.810546875,
0.1578369140625,
0.810546875,
0.1767578125,
0.72216796875,
0.1749267578125
],
[
0.1356201171875,
0.175537109375,
0.328369140625,
0.175537109375,
0.328369140625,
0.189453125,
0.1356201171875,
0.189453125
],
[
0.393798828125,
0.175537109375,
0.60546875,
0.175537109375,
0.60546875,
0.1900634765625,
0.393798828125,
0.1900634765625
],
[
0.69677734375,
0.1749267578125,
0.83837890625,
0.1749267578125,
0.83837890625,
0.1881103515625,
0.69677734375,
0.1881103515625
],
[
0.420654296875,
0.189453125,
0.57861328125,
0.1900634765625,
0.57861328125,
0.2052001953125,
0.420654296875,
0.203857421875
],
[
0.64794921875,
0.189453125,
0.88623046875,
0.189453125,
0.88623046875,
0.20263671875,
0.64794921875,
0.20263671875
],
[
0.1298828125,
0.20263671875,
0.33251953125,
0.203857421875,
0.33251953125,
0.2215576171875,
0.1298828125,
0.2203369140625
],
[
0.1544189453125,
0.1900634765625,
0.30810546875,
0.1900634765625,
0.30810546875,
0.203857421875,
0.1544189453125,
0.203857421875
],
[
0.400390625,
0.205810546875,
0.60107421875,
0.205810546875,
0.60107421875,
0.2197265625,
0.400390625,
0.2197265625
],
[
0.66650390625,
0.205810546875,
0.869140625,
0.205810546875,
0.869140625,
0.2197265625,
0.66650390625,
0.2197265625
],
[
0.3203125,
0.2310791015625,
0.41015625,
0.2310791015625,
0.41015625,
0.248779296875,
0.3203125,
0.248779296875
],
[
0.60302734375,
0.231689453125,
0.6650390625,
0.231689453125,
0.6650390625,
0.2474365234375,
0.60302734375,
0.2474365234375
],
[
0.271240234375,
0.248779296875,
0.4599609375,
0.248779296875,
0.4599609375,
0.26318359375,
0.271240234375,
0.26318359375
],
[
0.5810546875,
0.2474365234375,
0.6845703125,
0.2474365234375,
0.6845703125,
0.261962890625,
0.5810546875,
0.261962890625
],
[
0.28759765625,
0.264404296875,
0.44287109375,
0.264404296875,
0.44287109375,
0.27783203125,
0.28759765625,
0.27783203125
],
[
0.56640625,
0.26318359375,
0.701171875,
0.26318359375,
0.701171875,
0.27783203125,
0.56640625,
0.27783203125
],
[
0.2646484375,
0.279052734375,
0.465576171875,
0.279052734375,
0.465576171875,
0.29345703125,
0.2646484375,
0.29345703125
],
[
0.0196075439453125,
0.26513671875,
0.058013916015625,
0.26513671875,
0.060455322265625,
0.373779296875,
0.0228729248046875,
0.373779296875
],
[
0.5537109375,
0.279052734375,
0.71337890625,
0.279052734375,
0.71337890625,
0.292236328125,
0.5537109375,
0.292236328125
],
[
0.08905029296875,
0.304931640625,
0.181396484375,
0.304931640625,
0.181396484375,
0.31640625,
0.08905029296875,
0.31640625
],
[
0.521484375,
0.30615234375,
0.5390625,
0.30615234375,
0.5390625,
0.314453125,
0.521484375,
0.314453125
],
[
0.53759765625,
0.306884765625,
0.54345703125,
0.306884765625,
0.54345703125,
0.3125,
0.53759765625,
0.3125
],
[
0.5400390625,
0.3037109375,
0.68701171875,
0.3037109375,
0.68701171875,
0.317626953125,
0.5400390625,
0.317626953125
],
[
0.0849609375,
0.3212890625,
0.48046875,
0.3212890625,
0.48046875,
0.335205078125,
0.0849609375,
0.335205078125
],
[
0.51708984375,
0.3212890625,
0.91259765625,
0.3212890625,
0.91259765625,
0.335205078125,
0.51708984375,
0.335205078125
],
[
0.08660888671875,
0.3359375,
0.4794921875,
0.3359375,
0.4794921875,
0.3466796875,
0.08660888671875,
0.3466796875
],
[
0.51708984375,
0.335205078125,
0.9140625,
0.3359375,
0.9140625,
0.349853515625,
0.51708984375,
0.348388671875
],
[
0.08660888671875,
0.349853515625,
0.48193359375,
0.349853515625,
0.48193359375,
0.363037109375,
0.08660888671875,
0.363037109375
],
[
0.5185546875,
0.349853515625,
0.91259765625,
0.349853515625,
0.91259765625,
0.363037109375,
0.5185546875,
0.363037109375
],
[
0.02532958984375,
0.3623046875,
0.056365966796875,
0.3623046875,
0.056365966796875,
0.399658203125,
0.02532958984375,
0.399658203125
],
[
0.08660888671875,
0.363037109375,
0.48193359375,
0.363037109375,
0.48193359375,
0.376953125,
0.08660888671875,
0.376953125
],
[
0.5185546875,
0.363037109375,
0.91259765625,
0.363037109375,
0.91259765625,
0.376953125,
0.5185546875,
0.376953125
],
[
0.0882568359375,
0.37744140625,
0.4794921875,
0.37744140625,
0.4794921875,
0.38818359375,
0.0882568359375,
0.38818359375
],
[
0.5185546875,
0.37744140625,
0.91259765625,
0.37744140625,
0.91259765625,
0.391357421875,
0.5185546875,
0.391357421875
],
[
0.5185546875,
0.391357421875,
0.9111328125,
0.391357421875,
0.9111328125,
0.401611328125,
0.5185546875,
0.401611328125
],
[
0.037567138671875,
0.37890625,
0.043304443359375,
0.38330078125,
0.037567138671875,
0.386962890625,
0.0318603515625,
0.38330078125
],
[
0.08660888671875,
0.392578125,
0.48046875,
0.392578125,
0.48046875,
0.40283203125,
0.08660888671875,
0.40283203125
],
[
0.0228729248046875,
0.404052734375,
0.062103271484375,
0.404052734375,
0.062103271484375,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.0882568359375,
0.406005859375,
0.48046875,
0.406005859375,
0.48046875,
0.416748046875,
0.0882568359375,
0.416748046875
],
[
0.51708984375,
0.40478515625,
0.9140625,
0.40478515625,
0.9140625,
0.41845703125,
0.51708984375,
0.41845703125
],
[
0.0849609375,
0.41845703125,
0.48193359375,
0.41845703125,
0.48193359375,
0.431884765625,
0.0849609375,
0.431884765625
],
[
0.51708984375,
0.41845703125,
0.9140625,
0.41845703125,
0.9140625,
0.431884765625,
0.51708984375,
0.431884765625
],
[
0.08660888671875,
0.434326171875,
0.4794921875,
0.434326171875,
0.4794921875,
0.4443359375,
0.08660888671875,
0.4443359375
],
[
0.5185546875,
0.434326171875,
0.77783203125,
0.434326171875,
0.77783203125,
0.4443359375,
0.5185546875,
0.4443359375
],
[
0.08660888671875,
0.447509765625,
0.48046875,
0.447509765625,
0.48046875,
0.458251953125,
0.08660888671875,
0.458251953125
],
[
0.53125,
0.44580078125,
0.9140625,
0.4443359375,
0.9140625,
0.458984375,
0.53125,
0.460205078125
],
[
0.08660888671875,
0.460205078125,
0.483642578125,
0.460205078125,
0.483642578125,
0.473388671875,
0.08660888671875,
0.473388671875
],
[
0.51708984375,
0.460205078125,
0.9140625,
0.460205078125,
0.9140625,
0.473388671875,
0.51708984375,
0.473388671875
],
[
0.08660888671875,
0.47607421875,
0.4794921875,
0.47607421875,
0.4794921875,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.5185546875,
0.47607421875,
0.91259765625,
0.47607421875,
0.91259765625,
0.486083984375,
0.5185546875,
0.486083984375
],
[
0.08660888671875,
0.4892578125,
0.483642578125,
0.4892578125,
0.483642578125,
0.5,
0.08660888671875,
0.5
],
[
0.5185546875,
0.4892578125,
0.91259765625,
0.4892578125,
0.91259765625,
0.5,
0.5185546875,
0.5
],
[
0.5185546875,
0.501953125,
0.9140625,
0.501953125,
0.9140625,
0.51513671875,
0.5185546875,
0.51513671875
],
[
0.0187835693359375,
0.49365234375,
0.058013916015625,
0.492431640625,
0.060455322265625,
0.58056640625,
0.021240234375,
0.58203125
],
[
0.08660888671875,
0.50390625,
0.48046875,
0.50390625,
0.48046875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.0849609375,
0.5166015625,
0.48193359375,
0.5166015625,
0.48193359375,
0.52978515625,
0.0849609375,
0.52978515625
],
[
0.5185546875,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.52783203125,
0.5185546875,
0.52783203125
],
[
0.08416748046875,
0.52978515625,
0.483642578125,
0.52783203125,
0.483642578125,
0.54150390625,
0.08416748046875,
0.54345703125
],
[
0.515625,
0.52978515625,
0.9140625,
0.52783203125,
0.9140625,
0.54150390625,
0.515625,
0.54345703125
],
[
0.0849609375,
0.54345703125,
0.48193359375,
0.54345703125,
0.48193359375,
0.556640625,
0.0849609375,
0.556640625
],
[
0.51708984375,
0.54345703125,
0.9140625,
0.54345703125,
0.9140625,
0.556640625,
0.51708984375,
0.556640625
],
[
0.08660888671875,
0.5595703125,
0.48046875,
0.5595703125,
0.48046875,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.51708984375,
0.5595703125,
0.7138671875,
0.5595703125,
0.7138671875,
0.5693359375,
0.51708984375,
0.5693359375
],
[
0.0882568359375,
0.57275390625,
0.48046875,
0.57275390625,
0.48046875,
0.58349609375,
0.0882568359375,
0.58349609375
],
[
0.53173828125,
0.57080078125,
0.91259765625,
0.57080078125,
0.91259765625,
0.583984375,
0.53173828125,
0.583984375
],
[
0.037567138671875,
0.57568359375,
0.049835205078125,
0.57568359375,
0.049835205078125,
0.5791015625,
0.037567138671875,
0.5791015625
],
[
0.0196075439453125,
0.57763671875,
0.058013916015625,
0.57763671875,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.0318603515625,
0.57470703125,
0.037567138671875,
0.57470703125,
0.037567138671875,
0.5869140625,
0.0318603515625,
0.5869140625
],
[
0.08660888671875,
0.583984375,
0.48193359375,
0.583984375,
0.48193359375,
0.59765625,
0.08660888671875,
0.59765625
],
[
0.5185546875,
0.5859375,
0.91259765625,
0.5859375,
0.91259765625,
0.5966796875,
0.5185546875,
0.5966796875
],
[
0.08416748046875,
0.5986328125,
0.4248046875,
0.5966796875,
0.4248046875,
0.60986328125,
0.08416748046875,
0.6123046875
],
[
0.51708984375,
0.59765625,
0.9140625,
0.59765625,
0.9140625,
0.611328125,
0.51708984375,
0.611328125
],
[
0.51708984375,
0.6123046875,
0.91259765625,
0.6123046875,
0.91259765625,
0.62548828125,
0.51708984375,
0.62548828125
],
[
0.0882568359375,
0.62548828125,
0.188720703125,
0.62548828125,
0.188720703125,
0.63720703125,
0.0882568359375,
0.63720703125
],
[
0.5185546875,
0.62744140625,
0.91259765625,
0.62744140625,
0.91259765625,
0.63818359375,
0.5185546875,
0.63818359375
],
[
0.0849609375,
0.64306640625,
0.48193359375,
0.64306640625,
0.48193359375,
0.65771484375,
0.0849609375,
0.65771484375
],
[
0.51708984375,
0.64013671875,
0.9150390625,
0.64013671875,
0.9150390625,
0.65380859375,
0.51708984375,
0.65380859375
],
[
0.0849609375,
0.65771484375,
0.27294921875,
0.65771484375,
0.27294921875,
0.671875,
0.0849609375,
0.671875
],
[
0.5185546875,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.66650390625,
0.5185546875,
0.66650390625
],
[
0.5185546875,
0.66943359375,
0.9111328125,
0.66943359375,
0.9111328125,
0.6796875,
0.5185546875,
0.6796875
],
[
0.0849609375,
0.6767578125,
0.2320556640625,
0.6767578125,
0.2320556640625,
0.68994140625,
0.0849609375,
0.68994140625
],
[
0.51708984375,
0.681640625,
0.9140625,
0.681640625,
0.9140625,
0.69580078125,
0.51708984375,
0.69580078125
],
[
0.0849609375,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.51708984375,
0.69580078125,
0.9140625,
0.69580078125,
0.9140625,
0.708984375,
0.51708984375,
0.708984375
],
[
0.0849609375,
0.7021484375,
0.48193359375,
0.70263671875,
0.48193359375,
0.71630859375,
0.0849609375,
0.71533203125
],
[
0.0849609375,
0.71533203125,
0.48193359375,
0.71533203125,
0.48193359375,
0.72900390625,
0.0849609375,
0.72900390625
],
[
0.51708984375,
0.71044921875,
0.9140625,
0.71044921875,
0.9140625,
0.7236328125,
0.51708984375,
0.7236328125
],
[
0.0849609375,
0.7265625,
0.48193359375,
0.72802734375,
0.48193359375,
0.7412109375,
0.0849609375,
0.74072265625
],
[
0.5185546875,
0.724609375,
0.91259765625,
0.724609375,
0.91259765625,
0.73486328125,
0.5185546875,
0.73486328125
],
[
0.08416748046875,
0.74072265625,
0.455810546875,
0.74072265625,
0.455810546875,
0.75390625,
0.08416748046875,
0.75390625
],
[
0.51953125,
0.73779296875,
0.91259765625,
0.73779296875,
0.91259765625,
0.74853515625,
0.51953125,
0.74853515625
],
[
0.51708984375,
0.75048828125,
0.91259765625,
0.75048828125,
0.91259765625,
0.7646484375,
0.51708984375,
0.7646484375
],
[
0.08660888671875,
0.7666015625,
0.200927734375,
0.7666015625,
0.200927734375,
0.7763671875,
0.08660888671875,
0.7763671875
],
[
0.5185546875,
0.7666015625,
0.91259765625,
0.7666015625,
0.91259765625,
0.7763671875,
0.5185546875,
0.7763671875
],
[
0.0849609375,
0.77587890625,
0.289306640625,
0.77587890625,
0.289306640625,
0.7890625,
0.0849609375,
0.7890625
],
[
0.51708984375,
0.77880859375,
0.9140625,
0.77880859375,
0.9140625,
0.79248046875,
0.51708984375,
0.79248046875
],
[
0.51708984375,
0.79248046875,
0.9140625,
0.79248046875,
0.9140625,
0.80615234375,
0.51708984375,
0.80615234375
],
[
0.0849609375,
0.80322265625,
0.48193359375,
0.80322265625,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.51708984375,
0.80615234375,
0.9140625,
0.80615234375,
0.9140625,
0.8193359375,
0.51708984375,
0.8193359375
],
[
0.0882568359375,
0.814453125,
0.48046875,
0.814453125,
0.48046875,
0.8212890625,
0.0882568359375,
0.8212890625
],
[
0.0849609375,
0.82373046875,
0.48193359375,
0.82373046875,
0.48193359375,
0.833984375,
0.0849609375,
0.833984375
],
[
0.5185546875,
0.8212890625,
0.9111328125,
0.8212890625,
0.9111328125,
0.83203125,
0.5185546875,
0.83203125
],
[
0.098876953125,
0.83349609375,
0.48193359375,
0.83349609375,
0.48193359375,
0.84326171875,
0.098876953125,
0.84326171875
],
[
0.5185546875,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.5185546875,
0.84765625
],
[
0.0849609375,
0.84130859375,
0.306396484375,
0.84228515625,
0.306396484375,
0.85400390625,
0.0849609375,
0.85302734375
],
[
0.09234619140625,
0.8359375,
0.101318359375,
0.8359375,
0.101318359375,
0.8388671875,
0.09234619140625,
0.8388671875
],
[
0.08660888671875,
0.85400390625,
0.27685546875,
0.85400390625,
0.27685546875,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.5185546875,
0.849609375,
0.9111328125,
0.849609375,
0.9111328125,
0.85986328125,
0.5185546875,
0.85986328125
],
[
0.0849609375,
0.86279296875,
0.47216796875,
0.86279296875,
0.47216796875,
0.876953125,
0.0849609375,
0.876953125
],
[
0.5185546875,
0.86279296875,
0.91259765625,
0.86279296875,
0.91259765625,
0.87353515625,
0.5185546875,
0.87353515625
],
[
0.08660888671875,
0.875,
0.295654296875,
0.875,
0.295654296875,
0.88525390625,
0.08660888671875,
0.88525390625
],
[
0.515625,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.515625,
0.8896484375
],
[
0.08416748046875,
0.88525390625,
0.273681640625,
0.8837890625,
0.273681640625,
0.89599609375,
0.08416748046875,
0.896484375
]
] |
[
"Expression SyntaxInformationBottleneck",
"for Math Word Problems",
"Jing Xiong",
"Chengming Li",
"Min Yang*",
"xiongj69@mail2.sysu.edu.cn",
"lichengming@mail.sysu.edu.cn",
"min.yang@siat.ac.cn",
"Sun Yat-sen University",
"SIAT, Chinese Academy of Sciences",
"Shenzhen, Guangdong, China",
"Sun Yat-sen University",
"Shenzhen, Guangdong, China",
" Shenzhen, Guangdong, China",
"Xiping Hu",
"Bin Hu",
"huxiping@mail.sysu.edu.cn",
"bh@lzu.edu.cn",
"Sun Yat-sen University",
"Lanzhou University",
"Shenzhen, Guangdong, China",
"1",
"Lanzhou, Gansu, China",
"ABSTRACT",
"1",
"",
"INTRODUCTION",
"Math Word Problems (MWP) aims to automatically solve mathe",
"Math Word Problems (MWP) is challenging and draws much atten-",
"matical questions given in texts.Previous studies tend to design",
"tion from researchers in the field of natural language processing",
"complex models to capture additional information in the original",
"[3, 25] and information retrieval (e.g.,mathematical understanding)",
"",
"text so as to enable the model to gain more comprehensivefeatures.",
"[8, 26]. MWP aims to automatically answer mathematical questions",
"In this paper,we turn our attention in the oppositedirection.and",
"given in a natural language, which requires the model not only",
"understand what facts arepresented in a text.but alsopossessthe",
"",
"work onhowto discard redundantfeaturescontaining spurious",
"['s]",
"correlationsforMWP.Tothis end, wedesign anExpressionSyntax",
"reasoning capability to answer the mathematical question. Table 1",
"Information Bottleneck method for MWP (called ESIB) based on",
"shows three examples of MWP with three mathematical problems",
"variationalinformationbottleneck,whichextractsessentialfeatures",
"and their solutionexpressionswithanswer",
"of expressionsyntaxtreewhilefilteringlatent-specificredundancy",
"Inspired by the success of deep learning [14, 21], the encoder-",
"containing syntax-irrelevant features. The key idea of ESIB is to en-",
"decoder framework with attention mechanisms [2] have been dom-",
"couragemultiplemodelstopredictthesameexpressionsyntaxtree",
"inated inMWP[18-20],whichbringthestate-of-the-arttoanew",
"fordifferentproblemrepresentations of the sameproblembymu-",
"level. The key idea is to use an encoder to learn representations",
"of problem text and employ a decoder to generate the correspond-",
"99",
"tual learning so as to capture consistent information of expression",
"syntax tree and discard latent-specific redundancy. To improve the",
"ing solution expressionand answer.Subsequently,several studies",
"generalization ability of the model and generate more diverse ex-",
"propose sequence-to-tree models, which explore the tree structure",
"pressions, we design a self-distillation loss to encourage the model",
"information presented in the text and improve the generation of",
"to relymoreontheexpression syntax information in the latent",
"solutionexpressions[23,24,28].",
"space.Experimentalresultsontwolarge-scalebenchmarksshow",
"However, the previous MWwP methods appear to rely on spurious",
"",
"OZAX",
"",
"that our model not only achieves state-of-the-art results but also",
"correlations between the shallow heuristics in problem and solu-",
"generates more diverse solutions.The code is available.!",
"tion expression [15]. For example, as shown in Table 1, previous",
"models may associate Problem 1 and Problem 2 with the mathe",
"KEYWORDS",
"matical formula\"xxy-z\",sincethesetwoproblemshavesimilar",
"Math Word Problems, Mutual learning, Spurious correlations, Vari-",
"semantic patterns, e.g., calculating the speed. Based on this asso-",
"ational information bottleneck",
"ciation, the models could generate wrong solution expression for",
"Problem3whichhassimilarsemanticproblemexpressionlikethe",
"ACM Reference Format:",
"text segment \"place A to place B\" in Problems 1-2. In particular,",
"Jing Xiong, Chengming Li, Min Yang, Xiping Hu, and Bin Hu. 2022. Expres-",
"the models that learn spurious correlations are more likely to gen-",
"sion Syntax Information Bottleneck for Math Word Problems. In Proceedings",
"of the 45th International ACM SIGIR Conference on Research and Development",
"erate wrong solution expression\"220× 25%-30%\",rather than",
"in Information Retrieval (SIGIR22),July11-15,2022,Madrid,Spain.ACM",
"220(25%+30%)\"forProblem3.Wedefinesuchafalseassociation",
"New York, NY, USA,6 pages.https://doi.org/10.1145/3477495.3531824",
"as spurious correlation. Some recent studies have revealed that",
"MWP solvers relying on spurious correlations could achieve high",
"'Corresponding authors",
"accuracy[9,15].Thesemodelscan evencomputecorrectanswers",
"1https://github.com/menik1126/math_ESIB",
"without paying attention to the question part in the problem such as",
"the text segment \"how many kilometers is the total length of the two",
"Permission tomakedigital orhard copies of part orall of thiswork forpersonal or",
"places?\" in Problem 3 calculating the distance. In addition, the solu-",
"classroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributed",
"for profit or commercial advantage and that copies bear this notice and the full citation",
"tion expression is sensitiveto the perturbed latent representations",
"the first page. Copyrights for third-party components of this work must be honored.",
"[1o], since the semantically similar mathematical problems, even",
"For all other uses, contact the owner/author(s)",
"17",
"SIGIR22,July11-15,2022,Madrid,Spain",
"withtotallydifferent solution expressionsand questions,canbe",
"@ 2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.",
"encoded closely in the latent space.We believe it is an evidence that",
"ACMISBN978-1-4503-8732-3/22/07...$15.00",
"redundant information containing spurious correlations is encoded",
"https://doi.org/10.1145/3477495.3531824"
] |
{
"arxiv_id": "2310.15664",
"language": "en",
"page_idx": 0
}
|
||
2310.15664-en-1
|
2310.15664-en-1.jpg
|
Table 1: Three examples of MWP.
<table>
<tbody>
<tr>
<td>Problem 1: From place A to place B, if a bicy per hour, it can be reached in 4 hours. If it on how many kilometers per hour does the car t</td>
<td>3,if a bicycle travels 16 kilometers urs. If it only takes 2 hours by car, s the car travel?</td>
</tr>
<tr>
<td>I accelerated the speed due to the task. It only the first place. What was the speed when I re</td>
<td>sk. It only took 4 hours to return to when I returned?</td>
</tr>
<tr>
<td>the first place. What was the speed when I re</td>
<td>when I returned?</td>
</tr>
<tr>
<td>4T</td>
<td>Answer:105</td>
</tr>
<tr>
<td>$\textbf{Solution Expression 3: 220( 25\% + 30\% ) }$</td>
<td>${\mathcal Z}_{0}+30\%)$ $\mathbf{Answer:400}$</td>
</tr>
<tr>
<td>$\mathbf{WrongSolutionExpression~3:220\times25\%}\div$</td>
<td>$20\times25\%\div30\%$</td>
</tr>
</tbody>
</table>
in the latent representation. Therefore, it is necessary to alleviate the spurious correlations by compressing the latent representations for math expressions while filtering latent-specific redundancy.
To solve the above challenges, we design a Expression Syntax Information Bottleneck method for MWP based on variational information bottleneck (VIB) [1], which aims to discard redundant information containing spurious correlations [7,13]. Our key idea is to encourage multiple models to predict the same math expression with different problem representations of the same problem so as to capture consistent information about expression syntax tree in expressions and discard latent-specific redundancy containing syntax-irrelevant information. In addition, we leverage mutual learning [29] for learning variational information bottleneck, which can effectively reduce the latent-specific redundancy. Inspired by the observation that there are usually multiple solutions to solve a problem, we also design a self-distillation loss which encourages the decoder to rely on the syntax information in latent space, enabling the model to generate diverse solutions.
We summarize our main contributions as follows.(1) We are the first to reduce spurious correlations so as to improve the performance of MWP. (2) We propose a novel expression syntax information bottleneck method for MWP, which extracts essential syntax information of math expression and filters redundant information containing spurious correlations. (3) We design a self-distillation loss to encourage the model to generate more diverse solution expressions. (4) Extensive experiments on two benchmark datasets show that our model outperforms the strong baselines in a noticeable margin.
# 2 метноросоєт
A math word problems (MWP) can be denoted by a projection $F:$ $x\mapsto z\mapsto y$, where $x=\{w_1,w_2,\ldots,w_m\}$ is the problem sequence with $m$ words, $z$ is the compressed representation of the original problem $x$ and $y=\{o_1,o_2,\ldots,o_n\}$ is the solution expression of the problem with $n$ words. The goal of MWP is to establish a model $F$ which generates a correct solution expression $y$ and calculates the correct answer for the problem $x.$
<FigureHere>
Figure 1: Overview of the proposed method ESIB
As illustrated in Figure 1, the proposed $ESIB$ is composed of a source network (denoted as SN) and a collaborator network (denoted as CN). The two networks are optimised collaboratively and capture consistent information across different problem representations throughout the training process. We use the deep variational information bottleneck (VIB) framework as the backbone of both SN and CN. The VIB aims to generate problem representation $z\in\mathbb{R}^s$ by compressing and discarding redundant information in latent representation $v\in\mathbb{R}^d$ without reducing essential information related to the target $y$, where $d$ denotes hidden size of decoder and s denotes the dimension of problem representation. So the above projection $F$ can be rewritten as $F:x\mapsto v\mapsto z\mapsto y$,where $v$ denotes latent representation and $z$ denotes problem representation sampled from $e^\mu(v)\sim N\left(e^\mu(v),e^\sigma(v)\right)$ (e denotes a dense layer).
2.1 Encoder-Compressor-Decoder Architecture
$Encoder.$ We adopt the RoBERTa model [11] as our encoder. We pass the problem sequence $x$ into the RoBERTa model and obtain latent representation $v$ ([CLS] vector output by the pre-trained language model (PLM) and its dimension is converted from 768 to $d$ by a dense layer). In order to model the relationship between the quantities in the PLM, we set up a learnable quantity embedding matrix $Q_t= \{ q_1, q_2, \ldots , q_n\} $,similar to the learnable position embedding in BERT [5]. When passing the sequence $x$ into the encoder, we replace the each quantity in the sequence $x$ (i.e., the numbers in the problem) with a embedding $q_i\in Q_t.$
Compressor. Our model takes the encoder output $v$ and feeds it into a variational information bottleneck [1] module which outputs a sampled vector $z.$ This part removes redundant information by optimizing a variational upper bound $\mathcal{V}_IB.$ We will detail how to optimize $\mathcal{V}_IB$ in Section 2.2.
$Decoder.$ Our decoder follows the GTS model [24]. We use the sampled representation $z$ to initialize the initial state of the decoder and the recursive operation of the decoder to construct $y$ by the order of pre-order traversal. First, the root node $q_\mathrm{root}$ (middle $\hat{\text{operator part) is first generated. Then, we generate the left child}}$
|
[
[
0.1707763671875,
0.10479736328125,
0.393798828125,
0.10479736328125,
0.393798828125,
0.11871337890625,
0.1707763671875,
0.11871337890625
],
[
0.67724609375,
0.106689453125,
0.74609375,
0.10797119140625,
0.74609375,
0.11932373046875,
0.67724609375,
0.11871337890625
],
[
0.09234619140625,
0.1363525390625,
0.468017578125,
0.1363525390625,
0.468017578125,
0.146484375,
0.09234619140625,
0.146484375
],
[
0.0931396484375,
0.146484375,
0.47216796875,
0.147705078125,
0.47216796875,
0.1610107421875,
0.0931396484375,
0.160400390625
],
[
0.7587890625,
0.143310546875,
0.76708984375,
0.143310546875,
0.76708984375,
0.1483154296875,
0.7587890625,
0.1483154296875
],
[
0.09478759765625,
0.1610107421875,
0.370849609375,
0.1610107421875,
0.370849609375,
0.1717529296875,
0.09478759765625,
0.1717529296875
],
[
0.63623046875,
0.1610107421875,
0.64892578125,
0.1610107421875,
0.64892578125,
0.17041015625,
0.63623046875,
0.17041015625
],
[
0.775390625,
0.1610107421875,
0.78173828125,
0.1666259765625,
0.775390625,
0.1717529296875,
0.7685546875,
0.1666259765625
],
[
0.09234619140625,
0.17236328125,
0.281005859375,
0.17236328125,
0.281005859375,
0.186279296875,
0.09234619140625,
0.186279296875
],
[
0.285888671875,
0.173583984375,
0.35693359375,
0.173583984375,
0.35693359375,
0.1849365234375,
0.285888671875,
0.1849365234375
],
[
0.6259765625,
0.173583984375,
0.6396484375,
0.173583984375,
0.6396484375,
0.184326171875,
0.6259765625,
0.184326171875
],
[
0.74658203125,
0.1749267578125,
0.75732421875,
0.1749267578125,
0.75732421875,
0.184326171875,
0.74658203125,
0.184326171875
],
[
0.7939453125,
0.1749267578125,
0.80810546875,
0.1749267578125,
0.80810546875,
0.184326171875,
0.7939453125,
0.184326171875
],
[
0.55810546875,
0.18310546875,
0.57861328125,
0.18310546875,
0.57861328125,
0.1912841796875,
0.55810546875,
0.1912841796875
],
[
0.69384765625,
0.18115234375,
0.7138671875,
0.18115234375,
0.7138671875,
0.184326171875,
0.69384765625,
0.184326171875
],
[
0.85302734375,
0.18310546875,
0.85888671875,
0.18310546875,
0.85888671875,
0.189453125,
0.85302734375,
0.189453125
],
[
0.54736328125,
0.1849365234375,
0.5537109375,
0.1849365234375,
0.5537109375,
0.189453125,
0.54736328125,
0.189453125
],
[
0.0931396484375,
0.1849365234375,
0.47216796875,
0.1849365234375,
0.47216796875,
0.1988525390625,
0.0931396484375,
0.1988525390625
],
[
0.54150390625,
0.1976318359375,
0.583984375,
0.1976318359375,
0.583984375,
0.2052001953125,
0.54150390625,
0.2052001953125
],
[
0.72119140625,
0.1976318359375,
0.8310546875,
0.199462890625,
0.8310546875,
0.2108154296875,
0.72119140625,
0.208984375
],
[
0.8505859375,
0.1962890625,
0.88134765625,
0.1962890625,
0.88134765625,
0.203857421875,
0.8505859375,
0.203857421875
],
[
0.6044921875,
0.199462890625,
0.6845703125,
0.199462890625,
0.6845703125,
0.210205078125,
0.6044921875,
0.210205078125
],
[
0.096435546875,
0.199462890625,
0.471435546875,
0.199462890625,
0.471435546875,
0.210205078125,
0.096435546875,
0.210205078125
],
[
0.09478759765625,
0.212158203125,
0.471435546875,
0.212158203125,
0.471435546875,
0.222900390625,
0.09478759765625,
0.222900390625
],
[
0.67724609375,
0.2108154296875,
0.73876953125,
0.2108154296875,
0.73876953125,
0.2183837890625,
0.67724609375,
0.2183837890625
],
[
0.56787109375,
0.2183837890625,
0.58837890625,
0.2183837890625,
0.58837890625,
0.2310791015625,
0.56787109375,
0.2310791015625
],
[
0.6259765625,
0.2215576171875,
0.64794921875,
0.2215576171875,
0.64794921875,
0.2342529296875,
0.6259765625,
0.2342529296875
],
[
0.83642578125,
0.2171630859375,
0.8544921875,
0.2171630859375,
0.8544921875,
0.23486328125,
0.83642578125,
0.23486328125
],
[
0.09478759765625,
0.2247314453125,
0.3798828125,
0.2247314453125,
0.3798828125,
0.23486328125,
0.09478759765625,
0.23486328125
],
[
0.78515625,
0.2247314453125,
0.80615234375,
0.2247314453125,
0.80615234375,
0.23291015625,
0.78515625,
0.23291015625
],
[
0.611328125,
0.231689453125,
0.6748046875,
0.231689453125,
0.6748046875,
0.246826171875,
0.611328125,
0.246826171875
],
[
0.75244140625,
0.2310791015625,
0.81396484375,
0.2310791015625,
0.81396484375,
0.24560546875,
0.75244140625,
0.24560546875
],
[
0.09234619140625,
0.238037109375,
0.279296875,
0.238037109375,
0.279296875,
0.248779296875,
0.09234619140625,
0.248779296875
],
[
0.291748046875,
0.238037109375,
0.361083984375,
0.238037109375,
0.361083984375,
0.248779296875,
0.291748046875,
0.248779296875
],
[
0.8349609375,
0.23291015625,
0.84716796875,
0.24365234375,
0.83837890625,
0.250732421875,
0.826171875,
0.240478515625
],
[
0.5849609375,
0.242431640625,
0.59326171875,
0.242431640625,
0.59326171875,
0.246826171875,
0.5849609375,
0.246826171875
],
[
0.0931396484375,
0.25,
0.4697265625,
0.25,
0.4697265625,
0.26318359375,
0.0931396484375,
0.26318359375
],
[
0.62841796875,
0.246826171875,
0.64208984375,
0.246826171875,
0.64208984375,
0.255615234375,
0.62841796875,
0.255615234375
],
[
0.6259765625,
0.255126953125,
0.6708984375,
0.2568359375,
0.6689453125,
0.272705078125,
0.625,
0.270751953125
],
[
0.75732421875,
0.2568359375,
0.80126953125,
0.2568359375,
0.80126953125,
0.271484375,
0.75732421875,
0.271484375
],
[
0.096435546875,
0.26318359375,
0.4697265625,
0.26318359375,
0.4697265625,
0.2734375,
0.096435546875,
0.2734375
],
[
0.096435546875,
0.27587890625,
0.471435546875,
0.27587890625,
0.471435546875,
0.285888671875,
0.096435546875,
0.285888671875
],
[
0.68310546875,
0.28271484375,
0.705078125,
0.285400390625,
0.70361328125,
0.292236328125,
0.68115234375,
0.290283203125
],
[
0.73779296875,
0.2822265625,
0.82275390625,
0.2822265625,
0.82275390625,
0.29345703125,
0.73779296875,
0.29345703125
],
[
0.60107421875,
0.2841796875,
0.6748046875,
0.2841796875,
0.6748046875,
0.291748046875,
0.60107421875,
0.291748046875
],
[
0.0931396484375,
0.287353515625,
0.2303466796875,
0.287353515625,
0.2303466796875,
0.29736328125,
0.0931396484375,
0.29736328125
],
[
0.0906982421875,
0.2998046875,
0.33251953125,
0.2998046875,
0.33251953125,
0.313232421875,
0.0906982421875,
0.313232421875
],
[
0.3447265625,
0.3017578125,
0.418212890625,
0.3017578125,
0.418212890625,
0.313232421875,
0.3447265625,
0.313232421875
],
[
0.7685546875,
0.29736328125,
0.79345703125,
0.29736328125,
0.79345703125,
0.304931640625,
0.7685546875,
0.304931640625
],
[
0.60107421875,
0.306884765625,
0.65185546875,
0.306884765625,
0.65185546875,
0.31005859375,
0.60107421875,
0.31005859375
],
[
0.7197265625,
0.3125,
0.8388671875,
0.3125,
0.8388671875,
0.32568359375,
0.7197265625,
0.32568359375
],
[
0.7587890625,
0.306884765625,
0.76953125,
0.306884765625,
0.76953125,
0.31005859375,
0.7587890625,
0.31005859375
],
[
0.09478759765625,
0.314453125,
0.366943359375,
0.314453125,
0.366943359375,
0.324462890625,
0.09478759765625,
0.324462890625
],
[
0.609375,
0.314453125,
0.67626953125,
0.314453125,
0.67626953125,
0.32568359375,
0.609375,
0.32568359375
],
[
0.55126953125,
0.3447265625,
0.8759765625,
0.3447265625,
0.8759765625,
0.35791015625,
0.55126953125,
0.35791015625
],
[
0.0849609375,
0.375732421875,
0.48046875,
0.375732421875,
0.48046875,
0.389404296875,
0.0849609375,
0.389404296875
],
[
0.53369140625,
0.380126953125,
0.9140625,
0.380126953125,
0.9140625,
0.39453125,
0.53369140625,
0.39453125
],
[
0.08660888671875,
0.39013671875,
0.4794921875,
0.39013671875,
0.4794921875,
0.40087890625,
0.08660888671875,
0.40087890625
],
[
0.51708984375,
0.39453125,
0.9140625,
0.39453125,
0.9140625,
0.40771484375,
0.51708984375,
0.40771484375
],
[
0.08660888671875,
0.40478515625,
0.465576171875,
0.40478515625,
0.465576171875,
0.415283203125,
0.08660888671875,
0.415283203125
],
[
0.515625,
0.4091796875,
0.9140625,
0.40771484375,
0.9140625,
0.421630859375,
0.515625,
0.42236328125
],
[
0.101318359375,
0.417236328125,
0.48193359375,
0.417236328125,
0.48193359375,
0.43115234375,
0.101318359375,
0.43115234375
],
[
0.515625,
0.42236328125,
0.9140625,
0.421630859375,
0.9140625,
0.43505859375,
0.515625,
0.436279296875
],
[
0.0849609375,
0.43115234375,
0.48193359375,
0.43115234375,
0.48193359375,
0.4443359375,
0.0849609375,
0.4443359375
],
[
0.51708984375,
0.436279296875,
0.9140625,
0.436279296875,
0.9140625,
0.449462890625,
0.51708984375,
0.449462890625
],
[
0.0849609375,
0.443115234375,
0.48046875,
0.4443359375,
0.48046875,
0.458251953125,
0.0849609375,
0.45703125
],
[
0.51708984375,
0.45068359375,
0.9140625,
0.45068359375,
0.9140625,
0.464111328125,
0.51708984375,
0.464111328125
],
[
0.08416748046875,
0.458984375,
0.48193359375,
0.458984375,
0.48193359375,
0.472900390625,
0.08416748046875,
0.472900390625
],
[
0.5185546875,
0.464111328125,
0.91259765625,
0.464111328125,
0.91259765625,
0.477783203125,
0.5185546875,
0.477783203125
],
[
0.0849609375,
0.472900390625,
0.48193359375,
0.472900390625,
0.48193359375,
0.486083984375,
0.0849609375,
0.486083984375
],
[
0.51708984375,
0.477783203125,
0.91259765625,
0.477783203125,
0.91259765625,
0.4912109375,
0.51708984375,
0.4912109375
],
[
0.0849609375,
0.486083984375,
0.48193359375,
0.486083984375,
0.48193359375,
0.5,
0.0849609375,
0.5
],
[
0.515625,
0.492431640625,
0.9140625,
0.4912109375,
0.9140625,
0.505859375,
0.515625,
0.5068359375
],
[
0.0849609375,
0.50048828125,
0.48193359375,
0.50048828125,
0.48193359375,
0.5146484375,
0.0849609375,
0.5146484375
],
[
0.51708984375,
0.5068359375,
0.9140625,
0.5068359375,
0.9140625,
0.52099609375,
0.51708984375,
0.52099609375
],
[
0.08660888671875,
0.51513671875,
0.48193359375,
0.51513671875,
0.48193359375,
0.52587890625,
0.08660888671875,
0.52587890625
],
[
0.5185546875,
0.521484375,
0.9111328125,
0.521484375,
0.9111328125,
0.5322265625,
0.5185546875,
0.5322265625
],
[
0.08416748046875,
0.52783203125,
0.48193359375,
0.5263671875,
0.48193359375,
0.54052734375,
0.08416748046875,
0.54150390625
],
[
0.51708984375,
0.53515625,
0.9140625,
0.53515625,
0.9140625,
0.548828125,
0.51708984375,
0.548828125
],
[
0.08660888671875,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.5537109375,
0.08660888671875,
0.5537109375
],
[
0.51708984375,
0.548828125,
0.9140625,
0.548828125,
0.9140625,
0.5625,
0.51708984375,
0.5625
],
[
0.0849609375,
0.55615234375,
0.48193359375,
0.55615234375,
0.48193359375,
0.5693359375,
0.0849609375,
0.5693359375
],
[
0.0849609375,
0.568359375,
0.48046875,
0.568359375,
0.48046875,
0.58203125,
0.0849609375,
0.58203125
],
[
0.51708984375,
0.5625,
0.90673828125,
0.5625,
0.90673828125,
0.57568359375,
0.51708984375,
0.57568359375
],
[
0.0849609375,
0.583984375,
0.48193359375,
0.583984375,
0.48193359375,
0.59765625,
0.0849609375,
0.59765625
],
[
0.51953125,
0.5888671875,
0.55615234375,
0.5888671875,
0.55615234375,
0.599609375,
0.51953125,
0.599609375
],
[
0.5498046875,
0.5869140625,
0.9111328125,
0.5869140625,
0.9111328125,
0.6015625,
0.5498046875,
0.6015625
],
[
0.0849609375,
0.59765625,
0.48193359375,
0.59765625,
0.48193359375,
0.611328125,
0.0849609375,
0.611328125
],
[
0.0849609375,
0.611328125,
0.32275390625,
0.611328125,
0.32275390625,
0.6220703125,
0.0849609375,
0.6220703125
],
[
0.53515625,
0.60595703125,
0.9140625,
0.60595703125,
0.9140625,
0.619140625,
0.53515625,
0.619140625
],
[
0.515625,
0.619140625,
0.9140625,
0.61865234375,
0.9140625,
0.63330078125,
0.515625,
0.6337890625
],
[
0.10296630859375,
0.62548828125,
0.48046875,
0.62548828125,
0.48046875,
0.63623046875,
0.10296630859375,
0.63623046875
],
[
0.0849609375,
0.63818359375,
0.483642578125,
0.63818359375,
0.483642578125,
0.6513671875,
0.0849609375,
0.6513671875
],
[
0.51708984375,
0.6337890625,
0.9140625,
0.6337890625,
0.9140625,
0.64794921875,
0.51708984375,
0.64794921875
],
[
0.515625,
0.646484375,
0.9140625,
0.64599609375,
0.9140625,
0.66015625,
0.515625,
0.6611328125
],
[
0.08660888671875,
0.65380859375,
0.48193359375,
0.65380859375,
0.48193359375,
0.6640625,
0.08660888671875,
0.6640625
],
[
0.08416748046875,
0.66552734375,
0.48193359375,
0.66650390625,
0.48193359375,
0.6796875,
0.08416748046875,
0.6787109375
],
[
0.51708984375,
0.66015625,
0.91259765625,
0.6611328125,
0.91259765625,
0.6748046875,
0.51708984375,
0.673828125
],
[
0.08416748046875,
0.6796875,
0.48193359375,
0.6796875,
0.48193359375,
0.693359375,
0.08416748046875,
0.693359375
],
[
0.5185546875,
0.6767578125,
0.91259765625,
0.6767578125,
0.91259765625,
0.6875,
0.5185546875,
0.6875
],
[
0.5185546875,
0.689453125,
0.9140625,
0.689453125,
0.9140625,
0.7041015625,
0.5185546875,
0.7041015625
],
[
0.08660888671875,
0.69580078125,
0.48046875,
0.69580078125,
0.48046875,
0.70556640625,
0.08660888671875,
0.70556640625
],
[
0.0849609375,
0.70849609375,
0.483642578125,
0.70849609375,
0.483642578125,
0.7216796875,
0.0849609375,
0.7216796875
],
[
0.51708984375,
0.70263671875,
0.9140625,
0.70263671875,
0.9140625,
0.71630859375,
0.51708984375,
0.71630859375
],
[
0.0849609375,
0.7216796875,
0.48193359375,
0.7216796875,
0.48193359375,
0.73486328125,
0.0849609375,
0.73486328125
],
[
0.51708984375,
0.71728515625,
0.9140625,
0.71728515625,
0.9140625,
0.73095703125,
0.51708984375,
0.73095703125
],
[
0.08416748046875,
0.73486328125,
0.48193359375,
0.73486328125,
0.48193359375,
0.74853515625,
0.08416748046875,
0.74853515625
],
[
0.515625,
0.72998046875,
0.75732421875,
0.73095703125,
0.75732421875,
0.74560546875,
0.515625,
0.744140625
],
[
0.0849609375,
0.75048828125,
0.160888671875,
0.75048828125,
0.160888671875,
0.76220703125,
0.0849609375,
0.76220703125
],
[
0.53515625,
0.75048828125,
0.9140625,
0.75048828125,
0.9140625,
0.7646484375,
0.53515625,
0.7646484375
],
[
0.51708984375,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.51708984375,
0.77880859375
],
[
0.10455322265625,
0.78076171875,
0.253173828125,
0.78076171875,
0.253173828125,
0.79541015625,
0.10455322265625,
0.79541015625
],
[
0.5185546875,
0.78076171875,
0.9111328125,
0.78076171875,
0.9111328125,
0.79150390625,
0.5185546875,
0.79150390625
],
[
0.51708984375,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.51708984375,
0.806640625
],
[
0.0849609375,
0.7822265625,
0.111083984375,
0.7822265625,
0.111083984375,
0.79443359375,
0.0849609375,
0.79443359375
],
[
0.0849609375,
0.7998046875,
0.483642578125,
0.7998046875,
0.483642578125,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.51708984375,
0.806640625,
0.68701171875,
0.806640625,
0.68701171875,
0.82080078125,
0.51708984375,
0.82080078125
],
[
0.0849609375,
0.814453125,
0.48193359375,
0.814453125,
0.48193359375,
0.8291015625,
0.0849609375,
0.8291015625
],
[
0.0849609375,
0.82763671875,
0.48193359375,
0.82763671875,
0.48193359375,
0.84130859375,
0.0849609375,
0.84130859375
],
[
0.53515625,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.53515625,
0.84033203125
],
[
0.0849609375,
0.84228515625,
0.48193359375,
0.84228515625,
0.48193359375,
0.85595703125,
0.0849609375,
0.85595703125
],
[
0.515625,
0.84228515625,
0.9140625,
0.84228515625,
0.9140625,
0.85595703125,
0.515625,
0.85595703125
],
[
0.0849609375,
0.85595703125,
0.48193359375,
0.85595703125,
0.48193359375,
0.869140625,
0.0849609375,
0.869140625
],
[
0.51708984375,
0.85595703125,
0.9140625,
0.85595703125,
0.9140625,
0.869140625,
0.51708984375,
0.869140625
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.51708984375,
0.869140625,
0.9140625,
0.869140625,
0.9140625,
0.8837890625,
0.51708984375,
0.8837890625
],
[
0.08660888671875,
0.88525390625,
0.289306640625,
0.88525390625,
0.289306640625,
0.89599609375,
0.08660888671875,
0.89599609375
],
[
0.51708984375,
0.8837890625,
0.9140625,
0.8837890625,
0.9140625,
0.89794921875,
0.51708984375,
0.89794921875
]
] |
[
"Table 1: Three examples of MWP.",
"Ground-truth",
"Problem1:From place Ato placeB,ifa bicycle travels 16kilometers",
"per hour, it can be reached in 4 hours. If it only takes 2 hours by car,",
"",
"how many kilometers per hour does the car travel?",
"",
"X",
"SolutionExpression1:16x4+2",
"Answer:32",
"D",
"0",
"0",
"arding",
"aA",
"is",
"",
"Problem 2: Uncle Jack drove from place A to place B,and it took 6",
"information",
"CollaboratorNetwork",
"aformatio",
"Source Network",
"hours to arrive at the speed of 7o kilometers per hour. When I returned,",
"I accelerated the speed due to the task. It only took 4 hours to return to",
"TreeDecode",
"hi",
"311",
"h",
"the first place. What was the speed when I returned?",
"Z2",
"Compressor",
"Compressor",
"SolutionExpression2:70x6-4",
"Answer:105",
"VSF",
"",
"Problem 3: A car travels 25% of the whole journey from place A to",
"Vi",
"Encoder",
"Encoder",
"place B in the first hour, 30% of thewhole journey in the second hour,a",
"total of 22o kilometers in two hours, how manykilometers is thetotal",
"terspet",
"Vehicle A travels NUM2]",
"ehicleAtravels 70kilot",
"lengthofthetwoplaces?",
"SolutionExpression3:220(25%+30%)",
"Answer:400",
"INUMI",
"Hovmamwhur",
"ReplaceQuantity Token",
"tine",
"WrongSolutionExpression3:220×25%+30%",
"MWPTextx",
"Figure 1: Overview of the proposed method ESIB",
"in the latent representation. Therefore, it is necessary to alleviate",
"As illustrated in Figure 1, the proposed ESIB is composed of a",
"thespurious correlationsbycompressingthelatentrepresentations",
"source network (denoted as SN) and a collaborator network (de-",
"formathexpressions whilefiltering latent-specificredundancy.",
"noted as CN).The twonetworks are optimised collaboratively and",
"To solve the above challenges, we design a Expression Syntax",
"captureconsistent information across different problem representa-",
"Information Bottleneck method for MWP based on variational in-",
"tions throughout the training process. We use the deep variational",
"formation bottleneck (VIB) [1], which aims to discard redundant",
"information bottleneck (VIB)framework as thebackbone ofboth SN",
"information containing spurious correlations [7, 13]. Our key idea",
"and CN. The VIB aims to generate problem representation z e Rs",
"is to encourage multiple models to predict the same math expres-",
"by compressing and discarding redundant information in latent",
"sion with different problem representations of the same problem",
"representation e Rd without reducing essential information re-",
"so as to capture consistent information about expression syntax",
"lated to the target y, where d denotes hidden size of decoder and",
"tree in expressions and discardlatent-specific redundancy contain-",
"sdenotesthedimensionof problemrepresentation.Sotheabove",
"ing syntax-irrelevant information. In addition, we leverage mutual",
"projection F can be rewritten as F : x - u z y, where u",
"learning[29]forlearningvariationalinformationbottleneck,which",
"denotes latent representation and z denotes problem representation",
"can effectively reduce the latent-specific redundancy. Inspired by",
"the observation that there are usually multiple solutions to solve a",
"sampled from e\"(o) ~ N (e(), e (o) (e denotes a dense layer).",
"problem, we also design a self-distillation loss which encourages the",
"2.1",
"Encoder-Compressor-Decoder Architecture",
"decoder to rely on the syntax information in latent space, enabling",
"the model to generate diverse solutions",
"Encoder We adopt the RoBERTa model [11] as our encoder. We",
"pass the problem sequence x into the RoBERTa model and obtain",
"We summarize our main contributions as follows. (1) Weare the",
"first to reduce spurious correlations so as to improve the perfor-",
"latent representation ([CLS] vector output by the pre-trained lan-",
"guage model (PLM) and its dimension is converted from 768 to d by",
"manceofMWP.(2)Weproposeanovelexpressionsyntaxinforma-",
"tion bottleneck method for MWP, which extracts essential syntax",
"a dense layer).In order to model the relationship between the quan",
"information of math expression and filters redundant information",
"tities in thePLM, we set up a learnablequantity embedding matrix",
"Q, = (qi,q2..+-,qn), similar to thelearnable position embedding",
"containing spurious correlations.(3)We design a self-distillation",
"loss to encourage the model to generate more diverse solution ex-",
"in BERT [5]. When passing the sequence x into the encoder, we",
"pressions. (4) Extensive experiments on two benchmark datasets",
"replace the each quantity in the sequence x (i.e., the numbers in",
"show that our model outperforms the strong baselines in a notice-",
"the problem) with a embedding qi e Qr",
"ablemargin",
"Compressor. Our model takes the encoder output u and feeds it",
"into a variational information bottleneck [1] module which outputs",
"METHODOLOGY",
"a sampled vectorz.Thispart removesredundant information by",
"optimizing a variational upper bound ViB.We will detail how to",
"2",
"A math word problems (MWP) can be denoted by a projection F :",
"optimize VrB in Section 2.2.",
"x z - y, where x = ( wi, w2...., wm) is the problem sequence",
"with m words, z is the compressed representation of the original",
"Decoder. Our decoder follows the GTS model [24]. We use the",
"problem x and y = (o1,02,-**,on) is the solution expression of the",
"sampled representation z to initialize the initial state of the de-",
"problem with n words. The goal of MWP is to establish a model F",
"coder and the recursive operation of the decoder to construct y by",
"which generates a correct solution expression y and calculates the",
"the order of pre-order traversal. First, the root node qroot (middle",
"correctanswerfortheproblemx.",
"operator part) is first generated. Then, we generate the left child"
] |
{
"arxiv_id": "2310.15664",
"language": "en",
"page_idx": 1
}
|
||
2310.15664-en-2
|
2310.15664-en-2.jpg
|
node $q_l.$ Finally, we generate the right child node $q_r.$ This process has been iterated until the leaf nodes are generated. The attention mechanism is applied to learn the global context vector $G_i$ which is utilized to generate the current node token $\hat{y}_i.$ Here we denote the digital embedding after being encoded by the encoder as $Q.$ Mathematically, we define the attention mechanism as follows:
(1)
$$G_i=\begin{cases}&\text{Attention}\:(x,q_{\text{root}},q_l),\:q_l\notin\emptyset.\\&\text{Attention}\:(x,q_{\text{root}},q_{sl}),\:q_{sl}\notin\emptyset.\\&\text{Attention}\:(x,q_{\text{root}}),\:q_l,q_{sl}\in\emptyset.\end{cases}$$
$$\hat{y}_{i},\hat{h}_{i}=\mathrm{Predict}(G_{i},Q),$$
(2)
where Predict(·) is prediction layer for producing tree nodes $\hat{y}_{i}$ and
hidden state of decoding step $i$ (denoted as $\hat{h}_i\in\mathbb{R}^d).$
If the current node is an operator, we will generate the left and right child nodes and push them into the stack in the tree decoder according to the top-down method.If it is a number, we will perform the merge operation until the leaf nodes in the stack pop out, and the result of the merge is pushed into the left child node stack for attention operation. The merge operation will pop the required node $q_\mathrm{op}$ and $q_\mathrm{subtree}$ from an embedding stack. Finally, the merge operation outputs the answer of the mathematical expression. This recursive construction process can be defined as follows:
(3)
(4)
$$q_{l}=\mathrm{Left}(\hat{G},\hat{y}_{i},q_{\mathrm{root}}).\\q_{r}=\mathrm{Right}(\hat{G},\hat{y}_{i},q_{\mathrm{root}}).\\q_{m}=\mathrm{Merge}(q_{\mathrm{op}},q_{\mathrm{subtree}},q_{m-1}).$$
(5)
## 2.2 $\textbf{Information Bottleneck}$
Our $ESIB$ method is driven by Information Bottleneck (IB) [1,17] that forces features to be concise by filtering the task-irrelevant information (i.e., syntax-irrelevant spurious correlations for MWP). Specifically, suppose we know the joint distribution $p(x,y)$,our goal is to learn a problem representation $z$ that maximizes its predictive power for generating $y$, subject to the constraints of the amount of information it carries about $x:$
$$\mathcal{L}_{IB}=I(y;z)-\lambda I(x;z),$$
(6)
where $I(\cdot;\cdot)$ denotes the mutual information.$\lambda$ is a Lagrangian multiplier that controls the trade-off between the sufficiency (the task performance quantified by $I(y;z))$ and the minimality (the complexity of the representation quantified by $I(x;z)).$ In this paper, we focus on compressing the redundancy in latent representation $v$ (denoted as a substitute for the problem $x).$ Following [1],Equation (6) can be variationally upper bounded by:
$$\mathcal{V}_{IB}=\frac{1}{N}\sum_{n=1}^{N}\left\{\lambda D_{\mathrm{KL}}\left(e\left(z\mid v^{n}\right)\parallel b(z)\right)-\mathbb{E}_{\epsilon}\log d\left(y^{n}\mid e\left(z^{n},\epsilon\right)\right)\right\},$$
where $N$ is the number of data, $e(z^{n},\epsilon)$ transforms the $z$ into initial
state of the decoder (denoted as $\bar{z}\in\mathbb{R}^m\cdot d)$ and $e\left(z\mid v^n\right)$ transforms the representation $v$ into two tensors: $e^\mu(v)$ denotes the featuresmean and $e^\sigma(v)$ denotes the diagonal features-covariance. We use the reparameterization to obtain the compressed representation $z=e(v,\epsilon)=e^{\mu}(v)+\epsilon e^{\sigma}(v)$ with $\epsilon\sim N(0,1).$ The prior distribution of $z$ (denoted as $b(z))$ is a standard normal distribution. The decoder $d$ converts the sampled representation $z$ into a mathematical expression $y$, and calculates the answer.
With the compression capability of VIB, it is possible to lose the necessary information about the target $y$ in the feature $z$ when optimizing the trade-off between the compression and redundancy. We will introduce how to identify the syntax-irrelevant information in Section 2.3.
# 2.3 Latent-specific Redundancy
In this section, we demonstrate that the latent-specific redundancy containing syntax-irrelevant information can be effectively reduced by using mutual learning [29]. Inspired by the method forcing two networks to learn from each other [29], we encourage two networks to produce compressed representation $z$ filtered latent-specific re- $\hat{\text{dundancy for keeping all predictive information about expression}}$ syntax trees in expression $y$ by passing an representation of expression syntax tree (denoted as $y_tree$ obtained by concatenating each $\hat{h}_i$ defined in equation(2) when decoder predicts the expression syntax tree of the math expression) to each other.
We attempt to learn a vector $z$ that contains expression syntax tree information about solution expression $y$ as much as possible and achieve this goal by optimizing the mutual information of $v$ and $z.$ We take CN as an example, and factorize the mutual information [6] between $v_1$ and $z_1$ as follows:
, (8)
$$I\left(v_{1};z_{1}\right)=\quad\underbrace{I\left(v_{1};z_{1}\mid v_{2}\right)}\quad+\quad\underbrace{I\left(v_{2};z_{1}\right)}\quad,$$
Latent-specific Redundancy Consistent Informatior
where $v_1$ and $z_1$ denote latent and problem representation for CN respectively; $v_2$ and $z_2$ denote that of SN. $I\left(v_1;z_1\mid v_2\right)$ indicates that the information contained in $z_1$ is unique to $v_1$ and cannot be inferred by representation $v_2$ [6]. We call $I\left(v_1;z_1\mid v_2\right)$ as latentspecific redundancy or syntax-irrelevant information. It can be discarded by minimizing $I\left(v_1;z_1\mid v_2\right)[6]$ which can be upper bounded by the following inequality:
$$\begin{aligned}&I\left(v_{1};z_{1}\mid v_{2}\right)=\mathbb{E}_{v_{1},v_{2}\sim S_{1}\left(v|x\right)}\mathbb{E}_{z_{1},z_{2}\sim S_{2}\left(z|v\right)}\left[\log\frac{p\left(z_{1}\mid v_{1}\right)}{p\left(z_{1}\mid v_{2}\right)}\right]\\&=\mathbb{E}_{v_{1},v_{2}\sim S_{1}\left(v|x\right)}\mathbb{E}_{z_{1},z_{2}\sim S_{2}\left(z|v\right)}\left[\log\frac{p\left(z_{1}\mid v_{1}\right)p\left(z_{2}\mid v_{2}\right)}{p\left(z_{2}\mid v_{2}\right)p\left(z_{1}\mid v_{2}\right)}\right]\\&=D_{KL}\left[p\left(z_{1}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right]-D_{KL}\left[p\left(z_{2}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right]\right]\\&\leq D_{KL}\left[p\left(z_{1}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right].\\&&\text{(9)}\\&\text{一}&\text{1}\\&\text{1}\\&\text{1}&\text{1}\\&\text{1}&\text{1}\\&&\text{1}\\&&\text{1}\\&&\text{1}\\&&&\leq D_{KL}\left[p\left(z_{1}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right]-D_{KL}\left[p\left(z_{2}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right]\\&&&\leq D_{KL}\left[p\left(z_{1}\mid v_{1}\right)\left\|p\left(z_{2}\mid v_{2}\right)\right]\right].\\&&&\text{(9)}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1}\\&&&\text{1$$
Inspired by [6], we approximate the upper bound above by re-
placing $z$ with $y_tree.$ Since $y_tree$ generated from $z$ contains all the information of the representation $z$ and all the latent-specific redundancy to be discarded. In addition, the parameters to be optimized about $z$ are included in the decoder. Considering the above points, we utilize $D_\mathrm{KL}\left(\mathbb{P}_{{z_1}}\|\mathbb{P}_{{z_2}}\right)\left(\mathbb{P}_{{z_1}}\text{ denotes }p\left(y_{tree_1}\mid z_1\right)\mathrm{~and}\right.$ $\mathbb{P}_{z_2}$ denotes $p\left(y_{tree_2}\mid z_2\right)$ right) as an upper bound to approximate $D_{KL}\left[p\left(z_{1}\mid v_{1}\right)\parallel p\left(z_{2}\mid v_{2}\right)\right].$Similarly, we can use $D_\mathrm{KL}\left(\mathbb{P}_{{z_{2}}}\|\mathbb{P}_{{z_{1}}}\right)$ to minimize $I\left(v_2;z_2\mid v_1\right)$for SN. We introduce the objective $\mathcal{L}_SKL$ to minimize the latent-specific redundancy for both $z_1$ and $z_2;$
(10)
$$\mathcal{L}_{SKL}=\min_{\theta,\phi}\mathbb{E}_{v_{1},v_{2}\sim E_{\theta}(v|x)}\mathbb{E}_{z_{1},z_{2}\sim E_{\phi}(z|v)}\left[D_{SKL}\left[\mathbb{P}_{z_{1}}\|\mathbb{P}_{z_{2}}\right]\right]$$
where $\theta$ and $\phi$ denote the parameters of CN and SN. The two networks are optimized alternately during training. $\mathbb{P}_{z_1}=p_\theta\left(y_1\mid z_1\right)$ and $\mathbb{P}_{z_{2}}=p_{\phi}\left(y_{2}\mid z_{2}\right)$denote the concatenation of the output distributions at each step of the model prediction of CN and SN
|
[
[
0.0849609375,
0.106689453125,
0.48193359375,
0.106689453125,
0.48193359375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.53173828125,
0.106689453125,
0.9140625,
0.106689453125,
0.9140625,
0.12060546875,
0.53173828125,
0.12060546875
],
[
0.08416748046875,
0.121826171875,
0.483642578125,
0.121826171875,
0.483642578125,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.515625,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.515625,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1351318359375,
0.9140625,
0.1351318359375,
0.9140625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.08660888671875,
0.15087890625,
0.48046875,
0.15087890625,
0.48046875,
0.1610107421875,
0.08660888671875,
0.1610107421875
],
[
0.515625,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.515625,
0.1634521484375
],
[
0.0849609375,
0.1622314453125,
0.48193359375,
0.1622314453125,
0.48193359375,
0.175537109375,
0.0849609375,
0.175537109375
],
[
0.51708984375,
0.1641845703125,
0.60302734375,
0.1641845703125,
0.60302734375,
0.175537109375,
0.51708984375,
0.175537109375
],
[
0.08416748046875,
0.1767578125,
0.465576171875,
0.1767578125,
0.465576171875,
0.1900634765625,
0.08416748046875,
0.1900634765625
],
[
0.5185546875,
0.1900634765625,
0.55615234375,
0.1900634765625,
0.55615234375,
0.2020263671875,
0.5185546875,
0.2020263671875
],
[
0.5498046875,
0.189453125,
0.78369140625,
0.189453125,
0.78369140625,
0.20263671875,
0.5498046875,
0.20263671875
],
[
0.205078125,
0.2083740234375,
0.404296875,
0.2083740234375,
0.404296875,
0.2259521484375,
0.205078125,
0.2259521484375
],
[
0.2034912109375,
0.1932373046875,
0.393798828125,
0.1925048828125,
0.393798828125,
0.2108154296875,
0.2034912109375,
0.212158203125
],
[
0.463134765625,
0.208984375,
0.484375,
0.208984375,
0.484375,
0.222900390625,
0.463134765625,
0.222900390625
],
[
0.51708984375,
0.2083740234375,
0.9111328125,
0.2083740234375,
0.9111328125,
0.2215576171875,
0.51708984375,
0.2215576171875
],
[
0.51708984375,
0.2215576171875,
0.9140625,
0.2215576171875,
0.9140625,
0.23486328125,
0.51708984375,
0.23486328125
],
[
0.15283203125,
0.210205078125,
0.181396484375,
0.210205078125,
0.181396484375,
0.2215576171875,
0.15283203125,
0.2215576171875
],
[
0.2034912109375,
0.222900390625,
0.398681640625,
0.222900390625,
0.398681640625,
0.240478515625,
0.2034912109375,
0.240478515625
],
[
0.51708984375,
0.23486328125,
0.91259765625,
0.23486328125,
0.91259765625,
0.248779296875,
0.51708984375,
0.248779296875
],
[
0.2156982421875,
0.24365234375,
0.3505859375,
0.24365234375,
0.3505859375,
0.25830078125,
0.2156982421875,
0.25830078125
],
[
0.463134765625,
0.2442626953125,
0.483642578125,
0.2442626953125,
0.483642578125,
0.2568359375,
0.463134765625,
0.2568359375
],
[
0.51708984375,
0.250732421875,
0.91259765625,
0.250732421875,
0.91259765625,
0.261474609375,
0.51708984375,
0.261474609375
],
[
0.0849609375,
0.260009765625,
0.483642578125,
0.260009765625,
0.483642578125,
0.2734375,
0.0849609375,
0.2734375
],
[
0.5146484375,
0.261962890625,
0.91259765625,
0.261962890625,
0.91259765625,
0.27587890625,
0.5146484375,
0.27587890625
],
[
0.08251953125,
0.274658203125,
0.400390625,
0.2734375,
0.400390625,
0.28857421875,
0.08251953125,
0.2890625
],
[
0.51708984375,
0.276611328125,
0.9140625,
0.276611328125,
0.9140625,
0.290283203125,
0.51708984375,
0.290283203125
],
[
0.101318359375,
0.2890625,
0.483642578125,
0.2890625,
0.483642578125,
0.302978515625,
0.101318359375,
0.302978515625
],
[
0.51708984375,
0.291748046875,
0.9140625,
0.291748046875,
0.9140625,
0.304931640625,
0.51708984375,
0.304931640625
],
[
0.08660888671875,
0.304931640625,
0.48046875,
0.304931640625,
0.48046875,
0.31494140625,
0.08660888671875,
0.31494140625
],
[
0.51708984375,
0.304931640625,
0.9140625,
0.304931640625,
0.9140625,
0.318115234375,
0.51708984375,
0.318115234375
],
[
0.0849609375,
0.31640625,
0.48193359375,
0.31640625,
0.48193359375,
0.330078125,
0.0849609375,
0.330078125
],
[
0.51708984375,
0.3193359375,
0.9140625,
0.3193359375,
0.9140625,
0.333251953125,
0.51708984375,
0.333251953125
],
[
0.0849609375,
0.330810546875,
0.483642578125,
0.330810546875,
0.483642578125,
0.3447265625,
0.0849609375,
0.3447265625
],
[
0.515625,
0.333251953125,
0.81787109375,
0.333251953125,
0.81787109375,
0.3466796875,
0.515625,
0.3466796875
],
[
0.0849609375,
0.3447265625,
0.48193359375,
0.3447265625,
0.48193359375,
0.35791015625,
0.0849609375,
0.35791015625
],
[
0.53125,
0.34521484375,
0.9140625,
0.3466796875,
0.9140625,
0.35986328125,
0.53125,
0.359130859375
],
[
0.08416748046875,
0.356689453125,
0.48046875,
0.356689453125,
0.48046875,
0.37060546875,
0.08416748046875,
0.37060546875
],
[
0.51708984375,
0.3623046875,
0.9111328125,
0.3623046875,
0.9111328125,
0.37255859375,
0.51708984375,
0.37255859375
],
[
0.0849609375,
0.37255859375,
0.48193359375,
0.37255859375,
0.48193359375,
0.386962890625,
0.0849609375,
0.386962890625
],
[
0.515625,
0.375,
0.9140625,
0.375,
0.9140625,
0.38818359375,
0.515625,
0.38818359375
],
[
0.08416748046875,
0.386474609375,
0.483642578125,
0.385009765625,
0.483642578125,
0.399658203125,
0.08416748046875,
0.40087890625
],
[
0.51708984375,
0.38818359375,
0.9140625,
0.38818359375,
0.9140625,
0.401611328125,
0.51708984375,
0.401611328125
],
[
0.08251953125,
0.40087890625,
0.42724609375,
0.40087890625,
0.42724609375,
0.4140625,
0.08251953125,
0.4140625
],
[
0.5185546875,
0.40283203125,
0.7197265625,
0.40283203125,
0.7197265625,
0.416748046875,
0.5185546875,
0.416748046875
],
[
0.2156982421875,
0.417236328125,
0.3505859375,
0.417236328125,
0.3505859375,
0.43505859375,
0.2156982421875,
0.43505859375
],
[
0.463134765625,
0.419189453125,
0.484375,
0.419189453125,
0.484375,
0.434326171875,
0.463134765625,
0.434326171875
],
[
0.6435546875,
0.421630859375,
0.72607421875,
0.421630859375,
0.72607421875,
0.436279296875,
0.6435546875,
0.436279296875
],
[
0.79345703125,
0.421630859375,
0.85107421875,
0.421630859375,
0.85107421875,
0.436279296875,
0.79345703125,
0.436279296875
],
[
0.88818359375,
0.42236328125,
0.9140625,
0.42236328125,
0.9140625,
0.43505859375,
0.88818359375,
0.43505859375
],
[
0.75732421875,
0.425537109375,
0.76318359375,
0.425537109375,
0.76318359375,
0.43115234375,
0.75732421875,
0.43115234375
],
[
0.54736328125,
0.423583984375,
0.61279296875,
0.423583984375,
0.61279296875,
0.43505859375,
0.54736328125,
0.43505859375
],
[
0.2099609375,
0.4375,
0.3544921875,
0.4375,
0.3544921875,
0.455078125,
0.2099609375,
0.455078125
],
[
0.461669921875,
0.438232421875,
0.484375,
0.438232421875,
0.484375,
0.45263671875,
0.461669921875,
0.45263671875
],
[
0.615234375,
0.447509765625,
0.8798828125,
0.447509765625,
0.8798828125,
0.458251953125,
0.615234375,
0.458251953125
],
[
0.181396484375,
0.455810546875,
0.38232421875,
0.455810546875,
0.38232421875,
0.473388671875,
0.181396484375,
0.473388671875
],
[
0.461669921875,
0.45703125,
0.484375,
0.45703125,
0.484375,
0.47021484375,
0.461669921875,
0.47021484375
],
[
0.515625,
0.46337890625,
0.91259765625,
0.46337890625,
0.91259765625,
0.4765625,
0.515625,
0.4765625
],
[
0.08416748046875,
0.481689453125,
0.3203125,
0.481689453125,
0.3203125,
0.49560546875,
0.08416748046875,
0.49560546875
],
[
0.51708984375,
0.477783203125,
0.91259765625,
0.477783203125,
0.91259765625,
0.4912109375,
0.51708984375,
0.4912109375
],
[
0.515625,
0.490478515625,
0.9140625,
0.4912109375,
0.9140625,
0.5048828125,
0.515625,
0.50390625
],
[
0.08660888671875,
0.501953125,
0.4794921875,
0.501953125,
0.4794921875,
0.51220703125,
0.08660888671875,
0.51220703125
],
[
0.5146484375,
0.5048828125,
0.91259765625,
0.5048828125,
0.91259765625,
0.51806640625,
0.5146484375,
0.51806640625
],
[
0.08416748046875,
0.51318359375,
0.48193359375,
0.5146484375,
0.48193359375,
0.52783203125,
0.08416748046875,
0.5263671875
],
[
0.51708984375,
0.51806640625,
0.9140625,
0.51806640625,
0.9140625,
0.5322265625,
0.51708984375,
0.5322265625
],
[
0.0849609375,
0.52880859375,
0.48046875,
0.52880859375,
0.48046875,
0.54248046875,
0.0849609375,
0.54248046875
],
[
0.515625,
0.5322265625,
0.9140625,
0.5322265625,
0.9140625,
0.546875,
0.515625,
0.546875
],
[
0.0849609375,
0.54248046875,
0.48193359375,
0.54248046875,
0.48193359375,
0.55615234375,
0.0849609375,
0.55615234375
],
[
0.5146484375,
0.54541015625,
0.68701171875,
0.54541015625,
0.68701171875,
0.5595703125,
0.5146484375,
0.5595703125
],
[
0.08660888671875,
0.55810546875,
0.48046875,
0.55810546875,
0.48046875,
0.568359375,
0.08660888671875,
0.568359375
],
[
0.08416748046875,
0.5693359375,
0.483642578125,
0.568359375,
0.483642578125,
0.58203125,
0.08416748046875,
0.58349609375
],
[
0.52392578125,
0.5693359375,
0.7861328125,
0.57373046875,
0.7861328125,
0.59130859375,
0.52392578125,
0.5869140625
],
[
0.80126953125,
0.56494140625,
0.87744140625,
0.56298828125,
0.87744140625,
0.57763671875,
0.80126953125,
0.580078125
],
[
0.78369140625,
0.57763671875,
0.7998046875,
0.57373046875,
0.8037109375,
0.58056640625,
0.7861328125,
0.58544921875
],
[
0.0849609375,
0.58544921875,
0.26953125,
0.58544921875,
0.26953125,
0.59521484375,
0.0849609375,
0.59521484375
],
[
0.8095703125,
0.58203125,
0.8759765625,
0.58203125,
0.8759765625,
0.59326171875,
0.8095703125,
0.59326171875
],
[
0.2115478515625,
0.60302734375,
0.35546875,
0.60302734375,
0.35546875,
0.6162109375,
0.2115478515625,
0.6162109375
],
[
0.463134765625,
0.60302734375,
0.483642578125,
0.60302734375,
0.483642578125,
0.61572265625,
0.463134765625,
0.61572265625
],
[
0.52685546875,
0.60302734375,
0.73876953125,
0.60302734375,
0.73876953125,
0.62353515625,
0.52685546875,
0.62353515625
],
[
0.7294921875,
0.59765625,
0.86376953125,
0.59765625,
0.86376953125,
0.611328125,
0.7294921875,
0.611328125
],
[
0.72802734375,
0.61279296875,
0.86376953125,
0.61279296875,
0.86376953125,
0.626953125,
0.72802734375,
0.626953125
],
[
0.0849609375,
0.6220703125,
0.48193359375,
0.6220703125,
0.48193359375,
0.63525390625,
0.0849609375,
0.63525390625
],
[
0.08416748046875,
0.6337890625,
0.48193359375,
0.63525390625,
0.48193359375,
0.6484375,
0.08416748046875,
0.64794921875
],
[
0.53125,
0.62890625,
0.90283203125,
0.62890625,
0.90283203125,
0.642578125,
0.53125,
0.642578125
],
[
0.0849609375,
0.6494140625,
0.48193359375,
0.6494140625,
0.48193359375,
0.66357421875,
0.0849609375,
0.66357421875
],
[
0.529296875,
0.64599609375,
0.72216796875,
0.64599609375,
0.72216796875,
0.6591796875,
0.529296875,
0.6591796875
],
[
0.08416748046875,
0.66357421875,
0.483642578125,
0.66357421875,
0.483642578125,
0.6767578125,
0.08416748046875,
0.6767578125
],
[
0.89453125,
0.66015625,
0.9150390625,
0.66015625,
0.9150390625,
0.67236328125,
0.89453125,
0.67236328125
],
[
0.0849609375,
0.6767578125,
0.48193359375,
0.6767578125,
0.48193359375,
0.68994140625,
0.0849609375,
0.68994140625
],
[
0.51708984375,
0.673828125,
0.9140625,
0.673828125,
0.9140625,
0.6875,
0.51708984375,
0.6875
],
[
0.08660888671875,
0.68994140625,
0.48193359375,
0.68994140625,
0.48193359375,
0.7041015625,
0.08660888671875,
0.7041015625
],
[
0.51708984375,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.51708984375,
0.7021484375
],
[
0.08660888671875,
0.70556640625,
0.34228515625,
0.70556640625,
0.34228515625,
0.71630859375,
0.08660888671875,
0.71630859375
],
[
0.51708984375,
0.7021484375,
0.91259765625,
0.7021484375,
0.91259765625,
0.71533203125,
0.51708984375,
0.71533203125
],
[
0.51708984375,
0.71533203125,
0.9140625,
0.71533203125,
0.9140625,
0.72900390625,
0.51708984375,
0.72900390625
],
[
0.09234619140625,
0.73193359375,
0.1356201171875,
0.7392578125,
0.131591796875,
0.751953125,
0.0882568359375,
0.744140625
],
[
0.1690673828125,
0.734375,
0.473876953125,
0.734375,
0.473876953125,
0.74755859375,
0.1690673828125,
0.74755859375
],
[
0.51708984375,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.7412109375,
0.51708984375,
0.7412109375
],
[
0.1282958984375,
0.74072265625,
0.1650390625,
0.736328125,
0.1690673828125,
0.751953125,
0.131591796875,
0.755859375
],
[
0.515625,
0.7421875,
0.9140625,
0.74365234375,
0.9140625,
0.75830078125,
0.515625,
0.7568359375
],
[
0.1502685546875,
0.75048828125,
0.1707763671875,
0.75048828125,
0.1707763671875,
0.7587890625,
0.1502685546875,
0.7587890625
],
[
0.463134765625,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.77099609375,
0.463134765625,
0.77099609375
],
[
0.515625,
0.7568359375,
0.91259765625,
0.755859375,
0.91259765625,
0.77099609375,
0.515625,
0.771484375
],
[
0.08660888671875,
0.77294921875,
0.48046875,
0.77294921875,
0.48046875,
0.78271484375,
0.08660888671875,
0.78271484375
],
[
0.5185546875,
0.771484375,
0.91259765625,
0.771484375,
0.91259765625,
0.78515625,
0.5185546875,
0.78515625
],
[
0.0849609375,
0.7841796875,
0.48193359375,
0.7841796875,
0.48193359375,
0.798828125,
0.0849609375,
0.798828125
],
[
0.515625,
0.7841796875,
0.91259765625,
0.78515625,
0.91259765625,
0.7998046875,
0.515625,
0.798828125
],
[
0.08660888671875,
0.80029296875,
0.48046875,
0.80029296875,
0.48046875,
0.81103515625,
0.08660888671875,
0.81103515625
],
[
0.515625,
0.798828125,
0.890625,
0.798828125,
0.890625,
0.8125,
0.515625,
0.8125
],
[
0.0849609375,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.8271484375,
0.0849609375,
0.8271484375
],
[
0.54150390625,
0.81640625,
0.87158203125,
0.81640625,
0.87158203125,
0.833984375,
0.54150390625,
0.833984375
],
[
0.8759765625,
0.8173828125,
0.91259765625,
0.8173828125,
0.91259765625,
0.8291015625,
0.8759765625,
0.8291015625
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.08416748046875,
0.84130859375,
0.483642578125,
0.84033203125,
0.483642578125,
0.85400390625,
0.08416748046875,
0.85498046875
],
[
0.5185546875,
0.84228515625,
0.9111328125,
0.84228515625,
0.9111328125,
0.85302734375,
0.5185546875,
0.85302734375
],
[
0.08416748046875,
0.85498046875,
0.483642578125,
0.85498046875,
0.483642578125,
0.86865234375,
0.08416748046875,
0.86865234375
],
[
0.5146484375,
0.85400390625,
0.91259765625,
0.85498046875,
0.91259765625,
0.869140625,
0.5146484375,
0.86865234375
],
[
0.08660888671875,
0.87060546875,
0.48193359375,
0.87060546875,
0.48193359375,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.51708984375,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.08416748046875,
0.88330078125,
0.328369140625,
0.88330078125,
0.328369140625,
0.896484375,
0.08416748046875,
0.896484375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"node qi.Finally, we generate the right child node qr.This process",
"With the compression capability of VIB, it is possible to lose",
"has been iterated until the leaf nodes are generated. The attention",
"the necessary information about the target y in the feature z when",
"mechanism is applied to learn the global context vector G, which",
"optimizing the trade-off between the compression and redundancy.",
"is utilized to generate the current node token i.Here we denote",
"We will introduce how to identify the syntax-irrelevant information",
"the digital embedding after being encoded by the encoder as Q",
"in Section2.3.",
"Mathematically, we define the attention mechanism as follows:",
"2.3",
"Latent-specificRedundancy",
"Attention (x, Proot st), Qsi 0.",
"Attention (x, qroot qi), q0.",
"(1)",
"In this section, we demonstrate that the latent-specific redundancy",
"containing syntax-irrelevant information can be effectively reduced",
"Gi=",
"Attention (x, qroot), qi, Qsi e 0.",
"by using mutual learning [29]. Inspired by the method forcing two",
"gi, hi = Predict(Gi, Q),",
"(2)",
"networkstolearnfromeachother[29],weencouragetwonetworks",
"where Predict( is prediction layer for producing tree nodes gi and",
"to produce compressed representation z filtered latent-specific re-",
"hidden state of decoding step i (denoted as hi e IRd).",
"dundancy for keeping all predictive information about expression",
"If the current node is an operator, we will generate the left and",
"syntax trees in expression y by passing an representation of expres-",
"rightchildnodesand pushthemintothestack inthetreedecoder",
"sion syntax tree (denoted as ytree obtained by concatenating each",
"according to the top-down method. If it is a number, we will perform",
"h; defined in equation(2) when decoder predicts the expression",
"the merge operation until the leaf nodes in the stack pop out, and",
"syntax tree of the math expression) to each other.",
"the result of the merge is pushed into the left child node stack for",
"We attempt to learn a vector z that contains expression syntax",
" attention operation. The merge operation will pop the required",
"treeinformation about solutionexpressionyas much as possible",
"node qop and qsubtree from an embedding stack. Finally, the merge",
"and achieve this goal by optimizing the mutual information of u and",
"operation outputs the answer of the mathematical expression. This",
"z. We take CN as an example, and factorize the mutual information",
"recursive construction process can be defined as follows:",
"[6] between u1 and zi as follows:",
"q1 = Left(G, gi, root).",
"(3)",
"I(u1;z1 /02)",
"I(02;z1)",
"(8)",
"+",
"I (U1; Z1) =",
"qr = Right(G, yi root).",
"(4)",
"Latent-specific RedundancyConsistent Information",
"qm = Merge(qop subtree, qm-1).",
"(5)",
"where ui and z1 denote latent and problem representation for CN",
"2.2 InformationBottleneck",
"respectively; 02 and z2 denote that of SN. I (u1;z1 | u2) indicates",
"that the information contained in zi is unique to ui and cannot be",
"OurESIBmethodisdrivenbyInformationBottleneck(IB)[1.,17]",
"inferred by representation 02 [6]. We call I (u1;z1 / 2) as latent",
"that forces features to be concise by filtering the task-irrelevant",
"specific redundancy or syntax-irrelevant information. It can be dis-",
"information (i.e., syntax-irrelevant spurious correlations for MWP)",
"carded by minimizing I (u1; z1 I 2)[6] which can be upper bounded",
"Specifically,suppose weknow the joint distribution p(x,y),our goal",
"by the following inequality:",
"is tolearnaproblemrepresentationzthatmaximizesits predictive",
"power for generating y, subject to the constraints of the amount of",
"I(01;z1 / 02) = Bo1,02~S(o|x)Bz1,22~S2(z/0)",
"p(z1 / u1)",
"Lo",
"information it carries about x:",
"p(z1 / 02)",
"LIB = I(y;z) - ^I(x;z),",
"(6)",
"=Bu,2~(ox)Bz1~S2(20) og",
"p(z1 [01) p(z2 /02)",
"p(z2 [02)p(z1 /02)",
"where I(;) denotes the mutual information. ^ is a Lagrangian",
"multiplier that controls the trade-off between the sufficiency (the",
"= DkL [p(z1 / u1) Ip (z2 [ 02)] - DkL [p (z2 / 01) lp (z2 / 02)]",
"task performance quantified by I(y; z)) and the minimality (the",
"≤ DkL [p (z1 / u1) I/p (z2 / u2)] ",
"complexity of the representation quantified byI(x; z).In this paper,",
"(9)",
"we focus on compressing the redundancy in latent representation ",
"Inspired by [6], we approximate the upper bound above by re-",
"(denoted as a substitute for the problem x). Following [1], Equation",
"placing z with ytree. Since ytree generated from z contains all the",
"(6)can bevariationally upper bounded by:",
"information of the representation z and all the latent-specific re-",
"dundancy to be discarded. In addition, the parameters to be opti-",
"VIB=",
"[αDkL (e (z [ u\") llb(z)) - Be logd (yn [e (zn,e))",
"mized about z are included in the decoder.Considering the above",
"N2",
"points, we utilize DkL (Pz,//Pzz) (Pz denotes p (ytreer I zi) and",
"n=1",
"(7)",
"Pz denotes p (ytree2 I z2) right) as an upperbound to approximate",
"whereN is the numberofdata,e(zn,e)transforms thezinto initial",
"DkL [p (z1 [ u1) lp (z2 / 02)]. Similarly, we can use DkL (Pz2 /Pz,)",
"state of the decoder (denoted as z e Rm-d) and e (z |un) transforms",
"to minimize I (02; z2 | vi) for SN. We introduce the objective LsKL",
"therepresentation intotwotensors:e(o)denotesthefeatures",
"to minimize the latent-specific redundancy for both z1 and z2:",
"mean and e (o) denotes the diagonal features-covariance. We use",
"LsKL = min Bur,02~Eg(o|x) Bzi2~-Eg(2l0) [DsKL [Pzi IPz2]]",
"(10)",
"the reparameterization to obtain the compressed representation",
"z =e(u,e)=e\"(u)+ee(o) withe~N(0,1).Thepriordistribution",
"where and @ denote the parameters of CNand SN.The twonet",
"of z (denoted as b(z) is a standard normal distribution. The de-",
"works are optimized alternately during training. Pz =po (yi |zi)",
"coderd convertsthesampled representationzintoamathematical",
"and Pz2 = P (y2 | z2) denote the concatenation of the output",
"expression y, and calculates the answer.",
" distributions at each step of the model prediction of CN and SN"
] |
{
"arxiv_id": "2310.15664",
"language": "en",
"page_idx": 2
}
|
||
2310.15664-en-3
|
2310.15664-en-3.jpg
|
respectively. $D_{SKL}$ denotes symmetrized KL divergence obtained by averaging the expected value of $D_\mathrm{KL}\left(\mathbb{P}_{z_1}\|\mathbb{P}_{z_2}\right)$ and $D_\mathrm{KL}\left(\mathbb{P}_{z_2}\|\mathbb{P}_{z_1}\right).$ We calculate this loss by mutual learning [29] between CN and SN. In the mutual learning setup, in an iteration, the model will compute $\mathcal{L} _{SKL_{1}}$ and $\mathcal{L} _{SKL_{2}}$ for CN and SN respectively. In addition, we need to maximize $I\left ( v_{2}; z_{1}\right )$ to ensure that the compressed representation $z_1$ has enough information to predict $y.$ We use the chain rule to decompose $I\left(v_2;z_1\right)$ into the following two terms:
(11)
$$I\left(v_{2};z_{1}\right)=\underbrace{I\left(v_{2};z_{1}\mid y\right)}_{\text{Redundancy}}+\underbrace{I\left(z_{1};y\right)}_{\text{Predictive Information}},$$
where $y$ represents ground-truth solution expression. In practice, we can maximize $I\left(z_1;y\right)=\mathbb{E}_\epsilon\log d\left(y^n\mid e\left(x^n,\epsilon\right)\text{(included in }\mathcal{V}_{IB}\right)$ which is calculated to compress redundant information $I\left(v_2;z_1\mid y\right)$ and indirectly maximize $I\left(v_2;z_1\right).$ Ideally, $I\left(v_2;z_1\mid y\right)$ should be zero. As suggested in [6], we minimize latent-specific redundancy by jointly minimizing $I\left(v_1;z_1\mid v_2\right)$ and maximizing $I\left(v_2;z_1\right).$ We define the redundancy terms in Eq.(8) and Eq.(11) as the syntaxirrelevant information.
# 2.4 Self-distillation Loss
In this section, we introduce a novel self-distillation loss to increase the diversity of generated expressions. Suggested by [7], the $I(v;z\mid$ $y)$ in conditional information bottleneck can be variationally upper bounded by:
(12)
$$D_{\mathrm{KL}}\left(e\left(z\mid v^{n}\right)\|b\left(z\mid y^{n}\right)\right),$$
where $e\left(z\mid v^{n}\right)$defined in equation (2)is moved towards the conditional marginal $b^\mu(y)\sim N\left(b^\mu(y),b^\sigma(y)\right).$ We modify equation (12)as:
(13)
$$\mathcal{V}_{SDL}=\frac{1}{N}\sum_{n=1}^{N}D_{SKL}\left(\bar{y}\|\bar{z}\right),$$
where $\bar{y}$ denotes that averaging the all $\hat{h}_i.$ Intuitively, $\mathcal V_SDL$ make the decoder more rely on latent space of $z$ which contains expression syntax tree information for all expressions when predicting expression $y.$ Benefiting from the randomness of $z$,the model can generate more diverse solution expressions. Finally, we calculate the loss functions $\mathcal{L}_1$ for SN and $\mathcal{L}_2$ for CN as follows:
$$\mathcal{L}_{1}=\mathcal{V}_{IB_{1}}+\mathcal{V}_{SDL_{1}}+\alpha\times\mathcal{L}_{SKL_{1}}.\\\mathcal{L}_{2}=\mathcal{V}_{IB_{2}}+\mathcal{V}_{SDL_{2}}+\alpha\times\mathcal{L}_{SKL_{2}}.$$
(14) (15)
where $\alpha$ is a proportional coefficient. $\mathcal{V}_IB_1,\mathcal{V}_{SDL_1},\mathcal{L}_{SKL_1}$ are the training objectives for SN. $\mathcal{V}_{IB_{2}},\mathcal{V}_{SDL_{2}},\mathcal{L}_{SKL_{2}}$ are the training objectives for CN. Based on empirical observation, although $\mathcal{V}_SDL$ can increase the diversity of solution expressions, it also reduces accuracy of the final answer.
# 3 EXPERIMENTAL SETUP
Datasets. We conduct experiments on two benchmark MWP
datasets: Math23k [22] and Ape210k [30]. Math23k contains 22162/1000
questions for training/testing, respectively. Ape210k is composed of 166,270 questions for training, 4,157 questions for validation, and 4,159 questions for testing
Implementation Details. The word embedding size of decoder is set to 1024 and proportional coefficient $\alpha$ in loss function is set to 0.005. We set the dimension of vectors $z$ to 50. When the encoder is BERT, we set the dimension of $z$ to 32. We adopt RoBERTa [11] as the problem encoder. Following RoBERTa's setting, the hidden size of the encoder is set to 768, and we set the hidden size of the decoder to 1024. We used Adamw [12] as the optimizer with the learning rate as 5e-5. The mini-batch size is set to 16. We adopt a beam search with the size of 5. Dropout (dropout rate = 0.5) is employed to avoid overfitting. For Ape210K, we set the maximum sequence length of questions as 150 and that of solution expressions as 50, similar to [23]. Our model takes 80 epochs on Math23k and 50 epochs on Ape210k for convergence.
Baselines. We compare our model with several strong baseline methods, including NumS2T [23], TSN-MD [27], MATH-EN [18], Multi-E/D[16],GTS[24],StackDecoder[4],KAS2T[23],Graph2tree [28],and Ape[30].
4 EXPERIMENTAL RESULTS
## 4.1 Main Results
The evaluation metric is answer accuracy. Table 2 show the performance comparison of our model with baseline methods on Math23K and Ape210k, respectively. In Table 2, since there is a trade-off between the variety of expressions and the correctness of the answer, we do not add $V_SDL$ into the source network (SN) and the collaborator network (CN). From Table 2 we can observe that our models (both CN and SN) achieve consistently and substantially better performance than the compared methods. In addition, the accuracy of CN is higher than that of SN. This is because, in one iteration, CN is provided with $y_tree$ predicted by SN when SN has not been trained by ground-truth then SN is provided with $y_{tree}$ predicted by CN when CN has been trained by ground-truth.
We also measure the accuracy of solution expression. We consider a solution expression as correct when the predicted expression exactly matches the ground truth solution. Generally, the mathematical expression generated by the tree decoder conforms to the syntactic specification. As long as the answer obtained by the expression operation is consistent with the ground-truth, then we $\dot{\text{consider the expression to be a valid solution. We take the subtrac-}}$ tion value between answer accuracy (denoted as Answer-Acc) and solution expression accuracy (denoted as Expression-Acc) as the $\dot{\text{diversity evaluation metric(denoted as Diversity) of the generated}}$ solution expressions. As shown in Table 3, our model can generate more diverse solution expressions that are not included in groundtruth expressions. As expected, the model with $\mathcal V_SDL$ has better diversity but lower answer accuracy.
## 4.2 Ablation Study
We conduct ablation test on Math23k to analyze the impact of different components in ESIB. Since the best results are produced by CN, we only conduct ablation study on CN. First, we remove the mutual learning, denoted as CN w/o MT. Second, we remove the VIB from SN and CN to evaluate the impact of VIB (denoted as CN w/o VIB). In addition, we report the results by removing both MT and VIB (denoted as CN w/o MT+VIB). To evaluate the impact of
|
[
[
0.0849609375,
0.106689453125,
0.47802734375,
0.106689453125,
0.47802734375,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.53173828125,
0.106689453125,
0.91259765625,
0.106689453125,
0.91259765625,
0.12060546875,
0.53173828125,
0.12060546875
],
[
0.08416748046875,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.515625,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.515625,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.483642578125,
0.1351318359375,
0.483642578125,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1351318359375,
0.9140625,
0.1351318359375,
0.9140625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.5185546875,
0.149658203125,
0.9111328125,
0.149658203125,
0.9111328125,
0.160400390625,
0.5185546875,
0.160400390625
],
[
0.08660888671875,
0.1634521484375,
0.483642578125,
0.1622314453125,
0.483642578125,
0.1767578125,
0.08660888671875,
0.177978515625
],
[
0.51708984375,
0.1622314453125,
0.9140625,
0.1622314453125,
0.9140625,
0.175537109375,
0.51708984375,
0.175537109375
],
[
0.0849609375,
0.1767578125,
0.48193359375,
0.1767578125,
0.48193359375,
0.1900634765625,
0.0849609375,
0.1900634765625
],
[
0.51708984375,
0.1767578125,
0.91259765625,
0.1767578125,
0.91259765625,
0.1900634765625,
0.51708984375,
0.1900634765625
],
[
0.08660888671875,
0.1912841796875,
0.48046875,
0.1912841796875,
0.48046875,
0.2020263671875,
0.08660888671875,
0.2020263671875
],
[
0.5185546875,
0.1912841796875,
0.91259765625,
0.1912841796875,
0.91259765625,
0.2020263671875,
0.5185546875,
0.2020263671875
],
[
0.08416748046875,
0.2052001953125,
0.3896484375,
0.2052001953125,
0.3896484375,
0.2183837890625,
0.08416748046875,
0.2183837890625
],
[
0.51708984375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51708984375,
0.2183837890625
],
[
0.51708984375,
0.2183837890625,
0.91259765625,
0.2183837890625,
0.91259765625,
0.231689453125,
0.51708984375,
0.231689453125
],
[
0.148681640625,
0.2279052734375,
0.30224609375,
0.2279052734375,
0.30224609375,
0.242431640625,
0.148681640625,
0.242431640625
],
[
0.330810546875,
0.2279052734375,
0.38232421875,
0.2279052734375,
0.38232421875,
0.242431640625,
0.330810546875,
0.242431640625
],
[
0.45751953125,
0.2279052734375,
0.483642578125,
0.2279052734375,
0.483642578125,
0.2412109375,
0.45751953125,
0.2412109375
],
[
0.515625,
0.23291015625,
0.9140625,
0.231689453125,
0.9140625,
0.24560546875,
0.515625,
0.246826171875
],
[
0.29736328125,
0.251953125,
0.4150390625,
0.251953125,
0.4150390625,
0.26513671875,
0.29736328125,
0.26513671875
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.51708984375,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.221435546875,
0.25244140625,
0.28515625,
0.25244140625,
0.28515625,
0.264404296875,
0.221435546875,
0.264404296875
],
[
0.0849609375,
0.2734375,
0.48193359375,
0.2734375,
0.48193359375,
0.287353515625,
0.0849609375,
0.287353515625
],
[
0.51708984375,
0.2734375,
0.75732421875,
0.2734375,
0.75732421875,
0.287353515625,
0.51708984375,
0.287353515625
],
[
0.0849609375,
0.28857421875,
0.48193359375,
0.28857421875,
0.48193359375,
0.3017578125,
0.0849609375,
0.3017578125
],
[
0.0849609375,
0.3017578125,
0.48046875,
0.3017578125,
0.48046875,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.53369140625,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.311279296875,
0.53369140625,
0.311279296875
],
[
0.08416748046875,
0.31640625,
0.48193359375,
0.31640625,
0.48193359375,
0.330078125,
0.08416748046875,
0.330078125
],
[
0.51708984375,
0.311279296875,
0.9140625,
0.311279296875,
0.9140625,
0.324462890625,
0.51708984375,
0.324462890625
],
[
0.0849609375,
0.328857421875,
0.48046875,
0.328857421875,
0.48046875,
0.34228515625,
0.0849609375,
0.34228515625
],
[
0.51708984375,
0.32568359375,
0.9150390625,
0.32568359375,
0.9150390625,
0.339111328125,
0.51708984375,
0.339111328125
],
[
0.08416748046875,
0.343505859375,
0.48193359375,
0.34228515625,
0.48193359375,
0.35595703125,
0.08416748046875,
0.356689453125
],
[
0.51953125,
0.341552734375,
0.63134765625,
0.341552734375,
0.63134765625,
0.3515625,
0.51953125,
0.3515625
],
[
0.0849609375,
0.35791015625,
0.483642578125,
0.35791015625,
0.483642578125,
0.371826171875,
0.0849609375,
0.371826171875
],
[
0.08416748046875,
0.371826171875,
0.2254638671875,
0.371826171875,
0.2254638671875,
0.385009765625,
0.08416748046875,
0.385009765625
],
[
0.51953125,
0.371826171875,
0.53369140625,
0.371826171875,
0.53369140625,
0.380126953125,
0.51953125,
0.380126953125
],
[
0.53759765625,
0.37255859375,
0.54345703125,
0.37255859375,
0.54345703125,
0.37744140625,
0.53759765625,
0.37744140625
],
[
0.5390625,
0.367431640625,
0.76708984375,
0.367431640625,
0.76708984375,
0.380615234375,
0.5390625,
0.380615234375
],
[
0.51708984375,
0.38818359375,
0.66650390625,
0.38818359375,
0.66650390625,
0.401611328125,
0.51708984375,
0.401611328125
],
[
0.0849609375,
0.399658203125,
0.294189453125,
0.399658203125,
0.294189453125,
0.41357421875,
0.0849609375,
0.41357421875
],
[
0.51708984375,
0.406005859375,
0.9140625,
0.406005859375,
0.9140625,
0.419189453125,
0.51708984375,
0.419189453125
],
[
0.08416748046875,
0.416748046875,
0.48193359375,
0.417236328125,
0.48193359375,
0.431884765625,
0.08416748046875,
0.43115234375
],
[
0.515625,
0.42041015625,
0.9140625,
0.419189453125,
0.9140625,
0.43310546875,
0.515625,
0.434326171875
],
[
0.0849609375,
0.431884765625,
0.48193359375,
0.431884765625,
0.48193359375,
0.44580078125,
0.0849609375,
0.44580078125
],
[
0.515625,
0.434326171875,
0.9140625,
0.434326171875,
0.9140625,
0.447509765625,
0.515625,
0.447509765625
],
[
0.0849609375,
0.4462890625,
0.483642578125,
0.4462890625,
0.483642578125,
0.460205078125,
0.0849609375,
0.460205078125
],
[
0.51708984375,
0.449462890625,
0.9140625,
0.449462890625,
0.9140625,
0.460205078125,
0.51708984375,
0.460205078125
],
[
0.0849609375,
0.460205078125,
0.16259765625,
0.460205078125,
0.16259765625,
0.4716796875,
0.0849609375,
0.4716796875
],
[
0.5185546875,
0.46337890625,
0.91259765625,
0.46337890625,
0.91259765625,
0.473388671875,
0.5185546875,
0.473388671875
],
[
0.2017822265625,
0.477783203125,
0.363525390625,
0.477783203125,
0.363525390625,
0.492431640625,
0.2017822265625,
0.492431640625
],
[
0.45751953125,
0.477783203125,
0.483642578125,
0.477783203125,
0.483642578125,
0.4912109375,
0.45751953125,
0.4912109375
],
[
0.51708984375,
0.47607421875,
0.9140625,
0.47607421875,
0.9140625,
0.4892578125,
0.51708984375,
0.4892578125
],
[
0.51708984375,
0.4892578125,
0.9140625,
0.4892578125,
0.9140625,
0.50244140625,
0.51708984375,
0.50244140625
],
[
0.0849609375,
0.5,
0.48193359375,
0.5,
0.48193359375,
0.51318359375,
0.0849609375,
0.51318359375
],
[
0.51708984375,
0.50244140625,
0.9140625,
0.50390625,
0.9140625,
0.517578125,
0.51708984375,
0.5166015625
],
[
0.0849609375,
0.51318359375,
0.483642578125,
0.51318359375,
0.483642578125,
0.5263671875,
0.0849609375,
0.5263671875
],
[
0.51708984375,
0.517578125,
0.9140625,
0.517578125,
0.9140625,
0.53076171875,
0.51708984375,
0.53076171875
],
[
0.08416748046875,
0.5263671875,
0.134033203125,
0.5263671875,
0.134033203125,
0.54150390625,
0.08416748046875,
0.54150390625
],
[
0.51708984375,
0.53076171875,
0.9140625,
0.53076171875,
0.9140625,
0.54541015625,
0.51708984375,
0.54541015625
],
[
0.273681640625,
0.54541015625,
0.28759765625,
0.54541015625,
0.28759765625,
0.5546875,
0.273681640625,
0.5546875
],
[
0.5185546875,
0.54541015625,
0.9140625,
0.54541015625,
0.9140625,
0.56005859375,
0.5185546875,
0.56005859375
],
[
0.26953125,
0.5517578125,
0.370849609375,
0.55615234375,
0.369384765625,
0.57470703125,
0.268798828125,
0.57080078125
],
[
0.2607421875,
0.5498046875,
0.26708984375,
0.5546875,
0.2607421875,
0.56005859375,
0.253173828125,
0.5546875
],
[
0.1993408203125,
0.5537109375,
0.250732421875,
0.56005859375,
0.2484130859375,
0.57470703125,
0.1968994140625,
0.568359375
],
[
0.455810546875,
0.55615234375,
0.484375,
0.55615234375,
0.484375,
0.57080078125,
0.455810546875,
0.57080078125
],
[
0.515625,
0.556640625,
0.82421875,
0.556640625,
0.82421875,
0.57080078125,
0.515625,
0.57080078125
],
[
0.25244140625,
0.56640625,
0.27685546875,
0.56640625,
0.27685546875,
0.57470703125,
0.25244140625,
0.57470703125
],
[
0.273681640625,
0.57470703125,
0.289306640625,
0.57470703125,
0.289306640625,
0.58203125,
0.273681640625,
0.58203125
],
[
0.53173828125,
0.57080078125,
0.9140625,
0.5712890625,
0.9140625,
0.5859375,
0.53173828125,
0.58544921875
],
[
0.0849609375,
0.59033203125,
0.48046875,
0.59033203125,
0.48046875,
0.60498046875,
0.0849609375,
0.60498046875
],
[
0.5185546875,
0.58544921875,
0.9140625,
0.58544921875,
0.9140625,
0.5986328125,
0.5185546875,
0.5986328125
],
[
0.0849609375,
0.60498046875,
0.48193359375,
0.60498046875,
0.48193359375,
0.61865234375,
0.0849609375,
0.61865234375
],
[
0.5185546875,
0.6015625,
0.91259765625,
0.6015625,
0.91259765625,
0.6123046875,
0.5185546875,
0.6123046875
],
[
0.51708984375,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.51708984375,
0.626953125
],
[
0.08660888671875,
0.62060546875,
0.48046875,
0.62060546875,
0.48046875,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.0849609375,
0.63330078125,
0.48193359375,
0.63330078125,
0.48193359375,
0.646484375,
0.0849609375,
0.646484375
],
[
0.5185546875,
0.62744140625,
0.91259765625,
0.62744140625,
0.91259765625,
0.6416015625,
0.5185546875,
0.6416015625
],
[
0.08660888671875,
0.6484375,
0.48046875,
0.6484375,
0.48046875,
0.6591796875,
0.08660888671875,
0.6591796875
],
[
0.5185546875,
0.64306640625,
0.91259765625,
0.64306640625,
0.91259765625,
0.65380859375,
0.5185546875,
0.65380859375
],
[
0.0849609375,
0.66015625,
0.416748046875,
0.66015625,
0.416748046875,
0.673828125,
0.0849609375,
0.673828125
],
[
0.51708984375,
0.65576171875,
0.9140625,
0.65576171875,
0.9140625,
0.66943359375,
0.51708984375,
0.66943359375
],
[
0.51708984375,
0.67041015625,
0.91259765625,
0.66943359375,
0.91259765625,
0.68115234375,
0.51708984375,
0.681640625
],
[
0.1805419921875,
0.68115234375,
0.386474609375,
0.68310546875,
0.386474609375,
0.70068359375,
0.1805419921875,
0.69873046875
],
[
0.455810546875,
0.6845703125,
0.484375,
0.6845703125,
0.484375,
0.69873046875,
0.455810546875,
0.69873046875
],
[
0.5185546875,
0.68505859375,
0.9111328125,
0.68505859375,
0.9111328125,
0.69580078125,
0.5185546875,
0.69580078125
],
[
0.1805419921875,
0.69873046875,
0.38818359375,
0.70068359375,
0.38818359375,
0.71826171875,
0.1805419921875,
0.71630859375
],
[
0.455810546875,
0.7021484375,
0.484375,
0.7021484375,
0.484375,
0.71630859375,
0.455810546875,
0.71630859375
],
[
0.51708984375,
0.6962890625,
0.9140625,
0.6962890625,
0.9140625,
0.71044921875,
0.51708984375,
0.71044921875
],
[
0.51708984375,
0.7109375,
0.9140625,
0.7109375,
0.9140625,
0.724609375,
0.51708984375,
0.724609375
],
[
0.08416748046875,
0.7236328125,
0.483642578125,
0.7236328125,
0.483642578125,
0.7412109375,
0.08416748046875,
0.7412109375
],
[
0.51708984375,
0.724609375,
0.9140625,
0.724609375,
0.9140625,
0.73779296875,
0.51708984375,
0.73779296875
],
[
0.0849609375,
0.7392578125,
0.48193359375,
0.7392578125,
0.48193359375,
0.75390625,
0.0849609375,
0.75390625
],
[
0.515625,
0.7373046875,
0.9140625,
0.73779296875,
0.9140625,
0.75244140625,
0.515625,
0.751953125
],
[
0.08416748046875,
0.75048828125,
0.48046875,
0.75244140625,
0.48046875,
0.767578125,
0.08416748046875,
0.76513671875
],
[
0.515625,
0.75048828125,
0.740234375,
0.751953125,
0.740234375,
0.7666015625,
0.515625,
0.76513671875
],
[
0.0849609375,
0.76806640625,
0.48046875,
0.76806640625,
0.48046875,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.08416748046875,
0.78076171875,
0.261474609375,
0.78076171875,
0.261474609375,
0.79443359375,
0.08416748046875,
0.79443359375
],
[
0.51708984375,
0.77978515625,
0.6806640625,
0.77978515625,
0.6806640625,
0.79443359375,
0.51708984375,
0.79443359375
],
[
0.515625,
0.79736328125,
0.9140625,
0.798828125,
0.9140625,
0.81298828125,
0.515625,
0.8125
],
[
0.0849609375,
0.81005859375,
0.111083984375,
0.81005859375,
0.111083984375,
0.8212890625,
0.0849609375,
0.8212890625
],
[
0.10455322265625,
0.8095703125,
0.314453125,
0.8095703125,
0.314453125,
0.82275390625,
0.10455322265625,
0.82275390625
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.81298828125,
0.9140625,
0.8271484375,
0.51708984375,
0.8271484375
],
[
0.101318359375,
0.8271484375,
0.48193359375,
0.8271484375,
0.48193359375,
0.84033203125,
0.101318359375,
0.84033203125
],
[
0.5185546875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.5009765625,
0.84130859375,
0.5009765625,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.08660888671875,
0.85693359375,
0.48046875,
0.85693359375,
0.48046875,
0.8671875,
0.08660888671875,
0.8671875
],
[
0.515625,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.515625,
0.86865234375
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.515625,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.515625,
0.8818359375
],
[
0.08251953125,
0.8818359375,
0.25,
0.8837890625,
0.25,
0.89794921875,
0.08251953125,
0.89599609375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"respectively.DsKL denotes symmetrized KLdivergence obtained by",
"Implementation Details.The word embedding size of decoder is",
"averaging the expected value of DkL (Pz, IIPz2) and DkL (Pz, lIPz))",
"set to 1024 and proportional coefficient α in loss function is set to",
"We calculate this loss by mutual learning [29] between CN and SN.",
"0.005. We set the dimension of vectors z to 50. When the encoder",
"In the mutual learning setup, in an iteration, the model will compute",
"is BERT,we set thedimension of zto 32.We adopt RoBERTa[11]",
"LsKLi and LsKL for CN and SNrespectively.In addition, we need",
"as the problem encoder.Following RoBERTa's setting, the hidden",
"to maximize I (u2; z) to ensure that the compressed representation",
"size of the encoder is set to 768, and we set the hidden size of the",
"zi has enough information to predict y. We use the chain rule to",
"decoderto 1024.We used Adamw[12] as the optimizerwith the",
"decompose I (u2; zi) into the following two terms:",
"learning rate as 5e-5. The mini-batch size is set to 16. We adopt",
"a beam search with the size of 5. Dropout (dropout rate = 0.5) is",
"I (U2; z1) = I(02; Z1 / y) +",
"I (z1;y)",
"(11)",
"employed to avoid overfitting.For Ape210K, we set the maximum",
"Predictive Information",
"sequencelength ofquestions as150 and that of solution expressions",
"as 50, similar to [23]. Our model takes 80 epochs on Math23k and",
"Redundancy",
"where y represents ground-truth solution expression. In practice, we",
"50epochs onApe210kfor convergence.",
"can maximize I (zi;y) = Ee log d (yn |e (xn,e) (included in VrB)",
"which is calculated to compress redundant information I (u2; z1 I y)",
"Baselines. We compare our model with several strong baseline",
"and indirectly maximize I (02;zi). Ideally, I (02;zi I y) should be",
"methods, including NumS2T [23], TSN-MD [27], MATH-EN [18],",
"zero. As suggested in [6], we minimize latent-specific redundancy",
"Multi-E/D[16],GTS[24], StackDecoder [4],KAS2T[23],Graph2tree",
"by jointly minimizing I (u1;z1|u2)and maximizingI(v2;zi).We",
"[28].andApe[30].",
"define the redundancy terms in Eq. (8) and Eq. (11) as the syntax-",
"irrelevant information.",
"4",
"",
"EXPERIMENTALRESULTS",
"4.1 Main Results",
"2.4 Self-distillation Loss",
"The evaluation metric is answer accuracy. Table 2 show the perfor-",
"In this section, we introduce a novel self-distillation loss to increase",
"mance comparison of ourmodel withbaseline methods on Math23K",
"the diversity of generated expressions. Suggested by [7], the I(u; z ]",
"and Ape210k, respectively. In Table 2, since there is a trade-off be-",
"y) in conditional information bottleneck can be variationally upper",
"tween the variety of expressions and the correctness of the answer,",
"bounded by:",
"wedonotaddVsDlintothesourcenetwork(SN)and thecollabo-",
"DkL (e (z /o\") ib (z /y\"))",
"(12)",
"rator network (CN). From Table 2 we can observe that our models",
"(both CN and SN) achieve consistently and substantially better per-",
"where e (z u\") defined in equation (2) is moved towards the con-",
"formance than the compared methods. In addition, the accuracy",
"ditional marginal b(y) ~ N (b\"(y), b (y)). We modify equation",
"of CN is higher than that of SN. This is because, in one iteration,",
"(12) as:",
"CN is provided with ytree predicted by SN when SN has not been",
"N",
"trained by ground-truth then SN is provided with ytree predicted",
"DsKL (gllz),",
"1",
"VsDL =",
"(13)",
"by CN when CN has been trained by ground-truth.",
"NL",
"7=1",
"We also measure the accuracy of solution expression.We con-",
"where y denotes that averaging the all hi. Intuitively, VsDL make",
"sider a solution expression as correct when the predicted expression",
"the decoder more rely on latent space of z which contains expres",
"exactly matches the ground truth solution.Generally,the math-",
"ematical expression generated by the tree decoder conforms to",
"sion syntax treeinformationfor all expressions when predicting",
"expression y. Benefiting from the randomness of z, the model can",
"the syntactic specification. As long as the answer obtained by the",
"generatemorediversesolutionexpressions.Finally,wecalculate",
"expression operationis consistentwiththeground-truth,thenwe",
"the loss functions Li for SN and L2 for CN as follows:",
"consider the expression to be a valid solution. We take the subtrac-",
"tionvaluebetween answeraccuracy(denotedasAnswer-Acc)and",
"Li = VIB + VsDL +α× LSKLI-",
"(14)",
"solution expression accuracy (denoted as Expression-Acc)as the",
"L2 =VIB + VsDL2 +α× LsKL2",
"(15)",
"diversity evaluation metric (denoted as Diversity) of the generated",
"solution expressions. As shown in Table 3, our model can generate",
"where α is a proportional coefficient.IB,VsDL,LsKL, are the",
"more diverse solution expressions that are not included in ground-",
"training objectives for SN. VIB2, VsDL, LsKL, are the training",
"truth expressions. As expected, the model with VsDL has better",
"objectives for CN.Based on empirical observation, although VsDL",
"diversity but lower answer accuracy.",
"can increase the diversity of solution expressions, it alsoreduces",
"accuracy of the final answer.",
"4.2Ablation Study",
"We conduct ablation test on Math23k to analyze the impact of",
"3",
"EXPERIMENTALSETUP",
"different components in ESIB. Since the best results are produced",
"Datasets. We conduct experiments on two benchmark MWP",
"byCN,weonly conduct ablation study on CN.First,weremovethe",
"datasets:Math23k[22] andApe210k[30].Math23kcontains22162/1000",
"mutual learning, denoted as CN w/o MT. Second, we remove the",
"questionsfortraining/testing,respectively.Ape21okiscomposed",
"VIB from SN and CN to evaluate the impact of VIB (denoted as CN",
"of166,270 questions for training, 4,157 questions for validation, and",
"w/o VIB). In addition, we report the results by removing both MT",
"4,159 questions for testing.",
"and VIB (denoted as CN w/o MT+VIB). To evaluate the impact of"
] |
{
"arxiv_id": "2310.15664",
"language": "en",
"page_idx": 3
}
|
||
2310.15664-en-4
|
2310.15664-en-4.jpg
|
Table 2: Solution accuracy of $ESIB$ and various baselines. Note
that Math23K denotes results on public test set.
<table>
<tbody>
<tr>
<th>$\mathbf{Models}$</th>
<th>$\mathbf{Math23k}$</th>
<th>$\mathbf{Ape210k}$</th>
</tr>
<tr>
<td>StackDecoder</td>
<td> </td>
<td>52.2</td>
</tr>
<tr>
<td>$GTS$</td>
<td>75.6</td>
<td>67.7</td>
</tr>
<tr>
<td>KAS2T</td>
<td>76.3</td>
<td>68.7</td>
</tr>
<tr>
<td>TSN- MD</td>
<td>77.4</td>
<td> </td>
</tr>
<tr>
<td>Graph2Tree</td>
<td>77.4</td>
<td> </td>
</tr>
<tr>
<td>Ape</td>
<td> </td>
<td>70.2</td>
</tr>
<tr>
<td>NumS2T</td>
<td>78.1</td>
<td>70.5</td>
</tr>
<tr>
<td>Multi- E/ D</td>
<td>78.4</td>
<td> </td>
</tr>
<tr>
<td>$SN$</td>
<td>84.2</td>
<td>76.3</td>
</tr>
<tr>
<td>$CN$</td>
<td>85.9</td>
<td>76.8</td>
</tr>
</tbody>
</table>
Table 3: Accuracy of diversity generation on Math23k with
Self-distillation Loss.
<table>
<tbody>
<tr>
<th>$\mathbf{Models}$</th>
<th>$\mathbf{Answer-Acc}$</th>
<th>$\mathbf{Equation-Acc}$</th>
<th>$\mathbf{Diversity}$</th>
</tr>
<tr>
<td>MATH-EN</td>
<td>66.7</td>
<td>60.1</td>
<td>6.6</td>
</tr>
<tr>
<td>$GTS$</td>
<td>75.6</td>
<td>64.8</td>
<td>10.8</td>
</tr>
<tr>
<td>TSN- MD</td>
<td>77.4</td>
<td>65.8</td>
<td>11.6</td>
</tr>
<tr>
<td>$\overline{{\mathrm{CN~with~}\mathcal{V}_{SDL}}}$</td>
<td>85.4</td>
<td>71.9</td>
<td>13.5</td>
</tr>
<tr>
<td>$CN$</td>
<td>85.9</td>
<td>73.5</td>
<td>12.4</td>
</tr>
</tbody>
</table>
Table 4: Ablation study on Math23k.
<table>
<tbody>
<tr>
<th>$\mathbf{Models}$</th>
<th>$\mathbf{Math23k}$</th>
</tr>
<tr>
<td>$CN$</td>
<td>85.9</td>
</tr>
<tr>
<td>$\mathbf{CNW/oMT}$</td>
<td>85.2</td>
</tr>
<tr>
<td>$\mathbf{CN}$w/oVIB</td>
<td>84.8</td>
</tr>
<tr>
<td>CN w/ o MT+ VIB</td>
<td>83.1</td>
</tr>
<tr>
<td>$\mathrm{CN}_{\mathrm{BERT}}$</td>
<td>84.3</td>
</tr>
<tr>
<td>$\mathrm{CN}_{\mathrm{BERT}}$ $\mathbf{W/oMT+VIB}$</td>
<td>82.4</td>
</tr>
</tbody>
</table>
# 4.3 Case Study
As an intuitive way to show the performance of $ESIB$, we randomly choose one problem form Math23k and show its solution expression generated by our model. As shown in Table 5, we observe that our model can produce two different but equivalent solutions. In addition, our model can correctly solve the problems in Table 1, which proves that our model does effectively reduce spurious correlations.
# 5 coNCLUSION
In this paper, we proposed an expression syntax information bottleneck method for MWP with mutual learning, which discards redundant features containing spurious correlations. Furthermore, we designed a self-distillation loss to encourage the model to rely more on the syntax information in the latent space. Extensive experiments on two benchmark MWP datasets demonstrated the effectiveness of our model.
# ACKNOWLEDGMENTS
This work was partially supported by Natural Science Foundation of Guangdong Province of China (No. 2021A1515011905), Shenzhen Science and Technology Innovation Program (Grant No. KQTD20190929172835662), Shenzhen Basic Research Foundation (No. JCYJ20210324115614039 and No. JCYJ20200109113441941)
# $\mathbf{REFERENCES}$
[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. 2016. Deep
variational information bottleneck. $arXivpreprintarXiv:1612.00410(2016).$
[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural ma-
chine translation by jointly learning to align and translate. arXiv preprint
$arXiv{:}1409.0473\left(2014\right).$
[3] Daniel G Bobrow. 1964. Natural language input for a computer problem solving
system. (1964).
[4] Ting-Rui Chiang and Yun-Nung Chen. 2018. Semantically-aligned equation
generation for solving and reasoning math word problems. arXiv preprint
$arXiv{:}1811.00720\left(2018\right).$
[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
Pre-training of deep bidirectional transformers for language understanding. arXiv
$preprintarXiv{: }18I\dot{0} . 04805$ (2018).
[6] Marco Federici, Anjan Dutta, Patrick Forré, Nate Kushman, and Zeynep Akata
2020. Learning robust representations via multi-view information bottleneck
arXiv preprint $arXiv: 2002. 07017$ (2020).
[7] Ian Fischer. 2020. The conditional entropy bottleneck. Entropy 22, 9 (2020), 999. [8] Zhenya Huang, Qi Liu, Weibo Gao, Jinze Wu, Yu Yin, Hao Wang, and Enhong
Chen. 2020. Neural mathematical solver with enhanced formula structure. In Proceedings of the 43rd International ACM SIGIR Conference on Research and $DevelopmentinInformationRetrieval.1729-1732.$
[9] Vivek Kumar, Rishabh Maheshwary, and Vikram Pudi. 2021. Adversarial Exam-
ples for Evaluating Math Word Problem Solvers. arXiv preprint arXiv:2109.05925
(2021).
[10] Zhenwen Liang and Xiangliang Zhang. 2021. Solving Math Word Problems with
Teacher Supervision. IfC$\tilde{A} I\left ( 2\tilde{0} 21\right ) .$
[11] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019).
[12] Ilya Loshchilov and Frank Hutter. 2018. Fixing weight decay regularization in
${\mathrm{adam.~(2018)}}.$
[13] Sudipto Mukherjee, Himanshu Asnani, and Sreeram Kannan. 2020. CCMI: Classi-
fier based conditional mutual information estimation. In Uncertainty in artificial
$intelligence.$ PMLR, 1083-1093.
[14] Yu Pan, Zeyong Su, Ao Liu, Jingquan Wang, Nannan Li, and Zenglin Xu. 2022
A Unified Weight Initialization Paradigm for Tensorial Convolutional Neural Networks. In $ICML$ $(Proceedingsof$ Machine Learning Research, Vol. 162).PMLR 17238-17257.
Table 5: Case study from Math23k. Demonstrates the ability
of the model to generate multiple solutions.
<table>
<tbody>
<tr>
<td>Problem Text: </td>
<td>A train leaves from place A at 7 o'clock and arrives at place B at 17 o'clock. The train travels 75 kilome ters per hour. How many kilome ters is the distance between the twa places?</td>
</tr>
<tr>
<td>Ground-truth</td>
<td>$(17-7)\times75$</td>
</tr>
<tr>
<td>Source Network:</td>
<td>$(17-7)\times75$</td>
</tr>
<tr>
<td>Collaborator Network</td>
<td>$17\times75-7\times75$</td>
</tr>
</tbody>
</table>
the pre-training model, we also replaced the RoBERTa encoder with BERT[5] (denoted as CN$_{\mathrm{BERT}}).$ We summarize the results in Table 4. Both the VIB strategy and mutual learning contribute greatly to the performance of $ESB.$
|
[
[
0.0849609375,
0.10479736328125,
0.48046875,
0.10479736328125,
0.48046875,
0.11871337890625,
0.0849609375,
0.11871337890625
],
[
0.5185546875,
0.10797119140625,
0.5439453125,
0.10797119140625,
0.5439453125,
0.11932373046875,
0.5185546875,
0.11932373046875
],
[
0.55810546875,
0.10797119140625,
0.646484375,
0.10797119140625,
0.646484375,
0.11932373046875,
0.55810546875,
0.11932373046875
],
[
0.08660888671875,
0.12060546875,
0.40185546875,
0.12060546875,
0.40185546875,
0.1307373046875,
0.08660888671875,
0.1307373046875
],
[
0.51708984375,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.51708984375,
0.1395263671875
],
[
0.51708984375,
0.1395263671875,
0.91259765625,
0.1395263671875,
0.91259765625,
0.15283203125,
0.51708984375,
0.15283203125
],
[
0.1829833984375,
0.1483154296875,
0.2344970703125,
0.1483154296875,
0.2344970703125,
0.160400390625,
0.1829833984375,
0.160400390625
],
[
0.2646484375,
0.1483154296875,
0.326904296875,
0.1483154296875,
0.326904296875,
0.160400390625,
0.2646484375,
0.160400390625
],
[
0.338134765625,
0.147705078125,
0.406005859375,
0.1458740234375,
0.406005859375,
0.160400390625,
0.338134765625,
0.1622314453125
],
[
0.5185546875,
0.154052734375,
0.91259765625,
0.154052734375,
0.91259765625,
0.1641845703125,
0.5185546875,
0.1641845703125
],
[
0.1641845703125,
0.1641845703125,
0.25244140625,
0.1641845703125,
0.25244140625,
0.1749267578125,
0.1641845703125,
0.1749267578125
],
[
0.35546875,
0.1610107421875,
0.386474609375,
0.1610107421875,
0.386474609375,
0.1767578125,
0.35546875,
0.1767578125
],
[
0.5185546875,
0.1685791015625,
0.9140625,
0.1685791015625,
0.9140625,
0.1787109375,
0.5185546875,
0.1787109375
],
[
0.279296875,
0.175537109375,
0.310546875,
0.175537109375,
0.310546875,
0.1912841796875,
0.279296875,
0.1912841796875
],
[
0.35546875,
0.175537109375,
0.386474609375,
0.175537109375,
0.386474609375,
0.1912841796875,
0.35546875,
0.1912841796875
],
[
0.51708984375,
0.1817626953125,
0.91259765625,
0.1817626953125,
0.91259765625,
0.1925048828125,
0.51708984375,
0.1925048828125
],
[
0.1846923828125,
0.189453125,
0.233642578125,
0.189453125,
0.233642578125,
0.203857421875,
0.1846923828125,
0.203857421875
],
[
0.193603515625,
0.1767578125,
0.2254638671875,
0.1767578125,
0.2254638671875,
0.189453125,
0.193603515625,
0.189453125
],
[
0.281005859375,
0.1912841796875,
0.310546875,
0.1912841796875,
0.310546875,
0.203857421875,
0.281005859375,
0.203857421875
],
[
0.35693359375,
0.1912841796875,
0.3857421875,
0.1912841796875,
0.3857421875,
0.203857421875,
0.35693359375,
0.203857421875
],
[
0.51708984375,
0.1944580078125,
0.9140625,
0.1944580078125,
0.9140625,
0.2083740234375,
0.51708984375,
0.2083740234375
],
[
0.1805419921875,
0.2052001953125,
0.236083984375,
0.2052001953125,
0.236083984375,
0.216552734375,
0.1805419921875,
0.216552734375
],
[
0.279296875,
0.203857421875,
0.31201171875,
0.203857421875,
0.31201171875,
0.2197265625,
0.279296875,
0.2197265625
],
[
0.36767578125,
0.208984375,
0.374267578125,
0.208984375,
0.374267578125,
0.21533203125,
0.36767578125,
0.21533203125
],
[
0.51708984375,
0.210205078125,
0.5947265625,
0.210205078125,
0.5947265625,
0.2215576171875,
0.51708984375,
0.2215576171875
],
[
0.17236328125,
0.2197265625,
0.2467041015625,
0.2197265625,
0.2467041015625,
0.2310791015625,
0.17236328125,
0.2310791015625
],
[
0.279296875,
0.2171630859375,
0.31201171875,
0.2171630859375,
0.31201171875,
0.23291015625,
0.279296875,
0.23291015625
],
[
0.36767578125,
0.2215576171875,
0.373291015625,
0.2215576171875,
0.373291015625,
0.2279052734375,
0.36767578125,
0.2279052734375
],
[
0.193603515625,
0.231689453125,
0.2254638671875,
0.231689453125,
0.2254638671875,
0.2474365234375,
0.193603515625,
0.2474365234375
],
[
0.35546875,
0.2310791015625,
0.386474609375,
0.2310791015625,
0.386474609375,
0.246826171875,
0.35546875,
0.246826171875
],
[
0.51953125,
0.24560546875,
0.53125,
0.24560546875,
0.53125,
0.255126953125,
0.51953125,
0.255126953125
],
[
0.54150390625,
0.242431640625,
0.6640625,
0.242431640625,
0.6640625,
0.255615234375,
0.54150390625,
0.255615234375
],
[
0.281005859375,
0.246826171875,
0.3095703125,
0.246826171875,
0.3095703125,
0.2587890625,
0.281005859375,
0.2587890625
],
[
0.1805419921875,
0.246826171875,
0.236083984375,
0.246826171875,
0.236083984375,
0.25830078125,
0.1805419921875,
0.25830078125
],
[
0.35693359375,
0.246826171875,
0.3857421875,
0.246826171875,
0.3857421875,
0.2587890625,
0.35693359375,
0.2587890625
],
[
0.176513671875,
0.260009765625,
0.240234375,
0.260009765625,
0.240234375,
0.271484375,
0.176513671875,
0.271484375
],
[
0.281005859375,
0.260009765625,
0.310546875,
0.260009765625,
0.310546875,
0.272705078125,
0.281005859375,
0.272705078125
],
[
0.51708984375,
0.261962890625,
0.91259765625,
0.261962890625,
0.91259765625,
0.272705078125,
0.51708984375,
0.272705078125
],
[
0.19775390625,
0.2734375,
0.221435546875,
0.2734375,
0.221435546875,
0.28857421875,
0.19775390625,
0.28857421875
],
[
0.27783203125,
0.2734375,
0.31201171875,
0.2734375,
0.31201171875,
0.2890625,
0.27783203125,
0.2890625
],
[
0.3544921875,
0.2734375,
0.386474609375,
0.2734375,
0.386474609375,
0.2890625,
0.3544921875,
0.2890625
],
[
0.51708984375,
0.274658203125,
0.9140625,
0.274658203125,
0.9140625,
0.28857421875,
0.51708984375,
0.28857421875
],
[
0.27783203125,
0.285888671875,
0.31201171875,
0.285888671875,
0.31201171875,
0.3017578125,
0.27783203125,
0.3017578125
],
[
0.3544921875,
0.285888671875,
0.38818359375,
0.285888671875,
0.38818359375,
0.3017578125,
0.3544921875,
0.3017578125
],
[
0.515625,
0.2890625,
0.9111328125,
0.2890625,
0.9111328125,
0.2998046875,
0.515625,
0.2998046875
],
[
0.51708984375,
0.3037109375,
0.9111328125,
0.3037109375,
0.9111328125,
0.314453125,
0.51708984375,
0.314453125
],
[
0.1968994140625,
0.287353515625,
0.22216796875,
0.287353515625,
0.22216796875,
0.300537109375,
0.1968994140625,
0.300537109375
],
[
0.515625,
0.31640625,
0.9140625,
0.31640625,
0.9140625,
0.330078125,
0.515625,
0.330078125
],
[
0.08416748046875,
0.328857421875,
0.48193359375,
0.330078125,
0.48193359375,
0.3447265625,
0.08416748046875,
0.343505859375
],
[
0.5146484375,
0.330078125,
0.9140625,
0.328857421875,
0.9140625,
0.343505859375,
0.5146484375,
0.3447265625
],
[
0.0849609375,
0.3447265625,
0.226318359375,
0.3447265625,
0.226318359375,
0.35791015625,
0.0849609375,
0.35791015625
],
[
0.51708984375,
0.3447265625,
0.6787109375,
0.3447265625,
0.6787109375,
0.354736328125,
0.51708984375,
0.354736328125
],
[
0.11767578125,
0.373779296875,
0.1707763671875,
0.373779296875,
0.1707763671875,
0.38818359375,
0.11767578125,
0.38818359375
],
[
0.2059326171875,
0.375732421875,
0.285888671875,
0.375732421875,
0.285888671875,
0.386474609375,
0.2059326171875,
0.386474609375
],
[
0.2998046875,
0.375732421875,
0.39208984375,
0.375732421875,
0.39208984375,
0.386962890625,
0.2998046875,
0.386962890625
],
[
0.40283203125,
0.37255859375,
0.47216796875,
0.375,
0.47216796875,
0.389404296875,
0.40283203125,
0.386962890625
],
[
0.51953125,
0.37744140625,
0.71142578125,
0.37744140625,
0.71142578125,
0.391357421875,
0.51953125,
0.391357421875
],
[
0.1094970703125,
0.389404296875,
0.17724609375,
0.389404296875,
0.17724609375,
0.40087890625,
0.1094970703125,
0.40087890625
],
[
0.2296142578125,
0.389404296875,
0.259033203125,
0.389404296875,
0.259033203125,
0.401611328125,
0.2296142578125,
0.401611328125
],
[
0.330810546875,
0.389404296875,
0.359375,
0.389404296875,
0.359375,
0.401611328125,
0.330810546875,
0.401611328125
],
[
0.4248046875,
0.389404296875,
0.450927734375,
0.389404296875,
0.450927734375,
0.401611328125,
0.4248046875,
0.401611328125
],
[
0.1282958984375,
0.401611328125,
0.16015625,
0.401611328125,
0.16015625,
0.4140625,
0.1282958984375,
0.4140625
],
[
0.51708984375,
0.396484375,
0.9140625,
0.396484375,
0.9140625,
0.410400390625,
0.51708984375,
0.410400390625
],
[
0.2303466796875,
0.40283203125,
0.2607421875,
0.40283203125,
0.2607421875,
0.415283203125,
0.2303466796875,
0.415283203125
],
[
0.330810546875,
0.40283203125,
0.359375,
0.40283203125,
0.359375,
0.415283203125,
0.330810546875,
0.415283203125
],
[
0.423095703125,
0.40283203125,
0.451904296875,
0.40283203125,
0.451904296875,
0.415283203125,
0.423095703125,
0.415283203125
],
[
0.330078125,
0.4140625,
0.36279296875,
0.4140625,
0.36279296875,
0.429931640625,
0.330078125,
0.429931640625
],
[
0.420654296875,
0.415283203125,
0.453369140625,
0.415283203125,
0.453369140625,
0.43115234375,
0.420654296875,
0.43115234375
],
[
0.51708984375,
0.410400390625,
0.9140625,
0.410400390625,
0.9140625,
0.423583984375,
0.51708984375,
0.423583984375
],
[
0.11517333984375,
0.416748046875,
0.17236328125,
0.416748046875,
0.17236328125,
0.427978515625,
0.11517333984375,
0.427978515625
],
[
0.2303466796875,
0.416748046875,
0.2607421875,
0.416748046875,
0.2607421875,
0.4287109375,
0.2303466796875,
0.4287109375
],
[
0.0931396484375,
0.427978515625,
0.1895751953125,
0.43115234375,
0.1895751953125,
0.44580078125,
0.0931396484375,
0.442626953125
],
[
0.515625,
0.423583984375,
0.9150390625,
0.42236328125,
0.9150390625,
0.436279296875,
0.515625,
0.4375
],
[
0.2279052734375,
0.429931640625,
0.261474609375,
0.429931640625,
0.261474609375,
0.44580078125,
0.2279052734375,
0.44580078125
],
[
0.328369140625,
0.429931640625,
0.363525390625,
0.429931640625,
0.363525390625,
0.44580078125,
0.328369140625,
0.44580078125
],
[
0.420654296875,
0.429931640625,
0.455078125,
0.429931640625,
0.455078125,
0.44580078125,
0.420654296875,
0.44580078125
],
[
0.328369140625,
0.443115234375,
0.36279296875,
0.443115234375,
0.36279296875,
0.458984375,
0.328369140625,
0.458984375
],
[
0.420654296875,
0.443115234375,
0.455078125,
0.443115234375,
0.455078125,
0.458984375,
0.420654296875,
0.458984375
],
[
0.51708984375,
0.438232421875,
0.9140625,
0.438232421875,
0.9140625,
0.451904296875,
0.51708984375,
0.451904296875
],
[
0.13232421875,
0.4443359375,
0.1568603515625,
0.4443359375,
0.1568603515625,
0.45703125,
0.13232421875,
0.45703125
],
[
0.2296142578125,
0.4443359375,
0.261474609375,
0.4443359375,
0.261474609375,
0.45703125,
0.2296142578125,
0.45703125
],
[
0.51708984375,
0.451904296875,
0.8037109375,
0.451904296875,
0.8037109375,
0.46533203125,
0.51708984375,
0.46533203125
],
[
0.16259765625,
0.4873046875,
0.40283203125,
0.4873046875,
0.40283203125,
0.50048828125,
0.16259765625,
0.50048828125
],
[
0.51708984375,
0.486083984375,
0.6337890625,
0.486083984375,
0.6337890625,
0.5,
0.51708984375,
0.5
],
[
0.521484375,
0.501953125,
0.9140625,
0.501953125,
0.9140625,
0.51513671875,
0.521484375,
0.51513671875
],
[
0.2197265625,
0.5166015625,
0.27294921875,
0.5166015625,
0.27294921875,
0.53076171875,
0.2197265625,
0.53076171875
],
[
0.54345703125,
0.51123046875,
0.8896484375,
0.51123046875,
0.8896484375,
0.5244140625,
0.54345703125,
0.5244140625
],
[
0.325927734375,
0.517578125,
0.390625,
0.517578125,
0.390625,
0.52880859375,
0.325927734375,
0.52880859375
],
[
0.521484375,
0.521484375,
0.9150390625,
0.521484375,
0.9150390625,
0.53515625,
0.521484375,
0.53515625
],
[
0.233642578125,
0.53076171875,
0.259033203125,
0.53076171875,
0.259033203125,
0.54443359375,
0.233642578125,
0.54443359375
],
[
0.339111328125,
0.52978515625,
0.373291015625,
0.52978515625,
0.373291015625,
0.54541015625,
0.339111328125,
0.54541015625
],
[
0.54150390625,
0.5322265625,
0.91259765625,
0.5341796875,
0.91259765625,
0.54443359375,
0.54150390625,
0.54248046875
],
[
0.2099609375,
0.54541015625,
0.283447265625,
0.54541015625,
0.283447265625,
0.556640625,
0.2099609375,
0.556640625
],
[
0.34228515625,
0.54541015625,
0.370849609375,
0.54541015625,
0.370849609375,
0.55810546875,
0.34228515625,
0.55810546875
],
[
0.5439453125,
0.54248046875,
0.64794921875,
0.54248046875,
0.64794921875,
0.55322265625,
0.5439453125,
0.55322265625
],
[
0.521484375,
0.5517578125,
0.9140625,
0.5517578125,
0.9140625,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.2091064453125,
0.5595703125,
0.283447265625,
0.5595703125,
0.283447265625,
0.57080078125,
0.2091064453125,
0.57080078125
],
[
0.34228515625,
0.5595703125,
0.371826171875,
0.5595703125,
0.371826171875,
0.5712890625,
0.34228515625,
0.5712890625
],
[
0.54345703125,
0.564453125,
0.61376953125,
0.5625,
0.61376953125,
0.57275390625,
0.54345703125,
0.57470703125
],
[
0.193603515625,
0.5712890625,
0.298095703125,
0.5712890625,
0.298095703125,
0.5859375,
0.193603515625,
0.5859375
],
[
0.340576171875,
0.5712890625,
0.371826171875,
0.5712890625,
0.371826171875,
0.5869140625,
0.340576171875,
0.5869140625
],
[
0.521484375,
0.5712890625,
0.9140625,
0.5712890625,
0.9140625,
0.58544921875,
0.521484375,
0.58544921875
],
[
0.2197265625,
0.5859375,
0.271240234375,
0.58837890625,
0.26953125,
0.60302734375,
0.2181396484375,
0.60107421875
],
[
0.34228515625,
0.5859375,
0.373291015625,
0.5859375,
0.373291015625,
0.6015625,
0.34228515625,
0.6015625
],
[
0.54345703125,
0.58349609375,
0.9140625,
0.58349609375,
0.9140625,
0.59326171875,
0.54345703125,
0.59326171875
],
[
0.5439453125,
0.59228515625,
0.65283203125,
0.59228515625,
0.65283203125,
0.60302734375,
0.5439453125,
0.60302734375
],
[
0.17724609375,
0.599609375,
0.31201171875,
0.5986328125,
0.31201171875,
0.6142578125,
0.17724609375,
0.61572265625
],
[
0.340576171875,
0.599609375,
0.373291015625,
0.599609375,
0.373291015625,
0.61572265625,
0.340576171875,
0.61572265625
],
[
0.521484375,
0.6015625,
0.9150390625,
0.6015625,
0.9150390625,
0.61572265625,
0.521484375,
0.61572265625
],
[
0.54150390625,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.62353515625,
0.54150390625,
0.62353515625
],
[
0.54345703125,
0.62353515625,
0.69287109375,
0.62353515625,
0.69287109375,
0.6337890625,
0.54345703125,
0.6337890625
],
[
0.521484375,
0.6318359375,
0.9140625,
0.6318359375,
0.9140625,
0.64599609375,
0.521484375,
0.64599609375
],
[
0.0849609375,
0.64306640625,
0.48046875,
0.64306640625,
0.48046875,
0.6572265625,
0.0849609375,
0.6572265625
],
[
0.54150390625,
0.642578125,
0.91259765625,
0.64306640625,
0.91259765625,
0.65478515625,
0.54150390625,
0.65380859375
],
[
0.54345703125,
0.6513671875,
0.7236328125,
0.6513671875,
0.7236328125,
0.662109375,
0.54345703125,
0.662109375
],
[
0.08660888671875,
0.6591796875,
0.378173828125,
0.6591796875,
0.378173828125,
0.66943359375,
0.08660888671875,
0.66943359375
],
[
0.52294921875,
0.66357421875,
0.91015625,
0.66357421875,
0.91015625,
0.673828125,
0.52294921875,
0.673828125
],
[
0.52294921875,
0.673828125,
0.91259765625,
0.673828125,
0.91259765625,
0.6845703125,
0.52294921875,
0.6845703125
],
[
0.1192626953125,
0.6875,
0.19775390625,
0.6875,
0.19775390625,
0.69873046875,
0.1192626953125,
0.69873046875
],
[
0.2548828125,
0.68798828125,
0.44677734375,
0.68798828125,
0.44677734375,
0.69873046875,
0.2548828125,
0.69873046875
],
[
0.5439453125,
0.68505859375,
0.9111328125,
0.68505859375,
0.9111328125,
0.6923828125,
0.5439453125,
0.6923828125
],
[
0.54345703125,
0.69140625,
0.91650390625,
0.69140625,
0.91650390625,
0.705078125,
0.54345703125,
0.705078125
],
[
0.256591796875,
0.70068359375,
0.4453125,
0.70068359375,
0.4453125,
0.7109375,
0.256591796875,
0.7109375
],
[
0.5439453125,
0.7041015625,
0.76953125,
0.7041015625,
0.76953125,
0.7138671875,
0.5439453125,
0.7138671875
],
[
0.256591796875,
0.71337890625,
0.44677734375,
0.71337890625,
0.44677734375,
0.7236328125,
0.256591796875,
0.7236328125
],
[
0.521484375,
0.7119140625,
0.91650390625,
0.7119140625,
0.91650390625,
0.72607421875,
0.521484375,
0.72607421875
],
[
0.256591796875,
0.724609375,
0.44775390625,
0.724609375,
0.44775390625,
0.73779296875,
0.256591796875,
0.73779296875
],
[
0.5400390625,
0.72265625,
0.9150390625,
0.72265625,
0.9150390625,
0.736328125,
0.5400390625,
0.736328125
],
[
0.25732421875,
0.73779296875,
0.4453125,
0.73779296875,
0.4453125,
0.74853515625,
0.25732421875,
0.74853515625
],
[
0.5439453125,
0.73486328125,
0.57421875,
0.73486328125,
0.57421875,
0.74365234375,
0.5439453125,
0.74365234375
],
[
0.515625,
0.7421875,
0.9140625,
0.7421875,
0.9140625,
0.755859375,
0.515625,
0.755859375
],
[
0.2548828125,
0.75048828125,
0.2998046875,
0.75048828125,
0.2998046875,
0.76220703125,
0.2548828125,
0.76220703125
],
[
0.5439453125,
0.75390625,
0.70751953125,
0.75390625,
0.70751953125,
0.7646484375,
0.5439453125,
0.7646484375
],
[
0.1192626953125,
0.7646484375,
0.1993408203125,
0.7646484375,
0.1993408203125,
0.77587890625,
0.1192626953125,
0.77587890625
],
[
0.25244140625,
0.76318359375,
0.330078125,
0.76318359375,
0.330078125,
0.7744140625,
0.25244140625,
0.7744140625
],
[
0.5185546875,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.7744140625,
0.5185546875,
0.7744140625
],
[
0.2548828125,
0.77587890625,
0.330078125,
0.77587890625,
0.330078125,
0.7861328125,
0.2548828125,
0.7861328125
],
[
0.54150390625,
0.77294921875,
0.9140625,
0.77294921875,
0.9140625,
0.7861328125,
0.54150390625,
0.7861328125
],
[
0.54345703125,
0.78271484375,
0.9150390625,
0.78271484375,
0.9150390625,
0.796875,
0.54345703125,
0.796875
],
[
0.12091064453125,
0.7763671875,
0.2132568359375,
0.7763671875,
0.2132568359375,
0.78857421875,
0.12091064453125,
0.78857421875
],
[
0.1192626953125,
0.7890625,
0.2442626953125,
0.7890625,
0.2442626953125,
0.7998046875,
0.1192626953125,
0.7998046875
],
[
0.256591796875,
0.7890625,
0.3505859375,
0.7890625,
0.3505859375,
0.7998046875,
0.256591796875,
0.7998046875
],
[
0.54150390625,
0.79443359375,
0.57861328125,
0.79443359375,
0.57861328125,
0.8046875,
0.54150390625,
0.8046875
],
[
0.51708984375,
0.80322265625,
0.9140625,
0.80322265625,
0.9140625,
0.81640625,
0.51708984375,
0.81640625
],
[
0.54150390625,
0.814453125,
0.60546875,
0.8125,
0.60546875,
0.82373046875,
0.54150390625,
0.82568359375
],
[
0.5146484375,
0.82275390625,
0.9111328125,
0.82275390625,
0.9111328125,
0.8359375,
0.5146484375,
0.8359375
],
[
0.54150390625,
0.83349609375,
0.9140625,
0.833984375,
0.9140625,
0.84521484375,
0.54150390625,
0.8447265625
],
[
0.0849609375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.5439453125,
0.8447265625,
0.69140625,
0.8447265625,
0.69140625,
0.85498046875,
0.5439453125,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.869140625,
0.0849609375,
0.869140625
],
[
0.51708984375,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86572265625,
0.51708984375,
0.86572265625
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.54150390625,
0.86572265625,
0.91259765625,
0.86572265625,
0.91259765625,
0.87548828125,
0.54150390625,
0.87548828125
],
[
0.5400390625,
0.87353515625,
0.9140625,
0.875,
0.9140625,
0.88623046875,
0.5400390625,
0.88525390625
],
[
0.0849609375,
0.88330078125,
0.2344970703125,
0.8818359375,
0.2344970703125,
0.89599609375,
0.0849609375,
0.896484375
],
[
0.54150390625,
0.88525390625,
0.60888671875,
0.88525390625,
0.60888671875,
0.89599609375,
0.54150390625,
0.89599609375
]
] |
[
"Table 2: Solution accuracy of ESIB and various baselines. Note",
"4.3",
"Case Study",
"thatMath23K denotes results onpublictest set.",
"As an intuitive way to show the performance of ESIB, we randomly",
"choose one problem form Math23k and show its solution expres",
"Models",
"Math23k",
"Ape210k",
"siongeneratedbyourmodel.AsshowninTable5,weobserve",
"StackDecoder",
"52.2",
"thatourmodel canproducetwodifferentbutequivalent solutions.",
"75.6",
"67.7",
"In addition, ourmodel cancorrectly solve theproblems in Table",
"KAS2T",
"GTS",
"76.3",
"68.7",
"1, which proves that our model does effectively reduce spurious",
"TSN-MD",
"77.4",
"-",
"correlations.",
"Graph2Tree",
"77.4",
"-",
"Ape",
"70.2",
"5",
"CONCLUSION",
"78.1",
"NumS2T",
"70.5",
"Multi-E/D",
"78.4",
"In this paper,weproposed an expression syntax information bot-",
"SN",
"84.2",
"76.3",
"tleneck method for MWP with mutual learning, which discards",
"85.9",
"76.8",
"redundantfeaturescontainingspuriouscorrelations.Furthermore",
"we designed a self-distillation loss to encourage the model to rely",
"CN",
"more on the syntax information in the latent space. Extensive ex-",
"Table 3: Accuracy of diversity generation on Math23k with",
"periments on two benchmark MWP datasets demonstrated the",
"Self-distillation Loss.",
"effectivenessofourmodel.",
"Models",
"Answer-Acc",
"Equation-Acc",
"Diversity",
"ACKNOWLEDGMENTS",
"MATH-EN",
"66.7",
"60.1",
"6.6",
"GTS",
"This work was partially supported by Natural Science Foundation",
"75.6",
"64.8",
"10.8",
"65.8",
"11.6",
"of Guangdong Province of China (No.2021A1515011905), Shenzhen",
"TSN-MD",
"77.4",
"CNwithVsDL",
"Science and Technology Innovation Program (Grant No. KQTD20190",
"85.4",
"71.9",
"13.5",
"73.5",
"12.4",
"929172835662),Shenzhen Basic Research Foundation (No.JCYJ20210",
"CN",
"85.9",
"324115614039andNo.JCYJ20200109113441941)",
"Table 4: Ablation study on Math23k.",
"REFERENCES",
"[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. 2016. Deep",
"Models",
"variational information bottleneck. arXiv preprint arXiv:1612.00410 (2016).",
"Math23k",
"[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural ma-",
"CN",
"85.9",
"chine translation by jointly learning to align and translate. arXiv preprint",
"CN w/o MT",
"85.2",
"arxiv:1409.0473(2014)",
"[3] Daniel G Bobrow. 1964. Natural language input for a computer problem solving",
"CN w/o VIB",
"84.8",
"system. (1964)",
"CN w/o MT+VIB",
"83.1",
"[4] Ting-Rui Chiang and Yun-Nung Chen. 2018. Semantically-aligned equation",
"CNBERT",
"84.3",
"generation for solving and reasoning math word problems.arXiv preprint",
"arXiv:1811.00720(2018)",
"CNBERT W/o MT+VIB",
"82.4",
"[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:",
"Pre-training of deep bidirectional transformers for language understanding. arXiv",
"preprint arXiv:1810.04805 (2018)",
"[6] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushman, and Zeynep Akata.",
"Table 5: Case study from Math23k.Demonstrates the ability",
"2020. Learning robust representations via multi-view information bottleneck",
"arXiv preprint arXiv:2002.07017(2020).",
"of themodel togeneratemultiplesolutions.",
"[7] Ian Fischer.2020.The conditional entropy bottleneck. Entropy 22, 9 (2020),999.",
"[8] Zhenya Huang, Qi Liu, Weibo Gao, Jinze Wu, Yu Yin, Hao Wang, and Enhong",
"Problem Text:",
"AtrainleavesfromplaceAat7",
"Chen.2020.Neuralmathematical solver with enhanced formula structure.In",
"Proceedings of the 43rd International ACM SIGIR Conference on Research and",
"o'clock and arrives at place B at 17",
"Development in Information Retrieval.1729-1732",
"o'clock.Thetrain travels 75kilome-",
"[9] Vivek Kumar, Rishabh Maheshwary, and Vikram Pudi. 2021. Adversarial Exam-",
"ters per hour. How many kilome-",
"ples for Evaluating Math Word Problem Solvers. arXiv preprint arXiv:2109.05925",
"ters is the distance between the two",
"(2021)",
"[10] Zhenwen Liang and Xiangliang Zhang. 2021. Solving Math Word Problems with",
"places?",
"Teacher Supervision. IJCAI (2021).",
"Ground-truth:",
"(17 7) × 75",
"[11] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer",
"(17 - 7) × 75",
"Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A",
"robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692",
"Source Network:",
"CollaboratorNetwork:",
"17×757×75",
"(2019).",
"[12] Ilya Loshchilov and Frank Hutter. 20i8. Fixing weight decay regularization in",
"adam.(2018)",
"[13] Sudipto Mukherjee, Himanshu Asnani, and Sreeram Kannan. 2020. CCMI: Classi",
"fier based conditional mutual information estimation. In Uncertainty in artificial",
"the pre-training model, we also replaced the RoBERTa encoder with",
"intelligence. PMLR, 1083-1093.",
"BERT [5] (denoted as CNBERT). We summarize the results in Table",
"[14] Yu Pan, Zeyong Su, Ao Liu, Jingquan Wang, Nannan Li, and Zenglin Xu. 2022",
"4. Both the VIB strategy and mutual learning contribute greatly to",
"A Unified Weight Initialization Paradigm for Tensorial Convolutional Neural",
"Networks. In ICML (Proceedings of Machine Learning Research, Vol. 162). PMLR,",
"the performance of ESIB",
"17238-17257."
] |
{
"arxiv_id": "2310.15664",
"language": "en",
"page_idx": 4
}
|
||
2402.16086-en-3
|
2402.16086-en-3.jpg
|
<FigureHere>
Figure 2: Diagram of our re-ranking process with the DHE network. The feature maps $f_q$ and $f_c$ of the query image $I_q$ and a candidate image $I_c$ are fed into the Similarity Matching Module to compute the similarity map $s_qc.$ Then the Homography Regression Module uses the $s_qc$ to yield the homography matrix $\mathbf{H}_{qc}$ for geometric verification of mutual matches
Berton et al. 2021). Inspired by this, this module first computes the 4 pairs of points (4-point correspondence) on the $\operatorname*{images}I_{q}$ and $I_{c}$,i.e., a 16D vector. This process is implemented by using 6 stacked transformer encoder layers followed by a fully connected layer, and a shortcut connection is used between the third encoder layer and the sixth encoder layer. The transformer encoder can provide a larger effective perceptual field on the similarity map. And it is denoted as
That is, if the most similar local feature in image $I_q$ of the feature $f_c(y)$ is $f_q(x)$,and the most similar feature in image $I_{c}$ of the feature $\boldsymbol f_q(x)$ is $\boldsymbol f_c(y)$,then $(x,y)\in\mathcal{MN}.$
The $x$ and $y$ are the indices of corresponding patches. Assume that each patch corresponds to the Cartesian coordinates of a 2D image point in the center of the patch. We denote the homogeneous coordinates of a pair of matched patches as $\mathbf{p}_q$ and p$_c.$ If the re-projection error is less than a threshold $\theta$, then it is an inlier. That is, an inlier satisfies
$$\|C(\mathbf{p}_c)-C(\mathbf{H}_{qc}\mathbf{p}_q)\|\leq\theta,$$
(8)
(3)
$$\begin{aligned}TransE(\boldsymbol{s}_{qc}+\boldsymbol{E}_{pos})&=[\underbrace{\boldsymbol{p}_{q1},\ldots,\boldsymbol{p}_{q4}}_{\boldsymbol{P}_{q}},\underbrace{\boldsymbol{p}_{c1},\ldots,\boldsymbol{p}_{c4}}_{\boldsymbol{P}_{c}}]\\&=[P_{q},P_{c}],\end{aligned}$$
where $C(\cdot)$ means converting homogeneous coordinates to Cartesian (inhomogeneous) coordinates. We re-rank the candidate images using the number of inliers between the query and candidates.
where $p_{q1},\ldots,p_{q4}$ and $p_c1,\ldots,p_{c4}$ are four points on $I_q$
and $I_c$,respectively. $[\boldsymbol P_q,\boldsymbol P_c]$ is their concatenation.
After obtaining 4 pairs of points on two images, we use the Direct Linear Transform (Hartley et al. 2003) to yield the homography $\mathbf{H}_{qc}$,which is represented as
# Training Strategy
$$\mathbf{H}_{qc}=DLT(\mathbf{P}_{q},\mathbf{P}_{c}).$$
(4)
We succinctly denote the whole process of this module as
$$\mathbf{H}_{qc}=R(s_{qc}).$$
(5)
The whole process of the DHE network is summarized as
(6)
$$\mathbf{H}_{qc}=DHE(\boldsymbol{f}_q,\boldsymbol{f}_c)=R\big(S(\boldsymbol{f}_q,\boldsymbol{f}_c)\big).$$
Re-ranking with Geometric Verification
In common deep homography works, an image can be aligned with another image after using the homography transformation because the two images present a (approximately) same planar scene. So, they can train networks by minimizing the distance between two aligned images. But this prior condition does not hold for place images in many scenes. Meanwhile, we also cannot directly obtain homography labels to supervise the training of our DHE network. One possible solution is to train the network with the la-
bel provided by RANSAC. However, we found that the network is difficult to converge whether using the homography parameters yielded by RANSAC for direct supervision, or using the selected inliers and their re-projection errors for supervision. We propose to use only the number of inliers provided by RANSAC as supervision information, let the network autonomously decide which matched pairs are inliers, and optimize the network by minimizing the average re-projection error of these“inliers”.This makes it easier for networks to converge. Meanwhile, the loss produced by this process has the potential to optimize the feature extractor through back-propagation, so that the extracted feature maps are more suitable for local matching and outperform the method directly using RANSAC.
We fırst initialize the backbone and the DHE network individually. The initialization of the backbone follows the
In the re-ranking stage, we first calculate the mutual nearest neighbor matching of local features through exhaustive cross-matching. Then the homography matrix yielded by the DHE network is used to check geometric verification and remove outliers. Finally, we use the number of inliers as the similarities of image pairs to re-rank the candidates.
$\begin{array}{c}\mathrm{The~similarity~map~}s_{qc}\text{between the query}I_{q}\text{ and a can-}\\\mathrm{didate~}I_{c}\text{ has been computed. We apply it to search mutual}\end{array}$
nearest neighbor matches set $\mathcal{MN}$,which is defined as
$$\begin{aligned}\mathcal{MN}=\{(x,y):&\:x=\arg\max_i\boldsymbol{s}_{qc}(i,y),\\&\:y=\arg\max_j\boldsymbol{s}_{qc}(x,j)\}.\end{aligned}$$
(7)
|
[
[
0.16259765625,
0.0706787109375,
0.2017822265625,
0.0706787109375,
0.2017822265625,
0.080810546875,
0.16259765625,
0.080810546875
],
[
0.3798828125,
0.07257080078125,
0.4755859375,
0.07257080078125,
0.4755859375,
0.08331298828125,
0.3798828125,
0.08331298828125
],
[
0.423095703125,
0.0953369140625,
0.435546875,
0.0953369140625,
0.435546875,
0.10162353515625,
0.423095703125,
0.10162353515625
],
[
0.44287109375,
0.0953369140625,
0.453369140625,
0.0953369140625,
0.453369140625,
0.0985107421875,
0.44287109375,
0.0985107421875
],
[
0.451904296875,
0.0985107421875,
0.48046875,
0.1029052734375,
0.47802734375,
0.11553955078125,
0.449462890625,
0.111083984375
],
[
0.459228515625,
0.09344482421875,
0.483642578125,
0.09344482421875,
0.483642578125,
0.10162353515625,
0.459228515625,
0.10162353515625
],
[
0.30810546875,
0.121826171875,
0.328369140625,
0.121826171875,
0.328369140625,
0.1337890625,
0.30810546875,
0.1337890625
],
[
0.400390625,
0.1275634765625,
0.410888671875,
0.1275634765625,
0.410888671875,
0.1319580078125,
0.400390625,
0.1319580078125
],
[
0.529296875,
0.1275634765625,
0.55224609375,
0.1275634765625,
0.55224609375,
0.1363525390625,
0.529296875,
0.1363525390625
],
[
0.564453125,
0.1307373046875,
0.62158203125,
0.1307373046875,
0.62158203125,
0.141357421875,
0.564453125,
0.141357421875
],
[
0.44775390625,
0.1319580078125,
0.521484375,
0.1319580078125,
0.521484375,
0.1427001953125,
0.44775390625,
0.1427001953125
],
[
0.2607421875,
0.1351318359375,
0.293212890625,
0.1351318359375,
0.293212890625,
0.1427001953125,
0.2607421875,
0.1427001953125
],
[
0.338134765625,
0.1300048828125,
0.390625,
0.1319580078125,
0.390625,
0.143310546875,
0.338134765625,
0.141357421875
],
[
0.340576171875,
0.1395263671875,
0.3896484375,
0.1427001953125,
0.38818359375,
0.154052734375,
0.339111328125,
0.15087890625
],
[
0.418212890625,
0.14013671875,
0.4345703125,
0.14013671875,
0.4345703125,
0.14453125,
0.418212890625,
0.14453125
],
[
0.453369140625,
0.1427001953125,
0.515625,
0.14013671875,
0.515625,
0.1514892578125,
0.453369140625,
0.154052734375
],
[
0.56396484375,
0.141357421875,
0.6259765625,
0.141357421875,
0.6259765625,
0.1514892578125,
0.56396484375,
0.1514892578125
],
[
0.63134765625,
0.14453125,
0.68701171875,
0.14453125,
0.68701171875,
0.1558837890625,
0.63134765625,
0.1558837890625
],
[
0.30810546875,
0.1546630859375,
0.334228515625,
0.154052734375,
0.334228515625,
0.1578369140625,
0.30810546875,
0.1590576171875
],
[
0.25732421875,
0.14453125,
0.301513671875,
0.14453125,
0.301513671875,
0.1514892578125,
0.25732421875,
0.1514892578125
],
[
0.63818359375,
0.1546630859375,
0.68115234375,
0.1546630859375,
0.68115234375,
0.1654052734375,
0.63818359375,
0.1654052734375
],
[
0.73046875,
0.1572265625,
0.81884765625,
0.1572265625,
0.81884765625,
0.1685791015625,
0.73046875,
0.1685791015625
],
[
0.7197265625,
0.1666259765625,
0.830078125,
0.1666259765625,
0.830078125,
0.179931640625,
0.7197265625,
0.179931640625
],
[
0.67236328125,
0.1962890625,
0.76123046875,
0.1962890625,
0.76123046875,
0.23291015625,
0.67236328125,
0.23291015625
],
[
0.1544189453125,
0.205810546875,
0.2156982421875,
0.205810546875,
0.2156982421875,
0.216552734375,
0.1544189453125,
0.216552734375
],
[
0.369384765625,
0.2171630859375,
0.461669921875,
0.2171630859375,
0.461669921875,
0.2291259765625,
0.369384765625,
0.2291259765625
],
[
0.365234375,
0.2291259765625,
0.4599609375,
0.2291259765625,
0.4599609375,
0.2392578125,
0.365234375,
0.2392578125
],
[
0.69287109375,
0.2310791015625,
0.75634765625,
0.2310791015625,
0.75634765625,
0.242431640625,
0.69287109375,
0.242431640625
],
[
0.80126953125,
0.2291259765625,
0.8486328125,
0.2291259765625,
0.8486328125,
0.236083984375,
0.80126953125,
0.236083984375
],
[
0.69775390625,
0.2412109375,
0.74853515625,
0.242431640625,
0.74853515625,
0.246826171875,
0.69775390625,
0.24560546875
],
[
0.8037109375,
0.2412109375,
0.82421875,
0.2412109375,
0.82421875,
0.24560546875,
0.8037109375,
0.24560546875
],
[
0.08660888671875,
0.255615234375,
0.91259765625,
0.255615234375,
0.91259765625,
0.26953125,
0.08660888671875,
0.26953125
],
[
0.0849609375,
0.270751953125,
0.9111328125,
0.270751953125,
0.9111328125,
0.2841796875,
0.0849609375,
0.2841796875
],
[
0.08660888671875,
0.2841796875,
0.83837890625,
0.2841796875,
0.83837890625,
0.298583984375,
0.08660888671875,
0.298583984375
],
[
0.0882568359375,
0.311279296875,
0.47802734375,
0.311279296875,
0.47802734375,
0.324462890625,
0.0882568359375,
0.324462890625
],
[
0.521484375,
0.311279296875,
0.91259765625,
0.311279296875,
0.91259765625,
0.324462890625,
0.521484375,
0.324462890625
],
[
0.08660888671875,
0.32568359375,
0.4794921875,
0.32568359375,
0.4794921875,
0.339111328125,
0.08660888671875,
0.339111328125
],
[
0.521484375,
0.32568359375,
0.91259765625,
0.32568359375,
0.91259765625,
0.34033203125,
0.521484375,
0.34033203125
],
[
0.0849609375,
0.339111328125,
0.47802734375,
0.339111328125,
0.47802734375,
0.352783203125,
0.0849609375,
0.352783203125
],
[
0.521484375,
0.34033203125,
0.865234375,
0.34033203125,
0.865234375,
0.354736328125,
0.521484375,
0.354736328125
],
[
0.0849609375,
0.352783203125,
0.47802734375,
0.352783203125,
0.47802734375,
0.3662109375,
0.0849609375,
0.3662109375
],
[
0.5361328125,
0.354736328125,
0.91015625,
0.354736328125,
0.91015625,
0.36865234375,
0.5361328125,
0.36865234375
],
[
0.08660888671875,
0.3662109375,
0.4794921875,
0.3662109375,
0.4794921875,
0.380126953125,
0.08660888671875,
0.380126953125
],
[
0.51953125,
0.36865234375,
0.9111328125,
0.36865234375,
0.9111328125,
0.3818359375,
0.51953125,
0.3818359375
],
[
0.08660888671875,
0.380615234375,
0.48046875,
0.380615234375,
0.48046875,
0.39453125,
0.08660888671875,
0.39453125
],
[
0.51953125,
0.3818359375,
0.91259765625,
0.3818359375,
0.91259765625,
0.396484375,
0.51953125,
0.396484375
],
[
0.0849609375,
0.39453125,
0.4794921875,
0.39453125,
0.4794921875,
0.40771484375,
0.0849609375,
0.40771484375
],
[
0.51953125,
0.396484375,
0.9140625,
0.396484375,
0.9140625,
0.410400390625,
0.51953125,
0.410400390625
],
[
0.0849609375,
0.4091796875,
0.471435546875,
0.4091796875,
0.471435546875,
0.42236328125,
0.0849609375,
0.42236328125
],
[
0.5185546875,
0.410400390625,
0.9140625,
0.4091796875,
0.9140625,
0.423583984375,
0.5185546875,
0.4248046875
],
[
0.0931396484375,
0.4267578125,
0.44775390625,
0.4287109375,
0.44775390625,
0.449462890625,
0.0931396484375,
0.447509765625
],
[
0.521484375,
0.425537109375,
0.88623046875,
0.425537109375,
0.88623046875,
0.436279296875,
0.521484375,
0.436279296875
],
[
0.62353515625,
0.44140625,
0.8095703125,
0.44140625,
0.8095703125,
0.455810546875,
0.62353515625,
0.455810546875
],
[
0.890625,
0.442626953125,
0.91259765625,
0.442626953125,
0.91259765625,
0.455078125,
0.890625,
0.455078125
],
[
0.390625,
0.45068359375,
0.410888671875,
0.45068359375,
0.410888671875,
0.464111328125,
0.390625,
0.464111328125
],
[
0.45751953125,
0.45068359375,
0.47802734375,
0.45068359375,
0.47802734375,
0.464111328125,
0.45751953125,
0.464111328125
],
[
0.51953125,
0.458984375,
0.9140625,
0.458984375,
0.9140625,
0.472900390625,
0.51953125,
0.472900390625
],
[
0.301513671875,
0.45263671875,
0.32275390625,
0.45263671875,
0.32275390625,
0.466552734375,
0.301513671875,
0.466552734375
],
[
0.2459716796875,
0.468505859375,
0.33251953125,
0.468505859375,
0.33251953125,
0.486083984375,
0.2459716796875,
0.486083984375
],
[
0.521484375,
0.474853515625,
0.91015625,
0.474853515625,
0.91015625,
0.48486328125,
0.521484375,
0.48486328125
],
[
0.5185546875,
0.486083984375,
0.91259765625,
0.4873046875,
0.91259765625,
0.501953125,
0.5185546875,
0.50048828125
],
[
0.0849609375,
0.492431640625,
0.47802734375,
0.492431640625,
0.47802734375,
0.51025390625,
0.0849609375,
0.51025390625
],
[
0.51708984375,
0.5,
0.62353515625,
0.50048828125,
0.62353515625,
0.51513671875,
0.51708984375,
0.5146484375
],
[
0.08416748046875,
0.5087890625,
0.42724609375,
0.5087890625,
0.42724609375,
0.52294921875,
0.08416748046875,
0.52294921875
],
[
0.10455322265625,
0.5244140625,
0.47802734375,
0.5244140625,
0.47802734375,
0.53515625,
0.10455322265625,
0.53515625
],
[
0.51953125,
0.52587890625,
0.65625,
0.52587890625,
0.65625,
0.54052734375,
0.51953125,
0.54052734375
],
[
0.08660888671875,
0.537109375,
0.4794921875,
0.537109375,
0.4794921875,
0.55029296875,
0.08660888671875,
0.55029296875
],
[
0.51953125,
0.54345703125,
0.91259765625,
0.54345703125,
0.91259765625,
0.556640625,
0.51953125,
0.556640625
],
[
0.0849609375,
0.55029296875,
0.38818359375,
0.55029296875,
0.38818359375,
0.564453125,
0.0849609375,
0.564453125
],
[
0.5185546875,
0.55615234375,
0.91259765625,
0.556640625,
0.91259765625,
0.5712890625,
0.5185546875,
0.57080078125
],
[
0.2059326171875,
0.57275390625,
0.35693359375,
0.57275390625,
0.35693359375,
0.5869140625,
0.2059326171875,
0.5869140625
],
[
0.45751953125,
0.57373046875,
0.4794921875,
0.57373046875,
0.4794921875,
0.5859375,
0.45751953125,
0.5859375
],
[
0.51953125,
0.5712890625,
0.9111328125,
0.5712890625,
0.9111328125,
0.58544921875,
0.51953125,
0.58544921875
],
[
0.51953125,
0.58544921875,
0.91259765625,
0.58544921875,
0.91259765625,
0.5986328125,
0.51953125,
0.5986328125
],
[
0.0849609375,
0.59326171875,
0.465576171875,
0.5947265625,
0.465576171875,
0.609375,
0.0849609375,
0.60791015625
],
[
0.521484375,
0.60107421875,
0.91259765625,
0.60107421875,
0.91259765625,
0.611328125,
0.521484375,
0.611328125
],
[
0.2303466796875,
0.6171875,
0.330810546875,
0.6171875,
0.330810546875,
0.6318359375,
0.2303466796875,
0.6318359375
],
[
0.45751953125,
0.6171875,
0.4794921875,
0.6171875,
0.4794921875,
0.630859375,
0.45751953125,
0.630859375
],
[
0.51953125,
0.61279296875,
0.91259765625,
0.61279296875,
0.91259765625,
0.626953125,
0.51953125,
0.626953125
],
[
0.521484375,
0.626953125,
0.91015625,
0.626953125,
0.91015625,
0.64013671875,
0.521484375,
0.64013671875
],
[
0.08660888671875,
0.6396484375,
0.464111328125,
0.6396484375,
0.464111328125,
0.65283203125,
0.08660888671875,
0.65283203125
],
[
0.51953125,
0.64013671875,
0.904296875,
0.64013671875,
0.904296875,
0.65380859375,
0.51953125,
0.65380859375
],
[
0.5361328125,
0.65380859375,
0.91259765625,
0.65380859375,
0.91259765625,
0.66748046875,
0.5361328125,
0.66748046875
],
[
0.146240234375,
0.66015625,
0.414306640625,
0.66015625,
0.414306640625,
0.67822265625,
0.146240234375,
0.67822265625
],
[
0.45751953125,
0.66357421875,
0.47802734375,
0.66357421875,
0.47802734375,
0.6767578125,
0.45751953125,
0.6767578125
],
[
0.51953125,
0.66748046875,
0.91259765625,
0.66748046875,
0.91259765625,
0.68115234375,
0.51953125,
0.68115234375
],
[
0.5185546875,
0.68115234375,
0.91259765625,
0.681640625,
0.91259765625,
0.69580078125,
0.5185546875,
0.6943359375
],
[
0.08660888671875,
0.68798828125,
0.39208984375,
0.68798828125,
0.39208984375,
0.70263671875,
0.08660888671875,
0.70263671875
],
[
0.51953125,
0.69580078125,
0.91259765625,
0.69580078125,
0.91259765625,
0.708984375,
0.51953125,
0.708984375
],
[
0.0849609375,
0.70703125,
0.47802734375,
0.70703125,
0.47802734375,
0.72021484375,
0.0849609375,
0.72021484375
],
[
0.5185546875,
0.70849609375,
0.91259765625,
0.708984375,
0.91259765625,
0.7236328125,
0.5185546875,
0.72265625
],
[
0.08416748046875,
0.72021484375,
0.4794921875,
0.7197265625,
0.4794921875,
0.73291015625,
0.08416748046875,
0.734375
],
[
0.5185546875,
0.7236328125,
0.9140625,
0.72265625,
0.9140625,
0.736328125,
0.5185546875,
0.7373046875
],
[
0.08660888671875,
0.734375,
0.4794921875,
0.734375,
0.4794921875,
0.74755859375,
0.08660888671875,
0.74755859375
],
[
0.51953125,
0.736328125,
0.9140625,
0.736328125,
0.9140625,
0.75,
0.51953125,
0.75
],
[
0.08660888671875,
0.74755859375,
0.476318359375,
0.74755859375,
0.476318359375,
0.76123046875,
0.08660888671875,
0.76123046875
],
[
0.521484375,
0.75244140625,
0.91015625,
0.75244140625,
0.91015625,
0.76318359375,
0.521484375,
0.76318359375
],
[
0.08660888671875,
0.76318359375,
0.47802734375,
0.76318359375,
0.47802734375,
0.7734375,
0.08660888671875,
0.7734375
],
[
0.5185546875,
0.7646484375,
0.9140625,
0.76513671875,
0.9140625,
0.77880859375,
0.5185546875,
0.77783203125
],
[
0.08416748046875,
0.7744140625,
0.4248046875,
0.7744140625,
0.4248046875,
0.78857421875,
0.08416748046875,
0.78857421875
],
[
0.51953125,
0.77880859375,
0.91259765625,
0.77880859375,
0.91259765625,
0.79248046875,
0.51953125,
0.79248046875
],
[
0.101318359375,
0.78857421875,
0.476318359375,
0.7890625,
0.476318359375,
0.8037109375,
0.101318359375,
0.80322265625
],
[
0.51953125,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80615234375,
0.51953125,
0.80615234375
],
[
0.0849609375,
0.8017578125,
0.4794921875,
0.80322265625,
0.4794921875,
0.8173828125,
0.0849609375,
0.81640625
],
[
0.51953125,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.0849609375,
0.8173828125,
0.449462890625,
0.8173828125,
0.449462890625,
0.83056640625,
0.0849609375,
0.83056640625
],
[
0.51953125,
0.82080078125,
0.9140625,
0.82080078125,
0.9140625,
0.833984375,
0.51953125,
0.833984375
],
[
0.51953125,
0.833984375,
0.9140625,
0.833984375,
0.9140625,
0.84765625,
0.51953125,
0.84765625
],
[
0.1396484375,
0.84033203125,
0.41259765625,
0.84228515625,
0.41259765625,
0.85693359375,
0.1396484375,
0.85498046875
],
[
0.51953125,
0.84765625,
0.7607421875,
0.84765625,
0.7607421875,
0.861328125,
0.51953125,
0.861328125
],
[
0.45751953125,
0.85986328125,
0.4794921875,
0.85986328125,
0.4794921875,
0.87353515625,
0.45751953125,
0.87353515625
],
[
0.53759765625,
0.861328125,
0.91015625,
0.861328125,
0.91015625,
0.875,
0.53759765625,
0.875
],
[
0.261474609375,
0.86865234375,
0.423095703125,
0.86865234375,
0.423095703125,
0.88330078125,
0.261474609375,
0.88330078125
],
[
0.51953125,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.51953125,
0.8896484375
],
[
0.318603515625,
0.88525390625,
0.324462890625,
0.88525390625,
0.324462890625,
0.89013671875,
0.318603515625,
0.89013671875
]
] |
[
"query1",
"The DHE network",
"",
"R.",
"Epos",
"eddins",
"Ja",
"Sa",
"Hac",
"Geometric",
"Homography",
"Featur",
"Similarity",
"Matching",
"",
"Regression",
"Verification",
"removing",
"",
"Extractor",
"outliers",
"Spatial Scoring",
"(Number of inliers)",
"",
"candidate I",
"Searching Mutual",
"Nearest Neighbor",
"Homography",
"encoderlaye",
"Raaressiol",
"EClav",
"Figure 2: Diagram of our re-ranking process with the DHE network. The feature maps f, and f. of the query image Ig and",
"a candidate image I. are fed into the Similarity Matching Module to compute the similarity map sgc. Then the Homography",
"Regression Module uses the Sge to yield the homography matrix Hge for geometric verification of mutual matches.",
"Berton et al. 2021). Inspired by this, this module first com-",
"That is, if the most similar local feature in image I. of the",
"putes the 4 pairs of points (4-point correspondence) on the",
"feature f.(y) is f.(r), and the most similar feature in image",
"images Iq and Ic, ie., a 16D vector. This process is imple-",
"I of the feature f.(r) is f.(y), then (r, y) E MN.",
"mented by using 6 stacked transformer encoder layers fol-",
"The r and y are the indices of corresponding patches. As-",
"lowed by a fully connected layer, and a shortcut connection",
"sume that each patch corresponds to the Cartesian coordi-",
"is used between the third encoder layer and the sixth encoder",
"nates of a 2D image point in the center of the patch. We",
"layer.The transformer encoder can provide a larger effective",
"denote the homogeneous coordinates of a pair of matched",
"perceptual field on the similarity map. And it is denoted as",
"patches as Pg and pe. If the re-projection error is less than a",
"TransE(sqe + Epos) = [Pq..., Pa4, Pei..., Pe4]",
"threshold , then it is an inlier.That is, an inlier satisfies",
"IIC(pc) -C(HqcPq)l≤ 0,",
"(8)",
"P.",
"(3)",
"where C() means converting homogeneous coordinates to",
"Pa",
"=[Pq, P],",
"Cartesian(inhomogeneous)coordinates.Were-rankthecan-",
"didate images using the number of inliers between the query",
"where Pq1,..., Pq4 and pe1,..., Pea are four points on Iq",
"and candidates.",
"and Ie, respectively. [P., P| is their concatenation.",
"After obtaining 4 pairs of points on two images, we use",
"Training Strategy",
"the Direct Linear Transform (Hartley et al. 2003) to yield",
"In common deep homography works, an image can be",
"the homography Hgc, which is represented as",
"aligned with another image after using the homography",
"Hgc = DLT(Pq,P.).",
"(4)",
"transformation because the two images present a (approxi-",
"mately) same planar scene. So, they can train networks by",
"We succinctly denote the whole process of this module as",
"minimizingthedistancebetween two aligned images.But",
"Hqc = R(sqc)",
"(5)",
"this prior condition does not hold for place images in many",
"scenes. Meanwhile, we also cannot directly obtain homog-",
"The whole process of the DHE network is summarized as",
"raphy labels to supervise the training of our DHE network.",
"One possible solution is to train the network with the la-",
"Hqe = DHE(fg, f)= R(S(fg, f)",
"(6)",
"bel provided by RANSAC. However, we found that the net-",
"workis difficult to converge whether usingthehomography",
"Re-ranking with Geometric Verification",
"parameters yielded by RANSAC for direct supervision, or",
"In the re-ranking stage, we first calculate the mutual near-",
"using the selected inliers and their re-projection errors for",
"est neighbor matching of local features through exhaustive",
"supervision.We proposeto use only the number of inliers",
"cross-matching. Then the homography matrix yielded by the",
"provided by RANSAC as supervision information, let the",
"DHE network is used to check geometric verification and re-",
"networkautonomouslydecidewhichmatchedpairsarein-",
"moveoutliers.Finally.weusethenumberof inliers asthe",
"liers,and optimizethenetworkbyminimizingtheaverage",
"similarities of image pairs to re-rank the candidates.",
"re-projection error of these “inliers\". This makes it easier",
"The similarity map sqe between the query Ig and a can-",
"for networks to converge. Meanwhile, the loss produced by",
"didate Ihas been computed. We apply it to search mutual",
"this process has the potential to optimize the feature extrac-",
"nearest neighbor matches set MN, which is defined as",
"tor through back-propagation, so that the extracted feature",
"maps are more suitable for local matching and outperform",
"MN = ((r,y) : r = argmax sge(i,y),",
"the method directly using RANSAC.",
"(7)",
"We first initialize the backbone and the DHE network in-",
"y = arg max sqc(a, j)).",
"dividually. The initialization of the backbone follows the",
"7"
] |
{
"arxiv_id": "2402.16086",
"language": "en",
"page_idx": 3
}
|
||
2108.08584-en-1
|
2108.08584-en-1.jpg
|
HOI task has two benefits: (1) the SG puts each object and human in a relation graph, which can provide contextual cues and benefit the scene understanding; and (2) information contained in relationships in the SG can be explicitly or implicitly transferred to more accurately identify the correct interactions. Therefore, our motivation is to develop a decoding method that transfers the knowledge encoded in the scene graph to the HOI graph. To this end, we decode an SG from two aspects: the global scene-level and regional relation-level.
For the scene-level, our goal is to extract the scenespecific cues from the detailed scene graph, because many previous works[16,25,27] have demonstrated that the interactions are scene-biased, that is, some interactions are highly correlated to specific scenes. Taking Figure l as an example, the image is a restaurant scene from the visual appearance, and it is more likely that the interaction is about eating. Therefore, accurate scene cues can benefit HOI recognition. Many existing works[16,25] have employed the global visual appearance as the scene clues. However, due to the coarseness of the appearance feature, their performance does not gain a significant improvement. Instead, we treat a scene graph as the scene-specific contextual cue and propose two components to embed it: scene graph layout encoding and attention-based relation fusion.
For the relation-level, we observe that the relations in the scene graph can explicitly or implicitly transfer to interactions. For instance, as shown in Figure l, since the three relation triples: <hand, of, man\_1>, <hand, hold, fork>, and <man\_1, near, table> simultaneously take place, we are more likely to infer that man\_l's interaction is eat\_at. Besides, based on the relation <man\_2, near, man\_1>, we could hypothesise that both of them probably have the similar interaction. Furthermore, based on his visual features, we could infer that the man is also eating at the table. Thus, the knowledge of the exact relations between object pairs makes the inference more certain. To this end, we develop a relation-aware message passing module to reason on the SG by gathering relation information from inter- and intraclass neighbors and refine their features.
In summary, our contribution is three-fold:
$\cdot$ We propose a novel Scene Graph to Human-Object In-
teraction (SG2HOI) detection network to bridge the gap between the two tasks. To the best of our knowledge, we are the first to utilize scene graphs for HOI detection.
· We design two components to decode the SG: scene graph embedding and relation-aware message-passing, to learn the environmental context and transfer SG relations to HOI interactions, respectively.
$\cdot$ We evaluate our approach on two popular HOI detec-
tion benchmark datasets: V-COCO and HICO-DET, in
terms of a wide range of evaluation metrics. Our evaluation shows that SG2HOI method outperforms stateof-the-art models on both two datasets.
## $2. \textbf{Related Work}$
Scene Graph Generation (SGG) [30] is to detect visual relationship of all the objects pairs in an image. This task has been widely studied for many years [2, 17, 19, 29, 32]. Different from HOI, SGG tries to detect relation triples, i.e. <subject, predicate, object>, where the subject is not limited to human, and therefore the combinations of the relation triples are much more versatile than HOI and the longtail becomes the challenging problem in SGG. A couple of recent works [10, 24, 31] proposed a series of techniques to address the imbalanced distribution problem.[31] first observed that the biased relation distribution and statistically showed the frequencies of predicates are long-tailed, and even using the frequency information as the prior can obtain stunning performance. Subsequently,[24] analyzed the cause effect in the inference stage and proposed a novel debiased strategy able to be seamlessly applied in other SGG models. [10] tackled this problem by feature hallucination and knowledge transferring from the head to the tail.
Human-Object Interaction (HOI) Detection aims at detecting and localizing the interactions of human-object pairs, and requires a deeper and detailed understanding of the scene. Generally, most of the previous works [4,5,7,14,21,25] consistently consists of two steps. The first is to utilize a pre-trained object detection network, e.g., Faster-RCNN [22], to generate all the human and object proposals and construct quadratic number of human-object pairs, and the second step feeds those pairs to an interaction classification module. Due to the fact that the frst stage is based on the off-the-shelf model, numerous works focus on the second stage and dedicate to exploring more visual and contextual information so that the interaction classifier can capture the essential hidden relationship patterns between humans and objects. Specifically, InterPoint[28] proposed a fully convolutional approach to simultaneously detect the interaction points and predict the interactions, which avoided the computation of all the human-object pairs and improve the computational efficiency. VSGNet [25] concentrated on the relative spatial and structural cues by deploying two modules: spatial attention network and interactiveness graph. TIK[13] proposed an interactiveness network to suppress the non-interactive human-object pairs and improved HOI models' performance. PD-Net[33] devised a Polysemy Deciphering Network to address the diverse semantic meanings of verbs. CHGNet[27] developed a homogeneous graph network to conduct message passing between homogeneous entities and heterogeneous entities, but their passing messages are transferred in an agnostic way. To address the insufficient and indistinguishable vi-
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.53515625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.53515625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.53369140625,
0.1060791015625,
0.89208984375,
0.106689453125,
0.89208984375,
0.121826171875,
0.53369140625,
0.12060546875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.53173828125,
0.121826171875,
0.79345703125,
0.12249755859375,
0.79345703125,
0.136962890625,
0.53173828125,
0.1363525390625
],
[
0.07843017578125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.50341796875,
0.1514892578125,
0.640625,
0.1514892578125,
0.640625,
0.1654052734375,
0.50341796875,
0.1654052734375
],
[
0.07843017578125,
0.1673583984375,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.52294921875,
0.1767578125,
0.89208984375,
0.1767578125,
0.89208984375,
0.1900634765625,
0.52294921875,
0.1900634765625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.1912841796875,
0.890625,
0.1912841796875,
0.890625,
0.2052001953125,
0.5009765625,
0.2052001953125
],
[
0.07843017578125,
0.1976318359375,
0.468017578125,
0.1962890625,
0.468017578125,
0.2108154296875,
0.07843017578125,
0.212158203125
],
[
0.07843017578125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.07843017578125,
0.2266845703125
],
[
0.50341796875,
0.208984375,
0.8896484375,
0.208984375,
0.8896484375,
0.2197265625,
0.50341796875,
0.2197265625
],
[
0.50244140625,
0.222900390625,
0.890625,
0.222900390625,
0.890625,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.080078125,
0.2298583984375,
0.17236328125,
0.2298583984375,
0.17236328125,
0.240478515625,
0.080078125,
0.240478515625
],
[
0.09722900390625,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.09722900390625,
0.2568359375
],
[
0.50341796875,
0.238037109375,
0.89208984375,
0.238037109375,
0.89208984375,
0.251953125,
0.50341796875,
0.251953125
],
[
0.50244140625,
0.25244140625,
0.89208984375,
0.25244140625,
0.89208984375,
0.266357421875,
0.50244140625,
0.266357421875
],
[
0.07843017578125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.07843017578125,
0.272705078125
],
[
0.080078125,
0.2734375,
0.4697265625,
0.2734375,
0.4697265625,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50244140625,
0.267578125,
0.8896484375,
0.267578125,
0.8896484375,
0.281005859375,
0.50244140625,
0.281005859375
],
[
0.50244140625,
0.28271484375,
0.89208984375,
0.28271484375,
0.89208984375,
0.296630859375,
0.50244140625,
0.296630859375
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50341796875,
0.2998046875,
0.890625,
0.2998046875,
0.890625,
0.31005859375,
0.50341796875,
0.31005859375
],
[
0.08087158203125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.31640625,
0.08087158203125,
0.31640625
],
[
0.50244140625,
0.313232421875,
0.89208984375,
0.313232421875,
0.89208984375,
0.326904296875,
0.50244140625,
0.326904296875
],
[
0.08087158203125,
0.32080078125,
0.465576171875,
0.32080078125,
0.465576171875,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.50244140625,
0.32763671875,
0.890625,
0.32763671875,
0.890625,
0.341552734375,
0.50244140625,
0.341552734375
],
[
0.080078125,
0.333984375,
0.4697265625,
0.333984375,
0.4697265625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.080078125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.080078125,
0.3623046875
],
[
0.50244140625,
0.343505859375,
0.89208984375,
0.343505859375,
0.89208984375,
0.356689453125,
0.50244140625,
0.356689453125
],
[
0.080078125,
0.363037109375,
0.4697265625,
0.363037109375,
0.4697265625,
0.376953125,
0.080078125,
0.376953125
],
[
0.50341796875,
0.35986328125,
0.890625,
0.35986328125,
0.890625,
0.37060546875,
0.50341796875,
0.37060546875
],
[
0.50244140625,
0.373779296875,
0.89208984375,
0.373779296875,
0.89208984375,
0.386962890625,
0.50244140625,
0.386962890625
],
[
0.080078125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.5009765625,
0.38818359375,
0.890625,
0.38818359375,
0.890625,
0.401611328125,
0.5009765625,
0.401611328125
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.404052734375,
0.89208984375,
0.404052734375,
0.89208984375,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.07843017578125,
0.4248046875,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.07843017578125,
0.438232421875
],
[
0.50341796875,
0.42041015625,
0.890625,
0.42041015625,
0.890625,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.080078125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.50244140625,
0.434326171875,
0.84716796875,
0.434326171875,
0.84716796875,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.07843017578125,
0.455078125,
0.378173828125,
0.455078125,
0.378173828125,
0.468505859375,
0.07843017578125,
0.468505859375
],
[
0.52392578125,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.4609375,
0.52392578125,
0.4609375
],
[
0.098876953125,
0.4697265625,
0.468017578125,
0.4697265625,
0.468017578125,
0.48291015625,
0.098876953125,
0.48291015625
],
[
0.50244140625,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50244140625,
0.477783203125
],
[
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
],
[
0.50341796875,
0.49560546875,
0.61669921875,
0.49560546875,
0.61669921875,
0.505859375,
0.50341796875,
0.505859375
],
[
0.609375,
0.494384765625,
0.89208984375,
0.494384765625,
0.89208984375,
0.50830078125,
0.609375,
0.50830078125
],
[
0.07843017578125,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.07843017578125,
0.5146484375
],
[
0.50341796875,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50341796875,
0.52294921875
],
[
0.080078125,
0.51513671875,
0.4697265625,
0.51513671875,
0.4697265625,
0.52880859375,
0.080078125,
0.52880859375
],
[
0.5009765625,
0.5234375,
0.89208984375,
0.5244140625,
0.89208984375,
0.53857421875,
0.5009765625,
0.537109375
],
[
0.07843017578125,
0.52978515625,
0.4697265625,
0.52978515625,
0.4697265625,
0.54345703125,
0.07843017578125,
0.54345703125
],
[
0.50244140625,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.08087158203125,
0.546875,
0.46728515625,
0.546875,
0.46728515625,
0.556640625,
0.08087158203125,
0.556640625
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.08087158203125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57275390625,
0.08087158203125,
0.57275390625
],
[
0.08087158203125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.50341796875,
0.5712890625,
0.890625,
0.5712890625,
0.890625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.080078125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50244140625,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.080078125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50341796875,
0.6015625,
0.890625,
0.6015625,
0.890625,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.5009765625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.080078125,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.07843017578125,
0.63525390625,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.07843017578125,
0.6484375
],
[
0.50341796875,
0.6318359375,
0.890625,
0.6318359375,
0.890625,
0.642578125,
0.50341796875,
0.642578125
],
[
0.5009765625,
0.64453125,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.5009765625,
0.65771484375
],
[
0.080078125,
0.6513671875,
0.4697265625,
0.6513671875,
0.4697265625,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.080078125,
0.6796875
],
[
0.080078125,
0.68115234375,
0.3447265625,
0.68115234375,
0.3447265625,
0.6943359375,
0.080078125,
0.6943359375
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.09722900390625,
0.69580078125,
0.381591796875,
0.6943359375,
0.381591796875,
0.708984375,
0.09722900390625,
0.71044921875
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.5009765625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.098876953125,
0.71826171875,
0.46728515625,
0.71826171875,
0.46728515625,
0.73193359375,
0.098876953125,
0.73193359375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50244140625,
0.734375
],
[
0.111083984375,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.111083984375,
0.74755859375
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.111083984375,
0.74853515625,
0.468017578125,
0.74755859375,
0.468017578125,
0.76220703125,
0.111083984375,
0.76318359375
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.11273193359375,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.11273193359375,
0.77783203125
],
[
0.5009765625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.11273193359375,
0.77978515625,
0.1805419921875,
0.77978515625,
0.1805419921875,
0.79150390625,
0.11273193359375,
0.79150390625
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.09722900390625,
0.80029296875,
0.468017578125,
0.8017578125,
0.468017578125,
0.81640625,
0.09722900390625,
0.81494140625
],
[
0.50341796875,
0.79736328125,
0.8896484375,
0.79736328125,
0.8896484375,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.111083984375,
0.818359375,
0.468017578125,
0.818359375,
0.468017578125,
0.83203125,
0.111083984375,
0.83203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.111083984375,
0.83349609375,
0.4697265625,
0.83349609375,
0.4697265625,
0.8466796875,
0.111083984375,
0.8466796875
],
[
0.1094970703125,
0.8466796875,
0.37744140625,
0.84765625,
0.37744140625,
0.8623046875,
0.1094970703125,
0.861328125
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09722900390625,
0.86865234375,
0.471435546875,
0.869140625,
0.471435546875,
0.88720703125,
0.09722900390625,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.111083984375,
0.88623046875,
0.471435546875,
0.88623046875,
0.471435546875,
0.90087890625,
0.111083984375,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"HOI task has two benefits: (1) the SG puts each object and",
"terms of a wide range of evaluation metrics. Our eval-",
"human in a relation graph, which can provide contextual",
"uation shows that SG2HOI method outperforms state-",
"cues and benefitthe sceneunderstanding:and (2)informa-",
"of-the-art models on both two datasets.",
"tion contained in relationships in the SG can be explicitly",
"orimplicitlytransferredtomoreaccuratelyidentifythecor-",
"2.Related Work",
"rect interactions.Therefore,our motivation is to develop a",
"Scene Graph Generation (SGG) [3OJ is to detect visual",
"decoding method that transfers the knowledge encoded in",
"relationship of all the objects pairs in an image. This task",
"the scene graph to the HOI graph. To this end, we decode",
"an SG from two aspects: the global scene-level and regional",
"hasbeen widely studied formanyyears[2,17,19,29,32].",
"Different from HOI, SGG tries to detect relation triples, i.e.",
"relation-level.",
"For the scene-level, our goal is to extract the scene-",
"<subject, predicate, object>, where the subject is not lim-",
"ited to human, and therefore the combinations of the rela-",
"specific cues from the detailed scene graph, because many",
"previous works [16, 25, 27] have demonstrated that the in-",
"tion triples are much more versatile than HOI and the long-",
"tail becomes the challenging problem in SGG. A couple of",
"teractions are scene-biased, that is, some interactions are",
"recentworks[10,24.31lproposed aseriesoftechniquesto",
"highlycorrelated tospecific scenes.TakingFigure Ias an",
"address the imbalanced distribution problem. [31] first ob-",
"example.theimageisarestaurant scenefromthevisual ap-",
"served that the biased relation distribution and statistically",
"pearance, and it is more likely that the interaction is about",
"eat ing. Therefore, accurate scene cues can benefit HOI",
"showed the frequencies of predicates are long-tailed, and",
"recognition.Many existing works [16,25J have employed",
"evenusingthefrequencyinformation asthepriorcanob-",
"tain stunning performance. Subsequently, [24] analyzed the",
"the global visual appearance as the scene clues. However,",
"cause effect in the inference stage and proposed a novel de-",
"due to the coarseness of the appearance feature, their perfor-",
"biased strategy able to be seamlessly applied in other SGG",
"mance does not gain a significant improvement. Instead, we",
"treat a scene graph as the scene-specific contextual cue and",
"models.[lotackledthis problembyfeaturehallucination",
"propose two components to embed it: scene graph layout",
"and knowledge transferring from the head to the tail.",
"encoding and attention-based relation fusion.",
"Human-ObjectInteraction (HOD)Detection aims at",
"For the relation-level, we observe that the relations in the",
"detecting and localizing the interactions of human-object",
"scene graph can explicitly or implicitly transfer to interac-",
"pairs, and requires a deeper and detailed understanding",
"of the scene..",
"Generally, most of the previous works",
"tions. For instance, as shown in Figure 1, since the three",
"[4, 5, 7, 14, 21, 25] consistently consists of two steps. The",
"relation triples: <hand, of, man_1>, <hand, hold, fork>,",
"first is to utilize a pre-trained object detection network, e.g.",
"and <man_1, near, table> simultaneously take place, we",
"Faster-RCNN [22], to generate all the human and object",
"aremorelikelytoinferthatman_l'sinteractioniseat_at.",
"proposals and construct quadratic number of human-object",
"Besides,based on therelation <man_2,near,man_1 >,we",
"couldhypothesisethatbothofthemprobablyhavethesim-",
"pairs,and thesecond stepfeeds thosepairsto an interaction",
"ilar interaction. Furthermore, based on his visual features,",
"classification module. Due to the fact that the first stage",
"we could infer that the man is also eating at the table. Thus,",
"is based on the off-the-shelfmodel,numerous worksfocus",
"on the second stage and dedicate to exploring more visual",
"theknowledgeoftheexactrelationsbetween objectpairs",
"makes the inference more certain. To this end, we develop",
"and contextual information so that the interaction classi-",
"fier can capture the essential hidden relationship patterns",
"a relation-aware message passing module to reason on the",
"between humans and objects. Specifically, InterPoint [28]",
"SG by gathering relation information from inter- and intra-",
"class neighbors and refine their features.",
"proposed a fully convolutional approach to simultaneously",
"In summary, our contribution is three-fold:",
"detect the interaction points and predict the interactions,",
"which avoided the computation of all the human-object",
".We propose a novel Scene Graph to Human-Object In-",
"pairs and improve the computational efficiency. VSGNet",
"teraction (SG2HOI) detection network to bridge the",
"[25concentrated ontherelativespatial and structural cues",
"gap between the two tasks. To the best of our knowl-",
"by deploying two modules: spatial attention network and",
"edge, we are the first to utilize scene graphs for HOI",
"interactiveness graph. TIK [13] proposed an interactiveness",
"detection.",
"networkto suppressthe non-interactivehuman-objectpairs",
". We design two components to decode the SG: scene",
"and improvedHOImodelsperformance.PD-Net[331de-",
"vised a Polysemy Deciphering Network to address the di-",
"graph embedding and relation-aware message-passing,",
"verse semantic meanings of verbs. CHGNet [27] developed",
"to learn the environmental context and transfer SG re-",
"lations to HOI interactions, respectively.",
"a homogeneous graph network to conduct message passing",
"between homogeneous entities and heterogeneous entities,",
"• We evaluate our approach on two popular HOI detec-",
"but their passing messages are transferred in an agnostic",
"tion benchmark datasets: V-COCO and HICO-DET, in",
"way. To address the insufficient and indistinguishable vi-"
] |
{
"arxiv_id": "2108.08584",
"language": "en",
"page_idx": 1
}
|
||
2108.08584-en-4
|
2108.08584-en-4.jpg
|
reasoning module where we can aggregate relation information from neighbors and then learn the intrinsic correlation between relationships and interactions. Toward this end, we devise a relation-aware message passing strategy able to aggregate and propagate relation signals on the scene graph and refine the visual feature of humans and objects.
Although previous works [21,27] have proposed message passing mechanisms, the messages in these methods are homogeneous, i.e., they are not aware of the specific relations present in an image. This causes the gathered information to be ambiguous and unrepresentative. To address this issue, we propose to use the explicit relations in the scene graph as messages to refine the human and object visual features. We define two types of messages for feature refinement: inter- and intra-class messages, to capture the different roles humans and objects play in the interaction. Inter-class refinement. For the inter-class refinement, human nodes only aggregate the messages from object (i.e. non-human) nodes, aimed at updating humans' feature by receiving relation messages from object neighbors. For example, in Figure l, by incorporating the relation message of from hand, the model could reason that man\_1 is holding a fork.
Given the visual feature of human f$_h^i$ and object f$_o^j$,we refine the human feature by aggregating the relationship messages from its object neighbors. The inter-messages are calculated by:
$$\mathring{\mathbf{f}}_h^i=\mathbf{W}_{o\to h}\left(\sum_{j\in\mathcal{N}^i}\mathbf{W}^o\cdot\mathbf{f}_o^j\otimes\alpha_{ji}^{o\to h}\right)$$
(7)
where $\mathcal{N}^i$ denotes the neighbors of human $i$ in the scene graph $\mathcal{S};\alpha_{ji}^{o\to h}$ is the word embedding of relation r$_ji$ from object to human as described in Eq.(3); and $\dot{\mathbf{W}}^o$ and $\mathbf{W}_{o\to h}$ are two transformation layers.
Intra-class refinement. On the other hand, a human could interact with several other humans, and their relationships could provide essential clues. Different from the inter-class refinement strategy, the intra-class messages are calculated from human neighbors only:
$$\overline{\mathbf{f}}_h^i=\mathbf{W}_{h\to h}\left(\sum_{j\in\mathcal{N}^i}\mathbf{W}^h\cdot\mathbf{f}_h^j\otimes\alpha_{ji}^{h\to h}\right)$$
(8)
Then, the refined human feature is formulated as:
$$\tilde{\mathbf{f}}_h^i=\mathbf{f}_h^i+\overline{\mathbf{f}}_h^i+\mathring{\mathbf{f}}_h^i$$
(9)
For iterative refinement, we only let f$_h^i$ be updated as $\tilde{\mathbf{f}}_h^i$
and repeat the calculation following Eq. (7), (8), and (9).
Note that the object feature refinement also has inter- and intra- parts and is similar to the human refinement, and we omit their details for brevity reasons.
## 3.3. HOI Predication
After embedding the scene graph and refining visual features, we have the global scene graph representation $\tilde{\mathbf{g}}_k$ for the image $k$ and the refined human feature $\tilde{\mathbf{f}}_h^i$ and object feature $\tilde{\tilde{\tilde{f}}}_o^i.$ In addition, we also extract the spatial mask of each human-object pair with size $64\times64$ as the auxiliary feature $\mathbf{f}_s^{ij}$, following [25,27]. The final prediction of HOI is combined from two branches: visual appearance features and refined features by message passing. For the human $i$ and object $j$, the visual branch is predicted by:
$$\mathbf{p}_v=\delta(\mathbf{W}_v[\mathbf{f}_s^{ij}\otimes[\mathbf{f}_h^i;\mathbf{f}_o^j]])$$
(10)
where $\delta$ is the sigmoid active function and W$_v$ is the classi-
fier parameters for visual features.
The message passing branch is calculated as below:
$$\mathbf{p}_{m}=\delta(\mathbf{W}_{m}[\tilde{\mathbf{g}}_{k};\tilde{\mathbf{f}}_{h}^{i};\tilde{\mathbf{f}}_{o}^{j}])$$
(11)
where $\mathbf{W}_m$ is the classifer parameters for refined features.
Finally, we combine the two predicted scores as:
$$\mathbf{p}^{ij}=\lambda^{ij}\cdot\mathbf{p}_{v}\cdot\mathbf{p}_{m}$$
(12)
where $\lambda^{ij}$ is the multiplication of the detected human and
object score normalized by [13].
In the training stage, we optimize the Binary Cross-
Entropy (BCE) loss on $\mathbf{p}^{ij}$ to optimize the HOI model.
## $4. \textbf{Experiments}$
In this section, we first briefly describe the datasets, evaluation metrics and baseline models, as well as the implementation details. Next, we compare our model with the state-of-art methods and further conduct a series of ablation studies. Last, we show several qualitative results on the both tasks: SGG and HOI.
### 4.1. Datasets and Metrics
Datasets: We evaluate our model's performance on two standard benchmarks: the V-COCO[6] and HICO-DET[1] datasets. The training, validation and test sets of V-COCO consist of 2,533,2,867 and 4,946 images, respectively. Following previous works[25,28], we also use the training and validation set, containing 5,400 images in total, to train our model. Each human-object pair in V-COCO is labeled with a 29-D one-hot vector. Ît is worth noting that among the 29 actions, three (cut, hit, eat) have no interaction objects. HICO-Det is a large dataset and is split into two sets: 38,118 training and 9,658 testing images. In total, HICO-Det is annotated with 600 human-object interaction classes, 80 object classes and 117 actions, including a no-interaction class. Following previous works[1,28], we categorize the interactions into three groups: Full, Rare
|
[
[
0.080078125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.6650390625,
0.091552734375,
0.6650390625,
0.1060791015625,
0.5009765625,
0.10479736328125
],
[
0.080078125,
0.106689453125,
0.4697265625,
0.106689453125,
0.4697265625,
0.12060546875,
0.080078125,
0.12060546875
],
[
0.521484375,
0.11553955078125,
0.890625,
0.11553955078125,
0.890625,
0.1287841796875,
0.521484375,
0.1287841796875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.07843017578125,
0.1363525390625,
0.46728515625,
0.1383056640625,
0.46728515625,
0.15283203125,
0.07843017578125,
0.15087890625
],
[
0.50244140625,
0.1307373046875,
0.890625,
0.1307373046875,
0.890625,
0.14453125,
0.50244140625,
0.14453125
],
[
0.5009765625,
0.143310546875,
0.890625,
0.1458740234375,
0.890625,
0.160400390625,
0.5009765625,
0.1578369140625
],
[
0.080078125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.080078125,
0.1673583984375,
0.416748046875,
0.1673583984375,
0.416748046875,
0.18115234375,
0.080078125,
0.18115234375
],
[
0.50341796875,
0.1622314453125,
0.89208984375,
0.1622314453125,
0.89208984375,
0.175537109375,
0.50341796875,
0.175537109375
],
[
0.098876953125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.098876953125,
0.1962890625
],
[
0.50341796875,
0.1787109375,
0.8896484375,
0.1787109375,
0.8896484375,
0.189453125,
0.50341796875,
0.189453125
],
[
0.50244140625,
0.1925048828125,
0.890625,
0.1925048828125,
0.890625,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.0767822265625,
0.1988525390625,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.0767822265625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2083740234375,
0.89208984375,
0.2083740234375,
0.89208984375,
0.2215576171875,
0.50244140625,
0.2215576171875
],
[
0.080078125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.222900390625,
0.8935546875,
0.222900390625,
0.8935546875,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.50244140625,
0.2374267578125,
0.8095703125,
0.2374267578125,
0.8095703125,
0.251953125,
0.50244140625,
0.251953125
],
[
0.080078125,
0.24365234375,
0.4697265625,
0.24365234375,
0.4697265625,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.60546875,
0.268310546875,
0.7861328125,
0.268310546875,
0.7861328125,
0.28271484375,
0.60546875,
0.28271484375
],
[
0.859375,
0.267578125,
0.8935546875,
0.267578125,
0.8935546875,
0.28271484375,
0.859375,
0.28271484375
],
[
0.080078125,
0.2890625,
0.4697265625,
0.2890625,
0.4697265625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.5048828125,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.304931640625,
0.5048828125,
0.304931640625
],
[
0.07843017578125,
0.302978515625,
0.4697265625,
0.3037109375,
0.4697265625,
0.318115234375,
0.07843017578125,
0.317626953125
],
[
0.50341796875,
0.306884765625,
0.73046875,
0.306884765625,
0.73046875,
0.32080078125,
0.50341796875,
0.32080078125
],
[
0.08087158203125,
0.32080078125,
0.455078125,
0.32080078125,
0.455078125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.52294921875,
0.3212890625,
0.86279296875,
0.3212890625,
0.86279296875,
0.3359375,
0.52294921875,
0.3359375
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.080078125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.61279296875,
0.349853515625,
0.779296875,
0.349853515625,
0.779296875,
0.3642578125,
0.61279296875,
0.3642578125
],
[
0.859375,
0.349853515625,
0.8935546875,
0.349853515625,
0.8935546875,
0.365478515625,
0.859375,
0.365478515625
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.08087158203125,
0.380615234375,
0.46728515625,
0.380615234375,
0.46728515625,
0.391357421875,
0.08087158203125,
0.391357421875
],
[
0.50341796875,
0.37890625,
0.89208984375,
0.37890625,
0.89208984375,
0.392578125,
0.50341796875,
0.392578125
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.39453125,
0.82177734375,
0.39453125,
0.82177734375,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.62744140625,
0.417236328125,
0.7646484375,
0.42041015625,
0.7646484375,
0.438232421875,
0.62744140625,
0.43505859375
],
[
0.08087158203125,
0.425537109375,
0.160888671875,
0.425537109375,
0.160888671875,
0.4375,
0.08087158203125,
0.4375
],
[
0.861328125,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.43505859375,
0.861328125,
0.43505859375
],
[
0.09722900390625,
0.439453125,
0.468017578125,
0.439453125,
0.468017578125,
0.45263671875,
0.09722900390625,
0.45263671875
],
[
0.080078125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.50341796875,
0.449462890625,
0.89208984375,
0.449462890625,
0.89208984375,
0.46337890625,
0.50341796875,
0.46337890625
],
[
0.080078125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.080078125,
0.484130859375
],
[
0.50244140625,
0.466552734375,
0.7158203125,
0.466552734375,
0.7158203125,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.08087158203125,
0.486083984375,
0.1748046875,
0.486083984375,
0.1748046875,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.521484375,
0.48095703125,
0.89208984375,
0.48095703125,
0.89208984375,
0.494384765625,
0.521484375,
0.494384765625
],
[
0.5009765625,
0.494384765625,
0.86376953125,
0.49560546875,
0.86376953125,
0.51025390625,
0.5009765625,
0.5087890625
],
[
0.1422119140625,
0.5234375,
0.2296142578125,
0.5234375,
0.2296142578125,
0.54150390625,
0.1422119140625,
0.54150390625
],
[
0.265625,
0.521484375,
0.369384765625,
0.5244140625,
0.36767578125,
0.54248046875,
0.2646484375,
0.5390625
],
[
0.44677734375,
0.5263671875,
0.4697265625,
0.5263671875,
0.4697265625,
0.54052734375,
0.44677734375,
0.54052734375
],
[
0.50341796875,
0.52587890625,
0.63134765625,
0.52587890625,
0.63134765625,
0.54052734375,
0.50341796875,
0.54052734375
],
[
0.238525390625,
0.54345703125,
0.273681640625,
0.54345703125,
0.273681640625,
0.55615234375,
0.238525390625,
0.55615234375
],
[
0.373291015625,
0.52783203125,
0.384033203125,
0.52783203125,
0.384033203125,
0.53076171875,
0.373291015625,
0.53076171875
],
[
0.52294921875,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.52294921875,
0.564453125
],
[
0.080078125,
0.56640625,
0.468017578125,
0.5673828125,
0.468017578125,
0.58203125,
0.080078125,
0.58056640625
],
[
0.50244140625,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.580078125,
0.50244140625,
0.580078125
],
[
0.080078125,
0.58349609375,
0.4697265625,
0.58349609375,
0.4697265625,
0.59765625,
0.080078125,
0.59765625
],
[
0.5009765625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.5009765625,
0.5947265625
],
[
0.08087158203125,
0.5986328125,
0.471435546875,
0.5986328125,
0.471435546875,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.50341796875,
0.59765625,
0.890625,
0.59765625,
0.890625,
0.60791015625,
0.50341796875,
0.60791015625
],
[
0.08087158203125,
0.6142578125,
0.330078125,
0.6142578125,
0.330078125,
0.62890625,
0.08087158203125,
0.62890625
],
[
0.50244140625,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.08087158203125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.08087158203125,
0.642578125
],
[
0.50244140625,
0.626953125,
0.6435546875,
0.626953125,
0.6435546875,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50341796875,
0.6513671875,
0.69921875,
0.6513671875,
0.69921875,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.08087158203125,
0.6611328125,
0.468017578125,
0.6611328125,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
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
],
[
0.080078125,
0.689453125,
0.268798828125,
0.689453125,
0.268798828125,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.26708984375,
0.724609375,
0.366943359375,
0.72900390625,
0.365234375,
0.74560546875,
0.265625,
0.7412109375
],
[
0.50244140625,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50244140625,
0.734375
],
[
0.44677734375,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.44677734375,
0.744140625
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.136474609375,
0.72607421875,
0.233642578125,
0.72802734375,
0.233642578125,
0.7470703125,
0.136474609375,
0.744140625
],
[
0.24267578125,
0.75,
0.273681640625,
0.74560546875,
0.27685546875,
0.75830078125,
0.2459716796875,
0.76220703125
],
[
0.50341796875,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.08087158203125,
0.77099609375,
0.40283203125,
0.77099609375,
0.40283203125,
0.7841796875,
0.08087158203125,
0.7841796875
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.50244140625,
0.78076171875,
0.8935546875,
0.78076171875,
0.8935546875,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.2115478515625,
0.796875,
0.334228515625,
0.796875,
0.334228515625,
0.814453125,
0.2115478515625,
0.814453125
],
[
0.44677734375,
0.7998046875,
0.4697265625,
0.7998046875,
0.4697265625,
0.81298828125,
0.44677734375,
0.81298828125
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.50244140625,
0.8095703125,
0.89208984375,
0.81005859375,
0.89208984375,
0.82470703125,
0.50244140625,
0.82373046875
],
[
0.098876953125,
0.82568359375,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.098876953125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.450927734375,
0.84130859375,
0.450927734375,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85302734375,
0.5048828125,
0.85302734375
],
[
0.098876953125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.86865234375,
0.098876953125,
0.86865234375
],
[
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
],
[
0.07843017578125,
0.88623046875,
0.325927734375,
0.88720703125,
0.325927734375,
0.90087890625,
0.07843017578125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"reasoning module where we can aggregate relation informa-",
"3.3.HOIPredication",
"tion from neighbors and then learn the intrinsic correlation",
"After embedding the scene graph and refining visual fea-",
"between relationships and interactions. Toward this end, we",
"devise a relation-aware message passing strategy able to ag-",
"tures, we have the global scene graph representation gk for",
"the image k and the refined human feature fi and object",
"gregate and propagate relation signals on the scene graph",
"and refine the visual feature of humans and objects.",
"feature fi. In addition, we also extract the spatial mask of",
"Although previous works [21, 27] have proposed mes-",
"eachhuman-objectpairwithsize64x64 astheauxiliary",
"feature fij, following [25, 27]. The final prediction of HOI",
"sagepassing mechanisms, the messages in these methods",
"are homogeneous, i.e., they are not aware of the specific re-",
"is combined from two branches: visual appearance features",
"lations present in an image. This causes the gathered infor-",
"and refined features by message passing. For the human i",
"and object j, the visual branch is predicted by:",
"mation to be ambiguous and unrepresentative. To address",
"this issue, we propose to use the explicit relations in the",
"Pu = s(W,[fii [fi;fj]l)",
"(10)",
"sual features. We define two types of messages for feature",
"scene graph as messages to refine the human and object vi-",
"where is the sigmoid active function and W, is the classi-",
"refinement: inter- and intra-class messages, to capture the",
"fier parameters for visual features.",
"differentroleshumansandobjectsplayintheinteraction.",
"The message passing branch is calculated as below:",
"Inter-class refinement. For the inter-class refinement, hu-",
"man nodes only aggregate the messages from object (i.e.",
"Pm = o(Wm[gk;fi;fil)",
"(11)",
"non-human) nodes, aimed at updating humans' feature by",
"receivingrelationmessagesfromobjectneighbors.Forex-",
"where Wm is the classifier parameters for refined features.",
"ample, in Figure 1, by incorporating the relation message",
"Finally, we combine the two predicted scores as:",
"of from hand, the model could reason that man.i is hold-",
"p=.puPm",
"inga fork.",
"(12)",
"Given the visual feature of human fj and object fi, we",
"refine the human feature by aggregating the relationship",
"where Jzj is the multiplication of the detected human and",
"messages from its object neighbors. The inter-messages are",
"object score normalized by [13].",
"calculated by:",
"In the training stage, we optimize the Binary Cross-",
"Entropy (BCE) loss on pj to optimize the HOI model.",
"fh= Woh",
"wo.fi@ag",
"(7)",
"4.Experiments",
"JEN",
"",
"In this section, we first briefly describe the datasets, eval-",
"where Ni denotes the neighbors of human i in the scene",
"uation metrics and baseline models, as well as the imple-",
"graph S; α-h is the word embedding of relation rji from",
"mentation details. Next, we compare our model with the",
"object to human as described in Eq. (3); and Wo and",
"state-of-artmethodsandfurtherconducta seriesofablation",
"Wo→h are two transformation layers.",
"studies. Last, we show several qualitative results on the both",
"Intra-class refinement. On the other hand, a human could",
"tasks: SGG and HOI.",
"interact with several other humans, and their relationships",
"4.1. Datasets and Metrics",
"couldprovideessential clues.Differentfromtheinter-class",
"refinement strategy, the intra-class messages are calculated",
"Datasets: We evaluate our model's performance on two",
"from human neighbors only:",
"standard benchmarks: the V-COCO [6] and HICO-DET []]",
"datasets. The training, validation and test sets of V-COCO",
"wh.fi@a",
"consist of 2,533, 2, 867 and 4, 946 images, respectively.",
"(8)",
"Following previous works [25, 28], we also use the train-",
"fi = Wh-h",
"jEN",
"ing and validation set, containing 5, 400 images in total, to",
"Then, the refined human feature is formulated as:",
"train our model. Each human-object pair in V-COCO is la-",
"beled with a 29-D one-hot vector. It is worth noting that",
"fh=fh+fh+fi",
"(9)",
"among the 29 actions, three (cut, hit, eat) have no interac-",
"tion objects. HICO-Det is a large dataset and is split into",
"For iterative refinement, we only let fi, be updated as fi",
"two sets: 38, 118 training and 9, 658 testing images. In to-",
"and repeat the calculation following Eq. (7), (8), and (9).",
"tal.HICO-Det is annotatedwith 600human-object interac-",
"Note that the objectfeaturerefinement also has inter-and",
"tionclasses,80objectclasses and117actions,including",
"intra- parts and is similar to the human refinement, and we",
"a no-interaction class. Following previous works [1, 28],",
"omittheirdetailsforbrevityreasons.",
"we categorize the interactions into three groups: Full, Rare"
] |
{
"arxiv_id": "2108.08584",
"language": "en",
"page_idx": 4
}
|
||
2202.01361-en-1
|
2202.01361-en-1.jpg
|
Generative Flow Networks for Discrete Probabilistic Modeling
<FigureHere>
Figure 2. The state space $S$ and the GFlowNet's forward modeling process in a 9-dimensional discrete data space. The states are the vertices of a DAG whose edges are the transitions- actions of painting a grey pixel into black (1) or white (0).
modes without having to travel long low-probability paths in between (Fig. 1). Instead of having to exploit a priori structure to design long jumps in the MCMC, GFlowNets can discover that structure when it is not explicitly known.
This paper makes the following key contributions:
(1) We cast a non-autoregressive sequential generation
model for high-dimensional discrete data as a gener-
ative flow network (§3.1).
(2) We introduce a GFlowNet-based MCMC proposal en-
abling efficient large jumps with low probability of rejection, taking advantage of the compositional structure learned by the GFlowNet generative policy (§3.3).
(3) We describe a procedure based on such proposals for
jointly training the GFlowNet sampler with an energy model given a dataset and state conditions under which this estimates the true data log-likelihood gradient with respect to the energy function's parameters (§3.2, §3.3).
(4) We test the algorithm on a variety of synthetic and real
tasks, achieving competitive results (§4).
## $2. \textbf{ Preliminaries}$ $2. 1. \textbf{GFlowNets}$
Generative flow networks, or GFlowNets for short (Bengio et al., 2021a;b), are trainable generative policies on which this paper is built. They model the generation process of objects $\mathbf{x}\in\mathcal{X}$ by a sequence of discrete actions that incrementally modify a partially constructed object $(state).$ The space of possible action sequences is represented by $\hat{\text{a directed acyclic graph}}\left(\mathrm{DAG,~see~Fig.~2)~}\hat{G}=\left(\mathcal{S},\mathcal{A}\right),\:=\:(\mathcal{S},\mathcal{A}),\:=\:\hat{\text{directed acyclic graph}}\left(\mathrm{DAG,~see~Fig.~2}\right)~\hat{G}=\left(\mathcal{S},\mathcal{A}\right),\:=\:\hat{\text{directed acyclic graph}}\left(\mathrm{DAG,~seee~Fig.~2}\right)~\hat{G}=\left(\mathcal{S},\mathcal{A}\right),\:=\:\hat{\text{directed acyclic gra$ where the vertices in $\mathcal{S}$ are states and the edges in $\mathcal{A}$ are actions that modify one state to another. We use the terms parents and $children$, respectively, for the tails of incoming edges and the heads of outgoing edges of a state. Generation begins at a special initial state s$_0$ and terminates upon a transition to any terminal state, which is a state with no outgoing actions. The set of terminal states is identified with the target space $\mathcal{X}.$ Note that multiple possible action sequences may lead to the same terminal state.
A complete trajectory is a sequence of states $\mathbf{s}_{0}\to\mathbf{s}_{1}\to\ldots\to\mathbf{s}_{n}$,where each transition $\mathbf{s} _{t}\to \mathbf{s} _{t+ 1}$ is an action in $\mathcal{A}$ and s$_n\in\mathcal{X}$ is terminal. A trajectory flow is a measure (unnormalized density) on the set of all complete trajectories $\mathcal{T},i.e.$, a non-negative function $F:{\mathcal{T}}\to{\mathbb{R}}_{\geq0}$,which can be thought of as a number of particles flowing from s$_0$ to terminal states along each route. The flow is called Markovian if there exist distributions $P_F(\cdot|\mathbf{s})$ over the children of every nonterminal state s, and a constant $Z$, such that for any complete trajectory $\tau=(\mathbf{s}_0\to\mathbf{s}_1\to\ldots\to\ldots\to\mathbf{s}_n)$, we have $P_F(\tau)=F(\tau)/Z$ with
(1)
$$P_{F}(\tau)=P_{F}(\mathbf{s}_{1}|\mathbf{s}_{0})P_{F}(\mathbf{s}_{2}|\mathbf{s}_{1})\ldots P_{F}(\mathbf{s}_{n}|\mathbf{s}_{n-1}).$$
In this case, $P_F(\mathbf{s}_{t+1}|\mathbf{s}_{t})$ is called a forward policy, and can be used to sample complete trajectories from the density $F$ and thus also their terminal states, objects $\mathbf{x}\in\mathcal{X}.^1$ We write $P_T(\mathbf{x})$ for the probability that a trajectory sampled from $P_F$ terminates in x.
Past work on GFlowNets (Bengio et al., 2021a;b; Malkin et al., 2022; Jain et al., 2022) has considered the problem of fitting a Markovian flow to a fixed reward function on $\mathcal{X}.$ Given a non-negative reward function $R:\mathcal{X}\to\mathbb{R}_\geq0$, one seeks a Markovian flow $F$ such that the likelihood of a trajectory sampled from $F$ terminating in a given x $\in$ $\mathcal{X}$ is proportional to $R(\mathbf{x}),i.e.,P_T(\mathbf{x})\propto R(\mathbf{x}).$ This is accomplished by the reward matching condition:
(2)
$$R(\mathbf{x})=\sum_{\tau=(\mathbf{s}_0\to\ldots\to\mathbf{s}_n),\mathbf{s}_n=\mathbf{x}}F(\tau).$$
By Eq. (1), one can specify a Markovian flow by a learned scalar $\log Z_{\theta}$ and a neural net with parameters $\theta$ that outputs $P_{F}(\cdot|\mathbf{s};\boldsymbol{\theta})$ for any input state s. Algorithms for learning $P_F$ from maximum-entropy reinforcement learning (Haarnoja et al., 2017) would solve this problem if there was only one way to construct an object x (the DAG is a tree) while GFlowNets are applicable in the more general DAG setting (Bengio et al., 2021a). The solution originally proposed by Bengio et al. (2021a) recasts an unnormalized forward policy as a network flow on $G$ (in the classical sense of Ford & Fulkerson (1956)) and performs gradient descent on the error in a flow conservation constraint at states sampled from a training policy, amounting to a generalization of temporal difference objectives from Sutton (2005).
Alternative objectives proposed by Bengio et al. (2021b); Malkin et al. (2022) require a model to produce 3 outputs: the scalar $\log Z_{\boldsymbol{\theta}}$, a forward policy $P_F(\cdot|\cdot;\boldsymbol{\theta})$, and a back- $ward\textit{policy that produces distributions }P_{B}( \cdot | \mathbf{s} ; \boldsymbol{\theta })$ over the parents of an input state s. (When action sequences are sampled in reverse from $P_B$, we will use dashed arrows s'---s
It is helpful to think of the particle analogy. For a Markovian flow, the distribution over complete trajectories $(P_F(\tau)\propto F(\tau))$ satisfes a history-independence property: a particle's choice of route after reaching a state s is independent of how it reached s.
|
[
[
0.291748046875,
0.05682373046875,
0.68310546875,
0.05682373046875,
0.68310546875,
0.0706787109375,
0.291748046875,
0.0706787109375
],
[
0.41015625,
0.08587646484375,
0.42724609375,
0.08587646484375,
0.42724609375,
0.0947265625,
0.41015625,
0.0947265625
],
[
0.5,
0.08587646484375,
0.88623046875,
0.08587646484375,
0.88623046875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.5009765625,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.5009765625,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.5009765625,
0.1300048828125
],
[
0.5,
0.1307373046875,
0.88623046875,
0.1307373046875,
0.88623046875,
0.14453125,
0.5,
0.14453125
],
[
0.40283203125,
0.1458740234375,
0.43310546875,
0.143310546875,
0.4345703125,
0.1546630859375,
0.404296875,
0.1572265625
],
[
0.5009765625,
0.146484375,
0.88623046875,
0.146484375,
0.88623046875,
0.160400390625,
0.5009765625,
0.160400390625
],
[
0.50244140625,
0.160400390625,
0.8857421875,
0.160400390625,
0.8857421875,
0.1749267578125,
0.50244140625,
0.1749267578125
],
[
0.4150390625,
0.1673583984375,
0.4248046875,
0.1673583984375,
0.4248046875,
0.1787109375,
0.4150390625,
0.1787109375
],
[
0.5009765625,
0.175537109375,
0.8857421875,
0.175537109375,
0.8857421875,
0.189453125,
0.5009765625,
0.189453125
],
[
0.5009765625,
0.1900634765625,
0.88623046875,
0.1900634765625,
0.88623046875,
0.203857421875,
0.5009765625,
0.203857421875
],
[
0.5,
0.203857421875,
0.88623046875,
0.205810546875,
0.88623046875,
0.2203369140625,
0.5,
0.2183837890625
],
[
0.498291015625,
0.2197265625,
0.88623046875,
0.2203369140625,
0.88623046875,
0.23486328125,
0.498291015625,
0.2342529296875
],
[
0.0882568359375,
0.2291259765625,
0.473876953125,
0.2291259765625,
0.473876953125,
0.242431640625,
0.0882568359375,
0.242431640625
],
[
0.5,
0.236083984375,
0.88623046875,
0.236083984375,
0.88623046875,
0.25,
0.5,
0.25
],
[
0.0882568359375,
0.24365234375,
0.47216796875,
0.24365234375,
0.47216796875,
0.2568359375,
0.0882568359375,
0.2568359375
],
[
0.498291015625,
0.25,
0.689453125,
0.248779296875,
0.689453125,
0.26318359375,
0.498291015625,
0.264404296875
],
[
0.08905029296875,
0.25830078125,
0.473876953125,
0.25830078125,
0.473876953125,
0.268310546875,
0.08905029296875,
0.268310546875
],
[
0.08660888671875,
0.270751953125,
0.371826171875,
0.26953125,
0.371826171875,
0.2841796875,
0.08660888671875,
0.285400390625
],
[
0.525390625,
0.271484375,
0.8447265625,
0.271484375,
0.8447265625,
0.285888671875,
0.525390625,
0.285888671875
],
[
0.85107421875,
0.272705078125,
0.8857421875,
0.272705078125,
0.8857421875,
0.285400390625,
0.85107421875,
0.285400390625
],
[
0.08660888671875,
0.29345703125,
0.473876953125,
0.294189453125,
0.473876953125,
0.309326171875,
0.08660888671875,
0.30810546875
],
[
0.498291015625,
0.291748046875,
0.88623046875,
0.292236328125,
0.88623046875,
0.306884765625,
0.498291015625,
0.30615234375
],
[
0.0882568359375,
0.31005859375,
0.473876953125,
0.31005859375,
0.473876953125,
0.323974609375,
0.0882568359375,
0.323974609375
],
[
0.5,
0.30810546875,
0.8857421875,
0.30810546875,
0.8857421875,
0.3212890625,
0.5,
0.3212890625
],
[
0.08660888671875,
0.32568359375,
0.473876953125,
0.323974609375,
0.473876953125,
0.337158203125,
0.08660888671875,
0.339111328125
],
[
0.5009765625,
0.322509765625,
0.8857421875,
0.322509765625,
0.8857421875,
0.3359375,
0.5009765625,
0.3359375
],
[
0.0882568359375,
0.34033203125,
0.4697265625,
0.34033203125,
0.4697265625,
0.354248046875,
0.0882568359375,
0.354248046875
],
[
0.5,
0.33837890625,
0.88623046875,
0.33837890625,
0.88623046875,
0.3515625,
0.5,
0.3515625
],
[
0.5009765625,
0.354248046875,
0.66845703125,
0.354248046875,
0.66845703125,
0.367431640625,
0.5009765625,
0.367431640625
],
[
0.0882568359375,
0.361083984375,
0.419189453125,
0.361083984375,
0.419189453125,
0.375732421875,
0.0882568359375,
0.375732421875
],
[
0.5009765625,
0.375,
0.88623046875,
0.375,
0.88623046875,
0.38818359375,
0.5009765625,
0.38818359375
],
[
0.08905029296875,
0.386474609375,
0.473876953125,
0.386474609375,
0.473876953125,
0.399658203125,
0.08905029296875,
0.399658203125
],
[
0.5,
0.39013671875,
0.88623046875,
0.39013671875,
0.88623046875,
0.404052734375,
0.5,
0.404052734375
],
[
0.11517333984375,
0.40087890625,
0.4755859375,
0.40087890625,
0.4755859375,
0.4140625,
0.11517333984375,
0.4140625
],
[
0.498291015625,
0.404052734375,
0.88623046875,
0.40478515625,
0.88623046875,
0.419189453125,
0.498291015625,
0.41845703125
],
[
0.11517333984375,
0.416748046875,
0.285888671875,
0.416748046875,
0.285888671875,
0.429931640625,
0.11517333984375,
0.429931640625
],
[
0.5,
0.41845703125,
0.88623046875,
0.419189453125,
0.88623046875,
0.434326171875,
0.5,
0.43310546875
],
[
0.08905029296875,
0.43505859375,
0.473876953125,
0.43505859375,
0.473876953125,
0.448974609375,
0.08905029296875,
0.448974609375
],
[
0.5,
0.43505859375,
0.8857421875,
0.43505859375,
0.8857421875,
0.448974609375,
0.5,
0.448974609375
],
[
0.11517333984375,
0.45068359375,
0.4755859375,
0.45068359375,
0.4755859375,
0.464111328125,
0.11517333984375,
0.464111328125
],
[
0.5,
0.45068359375,
0.8857421875,
0.45068359375,
0.8857421875,
0.464111328125,
0.5,
0.464111328125
],
[
0.1168212890625,
0.466552734375,
0.473876953125,
0.466552734375,
0.473876953125,
0.479736328125,
0.1168212890625,
0.479736328125
],
[
0.5009765625,
0.46533203125,
0.88623046875,
0.46533203125,
0.88623046875,
0.479248046875,
0.5009765625,
0.479248046875
],
[
0.11517333984375,
0.48095703125,
0.4453125,
0.48095703125,
0.4453125,
0.494384765625,
0.11517333984375,
0.494384765625
],
[
0.498291015625,
0.479736328125,
0.8203125,
0.479736328125,
0.8203125,
0.494384765625,
0.498291015625,
0.494384765625
],
[
0.08905029296875,
0.5,
0.473876953125,
0.5,
0.473876953125,
0.51318359375,
0.08905029296875,
0.51318359375
],
[
0.580078125,
0.501953125,
0.6337890625,
0.501953125,
0.6337890625,
0.517578125,
0.580078125,
0.517578125
],
[
0.6845703125,
0.498779296875,
0.71728515625,
0.498779296875,
0.71728515625,
0.52099609375,
0.6845703125,
0.52099609375
],
[
0.7587890625,
0.501953125,
0.80615234375,
0.501953125,
0.80615234375,
0.517578125,
0.7587890625,
0.517578125
],
[
0.861328125,
0.501953125,
0.88818359375,
0.501953125,
0.88818359375,
0.51806640625,
0.861328125,
0.51806640625
],
[
0.11517333984375,
0.51513671875,
0.473876953125,
0.51513671875,
0.473876953125,
0.52880859375,
0.11517333984375,
0.52880859375
],
[
0.6396484375,
0.52587890625,
0.6953125,
0.52587890625,
0.6953125,
0.52978515625,
0.6396484375,
0.52978515625
],
[
0.70361328125,
0.5234375,
0.75634765625,
0.5234375,
0.75634765625,
0.53076171875,
0.70361328125,
0.53076171875
],
[
0.11517333984375,
0.53076171875,
0.473876953125,
0.53076171875,
0.473876953125,
0.54443359375,
0.11517333984375,
0.54443359375
],
[
0.5,
0.53857421875,
0.88623046875,
0.537109375,
0.88623046875,
0.5517578125,
0.5,
0.55322265625
],
[
0.1168212890625,
0.54541015625,
0.473876953125,
0.54541015625,
0.473876953125,
0.5595703125,
0.1168212890625,
0.5595703125
],
[
0.5009765625,
0.5537109375,
0.88623046875,
0.5537109375,
0.88623046875,
0.5673828125,
0.5009765625,
0.5673828125
],
[
0.11517333984375,
0.56103515625,
0.4755859375,
0.56103515625,
0.4755859375,
0.57470703125,
0.11517333984375,
0.57470703125
],
[
0.5009765625,
0.568359375,
0.88232421875,
0.5693359375,
0.88232421875,
0.58349609375,
0.5009765625,
0.58203125
],
[
0.0882568359375,
0.580078125,
0.473876953125,
0.580078125,
0.473876953125,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.5009765625,
0.583984375,
0.8857421875,
0.583984375,
0.8857421875,
0.59765625,
0.5009765625,
0.59765625
],
[
0.11358642578125,
0.5947265625,
0.38232421875,
0.5947265625,
0.38232421875,
0.60791015625,
0.11358642578125,
0.60791015625
],
[
0.50244140625,
0.5986328125,
0.8857421875,
0.5986328125,
0.8857421875,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.5009765625,
0.6142578125,
0.8857421875,
0.6142578125,
0.8857421875,
0.62744140625,
0.5009765625,
0.62744140625
],
[
0.08905029296875,
0.62548828125,
0.223876953125,
0.62548828125,
0.223876953125,
0.64013671875,
0.08905029296875,
0.64013671875
],
[
0.5,
0.626953125,
0.8857421875,
0.62744140625,
0.8857421875,
0.642578125,
0.5,
0.6416015625
],
[
0.08905029296875,
0.64599609375,
0.1993408203125,
0.64599609375,
0.1993408203125,
0.6591796875,
0.08905029296875,
0.6591796875
],
[
0.5009765625,
0.64306640625,
0.8857421875,
0.64306640625,
0.8857421875,
0.6572265625,
0.5009765625,
0.6572265625
],
[
0.5,
0.65771484375,
0.8857421875,
0.65771484375,
0.8857421875,
0.671875,
0.5,
0.671875
],
[
0.0882568359375,
0.66748046875,
0.4755859375,
0.66748046875,
0.4755859375,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.5009765625,
0.673828125,
0.88623046875,
0.673828125,
0.88623046875,
0.6875,
0.5009765625,
0.6875
],
[
0.0906982421875,
0.6845703125,
0.47216796875,
0.6845703125,
0.47216796875,
0.6943359375,
0.0906982421875,
0.6943359375
],
[
0.5009765625,
0.68798828125,
0.88623046875,
0.68798828125,
0.88623046875,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.0882568359375,
0.69775390625,
0.4755859375,
0.69775390625,
0.4755859375,
0.7109375,
0.0882568359375,
0.7109375
],
[
0.5009765625,
0.7041015625,
0.88623046875,
0.7041015625,
0.88623046875,
0.71728515625,
0.5009765625,
0.71728515625
],
[
0.0882568359375,
0.7119140625,
0.4755859375,
0.7119140625,
0.4755859375,
0.72607421875,
0.0882568359375,
0.72607421875
],
[
0.498291015625,
0.71728515625,
0.8857421875,
0.71728515625,
0.8857421875,
0.73095703125,
0.498291015625,
0.73095703125
],
[
0.0882568359375,
0.72802734375,
0.473876953125,
0.72802734375,
0.473876953125,
0.7412109375,
0.0882568359375,
0.7412109375
],
[
0.5009765625,
0.73486328125,
0.8310546875,
0.73486328125,
0.8310546875,
0.74560546875,
0.5009765625,
0.74560546875
],
[
0.0882568359375,
0.7421875,
0.473876953125,
0.7421875,
0.473876953125,
0.755859375,
0.0882568359375,
0.755859375
],
[
0.0882568359375,
0.7568359375,
0.473876953125,
0.7568359375,
0.473876953125,
0.77099609375,
0.0882568359375,
0.77099609375
],
[
0.5009765625,
0.755859375,
0.88623046875,
0.755859375,
0.88623046875,
0.76953125,
0.5009765625,
0.76953125
],
[
0.0882568359375,
0.77294921875,
0.4755859375,
0.77294921875,
0.4755859375,
0.7861328125,
0.0882568359375,
0.7861328125
],
[
0.5009765625,
0.77099609375,
0.88623046875,
0.77099609375,
0.88623046875,
0.7841796875,
0.5009765625,
0.7841796875
],
[
0.0882568359375,
0.787109375,
0.473876953125,
0.787109375,
0.473876953125,
0.80029296875,
0.0882568359375,
0.80029296875
],
[
0.5,
0.7841796875,
0.8857421875,
0.78515625,
0.8857421875,
0.7998046875,
0.5,
0.798828125
],
[
0.08905029296875,
0.8037109375,
0.47216796875,
0.8037109375,
0.47216796875,
0.814453125,
0.08905029296875,
0.814453125
],
[
0.5,
0.80029296875,
0.8857421875,
0.80029296875,
0.8857421875,
0.814453125,
0.5,
0.814453125
],
[
0.0882568359375,
0.8173828125,
0.4755859375,
0.8173828125,
0.4755859375,
0.83056640625,
0.0882568359375,
0.83056640625
],
[
0.5009765625,
0.81494140625,
0.88623046875,
0.81494140625,
0.88623046875,
0.8291015625,
0.5009765625,
0.8291015625
],
[
0.0882568359375,
0.83203125,
0.473876953125,
0.83203125,
0.473876953125,
0.84521484375,
0.0882568359375,
0.84521484375
],
[
0.5,
0.83056640625,
0.88623046875,
0.83056640625,
0.88623046875,
0.8447265625,
0.5,
0.8447265625
],
[
0.0882568359375,
0.84765625,
0.4755859375,
0.84765625,
0.4755859375,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.521484375,
0.85302734375,
0.8857421875,
0.85400390625,
0.8857421875,
0.86865234375,
0.521484375,
0.8671875
],
[
0.08660888671875,
0.8623046875,
0.473876953125,
0.861328125,
0.473876953125,
0.875,
0.08660888671875,
0.87548828125
],
[
0.5,
0.8671875,
0.88623046875,
0.8671875,
0.88623046875,
0.880859375,
0.5,
0.880859375
],
[
0.0882568359375,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.5,
0.8798828125,
0.88623046875,
0.8798828125,
0.88623046875,
0.89306640625,
0.5,
0.89306640625
],
[
0.0882568359375,
0.892578125,
0.396240234375,
0.892578125,
0.396240234375,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.498291015625,
0.892578125,
0.87841796875,
0.892578125,
0.87841796875,
0.90576171875,
0.498291015625,
0.90576171875
]
] |
[
"Generative Flow Networks for Discrete Probabilistic Modeling",
"SD",
"A complete trajectory is a sequence of states",
"So-→s1-→...-→sn, where each transition St-→st+1 is",
"an action in A and sn E X is terminal. A trajectory",
"flow is a measure (unnormalized density) on the set of",
"",
"all complete trajectories T, i.e., a non-negative function",
"F : T → R>o, which can be thought of as a number",
"...",
"of particles flowing from so to terminal states along",
"each route. The flow is called Markovian if there exist",
"distributions Pr(s) over the children of every non-",
"terminal state s, and a constant Z, such that for any",
"Figure 2. The state space S and the GFlowNet's forward modeling",
"complete trajectory T = (so-→si→..-→...-→sn), we",
"process in a 9-dimensional discrete data space. The states are the",
"have PF() = F(-)/Z with",
"vertices of a DAG whose edges are the transitions--actions of",
"painting a grey pixel into black (1) or white (0).",
"PF(T) = PF(s1so)PF(s2|s1)...PF(sn|sn-1).",
"(1)",
"modes without having to travel long low-probability paths",
"In this case, Pe (st+ist) is called a forward policy, and can",
"in between (Fig. I). Instead of having to exploit a priori",
"be used to sample complete trajectories from the density",
"structuretodesignlong jumps intheMCMC,GFlowNets",
"F and thus also their terminal states, objects x E X.1 We",
"can discover that structure when it is not explicitly known",
"write Pr(x) for the probability that a trajectory sampled",
"from PF terminates in x.",
"This paper makes the following key contributions:",
"Past work on GFlowNets (Bengio et al., 2021a;b; Malkin",
"(1) We cast a non-autoregressive sequential generation",
"et al., 2022; Jain et al., 2022) has considered the problem",
"model for high-dimensional discrete data as a gener-",
"of fitting a Markovian flow to a fixed reward function on",
"ative flownetwork (s3.1).",
"X.Given a non-negative reward function R:X →R>o,",
"(2) We introduce a GFlowNet-based MCMC proposal en",
"one seeks a Markovian flow F such that the likelihood of",
"abling efficient large jumps with low probability of re-",
"a trajectory sampled from F terminating in a given x E",
"jection, taking advantage of the compositional structure",
"X is proportional to R(x), i.e., Pr(x) α R(x). This is",
"learned by the GFlowNet generative policy ($3.3).",
"accomplished by the reward matching condition:",
"(3) We describe a procedure based on such proposals for",
"R(x) =",
"Z",
"F(T).",
"(2)",
"jointly training the GFlowNet sampler with an energy",
"So",
">Sm.Sn=",
"model given a dataset and state conditions under which",
"By Eq. (1), one can specify a Markovian flow by a learned",
"this estimates the true data log-likelihood gradient with",
"scalar log Ze and a neural net with parameters that outputs",
"respect to the energy function's parameters (s3.2, $3.3).",
"Pe(s; O) for any input state s. Algorithms for learning PF",
"(4) We test the algorithm on a variety of synthetic and real",
"from maximum-entropy reinforcement learning (Haarnoja",
"tasks, achieving competitive results (s4).",
"et al., 2017) would solve this problem if there was only",
"one way to construct an object x (the DAG is a tree) while",
"2. Preliminaries",
"GFlowNets are applicable in the more general DAG setting",
"2.1. GFlowNets",
"(Bengio et al., 202la). The solution originally proposed",
"by Bengio et al.(2021a)recasts an unnormalized forward",
"Generative flow networks, or GFlowNets for short (Bengio",
"policy as a network flow on G (in the classical sense of",
"etal.,202la:b),aretrainablegenerativepoliciesonwhich",
"Ford & Fulkerson (1956) and performs gradient descent on",
"this paper is built. They model the generation process of",
"the error in a flow conservation constraint at states sampled",
"objects x E x by a sequence of discrete actions that in-",
"from a training policy, amounting to a generalization of",
"crementally modify a partially constructed object (state)",
"temporaldifferenceobjectivesfromSutton(2005)",
"The space of possible action sequences is represented by",
"a directed acyclic graph (DAG, see Fig. 2) G = (S,A),",
"Alternative objectives proposed by Bengio et al. (2021b):",
"where the vertices in S are states and the edges in A are",
"Malkin et al. (2022) require a model to produce 3 outputs:",
"actions that modify one state to another. We use the terms",
"the scalar log Ze, a forward policy P(-:; O), and a back",
"parents and children,respectively,forthetails of incoming",
"ward policy that produces distributions PB(s; O) over the",
"edges and the heads of outgoing edges of a state. Genera-",
"parents of an input state s. (When action sequences are sam-",
"tion begins at a special initial state so and terminates upon",
"pled in reverse from PB, we will use dashed arrows s'--+s",
"a transition to any terminal state, which is a state with no",
"'it is helpful to think of the particle analogy. For a Markovian",
"outgoing actions.The set of terminal states is identified",
"flow, the distribution over complete trajectories (PF() α F(←))",
"with the target space X. Note that multiple possible action",
"satisfies a history-independence property: a particle's choice of",
"sequences may lead to the same terminal state.",
"route after reaching a state s is independent of how it reached s."
] |
{
"arxiv_id": "2202.01361",
"language": "en",
"page_idx": 1
}
|
||
2202.01361-en-2
|
2202.01361-en-2.jpg
|
Generative Flow Networks for Discrete Probabilistic Modeling
to indicate that the action s$\to\mathbf{s}^\prime$ has been sampled against the direction of the DAG edges. The $\theta$ will be dropped from $P_{F}$ and $P_B$ notation when it does not cause ambiguity.)
In modeling tasks with a fixed reward (Bengio et al., 2021a; Malkin et al., 2022), only the forward policy $P_F$ is used for sampling from the GFlowNet, and the backward policy $P_B$ is simply an training artifact. However, our approach makes use of the backward policy to perform local exploration in a contrastive divergence-like algorithm (§3.3, Figure 1).
## 2.2. Energy-based models
Energy-based models (EBMs) (LeCun et al., 2006; Song & Kingma, 2021) are a popular approach for probabilistic inference and modeling. An EBM specifies a distribution $\operatorname { \text{over a space }} \mathcal{X} \operatorname { \text{by a density }} p_{\phi }( \mathbf{x} ) = \frac 1{Z_{\phi }}\exp ( - \mathcal{E} _{\phi }( \mathbf{x} ) ) $, where $\mathcal{E}_\phi$ is the energy function, with model parameter $\phi$, and $Z_\phi$ is a normalizing constant independent of $\mathbf{x}.$ The normalizing factor is not explicitly parametrized, but, for finite spaces, can theoretically be computed as $Z_{\phi}=\sum_{\mathbf{x}\in\mathcal{X}}\exp(-\mathcal{E}_{\phi}(\mathbf{x})).$This summation is finite, and thus the energy function $\mathcal{E}_\phi(\mathbf{x})$ can take an arbitrary form. However, this flexibility comes with a price: calculating $Z_\phi$ can involve an exponentially large summation, making evaluation of the exact probability intractable. To avoid such expensive calculations, it is desirable to amortize sampling from such models by training a generative model.
We can train $\mathcal{E}_\phi$ through maximum likelihood estima-
$\operatorname{tion}($MLE$),i.e.$seeking$\operatorname { arg} \max _\phi \mathbb{E} _\mathbf{x} \sim p_{\mathrm{data}}( \mathbf{x} ) \left [ \log p_{\mathbf{\phi } }( \mathbf{x} ) \right ] $, where $p_{\mathrm{data}}$ is the empirical distribution associated with the training data. The gradient of negative log-likelihood (NLL) with respect to the model parameter $\phi$ is given by
$$-\nabla_{\phi}\log p_{\phi}(\mathbf{x})=\nabla_{\phi}\mathcal{E}_{\phi}(\mathbf{x})+\nabla_{\phi}\log Z_{\phi}\\=\nabla_{\phi}\mathcal{E}_{\phi}(\mathbf{x})-\mathbb{E}_{\mathbf{x}^{\prime}\sim p_{\phi}(\mathbf{x}^{\prime})}[\nabla_{\phi}\mathcal{E}_{\phi}(\mathbf{x}^{\prime})].$$
(3)
Evaluating or estimating the second term in (3) involves taking negative samples x' from the EBM distribution, which can require an expensive search. The classical Contrastive Divergence (CD) algorithm (Hinton, 2002) approximates this gradient update by changing the energy function parameter with the following stochastic approximation:
$$\mathbb{E}_{\mathbf{x}\sim p_{\mathrm{data}}(\mathbf{x})}[\nabla_{\boldsymbol{\phi}}\mathcal{E}_{\boldsymbol{\phi}}(\mathbf{x})-\mathbb{E}_{\mathbf{x}^{\prime}\sim q_{K}(\mathbf{x}^{\prime}|\mathbf{x})}\nabla_{\boldsymbol{\phi}}\mathcal{E}_{\boldsymbol{\phi}}(\mathbf{x}^{\prime})]$$
(4)
where $q_K(\mathbf{x}^{\prime}|\mathbf{x})$ is the distribution obtained by using a $K$- step MCMC initialized at x to approximately sample from $p_{\phi}.$ For example, $\mathbf{x}^\prime$ can be taken from a $K$-step MetropolisHastings chain starting at a true data sample $\mathbf{x} \sim p_{\mathrm{data}}( \mathbf{x} ) $, and the parameters updated with $\nabla_\phi(\mathcal{E}_\phi(\mathbf{x})-\mathcal{E}_\phi(\mathbf{x}^{\prime})).$ As $K\to\infty$, assuming mixing of MCMC chains, the distribution over negative samples $q_K(\mathbf{x}^{\prime}|\mathbf{x})$ converges to $p_\phi(\mathbf{x}^{\prime})$, and we recover, in expectation, the true gradient (3).
Tieleman (2008) later proposed persistent CD (PCD), where the MCMC chains that give negative samples do not restart
<FigureHere>
Figure 3. An illustration of the forward and backward GFlowNet policies in a 9-dimensional discrete space of the kind studied here. The forward policy transforms a state s$_t$ into s$_t+1$, while the backward policy does the opposite operation. We represent 0/1 with black / white patches, and use grey patches to denote unspecified entries $\oslash$ in incomplete (non-terminal) states.
at true data at every training step, but are initialized with a previous state. While CD-type algorithms are efficient in computation and can accelerate learning, their gradient estimation is biased and thus may not model the true data distribution faithfully (Nijkamp et al., 2020).
# 3. Methodology
3.1. GFlowNet generative process
In this work, we aim to model a target distribution over discrete data with a GFlowNet. In the domains we consider, the data is in $D$-dimensional binary space, $i.e.,\mathbf{x}\in\mathcal{X}\triangleq$ $\{0,1\}^D.$ As an example, x could be an image with $D$ pixels taking binary values. We model the generation of vectors in $\mathcal{X}$ by a GFlowNet. The state space $\mathcal{S}$ of the GFlowNet consists of vectors of length $d$ with entries in $\{0,1,\emptyset\}$, where the void symbol $\oslash$ represents a yet unspecified entry that may be turned to 0 or 1 by a future action. To be precise:
$$\mathcal{S}\triangleq\{(\mathbf{s}^1,\ldots,\mathbf{s}^D)\mid\mathbf{s}^d\in\{0,1,\otimes\},d=1,\ldots,D\}.$$
(5)
The DAG structure on $S$ is the $D$-th Cartesian power of the DAG with states $\{\varnothing,0,1\}$, where 0 and 1 are children of $\oslash.$ Concretely, the children of a state s$=(\mathbf{s}^1,\ldots,\mathbf{s}^D)$ are vectors that can be obtained from s by changing any one entry s$^d$ from $\oslash$ to 0 or 1, and its parents are states that can be obtained by changing a single entry s$^d\in\{0,1\}$ to$\otimes.$
We define $|\mathbf{s}|\triangleq\#\{\mathbf{s}^d\mid\mathbf{s}^d\in\{0,1\},d=1,\ldots,D\}$,the number of non-void entries in s, so $\chi$ is naturally identified with $\{\mathbf{s}\in\mathcal{S}:|\mathbf{s}|=D\}.$ There is an initial state $\mathbf{s}_0\triangleq$ $(\emptyset,\emptyset,\ldots,\emptyset).$ Any trajectory from s$_0$ to $\mathbf{x}\in\mathcal{X}$ has exactly $D\text{ actions. A choice of trajectory from s}_{0}$ to x amounts to a choice of the order in which the entries of x are assigned. In this setting, the forward policy $P_F(\cdot|\mathbf{s};\boldsymbol{\theta})$ of a GFlowNet introduced in §2.1, is a distribution over all ways to select a position with a void entry in s and a value (0 or 1) to assign to this entry. Thus the action space for a state s has size $2(D-|\mathbf{s}|).$ Correspondingly, the backward policy $P_{B}(\cdot|\mathbf{s};\boldsymbol{\theta})$ is a distribution over the $|\mathbf{s}|$ ways to select a position with a nonvoid entry in s. We illustrate the mechanism of the forward and backward policies in Figure 3.
In our experiments, we take $P_F(\cdot|\mathbf{s};\boldsymbol{\theta})$ and $P_B(\cdot|\mathbf{s};\boldsymbol{\theta})$ to
|
[
[
0.291748046875,
0.05682373046875,
0.68310546875,
0.05682373046875,
0.68310546875,
0.0706787109375,
0.291748046875,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.473876953125,
0.08587646484375,
0.473876953125,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.6845703125,
0.0889892578125,
0.775390625,
0.0889892578125,
0.775390625,
0.10101318359375,
0.6845703125,
0.10101318359375
],
[
0.81640625,
0.0889892578125,
0.84619140625,
0.0889892578125,
0.84619140625,
0.10101318359375,
0.81640625,
0.10101318359375
],
[
0.0906982421875,
0.1029052734375,
0.47216796875,
0.1029052734375,
0.47216796875,
0.11297607421875,
0.0906982421875,
0.11297607421875
],
[
0.53759765625,
0.09027099609375,
0.55224609375,
0.09027099609375,
0.55224609375,
0.0997314453125,
0.53759765625,
0.0997314453125
],
[
0.08905029296875,
0.11614990234375,
0.453369140625,
0.11614990234375,
0.453369140625,
0.1300048828125,
0.08905029296875,
0.1300048828125
],
[
0.8349609375,
0.11553955078125,
0.84228515625,
0.11553955078125,
0.84228515625,
0.11932373046875,
0.8349609375,
0.11932373046875
],
[
0.82861328125,
0.12249755859375,
0.85107421875,
0.12249755859375,
0.85107421875,
0.1300048828125,
0.82861328125,
0.1300048828125
],
[
0.08905029296875,
0.1383056640625,
0.4755859375,
0.1383056640625,
0.4755859375,
0.1514892578125,
0.08905029296875,
0.1514892578125
],
[
0.53759765625,
0.15283203125,
0.5556640625,
0.15283203125,
0.5556640625,
0.1610107421875,
0.53759765625,
0.1610107421875
],
[
0.67724609375,
0.15087890625,
0.771484375,
0.15087890625,
0.771484375,
0.1654052734375,
0.67724609375,
0.1654052734375
],
[
0.82275390625,
0.15283203125,
0.8447265625,
0.15283203125,
0.8447265625,
0.1610107421875,
0.82275390625,
0.1610107421875
],
[
0.08905029296875,
0.1685791015625,
0.47216796875,
0.1685791015625,
0.47216796875,
0.1817626953125,
0.08905029296875,
0.1817626953125
],
[
0.08905029296875,
0.154052734375,
0.473876953125,
0.154052734375,
0.473876953125,
0.1673583984375,
0.08905029296875,
0.1673583984375
],
[
0.50244140625,
0.1717529296875,
0.8837890625,
0.1717529296875,
0.8837890625,
0.1817626953125,
0.50244140625,
0.1817626953125
],
[
0.0882568359375,
0.184326171875,
0.473876953125,
0.184326171875,
0.473876953125,
0.1976318359375,
0.0882568359375,
0.1976318359375
],
[
0.5009765625,
0.184326171875,
0.88623046875,
0.184326171875,
0.88623046875,
0.1976318359375,
0.5009765625,
0.1976318359375
],
[
0.0882568359375,
0.1988525390625,
0.473876953125,
0.1988525390625,
0.473876953125,
0.212158203125,
0.0882568359375,
0.212158203125
],
[
0.5,
0.1988525390625,
0.8857421875,
0.1988525390625,
0.8857421875,
0.212158203125,
0.5,
0.212158203125
],
[
0.0882568359375,
0.21337890625,
0.44287109375,
0.21337890625,
0.44287109375,
0.2266845703125,
0.0882568359375,
0.2266845703125
],
[
0.5009765625,
0.212158203125,
0.8857421875,
0.21337890625,
0.8857421875,
0.2247314453125,
0.5009765625,
0.2235107421875
],
[
0.5,
0.2247314453125,
0.88623046875,
0.2259521484375,
0.88623046875,
0.240478515625,
0.5,
0.2392578125
],
[
0.08905029296875,
0.24365234375,
0.268798828125,
0.24365234375,
0.268798828125,
0.2568359375,
0.08905029296875,
0.2568359375
],
[
0.498291015625,
0.240478515625,
0.8388671875,
0.240478515625,
0.8388671875,
0.25390625,
0.498291015625,
0.25390625
],
[
0.0882568359375,
0.264404296875,
0.473876953125,
0.264404296875,
0.473876953125,
0.27783203125,
0.0882568359375,
0.27783203125
],
[
0.498291015625,
0.267578125,
0.8857421875,
0.266357421875,
0.8857421875,
0.281005859375,
0.498291015625,
0.2822265625
],
[
0.0882568359375,
0.27783203125,
0.47216796875,
0.27783203125,
0.47216796875,
0.292236328125,
0.0882568359375,
0.292236328125
],
[
0.5,
0.28271484375,
0.8857421875,
0.28271484375,
0.8857421875,
0.29736328125,
0.5,
0.29736328125
],
[
0.08905029296875,
0.294189453125,
0.473876953125,
0.294189453125,
0.473876953125,
0.30810546875,
0.08905029296875,
0.30810546875
],
[
0.5009765625,
0.2998046875,
0.8837890625,
0.2998046875,
0.8837890625,
0.31005859375,
0.5009765625,
0.31005859375
],
[
0.0882568359375,
0.31005859375,
0.4755859375,
0.31005859375,
0.4755859375,
0.324462890625,
0.0882568359375,
0.324462890625
],
[
0.498291015625,
0.313232421875,
0.8857421875,
0.313232421875,
0.8857421875,
0.326904296875,
0.498291015625,
0.326904296875
],
[
0.08660888671875,
0.324462890625,
0.473876953125,
0.32568359375,
0.473876953125,
0.34033203125,
0.08660888671875,
0.339111328125
],
[
0.5,
0.32763671875,
0.79345703125,
0.32763671875,
0.79345703125,
0.34228515625,
0.5,
0.34228515625
],
[
0.0882568359375,
0.341552734375,
0.473876953125,
0.341552734375,
0.473876953125,
0.354736328125,
0.0882568359375,
0.354736328125
],
[
0.08905029296875,
0.35791015625,
0.473876953125,
0.35791015625,
0.473876953125,
0.36865234375,
0.08905029296875,
0.36865234375
],
[
0.498291015625,
0.354248046875,
0.6298828125,
0.356689453125,
0.6298828125,
0.37255859375,
0.498291015625,
0.369384765625
],
[
0.08905029296875,
0.371826171875,
0.473876953125,
0.371826171875,
0.473876953125,
0.385009765625,
0.08905029296875,
0.385009765625
],
[
0.5,
0.376953125,
0.734375,
0.376953125,
0.734375,
0.39013671875,
0.5,
0.39013671875
],
[
0.08905029296875,
0.385009765625,
0.4755859375,
0.385009765625,
0.4755859375,
0.40283203125,
0.08905029296875,
0.40283203125
],
[
0.08905029296875,
0.401611328125,
0.47216796875,
0.401611328125,
0.47216796875,
0.415283203125,
0.08905029296875,
0.415283203125
],
[
0.5,
0.3984375,
0.88623046875,
0.3984375,
0.88623046875,
0.412353515625,
0.5,
0.412353515625
],
[
0.08905029296875,
0.416748046875,
0.4755859375,
0.416748046875,
0.4755859375,
0.429931640625,
0.08905029296875,
0.429931640625
],
[
0.5009765625,
0.415283203125,
0.8857421875,
0.415283203125,
0.8857421875,
0.425537109375,
0.5009765625,
0.425537109375
],
[
0.08905029296875,
0.43115234375,
0.473876953125,
0.43115234375,
0.473876953125,
0.4443359375,
0.08905029296875,
0.4443359375
],
[
0.498291015625,
0.4267578125,
0.88623046875,
0.425537109375,
0.88623046875,
0.443115234375,
0.498291015625,
0.4443359375
],
[
0.08905029296875,
0.447509765625,
0.473876953125,
0.447509765625,
0.473876953125,
0.458251953125,
0.08905029296875,
0.458251953125
],
[
0.50244140625,
0.443115234375,
0.8857421875,
0.443115234375,
0.8857421875,
0.45703125,
0.50244140625,
0.45703125
],
[
0.08905029296875,
0.4609375,
0.473876953125,
0.4609375,
0.473876953125,
0.474853515625,
0.08905029296875,
0.474853515625
],
[
0.50244140625,
0.460205078125,
0.8837890625,
0.460205078125,
0.8837890625,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.08905029296875,
0.47607421875,
0.4150390625,
0.47607421875,
0.4150390625,
0.490478515625,
0.08905029296875,
0.490478515625
],
[
0.5,
0.473388671875,
0.8857421875,
0.473388671875,
0.8857421875,
0.4873046875,
0.5,
0.4873046875
],
[
0.5,
0.4892578125,
0.88623046875,
0.4892578125,
0.88623046875,
0.50244140625,
0.5,
0.50244140625
],
[
0.0882568359375,
0.49755859375,
0.473876953125,
0.49755859375,
0.473876953125,
0.51123046875,
0.0882568359375,
0.51123046875
],
[
0.5,
0.50244140625,
0.8857421875,
0.50390625,
0.8857421875,
0.517578125,
0.5,
0.5166015625
],
[
0.0849609375,
0.51123046875,
0.4755859375,
0.51318359375,
0.4755859375,
0.53076171875,
0.0849609375,
0.52880859375
],
[
0.5,
0.51806640625,
0.88818359375,
0.51806640625,
0.88818359375,
0.5322265625,
0.5,
0.5322265625
],
[
0.08660888671875,
0.52783203125,
0.473876953125,
0.5263671875,
0.473876953125,
0.54150390625,
0.08660888671875,
0.54248046875
],
[
0.0882568359375,
0.54345703125,
0.473876953125,
0.54345703125,
0.473876953125,
0.556640625,
0.0882568359375,
0.556640625
],
[
0.5107421875,
0.54345703125,
0.8837890625,
0.54345703125,
0.8837890625,
0.55810546875,
0.5107421875,
0.55810546875
],
[
0.0882568359375,
0.55810546875,
0.4150390625,
0.55810546875,
0.4150390625,
0.5712890625,
0.0882568359375,
0.5712890625
],
[
0.5,
0.568359375,
0.8837890625,
0.5693359375,
0.8837890625,
0.583984375,
0.5,
0.58349609375
],
[
0.11517333984375,
0.58056640625,
0.39208984375,
0.58056640625,
0.39208984375,
0.59521484375,
0.11517333984375,
0.59521484375
],
[
0.5009765625,
0.58544921875,
0.8857421875,
0.58544921875,
0.8857421875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.1732177734375,
0.599609375,
0.43310546875,
0.599609375,
0.43310546875,
0.6171875,
0.1732177734375,
0.6171875
],
[
0.4306640625,
0.6015625,
0.47216796875,
0.6015625,
0.47216796875,
0.61279296875,
0.4306640625,
0.61279296875
],
[
0.5009765625,
0.60107421875,
0.8857421875,
0.60107421875,
0.8857421875,
0.6142578125,
0.5009765625,
0.6142578125
],
[
0.5009765625,
0.61572265625,
0.8857421875,
0.61572265625,
0.8857421875,
0.62890625,
0.5009765625,
0.62890625
],
[
0.42724609375,
0.60498046875,
0.43310546875,
0.60498046875,
0.43310546875,
0.611328125,
0.42724609375,
0.611328125
],
[
0.08905029296875,
0.62548828125,
0.473876953125,
0.62548828125,
0.473876953125,
0.6396484375,
0.08905029296875,
0.6396484375
],
[
0.5,
0.6298828125,
0.88623046875,
0.6298828125,
0.88623046875,
0.64453125,
0.5,
0.64453125
],
[
0.0882568359375,
0.64013671875,
0.473876953125,
0.6396484375,
0.473876953125,
0.65380859375,
0.0882568359375,
0.65478515625
],
[
0.5009765625,
0.64599609375,
0.87109375,
0.64599609375,
0.87109375,
0.6591796875,
0.5009765625,
0.6591796875
],
[
0.0882568359375,
0.65576171875,
0.473876953125,
0.65576171875,
0.473876953125,
0.66943359375,
0.0882568359375,
0.66943359375
],
[
0.0882568359375,
0.66943359375,
0.473876953125,
0.67041015625,
0.473876953125,
0.68505859375,
0.0882568359375,
0.6845703125
],
[
0.5009765625,
0.66748046875,
0.8857421875,
0.66748046875,
0.8857421875,
0.68115234375,
0.5009765625,
0.68115234375
],
[
0.08905029296875,
0.68603515625,
0.473876953125,
0.68603515625,
0.473876953125,
0.69970703125,
0.08905029296875,
0.69970703125
],
[
0.5009765625,
0.681640625,
0.8857421875,
0.681640625,
0.8857421875,
0.69580078125,
0.5009765625,
0.69580078125
],
[
0.0882568359375,
0.70068359375,
0.410888671875,
0.70068359375,
0.410888671875,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.5009765625,
0.69775390625,
0.8857421875,
0.69775390625,
0.8857421875,
0.7109375,
0.5009765625,
0.7109375
],
[
0.498291015625,
0.7119140625,
0.8857421875,
0.7109375,
0.8857421875,
0.72607421875,
0.498291015625,
0.7265625
],
[
0.11273193359375,
0.72265625,
0.431396484375,
0.7216796875,
0.431396484375,
0.7392578125,
0.11273193359375,
0.74072265625
],
[
0.449462890625,
0.72607421875,
0.47216796875,
0.72607421875,
0.47216796875,
0.7373046875,
0.449462890625,
0.7373046875
],
[
0.50244140625,
0.7265625,
0.8857421875,
0.7265625,
0.8857421875,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.5,
0.7421875,
0.8759765625,
0.7421875,
0.8759765625,
0.755859375,
0.5,
0.755859375
],
[
0.0882568359375,
0.74853515625,
0.47216796875,
0.74853515625,
0.47216796875,
0.76220703125,
0.0882568359375,
0.76220703125
],
[
0.0882568359375,
0.76513671875,
0.4755859375,
0.76513671875,
0.4755859375,
0.77880859375,
0.0882568359375,
0.77880859375
],
[
0.5,
0.7646484375,
0.88623046875,
0.7646484375,
0.88623046875,
0.77783203125,
0.5,
0.77783203125
],
[
0.08660888671875,
0.77978515625,
0.476318359375,
0.77783203125,
0.476318359375,
0.79248046875,
0.08660888671875,
0.79443359375
],
[
0.5,
0.78076171875,
0.8857421875,
0.78076171875,
0.8857421875,
0.79150390625,
0.5,
0.79150390625
],
[
0.0882568359375,
0.79345703125,
0.4755859375,
0.79443359375,
0.4755859375,
0.8095703125,
0.0882568359375,
0.80810546875
],
[
0.5,
0.79443359375,
0.88623046875,
0.79443359375,
0.88623046875,
0.80810546875,
0.5,
0.80810546875
],
[
0.08905029296875,
0.81005859375,
0.4755859375,
0.81005859375,
0.4755859375,
0.82373046875,
0.08905029296875,
0.82373046875
],
[
0.5009765625,
0.81005859375,
0.8857421875,
0.81005859375,
0.8857421875,
0.82373046875,
0.5009765625,
0.82373046875
],
[
0.08905029296875,
0.82470703125,
0.473876953125,
0.82470703125,
0.473876953125,
0.83837890625,
0.08905029296875,
0.83837890625
],
[
0.5,
0.82275390625,
0.8857421875,
0.82470703125,
0.8857421875,
0.8388671875,
0.5,
0.8369140625
],
[
0.0882568359375,
0.8388671875,
0.47802734375,
0.8388671875,
0.47802734375,
0.85400390625,
0.0882568359375,
0.85400390625
],
[
0.5009765625,
0.8388671875,
0.88623046875,
0.8388671875,
0.88623046875,
0.85302734375,
0.5009765625,
0.85302734375
],
[
0.0882568359375,
0.85498046875,
0.43310546875,
0.85498046875,
0.43310546875,
0.86865234375,
0.0882568359375,
0.86865234375
],
[
0.50244140625,
0.85595703125,
0.8857421875,
0.85595703125,
0.8857421875,
0.8662109375,
0.50244140625,
0.8662109375
],
[
0.5,
0.869140625,
0.826171875,
0.869140625,
0.826171875,
0.88330078125,
0.5,
0.88330078125
],
[
0.0882568359375,
0.87548828125,
0.473876953125,
0.87548828125,
0.473876953125,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.0882568359375,
0.892578125,
0.4755859375,
0.892578125,
0.4755859375,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.5,
0.8916015625,
0.8857421875,
0.8916015625,
0.8857421875,
0.90478515625,
0.5,
0.90478515625
]
] |
[
"Generative Flow Networks for Discrete Probabilistic Modeling",
"to indicate that the action s→s' has been sampled against",
"PE (St+1|St)",
"SD",
"thedirectionoftheDAGedges.Thewillbedroppedfrom",
"So",
"Pe and PB notation when it does not cause ambiguity.)",
"",
"",
"In modeling tasks with a fixed reward (Bengio et al., 202la;",
"So",
"PB (st /St+1",
"SD",
"sampling from the GFlowNet, and the backward policy PB",
"Malkin et al., 2022), only the forward policy PF is used for",
"Figure3.Anillustrationof theforward andbackwardGFlowNet",
"is simply an training artifact. However, our approach makes",
"policies in a 9-dimensional discrete space of the kind studied",
"use of the backward policy to perform local exploration in a",
"here. The forward policy transforms a state st into st+1, while",
"contrastive divergence-like algorithm (s3.3, Figure 1).",
"thebackward policy does theoppositeoperation.Werepresent",
"0/1 with black / white patches, and use grey patches to denote",
"2.2.Energy-based models",
"unspecified entries in incomplete (non-terminal) states.",
"Energy-based models (EBMs) (LeCun et al., 2006; Song",
"at true data at every training step, but are initialized with",
"&Kingma,2021) are a popular approach forprobabilistic",
"a previous state. While CD-type algorithms are efficient",
"inference and modeling. An EBM specifies a distribution",
"incomputationandcanacceleratelearning.theirgradient",
"over a space X by a density p(x) = 云 exp(-&(x),",
"estimation is biased and thus may not model the true data",
"where E is the energy function, with model parame-",
"distribution faithfully (Nijkamp et al., 2020).",
"ter Φ, and Z is a normalizing constant independent of",
"x. The normalizing factor is not explicitly parametrized.",
"3. Methodology",
"but, for finite spaces, can theoretically be computed as",
"3.1. GFlowNet generative process",
"Z = xex exp(-E(x). This summation is finite, and",
"thus the energy function &α(x) can take an arbitrary form",
"In this work, we aim to model a target distribution over",
"However, this flexibility comes with a price: calculating",
"discretedatawithaGFlowNet.Inthedomainsweconsider",
"Z can involve an exponentially large summation, making",
"the data is in D-dimensional binary space, i.e., x e ",
"evaluationoftheexactprobabilityintractable.Toavoid such",
"[0, 1 D. As an example, x could be an image with D pixels",
"expensive calculations, it is desirable to amortize sampling",
"takingbinary values.Wemodel thegeneration of vectors",
"from such models by training a generative model.",
"in X by a GFlowNet. The state space S of the GFlowNet",
"consists of vectors of length d with entries in [0,1, O],",
"We can train E through maximum likelihood estima-",
"where the void symbol represents a yet unspecified entry",
"tion (MLE), i.e. seeking arg maxg Ex~pa(x) [log Ps(x)],",
"that may be turned to O or 1 by a future action. To be precise:",
"where pdata is the empirical distribution associated with the",
"training data. The gradient of negative log-likelihood (NLL)",
"S 4 (si..... sD) I sd e{0,1.Ol,d = 1..... D). (5)",
"with respect to the model parameter is given by",
"The DAG structure on S is the D-th Cartesian power of the",
"VlogPo(x)=Ve(x) +VlogZs",
"DAG with states f, 0, 11, where 0 and 1 are children of",
"=&g(x) -Ex~ps(x)[VE(x)].",
"(3)",
". Concretely, the children of a state s = (s',..., sD) are",
"vectors that can be obtained from s by changing any one",
"",
"Evaluating or estimating the second term in (3) involves tak-",
"entry sd from to 0 or 1, and its parents are states that can",
"ing negative samples x' from the EBM distribution, which",
"be obtained by changing a single entry sd E (0, 1] to O.",
"can require an expensive search. The classical Contrastive",
"Divergence (CD) algorithm (Hinton, 2002) approximates",
"We define |s #(sd [ sd E {0, 1], d = 1, ..., D], the",
"this gradient update by changing the energy function param",
"number of non-void entries in s, so X is naturally identified",
"eter with the following stochastic approximation:",
"with (s E S : sl = D]. There is an initial state So ",
"(, O,..., ). Any trajectory from so to x eX has exactly",
"Ex~pa(x)[V(x) -Ex~gr(x*/x) VE(x)]",
"(4)",
"D actions. A choice of trajectory from so to x amounts to a",
"choice of the order in which the entries of x are assigned.",
"where qk(x'x) is the distribution obtained by using a K",
"step MCMC initialized at x to approximately sample from",
"In this setting, the forward policy Pr(s; )of a GFlowNet,",
"Po. For example, x' can be taken from a K-step Metropolis-",
"introduced in s2.1,is a distribution over all ways to select",
"Hastings chain starting at a true data sample x ~ pdata(x),",
"a position with a void entry in s and a value (O or 1) to",
"and the parameters updated with β(&(x) - Eβ(x'). As",
"assign to this entry. Thus the action space for a state s",
"K → oo, assuming mixing of MCMC chains, the distribu-",
"has size 2(D -s/).Correspondingly, the backward policy",
"tion over negative samples qk(xx) converges to po(x'),",
"P(s; O) is a distribution over the s ways to select a posi-",
"and we recover, in expectation, the true gradient (3).",
"tionwithanonvoidentryins.Weillustratethemechanism",
"of the forward and backward policies in Figure 3.",
"Tieleman (2008) later proposed persistent CD (PCD),where",
"the MCMC chains that give negative samples do not restart",
"In our experiments, we take Pe(s; 0) and PB(s: 0) to"
] |
{
"arxiv_id": "2202.01361",
"language": "en",
"page_idx": 2
}
|
||
1905.10861-en-0
|
1905.10861-en-0.jpg
|
$\textbf{Temporal Attentive Alignment for Video Domain Adaptation}$
Zsolt Kira
Min-Hung Chen
Ghassan AlRegib
Georgia Institute of Technology
$\{$cmhungsteve, zkira, alregib$\}$gatech.edu
$ଶ୍ନି$
八
## $\mathbf{Abstract}$
Although various image-based domain adaptation (DA) techniques have been proposed in recent years, domain shift
$\begin{array}{lll}&\text{in videos is still not well-explored. Most previous works}\\&\text{only evaluate performance on small-scale datasets which}\\&\text{ara saturatad Tharafora wa frest nronosa a laraars scal}\end{array}$
are saturated. Therefore, we first propose a larger-scale dataset with larger domain discrepancy: UCF-HMDB full. Second, we investigate different DA integration methods for
$\sum_{\begin{array}{cc}{second,we~mvesigarent~DA~mlegranton~methous~jor}\\&videos,~and~show~that~simultaneously~aligning~and~learning}\\&temporal~dynamics~achieves~effective~alignment~even~with-\\&out~sonhisticated~DA~methods~Finallv~we~nonose~Tem-~an~an~an~an~an~an~an~b$ $\begin{array}{ccc}{out~sophisticated~DA~methods.~Finally,~we~propose~Tem-}\\{\textcircled{0}}&{poral~Attentive~Adversarial~Adaptation~Network~(TA^{3}N),}\\{0}&{which~avnlicith,~attands~the~tamnoral~do~tamanics~usina}\end{array}$
which explicitly attends to the temporal dynamics using domain discrepancy for more effective domain alignment, achieving state-of-the-art performance on three video DA datasets. The code and data are released at http: // github. com/ cmhungsteve/ TA 3N.
## $1. \textbf{Introduction}$
Unsupervised Domain adaptation $(DA)$ [1] has been studied extensively in recent years to address the domain shift problem [9], which means the models trained on source labeled datasets do not generalize well to target datasets and tasks, without access to any target labels. While many DA approaches are able to diminish the distribution gap between source and target domains while learning discriminative deep features [2,8,7,10], most methods have been developed only for images and not videos.
Furthermore, unlike image-based DA work, there do not exist well-organized datasets to evaluate and benchmark the performance of DA algorithms for videos. The most common datasets are $UCF-Olympic$ and $UCF-HMDB_{small}$ [12, 14, 4], which have only a few overlapping categories between source and target domains. This introduces limited domain discrepancy so that a deep CNN architecture can achieve nearly perfect performance even without any DA method (details in Section 3.2 and Table 2). Therefore, we propose a larger-scale video DA dataset, $UCF-HMDB_{full}$, by collecting all relevant and overlapping categories be-
<FigureHere>
Figure 1: An overview of proposed TA$^3$N for video DA. In addition to spatial discrepancy between frame images, videos also suffer from temporal discrepancy between sets of time-ordered frames that contain multiple local temporal dynamics with different contributions to the overall domain shift, as indicated by the thickness of green dashed arrows. Therefore, we propose to focus on aligning the temporal dynamics which have higher domain discrepancy using a learned attention mechanism to effectively align the temporal-embedded feature spaces for videos. Here we use the action basketball as the example.
tween UCF101[11] and HMDB51[6], leading to three times larger than previous two datasets, and contains larger domain discrepancy (details in Section 3.2 and Table 3).
Videos can suffer from domain discrepancy along both the spatial and temporal directions, bringing the need of alignment for embedded feature spaces along both directions, as shown in Figure l. However, most DA approaches have not explicitly addressed the domain shift problem in the temporal direction. Therefore, we first investigate different DA integration methods for video classification and
1
|
[
[
0.1829833984375,
0.1337890625,
0.78759765625,
0.1337890625,
0.78759765625,
0.1483154296875,
0.1829833984375,
0.1483154296875
],
[
0.24267578125,
0.18310546875,
0.375732421875,
0.18310546875,
0.375732421875,
0.1976318359375,
0.24267578125,
0.1976318359375
],
[
0.439453125,
0.18310546875,
0.52392578125,
0.18310546875,
0.52392578125,
0.1976318359375,
0.439453125,
0.1976318359375
],
[
0.5888671875,
0.18310546875,
0.72802734375,
0.18310546875,
0.72802734375,
0.1976318359375,
0.5888671875,
0.1976318359375
],
[
0.359375,
0.2008056640625,
0.609375,
0.2008056640625,
0.609375,
0.21533203125,
0.359375,
0.21533203125
],
[
0.310546875,
0.2197265625,
0.66015625,
0.2197265625,
0.66015625,
0.2342529296875,
0.310546875,
0.2342529296875
],
[
0.0196075439453125,
0.26953125,
0.058013916015625,
0.26953125,
0.060455322265625,
0.369384765625,
0.0228729248046875,
0.369384765625
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.57861328125,
0.272705078125,
0.63623046875,
0.272705078125,
0.63623046875,
0.28857421875,
0.57861328125,
0.28857421875
],
[
0.76318359375,
0.2734375,
0.81787109375,
0.2734375,
0.81787109375,
0.2890625,
0.76318359375,
0.2890625
],
[
0.098876953125,
0.3017578125,
0.46728515625,
0.3017578125,
0.46728515625,
0.31494140625,
0.098876953125,
0.31494140625
],
[
0.564453125,
0.3037109375,
0.62744140625,
0.30615234375,
0.62744140625,
0.317626953125,
0.564453125,
0.31494140625
],
[
0.775390625,
0.304931640625,
0.83837890625,
0.304931640625,
0.83837890625,
0.31640625,
0.775390625,
0.31640625
],
[
0.08087158203125,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.5703125,
0.32763671875,
0.59619140625,
0.330810546875,
0.59326171875,
0.3447265625,
0.56787109375,
0.341552734375
],
[
0.8037109375,
0.32763671875,
0.81396484375,
0.32763671875,
0.81396484375,
0.330810546875,
0.8037109375,
0.330810546875
],
[
0.6611328125,
0.3447265625,
0.734375,
0.3447265625,
0.734375,
0.35595703125,
0.6611328125,
0.35595703125
],
[
0.080078125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.080078125,
0.3466796875
],
[
0.080078125,
0.347900390625,
0.4697265625,
0.347900390625,
0.4697265625,
0.361083984375,
0.080078125,
0.361083984375
],
[
0.8095703125,
0.35107421875,
0.81884765625,
0.35107421875,
0.81884765625,
0.354736328125,
0.8095703125,
0.354736328125
],
[
0.080078125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.080078125,
0.376953125
],
[
0.65625,
0.359130859375,
0.74169921875,
0.359130859375,
0.74169921875,
0.373779296875,
0.65625,
0.373779296875
],
[
0.02777099609375,
0.369384765625,
0.052276611328125,
0.369384765625,
0.052276611328125,
0.38818359375,
0.02777099609375,
0.38818359375
],
[
0.611328125,
0.37255859375,
0.6396484375,
0.37255859375,
0.6396484375,
0.386962890625,
0.611328125,
0.386962890625
],
[
0.833984375,
0.375732421875,
0.84716796875,
0.373779296875,
0.8486328125,
0.37890625,
0.8349609375,
0.380615234375
],
[
0.7685546875,
0.3818359375,
0.78369140625,
0.3818359375,
0.78369140625,
0.389404296875,
0.7685546875,
0.389404296875
],
[
0.080078125,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.406005859375,
0.080078125,
0.406005859375
],
[
0.07843017578125,
0.376953125,
0.46728515625,
0.37744140625,
0.46728515625,
0.392578125,
0.07843017578125,
0.391357421875
],
[
0.02532958984375,
0.397705078125,
0.056365966796875,
0.397705078125,
0.056365966796875,
0.44580078125,
0.02532958984375,
0.44580078125
],
[
0.080078125,
0.40771484375,
0.468017578125,
0.40771484375,
0.468017578125,
0.421630859375,
0.080078125,
0.421630859375
],
[
0.07843017578125,
0.42236328125,
0.468017578125,
0.42236328125,
0.468017578125,
0.436279296875,
0.07843017578125,
0.436279296875
],
[
0.57861328125,
0.425537109375,
0.6298828125,
0.425537109375,
0.6298828125,
0.4375,
0.57861328125,
0.4375
],
[
0.76318359375,
0.425537109375,
0.82177734375,
0.425537109375,
0.82177734375,
0.4375,
0.76318359375,
0.4375
],
[
0.07843017578125,
0.438232421875,
0.46728515625,
0.438232421875,
0.46728515625,
0.451904296875,
0.07843017578125,
0.451904296875
],
[
0.5908203125,
0.438232421875,
0.615234375,
0.438232421875,
0.615234375,
0.45068359375,
0.5908203125,
0.45068359375
],
[
0.7587890625,
0.44140625,
0.78173828125,
0.44580078125,
0.77783203125,
0.458251953125,
0.7548828125,
0.453857421875
],
[
0.79150390625,
0.4462890625,
0.80224609375,
0.4462890625,
0.80224609375,
0.45068359375,
0.79150390625,
0.45068359375
],
[
0.67333984375,
0.449462890625,
0.71826171875,
0.449462890625,
0.71826171875,
0.4609375,
0.67333984375,
0.4609375
],
[
0.0767822265625,
0.45263671875,
0.46728515625,
0.451904296875,
0.46728515625,
0.466552734375,
0.0767822265625,
0.46728515625
],
[
0.03350830078125,
0.443115234375,
0.0555419921875,
0.443115234375,
0.0555419921875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.619140625,
0.45703125,
0.63818359375,
0.45703125,
0.63818359375,
0.46533203125,
0.619140625,
0.46533203125
],
[
0.75244140625,
0.460205078125,
0.77392578125,
0.460205078125,
0.77392578125,
0.473388671875,
0.75244140625,
0.473388671875
],
[
0.80615234375,
0.4609375,
0.8203125,
0.4609375,
0.8203125,
0.47021484375,
0.80615234375,
0.47021484375
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.6650390625,
0.462158203125,
0.7255859375,
0.462158203125,
0.7255859375,
0.472900390625,
0.6650390625,
0.472900390625
],
[
0.080078125,
0.484130859375,
0.46728515625,
0.484130859375,
0.46728515625,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.78173828125,
0.481689453125,
0.80126953125,
0.481689453125,
0.80126953125,
0.4892578125,
0.78173828125,
0.4892578125
],
[
0.6435546875,
0.49365234375,
0.7490234375,
0.49365234375,
0.7490234375,
0.50830078125,
0.6435546875,
0.50830078125
],
[
0.0228729248046875,
0.48486328125,
0.05963134765625,
0.48486328125,
0.05963134765625,
0.6318359375,
0.0228729248046875,
0.6318359375
],
[
0.080078125,
0.498779296875,
0.468017578125,
0.498779296875,
0.468017578125,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.07843017578125,
0.5146484375,
0.468017578125,
0.5146484375,
0.468017578125,
0.52783203125,
0.07843017578125,
0.52783203125
],
[
0.080078125,
0.52880859375,
0.35302734375,
0.52783203125,
0.35302734375,
0.54248046875,
0.080078125,
0.54345703125
],
[
0.50341796875,
0.53857421875,
0.88818359375,
0.53857421875,
0.88818359375,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.5009765625,
0.55322265625,
0.890625,
0.5537109375,
0.890625,
0.568359375,
0.5009765625,
0.5673828125
],
[
0.08251953125,
0.57373046875,
0.20751953125,
0.57373046875,
0.20751953125,
0.58837890625,
0.08251953125,
0.58837890625
],
[
0.50341796875,
0.57080078125,
0.890625,
0.57080078125,
0.890625,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.50244140625,
0.583984375,
0.8896484375,
0.583984375,
0.8896484375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.1004638671875,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.61279296875,
0.1004638671875,
0.61279296875
],
[
0.50341796875,
0.60107421875,
0.8896484375,
0.60107421875,
0.8896484375,
0.611328125,
0.50341796875,
0.611328125
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50341796875,
0.6162109375,
0.890625,
0.6162109375,
0.890625,
0.626953125,
0.50341796875,
0.626953125
],
[
0.02777099609375,
0.62744140625,
0.0555419921875,
0.62744140625,
0.0555419921875,
0.66650390625,
0.02777099609375,
0.66650390625
],
[
0.03350830078125,
0.625,
0.037567138671875,
0.625,
0.037567138671875,
0.63330078125,
0.03350830078125,
0.63330078125
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.080078125,
0.64599609375,
0.4697265625,
0.64599609375,
0.4697265625,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.50244140625,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.6591796875,
0.50244140625,
0.6591796875
],
[
0.0228729248046875,
0.6572265625,
0.0555419921875,
0.65576171875,
0.058013916015625,
0.70263671875,
0.02532958984375,
0.7041015625
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.6748046875,
0.46728515625,
0.6748046875,
0.46728515625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.08087158203125,
0.69140625,
0.46728515625,
0.69140625,
0.46728515625,
0.7021484375,
0.08087158203125,
0.7021484375
],
[
0.50244140625,
0.68994140625,
0.74609375,
0.68994140625,
0.74609375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.36279296875,
0.72021484375,
0.36279296875,
0.734375,
0.080078125,
0.734375
],
[
0.1004638671875,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.1004638671875,
0.74755859375
],
[
0.50244140625,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.5009765625,
0.74853515625,
0.890625,
0.75,
0.890625,
0.7646484375,
0.5009765625,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.7646484375,
0.86767578125,
0.7646484375,
0.86767578125,
0.77783203125,
0.5009765625,
0.77783203125
],
[
0.0767822265625,
0.77783203125,
0.4697265625,
0.77978515625,
0.4697265625,
0.79736328125,
0.0767822265625,
0.79541015625
],
[
0.521484375,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.08251953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08251953125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.8935546875,
0.796875,
0.8935546875,
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,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
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,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.5009765625,
0.85595703125,
0.890625,
0.85498046875,
0.890625,
0.869140625,
0.5009765625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.484375,
0.92529296875,
0.491943359375,
0.9287109375,
0.486083984375,
0.9345703125,
0.4794921875,
0.93115234375
]
] |
[
"Temporal Attentive Alignment for Video Domain Adaptation",
"Min-Hung Chen",
"Zsolt Kira",
"GhassanAlRegib",
"Georgia Institute of Technology",
"{cmhungsteve,zkira,alregib)@gatech.edu",
"610z un",
"Abstract",
"Source",
"Target",
"Although various image-based domain adaptation (DA)",
"Temporal",
"Temporal",
"techniques have been proposed in recent years, domain shift",
"",
"",
"Temporal",
"in videos is still not well-explored. Most previous works",
"only evaluate performance on small-scale datasets which",
"",
"are saturated. Therefore, we first propose a larger-scale",
"Alignment",
"7",
"牌",
"",
"",
"Second, we investigate different DA integration methods for",
"dataset with larger domain discrepancy: UCF-HMDBfull.",
"LA",
"videos, and show that simultaneously aligning and learning",
"temporal dynamics achieves effective alignment even with-",
"Spatial",
"Spatial",
"out sophisticated DA methods. Finally, we propose Tem-",
"",
"",
"",
"Spatial",
"poral Attentive Adversarial Adaptation Network (TA3N),",
"'SO",
"",
"店",
"中",
"which explicitly attends to the temporal dynamics using",
"Alignment",
"domain discrepancy for more effective domain alignment,",
"",
"Domain Shift",
"SA1980S061",
"achieving state-of-the-art performance on three video DA",
"datasets. The code and data are released at http://",
"github.com/cmhungsteve/TA3N.",
"Figure 1: An overview of proposed TA3N for video DA",
"In addition to spatial discrepancy between frame images,",
"1. Introduction",
"videosalsosufferfromtemporaldiscrepancybetweensets",
"of time-ordered frames that contain multiple local tempo-",
"Unsupervised Domain adaptation (DA) [I] has been",
"ral dynamics withdifferent contributionstothe overall do-",
"studied extensively in recent years to address the domain",
"main shift, as indicated by the thickness of green dashed",
"[:A!",
"",
"shift problem [9], which means the models trained on source",
"arrows. Therefore, we propose to focus on aligning the tem-",
"labeled datasets do not generalize well to target datasets",
"poral dynamics which have higher domain discrepancy us-",
"!Xe",
"and tasks, without access to any target labels. While many",
"ing a learned attention mechanism to effectively align the",
"DA approaches are able to diminish the distribution gap be-",
"temporal-embedded feature spaces for videos. Here we use",
"tween source and target domains while learning discrimi-",
"the action basketball as the example.",
"native deep features [2, 8, 7, 10], most methods have been",
"developed only for images and not videos.",
"Furthermore.unlikeimage-based DA work.there donot",
"tween UCF101[1I]andHMDB51[6],leadingto three",
"exist well-organized datasets to evaluate and benchmark the",
"times larger than previous two datasets, and contains larger",
"performance of DA algorithms for videos. The most com-",
"domain discrepancy (details in Section 3.2 and Table 3)",
"mon datasets are UCF-Olympic and UCF-HMDBsmalu [12,",
"Videos can suffer from domain discrepancy along both",
"14, 4l, which have only a few overlapping categories be-",
"the spatial and temporal directions, bringing the need of",
"tween source and target domains. This introduces limited",
"alignment for embedded feature spaces along both direc-",
"domain discrepancy so that a deep CNN architecture can",
"tions, as shown in Figure I. However, most DA approaches",
"achieve nearly perfect performance even without any DA",
"have not explicitly addressed the domain shift problem in",
"method (details in Section 3.2 and Table 2). Therefore, we",
"the temporal direction. Therefore, we first investigate dif-",
"propose a larger-scale video DA dataset, UCF-HMDB full,",
"ferent DA integration methods for video classification and",
"by collecting all relevant and overlapping categories be-",
"propose Temporal Adversarial Adaptation Network (TA?N)",
"1"
] |
{
"arxiv_id": "1905.10861",
"language": "en",
"page_idx": 0
}
|
||
1905.10861-en-1
|
1905.10861-en-1.jpg
|
<FigureHere>
Figure 3: The domain attention mechanism in TA$^{3}$N.
Thicker arrows corresponds to larger attention weights.
which simultaneously aligns and learns temporal dynamics, to effectively align domains spatio-temporally, outperforming other approaches which naively apply more sophisticated image-based DA methods for videos. To further achieve more effective temporal alignment, we propose to focus more on aligning those which have higher contribution to the overall domain shift, such as the local temporal features connected by thicker green arrows shown in Figure l. Therefore, we propose Temporal Attentive Adversarial Adaptation Network (TA$^3$N) to explicitly attend to the temporal dynamics by taking into account the domain distribution discrepancy, achieving state-of-the-art performance on all three investigated video DA datasets. (GitHub)
# $2. \textbf{ Technical Approach}$
Here we introduce our baseline model (Section 2.1) and
proposed methods for video DA (Sections 2.2 and 2.3).
## $2. 1. \textbf{Baseline Model}$
Given the recent success of large-scale video classification using CNNs [5], we build our baseline on such architectures, as shown in the lower part of Figure 2.
<FigureHere>
Figure 2: Baseline architecture (TemPooling) with the adversarial discriminators $\hat{G}_{sd}$ and $\hat{G}_{td}.\mathcal{L}_{y}$ is the class prediction loss, and $\mathcal{L}_{sd}$ and $\mathcal{L}_{td}$ are the domain losses.
We first feed the frame-level feature representations extracted from ResNet[3] pre-trained on ImageNet into our model, which can be divided into two parts: 1) Spatial mod- $uleG_{sf}$, which consists of multilayer perceptrons (MLP) that aims to convert the general-purpose feature vectors into task-driven feature vectors, where the task is video classification in this paper; 2) Temporal module $G_{tf}$, which aggregates the frame-level feature representations to form a single video-level feature representation for each video. Our baseline architecture conducts mean-pooling along the temporal direction in $G_{tf}$, so we note it as TemPooling.
To address domain shift, we integrate TemPooling with $Adversarial\textit{ Discriminator }\hat{G} _{d}$,which is the combination of a gradient reversal layer (GRL) and a domain classifier $G_d$, inspired by[2]. Through adversarial training, $G_d$ is optimized to discriminate data across domains, while the feature generator is optimized to gradually align the feature distributions across domains. $\hat{G}_d$ is integrated in two ways:
$1)\hat{G}_{sd}\colon$ show how directly applying image-based DA approaches can benefit video DA; 2) $\hat{G}_{td}\colon$ indicate how DA on temporal-dynamics-encoded features benefits video DA.
$2. 2. \textbf{ Integration of Temporal Dynamics with DA}$
One main drawback of directly integrating image-based DA approaches into TemPooling is that the relation between frames is missing. Therefore, we would like to address this question: Does the video DA problem benefit from encoding temporal dynamics into features ?
Given the fact that humans can recognize actions by reasoning the observations across time, we propose the TemRelation architecture by replacing the temporal pooling mechanism in $G_{tf}$ with the Temporal Relation module, which is modified from[15], as shown in Figure 4. Specifically, We fuse the time-ordered feature representations with MLPs into $n$-frame relation $R_n$, and then sum up all $R_n$ into the final video representation to capture temporal relations at multiple time scales. To align and encode temporal dynamics simultaneously instead of solely modifying the Temporal module, we propose Temporal Adversarial Adaptation Network (TA$^2\mathbf{N})$, which integrates relation discriminators $\hat{G}_{rd}^{n}$ inside the Temporal module with corresponding $n$-frame relations to properly align different temporal relations, outperforming those which are extended from sophisticated image-based DA approaches although TA$^2$N is adopted from a simpler DA method (details in Table 3).
$2. 3. \textbf{ Temporal Attentive Alignment for Videos}$
Although aligning temporal features across domains benefits video DA, not all the features equally contribute to the overall domain shift. Therefore, we propose to focus more on aligning those which have larger domain discrepancy and assign them with larger attention weighting values, as shown in Figure 3. The main question becomes: How to incorporate domain discrepancy for attention?
To address this, we propose Temporal Attentive Adversarial Adaptation Network(TA$^3\mathbf{N})$, as shown in Figure 4, by introducing the domain attention mechanism, which assigns higher domain attention values to $n$-frame relation
2
|
[
[
0.08087158203125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.5390625,
0.091552734375,
0.583984375,
0.091552734375,
0.583984375,
0.10162353515625,
0.5390625,
0.10162353515625
],
[
0.81396484375,
0.0921630859375,
0.8544921875,
0.0921630859375,
0.8544921875,
0.10418701171875,
0.81396484375,
0.10418701171875
],
[
0.07843017578125,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.52294921875,
0.11297607421875,
0.5390625,
0.11614990234375,
0.53515625,
0.126953125,
0.5185546875,
0.12371826171875
],
[
0.58642578125,
0.10797119140625,
0.59228515625,
0.10797119140625,
0.59228515625,
0.1142578125,
0.58642578125,
0.1142578125
],
[
0.85693359375,
0.11553955078125,
0.8759765625,
0.11553955078125,
0.8759765625,
0.1307373046875,
0.85693359375,
0.1307373046875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.080078125,
0.136962890625
],
[
0.85693359375,
0.1287841796875,
0.8759765625,
0.1287841796875,
0.8759765625,
0.1458740234375,
0.85693359375,
0.1458740234375
],
[
0.07843017578125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.07843017578125,
0.1514892578125
],
[
0.6640625,
0.1395263671875,
0.72607421875,
0.1395263671875,
0.72607421875,
0.15087890625,
0.6640625,
0.15087890625
],
[
0.52294921875,
0.1458740234375,
0.53759765625,
0.1458740234375,
0.53759765625,
0.154052734375,
0.52294921875,
0.154052734375
],
[
0.8544921875,
0.141357421875,
0.8759765625,
0.141357421875,
0.8759765625,
0.1622314453125,
0.8544921875,
0.1622314453125
],
[
0.07843017578125,
0.1514892578125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1673583984375,
0.07843017578125,
0.1666259765625
],
[
0.66259765625,
0.15283203125,
0.73193359375,
0.15283203125,
0.73193359375,
0.1641845703125,
0.66259765625,
0.1641845703125
],
[
0.51953125,
0.160400390625,
0.53759765625,
0.160400390625,
0.53759765625,
0.169189453125,
0.51953125,
0.169189453125
],
[
0.85693359375,
0.1572265625,
0.8759765625,
0.1572265625,
0.8759765625,
0.17236328125,
0.85693359375,
0.17236328125
],
[
0.08087158203125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.5185546875,
0.17236328125,
0.54150390625,
0.17236328125,
0.54150390625,
0.186279296875,
0.5185546875,
0.186279296875
],
[
0.8544921875,
0.17041015625,
0.87841796875,
0.17041015625,
0.87841796875,
0.1881103515625,
0.8544921875,
0.1881103515625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.74267578125,
0.1787109375,
0.76123046875,
0.1787109375,
0.76123046875,
0.186279296875,
0.74267578125,
0.186279296875
],
[
0.6337890625,
0.189453125,
0.650390625,
0.189453125,
0.650390625,
0.1925048828125,
0.6337890625,
0.1925048828125
],
[
0.740234375,
0.186279296875,
0.76708984375,
0.186279296875,
0.76708984375,
0.1944580078125,
0.740234375,
0.1944580078125
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.56640625,
0.203857421875,
0.62353515625,
0.203857421875,
0.62353515625,
0.2139892578125,
0.56640625,
0.2139892578125
],
[
0.84228515625,
0.20263671875,
0.8896484375,
0.20263671875,
0.8896484375,
0.21337890625,
0.84228515625,
0.21337890625
],
[
0.77392578125,
0.2052001953125,
0.82666015625,
0.2052001953125,
0.82666015625,
0.212158203125,
0.77392578125,
0.212158203125
],
[
0.50634765625,
0.2052001953125,
0.5498046875,
0.2052001953125,
0.5498046875,
0.212158203125,
0.50634765625,
0.212158203125
],
[
0.080078125,
0.21337890625,
0.46728515625,
0.21337890625,
0.46728515625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.51318359375,
0.21337890625,
0.54736328125,
0.21337890625,
0.54736328125,
0.222900390625,
0.51318359375,
0.222900390625
],
[
0.576171875,
0.212158203125,
0.61376953125,
0.212158203125,
0.61376953125,
0.222900390625,
0.576171875,
0.222900390625
],
[
0.779296875,
0.2139892578125,
0.81640625,
0.2139892578125,
0.81640625,
0.2215576171875,
0.779296875,
0.2215576171875
],
[
0.8505859375,
0.21337890625,
0.8857421875,
0.21337890625,
0.8857421875,
0.2203369140625,
0.8505859375,
0.2203369140625
],
[
0.07843017578125,
0.2266845703125,
0.4697265625,
0.2279052734375,
0.4697265625,
0.242431640625,
0.07843017578125,
0.2412109375
],
[
0.5009765625,
0.2342529296875,
0.8896484375,
0.23291015625,
0.8896484375,
0.2474365234375,
0.5009765625,
0.248779296875
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50244140625,
0.250732421875,
0.86279296875,
0.250732421875,
0.86279296875,
0.264404296875,
0.50244140625,
0.264404296875
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.07843017578125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.07843017578125,
0.287353515625
],
[
0.50341796875,
0.290283203125,
0.890625,
0.292236328125,
0.890625,
0.306884765625,
0.50341796875,
0.304931640625
],
[
0.08087158203125,
0.304931640625,
0.268798828125,
0.304931640625,
0.268798828125,
0.3193359375,
0.08087158203125,
0.3193359375
],
[
0.50341796875,
0.30810546875,
0.89208984375,
0.30810546875,
0.89208984375,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.50244140625,
0.323974609375,
0.890625,
0.323974609375,
0.890625,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.098876953125,
0.330810546875,
0.468017578125,
0.330810546875,
0.468017578125,
0.3447265625,
0.098876953125,
0.3447265625
],
[
0.080078125,
0.3466796875,
0.441162109375,
0.3466796875,
0.441162109375,
0.35986328125,
0.080078125,
0.35986328125
],
[
0.50341796875,
0.347900390625,
0.86767578125,
0.347900390625,
0.86767578125,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.080078125,
0.373779296875,
0.2303466796875,
0.373779296875,
0.2303466796875,
0.386962890625,
0.080078125,
0.386962890625
],
[
0.521484375,
0.369384765625,
0.890625,
0.369384765625,
0.890625,
0.38330078125,
0.521484375,
0.38330078125
],
[
0.50341796875,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.50341796875,
0.3984375
],
[
0.098876953125,
0.396484375,
0.468017578125,
0.396484375,
0.468017578125,
0.410888671875,
0.098876953125,
0.410888671875
],
[
0.50244140625,
0.40087890625,
0.89208984375,
0.40087890625,
0.89208984375,
0.4140625,
0.50244140625,
0.4140625
],
[
0.07843017578125,
0.41357421875,
0.468017578125,
0.410888671875,
0.468017578125,
0.4248046875,
0.07843017578125,
0.4267578125
],
[
0.5009765625,
0.416748046875,
0.89208984375,
0.415283203125,
0.89208984375,
0.4287109375,
0.5009765625,
0.429931640625
],
[
0.080078125,
0.427978515625,
0.39208984375,
0.427978515625,
0.39208984375,
0.44140625,
0.080078125,
0.44140625
],
[
0.5009765625,
0.43115234375,
0.7255859375,
0.43115234375,
0.7255859375,
0.4443359375,
0.5009765625,
0.4443359375
],
[
0.521484375,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.521484375,
0.460205078125
],
[
0.50341796875,
0.462158203125,
0.89208984375,
0.462158203125,
0.89208984375,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.1502685546875,
0.477783203125,
0.16259765625,
0.477783203125,
0.16259765625,
0.48095703125,
0.1502685546875,
0.48095703125
],
[
0.50341796875,
0.4765625,
0.89208984375,
0.4765625,
0.89208984375,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.214111328125,
0.48486328125,
0.2418212890625,
0.48486328125,
0.2418212890625,
0.49365234375,
0.214111328125,
0.49365234375
],
[
0.2115478515625,
0.496826171875,
0.2442626953125,
0.496826171875,
0.2442626953125,
0.50390625,
0.2115478515625,
0.50390625
],
[
0.50341796875,
0.4912109375,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.50341796875,
0.5048828125
],
[
0.5048828125,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.51806640625,
0.5048828125,
0.51806640625
],
[
0.374267578125,
0.52099609375,
0.40185546875,
0.52099609375,
0.40185546875,
0.5244140625,
0.374267578125,
0.5244140625
],
[
0.400390625,
0.51953125,
0.449462890625,
0.51953125,
0.449462890625,
0.53515625,
0.400390625,
0.53515625
],
[
0.50341796875,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.256591796875,
0.5234375,
0.2607421875,
0.5234375,
0.2607421875,
0.52978515625,
0.256591796875,
0.52978515625
],
[
0.1478271484375,
0.52294921875,
0.1968994140625,
0.52294921875,
0.1968994140625,
0.5341796875,
0.1478271484375,
0.5341796875
],
[
0.50341796875,
0.537109375,
0.89208984375,
0.537109375,
0.89208984375,
0.55029296875,
0.50341796875,
0.55029296875
],
[
0.50341796875,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.08087158203125,
0.56494140625,
0.46728515625,
0.56494140625,
0.46728515625,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.50341796875,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.5791015625,
0.50341796875,
0.5791015625
],
[
0.080078125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.5947265625,
0.080078125,
0.5947265625
],
[
0.50244140625,
0.58203125,
0.89208984375,
0.58203125,
0.89208984375,
0.59521484375,
0.50244140625,
0.59521484375
],
[
0.080078125,
0.5966796875,
0.42236328125,
0.5966796875,
0.42236328125,
0.60986328125,
0.080078125,
0.60986328125
],
[
0.50244140625,
0.5966796875,
0.89208984375,
0.5966796875,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.50341796875,
0.6123046875,
0.89208984375,
0.6123046875,
0.89208984375,
0.62548828125,
0.50341796875,
0.62548828125
],
[
0.1004638671875,
0.62744140625,
0.46728515625,
0.62744140625,
0.46728515625,
0.6416015625,
0.1004638671875,
0.6416015625
],
[
0.50244140625,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
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,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.6591796875,
0.8896484375,
0.6591796875,
0.8896484375,
0.66943359375,
0.50341796875,
0.66943359375
],
[
0.07843017578125,
0.673828125,
0.468017578125,
0.67236328125,
0.468017578125,
0.6875,
0.07843017578125,
0.68798828125
],
[
0.50244140625,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.080078125,
0.69140625,
0.468017578125,
0.69140625,
0.468017578125,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.50244140625,
0.6875,
0.865234375,
0.6875,
0.865234375,
0.70068359375,
0.50244140625,
0.70068359375
],
[
0.07843017578125,
0.7041015625,
0.46728515625,
0.7041015625,
0.46728515625,
0.71728515625,
0.07843017578125,
0.71728515625
],
[
0.5009765625,
0.708984375,
0.85546875,
0.71044921875,
0.85546875,
0.72802734375,
0.5009765625,
0.7265625
],
[
0.080078125,
0.7197265625,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.52294921875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.52294921875,
0.74853515625
],
[
0.08087158203125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.07843017578125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.07843017578125,
0.77880859375
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.07843017578125,
0.77978515625,
0.418212890625,
0.77978515625,
0.418212890625,
0.79443359375,
0.07843017578125,
0.79443359375
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.098876953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.098876953125,
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,
0.810546875,
0.82470703125,
0.810546875,
0.8388671875,
0.5009765625,
0.84033203125
],
[
0.07843017578125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.07843017578125,
0.85498046875
],
[
0.521484375,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.07843017578125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.07843017578125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
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
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"which simultaneously aligns and learns temporal dynam-",
"Source",
"Target",
"ics, to effectively align domains spatio-temporally, outper-",
"福",
"S",
"",
"forming other approaches which naively apply more sophis-",
"厦",
"ticated image-based DA methods for videos.To further",
"Temporal",
"车",
"巨",
"achieve more effective temporal alignment,we propose to",
"Alignment",
"物",
"高",
"focusmore onaligningthose whichhavehighercontribu-",
"祥",
"百",
"tion to the overall domain shift, such as the local temporal",
"videc",
"atur",
"feature",
"features connected by thicker green arrows shown in Fig-",
"local temporal",
"frame-level",
"ocal temporal",
"frame-level",
"ure 1. Therefore, we propose Temporal Attentive Adver-",
"features",
"features",
"features",
"features",
"sarial Adaptation Network (TA3N) to explicitly attend to",
"Figure 3: The domain attention mechanism in TA3N.",
"the temporal dynamics by taking into account the domain",
"Thicker arrows corresponds to larger attention weights.",
"distribution discrepancy, achieving state-of-the-art perfor-",
"mance on all three investigated video DA datasets. (GitHub)",
"1) Gsd: show how directly applying image-based DA ap-",
"2. Technical Approach",
"proaches can benefit video DA; 2) Gtd: indicate how DA",
"on temporal-dynamics-encoded features benefits video DA.",
"Here we introduce our baseline model (Section 2.1) and",
"proposed methods for video DA (Sections 2.2 and 2.3).",
"2.2. Integration of Temporal Dynamics with DA",
"2.1. Baseline Model",
"One main drawback of directly integrating image-based",
"DA approaches into TemPooling is that the relation between",
"Given the recent success of large-scale video classifica-",
"frames is missing. Therefore, we would like to address this",
"tion using CNNs [5], we build our baseline on such archi-",
"question:Does the video DA problem benefit from encoding",
"tectures, as shown in the lower part of Figure 2.",
"temporal dynamics into features?",
"Given the fact that humans can recognize actions by rea-",
"soning the observations across time, we propose the TemRe-",
"anr",
"lation architecture by replacing the temporal pooling mech-",
"Frame",
"feature",
"anism in Gtf with the Temporal Relation module, which is",
"modifiedfrom[15l,asshowninFigure4.Specifically.We",
"Telace",
"",
"fuse the time-ordered feature representations with MLPs",
"",
"ConvNet",
"into n-frame relation Rn, and then sum up all Rn into the",
"final video representation to capture temporal relations at",
"Figure 2: Baseline architecture (TemPooling) with the ad-",
"multipletime scales.Toalignand encodetemporal dynam-",
"versarial discriminators Gsd and Gtd. Cy is the class pre-",
"ics simultaneously instead of solely modifying the Tempo-",
"diction loss, and Lsd and Ltd are the domain losses.",
"ral module, we propose Temporal Adversarial Adapta-",
"tion Network (TA2N), which integrates relation discrim-",
"We first feed the frame-level feature representations ex-",
"inators Gnd inside the Temporal module with correspond-",
"tracted from ResNet [3] pre-trained on ImageNet into our",
"ing n-frame relations to properly align different temporal",
"model, which can be divided into two parts: 1) Spatial mod-",
"relations,outperformingthosewhichareextendedfromso-",
"ule Gsf,which consists of multilayer perceptrons (MLP)",
"phisticated image-based DA approaches although TA2N is",
"that aims toconvertthe general-purposefeature vectors into",
"adopted from a simpler DA method (details in Table 3).",
"task-driven feature vectors, where the task is video classifi-",
"2.3. Temporal Attentive Alignment for Videos",
"cation in this paper; 2) Temporal module Gtf, which aggre-",
"gates the frame-level feature representations to form a sin-",
"Although aligning temporal features across domains",
"gle video-level feature representation for each video. Our",
"benefits video DA, not all the features equally contribute",
"baseline architecture conducts mean-pooling along the tem-",
"to the overall domain shift. Therefore, we propose to focus",
"poral direction in Gtf, so we note it as TemPooling.",
"more on aligning those whichhavelarger domain discrep-",
"To address domain shift, we integrate TemPooling with",
"ancy and assign them with larger attention weighting values,",
"Adversarial Discriminator Gd, which is the combination of",
"as shown in Figure 3. The main question becomes: How to",
"a gradient reversal layer (GRL) and a domain classifier Gd,",
"incorporate domain discrepancy for attention?",
"inspired by [2]. Through adversarial training, Ga is opti-",
"To address this, we propose Temporal Attentive Adver-",
"mized to discriminate data across domains, while the fea-",
"sarial Adaptation Network (TA3N), as shown in Figure 4,",
"ture generator is optimized to gradually align the feature",
"by introducing the domain attention mechanism, which as-",
"distributions across domains. Gd is integrated in two ways:",
"signs higher domain attention values to n-frame relation",
"2"
] |
{
"arxiv_id": "1905.10861",
"language": "en",
"page_idx": 1
}
|
||
1905.10861-en-2
|
1905.10861-en-2.jpg
|
<FigureHere>
Figure 4: The overall architecture of the proposed TA$^3$N. In the temporal relation module, time-ordered frames are used to generate $K$-1 n-frame relation feature representations $\boldsymbol R=\{R_2,...,R_K\}.$ The attentive entropy loss $\mathcal{L}_ae$, which is calculated by domain entropy $H(\hat{d})$ and class entropy $H(\hat{y})$, aims to enhance the certainty of those videos that are more similar across domains.
<table>
<tbody>
<tr>
<th> </th>
<th>UCF- HMDB small</th>
<th>UCF- Olympic</th>
<th>$UCF-HMDB$ $ful.$</th>
</tr>
<tr>
<td>length (sec.) </td>
<td>1-21</td>
<td>1-39</td>
<td>$1\cdot33$</td>
</tr>
<tr>
<td>resolution</td>
<td colspan="3">TICH varv x'lat</td>
</tr>
<tr>
<td>class #</td>
<td>5</td>
<td>$\overline{6}$</td>
<td>12</td>
</tr>
<tr>
<td>video #</td>
<td>1171</td>
<td>1145</td>
<td>3209</td>
</tr>
</tbody>
</table>
Table 1: The comparison of the video DA datasets.
<table>
<tbody>
<tr>
<th>Source $\rightarrow$ $\textbf{Target}$</th>
<th>$U$ $\rightarrow$ 0</th>
<th>$0\to$ $U$</th>
<th>$U\to$ $H$</th>
<th>$H$ $[\rightarrow$ $U$</th>
</tr>
<tr>
<td>W. Sultani et al.[12]</td>
<td>33.33</td>
<td>47.91</td>
<td>68.70</td>
<td>68.67</td>
</tr>
<tr>
<td>T.Xu et al. [14]</td>
<td>87.00</td>
<td>75.00</td>
<td>82.00</td>
<td>82.00</td>
</tr>
<tr>
<td>AMLS $\mathbf{S} \left ( \mathrm{GFK}\right ) [ 4\mathbf{T}$</td>
<td>84.65</td>
<td>86.44</td>
<td>89.53</td>
<td>95.36</td>
</tr>
<tr>
<td>AMLS$\left ( \mathrm{SA}\right ) [ 4]$</td>
<td>83.92</td>
<td>86.07</td>
<td>90.25</td>
<td>94.40</td>
</tr>
<tr>
<td>DAAA[ 4] </td>
<td>91.60</td>
<td>89.96</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>TemPooling</td>
<td>96.30</td>
<td>87.08</td>
<td>98.67</td>
<td>97.35</td>
</tr>
<tr>
<td>TemPooling+RevGrad [2]</td>
<td>98.15</td>
<td>90.00</td>
<td>99.33</td>
<td>98.41</td>
</tr>
<tr>
<td>$\overline{\mathrm{Ours~}(\mathrm{TA}^{2}\mathrm{N})}$</td>
<td>98.15</td>
<td>91.67</td>
<td>99.33</td>
<td>99.47</td>
</tr>
<tr>
<td>Ours $( \mathrm{TA}^{3}$N)</td>
<td>98.15</td>
<td>92.92</td>
<td>99.33</td>
<td>99.47</td>
</tr>
</tbody>
</table>
Table 2: The accuracy $(\%)$ comparison on UCF-Olympic and UCF-HMDB$_small$ (U: UCF, O: Olympic, H: HMDB).
features with lower domain entropy, which correspond to higher domains discrepancy. Morevoer, we minimize the class entropy for the videos with low domain discrepancy to refine the classifier adaptation.
The overall loss of TA$^3$N can be expressed as follows:
$$\begin{aligned}&\mathcal{L}=\frac{1}{N_{S}}\sum_{i=1}^{N_{S}}\mathcal{L}_{y}^{i}+\frac{1}{N_{S\cup T}}\sum_{i=1}^{N_{S\cup T}}\gamma\mathcal{L}_{ae}^{i}\\&-\:\frac{1}{N_{S\cup T}}\sum_{i=1}^{N_{S\cup T}}(\lambda^{s}\mathcal{L}_{sd}^{i}+\lambda^{r}\mathcal{L}_{rd}^{i}+\lambda^{t}\mathcal{L}_{td}^{i})\end{aligned}$$
(1)
where $N_S$ is the total number of source data, $N_{S\cup T}$ equals the number of all data. $i$ represents the $i$-th video. $\lambda^s,\lambda^r$ and $\lambda^t$ is the trade-off weighting for each domain loss $\mathcal{L}_sd$, $\mathcal{L}_{rd}$ and $\mathcal{L}_td.~\gamma$ is the weighting for the attentive entropy loss $\mathcal{L}_ae.$ All the weightings are chosen via grid search.
Our proposed TA$^{3}$N and TADA[13] both utilize entropy functions for attention but with different perspectives. TADA aims to focus on the foreground objects for image DA, while TA$^3$N aims to find important and discriminative parts of temporal dynamics to align for video DA.
## $3. \textbf{Experiments}$
We evaluate on three datasets: UCF-Olympic, UCFHMDB$_{small}$ and UCF-HMDB$_full$, as shown in Table 1. The adaptation setting is noted as “Source → Target”.
# 3.1. Datasets and Setup
UCF-Olympic and UCF-HMDB$_{small}:$ First, we evaluate our approaches on two publicly used datasets, UCFOlympic and UCF-HMDB $_small$, and compare with all other works that also evaluate on these two datasets[12,14,4].
UCF-HMDB$_full:$ To ensure large domain discrepancy to investigate DA approaches, we build UCF-HMDB$_full$, which is around 3 times larger than UCF-HMDB$_small$ and UCF-Olympic, as shown in Table 1. To compare with image-based DA approaches, we extend several state-ofthe-art methods [2,8,7,10] for video DA with our TemPooling and TemRelation architectures, as shown in Table 3. The difference between the“Target only”and Source only" settings is the domain used for training.
3.2. Experimental Results and Discussion
UCF-Olympic and UCF-HMDB$_{small}:$ In these two datasets, our approach outperforms all the previous methods, as shown in Table 2 (e.g. at least 9% absolute difference on “U→H”). These results also show that the performance on these datasets is saturated. With a strong CNN as the backbone architecture, even our baseline architecture (TemPooling) can achieve high accuracy without any DA method. This suggests that these two datasets are not enough to evaluate more sophisticated DA approaches, so larger-scale datasets for video DA are needed.
$\mathbf{UCF-HMDB}_{full}{:}$ It is worth noting that the“Source only” accuracy on UCF-HMDB $_{full}$ is much lower than UCFHMDB$_small$, which implies much larger domain discrepancy in UCF-HMDB$_full. \textbf{We now answer the question in}$ Section 2.2: Does the video DA problem benefit from encoding temporal dynamics into features? (details in Table 3)
For the same DA method, TemRelation outperforms TemPooling in most cases, especially for the gain value (e.g. on“U$\to$H”,“TemRelation+RevGrad”reaches 2.77% absolute accuracy gain while “TemPooling+RevGrad” only reaches $0.83\%$ gain). By simultaneously aligning and encoding temporal dynamics, TA$^2$N outperforms other approaches which are extended from more sophisticated
3
|
[
[
0.59326171875,
0.08709716796875,
0.6708984375,
0.08837890625,
0.6708984375,
0.0997314453125,
0.59326171875,
0.0985107421875
],
[
0.70166015625,
0.0889892578125,
0.77392578125,
0.0889892578125,
0.77392578125,
0.0997314453125,
0.70166015625,
0.0997314453125
],
[
0.7890625,
0.08587646484375,
0.87158203125,
0.0889892578125,
0.87158203125,
0.10101318359375,
0.7890625,
0.09783935546875
],
[
0.2091064453125,
0.10101318359375,
0.285888671875,
0.10101318359375,
0.285888671875,
0.10479736328125,
0.2091064453125,
0.10479736328125
],
[
0.51123046875,
0.0997314453125,
0.57666015625,
0.0997314453125,
0.57666015625,
0.10986328125,
0.51123046875,
0.10986328125
],
[
0.63134765625,
0.10101318359375,
0.654296875,
0.10101318359375,
0.654296875,
0.10919189453125,
0.63134765625,
0.10919189453125
],
[
0.72802734375,
0.10101318359375,
0.7548828125,
0.10101318359375,
0.7548828125,
0.10919189453125,
0.72802734375,
0.10919189453125
],
[
0.8203125,
0.10101318359375,
0.84716796875,
0.10101318359375,
0.84716796875,
0.10797119140625,
0.8203125,
0.10797119140625
],
[
0.6044921875,
0.10919189453125,
0.86767578125,
0.10919189453125,
0.86767578125,
0.12249755859375,
0.6044921875,
0.12249755859375
],
[
0.52294921875,
0.111083984375,
0.57421875,
0.111083984375,
0.57421875,
0.121826171875,
0.52294921875,
0.121826171875
],
[
0.3349609375,
0.1060791015625,
0.35302734375,
0.1060791015625,
0.35302734375,
0.10986328125,
0.3349609375,
0.10986328125
],
[
0.20751953125,
0.11236572265625,
0.240234375,
0.11236572265625,
0.240234375,
0.1300048828125,
0.20751953125,
0.1300048828125
],
[
0.529296875,
0.12060546875,
0.56787109375,
0.12060546875,
0.56787109375,
0.1307373046875,
0.529296875,
0.1307373046875
],
[
0.63818359375,
0.12371826171875,
0.64453125,
0.12371826171875,
0.64453125,
0.1275634765625,
0.63818359375,
0.1275634765625
],
[
0.73388671875,
0.12371826171875,
0.74267578125,
0.12371826171875,
0.74267578125,
0.1287841796875,
0.73388671875,
0.1287841796875
],
[
0.82666015625,
0.11932373046875,
0.84228515625,
0.12249755859375,
0.83837890625,
0.1307373046875,
0.82275390625,
0.1275634765625
],
[
0.348876953125,
0.1300048828125,
0.37744140625,
0.1300048828125,
0.37744140625,
0.141357421875,
0.348876953125,
0.141357421875
],
[
0.62841796875,
0.1300048828125,
0.65283203125,
0.1300048828125,
0.65283203125,
0.1427001953125,
0.62841796875,
0.1427001953125
],
[
0.72607421875,
0.1307373046875,
0.75244140625,
0.1307373046875,
0.75244140625,
0.1427001953125,
0.72607421875,
0.1427001953125
],
[
0.81884765625,
0.1307373046875,
0.84619140625,
0.1307373046875,
0.84619140625,
0.1427001953125,
0.81884765625,
0.1427001953125
],
[
0.529296875,
0.1319580078125,
0.56787109375,
0.1319580078125,
0.56787109375,
0.1427001953125,
0.529296875,
0.1427001953125
],
[
0.0931396484375,
0.1363525390625,
0.1192626953125,
0.14013671875,
0.1168212890625,
0.147705078125,
0.0906982421875,
0.143310546875
],
[
0.236083984375,
0.1319580078125,
0.2607421875,
0.1319580078125,
0.2607421875,
0.14013671875,
0.236083984375,
0.14013671875
],
[
0.146240234375,
0.141357421875,
0.156005859375,
0.141357421875,
0.156005859375,
0.1458740234375,
0.146240234375,
0.1458740234375
],
[
0.16015625,
0.1395263671875,
0.176513671875,
0.1427001953125,
0.1732177734375,
0.1483154296875,
0.1568603515625,
0.1458740234375
],
[
0.1993408203125,
0.143310546875,
0.2900390625,
0.143310546875,
0.2900390625,
0.154052734375,
0.1993408203125,
0.154052734375
],
[
0.398681640625,
0.146484375,
0.4306640625,
0.1483154296875,
0.428955078125,
0.1590576171875,
0.39794921875,
0.1572265625
],
[
0.1478271484375,
0.154052734375,
0.1748046875,
0.1578369140625,
0.17236328125,
0.1654052734375,
0.146240234375,
0.1622314453125
],
[
0.1422119140625,
0.147705078125,
0.1690673828125,
0.14453125,
0.17236328125,
0.1546630859375,
0.1446533203125,
0.1578369140625
],
[
0.52783203125,
0.154052734375,
0.86376953125,
0.1546630859375,
0.86376953125,
0.169189453125,
0.52783203125,
0.1685791015625
],
[
0.1478271484375,
0.169189453125,
0.1707763671875,
0.169189453125,
0.1707763671875,
0.18115234375,
0.1478271484375,
0.18115234375
],
[
0.365234375,
0.1634521484375,
0.39453125,
0.1654052734375,
0.393798828125,
0.177978515625,
0.363525390625,
0.175537109375
],
[
0.214111328125,
0.1717529296875,
0.2459716796875,
0.1717529296875,
0.2459716796875,
0.179931640625,
0.214111328125,
0.179931640625
],
[
0.09234619140625,
0.186279296875,
0.1192626953125,
0.1912841796875,
0.11517333984375,
0.20263671875,
0.0882568359375,
0.1976318359375
],
[
0.53759765625,
0.1849365234375,
0.63134765625,
0.1875,
0.63134765625,
0.1988525390625,
0.53759765625,
0.1962890625
],
[
0.66650390625,
0.186279296875,
0.72607421875,
0.186279296875,
0.72607421875,
0.1976318359375,
0.66650390625,
0.1976318359375
],
[
0.72216796875,
0.186279296875,
0.77783203125,
0.186279296875,
0.77783203125,
0.1976318359375,
0.72216796875,
0.1976318359375
],
[
0.7900390625,
0.1881103515625,
0.8056640625,
0.1881103515625,
0.8056640625,
0.1956787109375,
0.7900390625,
0.1956787109375
],
[
0.80615234375,
0.1881103515625,
0.833984375,
0.1881103515625,
0.833984375,
0.1962890625,
0.80615234375,
0.1962890625
],
[
0.84619140625,
0.1881103515625,
0.86669921875,
0.1881103515625,
0.86669921875,
0.1962890625,
0.84619140625,
0.1962890625
],
[
0.86376953125,
0.1881103515625,
0.890625,
0.1881103515625,
0.890625,
0.1956787109375,
0.86376953125,
0.1956787109375
],
[
0.221435546875,
0.1932373046875,
0.2459716796875,
0.1932373046875,
0.2459716796875,
0.2052001953125,
0.221435546875,
0.2052001953125
],
[
0.1478271484375,
0.1925048828125,
0.1707763671875,
0.1925048828125,
0.1707763671875,
0.205810546875,
0.1478271484375,
0.205810546875
],
[
0.361083984375,
0.1932373046875,
0.398681640625,
0.1932373046875,
0.398681640625,
0.203857421875,
0.361083984375,
0.203857421875
],
[
0.529296875,
0.1976318359375,
0.6396484375,
0.199462890625,
0.6396484375,
0.2108154296875,
0.529296875,
0.208984375
],
[
0.67724609375,
0.1988525390625,
0.7138671875,
0.1988525390625,
0.7138671875,
0.210205078125,
0.67724609375,
0.210205078125
],
[
0.734375,
0.1988525390625,
0.771484375,
0.1988525390625,
0.771484375,
0.210205078125,
0.734375,
0.210205078125
],
[
0.7939453125,
0.1988525390625,
0.8310546875,
0.1988525390625,
0.8310546875,
0.210205078125,
0.7939453125,
0.210205078125
],
[
0.8505859375,
0.1988525390625,
0.88818359375,
0.1988525390625,
0.88818359375,
0.210205078125,
0.8505859375,
0.210205078125
],
[
0.6787109375,
0.210205078125,
0.71337890625,
0.210205078125,
0.71337890625,
0.222900390625,
0.6787109375,
0.222900390625
],
[
0.73779296875,
0.2083740234375,
0.77294921875,
0.210205078125,
0.771484375,
0.222900390625,
0.736328125,
0.2203369140625
],
[
0.7958984375,
0.210205078125,
0.8310546875,
0.210205078125,
0.8310546875,
0.222900390625,
0.7958984375,
0.222900390625
],
[
0.8544921875,
0.210205078125,
0.8896484375,
0.210205078125,
0.8896484375,
0.222900390625,
0.8544921875,
0.222900390625
],
[
0.6787109375,
0.2215576171875,
0.71337890625,
0.2215576171875,
0.71337890625,
0.2342529296875,
0.6787109375,
0.2342529296875
],
[
0.5400390625,
0.2108154296875,
0.63232421875,
0.2108154296875,
0.63232421875,
0.222900390625,
0.5400390625,
0.222900390625
],
[
0.5390625,
0.222900390625,
0.63134765625,
0.222900390625,
0.63134765625,
0.2342529296875,
0.5390625,
0.2342529296875
],
[
0.736328125,
0.222900390625,
0.771484375,
0.222900390625,
0.771484375,
0.2342529296875,
0.736328125,
0.2342529296875
],
[
0.7958984375,
0.222900390625,
0.8310546875,
0.222900390625,
0.8310546875,
0.2342529296875,
0.7958984375,
0.2342529296875
],
[
0.85302734375,
0.222900390625,
0.8896484375,
0.222900390625,
0.8896484375,
0.2342529296875,
0.85302734375,
0.2342529296875
],
[
0.146240234375,
0.2298583984375,
0.1707763671875,
0.2298583984375,
0.1707763671875,
0.242431640625,
0.146240234375,
0.242431640625
],
[
0.2132568359375,
0.2298583984375,
0.2459716796875,
0.2298583984375,
0.2459716796875,
0.242431640625,
0.2132568359375,
0.242431640625
],
[
0.67724609375,
0.2342529296875,
0.71337890625,
0.2342529296875,
0.71337890625,
0.246826171875,
0.67724609375,
0.246826171875
],
[
0.736328125,
0.2342529296875,
0.771484375,
0.2342529296875,
0.771484375,
0.246826171875,
0.736328125,
0.246826171875
],
[
0.7939453125,
0.231689453125,
0.83251953125,
0.2342529296875,
0.8310546875,
0.25,
0.79345703125,
0.2474365234375
],
[
0.8544921875,
0.2342529296875,
0.89208984375,
0.2342529296875,
0.89208984375,
0.2474365234375,
0.8544921875,
0.2474365234375
],
[
0.55615234375,
0.246826171875,
0.61376953125,
0.246826171875,
0.61376953125,
0.25830078125,
0.55615234375,
0.25830078125
],
[
0.54345703125,
0.23486328125,
0.62841796875,
0.23486328125,
0.62841796875,
0.246826171875,
0.54345703125,
0.246826171875
],
[
0.67724609375,
0.246826171875,
0.71337890625,
0.246826171875,
0.71337890625,
0.25830078125,
0.67724609375,
0.25830078125
],
[
0.736328125,
0.246826171875,
0.77294921875,
0.246826171875,
0.77294921875,
0.25830078125,
0.736328125,
0.25830078125
],
[
0.67724609375,
0.2587890625,
0.71337890625,
0.2587890625,
0.71337890625,
0.271484375,
0.67724609375,
0.271484375
],
[
0.736328125,
0.25830078125,
0.771484375,
0.260009765625,
0.76953125,
0.271484375,
0.734375,
0.26953125
],
[
0.7958984375,
0.2587890625,
0.830078125,
0.2587890625,
0.830078125,
0.271484375,
0.7958984375,
0.271484375
],
[
0.8544921875,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.270751953125,
0.8544921875,
0.270751953125
],
[
0.080078125,
0.267578125,
0.46728515625,
0.266357421875,
0.46728515625,
0.281005859375,
0.080078125,
0.2822265625
],
[
0.55224609375,
0.260009765625,
0.619140625,
0.260009765625,
0.619140625,
0.271484375,
0.55224609375,
0.271484375
],
[
0.51123046875,
0.270751953125,
0.65869140625,
0.270751953125,
0.65869140625,
0.2841796875,
0.51123046875,
0.2841796875
],
[
0.6787109375,
0.270751953125,
0.71142578125,
0.270751953125,
0.71142578125,
0.28271484375,
0.6787109375,
0.28271484375
],
[
0.79736328125,
0.270751953125,
0.830078125,
0.270751953125,
0.830078125,
0.28271484375,
0.79736328125,
0.28271484375
],
[
0.8544921875,
0.26953125,
0.88818359375,
0.271484375,
0.88623046875,
0.2841796875,
0.85302734375,
0.2822265625
],
[
0.080078125,
0.28271484375,
0.468017578125,
0.28271484375,
0.468017578125,
0.296630859375,
0.080078125,
0.296630859375
],
[
0.73779296875,
0.271484375,
0.771484375,
0.271484375,
0.771484375,
0.28271484375,
0.73779296875,
0.28271484375
],
[
0.5498046875,
0.285888671875,
0.62109375,
0.285888671875,
0.62109375,
0.29736328125,
0.5498046875,
0.29736328125
],
[
0.6787109375,
0.285888671875,
0.71142578125,
0.285888671875,
0.71142578125,
0.298583984375,
0.6787109375,
0.298583984375
],
[
0.736328125,
0.285888671875,
0.76953125,
0.285888671875,
0.76953125,
0.29736328125,
0.736328125,
0.29736328125
],
[
0.7958984375,
0.285888671875,
0.830078125,
0.285888671875,
0.830078125,
0.29736328125,
0.7958984375,
0.29736328125
],
[
0.8544921875,
0.285888671875,
0.88818359375,
0.285888671875,
0.88818359375,
0.29736328125,
0.8544921875,
0.29736328125
],
[
0.07843017578125,
0.29736328125,
0.468017578125,
0.298583984375,
0.468017578125,
0.3125,
0.07843017578125,
0.311279296875
],
[
0.54833984375,
0.29736328125,
0.62158203125,
0.29736328125,
0.62158203125,
0.309326171875,
0.54833984375,
0.309326171875
],
[
0.6787109375,
0.29736328125,
0.71337890625,
0.29736328125,
0.71337890625,
0.309326171875,
0.6787109375,
0.309326171875
],
[
0.736328125,
0.29736328125,
0.771484375,
0.29736328125,
0.771484375,
0.309326171875,
0.736328125,
0.309326171875
],
[
0.7958984375,
0.29736328125,
0.8310546875,
0.29736328125,
0.8310546875,
0.309326171875,
0.7958984375,
0.309326171875
],
[
0.8544921875,
0.29736328125,
0.8896484375,
0.29736328125,
0.8896484375,
0.309326171875,
0.8544921875,
0.309326171875
],
[
0.0767822265625,
0.3125,
0.4697265625,
0.314453125,
0.4697265625,
0.328857421875,
0.0767822265625,
0.326904296875
],
[
0.50244140625,
0.322509765625,
0.89208984375,
0.323974609375,
0.89208984375,
0.33837890625,
0.50244140625,
0.337158203125
],
[
0.080078125,
0.330078125,
0.4697265625,
0.330078125,
0.4697265625,
0.3447265625,
0.080078125,
0.3447265625
],
[
0.50244140625,
0.339111328125,
0.8857421875,
0.339111328125,
0.8857421875,
0.354248046875,
0.50244140625,
0.354248046875
],
[
0.080078125,
0.34521484375,
0.471435546875,
0.34521484375,
0.471435546875,
0.359130859375,
0.080078125,
0.359130859375
],
[
0.08087158203125,
0.3623046875,
0.324462890625,
0.3623046875,
0.324462890625,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.50341796875,
0.3818359375,
0.890625,
0.3818359375,
0.890625,
0.396484375,
0.50341796875,
0.396484375
],
[
0.498291015625,
0.392578125,
0.89453125,
0.39453125,
0.89453125,
0.415283203125,
0.498291015625,
0.41357421875
],
[
0.08087158203125,
0.401611328125,
0.4697265625,
0.401611328125,
0.4697265625,
0.415283203125,
0.08087158203125,
0.415283203125
],
[
0.5,
0.4091796875,
0.8935546875,
0.410888671875,
0.8935546875,
0.4287109375,
0.5,
0.4267578125
],
[
0.080078125,
0.417236328125,
0.468017578125,
0.417236328125,
0.468017578125,
0.43115234375,
0.080078125,
0.43115234375
],
[
0.50341796875,
0.427978515625,
0.89208984375,
0.427978515625,
0.89208984375,
0.44140625,
0.50341796875,
0.44140625
],
[
0.08087158203125,
0.434326171875,
0.46728515625,
0.434326171875,
0.46728515625,
0.4443359375,
0.08087158203125,
0.4443359375
],
[
0.07843017578125,
0.447509765625,
0.298095703125,
0.447509765625,
0.298095703125,
0.4609375,
0.07843017578125,
0.4609375
],
[
0.50244140625,
0.443115234375,
0.890625,
0.443115234375,
0.890625,
0.45703125,
0.50244140625,
0.45703125
],
[
0.50341796875,
0.458251953125,
0.62158203125,
0.458251953125,
0.62158203125,
0.4716796875,
0.50341796875,
0.4716796875
],
[
0.61376953125,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.61376953125,
0.4716796875
],
[
0.50244140625,
0.473388671875,
0.8935546875,
0.473388671875,
0.8935546875,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.098876953125,
0.462158203125,
0.455078125,
0.462158203125,
0.455078125,
0.47607421875,
0.098876953125,
0.47607421875
],
[
0.2017822265625,
0.48486328125,
0.2254638671875,
0.48486328125,
0.2254638671875,
0.49755859375,
0.2017822265625,
0.49755859375
],
[
0.3056640625,
0.48291015625,
0.340576171875,
0.4873046875,
0.338134765625,
0.5,
0.30224609375,
0.49560546875
],
[
0.281005859375,
0.492431640625,
0.289306640625,
0.492431640625,
0.289306640625,
0.501953125,
0.281005859375,
0.501953125
],
[
0.50244140625,
0.488037109375,
0.8935546875,
0.488037109375,
0.8935546875,
0.501953125,
0.50244140625,
0.501953125
],
[
0.2034912109375,
0.49560546875,
0.25244140625,
0.498779296875,
0.250732421875,
0.51123046875,
0.2017822265625,
0.50830078125
],
[
0.338134765625,
0.49560546875,
0.37744140625,
0.5,
0.373291015625,
0.51513671875,
0.3349609375,
0.51123046875
],
[
0.1805419921875,
0.49365234375,
0.19287109375,
0.490478515625,
0.19775390625,
0.50048828125,
0.1854248046875,
0.50390625
],
[
0.1519775390625,
0.49755859375,
0.1641845703125,
0.50244140625,
0.1568603515625,
0.51123046875,
0.1446533203125,
0.505859375
],
[
0.310546875,
0.49755859375,
0.334228515625,
0.49755859375,
0.334228515625,
0.5087890625,
0.310546875,
0.5087890625
],
[
0.261474609375,
0.505859375,
0.306396484375,
0.5087890625,
0.3056640625,
0.52099609375,
0.2607421875,
0.517578125
],
[
0.50244140625,
0.501953125,
0.859375,
0.50244140625,
0.859375,
0.517578125,
0.50244140625,
0.5166015625
],
[
0.2034912109375,
0.5146484375,
0.22216796875,
0.5146484375,
0.22216796875,
0.521484375,
0.2034912109375,
0.521484375
],
[
0.17724609375,
0.5068359375,
0.2034912109375,
0.5068359375,
0.2034912109375,
0.51806640625,
0.17724609375,
0.51806640625
],
[
0.313720703125,
0.5146484375,
0.334228515625,
0.5146484375,
0.334228515625,
0.521484375,
0.313720703125,
0.521484375
],
[
0.449462890625,
0.51806640625,
0.471435546875,
0.51806640625,
0.471435546875,
0.5322265625,
0.449462890625,
0.5322265625
],
[
0.20751953125,
0.52587890625,
0.2418212890625,
0.52783203125,
0.240234375,
0.5390625,
0.2059326171875,
0.537109375
],
[
0.1805419921875,
0.5322265625,
0.191162109375,
0.5322265625,
0.191162109375,
0.54248046875,
0.1805419921875,
0.54248046875
],
[
0.50341796875,
0.52783203125,
0.8203125,
0.52978515625,
0.8203125,
0.54443359375,
0.50341796875,
0.54248046875
],
[
0.2091064453125,
0.53515625,
0.404296875,
0.537109375,
0.404296875,
0.5546875,
0.2091064453125,
0.55322265625
],
[
0.1641845703125,
0.54541015625,
0.2034912109375,
0.548828125,
0.2017822265625,
0.56103515625,
0.16259765625,
0.55810546875
],
[
0.5009765625,
0.5498046875,
0.8935546875,
0.5517578125,
0.8935546875,
0.5693359375,
0.5009765625,
0.5673828125
],
[
0.2156982421875,
0.55615234375,
0.233642578125,
0.55615234375,
0.233642578125,
0.56298828125,
0.2156982421875,
0.56298828125
],
[
0.5009765625,
0.5673828125,
0.890625,
0.568359375,
0.890625,
0.58349609375,
0.5009765625,
0.58203125
],
[
0.080078125,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.080078125,
0.5888671875,
0.46728515625,
0.5888671875,
0.46728515625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.0767822265625,
0.60302734375,
0.468017578125,
0.60400390625,
0.468017578125,
0.61865234375,
0.0767822265625,
0.6171875
],
[
0.50244140625,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.08251953125,
0.619140625,
0.468017578125,
0.619140625,
0.468017578125,
0.63330078125,
0.08251953125,
0.63330078125
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.6337890625,
0.44287109375,
0.6337890625,
0.44287109375,
0.64794921875,
0.080078125,
0.64794921875
],
[
0.50341796875,
0.630859375,
0.8896484375,
0.630859375,
0.8896484375,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.1004638671875,
0.6494140625,
0.468017578125,
0.6494140625,
0.468017578125,
0.66357421875,
0.1004638671875,
0.66357421875
],
[
0.50244140625,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.5009765625,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.5009765625,
0.671875
],
[
0.08087158203125,
0.66650390625,
0.46728515625,
0.66650390625,
0.46728515625,
0.6767578125,
0.08087158203125,
0.6767578125
],
[
0.07843017578125,
0.67822265625,
0.4697265625,
0.6787109375,
0.4697265625,
0.693359375,
0.07843017578125,
0.6923828125
],
[
0.50244140625,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.080078125,
0.6943359375,
0.4697265625,
0.6943359375,
0.4697265625,
0.70849609375,
0.080078125,
0.70849609375
],
[
0.50244140625,
0.689453125,
0.80615234375,
0.689453125,
0.80615234375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.080078125,
0.71044921875,
0.406982421875,
0.71044921875,
0.406982421875,
0.7236328125,
0.080078125,
0.7236328125
],
[
0.50341796875,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.5009765625,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.736328125,
0.5009765625,
0.736328125
],
[
0.08087158203125,
0.7373046875,
0.20751953125,
0.7373046875,
0.20751953125,
0.751953125,
0.08087158203125,
0.751953125
],
[
0.50244140625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.50244140625,
0.75048828125,
0.89208984375,
0.75048828125,
0.89208984375,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.1004638671875,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.77783203125,
0.1004638671875,
0.77783203125
],
[
0.50244140625,
0.76513671875,
0.890625,
0.76513671875,
0.890625,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.0767822265625,
0.77587890625,
0.4697265625,
0.7763671875,
0.4697265625,
0.79443359375,
0.0767822265625,
0.79345703125
],
[
0.50244140625,
0.78076171875,
0.8759765625,
0.78076171875,
0.8759765625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.08087158203125,
0.79443359375,
0.431396484375,
0.79443359375,
0.431396484375,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
0.521484375,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.080078125,
0.81640625,
0.2607421875,
0.8173828125,
0.2607421875,
0.83203125,
0.080078125,
0.83056640625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.50341796875,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.07843017578125,
0.8369140625,
0.471435546875,
0.8388671875,
0.471435546875,
0.85693359375,
0.07843017578125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.5009765625,
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
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.455810546875,
0.88720703125,
0.455810546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"UCF-HMDBam",
"UCF-Olympic",
"UCF-HMDBful",
"nain-Aentionhlork",
"length (sec.)",
"1-21",
"1-39",
"1-.33",
"UCF:320×240/Olympic:vary/HMDB:vary×240",
"resolution",
"te",
"Gra",
"class#",
"5",
"6",
"12",
"H(a)",
"1171",
"1145",
"3209",
"video#",
"Gsd",
"H(d)",
"ten",
"oora",
"DomainAttention",
"Lae",
"module",
"Relatio",
"Table 1: The comparison of the video DA datasets.",
"R2",
"H()",
"A2",
"Gsf",
"Source→Target",
"U-→0",
"1O-U",
"U",
"→H",
"H",
"-U",
"A3",
"R3",
"class pre",
"W. Sultani et al. [12]",
"33.33",
"47.91",
"68.70",
"68.67",
"87.00",
"75.00",
"82.00",
"82.00",
"84.65",
"T. Xu et al. [14]",
"AMLS (GFK) [4)",
"86.44",
"89.53",
"95.36",
"RK",
"AK",
"83.92",
"86.07",
"90.25",
"94.40",
"DAAA[4]",
"AMLS (SA) [4]",
"91.60",
"89.96",
"96.30",
"87.08",
"98.67",
"97.35",
"Figure 4: The overall architecture of the proposed TA3N.",
"TemPooling",
"TemPooling + RevGrad [2]",
"98.15",
"99.33",
"98.41",
"In the temporal relation module, time-ordered frames are",
"90.00",
"Ours (TA?N)",
"98.15",
"91.67",
"99.33",
"99.47",
"used to generate K-1 n-frame relation feature representa-",
"Ours (TA3N)",
"98.15",
"92.92",
"99.33",
"99.47",
"tions R - {R2,.., Rk]. The attentive entropy loss Cae,",
"Table 2:The accuracy (%)comparison on UCF-Olympic",
"which is calculated by domain entropy H(d) and class en-",
"and UCF-HMDB small (U: UCF, O: Olympic, H: HMDB).",
"tropy H(g), aims to enhance the certainty of those videos",
"that are more similar across domains.",
"UCF-HMDBfull: To ensure large domain discrepancy",
"to investigate DA approaches, we build UCF-HMDB f ull,",
"features with lower domain entropy, which correspond to",
"which is around 3 times larger than UCF-HMDB smalu and",
"higher domains discrepancy. Morevoer, we minimize the",
"UCF-Olympic, as shown in Table 1. To compare with",
"class entropyforthe videos with lowdomaindiscrepancy",
"to refine the classifier adaptation.",
"image-based DA approaches, we extend several state-of-",
"the-art methods",
"[2, 8, 7,10] for video DA with our",
"TemPooling and TemRelation architectures, as shown in",
"The overall loss of TA3N can be expressed as follows:",
"Ns",
"NSUT",
"1",
"Table 3.The difference between the \"Target only\" and",
"ci",
"Cae",
"1",
"C",
"",
"NSUT",
"\"Source only\" settings is the domain used for training.",
"1",
"Ns.",
"1=1",
"(1)",
"NSUT",
"1",
"3.2. Experimental Results and Discussion",
" ('Csd +X\"Crd+X'cia)",
"NSUT",
"UCF-Olympic and UCF-HMDBsmall: In these two",
"i=1",
"datasets, our approach outperforms all the previous meth-",
"where Ns is the total number of source data, NsuT equals",
"the number of all data. i represents the i-th video. s, Ar",
"ods, as shown in Table 2 (e.g. at least 9% absolute differ-",
"and At is the trade-off weighting for each domain loss Lsd,",
"ence on \"U → H\"). These results also show that the perfor-",
"Crd and Ltd. is the weighting for the attentive entropy",
"mance on these datasets is saturated. With a strong CNN",
"loss Lae. All the weightings are chosen via grid search.",
"as the backbone architecture, even our baseline architec-",
"Our proposed TA3N and TADA [13] both utilize en-",
"ture (TemPooling) can achieve high accuracy without any",
"DA method. This suggests that these two datasets are not",
"tropyfunctionsforattentionbutwithdifferentperspectives.",
"TADA aims to focus on the foreground objects for image",
"enough to evaluate more sophisticated DA approaches, so",
"DA, while TA3N aims to find important and discriminative",
"larger-scale datasets for video DA are needed.",
"parts of temporal dynamics to align for video DA.",
"UCF-HMDB ful: It is worth noting that the \"Source only\"",
"accuracy on UCF-HMDBfull is much lower than UCF-",
"3.Experiments",
"HMDBsmall, which implies much larger domain discrep-",
"ancy in UCF-HMDB full. We now answer the question in",
"We evaluate on three datasets: UCF-Olympic, UCF-",
"Section 2.2: Does the video DA problem benefit from encod-",
"HMDBsmall and UCF-HMDBfull, as shown in Table 1.",
"ing temporal dynamics into features? (details in Table 3)",
"The adaptation setting is noted as \"Source → Target\"",
"For the same DA method, TemRelation outperforms",
"3.1. Datasets and Setup",
"TemPooling in most cases, especially for the gain value",
"(e.g. on \"U → H\", \"TemRelation+RevGrad\" reaches 2.77%",
"UCF-Olympic and UCF-HMDBsmall: First, we evalu-",
"absolute accuracy gain while \"TemPooling+RevGrad\"only",
"ate our approaches on two publicly used datasets, UCF-",
"reaches 0.83% gain). By simultaneously aligning and",
"Olympic and UCF-HMDBsmall, and compare with all other",
"encoding temporal dynamics, TA?N outperforms other",
"works that also evaluate on these two datasets [12, 14, 4].",
"approaches which are extended from more sophisticated",
"3"
] |
{
"arxiv_id": "1905.10861",
"language": "en",
"page_idx": 2
}
|
||
2003.13911-en-0
|
2003.13911-en-0.jpg
|
$\textbf{Proxy Anchor Loss for Deep Metric Learning}$
Sungyeon Kim
Minsu Cho
Suha Kwak
Dongwon Kim
$\bar{\text{POSTECH, Pohang, Korea}}$
$\{$tjddus9597, kdwon, mscho, suha.kwak$\}$@postech.ac.kr
## Abstract
Existing metric learning losses can be categorized into two classes: pair- based and proxy- based losses. The former class can leverage fine- grained semantic relations between $datapoints,butslowsconvergenceingeneralduetoitshigh$ training $complexity. In$ $contrast, the$ latter class enables fast and reliable convergence, but cannot consider the rich datato-data relations. This paper presents a new proxy-based loss that takes advantages of both pair- and proxy-based methods and overcomes their limitations. Thanks to the use of proxies, our loss boosts the speed of convergence and is robust against noisy labels and outliers. At the same time, it allows embedding vectors of data to interact with each other through its gradients to exploit data-to-data relations. Our method is evaluated on four public benchmarks, where a standard network trained with our loss achieves state-ofthe-art performance and most quickly converges.
## $1. \textbf{Introduction}$
Learning a semantic distance metric has been a crucial step for many applications such as content-based image retrieval [14,21,27,29], face verification [18,25], person re-identification [3, 38], few-shot learning [24, 26, 30], and representation learning [14,33,41]. Following their great success in visual recognition, deep neural networks have been employed recently for metric learning. The networks are trained to project data onto an embedding space in which semantically similar data $(e.g.$, images of the same class) are closely grouped together. Such a quality of the embedding space is given mainly by loss functions used for training the networks, and most of the losses are categorized into two classes: pair-based and proxy-based.
The pair-based losses are built upon pairwise distances between data in the embedding space. A seminal example is Contrastive loss [4,9], which aims to minimize the distance between a pair of data if their class labels are identical and to separate them otherwise. Recent pair-based losses consider a group of pairwise distances to handle relations between more than two data[14,25,27,29,32,34,35,39].
<FigureHere>
Figure 1. Accuracy in Recall@ 1 versus training time on the Cars- 196[17] dataset. Note that all methods were trained with batch size of 150 on a single Titan Xp GPU. Our loss enables to achieve the highest accuracy, and converge faster than the baselines in terms of both the number of epochs and the actual training time.
These losses provide rich supervisory signals for training embedding networks by comparing data to data and examining fine-grained relations between them, $i.e.,data$-to-data relations. However, since they take a tuple of data as a unit input, the losses cause prohibitively high training complex- ${\mathrm{ity}^{1},O(M^{2})\mathrm{~or~}O(M^{3})\mathrm{~where~}M\mathrm{~is~the~number~of~training}}$ data, thus slow convergence. Furthermore, some tuples do not contribute to training or even degrade the quality of the learned embedding space. To resolve these issues, learning with the pair-based losses often entails tuple sampling techniques [10,25,37,40], which however have to be tuned by hand and may increase the risk of overfitting.
The proxy-based losses resolve the above complexity issue by introducing proxies[1,21,23]. A proxy is a representative of a subset of training data and learned as a part of the network parameters. Existing losses in this category consider each data point as an anchor, associate it with proxies instead of other images, and encourage the anchor to be close to proxies of the same class and far apart from those of different classes. Proxy-based losses reduce the training
$^{1}$The training complexity indicates the amount of computation required
to address the entire training dataset [1,6,10,23,35].
|
[
[
0.25732421875,
0.1319580078125,
0.71337890625,
0.1331787109375,
0.71337890625,
0.15087890625,
0.25732421875,
0.149658203125
],
[
0.53125,
0.179931640625,
0.62744140625,
0.1817626953125,
0.62744140625,
0.199462890625,
0.53125,
0.1976318359375
],
[
0.66650390625,
0.1817626953125,
0.7587890625,
0.1817626953125,
0.7587890625,
0.1962890625,
0.66650390625,
0.1962890625
],
[
0.374267578125,
0.184326171875,
0.494384765625,
0.184326171875,
0.494384765625,
0.1976318359375,
0.374267578125,
0.1976318359375
],
[
0.2132568359375,
0.184326171875,
0.33251953125,
0.184326171875,
0.33251953125,
0.1976318359375,
0.2132568359375,
0.1976318359375
],
[
0.37744140625,
0.1976318359375,
0.5947265625,
0.1988525390625,
0.5947265625,
0.216552734375,
0.37744140625,
0.21533203125
],
[
0.26318359375,
0.2197265625,
0.7099609375,
0.2197265625,
0.7099609375,
0.2342529296875,
0.26318359375,
0.2342529296875
],
[
0.0236968994140625,
0.2587890625,
0.058013916015625,
0.2587890625,
0.058013916015625,
0.3662109375,
0.0236968994140625,
0.3662109375
],
[
0.236083984375,
0.267578125,
0.314453125,
0.267578125,
0.314453125,
0.285400390625,
0.236083984375,
0.285400390625
],
[
0.52392578125,
0.268310546875,
0.5498046875,
0.268310546875,
0.5498046875,
0.2822265625,
0.52392578125,
0.2822265625
],
[
0.52294921875,
0.29345703125,
0.5498046875,
0.29345703125,
0.5498046875,
0.31005859375,
0.52294921875,
0.31005859375
],
[
0.1004638671875,
0.3017578125,
0.468017578125,
0.3017578125,
0.468017578125,
0.31494140625,
0.1004638671875,
0.31494140625
],
[
0.08087158203125,
0.317626953125,
0.4697265625,
0.317626953125,
0.4697265625,
0.330810546875,
0.08087158203125,
0.330810546875
],
[
0.52392578125,
0.3212890625,
0.54736328125,
0.3212890625,
0.54736328125,
0.335205078125,
0.52392578125,
0.335205078125
],
[
0.08087158203125,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.5048828125,
0.34521484375,
0.5498046875,
0.34521484375,
0.5498046875,
0.361083984375,
0.5048828125,
0.361083984375
],
[
0.66015625,
0.34521484375,
0.71142578125,
0.34521484375,
0.71142578125,
0.356689453125,
0.66015625,
0.356689453125
],
[
0.77783203125,
0.34521484375,
0.869140625,
0.34521484375,
0.869140625,
0.356689453125,
0.77783203125,
0.356689453125
],
[
0.80224609375,
0.356689453125,
0.84326171875,
0.356689453125,
0.84326171875,
0.371826171875,
0.80224609375,
0.371826171875
],
[
0.080078125,
0.347900390625,
0.4697265625,
0.347900390625,
0.4697265625,
0.361083984375,
0.080078125,
0.361083984375
],
[
0.50732421875,
0.35986328125,
0.52294921875,
0.35986328125,
0.52294921875,
0.37060546875,
0.50732421875,
0.37060546875
],
[
0.65673828125,
0.35791015625,
0.77685546875,
0.35791015625,
0.77685546875,
0.371826171875,
0.65673828125,
0.371826171875
],
[
0.0269622802734375,
0.367431640625,
0.056365966796875,
0.367431640625,
0.056365966796875,
0.395751953125,
0.0269622802734375,
0.395751953125
],
[
0.080078125,
0.363037109375,
0.4697265625,
0.363037109375,
0.4697265625,
0.376953125,
0.080078125,
0.376953125
],
[
0.525390625,
0.37255859375,
0.54833984375,
0.37255859375,
0.54833984375,
0.386474609375,
0.525390625,
0.386474609375
],
[
0.66015625,
0.37255859375,
0.708984375,
0.37255859375,
0.708984375,
0.3837890625,
0.66015625,
0.3837890625
],
[
0.8037109375,
0.37255859375,
0.84228515625,
0.37255859375,
0.84228515625,
0.3837890625,
0.8037109375,
0.3837890625
],
[
0.80224609375,
0.385009765625,
0.8447265625,
0.385009765625,
0.8447265625,
0.399658203125,
0.80224609375,
0.399658203125
],
[
0.080078125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.080078125,
0.391357421875
],
[
0.080078125,
0.392578125,
0.4697265625,
0.392578125,
0.4697265625,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.66015625,
0.386962890625,
0.74853515625,
0.386962890625,
0.74853515625,
0.3984375,
0.66015625,
0.3984375
],
[
0.525390625,
0.40087890625,
0.5498046875,
0.40087890625,
0.5498046875,
0.41357421875,
0.525390625,
0.41357421875
],
[
0.65869140625,
0.40087890625,
0.78515625,
0.40087890625,
0.78515625,
0.4140625,
0.65869140625,
0.4140625
],
[
0.80224609375,
0.399658203125,
0.84326171875,
0.399658203125,
0.84326171875,
0.4140625,
0.80224609375,
0.4140625
],
[
0.02532958984375,
0.406005859375,
0.0555419921875,
0.406005859375,
0.0555419921875,
0.449462890625,
0.02532958984375,
0.449462890625
],
[
0.08087158203125,
0.4091796875,
0.46728515625,
0.4091796875,
0.46728515625,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.66015625,
0.415283203125,
0.7236328125,
0.415283203125,
0.7236328125,
0.4267578125,
0.66015625,
0.4267578125
],
[
0.80224609375,
0.41357421875,
0.84326171875,
0.41357421875,
0.84326171875,
0.4287109375,
0.80224609375,
0.4287109375
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.521484375,
0.427978515625,
0.54736328125,
0.4248046875,
0.5498046875,
0.4375,
0.52392578125,
0.43994140625
],
[
0.6708984375,
0.43310546875,
0.69384765625,
0.43310546875,
0.69384765625,
0.4462890625,
0.6708984375,
0.4462890625
],
[
0.736328125,
0.43310546875,
0.75732421875,
0.43310546875,
0.75732421875,
0.4462890625,
0.736328125,
0.4462890625
],
[
0.80126953125,
0.43310546875,
0.82275390625,
0.43310546875,
0.82275390625,
0.4462890625,
0.80126953125,
0.4462890625
],
[
0.859375,
0.431884765625,
0.89208984375,
0.431884765625,
0.89208984375,
0.447509765625,
0.859375,
0.447509765625
],
[
0.07843017578125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.07843017578125,
0.451904296875
],
[
0.60693359375,
0.434326171875,
0.62841796875,
0.434326171875,
0.62841796875,
0.4462890625,
0.60693359375,
0.4462890625
],
[
0.03350830078125,
0.449462890625,
0.0555419921875,
0.449462890625,
0.0555419921875,
0.479248046875,
0.03350830078125,
0.479248046875
],
[
0.07843017578125,
0.45263671875,
0.46728515625,
0.45263671875,
0.46728515625,
0.466552734375,
0.07843017578125,
0.466552734375
],
[
0.6396484375,
0.448974609375,
0.78759765625,
0.448974609375,
0.78759765625,
0.46337890625,
0.6396484375,
0.46337890625
],
[
0.080078125,
0.468505859375,
0.4697265625,
0.468505859375,
0.4697265625,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.48095703125,
0.50244140625,
0.48095703125
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50244140625,
0.48095703125,
0.89208984375,
0.48095703125,
0.89208984375,
0.494384765625,
0.50244140625,
0.494384765625
],
[
0.080078125,
0.498779296875,
0.4697265625,
0.498779296875,
0.4697265625,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50244140625,
0.49560546875,
0.89208984375,
0.49560546875,
0.89208984375,
0.5087890625,
0.50244140625,
0.5087890625
],
[
0.0310516357421875,
0.50390625,
0.052276611328125,
0.50390625,
0.052276611328125,
0.52783203125,
0.0310516357421875,
0.52783203125
],
[
0.07843017578125,
0.51318359375,
0.46728515625,
0.51318359375,
0.46728515625,
0.52783203125,
0.07843017578125,
0.52783203125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.50244140625,
0.52294921875
],
[
0.07843017578125,
0.52783203125,
0.400390625,
0.52880859375,
0.400390625,
0.54345703125,
0.07843017578125,
0.54248046875
],
[
0.50244140625,
0.52294921875,
0.8798828125,
0.52294921875,
0.8798828125,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.02532958984375,
0.54150390625,
0.056365966796875,
0.54150390625,
0.056365966796875,
0.580078125,
0.02532958984375,
0.580078125
],
[
0.50341796875,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57470703125,
0.50341796875,
0.57470703125
],
[
0.03350830078125,
0.57470703125,
0.037567138671875,
0.57470703125,
0.037567138671875,
0.58203125,
0.03350830078125,
0.58203125
],
[
0.041656494140625,
0.57568359375,
0.0482177734375,
0.57568359375,
0.0482177734375,
0.5791015625,
0.041656494140625,
0.5791015625
],
[
0.08087158203125,
0.57373046875,
0.20751953125,
0.57373046875,
0.20751953125,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.50244140625,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.02532958984375,
0.5869140625,
0.056365966796875,
0.5869140625,
0.056365966796875,
0.6337890625,
0.02532958984375,
0.6337890625
],
[
0.50244140625,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.60498046875,
0.50244140625,
0.60498046875
],
[
0.09722900390625,
0.59765625,
0.468017578125,
0.5986328125,
0.468017578125,
0.61279296875,
0.09722900390625,
0.6123046875
],
[
0.50244140625,
0.60595703125,
0.89208984375,
0.60595703125,
0.89208984375,
0.619140625,
0.50244140625,
0.619140625
],
[
0.07843017578125,
0.61572265625,
0.4697265625,
0.61572265625,
0.4697265625,
0.62890625,
0.07843017578125,
0.62890625
],
[
0.50244140625,
0.62060546875,
0.89208984375,
0.62060546875,
0.89208984375,
0.6337890625,
0.50244140625,
0.6337890625
],
[
0.02777099609375,
0.626953125,
0.0555419921875,
0.626953125,
0.0555419921875,
0.64453125,
0.02777099609375,
0.64453125
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.0228729248046875,
0.63623046875,
0.056365966796875,
0.63525390625,
0.05963134765625,
0.708984375,
0.02532958984375,
0.71044921875
],
[
0.5009765625,
0.6337890625,
0.89208984375,
0.63623046875,
0.89208984375,
0.6494140625,
0.5009765625,
0.64794921875
],
[
0.080078125,
0.64453125,
0.46728515625,
0.64453125,
0.46728515625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.6513671875,
0.89208984375,
0.6513671875,
0.89208984375,
0.66552734375,
0.50244140625,
0.66552734375
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.66650390625,
0.89208984375,
0.66650390625,
0.89208984375,
0.6796875,
0.50244140625,
0.6796875
],
[
0.07843017578125,
0.6748046875,
0.468017578125,
0.673828125,
0.468017578125,
0.68798828125,
0.07843017578125,
0.689453125
],
[
0.50244140625,
0.681640625,
0.89208984375,
0.681640625,
0.89208984375,
0.69580078125,
0.50244140625,
0.69580078125
],
[
0.080078125,
0.69140625,
0.46728515625,
0.69140625,
0.46728515625,
0.7021484375,
0.080078125,
0.7021484375
],
[
0.50244140625,
0.6962890625,
0.89208984375,
0.6962890625,
0.89208984375,
0.7109375,
0.50244140625,
0.7109375
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.7119140625,
0.89208984375,
0.7119140625,
0.89208984375,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.5009765625,
0.72607421875,
0.79833984375,
0.72802734375,
0.79833984375,
0.7412109375,
0.5009765625,
0.7392578125
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.521484375,
0.7412109375,
0.890625,
0.7421875,
0.890625,
0.7568359375,
0.521484375,
0.755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75830078125,
0.89208984375,
0.75830078125,
0.89208984375,
0.771484375,
0.50244140625,
0.771484375
],
[
0.080078125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.50244140625,
0.7734375,
0.89208984375,
0.7734375,
0.89208984375,
0.787109375,
0.50244140625,
0.787109375
],
[
0.07843017578125,
0.78076171875,
0.381591796875,
0.78076171875,
0.381591796875,
0.79443359375,
0.07843017578125,
0.79443359375
],
[
0.5009765625,
0.787109375,
0.890625,
0.78857421875,
0.890625,
0.80322265625,
0.5009765625,
0.8017578125
],
[
0.098876953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.098876953125,
0.81005859375
],
[
0.50244140625,
0.8037109375,
0.890625,
0.8037109375,
0.890625,
0.8173828125,
0.50244140625,
0.8173828125
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.818359375,
0.89208984375,
0.818359375,
0.89208984375,
0.83203125,
0.50244140625,
0.83203125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.833984375,
0.89208984375,
0.833984375,
0.89208984375,
0.84765625,
0.50244140625,
0.84765625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.8486328125,
0.89208984375,
0.8486328125,
0.89208984375,
0.8623046875,
0.50244140625,
0.8623046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88525390625,
0.07843017578125,
0.88525390625
],
[
0.51708984375,
0.87353515625,
0.8935546875,
0.875,
0.8935546875,
0.8896484375,
0.51708984375,
0.88818359375
],
[
0.080078125,
0.88818359375,
0.46728515625,
0.88818359375,
0.46728515625,
0.8984375,
0.080078125,
0.8984375
],
[
0.50244140625,
0.88720703125,
0.78759765625,
0.88720703125,
0.78759765625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Proxy Anchor Loss for Deep Metric Learning",
"Minsu Cho",
"Suha Kwak",
"Dongwon Kim",
"Sungyeon Kim",
"POSTECH, Pohang, Korea",
"{tjddus9597,kdwon,mscho,suha.kwak]@postech.ac.kr",
"",
"Abstract",
"0.9",
"0.8",
"Existing metric learning losses can be categorized into",
"two classes: pair-based and proxy-based losses. The former",
"0.7",
"class can leverage fine-grained semantic relations between",
"@0.6",
"Method",
"TimePerEpoch",
"27.10s",
"data points, but slows convergence in general due to its high",
"R",
"Proxy-Anchor (Ours)",
"3",
"training complexity. In contrast, the latter class enables fast",
"0.5",
"MS [34]",
"28.43s",
"27.41s",
"and reliable convergence, but cannot consider the rich data-",
"to-data relations. This paper presents a new proxy-based",
"Proxy-NCA [21]",
"0.4",
"Semi-Hard Triplet [25]",
"29.97s",
"LAD",
"loss that takes advantages of both pair- and proxy-based",
"N-Pair [27]",
"28.41s",
"methods and overcomes their limitations. Thanks to the use",
"0.3",
"40",
"60",
"80",
"100",
"of proxies, our loss boosts the speed of convergence and is",
"20",
"'SO",
"robust against noisy labels and outliers. At the same time,",
"Training Time (Min)",
"it allows embedding vectors of data to interact with each",
"Figure 1. Accuracy in Recall@ I versus training time on the Cars-",
"other through its gradients to exploit data-to-data relations.",
"196 [17] dataset. Note that all methods were trained with batch",
"Our method is evaluated on four public benchmarks, where",
"size of 150 on a single Titan Xp GPU. Our loss enables to achieve",
"V",
"a standard network trained with our loss achieves state-of-",
"the highest accuracy, and converge faster than the baselines in",
"the-art performance and most quickly converges.",
"terms of both the number of epochs and the actual training time.",
"160",
"These losses provide rich supervisory signals for training",
"",
"",
"1. Introduction",
"embedding networks by comparing data to data and exam-",
"00",
"ining fine-grained relations between them, i.e., data-to-data",
"Learning a semantic distance metric has been a crucial",
"relations. However, since they take a tuple of data as a unit",
"step for many applications such as content-based image",
"input, the losses cause prohibitively high training complex-",
"2",
"retrieval [14, 21, 27, 29], face verification [18, 25], per-",
"ZAIIE",
"ityl.O(M2) or O(M3)whereM is the number of training",
"son re-identification [3, 38], few-shot learning [24, 26, 30]",
"data, thus slow convergence. Furthermore, some tuples do",
"and representation learning [14, 33, 41]. Following their",
"not contribute to training or even degrade the quality of the",
"great success in visual recognition, deep neural networks",
"learned embedding space. To resolve these issues, learning",
"havebeenemployedrecentlyformetriclearning.The net-",
"with the pair-based losses often entails tuple sampling tech-",
"works are trained to project data onto an embedding space",
"niques [10, 25, 37, 40], which however have to be tuned by",
"in which semantically similar data (e.g., images of the same",
"hand and may increase the risk of overfitting.",
"class) are closely grouped together. Such a quality of the",
"The proxy-based losses resolve the above complexity is-",
"embedding space is given mainly by loss functions used for",
"sue by introducing proxies [1, 21, 23]. A proxy is a repre-",
"training the networks, and most of the losses are categorized",
"sentative of a subset of training data and learned as a part",
"into two classes: pair-based and proxy-based.",
"of the network parameters. Existing losses in this category",
"The pair-based losses are built upon pairwise distances",
"consider each data point as an anchor, associate it with prox-",
"between data in the embedding space. A seminal example",
"ies instead of other images, and encourage the anchor to be",
"is Contrastive loss [4, 9], which aims to minimize the dis-",
"close to proxies of the same class and far apart from those",
"tance between a pair of data if their class labels are identical",
"of different classes. Proxy-based losses reduce the training",
"and to separate them otherwise. Recent pair-based losses",
"consider a group of pairwise distances to handle relations",
"IThe training complexity indicates the amount of computation required",
"between more than two data[14.25,27.29.32.34.35.391.",
"to address the entire training dataset [1, 6, 10, 23, 35]."
] |
{
"arxiv_id": "2003.13911",
"language": "en",
"page_idx": 0
}
|
||
2003.13911-en-1
|
2003.13911-en-1.jpg
|
<table>
<tbody>
<tr>
<td>$\square$ $O$</td>
<td>$\mathbf{O}$ $O$ 1</td>
<td>口 $O$</td>
<td>$O$ $\square$</td>
<td>T $\frac{1}{2}$ $\frac{1}{2}$</td>
<td>O 口</td>
</tr>
<tr>
<td>口 $O$ 口</td>
<td>$O$ $\square$ 口</td>
<td>口 $O$ 口</td>
<td>$\square$ 口</td>
<td>二 口</td>
<td>口 口</td>
</tr>
<tr>
<td>$\hat{C}$ 口 $(a)$ Triplet</td>
<td>$L$ $( b)$ $N- pair$</td>
<td>(c) Lifted Structure</td>
<td>口 $Proxv-NCA$</td>
<td>(e)Ours</td>
<td>(e) Ours</td>
</tr>
<tr>
<td>$(a)$ Triplet</td>
<td>(b$)\cap$ N-pair</td>
<td>Lifted Structure (c)</td>
<td>$Proxv-NCA$</td>
<td>(e) $10urs$</td>
<td>(e)Ours T</td>
</tr>
</tbody>
</table>
Figure 2. Comparison between popular metric learning losses and ours. Small nodes are embedding vectors of data in a batch, and black ones indicate proxies; their different shapes represent distinct classes. The associations defined by the losses are expressed by edges, and thicker edges get larger gradients. Also, embedding vectors associated with the anchor are colored in red if they are of the same class of the anchor $(i.e.$, positive) and in blue otherwise $(i.e.$, negative). (a) Triplet loss[25,32] associates each anchor with a positive and a negative data point without considering their hardness. (b) $N$-pair loss[27] and (c) Lifted Structure loss[29] reflect hardness of data, but do not utilize all data in the batch. (d) Proxy-NCA loss [21] cannot exploit data-to-data relations since it associates each data point only with proxies. (e) Our loss handles entire data in the batch, and associates them with each proxy with consideration of their relative hardness determined by data-to-data relations. See the text for more details.
data sampling; its convergence is even faster than those
of Proxy-NCA[21] and Multi-Similarity loss[34].
## $2. \textbf{Related Work}$
In this section, we categorize metric learning losses into two classes, pair-based and proxy-based losses, then review relevant methods for each category.
# $2. 1. \textbf{Pair- based Losses}$
complexity and enable faster convergence since the number of proxies is substantially smaller than that of training data in general. Further, these losses tend to be more robust against label noises and outliers. However, since they associate each data point only with proxies, proxy-based losses can leverage only data-to-proxy relations, which are impoverished compared to the rich data-to-data relations available for pair-based losses.
In this paper, we propose a novel proxy-based loss called Proxy-Anchor loss, which takes good points of both proxybased and pair-based losses while correcting their defects. Unlike the existing proxy-based losses, the proposed loss utilizes each proxy as an anchor and associates it with all data in a batch. Specifically, for each proxy, the loss aims to pull data of the same class close to the proxy and to push others away in the embedding space. Due to the use of proxies, our loss boosts the speed of convergence with no hyperparameter for tuple sampling, and is robust against noisy labels and outliers. At the same time, it can take data-to-data relations into account like pair-based losses; this property is given by associating all data in a batch with each proxy so that the gradients with respect to a data point are weighted by its relative proximity to the proxy $(i.e.$, relative hardness) affected by the other data in the batch. Thanks to the above advantages, a standard embedding network trained with our loss achieves state-of-the-art accuracy and most quickly converges as shown in Figure 1. The contribution of this paper is three-fold:
$\bullet$ We propose a novel metric learning loss that takes ad-
vantages of both pair-based and proxy-based methods; it leverages rich data-to-data relations and enables fast and reliable convergence.
· A standard embedding network trained with our loss
achieves state-of-the-art performance on the four public
benchmarks for metric learning [17,19,29,36].
$\bullet$ Our loss speeds up convergence greatly without careful
Contrastive loss [2,4,9] and Triplet loss [25,32] are seminal examples of loss functions for deep metric learning. Contrastive loss takes a pair of embedding vectors as input, and aims to pull them together if they are of the same class and push them apart otherwise. Triplet loss considers a data point as an anchor, associates it with a positive and a negative data point, and constrains the distance of the anchorpositive pair to be smaller than that of the anchor-negative pair in the embedding space as illustrated in Figure 2(a).
Recent pair-based losses aim to leverage higher order relations between data and reflect their hardness for further enhancement. As generalizations of Triplet loss, $N$-pair loss [27] and Lifted Structure loss [29] associate an anchor with a single positive and multiple negative data points, and pull the positive to the anchor and push the negatives away from the anchor while considering their hardness. As shown in Figure 2(b) and 2(c), however, these losses do not utilize entire data in a batch since they sample the same number of data per negative class, thus may drop informative examples during training. In contrast, Ranked List loss [35] takes into account all positive and negative data in a batch and aims to separate the positive and negative sets. MultiSimilarity loss[34] also considers every pair of data in a batch, and assigns a weight to each pair according to three complementary types of similarity to focus more on useful pairs for improving performance and convergence speed.
|
[
[
0.1871337890625,
0.0985107421875,
0.2017822265625,
0.0985107421875,
0.2017822265625,
0.10919189453125,
0.1871337890625,
0.10919189453125
],
[
0.339111328125,
0.0997314453125,
0.35546875,
0.0997314453125,
0.35546875,
0.10919189453125,
0.339111328125,
0.10919189453125
],
[
0.646484375,
0.0997314453125,
0.6611328125,
0.0997314453125,
0.6611328125,
0.10986328125,
0.646484375,
0.10986328125
],
[
0.7998046875,
0.0997314453125,
0.814453125,
0.0997314453125,
0.814453125,
0.10919189453125,
0.7998046875,
0.10919189453125
],
[
0.52783203125,
0.10418701171875,
0.53759765625,
0.11297607421875,
0.53173828125,
0.11614990234375,
0.52294921875,
0.106689453125
],
[
0.11767578125,
0.1060791015625,
0.1356201171875,
0.1060791015625,
0.1356201171875,
0.11932373046875,
0.11767578125,
0.11932373046875
],
[
0.2197265625,
0.1060791015625,
0.233642578125,
0.1060791015625,
0.233642578125,
0.11553955078125,
0.2197265625,
0.11553955078125
],
[
0.26953125,
0.1060791015625,
0.28759765625,
0.1060791015625,
0.28759765625,
0.11932373046875,
0.26953125,
0.11932373046875
],
[
0.373291015625,
0.1060791015625,
0.3857421875,
0.1060791015625,
0.3857421875,
0.11553955078125,
0.373291015625,
0.11553955078125
],
[
0.491943359375,
0.10101318359375,
0.50732421875,
0.10101318359375,
0.50732421875,
0.10919189453125,
0.491943359375,
0.10919189453125
],
[
0.67626953125,
0.1060791015625,
0.689453125,
0.1060791015625,
0.689453125,
0.1142578125,
0.67626953125,
0.1142578125
],
[
0.57861328125,
0.10797119140625,
0.59326171875,
0.10797119140625,
0.59326171875,
0.117431640625,
0.57861328125,
0.117431640625
],
[
0.4248046875,
0.106689453125,
0.438720703125,
0.106689453125,
0.438720703125,
0.117431640625,
0.4248046875,
0.117431640625
],
[
0.6357421875,
0.12249755859375,
0.64794921875,
0.12249755859375,
0.64794921875,
0.1307373046875,
0.6357421875,
0.1307373046875
],
[
0.2059326171875,
0.146484375,
0.223876953125,
0.146484375,
0.223876953125,
0.160400390625,
0.2059326171875,
0.160400390625
],
[
0.358642578125,
0.146484375,
0.375732421875,
0.146484375,
0.375732421875,
0.160400390625,
0.358642578125,
0.160400390625
],
[
0.6650390625,
0.147705078125,
0.6787109375,
0.147705078125,
0.6787109375,
0.1590576171875,
0.6650390625,
0.1590576171875
],
[
0.178955078125,
0.1641845703125,
0.19287109375,
0.1641845703125,
0.19287109375,
0.173583984375,
0.178955078125,
0.173583984375
],
[
0.33251953125,
0.1641845703125,
0.346435546875,
0.1641845703125,
0.346435546875,
0.173583984375,
0.33251953125,
0.173583984375
],
[
0.1568603515625,
0.1749267578125,
0.1748046875,
0.1749267578125,
0.1748046875,
0.1881103515625,
0.1568603515625,
0.1881103515625
],
[
0.3095703125,
0.1749267578125,
0.326904296875,
0.1749267578125,
0.326904296875,
0.1881103515625,
0.3095703125,
0.1881103515625
],
[
0.61669921875,
0.1767578125,
0.6298828125,
0.1767578125,
0.6298828125,
0.1875,
0.61669921875,
0.1875
],
[
0.771484375,
0.177978515625,
0.78369140625,
0.177978515625,
0.78369140625,
0.184326171875,
0.771484375,
0.184326171875
],
[
0.463134765625,
0.1787109375,
0.47216796875,
0.175537109375,
0.476318359375,
0.1817626953125,
0.46728515625,
0.1849365234375
],
[
0.11767578125,
0.179931640625,
0.134033203125,
0.177978515625,
0.136474609375,
0.1875,
0.12091064453125,
0.189453125
],
[
0.2279052734375,
0.1849365234375,
0.2344970703125,
0.1849365234375,
0.2344970703125,
0.1900634765625,
0.2279052734375,
0.1900634765625
],
[
0.273681640625,
0.1817626953125,
0.28515625,
0.1817626953125,
0.28515625,
0.186279296875,
0.273681640625,
0.186279296875
],
[
0.146240234375,
0.1962890625,
0.214111328125,
0.1962890625,
0.214111328125,
0.2108154296875,
0.146240234375,
0.2108154296875
],
[
0.2998046875,
0.1962890625,
0.366943359375,
0.1962890625,
0.366943359375,
0.2108154296875,
0.2998046875,
0.2108154296875
],
[
0.42236328125,
0.1962890625,
0.5498046875,
0.1962890625,
0.5498046875,
0.2108154296875,
0.42236328125,
0.2108154296875
],
[
0.5908203125,
0.1962890625,
0.68701171875,
0.1944580078125,
0.68701171875,
0.208984375,
0.5908203125,
0.2108154296875
],
[
0.76318359375,
0.1962890625,
0.82177734375,
0.1962890625,
0.82177734375,
0.2108154296875,
0.76318359375,
0.2108154296875
],
[
0.080078125,
0.2215576171875,
0.8896484375,
0.2215576171875,
0.8896484375,
0.23486328125,
0.080078125,
0.23486328125
],
[
0.07843017578125,
0.23486328125,
0.890625,
0.23486328125,
0.890625,
0.248779296875,
0.07843017578125,
0.248779296875
],
[
0.080078125,
0.25,
0.89208984375,
0.25,
0.89208984375,
0.26318359375,
0.080078125,
0.26318359375
],
[
0.080078125,
0.26318359375,
0.89208984375,
0.26318359375,
0.89208984375,
0.276611328125,
0.080078125,
0.276611328125
],
[
0.080078125,
0.279052734375,
0.890625,
0.279052734375,
0.890625,
0.2890625,
0.080078125,
0.2890625
],
[
0.080078125,
0.292236328125,
0.890625,
0.292236328125,
0.890625,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.89208984375,
0.304931640625,
0.89208984375,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.080078125,
0.32080078125,
0.4697265625,
0.32080078125,
0.4697265625,
0.330810546875,
0.080078125,
0.330810546875
],
[
0.080078125,
0.3515625,
0.468017578125,
0.3515625,
0.468017578125,
0.365478515625,
0.080078125,
0.365478515625
],
[
0.51953125,
0.352783203125,
0.890625,
0.352783203125,
0.890625,
0.363037109375,
0.51953125,
0.363037109375
],
[
0.07843017578125,
0.365478515625,
0.468017578125,
0.3662109375,
0.468017578125,
0.380615234375,
0.07843017578125,
0.380126953125
],
[
0.51708984375,
0.3662109375,
0.8505859375,
0.3662109375,
0.8505859375,
0.380126953125,
0.51708984375,
0.380126953125
],
[
0.080078125,
0.3818359375,
0.4697265625,
0.3818359375,
0.4697265625,
0.395751953125,
0.080078125,
0.395751953125
],
[
0.080078125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.50341796875,
0.3984375,
0.640625,
0.3984375,
0.640625,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.07843017578125,
0.412353515625,
0.471435546875,
0.412353515625,
0.471435546875,
0.425537109375,
0.07843017578125,
0.425537109375
],
[
0.08087158203125,
0.4287109375,
0.465576171875,
0.4287109375,
0.465576171875,
0.439453125,
0.08087158203125,
0.439453125
],
[
0.521484375,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.438232421875,
0.521484375,
0.438232421875
],
[
0.08087158203125,
0.443115234375,
0.46728515625,
0.443115234375,
0.46728515625,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.50244140625,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.080078125,
0.45703125,
0.221435546875,
0.45703125,
0.221435546875,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.5009765625,
0.453857421875,
0.736328125,
0.455810546875,
0.736328125,
0.47021484375,
0.5009765625,
0.468505859375
],
[
0.09722900390625,
0.472900390625,
0.4697265625,
0.472900390625,
0.4697265625,
0.486083984375,
0.09722900390625,
0.486083984375
],
[
0.080078125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.50244140625,
0.481689453125,
0.67333984375,
0.48291015625,
0.67333984375,
0.49755859375,
0.50244140625,
0.496826171875
],
[
0.08087158203125,
0.50390625,
0.46728515625,
0.50390625,
0.46728515625,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.521484375,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.521484375,
0.52099609375
],
[
0.080078125,
0.517578125,
0.4697265625,
0.517578125,
0.4697265625,
0.5322265625,
0.080078125,
0.5322265625
],
[
0.50244140625,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.537109375,
0.50244140625,
0.537109375
],
[
0.080078125,
0.53271484375,
0.468017578125,
0.53271484375,
0.468017578125,
0.546875,
0.080078125,
0.546875
],
[
0.5009765625,
0.537109375,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.5009765625,
0.5517578125
],
[
0.08087158203125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.56005859375,
0.08087158203125,
0.56005859375
],
[
0.5009765625,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.5009765625,
0.56640625
],
[
0.080078125,
0.56298828125,
0.4697265625,
0.56298828125,
0.4697265625,
0.5771484375,
0.080078125,
0.5771484375
],
[
0.50244140625,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.5791015625,
0.4697265625,
0.5791015625,
0.4697265625,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50244140625,
0.583984375,
0.8896484375,
0.583984375,
0.8896484375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.59326171875,
0.468017578125,
0.60693359375,
0.080078125,
0.60693359375
],
[
0.50244140625,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.08087158203125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62060546875,
0.08087158203125,
0.62060546875
],
[
0.5009765625,
0.61279296875,
0.890625,
0.61279296875,
0.890625,
0.626953125,
0.5009765625,
0.626953125
],
[
0.07843017578125,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.07843017578125,
0.63720703125
],
[
0.50244140625,
0.6298828125,
0.87109375,
0.6298828125,
0.87109375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.080078125,
0.63818359375,
0.4697265625,
0.63818359375,
0.4697265625,
0.6513671875,
0.080078125,
0.6513671875
],
[
0.52294921875,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.52294921875,
0.6572265625
],
[
0.080078125,
0.65380859375,
0.4697265625,
0.65380859375,
0.4697265625,
0.66748046875,
0.080078125,
0.66748046875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.66845703125,
0.4697265625,
0.66845703125,
0.4697265625,
0.681640625,
0.080078125,
0.681640625
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.080078125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.080078125,
0.69775390625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.69873046875,
0.4697265625,
0.69873046875,
0.4697265625,
0.7119140625,
0.080078125,
0.7119140625
],
[
0.50244140625,
0.70556640625,
0.8935546875,
0.70556640625,
0.8935546875,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.07843017578125,
0.71337890625,
0.468017578125,
0.71337890625,
0.468017578125,
0.7265625,
0.07843017578125,
0.7265625
],
[
0.5009765625,
0.7197265625,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.5009765625,
0.734375
],
[
0.080078125,
0.72900390625,
0.4697265625,
0.72900390625,
0.4697265625,
0.7421875,
0.080078125,
0.7421875
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.08087158203125,
0.74560546875,
0.46728515625,
0.74560546875,
0.46728515625,
0.755859375,
0.08087158203125,
0.755859375
],
[
0.50244140625,
0.751953125,
0.89208984375,
0.751953125,
0.89208984375,
0.76513671875,
0.50244140625,
0.76513671875
],
[
0.07843017578125,
0.7587890625,
0.253173828125,
0.75830078125,
0.253173828125,
0.77294921875,
0.07843017578125,
0.7734375
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.08087158203125,
0.77783203125,
0.46728515625,
0.77783203125,
0.46728515625,
0.79150390625,
0.08087158203125,
0.79150390625
],
[
0.50244140625,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.09478759765625,
0.79248046875,
0.4697265625,
0.79150390625,
0.4697265625,
0.80615234375,
0.09478759765625,
0.806640625
],
[
0.50244140625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.096435546875,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.096435546875,
0.8212890625
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.096435546875,
0.82470703125,
0.236083984375,
0.82470703125,
0.236083984375,
0.83544921875,
0.096435546875,
0.83544921875
],
[
0.5009765625,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.080078125,
0.8388671875,
0.4697265625,
0.8388671875,
0.4697265625,
0.85302734375,
0.080078125,
0.85302734375
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.09478759765625,
0.85498046875,
0.4697265625,
0.85498046875,
0.4697265625,
0.86865234375,
0.09478759765625,
0.86865234375
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.09478759765625,
0.87060546875,
0.41259765625,
0.87060546875,
0.41259765625,
0.8837890625,
0.09478759765625,
0.8837890625
],
[
0.5009765625,
0.8720703125,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.5009765625,
0.88623046875
],
[
0.08087158203125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.875,
0.88720703125,
0.875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"?",
"?",
"O",
"O",
"",
"0",
"■",
"0",
"园",
"O",
"",
"O",
"O",
"一",
"口",
"口",
"口",
"C",
"C",
"A",
"",
"C",
"福",
"图",
"C",
"",
"",
"(a) Triplet",
"(b) N-pair",
"(c) Lifted Structure",
"(d) Proxy-NCA",
"(e) Ours",
"Figure 2. Comparison between popular metric learning losses and ours. Small nodes are embedding vectors of data in a batch, and black",
"ones indicate proxies; their different shapes represent distinct classes. The associations defined by the losses are expressed by edges, and",
"thicker edges get larger gradients. Also, embedding vectors associated with the anchor are colored in red if they are of the same class of the",
"anchor (i.e., positive) and in blue otherwise (i.e., negative). (a) Triplet loss [25, 32] associates each anchor with a positive and a negative",
"datapoint without consideringtheirhardness.(b)N-pairloss[27|and (c)LiftedStructureloss[29]reflecthardness of data,butdonot",
"utilize all data in the batch. (d) Proxy-NCA loss [2l cannot exploit data-to-data relations since it associates each data point only with",
"proxies. (e) Our loss handles entire data in the batch, and associates them with each proxy with consideration of their relative hardness",
"determined by data-to-data relations.See the text for more details.",
"complexity and enable faster convergence since the num-",
"data sampling:its convergence is even faster than those",
"ber of proxies is substantially smaller than that of training",
"of Proxy-NCA [2]] and Multi-Similarity loss [34]]",
"data in general. Further, these losses tend to be more robust",
"against label noises and outliers. However, since they asso-",
"2. Related Work",
"ciate each data point only with proxies, proxy-based losses",
"canleverageonlydata-to-proxyrelations,which areimpov-",
"In this section, we categorize metric learning losses into",
"erished compared totherich data-to-datarelations available",
"two classes, pair-based and proxy-based losses, then review",
"for pair-based losses.",
"relevant methods foreach category.",
"In this paper, we propose a novel proxy-based loss called",
"Proxy-Anchor loss, which takes good points of both proxy-",
"2.1. Pair-based Losses",
"based and pair-based losses while correcting their defects.",
"Contrastive loss [2, 4, 9] and Triplet loss [25, 32] are",
"Unlike the existing proxy-based losses, the proposed loss",
"seminal examples of loss functions for deep metric learning.",
"utilizes each proxy as an anchor and associates it with all",
"Contrastive loss takes a pair of embedding vectors as input,",
"datain abatch.Specifically,for eachproxy,the loss aims",
"and aims to pull them together if they are of the same class",
"to pull data of the same class close to the proxy and to push",
"and push them apart otherwise. Triplet loss considers a data",
"others away in the embedding space. Due to the use of prox-",
"point as an anchor, associates it with a positive and a neg-",
"ies, our loss boosts the speed of convergence with no hyper-",
"ative data point, and constrains the distance of the anchor-",
"parameterfortuple sampling,and isrobustagainst noisyla-",
"positive pair to be smaller than that of the anchor-negative",
"bels and outliers. At the same time, it can take data-to-data",
"pair in the embedding space as illustrated in Figure 2(a).",
"relations into account like pair-based losses; this property is",
"Recentpair-based losses aimto leveragehigherorder re-",
"given by associating all data in a batch with each proxy so",
"lations between data and reflect their hardness for further",
"that the gradients with respect to a data point are weighted",
"enhancement.As generalizations of Triplet loss,N-pair",
"by its relative proximity to the proxy (i.e., relative hard-",
"loss [27] and Lifted Structure loss [29] associate an anchor",
"ness) affected by the other data in the batch. Thanks to the",
"with a single positive and multiple negative data points, and",
"above advantages, a standard embedding network trained",
"pull the positive to the anchor and push the negatives away",
"with our loss achieves state-of-the-art accuracy and most",
"fromtheanchor whileconsideringtheirhardness.Asshown",
"quicklyconverges as shown in Figurel.The contribution",
"in Figure 2(b) and 2(c), however, these losses do not utilize",
"of this paper is three-fold:",
"entire data in a batch since they sample the same number",
" We propose a novel metric learning loss that takes ad-",
"of data per negative class, thus may drop informative ex-",
"vantages of both pair-based and proxy-based methods; it",
"amples during training. In contrast, Ranked List loss [35]",
"leverages rich data-to-data relations and enables fast and",
"takes into account all positive and negative data in a batch",
"reliableconvergence.",
"and aims to separate the positive and negative sets. Multi-",
". A standard embedding network trained with our loss",
"Similarity loss [34] also considers every pair of data in a",
"achieves state-of-the-art performance on the four public",
"batch, and assigns a weight to each pair according to three",
"benchmarks for metric learning [17, 19, 29, 36].",
"complementary types of similarity to focus more on useful",
" Our loss speeds up convergence greatly without careful",
"pairs for improving performance and convergence speed."
] |
{
"arxiv_id": "2003.13911",
"language": "en",
"page_idx": 1
}
|
||
2003.13911-en-2
|
2003.13911-en-2.jpg
|
Pair-based losses enjoy rich and fine-grained data-todata relations as they examine tuples $(i.e.$, data pairs or their combinations) during training. However, since the number of tuples increases polynomially with the number of training data, their training complexity is prohibitively high and convergence is slow. In addition, a large amount of tuples are not effective and sometimes even degrade the quality of the learned embedding space[25,37]. To address this issue, most pair-based losses entail tuple sampling techniques [10, 25, 37, 40] to select and utilize tuples that will contribute to training. However, these techniques involve hyperparameters that have to be tuned carefully, and may increase the risk of overfitting since they rely mostly on local pairwise relations within a batch. Another way to alleviating the complexity issue is to assign larger weights to more useful pairs during training as in [34], which however also incorporates a sampling technique.
Our loss resolves this complexity issue by adopting proxies, which enables faster and more reliable convergence compared to pair-based losses. Furthermore, it demands no additional hyperparameter for tuple sampling.
## $2. 2. \textbf{Proxy- based Losses}$
Proxy-based metric learning is a relatively new approach that can address the complexity issue of the pair-based losses. A proxy means a representative of a subset of training data and is estimated as a part of the embedding network parameters. The common idea of the methods in this category is to infer a small set of proxies that capture the global structure of an embedding space and relate each data point with the proxies instead of the other data points during training. Since the number of proxies is significantly smaller than that of training data, the training complexity can be reduced substantially.
The first proxy-based loss is Proxy-NCA[21],which is an approximation of Neighborhood Component Analysis (NCA)[8] using proxies. In its standard setting, ProxyNCA loss assigns a single proxy for each class, associates a data point with proxies, and encourages the positive pair to be close and negative pairs to be far apart, as illustrated in Figure 2(d). SoftTriple loss [23], an extension of SoftMax loss for classification, is similar to Proxy-NCA yet assigns multiple proxies to each class to reflect intra-class variance. Manifold Proxy loss[1] is an extension of $N$-pair loss using proxies, and improves the performance by adopting a manifold-aware distance instead of Euclidean distance to measure the semantic distance in the embedding space.
Using proxies in these losses helps improve training convergence greatly, but has an inherent limitation as a side effect: Since each data point is associated only with proxies, the rich data-to-data relations that are available for the pair-based methods are not accessible anymore. Our loss can overcome this limitation since its gradients reflect rela-
tive hardness of data and allow their embedding vectors to
interact with each other during training.
## 3.0ur Method
We propose a new metric learning loss called ProxyAnchor loss to overcome the inherent limitations of the previous methods. The loss employs proxies that enable fast and reliable convergence as in proxy-based losses. Also, although it is built upon data-proxy relations, our loss can utilize data-to-data relations during training like pair-based losses since it enables embedding vectors of data points to be affected by each other through its gradients. This property of our loss improves the quality of the learned embedding space substantially.
In this section, we first review Proxy-NCA loss [21], a representative proxy-based loss, for comparison to our Proxy-Anchor loss. We then describe our Proxy-Anchor loss in detail and analyze its training complexity
## $3. 1. \textbf{Review of Proxy- NCA Loss}$
In the standard setting, Proxy-NCA loss [21] assigns a proxy to each class so that the number of proxies is the same with that of class labels. Given an input data point as an anchor, the proxy of the same class of the input is regarded as positive and the other proxies are negative. Let $x$ denote the embedding vector of the input, $p^+$ be the positive proxy, and $p^-$ be a negative proxy. The loss is then given by
(1)
$$\begin{aligned}\ell(X)&=\sum_{x\in X}-\log\frac{e^{s(x,p^{+})}}{\sum_{p^{-}\in P^{-}}e^{s(x,p^{-})}}\\&=\sum_{x\in X}\Big\{\:-\:s(x,p^{+})+\mathop{\mathrm{LSE}}_{p^{-}\in P^{-}}s(x,p^{-})\Big\},\end{aligned}$$
(2
where $X$ is a batch of embedding vectors, $P^-$ is the set of negative proxies, and $s(\cdot,\cdot)$ denotes the cosine similarity between two vectors. In addition, LSE in Eq. (2) means the Log-Sum-Exp function, a smooth approximation to the max function. The gradient of Proxy-NCA loss with respect to $s(x,p)$ is given by
(3)
$$\dfrac{\partial\ell(X)}{\partial s(x,p)}=\begin{cases}-1,&\text{if}p=p^+,\\\dfrac{e^{s(x,p)}}{\sum\limits_{p^-\in P^-}e^{s(x,p^-)}},&\text{otherwise.}\end{cases}$$
Eq.(3) shows that minimizing the loss encourages $x$ and $p^{+}$ to be close to each other, and $x$ and $p^-$ to be far away. In particular, $x$ and $p^+$ are pulled together by the constant power, while $x$ and $p^-$ closer to each other $(i.e.$,harder negative) are more strongly pushed away.
Proxy-NCA loss enables fast convergence thanks to its low training complexity, $O(MC)$ where $M$ is the number
|
[
[
0.098876953125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.098876953125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.5009765625,
0.106689453125,
0.7646484375,
0.10797119140625,
0.7646484375,
0.121826171875,
0.5009765625,
0.12060546875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.080078125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50244140625,
0.1351318359375,
0.6298828125,
0.1351318359375,
0.6298828125,
0.149658203125,
0.50244140625,
0.149658203125
],
[
0.080078125,
0.15283203125,
0.4697265625,
0.15283203125,
0.4697265625,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.521484375,
0.160400390625,
0.890625,
0.160400390625,
0.890625,
0.1749267578125,
0.521484375,
0.1749267578125
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08087158203125,
0.179931640625
],
[
0.50244140625,
0.175537109375,
0.89208984375,
0.175537109375,
0.89208984375,
0.189453125,
0.50244140625,
0.189453125
],
[
0.080078125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.5009765625,
0.189453125,
0.89208984375,
0.1900634765625,
0.89208984375,
0.2052001953125,
0.5009765625,
0.203857421875
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.205810546875,
0.89208984375,
0.205810546875,
0.89208984375,
0.2197265625,
0.50244140625,
0.2197265625
],
[
0.07843017578125,
0.21337890625,
0.46728515625,
0.21337890625,
0.46728515625,
0.2266845703125,
0.07843017578125,
0.2266845703125
],
[
0.50244140625,
0.2203369140625,
0.89208984375,
0.2203369140625,
0.89208984375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.07843017578125,
0.2291259765625,
0.468017578125,
0.2266845703125,
0.468017578125,
0.240478515625,
0.07843017578125,
0.242431640625
],
[
0.50341796875,
0.2374267578125,
0.8896484375,
0.2374267578125,
0.8896484375,
0.2474365234375,
0.50341796875,
0.2474365234375
],
[
0.080078125,
0.24365234375,
0.471435546875,
0.24365234375,
0.471435546875,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.25244140625,
0.890625,
0.25244140625,
0.890625,
0.26318359375,
0.50341796875,
0.26318359375
],
[
0.080078125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.266357421875,
0.890625,
0.266357421875,
0.890625,
0.27978515625,
0.50244140625,
0.27978515625
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.28271484375,
0.8896484375,
0.28271484375,
0.8896484375,
0.29345703125,
0.50341796875,
0.29345703125
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50244140625,
0.296630859375,
0.6640625,
0.296630859375,
0.6640625,
0.31005859375,
0.50244140625,
0.31005859375
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.521484375,
0.311279296875,
0.890625,
0.311279296875,
0.890625,
0.324462890625,
0.521484375,
0.324462890625
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.50341796875,
0.326904296875,
0.890625,
0.326904296875,
0.890625,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.080078125,
0.333984375,
0.338134765625,
0.333984375,
0.338134765625,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.5009765625,
0.34033203125,
0.890625,
0.34033203125,
0.890625,
0.354736328125,
0.5009765625,
0.354736328125
],
[
0.098876953125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.3642578125,
0.098876953125,
0.3642578125
],
[
0.5009765625,
0.35595703125,
0.82275390625,
0.356689453125,
0.82275390625,
0.371826171875,
0.5009765625,
0.37060546875
],
[
0.07843017578125,
0.365478515625,
0.4697265625,
0.365478515625,
0.4697265625,
0.37890625,
0.07843017578125,
0.37890625
],
[
0.080078125,
0.380615234375,
0.468017578125,
0.380615234375,
0.468017578125,
0.39453125,
0.080078125,
0.39453125
],
[
0.50341796875,
0.380126953125,
0.74609375,
0.380126953125,
0.74609375,
0.39453125,
0.50341796875,
0.39453125
],
[
0.0767822265625,
0.39453125,
0.381591796875,
0.396484375,
0.381591796875,
0.410888671875,
0.0767822265625,
0.4091796875
],
[
0.521484375,
0.40283203125,
0.89208984375,
0.40283203125,
0.89208984375,
0.416748046875,
0.521484375,
0.416748046875
],
[
0.080078125,
0.42236328125,
0.2646484375,
0.42236328125,
0.2646484375,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.50341796875,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.50341796875,
0.431884765625
],
[
0.50341796875,
0.43505859375,
0.8896484375,
0.43505859375,
0.8896484375,
0.44580078125,
0.50341796875,
0.44580078125
],
[
0.096435546875,
0.44580078125,
0.4697265625,
0.4462890625,
0.4697265625,
0.4609375,
0.096435546875,
0.460205078125
],
[
0.5009765625,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.5009765625,
0.462158203125
],
[
0.080078125,
0.462158203125,
0.4697265625,
0.462158203125,
0.4697265625,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.5009765625,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.5009765625,
0.477783203125
],
[
0.080078125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.080078125,
0.490478515625
],
[
0.5009765625,
0.477783203125,
0.89208984375,
0.479248046875,
0.89208984375,
0.49365234375,
0.5009765625,
0.492431640625
],
[
0.07843017578125,
0.492431640625,
0.468017578125,
0.492431640625,
0.468017578125,
0.505859375,
0.07843017578125,
0.505859375
],
[
0.50244140625,
0.494384765625,
0.85107421875,
0.494384765625,
0.85107421875,
0.50830078125,
0.50244140625,
0.50830078125
],
[
0.080078125,
0.5068359375,
0.468017578125,
0.5068359375,
0.468017578125,
0.52099609375,
0.080078125,
0.52099609375
],
[
0.080078125,
0.52294921875,
0.4697265625,
0.52294921875,
0.4697265625,
0.5361328125,
0.080078125,
0.5361328125
],
[
0.69140625,
0.517578125,
0.74658203125,
0.5146484375,
0.74853515625,
0.52978515625,
0.69287109375,
0.53271484375
],
[
0.5400390625,
0.5263671875,
0.67236328125,
0.5263671875,
0.67236328125,
0.54443359375,
0.5400390625,
0.54443359375
],
[
0.86767578125,
0.52783203125,
0.89453125,
0.52783203125,
0.89453125,
0.54443359375,
0.86767578125,
0.54443359375
],
[
0.080078125,
0.53857421875,
0.4697265625,
0.53857421875,
0.4697265625,
0.5517578125,
0.080078125,
0.5517578125
],
[
0.68115234375,
0.54345703125,
0.69677734375,
0.54345703125,
0.69677734375,
0.5537109375,
0.68115234375,
0.5537109375
],
[
0.7099609375,
0.53857421875,
0.76953125,
0.53515625,
0.771484375,
0.5498046875,
0.71142578125,
0.55322265625
],
[
0.59619140625,
0.54541015625,
0.6298828125,
0.54541015625,
0.6298828125,
0.55810546875,
0.59619140625,
0.55810546875
],
[
0.07843017578125,
0.55322265625,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.07843017578125,
0.56640625
],
[
0.66845703125,
0.56005859375,
0.70751953125,
0.556640625,
0.708984375,
0.568359375,
0.6689453125,
0.5712890625
],
[
0.080078125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.080078125,
0.58203125
],
[
0.63818359375,
0.57568359375,
0.833984375,
0.57568359375,
0.833984375,
0.5966796875,
0.63818359375,
0.5966796875
],
[
0.5986328125,
0.58203125,
0.61279296875,
0.5791015625,
0.61669921875,
0.5859375,
0.60302734375,
0.59033203125
],
[
0.87158203125,
0.58203125,
0.88623046875,
0.58203125,
0.88623046875,
0.59033203125,
0.87158203125,
0.59033203125
],
[
0.5986328125,
0.59765625,
0.625,
0.59765625,
0.625,
0.60595703125,
0.5986328125,
0.60595703125
],
[
0.07843017578125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.07843017578125,
0.5966796875
],
[
0.080078125,
0.599609375,
0.26708984375,
0.599609375,
0.26708984375,
0.60986328125,
0.080078125,
0.60986328125
],
[
0.09722900390625,
0.61279296875,
0.4697265625,
0.61279296875,
0.4697265625,
0.62744140625,
0.09722900390625,
0.62744140625
],
[
0.50341796875,
0.6162109375,
0.89208984375,
0.6162109375,
0.89208984375,
0.6298828125,
0.50341796875,
0.6298828125
],
[
0.080078125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.642578125,
0.080078125,
0.642578125
],
[
0.50244140625,
0.6318359375,
0.890625,
0.6318359375,
0.890625,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.080078125,
0.64453125,
0.468017578125,
0.64453125,
0.468017578125,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.646484375,
0.89208984375,
0.646484375,
0.89208984375,
0.66015625,
0.50244140625,
0.66015625
],
[
0.07843017578125,
0.65771484375,
0.471435546875,
0.6591796875,
0.471435546875,
0.673828125,
0.07843017578125,
0.67236328125
],
[
0.50244140625,
0.662109375,
0.89208984375,
0.662109375,
0.89208984375,
0.67529296875,
0.50244140625,
0.67529296875
],
[
0.080078125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.080078125,
0.6875
],
[
0.5009765625,
0.67529296875,
0.89208984375,
0.6767578125,
0.89208984375,
0.69140625,
0.5009765625,
0.68994140625
],
[
0.08087158203125,
0.69140625,
0.468017578125,
0.69140625,
0.468017578125,
0.7021484375,
0.08087158203125,
0.7021484375
],
[
0.50341796875,
0.6923828125,
0.625,
0.6923828125,
0.625,
0.70703125,
0.50341796875,
0.70703125
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.65283203125,
0.71728515625,
0.67236328125,
0.71728515625,
0.67236328125,
0.72998046875,
0.65283203125,
0.72998046875
],
[
0.76953125,
0.71630859375,
0.8408203125,
0.71630859375,
0.8408203125,
0.73095703125,
0.76953125,
0.73095703125
],
[
0.55810546875,
0.734375,
0.60693359375,
0.734375,
0.60693359375,
0.75,
0.55810546875,
0.75
],
[
0.080078125,
0.7197265625,
0.4697265625,
0.7197265625,
0.4697265625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.67626953125,
0.73486328125,
0.72216796875,
0.73486328125,
0.72216796875,
0.75,
0.67626953125,
0.75
],
[
0.76953125,
0.7470703125,
0.84228515625,
0.7470703125,
0.84228515625,
0.75830078125,
0.76953125,
0.75830078125
],
[
0.869140625,
0.744140625,
0.8935546875,
0.744140625,
0.8935546875,
0.75830078125,
0.869140625,
0.75830078125
],
[
0.55224609375,
0.75048828125,
0.611328125,
0.75244140625,
0.609375,
0.76806640625,
0.55126953125,
0.7666015625
],
[
0.69677734375,
0.755859375,
0.75244140625,
0.755859375,
0.75244140625,
0.767578125,
0.69677734375,
0.767578125
],
[
0.080078125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.080078125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.080078125,
0.77978515625,
0.4697265625,
0.77978515625,
0.4697265625,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.07843017578125,
0.79443359375,
0.44287109375,
0.79541015625,
0.44287109375,
0.8095703125,
0.07843017578125,
0.80810546875
],
[
0.5009765625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.5009765625,
0.8095703125
],
[
0.098876953125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.098876953125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.75341796875,
0.85693359375,
0.75341796875,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.521484375,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.521484375,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Pair-based losses enjoy rich and fine-grained data-to-",
"tive hardness of data and allow their embedding vectors to",
"data relations as they examine tuples (i.e., data pairs or their",
"interact with each other during training.",
"combinations)during training.However,since thenumber",
"of tuples increases polynomially with the number of train-",
"3. Our Method",
"ing data, their training complexity is prohibitively high and",
"We propose a new metric learning loss called Proxy-",
"convergence is slow.In addition,alargeamount of tuples",
"Anchor loss to overcome the inherent limitations of the pre-",
"are not effective and sometimes even degrade the quality",
"vious methods. The loss employs proxies that enable fast",
"of thelearned embedding space[25,37].Toaddressthis",
"and reliable convergence as in proxy-based losses. Also,",
"issue, most pair-based losses entail tuple sampling tech-",
"although it is built upon data-proxy relations, our loss can",
"niques [10, 25,37, 40] to select and utilize tuples that will",
"utilize data-to-data relations during training like pair-based",
"contribute to training. However, these techniques involve",
"losses sinceitenables embedding vectors ofdatapointsto",
"hyperparameters that have to be tuned carefully, and may",
"be affected by each other through its gradients. This prop-",
"increase the risk of overfitting since they rely mostly on lo-",
"erty of ourloss improvesthequality of thelearned embed-",
"cal pairwise relations within a batch. Another way to alle-",
"ding space substantially.",
"viating the complexity issue is to assign larger weights to",
"In this section, we first review Proxy-NCA loss [21],",
"more useful pairs during training as in [34], which however",
"a representative proxy-based loss, for comparison to our",
"also incorporates a sampling technique.",
"Proxy-Anchor loss. We then describe our Proxy-Anchor",
"Our loss resolves this complexity issue by adopting prox-",
"loss in detail and analyze its training complexity.",
"ies, which enables faster and more reliable convergence",
"compared to pair-based losses. Furthermore, it demands no",
"3.1. Review of Proxy-NCA Loss",
"additional hyperparameter for tuple sampling.",
"In the standard setting, Proxy-NCA loss [21] assigns a",
"2.2. Proxy-based Losses",
"proxy to each class so that the number of proxies is the same",
"with that of class labels. Given an input data point as an",
"Proxy-based metric learning is a relatively new approach",
"anchor, the proxy of the same class of the input is regarded",
"that can address the complexity issue of the pair-based",
"as positive and the other proxies are negative. Let r denote",
"losses. A proxy means a representative of a subset of train-",
"the embedding vector of the input, p+ be the positive proxy,",
"ing data and is estimated as a part of the embedding net-",
"and p- be a negative proxy. The loss is then given by",
"work parameters. The common idea of the methods in this",
"category is to infer a small set of proxies that capture the",
"es(r,pt)",
"l(X) = - log ",
"(1)",
"global structure of an embedding space and relate each data",
">",
"es(r,p-)",
"TEX",
"point with the proxies instead of the other data points dur-",
"P-EP",
"ing training. Since the number of proxies is significantly",
"- s(r,p+) + LSE s(a,p-)",
"",
"(2",
"EEX",
"smaller than that of training data, the training complexity",
"canbereduced substantially",
"The first proxy-based loss is Proxy-NCA [2]], which",
"where X is a batch of embedding vectors, P- is the set",
"is an approximation of Neighborhood Component Analy-",
"of negative proxies, and s(,) denotes the cosine similarity",
"sis (NCA) [8] using proxies. In its standard setting, Proxy-",
"between two vectors. In addition, LSE in Eq. (2) means the",
"NCA loss assigns a single proxy for each class, associates a",
"Log-Sum-Exp function, a smooth approximation to the max",
"data point with proxies, and encourages the positive pair to",
"function. The gradient of Proxy-NCA loss with respect to",
"be close and negative pairs to be far apart, as illustrated in",
"s(α,p) is given by",
"Figure 2(d). SoftTriple loss [23], an extension of SoftMax",
"-1",
"if p=p+,",
"al(x)",
"loss for classification, is similar to Proxy-NCA yet assigns",
"multiple proxies to each class to reflect intra-class variance.",
"es(r,p)",
"otherwise.",
"(3)",
"as(r,p)",
"es(z,p-)",
"ing proxies, and improves the performance by adopting a",
"Manifold Proxy loss [ll is an extension of N-pair loss us-",
"manifold-aware distance instead of Euclidean distance to",
"measure the semantic distance in the embedding space.",
"Eq. (3) shows that minimizing the loss encourages and",
"Using proxies in these losses helps improve training con-",
"p+ to be close to each other, and and p to be far away.",
"vergence greatly, but has an inherent limitation as a side",
"In particular, and p+ are pulled together by the constant",
"effect: Since each data point is associated only with prox-",
"power, while and p- closer to each other (i.e., harder neg-",
"ies, the rich data-to-data relations that are available for the",
"ative) are more strongly pushed away.",
"pair-based methods are not accessible anymore. Our loss",
"Proxy-NCA loss enables fast convergence thanks to its",
"can overcome this limitation since its gradients reflect rela-",
"low training complexity, O(MC) where M is the number"
] |
{
"arxiv_id": "2003.13911",
"language": "en",
"page_idx": 2
}
|
||
2003.13911-en-3
|
2003.13911-en-3.jpg
|
of training data and $C$ is that of classes, which is substantially lower than $O(M^2)$ or $O(M^3)$ of pair-based losses since $C\ll M;$ refer to Section 3.3 for details. Also, proxies are robust against outliers and noisy labels since they are trained to represent groups of data. However, since the loss associates each embedding vector only with proxies, it cannot exploit fine-grained data-to-data relations. This drawback limits the capability of embedding networks trained with Proxy-NCA loss.
### $3. 2. \textbf{Proxy- Anchor Loss}$
Our Proxy-Anchor loss is designed to overcome the limitation of Proxy-NCA while keeping the low training complexity. The main idea is to take each proxy as an anchor and associate it with entire data in a batch, as illustrated in Figure 2(e), so that the data interact with each other through the proxy anchor during training. Our loss assigns a proxy for each class following the standard proxy assignment setting of Proxy-NCA, and is formulated as
(4)
$$\begin{aligned}\ell(X)=&\frac{1}{|P^{+}|}\sum_{p\in P^{+}}\log\left(1+\sum_{x\in X_{p}^{+}}e^{-\alpha(s(x,p)-\delta)}\right)\\&+\frac{1}{|P|}\sum_{p\in P}\log\bigg(1+\sum_{x\in X_{p}^{-}}e^{\alpha(s(x,p)+\delta)}\bigg),\end{aligned}$$
where $\delta>0$ is a margin, $\alpha>0$ is a scaling factor, $P$ indicates the set of all proxies, and $P^+$ denotes the set of positive proxies of data in the batch. Also, for each proxy $\hat{p}$, a batch of embedding vectors $X$ is divided into two sets: $X_{p}^{+}$, the set of positive embedding vectors of $p$,and $X_p^{-}=$ $X^{r}-X_{p}^{+}.$ The proposed loss can be rewritten in an easierto-interpret form as
$$\begin{aligned}\ell(X)=&\frac{1}{|P^{+}|}\sum_{p\in P^{+}}\left[\mathrm{Softplus}\Big(\mathrm{LSE}_{x\in X_{p}^{+}}-\alpha(s(x,p)-\delta)\Big)\right]\\&+\frac{1}{|P|}\sum_{p\in P}\left[\mathrm{Softplus}\Big(\mathrm{LSE}_{x\in X_{p}^{-}}\alpha(s(x,p)+\delta)\Big)\right],\end{aligned}$$
where Softplus$(z)=\log{(1+e^{z})},\forall z\in\mathbb{R}$,and is a smooth
approximation of ReLU.
How it works: Regarding Log-Sum-Exp as the max function, it is easy to notice that the loss aims to pull $p$ and its most dissimilar positive example $(i.e.$, hardest positive example) together, and to push $p$ and its most similar negative example $(i.e.$, hardest negative example) apart. Due to the nature of Log-Sum-Exp, the loss in practice pulls and pushes all embedding vectors in the batch, but with different degrees of strength that are determined by their relative hardness. This characteristic is demonstrated by the gradi-
ent of our loss with respect to $s(x,p)$, which is given by
$$\dfrac{\partial\ell(X)}{\partial s(x,p)}=\begin{cases}\dfrac{1}{|P^+|}\dfrac{-\alpha\:h_p^+(x)}{1+\sum\limits_{x'\in X_p^+}h_p^+(x')},&\forall x\in X_p^+,\\[2ex]\dfrac{1}{|P|}\dfrac{\alpha\:h_p^-(x)}{1+\sum\limits_{x'\in X_p^-}h_p^-(x')},&\forall x\in X_p^-,\end{cases}$$
(6)
where $h_p^+(x)=e^{-\alpha(s(x,p)-\delta)}$ and $h_p^-(x)=e^{\alpha(s(x,p)+\delta)}$ are positive and negative hardness metrics for embedding vector $x$ given proxy $p$, respectively; $h_p^+(x)$ is large when the positive embedding vector $x$ is far from $p$,and $h_p^-(x)$ is large when the negative embedding vector $x$ is close to $p.$ The scaling parameter $\alpha$ and margin $\delta$ control the relative hardness of data points, and in consequence, determine how strongly pull or push their embedding vectors.
As shown in the above equations, the gradient for $s(x,p)$ is affected by not only $x$ but also other embedding vectors in the batch; the gradient becomes larger when $x$ is harder than the others. In this way, our loss enables embedding vectors in the batch to interact with each other and reflects their relative hardness through the gradients, which helps enhance the quality of the learned embedding space.
Comparison to Proxy-NCA: The key difference and advantage of Proxy-Anchor over Proxy-NCA is the active consideration of relative hardness based on data-to-data relations. This property enables Proxy-Anchor loss to provide richer supervisory signals to embedding networks during training. The gradients of the two losses demonstrate this clearly. In Proxy-NCA loss, the scale of the gradient is constant for every positive example and that of a negative example is calculated by taking only few proxies into account as shown in Eq. (3). In particular, the constant gradient scale for positive examples damages the flexibility and generalizability of embedding networks [37]. In contrast, Proxy-Anchor loss determines the scale of the gradient by taking relative hardness into consideration for both positive and negative examples as shown in Eq.(6). This feature of our loss allows the embedding network to consider data-to-data relations that are ignored in Proxy-NCA and observe much larger area of the embedding space during training than Proxy-NCA. Figure 3 illustrates these differences between the two losses in terms of handling the relative hardness of embedding vectors. In addition, unlike Proxy-Anchor loss, the margin imposed in our loss leads to intra-class compactness and inter-class separability, resulting in a more discriminative embedding space.
## 3.3. Training Complexity Analysis
Let $M,C,B$,and $U$ denote the numbers of training samples, classes, batches per epoch, and proxies held by each
|
[
[
0.08087158203125,
0.09027099609375,
0.46728515625,
0.09027099609375,
0.46728515625,
0.10479736328125,
0.08087158203125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.86767578125,
0.0921630859375,
0.86767578125,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.106689453125,
0.4697265625,
0.106689453125,
0.4697265625,
0.12060546875,
0.080078125,
0.12060546875
],
[
0.619140625,
0.12060546875,
0.63134765625,
0.12060546875,
0.63134765625,
0.1300048828125,
0.619140625,
0.1300048828125
],
[
0.67236328125,
0.11614990234375,
0.74609375,
0.11614990234375,
0.74609375,
0.1337890625,
0.67236328125,
0.1337890625
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.7861328125,
0.126953125,
0.861328125,
0.126953125,
0.861328125,
0.14453125,
0.7861328125,
0.14453125
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.60888671875,
0.1331787109375,
0.775390625,
0.136962890625,
0.77392578125,
0.1590576171875,
0.60693359375,
0.1546630859375
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.51708984375,
0.1546630859375,
0.56640625,
0.1572265625,
0.564453125,
0.173583984375,
0.515625,
0.1717529296875
],
[
0.6748046875,
0.1572265625,
0.72119140625,
0.1572265625,
0.72119140625,
0.17236328125,
0.6748046875,
0.17236328125
],
[
0.080078125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.86767578125,
0.1654052734375,
0.89453125,
0.1654052734375,
0.89453125,
0.1817626953125,
0.86767578125,
0.1817626953125
],
[
0.51318359375,
0.173583984375,
0.57177734375,
0.175537109375,
0.5703125,
0.1912841796875,
0.51123046875,
0.189453125
],
[
0.61767578125,
0.179931640625,
0.625,
0.179931640625,
0.625,
0.1875,
0.61767578125,
0.1875
],
[
0.66650390625,
0.175537109375,
0.7294921875,
0.175537109375,
0.7294921875,
0.1912841796875,
0.66650390625,
0.1912841796875
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.7890625,
0.179931640625,
0.86376953125,
0.1875,
0.861328125,
0.2052001953125,
0.7861328125,
0.1976318359375
],
[
0.6044921875,
0.189453125,
0.7646484375,
0.1925048828125,
0.7646484375,
0.2171630859375,
0.6044921875,
0.2139892578125
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.080078125,
0.21337890625,
0.2279052734375,
0.21337890625,
0.2279052734375,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.66845703125,
0.2108154296875,
0.70751953125,
0.216552734375,
0.705078125,
0.231689453125,
0.6640625,
0.2266845703125
],
[
0.07843017578125,
0.2392578125,
0.26318359375,
0.240478515625,
0.26318359375,
0.255126953125,
0.07843017578125,
0.25390625
],
[
0.5009765625,
0.238037109375,
0.8935546875,
0.238037109375,
0.8935546875,
0.2587890625,
0.5009765625,
0.2587890625
],
[
0.5009765625,
0.25830078125,
0.89208984375,
0.2568359375,
0.89208984375,
0.271484375,
0.5009765625,
0.272705078125
],
[
0.098876953125,
0.26513671875,
0.468017578125,
0.26513671875,
0.468017578125,
0.279052734375,
0.098876953125,
0.279052734375
],
[
0.07843017578125,
0.279052734375,
0.468017578125,
0.27978515625,
0.468017578125,
0.294189453125,
0.07843017578125,
0.29345703125
],
[
0.5009765625,
0.2734375,
0.89208984375,
0.272705078125,
0.89208984375,
0.287353515625,
0.5009765625,
0.28857421875
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.29541015625,
0.4697265625,
0.29541015625,
0.4697265625,
0.309326171875,
0.080078125,
0.309326171875
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.080078125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.323974609375,
0.080078125,
0.323974609375
],
[
0.50341796875,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50341796875,
0.33203125
],
[
0.080078125,
0.32568359375,
0.468017578125,
0.32568359375,
0.468017578125,
0.339111328125,
0.080078125,
0.339111328125
],
[
0.07843017578125,
0.339111328125,
0.468017578125,
0.34033203125,
0.468017578125,
0.354736328125,
0.07843017578125,
0.354248046875
],
[
0.50244140625,
0.333984375,
0.89208984375,
0.333984375,
0.89208984375,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.50244140625,
0.349853515625,
0.80810546875,
0.349853515625,
0.80810546875,
0.363037109375,
0.50244140625,
0.363037109375
],
[
0.080078125,
0.35595703125,
0.4697265625,
0.35595703125,
0.4697265625,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.521484375,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.37744140625,
0.521484375,
0.37744140625
],
[
0.07843017578125,
0.37060546875,
0.3505859375,
0.369384765625,
0.3505859375,
0.3837890625,
0.07843017578125,
0.385009765625
],
[
0.5048828125,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.391357421875,
0.5048828125,
0.391357421875
],
[
0.19287109375,
0.40087890625,
0.256591796875,
0.4072265625,
0.253173828125,
0.425537109375,
0.1895751953125,
0.419189453125
],
[
0.5048828125,
0.395751953125,
0.8896484375,
0.395751953125,
0.8896484375,
0.406005859375,
0.5048828125,
0.406005859375
],
[
0.101318359375,
0.4091796875,
0.1405029296875,
0.4091796875,
0.1405029296875,
0.42041015625,
0.101318359375,
0.42041015625
],
[
0.26318359375,
0.404052734375,
0.423095703125,
0.396484375,
0.4248046875,
0.417236328125,
0.2646484375,
0.4248046875
],
[
0.1544189453125,
0.415283203125,
0.1871337890625,
0.415283203125,
0.1871337890625,
0.4267578125,
0.1544189453125,
0.4267578125
],
[
0.5009765625,
0.410400390625,
0.8896484375,
0.410400390625,
0.8896484375,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.285888671875,
0.425537109375,
0.328369140625,
0.425537109375,
0.328369140625,
0.43994140625,
0.285888671875,
0.43994140625
],
[
0.50341796875,
0.4248046875,
0.89208984375,
0.423583984375,
0.89208984375,
0.4375,
0.50341796875,
0.438232421875
],
[
0.44677734375,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.447509765625,
0.44677734375,
0.447509765625
],
[
0.19287109375,
0.4267578125,
0.2197265625,
0.4267578125,
0.2197265625,
0.434326171875,
0.19287109375,
0.434326171875
],
[
0.50244140625,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.1953125,
0.4462890625,
0.41259765625,
0.443115234375,
0.41259765625,
0.46728515625,
0.1953125,
0.47021484375
],
[
0.5009765625,
0.453857421875,
0.84619140625,
0.455078125,
0.84619140625,
0.4697265625,
0.5009765625,
0.468505859375
],
[
0.17236328125,
0.462158203125,
0.1953125,
0.45703125,
0.1993408203125,
0.468505859375,
0.176513671875,
0.473388671875
],
[
0.200927734375,
0.4697265625,
0.2181396484375,
0.472900390625,
0.21728515625,
0.479248046875,
0.19775390625,
0.47607421875
],
[
0.291748046875,
0.4716796875,
0.324462890625,
0.4716796875,
0.324462890625,
0.48291015625,
0.291748046875,
0.48291015625
],
[
0.50341796875,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.080078125,
0.49755859375,
0.4697265625,
0.49755859375,
0.4697265625,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.5009765625,
0.492431640625,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.5009765625,
0.505859375
],
[
0.50244140625,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.080078125,
0.51318359375,
0.471435546875,
0.51318359375,
0.471435546875,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.5009765625,
0.52294921875,
0.890625,
0.52294921875,
0.890625,
0.5361328125,
0.5009765625,
0.5361328125
],
[
0.080078125,
0.52880859375,
0.468017578125,
0.52880859375,
0.468017578125,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.50244140625,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.55029296875,
0.50244140625,
0.55029296875
],
[
0.080078125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.080078125,
0.556640625
],
[
0.50244140625,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50244140625,
0.56494140625
],
[
0.08087158203125,
0.55810546875,
0.46728515625,
0.556640625,
0.46728515625,
0.5712890625,
0.08087158203125,
0.57275390625
],
[
0.5009765625,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.580078125,
0.5009765625,
0.580078125
],
[
0.08251953125,
0.57373046875,
0.46728515625,
0.57373046875,
0.46728515625,
0.5869140625,
0.08251953125,
0.5869140625
],
[
0.080078125,
0.5888671875,
0.2115478515625,
0.5888671875,
0.2115478515625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50341796875,
0.58349609375,
0.8896484375,
0.58349609375,
0.8896484375,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.0906982421875,
0.625,
0.1356201171875,
0.625,
0.1356201171875,
0.6396484375,
0.0906982421875,
0.6396484375
],
[
0.2181396484375,
0.619140625,
0.43701171875,
0.6220703125,
0.43701171875,
0.642578125,
0.2181396484375,
0.64013671875
],
[
0.1396484375,
0.63330078125,
0.178955078125,
0.63330078125,
0.178955078125,
0.64453125,
0.1396484375,
0.64453125
],
[
0.50244140625,
0.62744140625,
0.890625,
0.626953125,
0.890625,
0.6416015625,
0.50244140625,
0.642578125
],
[
0.291748046875,
0.6396484375,
0.330810546875,
0.6396484375,
0.330810546875,
0.646484375,
0.291748046875,
0.646484375
],
[
0.50244140625,
0.642578125,
0.8896484375,
0.642578125,
0.8896484375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.50244140625,
0.65771484375,
0.890625,
0.6572265625,
0.890625,
0.671875,
0.50244140625,
0.67236328125
],
[
0.178955078125,
0.64453125,
0.2156982421875,
0.642578125,
0.21728515625,
0.65283203125,
0.1805419921875,
0.65478515625
],
[
0.2156982421875,
0.662109375,
0.435546875,
0.662109375,
0.435546875,
0.6845703125,
0.2156982421875,
0.6845703125
],
[
0.50341796875,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50341796875,
0.6875
],
[
0.295654296875,
0.681640625,
0.32275390625,
0.681640625,
0.32275390625,
0.689453125,
0.295654296875,
0.689453125
],
[
0.5009765625,
0.68798828125,
0.89208984375,
0.6875,
0.89208984375,
0.7021484375,
0.5009765625,
0.70263671875
],
[
0.44677734375,
0.69775390625,
0.4697265625,
0.69775390625,
0.4697265625,
0.71044921875,
0.44677734375,
0.71044921875
],
[
0.5048828125,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71533203125,
0.5048828125,
0.71533203125
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.7412109375,
0.08087158203125,
0.7412109375
],
[
0.50341796875,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.7421875,
0.24267578125,
0.7412109375,
0.24267578125,
0.755859375,
0.080078125,
0.7568359375
],
[
0.5009765625,
0.74755859375,
0.890625,
0.7470703125,
0.890625,
0.76123046875,
0.5009765625,
0.76220703125
],
[
0.080078125,
0.7646484375,
0.468017578125,
0.76513671875,
0.468017578125,
0.77978515625,
0.080078125,
0.77880859375
],
[
0.50244140625,
0.76318359375,
0.89208984375,
0.76318359375,
0.89208984375,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.80615234375,
0.50341796875,
0.80615234375
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.8095703125,
0.890625,
0.8095703125,
0.890625,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82470703125,
0.80810546875,
0.82470703125,
0.80810546875,
0.83837890625,
0.5009765625,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84765625,
0.765625,
0.84765625,
0.765625,
0.8623046875,
0.50244140625,
0.8623046875
],
[
0.08087158203125,
0.85791015625,
0.465576171875,
0.85791015625,
0.465576171875,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.51953125,
0.869140625,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.51953125,
0.88330078125
],
[
0.07843017578125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.07843017578125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"of training data and C is that of classes, which is substan-",
"ent of our loss with respect to s(r, p), which is given by",
"tially lower than O(M2) or O(M3) of pair-based losses",
"1",
"-α ht(r)",
"since C < M; refer to Section 3.3 for details. Also, prox-",
"VaeXt,",
"ies are robust against outliers and noisy labels since they are",
"[P+I1+Z (a)",
"trained to represent groups of data. However, since the loss",
"al(x)",
"t'ext",
"associates each embedding vector only with proxies, it can-",
"(6)",
"as(r,p)",
"1",
"αh(α)",
"not exploit fine-grained data-to-data relations. This draw-",
"VaEXp,",
"Pl i+ hp(al)",
"back limits the capability of embedding networks trained",
"with Proxy-NCA loss.",
"I'EX,",
"3.2. Proxy-Anchor Loss",
"where ht(a) = e-a(s(a,p)-0) and h:(r) = ea(a,p)+3)",
"are positive and negative hardness metrics for embedding",
"Our Proxy-Anchor loss is designed to overcome the lim-",
"itation of Proxy-NCA while keeping the low training com-",
"vector given proxy p, respectively; h(r) is large when",
"the positive embedding vector is far from p, and h, (c) is",
"plexity. The main idea is to take each proxy as an anchor",
"large when the negative embedding vector r is close to p.",
"and associate it with entire data in a batch, as illustrated in",
"The scaling parameter Q and margin control the relative",
"Figure 2(e), so that the data interact with each other through",
"the proxy anchor during training. Our loss assigns a proxy",
"hardness of data points, and in consequence, determine how",
"strongly pull or push their embedding vectors.",
"for each class following the standard proxy assignment set-",
"As shown in the above equations, the gradient for s(r, p)",
"ting of Proxy-NCA, and is formulated as",
"isaffectedbynotonlybutalsootherembeddingvectors",
" og",
"inthebatch:thegradientbecomeslargerwhen isharder",
"(X) :",
"1 + e-a(s(m,p)-8)",
"P+",
"than the others. In this way, our loss enables embedding",
"rext",
"vectors in the batch to interact with each other and reflects",
"(4)",
"DEP-",
"their relative hardness through the gradients, which helps",
"log (1+ ea(s(a,p)+o)",
"enhance the quality of the learned embedding space.",
"PI",
"nF",
"TEX,",
"Comparison to Proxy-NCA: The key difference and ad-",
"where > 0 is a margin, α > 0 is a scaling factor, P",
"vantageof Proxy-AnchoroverProxy-NCA isthe active",
"consideration of relative hardness based on data-to-data re-",
"indicates the set of all proxies, and P+ denotes the set of",
"lations. This property enables Proxy-Anchor loss to pro-",
"positive proxies of data in the batch. Also, for each proxy",
"vide richer supervisory signals to embedding networks dur-",
"p, a batch of embedding vectors X is divided into two sets:",
"ing training.The gradients of the two losses demonstrate",
"Xt, the set of positive embedding vectors of p, and X,",
"this clearly. In Proxy-NCA loss, the scale of the gradient",
"X - X+. The proposed loss can be rewritten in an easier-",
"to-interpret form as",
"is constant for every positive example and that of a nega-",
"tive example is calculated by taking only few proxies into",
"account as shown in Eq. (3). In particular, the constant",
"l(X) :",
"[Softplus( LSE -α(s(r,P) -8)",
"P+I",
"gradient scale for positive examples damages the flexibility",
"TEX+",
"and generalizability of embedding networks [37].In con-",
"trast, Proxy-Anchor loss determines the scale of the gradi-",
"PEP+",
"Softplus(LSE a(s(a,P) +0)",
"ent by taking relative hardness into consideration for both",
"TEX",
"positive and negative examples as shown in Eq. (6). This",
"(5)",
"featureofourlossallowstheembeddingnetworktocon-",
"sider data-to-data relations that are ignored in Proxy-NCA",
"where Softplus(z) = log (1 + e), Vz e IR, and is a smooth",
"and observe much larger area of the embedding space dur-",
"approximation of ReLU.",
"ing training than Proxy-NCA. Figure 3 illustrates these dif-",
"How it works: Regarding Log-Sum-Exp as the max func-",
"ferences between the two losses in terms of handling the",
"tion, it is easy to notice that the loss aims to pull p and its",
"relative hardness of embedding vectors. In addition, unlike",
"most dissimilar positive example (i.e., hardest positive ex-",
"Proxy-Anchorloss,themarginimposedinourlossleadsto",
"ample) together, and to push p and its most similar nega-",
"intra-class compactness and inter-class separability, result-",
"tive example (i.e., hardest negative example) apart. Due to",
"ing in a more discriminative embedding space.",
"the nature of Log-Sum-Exp, the loss in practice pulls and",
"3.3. Training Complexity Analysis",
"pushes all embedding vectors inthe batch, but with differ",
"ent degrees of strength that are determined by their relative",
"Let M.C.B, and U denote the numbers of training sam-",
"hardness. This characteristic is demonstrated by the gradi-",
"ples, classes, batches per epoch, and proxies held by each"
] |
{
"arxiv_id": "2003.13911",
"language": "en",
"page_idx": 3
}
|
||
2003.13911-en-4
|
2003.13911-en-4.jpg
|
Case of Positive Examples
Case of Negative Examples
<FigureHere>
Figure 3. Differences between Proxy-NCA and Proxy-Anchor in handling proxies and embedding vectors during training. Each proxy is colored in black and three different colors indicate distinct classes. The associations defined by the losses are expressed by edges, and thicker edges get larger gradients. (a) Gradients of Proxy-NCA loss with respect to positive examples have the same scale regardless of their hardness. (b) Proxy-Anchor loss dynamically determines gradient scales regarding relative hardness of all positive examples so as to pull harder positives more strongly.(c) In Proxy-NCA, each negative example is pushed only by a small number of proxies without considering the distribution of embedding vectors in fine details. (d) Proxy-Anchor loss considers the distribution of embedding vectors in more details as it has all negative examples affect each other in their gradients.
<table>
<tbody>
<tr>
<th>$\underline{{\mathrm{Type}}}$</th>
<th>Loss</th>
<th>${\mathrm{Training}}$ Complexity</th>
</tr>
<tr>
<td>Proxy</td>
<td>Proxy- Anchor ( Ours) Proxy- NCA $\Lambda[21]$ SoftTriple [23]</td>
<td>$O(MC)$ $O(MC)$ $O(MCU^{2}$</td>
</tr>
<tr>
<td>Pair</td>
<td>Contrastive [2,4,9] Triplet (Semi-Hard) )[25] Triplet(Smart)[10] $N-$pair $•$ [27] Lifted Structure [29]</td>
<td>${\overline{O(M^{2})}}$ $O(M^{3}/B^{2})$ $O(M^{2})$ $O(M^{3}$ $O(M^{3}$</td>
</tr>
</tbody>
</table>
Table 1. Comparison of training complexities.
class, respectively. $U$ is 1 thus ignored in most of proxybased losses including ours, but is nontrivial for those managing multiple proxies per class such as SoftTriple loss [23].
Table 1 compares the training complexity of our loss with those of popular pair- and proxy-based losses. The complexity of our loss is $O(MC)$ since it compares every proxy with all positive or all negative examples in a batch. $\hat{\text{More specifically, in Eq.(4), the complexity of the first sum-}}$ mation is $O(MC)$ and that of the second summation is also $O(MC)$, hence the total training complexity is $O(MC).$ The complexity of Proxy-NCA[21] is also $\dot{O}(MC)$ since each data point is associated with one positive proxy and $C-1$ negative proxies as can be seen in Eq.(2). On the other hand, SoftTriple loss [23], a modification of SoftMax using multiple proxies per class, associates each data point with $U$ positive proxies and $U(C-1)$ negative proxies. The total training complexity of this loss is thus $O(MCU^{2}).$ In conclusion, the complexity of our loss is the same with or even lower than that of other proxy-based losses.
The training complexity of pair-based losses is higher than that of proxy-based ones. Since Contrastive loss [2, 4,9] takes a pair of data as input, its training complexity is $O(M^{2}).$ On the other hand, Triplet loss that examines triplets of data has complexity $O(\hat{M}^{3})$, which can be reduced by triplet mining strategies. For example, semi-hard mining[25] reduces the complexity to $O(\tilde{M}^3/B^2)$ by selecting negative pairs that are located within a neighborhood of anchor but sufficiently far from it. Similarly, Smart mining [10] lowers the complexity to $O(M^2)$ by sampling
hard triplets using an approximated nearest neighbor index. $\hat{\text{However, even with these techniques, the training complex-}}$ ity of Triplet loss is still high. Like Triplet loss, $N$-pair loss [27] and Lifted Structure loss [29] that compare each positive pair of data to multiple negative pairs also have $\hat{\text{complexity}}O(M^{3}).$ The training complexity of these losses becomes prohibitively high as the number of training data $M$ increases, which slows down the speed of convergence as demonstrated in Figure l.
## $4. \textbf{Experiments}$
In this section, our method is evaluated and compared to current state-of-the-art on the four benchmark datasets for deep metric learning [17, 19, 29, 36]. We also investigate the effect of hyperparameters and embedding dimensionality of our loss to demonstrate its robustness.
|
[
[
0.181396484375,
0.1029052734375,
0.374267578125,
0.1029052734375,
0.374267578125,
0.11614990234375,
0.181396484375,
0.11614990234375
],
[
0.580078125,
0.10101318359375,
0.78369140625,
0.10162353515625,
0.78369140625,
0.11614990234375,
0.580078125,
0.11553955078125
],
[
0.125,
0.1319580078125,
0.1405029296875,
0.1395263671875,
0.131591796875,
0.1514892578125,
0.11517333984375,
0.14453125
],
[
0.1968994140625,
0.1319580078125,
0.2197265625,
0.1319580078125,
0.2197265625,
0.146484375,
0.1968994140625,
0.146484375
],
[
0.38818359375,
0.1319580078125,
0.408447265625,
0.1319580078125,
0.408447265625,
0.146484375,
0.38818359375,
0.146484375
],
[
0.6005859375,
0.1307373046875,
0.625,
0.1307373046875,
0.625,
0.1483154296875,
0.6005859375,
0.1483154296875
],
[
0.79150390625,
0.1319580078125,
0.81201171875,
0.1319580078125,
0.81201171875,
0.1458740234375,
0.79150390625,
0.1458740234375
],
[
0.52294921875,
0.1337890625,
0.54150390625,
0.1337890625,
0.54150390625,
0.15087890625,
0.52294921875,
0.15087890625
],
[
0.30810546875,
0.1351318359375,
0.328369140625,
0.1351318359375,
0.328369140625,
0.15087890625,
0.30810546875,
0.15087890625
],
[
0.71337890625,
0.1351318359375,
0.73193359375,
0.1351318359375,
0.73193359375,
0.15087890625,
0.71337890625,
0.15087890625
],
[
0.41015625,
0.1427001953125,
0.42724609375,
0.1427001953125,
0.42724609375,
0.1572265625,
0.41015625,
0.1572265625
],
[
0.62353515625,
0.1427001953125,
0.640625,
0.1427001953125,
0.640625,
0.1558837890625,
0.62353515625,
0.1558837890625
],
[
0.4794921875,
0.14453125,
0.484375,
0.14453125,
0.484375,
0.1514892578125,
0.4794921875,
0.1514892578125
],
[
0.221435546875,
0.143310546875,
0.2344970703125,
0.143310546875,
0.2344970703125,
0.1546630859375,
0.221435546875,
0.1546630859375
],
[
0.1396484375,
0.149658203125,
0.15283203125,
0.149658203125,
0.15283203125,
0.1590576171875,
0.1396484375,
0.1590576171875
],
[
0.330078125,
0.149658203125,
0.346435546875,
0.149658203125,
0.346435546875,
0.1590576171875,
0.330078125,
0.1590576171875
],
[
0.5439453125,
0.1483154296875,
0.55810546875,
0.1483154296875,
0.55810546875,
0.1590576171875,
0.5439453125,
0.1590576171875
],
[
0.736328125,
0.149658203125,
0.74853515625,
0.149658203125,
0.74853515625,
0.1590576171875,
0.736328125,
0.1590576171875
],
[
0.1829833984375,
0.1558837890625,
0.1953125,
0.1558837890625,
0.1953125,
0.1654052734375,
0.1829833984375,
0.1654052734375
],
[
0.47802734375,
0.15283203125,
0.484375,
0.15283203125,
0.484375,
0.160400390625,
0.47802734375,
0.160400390625
],
[
0.5888671875,
0.1558837890625,
0.6005859375,
0.1558837890625,
0.6005859375,
0.1654052734375,
0.5888671875,
0.1654052734375
],
[
0.431396484375,
0.1578369140625,
0.450927734375,
0.1578369140625,
0.450927734375,
0.17236328125,
0.431396484375,
0.17236328125
],
[
0.2418212890625,
0.1590576171875,
0.2607421875,
0.1590576171875,
0.2607421875,
0.17236328125,
0.2418212890625,
0.17236328125
],
[
0.369384765625,
0.1572265625,
0.381591796875,
0.154052734375,
0.3857421875,
0.1641845703125,
0.373291015625,
0.1673583984375
],
[
0.48193359375,
0.160400390625,
0.488525390625,
0.1666259765625,
0.48193359375,
0.17041015625,
0.4755859375,
0.1641845703125
],
[
0.646484375,
0.1590576171875,
0.6640625,
0.1590576171875,
0.6640625,
0.17236328125,
0.646484375,
0.17236328125
],
[
0.71826171875,
0.1610107421875,
0.736328125,
0.1610107421875,
0.736328125,
0.175537109375,
0.71826171875,
0.175537109375
],
[
0.313720703125,
0.1610107421875,
0.33251953125,
0.1610107421875,
0.33251953125,
0.1767578125,
0.313720703125,
0.1767578125
],
[
0.12335205078125,
0.1622314453125,
0.1405029296875,
0.1622314453125,
0.1405029296875,
0.175537109375,
0.12335205078125,
0.175537109375
],
[
0.52685546875,
0.1622314453125,
0.54736328125,
0.1622314453125,
0.54736328125,
0.175537109375,
0.52685546875,
0.175537109375
],
[
0.48046875,
0.169189453125,
0.48779296875,
0.1749267578125,
0.483642578125,
0.177978515625,
0.476318359375,
0.17236328125
],
[
0.8408203125,
0.1641845703125,
0.8486328125,
0.1641845703125,
0.8486328125,
0.169189453125,
0.8408203125,
0.169189453125
],
[
0.58642578125,
0.1767578125,
0.60302734375,
0.1767578125,
0.60302734375,
0.186279296875,
0.58642578125,
0.186279296875
],
[
0.42236328125,
0.186279296875,
0.439453125,
0.186279296875,
0.439453125,
0.199462890625,
0.42236328125,
0.199462890625
],
[
0.61279296875,
0.1912841796875,
0.6259765625,
0.1912841796875,
0.6259765625,
0.20263671875,
0.61279296875,
0.20263671875
],
[
0.63818359375,
0.189453125,
0.65185546875,
0.189453125,
0.65185546875,
0.1988525390625,
0.63818359375,
0.1988525390625
],
[
0.5146484375,
0.1925048828125,
0.53173828125,
0.1925048828125,
0.53173828125,
0.205810546875,
0.5146484375,
0.205810546875
],
[
0.11273193359375,
0.1932373046875,
0.125,
0.1932373046875,
0.125,
0.203857421875,
0.11273193359375,
0.203857421875
],
[
0.20751953125,
0.1932373046875,
0.22216796875,
0.1932373046875,
0.22216796875,
0.2020263671875,
0.20751953125,
0.2020263671875
],
[
0.301513671875,
0.1925048828125,
0.318603515625,
0.1925048828125,
0.318603515625,
0.205810546875,
0.301513671875,
0.205810546875
],
[
0.398681640625,
0.1932373046875,
0.414306640625,
0.1932373046875,
0.414306640625,
0.2020263671875,
0.398681640625,
0.2020263671875
],
[
0.8037109375,
0.1932373046875,
0.81640625,
0.1932373046875,
0.81640625,
0.20263671875,
0.8037109375,
0.20263671875
],
[
0.71142578125,
0.1944580078125,
0.7197265625,
0.1976318359375,
0.7158203125,
0.203857421875,
0.70751953125,
0.2008056640625
],
[
0.1829833984375,
0.1956787109375,
0.1968994140625,
0.205810546875,
0.1829833984375,
0.21533203125,
0.1690673828125,
0.2052001953125
],
[
0.373291015625,
0.1944580078125,
0.386474609375,
0.2052001953125,
0.373291015625,
0.21533203125,
0.359375,
0.2052001953125
],
[
0.57861328125,
0.1976318359375,
0.59716796875,
0.1976318359375,
0.59716796875,
0.212158203125,
0.57861328125,
0.212158203125
],
[
0.64892578125,
0.199462890625,
0.6689453125,
0.199462890625,
0.6689453125,
0.212158203125,
0.64892578125,
0.212158203125
],
[
0.438720703125,
0.2008056640625,
0.450927734375,
0.2008056640625,
0.450927734375,
0.208984375,
0.438720703125,
0.208984375
],
[
0.2484130859375,
0.2020263671875,
0.2607421875,
0.2020263671875,
0.2607421875,
0.210205078125,
0.2484130859375,
0.210205078125
],
[
0.77294921875,
0.20263671875,
0.78125,
0.20263671875,
0.78125,
0.2083740234375,
0.77294921875,
0.2083740234375
],
[
0.8447265625,
0.203857421875,
0.8544921875,
0.203857421875,
0.8544921875,
0.2083740234375,
0.8447265625,
0.2083740234375
],
[
0.19775390625,
0.222900390625,
0.21728515625,
0.222900390625,
0.21728515625,
0.236083984375,
0.19775390625,
0.236083984375
],
[
0.38818359375,
0.222900390625,
0.408447265625,
0.2215576171875,
0.41015625,
0.236083984375,
0.3896484375,
0.2374267578125
],
[
0.5986328125,
0.222900390625,
0.61767578125,
0.2197265625,
0.62353515625,
0.23486328125,
0.6044921875,
0.238037109375
],
[
0.79150390625,
0.222900390625,
0.810546875,
0.222900390625,
0.810546875,
0.236083984375,
0.79150390625,
0.236083984375
],
[
0.30810546875,
0.2259521484375,
0.328369140625,
0.2259521484375,
0.328369140625,
0.240478515625,
0.30810546875,
0.240478515625
],
[
0.1192626953125,
0.2259521484375,
0.1380615234375,
0.2259521484375,
0.1380615234375,
0.240478515625,
0.1192626953125,
0.240478515625
],
[
0.52294921875,
0.2259521484375,
0.54150390625,
0.2259521484375,
0.54150390625,
0.2412109375,
0.52294921875,
0.2412109375
],
[
0.71337890625,
0.2259521484375,
0.73193359375,
0.2259521484375,
0.73193359375,
0.240478515625,
0.71337890625,
0.240478515625
],
[
0.5458984375,
0.23486328125,
0.56201171875,
0.240478515625,
0.55615234375,
0.25244140625,
0.5400390625,
0.248779296875
],
[
0.6259765625,
0.23486328125,
0.6396484375,
0.23486328125,
0.6396484375,
0.246826171875,
0.6259765625,
0.246826171875
],
[
0.81640625,
0.236083984375,
0.830078125,
0.236083984375,
0.830078125,
0.246826171875,
0.81640625,
0.246826171875
],
[
0.2181396484375,
0.2374267578125,
0.2320556640625,
0.23291015625,
0.23779296875,
0.24365234375,
0.223876953125,
0.2474365234375
],
[
0.48046875,
0.238037109375,
0.486083984375,
0.238037109375,
0.486083984375,
0.2442626953125,
0.48046875,
0.2442626953125
],
[
0.326904296875,
0.2392578125,
0.34716796875,
0.2392578125,
0.34716796875,
0.251953125,
0.326904296875,
0.251953125
],
[
0.416748046875,
0.2310791015625,
0.4306640625,
0.2412109375,
0.419189453125,
0.250732421875,
0.406005859375,
0.240478515625
],
[
0.1396484375,
0.240478515625,
0.1544189453125,
0.240478515625,
0.1544189453125,
0.250732421875,
0.1396484375,
0.250732421875
],
[
0.48046875,
0.246826171875,
0.48779296875,
0.246826171875,
0.48779296875,
0.25390625,
0.48046875,
0.25390625
],
[
0.736328125,
0.24365234375,
0.74609375,
0.240478515625,
0.74853515625,
0.24560546875,
0.73876953125,
0.248779296875
],
[
0.42724609375,
0.2474365234375,
0.451904296875,
0.2474365234375,
0.451904296875,
0.26513671875,
0.42724609375,
0.26513671875
],
[
0.181396484375,
0.2474365234375,
0.1953125,
0.2474365234375,
0.1953125,
0.2568359375,
0.181396484375,
0.2568359375
],
[
0.58642578125,
0.2474365234375,
0.6005859375,
0.2474365234375,
0.6005859375,
0.2568359375,
0.58642578125,
0.2568359375
],
[
0.78125,
0.248779296875,
0.7900390625,
0.248779296875,
0.7900390625,
0.255615234375,
0.78125,
0.255615234375
],
[
0.365234375,
0.25,
0.38232421875,
0.242431640625,
0.390625,
0.25390625,
0.373291015625,
0.261474609375
],
[
0.24267578125,
0.25,
0.2607421875,
0.25,
0.2607421875,
0.26318359375,
0.24267578125,
0.26318359375
],
[
0.64794921875,
0.250732421875,
0.6611328125,
0.250732421875,
0.6611328125,
0.261962890625,
0.64794921875,
0.261962890625
],
[
0.8388671875,
0.251953125,
0.85107421875,
0.251953125,
0.85107421875,
0.261962890625,
0.8388671875,
0.261962890625
],
[
0.313720703125,
0.25244140625,
0.33251953125,
0.25244140625,
0.33251953125,
0.267578125,
0.313720703125,
0.267578125
],
[
0.12335205078125,
0.25244140625,
0.1422119140625,
0.25244140625,
0.1422119140625,
0.267578125,
0.12335205078125,
0.267578125
],
[
0.52685546875,
0.25244140625,
0.54833984375,
0.25244140625,
0.54833984375,
0.266357421875,
0.52685546875,
0.266357421875
],
[
0.7197265625,
0.255126953125,
0.73388671875,
0.255126953125,
0.73388671875,
0.26318359375,
0.7197265625,
0.26318359375
],
[
0.369384765625,
0.26318359375,
0.39208984375,
0.267578125,
0.38818359375,
0.27978515625,
0.365234375,
0.27587890625
],
[
0.48193359375,
0.255615234375,
0.48779296875,
0.255615234375,
0.48779296875,
0.26318359375,
0.48193359375,
0.26318359375
],
[
0.1871337890625,
0.267578125,
0.19775390625,
0.267578125,
0.19775390625,
0.276611328125,
0.1871337890625,
0.276611328125
],
[
0.48046875,
0.267578125,
0.48779296875,
0.267578125,
0.48779296875,
0.2734375,
0.48046875,
0.2734375
],
[
0.5888671875,
0.267578125,
0.60302734375,
0.267578125,
0.60302734375,
0.27783203125,
0.5888671875,
0.27783203125
],
[
0.48193359375,
0.27587890625,
0.486083984375,
0.27587890625,
0.486083984375,
0.2822265625,
0.48193359375,
0.2822265625
],
[
0.236083984375,
0.2822265625,
0.2459716796875,
0.2822265625,
0.2459716796875,
0.2890625,
0.236083984375,
0.2890625
],
[
0.398681640625,
0.2822265625,
0.414306640625,
0.2822265625,
0.414306640625,
0.292236328125,
0.398681640625,
0.292236328125
],
[
0.42724609375,
0.2822265625,
0.43701171875,
0.2822265625,
0.43701171875,
0.2890625,
0.42724609375,
0.2890625
],
[
0.63818359375,
0.27978515625,
0.650390625,
0.27978515625,
0.650390625,
0.2890625,
0.63818359375,
0.2890625
],
[
0.80126953125,
0.2822265625,
0.833984375,
0.2822265625,
0.833984375,
0.29345703125,
0.80126953125,
0.29345703125
],
[
0.61279296875,
0.28271484375,
0.6259765625,
0.28271484375,
0.6259765625,
0.29345703125,
0.61279296875,
0.29345703125
],
[
0.2091064453125,
0.28271484375,
0.223876953125,
0.28271484375,
0.223876953125,
0.292236328125,
0.2091064453125,
0.292236328125
],
[
0.2998046875,
0.2841796875,
0.318603515625,
0.2841796875,
0.318603515625,
0.29736328125,
0.2998046875,
0.29736328125
],
[
0.515625,
0.285400390625,
0.53125,
0.285400390625,
0.53125,
0.29541015625,
0.515625,
0.29541015625
],
[
0.1829833984375,
0.287353515625,
0.1953125,
0.296630859375,
0.1846923828125,
0.30615234375,
0.1707763671875,
0.29541015625
],
[
0.11273193359375,
0.285888671875,
0.125,
0.285888671875,
0.125,
0.294189453125,
0.11273193359375,
0.294189453125
],
[
0.366943359375,
0.290283203125,
0.381591796875,
0.290283203125,
0.381591796875,
0.300537109375,
0.366943359375,
0.300537109375
],
[
0.57861328125,
0.2890625,
0.59716796875,
0.2890625,
0.59716796875,
0.3017578125,
0.57861328125,
0.3017578125
],
[
0.76953125,
0.291748046875,
0.78515625,
0.291748046875,
0.78515625,
0.2998046875,
0.76953125,
0.2998046875
],
[
0.43701171875,
0.292236328125,
0.451904296875,
0.292236328125,
0.451904296875,
0.300537109375,
0.43701171875,
0.300537109375
],
[
0.2467041015625,
0.292236328125,
0.261474609375,
0.292236328125,
0.261474609375,
0.300537109375,
0.2467041015625,
0.300537109375
],
[
0.64892578125,
0.292236328125,
0.66650390625,
0.292236328125,
0.66650390625,
0.300537109375,
0.64892578125,
0.300537109375
],
[
0.8388671875,
0.292236328125,
0.85888671875,
0.292236328125,
0.85888671875,
0.300537109375,
0.8388671875,
0.300537109375
],
[
0.318603515625,
0.311279296875,
0.426513671875,
0.311279296875,
0.426513671875,
0.324462890625,
0.318603515625,
0.324462890625
],
[
0.7236328125,
0.311279296875,
0.8310546875,
0.311279296875,
0.8310546875,
0.324462890625,
0.7236328125,
0.324462890625
],
[
0.5439453125,
0.3125,
0.6298828125,
0.3125,
0.6298828125,
0.323974609375,
0.5439453125,
0.323974609375
],
[
0.1396484375,
0.3125,
0.226318359375,
0.3125,
0.226318359375,
0.323974609375,
0.1396484375,
0.323974609375
],
[
0.07843017578125,
0.3359375,
0.890625,
0.337158203125,
0.890625,
0.3515625,
0.07843017578125,
0.35107421875
],
[
0.07843017578125,
0.3515625,
0.89208984375,
0.3515625,
0.89208984375,
0.365478515625,
0.07843017578125,
0.365478515625
],
[
0.08087158203125,
0.367431640625,
0.890625,
0.367431640625,
0.890625,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.080078125,
0.37890625,
0.890625,
0.37890625,
0.890625,
0.392578125,
0.080078125,
0.392578125
],
[
0.080078125,
0.393310546875,
0.890625,
0.393310546875,
0.890625,
0.4072265625,
0.080078125,
0.4072265625
],
[
0.08087158203125,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.07843017578125,
0.421630859375,
0.5458984375,
0.421630859375,
0.5458984375,
0.43505859375,
0.07843017578125,
0.43505859375
],
[
0.5107421875,
0.45703125,
0.5556640625,
0.4609375,
0.55224609375,
0.47607421875,
0.50732421875,
0.4716796875
],
[
0.7490234375,
0.458251953125,
0.89208984375,
0.458984375,
0.89208984375,
0.473388671875,
0.7490234375,
0.472900390625
],
[
0.63623046875,
0.460205078125,
0.67236328125,
0.460205078125,
0.67236328125,
0.472900390625,
0.63623046875,
0.472900390625
],
[
0.080078125,
0.458984375,
0.46728515625,
0.460205078125,
0.46728515625,
0.474853515625,
0.080078125,
0.473388671875
],
[
0.58251953125,
0.474853515625,
0.7236328125,
0.474853515625,
0.7236328125,
0.4892578125,
0.58251953125,
0.4892578125
],
[
0.79345703125,
0.473388671875,
0.85107421875,
0.473388671875,
0.85107421875,
0.4892578125,
0.79345703125,
0.4892578125
],
[
0.5107421875,
0.4873046875,
0.5537109375,
0.4892578125,
0.55224609375,
0.5048828125,
0.5087890625,
0.50244140625
],
[
0.080078125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.080078125,
0.4892578125
],
[
0.59619140625,
0.4892578125,
0.71142578125,
0.4892578125,
0.71142578125,
0.50390625,
0.59619140625,
0.50390625
],
[
0.79345703125,
0.4892578125,
0.85107421875,
0.4892578125,
0.85107421875,
0.50390625,
0.79345703125,
0.50390625
],
[
0.78125,
0.50390625,
0.861328125,
0.501953125,
0.861328125,
0.517578125,
0.78125,
0.51953125
],
[
0.080078125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.080078125,
0.5048828125
],
[
0.09722900390625,
0.505859375,
0.4697265625,
0.5068359375,
0.4697265625,
0.521484375,
0.09722900390625,
0.52099609375
],
[
0.60302734375,
0.5048828125,
0.705078125,
0.5048828125,
0.705078125,
0.51953125,
0.60302734375,
0.51953125
],
[
0.58642578125,
0.52099609375,
0.72119140625,
0.52099609375,
0.72119140625,
0.5341796875,
0.58642578125,
0.5341796875
],
[
0.7958984375,
0.51953125,
0.84716796875,
0.51953125,
0.84716796875,
0.5341796875,
0.7958984375,
0.5341796875
],
[
0.78173828125,
0.5341796875,
0.86376953125,
0.5341796875,
0.86376953125,
0.548828125,
0.78173828125,
0.548828125
],
[
0.080078125,
0.5234375,
0.4697265625,
0.5234375,
0.4697265625,
0.537109375,
0.080078125,
0.537109375
],
[
0.0767822265625,
0.537109375,
0.468017578125,
0.53857421875,
0.468017578125,
0.55322265625,
0.0767822265625,
0.5517578125
],
[
0.5703125,
0.53515625,
0.73388671875,
0.53515625,
0.73388671875,
0.548828125,
0.5703125,
0.548828125
],
[
0.51708984375,
0.5498046875,
0.5498046875,
0.5498046875,
0.5498046875,
0.56494140625,
0.51708984375,
0.56494140625
],
[
0.58837890625,
0.55029296875,
0.72119140625,
0.55029296875,
0.72119140625,
0.564453125,
0.58837890625,
0.564453125
],
[
0.7958984375,
0.5498046875,
0.8486328125,
0.5498046875,
0.8486328125,
0.564453125,
0.7958984375,
0.564453125
],
[
0.61279296875,
0.564453125,
0.69384765625,
0.56640625,
0.69384765625,
0.58056640625,
0.61279296875,
0.5791015625
],
[
0.080078125,
0.55322265625,
0.468017578125,
0.55322265625,
0.468017578125,
0.56640625,
0.080078125,
0.56640625
],
[
0.07843017578125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.07843017578125,
0.58203125
],
[
0.7939453125,
0.564453125,
0.84716796875,
0.5625,
0.8486328125,
0.57763671875,
0.7958984375,
0.580078125
],
[
0.583984375,
0.5791015625,
0.72216796875,
0.580078125,
0.72216796875,
0.5947265625,
0.583984375,
0.59326171875
],
[
0.7939453125,
0.580078125,
0.8486328125,
0.580078125,
0.8486328125,
0.5947265625,
0.7939453125,
0.5947265625
],
[
0.08087158203125,
0.599609375,
0.465576171875,
0.599609375,
0.465576171875,
0.60986328125,
0.08087158203125,
0.60986328125
],
[
0.080078125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.080078125,
0.5966796875
],
[
0.5595703125,
0.60498046875,
0.8310546875,
0.60595703125,
0.8310546875,
0.62060546875,
0.5595703125,
0.619140625
],
[
0.080078125,
0.61279296875,
0.4697265625,
0.61279296875,
0.4697265625,
0.626953125,
0.080078125,
0.626953125
],
[
0.080078125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.080078125,
0.642578125
],
[
0.080078125,
0.64306640625,
0.4697265625,
0.64306640625,
0.4697265625,
0.6572265625,
0.080078125,
0.6572265625
],
[
0.5009765625,
0.6416015625,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.080078125,
0.671875
],
[
0.50341796875,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.50341796875,
0.671875
],
[
0.08087158203125,
0.67529296875,
0.46728515625,
0.67529296875,
0.46728515625,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.5009765625,
0.671875,
0.890625,
0.67236328125,
0.890625,
0.6875,
0.5009765625,
0.68603515625
],
[
0.080078125,
0.689453125,
0.4697265625,
0.689453125,
0.4697265625,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.5009765625,
0.6875,
0.89208984375,
0.68798828125,
0.89208984375,
0.70263671875,
0.5009765625,
0.7021484375
],
[
0.07843017578125,
0.70263671875,
0.468017578125,
0.70068359375,
0.468017578125,
0.71728515625,
0.07843017578125,
0.7197265625
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.71826171875,
0.4697265625,
0.71826171875,
0.4697265625,
0.73193359375,
0.080078125,
0.73193359375
],
[
0.50244140625,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.07843017578125,
0.73291015625,
0.398681640625,
0.73291015625,
0.398681640625,
0.7470703125,
0.07843017578125,
0.7470703125
],
[
0.50244140625,
0.734375,
0.89208984375,
0.734375,
0.89208984375,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.1004638671875,
0.75,
0.46728515625,
0.75,
0.46728515625,
0.76318359375,
0.1004638671875,
0.76318359375
],
[
0.50244140625,
0.74755859375,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76123046875
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.5009765625,
0.76318359375,
0.68896484375,
0.76318359375,
0.68896484375,
0.7763671875,
0.5009765625,
0.7763671875
],
[
0.08087158203125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.07843017578125,
0.79443359375,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.07843017578125,
0.8095703125
],
[
0.50244140625,
0.79736328125,
0.63134765625,
0.798828125,
0.63134765625,
0.81298828125,
0.50244140625,
0.8125
],
[
0.07843017578125,
0.81005859375,
0.468017578125,
0.8095703125,
0.468017578125,
0.82373046875,
0.07843017578125,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.521484375,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.080078125,
0.84033203125,
0.46728515625,
0.84033203125,
0.46728515625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
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,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.79345703125,
0.88720703125,
0.79345703125,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Caseof PositiveExamples",
"Case of Negative Examples",
"O",
"?",
"?",
"?",
"?",
"O",
"O",
"O",
"W",
"W",
"-",
"W",
"3",
"3",
"3",
"3",
"",
"-",
"D",
"W",
"W",
"O",
"I",
"W",
"?",
"O",
"3",
"3",
"4",
"N",
"E",
"",
"E",
"E",
"®",
"®",
"E",
"®",
"E",
"E",
"3",
"®",
"®",
"®",
"O",
"E",
"E",
"8",
"",
"?",
"D",
"?",
"?",
"O",
"O",
"O",
"O",
"",
"W",
"W",
"W",
"1",
"",
"",
"3",
"I",
"2",
"W",
"C",
"O",
"J",
"",
"W",
"W",
"W",
"?",
"O",
"3",
"3",
"e",
"-",
"E",
"1",
"E",
"-",
"",
"E",
"E",
"E",
"E",
"E",
"E",
"®",
"O",
"®",
"O",
"8",
"O",
"8",
"",
"E",
"",
"",
"(b) Proxy-Anchor",
"(d) Proxy-Anchor",
"(c) Proxy-NCA",
"(a) Proxy-NCA",
"Figure 3. Differences between Proxy-NCA and Proxy-Anchor in handling proxies and embedding vectors during training. Each proxy",
"is colored in black and three different colors indicate distinct classes. The associations defined by the losses are expressed by edges, and",
"thickeredges getlargergradients.(a)Gradients ofProxy-NCA loss withrespectto positive examples havethe same scaleregardless of",
"their hardness. (b) Proxy-Anchor loss dynamically determines gradient scales regarding relative hardness of all positive examples so as",
"to pull harder positives more strongly. (c) In Proxy-NCA, each negative example is pushed only by a small number of proxies without",
"considering the distributionof embeddingvectors infine details.(d)Proxy-Anchorloss considers the distributionof embedding vectors in",
"more details as it has all negative examples affect each other in their gradients.",
"Type",
"Training Complexity",
"Loss",
"class, respectively. U is 1 thus ignored in most of proxy-",
"Proxy-Anchor (Ours)",
"O(MC)",
"Proxy",
"based losses including ours, but is nontrivial for those man-",
"Proxy-NCA [21]",
"O(MC)",
"O(MCU2)",
"aging multiple proxies per class such as SoftTriple loss [23].",
"Table1 compares the training complexity of our loss",
"SoftTriple [23]",
"Contrastive [2, 4, 9]",
"O(M?)",
"O(M3 /B2)",
"with those of popular pair- and proxy-based losses. The",
" complexity of our loss is O(MC) since it compares every",
"Triplet (Semi-Hard) [25]",
"Pair",
"Triplet (Smart) [10]",
"O(M?)",
"N-pair [27]",
"proxy with all positive or all negative examples in a batch.",
"More specifically, in Eq. (4), the complexity of the first sum-",
"0(M3)",
"Lifted Structure [29]",
"0(M3)",
"O(MC).hencethetotaltrainingcomplexityisO(MC)",
"mation is O(MC) and that of the second summation is also",
"Table 1. Comparison of training complexities.",
"The complexity of Proxy-NCA [21] is also O(MC) since",
"each data point is associated with one positive proxy and",
"C-1 negative proxies as can be seen in Eq. (2). On the",
"hard triplets using an approximated nearest neighbor index.",
"other hand, SoftTriple loss [23], a modification of SoftMax",
"However, even with these techniques, the training complex-",
"using multipleproxies per class,associates each datapoint",
"ity of Triplet loss is still high.Like Triplet loss, N-pair",
"with U positive proxies and U (C-1) negative proxies. The",
"loss [27] and Lifted Structure loss [29] that compare each",
"total training complexity of this loss is thus O(MCU2). In",
"positive pair of data to multiple negative pairs also have",
"conclusion, the complexity of our loss is the same with or",
"complexity O(M3).The training complexity of these losses",
"even lower than that of other proxy-based losses.",
"becomes prohibitively high as the number of training data",
"The training complexity of pair-based losses is higher",
"M increases,which slowsdownthespeed ofconvergence",
"than that of proxy-based ones. Since Contrastive loss [2,",
"as demonstrated in Figure 1.",
"4, 9] takes a pair of data as input, its training complexity",
"is O(M?). On the other hand, Triplet loss that examines",
"4.Experiments",
"triplets of data has complexity O(M3), which can be re-",
"duced by triplet mining strategies. For example, semi-hard",
"In this section, our method is evaluated and compared to",
"mining [25] reduces the complexity to O(M3 /B2) by se-",
"current state-of-the-art on the four benchmark datasets for",
"lecting negative pairs that are located within a neighbor-",
"deepmetriclearning[17,19,29,36.Wealsoinvestigate",
"hood of anchor but sufficiently far from it. Similarly, Smart",
"the effect of hyperparameters and embedding dimensional-",
"mining [10] lowers the complexity to O(M2) by sampling",
"ity of our loss to demonstrate its robustness."
] |
{
"arxiv_id": "2003.13911",
"language": "en",
"page_idx": 4
}
|
||
2009.05769-en-0
|
2009.05769-en-0.jpg
|
$\textbf{Removing the Background by Adding the Background: Towards Background}$
$\textbf{Robust Self- supervised Video Representation Learning}$
Jinpeng Wang$^{1,2}*$ Yuting Gao$^{2*}$ Ke Li$^{2}$ Yiqi Lin$^{1}$ Andy J.Ma$^{1\text{}}$ Hao Cheng $^2$ Pai Peng $^2$ Feiyue Huang$^2$ Rongrong Ji$^3$ Xing Sun$^2$
$^{1}$Sun Yat-sen University $^{2}$Tencent Youtu Lab$^{3}$Xia Men University
### $\mathbf{Abstract}$
<FigureHere>
Figure 1: Illustration of the background cheating. In the real open world, $an$ action can happen at various locations. Current models trained on the mainstream datasets tend to give predictions simply because it sees some background cues, neglecting the fact that motion pattern is what actually defines an“action”.
Self-supervised learning has shown great potentials in improving the video representation ability of deep neural
$\bigcap_{networksbygettingsupervisionfrom~the~data~itself.~How-}$
ever, some of the current methods tend to cheat from the background, i.e., the prediction is highly dependent on the video background instead of the motion, making the model vulnerable to background changes. To mitigate the model reliance towards the background, we propose to remove the background impact by adding the background. That is, given a video, we randomly select a static frame and add it to every other frames to construct a distracting video sample. Then we force the model to pull the feature of the distracting video and the feature of the original video closer, so that the model is explicitly restricted to resist the background influence, focusing more on the motion changes. We term our method as Background Erasing $( BE) . \textit{It is worth}$ noting that the implementation of our method is so simple and neat and can be added to most of the SOTA methods without much efforts. Specifically, BE brings 16.4% and 19. 1% improvements with MoCo on the severely biased datasets UCF101 and HMDB51, and 14.5% improvement on the less biased dataset Diving48.
more researchers' attention from the community to utilize off-the-shelf unlabeled data to improve the performance of CNNs by self-supervised learning.
## $1. \textbf{Introduction}$
Recently, self-supervised learning methods have been developed from the image field to the video field. However, there are big differences between the mainstream video dataset and the mainstream image dataset. Li et al.[29] and Girdhar et al.[14] point out that the current commonly used video datasets usually exist large implicit biases over scene and object structure , making the temporal structure become less important and the prediction tends to have a high dependence on the video background. We name this phenomenon as background cheating, as is shown in Figure 1. For example, a trained model may classify an action as playing soccer simply because it sees the field, without really understanding the cartwheel motion. As a result, the model is easily to overfit the training set, and the learned feature representation is likely to be scene-biased. Li et al.[29] reduce the bias by resampling the training set, and Wang et al.[53] propose to pull actions out of the context by training a binary classifer to explicitly distinguish action samples and conjugate samples that are contextually similar to action samples but contains different action.
Convolutional neural networks (CNNs) have achieved competitive accuracy on a variety of video understanding tasks, including action recognition [20], temporal action detection [63] and spatio-temporal action localization [55]. Such success relies heavily on manually annotated datasets, which are time-consuming and expensive to obtain. Meanwhile, there are numerous unlabeled data that are instantly available on the Internet, drawing more and
sThe first two authors contributed equally. This work was done when
Jinpeng was in Tencent Youtu Lab.
$^{\dagger}$Corresponding Author
1
|
[
[
0.1004638671875,
0.1337890625,
0.87109375,
0.1337890625,
0.87109375,
0.1483154296875,
0.1004638671875,
0.1483154296875
],
[
0.2091064453125,
0.1546630859375,
0.76123046875,
0.1572265625,
0.76123046875,
0.1749267578125,
0.2091064453125,
0.17236328125
],
[
0.1968994140625,
0.2020263671875,
0.78125,
0.199462890625,
0.78125,
0.2203369140625,
0.1968994140625,
0.222900390625
],
[
0.1895751953125,
0.222900390625,
0.301513671875,
0.222900390625,
0.301513671875,
0.2374267578125,
0.1895751953125,
0.2374267578125
],
[
0.291748046875,
0.2215576171875,
0.779296875,
0.2215576171875,
0.779296875,
0.2392578125,
0.291748046875,
0.2392578125
],
[
0.2115478515625,
0.255126953125,
0.7587890625,
0.255615234375,
0.7587890625,
0.2734375,
0.2115478515625,
0.272705078125
],
[
0.021240234375,
0.266357421875,
0.06451416015625,
0.266357421875,
0.06451416015625,
0.399658203125,
0.021240234375,
0.399658203125
],
[
0.236083984375,
0.30810546875,
0.314453125,
0.30810546875,
0.314453125,
0.32568359375,
0.236083984375,
0.32568359375
],
[
0.80224609375,
0.311279296875,
0.859375,
0.311279296875,
0.859375,
0.322509765625,
0.80224609375,
0.322509765625
],
[
0.8056640625,
0.32763671875,
0.87158203125,
0.32763671875,
0.87158203125,
0.33837890625,
0.8056640625,
0.33837890625
],
[
0.098876953125,
0.34228515625,
0.468017578125,
0.34228515625,
0.468017578125,
0.35595703125,
0.098876953125,
0.35595703125
],
[
0.80615234375,
0.34228515625,
0.8857421875,
0.34228515625,
0.8857421875,
0.349853515625,
0.80615234375,
0.349853515625
],
[
0.08087158203125,
0.35791015625,
0.468017578125,
0.35791015625,
0.468017578125,
0.371826171875,
0.08087158203125,
0.371826171875
],
[
0.08087158203125,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.08087158203125,
0.386962890625
],
[
0.814453125,
0.375,
0.85693359375,
0.375,
0.85693359375,
0.385009765625,
0.814453125,
0.385009765625
],
[
0.08087158203125,
0.386962890625,
0.46728515625,
0.386962890625,
0.46728515625,
0.40087890625,
0.08087158203125,
0.40087890625
],
[
0.8056640625,
0.38818359375,
0.84326171875,
0.38818359375,
0.84326171875,
0.3984375,
0.8056640625,
0.3984375
],
[
0.0228729248046875,
0.40283203125,
0.056365966796875,
0.401611328125,
0.05963134765625,
0.48291015625,
0.02532958984375,
0.484130859375
],
[
0.08251953125,
0.40478515625,
0.465576171875,
0.40478515625,
0.465576171875,
0.415283203125,
0.08251953125,
0.415283203125
],
[
0.08087158203125,
0.41845703125,
0.468017578125,
0.41845703125,
0.468017578125,
0.431884765625,
0.08087158203125,
0.431884765625
],
[
0.50341796875,
0.4248046875,
0.89208984375,
0.4248046875,
0.89208984375,
0.439453125,
0.50341796875,
0.439453125
],
[
0.08087158203125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.08087158203125,
0.447509765625
],
[
0.50341796875,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50341796875,
0.455078125
],
[
0.08087158203125,
0.448974609375,
0.468017578125,
0.448974609375,
0.468017578125,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.080078125,
0.46337890625,
0.468017578125,
0.46337890625,
0.468017578125,
0.4765625,
0.080078125,
0.4765625
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.08087158203125,
0.479248046875,
0.4697265625,
0.479248046875,
0.4697265625,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.50244140625,
0.486083984375,
0.890625,
0.486083984375,
0.890625,
0.5,
0.50244140625,
0.5
],
[
0.0269622802734375,
0.49755859375,
0.0555419921875,
0.49755859375,
0.0555419921875,
0.5966796875,
0.0269622802734375,
0.5966796875
],
[
0.080078125,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50244140625,
0.50048828125,
0.6337890625,
0.50048828125,
0.6337890625,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.5087890625,
0.46728515625,
0.5087890625,
0.46728515625,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.080078125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.080078125,
0.537109375
],
[
0.080078125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.035125732421875,
0.54736328125,
0.0400390625,
0.54736328125,
0.0400390625,
0.5546875,
0.035125732421875,
0.5546875
],
[
0.07843017578125,
0.5537109375,
0.468017578125,
0.55322265625,
0.468017578125,
0.5673828125,
0.07843017578125,
0.568359375
],
[
0.50244140625,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.50244140625,
0.564453125
],
[
0.08087158203125,
0.57080078125,
0.468017578125,
0.57080078125,
0.468017578125,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50341796875,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.50341796875,
0.580078125
],
[
0.0196075439453125,
0.58349609375,
0.058013916015625,
0.58349609375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.5009765625,
0.580078125,
0.73193359375,
0.58203125,
0.73193359375,
0.59521484375,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.5986328125,
0.46728515625,
0.5986328125,
0.46728515625,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.521484375,
0.599609375,
0.89208984375,
0.599609375,
0.89208984375,
0.61279296875,
0.521484375,
0.61279296875
],
[
0.08087158203125,
0.6162109375,
0.46728515625,
0.6162109375,
0.46728515625,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.50244140625,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50244140625,
0.62890625
],
[
0.08087158203125,
0.630859375,
0.4697265625,
0.630859375,
0.4697265625,
0.6416015625,
0.08087158203125,
0.6416015625
],
[
0.50244140625,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.64453125,
0.50244140625,
0.64453125
],
[
0.080078125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.5048828125,
0.646484375,
0.8896484375,
0.646484375,
0.8896484375,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.080078125,
0.6591796875,
0.316162109375,
0.6591796875,
0.316162109375,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.50341796875,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50341796875,
0.673828125
],
[
0.50244140625,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50244140625,
0.689453125
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08251953125,
0.705078125,
0.20751953125,
0.705078125,
0.20751953125,
0.7197265625,
0.08251953125,
0.7197265625
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.098876953125,
0.73095703125,
0.4697265625,
0.73095703125,
0.4697265625,
0.744140625,
0.098876953125,
0.744140625
],
[
0.50341796875,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50341796875,
0.75
],
[
0.08087158203125,
0.7470703125,
0.4697265625,
0.7470703125,
0.4697265625,
0.76025390625,
0.08087158203125,
0.76025390625
],
[
0.5048828125,
0.75244140625,
0.8896484375,
0.75244140625,
0.8896484375,
0.76318359375,
0.5048828125,
0.76318359375
],
[
0.080078125,
0.76220703125,
0.46728515625,
0.76220703125,
0.46728515625,
0.77587890625,
0.080078125,
0.77587890625
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.7763671875,
0.468017578125,
0.7763671875,
0.468017578125,
0.79052734375,
0.080078125,
0.79052734375
],
[
0.50341796875,
0.78271484375,
0.890625,
0.78271484375,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.79248046875,
0.4697265625,
0.79248046875,
0.4697265625,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.5009765625,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.08087158203125,
0.8095703125,
0.46728515625,
0.8095703125,
0.46728515625,
0.8193359375,
0.08087158203125,
0.8193359375
],
[
0.50341796875,
0.8125,
0.890625,
0.8125,
0.890625,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.080078125,
0.82275390625,
0.4697265625,
0.82275390625,
0.4697265625,
0.8359375,
0.080078125,
0.8359375
],
[
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,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.50341796875,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.09722900390625,
0.86279296875,
0.4697265625,
0.86279296875,
0.4697265625,
0.876953125,
0.09722900390625,
0.876953125
],
[
0.080078125,
0.87548828125,
0.268798828125,
0.87548828125,
0.268798828125,
0.8896484375,
0.080078125,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.096435546875,
0.88623046875,
0.2279052734375,
0.88720703125,
0.2279052734375,
0.9013671875,
0.096435546875,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.6689453125,
0.88720703125,
0.6689453125,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.483642578125,
0.92822265625,
0.48779296875,
0.92822265625,
0.48779296875,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Removing the Background by Adding the Background: Towards Background",
"Robust Self-supervised Video Representation Learning",
"Jinpeng Wangl,2 *Yuting Gao2 *Ke Li2Yiqi LinlAndy J. Ma 1 t",
"Hao Cheng 2",
"Pai Peng2Feiyue Huang? Rongrong Ji3Xing Sun?",
"1Sun Yat-sen University 2Tencent Youtu Lab3Xia Men University",
"",
"Abstract",
"Prediction",
"Cartwheel:0.02",
"Self-supervised learning has shown great potentials in",
"Playing.soccer:0.9",
"improving the video representation ability of deep neural",
"networks by getting supervision from the data itself. How-",
"ance:0.15",
"ever, some of the current methods tend to cheat from the",
"Run:0.8",
"['S]",
"backgroundi.e.,the prediction ishighly dependent on the",
"video background instead of the motion, making the model",
"Figure : Illustration of the background cheating. In the",
"vulnerable to background changes. To mitigate the model",
"real open world, an action can happen at various locations.",
"reliance towards the background, we propose to remove",
"Current models trained on the mainstream datasets tend to",
"the background impact by adding the background. That is,",
"give predictions simply because it sees some background",
"given a video, we randomly select a static frame and add it",
"cues, neglecting the fact that motion pattern is what actually",
"770",
"to every other frames to construct a distracting video sam-",
"defines an \"action\".",
"ple. Then we force the model to pull the feature of the dis-",
"tracting video and the feature of the original video closer",
"so that the model is explicitly restricted to resist the back-",
"",
"ground influence, focusing more on the motion changes. We",
"more researchers' attention from the community to utilize",
"termourmethodasBackgroundErasing(BE).Itisworth",
"off-the-shelf unlabeled data to improve the performance of",
"6002:AX",
"noting that the implementation of our method is so sim-",
"CNNs by self-supervised learning.",
"ple and neat and can be added to most of the SOTA meth-",
"Recently, self-supervised learning methods have been",
"ods without muchefforts.Specifically,BEbrings16.4%",
"developed from the image field to the video field. However,",
"and19.1%improvementswithMoCoontheseverelybiased",
"there are big differences between the mainstream video",
"datasets UCF101 and HMDB51, and 14.5% improvement",
"dataset and the mainstream image dataset.Li et al.[29and",
"on the less biased datasetDiving48.",
"Girdhar et al.[14] point out that the current commonly used",
"video datasets usually exist large implicit biases over scene",
"and object structure , making the temporal structure become",
"1. Introduction",
"less important and the prediction tends to have a high depen-",
"dence on the video background. We name this phenomenon",
"Convolutional neural networks (CNNs) have achieved",
"as background cheating, as is shown in Figure 1. For exam-",
"competitive accuracy on a variety of video understand-",
"ple,atrained modelmayclassifyanactionasplaying soccer",
"ing tasks, including action recognition [20], temporal ac-",
"simply because it sees the field, without really understand-",
"tion detection [63] and spatio-temporal action localization",
"ing the cartwheel motion.As a result, the modelis easilyto",
"[55]. Such success relies heavily on manually annotated",
"overfit the training set, and the learned feature representa-",
"datasets,which are time-consuming and expensive to ob-",
"tionislikelytobescene-biased.Lietal.[29lreducethebias",
"tain. Meanwhile, there are numerous unlabeled data that",
"by resampling the training set, and Wang et al.[53] propose",
"are instantly available on the Internet, drawing more and",
"to pull actions out of the context by training a binary clas-",
"sifier to explicitly distinguish action samples and conjugate",
"*The first two authors contributed equally. This work was done when",
"Jinpeng was in Tencent Youtu Lab.",
"samples that are contextually similar to action samples but",
"tCorresponding Author",
"contains different action.",
"-"
] |
{
"arxiv_id": "2009.05769",
"language": "en",
"page_idx": 0
}
|
||
2009.05769-en-1
|
2009.05769-en-1.jpg
|
mentum encoder and achieves great success. In contrast, Ye et al. [59] propose to use a mini batch to replace the memory bank. SimCLR[8] shows that the memory bank can be entirely replaced by a large batch size.
In this work, to hinder the model from background cheating and make the model generalize better, we present to reduce the impact of the background by adding the background and encourage the model to learn consistent feature w/ or w/o the operation. Specifically, given a video, we randomly select a static frame and add it to every other frames to construct a distracting video, as is shown in Figure 3. Then we force the model to pull the feature of the distracting video and the feature of the original video together by consistency regularization. In this way, we made a disturbance to the video background and require its feature to be consistent with the original video, achieving the purpose of making the model not be excessively dependent on the background, thereby alleviating the background cheating problem.
Experimental results demonstrate that the proposed method can effectively reduce the influence of the $back-$ ground cheating, and the extracted representation is more robust to the background bias and have stronger generalization ability. Our approach is simple and incorporate it into existing self-supervised video learning methods can bring significant gains. In summary, our main contributions are twofold:
· We propose a simple yet effective video representation
learning method that is robust to the background.
· The proposed approach can be easily incorporated with
existing self-supervised video representation learning methods, bringing further gains on UCF101[41], HMDB51[27] and Diving48[30] datasets.
## $2. \textbf{ Related Work}$
2.1. Self-supervised Learning for Image
Self-supervised learning is a generic learning framework which gets supervision from the data itself. Current methods can be grouped into two types of paradigms, $i.e.$, constructing pretext tasks or constructing contrastive learning. Pretext tasks. These methods focus on solving surrogate classification tasks with surrogate labels, including predicting the rotation angle of image[13], solving the jigsaw puzzle[35], coloring image[62] and predicting relative patches[35], etc. Recently, the type of image transformation also be used as a surrogate[61]. Contrastive learning. Another mainstream method is based on contrastive learning, which regards each instance as a category. Early work [11] directly used each sample in the dataset as a category to learn a linear classifier, but this method will become infeasible when the number of samples increases. To alleviate this problem, Wu et al. [56] replace the classifier with a memory bank storing previously computed representations and then use a noise contrastive estimation [15] to compare instances. MoCo[21] stores the representations from a mo-
2.2. Self-supervised Video Representation Learning
Recent years, self-supervised learning has been ex-
panded into the video domain and attracts a lot interests.
Pretext tasks. The majority of the prior work explore natural video properties as supervision signal. Among them, temporal order is one of the most widely-used property, such as, the arrow of time [54], the order of shuffled frames [34], the order of video clip [57] and the playback rate of the video[1,58]. Besides the temporal order, the spatiotemporal statistics are also used as supervision. For example, pxiel-wise geometry information [12], space-time cubic puzzles [26, 32] and the optical-flow and the appearance statistics[49]. In addition, DynamoNet[10] predicts future frames by learning dynamic motion filter, which is pre-trained on a large-scale dataset Youtube-8M. More recently, Buchler et al. [6] and ELO [38] propose to ensemble multiple pretext task based on reinforcement learning.
Contrastive learning. Contrastive learning is introduced into the field of video representation learning by TCN [40], which uses different camera views as positive samples. IIC [43] proposes an inter-intra mulit-modal contrastive framework based on the Contrastive Multiview Coding [44]. CoCLR[19] takes the advantage of the natural correlation between the RGB and the Optical Flow modalities to select the negative samples in the memory bank. GDT[33] achieves great success by using tens of millions data for pre-training with multi-modal constrastive leanring.
It is worth to mention that while all methods mentioned above focus on designing specific tasks, we present a generalized constraint term that can be integrated into any existing self-supervised video representation learning approach. $2. 3. \textbf{Background Biases in Video}$
Current widely used video datasets have serious bias towards the background[29, 14], which may misleads the model using just the static cues to achieve good results. For example, only using three frames during training, TSN[52] can achieve 85% accuracy on UCF101. Therefore, using these datasets for training can easily cause the model making background biased predictions
In order to mitigate the background bias, Li et al.[30] resample the original dataset to generate a less biased dataset Diving48 for the action recognition task. Wang et al.[53] use conjugate samples that are contextually similar to human action samples but do not contain the action to train a classifier to deliberately separate the action from the context. Choi et al.[9] propose to detect and mask actors with a human detector and further present a novel adversarial loss
2
|
[
[
0.098876953125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.098876953125,
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,
0.10797119140625,
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,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50244140625,
0.1363525390625
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.75634765625,
0.1383056640625,
0.75634765625,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.5009765625,
0.160400390625,
0.8935546875,
0.1610107421875,
0.8935546875,
0.1787109375,
0.5009765625,
0.177978515625
],
[
0.08251953125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.179931640625,
0.08251953125,
0.179931640625
],
[
0.080078125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.080078125,
0.1962890625
],
[
0.521484375,
0.186279296875,
0.8896484375,
0.186279296875,
0.8896484375,
0.2008056640625,
0.521484375,
0.2008056640625
],
[
0.08087158203125,
0.199462890625,
0.468017578125,
0.199462890625,
0.468017578125,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.5048828125,
0.203857421875,
0.869140625,
0.203857421875,
0.869140625,
0.2139892578125,
0.5048828125,
0.2139892578125
],
[
0.080078125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.2171630859375,
0.890625,
0.2171630859375,
0.890625,
0.2310791015625,
0.50244140625,
0.2310791015625
],
[
0.07843017578125,
0.2291259765625,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.07843017578125,
0.242431640625
],
[
0.50341796875,
0.23291015625,
0.890625,
0.23291015625,
0.890625,
0.246826171875,
0.50341796875,
0.246826171875
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.5009765625,
0.246826171875,
0.890625,
0.2474365234375,
0.890625,
0.261962890625,
0.5009765625,
0.261474609375
],
[
0.080078125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.080078125,
0.271484375
],
[
0.50341796875,
0.264404296875,
0.8896484375,
0.264404296875,
0.8896484375,
0.274658203125,
0.50341796875,
0.274658203125
],
[
0.080078125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.080078125,
0.287353515625
],
[
0.50341796875,
0.27783203125,
0.89208984375,
0.27783203125,
0.89208984375,
0.291748046875,
0.50341796875,
0.291748046875
],
[
0.07843017578125,
0.287353515625,
0.468017578125,
0.28857421875,
0.468017578125,
0.302978515625,
0.07843017578125,
0.3017578125
],
[
0.50244140625,
0.292236328125,
0.890625,
0.292236328125,
0.890625,
0.306884765625,
0.50244140625,
0.306884765625
],
[
0.07843017578125,
0.3037109375,
0.1683349609375,
0.3017578125,
0.1683349609375,
0.31640625,
0.07843017578125,
0.318115234375
],
[
0.5048828125,
0.309326171875,
0.890625,
0.309326171875,
0.890625,
0.3193359375,
0.5048828125,
0.3193359375
],
[
0.098876953125,
0.318115234375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.098876953125,
0.33203125
],
[
0.50341796875,
0.323974609375,
0.89208984375,
0.323974609375,
0.89208984375,
0.337158203125,
0.50341796875,
0.337158203125
],
[
0.080078125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.33837890625,
0.89208984375,
0.33837890625,
0.89208984375,
0.3515625,
0.50244140625,
0.3515625
],
[
0.080078125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.080078125,
0.363037109375
],
[
0.50341796875,
0.352783203125,
0.89208984375,
0.352783203125,
0.89208984375,
0.3662109375,
0.50341796875,
0.3662109375
],
[
0.07843017578125,
0.363037109375,
0.468017578125,
0.3642578125,
0.468017578125,
0.37890625,
0.07843017578125,
0.37744140625
],
[
0.50341796875,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.080078125,
0.380126953125,
0.4697265625,
0.380126953125,
0.4697265625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.50341796875,
0.38330078125,
0.8896484375,
0.38330078125,
0.8896484375,
0.396484375,
0.50341796875,
0.396484375
],
[
0.080078125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.080078125,
0.40771484375
],
[
0.50244140625,
0.3984375,
0.89208984375,
0.3984375,
0.89208984375,
0.412353515625,
0.50244140625,
0.412353515625
],
[
0.080078125,
0.410400390625,
0.4697265625,
0.410400390625,
0.4697265625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5009765625,
0.4140625,
0.85546875,
0.4140625,
0.85546875,
0.427978515625,
0.5009765625,
0.427978515625
],
[
0.080078125,
0.425537109375,
0.1380615234375,
0.425537109375,
0.1380615234375,
0.4375,
0.080078125,
0.4375
],
[
0.50341796875,
0.4287109375,
0.89208984375,
0.4287109375,
0.89208984375,
0.442626953125,
0.50341796875,
0.442626953125
],
[
0.50244140625,
0.443115234375,
0.8896484375,
0.443115234375,
0.8896484375,
0.45703125,
0.50244140625,
0.45703125
],
[
0.1004638671875,
0.449462890625,
0.468017578125,
0.449462890625,
0.468017578125,
0.464111328125,
0.1004638671875,
0.464111328125
],
[
0.111083984375,
0.464111328125,
0.4345703125,
0.46533203125,
0.4345703125,
0.479736328125,
0.111083984375,
0.479248046875
],
[
0.50341796875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.472900390625,
0.50341796875,
0.472900390625
],
[
0.50341796875,
0.474853515625,
0.89208984375,
0.474853515625,
0.89208984375,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.1004638671875,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.1004638671875,
0.5048828125
],
[
0.50244140625,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.111083984375,
0.5068359375,
0.46728515625,
0.5068359375,
0.46728515625,
0.52099609375,
0.111083984375,
0.52099609375
],
[
0.50341796875,
0.5048828125,
0.89208984375,
0.5048828125,
0.89208984375,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.111083984375,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.111083984375,
0.53515625
],
[
0.50244140625,
0.51953125,
0.89208984375,
0.51953125,
0.89208984375,
0.53271484375,
0.50244140625,
0.53271484375
],
[
0.111083984375,
0.53515625,
0.390625,
0.537109375,
0.390625,
0.5517578125,
0.111083984375,
0.5498046875
],
[
0.50341796875,
0.53515625,
0.89208984375,
0.53515625,
0.89208984375,
0.548828125,
0.50341796875,
0.548828125
],
[
0.50341796875,
0.55029296875,
0.89208984375,
0.55029296875,
0.89208984375,
0.564453125,
0.50341796875,
0.564453125
],
[
0.080078125,
0.56640625,
0.2181396484375,
0.56640625,
0.2181396484375,
0.58056640625,
0.080078125,
0.58056640625
],
[
0.50341796875,
0.56494140625,
0.7587890625,
0.56494140625,
0.7587890625,
0.5791015625,
0.50341796875,
0.5791015625
],
[
0.521484375,
0.580078125,
0.8935546875,
0.580078125,
0.8935546875,
0.59326171875,
0.521484375,
0.59326171875
],
[
0.080078125,
0.5888671875,
0.381591796875,
0.5888671875,
0.381591796875,
0.60400390625,
0.080078125,
0.60400390625
],
[
0.50244140625,
0.5947265625,
0.89208984375,
0.59521484375,
0.89208984375,
0.60986328125,
0.50244140625,
0.609375
],
[
0.098876953125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.098876953125,
0.62744140625
],
[
0.5048828125,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.5048828125,
0.62255859375
],
[
0.08087158203125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.08087158203125,
0.64306640625
],
[
0.50244140625,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.07843017578125,
0.64306640625,
0.468017578125,
0.64453125,
0.468017578125,
0.6591796875,
0.07843017578125,
0.65771484375
],
[
0.50341796875,
0.65087890625,
0.75244140625,
0.65087890625,
0.75244140625,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.080078125,
0.66015625,
0.463134765625,
0.66015625,
0.463134765625,
0.673828125,
0.080078125,
0.673828125
],
[
0.080078125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.689453125,
0.080078125,
0.689453125
],
[
0.521484375,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.521484375,
0.68798828125
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.5009765625,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.5009765625,
0.71826171875
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.734375,
0.080078125,
0.734375
],
[
0.50341796875,
0.7216796875,
0.8896484375,
0.7216796875,
0.8896484375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
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,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.08087158203125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.5009765625,
0.7646484375,
0.73193359375,
0.7646484375,
0.73193359375,
0.77783203125,
0.5009765625,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.521484375,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.08087158203125,
0.796875,
0.47216796875,
0.796875,
0.47216796875,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.8125,
0.468017578125,
0.8125,
0.468017578125,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84033203125,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5009765625,
0.85400390625
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88818359375,
0.89208984375,
0.88818359375,
0.89208984375,
0.8984375,
0.50341796875,
0.8984375
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"In this work, to hinder the model from background cheat-",
"mentum encoder and achieves great success. In contrast, Ye",
"ing and make the model generalize better, we present to",
"et al. [59] propose to use a mini batch to replace the mem-",
"reduce the impact of the background by adding the back-",
"ory bank. SimCLR [8] shows that the memory bank can be",
"ground and encourage the model to learn consistent fea-",
"entirely replaced by a large batch size.",
"ture wl or w/o the operation. Specifically, given a video,",
"2.2. Self-supervised Video Representation Learning",
"werandomlyselectastaticframeand add ittoeveryother",
"frames to construct a distracting video, as is shown in Fig-",
"Recent years, self-supervised learning has been ex-",
"ure3.Thenweforcethemodel topull the feature of the",
"pandedintothevideodomainandattractsalotinterests",
"distracting video and the feature of the original video to-",
"Pretext tasks. The majority of the prior work explore nat-",
"gether by consistency regularization. In this way,we made",
"ural video properties as supervision signal. Among them,",
"a disturbance to the video background and require its feature",
"temporal order is one of the most widely-used property,",
"to be consistent with the original video, achieving the pur-",
"suchas.thearrowoftime[54l,theorderofshuffled frames",
"pose of making the model not be excessively dependent on",
"[34], the order of video clip [57] and the playback rate of",
"the background, thereby alleviating the background cheat-",
"the video [1, 58]. Besides the temporal order, the spatio-",
"ingproblem.",
"temporal statistics arealsousedas supervision.Forexam-",
"Experimental results demonstrate that the proposed",
"ple, pxiel-wise geometry information [12], space-time cu-",
"method can effectively reduce the influence of the back-",
"bic puzzles [26, 32] and the optical-flow and the appear-",
"ground cheating, and the extracted representation is more",
"ance statistics [49]. In addition, DynamoNet[10] predicts",
"robust to the background bias and have stronger generaliza-",
"future frames by learning dynamic motion filter, which is",
"tion ability. Our approach is simple and incorporate it into",
"pre-trained on a large-scale dataset Youtube-8M. More re-",
"existing self-supervised video learning methods can bring",
"cently, Buchler et al. [6] and ELO [38] propose to ensemble",
"significant gains. In summary, our main contributions are",
"multiple pretext task based on reinforcement learning.",
"twofold:",
"Contrastive learning. Contrastive learning is introduced",
"into the field of video representation learning by TCN [40]",
".We propose a simple yet effective video representation",
"learning method that is robust to the background.",
"which uses different camera views as positive samples. IIC",
"[43] proposes an inter-intra mulit-modal contrastive frame-",
".The proposed approach can be easily incorporated with",
"work based on the Contrastive Multiview Coding [44]. Co-",
"existing self-supervised video representation learn-",
"CLR [19] takes the advantage of the natural correlation be-",
"ing methods, bringing further gains on UCF101[41],",
"tween the RGB and the Optical Flow modalities to select the",
"HMDB51 [27] and Diving48[30] datasets.",
"negative samples in the memory bank. GDT [33] achieves",
"great success by using tens of millions data for pre-training",
"2.Related Work",
"with multi-modal constrastive leanring",
"It is worth to mention that while all methods mentioned",
"2.1. Self-supervised Learning for Image",
"above focus on designing specific tasks, we present a gener-",
"Self-supervised learning is a generic learning framework",
"alizedconstrainttermthatcanbeintegratedintoanyexist-",
"which gets supervision from the data itself. Current meth-",
"ing self-supervised video representation learning approach.",
"ods can be grouped into two types of paradigms, i.e., con-",
"2.3. Background Biases in Video",
"structing pretext tasks or constructing contrastive learning.",
"Pretext tasks. These methods focus on solving surrogate",
"Current widely used video datasets have serious bias",
"classification tasks with surrogate labels, including pre-",
"towards the background[29, 14], which may misleads the",
"dicting the rotation angle of image[13], solving the jig-",
"model using just the static cues to achieve good results. For",
"saw puzzle[35], coloring image[62] and predicting relative",
"example,onlyusingthreeframes duringtraining,TSN[52",
"patches[35], etc. Recently, the type of image transforma-",
"can achieve 85% accuracy on UCF101. Therefore, using",
"tion also be used as a surrogate[6l]. Contrastive learning.",
"these datasets for training can easily cause the model mak-",
"Anothermainstreammethodisbased oncontrastivelearn-",
"ing background biased predictions.",
"ing, which regards each instance as a category. Early work",
"In order to mitigate the background bias, Li et al.[30] re-",
"[1l] directly used each sample in the dataset as a category to",
"sample the original dataset to generate a less biased dataset",
"learn alinearclassifier,butthis method will becomeinfea-",
"Diving48 for the action recognition task. Wang et al.[53]",
"sible when the number of samples increases. To alleviate",
"use conjugate samples that are contextually similar to hu-",
"this problem, Wu et al. [56] replace the classifier with a",
"man action samples but do not contain the action to train a",
"memory bank storing previously computed representations",
"classifierto deliberately separatetheaction from the con-",
"and then use a noise contrastive estimation [15] to compare",
"text. Choi et al.[9] propose to detect and mask actors with a",
"instances. MoCo [21] stores the representations from a mo-",
"humandetectorandfurtherpresentanovel adversarial loss",
"2"
] |
{
"arxiv_id": "2009.05769",
"language": "en",
"page_idx": 1
}
|
||
2009.05769-en-2
|
2009.05769-en-2.jpg
|
for debasing. In this work, we try to debias through consistency constraint, which is simple but effective and does not need additional costs.
# 3. Methodology
In this section we introduce the proposed Background Erasing (BE) method. We frst give an overall description of BE, and then introduce how to integrate BE into existing self-supervised methods.
## 3.1. Overall Architecture
<FigureHere>
Figure 2: The framework of the proposed method BE. A video is first randomly cropped spatially, then we generate the distracting video by adding a static frame upon other frames. The model is trained by a existing self-supervised task together with a consistency constraint, with the goal of pulling the feature of the original video and that of the distracting video closer. (Best viewed in color).
The framework of the proposed BE is shown in Figure 2. For each input video $x$, we first randomly crop two fixedlength clips from different spatial locations, denoted as $x^o$ and $x^v.$ Suppose we have a basic data augmentation set $\mathcal{A}$, from which we sample two specific operations $a^1$ and $a^2$, and operate on $x^o$ and $x^v$ respectively. In this way, the input clips have different distribution in the pixel level but are consistent in the semantic level. Afterwards, $x^o$ is directly fed into the 3D backbone to extract the feature representation and we denote this procedure as $F(x^o;\theta)$, where $\theta$ represents the backbone parameters. For $x^v$, we first generate a distracting counterpart $x^d$ for it, which has the interference of added static frame noise but the semantics remains the same. The output feature maps of $x^o$ and $x^d$ are rep-
resented by $f_x^o,f_{x^d}\in\mathbb{R}^{C\times T\times H\times W}.C$ is the number of channel and $T$ is the length of time dimension. $W$ and $H$ are spatial size. At last, the extracted features $f_{x^o},f_{x^d}$ are pulled closer within the existing self-supervised methods.
# 3.2. Background Erasing.
In the video representation learning, sometimes the statistical characteristics of the background will drown out the motion features of the moving subject. Thus it is easy for the model to make predictions based only on the background information. And the model is easy to overfit to the training set and has poor generalization on the new dataset.
Background Erasing(BE) is proposed to remove the negative impact of the background by adding the background. Specifically, for a video sequence $x$, we randomly select one static frame and add it as a spatial background noise to every other frames to generate a distracting video, in which each frame $\hat{x}$ is obtained by the following formula:
$$\hat{x}=(1-\lambda)\cdot x^{(j)}+\lambda\cdot x^{(k)},j\in[1,T]$$
(1)
where $\lambda$ is sampled from the uniform distribution $[0,\gamma],x^{(j)}$
means the $j$-th frame of $x,k$ denotes the index of the randomly selected frame and $T$ is the length of the video sequence $x.$ BE operation is applied to $x^v$, and the generated distracting video clip $x^d$ has a background perturbation on the spatial dimension, but the motion pattern is basically not changed, as shown in Figure 3.
Furthermore, it is easy to prove that the time derivative of $x^d$ is a linear transformation of the time derivative of $x^v$, formally:
$$\begin{aligned}\frac{\mathrm{d}((1-\lambda)x^v+\lambda\delta)}{\mathrm{d}t}=(1-\lambda)\frac{\mathrm{d}x^v}{\mathrm{d}t}\end{aligned}$$
(2)
where $\delta$ represents the result of repeating the selected frame $x^{(k)}T$ times along the time dimension. Previous works[3, 5,4,51] have shown that the time derivative of a video clip is an important information for action classification, thus, the property that BE maintains the linear transformation of such information is very crucial. Afterwards, we force the model to pull the feature of $x^o$ and the feature of $x^d$ closer, which will be introduced in details later. Since $x^o$ and $x^d$ resemble each other in the motion pattern but differentiate each other in spatial, when the features of $x^o$ and $x^d$ are brought closer, the model will be promoted to suppress the background noise, yielding video representations that are more sensitive to motion changes. We have tried a variety of ways to add background noise, results are shown in Table 4. Experimental results demonstrate that the intra-video static frame, $i.e.$, BE, works best overall. Meanwhile, we have also tried to add various data augmentations to the selected intra-video static frame to introduce more disturbance, but there is no positive gain.
3
|
[
[
0.495849609375,
0.083984375,
0.89453125,
0.083984375,
0.89453125,
0.10797119140625,
0.495849609375,
0.10797119140625
],
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.08087158203125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.50244140625,
0.106689453125,
0.890625,
0.106689453125,
0.890625,
0.12060546875,
0.50244140625,
0.12060546875
],
[
0.07843017578125,
0.12249755859375,
0.22216796875,
0.12249755859375,
0.22216796875,
0.1363525390625,
0.07843017578125,
0.1363525390625
],
[
0.50244140625,
0.12371826171875,
0.89208984375,
0.12371826171875,
0.89208984375,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.50244140625,
0.1383056640625,
0.8798828125,
0.1383056640625,
0.8798828125,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.1514892578125,
0.2099609375,
0.1514892578125,
0.2099609375,
0.1666259765625,
0.080078125,
0.1666259765625
],
[
0.5009765625,
0.1610107421875,
0.69921875,
0.1622314453125,
0.69921875,
0.179931640625,
0.5009765625,
0.1787109375
],
[
0.098876953125,
0.1767578125,
0.468017578125,
0.1767578125,
0.468017578125,
0.1900634765625,
0.098876953125,
0.1900634765625
],
[
0.521484375,
0.186279296875,
0.890625,
0.186279296875,
0.890625,
0.199462890625,
0.521484375,
0.199462890625
],
[
0.08087158203125,
0.1925048828125,
0.468017578125,
0.1925048828125,
0.468017578125,
0.205810546875,
0.08087158203125,
0.205810546875
],
[
0.080078125,
0.20703125,
0.468017578125,
0.20703125,
0.468017578125,
0.2203369140625,
0.080078125,
0.2203369140625
],
[
0.50244140625,
0.2020263671875,
0.89208984375,
0.2020263671875,
0.89208984375,
0.21533203125,
0.50244140625,
0.21533203125
],
[
0.080078125,
0.222900390625,
0.2459716796875,
0.222900390625,
0.2459716796875,
0.236083984375,
0.080078125,
0.236083984375
],
[
0.50244140625,
0.2171630859375,
0.890625,
0.2171630859375,
0.890625,
0.231689453125,
0.50244140625,
0.231689453125
],
[
0.50244140625,
0.23291015625,
0.890625,
0.23291015625,
0.890625,
0.246826171875,
0.50244140625,
0.246826171875
],
[
0.080078125,
0.246826171875,
0.27294921875,
0.246826171875,
0.27294921875,
0.260009765625,
0.080078125,
0.260009765625
],
[
0.50244140625,
0.248779296875,
0.89208984375,
0.2474365234375,
0.89208984375,
0.261474609375,
0.50244140625,
0.261962890625
],
[
0.50244140625,
0.26318359375,
0.8896484375,
0.26318359375,
0.8896484375,
0.276611328125,
0.50244140625,
0.276611328125
],
[
0.1953125,
0.27978515625,
0.326904296875,
0.27978515625,
0.326904296875,
0.29345703125,
0.1953125,
0.29345703125
],
[
0.52294921875,
0.279052734375,
0.890625,
0.279052734375,
0.890625,
0.292236328125,
0.52294921875,
0.292236328125
],
[
0.21728515625,
0.292236328125,
0.306396484375,
0.292236328125,
0.306396484375,
0.2998046875,
0.21728515625,
0.2998046875
],
[
0.50341796875,
0.29345703125,
0.8896484375,
0.29345703125,
0.8896484375,
0.306884765625,
0.50341796875,
0.306884765625
],
[
0.50244140625,
0.309326171875,
0.89208984375,
0.309326171875,
0.89208984375,
0.322509765625,
0.50244140625,
0.322509765625
],
[
0.2254638671875,
0.326904296875,
0.3056640625,
0.326904296875,
0.3056640625,
0.333984375,
0.2254638671875,
0.333984375
],
[
0.375732421875,
0.32763671875,
0.40283203125,
0.32763671875,
0.40283203125,
0.335205078125,
0.375732421875,
0.335205078125
],
[
0.50244140625,
0.323974609375,
0.89208984375,
0.323974609375,
0.89208984375,
0.337158203125,
0.50244140625,
0.337158203125
],
[
0.35546875,
0.33837890625,
0.381591796875,
0.33837890625,
0.381591796875,
0.37744140625,
0.35546875,
0.37744140625
],
[
0.5009765625,
0.33837890625,
0.890625,
0.33837890625,
0.890625,
0.3515625,
0.5009765625,
0.3515625
],
[
0.5009765625,
0.352783203125,
0.8388671875,
0.354248046875,
0.8388671875,
0.36865234375,
0.5009765625,
0.367431640625
],
[
0.1690673828125,
0.339111328125,
0.19287109375,
0.339111328125,
0.19287109375,
0.361083984375,
0.1690673828125,
0.361083984375
],
[
0.56640625,
0.380615234375,
0.82666015625,
0.380615234375,
0.82666015625,
0.395751953125,
0.56640625,
0.395751953125
],
[
0.1690673828125,
0.38818359375,
0.19287109375,
0.38818359375,
0.19287109375,
0.40087890625,
0.1690673828125,
0.40087890625
],
[
0.351318359375,
0.38818359375,
0.378173828125,
0.38818359375,
0.378173828125,
0.399658203125,
0.351318359375,
0.399658203125
],
[
0.869140625,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.395751953125,
0.869140625,
0.395751953125
],
[
0.15283203125,
0.397705078125,
0.19775390625,
0.397705078125,
0.19775390625,
0.40771484375,
0.15283203125,
0.40771484375
],
[
0.336669921875,
0.395751953125,
0.3896484375,
0.395751953125,
0.3896484375,
0.410400390625,
0.336669921875,
0.410400390625
],
[
0.50244140625,
0.4091796875,
0.89208984375,
0.4072265625,
0.89208984375,
0.4248046875,
0.50244140625,
0.4267578125
],
[
0.30224609375,
0.425537109375,
0.4345703125,
0.425537109375,
0.4345703125,
0.4443359375,
0.30224609375,
0.4443359375
],
[
0.50341796875,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50341796875,
0.43994140625
],
[
0.5048828125,
0.442626953125,
0.8896484375,
0.442626953125,
0.8896484375,
0.45263671875,
0.5048828125,
0.45263671875
],
[
0.160888671875,
0.4287109375,
0.19775390625,
0.4287109375,
0.19775390625,
0.439453125,
0.160888671875,
0.439453125
],
[
0.35546875,
0.448974609375,
0.37744140625,
0.448974609375,
0.37744140625,
0.47021484375,
0.35546875,
0.47021484375
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455078125,
0.89208984375,
0.4697265625,
0.50244140625,
0.47021484375
],
[
0.3349609375,
0.47021484375,
0.396240234375,
0.47021484375,
0.396240234375,
0.490478515625,
0.3349609375,
0.490478515625
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.50244140625,
0.48486328125,
0.8935546875,
0.486083984375,
0.8935546875,
0.50048828125,
0.50244140625,
0.5
],
[
0.143798828125,
0.474853515625,
0.1666259765625,
0.474853515625,
0.1666259765625,
0.48291015625,
0.143798828125,
0.48291015625
],
[
0.35546875,
0.4912109375,
0.371826171875,
0.4912109375,
0.371826171875,
0.5087890625,
0.35546875,
0.5087890625
],
[
0.238525390625,
0.50048828125,
0.27685546875,
0.50048828125,
0.27685546875,
0.50830078125,
0.238525390625,
0.50830078125
],
[
0.50244140625,
0.501953125,
0.70751953125,
0.501953125,
0.70751953125,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.2484130859375,
0.51123046875,
0.259033203125,
0.51123046875,
0.259033203125,
0.51513671875,
0.2484130859375,
0.51513671875
],
[
0.291748046875,
0.51513671875,
0.431396484375,
0.51513671875,
0.431396484375,
0.537109375,
0.291748046875,
0.537109375
],
[
0.52294921875,
0.517578125,
0.89208984375,
0.517578125,
0.89208984375,
0.53076171875,
0.52294921875,
0.53076171875
],
[
0.50341796875,
0.5322265625,
0.890625,
0.5322265625,
0.890625,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.50244140625,
0.54736328125,
0.56787109375,
0.54736328125,
0.56787109375,
0.5625,
0.50244140625,
0.5625
],
[
0.08087158203125,
0.564453125,
0.468017578125,
0.564453125,
0.468017578125,
0.5791015625,
0.08087158203125,
0.5791015625
],
[
0.58251953125,
0.57373046875,
0.7197265625,
0.57373046875,
0.7197265625,
0.58837890625,
0.58251953125,
0.58837890625
],
[
0.71142578125,
0.5791015625,
0.79150390625,
0.58349609375,
0.7900390625,
0.59765625,
0.7099609375,
0.59326171875
],
[
0.779296875,
0.57373046875,
0.810546875,
0.57373046875,
0.810546875,
0.5859375,
0.779296875,
0.5859375
],
[
0.08087158203125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.869140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.869140625,
0.5966796875
],
[
0.6337890625,
0.59130859375,
0.65673828125,
0.59130859375,
0.65673828125,
0.60498046875,
0.6337890625,
0.60498046875
],
[
0.7861328125,
0.59033203125,
0.8095703125,
0.59033203125,
0.8095703125,
0.60400390625,
0.7861328125,
0.60400390625
],
[
0.080078125,
0.611328125,
0.468017578125,
0.611328125,
0.468017578125,
0.625,
0.080078125,
0.625
],
[
0.080078125,
0.5947265625,
0.468017578125,
0.59521484375,
0.468017578125,
0.60986328125,
0.080078125,
0.609375
],
[
0.50341796875,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50341796875,
0.62890625
],
[
0.080078125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.5,
0.626953125,
0.89208984375,
0.62744140625,
0.89208984375,
0.64599609375,
0.5,
0.64453125
],
[
0.080078125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.50244140625,
0.64306640625,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.6572265625
],
[
0.080078125,
0.65576171875,
0.390625,
0.65576171875,
0.390625,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.5009765625,
0.6748046875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.098876953125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.705078125,
0.46728515625,
0.705078125,
0.46728515625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
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,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50244140625,
0.734375
],
[
0.080078125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.080078125,
0.75
],
[
0.50341796875,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.74755859375,
0.50341796875,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
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,
0.78076171875,
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
],
[
0.08087158203125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08087158203125,
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
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.82568359375,
0.468017578125,
0.82568359375,
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
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
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,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.6650390625,
0.88720703125,
0.6650390625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.923828125,
0.491943359375,
0.923828125,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"resented by fro, fa E RCXTxHxW. Cis the number of",
"for debasing. In this work, we try to debias through consis-",
"tency constraint.whichissimplebuteffectiveand doesnot",
"channel and T is the length of time dimension. W and H",
"need additional costs.",
"are spatial size. At last, the extracted features fro, frd are",
"pulled closer within the existing self-supervised methods.",
"3.Methodology",
"3.2. Background Erasing.",
"In this section we introduce the proposed Background",
"In the video representation learning, sometimes the sta-",
"Erasing (BE) method. We first give an overall description",
"of BE, and then introduce how to integrate BE into existing",
"tistical characteristics of the background will drown out the",
"self-supervised methods.",
"motion features of the moving subject. Thus it is easy",
"for the model to make predictions based only on the back-",
"3.1.Overall Architecture",
"ground information.And themodel is easy to overfitto the",
"training set and has poor generalization on the new dataset.",
"Consistency constraint with",
"Background Erasing(BE) is proposed to remove the neg-",
"self-supervisedtask",
"ative impact of the background by adding the background.",
"Specifically, for a video sequence , we randomly select",
"representations",
"",
"one static frame and add it as a spatial background noise to",
"介",
"every other frames to generate a distracting video, in which",
"each frame is obtained by the following formula:",
"介",
"= (1-)-rG) +>.r(k),jE[1,T",
"3D",
"3D",
"(1)",
"Backbon",
"Backbone",
"where 入is sampled from the uniform distribution [0, ],()",
"",
"means the j-th frame of c, k denotes the index of the ran-",
"domly selected frame and T is the length of the video se-",
"",
"↑",
"quence r. BE operation is applied to &, and the generated",
"Dornteainaydo",
"distracting video clip rd has a background perturbation on",
"the spatial dimension, but the motion pattern is basically not",
"Crop",
"个",
"Random",
"changed, as shown in Figure 3.",
"r",
"",
"Furthermore, it is easy to prove that the time derivative",
"of rd is a linear transformation of the time derivative of r,",
"formally:",
"Figure 2: The framework of the proposed method BE. A",
"d((1 - )r\" + ^)",
"(= (1-入)-",
"dau",
"video is first randomly cropped spatially, then we generate",
"(2)",
"dt",
"dt",
"frames. The model is trained by a existing self-supervised",
"the distracting video by adding a static frame upon other",
"where d represents the result of repeating the selected frame",
"task together with a consistency constraint, with the goal",
"(k) T times along the time dimension. Previous works[3,",
"of pulling the feature of the original video and that of the",
"5, 4, 51j have shown that the time derivative of a video clip",
"distracting video closer. (Best viewed in color).",
"is an important information for action classification, thus,",
"the property that BE maintains the linear transformation of",
"The framework of the proposed BE is shown in Figure 2.",
"such information is very crucial. Afterwards, we force the",
"For each input video r, we first randomly crop two fixed-",
"model to pull the feature of r and the feature of rd closer",
"length clips from different spatial locations, denoted as ro",
"which will be introduced in details later. Since ro and rd",
"and r\". Suppose we have a basic data augmentation set A",
"resembleeachother inthemotion patternbut differentiate",
"from which we sample two specific operations al and a?,",
"each other in spatial, when the features of ro and d are",
"and operate on r and \" respectively. In this way, the in-",
"brought closer, the model will be promoted to suppress the",
"put clips have different distribution in the pixel level but are",
"background noise,yielding video representations that are",
"consistent in the semantic level. Afterwards, r° is directly",
"more sensitive to motion changes. We have tried a variety of",
"fed into the 3D backbone to extract the feature representa-",
"ways to add background noise, results are shown in Table 4.",
"tion and we denote this procedure as F(r°; 0), where rep-",
"Experimental results demonstrate that the intra-video static",
"resents the backbone parameters. For \", we first generate",
"frame, i.e., BE, works best overall. Meanwhile, we have",
"a distracting counterpart rd for it, which has the interfer-",
"also tried to add various data augmentations to the selected",
"ence of added static frame noise but the semantics remains",
"intra-video static frame to introduce more disturbance, but",
"the same. The output feature maps of ro and rd are rep-",
"there is no positive gain.",
"3"
] |
{
"arxiv_id": "2009.05769",
"language": "en",
"page_idx": 2
}
|
||
2009.05769-en-3
|
2009.05769-en-3.jpg
|
<FigureHere>
Figure 3: Distracting Video Generation. One intra-video static frame is randomly selected and added to other frames as Noise. The background of the generated distracting video has changed, but the optical flow gradient is basically not changed, indicating that the motion pattern is retained.
## 3.3.Plug-and-Play
Using BE solely for optimization will make the model fall into a trivial solution easily. Therefore, we integrate BE into the existing self-supervised methods, specifically, we adopt two paradigms, handcrafted pretext and contrastive.
### 3.3.1 Pretext Task
Most pretext tasks can be formulated as a multi-category classification task and optimized with the cross-entropy loss. Specifically, each pretext will define a transformation set $R$ with $M$ operations. Given an input $x$, a transformation $r\in R$ is performed, then the convolutional neural network with parameters $\theta$ is required to distinguish which operation it is. The loss function is as follows:
(3)
$$\mathcal{L}_{p}=-\frac{1}{M}\sum_{r\in R}\mathcal{L}_{ce}(F(r(x);\theta),r),$$
where $\mathcal{L}_ce$ is Cross Entropy.
Plugged-in BE. For handcrafted pretext task, we use a consistency regularization term to pull the feature of $x^o$ closer to the feature of $x^d$, and make them consistent in the temporal dimension. Formally,
$$\mathcal{L}_{be}=\left\|\psi(f_{x^{o}})-\psi(f_{x^{d}})\right\|^{2}$$
(4)
where $\psi$ is an explicit feature mapping function that project features from $C\times T\times H\times W$ to $C\times T.$ We use spatial
global max pooling since $x^o$ and $x^d$ have different pixel distribution due to random cropping. In this fashion, we force the max response at each time dimension being consistent. And the final loss is:
(5)
$$\mathcal{L}=\mathcal{L}_{p}+\beta\mathcal{L}_{be}$$
where $\beta$ is a hyperparameter that controls the importance of
the regularization term. In our experiments, $\beta$ is set to 1.
## 3.3.2 Contrastive Learning
Contrastive learning [16] aims to learn an invariant representation for each sample, which is achieved by maximizing similarity of similar pairs over dissimilar pairs.
Plugged-in BE. Given a video dataset $\mathcal{D}$ with $N$ videos $\mathcal{D}=\{x_{1},x_{2},...,x_{N}\}$,for each video $x_i$,we randomly sample once in each epoch, obtaining $x_i^o$ and $x_i^d.$ In order to $\hat{\text{add a consistency constraint between }x^o}$ and $x^d$,we directly treat their features $f(x^o)$ and $f(x^d)$ as positive pairs instead of using MSE loss. Specifically, assume there is a projection function $\phi$, which consists of a spatio-temporal max pooling and a fully connected layer with $D$ dimension. Then the high level feature can be encoded by $z_x=\phi(f(x)).$ Given a particular video $x_i$ and clip sampling function $s$, the negative set $\mathcal{N}_1i$ is defined as: $\mathcal{N}_1i=\{s(x_n)|\forall n\neq i\}$,each element in $\mathcal{N}_1i$ is a clip and represents an identity, then the InfoNCE[36] loss is improved as follows:
$$\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N}log\frac{exp(z_{x_{i}^{o}}\cdot z_{x_{i}^{d}})}{exp(z_{x_{i}^{o}}\cdot z_{x_{i}^{d}})+\sum_{n\in\mathcal{N}_{1i}}exp(z_{x_{i}^{o}},z_{n})}$$
where $\cdot$ denotes the dot product. In this way, the optimiza-
tion goal is video-level discrimination in essence.
However, in order to discriminate each instance in $\mathcal{D}$, there may exist many spatial details. In order to make the objective more challenge, we introduce hard negatives, the different video clips with augmentation $a^1$ but from the same video. In this way, the optimization goal changes from video-level into clip-level, which is based on the observation that different clips of the same video contain different motion patterns but similar background. The hard negative set $\mathcal{N}_{2i}$ for $x_i$ is defined as: $\mathcal{N}_{2i}=\{x_{i}^{h}|x_{i}^{h}\neq x_{i}^{o},\bar{x}_{i}^{h}\in x_{i}\}$, and the overall negative set is $\mathcal{N}_i=\{\mathcal{N}_{1i}\cup\mathcal{N}_{2i}\}.$ Then the final objective function is:
$$\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N}log\frac{exp(z_{x_{i}^{o}}\cdot z_{x_{i}^{d}})}{exp(z_{x_{i}^{o}}\cdot z_{x_{i}^{d}})+\sum_{n\in\mathcal{N}_{i}}exp(z_{x_{i}^{o}}\cdot z_{n})}$$
(7)
For efficiency, we randomly select one hard negative sam-
ple from $\mathcal{N}_2i$ each iteration and we find more hard negative
samples have a similar result experimentally.
4
|
[
[
0.50341796875,
0.091552734375,
0.890625,
0.091552734375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.50341796875,
0.10797119140625,
0.89208984375,
0.10797119140625,
0.89208984375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.10540771484375,
0.12249755859375,
0.1519775390625,
0.12249755859375,
0.1519775390625,
0.1300048828125,
0.10540771484375,
0.1300048828125
],
[
0.1094970703125,
0.1307373046875,
0.1502685546875,
0.1307373046875,
0.1502685546875,
0.1383056640625,
0.1094970703125,
0.1383056640625
],
[
0.50341796875,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.50244140625,
0.1383056640625,
0.6396484375,
0.1383056640625,
0.6396484375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.6435546875,
0.1673583984375,
0.7490234375,
0.1673583984375,
0.7490234375,
0.1817626953125,
0.6435546875,
0.1817626953125
],
[
0.869140625,
0.1685791015625,
0.89208984375,
0.1685791015625,
0.89208984375,
0.1817626953125,
0.869140625,
0.1817626953125
],
[
0.1217041015625,
0.1849365234375,
0.1380615234375,
0.1849365234375,
0.1380615234375,
0.1932373046875,
0.1217041015625,
0.1932373046875
],
[
0.50244140625,
0.1988525390625,
0.89208984375,
0.1988525390625,
0.89208984375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.5009765625,
0.21337890625,
0.87158203125,
0.21337890625,
0.87158203125,
0.2266845703125,
0.5009765625,
0.2266845703125
],
[
0.17724609375,
0.2279052734375,
0.223876953125,
0.2279052734375,
0.223876953125,
0.23486328125,
0.17724609375,
0.23486328125
],
[
0.406005859375,
0.2279052734375,
0.418212890625,
0.2279052734375,
0.418212890625,
0.2310791015625,
0.406005859375,
0.2310791015625
],
[
0.1953125,
0.248779296875,
0.223876953125,
0.248779296875,
0.223876953125,
0.255615234375,
0.1953125,
0.255615234375
],
[
0.28759765625,
0.25,
0.3203125,
0.25,
0.3203125,
0.25390625,
0.28759765625,
0.25390625
],
[
0.374267578125,
0.25,
0.3857421875,
0.25,
0.3857421875,
0.25390625,
0.374267578125,
0.25390625
],
[
0.50244140625,
0.251953125,
0.70361328125,
0.25244140625,
0.70361328125,
0.267578125,
0.50244140625,
0.266357421875
],
[
0.11767578125,
0.2734375,
0.1422119140625,
0.2734375,
0.1422119140625,
0.285400390625,
0.11767578125,
0.285400390625
],
[
0.50244140625,
0.27978515625,
0.89208984375,
0.27978515625,
0.89208984375,
0.29345703125,
0.50244140625,
0.29345703125
],
[
0.50244140625,
0.29541015625,
0.89208984375,
0.29541015625,
0.89208984375,
0.309326171875,
0.50244140625,
0.309326171875
],
[
0.50244140625,
0.309326171875,
0.8095703125,
0.31005859375,
0.8095703125,
0.324462890625,
0.50244140625,
0.323974609375
],
[
0.50341796875,
0.326904296875,
0.89208984375,
0.326904296875,
0.89208984375,
0.34033203125,
0.50341796875,
0.34033203125
],
[
0.10455322265625,
0.333251953125,
0.156005859375,
0.333251953125,
0.156005859375,
0.343505859375,
0.10455322265625,
0.343505859375
],
[
0.111083984375,
0.34228515625,
0.148681640625,
0.34228515625,
0.148681640625,
0.349853515625,
0.111083984375,
0.349853515625
],
[
0.50244140625,
0.34033203125,
0.890625,
0.341552734375,
0.890625,
0.35595703125,
0.50244140625,
0.354736328125
],
[
0.50341796875,
0.356689453125,
0.89208984375,
0.356689453125,
0.89208984375,
0.37060546875,
0.50341796875,
0.37060546875
],
[
0.5009765625,
0.371826171875,
0.890625,
0.37060546875,
0.890625,
0.385009765625,
0.5009765625,
0.386474609375
],
[
0.08087158203125,
0.386474609375,
0.468017578125,
0.386474609375,
0.468017578125,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.50244140625,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50244140625,
0.40087890625
],
[
0.080078125,
0.401611328125,
0.4697265625,
0.401611328125,
0.4697265625,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.50244140625,
0.401611328125,
0.89208984375,
0.401611328125,
0.89208984375,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.08087158203125,
0.41845703125,
0.46728515625,
0.41845703125,
0.46728515625,
0.4287109375,
0.08087158203125,
0.4287109375
],
[
0.50244140625,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.50244140625,
0.43115234375
],
[
0.080078125,
0.431884765625,
0.4697265625,
0.431884765625,
0.4697265625,
0.44580078125,
0.080078125,
0.44580078125
],
[
0.5009765625,
0.431884765625,
0.89208984375,
0.43115234375,
0.89208984375,
0.44580078125,
0.5009765625,
0.4462890625
],
[
0.0767822265625,
0.4462890625,
0.43701171875,
0.44580078125,
0.43701171875,
0.460205078125,
0.0767822265625,
0.4609375
],
[
0.5009765625,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.4609375,
0.5009765625,
0.4609375
],
[
0.50244140625,
0.462158203125,
0.890625,
0.4609375,
0.890625,
0.47607421875,
0.50244140625,
0.4765625
],
[
0.50244140625,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.50244140625,
0.4912109375
],
[
0.080078125,
0.48486328125,
0.221435546875,
0.48486328125,
0.221435546875,
0.498779296875,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.098876953125,
0.50830078125,
0.468017578125,
0.50830078125,
0.468017578125,
0.521484375,
0.098876953125,
0.521484375
],
[
0.5009765625,
0.5068359375,
0.77783203125,
0.505859375,
0.77783203125,
0.52099609375,
0.5009765625,
0.521484375
],
[
0.08087158203125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.08087158203125,
0.537109375
],
[
0.080078125,
0.5390625,
0.4697265625,
0.5390625,
0.4697265625,
0.55322265625,
0.080078125,
0.55322265625
],
[
0.580078125,
0.537109375,
0.5947265625,
0.537109375,
0.5947265625,
0.546875,
0.580078125,
0.546875
],
[
0.7060546875,
0.5390625,
0.80126953125,
0.5390625,
0.80126953125,
0.5537109375,
0.7060546875,
0.5537109375
],
[
0.51123046875,
0.5517578125,
0.53125,
0.5517578125,
0.53125,
0.56005859375,
0.51123046875,
0.56005859375
],
[
0.5947265625,
0.55029296875,
0.62158203125,
0.55029296875,
0.62158203125,
0.5625,
0.5947265625,
0.5625
],
[
0.61767578125,
0.556640625,
0.8837890625,
0.55810546875,
0.8837890625,
0.5771484375,
0.61767578125,
0.57568359375
],
[
0.5537109375,
0.55810546875,
0.58837890625,
0.55615234375,
0.5888671875,
0.5673828125,
0.5556640625,
0.5693359375
],
[
0.07843017578125,
0.55322265625,
0.459228515625,
0.5537109375,
0.459228515625,
0.568359375,
0.07843017578125,
0.5673828125
],
[
0.580078125,
0.568359375,
0.5986328125,
0.568359375,
0.5986328125,
0.57568359375,
0.580078125,
0.57568359375
],
[
0.87158203125,
0.5791015625,
0.8896484375,
0.5791015625,
0.8896484375,
0.58837890625,
0.87158203125,
0.58837890625
],
[
0.080078125,
0.5859375,
0.2197265625,
0.5859375,
0.2197265625,
0.599609375,
0.080078125,
0.599609375
],
[
0.50341796875,
0.59033203125,
0.890625,
0.59033203125,
0.890625,
0.60400390625,
0.50341796875,
0.60400390625
],
[
0.080078125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.60595703125,
0.82421875,
0.60595703125,
0.82421875,
0.619140625,
0.50244140625,
0.619140625
],
[
0.07843017578125,
0.62353515625,
0.468017578125,
0.625,
0.468017578125,
0.6396484375,
0.07843017578125,
0.63818359375
],
[
0.52294921875,
0.62353515625,
0.890625,
0.62353515625,
0.890625,
0.6337890625,
0.52294921875,
0.6337890625
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50244140625,
0.63720703125,
0.89208984375,
0.63720703125,
0.89208984375,
0.6513671875,
0.50244140625,
0.6513671875
],
[
0.080078125,
0.65576171875,
0.4697265625,
0.65576171875,
0.4697265625,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50341796875,
0.65380859375,
0.890625,
0.65380859375,
0.890625,
0.6640625,
0.50341796875,
0.6640625
],
[
0.080078125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.50244140625,
0.66748046875,
0.89208984375,
0.66748046875,
0.89208984375,
0.68115234375,
0.50244140625,
0.68115234375
],
[
0.080078125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50341796875,
0.68310546875,
0.89208984375,
0.68310546875,
0.89208984375,
0.6962890625,
0.50341796875,
0.6962890625
],
[
0.080078125,
0.7021484375,
0.36767578125,
0.7021484375,
0.36767578125,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.5048828125,
0.69970703125,
0.890625,
0.69970703125,
0.890625,
0.71044921875,
0.5048828125,
0.71044921875
],
[
0.50244140625,
0.71337890625,
0.89208984375,
0.71337890625,
0.89208984375,
0.7265625,
0.50244140625,
0.7265625
],
[
0.2484130859375,
0.724609375,
0.39208984375,
0.7265625,
0.39208984375,
0.744140625,
0.2484130859375,
0.7421875
],
[
0.5009765625,
0.72802734375,
0.89208984375,
0.7265625,
0.89208984375,
0.7412109375,
0.5009765625,
0.7421875
],
[
0.44677734375,
0.72900390625,
0.4697265625,
0.72900390625,
0.4697265625,
0.7412109375,
0.44677734375,
0.7412109375
],
[
0.16015625,
0.72900390625,
0.19287109375,
0.72900390625,
0.19287109375,
0.7412109375,
0.16015625,
0.7412109375
],
[
0.20751953125,
0.736328125,
0.2459716796875,
0.73291015625,
0.2467041015625,
0.744140625,
0.2099609375,
0.74853515625
],
[
0.2467041015625,
0.734375,
0.2548828125,
0.73193359375,
0.256591796875,
0.73486328125,
0.2484130859375,
0.7373046875
],
[
0.2344970703125,
0.744140625,
0.26318359375,
0.744140625,
0.26318359375,
0.7568359375,
0.2344970703125,
0.7568359375
],
[
0.50341796875,
0.7421875,
0.890625,
0.7421875,
0.890625,
0.7568359375,
0.50341796875,
0.7568359375
],
[
0.080078125,
0.76220703125,
0.265625,
0.76318359375,
0.265625,
0.77783203125,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.7587890625,
0.89208984375,
0.7587890625,
0.89208984375,
0.77294921875,
0.50244140625,
0.77294921875
],
[
0.080078125,
0.77783203125,
0.468017578125,
0.7763671875,
0.468017578125,
0.79150390625,
0.080078125,
0.79248046875
],
[
0.50341796875,
0.7744140625,
0.6748046875,
0.7744140625,
0.6748046875,
0.78515625,
0.50341796875,
0.78515625
],
[
0.080078125,
0.79345703125,
0.4697265625,
0.79345703125,
0.4697265625,
0.806640625,
0.080078125,
0.806640625
],
[
0.576171875,
0.80029296875,
0.5986328125,
0.80029296875,
0.5986328125,
0.81298828125,
0.576171875,
0.81298828125
],
[
0.080078125,
0.8095703125,
0.4697265625,
0.8095703125,
0.4697265625,
0.82275390625,
0.080078125,
0.82275390625
],
[
0.7060546875,
0.8046875,
0.80224609375,
0.8046875,
0.80224609375,
0.8193359375,
0.7060546875,
0.8193359375
],
[
0.51708984375,
0.814453125,
0.52783203125,
0.82080078125,
0.51953125,
0.8291015625,
0.5087890625,
0.8212890625
],
[
0.56396484375,
0.81103515625,
0.5703125,
0.81103515625,
0.5703125,
0.8173828125,
0.56396484375,
0.8173828125
],
[
0.58837890625,
0.814453125,
0.62744140625,
0.8173828125,
0.6259765625,
0.830078125,
0.58642578125,
0.8271484375
],
[
0.080078125,
0.82470703125,
0.259033203125,
0.82470703125,
0.259033203125,
0.83837890625,
0.080078125,
0.83837890625
],
[
0.55224609375,
0.82568359375,
0.5849609375,
0.8173828125,
0.5908203125,
0.8291015625,
0.55810546875,
0.8369140625
],
[
0.62158203125,
0.82373046875,
0.8857421875,
0.82373046875,
0.8857421875,
0.84130859375,
0.62158203125,
0.84130859375
],
[
0.57666015625,
0.83203125,
0.60107421875,
0.83203125,
0.60107421875,
0.84326171875,
0.57666015625,
0.84326171875
],
[
0.87158203125,
0.8388671875,
0.89208984375,
0.84228515625,
0.88818359375,
0.85595703125,
0.86767578125,
0.85302734375
],
[
0.178955078125,
0.84521484375,
0.36767578125,
0.84326171875,
0.36767578125,
0.861328125,
0.178955078125,
0.86279296875
],
[
0.44677734375,
0.84765625,
0.4697265625,
0.84765625,
0.4697265625,
0.861328125,
0.44677734375,
0.861328125
],
[
0.50244140625,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.79736328125,
0.88720703125,
0.79736328125,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"global max pooling since r and rd have different pixel dis-",
"tribution due to random cropping. In this fashion, we force",
"Optical-flov",
"Gradient",
"themaxresponse at each timedimension being consistent.",
"And the final loss is:",
"C=Cp+BCbe",
"(5)",
"RGB",
"where β is a hyperparameter that controls the importance of",
"the regularization term. In our experiments, β is set to 1.",
"StaticNoi",
"D0",
"",
"",
"",
"3.3.2ContrastiveLearning",
"RGB",
"Contrastive learning [16] aims to learn an invariant repre-",
"sentation for each sample, which is achieved by maximizing",
"similarity of similar pairs over dissimilar pairs.",
"Plugged-in BE. Given a video dataset D with N videos",
"Optical-flow",
"Gradient",
"D = {ri, r2,.., &l, for each video ri, we randomly sam-",
"ple once in each epoch, obtaining rg and g. In order to",
"add a consistency constraint between r and ad, we directly",
"Figure 3: Distracting Video Generation. One intra-video",
"treat their features f(r) and f(rd) as positive pairs instead",
"static frame is randomly selected and added to other frames",
"of using MSE loss. Specifically, assume there is a projection",
"asNoise.Thebackground ofthegenerated distracting video",
"function p, which consists of a spatio-temporal max pool-",
"has changed, but the optical flow gradient is basically not",
"ing and a fully connected layer with D dimension. Then the",
" changed, indicating that the motion pattern is retained.",
"high level feature can be encoded by z = @(f(α)). Given",
"a particular video i and clip sampling function s, the neg-",
"ative set Nii is defined as: Nii = (s(n)/Vn ≠ i), each",
"3.3.Plug-and-Play",
"element in Ni; is a clip and represents an identity, then the",
"Using BE solely for optimization will make the model",
"InfoNCE[36] loss is improved as follows:",
"fall into a trivial solution easily. Therefore, we integrate BE",
"into the existing self-supervised methods, specifically, we",
"N",
"erp(zro : zrd)",
"",
"log",
"erp(zmzg)+EneNi.eap(za,zn)",
"NZ",
"adopt two paradigms, handcrafted pretext and contrastive.",
"=1",
"(6)",
"3.3.1Pretext Task",
"where • denotes the dot product. In this way, the optimiza-",
"Most pretext tasks can be formulated as a multi-category",
"tion goal is video-level discrimination in essence.",
"classification task and optimized with the cross-entropy",
"However. in orderto discriminate each instance in D",
"loss. Specifically, each pretext will define a transformation",
"there may exist many spatial details. In order to make the",
"set R with M operations. Given an input , a transforma-",
"objectivemorechallenge,weintroducehardnegatives,the",
"tion r E R is performed, then the convolutional neural net-",
"different video clips with augmentation al but from the",
"work with parameters is required to distinguish which op-",
"same video. In this way, the optimization goal changes from",
"eration it is. The loss function is as follows:",
"video-level intoclip-level,whichisbasedontheobservation",
"that different clips of the same video contain different mo-",
"Cce(F(r(c); 0), r),",
"tion patterns but similar background. The hard negative set",
"(3)",
"Cp=",
"MZ",
"",
"TER",
"N2i for ris defined as: N2i = ahch 手 ri,rh e cil,",
"where Cee is Cross Entropy.",
"and the overall negative set is N, = [Ni, U N2i]. Then the",
"Plugged-in BE.For handcrafted pretext task, we use a con-",
"final objectivefunctionis:",
"sistency regularization term to pull the feature of ro closer",
"N",
"to the feature of rd, and make them consistent in the tem-",
"erp(zro .zra)",
"",
"L",
"log",
"poral dimension. Formally,",
"N",
"eap(zazag)+EneN eap(zr:zn)",
"i=1",
"(7)",
"Cbe = lb(fro)- b(faa)2",
"(4)",
"For efficiency, we randomly select one hard negative sam-",
"where b is an explicit feature mapping function that project",
"ple from N2i each iteration and we find more hard negative",
"features from C × T × H × W to C × T. We use spatial",
"samples have a similar result experimentally.",
"4"
] |
{
"arxiv_id": "2009.05769",
"language": "en",
"page_idx": 3
}
|
||
2009.05769-en-4
|
2009.05769-en-4.jpg
|
<table>
<tbody>
<tr>
<th colspan="2">$\mathbf{Method}$</th>
<th rowspan="2">Depth</th>
<th rowspan="2">${\mathbf{D}}$ataset(duration)</th>
<th colspan="2">$\mathbf{Pretrain}$</th>
<th rowspan="2">$\mathbf{Single-Mod}$</th>
<th> </th>
<th colspan="2">$\mathbf{Fine-tune}$</th>
</tr>
<tr>
<th>$\mathbf{Method}(\mathbf{vear})$</th>
<th>$\mathbf{Backbone}$</th>
<th>Frame</th>
<th>Res</th>
<th>$C/P$</th>
<th>$\mathbf{UCF101}$</th>
<th>$\mathbf{HMDB51}$</th>
</tr>
<tr>
<td>Supervised</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Random Init</td>
<td>$I3D$</td>
<td>22</td>
<td>$X$</td>
<td> </td>
<td>224</td>
<td> </td>
<td>-</td>
<td>60.5</td>
<td>21.2</td>
</tr>
<tr>
<td>ImageNet Supervised</td>
<td>$I3D$</td>
<td>22</td>
<td>ImageNet</td>
<td> </td>
<td>224</td>
<td> </td>
<td>-</td>
<td>67.1</td>
<td>28.5</td>
</tr>
<tr>
<td>K400 Supervised</td>
<td>$I3D$</td>
<td>22</td>
<td>$\mathbf{K}400(28$d)</td>
<td> </td>
<td>224</td>
<td> </td>
<td>-</td>
<td>96.8</td>
<td>74.5</td>
</tr>
<tr>
<td>Self-supervised</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Shuffle[34] [ECCV, 2016]</td>
<td>AlexNet</td>
<td>8</td>
<td>UCF101( 1d) </td>
<td> </td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>50.2</td>
<td>18.1</td>
</tr>
<tr>
<td>VGAN[ 47] [NeurlPS, 2016]</td>
<td>$VGAN$</td>
<td>22</td>
<td>UCF101( 1d) </td>
<td> </td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>52.1</td>
<td> </td>
</tr>
<tr>
<td>OPN [28]| [ICCV,2017T</td>
<td>Caffe Net</td>
<td>14</td>
<td>UCF101( 1d) </td>
<td> </td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>56.3</td>
<td>22.1</td>
</tr>
<tr>
<td>Geometry[12] 1 [CVPR,2018]</td>
<td>Flow Net</td>
<td>56</td>
<td>UCF101( 1d) </td>
<td>16</td>
<td>112</td>
<td>$X$</td>
<td>$P$</td>
<td>55.1</td>
<td>23.3</td>
</tr>
<tr>
<td>IIC$\left[43\right]\left[\mathrm{ACM~MM,~2020}\right]$</td>
<td>$C3D$</td>
<td>10</td>
<td>UCF101( 1d) </td>
<td>16</td>
<td>112</td>
<td>$X$</td>
<td>$C$</td>
<td>72.7</td>
<td>36.8</td>
</tr>
<tr>
<td>Pace [49] [ECCV. 2020]</td>
<td>$\mathbf{R}(2+1)\mathbf{D}$</td>
<td>23</td>
<td>$\mathbf{K}400(28$d)</td>
<td>16</td>
<td>112</td>
<td> </td>
<td>$C$</td>
<td>77.1</td>
<td>36.6</td>
</tr>
<tr>
<td>3D RotNet[23] [2018]</td>
<td>$C3D$</td>
<td>10</td>
<td>$\mathbf{K}400(28$d)</td>
<td>16</td>
<td>$\overline{1}1\overline{2}$</td>
<td> </td>
<td>$P$</td>
<td>62.9</td>
<td>33.7</td>
</tr>
<tr>
<td>$3D$RotNet$+BE$</td>
<td>$C3D$</td>
<td>10</td>
<td>$\mathbf{K}400(28$d)</td>
<td>16</td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>$65.4(2.5\uparrow)$</td>
<td>$37.4(3.7\uparrow)$</td>
</tr>
<tr>
<td>$\bar{\mathrm{STPuzzles}}$ ;[26] [AAAI] 20191</td>
<td>$\bar{\mathrm{C}}3\bar{\mathrm{D}}$</td>
<td>$1\overline{0}$</td>
<td>$\bar{\mathrm{UCF1}}\overline{0}1\overline{(1}\overline{\mathrm{d}})$</td>
<td>48</td>
<td>$\bar{1}1\overline{2}$</td>
<td> </td>
<td>$P$</td>
<td>$6\overline{0}.\overline{6}$</td>
<td>$\bar{28}.\overline{3}$</td>
</tr>
<tr>
<td>ST Puzzles+BE</td>
<td>$C3D$</td>
<td>10</td>
<td>UCF101( 1d) </td>
<td>48</td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>$63.7(3.1\uparrow)$</td>
<td>$30.8(2.5\uparrow)$</td>
</tr>
<tr>
<td>$\bar{\mathrm{Clip}}\bar{\mathrm{~}}\bar{\mathrm{Or}}\bar{\mathrm{der}}[\bar{5}7]\bar{\mathrm{[CVPR}}$ 2019]</td>
<td>$\bar{\mathrm{C}}3\bar{\mathrm{D}}$</td>
<td>$1\overline{0}$</td>
<td>UCF101( 1d) </td>
<td>64</td>
<td>$\bar{1}1\overline{2}$</td>
<td> </td>
<td>$P$</td>
<td>$6\overline{5}.\overline{6}$</td>
<td>28.4</td>
</tr>
<tr>
<td>$\mathbf{Clip}$Order$+\mathbf{BE}$</td>
<td>$C3D$</td>
<td>10</td>
<td>UCF101( 1d) </td>
<td>64</td>
<td>112</td>
<td> </td>
<td>$P$</td>
<td>$68.5(2.9\uparrow)$</td>
<td>$32.8(4.4\uparrow)$</td>
</tr>
<tr>
<td>$\overline{\mathrm{MoCo}}$ [21][CVPR $\overline{2}0\overline{2}0\overline{1}$</td>
<td>$\bar{\mathrm{C}}3\bar{\mathrm{D}}$</td>
<td>$1\overline{0}$</td>
<td>$\bar{\mathrm{UCF1}}\overline{0}1\overline{(1}\overline{\mathrm{d}})$</td>
<td>176</td>
<td>$\overline{1}1\overline{2}$</td>
<td> </td>
<td>$\dot{C}$</td>
<td>$6\overline{0}.5$</td>
<td>$\bar{27}.\overline{2}$</td>
</tr>
<tr>
<td>$\mathbf{MoCo}+\mathbf{BE}$</td>
<td>$C3D$</td>
<td>10</td>
<td>UCF101( 1d) </td>
<td>16</td>
<td>112</td>
<td> </td>
<td>$C$</td>
<td>$72.4(11.9\uparrow)$</td>
<td>$42.3(14.1\uparrow)$</td>
</tr>
<tr>
<td>$\bar{\mathrm{Co}}\bar{\mathrm{CLR}}[1\bar{9}]$ [NeuIPS, 2020]</td>
<td>$\bar{\mathrm{R}}3\bar{\mathrm{D}}$</td>
<td>$2\overline{3}$</td>
<td>$\mathbf{K}400(28$d)</td>
<td>$3\overline{2}$</td>
<td>$\overline{1}2\overline{8}$</td>
<td>$X$</td>
<td>$\dot{C}$</td>
<td>$8\overline{7}.\overline{9}$</td>
<td>$\bar{5}4.\overline{6}$</td>
</tr>
<tr>
<td>DPC$\left [ 17\right ]$ICCW,2019</td>
<td>$R3D$</td>
<td>34</td>
<td>$\mathbf{K}400(28$d)</td>
<td>64</td>
<td>224</td>
<td> </td>
<td>$P$</td>
<td>75.7</td>
<td>35.7</td>
</tr>
<tr>
<td>AoT [54][CVPR,2018</td>
<td>T- CAM</td>
<td> </td>
<td>$\mathbf{K}400(28$d)</td>
<td>64</td>
<td>224</td>
<td> </td>
<td>$P$</td>
<td>79.4</td>
<td> </td>
</tr>
<tr>
<td>Pace [49] [ECCV. 2020]</td>
<td>S3D- G</td>
<td>23</td>
<td>$\mathbf{K}400(28$d)</td>
<td>64</td>
<td>224</td>
<td> </td>
<td>$C$</td>
<td>87.1</td>
<td>52.6</td>
</tr>
<tr>
<td>SpeedNet[1][CVPR,2020]</td>
<td>S3D- G</td>
<td>23</td>
<td>$\mathbf{K}400(28$d)</td>
<td>64</td>
<td>224</td>
<td> </td>
<td>$P$</td>
<td>81.1</td>
<td>48.8</td>
</tr>
<tr>
<td>SpeedNet[1][CVPR,2020]</td>
<td>$I3D$</td>
<td>22</td>
<td>$\mathbf{K}400(28$d)</td>
<td>64</td>
<td>224</td>
<td> </td>
<td>$P$</td>
<td>66.7</td>
<td>43.7</td>
</tr>
<tr>
<td>$\overline{\mathrm{MoCo}}$ [21][CVPR $\overline{2}0\overline{2}0\overline{1}$</td>
<td>$I3D$</td>
<td>$\bar{2}\overline{2}$</td>
<td>$\bar{\mathrm{K}}4\overline{00}(\overline{28}\overline{\mathrm{d}})$</td>
<td>176</td>
<td>$\overline{2}2\overline{4}$</td>
<td> </td>
<td>$\dot{C}$</td>
<td>$7\overline{0}.\overline{4}$</td>
<td>$\bar{3}6.\overline{3}$</td>
</tr>
<tr>
<td>$\mathbf{MoCo}+\mathbf{BE}$</td>
<td>$I3D$</td>
<td>22</td>
<td>$\mathbf{K}400(28$d)</td>
<td>16</td>
<td>224</td>
<td> </td>
<td>$C$</td>
<td>$86.8(16.4\uparrow)$</td>
<td>$55.4(19.1\uparrow)$</td>
</tr>
<tr>
<td>$\mathbf{MoCo}+\mathbf{BE}$</td>
<td>$I3D$</td>
<td>$2\overline{2}$</td>
<td>UCF101( 1d) </td>
<td>$1\overline{6}$</td>
<td>$\overline{2}2\overline{4}$</td>
<td> </td>
<td>$\bar{C}$</td>
<td>82.4</td>
<td>$5\overline{2}.\overline{9}$</td>
</tr>
<tr>
<td>$\mathbf{MoCo}+\mathbf{BE}$</td>
<td>$R3D$</td>
<td>34</td>
<td>UCF101( 1d) </td>
<td>16</td>
<td>224</td>
<td> </td>
<td>$C$</td>
<td>83.4</td>
<td>53.7</td>
</tr>
<tr>
<td>$\mathbf{MoCo}+\mathbf{BE}$</td>
<td>$R3D$</td>
<td>34</td>
<td>$\mathbf{K}400(28$d)</td>
<td>16</td>
<td>224</td>
<td> </td>
<td>$C$</td>
<td>87.1</td>
<td>56.2</td>
</tr>
</tbody>
</table>
Table 1: Top-1 accuracy (%) of integrating BE as a regularization term to four existing approaches and compared with previ-
ous methods on the UCF101 and HMDB51 dataset. Single-Mod denotes Single-Modality, C/P represents Contrastive/Pretext
task,◇means our implementation, K400 is short for Kinetics-400 and d represents day.
## $4. \textbf{Experiments}$
4.1. Implementation Details
original video with a temporal stride of 4. Then the sampled clip is resized to $16\times3\times112\times112$ or $16\times3\times224\times224.$ The $\gamma$ of Background Erasing is experimentally set to 0.3, and a larger value may result in excessive blur. The choice of temporal stride and $\gamma$ is analysed in the supplementary. The basic augmentation set $\mathcal{A}$ contains random rotation less than 10 degrees and color jittering, and all these operations are applied in a temporal consistent way, that is, each frame of a video uses the same augmentation. The vector dimension $D$ is 128.
Datasets. All the experiments are conducted on four video datasets, UCF101 [41], HMDB51 [27], Kinetics [25] and Diving48[30]. The first three contain prominent bias, while Diving48 is less biased. UCF101 is a realistic video dataset with 13,320 videos of 101 action categories. HMDB51 contains 6,849 clips of 51 action categories. Kinetics is a large scale action recognition dataset that contains 246k/20k train/val video clips of 400 classes. Diving48 consists of 18k trimmed video clips of 48 diving sequences.
Networks. We use C3D[45], R3D[20] and I3D[7] as base encoders followed by a spatio-temporal max pooling layer. Default Settings. All the experiments are conducted on 8 Tesla V100 GPUs with a batch size of 64 under PyTorch[37] framework. We adopt SGD as our optimizer with momentum of 0.9 and weight decay of 5e-4.
Self-supervised Pre-training Settings. We pre-train the network for 50 epochs with the learning rate initialized as 0.01 and decreased to 1/10 every 10 epochs. The input clip consists of 16 frames, which is uniformly sampled from the
Supervised Fine-tuning Settings. After pre-training, we transfer the weights of the base encoder to two downstream tasks, $i.e.$, action recognition and video retrieval, with the last fully connected layer randomly initialized. We fine-tune the network for 45 epochs. The learning rate is initialized as 0.05 and decreases to 1/10 every 10 epochs.
Evaluation Settings. For action recognition, following common practice[57], the final result of a video is the average of the results of 10 clips that are uniformly sampled from it during testing time.
5
|
[
[
0.2132568359375,
0.091552734375,
0.259033203125,
0.091552734375,
0.259033203125,
0.1029052734375,
0.2132568359375,
0.1029052734375
],
[
0.5361328125,
0.091552734375,
0.5849609375,
0.091552734375,
0.5849609375,
0.1029052734375,
0.5361328125,
0.1029052734375
],
[
0.78173828125,
0.091552734375,
0.8408203125,
0.091552734375,
0.8408203125,
0.1029052734375,
0.78173828125,
0.1029052734375
],
[
0.08416748046875,
0.1029052734375,
0.1683349609375,
0.10418701171875,
0.1683349609375,
0.11553955078125,
0.08416748046875,
0.1142578125
],
[
0.26953125,
0.10418701171875,
0.328369140625,
0.10418701171875,
0.328369140625,
0.11553955078125,
0.26953125,
0.11553955078125
],
[
0.343017578125,
0.10418701171875,
0.38232421875,
0.10418701171875,
0.38232421875,
0.11553955078125,
0.343017578125,
0.11553955078125
],
[
0.57275390625,
0.10418701171875,
0.59619140625,
0.10418701171875,
0.59619140625,
0.11553955078125,
0.57275390625,
0.11553955078125
],
[
0.61279296875,
0.10418701171875,
0.6787109375,
0.10418701171875,
0.6787109375,
0.11553955078125,
0.61279296875,
0.11553955078125
],
[
0.69677734375,
0.10418701171875,
0.72119140625,
0.10418701171875,
0.72119140625,
0.11553955078125,
0.69677734375,
0.11553955078125
],
[
0.734375,
0.10418701171875,
0.78515625,
0.10418701171875,
0.78515625,
0.11553955078125,
0.734375,
0.11553955078125
],
[
0.81884765625,
0.10418701171875,
0.87353515625,
0.10418701171875,
0.87353515625,
0.11553955078125,
0.81884765625,
0.11553955078125
],
[
0.51708984375,
0.10479736328125,
0.55810546875,
0.10479736328125,
0.55810546875,
0.11614990234375,
0.51708984375,
0.11614990234375
],
[
0.39453125,
0.10479736328125,
0.5009765625,
0.10479736328125,
0.5009765625,
0.11553955078125,
0.39453125,
0.11553955078125
],
[
0.0849609375,
0.11614990234375,
0.156005859375,
0.1142578125,
0.156005859375,
0.1287841796875,
0.0849609375,
0.1307373046875
],
[
0.34228515625,
0.126953125,
0.36279296875,
0.126953125,
0.36279296875,
0.1427001953125,
0.34228515625,
0.1427001953125
],
[
0.57177734375,
0.1256103515625,
0.5986328125,
0.1256103515625,
0.5986328125,
0.1427001953125,
0.57177734375,
0.1427001953125
],
[
0.736328125,
0.1275634765625,
0.765625,
0.1275634765625,
0.765625,
0.14013671875,
0.736328125,
0.14013671875
],
[
0.81787109375,
0.1275634765625,
0.84619140625,
0.1275634765625,
0.84619140625,
0.14013671875,
0.81787109375,
0.14013671875
],
[
0.268798828125,
0.1275634765625,
0.294189453125,
0.1275634765625,
0.294189453125,
0.14013671875,
0.268798828125,
0.14013671875
],
[
0.08905029296875,
0.1287841796875,
0.158447265625,
0.1287841796875,
0.158447265625,
0.1395263671875,
0.08905029296875,
0.1395263671875
],
[
0.39794921875,
0.1287841796875,
0.408447265625,
0.1287841796875,
0.408447265625,
0.1383056640625,
0.39794921875,
0.1383056640625
],
[
0.640625,
0.1287841796875,
0.65185546875,
0.1287841796875,
0.65185546875,
0.141357421875,
0.640625,
0.141357421875
],
[
0.339111328125,
0.1363525390625,
0.365234375,
0.1363525390625,
0.365234375,
0.1666259765625,
0.339111328125,
0.1666259765625
],
[
0.5146484375,
0.1383056640625,
0.521484375,
0.1307373046875,
0.52685546875,
0.1337890625,
0.51953125,
0.141357421875
],
[
0.56787109375,
0.1383056640625,
0.59619140625,
0.1337890625,
0.6005859375,
0.1514892578125,
0.57177734375,
0.1558837890625
],
[
0.63818359375,
0.1383056640625,
0.654296875,
0.1383056640625,
0.654296875,
0.1641845703125,
0.63818359375,
0.1641845703125
],
[
0.734375,
0.1383056640625,
0.765625,
0.1383056640625,
0.765625,
0.154052734375,
0.734375,
0.154052734375
],
[
0.814453125,
0.1383056640625,
0.8486328125,
0.1383056640625,
0.8486328125,
0.1546630859375,
0.814453125,
0.1546630859375
],
[
0.26708984375,
0.1395263671875,
0.295654296875,
0.1395263671875,
0.295654296875,
0.154052734375,
0.26708984375,
0.154052734375
],
[
0.0882568359375,
0.1395263671875,
0.205078125,
0.14013671875,
0.205078125,
0.15283203125,
0.0882568359375,
0.1514892578125
],
[
0.39453125,
0.14013671875,
0.453369140625,
0.14013671875,
0.453369140625,
0.1514892578125,
0.39453125,
0.1514892578125
],
[
0.5146484375,
0.141357421875,
0.525390625,
0.141357421875,
0.525390625,
0.154052734375,
0.5146484375,
0.154052734375
],
[
0.701171875,
0.141357421875,
0.70751953125,
0.146484375,
0.701171875,
0.1514892578125,
0.69384765625,
0.146484375
],
[
0.26708984375,
0.1483154296875,
0.294189453125,
0.15087890625,
0.293212890625,
0.1641845703125,
0.2646484375,
0.1610107421875
],
[
0.73388671875,
0.149658203125,
0.76708984375,
0.149658203125,
0.76708984375,
0.1654052734375,
0.73388671875,
0.1654052734375
],
[
0.814453125,
0.149658203125,
0.84716796875,
0.149658203125,
0.84716796875,
0.1654052734375,
0.814453125,
0.1654052734375
],
[
0.51708984375,
0.15087890625,
0.52783203125,
0.154052734375,
0.52294921875,
0.1641845703125,
0.51123046875,
0.1610107421875
],
[
0.08660888671875,
0.15087890625,
0.181396484375,
0.15087890625,
0.181396484375,
0.1622314453125,
0.08660888671875,
0.1622314453125
],
[
0.396240234375,
0.1514892578125,
0.459228515625,
0.1514892578125,
0.459228515625,
0.1634521484375,
0.396240234375,
0.1634521484375
],
[
0.57177734375,
0.15087890625,
0.59716796875,
0.15087890625,
0.59716796875,
0.1634521484375,
0.57177734375,
0.1634521484375
],
[
0.69775390625,
0.1546630859375,
0.70361328125,
0.1546630859375,
0.70361328125,
0.1610107421875,
0.69775390625,
0.1610107421875
],
[
0.08251953125,
0.1641845703125,
0.176513671875,
0.1641845703125,
0.176513671875,
0.175537109375,
0.08251953125,
0.175537109375
],
[
0.348876953125,
0.17236328125,
0.36279296875,
0.18310546875,
0.3505859375,
0.1912841796875,
0.338134765625,
0.1817626953125
],
[
0.39453125,
0.17236328125,
0.4697265625,
0.1749267578125,
0.4697265625,
0.189453125,
0.39453125,
0.1875
],
[
0.57275390625,
0.175537109375,
0.59716796875,
0.175537109375,
0.59716796875,
0.1881103515625,
0.57275390625,
0.1881103515625
],
[
0.69677734375,
0.175537109375,
0.7099609375,
0.175537109375,
0.7099609375,
0.1875,
0.69677734375,
0.1875
],
[
0.73779296875,
0.175537109375,
0.765625,
0.175537109375,
0.765625,
0.1881103515625,
0.73779296875,
0.1881103515625
],
[
0.81787109375,
0.175537109375,
0.8447265625,
0.175537109375,
0.8447265625,
0.1881103515625,
0.81787109375,
0.1881103515625
],
[
0.26953125,
0.1767578125,
0.31787109375,
0.1767578125,
0.31787109375,
0.1881103515625,
0.26953125,
0.1881103515625
],
[
0.08905029296875,
0.1749267578125,
0.233642578125,
0.1749267578125,
0.233642578125,
0.1881103515625,
0.08905029296875,
0.1881103515625
],
[
0.51708984375,
0.1817626953125,
0.52294921875,
0.1817626953125,
0.52294921875,
0.1875,
0.51708984375,
0.1875
],
[
0.6396484375,
0.1767578125,
0.65185546875,
0.1767578125,
0.65185546875,
0.1881103515625,
0.6396484375,
0.1881103515625
],
[
0.63623046875,
0.184326171875,
0.65673828125,
0.184326171875,
0.65673828125,
0.21533203125,
0.63623046875,
0.21533203125
],
[
0.70166015625,
0.18310546875,
0.7158203125,
0.1932373046875,
0.70361328125,
0.20263671875,
0.689453125,
0.1925048828125
],
[
0.734375,
0.186279296875,
0.765625,
0.186279296875,
0.765625,
0.2020263671875,
0.734375,
0.2020263671875
],
[
0.34228515625,
0.186279296875,
0.36279296875,
0.186279296875,
0.36279296875,
0.2008056640625,
0.34228515625,
0.2008056640625
],
[
0.08905029296875,
0.1875,
0.2442626953125,
0.1875,
0.2442626953125,
0.2008056640625,
0.08905029296875,
0.2008056640625
],
[
0.39453125,
0.1875,
0.4697265625,
0.1875,
0.4697265625,
0.2020263671875,
0.39453125,
0.2020263671875
],
[
0.57275390625,
0.1875,
0.59716796875,
0.1875,
0.59716796875,
0.2008056640625,
0.57275390625,
0.2008056640625
],
[
0.515625,
0.189453125,
0.525390625,
0.189453125,
0.525390625,
0.2008056640625,
0.515625,
0.2008056640625
],
[
0.26953125,
0.1881103515625,
0.310546875,
0.1881103515625,
0.310546875,
0.199462890625,
0.26953125,
0.199462890625
],
[
0.343017578125,
0.1988525390625,
0.36279296875,
0.1988525390625,
0.36279296875,
0.21337890625,
0.343017578125,
0.21337890625
],
[
0.396240234375,
0.1988525390625,
0.468017578125,
0.2008056640625,
0.468017578125,
0.21337890625,
0.396240234375,
0.2108154296875
],
[
0.57275390625,
0.199462890625,
0.59716796875,
0.199462890625,
0.59716796875,
0.212158203125,
0.57275390625,
0.212158203125
],
[
0.69384765625,
0.1988525390625,
0.71142578125,
0.1988525390625,
0.71142578125,
0.212158203125,
0.69384765625,
0.212158203125
],
[
0.736328125,
0.199462890625,
0.765625,
0.199462890625,
0.765625,
0.212158203125,
0.736328125,
0.212158203125
],
[
0.81787109375,
0.199462890625,
0.84619140625,
0.199462890625,
0.84619140625,
0.212158203125,
0.81787109375,
0.212158203125
],
[
0.268798828125,
0.2008056640625,
0.324462890625,
0.2008056640625,
0.324462890625,
0.212158203125,
0.268798828125,
0.212158203125
],
[
0.0906982421875,
0.2008056640625,
0.2156982421875,
0.2008056640625,
0.2156982421875,
0.2108154296875,
0.0906982421875,
0.2108154296875
],
[
0.08905029296875,
0.2108154296875,
0.2484130859375,
0.2108154296875,
0.2484130859375,
0.2247314453125,
0.08905029296875,
0.2247314453125
],
[
0.343017578125,
0.2108154296875,
0.36279296875,
0.2108154296875,
0.36279296875,
0.2259521484375,
0.343017578125,
0.2259521484375
],
[
0.5146484375,
0.2108154296875,
0.53515625,
0.2108154296875,
0.53515625,
0.2247314453125,
0.5146484375,
0.2247314453125
],
[
0.57275390625,
0.2108154296875,
0.59716796875,
0.2108154296875,
0.59716796875,
0.2247314453125,
0.57275390625,
0.2247314453125
],
[
0.69384765625,
0.2108154296875,
0.71142578125,
0.2108154296875,
0.71142578125,
0.2247314453125,
0.69384765625,
0.2247314453125
],
[
0.39794921875,
0.212158203125,
0.468017578125,
0.212158203125,
0.468017578125,
0.2235107421875,
0.39794921875,
0.2235107421875
],
[
0.268798828125,
0.212158203125,
0.32275390625,
0.212158203125,
0.32275390625,
0.2235107421875,
0.268798828125,
0.2235107421875
],
[
0.736328125,
0.212158203125,
0.7646484375,
0.212158203125,
0.7646484375,
0.2247314453125,
0.736328125,
0.2247314453125
],
[
0.81640625,
0.212158203125,
0.84619140625,
0.212158203125,
0.84619140625,
0.2247314453125,
0.81640625,
0.2247314453125
],
[
0.396240234375,
0.2215576171875,
0.468017578125,
0.2235107421875,
0.468017578125,
0.238037109375,
0.396240234375,
0.236083984375
],
[
0.63232421875,
0.2139892578125,
0.64892578125,
0.20703125,
0.65869140625,
0.2215576171875,
0.64208984375,
0.2291259765625
],
[
0.343017578125,
0.222900390625,
0.36279296875,
0.222900390625,
0.36279296875,
0.2374267578125,
0.343017578125,
0.2374267578125
],
[
0.515625,
0.222900390625,
0.53515625,
0.222900390625,
0.53515625,
0.236083984375,
0.515625,
0.236083984375
],
[
0.6396484375,
0.222900390625,
0.654296875,
0.222900390625,
0.654296875,
0.24365234375,
0.6396484375,
0.24365234375
],
[
0.69384765625,
0.222900390625,
0.71142578125,
0.222900390625,
0.71142578125,
0.2374267578125,
0.69384765625,
0.2374267578125
],
[
0.57275390625,
0.2235107421875,
0.59716796875,
0.2235107421875,
0.59716796875,
0.236083984375,
0.57275390625,
0.236083984375
],
[
0.268798828125,
0.2235107421875,
0.298095703125,
0.2235107421875,
0.298095703125,
0.236083984375,
0.268798828125,
0.236083984375
],
[
0.736328125,
0.2235107421875,
0.765625,
0.2235107421875,
0.765625,
0.236083984375,
0.736328125,
0.236083984375
],
[
0.81787109375,
0.2235107421875,
0.84619140625,
0.2235107421875,
0.84619140625,
0.236083984375,
0.81787109375,
0.236083984375
],
[
0.70166015625,
0.231689453125,
0.7138671875,
0.240478515625,
0.70166015625,
0.251953125,
0.689453125,
0.24365234375
],
[
0.08905029296875,
0.2247314453125,
0.233642578125,
0.2247314453125,
0.233642578125,
0.23486328125,
0.08905029296875,
0.23486328125
],
[
0.34228515625,
0.2342529296875,
0.36279296875,
0.2342529296875,
0.36279296875,
0.248779296875,
0.34228515625,
0.248779296875
],
[
0.734375,
0.2342529296875,
0.76708984375,
0.2342529296875,
0.76708984375,
0.25,
0.734375,
0.25
],
[
0.81640625,
0.2342529296875,
0.84619140625,
0.236083984375,
0.8447265625,
0.248779296875,
0.814453125,
0.246826171875
],
[
0.26708984375,
0.23486328125,
0.32177734375,
0.23486328125,
0.32177734375,
0.25,
0.26708984375,
0.25
],
[
0.0882568359375,
0.23486328125,
0.2197265625,
0.23486328125,
0.2197265625,
0.248779296875,
0.0882568359375,
0.248779296875
],
[
0.396240234375,
0.23486328125,
0.4599609375,
0.23486328125,
0.4599609375,
0.25,
0.396240234375,
0.25
],
[
0.515625,
0.23486328125,
0.5361328125,
0.23486328125,
0.5361328125,
0.248779296875,
0.515625,
0.248779296875
],
[
0.57275390625,
0.23486328125,
0.59716796875,
0.23486328125,
0.59716796875,
0.248779296875,
0.57275390625,
0.248779296875
],
[
0.268798828125,
0.246826171875,
0.298095703125,
0.248779296875,
0.29736328125,
0.261962890625,
0.26708984375,
0.260009765625
],
[
0.70166015625,
0.24365234375,
0.7138671875,
0.25244140625,
0.70166015625,
0.26318359375,
0.689453125,
0.25390625
],
[
0.340576171875,
0.2474365234375,
0.36279296875,
0.2474365234375,
0.36279296875,
0.261474609375,
0.340576171875,
0.261474609375
],
[
0.0849609375,
0.2474365234375,
0.2132568359375,
0.2474365234375,
0.2132568359375,
0.261474609375,
0.0849609375,
0.261474609375
],
[
0.51708984375,
0.2474365234375,
0.53515625,
0.2474365234375,
0.53515625,
0.261474609375,
0.51708984375,
0.261474609375
],
[
0.57177734375,
0.2474365234375,
0.59716796875,
0.2474365234375,
0.59716796875,
0.261474609375,
0.57177734375,
0.261474609375
],
[
0.734375,
0.2474365234375,
0.765625,
0.2474365234375,
0.765625,
0.260009765625,
0.734375,
0.260009765625
],
[
0.81640625,
0.2474365234375,
0.8447265625,
0.2474365234375,
0.8447265625,
0.260009765625,
0.81640625,
0.260009765625
],
[
0.814453125,
0.2568359375,
0.87744140625,
0.2587890625,
0.87744140625,
0.2734375,
0.814453125,
0.271484375
],
[
0.39794921875,
0.248779296875,
0.459228515625,
0.248779296875,
0.459228515625,
0.260009765625,
0.39794921875,
0.260009765625
],
[
0.343017578125,
0.2587890625,
0.36279296875,
0.2587890625,
0.36279296875,
0.272705078125,
0.343017578125,
0.272705078125
],
[
0.515625,
0.2587890625,
0.5361328125,
0.2587890625,
0.5361328125,
0.272705078125,
0.515625,
0.272705078125
],
[
0.57275390625,
0.2587890625,
0.59716796875,
0.2587890625,
0.59716796875,
0.272705078125,
0.57275390625,
0.272705078125
],
[
0.26953125,
0.260009765625,
0.2998046875,
0.260009765625,
0.2998046875,
0.272705078125,
0.26953125,
0.272705078125
],
[
0.08905029296875,
0.260009765625,
0.1829833984375,
0.260009765625,
0.1829833984375,
0.271484375,
0.08905029296875,
0.271484375
],
[
0.39453125,
0.260009765625,
0.459228515625,
0.260009765625,
0.459228515625,
0.271484375,
0.39453125,
0.271484375
],
[
0.6953125,
0.260009765625,
0.708984375,
0.260009765625,
0.708984375,
0.271484375,
0.6953125,
0.271484375
],
[
0.734375,
0.260009765625,
0.7958984375,
0.260009765625,
0.7958984375,
0.271484375,
0.734375,
0.271484375
],
[
0.08660888671875,
0.26953125,
0.253173828125,
0.270751953125,
0.253173828125,
0.285400390625,
0.08660888671875,
0.2841796875
],
[
0.268798828125,
0.271484375,
0.298095703125,
0.271484375,
0.298095703125,
0.2841796875,
0.268798828125,
0.2841796875
],
[
0.34228515625,
0.271484375,
0.36279296875,
0.271484375,
0.36279296875,
0.285400390625,
0.34228515625,
0.285400390625
],
[
0.515625,
0.271484375,
0.53515625,
0.271484375,
0.53515625,
0.285400390625,
0.515625,
0.285400390625
],
[
0.57275390625,
0.271484375,
0.59716796875,
0.271484375,
0.59716796875,
0.2841796875,
0.57275390625,
0.2841796875
],
[
0.69140625,
0.271484375,
0.71142578125,
0.271484375,
0.71142578125,
0.2841796875,
0.69140625,
0.2841796875
],
[
0.81640625,
0.271484375,
0.8447265625,
0.271484375,
0.8447265625,
0.2841796875,
0.81640625,
0.2841796875
],
[
0.734375,
0.272705078125,
0.765625,
0.272705078125,
0.765625,
0.2841796875,
0.734375,
0.2841796875
],
[
0.39794921875,
0.272705078125,
0.468017578125,
0.272705078125,
0.468017578125,
0.2841796875,
0.39794921875,
0.2841796875
],
[
0.26953125,
0.2841796875,
0.2998046875,
0.2841796875,
0.2998046875,
0.296630859375,
0.26953125,
0.296630859375
],
[
0.343017578125,
0.2841796875,
0.363525390625,
0.2841796875,
0.363525390625,
0.29736328125,
0.343017578125,
0.29736328125
],
[
0.515625,
0.2841796875,
0.5361328125,
0.2841796875,
0.5361328125,
0.29736328125,
0.515625,
0.29736328125
],
[
0.57275390625,
0.2841796875,
0.59716796875,
0.2841796875,
0.59716796875,
0.29736328125,
0.57275390625,
0.29736328125
],
[
0.6953125,
0.2841796875,
0.708984375,
0.2841796875,
0.708984375,
0.29541015625,
0.6953125,
0.29541015625
],
[
0.39453125,
0.285400390625,
0.468017578125,
0.285400390625,
0.468017578125,
0.296630859375,
0.39453125,
0.296630859375
],
[
0.08905029296875,
0.285400390625,
0.1846923828125,
0.285400390625,
0.1846923828125,
0.296630859375,
0.08905029296875,
0.296630859375
],
[
0.734375,
0.285400390625,
0.79736328125,
0.285400390625,
0.79736328125,
0.296630859375,
0.734375,
0.296630859375
],
[
0.81884765625,
0.285400390625,
0.87744140625,
0.285400390625,
0.87744140625,
0.296630859375,
0.81884765625,
0.296630859375
],
[
0.08660888671875,
0.294189453125,
0.253173828125,
0.29541015625,
0.253173828125,
0.31005859375,
0.08660888671875,
0.309326171875
],
[
0.268798828125,
0.296630859375,
0.298095703125,
0.296630859375,
0.298095703125,
0.309326171875,
0.268798828125,
0.309326171875
],
[
0.34228515625,
0.296630859375,
0.36279296875,
0.296630859375,
0.36279296875,
0.31005859375,
0.34228515625,
0.31005859375
],
[
0.5146484375,
0.296630859375,
0.53515625,
0.296630859375,
0.53515625,
0.309326171875,
0.5146484375,
0.309326171875
],
[
0.57275390625,
0.296630859375,
0.59716796875,
0.296630859375,
0.59716796875,
0.309326171875,
0.57275390625,
0.309326171875
],
[
0.69287109375,
0.296630859375,
0.71142578125,
0.296630859375,
0.71142578125,
0.309326171875,
0.69287109375,
0.309326171875
],
[
0.734375,
0.296630859375,
0.765625,
0.296630859375,
0.765625,
0.309326171875,
0.734375,
0.309326171875
],
[
0.81787109375,
0.296630859375,
0.8447265625,
0.296630859375,
0.8447265625,
0.309326171875,
0.81787109375,
0.309326171875
],
[
0.69775390625,
0.30615234375,
0.71142578125,
0.31005859375,
0.7060546875,
0.32080078125,
0.69287109375,
0.31640625
],
[
0.396240234375,
0.29736328125,
0.46728515625,
0.29736328125,
0.46728515625,
0.309326171875,
0.396240234375,
0.309326171875
],
[
0.73193359375,
0.30615234375,
0.79736328125,
0.30810546875,
0.79736328125,
0.322509765625,
0.73193359375,
0.32080078125
],
[
0.81787109375,
0.30615234375,
0.875,
0.30810546875,
0.87353515625,
0.322509765625,
0.81640625,
0.32080078125
],
[
0.268798828125,
0.306884765625,
0.301513671875,
0.306884765625,
0.301513671875,
0.322509765625,
0.268798828125,
0.322509765625
],
[
0.08905029296875,
0.30810546875,
0.1846923828125,
0.30810546875,
0.1846923828125,
0.3193359375,
0.08905029296875,
0.3193359375
],
[
0.343017578125,
0.30810546875,
0.36279296875,
0.30810546875,
0.36279296875,
0.3212890625,
0.343017578125,
0.3212890625
],
[
0.393798828125,
0.30810546875,
0.4697265625,
0.30810546875,
0.4697265625,
0.322509765625,
0.393798828125,
0.322509765625
],
[
0.515625,
0.30810546875,
0.53515625,
0.30810546875,
0.53515625,
0.3212890625,
0.515625,
0.3212890625
],
[
0.57275390625,
0.30810546875,
0.59716796875,
0.30810546875,
0.59716796875,
0.3212890625,
0.57275390625,
0.3212890625
],
[
0.08660888671875,
0.318115234375,
0.23779296875,
0.318115234375,
0.23779296875,
0.333251953125,
0.08660888671875,
0.333251953125
],
[
0.26708984375,
0.3193359375,
0.2998046875,
0.3193359375,
0.2998046875,
0.335205078125,
0.26708984375,
0.335205078125
],
[
0.340576171875,
0.32080078125,
0.36279296875,
0.32080078125,
0.36279296875,
0.333984375,
0.340576171875,
0.333984375
],
[
0.51708984375,
0.32080078125,
0.5361328125,
0.32080078125,
0.5361328125,
0.335205078125,
0.51708984375,
0.335205078125
],
[
0.57275390625,
0.32080078125,
0.59716796875,
0.32080078125,
0.59716796875,
0.333251953125,
0.57275390625,
0.333251953125
],
[
0.734375,
0.32080078125,
0.765625,
0.32080078125,
0.765625,
0.333251953125,
0.734375,
0.333251953125
],
[
0.81787109375,
0.32080078125,
0.8447265625,
0.32080078125,
0.8447265625,
0.333251953125,
0.81787109375,
0.333251953125
],
[
0.689453125,
0.324462890625,
0.705078125,
0.317626953125,
0.7138671875,
0.328857421875,
0.69921875,
0.3359375
],
[
0.396240234375,
0.3212890625,
0.46728515625,
0.3212890625,
0.46728515625,
0.333251953125,
0.396240234375,
0.333251953125
],
[
0.69677734375,
0.330078125,
0.71142578125,
0.333251953125,
0.70751953125,
0.34521484375,
0.69287109375,
0.34228515625
],
[
0.736328125,
0.330078125,
0.80224609375,
0.33203125,
0.80224609375,
0.3466796875,
0.736328125,
0.3447265625
],
[
0.396240234375,
0.33203125,
0.4697265625,
0.33203125,
0.4697265625,
0.3466796875,
0.396240234375,
0.3466796875
],
[
0.343017578125,
0.33203125,
0.36279296875,
0.33203125,
0.36279296875,
0.34521484375,
0.343017578125,
0.34521484375
],
[
0.515625,
0.33203125,
0.53515625,
0.33203125,
0.53515625,
0.34521484375,
0.515625,
0.34521484375
],
[
0.57275390625,
0.33203125,
0.59716796875,
0.33203125,
0.59716796875,
0.34521484375,
0.57275390625,
0.34521484375
],
[
0.26953125,
0.333251953125,
0.298095703125,
0.333251953125,
0.298095703125,
0.3447265625,
0.26953125,
0.3447265625
],
[
0.08905029296875,
0.333984375,
0.1568603515625,
0.333984375,
0.1568603515625,
0.3447265625,
0.08905029296875,
0.3447265625
],
[
0.6435546875,
0.3359375,
0.64892578125,
0.3359375,
0.64892578125,
0.341552734375,
0.6435546875,
0.341552734375
],
[
0.81884765625,
0.333984375,
0.88232421875,
0.333984375,
0.88232421875,
0.34521484375,
0.81884765625,
0.34521484375
],
[
0.0882568359375,
0.3447265625,
0.238525390625,
0.343505859375,
0.238525390625,
0.356689453125,
0.0882568359375,
0.35791015625
],
[
0.34228515625,
0.3447265625,
0.36279296875,
0.3447265625,
0.36279296875,
0.35791015625,
0.34228515625,
0.35791015625
],
[
0.39453125,
0.34228515625,
0.459228515625,
0.3447265625,
0.459228515625,
0.359130859375,
0.39453125,
0.356689453125
],
[
0.5146484375,
0.3447265625,
0.53369140625,
0.3447265625,
0.53369140625,
0.35791015625,
0.5146484375,
0.35791015625
],
[
0.57275390625,
0.3447265625,
0.59619140625,
0.3447265625,
0.59619140625,
0.356689453125,
0.57275390625,
0.356689453125
],
[
0.736328125,
0.34521484375,
0.765625,
0.34521484375,
0.765625,
0.35791015625,
0.736328125,
0.35791015625
],
[
0.81787109375,
0.34521484375,
0.84619140625,
0.34521484375,
0.84619140625,
0.35986328125,
0.81787109375,
0.35986328125
],
[
0.640625,
0.347900390625,
0.65185546875,
0.347900390625,
0.65185546875,
0.35791015625,
0.640625,
0.35791015625
],
[
0.268798828125,
0.34521484375,
0.298095703125,
0.34521484375,
0.298095703125,
0.35791015625,
0.268798828125,
0.35791015625
],
[
0.0882568359375,
0.356689453125,
0.21728515625,
0.356689453125,
0.21728515625,
0.37060546875,
0.0882568359375,
0.37060546875
],
[
0.343017578125,
0.356689453125,
0.36279296875,
0.356689453125,
0.36279296875,
0.37060546875,
0.343017578125,
0.37060546875
],
[
0.5146484375,
0.356689453125,
0.53515625,
0.356689453125,
0.53515625,
0.37060546875,
0.5146484375,
0.37060546875
],
[
0.57275390625,
0.356689453125,
0.59716796875,
0.356689453125,
0.59716796875,
0.37060546875,
0.57275390625,
0.37060546875
],
[
0.69140625,
0.35107421875,
0.705078125,
0.34228515625,
0.7138671875,
0.35107421875,
0.701171875,
0.359130859375
],
[
0.268798828125,
0.35791015625,
0.298095703125,
0.35791015625,
0.298095703125,
0.369384765625,
0.268798828125,
0.369384765625
],
[
0.39794921875,
0.35791015625,
0.45751953125,
0.35791015625,
0.45751953125,
0.369384765625,
0.39794921875,
0.369384765625
],
[
0.6953125,
0.35791015625,
0.708984375,
0.35791015625,
0.708984375,
0.369384765625,
0.6953125,
0.369384765625
],
[
0.736328125,
0.35791015625,
0.765625,
0.35791015625,
0.765625,
0.37060546875,
0.736328125,
0.37060546875
],
[
0.81787109375,
0.356689453125,
0.84619140625,
0.356689453125,
0.84619140625,
0.371826171875,
0.81787109375,
0.371826171875
],
[
0.69921875,
0.365478515625,
0.71337890625,
0.37060546875,
0.705078125,
0.38330078125,
0.69140625,
0.37744140625
],
[
0.640625,
0.359130859375,
0.65185546875,
0.359130859375,
0.65185546875,
0.37060546875,
0.640625,
0.37060546875
],
[
0.26708984375,
0.36865234375,
0.314453125,
0.36865234375,
0.314453125,
0.38330078125,
0.26708984375,
0.38330078125
],
[
0.396240234375,
0.36865234375,
0.4599609375,
0.36865234375,
0.4599609375,
0.38330078125,
0.396240234375,
0.38330078125
],
[
0.5146484375,
0.369384765625,
0.53515625,
0.369384765625,
0.53515625,
0.38330078125,
0.5146484375,
0.38330078125
],
[
0.57275390625,
0.369384765625,
0.59716796875,
0.369384765625,
0.59716796875,
0.38330078125,
0.57275390625,
0.38330078125
],
[
0.736328125,
0.369384765625,
0.765625,
0.369384765625,
0.765625,
0.3818359375,
0.736328125,
0.3818359375
],
[
0.640625,
0.37060546875,
0.65185546875,
0.37060546875,
0.65185546875,
0.38330078125,
0.640625,
0.38330078125
],
[
0.0906982421875,
0.37060546875,
0.2181396484375,
0.37060546875,
0.2181396484375,
0.380615234375,
0.0906982421875,
0.380615234375
],
[
0.340576171875,
0.37890625,
0.363525390625,
0.37890625,
0.363525390625,
0.397705078125,
0.340576171875,
0.397705078125
],
[
0.5146484375,
0.380126953125,
0.53515625,
0.380126953125,
0.53515625,
0.39453125,
0.5146484375,
0.39453125
],
[
0.81640625,
0.380126953125,
0.84716796875,
0.380126953125,
0.84716796875,
0.395751953125,
0.81640625,
0.395751953125
],
[
0.57275390625,
0.380615234375,
0.59716796875,
0.380615234375,
0.59716796875,
0.39453125,
0.57275390625,
0.39453125
],
[
0.0882568359375,
0.380615234375,
0.2197265625,
0.380615234375,
0.2197265625,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.640625,
0.380615234375,
0.65185546875,
0.380615234375,
0.65185546875,
0.393310546875,
0.640625,
0.393310546875
],
[
0.69384765625,
0.380615234375,
0.71142578125,
0.380615234375,
0.71142578125,
0.39453125,
0.69384765625,
0.39453125
],
[
0.39794921875,
0.3818359375,
0.459228515625,
0.3818359375,
0.459228515625,
0.393310546875,
0.39794921875,
0.393310546875
],
[
0.268798828125,
0.3818359375,
0.310546875,
0.3818359375,
0.310546875,
0.393310546875,
0.268798828125,
0.393310546875
],
[
0.736328125,
0.3818359375,
0.7646484375,
0.3818359375,
0.7646484375,
0.39453125,
0.736328125,
0.39453125
],
[
0.0882568359375,
0.392578125,
0.240234375,
0.391357421875,
0.240234375,
0.406005859375,
0.0882568359375,
0.4072265625
],
[
0.26708984375,
0.392578125,
0.31201171875,
0.392578125,
0.31201171875,
0.4072265625,
0.26708984375,
0.4072265625
],
[
0.34228515625,
0.392578125,
0.36279296875,
0.392578125,
0.36279296875,
0.4072265625,
0.34228515625,
0.4072265625
],
[
0.396240234375,
0.392578125,
0.4599609375,
0.392578125,
0.4599609375,
0.4072265625,
0.396240234375,
0.4072265625
],
[
0.5146484375,
0.392578125,
0.53515625,
0.392578125,
0.53515625,
0.4072265625,
0.5146484375,
0.4072265625
],
[
0.57275390625,
0.392578125,
0.59716796875,
0.392578125,
0.59716796875,
0.406005859375,
0.57275390625,
0.406005859375
],
[
0.69384765625,
0.392578125,
0.71142578125,
0.392578125,
0.71142578125,
0.406005859375,
0.69384765625,
0.406005859375
],
[
0.734375,
0.392578125,
0.76708984375,
0.392578125,
0.76708984375,
0.40771484375,
0.734375,
0.40771484375
],
[
0.81640625,
0.392578125,
0.84716796875,
0.392578125,
0.84716796875,
0.40771484375,
0.81640625,
0.40771484375
],
[
0.26708984375,
0.40283203125,
0.294189453125,
0.40478515625,
0.293212890625,
0.41845703125,
0.2646484375,
0.416748046875
],
[
0.640625,
0.393310546875,
0.65185546875,
0.393310546875,
0.65185546875,
0.406005859375,
0.640625,
0.406005859375
],
[
0.34228515625,
0.40283203125,
0.36279296875,
0.40283203125,
0.36279296875,
0.41845703125,
0.34228515625,
0.41845703125
],
[
0.814453125,
0.404052734375,
0.84619140625,
0.404052734375,
0.84619140625,
0.419189453125,
0.814453125,
0.419189453125
],
[
0.5146484375,
0.40478515625,
0.5361328125,
0.40478515625,
0.5361328125,
0.41845703125,
0.5146484375,
0.41845703125
],
[
0.0882568359375,
0.40478515625,
0.240234375,
0.40478515625,
0.240234375,
0.41845703125,
0.0882568359375,
0.41845703125
],
[
0.57275390625,
0.40478515625,
0.59716796875,
0.40478515625,
0.59716796875,
0.417236328125,
0.57275390625,
0.417236328125
],
[
0.69384765625,
0.40478515625,
0.71142578125,
0.40478515625,
0.71142578125,
0.41845703125,
0.69384765625,
0.41845703125
],
[
0.734375,
0.40478515625,
0.765625,
0.40478515625,
0.765625,
0.417236328125,
0.734375,
0.417236328125
],
[
0.6435546875,
0.4072265625,
0.64892578125,
0.4072265625,
0.64892578125,
0.416748046875,
0.6435546875,
0.416748046875
],
[
0.39794921875,
0.406005859375,
0.459228515625,
0.406005859375,
0.459228515625,
0.417236328125,
0.39794921875,
0.417236328125
],
[
0.0882568359375,
0.416748046875,
0.238525390625,
0.416748046875,
0.238525390625,
0.429931640625,
0.0882568359375,
0.429931640625
],
[
0.340576171875,
0.416748046875,
0.36279296875,
0.416748046875,
0.36279296875,
0.429931640625,
0.340576171875,
0.429931640625
],
[
0.393798828125,
0.415283203125,
0.4599609375,
0.417236328125,
0.4599609375,
0.431884765625,
0.393798828125,
0.429931640625
],
[
0.57275390625,
0.416748046875,
0.59716796875,
0.416748046875,
0.59716796875,
0.4287109375,
0.57275390625,
0.4287109375
],
[
0.51708984375,
0.417236328125,
0.5361328125,
0.417236328125,
0.5361328125,
0.429931640625,
0.51708984375,
0.429931640625
],
[
0.26708984375,
0.417236328125,
0.294189453125,
0.417236328125,
0.294189453125,
0.429931640625,
0.26708984375,
0.429931640625
],
[
0.69287109375,
0.417236328125,
0.71142578125,
0.417236328125,
0.71142578125,
0.429931640625,
0.69287109375,
0.429931640625
],
[
0.73779296875,
0.417236328125,
0.765625,
0.417236328125,
0.765625,
0.429931640625,
0.73779296875,
0.429931640625
],
[
0.81640625,
0.417236328125,
0.8447265625,
0.417236328125,
0.8447265625,
0.429931640625,
0.81640625,
0.429931640625
],
[
0.343017578125,
0.427978515625,
0.361083984375,
0.427978515625,
0.361083984375,
0.442626953125,
0.343017578125,
0.442626953125
],
[
0.69384765625,
0.427978515625,
0.7099609375,
0.427978515625,
0.7099609375,
0.443115234375,
0.69384765625,
0.443115234375
],
[
0.734375,
0.4267578125,
0.8056640625,
0.429931640625,
0.8037109375,
0.4443359375,
0.73388671875,
0.44140625
],
[
0.81787109375,
0.427978515625,
0.88232421875,
0.429931640625,
0.88232421875,
0.443115234375,
0.81787109375,
0.44140625
],
[
0.57275390625,
0.4287109375,
0.59619140625,
0.4287109375,
0.59619140625,
0.442626953125,
0.57275390625,
0.442626953125
],
[
0.08905029296875,
0.429931640625,
0.16015625,
0.429931640625,
0.16015625,
0.44140625,
0.08905029296875,
0.44140625
],
[
0.268798828125,
0.4287109375,
0.294189453125,
0.4287109375,
0.294189453125,
0.442626953125,
0.268798828125,
0.442626953125
],
[
0.396240234375,
0.429931640625,
0.459228515625,
0.429931640625,
0.459228515625,
0.44140625,
0.396240234375,
0.44140625
],
[
0.5185546875,
0.43115234375,
0.53369140625,
0.43115234375,
0.53369140625,
0.44140625,
0.5185546875,
0.44140625
],
[
0.340576171875,
0.43994140625,
0.36279296875,
0.43994140625,
0.36279296875,
0.455078125,
0.340576171875,
0.455078125
],
[
0.39208984375,
0.439453125,
0.4697265625,
0.44140625,
0.4697265625,
0.455810546875,
0.39208984375,
0.453857421875
],
[
0.69287109375,
0.43994140625,
0.71337890625,
0.43994140625,
0.71337890625,
0.4697265625,
0.69287109375,
0.4697265625
],
[
0.73388671875,
0.43994140625,
0.76708984375,
0.43994140625,
0.76708984375,
0.455810546875,
0.73388671875,
0.455810546875
],
[
0.814453125,
0.43994140625,
0.84619140625,
0.43994140625,
0.84619140625,
0.455810546875,
0.814453125,
0.455810546875
],
[
0.57275390625,
0.44140625,
0.59716796875,
0.44140625,
0.59716796875,
0.455078125,
0.57275390625,
0.455078125
],
[
0.268798828125,
0.44140625,
0.29736328125,
0.44140625,
0.29736328125,
0.455810546875,
0.268798828125,
0.455810546875
],
[
0.63623046875,
0.44140625,
0.65625,
0.44140625,
0.65625,
0.479736328125,
0.63623046875,
0.479736328125
],
[
0.515625,
0.442626953125,
0.5361328125,
0.442626953125,
0.5361328125,
0.455078125,
0.515625,
0.455078125
],
[
0.08905029296875,
0.442626953125,
0.16015625,
0.442626953125,
0.16015625,
0.453857421875,
0.08905029296875,
0.453857421875
],
[
0.57177734375,
0.451904296875,
0.5986328125,
0.451904296875,
0.5986328125,
0.4697265625,
0.57177734375,
0.4697265625
],
[
0.81787109375,
0.449462890625,
0.84619140625,
0.45263671875,
0.84326171875,
0.47021484375,
0.81396484375,
0.46728515625
],
[
0.343017578125,
0.45263671875,
0.36279296875,
0.45263671875,
0.36279296875,
0.468505859375,
0.343017578125,
0.468505859375
],
[
0.26708984375,
0.45263671875,
0.2998046875,
0.45263671875,
0.2998046875,
0.468505859375,
0.26708984375,
0.468505859375
],
[
0.39453125,
0.453857421875,
0.468017578125,
0.453857421875,
0.468017578125,
0.468505859375,
0.39453125,
0.468505859375
],
[
0.515625,
0.453857421875,
0.5361328125,
0.453857421875,
0.5361328125,
0.468505859375,
0.515625,
0.468505859375
],
[
0.734375,
0.45263671875,
0.76708984375,
0.45263671875,
0.76708984375,
0.468505859375,
0.734375,
0.468505859375
],
[
0.343017578125,
0.464111328125,
0.36279296875,
0.464111328125,
0.36279296875,
0.479248046875,
0.343017578125,
0.479248046875
],
[
0.0906982421875,
0.455078125,
0.16015625,
0.455078125,
0.16015625,
0.466552734375,
0.0906982421875,
0.466552734375
],
[
0.69384765625,
0.464111328125,
0.71142578125,
0.464111328125,
0.71142578125,
0.479248046875,
0.69384765625,
0.479248046875
],
[
0.73388671875,
0.464111328125,
0.765625,
0.464111328125,
0.765625,
0.479736328125,
0.73388671875,
0.479736328125
],
[
0.81640625,
0.464111328125,
0.84716796875,
0.464111328125,
0.84716796875,
0.479736328125,
0.81640625,
0.479736328125
],
[
0.57177734375,
0.46533203125,
0.59716796875,
0.46533203125,
0.59716796875,
0.479248046875,
0.57177734375,
0.479248046875
],
[
0.0882568359375,
0.466552734375,
0.16015625,
0.466552734375,
0.16015625,
0.477783203125,
0.0882568359375,
0.477783203125
],
[
0.26708984375,
0.466552734375,
0.298095703125,
0.466552734375,
0.298095703125,
0.479248046875,
0.26708984375,
0.479248046875
],
[
0.396240234375,
0.466552734375,
0.459228515625,
0.466552734375,
0.459228515625,
0.477783203125,
0.396240234375,
0.477783203125
],
[
0.515625,
0.46533203125,
0.5361328125,
0.46533203125,
0.5361328125,
0.479736328125,
0.515625,
0.479736328125
],
[
0.08087158203125,
0.494384765625,
0.88818359375,
0.494384765625,
0.88818359375,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.080078125,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.52294921875,
0.080078125,
0.52294921875
],
[
0.080078125,
0.5244140625,
0.65283203125,
0.5244140625,
0.65283203125,
0.53857421875,
0.080078125,
0.53857421875
],
[
0.080078125,
0.564453125,
0.2091064453125,
0.56494140625,
0.2091064453125,
0.580078125,
0.080078125,
0.5791015625
],
[
0.50244140625,
0.5673828125,
0.89208984375,
0.5673828125,
0.89208984375,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.50244140625,
0.58203125,
0.890625,
0.58203125,
0.890625,
0.59521484375,
0.50244140625,
0.59521484375
],
[
0.080078125,
0.59033203125,
0.293212890625,
0.59033203125,
0.293212890625,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.080078125,
0.6142578125,
0.4697265625,
0.6142578125,
0.4697265625,
0.62744140625,
0.080078125,
0.62744140625
],
[
0.50244140625,
0.6123046875,
0.8935546875,
0.6123046875,
0.8935546875,
0.62548828125,
0.50244140625,
0.62548828125
],
[
0.080078125,
0.62890625,
0.4697265625,
0.62890625,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.080078125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50341796875,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65478515625,
0.50341796875,
0.65478515625
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.07843017578125,
0.673828125,
0.468017578125,
0.67236328125,
0.468017578125,
0.6875,
0.07843017578125,
0.68798828125
],
[
0.5009765625,
0.67236328125,
0.8935546875,
0.671875,
0.8935546875,
0.68603515625,
0.5009765625,
0.6875
],
[
0.07843017578125,
0.689453125,
0.471435546875,
0.689453125,
0.471435546875,
0.70263671875,
0.07843017578125,
0.70263671875
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.7021484375,
0.56640625,
0.7021484375,
0.56640625,
0.71630859375,
0.50244140625,
0.71630859375
],
[
0.080078125,
0.7197265625,
0.471435546875,
0.7197265625,
0.471435546875,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.08087158203125,
0.734375,
0.39453125,
0.734375,
0.39453125,
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,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.080078125,
0.76318359375
],
[
0.50244140625,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.465576171875,
0.7666015625,
0.465576171875,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79443359375,
0.89208984375,
0.79443359375,
0.89208984375,
0.80810546875,
0.50244140625,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.81005859375,
0.8056640625,
0.81005859375,
0.8056640625,
0.82373046875,
0.5009765625,
0.82373046875
],
[
0.080078125,
0.8271484375,
0.32177734375,
0.8271484375,
0.32177734375,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84228515625,
0.65283203125,
0.84228515625,
0.65283203125,
0.85595703125,
0.50244140625,
0.85595703125
],
[
0.65673828125,
0.84130859375,
0.89208984375,
0.84228515625,
0.89208984375,
0.85595703125,
0.65673828125,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.07843017578125,
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,
0.68115234375,
0.88720703125,
0.68115234375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.4794921875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"Method",
"Pretrain",
"Fine-tune",
"Method(year)",
"Backbone",
"Depth",
"Res",
"Single-Mod",
"C/P",
"UCF101",
"HMDB51",
"Frame",
"Dataset(duration)",
"Supervised",
"22",
"224",
"60.5",
"21.2",
"I3D",
"Random Init",
"X",
"V",
"122",
"",
"224",
"",
"67.1",
"28.5",
"I3D",
"ImageNet Supervised",
"ImageNet",
"！",
".",
"13D",
"96.8",
"74.5",
"-",
"K400Supervised",
"K400(28d)",
"224",
"-",
"Self-supervised",
"9",
"UCF101(1d)",
"112",
"P",
"50.2",
"18.1",
"AlexNet",
"Shuffle [34] [ECCV. 2016]",
"-",
"V",
"SS",
"R",
"52.1",
"22",
"VGAN [47] [NeurlPS, 2016]",
"UCF101(1d)",
"112",
"-",
"VGAN",
"14",
"UCF101(1d)",
"112",
"P",
"56.3",
"22.1",
"Caffe Net",
"OPN[28][ICCV.2017]",
"Geometry [12] [CVPR, 2018]",
"56",
"16",
"112",
"P",
"UCF101(1d)",
"FlowNet",
"55.1",
"23.3",
"UCF101(1d)",
"+",
"10",
"16",
"×",
"c",
"112",
"C3D",
"72.7",
"36.8",
"C",
"IIC[431[ACMMM,2020]",
"23",
"77.1",
"36.6",
"R(2+1)D",
"Pace [49] [ECCV, 2020]",
"K400(28d)",
"16",
"112",
"C3D",
"R",
"10",
"3D RotNet [23] [2018]",
"16",
"112",
"62.9",
"33.7",
"37.4(3.71)",
"K400(28d)",
"10",
"16",
"112",
"C3D",
"3DRotNet+BE",
"K400(28d)",
"P",
"65.4(2.5↑)",
"ST Puzzles [26] [AAAT, 2019]",
"C3D",
"10",
"48",
"112",
"p",
"28.3",
"60.6",
"UCF101(1d)",
"C3D",
"10",
"48",
"112",
"P",
"UCF101(1d)",
"ST Puzzles + BE",
"63.7(3.1个)",
"30.8(2.51)",
"Clip Order [57] [CVPR, 2019]",
"C3D",
"10",
"64",
"112",
"P",
"65.6",
"28.4",
"P",
"UCF101(1d)",
"68.5(2.91)",
"32.8(4.41)",
"C3D ",
"Clip Order+BE",
"10",
"UCF101(1d)",
"64",
"112",
"MoCo[2] (CVPR,2020]",
"C3D",
"10",
"16",
"112",
"60.5",
"27.2",
"C",
"UCF101(1d)",
"c",
"72.4(11.91)",
"UCF101(1d)",
"10",
"16",
"112",
"C3D",
"MoCo+BE",
"",
"42.3(14.1)",
"CoCLR[19] [NeulPS, 2020]",
"23",
"K400(28d)",
"32",
"128",
"87.9",
"54.6",
"x",
"R3D",
"DPC [17][ICCW, 2019]",
"34",
"64",
"224",
"O",
"R3D",
"K400(28d)",
"P",
"75.7",
"35.7",
"P",
"V",
"T-CAM",
"K400(28d)",
"64",
"224",
"79.4",
"V",
"AoT[541[CVPR,2018]",
"23",
"64",
"52.6",
"224",
"Pace [49] [ECCV, 2020]",
"V",
"c",
"K400(28d)",
"S3D-G",
"87.1",
"SpeedNet [1] [CVPR, 2020]",
"S3D-G",
"23",
"K400(28d)",
"64",
"224",
"P",
"81.1",
"48.8",
"I3D",
"V",
"22",
"43.7",
"64",
"SpeedNet [1] [CVPR, 2020]",
"224",
"P",
"66.7",
"",
"K400(28d)",
"MoCo[2]](CVPR,2020]",
"22",
"K400(28d)",
"224",
"16",
"13D",
"c",
"70.4",
"36.3",
"22",
"c",
"86.8(16.41)",
"55.4(19.11)",
"224",
"MoCo+BE",
"I3D",
"K400(28d)",
"16",
"22",
"UCF101(1d)",
"lcc",
"82.4",
"52.9",
"224",
"13D",
"S>S",
"16",
"MoCo+BE",
"224",
"53.7",
"34",
"R3D",
"UCF101(1d)",
"16",
"83.4",
"34",
"MoCo+BE",
"c",
"87.1",
"56.2",
"224",
"MoCo +BE",
"R3D",
"K400(28d)",
"16",
"Table 1: Top-1 accuracy (%) of integrating BE as a regularization term to four existing approaches and compared with previ-",
"ous methods on the UCF101 and HMDB51 dataset. Single-Mod denotes Single-Modality, C/P represents Contrastive/Pretext",
"task, means our implementation, K400 is short for Kinetics-400 and d represents day.",
"4.Experiments",
"original video with a temporal stride of 4. Then the sampled",
"clipisresizedto16×3×112×112or16×3×224×224.The",
"4.1. Implementation Details",
" of Background Erasing is experimentally set to 0.3, and",
"Datasets. All the experiments are conducted on four video",
"a larger value may result in excessive blur. The choice of",
"datasets, UCF101 [41], HMDB51 [27], Kinetics [25] and",
"temporal stride and is analysed in the supplementary. The",
"Diving48 [30]. The first three contain prominent bias, while",
"basicaugmentationsetAcontainsrandomrotationlessthan",
"Diving48 is less biased. UCF101 is a realistic video dataset",
"10 degrees and color jittering, and all these operations are",
"with 13,320 videos of 101 action categories.HMDB51",
"applied in a temporal consistent way, that is, each frame of",
"contains 6,849 clips of 51 action categories. Kinetics is a",
"a video uses the same augmentation. The vector dimension",
"large scale action recognition dataset that contains 246k/20k",
"D is 128.",
"train/val video clips of 400 classes. Diving48 consists of",
"18k trimmed video clips of 48 diving sequences",
"Supervised Fine-tuning Settings. After pre-training, we",
"Networks. We use C3D [45], R3D [20] and I3D [7] as base",
"transfer the weights of thebase encoderto two downstream",
"encoders followed by a spatio-temporal max pooling layer.",
"tasks, i.e., action recognition and video retrieval, with the",
"Default Settings. All the experiments are conducted on 8",
"last fully connected layer randomly initialized. We fine-tune",
"Tesla V100 GPUs with a batch size of 64 under PyTorch[37]",
"the network for 45 epochs.The learning rate is initialized",
"framework. We adopt SGD as our optimizer with momen-",
"as 0.05 and decreases to 1/10 every 10 epochs.",
"tum of 0.9 and weight decay of 5e-4.",
"Self-supervised Pre-training Settings. We pre-train the",
"Evaluation Settings.",
"For action recognition,following",
"network for 50 epochs with the learning rate initialized as",
"common practice[57], the final result of a video is the av-",
"0.01 and decreased to 1/10 every 10 epochs. The input clip",
"erage of the results of 10 clips that are uniformly sampled",
"consists of 16 frames, which is uniformly sampled from the",
"from it during testing time.",
"5"
] |
{
"arxiv_id": "2009.05769",
"language": "en",
"page_idx": 4
}
|
||
2404.09709-en-0
|
2404.09709-en-0.jpg
|
$ন্ল$
$\textbf{Scenario- Adaptive Fine- Grained Personalization Network: }$
Tailoring User Behavior Representation to the Scenario Context
Moyu Zhang Lazada Group
Beijing, Beijing, China zhangmoyu@bupt.cn
Yongxiang Tang
$\bar{\mathrm{Unaffiliated}}$
Beijing, Beijing, China
tangyongxiang94@gmail.com
$\operatorname{Jinxin}\text{Hu}^*$ Lazada Group
Beijing, Beijing, China jinxin. hjx@ lazada. com
Yu Zhang Lazada Group
Beijing, Beijing, China
daoji@lazada.com
# $\mathbf{ABSTRACT}$
# $\mathbf{KEYWORDS}$
Multi-Scenario Recommendation, Scenario-Specific Behavior Rep-
resentation, Recommender System
ACM Reference Format:
Moyu Zhang, Yongxiang Tang, Jinxin Hu, and Yu Zhang. 2018. ScenarioAdaptive Fine-Grained Personalization Network: Tailoring User Behavior Representation to the Scenario Context. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), fuly 14-18, 2024, Washington D.C., USA. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/1122445.1122456
# 1 INTRODUCTION
As e- commerce has evolved, modern large- scale commercial plat- forms now accommodate various scenarios to cater to the diverse shopping preferences of users. To conserve resources, current multi-
scenario methods often utilize a unified framework to deliver personalized recommendations across various scenarios. Given the overlap of users and items in multiple scenarios on commercial platforms, current methods typically employ shared bottom representations, capturing similarities and differences between scenarios through adaptive adjustments. However, existing methods often adjust representations adaptively only after aggregating user behavior sequences. This coarse-grained approach to re-weighting the entire user sequence hampers the model's ability to accurately model the user interest migration across different scenarios. To enhance the model's capacity to capture user interests from historical behavior sequences in each scenario, we develop a ranking framework named the Scenario-Adaptive Fine-Grained Personalization Network (SFPNet), which designs a kind of fine-grained method for multi-scenario personalized recommendations. Specifically, SFPNet comprises a series of blocks named as Scenario-Tailoring Block, stacked sequentially. Each block initially deploys a parameter personalization unit to integrate scenario information at a coarsegrained level by redefining fundamental features. Subsequently, we consolidate scenario-adaptively adjusted feature representations to serve as context information. By employing residual connection, we incorporate this context into the representation of each historical behavior, allowing for context-aware fine-grained customization of the behavior representations at the scenario-level, which in turn supports scenario-aware user interest modeling. Ultimately, the effectiveness of our proposed method is strongly substantiated by extensive experiments and online A/B testing.
In recent years, with the rise of electronic shopping as a dominant trend, a multitude of e-commerce platforms have emerged, including Amazon, Alibaba's Taobao, Lazada, TikTok, and others [3,12,18,26,27,30]. To cater to the growing diversity of user shopping preferences, contemporary commercial platforms must offer personalized services tailored to a variety of shopping contexts. For instance, the online e-commerce platform depicted in Figure 1 offers multiple item recommendation scenarios. The "Homepage" scenario primarily suggests a broad range of products potentially appealing to the user, based on their past behavior, whereas the "My Cart" and "My Orders" scenarios depend more on the user's historical purchase records to generate recommendations. Compared to the "Homepage" scenario, user interests in the "My Cart" and My Orders" scenarios become more specific. Consequently, these multi-scenario platforms present a challenge to existing recommendation models: how to devise personalized recommendations that consider the distinct characteristics of each scenario?
Traditional methods typically employ scenario-specific data to train distinct models for each scenario, enhancing scenario differentiation for online services. However, this isolated model training approach is resource-intensive, often inefficient, and fails to capture inter-scenario correlations, leading to suboptimal performance in data-sparse scenarios. In practice, most current e-commerce platforms experience significant user and item overlap across various scenarios. Consequently, despite the differences in user and item performance across scenarios, substantial correlations persist. Leveraging the correlations across scenarios, we can incorporate data from other scenarios into the training process, alleviating the challenge of data sparsity in individual scenarios. To model
$^{*}$Corresponding Author
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.
Woodstock'18, fune 03-05, 2018, Woodstock, NY $\odot$ 2018 Association for Computing Machinery. ACM ISBN 978-1-4503-XXXX-X/18/06...$15.00 https://doi.org/10.1145/1122445.1122456
|
[
[
0.1298828125,
0.10479736328125,
0.87158203125,
0.10479736328125,
0.87158203125,
0.12249755859375,
0.1298828125,
0.12249755859375
],
[
0.0906982421875,
0.1287841796875,
0.91015625,
0.1287841796875,
0.91015625,
0.149658203125,
0.0906982421875,
0.149658203125
],
[
0.2442626953125,
0.160400390625,
0.3544921875,
0.1634521484375,
0.35302734375,
0.1817626953125,
0.24267578125,
0.1787109375
],
[
0.6337890625,
0.1634521484375,
0.7685546875,
0.1634521484375,
0.7685546875,
0.18115234375,
0.6337890625,
0.18115234375
],
[
0.2467041015625,
0.1767578125,
0.348876953125,
0.1787109375,
0.348876953125,
0.1962890625,
0.2467041015625,
0.1944580078125
],
[
0.66015625,
0.179931640625,
0.74267578125,
0.179931640625,
0.74267578125,
0.1944580078125,
0.66015625,
0.1944580078125
],
[
0.22216796875,
0.1956787109375,
0.374267578125,
0.1956787109375,
0.374267578125,
0.210205078125,
0.22216796875,
0.210205078125
],
[
0.6259765625,
0.1956787109375,
0.77685546875,
0.1956787109375,
0.77685546875,
0.210205078125,
0.6259765625,
0.210205078125
],
[
0.226318359375,
0.2108154296875,
0.373291015625,
0.2108154296875,
0.373291015625,
0.2247314453125,
0.226318359375,
0.2247314453125
],
[
0.5986328125,
0.210205078125,
0.80224609375,
0.208984375,
0.80224609375,
0.2235107421875,
0.5986328125,
0.2247314453125
],
[
0.253173828125,
0.236083984375,
0.340576171875,
0.236083984375,
0.340576171875,
0.25390625,
0.253173828125,
0.25390625
],
[
0.65673828125,
0.2342529296875,
0.74267578125,
0.236083984375,
0.74267578125,
0.255126953125,
0.65673828125,
0.25244140625
],
[
0.250732421875,
0.25390625,
0.346435546875,
0.25390625,
0.346435546875,
0.268310546875,
0.250732421875,
0.268310546875
],
[
0.650390625,
0.251953125,
0.74853515625,
0.25390625,
0.74853515625,
0.268310546875,
0.650390625,
0.266357421875
],
[
0.0187835693359375,
0.268310546875,
0.062103271484375,
0.267578125,
0.066162109375,
0.3818359375,
0.0228729248046875,
0.38330078125
],
[
0.22216796875,
0.26953125,
0.373291015625,
0.26953125,
0.373291015625,
0.28271484375,
0.22216796875,
0.28271484375
],
[
0.625,
0.268310546875,
0.77685546875,
0.268310546875,
0.77685546875,
0.2822265625,
0.625,
0.2822265625
],
[
0.221435546875,
0.2841796875,
0.375732421875,
0.2841796875,
0.375732421875,
0.29736328125,
0.221435546875,
0.29736328125
],
[
0.63623046875,
0.28271484375,
0.76318359375,
0.2841796875,
0.76318359375,
0.298583984375,
0.63623046875,
0.29736328125
],
[
0.0882568359375,
0.31005859375,
0.1829833984375,
0.31005859375,
0.1829833984375,
0.3212890625,
0.0882568359375,
0.3212890625
],
[
0.515625,
0.306884765625,
0.62158203125,
0.30810546875,
0.62158203125,
0.322509765625,
0.515625,
0.3212890625
],
[
0.0849609375,
0.326904296875,
0.48193359375,
0.326904296875,
0.48193359375,
0.34033203125,
0.0849609375,
0.34033203125
],
[
0.515625,
0.326904296875,
0.91259765625,
0.326904296875,
0.91259765625,
0.34033203125,
0.515625,
0.34033203125
],
[
0.08660888671875,
0.341552734375,
0.4794921875,
0.341552734375,
0.4794921875,
0.3515625,
0.08660888671875,
0.3515625
],
[
0.51708984375,
0.341552734375,
0.7294921875,
0.341552734375,
0.7294921875,
0.354736328125,
0.51708984375,
0.354736328125
],
[
0.08660888671875,
0.354736328125,
0.484375,
0.354736328125,
0.484375,
0.36865234375,
0.08660888671875,
0.36865234375
],
[
0.02532958984375,
0.37060546875,
0.058013916015625,
0.37060546875,
0.058013916015625,
0.4072265625,
0.02532958984375,
0.4072265625
],
[
0.0882568359375,
0.37060546875,
0.48046875,
0.37060546875,
0.48046875,
0.380615234375,
0.0882568359375,
0.380615234375
],
[
0.51708984375,
0.365478515625,
0.6640625,
0.365478515625,
0.6640625,
0.37890625,
0.51708984375,
0.37890625
],
[
0.5185546875,
0.37744140625,
0.91259765625,
0.37744140625,
0.91259765625,
0.38818359375,
0.5185546875,
0.38818359375
],
[
0.08660888671875,
0.3837890625,
0.48046875,
0.3837890625,
0.48046875,
0.39453125,
0.08660888671875,
0.39453125
],
[
0.515625,
0.39013671875,
0.9140625,
0.39013671875,
0.9140625,
0.404052734375,
0.515625,
0.404052734375
],
[
0.08660888671875,
0.396484375,
0.483642578125,
0.396484375,
0.483642578125,
0.410400390625,
0.08660888671875,
0.410400390625
],
[
0.515625,
0.40283203125,
0.9140625,
0.40283203125,
0.9140625,
0.416748046875,
0.515625,
0.416748046875
],
[
0.0228729248046875,
0.41357421875,
0.058013916015625,
0.412353515625,
0.060455322265625,
0.484130859375,
0.02532958984375,
0.48486328125
],
[
0.08660888671875,
0.412353515625,
0.48193359375,
0.412353515625,
0.48193359375,
0.42236328125,
0.08660888671875,
0.42236328125
],
[
0.51708984375,
0.415283203125,
0.9140625,
0.415283203125,
0.9140625,
0.4287109375,
0.51708984375,
0.4287109375
],
[
0.0849609375,
0.4248046875,
0.48193359375,
0.4248046875,
0.48193359375,
0.438232421875,
0.0849609375,
0.438232421875
],
[
0.5185546875,
0.427978515625,
0.9140625,
0.427978515625,
0.9140625,
0.44140625,
0.5185546875,
0.44140625
],
[
0.0849609375,
0.438232421875,
0.48193359375,
0.438232421875,
0.48193359375,
0.451904296875,
0.0849609375,
0.451904296875
],
[
0.51708984375,
0.44140625,
0.83642578125,
0.44140625,
0.83642578125,
0.455078125,
0.51708984375,
0.455078125
],
[
0.08660888671875,
0.453857421875,
0.4794921875,
0.453857421875,
0.4794921875,
0.464111328125,
0.08660888671875,
0.464111328125
],
[
0.0849609375,
0.466552734375,
0.48193359375,
0.466552734375,
0.48193359375,
0.479736328125,
0.0849609375,
0.479736328125
],
[
0.08660888671875,
0.48095703125,
0.4794921875,
0.48095703125,
0.4794921875,
0.4912109375,
0.08660888671875,
0.4912109375
],
[
0.5390625,
0.4765625,
0.68896484375,
0.4765625,
0.68896484375,
0.490478515625,
0.5390625,
0.490478515625
],
[
0.0228729248046875,
0.4912109375,
0.060455322265625,
0.4912109375,
0.060455322265625,
0.5947265625,
0.0228729248046875,
0.5947265625
],
[
0.0849609375,
0.49365234375,
0.483642578125,
0.49365234375,
0.483642578125,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.51708984375,
0.49560546875,
0.9140625,
0.49560546875,
0.9140625,
0.5087890625,
0.51708984375,
0.5087890625
],
[
0.0849609375,
0.5068359375,
0.483642578125,
0.5068359375,
0.483642578125,
0.521484375,
0.0849609375,
0.521484375
],
[
0.51708984375,
0.51025390625,
0.9140625,
0.51025390625,
0.9140625,
0.5234375,
0.51708984375,
0.5234375
],
[
0.08660888671875,
0.52294921875,
0.48046875,
0.52294921875,
0.48046875,
0.53271484375,
0.08660888671875,
0.53271484375
],
[
0.51708984375,
0.5234375,
0.9140625,
0.5234375,
0.9140625,
0.537109375,
0.51708984375,
0.537109375
],
[
0.0849609375,
0.5341796875,
0.48046875,
0.5341796875,
0.48046875,
0.54736328125,
0.0849609375,
0.54736328125
],
[
0.5185546875,
0.537109375,
0.9140625,
0.537109375,
0.9140625,
0.55029296875,
0.5185546875,
0.55029296875
],
[
0.0849609375,
0.548828125,
0.48193359375,
0.548828125,
0.48193359375,
0.5625,
0.0849609375,
0.5625
],
[
0.51708984375,
0.5517578125,
0.91259765625,
0.5517578125,
0.91259765625,
0.56494140625,
0.51708984375,
0.56494140625
],
[
0.08660888671875,
0.564453125,
0.48046875,
0.564453125,
0.48046875,
0.57470703125,
0.08660888671875,
0.57470703125
],
[
0.51708984375,
0.56494140625,
0.9140625,
0.56494140625,
0.9140625,
0.5791015625,
0.51708984375,
0.5791015625
],
[
0.0849609375,
0.5771484375,
0.483642578125,
0.5771484375,
0.483642578125,
0.59033203125,
0.0849609375,
0.59033203125
],
[
0.51708984375,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.51708984375,
0.59228515625
],
[
0.0196075439453125,
0.58544921875,
0.058013916015625,
0.58544921875,
0.060455322265625,
0.71337890625,
0.0228729248046875,
0.71337890625
],
[
0.041656494140625,
0.5859375,
0.0482177734375,
0.5859375,
0.0482177734375,
0.59130859375,
0.041656494140625,
0.59130859375
],
[
0.0849609375,
0.59033203125,
0.483642578125,
0.59033203125,
0.483642578125,
0.60400390625,
0.0849609375,
0.60400390625
],
[
0.51708984375,
0.59130859375,
0.91259765625,
0.59130859375,
0.91259765625,
0.60498046875,
0.51708984375,
0.60498046875
],
[
0.08660888671875,
0.60595703125,
0.48046875,
0.60595703125,
0.48046875,
0.6162109375,
0.08660888671875,
0.6162109375
],
[
0.51708984375,
0.60693359375,
0.91259765625,
0.60693359375,
0.91259765625,
0.62060546875,
0.51708984375,
0.62060546875
],
[
0.08660888671875,
0.6171875,
0.48046875,
0.6171875,
0.48046875,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.51708984375,
0.62060546875,
0.91259765625,
0.62060546875,
0.91259765625,
0.6337890625,
0.51708984375,
0.6337890625
],
[
0.08660888671875,
0.63330078125,
0.4794921875,
0.63330078125,
0.4794921875,
0.64306640625,
0.08660888671875,
0.64306640625
],
[
0.51708984375,
0.6337890625,
0.9140625,
0.6337890625,
0.9140625,
0.64794921875,
0.51708984375,
0.64794921875
],
[
0.0849609375,
0.64599609375,
0.48193359375,
0.64599609375,
0.48193359375,
0.6591796875,
0.0849609375,
0.6591796875
],
[
0.5185546875,
0.6484375,
0.91259765625,
0.6484375,
0.91259765625,
0.6591796875,
0.5185546875,
0.6591796875
],
[
0.0849609375,
0.66015625,
0.483642578125,
0.66015625,
0.483642578125,
0.673828125,
0.0849609375,
0.673828125
],
[
0.51708984375,
0.662109375,
0.9140625,
0.662109375,
0.9140625,
0.67529296875,
0.51708984375,
0.67529296875
],
[
0.0849609375,
0.67236328125,
0.48193359375,
0.67236328125,
0.48193359375,
0.68603515625,
0.0849609375,
0.68603515625
],
[
0.5185546875,
0.67529296875,
0.9140625,
0.67529296875,
0.9140625,
0.689453125,
0.5185546875,
0.689453125
],
[
0.08660888671875,
0.6875,
0.4794921875,
0.6875,
0.4794921875,
0.69775390625,
0.08660888671875,
0.69775390625
],
[
0.515625,
0.68798828125,
0.9140625,
0.689453125,
0.9140625,
0.7041015625,
0.515625,
0.70263671875
],
[
0.0849609375,
0.7021484375,
0.48193359375,
0.7021484375,
0.48193359375,
0.71533203125,
0.0849609375,
0.71533203125
],
[
0.51708984375,
0.7041015625,
0.9140625,
0.7041015625,
0.9140625,
0.71728515625,
0.51708984375,
0.71728515625
],
[
0.08660888671875,
0.71533203125,
0.48193359375,
0.71533203125,
0.48193359375,
0.72900390625,
0.08660888671875,
0.72900390625
],
[
0.51708984375,
0.71728515625,
0.8408203125,
0.71728515625,
0.8408203125,
0.73095703125,
0.51708984375,
0.73095703125
],
[
0.08416748046875,
0.7265625,
0.363525390625,
0.72900390625,
0.363525390625,
0.74365234375,
0.08416748046875,
0.7412109375
],
[
0.53369140625,
0.73095703125,
0.9140625,
0.73095703125,
0.9140625,
0.744140625,
0.53369140625,
0.744140625
],
[
0.5185546875,
0.7470703125,
0.9111328125,
0.7470703125,
0.9111328125,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.08660888671875,
0.7548828125,
0.1993408203125,
0.7548828125,
0.1993408203125,
0.76806640625,
0.08660888671875,
0.76806640625
],
[
0.515625,
0.75830078125,
0.9140625,
0.7587890625,
0.9140625,
0.77294921875,
0.515625,
0.771484375
],
[
0.51708984375,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.78515625
],
[
0.08660888671875,
0.7822265625,
0.48193359375,
0.7822265625,
0.48193359375,
0.79248046875,
0.08660888671875,
0.79248046875
],
[
0.515625,
0.78515625,
0.9140625,
0.7861328125,
0.9140625,
0.7998046875,
0.515625,
0.798828125
],
[
0.0882568359375,
0.79443359375,
0.134033203125,
0.79443359375,
0.134033203125,
0.8017578125,
0.0882568359375,
0.8017578125
],
[
0.1298828125,
0.79150390625,
0.48193359375,
0.79052734375,
0.48193359375,
0.8037109375,
0.1298828125,
0.8046875
],
[
0.0849609375,
0.80322265625,
0.48193359375,
0.80322265625,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.51708984375,
0.80029296875,
0.91259765625,
0.80029296875,
0.91259765625,
0.814453125,
0.51708984375,
0.814453125
],
[
0.08660888671875,
0.81298828125,
0.48193359375,
0.81298828125,
0.48193359375,
0.82373046875,
0.08660888671875,
0.82373046875
],
[
0.51708984375,
0.814453125,
0.91259765625,
0.814453125,
0.91259765625,
0.82763671875,
0.51708984375,
0.82763671875
],
[
0.08660888671875,
0.82373046875,
0.48193359375,
0.82373046875,
0.48193359375,
0.833984375,
0.08660888671875,
0.833984375
],
[
0.0849609375,
0.83203125,
0.48193359375,
0.83203125,
0.48193359375,
0.84521484375,
0.0849609375,
0.84521484375
],
[
0.51708984375,
0.82763671875,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.51708984375,
0.84130859375
],
[
0.0849609375,
0.84326171875,
0.334228515625,
0.84326171875,
0.334228515625,
0.85400390625,
0.0849609375,
0.85400390625
],
[
0.5185546875,
0.84326171875,
0.91259765625,
0.84326171875,
0.91259765625,
0.85400390625,
0.5185546875,
0.85400390625
],
[
0.0882568359375,
0.85400390625,
0.306396484375,
0.85400390625,
0.306396484375,
0.8642578125,
0.0882568359375,
0.8642578125
],
[
0.51708984375,
0.85595703125,
0.9140625,
0.85595703125,
0.9140625,
0.869140625,
0.51708984375,
0.869140625
],
[
0.08660888671875,
0.86572265625,
0.30224609375,
0.86572265625,
0.30224609375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.51708984375,
0.869140625,
0.9140625,
0.869140625,
0.9140625,
0.88330078125,
0.51708984375,
0.88330078125
],
[
0.08660888671875,
0.87548828125,
0.306396484375,
0.87548828125,
0.306396484375,
0.88623046875,
0.08660888671875,
0.88623046875
],
[
0.08416748046875,
0.88525390625,
0.27294921875,
0.8837890625,
0.27294921875,
0.89599609375,
0.08416748046875,
0.896484375
],
[
0.5185546875,
0.88525390625,
0.91259765625,
0.88525390625,
0.91259765625,
0.89599609375,
0.5185546875,
0.89599609375
]
] |
[
"Scenario-Adaptive Fine-Grained Personalization Network",
"Tailoring User Behavior Representation to the Scenario Context",
"Moyu Zhang",
"Yongxiang Tang",
" Lazada Group",
"Unaffiliated",
"Beijing, Beijing, China",
"Beijing, Beijing, China",
"zhangmoyu@bupt.cn",
"tangyongxiang94@gmail.com",
"Jinxin Hu*",
"Yu Zhang",
"Lazada Group",
"Lazada Group",
"",
"Beijing, Beijing, China",
"Beijing, Beijing, China",
"jinxin.hjx@lazada.com",
"daoji@lazada.com",
"ABSTRACT",
"KEYWORDS",
"As e-commerce has evolved, modern large-scale commercial plat-",
"Multi-Scenario Recommendation, Scenario-Specific Behavior Rep-",
"forms nowaccommodatevariousscenariosto cater tothediverse",
"resentation, Recommender System",
"shopping preferences of users. To conserve resources, current multi-",
"3",
"scenariomethodsoftenutilizea unifiedframeworkto deliverper-",
"ACMReferenceFormat:",
"Moyu Zhang,Yongxiang Tang,Jinxin Hu, and Yu Zhang.2018. Scenario-",
"sonalized recommendations across various scenarios.Given the",
"Adaptive Fine-Grained Personalization Network: Tailoring User Behavior",
"overlap of users and items in multiple scenarios on commercial",
"Representation to the Scenario Context In Proceedings of the 47th Interna-",
"['s]",
"platforms,currentmethodstypicallyemploysharedbottomrepre-",
"tional ACM SIGIR Conference on Research and Development in Information",
"sentations, capturing similarities and differences between scenarios",
"Retrieval (SIGIR), July 14-18, 2024, Washington D.C., USA. ACM, New York,",
"through adaptive adjustments. However, existing methods often",
"NY, USA, 11 pages. https://doi.org/10.1145/1122445.1122456",
"adjust representations adaptively onlyafter aggregating userbe",
"havior sequences. This coarse-grained approach to re-weighting",
"the entire user sequence hampers the model'sability to accurately",
"INTRODUCTION",
"A60260",
"model the user interest migration across different scenarios. To en-",
"In recent years, with the rise of electronic shopping as a domi-",
"hance the model's capacity to capture user interests from historical",
"nant trend, a multitude of e-commerce platforms have emerged,",
"behavior sequencesineach scenario,wedeveloparankingframe-",
"including Amazon, Alibaba's Taobao, Lazada, TikTok, and others",
"work named the Scenario-Adaptive Fine-Grained Personalization",
"[3, 12, 18, 26, 27, 30]. To cater to the growing diversity of user shop-",
"Network (SFPNet), which designs a kind of fine-grained method",
"ping preferences,contemporary commercial platforms must offer",
"formulti-scenariopersonalizedrecommendations.Specifically,SF",
"personalized services tailored to a variety of shopping contexts",
"PNet comprises a series of blocks named as Scenario-Tailoring Block,",
"For instance, the online e-commerce platform depicted in Figure 1",
"",
"",
"stacked sequentially. Each block initially deploys a parameter per-",
"offers multiple item recommendation scenarios. The\"Homepage\"",
"sonalization unit to integrate scenario information at a coarse-",
"scenario primarily suggests a broad range of products potentially",
"grained level by redefining fundamental features. Subsequently, we",
"appealing to the user, based on their past behavior, whereas the\"My",
"consolidatescenario-adaptivelyadjusted featurerepresentations to",
"Cart\" and \"My Orders\" scenarios depend more on the user's his-",
"serve as context information. By employing residual connection, we",
"torical purchaserecordstogeneraterecommendations.Compared",
"incorporate this context into the representation of each historical",
"to the\"Homepage\" scenario, user interests in the\"My Cart\"and",
"behavior, allowing for context-aware fine-grained customization of",
"\"My Orders\" scenarios become more specific. Consequently, these",
"the behavior representations atthe scenario-level, which inturn",
"multi-scenario platforms present a challenge to existing recommen-",
"supports scenario-aware user interest modeling. Ultimately, the",
"dation models: how to devise personalized recommendations that",
"effectiveness of our proposed method is strongly substantiated by",
"consider the distinct characteristics of each scenario?",
"extensive experiments and online A/B testing.",
"Traditional methods typically employ scenario-specific data to",
"train distinct modelsfor each scenario, enhancing scenariodiffer",
"Corresponding Author",
"entiation for online services.However, this isolated model training",
"approach is resource-intensive, often inefficient, and fails to cap-",
"Permission to make digital or hard copies of all or part of this work for personal or",
"ture inter-scenario correlations, leading to suboptimal performance",
"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",
"in data-sparse scenarios. In practice, most current e-commerce",
"on the first page. Copyrights for components of this work owned by others than ACM",
"platforms experience significant user and item overlap across var",
"must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,",
"to post on servers or toredistribute to lists, requires prior specific permission and/ora",
"ious scenarios. Consequently, despite the differences in user and",
"fee. Request permissions from permissions@acm.org",
"item performance across scenarios, substantial correlations persist.",
"Woodstock'18,June03-05,2018,Woodstock,NY",
"Leveraging the correlations across scenarios, we can incorporate",
"2018Association for Computing Machinery",
"data from other scenarios into the training process, alleviating",
"ACM ISBN978-1-4503-XXXX-X/18/06...$15.00",
"https://doi.org/10.1145/1122445.1122456",
"the challenge of data sparsity in individual scenarios.To model"
] |
{
"arxiv_id": "2404.09709",
"language": "en",
"page_idx": 0
}
|
||
2404.09709-en-2
|
2404.09709-en-2.jpg
|
Woodstock'18, June 03-05, 2018, Woodstock, NY
Scenario-Adaptive Fine-Grained Personalization Network: Tailoring User Behavior Representation to the Scenario Context
generate high-order interactions, subsequently tailoring the behavior representation of the sequence via residual computations. The integration of the above two modules allows SFPNet to incorporate scenario-related information into the feature representation vectors with high granularity, markedly enhancing the model's proficiency in discerning commonalities and distinctions across scenarios.
The contributions of our paper can be summarized as follows: $\bullet$ To our knowledge, SFPNet is the first multi-scenario work to offer fine-grained scenario-specific customization for individual user behaviors, thereby advancing the model's capacity to track shifts in user interests across various scenarios.
$\bullet$ SFPNet features a novel block comprising two key modules, which capitalizes on the context and scenario-specific information of each input instance to tailor the representation vectors of user behaviors in the sequence and associated context features of the input instance. By stacking the block, SFPNet constructs a deep network that progressively bolsters its ability to model the intricate interplay between scenarios and input instances.
$\bullet$ Evaluations using offline datasets and online A/B testing have demonstrated the superiority of the proposed SFPNet method over existing state-of-the-art multi-scenario approaches.
of feature extraction. M2M [31] utilizes expert networks to solve multi-scenario and multi-task problems based on a meta network to express the scenario information explicitly. AESM$^2\left[37\right]$proposes a novel expert network structure with automatic selection of fine granularity by calculating the KL divergence to select the most suitable sharing and exclusive experts. Beyond the MoE-based method, there are alternative approaches like STAR [22], which introduces an extra tower for each scenario, combining the parameters of scenario-specific tower with those of the shared tower's parameter.
However, scenario-specific approaches learn scenario information transfer in quite implicit ways and often overlook differences in underlying representations, which frequently constitute a bottleneck in recommendation models. Therefore, SASS[34] designs a scenario adaptive transfer module to select and fuse effective transfer information from whole scenario to individual scenario. Adasparse [29] learns adaptively sparse structure for each scenario, achieving better generalization across domains with lower computational cost. HiNet [36] achieves hierarchical extraction based on coarse-to-fine knowledge transfer scheme. DFFM [6] incorporates scenario-related information into the parameters of feature interaction and user behavior modules, allowing scenario-specific learning of these two aspects, which is essentially a scenario-adaptive parameterization method. 3MN [33] proposes a novel three meta networks-based solution to model the complicated task-task, scenario-scenario, and task-scenario interrelations. PEPNet [4] takes features with strong biases as input and dynamically scales the bottom-layer embeddings and the top-layer DNN hidden units in the model through a gate mechanism. MARIA [24] proposes to project the scenario semantic information into the bottom feature representation to derive more discriminative feature representations.
# 2 RELATED WORK
## 3 РROВLEМ FORMULATION $\begin{array}{l}\text{Given a set of scenarios }S=\left\{s_{m}\right\}_{m=1}^{N_{s}}\text{with a shared feature space}\\X\text{ and label space }y,\text{the multi-scenario recommendation task aims}\end{array}$
to devise a unified ranking formula $\mathcal{F}:X\to\mathcal{Y}$, to concurrently provide accurate, personalized recommendations across $M$ scenarios. The common feature space typically encompasses a variety of features that can be categorized into distinct fields, such as usercentric and item-centric features, which collectively represent an instance's comprehensive contextual information. Within this paper, the feature space $X$ is constructed as $\left[x_1,x_2,...,x_{N_f},\left\{b_i\right\}_{i=0}^{{\hat{N}_b}}\right]$, $\begin{array}{ll}\mathrm{where~for~clarity~in~subsequent~discussions,~\{b_{i}\}_{i=0}^{N_{b}}~denote~the}\\\mathrm{user's~historical~behavior~sequence~to~distinguish~it~from~other~}N_{f}\\\mathrm{features,~and~}N_{b}~\mathrm{represents~the~number~of~the~user's~historical~be-}\end{array}$ features, and $N_b$ represents the number of the user's historical behaviors. Mathematically, the multi-scenario recommendation task involves estimating the probability that the target user will interact with the target item in a given scenario s$_m$,utilizing the contextual features, as illustrated below:
With the evolution of e-commerce platforms and growing user engagement, a single scenario is insufficient to satisfy the expanding needs of users [9, 17, 32]. Consequently, modern online e-commerce platforms are required to cater to multiple scenarios simultaneously. Initially, recommendation systems typically trained a distinct model for each scenario to cater to its specific demands. However, with the proliferation of platform scenarios, the drawbacks of this independent training approach have become increasingly evident. This method is not only resource-intensive but also fails to leverage the correlations between scenarios, leading to suboptimal model performance. Recent advances in deep learning and the industry's emphasis on replicability- avoiding the integration of new elements with each additional scenario-have led to a preference for a unified model structure, employing a singular ranking framework for scenario-specific recommendations. Single model structures are primarily categorized into two approaches: 1) Scenario-specific network structures, and 2) Scenario-adaptive parameter network structures. Though both strive to train a unified model capable of serving multiple scenarios, their underlying philosophies differ.
Specifically, scenario-specific network structures draw inspiration from multi-task learning (MTL) principles. These approaches seek to harness the similarities between multi-task and multi-scenario applications by considering each scenario as an independent task. Mixture-of-Experts (MoE) [19] is proposed to significantly increase model capacity and capability and is widely used in the multiscenario area. MMoE [15] characterizes the task correlation and learns the function of specific tasks based on shared bottom feature representation. HMoE [13] takes advantage of MMoE to implicitly identify distinctions and commonalities between tasks, and improves the performance with a stacked model learning task relationships in the label space explicitly. Later, PLE [23] shares experts in the share layer and refines tasks uniquely to effectively alleviate the noise caused by other scenarios and improves the effectiveness
$$\hat{y}=P(y=1|\mathcal{X},s_{m})=\mathcal{F}(x_{1},x_{2},...,x_{N_{f}},\{b_{i}\}_{i=0}^{N_{b}},s_{m})$$
(1)
where $x_i\in\mathbb{R}^d$ denotes the embedding of the i-th feature, and $d$ denotes the embedding dimension of features. $b_i\in\mathbb{R}^d$ denotes the representation of $i$-th behavior in the behavior sequence, typically derived from a pooled embedding that combines the clicked item embedding and the embeddings of its associated attribute features $s_m$ represents the representation vector of $m$-th scenario features.
|
[
[
0.08660888671875,
0.0770263671875,
0.66650390625,
0.0770263671875,
0.66650390625,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6689453125,
0.0770263671875,
0.91259765625,
0.0770263671875,
0.91259765625,
0.08709716796875,
0.6689453125,
0.08709716796875
],
[
0.0882568359375,
0.10919189453125,
0.48193359375,
0.10919189453125,
0.48193359375,
0.11932373046875,
0.0882568359375,
0.11932373046875
],
[
0.5185546875,
0.10919189453125,
0.9111328125,
0.10919189453125,
0.9111328125,
0.11932373046875,
0.5185546875,
0.11932373046875
],
[
0.0849609375,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.0849609375,
0.1351318359375
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.0849609375,
0.1351318359375,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.0849609375,
0.1483154296875
],
[
0.51708984375,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.51708984375,
0.1483154296875
],
[
0.08660888671875,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.08660888671875,
0.1634521484375
],
[
0.51708984375,
0.149658203125,
0.91259765625,
0.149658203125,
0.91259765625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.0849609375,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.0849609375,
0.1767578125
],
[
0.51708984375,
0.1634521484375,
0.9140625,
0.1634521484375,
0.9140625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.0849609375,
0.1767578125,
0.4599609375,
0.1767578125,
0.4599609375,
0.1900634765625,
0.0849609375,
0.1900634765625
],
[
0.51708984375,
0.1767578125,
0.9140625,
0.1767578125,
0.9140625,
0.1900634765625,
0.51708984375,
0.1900634765625
],
[
0.10296630859375,
0.1912841796875,
0.4755859375,
0.1912841796875,
0.4755859375,
0.2052001953125,
0.10296630859375,
0.2052001953125
],
[
0.51708984375,
0.1912841796875,
0.9140625,
0.1912841796875,
0.9140625,
0.2052001953125,
0.51708984375,
0.2052001953125
],
[
0.08905029296875,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.08905029296875,
0.2183837890625
],
[
0.51708984375,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51708984375,
0.2183837890625
],
[
0.10296630859375,
0.2183837890625,
0.48193359375,
0.2183837890625,
0.48193359375,
0.231689453125,
0.10296630859375,
0.231689453125
],
[
0.51708984375,
0.2183837890625,
0.9140625,
0.2183837890625,
0.9140625,
0.231689453125,
0.51708984375,
0.231689453125
],
[
0.101318359375,
0.23291015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.246826171875,
0.101318359375,
0.246826171875
],
[
0.53173828125,
0.23291015625,
0.9140625,
0.23291015625,
0.9140625,
0.246826171875,
0.53173828125,
0.246826171875
],
[
0.10455322265625,
0.248779296875,
0.386474609375,
0.248779296875,
0.386474609375,
0.255615234375,
0.10455322265625,
0.255615234375
],
[
0.5185546875,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.08905029296875,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.08905029296875,
0.2734375
],
[
0.51708984375,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.10296630859375,
0.27587890625,
0.4794921875,
0.27587890625,
0.4794921875,
0.285888671875,
0.10296630859375,
0.285888671875
],
[
0.5185546875,
0.27587890625,
0.91259765625,
0.27587890625,
0.91259765625,
0.285888671875,
0.5185546875,
0.285888671875
],
[
0.10296630859375,
0.2890625,
0.48193359375,
0.2890625,
0.48193359375,
0.2998046875,
0.10296630859375,
0.2998046875
],
[
0.51953125,
0.2890625,
0.9111328125,
0.2890625,
0.9111328125,
0.296630859375,
0.51953125,
0.296630859375
],
[
0.10296630859375,
0.300537109375,
0.48046875,
0.300537109375,
0.48046875,
0.314453125,
0.10296630859375,
0.314453125
],
[
0.5185546875,
0.3017578125,
0.9140625,
0.3017578125,
0.9140625,
0.31494140625,
0.5185546875,
0.31494140625
],
[
0.10455322265625,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.10455322265625,
0.32763671875
],
[
0.521484375,
0.317626953125,
0.91259765625,
0.317626953125,
0.91259765625,
0.32763671875,
0.521484375,
0.32763671875
],
[
0.10296630859375,
0.330810546875,
0.48046875,
0.330810546875,
0.48046875,
0.341552734375,
0.10296630859375,
0.341552734375
],
[
0.51953125,
0.330810546875,
0.91259765625,
0.330810546875,
0.91259765625,
0.341552734375,
0.51953125,
0.341552734375
],
[
0.1004638671875,
0.34228515625,
0.44775390625,
0.34228515625,
0.44775390625,
0.35595703125,
0.1004638671875,
0.35595703125
],
[
0.51708984375,
0.34228515625,
0.91259765625,
0.34228515625,
0.91259765625,
0.35595703125,
0.51708984375,
0.35595703125
],
[
0.0882568359375,
0.35595703125,
0.48046875,
0.356689453125,
0.48046875,
0.37060546875,
0.0882568359375,
0.369384765625
],
[
0.51708984375,
0.356689453125,
0.9140625,
0.356689453125,
0.9140625,
0.37060546875,
0.51708984375,
0.37060546875
],
[
0.10296630859375,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.38330078125,
0.10296630859375,
0.38330078125
],
[
0.5185546875,
0.37255859375,
0.91259765625,
0.37255859375,
0.91259765625,
0.38330078125,
0.5185546875,
0.38330078125
],
[
0.101318359375,
0.385009765625,
0.44287109375,
0.385009765625,
0.44287109375,
0.3984375,
0.101318359375,
0.3984375
],
[
0.51708984375,
0.385009765625,
0.9140625,
0.385009765625,
0.9140625,
0.3984375,
0.51708984375,
0.3984375
],
[
0.515625,
0.399658203125,
0.9140625,
0.3984375,
0.9140625,
0.412353515625,
0.515625,
0.41357421875
],
[
0.51708984375,
0.412353515625,
0.9140625,
0.412353515625,
0.9140625,
0.425537109375,
0.51708984375,
0.425537109375
],
[
0.0882568359375,
0.4248046875,
0.1094970703125,
0.4248046875,
0.1094970703125,
0.43310546875,
0.0882568359375,
0.43310546875
],
[
0.10455322265625,
0.421630859375,
0.25732421875,
0.421630859375,
0.25732421875,
0.43505859375,
0.10455322265625,
0.43505859375
],
[
0.5185546875,
0.427978515625,
0.91259765625,
0.427978515625,
0.91259765625,
0.438232421875,
0.5185546875,
0.438232421875
],
[
0.0849609375,
0.439453125,
0.48193359375,
0.439453125,
0.48193359375,
0.45263671875,
0.0849609375,
0.45263671875
],
[
0.51708984375,
0.439453125,
0.9140625,
0.43994140625,
0.9140625,
0.453857421875,
0.51708984375,
0.45263671875
],
[
0.08660888671875,
0.453857421875,
0.48193359375,
0.453857421875,
0.48193359375,
0.46728515625,
0.08660888671875,
0.46728515625
],
[
0.51953125,
0.455078125,
0.91259765625,
0.455078125,
0.91259765625,
0.46533203125,
0.51953125,
0.46533203125
],
[
0.08660888671875,
0.4697265625,
0.4794921875,
0.4697265625,
0.4794921875,
0.479736328125,
0.08660888671875,
0.479736328125
],
[
0.51708984375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51708984375,
0.48095703125
],
[
0.0882568359375,
0.48291015625,
0.48193359375,
0.48291015625,
0.48193359375,
0.49365234375,
0.0882568359375,
0.49365234375
],
[
0.51708984375,
0.481689453125,
0.91259765625,
0.481689453125,
0.91259765625,
0.49560546875,
0.51708984375,
0.49560546875
],
[
0.08660888671875,
0.49560546875,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.08660888671875,
0.5087890625
],
[
0.5185546875,
0.496826171875,
0.91259765625,
0.496826171875,
0.91259765625,
0.5068359375,
0.5185546875,
0.5068359375
],
[
0.08660888671875,
0.51123046875,
0.48046875,
0.51123046875,
0.48046875,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.51708984375,
0.51123046875,
0.82666015625,
0.51123046875,
0.82666015625,
0.521484375,
0.51708984375,
0.521484375
],
[
0.08660888671875,
0.5244140625,
0.48193359375,
0.5244140625,
0.48193359375,
0.53515625,
0.08660888671875,
0.53515625
],
[
0.0849609375,
0.537109375,
0.483642578125,
0.537109375,
0.483642578125,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.51953125,
0.53857421875,
0.53515625,
0.53857421875,
0.53515625,
0.546875,
0.51953125,
0.546875
],
[
0.54150390625,
0.53515625,
0.76708984375,
0.53515625,
0.76708984375,
0.548828125,
0.54150390625,
0.548828125
],
[
0.08660888671875,
0.55322265625,
0.4794921875,
0.55322265625,
0.4794921875,
0.56298828125,
0.08660888671875,
0.56298828125
],
[
0.515625,
0.5517578125,
0.9140625,
0.5498046875,
0.9140625,
0.56640625,
0.515625,
0.568359375
],
[
0.08416748046875,
0.564453125,
0.48193359375,
0.56298828125,
0.48193359375,
0.5771484375,
0.08416748046875,
0.57763671875
],
[
0.5185546875,
0.5673828125,
0.91259765625,
0.5673828125,
0.91259765625,
0.58056640625,
0.5185546875,
0.58056640625
],
[
0.08660888671875,
0.5791015625,
0.4794921875,
0.5791015625,
0.4794921875,
0.5888671875,
0.08660888671875,
0.5888671875
],
[
0.51708984375,
0.58203125,
0.91259765625,
0.58203125,
0.91259765625,
0.59521484375,
0.51708984375,
0.59521484375
],
[
0.0849609375,
0.59228515625,
0.48193359375,
0.59228515625,
0.48193359375,
0.60595703125,
0.0849609375,
0.60595703125
],
[
0.5185546875,
0.59765625,
0.91015625,
0.59765625,
0.91015625,
0.60791015625,
0.5185546875,
0.60791015625
],
[
0.0849609375,
0.60595703125,
0.483642578125,
0.60595703125,
0.483642578125,
0.619140625,
0.0849609375,
0.619140625
],
[
0.5185546875,
0.611328125,
0.91259765625,
0.611328125,
0.91259765625,
0.6220703125,
0.5185546875,
0.6220703125
],
[
0.08660888671875,
0.6220703125,
0.48046875,
0.6220703125,
0.48046875,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.51708984375,
0.62353515625,
0.9140625,
0.62353515625,
0.9140625,
0.63720703125,
0.51708984375,
0.63720703125
],
[
0.08660888671875,
0.63525390625,
0.48046875,
0.63525390625,
0.48046875,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.5185546875,
0.6396484375,
0.91259765625,
0.6396484375,
0.91259765625,
0.6494140625,
0.5185546875,
0.6494140625
],
[
0.0849609375,
0.64794921875,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.5185546875,
0.65283203125,
0.9111328125,
0.65283203125,
0.9111328125,
0.66357421875,
0.5185546875,
0.66357421875
],
[
0.08660888671875,
0.66357421875,
0.48046875,
0.66357421875,
0.48046875,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.51708984375,
0.66748046875,
0.91259765625,
0.66748046875,
0.91259765625,
0.681640625,
0.51708984375,
0.681640625
],
[
0.08660888671875,
0.6767578125,
0.48193359375,
0.6767578125,
0.48193359375,
0.6875,
0.08660888671875,
0.6875
],
[
0.0849609375,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.0849609375,
0.7021484375
],
[
0.51708984375,
0.6845703125,
0.91259765625,
0.6845703125,
0.91259765625,
0.69873046875,
0.51708984375,
0.69873046875
],
[
0.5185546875,
0.69873046875,
0.9111328125,
0.69873046875,
0.9111328125,
0.7119140625,
0.5185546875,
0.7119140625
],
[
0.10296630859375,
0.705078125,
0.4794921875,
0.705078125,
0.4794921875,
0.71533203125,
0.10296630859375,
0.71533203125
],
[
0.0849609375,
0.71728515625,
0.48193359375,
0.71728515625,
0.48193359375,
0.73095703125,
0.0849609375,
0.73095703125
],
[
0.51708984375,
0.71337890625,
0.9111328125,
0.7119140625,
0.9111328125,
0.7236328125,
0.51708984375,
0.724609375
],
[
0.08660888671875,
0.73095703125,
0.490234375,
0.73095703125,
0.490234375,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.5185546875,
0.72802734375,
0.91259765625,
0.72802734375,
0.91259765625,
0.73779296875,
0.5185546875,
0.73779296875
],
[
0.0849609375,
0.744140625,
0.483642578125,
0.744140625,
0.483642578125,
0.75830078125,
0.0849609375,
0.75830078125
],
[
0.5185546875,
0.7421875,
0.91259765625,
0.7421875,
0.91259765625,
0.75244140625,
0.5185546875,
0.75244140625
],
[
0.0849609375,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.771484375,
0.0849609375,
0.771484375
],
[
0.51708984375,
0.75390625,
0.9140625,
0.75390625,
0.9140625,
0.767578125,
0.51708984375,
0.767578125
],
[
0.0849609375,
0.771484375,
0.48193359375,
0.771484375,
0.48193359375,
0.78515625,
0.0849609375,
0.78515625
],
[
0.51708984375,
0.767578125,
0.69775390625,
0.767578125,
0.69775390625,
0.78076171875,
0.51708984375,
0.78076171875
],
[
0.08660888671875,
0.78515625,
0.483642578125,
0.78515625,
0.483642578125,
0.798828125,
0.08660888671875,
0.798828125
],
[
0.55615234375,
0.7841796875,
0.87353515625,
0.7841796875,
0.87353515625,
0.80615234375,
0.55615234375,
0.80615234375
],
[
0.89453125,
0.7890625,
0.9140625,
0.7890625,
0.9140625,
0.8017578125,
0.89453125,
0.8017578125
],
[
0.0849609375,
0.7998046875,
0.48193359375,
0.7998046875,
0.48193359375,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.08660888671875,
0.8125,
0.48046875,
0.81298828125,
0.48046875,
0.8271484375,
0.08660888671875,
0.82568359375
],
[
0.5185546875,
0.8125,
0.91259765625,
0.8125,
0.91259765625,
0.82568359375,
0.5185546875,
0.82568359375
],
[
0.08660888671875,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.5185546875,
0.84033203125
],
[
0.0849609375,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51708984375,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.0849609375,
0.85498046875,
0.48193359375,
0.85498046875,
0.48193359375,
0.86865234375,
0.0849609375,
0.86865234375
],
[
0.51708984375,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0849609375,
0.86865234375,
0.48193359375,
0.86865234375,
0.48193359375,
0.8818359375,
0.0849609375,
0.8818359375
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.08416748046875,
0.8818359375,
0.483642578125,
0.88330078125,
0.483642578125,
0.896484375,
0.08416748046875,
0.89599609375
],
[
0.5146484375,
0.8837890625,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.5146484375,
0.89794921875
]
] |
[
"Scenario-Adaptive Fine-Grained Personalization Network: Tailoring User Behavior Representation to the Scenario Context",
"Woodstock18, June 03-05,2018, Woodstock,NY",
"generatehigh-orderinteractions,subsequentlytailoring thebehav-",
"offeatureextraction.M2M[31lutilizesexpertnetworksto solve",
"ior representation of the sequence via residual computations. The",
"multi-scenario and multi-task problems based on a meta network to",
"integration of the above two modules allows SFPNet to incorporate",
"express the scenario information explicitly. AESM? [37] proposes",
"scenario-related information into the feature representation vectors",
"a novel expert network structure with automatic selection of fine",
"with high granularity, markedly enhancing the model's proficiency",
"granularity by calculating the KL divergence to select the most suit-",
"in discerning commonalities and distinctions across scenarios.",
"able sharing and exclusive experts. Beyond the MoE-based method,",
"The contributions of our paper can be summarized as follows:",
"there are alternative approaches like STAR [22], which introduces",
". To our knowledge, SFPNet is the first multi-scenario work to",
"an extra tower for each scenario, combining the parameters of",
"offer fine-grained scenario-specific customization for individual",
"scenario-specific tower with those of the shared tower's parameter.",
"user behaviors, thereby advancing the model's capacity to track",
"However, scenario-specific approaches learn scenario informa-",
"shiftsinuserinterestsacrossvariousscenarios",
"tion transfer in quite implicit ways and often overlook differences",
" SFPNet features a novel block comprising two key modules,",
"in underlying representations, which frequently constitute a bottle-",
"whichcapitalizesonthecontextandscenario-specificinforma",
"neckinrecommendationmodels.Therefore,SASS[34]designsasce-",
"tion ofeach inputinstancetotailorthe representation vectors of",
"narioadantivetransfermoduletoselectandfuseeffectivetransfer",
"user behaviors in the sequence and associated context features",
"information from whole scenario to individual scenario. Adasparse",
"oftheinput instance.By stackingtheblock,SFPNet constructsa",
"[29]learnsadaptivelysparsestructureforeachscenario,achieving",
"deepnetworkthatprogressivelybolsters its abilitytomodel the",
"bettergeneralizationacrossdomainswithlowercomputational cost.",
"intricate interplay between scenarios and input instances",
"HiNet [36] achieves hierarchical extraction based on coarse-to-fine",
".Evaluations using offline datasets and online A/B testing have",
"knowledge transfer scheme. DFFM [6] incorporates scenario-related",
"demonstrated thesuperiority of theproposed SFPNet method",
"information into theparameters of feature interaction and user be-",
"over existing state-of-the-art multi-scenario approaches.",
"havior modules, allowing scenario-specific learning of these two",
"aspects, which is essentially a scenario-adaptive parameterization",
"method. 3MN [33] proposes a novel three meta networks-based",
"2",
"RELATEDWORK",
"solutiontomodelthecomplicatedtask-task,scenario-scenario,and",
"With the evolution of e-commerce platforms and growing user en-",
"task-scenario interrelations. PEPNet [4] takes features with strong",
"gagement, a single scenario is insufficient to satisfy the expanding",
"biasesas inputand dynamically scalesthebottom-layer embed",
"needsofusers[9,17,32].Consequently,modernonlinee-commerce",
"dings and the top-layer DNN hidden units in the model through",
"platforms arerequired tocatertomultiple scenarios simultaneously",
"a gate mechanism. MARIA [24] proposes to project the scenario",
"Initially,recommendation systems typically trained a distinct model",
"semanticinformationintothebottomfeaturerepresentationto",
"for each scenario to caterto its specific demands.However,with",
"derivemorediscriminativefeaturerepresentations.",
"theproliferation of platform scenarios, the drawbacks of this in-",
"dependent training approach have become increasingly evident.",
"3",
"PROBLEMFORMULATION",
"Thismethod is not only resource-intensivebut alsofails to leverage",
"Given a set of scenarios S = (sm)N-, with a shared feature space",
"the correlations between scenarios,leading to suboptimal model",
"X and label space Y, the multi-scenario recommendation task aims",
"performance.Recent advances in deep learning and the industry's",
"to devise a unified ranking formula F : X - y, to concurrently",
"emphasis on replicability-avoiding the integration of new elements",
"provideaccurate,personalized recommendations across M scenar",
"with each additional scenario-have led to a preference for a uni-",
"ios.The common feature spacetypically encompasses a varietyof",
"fiedmodelstructure,employingasingularrankingframeworkfor",
"features that can be categorized into distinct fields, such as user-",
"scenario-specific recommendations.Single model structures are",
"centric anditem-centricfeatures,whichcollectivelyrepresentan",
"primarily categorized into two approaches: 1) Scenario-specific",
"instance'scomprehensivecontextual information.Withinthispa",
"networkstructures,and2)Scenario-adaptiveparameternetwork",
"per, the feature space X is constructed as [xi, x2.., xNf. (bib]",
"structures.Though both strive to train a unified model capable of",
"serving multiple scenarios, their underlying philosophies differ.",
"where for clarity in subsequent discussions, (bib denote the",
"user's historical behavior sequence to distinguish it from other Ng",
"Specifically,scenario-specificnetworkstructuresdrawinspira",
"tion from multi-task learning (MTL) principles. These approaches",
"features,and Nrepresents thenumber of theuser's historicalbe",
"seek to harness the similarities between multi-task and multi-scenario",
"haviors.Mathematically,themulti-scenariorecommendationtask",
"applications by considering each scenario as an independent task.",
"involves estimating theprobability thatthetarget user will interact",
"Mixture-of-Experts (MoE) [19] is proposed to significantly increase",
"with the target item in a given scenario sm, utilizing the contextual",
"model capacity and capability and is widely used in the multi-",
"features, as illustrated below:",
"scenario area. MMoE [15] characterizes the task correlation and",
"9= P(y =1X,sm) =F(x1 2..Ng.(b)o m)",
"(1)",
"learns the function of specific tasks based on shared bottom feature",
"representation.HMoE [13] takes advantage of MMoE to implic-",
"where xi e Rd denotes the embedding of the i-th feature, and d",
"itlyidentifydistinctionsand commonalities between tasks,and",
"denotes the embedding dimension of features. b, e Rd denotes the",
"improves the performance with a stacked model learning task rela-",
"representation of i-th behavior in the behavior sequence, typically",
"tionships in the label space explicitly. Later, PLE [23] shares experts",
"derived from a pooled embedding that combines the clicked item",
"in the share layer and refines tasks uniquely to effectively alleviate",
"embeddingandtheembeddingsofitsassociatedattributefeatures",
"the noise caused by other scenarios and improves the effectiveness",
"Sm represents the representation vector of m-th scenario features."
] |
{
"arxiv_id": "2404.09709",
"language": "en",
"page_idx": 2
}
|
||
2404.09709-en-3
|
2404.09709-en-3.jpg
|
Woodstock'18, June 03-05, 2018, Woodstock, NY
Moyu Zhang, Yongxiang Tang, Jinxin Hu, and Yu Zhang
<FigureHere>
Figure 3: The network architecture of our proposed Scenario-Adaptive Fine-Grained Personalization Network (SFPNet).
# 4 мЕтнор
As previously discussed, contemporary multi-scenario models typically employ a unified framework to deliver personalized recommendations across various scenarios. While previous sceneadaptive parameter networks are lauded for their flexibility in adjusting the model's foundational representation to excel in multiscenario recommendations, they typically consider the entire user historical behavior sequence as a singular feature field and conduct scenario-based transformations within this sequence. This approach uniformly adjusts all behaviors with the same weight, thus overlooking the distinct semantic information that each behavior may signify in different scenarios. In reality, users' interests frequently shift with the scenario transition, causing each behavior in the sequence to represent varying information across different scenarios. Inadequate cross-scenario modeling precludes precise extraction of user preferences, reflected in individual behaviors, for each scenario, thereby compromising the accuracy of cross-scenario recommendations. To more effectively model the intricate interplay between user historical behaviors and scenarios, this paper introduces a novel framework named Scenario-Adaptive Fine-Grained Personalization Network (SFPNet). It is primarily built from stacked Scenario-Tailoring Block and comprises two main modules:
$\bullet$ Scenario-Adaptive Module (SAM), which employs a gating mechanism to integrate scenario-related information into the base representations of features, ensuring that the instance's contextual information is strongly guided by the scenario context.
$\bullet$ Residual-Tailoring Module (RTM), which aggregates scenarioaware contextual information and, via residual connections, disaggregates it into distinct representations for each behavior, allowing each behavior to capture contextual-cross information.
# 4.1 Scenario Adaptive Module (SAM)
Drawing inspiration from the LHUC algorithm [21], which focuses on learning the distinct contributions of each speaker's hidden units, a variety of scenario-adaptive parameterization methods have been developed. These methods involve integrating different scenario prior bias features to create personalized network parameters for each instance, thereby enabling distinction of features at the foundational level across different scenarios. Consequently, within each Scenario-Tailoring Block, we also incorporate a Scenario-Adaptive Module (SAM) that coarsely applies scenario prior information to diverse feature representations. This ensures that while serving as contextual information for instances, the distinctions between scenarios are consistently maintained. Specifically, the input to the SAM in the first Scenario-Tailoring Block is the original embedding vector of the features, and subsequent SAM inputs are typically the outputs from the Residual-Tailoring Module of the previous block. As the purpose of SAM is to produce the contextual information for the instance, we compress the user behavior sequence into a distinct feature field, enabling complete perception of the sequence's information during subsequent behavior customization. SAM operates through a two-step process: sequence pooling, and gate personalization. To elucidate, let's examine the computation within SAM of the $l$-th block, delineated as follows:
4.1.1 Sequence Pooling. Since that SAM's output serves solely as contextual information for the instance, our initial step is to pool the sequence. To avoid substantial increases in computational complexity from stacking blocks, we opt for computationally efficient aggregation methods-like summation, max pooling, or average pooling-for user sequences. This contrasts with the more complex and computationally intensive neural network structures used in DIN[35] to capture sequence information related to target items.
|
[
[
0.08660888671875,
0.0770263671875,
0.31787109375,
0.0770263671875,
0.31787109375,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.64794921875,
0.07574462890625,
0.9140625,
0.07574462890625,
0.9140625,
0.0889892578125,
0.64794921875,
0.0889892578125
],
[
0.314453125,
0.10919189453125,
0.3349609375,
0.10919189453125,
0.3349609375,
0.125,
0.314453125,
0.125
],
[
0.62158203125,
0.10986328125,
0.82421875,
0.11236572265625,
0.82421875,
0.126953125,
0.62158203125,
0.125
],
[
0.81884765625,
0.11236572265625,
0.869140625,
0.11236572265625,
0.869140625,
0.12371826171875,
0.81884765625,
0.12371826171875
],
[
0.318603515625,
0.12249755859375,
0.330078125,
0.12249755859375,
0.330078125,
0.1337890625,
0.318603515625,
0.1337890625
],
[
0.640625,
0.1275634765625,
0.66015625,
0.121826171875,
0.6689453125,
0.141357421875,
0.65185546875,
0.146484375
],
[
0.74658203125,
0.126953125,
0.7646484375,
0.12371826171875,
0.7685546875,
0.141357421875,
0.7509765625,
0.14453125
],
[
0.79736328125,
0.1287841796875,
0.81640625,
0.121826171875,
0.82666015625,
0.1383056640625,
0.80810546875,
0.1458740234375
],
[
0.271240234375,
0.1363525390625,
0.381591796875,
0.1363525390625,
0.381591796875,
0.149658203125,
0.271240234375,
0.149658203125
],
[
0.619140625,
0.143310546875,
0.6689453125,
0.1458740234375,
0.66845703125,
0.1610107421875,
0.61767578125,
0.1590576171875
],
[
0.73388671875,
0.1458740234375,
0.83251953125,
0.1458740234375,
0.83251953125,
0.160400390625,
0.73388671875,
0.160400390625
],
[
0.5556640625,
0.15283203125,
0.56201171875,
0.15087890625,
0.564453125,
0.1572265625,
0.55810546875,
0.1590576171875
],
[
0.57666015625,
0.147705078125,
0.61376953125,
0.147705078125,
0.61376953125,
0.1590576171875,
0.57666015625,
0.1590576171875
],
[
0.340576171875,
0.1666259765625,
0.374267578125,
0.160400390625,
0.378173828125,
0.173583984375,
0.3447265625,
0.1787109375
],
[
0.42724609375,
0.17236328125,
0.465576171875,
0.17236328125,
0.465576171875,
0.199462890625,
0.42724609375,
0.199462890625
],
[
0.68896484375,
0.17236328125,
0.70166015625,
0.17236328125,
0.70166015625,
0.18310546875,
0.68896484375,
0.18310546875
],
[
0.3095703125,
0.199462890625,
0.314453125,
0.199462890625,
0.314453125,
0.2083740234375,
0.3095703125,
0.2083740234375
],
[
0.371826171875,
0.199462890625,
0.378173828125,
0.199462890625,
0.378173828125,
0.203857421875,
0.371826171875,
0.203857421875
],
[
0.414306640625,
0.1962890625,
0.428955078125,
0.1962890625,
0.428955078125,
0.2298583984375,
0.414306640625,
0.2298583984375
],
[
0.486083984375,
0.199462890625,
0.49267578125,
0.2008056640625,
0.491943359375,
0.20703125,
0.484375,
0.205810546875
],
[
0.371826171875,
0.208984375,
0.378173828125,
0.208984375,
0.378173828125,
0.21533203125,
0.371826171875,
0.21533203125
],
[
0.48779296875,
0.208984375,
0.49267578125,
0.208984375,
0.49267578125,
0.21533203125,
0.48779296875,
0.21533203125
],
[
0.6640625,
0.205810546875,
0.72607421875,
0.205810546875,
0.72607421875,
0.2171630859375,
0.6640625,
0.2171630859375
],
[
0.370849609375,
0.216552734375,
0.3798828125,
0.216552734375,
0.3798828125,
0.2291259765625,
0.370849609375,
0.2291259765625
],
[
0.6640625,
0.21533203125,
0.72802734375,
0.21533203125,
0.72802734375,
0.2266845703125,
0.6640625,
0.2266845703125
],
[
0.12744140625,
0.2342529296875,
0.181396484375,
0.2342529296875,
0.181396484375,
0.2442626953125,
0.12744140625,
0.2442626953125
],
[
0.30810546875,
0.2203369140625,
0.313720703125,
0.2203369140625,
0.313720703125,
0.2279052734375,
0.30810546875,
0.2279052734375
],
[
0.12744140625,
0.2442626953125,
0.1846923828125,
0.2442626953125,
0.1846923828125,
0.255615234375,
0.12744140625,
0.255615234375
],
[
0.2459716796875,
0.2412109375,
0.42724609375,
0.2412109375,
0.42724609375,
0.255126953125,
0.2459716796875,
0.255126953125
],
[
0.136474609375,
0.255126953125,
0.1707763671875,
0.255126953125,
0.1707763671875,
0.267578125,
0.136474609375,
0.267578125
],
[
0.301513671875,
0.255615234375,
0.343017578125,
0.255615234375,
0.343017578125,
0.270751953125,
0.301513671875,
0.270751953125
],
[
0.1380615234375,
0.266357421875,
0.1666259765625,
0.266357421875,
0.1666259765625,
0.27783203125,
0.1380615234375,
0.27783203125
],
[
0.8505859375,
0.261962890625,
0.88134765625,
0.261962890625,
0.88134765625,
0.270751953125,
0.8505859375,
0.270751953125
],
[
0.2459716796875,
0.27587890625,
0.42724609375,
0.27587890625,
0.42724609375,
0.285888671875,
0.2459716796875,
0.285888671875
],
[
0.3056640625,
0.292236328125,
0.318603515625,
0.292236328125,
0.318603515625,
0.343505859375,
0.3056640625,
0.343505859375
],
[
0.193603515625,
0.298583984375,
0.1993408203125,
0.298583984375,
0.1993408203125,
0.30810546875,
0.193603515625,
0.30810546875
],
[
0.1895751953125,
0.30615234375,
0.2034912109375,
0.30615234375,
0.2034912109375,
0.343505859375,
0.1895751953125,
0.343505859375
],
[
0.2344970703125,
0.300537109375,
0.25,
0.300537109375,
0.25,
0.341552734375,
0.2344970703125,
0.341552734375
],
[
0.111083984375,
0.309326171875,
0.11517333984375,
0.309326171875,
0.11517333984375,
0.31494140625,
0.111083984375,
0.31494140625
],
[
0.374267578125,
0.309326171875,
0.378173828125,
0.311279296875,
0.374267578125,
0.31494140625,
0.370849609375,
0.313232421875
],
[
0.75634765625,
0.311279296875,
0.81201171875,
0.311279296875,
0.81201171875,
0.3212890625,
0.75634765625,
0.3212890625
],
[
0.418212890625,
0.324462890625,
0.423095703125,
0.324462890625,
0.423095703125,
0.33837890625,
0.418212890625,
0.33837890625
],
[
0.75634765625,
0.3193359375,
0.8203125,
0.3193359375,
0.8203125,
0.330078125,
0.75634765625,
0.330078125
],
[
0.373291015625,
0.32763671875,
0.378173828125,
0.32763671875,
0.378173828125,
0.337158203125,
0.373291015625,
0.337158203125
],
[
0.2607421875,
0.343505859375,
0.351318359375,
0.343505859375,
0.351318359375,
0.354248046875,
0.2607421875,
0.354248046875
],
[
0.650390625,
0.35107421875,
0.689453125,
0.35107421875,
0.689453125,
0.371826171875,
0.650390625,
0.371826171875
],
[
0.654296875,
0.3466796875,
0.6650390625,
0.3466796875,
0.6650390625,
0.359130859375,
0.654296875,
0.359130859375
],
[
0.73046875,
0.35595703125,
0.74658203125,
0.35595703125,
0.74658203125,
0.365478515625,
0.73046875,
0.365478515625
],
[
0.80126953125,
0.354736328125,
0.83837890625,
0.348388671875,
0.84326171875,
0.365478515625,
0.80615234375,
0.371826171875
],
[
0.56884765625,
0.35791015625,
0.58837890625,
0.352783203125,
0.59228515625,
0.361083984375,
0.57275390625,
0.365478515625
],
[
0.30810546875,
0.35791015625,
0.316162109375,
0.35791015625,
0.316162109375,
0.367431640625,
0.30810546875,
0.367431640625
],
[
0.87353515625,
0.35791015625,
0.896484375,
0.35791015625,
0.896484375,
0.365478515625,
0.87353515625,
0.365478515625
],
[
0.81396484375,
0.37060546875,
0.861328125,
0.37060546875,
0.861328125,
0.3818359375,
0.81396484375,
0.3818359375
],
[
0.193603515625,
0.359130859375,
0.1993408203125,
0.359130859375,
0.1993408203125,
0.3662109375,
0.193603515625,
0.3662109375
],
[
0.09722900390625,
0.371826171875,
0.131591796875,
0.371826171875,
0.131591796875,
0.380615234375,
0.09722900390625,
0.380615234375
],
[
0.351318359375,
0.37255859375,
0.439453125,
0.37255859375,
0.439453125,
0.38330078125,
0.351318359375,
0.38330078125
],
[
0.468017578125,
0.373779296875,
0.5146484375,
0.373779296875,
0.5146484375,
0.3837890625,
0.468017578125,
0.3837890625
],
[
0.6298828125,
0.373779296875,
0.66259765625,
0.373779296875,
0.66259765625,
0.385009765625,
0.6298828125,
0.385009765625
],
[
0.7138671875,
0.373779296875,
0.74609375,
0.373779296875,
0.74609375,
0.386474609375,
0.7138671875,
0.386474609375
],
[
0.291748046875,
0.375732421875,
0.326904296875,
0.375732421875,
0.326904296875,
0.38330078125,
0.291748046875,
0.38330078125
],
[
0.181396484375,
0.376953125,
0.2115478515625,
0.376953125,
0.2115478515625,
0.380615234375,
0.181396484375,
0.380615234375
],
[
0.221435546875,
0.375,
0.25732421875,
0.375,
0.25732421875,
0.3818359375,
0.221435546875,
0.3818359375
],
[
0.8056640625,
0.3818359375,
0.86669921875,
0.3818359375,
0.86669921875,
0.392578125,
0.8056640625,
0.392578125
],
[
0.3798828125,
0.385009765625,
0.471435546875,
0.385009765625,
0.471435546875,
0.392578125,
0.3798828125,
0.392578125
],
[
0.101318359375,
0.397705078125,
0.8935546875,
0.396484375,
0.8935546875,
0.410888671875,
0.101318359375,
0.412353515625
],
[
0.0849609375,
0.431884765625,
0.193603515625,
0.431884765625,
0.193603515625,
0.44580078125,
0.0849609375,
0.44580078125
],
[
0.5556640625,
0.431884765625,
0.826171875,
0.431884765625,
0.826171875,
0.4462890625,
0.5556640625,
0.4462890625
],
[
0.51953125,
0.434326171875,
0.56201171875,
0.434326171875,
0.56201171875,
0.4443359375,
0.51953125,
0.4443359375
],
[
0.098876953125,
0.37890625,
0.1298828125,
0.37890625,
0.1298828125,
0.39013671875,
0.098876953125,
0.39013671875
],
[
0.08416748046875,
0.449462890625,
0.483642578125,
0.45068359375,
0.483642578125,
0.46533203125,
0.08416748046875,
0.464111328125
],
[
0.51708984375,
0.45068359375,
0.91259765625,
0.45068359375,
0.91259765625,
0.464111328125,
0.51708984375,
0.464111328125
],
[
0.0849609375,
0.46533203125,
0.48193359375,
0.46533203125,
0.48193359375,
0.479248046875,
0.0849609375,
0.479248046875
],
[
0.51708984375,
0.46533203125,
0.9140625,
0.46533203125,
0.9140625,
0.479248046875,
0.51708984375,
0.479248046875
],
[
0.08660888671875,
0.479736328125,
0.48193359375,
0.479736328125,
0.48193359375,
0.490478515625,
0.08660888671875,
0.490478515625
],
[
0.5185546875,
0.479736328125,
0.91259765625,
0.479736328125,
0.91259765625,
0.490478515625,
0.5185546875,
0.490478515625
],
[
0.0849609375,
0.492431640625,
0.483642578125,
0.492431640625,
0.483642578125,
0.505859375,
0.0849609375,
0.505859375
],
[
0.51708984375,
0.492431640625,
0.9140625,
0.492431640625,
0.9140625,
0.505859375,
0.51708984375,
0.505859375
],
[
0.0849609375,
0.505859375,
0.483642578125,
0.505859375,
0.483642578125,
0.51953125,
0.0849609375,
0.51953125
],
[
0.5185546875,
0.50830078125,
0.91259765625,
0.50830078125,
0.91259765625,
0.51806640625,
0.5185546875,
0.51806640625
],
[
0.08660888671875,
0.52099609375,
0.48193359375,
0.52099609375,
0.48193359375,
0.5341796875,
0.08660888671875,
0.5341796875
],
[
0.5185546875,
0.521484375,
0.91259765625,
0.521484375,
0.91259765625,
0.5322265625,
0.5185546875,
0.5322265625
],
[
0.0849609375,
0.5341796875,
0.483642578125,
0.5341796875,
0.483642578125,
0.54736328125,
0.0849609375,
0.54736328125
],
[
0.51708984375,
0.5341796875,
0.9140625,
0.5341796875,
0.9140625,
0.54736328125,
0.51708984375,
0.54736328125
],
[
0.0849609375,
0.54736328125,
0.48193359375,
0.54736328125,
0.48193359375,
0.56103515625,
0.0849609375,
0.56103515625
],
[
0.5185546875,
0.5498046875,
0.9111328125,
0.5498046875,
0.9111328125,
0.56005859375,
0.5185546875,
0.56005859375
],
[
0.08416748046875,
0.5625,
0.483642578125,
0.56103515625,
0.483642578125,
0.57470703125,
0.08416748046875,
0.57568359375
],
[
0.51953125,
0.56298828125,
0.9111328125,
0.56298828125,
0.9111328125,
0.57373046875,
0.51953125,
0.57373046875
],
[
0.0849609375,
0.57568359375,
0.483642578125,
0.57568359375,
0.483642578125,
0.5888671875,
0.0849609375,
0.5888671875
],
[
0.51708984375,
0.57470703125,
0.91259765625,
0.57470703125,
0.91259765625,
0.58837890625,
0.51708984375,
0.58837890625
],
[
0.0849609375,
0.5888671875,
0.48193359375,
0.5888671875,
0.48193359375,
0.60302734375,
0.0849609375,
0.60302734375
],
[
0.51708984375,
0.5888671875,
0.9140625,
0.5888671875,
0.9140625,
0.60302734375,
0.51708984375,
0.60302734375
],
[
0.08660888671875,
0.60498046875,
0.48046875,
0.60498046875,
0.48046875,
0.61572265625,
0.08660888671875,
0.61572265625
],
[
0.5185546875,
0.60498046875,
0.9111328125,
0.60498046875,
0.9111328125,
0.61572265625,
0.5185546875,
0.61572265625
],
[
0.0849609375,
0.6171875,
0.48193359375,
0.6171875,
0.48193359375,
0.630859375,
0.0849609375,
0.630859375
],
[
0.51708984375,
0.6162109375,
0.9140625,
0.6162109375,
0.9140625,
0.6298828125,
0.51708984375,
0.6298828125
],
[
0.08660888671875,
0.63330078125,
0.48046875,
0.63330078125,
0.48046875,
0.64306640625,
0.08660888671875,
0.64306640625
],
[
0.51708984375,
0.630859375,
0.91259765625,
0.630859375,
0.91259765625,
0.64453125,
0.51708984375,
0.64453125
],
[
0.0849609375,
0.64453125,
0.48193359375,
0.64453125,
0.48193359375,
0.65771484375,
0.0849609375,
0.65771484375
],
[
0.51708984375,
0.64453125,
0.91259765625,
0.64453125,
0.91259765625,
0.65771484375,
0.51708984375,
0.65771484375
],
[
0.08660888671875,
0.66015625,
0.48046875,
0.66015625,
0.48046875,
0.67041015625,
0.08660888671875,
0.67041015625
],
[
0.51708984375,
0.65771484375,
0.91259765625,
0.65771484375,
0.91259765625,
0.671875,
0.51708984375,
0.671875
],
[
0.0849609375,
0.67041015625,
0.48193359375,
0.671875,
0.48193359375,
0.68603515625,
0.0849609375,
0.68505859375
],
[
0.5185546875,
0.673828125,
0.9111328125,
0.673828125,
0.9111328125,
0.6845703125,
0.5185546875,
0.6845703125
],
[
0.08416748046875,
0.68505859375,
0.483642578125,
0.68603515625,
0.483642578125,
0.69970703125,
0.08416748046875,
0.69873046875
],
[
0.5185546875,
0.68603515625,
0.91259765625,
0.68603515625,
0.91259765625,
0.69970703125,
0.5185546875,
0.69970703125
],
[
0.08660888671875,
0.7021484375,
0.48046875,
0.7021484375,
0.48046875,
0.7119140625,
0.08660888671875,
0.7119140625
],
[
0.5185546875,
0.70068359375,
0.91259765625,
0.70068359375,
0.91259765625,
0.7109375,
0.5185546875,
0.7109375
],
[
0.08660888671875,
0.71337890625,
0.48193359375,
0.71337890625,
0.48193359375,
0.7265625,
0.08660888671875,
0.7265625
],
[
0.515625,
0.7138671875,
0.9140625,
0.7138671875,
0.9140625,
0.72802734375,
0.515625,
0.72802734375
],
[
0.0849609375,
0.7265625,
0.455078125,
0.72900390625,
0.455078125,
0.7412109375,
0.0849609375,
0.7392578125
],
[
0.51708984375,
0.72802734375,
0.9140625,
0.72802734375,
0.9140625,
0.7412109375,
0.51708984375,
0.7412109375
],
[
0.08660888671875,
0.7412109375,
0.48193359375,
0.7412109375,
0.48193359375,
0.7548828125,
0.08660888671875,
0.7548828125
],
[
0.51708984375,
0.7412109375,
0.830078125,
0.7412109375,
0.830078125,
0.7548828125,
0.51708984375,
0.7548828125
],
[
0.08660888671875,
0.7568359375,
0.4794921875,
0.7568359375,
0.4794921875,
0.767578125,
0.08660888671875,
0.767578125
],
[
0.0849609375,
0.76806640625,
0.48046875,
0.76806640625,
0.48046875,
0.7822265625,
0.0849609375,
0.7822265625
],
[
0.0849609375,
0.7841796875,
0.416748046875,
0.7841796875,
0.416748046875,
0.79736328125,
0.0849609375,
0.79736328125
],
[
0.5185546875,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.5185546875,
0.798828125
],
[
0.08416748046875,
0.79541015625,
0.483642578125,
0.796875,
0.483642578125,
0.81103515625,
0.08416748046875,
0.81005859375
],
[
0.51708984375,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.08416748046875,
0.81005859375,
0.483642578125,
0.81103515625,
0.483642578125,
0.82470703125,
0.08416748046875,
0.82373046875
],
[
0.5185546875,
0.81494140625,
0.91259765625,
0.81494140625,
0.91259765625,
0.82568359375,
0.5185546875,
0.82568359375
],
[
0.0849609375,
0.82470703125,
0.483642578125,
0.82470703125,
0.483642578125,
0.83837890625,
0.0849609375,
0.83837890625
],
[
0.5185546875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0849609375,
0.83837890625,
0.418212890625,
0.83837890625,
0.418212890625,
0.8515625,
0.0849609375,
0.8515625
],
[
0.515625,
0.84130859375,
0.91259765625,
0.84033203125,
0.91259765625,
0.85498046875,
0.515625,
0.85595703125
],
[
0.51708984375,
0.85595703125,
0.9140625,
0.85595703125,
0.9140625,
0.869140625,
0.51708984375,
0.869140625
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.51708984375,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"Woodstock'18, June 03-05, 2018, Woodstock, NY",
"Moyu Zhang, Yongxiang Tang, Jinxin Hu, and Yu Zhang",
"回",
"thScenario-Tailoring",
"Block",
"1",
"热",
"",
"",
"Scenario-AwareDNN",
"\"-Tallor",
"Talor. Tallor",
"1",
"Tailor",
"H",
"IDIN",
"cl",
"",
"酒",
"",
"P",
"福",
"国",
"Aggregation",
"+",
"Interactiont",
"Scenario-",
"",
"Tailoring!",
"Residual-Tailoring Module (RTM)",
"Block",
"",
"(L*)",
"Gate!",
"Scenario-Adaptive Module (SAM)",
"14",
"-",
"←",
"",
"",
"福",
"Distribution",
"+",
"AwarePooling",
"",
"EmbeddingLayer",
"",
"服",
"",
"Ukg!",
"",
"T",
"Sm",
"Residual",
"",
"Scenario",
"Bchavior 1 Behavior 2",
"Behavior N,",
"RTM",
"SAM",
"Featurer",
"Featre",
"Feature2",
"Connection",
"Historical Bchavior Sequs",
"Figure 3: The network architecture of our proposed Scenario-Adaptive Fine-Grained Personalization Network (SFPNet).",
"4METHOD",
"Scenario Adaptive Module (SAM)",
"4.1",
"Feature",
"As previously discussed, contemporary multi-scenario models typ-",
"Drawing inspiration from the LHUC algorithm [21], which focuses",
"ically employ a unified framework to deliver personalized rec",
"on learning the distinct contributions of each speaker's hidden units,",
"ommendations across various scenarios.While previous scene-",
"a variety of scenario-adaptiveparameterization methods have been",
"adaptive parameter networks are lauded for their flexibility in",
"developed. These methods involve integrating different scenario",
"adjusting the model's foundational representation to excel in multi-",
"priorbiasfeaturestocreatepersonalizednetworkparametersfor",
"scenario recommendations, they typically consider the entire user",
"eachinstance,therebyenablingdistinctionoffeaturesatthefoun-",
"historical behavior sequence as a singular feature field and con-",
"dational level across different scenarios. Consequently, within each",
"duct scenario-based transformations within this sequence. This",
"Scenario-Tailoring Block,we also incorporatea Scenario-Adaptive",
"approach uniformly adjusts all behaviors with the same weight,",
"Module(SAM)thatcoarselyapplies scenariopriorinformationto",
"thus overlooking the distinct semantic information that each be-",
"diverse feature representations. This ensures that while serving",
"havior may signify in different scenarios. In reality, users' interests",
"as contextual information for instances, the distinctions between",
"frequently shift with the scenario transition, causing each behavior",
"scenarios are consistently maintained. Specifically,the input to the",
"in the sequence to represent varying information across different",
"SAM in the first Scenario-Tailoring Block is the original embedding",
"scenarios.Inadequatecross-scenario modeling precludes precise",
"vector of the features, and subsequent SAM inputs are typically",
"extraction of user preferences, reflected in individual behaviors,for",
"the outputs from the Residual-Tailoring Module of the previous",
"eachscenario,therebycompromisingtheaccuracyofcross-scenario",
"block. As the purpose of SAM is to produce the contextual infor-",
"recommendations. To more effectively model the intricate interplay",
"mationforthe instance,wecompresstheuserbehavior sequence",
"between user historical behaviors and scenarios, this paper intro-",
"into a distinct feature field, enabling complete perception of the",
"ducesanovelframeworknamedScenario-AdaptiveFine-Grained",
"sequence'sinformationduringsubsequentbehaviorcustomization",
"Personalization Network (SFPNet). It is primarily built from stacked",
"SAM operates through a two-step process: sequence pooling, and",
"Scenario-Tailoring Block and comprises twomain modules:",
"gate personalization. To elucidate, let's examine the computation",
". Scenario-Adaptive Module (SAM), which employs a gating",
"within SAM of the I-th block, delineated as follows:",
"mechanism to integrate scenario-related information into the base",
"representations of features, ensuring that the instance's contextual",
"information is strongly guided by the scenario context.",
"4.1.1 Sequence Pooling. Since that SAM's output serves solely as",
". Residual-Tailoring Module (RTM), which aggregates scenario-",
"contextual information for the instance, our initial step is to pool",
"aware contextual information and,viaresidual connections,disag-",
"the sequence.To avoid substantial increases in computational com-",
"gregates it into distinct representations for each behavior, allowing",
"plexityfromstackingblocks,weoptforcomputationallyefficient",
"each behavior to capture contextual-cross information.",
"aggregation methods-like summation, max pooling, or average",
"pooling-for user sequences. This contrasts with the more complex",
"and computationallyintensiveneuralnetworkstructuresusedin",
"DIN [35] to capture sequence information related to target items."
] |
{
"arxiv_id": "2404.09709",
"language": "en",
"page_idx": 3
}
|
||
1611.06722-en-2
|
1611.06722-en-2.jpg
|
### 2 $\textbf{Related Work}$
Transliteration research first associates with the field of orthographic similarity detections since sound similarities co-exist with orthographic similarities (Brew et al., 1996; Mann and Yarowsky, 2001; Dijkstra et al., 1999; Van Heuven et al., 1998; Kondrak, 2004; Chamizo Dominguez and Nerlich, 2002). This work shows reasonableness of character-based transliteration between close languages (i.e. languages sharing characters) but does not discuss on distant language pairs.
Similarly, work on cognate identification also focus on close language pairs (Simard et al., 1993; Inkpen et al., 2005; Schepens et al., 2013; Boada et al., 2013; Kolb, 2008; Kondrak and Dorr, 2004; Resnik, 2011). However, we believe multilingual transliterations contribute to even distant languages (e.g. English and Japanese) when handling OOV words and resolving ambiguities.
Further transliteration researches divide into two branches. One tries to study delicate sound changing rules of specific languages (Knight and Graehl, 1998; AbdulJaleel and Larkey, 2003; Suwanvisat and Prasitjutrakul, 1998; Gao et al., 2005; Virga and Khudanpur, 2003; Jagarlamudi and Daumé III, 2012; Hong et al., 2009). Especially, an excellent ideas of using Wikipedia external links is proposed in (Kirschenbaum and Wintner, 2009; Kirschenbaum and Wintner, 2010) and achieve promising results in English-Hebrew transliteration using Moses (Koehn et al., 2007). However, all these systems are supervised and require extra linguistic background knowledge during processing. Plus, only one among this work evaluates transliteration on up to 4 languages and it is hard to generalize for multiple languages.
The other branch learns from only sequence of characters. One of the great advantages against sound based transliteration is that multilingual texts are much easier to obtain. Al-Onaizan and Knight (2002) compares phonetic based systems with spelling based systems on transliterations between English and Arabic. Pouliquen et al. (2006) makes transliteration model based on similar spelling rules in close languages. Recent work of Durrani et al.(2014) is integrated in Moses as a module, providing an unsupervised characterbased transliteration training model. Matthews (2007) proposed a proper name transliteration model on several language pairs.
However,
we believe utilizing character-based transliteration
model can provide us with even more valuable in-
formation in natural language processing tasks.
# 3 Data Collection and Pre-processing
Kirschenbaum and Wintner (2009) inspire us to use Wikipedia external links to build an aligned multilingual corpus. However, their work requires language-specific knowledge, for instance, discarding vowels and filtering out junk data using pre-defined consonant matching. In our task, we use the names of entities (people and places) to create a training set for transliteration. (Francis et al., 2002) state that over 40% of the brands choose to create corresponding foreign names via transliterations, By querying Freebase in categories containing 3,388,225 entries, we create a precise multilingual transliteration dictionary through Wikipedia page titles. We then perform a rough clean up procedure to (1) unify punctuation by converting hyphens, dots, comma to underscores and, (2) remove entries which do not adhere to the (first name, last name) or whole name format Our final collection contains 576,403 English entries with multilingual mapping.
As an additional resource, we query Google translation API to get formal translations of certain English proper nouns to all 69 languages. To reduce machine translation error, we manually pick 1,373 entities without no multi-sense ambiguities from the names of people (Census, 2000), countries and capital cities (Wikipedia.org, 2014), resulting in more than 70,000 pairs of proper name transliteration from English. Table 1 shows statistics of final data size in each language. 80% of the final data will be used for training, 10% is for tuning and the remaining 10% is for testing.
<table>
<tbody>
<tr>
<th colspan="2">Largest</th>
<th colspan="2">Smallest</th>
</tr>
<tr>
<th>$\overline{\mathrm{Lang}}$</th>
<th>$\overline{\text{Count}}$</th>
<th>$\overline{\mathrm{Lang}}$</th>
<th>$\overline{\text{Count}}$</th>
</tr>
<tr>
<td>French</td>
<td>183,270</td>
<td>$\overline{{\mathrm{Khmer}}}$</td>
<td>1,585</td>
</tr>
<tr>
<td>German</td>
<td>178,715</td>
<td>Amharic</td>
<td>2,035</td>
</tr>
<tr>
<td>Italian</td>
<td>132,545</td>
<td>Gujarati</td>
<td>2,130</td>
</tr>
<tr>
<td>Polish</td>
<td>124,870</td>
<td>$\mathbf{Maltese}$</td>
<td>2,415</td>
</tr>
<tr>
<td>Spanish</td>
<td>107,790</td>
<td>Yiddish</td>
<td>2,835</td>
</tr>
<tr>
<td>Russian</td>
<td>100,085</td>
<td>Kannada</td>
<td>3,100</td>
</tr>
<tr>
<td>Swedish</td>
<td>91,125</td>
<td>Telugu</td>
<td>3,840</td>
</tr>
<tr>
<td>Dutch</td>
<td>87,870</td>
<td>Swahili</td>
<td>4,620</td>
</tr>
<tr>
<td>Portuguese</td>
<td>86,515</td>
<td>${\mathrm{Haitian}}$</td>
<td>5.245</td>
</tr>
<tr>
<td>Norwegian</td>
<td>74,790</td>
<td>Urdu</td>
<td>6,115</td>
</tr>
</tbody>
</table>
Table 1: Languages with the largest and smallest set of translated entities, i.e. reflecting the availability of training data.
|
[
[
0.12005615234375,
0.0748291015625,
0.27197265625,
0.0748291015625,
0.27197265625,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.51318359375,
0.07659912109375,
0.880859375,
0.07659912109375,
0.880859375,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.51318359375,
0.09320068359375,
0.8623046875,
0.09320068359375,
0.8623046875,
0.10687255859375,
0.51318359375,
0.10687255859375
],
[
0.12005615234375,
0.1051025390625,
0.48779296875,
0.1051025390625,
0.48779296875,
0.11517333984375,
0.12005615234375,
0.11517333984375
],
[
0.12005615234375,
0.119384765625,
0.490234375,
0.119384765625,
0.490234375,
0.133056640625,
0.12005615234375,
0.133056640625
],
[
0.51708984375,
0.127685546875,
0.5283203125,
0.127685546875,
0.5283203125,
0.1353759765625,
0.51708984375,
0.1353759765625
],
[
0.54150390625,
0.12469482421875,
0.8486328125,
0.1258544921875,
0.8486328125,
0.14013671875,
0.54150390625,
0.1395263671875
],
[
0.12005615234375,
0.135986328125,
0.48779296875,
0.135986328125,
0.48779296875,
0.149658203125,
0.12005615234375,
0.149658203125
],
[
0.118408203125,
0.15087890625,
0.489501953125,
0.1519775390625,
0.489501953125,
0.166259765625,
0.118408203125,
0.1650390625
],
[
0.51318359375,
0.152587890625,
0.8818359375,
0.152587890625,
0.8818359375,
0.166259765625,
0.51318359375,
0.166259765625
],
[
0.1175537109375,
0.166259765625,
0.489501953125,
0.16748046875,
0.489501953125,
0.1817626953125,
0.1175537109375,
0.18115234375
],
[
0.51123046875,
0.169189453125,
0.884765625,
0.17041015625,
0.884765625,
0.185302734375,
0.51123046875,
0.18408203125
],
[
0.12176513671875,
0.18408203125,
0.489501953125,
0.18408203125,
0.489501953125,
0.19775390625,
0.12176513671875,
0.19775390625
],
[
0.51318359375,
0.18701171875,
0.884765625,
0.18701171875,
0.884765625,
0.20068359375,
0.51318359375,
0.20068359375
],
[
0.12176513671875,
0.201904296875,
0.489501953125,
0.201904296875,
0.489501953125,
0.2120361328125,
0.12176513671875,
0.2120361328125
],
[
0.5146484375,
0.2037353515625,
0.88330078125,
0.2037353515625,
0.88330078125,
0.21728515625,
0.5146484375,
0.21728515625
],
[
0.118408203125,
0.2161865234375,
0.490234375,
0.2161865234375,
0.490234375,
0.2298583984375,
0.118408203125,
0.2298583984375
],
[
0.51318359375,
0.2197265625,
0.884765625,
0.2197265625,
0.884765625,
0.2327880859375,
0.51318359375,
0.2327880859375
],
[
0.118408203125,
0.2327880859375,
0.489501953125,
0.2327880859375,
0.489501953125,
0.2464599609375,
0.118408203125,
0.2464599609375
],
[
0.5146484375,
0.235107421875,
0.884765625,
0.235107421875,
0.884765625,
0.248779296875,
0.5146484375,
0.248779296875
],
[
0.118408203125,
0.248779296875,
0.4306640625,
0.248779296875,
0.4306640625,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.51318359375,
0.251708984375,
0.88330078125,
0.251708984375,
0.88330078125,
0.265380859375,
0.51318359375,
0.265380859375
],
[
0.1368408203125,
0.26416015625,
0.489501953125,
0.265380859375,
0.489501953125,
0.27978515625,
0.1368408203125,
0.279052734375
],
[
0.51123046875,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.51123046875,
0.27978515625
],
[
0.118408203125,
0.281982421875,
0.490234375,
0.281982421875,
0.490234375,
0.295654296875,
0.118408203125,
0.295654296875
],
[
0.51123046875,
0.281982421875,
0.8818359375,
0.281982421875,
0.8818359375,
0.295654296875,
0.51123046875,
0.295654296875
],
[
0.118408203125,
0.297607421875,
0.490234375,
0.297607421875,
0.490234375,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51318359375,
0.2998046875,
0.884765625,
0.2998046875,
0.884765625,
0.3134765625,
0.51318359375,
0.3134765625
],
[
0.12005615234375,
0.3134765625,
0.490234375,
0.3134765625,
0.490234375,
0.32666015625,
0.12005615234375,
0.32666015625
],
[
0.51318359375,
0.31640625,
0.8818359375,
0.31640625,
0.8818359375,
0.330078125,
0.51318359375,
0.330078125
],
[
0.118408203125,
0.330078125,
0.489501953125,
0.330078125,
0.489501953125,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.51123046875,
0.33203125,
0.88330078125,
0.330810546875,
0.88330078125,
0.345703125,
0.51123046875,
0.3466796875
],
[
0.1175537109375,
0.345703125,
0.490234375,
0.344482421875,
0.490234375,
0.359375,
0.1175537109375,
0.35986328125
],
[
0.51318359375,
0.3486328125,
0.8818359375,
0.3486328125,
0.8818359375,
0.3623046875,
0.51318359375,
0.3623046875
],
[
0.12005615234375,
0.3623046875,
0.489501953125,
0.3623046875,
0.489501953125,
0.3759765625,
0.12005615234375,
0.3759765625
],
[
0.51318359375,
0.364013671875,
0.884765625,
0.364013671875,
0.884765625,
0.37890625,
0.51318359375,
0.37890625
],
[
0.118408203125,
0.377197265625,
0.4072265625,
0.377685546875,
0.4072265625,
0.392578125,
0.118408203125,
0.391357421875
],
[
0.51123046875,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.393798828125,
0.51123046875,
0.393798828125
],
[
0.1376953125,
0.393798828125,
0.489501953125,
0.393798828125,
0.489501953125,
0.40673828125,
0.1376953125,
0.40673828125
],
[
0.51318359375,
0.396728515625,
0.88330078125,
0.396728515625,
0.88330078125,
0.410400390625,
0.51318359375,
0.410400390625
],
[
0.118408203125,
0.410888671875,
0.490234375,
0.410888671875,
0.490234375,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.51123046875,
0.412109375,
0.884765625,
0.412109375,
0.884765625,
0.42578125,
0.51123046875,
0.42578125
],
[
0.12005615234375,
0.427490234375,
0.489501953125,
0.427490234375,
0.489501953125,
0.441162109375,
0.12005615234375,
0.441162109375
],
[
0.51318359375,
0.4287109375,
0.88330078125,
0.4287109375,
0.88330078125,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.12005615234375,
0.443603515625,
0.490234375,
0.443603515625,
0.490234375,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.51318359375,
0.44482421875,
0.884765625,
0.44482421875,
0.884765625,
0.457763671875,
0.51318359375,
0.457763671875
],
[
0.1175537109375,
0.457763671875,
0.48779296875,
0.458984375,
0.48779296875,
0.473876953125,
0.1175537109375,
0.47265625
],
[
0.51318359375,
0.46142578125,
0.77099609375,
0.46142578125,
0.77099609375,
0.474365234375,
0.51318359375,
0.474365234375
],
[
0.12005615234375,
0.4755859375,
0.489501953125,
0.4755859375,
0.489501953125,
0.4892578125,
0.12005615234375,
0.4892578125
],
[
0.5322265625,
0.47802734375,
0.88330078125,
0.47802734375,
0.88330078125,
0.4921875,
0.5322265625,
0.4921875
],
[
0.118408203125,
0.4912109375,
0.489501953125,
0.4912109375,
0.489501953125,
0.5048828125,
0.118408203125,
0.5048828125
],
[
0.51318359375,
0.49462890625,
0.884765625,
0.49462890625,
0.884765625,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.118408203125,
0.50830078125,
0.490234375,
0.50830078125,
0.490234375,
0.521484375,
0.118408203125,
0.521484375
],
[
0.5146484375,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52490234375,
0.5146484375,
0.52490234375
],
[
0.118408203125,
0.52392578125,
0.489501953125,
0.52392578125,
0.489501953125,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.51318359375,
0.52685546875,
0.88330078125,
0.52685546875,
0.88330078125,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.118408203125,
0.5390625,
0.48779296875,
0.5390625,
0.48779296875,
0.552734375,
0.118408203125,
0.552734375
],
[
0.5146484375,
0.5419921875,
0.88330078125,
0.5419921875,
0.88330078125,
0.5556640625,
0.5146484375,
0.5556640625
],
[
0.1175537109375,
0.5546875,
0.486083984375,
0.5546875,
0.486083984375,
0.568359375,
0.1175537109375,
0.568359375
],
[
0.5146484375,
0.56005859375,
0.880859375,
0.56005859375,
0.880859375,
0.57080078125,
0.5146484375,
0.57080078125
],
[
0.1175537109375,
0.57177734375,
0.489501953125,
0.57080078125,
0.489501953125,
0.5849609375,
0.1175537109375,
0.5859375
],
[
0.51318359375,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.5888671875,
0.51318359375,
0.5888671875
],
[
0.118408203125,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6015625,
0.118408203125,
0.6015625
],
[
0.51318359375,
0.59130859375,
0.884765625,
0.59130859375,
0.884765625,
0.60498046875,
0.51318359375,
0.60498046875
],
[
0.118408203125,
0.60498046875,
0.489501953125,
0.60498046875,
0.489501953125,
0.6181640625,
0.118408203125,
0.6181640625
],
[
0.51123046875,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.619140625,
0.51123046875,
0.619140625
],
[
0.118408203125,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.51318359375,
0.62353515625,
0.88671875,
0.62353515625,
0.88671875,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.12005615234375,
0.63720703125,
0.489501953125,
0.63720703125,
0.489501953125,
0.65087890625,
0.12005615234375,
0.65087890625
],
[
0.51123046875,
0.63818359375,
0.88330078125,
0.63818359375,
0.88330078125,
0.65185546875,
0.51123046875,
0.65185546875
],
[
0.1175537109375,
0.65185546875,
0.459228515625,
0.65283203125,
0.459228515625,
0.66748046875,
0.1175537109375,
0.66650390625
],
[
0.509765625,
0.65283203125,
0.84033203125,
0.65380859375,
0.84033203125,
0.67138671875,
0.509765625,
0.67041015625
],
[
0.1368408203125,
0.66845703125,
0.489501953125,
0.66943359375,
0.489501953125,
0.68408203125,
0.1368408203125,
0.68310546875
],
[
0.118408203125,
0.68603515625,
0.489501953125,
0.68603515625,
0.489501953125,
0.69970703125,
0.118408203125,
0.69970703125
],
[
0.60693359375,
0.68603515625,
0.65576171875,
0.68603515625,
0.65576171875,
0.69775390625,
0.60693359375,
0.69775390625
],
[
0.75244140625,
0.68603515625,
0.8095703125,
0.68603515625,
0.8095703125,
0.69775390625,
0.75244140625,
0.69775390625
],
[
0.59765625,
0.69677734375,
0.63623046875,
0.69873046875,
0.634765625,
0.71044921875,
0.59619140625,
0.7080078125
],
[
0.66162109375,
0.69775390625,
0.7060546875,
0.69775390625,
0.7060546875,
0.708984375,
0.66162109375,
0.708984375
],
[
0.74658203125,
0.69677734375,
0.78515625,
0.69873046875,
0.7841796875,
0.71142578125,
0.74462890625,
0.708984375
],
[
0.7900390625,
0.69873046875,
0.8427734375,
0.69873046875,
0.8427734375,
0.708984375,
0.7900390625,
0.708984375
],
[
0.73486328125,
0.7080078125,
0.78076171875,
0.71044921875,
0.779296875,
0.72265625,
0.73291015625,
0.72021484375
],
[
0.12005615234375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.5859375,
0.71044921875,
0.63623046875,
0.71044921875,
0.63623046875,
0.7216796875,
0.5859375,
0.7216796875
],
[
0.6513671875,
0.71044921875,
0.70361328125,
0.71044921875,
0.70361328125,
0.7216796875,
0.6513671875,
0.7216796875
],
[
0.8017578125,
0.71044921875,
0.84228515625,
0.71044921875,
0.84228515625,
0.7216796875,
0.8017578125,
0.7216796875
],
[
0.118408203125,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.580078125,
0.7197265625,
0.63818359375,
0.7216796875,
0.63623046875,
0.736328125,
0.57861328125,
0.73388671875
],
[
0.6513671875,
0.7216796875,
0.70361328125,
0.7216796875,
0.70361328125,
0.73291015625,
0.6513671875,
0.73291015625
],
[
0.72705078125,
0.72265625,
0.78076171875,
0.72265625,
0.78076171875,
0.73388671875,
0.72705078125,
0.73388671875
],
[
0.8017578125,
0.7216796875,
0.84033203125,
0.7216796875,
0.84033203125,
0.73291015625,
0.8017578125,
0.73291015625
],
[
0.6513671875,
0.73291015625,
0.7021484375,
0.73291015625,
0.7021484375,
0.74462890625,
0.6513671875,
0.74462890625
],
[
0.8017578125,
0.73291015625,
0.84033203125,
0.73291015625,
0.84033203125,
0.7451171875,
0.8017578125,
0.7451171875
],
[
0.59033203125,
0.73388671875,
0.63818359375,
0.73388671875,
0.63818359375,
0.7451171875,
0.59033203125,
0.7451171875
],
[
0.12005615234375,
0.73388671875,
0.48779296875,
0.73388671875,
0.48779296875,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.72705078125,
0.73388671875,
0.78076171875,
0.73388671875,
0.78076171875,
0.7451171875,
0.72705078125,
0.7451171875
],
[
0.59033203125,
0.7451171875,
0.63623046875,
0.7451171875,
0.63623046875,
0.75732421875,
0.59033203125,
0.75732421875
],
[
0.64990234375,
0.7451171875,
0.70361328125,
0.7451171875,
0.70361328125,
0.75732421875,
0.64990234375,
0.75732421875
],
[
0.72900390625,
0.74365234375,
0.78271484375,
0.7451171875,
0.78076171875,
0.75830078125,
0.72705078125,
0.755859375
],
[
0.80029296875,
0.7451171875,
0.84033203125,
0.7451171875,
0.84033203125,
0.75732421875,
0.80029296875,
0.75732421875
],
[
0.12176513671875,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.7626953125,
0.12176513671875,
0.7626953125
],
[
0.5810546875,
0.75830078125,
0.63623046875,
0.75830078125,
0.63623046875,
0.76953125,
0.5810546875,
0.76953125
],
[
0.6484375,
0.75830078125,
0.705078125,
0.75830078125,
0.705078125,
0.7685546875,
0.6484375,
0.7685546875
],
[
0.73046875,
0.75830078125,
0.78271484375,
0.75830078125,
0.78271484375,
0.76953125,
0.73046875,
0.76953125
],
[
0.80029296875,
0.75830078125,
0.84033203125,
0.75830078125,
0.84033203125,
0.76953125,
0.80029296875,
0.76953125
],
[
0.118408203125,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.77978515625,
0.118408203125,
0.77978515625
],
[
0.5810546875,
0.76953125,
0.6396484375,
0.76953125,
0.6396484375,
0.78076171875,
0.5810546875,
0.78076171875
],
[
0.6484375,
0.76953125,
0.70361328125,
0.76953125,
0.70361328125,
0.78076171875,
0.6484375,
0.78076171875
],
[
0.724609375,
0.76953125,
0.78271484375,
0.76953125,
0.78271484375,
0.78076171875,
0.724609375,
0.78076171875
],
[
0.80029296875,
0.76953125,
0.84228515625,
0.76953125,
0.84228515625,
0.78076171875,
0.80029296875,
0.78076171875
],
[
0.57666015625,
0.78076171875,
0.63623046875,
0.78076171875,
0.63623046875,
0.7919921875,
0.57666015625,
0.7919921875
],
[
0.65576171875,
0.78076171875,
0.70361328125,
0.78076171875,
0.70361328125,
0.7919921875,
0.65576171875,
0.7919921875
],
[
0.73486328125,
0.78076171875,
0.78271484375,
0.78076171875,
0.78271484375,
0.7919921875,
0.73486328125,
0.7919921875
],
[
0.80029296875,
0.78076171875,
0.84228515625,
0.78076171875,
0.84228515625,
0.7919921875,
0.80029296875,
0.7919921875
],
[
0.591796875,
0.79345703125,
0.63623046875,
0.79345703125,
0.63623046875,
0.8046875,
0.591796875,
0.8046875
],
[
0.12005615234375,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.6572265625,
0.79345703125,
0.705078125,
0.79345703125,
0.705078125,
0.8046875,
0.6572265625,
0.8046875
],
[
0.73046875,
0.79345703125,
0.78271484375,
0.79345703125,
0.78271484375,
0.8046875,
0.73046875,
0.8046875
],
[
0.80029296875,
0.79345703125,
0.84228515625,
0.79345703125,
0.84228515625,
0.8046875,
0.80029296875,
0.8046875
],
[
0.55908203125,
0.802734375,
0.63623046875,
0.8046875,
0.63623046875,
0.8193359375,
0.55908203125,
0.81787109375
],
[
0.73291015625,
0.80419921875,
0.78271484375,
0.8056640625,
0.78076171875,
0.81787109375,
0.7314453125,
0.8154296875
],
[
0.80029296875,
0.80419921875,
0.84228515625,
0.80419921875,
0.84228515625,
0.81787109375,
0.80029296875,
0.81787109375
],
[
0.6572265625,
0.8046875,
0.70361328125,
0.8046875,
0.70361328125,
0.81640625,
0.6572265625,
0.81640625
],
[
0.118408203125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.55908203125,
0.81640625,
0.634765625,
0.81640625,
0.634765625,
0.82763671875,
0.55908203125,
0.82763671875
],
[
0.6572265625,
0.81640625,
0.7060546875,
0.81640625,
0.7060546875,
0.82763671875,
0.6572265625,
0.82763671875
],
[
0.74462890625,
0.81640625,
0.7841796875,
0.81640625,
0.7841796875,
0.82763671875,
0.74462890625,
0.82763671875
],
[
0.80029296875,
0.81640625,
0.84228515625,
0.81640625,
0.84228515625,
0.82763671875,
0.80029296875,
0.82763671875
],
[
0.1175537109375,
0.83056640625,
0.484375,
0.83056640625,
0.484375,
0.84423828125,
0.1175537109375,
0.84423828125
],
[
0.51318359375,
0.84033203125,
0.884765625,
0.84130859375,
0.884765625,
0.85546875,
0.51318359375,
0.85498046875
],
[
0.40478515625,
0.845703125,
0.490234375,
0.84716796875,
0.490234375,
0.8623046875,
0.40478515625,
0.85986328125
],
[
0.118408203125,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.118408203125,
0.8125
],
[
0.1175537109375,
0.84716796875,
0.4130859375,
0.84716796875,
0.4130859375,
0.86083984375,
0.1175537109375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.490234375,
0.86328125,
0.490234375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.85791015625,
0.88330078125,
0.85791015625,
0.88330078125,
0.87158203125,
0.51318359375,
0.87158203125
],
[
0.1175537109375,
0.87744140625,
0.39794921875,
0.87890625,
0.39794921875,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.5146484375,
0.87451171875,
0.68603515625,
0.87451171875,
0.68603515625,
0.88818359375,
0.5146484375,
0.88818359375
],
[
0.393798828125,
0.88525390625,
0.40478515625,
0.88525390625,
0.40478515625,
0.88818359375,
0.393798828125,
0.88818359375
],
[
0.4130859375,
0.88134765625,
0.48779296875,
0.88134765625,
0.48779296875,
0.89111328125,
0.4130859375,
0.89111328125
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
]
] |
[
"2 Related Work",
"model can provide us with even more valuable in-",
"formation in natural language processing tasks.",
"Transliteration research first associates with the",
"field of orthographic similarity detections since",
"3",
"Data Collection and Pre-processing",
"sound similarities co-exist with orthographic sim-",
"ilarities (Brew et al., 1996; Mann and Yarowsky,",
"Kirschenbaum and Wintner (2009) inspire us to",
"2001; Dijkstra et al., 1999; Van Heuven et al.,",
"use Wikipedia external links to build an aligned",
"1998; Kondrak, 2004; Chamizo Dominguez and",
"multilingual corpus. However, their work requires",
"Nerlich.2002).This workshowsreasonableness",
"language-specific knowledge, for instance, dis-",
"of character-based transliteration between close",
"carding vowels and filtering out junk data using",
"languages (i.e. languages sharing characters) but",
"pre-defined consonant matching. In our task, we",
"does not discuss on distant language pairs.",
"use the names of entities (people and places) to",
"Similarly, work on cognate identification also",
"create a training set for transliteration. (Francis",
"focus on close language pairs (Simard et al., 1993;",
"et al., 2002) state that over 40% of the brands",
"Inkpen et al., 2005; Schepens et al., 2013; Boada",
"choose to create corresponding foreign names",
"et al., 2013; Kolb, 2008; Kondrak and Dorr, 2004;",
"via transliterations, By querying Freebase in cat",
"Resnik, 2011). However, we believe multilin-",
"egories containing 3,388,225 entries, we create",
"gual transliterations contribute to even distant lan-",
"a precise multilingual transliteration dictionary",
"guages (e.g. English and Japanese) when handling",
"through Wikipedia page titles. We then perform",
"OOV words and resolving ambiguities.",
"a rough clean up procedure to (1) unify punctua-",
"Further transliteration researches divide into",
"tion by converting hyphens, dots, comma to under-",
"two branches. One tries to study delicate sound",
"scores and, (2) remove entries which do not adhere",
"changing rules of specific languages (Knight and",
"to the (first name, last name) or whole name for-",
"Graehl, 1998; AbdulJaleel and Larkey, 2003;",
"mat Our final collection contains 576,403 English",
"Suwanvisat and Prasitjutrakul, 1998; Gao et al.,",
"entries with multilingual mapping.",
"2005; Virga and Khudanpur, 2003; Jagarlamudi",
"As an additional resource, we query Google",
"and Daume IIl, 2012; Hong et al., 2009). Es-",
"translation API to get formal translations of certain",
"pecially, an excellent ideas of using Wikipedia",
"English proper nouns to all 69 languages. To re-",
"external links is proposed in (Kirschenbaum and",
"duce machine translation error, we manually pick",
"Wintner, 2009; Kirschenbaum and Wintner, 2010)",
"1,373 entities without no multi-sense ambiguities",
"and achieve promising results in English-Hebrew",
"fromthe names of people (Census,2000),coun-",
"transliteration using Moses (Koehn et al., 2007).",
"tries and capital cities (Wikipedia.org, 2014), re-",
"However, all these systems are supervised and re-",
"sulting in more than 70,000 pairs of proper name",
"quire extra linguistic background knowledge dur-",
"transliteration from English. Table 1 shows statis-",
"ing processing. Plus, only one among this work",
"tics of final data size in each language. 80% of",
"evaluates transliteration on up to 4 languages and",
"the final data will be used for training, 10% is for",
"it is hard to generalize for multiple languages.",
"tuning and the remaining 10% is for testing.",
"The other branch learns from only sequence of",
"characters. One of the great advantages against",
"Largest",
"Smallest",
"Lang",
"Count",
"Lang",
"Count",
"Khmer",
"sound based transliteration is that multilingual",
"French",
"183,270",
"1,585",
"texts are much easier to obtain. Al-Onaizan and",
"German",
"178.715",
"Amharic",
"2.035",
"132.545",
"2,130",
"Italian",
"Knight (2002) compares phonetic based systems",
"Gujarati",
"Polish",
"124,870",
"Maltese",
"2.415",
"withspelling basedsystemsontransliterations",
"Spanish",
"107,790",
"Yiddish",
"2,835",
"between English and Arabic.Pouliquen et al.",
"Russian",
"100,085",
"Kannada",
"3,100",
"Swedish",
"91,125",
"Telugu",
"3.840",
"Dutch",
"(2006) makes transliteration model based on simi-",
"87,870",
"Swahili",
"4.620",
"Portuguese",
"Haitian",
"5,245",
"86,515",
"of Durrani et al. (2014) is integrated in Moses as",
"Norwegian",
"74,790",
"Urdu",
"6,115",
"a module, providing an unsupervised character",
"Table l: Languages with the largest and smallest",
"Matthews",
"lar spelling rules in close languages. Recent work",
"based transliteration training model.",
"(2007) proposed a proper name transliteration",
"set of translated entities, i.e. reflecting the avail-",
"model on several language pairs.",
"ability of training data.",
"",
"However,",
"we believe utilizing character-based transliteration"
] |
{
"arxiv_id": "1611.06722",
"language": "en",
"page_idx": 2
}
|
||
1611.06722-en-3
|
1611.06722-en-3.jpg
|
### 4 Training Transliteration Model
The purpose of our training is to get a quantified measurement of sound similarities between any possible character strings in arbitrary scripts. We expect to learn pairwise word segmentations and n-gram statistics of correlated string pieces between different languages.
We maintain a cost matrix in which the cost of substituting any string $s_1$ in $Language_1$ with string $s_2$ in $Langauge_2$ will first be initialized to $len(s_1)+len(s_2)$,including empty strings. This way each training example has a fixed cost equal to the total length of two strings. After that we start an $R$ round iteration. In each round we go through all training examples and compute the minimum-cost segmentation matching. We keep tracking of all observations of matched n-grams during this round in observation table. We then adapt Bayesian setting mentioned in (Snyder and Barzilay, 2008) to update cost matrix according to probability calculated by observation table. Figure 2 illustrates the training procedure.
Data used in training example may be flawed as it might not reflect transliteration. Such training examples act as outliners during out training and we cannot find any reasonable matching even for partial string pieces. We here define “Dirtiness" to measure how many training examples are flawed in training a specific language. Table 2 shows 10 dirtiest among 69 languages. Big languages included in Table 2 (e.g. Chinese, Korean) are not problematic since we have plenty of training examples. However, we expect a bad performance on small languages like Khmer and Amharic due to lack of high quality data.
<table>
<tbody>
<tr>
<th colspan="2">Dirtiest</th>
<th colspan="2">${\mathrm{Cleanest}}$</th>
</tr>
<tr>
<th>Lang</th>
<th>Dirtiness</th>
<th>Lang</th>
<th>Dirtiness</th>
</tr>
<tr>
<td>Hungarian</td>
<td>$41.00\%$</td>
<td>Norwegian</td>
<td>$1.23\%$</td>
</tr>
<tr>
<td>Amharic</td>
<td>$36.09\%$</td>
<td>$\operatorname{Bulgarian}$</td>
<td>$1.80\%$</td>
</tr>
<tr>
<td>Vietnamese</td>
<td>$32.10\%$</td>
<td>$\mathbf{Macedonian}$</td>
<td>$1.83\%$</td>
</tr>
<tr>
<td>Khmer</td>
<td>$24.58\%$</td>
<td>Latvian</td>
<td>$1.99\%$</td>
</tr>
<tr>
<td>Thai</td>
<td>$24.50\%$</td>
<td>Russian</td>
<td>$2.20\%$</td>
</tr>
<tr>
<td>Chinese</td>
<td>$23.80\%$</td>
<td>Greek</td>
<td>$2.27\%$</td>
</tr>
<tr>
<td>Korean</td>
<td>$22.20\%$</td>
<td>Armenian</td>
<td>$2.41\%$</td>
</tr>
<tr>
<td>Malay</td>
<td>$20.11\%$</td>
<td>Georgian</td>
<td>$2.60\%$</td>
</tr>
<tr>
<td>Tamil</td>
<td>$18.72\%$</td>
<td>$\mathbf{C}_{\mathrm{Zech}}$</td>
<td>$2.95\%$</td>
</tr>
<tr>
<td>Japanese</td>
<td>$16.81\%$</td>
<td>Latin</td>
<td>$3.26\%$</td>
</tr>
</tbody>
</table>
Table 2: The 10 cleanest and dirtiest languages, defined according to the ratio of flawed examples in the training set.
Figure 3a and Fig. 3b present heatmaps generated from our cost matrix showing 1-1 match-
ing rules between characters in these languages. The highlighted diagonals indicate strong similarity between identical Latin characters as expected, making transliteration inside the same language family meaningless. However, these matrices also reflect language differences: e.g. the Spanish “y” more often acts as an English “j”than an English “y”; and the Spanish “b”is close to English“v” while the French“b” matches only with English $“b”.$
### 5 Experimental Results
In this section we evaluate the quality of our
transliteration model.
## 5.1 Baseline Model
We use the transliteration system described in (Durrani et al., 2014) as baseline method to compare our results. The Moses statistical machine translation system integrates their work as a module, and allows training unsupervised transliteration for OOV words. $^{2}$
## 5.2 Test Results
We first compare both systems trained on our Wikipedia dataset. We focus on the performance of phrase table, i.e. measurement of sound similarities between string pieces since our dataset does not contain corpus context, We generate the 100-best transliterations for entries in testing set on four languages of different language families: Chinese, Arabic, Hindi and Russian.
We repeated this test using third party datasets
to check consistency of training models.$^{3}$
We cleaned data and retained only name mapping to feed the model, since our model does not rely on context and target a generalized method for multiple languages. Note Moses provides several language-specific optimization methods, including weights optimizing (e.g. Mert) and Language Model Smoothing (e.g. Kneser-Ney) that
$^{2}$We use the following parameters when configuring Moses: maximum phrase length=3, language model $N-gramorder=3,languagemodelsmoothing\&inter-$ $polation=$Automatically Disabled, Interpolate; alignment $heuristic=$grow-diag-final; reordering=Monotone; and max- $imum\textit{distortion length= 0. The weights for the models are: }$ translation model (0.2,0.2,0.2,0.2,0.2,0.2), language model (0.5), and distortion Model (0.0),with word penalty=-1.
$^{3}Chinese:$ Chinese -English Name EntityList sv1.0(LDC2005T34).Arabic: $\bar{\text{Combination of }10001}$ Arabic Names(LDC2005G02) and (Cettolo et al., 2012) made available for IWSLT-13. $Hindi:$ Indian multi-parallel corpus (Post et al., 2012), and Russian: WMT-13 data (Bojar et al., 2014).
|
[
[
0.118408203125,
0.0748291015625,
0.418212890625,
0.0748291015625,
0.418212890625,
0.0885009765625,
0.118408203125,
0.0885009765625
],
[
0.51318359375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.51318359375,
0.0914306640625,
0.8818359375,
0.0926513671875,
0.8818359375,
0.10687255859375,
0.51318359375,
0.105712890625
],
[
0.118408203125,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.5146484375,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.12005615234375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.12890625,
0.12005615234375,
0.12890625
],
[
0.51318359375,
0.12353515625,
0.884765625,
0.12469482421875,
0.884765625,
0.1395263671875,
0.51318359375,
0.1383056640625
],
[
0.118408203125,
0.1341552734375,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.118408203125,
0.1478271484375
],
[
0.5146484375,
0.141357421875,
0.884765625,
0.141357421875,
0.884765625,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.118408203125,
0.1478271484375,
0.489501953125,
0.1484375,
0.489501953125,
0.163330078125,
0.118408203125,
0.162109375
],
[
0.51123046875,
0.1561279296875,
0.88330078125,
0.155029296875,
0.88330078125,
0.169189453125,
0.51123046875,
0.17041015625
],
[
0.12005615234375,
0.1650390625,
0.48779296875,
0.1650390625,
0.48779296875,
0.1787109375,
0.12005615234375,
0.1787109375
],
[
0.51123046875,
0.1728515625,
0.884765625,
0.171630859375,
0.884765625,
0.1859130859375,
0.51123046875,
0.18701171875
],
[
0.1175537109375,
0.1817626953125,
0.313232421875,
0.1817626953125,
0.313232421875,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51123046875,
0.188232421875,
0.88330078125,
0.1859130859375,
0.88330078125,
0.20068359375,
0.51123046875,
0.203125
],
[
0.1351318359375,
0.1953125,
0.490234375,
0.196533203125,
0.490234375,
0.21142578125,
0.1351318359375,
0.210205078125
],
[
0.51123046875,
0.204833984375,
0.88330078125,
0.204833984375,
0.88330078125,
0.218505859375,
0.51123046875,
0.218505859375
],
[
0.118408203125,
0.213134765625,
0.489501953125,
0.213134765625,
0.489501953125,
0.22802734375,
0.118408203125,
0.22802734375
],
[
0.51318359375,
0.218505859375,
0.54833984375,
0.22265625,
0.5458984375,
0.235107421875,
0.509765625,
0.23095703125
],
[
0.118408203125,
0.2298583984375,
0.490234375,
0.2298583984375,
0.490234375,
0.24462890625,
0.118408203125,
0.24462890625
],
[
0.118408203125,
0.2464599609375,
0.489501953125,
0.2464599609375,
0.489501953125,
0.260009765625,
0.118408203125,
0.260009765625
],
[
0.515625,
0.2493896484375,
0.52880859375,
0.2493896484375,
0.52880859375,
0.25830078125,
0.515625,
0.25830078125
],
[
0.5263671875,
0.25048828125,
0.54248046875,
0.25048828125,
0.54248046875,
0.25830078125,
0.5263671875,
0.25830078125
],
[
0.5380859375,
0.2476806640625,
0.7314453125,
0.2476806640625,
0.7314453125,
0.262451171875,
0.5380859375,
0.262451171875
],
[
0.1175537109375,
0.262451171875,
0.48779296875,
0.262451171875,
0.48779296875,
0.275634765625,
0.1175537109375,
0.275634765625
],
[
0.118408203125,
0.27783203125,
0.489501953125,
0.27783203125,
0.489501953125,
0.292236328125,
0.118408203125,
0.292236328125
],
[
0.51318359375,
0.2724609375,
0.88330078125,
0.2724609375,
0.88330078125,
0.2861328125,
0.51318359375,
0.2861328125
],
[
0.1175537109375,
0.29345703125,
0.48779296875,
0.29345703125,
0.48779296875,
0.30712890625,
0.1175537109375,
0.30712890625
],
[
0.51318359375,
0.289306640625,
0.67333984375,
0.289306640625,
0.67333984375,
0.302734375,
0.51318359375,
0.302734375
],
[
0.12176513671875,
0.312255859375,
0.48779296875,
0.312255859375,
0.48779296875,
0.322509765625,
0.12176513671875,
0.322509765625
],
[
0.51318359375,
0.314208984375,
0.67822265625,
0.314208984375,
0.67822265625,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.118408203125,
0.325927734375,
0.489501953125,
0.325927734375,
0.489501953125,
0.34033203125,
0.118408203125,
0.34033203125
],
[
0.51123046875,
0.334228515625,
0.8818359375,
0.334228515625,
0.8818359375,
0.347412109375,
0.51123046875,
0.347412109375
],
[
0.118408203125,
0.342529296875,
0.489501953125,
0.342529296875,
0.489501953125,
0.355712890625,
0.118408203125,
0.355712890625
],
[
0.51318359375,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.118408203125,
0.358154296875,
0.489501953125,
0.358154296875,
0.489501953125,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51123046875,
0.36767578125,
0.884765625,
0.366455078125,
0.884765625,
0.380126953125,
0.51123046875,
0.38134765625
],
[
0.118408203125,
0.374755859375,
0.489501953125,
0.37353515625,
0.489501953125,
0.38720703125,
0.118408203125,
0.388427734375
],
[
0.118408203125,
0.3896484375,
0.489501953125,
0.3896484375,
0.489501953125,
0.402587890625,
0.118408203125,
0.402587890625
],
[
0.5146484375,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.3955078125,
0.5146484375,
0.3955078125
],
[
0.51318359375,
0.399658203125,
0.88330078125,
0.399658203125,
0.88330078125,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.118408203125,
0.40673828125,
0.490234375,
0.40673828125,
0.490234375,
0.42041015625,
0.118408203125,
0.42041015625
],
[
0.51123046875,
0.41455078125,
0.68408203125,
0.413330078125,
0.68408203125,
0.427490234375,
0.51123046875,
0.4287109375
],
[
0.1175537109375,
0.421630859375,
0.3779296875,
0.4228515625,
0.3779296875,
0.43701171875,
0.1175537109375,
0.435791015625
],
[
0.1368408203125,
0.438232421875,
0.490234375,
0.438232421875,
0.490234375,
0.453125,
0.1368408203125,
0.453125
],
[
0.5146484375,
0.440673828125,
0.6533203125,
0.440673828125,
0.6533203125,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.1175537109375,
0.45361328125,
0.489501953125,
0.454833984375,
0.489501953125,
0.4697265625,
0.1175537109375,
0.468505859375
],
[
0.5146484375,
0.46142578125,
0.88671875,
0.46142578125,
0.88671875,
0.4755859375,
0.5146484375,
0.4755859375
],
[
0.118408203125,
0.471435546875,
0.489501953125,
0.471435546875,
0.489501953125,
0.485107421875,
0.118408203125,
0.485107421875
],
[
0.5146484375,
0.47802734375,
0.884765625,
0.47802734375,
0.884765625,
0.4912109375,
0.5146484375,
0.4912109375
],
[
0.1175537109375,
0.48681640625,
0.489501953125,
0.48681640625,
0.489501953125,
0.50048828125,
0.1175537109375,
0.50048828125
],
[
0.51318359375,
0.49462890625,
0.8818359375,
0.49462890625,
0.8818359375,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.1175537109375,
0.5029296875,
0.489501953125,
0.501953125,
0.489501953125,
0.5166015625,
0.1175537109375,
0.51708984375
],
[
0.51318359375,
0.51025390625,
0.88330078125,
0.51025390625,
0.88330078125,
0.52490234375,
0.51318359375,
0.52490234375
],
[
0.118408203125,
0.51953125,
0.489501953125,
0.51953125,
0.489501953125,
0.533203125,
0.118408203125,
0.533203125
],
[
0.51318359375,
0.52685546875,
0.88330078125,
0.52685546875,
0.88330078125,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.118408203125,
0.53515625,
0.489501953125,
0.53515625,
0.489501953125,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.515625,
0.5419921875,
0.88330078125,
0.5419921875,
0.88330078125,
0.5556640625,
0.515625,
0.5556640625
],
[
0.1175537109375,
0.55029296875,
0.486083984375,
0.55029296875,
0.486083984375,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.5146484375,
0.55810546875,
0.884765625,
0.55810546875,
0.884765625,
0.572265625,
0.5146484375,
0.572265625
],
[
0.12005615234375,
0.568359375,
0.489501953125,
0.568359375,
0.489501953125,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.5146484375,
0.57373046875,
0.7841796875,
0.57373046875,
0.7841796875,
0.58740234375,
0.5146484375,
0.58740234375
],
[
0.12005615234375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.53076171875,
0.5888671875,
0.88330078125,
0.5888671875,
0.88330078125,
0.6025390625,
0.53076171875,
0.6025390625
],
[
0.118408203125,
0.599609375,
0.489501953125,
0.599609375,
0.489501953125,
0.61328125,
0.118408203125,
0.61328125
],
[
0.51318359375,
0.60693359375,
0.8134765625,
0.60693359375,
0.8134765625,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.12176513671875,
0.61767578125,
0.48779296875,
0.61767578125,
0.48779296875,
0.62744140625,
0.12176513671875,
0.62744140625
],
[
0.53076171875,
0.62060546875,
0.8818359375,
0.62255859375,
0.8818359375,
0.63720703125,
0.53076171875,
0.63525390625
],
[
0.1175537109375,
0.63134765625,
0.32080078125,
0.63134765625,
0.32080078125,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.51318359375,
0.63818359375,
0.884765625,
0.63720703125,
0.884765625,
0.65185546875,
0.51318359375,
0.65283203125
],
[
0.193115234375,
0.6591796875,
0.2418212890625,
0.6591796875,
0.2418212890625,
0.67041015625,
0.193115234375,
0.67041015625
],
[
0.36279296875,
0.6591796875,
0.420654296875,
0.6591796875,
0.420654296875,
0.67041015625,
0.36279296875,
0.67041015625
],
[
0.51123046875,
0.65478515625,
0.884765625,
0.65380859375,
0.884765625,
0.66845703125,
0.51123046875,
0.66943359375
],
[
0.35693359375,
0.66943359375,
0.3955078125,
0.67333984375,
0.39208984375,
0.68408203125,
0.353515625,
0.6796875
],
[
0.51123046875,
0.67041015625,
0.88330078125,
0.67041015625,
0.88330078125,
0.68408203125,
0.51123046875,
0.68408203125
],
[
0.2283935546875,
0.67138671875,
0.29296875,
0.67138671875,
0.29296875,
0.681640625,
0.2283935546875,
0.681640625
],
[
0.1822509765625,
0.67138671875,
0.219970703125,
0.67138671875,
0.219970703125,
0.68310546875,
0.1822509765625,
0.68310546875
],
[
0.4013671875,
0.67138671875,
0.47021484375,
0.67138671875,
0.47021484375,
0.681640625,
0.4013671875,
0.681640625
],
[
0.1502685546875,
0.681640625,
0.21826171875,
0.68408203125,
0.21826171875,
0.6953125,
0.1502685546875,
0.693359375
],
[
0.2418212890625,
0.68310546875,
0.29296875,
0.68310546875,
0.29296875,
0.69482421875,
0.2418212890625,
0.69482421875
],
[
0.322509765625,
0.68408203125,
0.39208984375,
0.68408203125,
0.39208984375,
0.6953125,
0.322509765625,
0.6953125
],
[
0.426513671875,
0.68310546875,
0.47021484375,
0.68310546875,
0.47021484375,
0.69482421875,
0.426513671875,
0.69482421875
],
[
0.5146484375,
0.68701171875,
0.88330078125,
0.68701171875,
0.88330078125,
0.70166015625,
0.5146484375,
0.70166015625
],
[
0.2418212890625,
0.69482421875,
0.29296875,
0.69482421875,
0.29296875,
0.7060546875,
0.2418212890625,
0.7060546875
],
[
0.32666015625,
0.69482421875,
0.393798828125,
0.69482421875,
0.393798828125,
0.7060546875,
0.32666015625,
0.7060546875
],
[
0.4248046875,
0.69482421875,
0.466796875,
0.69482421875,
0.466796875,
0.7060546875,
0.4248046875,
0.7060546875
],
[
0.5146484375,
0.70361328125,
0.884765625,
0.70361328125,
0.884765625,
0.71728515625,
0.5146484375,
0.71728515625
],
[
0.162109375,
0.6953125,
0.21826171875,
0.6953125,
0.21826171875,
0.7060546875,
0.162109375,
0.7060546875
],
[
0.1427001953125,
0.705078125,
0.21826171875,
0.70703125,
0.21826171875,
0.71875,
0.1427001953125,
0.71630859375
],
[
0.2418212890625,
0.7060546875,
0.291259765625,
0.7060546875,
0.291259765625,
0.71728515625,
0.2418212890625,
0.71728515625
],
[
0.312255859375,
0.70703125,
0.393798828125,
0.70703125,
0.393798828125,
0.71875,
0.312255859375,
0.71875
],
[
0.4248046875,
0.7060546875,
0.466796875,
0.7060546875,
0.466796875,
0.71728515625,
0.4248046875,
0.71728515625
],
[
0.172119140625,
0.71875,
0.219970703125,
0.71875,
0.219970703125,
0.72998046875,
0.172119140625,
0.72998046875
],
[
0.2418212890625,
0.71630859375,
0.29296875,
0.71875,
0.291259765625,
0.73095703125,
0.240966796875,
0.728515625
],
[
0.341796875,
0.71875,
0.393798828125,
0.71875,
0.393798828125,
0.72998046875,
0.341796875,
0.72998046875
],
[
0.4248046875,
0.71875,
0.466796875,
0.71875,
0.466796875,
0.72998046875,
0.4248046875,
0.72998046875
],
[
0.5146484375,
0.7197265625,
0.884765625,
0.7197265625,
0.884765625,
0.73388671875,
0.5146484375,
0.73388671875
],
[
0.2418212890625,
0.72802734375,
0.291259765625,
0.72998046875,
0.28955078125,
0.7421875,
0.240966796875,
0.74072265625
],
[
0.339111328125,
0.72802734375,
0.397216796875,
0.72998046875,
0.3955078125,
0.74462890625,
0.337646484375,
0.7421875
],
[
0.4248046875,
0.72802734375,
0.466796875,
0.72998046875,
0.465087890625,
0.7421875,
0.424072265625,
0.74072265625
],
[
0.5322265625,
0.73828125,
0.88330078125,
0.7392578125,
0.88330078125,
0.75439453125,
0.5322265625,
0.7529296875
],
[
0.185546875,
0.72998046875,
0.219970703125,
0.72998046875,
0.219970703125,
0.7421875,
0.185546875,
0.7421875
],
[
0.2418212890625,
0.74072265625,
0.29296875,
0.7421875,
0.291259765625,
0.7548828125,
0.240966796875,
0.7529296875
],
[
0.4248046875,
0.7392578125,
0.466796875,
0.7412109375,
0.465087890625,
0.75439453125,
0.424072265625,
0.751953125
],
[
0.35107421875,
0.7421875,
0.393798828125,
0.7421875,
0.393798828125,
0.75439453125,
0.35107421875,
0.75439453125
],
[
0.166259765625,
0.7421875,
0.21826171875,
0.7421875,
0.21826171875,
0.75439453125,
0.166259765625,
0.75439453125
],
[
0.17041015625,
0.75439453125,
0.219970703125,
0.75439453125,
0.219970703125,
0.765625,
0.17041015625,
0.765625
],
[
0.2418212890625,
0.75439453125,
0.291259765625,
0.75439453125,
0.291259765625,
0.765625,
0.2418212890625,
0.765625
],
[
0.4248046875,
0.75439453125,
0.466796875,
0.75439453125,
0.466796875,
0.765625,
0.4248046875,
0.765625
],
[
0.51318359375,
0.7529296875,
0.884765625,
0.7529296875,
0.884765625,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.2418212890625,
0.76318359375,
0.29296875,
0.765625,
0.291259765625,
0.779296875,
0.240966796875,
0.77685546875
],
[
0.330078125,
0.7548828125,
0.393798828125,
0.7548828125,
0.393798828125,
0.765625,
0.330078125,
0.765625
],
[
0.173828125,
0.7666015625,
0.21826171875,
0.7666015625,
0.21826171875,
0.77783203125,
0.173828125,
0.77783203125
],
[
0.33154296875,
0.7666015625,
0.3955078125,
0.7666015625,
0.3955078125,
0.77783203125,
0.33154296875,
0.77783203125
],
[
0.424072265625,
0.76416015625,
0.468505859375,
0.76416015625,
0.468505859375,
0.779296875,
0.424072265625,
0.779296875
],
[
0.5146484375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77685546875,
0.5146484375,
0.77685546875
],
[
0.345947265625,
0.77685546875,
0.397216796875,
0.77685546875,
0.397216796875,
0.79150390625,
0.345947265625,
0.79150390625
],
[
0.42236328125,
0.775390625,
0.47021484375,
0.775390625,
0.47021484375,
0.79052734375,
0.42236328125,
0.79052734375
],
[
0.51318359375,
0.77685546875,
0.884765625,
0.77685546875,
0.884765625,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.2435302734375,
0.77783203125,
0.29296875,
0.77783203125,
0.29296875,
0.7890625,
0.2435302734375,
0.7890625
],
[
0.177978515625,
0.77783203125,
0.219970703125,
0.77783203125,
0.219970703125,
0.79150390625,
0.177978515625,
0.79150390625
],
[
0.353515625,
0.78759765625,
0.3955078125,
0.7890625,
0.393798828125,
0.8017578125,
0.3525390625,
0.7998046875
],
[
0.424072265625,
0.78759765625,
0.47021484375,
0.78759765625,
0.47021484375,
0.8017578125,
0.424072265625,
0.8017578125
],
[
0.51123046875,
0.78759765625,
0.88330078125,
0.7880859375,
0.88330078125,
0.8017578125,
0.51123046875,
0.80029296875
],
[
0.2435302734375,
0.7890625,
0.2939453125,
0.7890625,
0.2939453125,
0.80029296875,
0.2435302734375,
0.80029296875
],
[
0.162109375,
0.79052734375,
0.2166748046875,
0.79052734375,
0.2166748046875,
0.8017578125,
0.162109375,
0.8017578125
],
[
0.51318359375,
0.7998046875,
0.88330078125,
0.7998046875,
0.88330078125,
0.8134765625,
0.51318359375,
0.8134765625
],
[
0.1175537109375,
0.81396484375,
0.489501953125,
0.8154296875,
0.489501953125,
0.830078125,
0.1175537109375,
0.8291015625
],
[
0.51318359375,
0.8125,
0.884765625,
0.8125,
0.884765625,
0.826171875,
0.51318359375,
0.826171875
],
[
0.1175537109375,
0.830078125,
0.490234375,
0.83056640625,
0.490234375,
0.845703125,
0.1175537109375,
0.84423828125
],
[
0.51318359375,
0.826171875,
0.8505859375,
0.826171875,
0.8505859375,
0.8359375,
0.51318359375,
0.8359375
],
[
0.53564453125,
0.8359375,
0.599609375,
0.83837890625,
0.599609375,
0.849609375,
0.53564453125,
0.84716796875
],
[
0.615234375,
0.83837890625,
0.67333984375,
0.83837890625,
0.67333984375,
0.8486328125,
0.615234375,
0.8486328125
],
[
0.6943359375,
0.8359375,
0.884765625,
0.8359375,
0.884765625,
0.849609375,
0.6943359375,
0.849609375
],
[
0.1158447265625,
0.8466796875,
0.25439453125,
0.84716796875,
0.25439453125,
0.8623046875,
0.1158447265625,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.51318359375,
0.85986328125,
0.884765625,
0.85986328125,
0.884765625,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.51318359375,
0.87158203125,
0.884765625,
0.87158203125,
0.884765625,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.1368408203125,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.1368408203125,
0.8935546875
],
[
0.51123046875,
0.88427734375,
0.88330078125,
0.88427734375,
0.88330078125,
0.89794921875,
0.51123046875,
0.89794921875
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.89453125,
0.5556640625,
0.896484375,
0.55517578125,
0.9091796875,
0.51318359375,
0.90673828125
]
] |
[
"4Training Transliteration Model",
"ing rules between characters in these languages.",
"The highlighted diagonals indicate strong similar-",
"The purpose of our training is to get a quanti-",
"ity between identical Latin characters as expected,",
"fied measurement of sound similarities between",
"making transliteration inside the same language",
"any possible character strings in arbitrary scripts.",
"family meaningless. However, these matrices also",
"We expect to learn pairwise word segmentations",
"reflect language differences: e.g. the Spanish \"y\"",
"and n-gram statistics of correlated string pieces be-",
"more often acts as an English j\" than an English",
"tween different languages.",
"“y\"; and the Spanish “b\" is close to English \"y\"",
"We maintain a cost matrix in which the cost",
"while the French“b\"matches only with English",
"of substituting any string si in Language1 with",
"*\"b\".",
"string s2 in Langauge2 will first be initialized to",
"len(si) + len(s2), including empty strings. This",
"5",
"",
"Experimental Results",
"way each training example has a fixed cost equal",
"to the total length of two strings. After that we",
"In this section we evaluate the quality of our",
"start an R round iteration. In each round we go",
"transliteration model.",
"through alltraining examples and compute the",
"5.1Baseline Model",
"minimum-cost segmentation matching. We keep",
"We use the transliteration system described in",
"tracking of all observations of matched n-grams",
"(Durrani et al., 2014) as baseline method to com-",
"during this round in observation table.We then",
"pare our results. The Moses statistical machine",
"adapt Bayesian setting mentioned in (Snyder and",
"Barzilay,2008) to update cost matrix according to",
"translation system integrates theirwork as a mod-",
"ule, and allows training unsupervised translitera-",
"probability calculated by observation table. Figure",
"tion for OoV words. 2",
"2 illustrates the training procedure.",
"Data used in training example may be flawed as",
"5.2Test Results",
"it might not reflect transliteration. Such training",
"We first compare both systems trained on our",
"examples act as outliners during out training and",
"Wikipedia dataset. We focus on the performance",
"we cannot find any reasonable matching even for",
"of phrase table, i.e. measurement of sound sim-",
"partial string pieces. We here define\"Dirtiness\"to",
"ilarities between string pieces since our dataset",
"measure how many training examples are flawed",
"does not contain corpus context, We generate the",
"in training a specific language. Table 2 shows 10",
"100-best transliterations for entries in testing set",
"dirtiest among 69 languages. Big languages in-",
"on four languages of different language families:",
"cluded in Table 2 (e.g. Chinese, Korean) are not",
"Chinese, Arabic, Hindi and Russian.",
"problematic since we have plenty of training ex-",
"We repeated this test using third party datasets",
"amples. However, we expect a bad performance",
"to check consistency of training models.",
"onsmalllanguageslikeKhmer and Amharicdue",
"We cleaned data and retained only name map-",
"to lack of high quality data.",
"ping to feed the model, since our model does not",
"Dirtiest",
"Cleanest",
"rely on context and target a generalized method",
"Lang",
"for multiple languages. Note Moses provides sev-",
"Dirtiness",
"Lang",
"Dirtiness",
"Hungarian",
"41.00%",
"Norwegian",
"1.23%",
"eral language-specific optimization methods, in-",
"36.09%",
"Bulgarian",
"1.80%",
"cluding weights optimizing (e.g. Mert) and Lan-",
"Amharic",
"Vietnamese",
"32.10%",
"Macedonian",
"1.83%",
"Khmer",
"24.58%",
"Latvian",
"1.99%",
"guage Model Smoothing (e.g. Kneser-Ney) that",
"24.50%",
"Russian",
"2.20%",
"2we use the following parameters when configur-",
"Thai",
"23.80%",
"2.27%",
"Greek",
"Chinese",
"Korean",
"22.20%",
"2.41%",
"ing Moses: maximum phrase length=3, language model",
"20.11%",
"Armenian",
"Malay",
"Georgian",
"2.60%",
"N-gram order=3,languagemodel smoothing&inter-",
"Czech",
"2.95%",
"polation=Automatically Disabled, Interpolate; alignment",
"18.72%",
"Tamil",
"Latin",
"3.26%",
"heuristic=grow-diag-final; reordering=Monotone; and max-",
"16.81%",
"Japanese",
"imum distortion length=0. The weights for the models are:",
"Table 2: The 10 cleanest and dirtiest languages,",
"translation model (0.2, 0.2, 0.2, 0.2, 0.2), language model",
"defined according to the ratio of flawed examples",
"(0.5),and distortion Model (0.0),with word penalty=-1",
"3Chinese:",
"Chinese",
"English Name Entity List",
"in the training set.",
"sv1.0(LDC2005T34). Arabic:Combination of 10001 Ara-",
"bic Names(LDC2005G02) and (Cettolo et al., 2012) made",
"available for IWSLT-13. Hindi: Indian multi-parallel corpus",
"Figure 3a and Fig. 3b present heatmaps gen-",
"(Post et al., 2012), and Russian: WMT-13 data (Bojar et al.,",
"erated from our cost matrix showing 1-1 match-",
"2014)."
] |
{
"arxiv_id": "1611.06722",
"language": "en",
"page_idx": 3
}
|
||
2311.01882-en-0
|
2311.01882-en-0.jpg
|
Indicative Summarization of Long Discussions
Shahbaz Syed $^{+ * }$ $\textbf{Dominik Schwabe}^{+ * }$ $\textbf{Khalid Al- Khatib}^{+ }$ $\textbf{Martin Potthast}^{+ \text{if}* }$
$^{8}$ScaDS.AI
$^{\dagger}$Leipzig University
tUniversity of Groningen
shahbaz. syed@uni-leipzig. de
Abstract
Online forums encourage the exchange and discussion of different stances on many topics. Not only do they provide an opportunity to present one's own arguments, but may also gather a broad cross-section of others' arguments. However, the resulting long discussions are difficult to overview. This paper presents a novel unsupervised approach using large language models (LLMs) to generating indicative summaries for long discussions that basically serve as tables of contents. Our approach first clusters argument sentences, generates cluster labels as abstractive summaries, and classifies the generated cluster labels into argumentation frames resulting in a two-level summary. Based on an extensively optimized prompt engineering approach, we evaluate 19 LLMs for generative cluster labeling and frame classification. To evaluate the usefulness of our indicative summaries, we conduct a purpose-driven user study via a new visual interface called DISCUSSION EXPLORER: It shows that our proposed indicative summaries serve as a convenient navigation tool to explore long discussions.$^{1}$
exchanged, so it is still necessary to read most of them for a comprehensive overview. In this paper, we depart from previous approaches to summarizing long discussions by using indicative summaries instead of informative summaries.$^3$ Figure 1 illustrates our three-step approach: first, the sentences of the arguments are clustered according to their latent subtopics. Then, a large language model generates a concise abstractive summary for each cluster as its label. Finally, the argument frame (Chong and Druckman, 2007; Boydstun et al., 2014) of each cluster label is predicted as a generalizable operationalization of perspectives on a discussion's topic. From this, a hierarchical summary is created in the style of a table of contents, where frames act as headings and cluster labels as subheadings. To our knowledge, indicative summaries of this type have not been explored before (see Section 2).
Our four main contributions are: (1) A fully unsupervised approach to indicative summarization of long discussions (Section 3). We develop robust prompts for generative cluster labeling and frame assignment based on extensive empirical evaluation and best practices (Section 4). (2) A comprehensive evaluation of 19 state-of-the-art, prompt-based, large language models (LLMs) for both tasks, supported by quantitative and qualitative assessments (Section 5). (3) A user study of the usefulness of indicative summaries for exploring long discussions (Section 5). (4) DIscussION EXPLORER, an interactive visual interface for exploring the indicative summaries generated by our approach and the corresponding discussions.$^{4}$ Our results show that the GPT variants of OpenAI (GPT3.5, ChatGPT, and GPT4) outperform all other open source models at the time of writing. LLaMA and T0 perform well, but are not competitive with the GPT models. Regarding the usefulness of the summaries, users preferred our summaries to alternative views to explore long discussions with hundreds of arguments.
## 1 Introduction
Online discussion forums are a popular medium for discussing a wide range of topics. As the size of a community grows, so does the average length of the discussions held there, especially when current controversial topics are discussed. On ChangeMyView (CMV),$^2$for example, discussions often go into the hundreds of arguments covering many perspectives on the topics in question. Initiating participating in, or reading discussions generally has two goals: to learn more about others' views on a topic and/or to share one's own.
To help their users navigate large volumes of arguments in long discussions, many forums offer basic features to sort them, for example, by time of creation or popularity. However, these alternative views may not capture the full range of perspectives
$\overline{^3\text{Unlike an informative summary, an indicative summary does}}$ not capture as much information as possible from a text, but only its gist. This makes them particularly suitable for long documents like books in the form of tables of contents.
$^{4}$https://discussion-explorer.web.webis.de/
${\overline{*}}{\mathrm{Equal~contribution.}}$
$^{\mathrm{l}}$Code: https://github.com/webis-de/EMNLP-23
$^{2}$https://www.reddit.com/r/changemyview/
|
[
[
0.26025390625,
0.0872802734375,
0.73779296875,
0.0885009765625,
0.73779296875,
0.10394287109375,
0.26025390625,
0.10272216796875
],
[
0.1502685546875,
0.1175537109375,
0.85205078125,
0.11517333984375,
0.85205078125,
0.133056640625,
0.1502685546875,
0.1353759765625
],
[
0.2283935546875,
0.1436767578125,
0.39208984375,
0.1455078125,
0.39208984375,
0.163330078125,
0.2283935546875,
0.160888671875
],
[
0.429931640625,
0.144287109375,
0.640625,
0.146728515625,
0.640625,
0.160888671875,
0.429931640625,
0.1591796875
],
[
0.67822265625,
0.142578125,
0.77685546875,
0.144287109375,
0.77685546875,
0.162109375,
0.67822265625,
0.1602783203125
],
[
0.39794921875,
0.1644287109375,
0.60546875,
0.1644287109375,
0.60546875,
0.1776123046875,
0.39794921875,
0.1776123046875
],
[
0.2626953125,
0.203125,
0.343505859375,
0.203125,
0.343505859375,
0.21728515625,
0.2626953125,
0.21728515625
],
[
0.51123046875,
0.2037353515625,
0.880859375,
0.2037353515625,
0.880859375,
0.21728515625,
0.51123046875,
0.21728515625
],
[
0.51123046875,
0.2203369140625,
0.8818359375,
0.2203369140625,
0.8818359375,
0.2340087890625,
0.51123046875,
0.2340087890625
],
[
0.1435546875,
0.2286376953125,
0.459228515625,
0.2298583984375,
0.459228515625,
0.24462890625,
0.1435546875,
0.243408203125
],
[
0.51123046875,
0.236328125,
0.8818359375,
0.236328125,
0.8818359375,
0.2493896484375,
0.51123046875,
0.2493896484375
],
[
0.1435546875,
0.243408203125,
0.45751953125,
0.24462890625,
0.45751953125,
0.25830078125,
0.1435546875,
0.257080078125
],
[
0.51123046875,
0.2529296875,
0.8818359375,
0.2529296875,
0.8818359375,
0.2666015625,
0.51123046875,
0.2666015625
],
[
0.145263671875,
0.2587890625,
0.45751953125,
0.2587890625,
0.45751953125,
0.2724609375,
0.145263671875,
0.2724609375
],
[
0.145263671875,
0.273681640625,
0.45751953125,
0.273681640625,
0.45751953125,
0.287353515625,
0.145263671875,
0.287353515625
],
[
0.51123046875,
0.268310546875,
0.8818359375,
0.268310546875,
0.8818359375,
0.281982421875,
0.51123046875,
0.281982421875
],
[
0.145263671875,
0.287353515625,
0.459228515625,
0.287353515625,
0.459228515625,
0.300537109375,
0.145263671875,
0.300537109375
],
[
0.51123046875,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.145263671875,
0.300537109375,
0.459228515625,
0.300537109375,
0.459228515625,
0.314208984375,
0.145263671875,
0.314208984375
],
[
0.51123046875,
0.300537109375,
0.8818359375,
0.300537109375,
0.8818359375,
0.314208984375,
0.51123046875,
0.314208984375
],
[
0.0235137939453125,
0.308837890625,
0.060455322265625,
0.308837890625,
0.060455322265625,
0.43115234375,
0.0235137939453125,
0.43115234375
],
[
0.145263671875,
0.3154296875,
0.45751953125,
0.3154296875,
0.45751953125,
0.328857421875,
0.145263671875,
0.328857421875
],
[
0.51123046875,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.327880859375,
0.51123046875,
0.327880859375
],
[
0.145263671875,
0.330078125,
0.459228515625,
0.330078125,
0.459228515625,
0.34326171875,
0.145263671875,
0.34326171875
],
[
0.509765625,
0.333251953125,
0.8818359375,
0.33203125,
0.8818359375,
0.345703125,
0.509765625,
0.3466796875
],
[
0.1435546875,
0.344482421875,
0.45751953125,
0.34326171875,
0.45751953125,
0.35693359375,
0.1435546875,
0.358154296875
],
[
0.509765625,
0.347412109375,
0.8818359375,
0.3486328125,
0.8818359375,
0.363525390625,
0.509765625,
0.3623046875
],
[
0.145263671875,
0.358154296875,
0.45751953125,
0.358154296875,
0.45751953125,
0.371826171875,
0.145263671875,
0.371826171875
],
[
0.51123046875,
0.364013671875,
0.8818359375,
0.364013671875,
0.8818359375,
0.377685546875,
0.51123046875,
0.377685546875
],
[
0.145263671875,
0.372314453125,
0.459228515625,
0.372314453125,
0.459228515625,
0.385986328125,
0.145263671875,
0.385986328125
],
[
0.51318359375,
0.38134765625,
0.8818359375,
0.38134765625,
0.8818359375,
0.394287109375,
0.51318359375,
0.394287109375
],
[
0.145263671875,
0.38720703125,
0.459228515625,
0.38720703125,
0.459228515625,
0.40087890625,
0.145263671875,
0.40087890625
],
[
0.145263671875,
0.40087890625,
0.45751953125,
0.40087890625,
0.45751953125,
0.41455078125,
0.145263671875,
0.41455078125
],
[
0.51123046875,
0.39794921875,
0.8818359375,
0.39794921875,
0.8818359375,
0.410888671875,
0.51123046875,
0.410888671875
],
[
0.145263671875,
0.4150390625,
0.459228515625,
0.4150390625,
0.459228515625,
0.4287109375,
0.145263671875,
0.4287109375
],
[
0.509765625,
0.413330078125,
0.8818359375,
0.413330078125,
0.8818359375,
0.427001953125,
0.509765625,
0.427001953125
],
[
0.14697265625,
0.43115234375,
0.45751953125,
0.43115234375,
0.45751953125,
0.441162109375,
0.14697265625,
0.441162109375
],
[
0.509765625,
0.4287109375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.509765625,
0.4423828125
],
[
0.145263671875,
0.443603515625,
0.4609375,
0.443603515625,
0.4609375,
0.457275390625,
0.145263671875,
0.457275390625
],
[
0.0235137939453125,
0.43408203125,
0.0621337890625,
0.43408203125,
0.0621337890625,
0.51025390625,
0.0235137939453125,
0.51025390625
],
[
0.51123046875,
0.4453125,
0.8818359375,
0.4453125,
0.8818359375,
0.458984375,
0.51123046875,
0.458984375
],
[
0.145263671875,
0.457275390625,
0.459228515625,
0.457275390625,
0.459228515625,
0.47021484375,
0.145263671875,
0.47021484375
],
[
0.51123046875,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.51123046875,
0.4755859375
],
[
0.145263671875,
0.471435546875,
0.459228515625,
0.471435546875,
0.459228515625,
0.485107421875,
0.145263671875,
0.485107421875
],
[
0.509765625,
0.47802734375,
0.8486328125,
0.47802734375,
0.8486328125,
0.4912109375,
0.509765625,
0.4912109375
],
[
0.14697265625,
0.48681640625,
0.45751953125,
0.48681640625,
0.45751953125,
0.49755859375,
0.14697265625,
0.49755859375
],
[
0.53076171875,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.53076171875,
0.50732421875
],
[
0.1435546875,
0.49951171875,
0.459228515625,
0.498779296875,
0.459228515625,
0.51318359375,
0.1435546875,
0.51416015625
],
[
0.019317626953125,
0.51416015625,
0.060455322265625,
0.51416015625,
0.06378173828125,
0.7236328125,
0.0218353271484375,
0.7236328125
],
[
0.145263671875,
0.51416015625,
0.459228515625,
0.51416015625,
0.459228515625,
0.52783203125,
0.145263671875,
0.52783203125
],
[
0.51123046875,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.51123046875,
0.52392578125
],
[
0.14697265625,
0.529296875,
0.459228515625,
0.529296875,
0.459228515625,
0.5419921875,
0.14697265625,
0.5419921875
],
[
0.51123046875,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.51123046875,
0.5390625
],
[
0.14697265625,
0.54443359375,
0.45751953125,
0.54443359375,
0.45751953125,
0.5546875,
0.14697265625,
0.5546875
],
[
0.51318359375,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.51318359375,
0.5556640625
],
[
0.145263671875,
0.55712890625,
0.420654296875,
0.55712890625,
0.420654296875,
0.57080078125,
0.145263671875,
0.57080078125
],
[
0.51123046875,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51123046875,
0.57177734375
],
[
0.51318359375,
0.57373046875,
0.88330078125,
0.57373046875,
0.88330078125,
0.58740234375,
0.51318359375,
0.58740234375
],
[
0.12176513671875,
0.58837890625,
0.1318359375,
0.58837890625,
0.1318359375,
0.595703125,
0.12176513671875,
0.595703125
],
[
0.1435546875,
0.58447265625,
0.26025390625,
0.58447265625,
0.26025390625,
0.5986328125,
0.1435546875,
0.5986328125
],
[
0.51123046875,
0.59033203125,
0.88330078125,
0.59033203125,
0.88330078125,
0.60400390625,
0.51123046875,
0.60400390625
],
[
0.1175537109375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62353515625,
0.1175537109375,
0.62353515625
],
[
0.51123046875,
0.60693359375,
0.8818359375,
0.60693359375,
0.8818359375,
0.62060546875,
0.51123046875,
0.62060546875
],
[
0.1158447265625,
0.62451171875,
0.48779296875,
0.6259765625,
0.48779296875,
0.64013671875,
0.1158447265625,
0.6396484375
],
[
0.51123046875,
0.62255859375,
0.8818359375,
0.62255859375,
0.8818359375,
0.6357421875,
0.51123046875,
0.6357421875
],
[
0.1158447265625,
0.642578125,
0.48779296875,
0.642578125,
0.48779296875,
0.65625,
0.1158447265625,
0.65625
],
[
0.51123046875,
0.63818359375,
0.88330078125,
0.63818359375,
0.88330078125,
0.65185546875,
0.51123046875,
0.65185546875
],
[
0.1158447265625,
0.65673828125,
0.486083984375,
0.65673828125,
0.486083984375,
0.67041015625,
0.1158447265625,
0.67041015625
],
[
0.51123046875,
0.65478515625,
0.8818359375,
0.65478515625,
0.8818359375,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.1175537109375,
0.6748046875,
0.491943359375,
0.6748046875,
0.491943359375,
0.6884765625,
0.1175537109375,
0.6884765625
],
[
0.509765625,
0.67041015625,
0.884765625,
0.67041015625,
0.884765625,
0.68408203125,
0.509765625,
0.68408203125
],
[
0.1175537109375,
0.6904296875,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.1175537109375,
0.70361328125
],
[
0.509765625,
0.68701171875,
0.8818359375,
0.68701171875,
0.8818359375,
0.70068359375,
0.509765625,
0.70068359375
],
[
0.1175537109375,
0.70703125,
0.48779296875,
0.70703125,
0.48779296875,
0.72021484375,
0.1175537109375,
0.72021484375
],
[
0.5087890625,
0.70166015625,
0.880859375,
0.70166015625,
0.880859375,
0.71533203125,
0.5087890625,
0.71533203125
],
[
0.1158447265625,
0.72265625,
0.490234375,
0.7216796875,
0.490234375,
0.736328125,
0.1158447265625,
0.73681640625
],
[
0.509765625,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.509765625,
0.732421875
],
[
0.1175537109375,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.1175537109375,
0.7529296875
],
[
0.51318359375,
0.73388671875,
0.8818359375,
0.73388671875,
0.8818359375,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.1158447265625,
0.7548828125,
0.489501953125,
0.7548828125,
0.489501953125,
0.7685546875,
0.1158447265625,
0.7685546875
],
[
0.51318359375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.11419677734375,
0.76953125,
0.391357421875,
0.76953125,
0.391357421875,
0.783203125,
0.11419677734375,
0.783203125
],
[
0.51123046875,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.51123046875,
0.78076171875
],
[
0.1351318359375,
0.78759765625,
0.489501953125,
0.78759765625,
0.489501953125,
0.80029296875,
0.1351318359375,
0.80029296875
],
[
0.509765625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.509765625,
0.79638671875
],
[
0.1158447265625,
0.802734375,
0.489501953125,
0.8017578125,
0.489501953125,
0.81640625,
0.1158447265625,
0.81787109375
],
[
0.51318359375,
0.7998046875,
0.8818359375,
0.7998046875,
0.8818359375,
0.8134765625,
0.51318359375,
0.8134765625
],
[
0.118408203125,
0.82080078125,
0.486083984375,
0.82080078125,
0.486083984375,
0.83056640625,
0.118408203125,
0.83056640625
],
[
0.51123046875,
0.81640625,
0.88330078125,
0.81640625,
0.88330078125,
0.830078125,
0.51123046875,
0.830078125
],
[
0.1158447265625,
0.8349609375,
0.486083984375,
0.8349609375,
0.486083984375,
0.8486328125,
0.1158447265625,
0.8486328125
],
[
0.51318359375,
0.83203125,
0.88330078125,
0.83203125,
0.88330078125,
0.845703125,
0.51318359375,
0.845703125
],
[
0.1158447265625,
0.8515625,
0.48779296875,
0.8515625,
0.48779296875,
0.865234375,
0.1158447265625,
0.865234375
],
[
0.51123046875,
0.8525390625,
0.8818359375,
0.85400390625,
0.8818359375,
0.8681640625,
0.51123046875,
0.86767578125
],
[
0.1158447265625,
0.87255859375,
0.2435302734375,
0.87255859375,
0.2435302734375,
0.8857421875,
0.1158447265625,
0.8857421875
],
[
0.51708984375,
0.8681640625,
0.87890625,
0.8681640625,
0.87890625,
0.87890625,
0.51708984375,
0.87890625
],
[
0.5146484375,
0.87890625,
0.880859375,
0.87890625,
0.880859375,
0.892578125,
0.5146484375,
0.892578125
],
[
0.1158447265625,
0.88525390625,
0.411376953125,
0.8857421875,
0.411376953125,
0.8994140625,
0.1158447265625,
0.8984375
],
[
0.11419677734375,
0.89794921875,
0.3779296875,
0.8984375,
0.3779296875,
0.912109375,
0.11419677734375,
0.9111328125
],
[
0.51904296875,
0.892578125,
0.8505859375,
0.892578125,
0.8505859375,
0.90283203125,
0.51904296875,
0.90283203125
],
[
0.509765625,
0.90185546875,
0.771484375,
0.90283203125,
0.771484375,
0.91748046875,
0.509765625,
0.91650390625
]
] |
[
"Indicative Summarization of Long Discussions",
"Shahbaz Syed +*Dominik Schwabe +*Khalid Al-Khatib t Martin Potthast t$",
"tLeipzig University",
"University ofGroningen",
"$ScaDS.AI",
"shahbaz.syed@uni-leipzig.de",
"Abstract",
"exchanged, so it is still necessary to read most of",
"them for a comprehensive overview. In this paper.",
"Online forums encourage the exchange and",
"we depart from previous approaches to summariz-",
"discussion of different stances on many top",
"ing long discussions by using indicative summaries",
"ics. Not only do they provide an opportunity",
"to present one's own arguments, but may also",
"instead of informative summaries.3 Figure 1 illus-",
"gather a broad cross-section of others' argu-",
"trates our three-step approach: first, the sentences",
"ments. However, the resulting long discussions",
"of the arguments are clustered according to their",
"FZ AN",
"are difficult to overview. This paper presents",
"latent subtopics.Then.alargelanguagemodel gen",
"a novel unsupervised approach using large lan-",
"erates a concise abstractive summary for each clus-",
"guage models (LLMs) to generating indicative",
"ter as its label. Finally, the argument frame (Chong",
"summaries for long discussions that basically",
"and Druckman, 2007; Boydstun et al., 2014) of",
"serve as tables of contents. Our approach first",
"each cluster label is predicted as a generalizable",
"clusters argument sentences, generates cluster",
"labels as abstractive summaries, and classifies",
"operationalization of perspectives on a discussion's",
"the generated cluster labels into argumentation",
"topic. From this, a hierarchical summary is created",
"frames resulting ina two-level summary.Based",
"in the style of a table of contents, where frames act",
"on an extensively optimized prompt engineer-",
"['s]",
"as headings and cluster labels as subheadings. To",
"ing approach, we evaluate 19 LLMs for genera-",
"our knowledge, indicative summaries of this type",
"tive cluster labeling and frame classification",
"have not been explored before (see Section 2)",
"To evaluate the usefulness of our indicative",
"Our four main contributions are: (1) A fully un-",
"summaries, we conduct a purpose-driven user",
"17881017:1",
"study via a new visual interface called DISCUS-",
"supervised approach to indicative summarization",
"SION ExPLORER: It shows that our proposed",
"of long discussions (Section 3). We develop robust",
"indicativesummaries serve as a convenientnav-",
"prompts for generative cluster labeling and frame",
"igation tool to explore long discussions. ",
"assignment based on extensive empirical evaluation",
"and best practices (Section 4). (2) A comprehen-",
"1",
"Introduction",
"sive evaluation of 19 state-of-the-art, prompt-based,",
"Online discussion forums are a popular medium for",
"large language models (LLMs) for both tasks, sup-",
"discussing a wide range of topics. As the size of",
"ported by quantitative and qualitative assessments",
"a community grows, so does the average length of",
"(Section 5). (3) A user study of the usefulness of in-",
"the discussions held there, especially when current",
"dicative summaries for exploring long discussions",
"controversial topics are discussed. On Change-",
"(Section 5). (4) DISCUSSION EXPLORER, an inter-",
"MyView (CMV),for example, discussions often",
" active visual interface for exploring the indicative",
"go into the hundreds of arguments covering many",
" summaries generated by our approach and the cor",
"perspectives on the topics in question. Initiating,",
"responding discussions.4 Our results show that the",
"participating in, or reading discussions generally",
"GPT variants of OpenAI (GPT3.5, ChatGPT, and",
"has two goals: to learn more about others' views",
"GPT4) outperform all other open source models",
"on a topic and/or to share one's own.",
"at the time of writing. LLaMA and TO perform",
"To help their users navigate large volumes of",
"well, but are not competitive with the GPT models.",
"arguments in long discussions, many forums offer",
"Regarding the usefulness of the summaries, users",
"basicfeaturesto sortthem,forexample.bytimeof",
"preferred our summaries to alternative views to ex-",
"creation or popularity. However, these alternative",
"plore long discussions with hundreds of arguments.",
"views may not capture the full range of perspectives",
"3Unlike an informative summary,an indicative summary does",
"*Equal contribution.",
"not capture as much information as possiblefrom a text,but",
" only its gist. This makes them particularly suitable for long",
"'Code: https://github.com/webis-de/EMNLP-23",
"2https://www.reddit.com/r/changemyview/",
"documents like books in the form of tables of contents.",
"*https:/discussion-explorer.web.webis.de/"
] |
{
"arxiv_id": "2311.01882",
"language": "en",
"page_idx": 0
}
|
||
2311.01882-en-2
|
2311.01882-en-2.jpg
|
tree node's cluster. Misra et al. (2015) used summarization to identify arguments with similar aspects in dialogues from the Internet Argument Corpus (Walker et al., 2012). Similarly, Reimers et al. (2019) used agglomerative clustering of contextual embeddings and aspects to group sentence-level arguments. Bar-Haim et al. (2020a,b) examined the mapping of debate arguments to key points written by experts to serve as summaries.
Our approach clusters discussion units, but instead of a supervised selection of key cluster members, we use vanilla LLMs for abstractive summarization. Moreover, our summaries are hierarchical, using issue-generic frames as headings (Chong and Druckman, 2007; Boydstun et al., 2014) and generating concise abstractive summaries of corresponding clusters as subheadings. Thus our approach is unsupervised, facilitating a scalable and generalizable summarization of discussions.
## 2.3 Cluster Labeling
Cluster labeling involves assigning representative labels to document clusters to facilitate clustering exploration. Labeling approaches include comparing term distributions (Manning et al., 2008), selecting key terms closest to the cluster centroid (Role and Nadif, 2014), formulating key queries (Gollub et al., 2016), identify keywords through hypernym relationships (Poostchi and Piccardi, 2018), and weak supervision to generate topic labels Popa and Rebedea (2021). These approaches often select a small set of terms as labels that do not describe a cluster's contents in closed form. Our approach overcomes this limitation by treating cluster labeling as a zero-shot abstractive summarization task.
## 2.4 Frame Assignment
Framing involves emphasizing certain aspects of a topic for various purposes, such as persuasion (Entman, 1993; Chong and Druckman, 2007). Frame analysis for discussions provides insights into different perspectives on a topic (Morstatter et al., 2018; Liu et al., 2019). It also helps to identify biases in discussions resulting, e.g., from word choice (Hamborg et al., 2019b,a). Thus, frames can serve as valuable reference points for organizing long discussions. We use a predefined inventory of media frames (Boydstun et al., 2014) for discussion summarization. Instead of supervised frame assignment (Naderi and Hirst, 2017; Ajjour et al., 2019; Heinisch and Cimiano, 2021), we use prompt-based LLMs for more flexibility.
3 Indicative Discussion Summarization
Our indicative summarization approach takes the sentences of a discussion as input and generates a summary in the form of a table of contents, as shown in Figure 1. Its three steps consist of clustering discussion sentences, cluster labeling, and frame assignment to cluster labels.
## 3.1 Unit Clustering
Given a discussion, we extract its sentences as discussion units. The set of sentences is then clustered using the density-based hierarchical clustering algorithm HDBSCAN (Campello et al., 2013). Each sentence is embedded using SBERT (Reimers and Gurevych, 2019) and these embeddings are then mapped to a lower dimensionality using UMAP (McInnes et al., 2017).5 Unlike previous approaches that rank and filter clusters to generate informative summaries (Ernst et al., 2022; Syed et al., 2023), our summaries incorporate all clusters. The sentences of each cluster are ranked by centrality, which is determined by the $\lambda$ value of HDBSCAN. A number of central sentences per cluster are selected as input for cluster labeling by abstractive summarization.
Meta-sentence filtering Some sentences in a discussion do not contribute directly to the topic, but reflect the interaction between its participants. Examples include sentences such as“I agree with you." or “You are setting up a straw man." Pilot experiments have shown that such meta-sentences may cause our summarization approach to include them in the final summary. As these are irrelevant to our goal, we apply a corpus-specific and channel-specific meta-sentence filtering approach, respectively. Corpus-specific filtering is based on a small set of frequently used meta-sentences $M$ in a large corpus (e.g., on Reddit). It is bootstrapped during preprocessing, and all sentences in it are omitted by default.$^{6}$
Our pilot experiments revealed that some sentences in discussions are also channel-specific (e.g., for the ChangeMyView Subreddit). Therefore, we augment our sentence clustering approach by adding a random sample $M^{\prime}\subset M$ to the set of sentences $D$ of each individual discussion before clustering, where $|M^{\prime}|=\max\{300,|D|\}.$ The maximum value for the number of meta-sentences $|M^\prime|$ $\sqrt{5}$Implementation details are given in Appendix B.
$^{6}$The set is used like a typical stop word list, only for sentences.
|
[
[
0.1158447265625,
0.08489990234375,
0.496337890625,
0.0843505859375,
0.496337890625,
0.09857177734375,
0.1158447265625,
0.09979248046875
],
[
0.5087890625,
0.08489990234375,
0.8623046875,
0.08489990234375,
0.8623046875,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.1175537109375,
0.10394287109375,
0.48779296875,
0.10394287109375,
0.48779296875,
0.114013671875,
0.1175537109375,
0.114013671875
],
[
0.51123046875,
0.1092529296875,
0.880859375,
0.10986328125,
0.880859375,
0.12469482421875,
0.51123046875,
0.12353515625
],
[
0.1158447265625,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1158447265625,
0.1318359375
],
[
0.509765625,
0.1258544921875,
0.8818359375,
0.1270751953125,
0.8818359375,
0.141357421875,
0.509765625,
0.14013671875
],
[
0.1175537109375,
0.1341552734375,
0.489501953125,
0.1341552734375,
0.489501953125,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.1175537109375,
0.1484375,
0.48779296875,
0.149658203125,
0.48779296875,
0.1644287109375,
0.1175537109375,
0.163330078125
],
[
0.51123046875,
0.1436767578125,
0.8818359375,
0.1436767578125,
0.8818359375,
0.15673828125,
0.51123046875,
0.15673828125
],
[
0.509765625,
0.157958984375,
0.880859375,
0.157958984375,
0.880859375,
0.171630859375,
0.509765625,
0.171630859375
],
[
0.1175537109375,
0.166259765625,
0.490234375,
0.166259765625,
0.490234375,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.51123046875,
0.174560546875,
0.8818359375,
0.174560546875,
0.8818359375,
0.189453125,
0.51123046875,
0.189453125
],
[
0.1158447265625,
0.1817626953125,
0.48779296875,
0.18115234375,
0.48779296875,
0.1953125,
0.1158447265625,
0.196533203125
],
[
0.51123046875,
0.191162109375,
0.77099609375,
0.191162109375,
0.77099609375,
0.204833984375,
0.51123046875,
0.204833984375
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.1158447265625,
0.215576171875,
0.365234375,
0.215576171875,
0.365234375,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.51123046875,
0.2161865234375,
0.67822265625,
0.21728515625,
0.67822265625,
0.232177734375,
0.51123046875,
0.23095703125
],
[
0.1368408203125,
0.23095703125,
0.489501953125,
0.23095703125,
0.489501953125,
0.2452392578125,
0.1368408203125,
0.2452392578125
],
[
0.51318359375,
0.2381591796875,
0.8818359375,
0.2381591796875,
0.8818359375,
0.251708984375,
0.51318359375,
0.251708984375
],
[
0.1175537109375,
0.2476806640625,
0.490234375,
0.2476806640625,
0.490234375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.509765625,
0.254638671875,
0.88330078125,
0.254638671875,
0.88330078125,
0.268310546875,
0.509765625,
0.268310546875
],
[
0.1175537109375,
0.26318359375,
0.490234375,
0.26318359375,
0.490234375,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.509765625,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.283935546875,
0.509765625,
0.283935546875
],
[
0.1175537109375,
0.279052734375,
0.489501953125,
0.279052734375,
0.489501953125,
0.292236328125,
0.1175537109375,
0.292236328125
],
[
0.509765625,
0.28515625,
0.880859375,
0.28515625,
0.880859375,
0.2998046875,
0.509765625,
0.2998046875
],
[
0.1158447265625,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1158447265625,
0.308837890625
],
[
0.51123046875,
0.3017578125,
0.880859375,
0.3017578125,
0.880859375,
0.3154296875,
0.51123046875,
0.3154296875
],
[
0.1158447265625,
0.31005859375,
0.489501953125,
0.311279296875,
0.489501953125,
0.325927734375,
0.1158447265625,
0.32470703125
],
[
0.51123046875,
0.317626953125,
0.880859375,
0.317626953125,
0.880859375,
0.330810546875,
0.51123046875,
0.330810546875
],
[
0.1158447265625,
0.32666015625,
0.48779296875,
0.325927734375,
0.48779296875,
0.34033203125,
0.1158447265625,
0.341552734375
],
[
0.51318359375,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.1158447265625,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.1158447265625,
0.355712890625
],
[
0.5087890625,
0.349853515625,
0.8818359375,
0.350830078125,
0.8818359375,
0.365234375,
0.5087890625,
0.364013671875
],
[
0.1158447265625,
0.358154296875,
0.489501953125,
0.358154296875,
0.489501953125,
0.371826171875,
0.1158447265625,
0.371826171875
],
[
0.51123046875,
0.366455078125,
0.8818359375,
0.365234375,
0.8818359375,
0.380126953125,
0.51123046875,
0.38134765625
],
[
0.1158447265625,
0.3759765625,
0.376220703125,
0.3759765625,
0.376220703125,
0.3896484375,
0.1158447265625,
0.3896484375
],
[
0.509765625,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.509765625,
0.3955078125
],
[
0.1158447265625,
0.40087890625,
0.2939453125,
0.402099609375,
0.2939453125,
0.416259765625,
0.1158447265625,
0.4150390625
],
[
0.51318359375,
0.399658203125,
0.88330078125,
0.399658203125,
0.88330078125,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.1175537109375,
0.4228515625,
0.489501953125,
0.4228515625,
0.489501953125,
0.435791015625,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.427490234375,
0.51318359375,
0.427490234375
],
[
0.51123046875,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51123046875,
0.4453125
],
[
0.1175537109375,
0.439453125,
0.48779296875,
0.439453125,
0.48779296875,
0.453125,
0.1175537109375,
0.453125
],
[
0.509765625,
0.446533203125,
0.8818359375,
0.44775390625,
0.8818359375,
0.4619140625,
0.509765625,
0.46142578125
],
[
0.1175537109375,
0.454833984375,
0.490234375,
0.454833984375,
0.490234375,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.1175537109375,
0.471435546875,
0.490234375,
0.471435546875,
0.490234375,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.5087890625,
0.479736328125,
0.7109375,
0.479736328125,
0.7109375,
0.493408203125,
0.5087890625,
0.493408203125
],
[
0.1158447265625,
0.48681640625,
0.48779296875,
0.48681640625,
0.48779296875,
0.50048828125,
0.1158447265625,
0.50048828125
],
[
0.1158447265625,
0.5029296875,
0.486083984375,
0.5029296875,
0.486083984375,
0.5166015625,
0.1158447265625,
0.5166015625
],
[
0.509765625,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.509765625,
0.5185546875
],
[
0.118408203125,
0.52099609375,
0.486083984375,
0.52099609375,
0.486083984375,
0.53076171875,
0.118408203125,
0.53076171875
],
[
0.509765625,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.509765625,
0.53515625
],
[
0.1158447265625,
0.53369140625,
0.486083984375,
0.53369140625,
0.486083984375,
0.54736328125,
0.1158447265625,
0.54736328125
],
[
0.5087890625,
0.53759765625,
0.88330078125,
0.53759765625,
0.88330078125,
0.55029296875,
0.5087890625,
0.55029296875
],
[
0.1175537109375,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.1175537109375,
0.5654296875
],
[
0.5087890625,
0.5517578125,
0.880859375,
0.5517578125,
0.880859375,
0.5654296875,
0.5087890625,
0.5654296875
],
[
0.1158447265625,
0.56640625,
0.48779296875,
0.56787109375,
0.48779296875,
0.58203125,
0.1158447265625,
0.58056640625
],
[
0.50732421875,
0.568359375,
0.8818359375,
0.56787109375,
0.8818359375,
0.58203125,
0.50732421875,
0.5830078125
],
[
0.1175537109375,
0.58447265625,
0.484375,
0.58447265625,
0.484375,
0.59423828125,
0.1175537109375,
0.59423828125
],
[
0.51318359375,
0.5849609375,
0.880859375,
0.5849609375,
0.880859375,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.1175537109375,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.6142578125,
0.1175537109375,
0.6142578125
],
[
0.509765625,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.509765625,
0.615234375
],
[
0.1175537109375,
0.615234375,
0.490234375,
0.615234375,
0.490234375,
0.62890625,
0.1175537109375,
0.62890625
],
[
0.509765625,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.509765625,
0.6318359375
],
[
0.11419677734375,
0.63134765625,
0.48193359375,
0.63134765625,
0.48193359375,
0.64453125,
0.11419677734375,
0.64453125
],
[
0.509765625,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64794921875,
0.509765625,
0.64794921875
],
[
0.51123046875,
0.6484375,
0.8818359375,
0.6494140625,
0.8818359375,
0.66455078125,
0.51123046875,
0.6630859375
],
[
0.1158447265625,
0.65625,
0.310546875,
0.65771484375,
0.310546875,
0.6728515625,
0.1158447265625,
0.67041015625
],
[
0.51123046875,
0.66650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.6796875,
0.51123046875,
0.6796875
],
[
0.1175537109375,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.509765625,
0.681640625,
0.8818359375,
0.681640625,
0.8818359375,
0.6953125,
0.509765625,
0.6953125
],
[
0.11419677734375,
0.6953125,
0.49365234375,
0.69482421875,
0.49365234375,
0.708984375,
0.11419677734375,
0.71044921875
],
[
0.509765625,
0.69775390625,
0.8818359375,
0.69775390625,
0.8818359375,
0.7119140625,
0.509765625,
0.7119140625
],
[
0.1158447265625,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.1158447265625,
0.724609375
],
[
0.509765625,
0.71435546875,
0.8818359375,
0.71435546875,
0.8818359375,
0.72802734375,
0.509765625,
0.72802734375
],
[
0.1158447265625,
0.72802734375,
0.489501953125,
0.72802734375,
0.489501953125,
0.7412109375,
0.1158447265625,
0.7412109375
],
[
0.509765625,
0.72998046875,
0.66015625,
0.72998046875,
0.66015625,
0.74365234375,
0.509765625,
0.74365234375
],
[
0.1175537109375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.52880859375,
0.7451171875,
0.88330078125,
0.7451171875,
0.88330078125,
0.76025390625,
0.52880859375,
0.76025390625
],
[
0.1158447265625,
0.75830078125,
0.486083984375,
0.7587890625,
0.486083984375,
0.77392578125,
0.1158447265625,
0.7724609375
],
[
0.5087890625,
0.76123046875,
0.88671875,
0.76318359375,
0.88671875,
0.77783203125,
0.5087890625,
0.775390625
],
[
0.1175537109375,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.51318359375,
0.779296875,
0.88330078125,
0.779296875,
0.88330078125,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.509765625,
0.79443359375,
0.8818359375,
0.79443359375,
0.8818359375,
0.80859375,
0.509765625,
0.80859375
],
[
0.1175537109375,
0.80810546875,
0.490234375,
0.80810546875,
0.490234375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51318359375,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.1158447265625,
0.82373046875,
0.491943359375,
0.82373046875,
0.491943359375,
0.83837890625,
0.1158447265625,
0.83837890625
],
[
0.51318359375,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.51318359375,
0.84033203125
],
[
0.1175537109375,
0.84033203125,
0.489501953125,
0.84033203125,
0.489501953125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.5087890625,
0.8427734375,
0.8818359375,
0.84130859375,
0.8818359375,
0.85546875,
0.5087890625,
0.85693359375
],
[
0.118408203125,
0.85791015625,
0.486083984375,
0.85791015625,
0.486083984375,
0.8681640625,
0.118408203125,
0.8681640625
],
[
0.509765625,
0.859375,
0.880859375,
0.859375,
0.880859375,
0.87255859375,
0.509765625,
0.87255859375
],
[
0.1175537109375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.88134765625,
0.8134765625,
0.8818359375,
0.8134765625,
0.896484375,
0.509765625,
0.8955078125
],
[
0.11419677734375,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.11419677734375,
0.90185546875
],
[
0.51123046875,
0.89453125,
0.880859375,
0.8955078125,
0.880859375,
0.9091796875,
0.51123046875,
0.90771484375
],
[
0.1158447265625,
0.90478515625,
0.41650390625,
0.90478515625,
0.41650390625,
0.91845703125,
0.1158447265625,
0.91845703125
]
] |
[
"tree node's cluster. Misra et al. (2015) used sum-",
"3Indicative Discussion Summarization",
"marization to identify arguments with similar as-",
"Our indicative summarization approach takes the",
"pects in dialogues from the Internet Argument Cor-",
"sentences of a discussion as input and generates",
"pus (Walker et al., 2012). Similarly, Reimers et al.",
"(2019) used agglomerative clustering of contextual",
"a summary in the form of a table of contents, as",
"shown in Figure 1. Its three steps consist of clus",
"embeddings and aspects to group sentence-level ar-",
"tering discussion sentences, cluster labeling, and",
"guments. Bar-Haim et al. (2020a,b) examined the",
"frame assignment to cluster labels.",
"mapping of debate arguments to key points written",
"by experts to serve as summaries.",
"3.1 Unit Clustering",
"Our approach clusters discussion units, but in-",
"Given a discussion, we extract its sentences as dis-",
"stead of a supervised selection of key cluster mem-",
"cussion units. The set of sentences is then clustered",
"bers, we use vanilla LLMs for abstractive summa-",
"using the density-based hierarchical clustering al-",
"rization. Moreover, our summaries are hierarchical,",
"gorithm HDBSCAN (Campello et al., 2013). Each",
"using issue-generic frames as headings (Chong and",
"sentence is embedded using SBERT (Reimers and",
"Druckman, 2007; Boydstun et al., 2014) and gener-",
"Gurevych. 2019) and these embeddings are then",
"ating concise abstractive summaries of correspond-",
"mapped to a lower dimensionality using UMAP",
"ing clusters as subheadings. Thus our approach is",
"(McInnes et al., 2017).5Unlike previous ap",
"unsupervised, facilitating a scalable and generaliz-",
"proaches that rank and filter clusters to generate",
"able summarization of discussions.",
"informative summaries (Ernst et al.. 2022: Syed",
"2.3Cluster Labeling",
"et al., 2023), our summaries incorporate all clus-",
"Cluster labeling involves assigning representative",
"ters.The sentences of each cluster are ranked by",
"centrality, which is determined by the ^ value of",
"labels to document clusters to facilitate clustering",
"HDBSCAN. A number of central sentences per",
"exploration. Labeling approaches include compar-",
"cluster are selected as input for cluster labeling by",
"ing term distributions (Manning et al., 2008), select-",
"abstractive summarization.",
"ing key terms closest to the cluster centroid (Role",
"and Nadif, 2014), formulating key queries (Gollub",
"Meta-sentence filteringSome sentences in a dis-",
"etal.,2016).identifykeywordsthroughhypernym",
" cussion do not contribute directly to the topic, but",
"relationships (Poostchi and Piccardi. 2018), and",
"reflect the interaction between its participants.Ex-",
"weak supervision to generate topic labels Popa and",
"amples include sentences such as \"I agree with",
"Rebedea (2021). These approaches often select a",
"you.\" or “\"You are setting up a straw man.\" Pilot",
"small set of terms as labels that donot describe a",
"experiments have shown that such meta-sentences",
"cluster's contents in closed form. Our approach",
" may cause our summarization approach to include",
"overcomes this limitation by treating cluster label-",
"them in the final summary. As these are irrele",
"ing as a zero-shot abstractive summarization task.",
"vant to our goal, we apply a corpus-specific and",
"channel-specific meta-sentence filtering approach,",
"2.4Frame Assignment",
"respectively. Corpus-specific filtering is based on a",
"Framing involves emphasizing certain aspects of a",
" small set of frequently used meta-sentences M in",
"topic for various purposes, such as persuasion (Ent-",
"a large corpus (e.g., on Reddit). It is bootstrapped",
"man, 1993; Chong and Druckman, 2007). Frame",
" during preprocessing, and all sentences in it are",
"analysis for discussions provides insights into dif-",
"omitted by default.6",
"ferent perspectives on a topic (Morstatter et al..",
"Our pilot experiments revealed that some sen-",
"2018; Liu et al., 2019). It also helps to identify",
"tences in discussions are also channel-specific (e.g.,",
"biases in discussions resulting, e.g., from word",
"for the ChangeMyView Subreddit).Therefore,",
"choice (Hamborg et al., 2019b,a). Thus, frames",
"we augment our sentence clustering approach by",
"can serve as valuable reference points for organiz-",
"adding a random sample M' C M to the set of sen-",
"ing long discussions. We use a predefined inven-",
"tences D of each individual discussion before clus-",
"tory of media frames (Boydstun et al., 2014) for",
"tering, where M'|= max[300, |D|}. The maxi-",
"discussion summarization.Instead of supervised",
"mum value for the number of meta-sentences |M'l",
"frame assignment (Naderi and Hirst, 2017; Ajjour",
"'Implementation details are given in Appendix B.",
"et al., 2019; Heinisch and Cimiano, 2021), we use",
"6The set is used like a typical stop word list, only for sentences.",
"prompt-based LLMs for more flexibility."
] |
{
"arxiv_id": "2311.01882",
"language": "en",
"page_idx": 2
}
|
||
2311.01882-en-3
|
2311.01882-en-3.jpg
|
is chosen empirically, to maximize the likelihood that channel-specific meta-sentences are clustered with corpus-specific ones. After clustering the joint set of meta-sentences and discussion sentences $D\cup M^{\prime}$, we obtain the clustering $C.$ Let $m_C$ = $| C$ $\cap$ $M^{\prime }|$ denote the number of metasentences and $d_C=|C\cap D|$ the number of discussion sentences in a cluster $C\in\mathcal{C}.$ The proportion of meta-sentences in a cluster is then estimated as $P(M^{\prime}|C)=\frac{m_{C}}{m_{C}+d_{C}}.$
A cluster $C$ is classified as a meta-sentence cluster if $P(M^{\prime}|C)>\theta\cdot P(M^{\prime})$,where $P(M^{\prime})=\frac{|M^{\prime}|}{|D|}$ assumes that meta-sentences are independent of others in a discussion. The noise threshold $\theta=\frac{2}{3}$ was chosen empirically. Sentences in a discussion that either belong to a meta-sentence cluster or whose nearest cluster is considered to be one are omitted. In our evaluation, an average of 23% of sentences are filtered from discussions. Figure 2 illustrates the effect of meta-sentence filtering on a discussion's set of sentence.
## 3.2 Generative Cluster Labeling
Most cluster labeling approaches extract keywords or key phrases as labels, which limits their fluency. These approaches may also require training data acquisition for supervised learning. We formulate cluster labeling as an unsupervised abstractive summarization task. We experiment with prompt-based large language models in zero-shot and few-shot settings. This enables generalization across multiple domains, the elimination of supervised learning, and fluent cluster labels with higher readability in comparison to keywords or phrases.
We develop several prompt templates specifically tailored for different types of LLMs. For encoder-decoder models, we carefully develop appropriate prompts based on PromptSource (Bach et al., 2022), a toolkit that provides a comprehensive collection of natural language prompts for various tasks across 180 datasets. In particular, we analyze prompts for text summarization datasets with respect to (1) descriptive words for the generation of cluster labels using abstractive summarization, (2) commonly used separators to distinguish instructions from context, (3) the position of instructions within prompts, and (4) the granularity level of input data (full text, document title, or sentence). Since our task is about summarizing groups of sentences, we chose prompts that require the full text as input to ensure that enough contextual
<FigureHere>
Figure 2: Effect of meta-sentence filtering: (a and b) A
discussion's sentences $D$ are jointly clustered with a sample of meta-sentences $M^{\prime}\subset M.$ (c) Then each cluster is classified as a meta-sentence cluster based on its proportion of meta-sentences and its neighboring clusters. Meta-sentence clusters are omitted
information is provided (within the limits of each model's input size). Section 4.1 provides details on the prompt engineering process.
|
[
[
0.1158447265625,
0.08612060546875,
0.486083984375,
0.08612060546875,
0.486083984375,
0.09979248046875,
0.1158447265625,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.1175537109375,
0.1175537109375,
0.48779296875,
0.1175537109375,
0.48779296875,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.1175537109375,
0.1341552734375,
0.48779296875,
0.1341552734375,
0.48779296875,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.1175537109375,
0.149658203125,
0.48779296875,
0.149658203125,
0.48779296875,
0.163330078125,
0.1175537109375,
0.163330078125
],
[
0.1175537109375,
0.1650390625,
0.48779296875,
0.1650390625,
0.48779296875,
0.1787109375,
0.1175537109375,
0.1787109375
],
[
0.1175537109375,
0.18115234375,
0.489501953125,
0.18115234375,
0.489501953125,
0.19482421875,
0.1175537109375,
0.19482421875
],
[
0.1175537109375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.118408203125,
0.2161865234375,
0.48779296875,
0.2161865234375,
0.48779296875,
0.226806640625,
0.118408203125,
0.226806640625
],
[
0.118408203125,
0.232177734375,
0.21240234375,
0.232177734375,
0.21240234375,
0.243408203125,
0.118408203125,
0.243408203125
],
[
0.2166748046875,
0.2392578125,
0.2626953125,
0.2392578125,
0.2626953125,
0.2464599609375,
0.2166748046875,
0.2464599609375
],
[
0.1351318359375,
0.2464599609375,
0.48779296875,
0.2464599609375,
0.48779296875,
0.260009765625,
0.1351318359375,
0.260009765625
],
[
0.1158447265625,
0.26318359375,
0.45166015625,
0.26318359375,
0.45166015625,
0.276611328125,
0.1158447265625,
0.276611328125
],
[
0.455078125,
0.260009765625,
0.480224609375,
0.260009765625,
0.480224609375,
0.267333984375,
0.455078125,
0.267333984375
],
[
0.11419677734375,
0.27783203125,
0.484375,
0.27783203125,
0.484375,
0.29150390625,
0.11419677734375,
0.29150390625
],
[
0.51123046875,
0.279052734375,
0.87744140625,
0.27783203125,
0.87744140625,
0.29150390625,
0.51123046875,
0.292236328125
],
[
0.11419677734375,
0.29150390625,
0.486083984375,
0.29345703125,
0.486083984375,
0.311279296875,
0.11419677734375,
0.308837890625
],
[
0.1175537109375,
0.311279296875,
0.486083984375,
0.311279296875,
0.486083984375,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.52490234375,
0.30712890625,
0.6484375,
0.30712890625,
0.6484375,
0.317626953125,
0.52490234375,
0.317626953125
],
[
0.1175537109375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.1175537109375,
0.34033203125
],
[
0.705078125,
0.32373046875,
0.724609375,
0.32373046875,
0.724609375,
0.33740234375,
0.705078125,
0.33740234375
],
[
0.1175537109375,
0.34326171875,
0.48779296875,
0.34326171875,
0.48779296875,
0.35693359375,
0.1175537109375,
0.35693359375
],
[
0.7138671875,
0.339111328125,
0.72119140625,
0.339111328125,
0.72119140625,
0.345703125,
0.7138671875,
0.345703125
],
[
0.1158447265625,
0.359375,
0.48779296875,
0.358154296875,
0.48779296875,
0.372314453125,
0.1158447265625,
0.37353515625
],
[
0.1175537109375,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.1175537109375,
0.388427734375
],
[
0.1158447265625,
0.39013671875,
0.48779296875,
0.391357421875,
0.48779296875,
0.40625,
0.1158447265625,
0.405029296875
],
[
0.1175537109375,
0.407958984375,
0.32568359375,
0.407958984375,
0.32568359375,
0.421630859375,
0.1175537109375,
0.421630859375
],
[
0.1175537109375,
0.43408203125,
0.3837890625,
0.43408203125,
0.3837890625,
0.44775390625,
0.1175537109375,
0.44775390625
],
[
0.1175537109375,
0.454833984375,
0.48779296875,
0.454833984375,
0.48779296875,
0.4697265625,
0.1175537109375,
0.4697265625
],
[
0.1175537109375,
0.47021484375,
0.489501953125,
0.47021484375,
0.489501953125,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.1175537109375,
0.48681640625,
0.48779296875,
0.48681640625,
0.48779296875,
0.501953125,
0.1175537109375,
0.501953125
],
[
0.509765625,
0.496337890625,
0.8759765625,
0.49755859375,
0.8759765625,
0.51123046875,
0.509765625,
0.51025390625
],
[
0.1158447265625,
0.50439453125,
0.486083984375,
0.50439453125,
0.486083984375,
0.51708984375,
0.1158447265625,
0.51708984375
],
[
0.1175537109375,
0.51953125,
0.489501953125,
0.51953125,
0.489501953125,
0.533203125,
0.1175537109375,
0.533203125
],
[
0.5263671875,
0.52685546875,
0.5439453125,
0.52685546875,
0.5439453125,
0.53369140625,
0.5263671875,
0.53369140625
],
[
0.5400390625,
0.525390625,
0.66162109375,
0.525390625,
0.66162109375,
0.5361328125,
0.5400390625,
0.5361328125
],
[
0.1175537109375,
0.53515625,
0.486083984375,
0.53515625,
0.486083984375,
0.548828125,
0.1175537109375,
0.548828125
],
[
0.7060546875,
0.54345703125,
0.72119140625,
0.54345703125,
0.72119140625,
0.55419921875,
0.7060546875,
0.55419921875
],
[
0.1158447265625,
0.55029296875,
0.486083984375,
0.55029296875,
0.486083984375,
0.56396484375,
0.1158447265625,
0.56396484375
],
[
0.1175537109375,
0.56787109375,
0.489501953125,
0.56787109375,
0.489501953125,
0.58203125,
0.1175537109375,
0.58203125
],
[
0.1175537109375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.1158447265625,
0.5986328125,
0.486083984375,
0.5986328125,
0.486083984375,
0.6123046875,
0.1158447265625,
0.6123046875
],
[
0.118408203125,
0.61767578125,
0.382080078125,
0.61767578125,
0.382080078125,
0.62744140625,
0.118408203125,
0.62744140625
],
[
0.1368408203125,
0.6318359375,
0.490234375,
0.6318359375,
0.490234375,
0.6455078125,
0.1368408203125,
0.6455078125
],
[
0.1175537109375,
0.6484375,
0.489501953125,
0.6484375,
0.489501953125,
0.662109375,
0.1175537109375,
0.662109375
],
[
0.1158447265625,
0.66455078125,
0.48779296875,
0.6650390625,
0.48779296875,
0.6787109375,
0.1158447265625,
0.677734375
],
[
0.1158447265625,
0.68115234375,
0.48779296875,
0.68115234375,
0.48779296875,
0.69482421875,
0.1158447265625,
0.69482421875
],
[
0.1158447265625,
0.6953125,
0.490234375,
0.69775390625,
0.490234375,
0.71142578125,
0.1158447265625,
0.708984375
],
[
0.1158447265625,
0.71142578125,
0.490234375,
0.7119140625,
0.490234375,
0.72705078125,
0.1158447265625,
0.7255859375
],
[
0.51123046875,
0.71630859375,
0.85791015625,
0.71533203125,
0.85791015625,
0.72705078125,
0.51123046875,
0.72802734375
],
[
0.1158447265625,
0.728515625,
0.486083984375,
0.728515625,
0.486083984375,
0.7421875,
0.1158447265625,
0.7421875
],
[
0.51123046875,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.1175537109375,
0.7451171875,
0.48779296875,
0.7451171875,
0.48779296875,
0.7587890625,
0.1175537109375,
0.7587890625
],
[
0.5146484375,
0.7548828125,
0.880859375,
0.7548828125,
0.880859375,
0.765625,
0.5146484375,
0.765625
],
[
0.1175537109375,
0.76123046875,
0.490234375,
0.76123046875,
0.490234375,
0.77490234375,
0.1175537109375,
0.77490234375
],
[
0.51123046875,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.51123046875,
0.78076171875
],
[
0.1175537109375,
0.77685546875,
0.490234375,
0.77685546875,
0.490234375,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.51123046875,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79345703125,
0.51123046875,
0.79345703125
],
[
0.1175537109375,
0.79345703125,
0.48779296875,
0.79345703125,
0.48779296875,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.509765625,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.509765625,
0.81005859375
],
[
0.1158447265625,
0.80810546875,
0.490234375,
0.80859375,
0.490234375,
0.82373046875,
0.1158447265625,
0.822265625
],
[
0.5146484375,
0.8125,
0.81005859375,
0.8125,
0.81005859375,
0.822265625,
0.5146484375,
0.822265625
],
[
0.1175537109375,
0.826171875,
0.486083984375,
0.826171875,
0.486083984375,
0.8388671875,
0.1175537109375,
0.8388671875
],
[
0.1175537109375,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.509765625,
0.85107421875,
0.87890625,
0.85107421875,
0.87890625,
0.8642578125,
0.509765625,
0.8642578125
],
[
0.1175537109375,
0.85791015625,
0.489501953125,
0.85791015625,
0.489501953125,
0.87158203125,
0.1175537109375,
0.87158203125
],
[
0.509765625,
0.86767578125,
0.8818359375,
0.86767578125,
0.8818359375,
0.88134765625,
0.509765625,
0.88134765625
],
[
0.1175537109375,
0.87353515625,
0.489501953125,
0.87353515625,
0.489501953125,
0.88720703125,
0.1175537109375,
0.88720703125
],
[
0.1175537109375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.88427734375,
0.75146484375,
0.88427734375,
0.75146484375,
0.89794921875,
0.51123046875,
0.89794921875
]
] |
[
"is chosen empirically, to maximize the likelihood",
"that channel-specific meta-sentences are clustered",
"with corpus-specific ones..After clustering the",
"joint set of meta-sentences and discussion sen-",
"tences D U M', we obtain the clustering C. Let",
"mc = IC n M'l denote the number of meta-",
"sentences and dc = |C n D| the number of discus-",
"sion sentences in a cluster C E C. The proportion",
"ofmeta-sentencesinaclusteristhenestimatedas",
"P(M'IC) =",
"me+d",
"A cluster C is classified as a meta-sentence clus-",
"terif P(M'C) > .P(M'), where P(M') =",
"IM",
"assumes that meta-sentences are independent of",
"(a) Joint clustering of a discussion and meta-sentences DUM'",
"others in a discussion. The noise threshold = 3",
"was chosen empirically. Sentences in a discussion",
"CuratedNoiseSet",
"that either belong to a meta-sentence cluster or",
"8",
"whose nearest cluster is considered to be one are",
"e",
"omitted. In our evaluation, an average of 23% of",
"sentences are filtered from discussions. Figure 2",
"illustrates the effect of meta-sentence filtering on a",
"discussion's set of sentence.",
"3.2Generative Cluster Labeling",
"Most cluster labeling approaches extract keywords",
"or key phrases as labels, which limits their fluency.",
"These approaches may also require training data",
"(b) The sampled meta-sentences M' C M highlighted gray.",
"acquisition for supervised learning.We formulate",
"cluster labeling as an unsupervised abstractive sum-",
"·",
"DiscardedSentences",
"marization task. We experiment with prompt-based",
"",
"large language models in zero-shot and few-shot",
"settings. This enables generalization across multi-",
"ple domains, the elimination of supervised learning,",
"and fluent cluster labels with higher readability in",
"comparison to keywords or phrases.",
"We develop several prompt templates specifi-",
"cally tailored for different types of LLMs. For",
"encoder-decodermodels,wecarefullydevelopap-",
"propriate prompts based on PromptSource (Bach",
"et al., 2022), a toolkit that provides a comprehen-",
"sive collection of natural language prompts for var-",
"(c)Classification of meta-sentence clusters to be omitted.",
"ious tasks across 180 datasets. In particular, we",
"Figure 2: Effect of meta-sentence filtering: (a and b) A",
"analyze prompts for text summarization datasets",
"discussion's sentences D are jointly clustered with a",
"with respect to (1) descriptive words for the gener-",
"sample of meta-sentences M' C M. (c) Then each",
"ation of cluster labels using abstractive summariza-",
"cluster is classified as a meta-sentence cluster based",
"tion, (2) commonly used separators to distinguish",
" on its proportion of meta-sentences and its neighboring",
"instructions from context, (3) the position of in-",
"clusters.Meta-sentenceclustersareomitted",
"structions within prompts, and (4) the granularity",
"level of input data (full text, document title, or sen-",
"information is provided (within the limits of each",
"tence). Since our task is about summarizing groups",
"model's input size). Section 4.1 provides details on",
"of sentences, we chose prompts that require the",
"full text as input to ensure that enough contextual",
"the prompt engineering process."
] |
{
"arxiv_id": "2311.01882",
"language": "en",
"page_idx": 3
}
|
||
2311.01882-en-4
|
2311.01882-en-4.jpg
|
### 3.3 Frame Assignment
Any controversial topic can be discussed from different perspectives. For example,“the dangers of social media" can be discussed from a moral or a health perspective, among others. In our indicative summaries, we use argumentation frame labels as top-level headings to operationalize different perspectives. An argumentation frame may include one or more groups of relevant arguments. We assign frame labels from the issue-generic frame inventory shown in Table 1 (Boydstun et al., 2014) to each cluster label derived in the previous step.$^{7}$
We use prompt-based models in both zero-shot and few-shot settings for frame assignment. In our experiments with instruction-tuned models, we designed two types of instructions, shown in Figure 10, namely direct instructions for models trained on instruction-response samples, and dia- $\log$ instructions for chat models. The instructions are included along with the cluster labels in the prompts. Moreover, including the citation of the frame inventory used in our experiments has a positive effect on the effectiveness of some models (see Appendix D.1 for details).
## 3.4 Indicative Summary Presentation
Given the generated labels of all sentence clusters and the frame labels assigned to each cluster label, our indicative summary groups the cluster labels by their respective frame labels. The cluster label groups of each frame label are then ordered by cluster size. This results in a two-level indicative summary, as shown in Figures 1 and 4.
# 4 Prompt Engineering
Using prompt-based LLMs for generative cluster labeling and frame assignment requires modelspecific prompt engineering as a preliminary step. We explored the 19 model variants listed in Table 2. To select the most appropriate models for our task, we consulted the HELM benchmark (Liang et al., 2022), which compares the effectiveness of different LLMs for different tasks. Further, we have included various recently released open source models (with optimized instructions) as they were released. Since many of them were released during our research, we reuse prompts previously optimized prompts for the newer models.$^{8}$
$\overline{{^7}}$For detailed label descriptions see Table 6 in the Appendix.
$^{8}$See Appendices C and D for details.
<table>
<tbody>
<tr>
<th colspan="2">Frame Inventory</th>
</tr>
<tr>
<th>CapacityResources</th>
<th>Fairness & Equality</th>
</tr>
<tr>
<td>Constitutionality &</td>
<td>Health& Safety</td>
</tr>
<tr>
<td>Jurisprudence</td>
<td>$\mathbf{Morality}$</td>
</tr>
<tr>
<td>Crime & Punishment</td>
<td>Policy Prescription & Evaluation</td>
</tr>
<tr>
<td>Cultural Identity</td>
<td>Political</td>
</tr>
<tr>
<td>Economic</td>
<td>Public Opinion</td>
</tr>
<tr>
<td>External Regulation &</td>
<td>Quality of Life</td>
</tr>
<tr>
<td>Reputation</td>
<td>Security & Defense</td>
</tr>
</tbody>
</table>
Table 1: Inventory of frames proposed by Boydstun et al.
(2014) to track the media's framing of policy issues.
<table>
<tbody>
<tr>
<th>$\mathbf{Model}$ Variants</th>
<th>$\mathbf{Description}$</th>
</tr>
<tr>
<th> </th>
<th>Pre-InstructGPT</th>
</tr>
<tr>
<td>$T0$ vanilla</td>
<td>Encoder-decoder model trained on datasets transformed as task-specific prompts.</td>
</tr>
<tr>
<td>BLOOM vanilla</td>
<td>A multlingual autoregressive model with 176B parameters for prompt-based text completion</td>
</tr>
<tr>
<td>$GPT-NeoX$ $20B$</td>
<td>Open source alternative to GPT-3.</td>
</tr>
<tr>
<td>$OPT$</td>
<td>Autoregressive M da sarith similar effective</td>
</tr>
<tr>
<td>$66B$</td>
<td>ness to GPT-3, but more efficient data collec tion and training.</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td> </td>
<td>Direct Instruction</td>
</tr>
<tr>
<td>LLaMA-CoT vanilla</td>
<td>LLaMA-30B fine-tuned on chain-of-thought and reasoning samples (Si and Lin, 2023).</td>
</tr>
<tr>
<td>Alpaca $7B$</td>
<td>LLaMA-7B fine-tuned based on 52k self instruct responses (Wang et al., 2022).</td>
</tr>
<tr>
<td>$0ASST$ vanilla</td>
<td>LLaMA-30B fine-tuned on the OpenAssistant Conversations dataset (Köpf et al., 2023) using reinforcement learning.</td>
</tr>
<tr>
<td>Pythia $12B$</td>
<td>Suite of LLMs trained on public data to in vestigate the effects of training and scaling on various model properties.</td>
</tr>
<tr>
<td>GPTx 3.5,Chat,4</td>
<td>OpenAI models GPT3.5 (text-davinci-003) ChatGPT(gpt-3.5-turbo),andGPT4.</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>LLaMA $30\mathbf{B},65\mathbf{B}$</td>
<td>Suite of open-source LLMs from Meta AI trained on public datasets.</td>
</tr>
<tr>
<td>Baize $7\mathbf{B},13\mathbf{B}$</td>
<td>Open source chat model trained on 100k c alogues generated by letting ChatGPT (GPT 3.5-turbo) talk to itself.</td>
</tr>
<tr>
<td>Falcon</td>
<td>Trained on the RefinedWeb corpus (Penedo</td>
</tr>
<tr>
<td>$40B$</td>
<td>////// ilterins</td>
</tr>
<tr>
<td>$40\mathbf{B}-$Instruct</td>
<td>web data Ant def</td>
</tr>
</tbody>
</table>
Table 2: LLMs studied for cluster labeling and frame assignment. Older models are listed by Pre-InstructGPT (prior to GPT3.5) and newer models are listed by their respective prompt types investigated( Direct Dialogue ). See Appendices C and D for details.
|
[
[
0.1175537109375,
0.08612060546875,
0.30908203125,
0.08612060546875,
0.30908203125,
0.09979248046875,
0.1175537109375,
0.09979248046875
],
[
0.63818359375,
0.09027099609375,
0.75390625,
0.0914306640625,
0.75390625,
0.105712890625,
0.63818359375,
0.1051025390625
],
[
0.1158447265625,
0.10687255859375,
0.48779296875,
0.10687255859375,
0.48779296875,
0.12054443359375,
0.1158447265625,
0.12054443359375
],
[
0.51123046875,
0.1134033203125,
0.6533203125,
0.1134033203125,
0.6533203125,
0.1270751953125,
0.51123046875,
0.1270751953125
],
[
0.67822265625,
0.1134033203125,
0.8037109375,
0.1134033203125,
0.8037109375,
0.1270751953125,
0.67822265625,
0.1270751953125
],
[
0.1175537109375,
0.12353515625,
0.48779296875,
0.12353515625,
0.48779296875,
0.13720703125,
0.1175537109375,
0.13720703125
],
[
0.51318359375,
0.12890625,
0.63037109375,
0.12890625,
0.63037109375,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.6767578125,
0.12469482421875,
0.779296875,
0.1270751953125,
0.779296875,
0.141357421875,
0.6767578125,
0.1395263671875
],
[
0.1175537109375,
0.1395263671875,
0.48779296875,
0.1395263671875,
0.48779296875,
0.152587890625,
0.1175537109375,
0.152587890625
],
[
0.5283203125,
0.142578125,
0.615234375,
0.142578125,
0.615234375,
0.15380859375,
0.5283203125,
0.15380859375
],
[
0.67822265625,
0.142578125,
0.73486328125,
0.142578125,
0.73486328125,
0.15380859375,
0.67822265625,
0.15380859375
],
[
0.51123046875,
0.155029296875,
0.6455078125,
0.155029296875,
0.6455078125,
0.168701171875,
0.51123046875,
0.168701171875
],
[
0.6767578125,
0.155029296875,
0.8818359375,
0.155029296875,
0.8818359375,
0.168701171875,
0.6767578125,
0.168701171875
],
[
0.51318359375,
0.17041015625,
0.61376953125,
0.17041015625,
0.61376953125,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.1158447265625,
0.1561279296875,
0.489501953125,
0.1561279296875,
0.489501953125,
0.169189453125,
0.1158447265625,
0.169189453125
],
[
0.1158447265625,
0.171630859375,
0.489501953125,
0.171630859375,
0.489501953125,
0.185302734375,
0.1158447265625,
0.185302734375
],
[
0.67822265625,
0.17041015625,
0.73291015625,
0.17041015625,
0.73291015625,
0.1817626953125,
0.67822265625,
0.1817626953125
],
[
0.51123046875,
0.185302734375,
0.57666015625,
0.185302734375,
0.57666015625,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.6767578125,
0.1817626953125,
0.77490234375,
0.18408203125,
0.77490234375,
0.198974609375,
0.6767578125,
0.196533203125
],
[
0.509765625,
0.19775390625,
0.6513671875,
0.19775390625,
0.6513671875,
0.21142578125,
0.509765625,
0.21142578125
],
[
0.1158447265625,
0.18701171875,
0.490234375,
0.18701171875,
0.490234375,
0.20068359375,
0.1158447265625,
0.20068359375
],
[
0.11419677734375,
0.2037353515625,
0.486083984375,
0.2037353515625,
0.486083984375,
0.21728515625,
0.11419677734375,
0.21728515625
],
[
0.67919921875,
0.198974609375,
0.771484375,
0.198974609375,
0.771484375,
0.210205078125,
0.67919921875,
0.210205078125
],
[
0.5283203125,
0.213134765625,
0.59765625,
0.213134765625,
0.59765625,
0.2244873046875,
0.5283203125,
0.2244873046875
],
[
0.6767578125,
0.2120361328125,
0.8017578125,
0.2120361328125,
0.8017578125,
0.2257080078125,
0.6767578125,
0.2257080078125
],
[
0.11419677734375,
0.2197265625,
0.48779296875,
0.21728515625,
0.48779296875,
0.232177734375,
0.11419677734375,
0.2340087890625
],
[
0.1175537109375,
0.235107421875,
0.48779296875,
0.235107421875,
0.48779296875,
0.2493896484375,
0.1175537109375,
0.2493896484375
],
[
0.51123046875,
0.2410888671875,
0.8818359375,
0.2410888671875,
0.8818359375,
0.254638671875,
0.51123046875,
0.254638671875
],
[
0.1175537109375,
0.251708984375,
0.48779296875,
0.251708984375,
0.48779296875,
0.265380859375,
0.1175537109375,
0.265380859375
],
[
0.509765625,
0.255859375,
0.861328125,
0.255859375,
0.861328125,
0.26953125,
0.509765625,
0.26953125
],
[
0.11419677734375,
0.268310546875,
0.480224609375,
0.267333984375,
0.480224609375,
0.28076171875,
0.11419677734375,
0.281982421875
],
[
0.1368408203125,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.1368408203125,
0.297607421875
],
[
0.51123046875,
0.283935546875,
0.56103515625,
0.283935546875,
0.56103515625,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.603515625,
0.2861328125,
0.6826171875,
0.2861328125,
0.6826171875,
0.297607421875,
0.603515625,
0.297607421875
],
[
0.1175537109375,
0.3017578125,
0.486083984375,
0.3017578125,
0.486083984375,
0.312255859375,
0.1175537109375,
0.312255859375
],
[
0.51123046875,
0.297607421875,
0.56494140625,
0.297607421875,
0.56494140625,
0.308837890625,
0.51123046875,
0.308837890625
],
[
0.1158447265625,
0.31640625,
0.486083984375,
0.31640625,
0.486083984375,
0.32666015625,
0.1158447265625,
0.32666015625
],
[
0.634765625,
0.318359375,
0.75830078125,
0.318359375,
0.75830078125,
0.333251953125,
0.634765625,
0.333251953125
],
[
0.1175537109375,
0.333251953125,
0.48779296875,
0.333251953125,
0.48779296875,
0.3466796875,
0.1175537109375,
0.3466796875
],
[
0.51123046875,
0.33740234375,
0.53076171875,
0.33740234375,
0.53076171875,
0.3466796875,
0.51123046875,
0.3466796875
],
[
0.603515625,
0.33837890625,
0.87890625,
0.33837890625,
0.87890625,
0.3486328125,
0.603515625,
0.3486328125
],
[
0.118408203125,
0.349853515625,
0.486083984375,
0.349853515625,
0.486083984375,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.509765625,
0.349853515625,
0.5576171875,
0.349853515625,
0.5576171875,
0.361083984375,
0.509765625,
0.361083984375
],
[
0.60107421875,
0.3486328125,
0.82861328125,
0.349853515625,
0.82861328125,
0.361083984375,
0.60107421875,
0.35986328125
],
[
0.1158447265625,
0.363525390625,
0.48779296875,
0.363525390625,
0.48779296875,
0.377197265625,
0.1158447265625,
0.377197265625
],
[
0.51123046875,
0.365234375,
0.5556640625,
0.365234375,
0.5556640625,
0.377197265625,
0.51123046875,
0.377197265625
],
[
0.60205078125,
0.365234375,
0.880859375,
0.365234375,
0.880859375,
0.37890625,
0.60205078125,
0.37890625
],
[
0.51123046875,
0.377685546875,
0.5576171875,
0.377685546875,
0.5576171875,
0.3896484375,
0.51123046875,
0.3896484375
],
[
0.603515625,
0.37890625,
0.87744140625,
0.37890625,
0.87744140625,
0.3896484375,
0.603515625,
0.3896484375
],
[
0.509765625,
0.39013671875,
0.580078125,
0.392578125,
0.580078125,
0.40673828125,
0.509765625,
0.405029296875
],
[
0.1158447265625,
0.38134765625,
0.489501953125,
0.38134765625,
0.489501953125,
0.394287109375,
0.1158447265625,
0.394287109375
],
[
0.1158447265625,
0.396728515625,
0.48779296875,
0.396728515625,
0.48779296875,
0.410400390625,
0.1158447265625,
0.410400390625
],
[
0.60205078125,
0.393798828125,
0.81201171875,
0.393798828125,
0.81201171875,
0.40673828125,
0.60205078125,
0.40673828125
],
[
0.51318359375,
0.40625,
0.5439453125,
0.40625,
0.5439453125,
0.418701171875,
0.51318359375,
0.418701171875
],
[
0.11419677734375,
0.412109375,
0.48779296875,
0.410888671875,
0.48779296875,
0.42578125,
0.11419677734375,
0.427001953125
],
[
0.51318359375,
0.421630859375,
0.54150390625,
0.421630859375,
0.54150390625,
0.43408203125,
0.51318359375,
0.43408203125
],
[
0.60205078125,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435302734375,
0.60205078125,
0.435302734375
],
[
0.1175537109375,
0.429931640625,
0.489501953125,
0.429931640625,
0.489501953125,
0.443603515625,
0.1175537109375,
0.443603515625
],
[
0.51123046875,
0.43408203125,
0.5439453125,
0.43408203125,
0.5439453125,
0.446533203125,
0.51123046875,
0.446533203125
],
[
0.60205078125,
0.432861328125,
0.88330078125,
0.432861328125,
0.88330078125,
0.44775390625,
0.60205078125,
0.44775390625
],
[
0.1175537109375,
0.4453125,
0.48779296875,
0.4453125,
0.48779296875,
0.458984375,
0.1175537109375,
0.458984375
],
[
0.603515625,
0.44775390625,
0.70947265625,
0.44775390625,
0.70947265625,
0.458984375,
0.603515625,
0.458984375
],
[
0.1158447265625,
0.46142578125,
0.30908203125,
0.46142578125,
0.30908203125,
0.474365234375,
0.1158447265625,
0.474365234375
],
[
0.62451171875,
0.46728515625,
0.76904296875,
0.46728515625,
0.76904296875,
0.482177734375,
0.62451171875,
0.482177734375
],
[
0.51318359375,
0.486328125,
0.58447265625,
0.486328125,
0.58447265625,
0.49755859375,
0.51318359375,
0.49755859375
],
[
0.60205078125,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.498779296875,
0.60205078125,
0.498779296875
],
[
0.60205078125,
0.49755859375,
0.861328125,
0.49755859375,
0.861328125,
0.51123046875,
0.60205078125,
0.51123046875
],
[
0.1158447265625,
0.48681640625,
0.42236328125,
0.48681640625,
0.42236328125,
0.50048828125,
0.1158447265625,
0.50048828125
],
[
0.51318359375,
0.498779296875,
0.5576171875,
0.498779296875,
0.5576171875,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.1175537109375,
0.50830078125,
0.48779296875,
0.50830078125,
0.48779296875,
0.521484375,
0.1175537109375,
0.521484375
],
[
0.509765625,
0.51318359375,
0.5634765625,
0.51318359375,
0.5634765625,
0.52783203125,
0.509765625,
0.52783203125
],
[
0.60205078125,
0.51318359375,
0.88330078125,
0.51318359375,
0.88330078125,
0.52685546875,
0.60205078125,
0.52685546875
],
[
0.1158447265625,
0.52490234375,
0.489501953125,
0.52490234375,
0.489501953125,
0.5380859375,
0.1158447265625,
0.5380859375
],
[
0.51318359375,
0.52783203125,
0.53564453125,
0.52783203125,
0.53564453125,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.603515625,
0.52783203125,
0.83642578125,
0.52783203125,
0.83642578125,
0.5380859375,
0.603515625,
0.5380859375
],
[
0.11419677734375,
0.54052734375,
0.48779296875,
0.5390625,
0.48779296875,
0.55419921875,
0.11419677734375,
0.5546875
],
[
0.509765625,
0.54150390625,
0.5556640625,
0.54150390625,
0.5556640625,
0.5556640625,
0.509765625,
0.5556640625
],
[
0.60205078125,
0.54150390625,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.60205078125,
0.5546875
],
[
0.509765625,
0.5546875,
0.5576171875,
0.5546875,
0.5576171875,
0.56640625,
0.509765625,
0.56640625
],
[
0.60107421875,
0.552734375,
0.8818359375,
0.55419921875,
0.8818359375,
0.568359375,
0.60107421875,
0.56787109375
],
[
0.60205078125,
0.5654296875,
0.74462890625,
0.56640625,
0.74462890625,
0.580078125,
0.60205078125,
0.5791015625
],
[
0.1158447265625,
0.55712890625,
0.48779296875,
0.55712890625,
0.48779296875,
0.57080078125,
0.1158447265625,
0.57080078125
],
[
0.1158447265625,
0.572265625,
0.48779296875,
0.57177734375,
0.48779296875,
0.5859375,
0.1158447265625,
0.58740234375
],
[
0.51123046875,
0.5830078125,
0.5634765625,
0.5830078125,
0.5634765625,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.60205078125,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.60205078125,
0.595703125
],
[
0.1158447265625,
0.5888671875,
0.48779296875,
0.5888671875,
0.48779296875,
0.6025390625,
0.1158447265625,
0.6025390625
],
[
0.51318359375,
0.59423828125,
0.54248046875,
0.59423828125,
0.54248046875,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.60205078125,
0.59326171875,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.60205078125,
0.60693359375
],
[
0.1158447265625,
0.60498046875,
0.403076171875,
0.60498046875,
0.403076171875,
0.6181640625,
0.1158447265625,
0.6181640625
],
[
0.60205078125,
0.60791015625,
0.75830078125,
0.60791015625,
0.75830078125,
0.6181640625,
0.60205078125,
0.6181640625
],
[
0.51318359375,
0.62255859375,
0.54736328125,
0.62255859375,
0.54736328125,
0.63427734375,
0.51318359375,
0.63427734375
],
[
0.60205078125,
0.62158203125,
0.8818359375,
0.62158203125,
0.8818359375,
0.6357421875,
0.60205078125,
0.6357421875
],
[
0.118408203125,
0.6328125,
0.151123046875,
0.6328125,
0.151123046875,
0.64453125,
0.118408203125,
0.64453125
],
[
0.1427001953125,
0.63134765625,
0.32568359375,
0.6328125,
0.32568359375,
0.64794921875,
0.1427001953125,
0.6455078125
],
[
0.5146484375,
0.63525390625,
0.587890625,
0.63525390625,
0.587890625,
0.646484375,
0.5146484375,
0.646484375
],
[
0.60107421875,
0.6328125,
0.82861328125,
0.6328125,
0.82861328125,
0.646484375,
0.60107421875,
0.646484375
],
[
0.1175537109375,
0.65673828125,
0.489501953125,
0.65673828125,
0.489501953125,
0.67138671875,
0.1175537109375,
0.67138671875
],
[
0.6171875,
0.65771484375,
0.7734375,
0.65625,
0.7734375,
0.66943359375,
0.6171875,
0.67138671875
],
[
0.1158447265625,
0.67333984375,
0.489501953125,
0.6728515625,
0.489501953125,
0.68701171875,
0.1158447265625,
0.6884765625
],
[
0.51318359375,
0.67578125,
0.55517578125,
0.67578125,
0.55517578125,
0.68701171875,
0.51318359375,
0.68701171875
],
[
0.60205078125,
0.6748046875,
0.8818359375,
0.6748046875,
0.8818359375,
0.6884765625,
0.60205078125,
0.6884765625
],
[
0.51318359375,
0.6884765625,
0.57421875,
0.6884765625,
0.57421875,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.60205078125,
0.68701171875,
0.765625,
0.68701171875,
0.765625,
0.70068359375,
0.60205078125,
0.70068359375
],
[
0.51318359375,
0.70361328125,
0.5615234375,
0.70361328125,
0.5615234375,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.1175537109375,
0.6904296875,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.1175537109375,
0.70361328125
],
[
0.1158447265625,
0.7060546875,
0.489501953125,
0.7060546875,
0.489501953125,
0.7197265625,
0.1158447265625,
0.7197265625
],
[
0.603515625,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71435546875,
0.603515625,
0.71435546875
],
[
0.51318359375,
0.71630859375,
0.56689453125,
0.71630859375,
0.56689453125,
0.72802734375,
0.51318359375,
0.72802734375
],
[
0.60107421875,
0.71533203125,
0.88330078125,
0.71533203125,
0.88330078125,
0.728515625,
0.60107421875,
0.728515625
],
[
0.11419677734375,
0.7216796875,
0.486083984375,
0.7216796875,
0.486083984375,
0.7353515625,
0.11419677734375,
0.7353515625
],
[
0.603515625,
0.72998046875,
0.63623046875,
0.72998046875,
0.63623046875,
0.73828125,
0.603515625,
0.73828125
],
[
0.1158447265625,
0.73828125,
0.489501953125,
0.73828125,
0.489501953125,
0.751953125,
0.1158447265625,
0.751953125
],
[
0.51123046875,
0.74365234375,
0.55517578125,
0.74365234375,
0.55517578125,
0.7548828125,
0.51123046875,
0.7548828125
],
[
0.603515625,
0.74462890625,
0.880859375,
0.74462890625,
0.880859375,
0.7548828125,
0.603515625,
0.7548828125
],
[
0.51123046875,
0.75439453125,
0.56689453125,
0.75439453125,
0.56689453125,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.60205078125,
0.75439453125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76708984375,
0.60205078125,
0.76708984375
],
[
0.60546875,
0.7685546875,
0.74462890625,
0.7685546875,
0.74462890625,
0.779296875,
0.60546875,
0.779296875
],
[
0.1175537109375,
0.7548828125,
0.490234375,
0.7548828125,
0.490234375,
0.7685546875,
0.1175537109375,
0.7685546875
],
[
0.1158447265625,
0.77099609375,
0.489501953125,
0.77099609375,
0.489501953125,
0.78369140625,
0.1158447265625,
0.78369140625
],
[
0.51123046875,
0.78369140625,
0.5634765625,
0.78369140625,
0.5634765625,
0.7958984375,
0.51123046875,
0.7958984375
],
[
0.60205078125,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.60205078125,
0.79638671875
],
[
0.60205078125,
0.7958984375,
0.8818359375,
0.7958984375,
0.8818359375,
0.80859375,
0.60205078125,
0.80859375
],
[
0.118408203125,
0.7880859375,
0.48779296875,
0.7880859375,
0.48779296875,
0.798828125,
0.118408203125,
0.798828125
],
[
0.51123046875,
0.79638671875,
0.5439453125,
0.79638671875,
0.5439453125,
0.80810546875,
0.51123046875,
0.80810546875
],
[
0.1175537109375,
0.802734375,
0.490234375,
0.802734375,
0.490234375,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.509765625,
0.80712890625,
0.603515625,
0.80859375,
0.603515625,
0.82080078125,
0.509765625,
0.818359375
],
[
0.60107421875,
0.80859375,
0.82958984375,
0.80859375,
0.82958984375,
0.8193359375,
0.60107421875,
0.8193359375
],
[
0.1175537109375,
0.81787109375,
0.486083984375,
0.818359375,
0.486083984375,
0.8330078125,
0.1175537109375,
0.83203125
],
[
0.1175537109375,
0.8349609375,
0.490234375,
0.8349609375,
0.490234375,
0.8486328125,
0.1175537109375,
0.8486328125
],
[
0.509765625,
0.83740234375,
0.88330078125,
0.83740234375,
0.88330078125,
0.85107421875,
0.509765625,
0.85107421875
],
[
0.11419677734375,
0.85107421875,
0.399658203125,
0.849609375,
0.399658203125,
0.8642578125,
0.11419677734375,
0.865234375
],
[
0.51318359375,
0.8515625,
0.87890625,
0.8515625,
0.87890625,
0.865234375,
0.51318359375,
0.865234375
],
[
0.51123046875,
0.8662109375,
0.8818359375,
0.8662109375,
0.8818359375,
0.8798828125,
0.51123046875,
0.8798828125
],
[
0.51123046875,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.89453125,
0.51123046875,
0.89453125
],
[
0.1158447265625,
0.89013671875,
0.483642578125,
0.89111328125,
0.483642578125,
0.90625,
0.1158447265625,
0.90478515625
],
[
0.5146484375,
0.8955078125,
0.84814453125,
0.8955078125,
0.84814453125,
0.91015625,
0.5146484375,
0.91015625
],
[
0.118408203125,
0.90478515625,
0.344970703125,
0.90478515625,
0.344970703125,
0.91845703125,
0.118408203125,
0.91845703125
]
] |
[
"3.3 Frame Assignment",
"Frame Inventory",
"Any controversial topic can be discussed from dif-",
"Capacity & Resources",
"Fairness & Equality",
"ferent perspectives. For example, \"the dangers of",
"Constitutionality&",
"Health & Safety",
"social media can be discussed from a moral or a",
"Jurisprudence",
"Morality",
"Crime & Punishment",
"Policy Prescription & Evaluation",
"Cultural Identity",
"health perspective, among others. In our indicative",
"summaries, we use argumentation frame labels as",
"Political",
"Economic",
"Public Opinion",
"External Regulation &",
"top-level headings to operationalize different per-",
"spectives. An argumentation frame may include",
"Quality of Life",
"Reputation",
"Security & Defense",
"one or more groups of relevant arguments.We",
"assign frame labels from the issue-generic frame",
"Table 1: Inventory of frames proposed by Boydstun et al.",
"inventory shown in Table 1 (Boydstun et al., 2014)",
"(2014) to track the media's framing of policy issues.",
"to each cluster label derived in the previous step.7",
"We use prompt-based models in both zero-shot",
"Model",
"Description",
"and few-shot settings for frame assignment.In",
"Variants",
"our experiments with instruction-tuned models.",
"Pre-InstructGPT",
"we designed two types of instructions, shown in",
"To",
"Encoder-decoder model trained on datasets",
"Figure 1O,namely directinstructionsfor models",
"vanilla",
"transformed as task-specific prompts.",
"trained on instruction-response samples, and dia-",
"BLOOM",
"A multlingual autoregressive model with 176B",
"vanilla",
"parameters forprompt-based text completion.",
"GPT-NeoX",
"log instructions for chat models. The instructions",
"are included along with the cluster labels in the",
"Open source alternative to GPT-3.",
"20B",
"prompts. Moreover, including the citation of the",
"OPT",
"Autoregressive model with similar effective-",
"frame inventory used in our experiments has a posi-",
"66B",
"ness to GPT-3, but more effcient data collec-",
"tive effect on the effectiveness of some models (see",
"tion and training.",
"Appendix D.1 for details)",
"Direct Instruction",
"LLaMA-CoT",
"LLaMA-30B fine-tuned on chain-of-thought",
"and reasoning samples (Si and Lin, 2023).",
"3.4Indicative Summary Presentation",
"vanilla",
"Given the generated labels of all sentence clusters",
"Alpaca",
"LLaMA-7B fine-tuned based on 52k self-",
"and the frame labels assigned to each cluster label,",
"7B",
"instruct responses (Wang et al., 2022).",
"our indicative summary groups the cluster labels",
"OASST",
"LLaMA-30B fine-tuned on the OpenAssistant",
" vanilla",
"Conversations dataset (Kopf et al., 2023) using",
"reinforcement learning",
"by their respective frame labels. The cluster label",
"groups of each frame label are then ordered by",
"Pythia",
"Suite of LLMs trained on public data to in-",
"cluster size. This results in a two-level indicative",
"12B",
"vestigate the effects of training and scaling on",
"summary, as shown in Figures 1 and 4.",
"various model properties.",
"GPT*",
"OpenAI models GPT3.5 (text-davinci-003),",
"4",
"PromptEngineering",
"3.5, Chat, 4",
"ChatGPT (gpt-3.5-turbo), and GPT4.",
"Using prompt-based LLMs for generative clus-",
"Dialogue Instruction",
"ter labeling and frame assignment requires model-",
"LLaMA",
"Suite of open-source LLMs from Meta AI",
"30B,65B",
"trained on public datasets.",
"Vicuna",
"specific prompt engineering as a preliminary step.",
"We explored the 19 model variants listed in Table 2.",
"LLaMA models fine-tuned using conversa-",
"7B, 13B",
"tions collected by ShareGPT (https://sharegpt.",
"To select the most appropriate models for our task",
"com)",
"we consulted the HELM benchmark (Liang et al.,",
"Baize",
"Open source chat model trained on 100k di-",
"7B, 13B",
"alogues generated by letting ChatGPT (GPT",
"3.5-turbo)talk to itself.",
"2022), which compares the effectiveness of differ-",
"ent LLMs for different tasks. Further, we have in-",
"Falcon",
"Trained on the RefinedWeb corpus (Penedo",
"et al., 2023), which was obtained by filtering",
"cluded variousrecentlyreleased open sourcemod-",
"40B,",
"els (with optimized instructions) as they were re-",
"40B-Instruct",
"and deduplication of publicweb data",
"leased. Since many of them were released during",
"our research, we reuse prompts previously opti-",
"Table 2: LLMs studied for cluster labeling and frame as-",
"mized prompts for the newer models.8",
"signment. Older models are listed byPre-InstructGPT",
"(prior to GPT3.5) and newer models are listed by",
"their respective prompt types investigated (Direct/",
"7For detailed label descriptions see Table 6 in the Appendix.",
"Dialogue). See Appendices C and D for details.",
"° See Appendices C and D for details."
] |
{
"arxiv_id": "2311.01882",
"language": "en",
"page_idx": 4
}
|
||
2010.14678-en-0
|
2010.14678-en-0.jpg
|
# $\textbf{What Does This Acronym Mean? Introducing a New Dataset for}$ $\textbf{Acronym Identification and Disambiguation}$
Amir Pouran Ben Veyseh$^1$,Franck Dernoncourt$^2,\textbf{Quan Hung Tran}^2$,
## and Thien Huu Nguyen
$^{1}$ Department of Computer and Information Science, University of Oregon, USA
$^{2}$Adobe Research, San Jose, CA, USA
$\{$apouranb,thien$\}$cs.uoregon.edu,
$\{$franck.dernoncourt,qtran$\}$eadobe.com
## Abstract
Acronyms are the short forms of phrases that facilitate conveying lengthy sentences in documents and serve as one of the mainstays of writing. Due to their importance, identifying acronyms and corresponding phrases (i.e., acronym identification (AI)) and finding the correct meaning of each acronym (i.e., acronym disambiguation (AD)) are crucial for text understanding. Despite the recent progress on this task, there are some limitations in the existing datasets which hinder further improvement. More specifically, limited size of manually annotated AI datasets or noises in the automatically created acronym identification datasets obstruct designing advanced highperforming acronym identification models. Moreover, the existing datasets are mostly limited to the medical domain and ignore other domains. In order to address these two limitations, we first create a manually annotated large AI dataset for scientific domain. This dataset contains 17,506 sentences which is substantially larger than previous scientific AI datasets. Next, we prepare an AD dataset for scientific domain with 62,441 samples which is significantly larger than previous scientific AD dataset. Our experiments show that the existing state-of-the-art models fall far behind human-level performance on both datasets proposed by this work. In addition, we propose a new deep learning model which utilizes the syntactical structure of the sentence to expand an ambiguous acronym in a sentence. The proposed model outperforms the state-of-the-art models on the new AD dataset, providing a strong baseline for future research on this dataset $^{1}.$
## 1 Introduction
Acronyms are shortened forms of a longer phrase. As a running example, in the sentence“The main key performance indicator, herein referred to as KPI, is the E2E throughput" there are two acronyms KPI and $E2E$. Also, the acronym $KPI$ refers to the phrase key performance indicator (a.k.a. the long form of the acronym KPI). In written language, acronyms are prevalent in technical documents that helps to avoid the repetition of long and cumbersome terms, thus saving text space. For instance, about 15% of PubMed queries include abbreviations, and about $14.8\%$ of all tokens in a clinical note dataset are abbreviations (Islamaj Dogan et al., 2009; Xu et al., 2007; Jin et al., 2019).
Considering the widespread use of acronyms in texts, a text processing application, such as question answering or document retrieval, should be able to correctly process the acronyms in the text and find their meanings. To this end, two sub-tasks should be solved: Acronym Identification (AI): to find the acronyms and the phrases that have been abbreviated by the acronyms in the document. In the running example, the acronyms $KPI$ and $E2E$ and the phrase key performance indicator should be extracted. Acronym Disambiguation (AD): to find the right meaning for a given acronym in text. In the running example, the systems should be able to find the right meanings of the two acronyms $KPI$ and $E2E.$ Note that while the meaning of $KPI$ is found in the sentence, the meaning of $E2E$ should be inferred from its occurrences either in the previous text of the same document or external resources (e.g., dictionaries). Acronym Identification and disambiguation can be used in many downstream applications including slotfilling (Veyseh et al., 2020), definition extraction (Kang et al., 2020; Veyseh et al., 2020), and question
l Dataset for AI is available at https://github.com/amirveyseh/AAAAI-21-SDU-shared-task-1-AI and dataset for AD is avail-
able at https://github.com/amirveyseh/AAAI-21-SDU-shared-task-2-AD
|
[
[
0.17041015625,
0.07598876953125,
0.8310546875,
0.07598876953125,
0.8310546875,
0.09027099609375,
0.17041015625,
0.09027099609375
],
[
0.273681640625,
0.0955810546875,
0.72900390625,
0.0955810546875,
0.72900390625,
0.10986328125,
0.273681640625,
0.10986328125
],
[
0.208251953125,
0.122314453125,
0.7958984375,
0.122314453125,
0.7958984375,
0.13720703125,
0.208251953125,
0.13720703125
],
[
0.39794921875,
0.14013671875,
0.60546875,
0.14013671875,
0.60546875,
0.155029296875,
0.39794921875,
0.155029296875
],
[
0.17626953125,
0.152587890625,
0.82958984375,
0.15380859375,
0.82958984375,
0.171630859375,
0.17626953125,
0.17041015625
],
[
0.343505859375,
0.171630859375,
0.6630859375,
0.1728515625,
0.6630859375,
0.18701171875,
0.343505859375,
0.1859130859375
],
[
0.316650390625,
0.190673828125,
0.69189453125,
0.190673828125,
0.69189453125,
0.204833984375,
0.316650390625,
0.204833984375
],
[
0.28955078125,
0.2060546875,
0.72119140625,
0.207275390625,
0.72119140625,
0.221435546875,
0.28955078125,
0.2203369140625
],
[
0.4609375,
0.24462890625,
0.54248046875,
0.24462890625,
0.54248046875,
0.262451171875,
0.4609375,
0.262451171875
],
[
0.1485595703125,
0.27490234375,
0.85400390625,
0.27490234375,
0.85400390625,
0.289306640625,
0.1485595703125,
0.289306640625
],
[
0.14697265625,
0.290283203125,
0.85546875,
0.29150390625,
0.85546875,
0.305908203125,
0.14697265625,
0.3046875
],
[
0.0251922607421875,
0.312255859375,
0.060455322265625,
0.312255859375,
0.060455322265625,
0.438232421875,
0.0251922607421875,
0.438232421875
],
[
0.1485595703125,
0.30810546875,
0.85400390625,
0.30810546875,
0.85400390625,
0.322509765625,
0.1485595703125,
0.322509765625
],
[
0.14697265625,
0.32470703125,
0.85400390625,
0.32470703125,
0.85400390625,
0.33837890625,
0.14697265625,
0.33837890625
],
[
0.1485595703125,
0.34033203125,
0.85400390625,
0.34033203125,
0.85400390625,
0.35400390625,
0.1485595703125,
0.35400390625
],
[
0.145263671875,
0.355224609375,
0.85546875,
0.35693359375,
0.85546875,
0.37060546875,
0.145263671875,
0.3681640625
],
[
0.14697265625,
0.372314453125,
0.85400390625,
0.372314453125,
0.85400390625,
0.385986328125,
0.14697265625,
0.385986328125
],
[
0.1485595703125,
0.388427734375,
0.8564453125,
0.388427734375,
0.8564453125,
0.402099609375,
0.1485595703125,
0.402099609375
],
[
0.1502685546875,
0.40625,
0.85400390625,
0.40625,
0.85400390625,
0.416259765625,
0.1502685546875,
0.416259765625
],
[
0.14697265625,
0.42041015625,
0.85546875,
0.42041015625,
0.85546875,
0.43408203125,
0.14697265625,
0.43408203125
],
[
0.0218353271484375,
0.43408203125,
0.06463623046875,
0.435302734375,
0.0621337890625,
0.5185546875,
0.019317626953125,
0.51708984375
],
[
0.1485595703125,
0.435791015625,
0.85546875,
0.43701171875,
0.85546875,
0.45068359375,
0.1485595703125,
0.449462890625
],
[
0.14697265625,
0.45068359375,
0.85400390625,
0.451904296875,
0.85400390625,
0.466064453125,
0.14697265625,
0.465576171875
],
[
0.1502685546875,
0.47021484375,
0.85205078125,
0.47021484375,
0.85205078125,
0.48095703125,
0.1502685546875,
0.48095703125
],
[
0.1485595703125,
0.485107421875,
0.85400390625,
0.485107421875,
0.85400390625,
0.498779296875,
0.1485595703125,
0.498779296875
],
[
0.1485595703125,
0.501953125,
0.85546875,
0.501953125,
0.85546875,
0.515625,
0.1485595703125,
0.515625
],
[
0.0218353271484375,
0.5185546875,
0.0621337890625,
0.5185546875,
0.0621337890625,
0.728515625,
0.0218353271484375,
0.728515625
],
[
0.1485595703125,
0.51708984375,
0.85400390625,
0.51708984375,
0.85400390625,
0.53076171875,
0.1485595703125,
0.53076171875
],
[
0.14697265625,
0.533203125,
0.78515625,
0.533203125,
0.78515625,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.1292724609375,
0.55859375,
0.26123046875,
0.56005859375,
0.26123046875,
0.57470703125,
0.1292724609375,
0.57373046875
],
[
0.124267578125,
0.5654296875,
0.130126953125,
0.5654296875,
0.130126953125,
0.57080078125,
0.124267578125,
0.57080078125
],
[
0.12005615234375,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.60107421875,
0.12005615234375,
0.60107421875
],
[
0.1175537109375,
0.6025390625,
0.88330078125,
0.6015625,
0.88330078125,
0.6162109375,
0.1175537109375,
0.61767578125
],
[
0.118408203125,
0.61767578125,
0.88330078125,
0.6181640625,
0.88330078125,
0.6328125,
0.118408203125,
0.6318359375
],
[
0.118408203125,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64794921875,
0.118408203125,
0.64794921875
],
[
0.12005615234375,
0.65087890625,
0.8818359375,
0.65087890625,
0.8818359375,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.12005615234375,
0.66650390625,
0.88330078125,
0.66650390625,
0.88330078125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.12005615234375,
0.68310546875,
0.6689453125,
0.68310546875,
0.6689453125,
0.69677734375,
0.12005615234375,
0.69677734375
],
[
0.1368408203125,
0.69970703125,
0.88330078125,
0.69970703125,
0.88330078125,
0.71337890625,
0.1368408203125,
0.71337890625
],
[
0.12005615234375,
0.71630859375,
0.88330078125,
0.71630859375,
0.88330078125,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.118408203125,
0.732421875,
0.88330078125,
0.732421875,
0.88330078125,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.118408203125,
0.74755859375,
0.88330078125,
0.74755859375,
0.88330078125,
0.7626953125,
0.118408203125,
0.7626953125
],
[
0.12005615234375,
0.76416015625,
0.88330078125,
0.76416015625,
0.88330078125,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.118408203125,
0.779296875,
0.884765625,
0.77978515625,
0.884765625,
0.79443359375,
0.118408203125,
0.79345703125
],
[
0.12005615234375,
0.7958984375,
0.88330078125,
0.7958984375,
0.88330078125,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.12005615234375,
0.8125,
0.88330078125,
0.8125,
0.88330078125,
0.826171875,
0.12005615234375,
0.826171875
],
[
0.12005615234375,
0.82763671875,
0.8818359375,
0.82763671875,
0.8818359375,
0.84130859375,
0.12005615234375,
0.84130859375
],
[
0.12176513671875,
0.845703125,
0.880859375,
0.845703125,
0.880859375,
0.85546875,
0.12176513671875,
0.85546875
],
[
0.12005615234375,
0.85986328125,
0.88330078125,
0.85986328125,
0.88330078125,
0.87353515625,
0.12005615234375,
0.87353515625
],
[
0.14111328125,
0.8818359375,
0.8818359375,
0.8828125,
0.8818359375,
0.89794921875,
0.14111328125,
0.896484375
],
[
0.12176513671875,
0.89794921875,
0.5556640625,
0.89794921875,
0.5556640625,
0.90771484375,
0.12176513671875,
0.90771484375
]
] |
[
"What Does This Acronym Mean? Introducing a New Dataset for",
"Acronym Identification and Disambiguation",
"Amir Pouran Ben Veyseh', Franck Dernoncourt?, Quan Hung Tran?",
"and Thien Huu Nguyen'",
"I Department of Computer and Information Science, University of Oregon, USA",
"2 Adobe Research, San Jose, CA, USA",
"[apouranb,thien]@cs.uoregon.edu,",
"{franck.dernoncourt,qtran}@adobe.com",
"Abstract",
"Acronyms are the short forms of phrases that facilitate conveying lengthy sentences in documents",
"and serve as one of the mainstays of writing. Due to their importance, identifying acronyms and",
"201008",
"corresponding phrases (i.e., acronym identification (AD)) and finding the correct meaning of each",
"acronym (i.e., acronym disambiguation (AD)) are crucial for text understanding. Despite the",
"recent progress on this task, there are some limitations in the existing datasets which hinder",
"further improvement.More specifically,limited size ofmanuallyannotated AI datasets or noises",
"in the automatically created acronym identification datasets obstruct designing advanced high-",
"performing acronym identification models. Moreover, the existing datasets are mostly limited to",
"themedicaldomainandignoreotherdomains.Inordertoaddressthesetwolimitations.wefirst",
"create a manually annotated large AI dataset for scientific domain. This dataset contains 17,506",
"['s]",
"sentences which is substantially larger than previous scientific AI datasets. Next, we prepare an",
"AD datasetfor scientific domain with 62,441 samples which is significantly larger than previous",
"scientific AD dataset.Our experiments show thatthe existing state-of-the-artmodels fallfar",
"behind human-level performance on both datasets proposed by this work. In addition, we propose",
"a new deep learning model which utilizes the syntactical structure of the sentence to expand an",
"18101020",
"ambiguous acronym in a sentence. The proposed model outperforms the state-of-the-art models",
"on the new AD dataset, providing a strong baseline for future research on this dataset 1",
"Introduction",
"",
"Acronyms are shortened forms of a longer phrase. As a running example, in the sentence “The main key",
"performance indicator, herein referred to as KPI, is the E2E throughput' there are two acronyms KPI",
"and E2E. Also, the acronym KPI refers to the phrase key performance indicator (a.k.a. the long form",
"of the acronym KPl). In written language, acronyms are prevalent in technical documents that helps to",
"avoid the repetition of long and cumbersome terms, thus saving text space. For instance, about 15%",
"of PubMed queries include abbreviations, and about 14.8% of all tokens in a clinical note dataset are",
"abbreviations (Islamaj Dogan et al., 2009; Xu et al., 2007; Jin et al., 2019).",
"Considering the widespread use of acronyms in texts, a text processing application, such as question",
"answering or document retrieval, should be able to correctly process the acronyms in the text and find",
"their meanings. To this end, two sub-tasks should be solved: Acronym Identification (AD): to find the",
"acronyms and the phrases that have been abbreviated by the acronyms in the document. In the running",
"example, the acronyms KPI and E2E and the phrase key performance indicator should be extracted.",
"Acronym Disambiguation (AD): to find the right meaning for a given acronym in text. In the running",
"example, the systems should be able to find the right meanings of the two acronyms KPI and E2E. Note",
"that while the meaning of KPI is found in the sentence, the meaning of E2E should be inferred from its",
"occurrences either in the previous text of the same document or external resources (e.g., dictionaries).",
"AcronymIdentificationanddisambiguationcanbeusedinmanydownstreamapplicationsincludingslot-",
"filling (Veyseh et al., 2020), definition extraction (Kang et al., 2020; Veyseh et al., 2020), and question",
"'Dataset for AI is available at https://github.com/amirveyseh/AAAI-21-SDU-shared-task-1-AI and dataset for AD is avail-",
"ableathttps://github.com/amirveyseh/AAAI-21-SDU-shared-task-2-AD"
] |
{
"arxiv_id": "2010.14678",
"language": "en",
"page_idx": 0
}
|
||
2010.14678-en-1
|
2010.14678-en-1.jpg
|
answering (Ackermann et al., 2020; Veyseh, 2016)
Over the past two decades, several approaches and resources have been proposed to solve the two subtasks for acronyms. These approaches extend from rule-based methods for AI and feature-based models for AD (i.e., SVM and Naive Bayes) (Schwartz and Hearst, 2002; Nadeau and Turney, 2005; Okazaki and Ananiadou, 2006; Yu et al., 2007) to the recent deep learning methods (Li et al., 2015; Charbonnien and Wartena, 2018; Wu et al., 2015; Ciosici et al., 2019; Jin et al., 2019; Li et al., 2019). While the prior work has made substantial progress on this task by providing new approaches and datasets, there are some limitations in the existing datasets which hinder further improvement. First, most of the existing datasets for AI are either limited in their sizes or created using simple rule-based methods (i.e., not humnanannotated). For instance, Ciosici (2019) exploits the rules proposed by Schwartz (2002) to generate a corpus for acronym disambiguation from Wikipedia. This is unfortunate as rules are in general not able to capture all the diverse forms to express acronyms and their long forms in text (Harris and Srinivasan, 2019). This limitation not only restricts the coverage of the previous methods and datasets for AI, but also imposes some bias to evaluation of AI using these datasets. Second, most of the existing datasets are in the medical domain, ignoring the challenges in other scientific domains. While there are a few datasets for general domain (e.g., Wikipedia), online forums, news and scientific documents (Thakker et al., 2017; Li et al., 2018; Charbonnier and Wartena, 2018; Liu et al., 2011; Harris and Srinivasan, 2019), they still suffer from either noisy examples inevitable in the heuristically generated datasets (Charbonnier and Wartena, 2018; Ciosici et al., 2019; Thakker et al., 2017) or their small size, which makes them inappropriate for training advanced methods (e.g., deep neural networks) (Prokofyev et al., 2013; Harris and Srinivasan, 2019; Nautial et al., 2014).
In order to address the limitations of the existing resources and push forward the research on acronym identification and disambiguation, especially in scientific domain, this paper introduces two new datasets for AI and AD. Notably, our datasets are annotated by human to achieve high quality and have substantially larger numbers of examples than the existing AI datasets in the non-medical domain (see Appendix B.3). Moreover, based on acronyms and long forms obtained by AI annotations, we introduce a new AD dataset that is also substantially larger than the existing AD datasets for the scientific domain. For more details about the datasets, see Appendices B.2 and B.3. Finally, we conduct extensive experiments to evaluate the state-of-the-art methods for AI and AD on the proposed datasets. The experiments show that the existing models fail to match the human-level performance on the proposed datasets
Motivated by the unsatisfactory performance of the existing AD models, we introduce a new deep learning model for acronym disambiguation. In particular, one of the limitations of the prior AD models is that they fail to efficiently encode the context of the ambiguous acronyms in the given sentences On the one hand, the traditional feature-based models for AD are limited in their ability to effective representations for the contexts due to the use of hand-designed features. On the other hand, the current deep learning models for AD, based on either language models or sentence encoders (e.g., LSTMs) to capture the contextual information, cannot effectively encode the long dependencies between words in the sentences. These long dependencies are important for acronym disambiguation as the meanings of the acronyms might depend on some words that are sequentially far from the acronyms of interest. To address the issue of long dependencies in sentences, prior work on other natural language processing (NLP) applications has shown that dependency trees can be employed to capture dependencies between words that are sequentially far from each other (Veyseh et al., 2020). As none of the recent AD models employs the dependency trees to encode the sentence context for acronyms, we propose a novel deep learning model for AD that exploits the syntactic dependency structures of the sentences using graph convolutional neural networks (GCN) (Kipf and Welling, 2016). In this work, the structure information induced by the GCN model would be employed as the structural context for acronym disambiguation. Our experiments show that the proposed model outperforms the existing state-of-the-art AD models, providing a strong baseline for the future work on the proposed dataset for AD.
In summary, in this paper, we make the following contributions:
· We release the first publicly available and the largest manually annotated acronym identification dataset in scientific domain. Moreover, we also release the largest acronym disambiguation dataset
|
[
[
0.12005615234375,
0.07598876953125,
0.491943359375,
0.07598876953125,
0.491943359375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.1376953125,
0.0926513671875,
0.8818359375,
0.0926513671875,
0.8818359375,
0.10687255859375,
0.1376953125,
0.10687255859375
],
[
0.118408203125,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.12005615234375,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1383056640625,
0.12005615234375,
0.1383056640625
],
[
0.12005615234375,
0.1395263671875,
0.8818359375,
0.1395263671875,
0.8818359375,
0.152587890625,
0.12005615234375,
0.152587890625
],
[
0.12005615234375,
0.15673828125,
0.88330078125,
0.15673828125,
0.88330078125,
0.17041015625,
0.12005615234375,
0.17041015625
],
[
0.12005615234375,
0.17333984375,
0.88330078125,
0.17333984375,
0.88330078125,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.12005615234375,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.12005615234375,
0.203125
],
[
0.118408203125,
0.2060546875,
0.8818359375,
0.2060546875,
0.8818359375,
0.2197265625,
0.118408203125,
0.2197265625
],
[
0.12005615234375,
0.221435546875,
0.884765625,
0.221435546875,
0.884765625,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.1175537109375,
0.2369384765625,
0.884765625,
0.236328125,
0.884765625,
0.25048828125,
0.1175537109375,
0.251708984375
],
[
0.12005615234375,
0.254150390625,
0.88330078125,
0.254150390625,
0.88330078125,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.12005615234375,
0.26953125,
0.88330078125,
0.26953125,
0.88330078125,
0.283203125,
0.12005615234375,
0.283203125
],
[
0.12176513671875,
0.2861328125,
0.880859375,
0.2861328125,
0.880859375,
0.29638671875,
0.12176513671875,
0.29638671875
],
[
0.12005615234375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.12005615234375,
0.318359375,
0.88330078125,
0.318359375,
0.88330078125,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.12005615234375,
0.333251953125,
0.880859375,
0.333251953125,
0.880859375,
0.3466796875,
0.12005615234375,
0.3466796875
],
[
0.12005615234375,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.12005615234375,
0.364013671875
],
[
0.118408203125,
0.366455078125,
0.8818359375,
0.366455078125,
0.8818359375,
0.380126953125,
0.118408203125,
0.380126953125
],
[
0.118408203125,
0.383056640625,
0.88330078125,
0.383056640625,
0.88330078125,
0.396728515625,
0.118408203125,
0.396728515625
],
[
0.12005615234375,
0.3984375,
0.43408203125,
0.3984375,
0.43408203125,
0.412109375,
0.12005615234375,
0.412109375
],
[
0.1368408203125,
0.41455078125,
0.88330078125,
0.41455078125,
0.88330078125,
0.4287109375,
0.1368408203125,
0.4287109375
],
[
0.118408203125,
0.43115234375,
0.88330078125,
0.43115234375,
0.88330078125,
0.4453125,
0.118408203125,
0.4453125
],
[
0.118408203125,
0.44775390625,
0.88330078125,
0.44775390625,
0.88330078125,
0.46142578125,
0.118408203125,
0.46142578125
],
[
0.1175537109375,
0.4619140625,
0.88330078125,
0.463134765625,
0.88330078125,
0.47802734375,
0.1175537109375,
0.476806640625
],
[
0.12005615234375,
0.478515625,
0.884765625,
0.478515625,
0.884765625,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.12005615234375,
0.495361328125,
0.884765625,
0.495361328125,
0.884765625,
0.5087890625,
0.12005615234375,
0.5087890625
],
[
0.12005615234375,
0.51123046875,
0.88330078125,
0.51123046875,
0.88330078125,
0.52490234375,
0.12005615234375,
0.52490234375
],
[
0.118408203125,
0.52685546875,
0.88330078125,
0.52685546875,
0.88330078125,
0.54150390625,
0.118408203125,
0.54150390625
],
[
0.118408203125,
0.54345703125,
0.8017578125,
0.54345703125,
0.8017578125,
0.55712890625,
0.118408203125,
0.55712890625
],
[
0.1368408203125,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.1368408203125,
0.572265625
],
[
0.118408203125,
0.576171875,
0.88330078125,
0.576171875,
0.88330078125,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.118408203125,
0.5927734375,
0.8818359375,
0.5927734375,
0.8818359375,
0.60546875,
0.118408203125,
0.60546875
],
[
0.12005615234375,
0.60791015625,
0.8818359375,
0.60791015625,
0.8818359375,
0.62158203125,
0.12005615234375,
0.62158203125
],
[
0.12005615234375,
0.62451171875,
0.88330078125,
0.62451171875,
0.88330078125,
0.63818359375,
0.12005615234375,
0.63818359375
],
[
0.118408203125,
0.64013671875,
0.88330078125,
0.64013671875,
0.88330078125,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.1175537109375,
0.65625,
0.88330078125,
0.65478515625,
0.88330078125,
0.66943359375,
0.1175537109375,
0.67041015625
],
[
0.118408203125,
0.6728515625,
0.884765625,
0.6728515625,
0.884765625,
0.68603515625,
0.118408203125,
0.68603515625
],
[
0.118408203125,
0.6884765625,
0.88330078125,
0.6884765625,
0.88330078125,
0.703125,
0.118408203125,
0.703125
],
[
0.12005615234375,
0.705078125,
0.88330078125,
0.705078125,
0.88330078125,
0.7197265625,
0.12005615234375,
0.7197265625
],
[
0.12176513671875,
0.72265625,
0.88330078125,
0.72265625,
0.88330078125,
0.73291015625,
0.12176513671875,
0.73291015625
],
[
0.12005615234375,
0.73681640625,
0.88330078125,
0.73681640625,
0.88330078125,
0.75048828125,
0.12005615234375,
0.75048828125
],
[
0.12005615234375,
0.75439453125,
0.88330078125,
0.75439453125,
0.88330078125,
0.76708984375,
0.12005615234375,
0.76708984375
],
[
0.118408203125,
0.7685546875,
0.88330078125,
0.76953125,
0.88330078125,
0.783203125,
0.118408203125,
0.7822265625
],
[
0.12005615234375,
0.78515625,
0.88330078125,
0.78515625,
0.88330078125,
0.798828125,
0.12005615234375,
0.798828125
],
[
0.12176513671875,
0.802734375,
0.880859375,
0.802734375,
0.880859375,
0.8134765625,
0.12176513671875,
0.8134765625
],
[
0.12005615234375,
0.81787109375,
0.8818359375,
0.81787109375,
0.8818359375,
0.83056640625,
0.12005615234375,
0.83056640625
],
[
0.118408203125,
0.83203125,
0.7021484375,
0.8330078125,
0.7021484375,
0.84716796875,
0.118408203125,
0.8466796875
],
[
0.1368408203125,
0.849609375,
0.611328125,
0.849609375,
0.611328125,
0.8642578125,
0.1368408203125,
0.8642578125
],
[
0.14111328125,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.14111328125,
0.8935546875
],
[
0.154541015625,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.91015625,
0.154541015625,
0.91015625
]
] |
[
"answering (Ackermann et al., 2020; Veyseh, 2016)",
"Over the past two decades, several approaches and resources have been proposed to solve the two sub-",
"tasks for acronyms. These approaches extend from rule-based methods for AI and feature-based models",
"for AD (i.e., SVM and Naive Bayes) (Schwartz and Hearst, 2002; Nadeau and Turney, 2005; Okazaki",
"and Ananiadou, 2006: Yu et al., 2007) to the recent deep learning methods (Li et al.. 2015: Charbonnier",
"and Wartena, 2018; Wu et al., 2015; Ciosici et al., 2019; Jin et al., 2019; Li et al., 2019). While the prior",
"work has made substantial progress on this task by providing new approaches and datasets, there are some",
"limitations in the existing datasets which hinder further improvement. First, most of the existing datasets",
"for Al are either limited in their sizes or created using simple rule-based methods (i.e., not humnan-",
"annotated). For instance, Ciosici (2019) exploits the rules proposed by Schwartz (2002) to generate a",
"corpus for acronym disambiguation from Wikipedia. This is unfortunate as rules are in general not able",
"to capture all the diverse forms to express acronyms and their long forms in text (Harris and Srinivasan,",
"2019). This limitation not only restricts the coverage of the previous methods and datasets for AI, but",
"also imposes somebiasto evaluation of AIusing these datasets.Second.most of the existing datasets",
"are in the medical domain, ignoring the challenges in other scientific domains. While there are a few",
"datasets for general domain (e.g., Wikipedia), online forums, news and scientific documents (Thakker et",
"al., 2017: Li et al., 2018: Charbonnier and Wartena, 2018; Liu et al., 2011: Harris and Srinivasan, 2019).",
"they still suffer from either noisy examples inevitable in the heuristically generated datasets (Charbonnier",
"and Wartena, 2018; Ciosici et al., 2019; Thakker et al., 2017) or their small size, which makes them",
"inappropriate for training advanced methods (e.g., deep neural networks) (Prokofyev et al., 2013; Harris",
"and Srinivasan, 2019; Nautial et al., 2014).",
"In order to address the limitations of the existing resources and push forward the research on acronym",
"identification and disambiguation, especially in scientific domain, this paper introduces two new datasets",
"for AI and AD. Notably, our datasets are annotated by human to achieve high quality and have substan-",
"tially larger numbers of examples than the existing AI datasets in the non-medical domain (see Appendix",
"B.3). Moreover, based on acronyms and long forms obtained by AI annotations, we introduce a new AD",
"dataset that is also substantially larger than the existing AD datasets for the scientific domain. For more",
"details about the datasets, see Appendices B.2 and B.3. Finally, we conduct extensive experiments to",
"evaluate the state-of-the-art methods for AI and AD on the proposed datasets. The experiments show",
"that the existing models fail to match the human-level performance on the proposed datasets.",
"Motivated by the unsatisfactory performance of the existing AD models, we introduce a new deep",
"learning model for acronym disambiguation. In particular, one of the limitations of the prior AD models",
"is that they fail to efficiently encode the context of the ambiguous acronyms in the given sentences.",
"On the one hand, the traditional feature-based models for AD are limited in their ability to effective",
"representations for the contexts due to the use of hand-designed features. On the other hand, the current",
"deep learning models for AD, based on either language models or sentence encoders (e.g., LSTMs) to",
"capture the contextual information, cannot effectively encode the long dependencies between words in",
"the sentences. These long dependencies are important for acronym disambiguation as the meanings of",
"the acronyms might depend on some words that are sequentially far from the acronyms of interest. To",
"address the issue of long dependencies in sentences, prior work on other natural language processing",
"(NLP)applicationshasshownthatdependencytreescanbeemployedtocapturedependenciesbetween",
"words that are sequentially far from each other (Veyseh et al., 202O). As none of the recent AD models",
"employs the dependency trees to encode the sentence context for acronyms, we propose a novel deep",
"learning model for AD that exploits the syntactic dependency structures of the sentences using graph",
"convolutional neural networks (GCN) (Kipf and Welling, 2016). In this work, the structure information",
"inducedbytheGCNmodel wouldbeemployed as the structural contextforacronym disambiguation.",
"Our experiments show that the proposed model outperforms the existing state-of-the-art AD models.",
"providing a strong baseline for the future work on the proposed dataset for AD.",
"In summary, in this paper, we make the following contributions:",
". We release the first publicly available and the largest manually annotated acronym identification",
"dataset in scientific domain. Moreover, we also release the largest acronym disambiguation dataset"
] |
{
"arxiv_id": "2010.14678",
"language": "en",
"page_idx": 1
}
|
||
2010.14678-en-2
|
2010.14678-en-2.jpg
|
in scientific domain which is created from humanly curated dictionary of ambiguous acronyms
· We conduct extensive experiments on the proposed datasets and compare the performance of the
state-of-the-art acronym identification and disambiguation systems.
· We propose a new graph-based acronym disambiguation method that employs the syntactic structure of the sentence to predict the expanded form of the ambiguous acronym. This baseline outperforms existing state-of-the-art models for acronym disambiguation.
# 2 Data Collection and Annotation
In order to prepare a corpus for acronym annotation, we collect a corpus of 6,786 English papers from arXiv. These papers consist of 2,031,592 sentences that would be used for data annotation for AI and AD in this work.
# 2.1 AI Dataset
AI aims to identify acronyms (i.e., short forms) and phrases that have been abbreviated by the acronyms in text (i.e., long forms). To create a dataset for this task, we manually annotate the sentences from the collected papers. In order to improve the efficiency of the data annotation, we seek to annotate only a set of sentences $S=\{s_1,s_2,...,s_n\}$ that have a high chance to host short forms or long forms for acronyms. As such, each sentence $s_i$ for annotation needs to contain at least one word $w_a$ in which more than half of the characters in $w_a$ are capital letters (i.e., acronym candidates). Afterward, for each sentence $s_i$ in $S$, we search for a sub-sequence of words $W_t=\{w_t,w_{t+1},...,w_{t+k}\}$ in $s_i$ in which the concatenation of the first one, two or three characters of the words $w_j\in W_t$ (in the order of the words in the sub-sequence $W_t)$ could form an acronym candidate $w_a$ in $s_i^2.$ We call$W_t$ as long form candidate. If we cannot find any long form candidate $W_t$ in sentence $s_i$, we remove sentence $s_i$ from $S.$ Using this process, we end up with 17,506 sentences in $S$ to be annotated manually by the annotators from Amazon Mechanical Turk (MTurk). In particular, we create a HIT for each sentence and ask the workers to annotate the short forms and the long forms in the sentence. Also, we ask them to map each long form to its corresponding short form. In order to work on these HITs, workers should have an approval rate of more than 99% and pass a qualification test. Moreover, we remove the annotations from workers who submit the HIT sooner than a preset time or have more than $10\%$ of their annotations different from with the others'. We hire three workers per HIT and pay each of them $0.05. In case of disagreements, if two out of three workers agree on an annotation, we use majority voting to decide the correct annotation. Otherwise, a fourth annotator is hired to resolve the conflict. The inter-annotator agreement (IAA) using Krippendorff’s alpha (Krippendorff, 2011) with the MASI distance metric (Passonneau, 2006) for short-forms (i.e., acronyms) is 0.80 and for long-forms (i.e., phrases) is 0.86. Out of the 17,506 annotated sentences, 1% and 24% of the sentences do not contain any short form or long form, respectively. We call this dataset $\mathbf{SciAI}$ (scientific acronym identification). In this work, the AI task is formulated as sequence labeling where we provide the boundaries for the acronyms and long forms in the sentence using BIO format (i.e., label set includes B-acronym, I-acronym, B-long, I-long and $O).$
# 2.2 $\mathbf{AD}$ dataset
Using the AI dataset annotated in the previous section, we first create a dictionary of acronym meanings (i.e., a mapping from short forms to all of the possible long forms). First, we create a meaning candidate set$^{3}$ for each acronym based on the annotations in AI dataset. As some of the long forms in the meaning set might be different variants of a single meaning, we need to normalize the long forms of each acronym to find the canonical set of long forms of the acronym. To achieve this goal, we use the normalization approach proposed by the prior work (Ciosici et al., 2019). The main idea is to compute the Levenshtein string edit distance between any pair of long forms, and if it is smaller than a threshold, we replace
$^{2}$Note that we also allow some words $w_j\in W$ to be omitted when we make the concatenation of the characters. Also, note
that $w_a$ could be any acronym candidate in $s_i$
$^{3}$In this paper, we use meaning and long form interchangeably
|
[
[
0.154541015625,
0.07659912109375,
0.85205078125,
0.07659912109375,
0.85205078125,
0.09027099609375,
0.154541015625,
0.09027099609375
],
[
0.139404296875,
0.1015625,
0.8818359375,
0.10394287109375,
0.8818359375,
0.1181640625,
0.139404296875,
0.11639404296875
],
[
0.154541015625,
0.12054443359375,
0.6533203125,
0.12054443359375,
0.6533203125,
0.1341552734375,
0.154541015625,
0.1341552734375
],
[
0.14111328125,
0.1478271484375,
0.88330078125,
0.1478271484375,
0.88330078125,
0.162109375,
0.14111328125,
0.162109375
],
[
0.154541015625,
0.163330078125,
0.88330078125,
0.163330078125,
0.88330078125,
0.177001953125,
0.154541015625,
0.177001953125
],
[
0.154541015625,
0.18115234375,
0.603515625,
0.18115234375,
0.603515625,
0.19482421875,
0.154541015625,
0.19482421875
],
[
0.118408203125,
0.207275390625,
0.424072265625,
0.2078857421875,
0.424072265625,
0.22265625,
0.118408203125,
0.221435546875
],
[
0.118408203125,
0.2340087890625,
0.8818359375,
0.2340087890625,
0.8818359375,
0.248779296875,
0.118408203125,
0.248779296875
],
[
0.12005615234375,
0.25048828125,
0.8818359375,
0.25048828125,
0.8818359375,
0.26416015625,
0.12005615234375,
0.26416015625
],
[
0.1175537109375,
0.265380859375,
0.2451171875,
0.2666015625,
0.2451171875,
0.28076171875,
0.1175537109375,
0.27978515625
],
[
0.118408203125,
0.29345703125,
0.2476806640625,
0.29345703125,
0.2476806640625,
0.30712890625,
0.118408203125,
0.30712890625
],
[
0.12005615234375,
0.3154296875,
0.88330078125,
0.3154296875,
0.88330078125,
0.328857421875,
0.12005615234375,
0.328857421875
],
[
0.118408203125,
0.33203125,
0.88330078125,
0.33203125,
0.88330078125,
0.345703125,
0.118408203125,
0.345703125
],
[
0.118408203125,
0.347412109375,
0.88330078125,
0.3466796875,
0.88330078125,
0.361083984375,
0.118408203125,
0.3623046875
],
[
0.12005615234375,
0.364013671875,
0.8818359375,
0.364013671875,
0.8818359375,
0.37890625,
0.12005615234375,
0.37890625
],
[
0.12005615234375,
0.380126953125,
0.88671875,
0.380126953125,
0.88671875,
0.393798828125,
0.12005615234375,
0.393798828125
],
[
0.12005615234375,
0.3955078125,
0.88330078125,
0.3955078125,
0.88330078125,
0.410400390625,
0.12005615234375,
0.410400390625
],
[
0.12005615234375,
0.412109375,
0.884765625,
0.412109375,
0.884765625,
0.427001953125,
0.12005615234375,
0.427001953125
],
[
0.12005615234375,
0.427490234375,
0.884765625,
0.427490234375,
0.884765625,
0.441162109375,
0.12005615234375,
0.441162109375
],
[
0.12005615234375,
0.44482421875,
0.88330078125,
0.44482421875,
0.88330078125,
0.457763671875,
0.12005615234375,
0.457763671875
],
[
0.12005615234375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.473876953125,
0.12005615234375,
0.473876953125
],
[
0.12005615234375,
0.476806640625,
0.88330078125,
0.476806640625,
0.88330078125,
0.490478515625,
0.12005615234375,
0.490478515625
],
[
0.118408203125,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.505859375,
0.118408203125,
0.505859375
],
[
0.1175537109375,
0.50732421875,
0.88330078125,
0.50830078125,
0.88330078125,
0.5224609375,
0.1175537109375,
0.521484375
],
[
0.12005615234375,
0.52490234375,
0.88330078125,
0.52490234375,
0.88330078125,
0.5380859375,
0.12005615234375,
0.5380859375
],
[
0.118408203125,
0.54052734375,
0.88330078125,
0.5390625,
0.88330078125,
0.55419921875,
0.118408203125,
0.5546875
],
[
0.1175537109375,
0.5556640625,
0.88330078125,
0.5546875,
0.88330078125,
0.5693359375,
0.1175537109375,
0.57080078125
],
[
0.118408203125,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.12005615234375,
0.58837890625,
0.884765625,
0.58837890625,
0.884765625,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.12005615234375,
0.60498046875,
0.884765625,
0.60498046875,
0.884765625,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.12005615234375,
0.62060546875,
0.88330078125,
0.62060546875,
0.88330078125,
0.63525390625,
0.12005615234375,
0.63525390625
],
[
0.118408203125,
0.6357421875,
0.8818359375,
0.6357421875,
0.8818359375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.12176513671875,
0.65380859375,
0.8818359375,
0.65380859375,
0.8818359375,
0.66455078125,
0.12176513671875,
0.66455078125
],
[
0.12005615234375,
0.66845703125,
0.8818359375,
0.66845703125,
0.8818359375,
0.681640625,
0.12005615234375,
0.681640625
],
[
0.118408203125,
0.68408203125,
0.8818359375,
0.68505859375,
0.8818359375,
0.69970703125,
0.118408203125,
0.69873046875
],
[
0.1175537109375,
0.69970703125,
0.58837890625,
0.70068359375,
0.58837890625,
0.71533203125,
0.1175537109375,
0.71435546875
],
[
0.12005615234375,
0.72802734375,
0.2509765625,
0.72802734375,
0.2509765625,
0.7421875,
0.12005615234375,
0.7421875
],
[
0.12005615234375,
0.7490234375,
0.8818359375,
0.7490234375,
0.8818359375,
0.76318359375,
0.12005615234375,
0.76318359375
],
[
0.12005615234375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.12005615234375,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.12005615234375,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.12005615234375,
0.830078125,
0.88330078125,
0.830078125,
0.88330078125,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.12005615234375,
0.8466796875,
0.884765625,
0.8466796875,
0.884765625,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.139404296875,
0.87060546875,
0.88330078125,
0.87060546875,
0.88330078125,
0.88427734375,
0.139404296875,
0.88427734375
],
[
0.118408203125,
0.8828125,
0.3896484375,
0.8828125,
0.3896484375,
0.896484375,
0.118408203125,
0.896484375
],
[
0.1427001953125,
0.8955078125,
0.515625,
0.8955078125,
0.515625,
0.9091796875,
0.1427001953125,
0.9091796875
]
] |
[
"in scientific domain which is created from humanly curated dictionary of ambiguous acronyms",
". We conduct extensive experiments on the proposed datasets and compare the performance of the",
"state-of-the-art acronym identification and disambiguation systems.",
". We propose a new graph-based acronym disambiguation method that employs the syntactic structure",
"of the sentence to predict the expanded form of the ambiguous acronym. This baseline outperforms",
"existing state-of-the-art models for acronym disambiguation.",
"2Data Collection and Annotation",
"In order to prepare a corpus for acronym annotation, we collect a corpus of 6,786 English papers from",
"arXiv. These papers consist of 2,031,592 sentences that would be used for data annotation for AI and",
"ADin this work.",
"2.1AI Dataset",
"AI aims to identify acronyms (i.e., short forms) and phrases that have been abbreviated by the acronyms",
"in text (i.e., long forms). To create a dataset for this task, we manually annotate the sentences from the",
"collected papers. In order to improve the efficiency of the data annotation, we seek to annotate only a set",
"of sentences S = [si, s2, ..., Sn} that have a high chance to host short forms or long forms for acronyms.",
"As such, each sentence si for annotation needs to contain at least one word wa in which more than half",
"of the characters in wa are capital letters (i.e., acronym candidates). Afterward, for each sentence s, in",
"S, we search for a sub-sequence of words Wt = [wt, Wt+1, .., wt+k] in s, in which the concatenation of",
"the first one, two or three characters of the words w; E Wt (in the order of the words in the sub-sequence",
"Wt) could form an acronym candidate wa in si?. We call Wt as long form candidate. If we cannot find",
"any long form candidate Wt in sentence si, we remove sentence Si from S. Using this process, we end",
"up with 17,506 sentences in S to be annotated manually by the annotators from Amazon Mechanical",
"Turk (MTurk). In particular, we create a HIT for each sentence and ask the workers to annotate the short",
"forms and the long forms in the sentence. Also, we ask them to map each long form to its corresponding",
"short form. In order to work on these HITs, workers should have an approval rate of more than 99% and",
"pass a qualification test. Moreover, we remove the annotations from workers who submit the HIT sooner",
"than a preset time or have more than 10% of their annotations different from with the others'. We hire",
"three workers per HIT and pay each of them $o.05. In case of disagreements, if two out of three workers",
"agree on an annotation, we use majority voting to decide the correct annotation. Otherwise, a fourth",
"annotator is hired to resolve the conflict. The inter-annotator agreement (IAA) using Krippendorff's",
"alpha (Krippendorff, 2011) with the MASI distance metric (Passonneau, 2006) for short-forms (i.e.,",
"acronyms) is 0.80 and for long-forms (i.e., phrases) is 0.86. Out of the 17,506 annotated sentences, 1%",
"and 24% of the sentences do not contain any short form or long form. respectively. We call this dataset",
"SciAI (scientific acronym identification). In this work, the AI task is formulated as sequence labeling",
"where we provide the boundaries for the acronyms and long forms in the sentence using BIO format (i.e.,",
"label set includes B-acronym, I-acronym, B-long, I-long and O).",
"2.2ADdataset",
"Using the AI dataset annotated in the previous section, we first create a dictionary of acronym meanings",
"(i.e., a mapping from short forms to all of the possible long forms). First, we create a meaning candidate",
"set’ for each acronym based on the annotations in AI dataset. As some of the long forms in the meaning",
"set might be different variants of a single meaning, we need to normalize the long forms of each acronym",
"to find the canonical set of long forms of the acronym. To achieve this goal, we use the normalization",
"approach proposed by the prior work (Ciosici et al., 2019). The main idea is to compute the Levenshtein",
"string edit distance between any pair of long forms, and if it is smaller than a threshold, we replace",
"2Note that we also allow some words wj e W to be omitted when we make the concatenation of the characters. Also, note",
"that wa could be any acronym candidate in s",
"In this paper, we use meaning and long form interchangeably"
] |
{
"arxiv_id": "2010.14678",
"language": "en",
"page_idx": 2
}
|
||
2010.14678-en-3
|
2010.14678-en-3.jpg
|
the less common long form with the more frequent one. After creating the canonical set of long forms for the acronyms, we remove unambiguous acronyms from the mapping/dictionary (i.e., acronyms with only one possible long form), leading to a dictionary $\mathcal{D}$ of ambiguous acronyms. Furthermore, in order to improve the quality of the mappings in $\mathcal{D}$, a graduate student curates the dictionary $\mathcal{D}$ to verify the canonical forms and resolve any potential error. Afterwards, we employ $\mathcal{D}$ to create an AD dataset. To this end, we assume that if an acronym is defined in a paper, its meaning will not change throughout that paper, i.e., the one-sense-per-discourse assumption. Using this criterion, for each acronym $a$ in the dictionary $\mathcal{D}$, we look up all sentences $s_i$ in $S$, i.e., the set of sentences annotated for AI, that contains $a$ and its corresponding long form $l$ (indicated by the annotation for SciAI). Note that here $l$, or its more frequent variant, is already in the canonical long form set of $\alpha$ in $D.$ Then, we assign the long form $l$ to all the occurrences of $a$ in the document that contains $s_i.$ This approach leads to a dataset of 62,441 samples where each sample involves a sentence, an ambiguous acronym, and its correct meaning (i.e., one of the meanings of the acronym recorded by the dictionary $\mathcal{D}).$ We call this dataset SciAD (scientific acronym disambiguation). Consequently, AD is formulated as a classification problem where given a sentence and an acronym, we need to predict the long form of the acronym in a given candidate set.
### 3 $\mathbf{Model}$
In this section we describe the details of the proposed model for AD. As mentioned in section 2.2, we formulate this task as sequence classification problem. Formally, given an input sentence $W=$ $w_1,w_2,...,w_n$ and the position of the acronym, i.e., $p$, the goal is to disambiguate the acronym $w_p$, that is, predicting the long form $l$ from all the possible long forms of $w_p$ specified in the dictionary $\mathcal{D}.$ The proposed model consists of three major components: (1) Sentence Encoder: This component encodes the input sentence into a sequence of representation vectors for the words,(2) Context Encoder: In order to augment the representation of the acronym $w_p$, the context encoder component is designed to leverage the syntactical structure of the sentence to generate the context representation vector for $w_p$, and (3) Prediction: This component consumes the representations obtained from previous components to predict the long form $l^{\prime}$ for the acronym $w_p$ in the sentence. In this section, we describes these components in details.
## 3.1 Sentence Encoder
We represent the word $w_i$ of $W$ using its corresponding pre-trained word embedding. Furthermore, the word embedding $w_i$ is concatenated with the POS tag embedding of $w_i$ to obtain the representation vector $e_i$ for $w_i.$ In order to create more abstract representations of the words and to encode the sequential order of the words in the sentence, we exploit the popular recurrent architecture encoder (i.e., Bidirec- Bidirec- bires (1,2,2,3) tional Long Short-Term Memory (BiLSTM)) to consume the word representations $E=e_1,e_2,\ldots,e_n$ to generate abstract representations $H=h_1,h_2,\ldots,h_n.$ More specifically, the abstract representation $h_i$ is obtained by concatenating the corresponding hidden states of the forward and backward LSTMs: $h_{i}=[\overrightarrow{h_{i}}:\overleftarrow{h_{i}}].$ The abstract representation $H$ will be consumed by the next components in our model.
## 3.2 Context Encoder
While the representation generated by recurrent encoders such as BiLSTM could potentially capture the context of the entire sentence to represent each word $w_i$, in practice, these models fail to encode long dependencies due to vanishing gradient issue. This is problematic as the actual meaning of the words, especially the acronym $w_p$, might depends on some words appeared in far distances from the word itself. In order to address this limitation, prior work in other NLP tasks has shown that the syntactical structure (e.g., dependency trees) could substantially improve the range of contextual information encoded in word representations. Unfortunately, none of the existing deep learning models for AD exploits the dependency trees of the sentences to enrich the contextual information encoded in the word representations. Thus, in this work, we propose to leverage the syntactical structure available in the dependency tree to augment the representations of the words with syntactical contextual information. More specifically, the dependency tree of $W$ is modeled as an undirected graph with adjacency matrix $A$ where $A_{i,j}$
|
[
[
0.12005615234375,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.12005615234375,
0.0914306640625,
0.8818359375,
0.0914306640625,
0.8818359375,
0.105712890625,
0.12005615234375,
0.105712890625
],
[
0.118408203125,
0.1092529296875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.118408203125,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.12005615234375,
0.141357421875,
0.88330078125,
0.141357421875,
0.88330078125,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.12005615234375,
0.15673828125,
0.88330078125,
0.15673828125,
0.88330078125,
0.17041015625,
0.12005615234375,
0.17041015625
],
[
0.118408203125,
0.17333984375,
0.88330078125,
0.17333984375,
0.88330078125,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.118408203125,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.118408203125,
0.203125
],
[
0.12005615234375,
0.204833984375,
0.884765625,
0.204833984375,
0.884765625,
0.218505859375,
0.12005615234375,
0.218505859375
],
[
0.118408203125,
0.221435546875,
0.884765625,
0.221435546875,
0.884765625,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.118408203125,
0.2369384765625,
0.8818359375,
0.2369384765625,
0.8818359375,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.12005615234375,
0.254150390625,
0.88330078125,
0.254150390625,
0.88330078125,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.12005615234375,
0.26953125,
0.88330078125,
0.26953125,
0.88330078125,
0.283203125,
0.12005615234375,
0.283203125
],
[
0.12005615234375,
0.287353515625,
0.88330078125,
0.287353515625,
0.88330078125,
0.297607421875,
0.12005615234375,
0.297607421875
],
[
0.12005615234375,
0.300537109375,
0.84619140625,
0.300537109375,
0.84619140625,
0.314208984375,
0.12005615234375,
0.314208984375
],
[
0.11248779296875,
0.330078125,
0.210693359375,
0.327880859375,
0.210693359375,
0.345703125,
0.11248779296875,
0.347412109375
],
[
0.1175537109375,
0.358154296875,
0.8818359375,
0.358154296875,
0.8818359375,
0.371826171875,
0.1175537109375,
0.371826171875
],
[
0.118408203125,
0.374755859375,
0.86572265625,
0.374755859375,
0.86572265625,
0.388427734375,
0.118408203125,
0.388427734375
],
[
0.118408203125,
0.3896484375,
0.8818359375,
0.3896484375,
0.8818359375,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.118408203125,
0.40625,
0.88330078125,
0.40625,
0.88330078125,
0.42041015625,
0.118408203125,
0.42041015625
],
[
0.12005615234375,
0.421630859375,
0.88330078125,
0.421630859375,
0.88330078125,
0.435791015625,
0.12005615234375,
0.435791015625
],
[
0.118408203125,
0.438232421875,
0.88671875,
0.43701171875,
0.88671875,
0.451904296875,
0.118408203125,
0.453125
],
[
0.12005615234375,
0.454833984375,
0.884765625,
0.454833984375,
0.884765625,
0.468505859375,
0.12005615234375,
0.468505859375
],
[
0.1175537109375,
0.4697265625,
0.88330078125,
0.471435546875,
0.88330078125,
0.486328125,
0.1175537109375,
0.48388671875
],
[
0.118408203125,
0.48681640625,
0.88330078125,
0.48681640625,
0.88330078125,
0.50048828125,
0.118408203125,
0.50048828125
],
[
0.118408203125,
0.5029296875,
0.88330078125,
0.5029296875,
0.88330078125,
0.51708984375,
0.118408203125,
0.51708984375
],
[
0.118408203125,
0.51953125,
0.287109375,
0.51953125,
0.287109375,
0.533203125,
0.118408203125,
0.533203125
],
[
0.1175537109375,
0.5458984375,
0.30322265625,
0.5458984375,
0.30322265625,
0.56005859375,
0.1175537109375,
0.56005859375
],
[
0.12005615234375,
0.5693359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.5830078125,
0.12005615234375,
0.5830078125
],
[
0.12176513671875,
0.58740234375,
0.87890625,
0.58740234375,
0.87890625,
0.59716796875,
0.12176513671875,
0.59716796875
],
[
0.12005615234375,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.12005615234375,
0.615234375
],
[
0.12005615234375,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.12005615234375,
0.6318359375
],
[
0.1175537109375,
0.6318359375,
0.8818359375,
0.63427734375,
0.8818359375,
0.6484375,
0.1175537109375,
0.646484375
],
[
0.12005615234375,
0.6494140625,
0.88330078125,
0.6494140625,
0.88330078125,
0.66455078125,
0.12005615234375,
0.66455078125
],
[
0.12176513671875,
0.66650390625,
0.880859375,
0.66650390625,
0.880859375,
0.6767578125,
0.12176513671875,
0.6767578125
],
[
0.12005615234375,
0.68310546875,
0.869140625,
0.68310546875,
0.869140625,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.12005615234375,
0.71142578125,
0.295654296875,
0.71142578125,
0.295654296875,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.12005615234375,
0.73291015625,
0.880859375,
0.73291015625,
0.880859375,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.118408203125,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76416015625,
0.118408203125,
0.76416015625
],
[
0.12005615234375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.12005615234375,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.118408203125,
0.79736328125,
0.8818359375,
0.79736328125,
0.8818359375,
0.8125,
0.118408203125,
0.8125
],
[
0.1175537109375,
0.81396484375,
0.88330078125,
0.8134765625,
0.88330078125,
0.82763671875,
0.1175537109375,
0.8291015625
],
[
0.118408203125,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.118408203125,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.12005615234375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.118408203125,
0.8935546875,
0.8818359375,
0.8955078125,
0.8818359375,
0.91015625,
0.118408203125,
0.90771484375
]
] |
[
"the less common long form with the more frequent one. After creating the canonical set of longforms",
"for the acronyms, we remove unambiguous acronyms from the mapping/dictionary (i.e., acronyms with",
"only one possible long form), leading to a dictionary D of ambiguous acronyms. Furthermore, in order",
"to improve the quality of the mappings in D, a graduate student curates the dictionary D to verify the",
"canonical forms and resolve any potential error. Afterwards, we employ D to create an AD dataset. To",
"this end, we assume that if an acronym is defined in a paper, its meaning will not change throughout",
"that paper, i.e., the one-sense-per-discourse assumption. Using this criterion, for each acronym a in the",
"dictionary D, we look up all sentences si in S, i.e., the set of sentences annotated for Al, that contains",
"a and its corresponding long form I (indicated by the annotation for SciAI). Note that here l, or its more",
"frequent variant, is already in the canonical long form set of a in D. Then, we assign the long form l",
"to all the occurrences of a in the document that contains si. This approach leads to a dataset of 62,441",
"samples where each sample involves a sentence, an ambiguous acronym, and its correct meaning (i.e.,",
"one of the meanings of the acronym recorded by the dictionary D). We call this dataset SciAD (scientific",
"acronym disambiguation).Consequently,ADisformulatedasaclassificationproblemwhere givena",
"sentence and an acronym, we need to predict the long form of the acronym in a given candidate set",
"3Model",
"In this section we describe the details of the proposed model for AD. As mentioned in section 2.2,",
"we formulate this task as sequence classification problem. Formally, given an input sentence W",
"Wi, w2, ..., Wn and the position of the acronym, i.e.. p, the goal is to disambiguate the acronym Wp, that",
"is, predicting the long form I from all the possible long forms of wp specified in the dictionary D. The",
"proposed model consists of three major components: (1) Sentence Encoder: This component encodes",
"the input sentence into a sequence of representation vectors for the words, (2) Context Encoder: In",
"order to augment the representation of the acronym wp, the context encoder component is designed to",
"leverage the syntactical structure of the sentence to generate the context representation vector for Wp,",
"and (3) Prediction: This component consumes the representations obtained from previous components",
"to predict the long form I' for the acronym wp in the sentence. In this section, we describes these",
"components indetails.",
"3.1SentenceEncoder",
"We represent the word wi of W using its corresponding pre-trained word embedding. Furthermore, the",
"wordembeddingw,isconcatenatedwiththePOStagembeddingofwtoobtaintherepresentationvec",
"tor e; for wi. In order to create more abstract representations of the words and to encode the sequential",
"order of the words in the sentence, we exploit the popular recurrent architecture encoder (i.e., Bidirec-",
"tional Long Short-Term Memory (BiLSTM) to consume the word representations E = ei,e2,..., en",
"to generate abstract representations H = hi, h2, ..., hn. More specifically, the abstract representation",
"h:is obtainedbyconcatenating thecorrespondinghidden states oftheforward andbackward LSTMs:",
"hi = [hi : hi]. The abstract representation H will be consumed by the next components in our model.",
"3.2ContextEncoder",
"While the representation generated by recurrent encoders such as BiLSTM could potentially capture the",
"context of the entire sentence to represent each word wi, in practice, these models fail to encode long",
"dependencies due to vanishing gradient issue. This is problematic as the actual meaning of the words.",
"especially the acronym wp, might depends on some words appeared in far distances from the word itself.",
"In order to address this limitation, prior work in other NLP tasks has shown that the syntactical struc-",
"ture (e.g., dependency trees) could substantially improve the range of contextual information encoded",
"in word representations. Unfortunately, none of the existing deep learning models for AD exploits the",
"dependency trees of the sentences to enrich the contextual information encoded in the word representa-",
"tions. Thus, in this work, we propose to leverage the syntactical structure available in the dependency",
"tree to augment the representations of the words with syntactical contextual information. More specifi-",
"cally, the dependency tree of W is modeled as an undirected graph with adjacency matrix A where Aij"
] |
{
"arxiv_id": "2010.14678",
"language": "en",
"page_idx": 3
}
|
||
2010.14678-en-4
|
2010.14678-en-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>SciAD</th>
<th>$UAD$</th>
<th>SciUAD</th>
</tr>
<tr>
<td>Number of acronyms</td>
<td>732</td>
<td>567</td>
<td>538</td>
</tr>
<tr>
<td>average number of long form per acronym</td>
<td>3.1</td>
<td>2.3</td>
<td>1.8</td>
</tr>
<tr>
<td>overlap between sentence and long forms</td>
<td>0.32</td>
<td>0.01</td>
<td>0.31</td>
</tr>
<tr>
<td>average sentence length</td>
<td>30.7</td>
<td>18.55</td>
<td>50.30</td>
</tr>
</tbody>
</table>
Table 1: Comparison of different AD datasets. Note that the third row shows the ratio of sentences that
have at least one word in common with the long forms of the acronyms appearing in the sentence.
is l if the word $w_i$ is the head or one of the children of the word $w_j$ in the dependency tree. Note that we add self-loops to the dependency tree by adding the identity matrix $I$ to $A{:}\hat{A}=A+I.$ Afterwards, in order to encode the structure $A$ into the node (i.e., word) representations, we employ graph convolutional neural networks (GCN) (Kipf and Welling, 2016; Pouran Ben Veyseh et al., 2019). Concretely, we first initialize the node representations with the vectors $H$ generated by BiLSTM and then at the $m$-th layer of the GCN, we update the representation of $i$-th node by: $h_i^m=\sigma(W_m\cdot\frac1{\deg(i)}\Sigma_{j\in\mathcal{N}(i)}h_j^{m-1})$, where $W_m$ is the weight matrix at layer $m,\sigma$ is a non-linear activation function and $\tilde{\mathcal{N}}(i)$ and deg$(i)$ are the set of neighbors and the degree of the $i$ -th node in the adjacency matrix $\hat{A}$, respectively. The word representations at the final layer of the GCN, i.e., $H^s=h_1^s,h_2^s,\ldots,h_n^s$, are then employed as the structure-aware representations to be consumed by the next component.
# 3.3 Prediction
The prediction component is the last layer in our model which employs the representation obtained from the BiLSTM, i.e., $H$, and the GCN, i.e., $H^s$, to predict the true long form of the acronym $w_p$ in sentence $W.$ To this end, the representation of the acronym $w_p$ from BiLSTM, i.e., $h_i$, and GCN, i.e., $h_i^c$, are concatenated to capture both the sequential and structural context of the word $w_p.$ In addition, we also concatenate the representations of the entire sentence obtained from BiLSTM and GCN to enrich the acronym representation. To create the sentence representations, we employ max pooling over all words: $H_{sent}= MAX\_ POOL( h_{1}, h_{2}, . . . , h_{n})$ and $H_{sent}^{s}=MAX\_POOL(h_{1}^{s},h_{2}^{s},...,h_{n}^{s}).$ Thus, the final representation vector for prediction will be: $V=[h_p:h_p^s:H_{sent}:H_{sent}^s]$, where “”indicates concatenation. Finally, a two-layer feed forward classifier is employed to predict long form $l^{\prime}.$ Note that the number of neurons in the last layer of the feed forward classifier is equal to the total number of long forms of all acronyms in $\mathcal{D}.$ We use negative log-likelihood as the loss function to train the model: $L=-P(l|W,p).$ We call this model graph-based acronym disambiguation (GAD).
## 4 Experiments
Datasets & Evaluation Metrics: We evaluate the performance of the state-of-the-art acronym identiffcation and disambiguation models on SciAI and SciAD, respectively. For AD, in addition to SciAD, we evaluate the performance of the models on the UAD Wikipedia dataset proposed by Ciosici (2019). UAD consists of sentences in general domain (i.e., Wikipedia). As the domain difference between SciAD and UAD (i.e., scientific papers vs Wikipedia articles) could effect on the direct comparison between the performances of the models, we also prepare SciUAD dataset for acronym disambiguation$^{4}$. SciUAD employs the same corpus as we use to create SciAD but the acronyms, long forms, their mappings, and the ambiguous use of acronyms in corpus are all extracted using the unsupervised method proposed by UAD (Ciosici et al., 2019). Table l compares all AD datasets using different criteria. There are several observations from this table. First, SciAD supports more ambiguous acronyms and the level of ambiguity is higher in SciAD as there are more long forms per acronym. This is significant, specially considering the fact that SciAD is prepared from a smaller corpus than UAD. Also, comparing the level of ambiguity in SciAD and SciUAD indicates that our dataset preparation is more effective to capture acronym ambiguity in the given corpus than UAD. Second, comparing the ratio of sentences that have overlap with the
$^{4}$Note that we do not use NOA dataset (Charbonnier and Wartena, 2018) as it has only 4 samples per each long form on
average and it is not suitable for supervised baselines. So, it cannot be comparable with the other datasets in our experiments
|
[
[
0.56103515625,
0.0736083984375,
0.61865234375,
0.0736083984375,
0.61865234375,
0.0885009765625,
0.56103515625,
0.0885009765625
],
[
0.6337890625,
0.0736083984375,
0.68115234375,
0.0736083984375,
0.68115234375,
0.0885009765625,
0.6337890625,
0.0885009765625
],
[
0.6962890625,
0.0736083984375,
0.765625,
0.0736083984375,
0.765625,
0.0885009765625,
0.6962890625,
0.0885009765625
],
[
0.640625,
0.09027099609375,
0.6767578125,
0.09027099609375,
0.6767578125,
0.105712890625,
0.640625,
0.105712890625
],
[
0.7138671875,
0.09027099609375,
0.748046875,
0.09027099609375,
0.748046875,
0.105712890625,
0.7138671875,
0.105712890625
],
[
0.57861328125,
0.0914306640625,
0.609375,
0.0914306640625,
0.609375,
0.105712890625,
0.57861328125,
0.105712890625
],
[
0.231689453125,
0.0914306640625,
0.391357421875,
0.0914306640625,
0.391357421875,
0.1051025390625,
0.231689453125,
0.1051025390625
],
[
0.57666015625,
0.105712890625,
0.60791015625,
0.105712890625,
0.60791015625,
0.122314453125,
0.57666015625,
0.122314453125
],
[
0.64404296875,
0.1051025390625,
0.67333984375,
0.1051025390625,
0.67333984375,
0.122314453125,
0.64404296875,
0.122314453125
],
[
0.71533203125,
0.105712890625,
0.748046875,
0.105712890625,
0.748046875,
0.122314453125,
0.71533203125,
0.122314453125
],
[
0.57275390625,
0.122314453125,
0.61279296875,
0.122314453125,
0.61279296875,
0.1383056640625,
0.57275390625,
0.1383056640625
],
[
0.2333984375,
0.10809326171875,
0.54736328125,
0.10809326171875,
0.54736328125,
0.1217041015625,
0.2333984375,
0.1217041015625
],
[
0.6396484375,
0.122314453125,
0.6767578125,
0.122314453125,
0.6767578125,
0.1383056640625,
0.6396484375,
0.1383056640625
],
[
0.7109375,
0.122314453125,
0.75,
0.122314453125,
0.75,
0.1383056640625,
0.7109375,
0.1383056640625
],
[
0.57275390625,
0.1383056640625,
0.611328125,
0.1383056640625,
0.611328125,
0.15380859375,
0.57275390625,
0.15380859375
],
[
0.2333984375,
0.12353515625,
0.54150390625,
0.12353515625,
0.54150390625,
0.1383056640625,
0.2333984375,
0.1383056640625
],
[
0.63623046875,
0.1383056640625,
0.6826171875,
0.1383056640625,
0.6826171875,
0.15380859375,
0.63623046875,
0.15380859375
],
[
0.7060546875,
0.1383056640625,
0.755859375,
0.1383056640625,
0.755859375,
0.15380859375,
0.7060546875,
0.15380859375
],
[
0.12005615234375,
0.166259765625,
0.88330078125,
0.166259765625,
0.88330078125,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.2333984375,
0.14013671875,
0.414794921875,
0.14013671875,
0.414794921875,
0.15380859375,
0.2333984375,
0.15380859375
],
[
0.12005615234375,
0.182861328125,
0.8388671875,
0.182861328125,
0.8388671875,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.12005615234375,
0.221435546875,
0.88330078125,
0.221435546875,
0.88330078125,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.118408203125,
0.2381591796875,
0.88330078125,
0.2381591796875,
0.88330078125,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.118408203125,
0.254638671875,
0.88671875,
0.254638671875,
0.88671875,
0.268310546875,
0.118408203125,
0.268310546875
],
[
0.1175537109375,
0.26953125,
0.88671875,
0.270751953125,
0.88671875,
0.28515625,
0.1175537109375,
0.283935546875
],
[
0.1175537109375,
0.28515625,
0.88330078125,
0.2861328125,
0.88330078125,
0.300537109375,
0.1175537109375,
0.2998046875
],
[
0.1158447265625,
0.2998046875,
0.88818359375,
0.2998046875,
0.88818359375,
0.32373046875,
0.1158447265625,
0.32373046875
],
[
0.12005615234375,
0.320556640625,
0.884765625,
0.320556640625,
0.884765625,
0.3349609375,
0.12005615234375,
0.3349609375
],
[
0.12005615234375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.12005615234375,
0.350830078125
],
[
0.118408203125,
0.3515625,
0.88330078125,
0.352783203125,
0.88330078125,
0.36767578125,
0.118408203125,
0.366455078125
],
[
0.118408203125,
0.369384765625,
0.52880859375,
0.369384765625,
0.52880859375,
0.383056640625,
0.118408203125,
0.383056640625
],
[
0.118408203125,
0.396728515625,
0.246826171875,
0.396728515625,
0.246826171875,
0.410400390625,
0.118408203125,
0.410400390625
],
[
0.118408203125,
0.41748046875,
0.88330078125,
0.41748046875,
0.88330078125,
0.431640625,
0.118408203125,
0.431640625
],
[
0.118408203125,
0.432861328125,
0.88330078125,
0.43408203125,
0.88330078125,
0.448974609375,
0.118408203125,
0.44775390625
],
[
0.12005615234375,
0.449462890625,
0.8818359375,
0.449462890625,
0.8818359375,
0.46435546875,
0.12005615234375,
0.46435546875
],
[
0.12176513671875,
0.466064453125,
0.8818359375,
0.466064453125,
0.8818359375,
0.479736328125,
0.12176513671875,
0.479736328125
],
[
0.12176513671875,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49462890625,
0.12176513671875,
0.49462890625
],
[
0.118408203125,
0.498779296875,
0.88330078125,
0.498779296875,
0.88330078125,
0.5126953125,
0.118408203125,
0.5126953125
],
[
0.118408203125,
0.51318359375,
0.55517578125,
0.51318359375,
0.55517578125,
0.53076171875,
0.118408203125,
0.53076171875
],
[
0.5517578125,
0.51416015625,
0.88330078125,
0.51416015625,
0.88330078125,
0.529296875,
0.5517578125,
0.529296875
],
[
0.118408203125,
0.529296875,
0.884765625,
0.529296875,
0.884765625,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.118408203125,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56103515625,
0.118408203125,
0.56103515625
],
[
0.12005615234375,
0.5634765625,
0.884765625,
0.5634765625,
0.884765625,
0.57666015625,
0.12005615234375,
0.57666015625
],
[
0.12005615234375,
0.5791015625,
0.88330078125,
0.5791015625,
0.88330078125,
0.5927734375,
0.12005615234375,
0.5927734375
],
[
0.12005615234375,
0.595703125,
0.73486328125,
0.595703125,
0.73486328125,
0.609375,
0.12005615234375,
0.609375
],
[
0.12005615234375,
0.62353515625,
0.26123046875,
0.62353515625,
0.26123046875,
0.63720703125,
0.12005615234375,
0.63720703125
],
[
0.118408203125,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.6630859375,
0.118408203125,
0.6630859375
],
[
0.12005615234375,
0.66650390625,
0.88330078125,
0.66650390625,
0.88330078125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.118408203125,
0.681640625,
0.88330078125,
0.681640625,
0.88330078125,
0.6953125,
0.118408203125,
0.6953125
],
[
0.118408203125,
0.69873046875,
0.88330078125,
0.69873046875,
0.88330078125,
0.7119140625,
0.118408203125,
0.7119140625
],
[
0.12005615234375,
0.71337890625,
0.8818359375,
0.71337890625,
0.8818359375,
0.72802734375,
0.12005615234375,
0.72802734375
],
[
0.12005615234375,
0.73095703125,
0.88330078125,
0.73095703125,
0.88330078125,
0.74462890625,
0.12005615234375,
0.74462890625
],
[
0.118408203125,
0.7451171875,
0.88330078125,
0.74658203125,
0.88330078125,
0.76123046875,
0.118408203125,
0.76025390625
],
[
0.12005615234375,
0.7626953125,
0.8818359375,
0.7626953125,
0.8818359375,
0.775390625,
0.12005615234375,
0.775390625
],
[
0.12005615234375,
0.779296875,
0.88330078125,
0.779296875,
0.88330078125,
0.7919921875,
0.12005615234375,
0.7919921875
],
[
0.12005615234375,
0.79443359375,
0.88330078125,
0.79443359375,
0.88330078125,
0.80859375,
0.12005615234375,
0.80859375
],
[
0.118408203125,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82470703125,
0.118408203125,
0.82470703125
],
[
0.12005615234375,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.12005615234375,
0.84033203125
],
[
0.1175537109375,
0.84130859375,
0.8818359375,
0.8427734375,
0.8818359375,
0.85693359375,
0.1175537109375,
0.85546875
],
[
0.12005615234375,
0.859375,
0.8818359375,
0.859375,
0.8818359375,
0.87353515625,
0.12005615234375,
0.87353515625
],
[
0.1376953125,
0.8818359375,
0.88330078125,
0.8828125,
0.88330078125,
0.89794921875,
0.1376953125,
0.896484375
],
[
0.12176513671875,
0.89794921875,
0.8759765625,
0.89794921875,
0.8759765625,
0.90771484375,
0.12176513671875,
0.90771484375
]
] |
[
"SciAD",
"UAD",
"SciUAD",
"567",
"538",
"732",
"Number of acronyms",
"3.1",
"2.3",
"1.8",
"0.32",
"average number of long form per acronym",
"0.01",
"0.31",
"30.7",
"overlap between sentence and long forms",
"18.55",
"50.30",
"Table l: Comparison of different AD datasets. Note that the third row shows the ratio of sentences that",
"average sentence length",
"have at least one word in common with the long forms of the acronyms appearing in the sentence.",
"is 1 if the word w; is the head or one of the children of the word w; in the dependency tree. Note that we",
"add self-loops to the dependency tree by adding the identity matrix I to A: A = A + I. Afterwards, in",
"order to encode the structure A into the node (i.e., word) representations, we employ graph convolutional",
"neural networks (GCN) (Kipf and Welling, 2016; Pouran Ben Veyseh et al., 2019). Concretely, we first",
"initialize the node representations with the vectors H generated by BiLSTM and then at the m-th layer",
"of the GCN, we update the representation of i-th node by: h = o(Wm ' deg(,jeN(a) hm-1), where",
"Wm is the weight matrix at layer m, is a non-linear activation function and N(i) and deg(i) are the set",
"of neighbors and the degree of the i-th node in the adjacency matrix A, respectively. The word represen-",
"tations at the final layer of the GCN, i.e., Hs = hi, h, ..., hn, are then employed as the structure-aware",
"representations to be consumed by the next component.",
"3.3Prediction",
"The prediction component is the last layer in our model which employs the representation obtained",
"from the BiLSTM, i.e., H, and the GCN, i.e., Hs, to predict the true long form of the acronym wp in",
"sentence W. To this end, the representation of the acronym wp from BiLSTM, i.e., hi, and GCN, i.e..",
"hg, are concatenated to capture both the sequential and structural context of the word wp. In addition, we",
"also concatenatetherepresentations oftheentire sentenceobtainedfromBiLSTM and GCNtoenrich",
"the acronym representation. To create the sentence representations, we employ max pooling over all",
"words: Hsent = MAX_POOL(hi, h2,.hn) and Hsent",
"= MAX_POOL(hi, h2, .., hs). Thus, the",
"final representation vector for prediction will be: V= [hp : hg : Hsent : Hsentl], where “:\" indicates",
"concatenation. Finally, a two-layer feed forward classifier is employed to predict long form I'. Note",
"that the number of neurons in the last layer of the feed forward classifier is equal to the total number of",
"long forms of all acronyms in D. We use negative log-likelihood as the loss function to train the model:",
"L = -P(1|W, p). We call this model graph-based acronym disambiguation (GAD).",
"4Experiments",
"Datasets & Evaluation Metrics: We evaluate the performance of the state-of-the-art acronym identifi-",
"cation and disambiguation models on SciAI and SciAD, respectively. For AD, in addition to SciAD, we",
"evaluate the performance of the models on the UAD Wikipedia dataset proposed by Ciosici (2019). UAD",
"consists of sentences in general domain (i.e., Wikipedia). As the domain difference between SciAD and",
"UAD (i.e., scientific papers vs Wikipedia articles) could effect on the direct comparison between the",
"performances of the models, we also prepare SciUAD dataset for acronym disambiguation. SciUAD",
"employs the same corpus as we use to create SciAD but the acronyms, long forms, their mappings, and",
"the ambiguous use of acronyms in corpus are all extracted using the unsupervised method proposed by",
"UAD (Ciosici et al., 2019). Table 1 compares all AD datasets using different criteria. There are several",
"observations from this table. First, SciAD supports more ambiguous acronyms and the level of ambiguity",
"is higher in SciAD as there are more long forms per acronym. This is significant, specially considering",
"the fact that SciAD is prepared from a smaller corpus than UAD. Also, comparing the level of ambiguity",
"in SciAD and SciUAD indicates that our dataset preparation is more effective to capture acronym ambi-",
"guity in the given corpus than UAD. Second, comparing the ratio of sentences that have overlap with the",
"4Note that we do not use NOA dataset (Charbonnier and Wartena, 2018) as it has only 4 samples per each long form on",
"averageand it is not suitable for supervised baselines.So, it cannot be comparablewith the other datasets in our experiments."
] |
{
"arxiv_id": "2010.14678",
"language": "en",
"page_idx": 4
}
|
||
2004.07499-en-1
|
2004.07499-en-1.jpg
|
forts. However, these recommendations are provided by a pre-trained model or via dictionary look-ups. This methodology of providing recommendations often proves to be unhelpful when little annotated data exists for pre-training, as is usually the case for natural language tasks being applied to domain-specific or user-provided corpora.
To resolve this issue, AlpacaTag, an annotation framework for sequence labeling (Lin et al., 2019) attempts to provide annotation recommendations from a learned sequence labeling model that is incrementally updated by batches of incoming human annotations. Its model training follows an active learning strategy (Shen et al., 2017), which is shown to be a label-efficient, thus it attempts to minimize human annotation efforts. AlpacaTag selects the most informative batches of documents for humans to annotate and thus achieves a more cost-effective way of using human efforts. While active learning allows the model to achieve higher performance earlier in the learning process, model performance could be improved if additional supervision existed. It is imperative that provided annotation recommendations be as accurate as possible, as inaccurate annotation recommendations from the framework can push users towards generating noisy data, hindering instead of aiding the model training process.
Our effort to prevent this problem is centered around allowing annotators to provide additional supervision by capturing labeling explanations, while still taking advantage of the costeffectiveness of active learning. Specifically, as shown in Fig. 1, we allow annotators to provide explanations for their decisions in natural language or by selecting triggersnearby phrases that provide helpful context for their decisions. These enhanced annotations allow for model training over both user-provided labels, as well as weakly labeled data created by parsing explanations into high precision labeling rules. We therefore make attempts to ameliorate the erroneous recommendation problem by a performance-boosting training strategy that incorporates both labeled and unlabeled data.
Our work is also similar to recent attempts that exploit explanations for an improved training process (Srivastava et al., 2017; Hancock et al., 2018; Zhou et al., 2020; Qin et al., 2020), but with two main differences. First, we embed this improved training process in a practical application and sec-
<FigureHere>
Figure 2: System Architecture.
ond, we design task specific architectures to incorporate the now captured explanations into training.
To the best of our knowledge, there is no existing open-source, easy-to-use, recommendationproviding, online-learning annotation framework that can also capture explanations. LEAN-LIFE is the frst framework to capture and leverage explanations for improved model training and performance, while still inheriting the advantages of existing tools. We summarize our contributions as: $\bullet$ Improved Model Training: Our recommendation models use a performance improving training process that leverages explanations to weakly label unlabeled instances. Our models improve on competitive baseline F-1 scores by more than 5-10 percentage points, while using 2X less data.
· Multiple Supported Tasks: Our framework supports both sequence labeling (as in NER) and sequence classification (as in RE, SA).
Explanation Dataset Creation: We make it easy to build a new type of dataset, one that consists of triples of: text, labels and labeling explanations. The exporting of this captured data is available in two common data formats, CSV and JSON.
# 2 System Overview
As shown in Fig. 2, our framework consists of two main components, a user-friendly web-UI that can capture labels and explanations for labeling decisions, and a weak supervision framework that parses explanations for the creation of weakly labeled data. The framework then uses this weakly labeled data in conjunction with user-provided labels to train models for improved annotation recommendations. Our UI shows annotators unlabeled instances (can be sampled using active learning), along with annotation recommendations in an effort to reduce annotation costs. We use PyTorch to build our models and implement an API
|
[
[
0.118408203125,
0.07598876953125,
0.486083984375,
0.07598876953125,
0.486083984375,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.12005615234375,
0.0926513671875,
0.486083984375,
0.0926513671875,
0.486083984375,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.5693359375,
0.0885009765625,
0.59619140625,
0.0885009765625,
0.59619140625,
0.1009521484375,
0.5693359375,
0.1009521484375
],
[
0.5576171875,
0.10272216796875,
0.603515625,
0.10272216796875,
0.603515625,
0.1134033203125,
0.5576171875,
0.1134033203125
],
[
0.75390625,
0.10394287109375,
0.81201171875,
0.10394287109375,
0.81201171875,
0.114013671875,
0.75390625,
0.114013671875
],
[
0.73779296875,
0.10986328125,
0.8271484375,
0.11102294921875,
0.8271484375,
0.1217041015625,
0.73779296875,
0.12054443359375
],
[
0.55908203125,
0.11102294921875,
0.60107421875,
0.11102294921875,
0.60107421875,
0.1217041015625,
0.55908203125,
0.1217041015625
],
[
0.118408203125,
0.10809326171875,
0.48779296875,
0.10809326171875,
0.48779296875,
0.1217041015625,
0.118408203125,
0.1217041015625
],
[
0.1175537109375,
0.12469482421875,
0.486083984375,
0.12469482421875,
0.486083984375,
0.1383056640625,
0.1175537109375,
0.1383056640625
],
[
0.62890625,
0.122314453125,
0.6826171875,
0.122314453125,
0.6826171875,
0.133056640625,
0.62890625,
0.133056640625
],
[
0.57421875,
0.133056640625,
0.58837890625,
0.133056640625,
0.58837890625,
0.141357421875,
0.57421875,
0.141357421875
],
[
0.6455078125,
0.1341552734375,
0.6708984375,
0.1341552734375,
0.6708984375,
0.142578125,
0.6455078125,
0.142578125
],
[
0.12005615234375,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.744140625,
0.135986328125,
0.82275390625,
0.135986328125,
0.82275390625,
0.1591796875,
0.744140625,
0.1591796875
],
[
0.51318359375,
0.1478271484375,
0.54248046875,
0.1478271484375,
0.54248046875,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.118408203125,
0.1561279296875,
0.486083984375,
0.1561279296875,
0.486083984375,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.51318359375,
0.157958984375,
0.5439453125,
0.157958984375,
0.5439453125,
0.169189453125,
0.51318359375,
0.169189453125
],
[
0.6455078125,
0.157958984375,
0.6689453125,
0.157958984375,
0.6689453125,
0.1650390625,
0.6455078125,
0.1650390625
],
[
0.118408203125,
0.1728515625,
0.486083984375,
0.1728515625,
0.486083984375,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.79833984375,
0.1787109375,
0.84228515625,
0.1787109375,
0.84228515625,
0.190673828125,
0.79833984375,
0.190673828125
],
[
0.1351318359375,
0.1923828125,
0.48779296875,
0.193603515625,
0.48779296875,
0.2078857421875,
0.1351318359375,
0.207275390625
],
[
0.118408203125,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.118408203125,
0.223876953125
],
[
0.12005615234375,
0.2257080078125,
0.48779296875,
0.2257080078125,
0.48779296875,
0.2392578125,
0.12005615234375,
0.2392578125
],
[
0.58447265625,
0.221435546875,
0.66015625,
0.221435546875,
0.66015625,
0.232177734375,
0.58447265625,
0.232177734375
],
[
0.118408203125,
0.243408203125,
0.486083984375,
0.243408203125,
0.486083984375,
0.254150390625,
0.118408203125,
0.254150390625
],
[
0.12005615234375,
0.25830078125,
0.489501953125,
0.25830078125,
0.489501953125,
0.271484375,
0.12005615234375,
0.271484375
],
[
0.58837890625,
0.254638671875,
0.80615234375,
0.254638671875,
0.80615234375,
0.268310546875,
0.58837890625,
0.268310546875
],
[
0.12005615234375,
0.27490234375,
0.490234375,
0.27490234375,
0.490234375,
0.2880859375,
0.12005615234375,
0.2880859375
],
[
0.509765625,
0.275634765625,
0.8818359375,
0.275634765625,
0.8818359375,
0.289306640625,
0.509765625,
0.289306640625
],
[
0.118408203125,
0.289306640625,
0.490234375,
0.290283203125,
0.490234375,
0.3046875,
0.118408203125,
0.303955078125
],
[
0.51123046875,
0.29150390625,
0.880859375,
0.29150390625,
0.880859375,
0.3046875,
0.51123046875,
0.3046875
],
[
0.12005615234375,
0.30712890625,
0.486083984375,
0.30712890625,
0.486083984375,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.52880859375,
0.305908203125,
0.88330078125,
0.30712890625,
0.88330078125,
0.32177734375,
0.52880859375,
0.320556640625
],
[
0.1175537109375,
0.32177734375,
0.48779296875,
0.32177734375,
0.48779296875,
0.3349609375,
0.1175537109375,
0.3349609375
],
[
0.51318359375,
0.32470703125,
0.88330078125,
0.32470703125,
0.88330078125,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.118408203125,
0.339111328125,
0.489501953125,
0.339111328125,
0.489501953125,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.5146484375,
0.34033203125,
0.884765625,
0.34033203125,
0.884765625,
0.35400390625,
0.5146484375,
0.35400390625
],
[
0.118408203125,
0.355224609375,
0.490234375,
0.355224609375,
0.490234375,
0.3681640625,
0.118408203125,
0.3681640625
],
[
0.51318359375,
0.35693359375,
0.884765625,
0.35693359375,
0.884765625,
0.37060546875,
0.51318359375,
0.37060546875
],
[
0.1175537109375,
0.37060546875,
0.491943359375,
0.37060546875,
0.491943359375,
0.38427734375,
0.1175537109375,
0.38427734375
],
[
0.51318359375,
0.372314453125,
0.88330078125,
0.372314453125,
0.88330078125,
0.385986328125,
0.51318359375,
0.385986328125
],
[
0.118408203125,
0.385498046875,
0.490234375,
0.385986328125,
0.490234375,
0.40087890625,
0.118408203125,
0.399658203125
],
[
0.51318359375,
0.3896484375,
0.88330078125,
0.3896484375,
0.88330078125,
0.402587890625,
0.51318359375,
0.402587890625
],
[
0.12005615234375,
0.402587890625,
0.489501953125,
0.402587890625,
0.489501953125,
0.416259765625,
0.12005615234375,
0.416259765625
],
[
0.51318359375,
0.405029296875,
0.88671875,
0.405029296875,
0.88671875,
0.418701171875,
0.51318359375,
0.418701171875
],
[
0.118408203125,
0.418701171875,
0.489501953125,
0.41748046875,
0.489501953125,
0.431640625,
0.118408203125,
0.432861328125
],
[
0.51318359375,
0.42041015625,
0.88671875,
0.42041015625,
0.88671875,
0.43408203125,
0.51318359375,
0.43408203125
],
[
0.12005615234375,
0.435302734375,
0.489501953125,
0.435302734375,
0.489501953125,
0.448974609375,
0.12005615234375,
0.448974609375
],
[
0.51123046875,
0.43701171875,
0.8818359375,
0.43701171875,
0.8818359375,
0.45068359375,
0.51123046875,
0.45068359375
],
[
0.12005615234375,
0.451904296875,
0.48779296875,
0.451904296875,
0.48779296875,
0.465576171875,
0.12005615234375,
0.465576171875
],
[
0.51318359375,
0.45361328125,
0.884765625,
0.45361328125,
0.884765625,
0.46728515625,
0.51318359375,
0.46728515625
],
[
0.12005615234375,
0.46728515625,
0.489501953125,
0.46728515625,
0.489501953125,
0.48095703125,
0.12005615234375,
0.48095703125
],
[
0.5146484375,
0.4697265625,
0.8818359375,
0.4697265625,
0.8818359375,
0.482666015625,
0.5146484375,
0.482666015625
],
[
0.12005615234375,
0.482666015625,
0.489501953125,
0.482666015625,
0.489501953125,
0.496337890625,
0.12005615234375,
0.496337890625
],
[
0.51318359375,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.498779296875,
0.51318359375,
0.498779296875
],
[
0.12005615234375,
0.49951171875,
0.489501953125,
0.49951171875,
0.489501953125,
0.51318359375,
0.12005615234375,
0.51318359375
],
[
0.51123046875,
0.50048828125,
0.88330078125,
0.49951171875,
0.88330078125,
0.51416015625,
0.51123046875,
0.515625
],
[
0.12005615234375,
0.5166015625,
0.3720703125,
0.5166015625,
0.3720703125,
0.52978515625,
0.12005615234375,
0.52978515625
],
[
0.51318359375,
0.51708984375,
0.8388671875,
0.51708984375,
0.8388671875,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.1368408203125,
0.5361328125,
0.484375,
0.5361328125,
0.484375,
0.5498046875,
0.1368408203125,
0.5498046875
],
[
0.5146484375,
0.53369140625,
0.88330078125,
0.53369140625,
0.88330078125,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.118408203125,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.118408203125,
0.5654296875
],
[
0.51123046875,
0.5498046875,
0.88330078125,
0.548828125,
0.88330078125,
0.5634765625,
0.51123046875,
0.56396484375
],
[
0.12005615234375,
0.568359375,
0.484375,
0.568359375,
0.484375,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.51123046875,
0.56640625,
0.79833984375,
0.5654296875,
0.79833984375,
0.5791015625,
0.51123046875,
0.580078125
],
[
0.480224609375,
0.572265625,
0.484375,
0.572265625,
0.484375,
0.57763671875,
0.480224609375,
0.57763671875
],
[
0.118408203125,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.118408203125,
0.59716796875
],
[
0.5146484375,
0.58056640625,
0.88330078125,
0.58056640625,
0.88330078125,
0.59423828125,
0.5146484375,
0.59423828125
],
[
0.1175537109375,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.1175537109375,
0.61328125
],
[
0.51318359375,
0.5986328125,
0.88330078125,
0.5986328125,
0.88330078125,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.118408203125,
0.6162109375,
0.489501953125,
0.6162109375,
0.489501953125,
0.6298828125,
0.118408203125,
0.6298828125
],
[
0.51318359375,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.51318359375,
0.62890625
],
[
0.118408203125,
0.6328125,
0.489501953125,
0.6328125,
0.489501953125,
0.646484375,
0.118408203125,
0.646484375
],
[
0.51318359375,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.12005615234375,
0.6494140625,
0.489501953125,
0.6494140625,
0.489501953125,
0.6630859375,
0.12005615234375,
0.6630859375
],
[
0.51123046875,
0.6455078125,
0.8818359375,
0.6455078125,
0.8818359375,
0.6591796875,
0.51123046875,
0.6591796875
],
[
0.12005615234375,
0.6650390625,
0.489501953125,
0.6650390625,
0.489501953125,
0.6787109375,
0.12005615234375,
0.6787109375
],
[
0.5224609375,
0.6767578125,
0.5322265625,
0.68310546875,
0.5263671875,
0.6884765625,
0.5146484375,
0.68310546875
],
[
0.5380859375,
0.6748046875,
0.69287109375,
0.6748046875,
0.69287109375,
0.6884765625,
0.5380859375,
0.6884765625
],
[
0.118408203125,
0.681640625,
0.489501953125,
0.681640625,
0.489501953125,
0.6953125,
0.118408203125,
0.6953125
],
[
0.118408203125,
0.69775390625,
0.489501953125,
0.69775390625,
0.489501953125,
0.71142578125,
0.118408203125,
0.71142578125
],
[
0.51318359375,
0.70068359375,
0.8818359375,
0.70068359375,
0.8818359375,
0.71435546875,
0.51318359375,
0.71435546875
],
[
0.118408203125,
0.7119140625,
0.490234375,
0.71337890625,
0.490234375,
0.72802734375,
0.118408203125,
0.72705078125
],
[
0.51123046875,
0.71728515625,
0.88330078125,
0.71630859375,
0.88330078125,
0.73095703125,
0.51123046875,
0.732421875
],
[
0.1175537109375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7421875,
0.1175537109375,
0.7421875
],
[
0.51318359375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.51318359375,
0.74755859375
],
[
0.118408203125,
0.74462890625,
0.48779296875,
0.74462890625,
0.48779296875,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.51318359375,
0.75048828125,
0.880859375,
0.75048828125,
0.880859375,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.12005615234375,
0.76123046875,
0.489501953125,
0.76123046875,
0.489501953125,
0.775390625,
0.12005615234375,
0.775390625
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.118408203125,
0.77685546875,
0.489501953125,
0.775390625,
0.489501953125,
0.79052734375,
0.118408203125,
0.79150390625
],
[
0.51318359375,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.1175537109375,
0.79150390625,
0.202392578125,
0.79345703125,
0.202392578125,
0.80810546875,
0.1175537109375,
0.8056640625
],
[
0.51123046875,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.51123046875,
0.8125
],
[
0.1368408203125,
0.81396484375,
0.490234375,
0.81396484375,
0.490234375,
0.82763671875,
0.1368408203125,
0.82763671875
],
[
0.51123046875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51123046875,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.51123046875,
0.84423828125
],
[
0.118408203125,
0.8466796875,
0.49365234375,
0.8466796875,
0.49365234375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51123046875,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.12176513671875,
0.8642578125,
0.490234375,
0.8642578125,
0.490234375,
0.87451171875,
0.12176513671875,
0.87451171875
],
[
0.51318359375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1175537109375,
0.87744140625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.51318359375,
0.892578125
],
[
0.12005615234375,
0.8955078125,
0.490234375,
0.8955078125,
0.490234375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.89453125,
0.88330078125,
0.8935546875,
0.88330078125,
0.90771484375,
0.51318359375,
0.9091796875
]
] |
[
"forts.However, these recommendations are pro-",
"vided by a pre-trained model or via dictionary",
"",
"Unlabeled",
"Explanation",
"Parsing/Encoding",
"Instances",
"look-ups. This methodology of providing recom-",
"mendations often proves to be unhelpful when lit",
"Explanation",
"UI",
"FO-",
"tle annotated data exists for pre-training, as is usu-",
"SoftyMasing",
"",
"ally the case for natural language tasks being ap-",
"USER",
"Label",
"plied to domain-specific or user-provided corpora.",
"Weak Lab",
"To resolve this issue, AlpacaTag, an annota-",
"tion framework for sequence labeling (Lin et al.,",
"2019) attempts to provide annotation recommen-",
"Recommendation",
"dations from a learned sequence labeling model",
"that is incrementally updated by batches of in-",
"Figure 2: System Architecture.",
"coming human annotations. Its model training",
" ond, we design task specific architectures to incor-",
"follows an active learning strategy (Shen et al.,",
"porate the now captured explanations into training.",
"2017). which is shown to be a label-efficient. thus",
"To the best of our knowledge, there is no ex-",
"it attempts to minimize human annotation efforts.",
"isting open-source, easy-to-use, recommendation-",
"AlpacaTag selects the most informative batches",
"providing, online-learning annotation framework",
"of documents for humans to annotate and thus",
"that can also capture explanations. LEAN-LIFE",
"achieves a more cost-effective way of using human",
"is the first framework to capture and leverage ex-",
"efforts. While active learning allows the model to",
"planations for improved model training and per-",
"achieve higher performance earlier in the learning",
"formance, while still inheriting the advantages of",
"process, model performance could be improved if",
"existing tools. We summarize our contributions as:",
"additional supervision existed. It is imperative that",
": Improved Model Training: Our recommenda-",
"provided annotation recommendations be as accu-",
"tion models use a performance improving training",
"rate as possible, as inaccurate annotation recom-",
"process that leverages explanations to weakly la-",
"mendations from the framework can push users to-",
"bel unlabeled instances. Our models improve on",
"wards generating noisy data, hindering instead of",
"competitive baseline F-1 scores by more than 5-10",
"aiding the model training process.",
"percentage points, while using 2X less data.",
"Our effort to prevent this problem is cen-",
": Multiple Supported Tasks: Our framework",
"tered around allowing annotators to provide ad-",
"supports both sequence labeling (as in NER) and",
"ditional supervision by capturing labeling expla-",
"sequence classification (as in RE, SA)",
"I",
"nations, while still taking advantage of the cost-",
". Explanation Dataset Creation: We make it",
"effectiveness of active learning.Specifically, as",
"easy to build a new type of dataset, one that con-",
"shown in Fig. 1, we allow annotators to pro-",
"sists of triples of: text, labels and labeling explana-",
"vide explanations for their decisions in natural lan-",
"tions. The exporting of this captured data is avail-",
"guage or by selecting triggersnearby phrases that",
"able in two common data formats, CSV and JSON.",
"provide helpful context for their decisions. These",
"",
"System Overview",
"enhanced annotations allow for model training",
"over both user-provided labels, as well as weakly",
"As shown in Fig. 2, our framework consists of",
"labeled data created by parsing explanations into",
"two main components, a user-friendly web-UI that",
"high precision labeling rules. We therefore make",
"can capture labels and explanations for labeling",
"attempts to ameliorate the erroneous recommenda-",
"decisions.and a weak supervision framework that",
"tion problem by a performance-boosting training",
"parses explanations for the creation of weakly la-",
"strategy that incorporates both labeled and unla-",
"beled data. The framework then uses this weakly",
"beled data.",
"labeled data in conjunction with user-provided la-",
"Our work is also similar to recent attempts that",
"bels to train models for improved annotation rec-",
"exploit explanations for an improved training pro-",
"ommendations. Our UI shows annotators unla-",
"cess (Srivastava et al., 2017; Hancock et al., 2018;",
" beled instances (can be sampled using active learn-",
"Zhou et al.. 2020: Oin et al.. 2020), but with two",
"ing), along with annotation recommendations in",
"main differences. First, we embed this improved",
"an effort to reduce annotation costs. We use Py-",
"training process in a practical application and sec-",
"Torch to build our models and implement an API"
] |
{
"arxiv_id": "2004.07499",
"language": "en",
"page_idx": 1
}
|
||
2112.04016-en-0
|
2112.04016-en-0.jpg
|
## $\textbf{DeepFace- EMD: Re- ranking Using Patch- wise Earth Mover' s Distance}$ $\textbf{Improves Out- Of- Distribution Face Identification}$
Anh Nguyen$^{1}$
anh.ng8@gmail.com
Hai Phan$^{1,2}$
pthail204@gmail.com
$^{2}$Carnegie Mellon University
$^{1}$Auburn University
$ଟ୍ରି$
$പ്റെ$
<FigureHere>
$উু$
Figure 1. Traditional face identification ranks gallery images based on their cosine distance with the query (top row) at the image-level embedding, which yields large errors upon out-of-distribution changes in the input $(e.g.$ masks or sunglasses; b-d). We find that re-ranking
the top-$k$ shortlisted faces from Stage 1 (leftmost column) using their patch-wise EMD similarity w.r.t. the query substantially improves the precision (Stage 2) on challenging cases (b-d). The “Flow”visualization intuitively shows the patch-wise reconstruction of the query face using the most similar patches $(i.e.$ highest flow) from the retrieved face. See Fig. S4 for a full figure with the top-5 candidates
Abstract
while obtaining similar results on in-distribution images.
arXiv:2112.
Face identifrcation (FI) is ubiquitous and drives many high-stake decisions made by the law enforcement. A common FI approach compares two images by taking the cosine similarity between their image embeddings. Yet, such approach suffers from poor out-of-distribution (OOD) generalization to new types of images (e.g., when a query face is masked, cropped or rotated) not included in the training set or the gallery. Here, we propose a re-ranking approach that compares two faces using the Earth Mover's Distance on the deep, spatial features of image patches. Our extra comparison stage explicitly examines image similarity at a fine-grained level (e.g., eyes to eyes) and is more robust to OOD perturbations and occlusions than traditional $FI. \textit{ Interestingly, without finetuning feature extractors, our}$ method consistently improves the accuracy on all tested $OOD$ queries: masked, cropped, rotated, and adversarial
### $1. \textbf{Introduction}$
Who shoplifted from the Shinola luxury store in Detroit[4]? Who are you to receive unemployment benefits[5] or board an airplane [1]? Face identification (FI) today is behind the answers to such life-critical questions. Yet, the technology can make errors, leading to severe consequences, e.g. people wrongly denied of unemployment benefits[5] or falsely arrested[2-4,7]. Identifying the person in a single photo remains challenging because, in many cases, the problem is a zero-shot and ill-posed image retrieval task. First, a deep feature extractor may not have seen a normal, non-celebrity person before during training. Second, there may be too few photos of a person in the database for FI systems to make reliable decisions. Third, it is harder to identify when a face in the wild $(e.g.$ from surveillance cameras) is occluded [44,54] (e.g. wearing masks), distant or
|
[
[
0.13232421875,
0.1331787109375,
0.8388671875,
0.1331787109375,
0.8388671875,
0.147705078125,
0.13232421875,
0.147705078125
],
[
0.238525390625,
0.1572265625,
0.734375,
0.1572265625,
0.734375,
0.1717529296875,
0.238525390625,
0.1717529296875
],
[
0.324462890625,
0.203857421875,
0.42236328125,
0.199462890625,
0.423095703125,
0.2183837890625,
0.325927734375,
0.222900390625
],
[
0.5498046875,
0.2052001953125,
0.66015625,
0.2052001953125,
0.66015625,
0.2197265625,
0.5498046875,
0.2197265625
],
[
0.289306640625,
0.2247314453125,
0.4599609375,
0.2247314453125,
0.4599609375,
0.238037109375,
0.289306640625,
0.238037109375
],
[
0.529296875,
0.2247314453125,
0.68115234375,
0.2247314453125,
0.68115234375,
0.238037109375,
0.529296875,
0.238037109375
],
[
0.293212890625,
0.246826171875,
0.44677734375,
0.2474365234375,
0.44677734375,
0.26318359375,
0.293212890625,
0.261962890625
],
[
0.491943359375,
0.24560546875,
0.7197265625,
0.2474365234375,
0.7197265625,
0.26513671875,
0.491943359375,
0.26318359375
],
[
0.0196075439453125,
0.2568359375,
0.056365966796875,
0.2568359375,
0.05963134765625,
0.486083984375,
0.0228729248046875,
0.486083984375
],
[
0.156005859375,
0.296630859375,
0.2115478515625,
0.294189453125,
0.2132568359375,
0.31005859375,
0.1568603515625,
0.3125
],
[
0.32275390625,
0.29736328125,
0.43701171875,
0.29736328125,
0.43701171875,
0.311279296875,
0.32275390625,
0.311279296875
],
[
0.51953125,
0.29736328125,
0.65185546875,
0.29736328125,
0.65185546875,
0.311279296875,
0.51953125,
0.311279296875
],
[
0.72607421875,
0.298583984375,
0.82666015625,
0.298583984375,
0.82666015625,
0.31005859375,
0.72607421875,
0.31005859375
],
[
0.4345703125,
0.4091796875,
0.461669921875,
0.4091796875,
0.461669921875,
0.43994140625,
0.4345703125,
0.43994140625
],
[
0.0196075439453125,
0.484130859375,
0.058013916015625,
0.484130859375,
0.060455322265625,
0.7109375,
0.0228729248046875,
0.7109375
],
[
0.5,
0.494384765625,
0.53369140625,
0.494384765625,
0.53369140625,
0.505859375,
0.5,
0.505859375
],
[
0.62158203125,
0.494384765625,
0.66015625,
0.494384765625,
0.66015625,
0.505859375,
0.62158203125,
0.505859375
],
[
0.68896484375,
0.494384765625,
0.7255859375,
0.494384765625,
0.7255859375,
0.505859375,
0.68896484375,
0.505859375
],
[
0.7607421875,
0.494384765625,
0.7861328125,
0.494384765625,
0.7861328125,
0.505859375,
0.7607421875,
0.505859375
],
[
0.431396484375,
0.49560546875,
0.463134765625,
0.49560546875,
0.463134765625,
0.50390625,
0.431396484375,
0.50390625
],
[
0.188720703125,
0.49560546875,
0.20751953125,
0.49560546875,
0.20751953125,
0.50390625,
0.188720703125,
0.50390625
],
[
0.57275390625,
0.49560546875,
0.59228515625,
0.49560546875,
0.59228515625,
0.50390625,
0.57275390625,
0.50390625
],
[
0.81396484375,
0.49560546875,
0.8505859375,
0.49560546875,
0.8505859375,
0.505859375,
0.81396484375,
0.505859375
],
[
0.2418212890625,
0.496826171875,
0.27294921875,
0.496826171875,
0.27294921875,
0.5048828125,
0.2418212890625,
0.5048828125
],
[
0.1168212890625,
0.496826171875,
0.1478271484375,
0.496826171875,
0.1478271484375,
0.5048828125,
0.1168212890625,
0.5048828125
],
[
0.30810546875,
0.496826171875,
0.338134765625,
0.496826171875,
0.338134765625,
0.5048828125,
0.30810546875,
0.5048828125
],
[
0.381591796875,
0.496826171875,
0.39794921875,
0.496826171875,
0.39794921875,
0.50390625,
0.381591796875,
0.50390625
],
[
0.08087158203125,
0.5166015625,
0.328369140625,
0.5166015625,
0.328369140625,
0.52978515625,
0.08087158203125,
0.52978515625
],
[
0.441162109375,
0.51806640625,
0.5361328125,
0.5166015625,
0.5361328125,
0.5263671875,
0.441162109375,
0.52880859375
],
[
0.619140625,
0.5166015625,
0.89208984375,
0.5166015625,
0.89208984375,
0.52978515625,
0.619140625,
0.52978515625
],
[
0.55615234375,
0.51953125,
0.625,
0.51953125,
0.625,
0.5263671875,
0.55615234375,
0.5263671875
],
[
0.32275390625,
0.51953125,
0.34228515625,
0.51953125,
0.34228515625,
0.5263671875,
0.32275390625,
0.5263671875
],
[
0.08087158203125,
0.52978515625,
0.890625,
0.52978515625,
0.890625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.080078125,
0.54345703125,
0.89208984375,
0.54345703125,
0.89208984375,
0.556640625,
0.080078125,
0.556640625
],
[
0.080078125,
0.556640625,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.080078125,
0.57080078125
],
[
0.08087158203125,
0.5712890625,
0.85302734375,
0.5712890625,
0.85302734375,
0.58544921875,
0.08087158203125,
0.58544921875
],
[
0.236083984375,
0.5966796875,
0.313720703125,
0.5986328125,
0.313720703125,
0.6142578125,
0.236083984375,
0.6123046875
],
[
0.50341796875,
0.60107421875,
0.875,
0.60107421875,
0.875,
0.6142578125,
0.50341796875,
0.6142578125
],
[
0.50341796875,
0.6484375,
0.63134765625,
0.6484375,
0.63134765625,
0.662109375,
0.50341796875,
0.662109375
],
[
0.1004638671875,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.1004638671875,
0.673828125
],
[
0.08251953125,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.6875,
0.08251953125,
0.6875
],
[
0.52294921875,
0.6748046875,
0.8896484375,
0.6748046875,
0.8896484375,
0.68798828125,
0.52294921875,
0.68798828125
],
[
0.08087158203125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.08087158203125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.08087158203125,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.080078125,
0.73486328125
],
[
0.50244140625,
0.7197265625,
0.89208984375,
0.7197265625,
0.89208984375,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.08087158203125,
0.75
],
[
0.5048828125,
0.7373046875,
0.8896484375,
0.7373046875,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.080078125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50341796875,
0.751953125,
0.8896484375,
0.751953125,
0.8896484375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.08087158203125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5048828125,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.08087158203125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81103515625,
0.08087158203125,
0.81103515625
],
[
0.50244140625,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
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,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
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
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8935546875,
0.88720703125,
0.8935546875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"DeepFace-EMD: Re-ranking Using Patch-wise Earth Mover's Distance",
"Improves Out-Of-Distribution Face Identification",
"Hai Phan1,2",
"Anh Nguyeni",
"pthai1204@gmail.com",
"anh.ng8@gmail.com",
"1Auburn University",
"2Carnegie Mellon University",
"]",
"(a) LFW",
"(b) Masked (LFW)",
"(c) Sunglasses (LFW)",
"(d)Profile(CFP)",
"酒",
"Z91011Z:A1",
"Stage1",
"Stage 2",
"Stage1",
"Flow",
"Stage",
"Flo",
"Flow",
"Stage 2",
"Stage2",
"Stage",
"Stage",
"Flov",
"Figure 1.Traditional face identification",
"sbasedontheir",
"ce with the query (top row) at the image-level",
"sinedistanc",
"ran",
"embedding, which yields large errors upon out-of-distribution changes in the input (e.g. masks or sunglasses; b-d). We find that re-ranking",
"the top-k shortlisted faces from Stage 1 (leftmost column) using their patch-wise EMD similarity w.r.t. the query substantially improves",
"the precision (Stage 2) on challenging cases (b-d).TheFlow\"visualization intuitively shows the patch-wise reconstruction of the query",
"face using the most similar patches (i.e. highest flow) from the retrieved face. See Fig. S4 for a full figure with the top-5 candidates.",
"Abstract",
"while obtaining similar results on in-distribution images.",
"1.Introduction",
"Face identification (FI) is ubiquitous and drives many",
"high-stakedecisionsmadebythelawenforcement.Acom-",
"Who shoplifted from the Shinola luxury store in De-",
"mon FI approach compares two images by taking the co-",
"troit [4]? Who are you to receive unemployment benefits [5]",
"sine similarity between their image embeddings. Yet, such",
"or board an airplane []? Face identification (FI) today",
"approach suffers from poor out-of-distribution (OOD) gen-",
"is behind the answers to such life-critical questions. Yet.",
"eralization to new types of images (e.g., when a query face",
"thetechnologycan make errors,leadingto severeconse-",
"is masked, cropped or rotated) not included in the training",
"quences, e.g. people wrongly denied of unemployment ben-",
"set or the gallery. Here, we propose a re-ranking approach",
"efits [5] or falsely arrested [2-4, 7]. Identifying the person in",
"that compares two faces using the Earth Mover's Distance",
"asinglephotoremainschallengingbecause,inmanycases,",
"on the deep, spatial features of image patches. Our ex-",
"the problem is a zero-shot and ill-posed image retrieval task.",
"tra comparison stage explicitly examines image similarity",
"First, a deep feature extractor may not have seen a normal,",
"at a fine-grained level (e.g., eyes to eyes) and is more ro-",
"non-celebrity person before during training. Second, there",
"busttoOODperturbationsandocclusionsthantraditional",
"may be too few photos of a person in the database for FI",
"Fl. Interestingly, without finetuning feature extractors, our",
"systems to makereliable decisions.Third,itis harder to",
"method consistently improves the accuracy on all tested",
"identify when a face in the wild (e.g.from surveillance cam-",
"OOD queries: masked, cropped, rotated, and adversarial",
"eras) is occluded [44, 54] (e.g. wearing masks), distant or"
] |
{
"arxiv_id": "2112.04016",
"language": "en",
"page_idx": 0
}
|
||
2112.04016-en-1
|
2112.04016-en-1.jpg
|
cropped, yielding a new type of photo not in both the training set of deep networks and the retrieval database—i.e., out-of-distribution (OOD) data. For example, face verification accuracy may notoriously drop significantly (from 99.38% to 81.12% on LFW) given an occluded query face ( Fig. 1b- d) [ 44] or adversarial queries [ 8, 73] .
In this paper, we propose to evaluate the performance of state-of-the-art facial feature extractors (ArcFace [19],CosFace [61], and FaceNet [47]) on OOD face identification tests. That is, our main task is to recognize the person in a query image given a gallery of known faces. Besides indistribution (ID) query images, we also test FI models on OOD queries that contain (1) common occlusions, $i.e.$ random crops, faces with masks or sunglasses; and (2) adversarial perturbations[73]. Our main findings are:$^{\mathrm{l}}$
$\bullet$ Interestingly, the OOD accuracy can be substantially improved via a 2-stage approach (see Fig. 2): First, identify a set of the most globally-similar faces from the gallery using cosine distance and then, re-rank these shortlisted candidates by comparing them with the query at the patch-embedding level using the Earth Mover's Distance (EMD) [45] (Sec. 3 & Sec. 4).
$\bullet$ Across three different models (ArcFace, CosFace, and FaceNet), our re-ranking approach consistently improves the original precision (under all metrics: P@1, R-Precision, and MAP@R) without finetuning (Sec. 4). That is, interestingly, the spatial features extracted from these models can be leveraged to compare images patch-wise (in addition to image-wise) and further improve FI accuracy.
$\bullet$ On masked images [59], our re-ranking method(no training) rivals the ArcFace models finetuned directly on masked images (Sec. 4.3).
To our knowledge, our work is the first to demonstrate the remarkable effectiveness of EMD for comparing OOD, occluded and adversarial images at the deep feature level.
## $2. \textbf{ Methods}$
## $2. 1. \textbf{Problem formulation}$
To demonstrate the generality of our method, we adopt the following simple FI formulation as in [19,33,71]: Identify the person in a query image by ranking all gallery images based on their pair-wise similarity with the query. After ranking (Stage 1) or re-ranking (Stage 2), we take the identity of the top-1 nearest image as the predicted identity. $\dot{\textbf{Evaluation Following[38,71],we use three common eval-}}$ uation metrics: Precision@1 (P@ 1), R-Precision (RP), and MAP@R (M@R). See their definitions in Sec. B1 in [71].
$^{\mathrm{l}}$Code, demo and data are available at https://github.com/
anguyen8/deepface-emd
# 2.2. Networks
Pre-trained models We use three state-of-the-art PyTorch models of ArcFace, FaceNet, and CosFace pretrained on CASIA [65], VGGFace2 [14], and CASIA,
Their architectures are ResNet-18 [24],
respectively.
Inception-ResNet-v1[56], and 20-layer SphereFace[33], respectively. See Sec. S1 for more details on network architectures and implementation in PyTorch.
Image pre-processing For all networks, we align and crop input images following the 3D facial alignment in [11] (which uses 5 reference points, 0.7 and 0.6 crop ratios for width and height, and Similarity transformation). All images shown in this paper $(e.g.$ Fig. 1) are pre-processed. Using MTCNN, the default pre-processing of all three networks, does not change the results substantially (Sec. S5).
2.3. 2-stage hierarchical face identification
Stage-1: Ranking A common l-stage face identification [33,47,61] ranks gallery images based on their pairwise cosine similarity with a given query in the last-linearlayer feature space of a pre-trained feature extractor (Fig. 2). $\dot{\text{Here, our image embeddings are extracted from the }last}lin-$ ear layer of all three models and are all$\in\mathbb{R}^512.$
Stage-2: Re-ranking We re-rank the top-$k$ (where the $\bar{\mathrm{optimal~}k=100}\bar{\text{candidates from Stage }1}\bar{\text{by computing the}}$ patch-wise similarity for an image pair using EMD. Overall, we compare faces in two hierarchical stages (Fig. 2), first at a coarse, image level and then a fine-grained, patch level.
Via an ablation study (Sec. 3), we find our 2-stage approach (a.k.a. DeepFace-EMD) more accurate than Stage 1 alone $(i.e.$ no patch-wise re-ranking) and also Stage 2 alone $(i.e.$ sorting the entire gallery using patch-wise similarity).
### $2. 4. \textbf{Earth Mover' s Distance ( EMD) }$
EMD is an edit distance between two set of weighted objects or distributions [45]. Its effectiveness was first demonstrated in measuring pair-wise image similarity based on color histograms and texture frequencies [45] for image retrieval. Yet, EMD is also an effective distance between two text documents [29], probability distributions (where EMD is equivalent to Wasserstein, $i.e.$ Mallows distance)[30], and distributions in many other domains [27,34,43]. Here, we propose to harness EMD as a distance between two faces, $i.e.$ two sets of weighted facial features.
Let $Q=\{(q_{1},w_{q_{1}}),...,(q_{N},w_{q_{N}})\}$ be a set of $N$ (facial feature, weight) pairs describing a query face where $q_i$ is
feature $(e.g.$ left eye or nose) and the corresponding $w_q_i$ indicates how important the feature $q_i$ is in FI. The fow between $Q$ and the set of weighted features of a gallery face $\mathcal{G}=\{(g_{1},w_{g_{1}}),...,(g_{N},w_{g_{N}})\}$ is any matrix $\boldsymbol F=(f_{ij})\in$ $\mathbb{R}^{N\times N}.$ Intuitively, $f_ij$ is the amount of importance weight at $q_i$ that is matched to the weight at $g_j.$ Let $d_ij$ be a ground
|
[
[
0.080078125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50341796875,
0.091552734375,
0.611328125,
0.091552734375,
0.611328125,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.07843017578125,
0.106689453125,
0.4697265625,
0.10797119140625,
0.4697265625,
0.121826171875,
0.07843017578125,
0.12060546875
],
[
0.50244140625,
0.11553955078125,
0.88818359375,
0.11553955078125,
0.88818359375,
0.1287841796875,
0.50244140625,
0.1287841796875
],
[
0.080078125,
0.12249755859375,
0.4697265625,
0.12249755859375,
0.4697265625,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.5009765625,
0.1287841796875,
0.890625,
0.1307373046875,
0.890625,
0.1458740234375,
0.5009765625,
0.143310546875
],
[
0.080078125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.080078125,
0.1514892578125
],
[
0.080078125,
0.15087890625,
0.4697265625,
0.15283203125,
0.4697265625,
0.1673583984375,
0.080078125,
0.1654052734375
],
[
0.50244140625,
0.1458740234375,
0.890625,
0.1458740234375,
0.890625,
0.1590576171875,
0.50244140625,
0.1590576171875
],
[
0.50341796875,
0.1634521484375,
0.5908203125,
0.1634521484375,
0.5908203125,
0.173583984375,
0.50341796875,
0.173583984375
],
[
0.6044921875,
0.1610107421875,
0.890625,
0.1610107421875,
0.890625,
0.1749267578125,
0.6044921875,
0.1749267578125
],
[
0.08087158203125,
0.1685791015625,
0.378173828125,
0.1685791015625,
0.378173828125,
0.1817626953125,
0.08087158203125,
0.1817626953125
],
[
0.6044921875,
0.1654052734375,
0.60888671875,
0.1654052734375,
0.60888671875,
0.17041015625,
0.6044921875,
0.17041015625
],
[
0.096435546875,
0.1817626953125,
0.4697265625,
0.18115234375,
0.4697265625,
0.1956787109375,
0.096435546875,
0.1962890625
],
[
0.50244140625,
0.1767578125,
0.890625,
0.1767578125,
0.890625,
0.1900634765625,
0.50244140625,
0.1900634765625
],
[
0.50244140625,
0.1912841796875,
0.890625,
0.1912841796875,
0.890625,
0.2052001953125,
0.50244140625,
0.2052001953125
],
[
0.08087158203125,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08087158203125,
0.210205078125
],
[
0.50244140625,
0.20703125,
0.79150390625,
0.20703125,
0.79150390625,
0.2203369140625,
0.50244140625,
0.2203369140625
],
[
0.080078125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.080078125,
0.2266845703125
],
[
0.50244140625,
0.222900390625,
0.89208984375,
0.222900390625,
0.89208984375,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.5009765625,
0.2374267578125,
0.89208984375,
0.236083984375,
0.89208984375,
0.250732421875,
0.5009765625,
0.251953125
],
[
0.080078125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.080078125,
0.2568359375
],
[
0.50341796875,
0.25244140625,
0.89208984375,
0.25244140625,
0.89208984375,
0.266357421875,
0.50341796875,
0.266357421875
],
[
0.080078125,
0.2587890625,
0.46728515625,
0.2587890625,
0.46728515625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.08087158203125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.08087158203125,
0.287353515625
],
[
0.50341796875,
0.268310546875,
0.890625,
0.268310546875,
0.890625,
0.2822265625,
0.50341796875,
0.2822265625
],
[
0.080078125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.50341796875,
0.2841796875,
0.890625,
0.2841796875,
0.890625,
0.29736328125,
0.50341796875,
0.29736328125
],
[
0.50244140625,
0.296630859375,
0.890625,
0.298583984375,
0.890625,
0.3125,
0.50244140625,
0.31005859375
],
[
0.080078125,
0.3037109375,
0.39794921875,
0.3037109375,
0.39794921875,
0.317626953125,
0.080078125,
0.317626953125
],
[
0.50244140625,
0.313232421875,
0.88232421875,
0.313232421875,
0.88232421875,
0.32763671875,
0.50244140625,
0.32763671875
],
[
0.098876953125,
0.32763671875,
0.465576171875,
0.32763671875,
0.465576171875,
0.34228515625,
0.098876953125,
0.34228515625
],
[
0.50244140625,
0.33837890625,
0.82666015625,
0.34033203125,
0.82666015625,
0.354736328125,
0.50244140625,
0.352783203125
],
[
0.111083984375,
0.3447265625,
0.46728515625,
0.3447265625,
0.46728515625,
0.35791015625,
0.111083984375,
0.35791015625
],
[
0.11273193359375,
0.35791015625,
0.468017578125,
0.35791015625,
0.468017578125,
0.371826171875,
0.11273193359375,
0.371826171875
],
[
0.50244140625,
0.3642578125,
0.6435546875,
0.3642578125,
0.6435546875,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.6357421875,
0.3642578125,
0.89208984375,
0.3642578125,
0.89208984375,
0.37744140625,
0.6357421875,
0.37744140625
],
[
0.11273193359375,
0.373779296875,
0.468017578125,
0.373779296875,
0.468017578125,
0.386962890625,
0.11273193359375,
0.386962890625
],
[
0.50244140625,
0.380126953125,
0.89208984375,
0.380126953125,
0.89208984375,
0.393310546875,
0.50244140625,
0.393310546875
],
[
0.111083984375,
0.38818359375,
0.46728515625,
0.38818359375,
0.46728515625,
0.401611328125,
0.111083984375,
0.401611328125
],
[
0.50341796875,
0.395751953125,
0.890625,
0.395751953125,
0.890625,
0.4091796875,
0.50341796875,
0.4091796875
],
[
0.11273193359375,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.11273193359375,
0.417236328125
],
[
0.50244140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.111083984375,
0.41845703125,
0.431396484375,
0.41845703125,
0.431396484375,
0.431884765625,
0.111083984375,
0.431884765625
],
[
0.5009765625,
0.4248046875,
0.89208984375,
0.423583984375,
0.89208984375,
0.438232421875,
0.5009765625,
0.439453125
],
[
0.098876953125,
0.44140625,
0.468017578125,
0.442626953125,
0.468017578125,
0.45703125,
0.098876953125,
0.455810546875
],
[
0.50244140625,
0.43994140625,
0.80810546875,
0.438232421875,
0.80810546875,
0.45263671875,
0.50244140625,
0.455078125
],
[
0.111083984375,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.472900390625,
0.111083984375,
0.472900390625
],
[
0.50244140625,
0.455810546875,
0.89208984375,
0.455810546875,
0.89208984375,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.11273193359375,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.48486328125,
0.11273193359375,
0.48486328125
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.111083984375,
0.486083984375,
0.4697265625,
0.488037109375,
0.4697265625,
0.50244140625,
0.111083984375,
0.50048828125
],
[
0.50244140625,
0.486083984375,
0.890625,
0.486083984375,
0.890625,
0.5,
0.50244140625,
0.5
],
[
0.11358642578125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.517578125,
0.11358642578125,
0.517578125
],
[
0.50341796875,
0.501953125,
0.89208984375,
0.501953125,
0.89208984375,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.11273193359375,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.11273193359375,
0.5322265625
],
[
0.50244140625,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50244140625,
0.53076171875
],
[
0.11273193359375,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.11273193359375,
0.54736328125
],
[
0.521484375,
0.5322265625,
0.890625,
0.53271484375,
0.890625,
0.546875,
0.521484375,
0.54541015625
],
[
0.111083984375,
0.548828125,
0.281982421875,
0.548828125,
0.281982421875,
0.5625,
0.111083984375,
0.5625
],
[
0.50341796875,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.50341796875,
0.56103515625
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.09722900390625,
0.57080078125,
0.468017578125,
0.57275390625,
0.468017578125,
0.5859375,
0.09722900390625,
0.583984375
],
[
0.50244140625,
0.5791015625,
0.8837890625,
0.5791015625,
0.8837890625,
0.59228515625,
0.50244140625,
0.59228515625
],
[
0.1094970703125,
0.5859375,
0.468017578125,
0.5869140625,
0.468017578125,
0.6015625,
0.1094970703125,
0.60107421875
],
[
0.111083984375,
0.60302734375,
0.3056640625,
0.60302734375,
0.3056640625,
0.6162109375,
0.111083984375,
0.6162109375
],
[
0.50244140625,
0.60400390625,
0.775390625,
0.60400390625,
0.775390625,
0.6171875,
0.50244140625,
0.6171875
],
[
0.098876953125,
0.626953125,
0.4697265625,
0.626953125,
0.4697265625,
0.6416015625,
0.098876953125,
0.6416015625
],
[
0.521484375,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.521484375,
0.642578125
],
[
0.08087158203125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.65576171875,
0.08087158203125,
0.65576171875
],
[
0.50244140625,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.65771484375,
0.45751953125,
0.65771484375,
0.45751953125,
0.671875,
0.080078125,
0.671875
],
[
0.50244140625,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.50244140625,
0.673828125
],
[
0.50341796875,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.080078125,
0.68505859375,
0.1748046875,
0.68505859375,
0.1748046875,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50244140625,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.07843017578125,
0.70849609375,
0.275390625,
0.708984375,
0.275390625,
0.7236328125,
0.07843017578125,
0.72265625
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.50244140625,
0.7197265625,
0.890625,
0.7197265625,
0.890625,
0.73291015625,
0.50244140625,
0.73291015625
],
[
0.09722900390625,
0.73095703125,
0.468017578125,
0.73193359375,
0.468017578125,
0.7470703125,
0.09722900390625,
0.74560546875
],
[
0.50341796875,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50341796875,
0.74853515625
],
[
0.07843017578125,
0.7470703125,
0.468017578125,
0.74755859375,
0.468017578125,
0.76220703125,
0.07843017578125,
0.76123046875
],
[
0.50341796875,
0.75,
0.89208984375,
0.75,
0.89208984375,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.76513671875,
0.8037109375,
0.76513671875,
0.8037109375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.080078125,
0.77880859375,
0.468017578125,
0.77880859375,
0.468017578125,
0.79248046875,
0.080078125,
0.79248046875
],
[
0.51953125,
0.77978515625,
0.8935546875,
0.77978515625,
0.8935546875,
0.79736328125,
0.51953125,
0.79736328125
],
[
0.080078125,
0.79345703125,
0.468017578125,
0.79345703125,
0.468017578125,
0.806640625,
0.080078125,
0.806640625
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.81005859375,
0.465576171875,
0.81005859375,
0.465576171875,
0.82080078125,
0.08087158203125,
0.82080078125
],
[
0.5,
0.806640625,
0.8896484375,
0.81005859375,
0.8896484375,
0.82763671875,
0.5,
0.82470703125
],
[
0.080078125,
0.82373046875,
0.468017578125,
0.82373046875,
0.468017578125,
0.8369140625,
0.080078125,
0.8369140625
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.83837890625,
0.4697265625,
0.83837890625,
0.4697265625,
0.8515625,
0.080078125,
0.8515625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.07843017578125,
0.85302734375,
0.461669921875,
0.85498046875,
0.461669921875,
0.86865234375,
0.07843017578125,
0.8662109375
],
[
0.5009765625,
0.85400390625,
0.89208984375,
0.85498046875,
0.89208984375,
0.87255859375,
0.5009765625,
0.8720703125
],
[
0.5,
0.8671875,
0.8935546875,
0.869140625,
0.8935546875,
0.88818359375,
0.5,
0.88623046875
],
[
0.096435546875,
0.87353515625,
0.468017578125,
0.875,
0.468017578125,
0.8896484375,
0.096435546875,
0.88818359375
],
[
0.080078125,
0.88818359375,
0.2484130859375,
0.88818359375,
0.2484130859375,
0.9013671875,
0.080078125,
0.9013671875
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"cropped, yielding a new type of photo not in both the train-",
"2.2.Networks",
"ing set of deep networks and the retrieval databasei.e.,",
"Pre-trained models We use three state-of-the-art Py-",
"out-of-distribution (OOD) data. For example, face verifi-",
"Torch models of ArcFace, FaceNet, and CosFace pre-",
"cation accuracy may notoriously drop significantly (from",
"99.38% to 81.12% on LFW) given an occluded query face",
"trained on CASIA [65], VGGFace2 [14], and CASIA,",
"respectively.",
"Their architectures are ResNet-18 [24],",
"(Fig. 1b-d) [44] or adversarial queries [8, 73]",
"",
"In this paper,wepropose to evaluate the performance of",
"Inception-ResNet-v1 [56], and 20-layer SphereFace [33],",
"respectively. See Sec. Sl for more details on network ar-",
"state-of-the-artfacialfeatureextractors (ArcFace[19l,Cos-",
"chitectures and implementation in PyTorch.",
"Face [6]], and FaceNet [47]] on OOD face identification",
"Image pre-processingFor all networks, we align and",
"tests.That is, our main task is to recognize the person in",
"crop input images following the 3D facial alignment in [1]]",
"a query image given a gallery of known faces. Besides in-",
"(which uses 5 reference points, 0.7 and 0.6 crop ratios for",
"distribution (ID) query images, we also test FI models on",
"OOD queries that contain (1) common occlusions, i.e. ran-",
"width and height, and Similarity transformation). All im-",
"dom crops, faces with masks or sunglasses; and (2) adver-",
"ages shown in this paper (e.g. Fig. I) are pre-processed.",
"Using MTCNN,the default pre-processing of all three net-",
"sarial perturbations [73]. Our main findings are:",
"works, does not change the results substantially (Sec. S5).",
". Interestingly, the OOD accuracy can be substantially",
"2.3. 2-stage hierarchical face identification",
"improved via a 2-stage approach (see Fig. 2): First,",
"identify a set of the most globally-similar faces from",
"Stage-l: Ranking",
"A common 1-stagefaceidentifica-",
"the gallery using cosine distance and then, re-rank",
"tion [33, 47, 61] ranks gallery images based on their pair-",
"these shortlisted candidates by comparing them with",
"wise cosine similarity with a given query in the last-linear-",
"the query at the patch-embedding level using the Earth",
"layer feature space of a pre-trained feature extractor (Fig.2).",
"Mover's Distance (EMD) [45] (Sec. 3 & Sec. 4).",
"Here, our image embeddings are extracted from the last lin-",
"· Across three different models (ArcFace, CosFace,",
"ear layer of all three models and are all e IR512",
"and FaceNet), our re-ranking approach consistently",
"Stage-2: Re-rankingWe re-rank the top-k (where the",
"improves the original precision (under all metrics:",
"optimal k = 100) candidates from Stage 1 by computing the",
"P@1, R-Precision, and MAP@R) without finetuning",
"patch-wise similarity for an image pair using EMD. Overall,",
"(Sec. 4). That is, interestingly, the spatial features ex-",
"we compare faces in two hierarchical stages (Fig. 2), first",
"tracted from these models can be leveraged to compare",
"at a coarse, image level and then a fine-grained, patch level.",
"images patch-wise (in addition to image-wise) and fur-",
"Via an ablation study (Sec.3), we find our 2-stage ap-",
"ther improve FI accuracy.",
"proach (a.k.a. DeepFace-EMD) more accurate than Stage 1",
"alone (i.e. no patch-wise re-ranking) and also Stage 2 alone",
".On masked images [59],our re-ranking method (no",
"(i.e. sorting the entire gallery using patch-wise similarity).",
"training) rivals the ArcFace models finetuned directly",
"on masked images (Sec. 4.3).",
"2.4. Earth Mover's Distance (EMD)",
"To our knowledge, our work is the first to demonstrate",
"EMD is an edit distance between two set of weighted ob-",
"the remarkable effectiveness of EMD for comparing OOD,",
"jects or distributions [45]. Its effectiveness was first demon-",
"occluded and adversarial images at the deep feature level.",
"strated in measuring pair-wise image similarity based on",
"color histograms and texture frequencies [45] for image re-",
"2. Methods",
"trieval. Yet, EMD is also an effective distance between two",
"2.1.Problem formulation",
"text documents [29], probability distributions (where EMD",
"is equivalent to Wasserstein, i.e. Mallows distance) [30],",
"Todemonstrate thegenerality of our method, we adopt",
"and distributions in many other domains [27, 34, 43]. Here,",
"the following simple FI formulation as in [19,33, 71]: Iden-",
"we propose to harness EMD as a distance between two",
"tify the person in a query image by ranking all gallery im-",
"faces, i.e. two sets of weighted facial features.",
"ages based on their pair-wise similarity with the query. Af-",
"Let Q = ((q1, Wqi),.., (qN, Wan )) be a set of N (facial",
"ter ranking (Stage 1) or re-ranking (Stage 2), we take the",
"feature, weight) pairs describing a query face where qi is",
"identityof thetop-1nearestimageasthepredicted identity",
"a feature (e.g. left eye or nose) and the corresponding wqt",
"Evaluation Following [38,71], we use three common eval-",
"indicates how important the feature qi is in FI. The flow",
"uation metrics: Precision@1 (P@1), R-Precision (RP), and",
"between Q and the set of weighted features of a gallery face",
"MAP@R (M@R).See their definitions in Sec.B1 in [71].",
"G = (gi, Wgi),., (gN, Wgn)) is any matrix F = (fuj) E",
"RN xN Intuitively, fi, is the amount of importance weight",
"ICode, demo and data are available at https : / /github.com/",
"anguyen8/deepface-emd",
"at qi that is matched to the weight at gj. Let di, be a ground"
] |
{
"arxiv_id": "2112.04016",
"language": "en",
"page_idx": 1
}
|
||
2112.04016-en-2
|
2112.04016-en-2.jpg
|
<FigureHere>
Optimal flow
Ground distance matrix
Figure 2. Our 2-stage face identification pipeline. Stage 1 ranks gallery images based on their cosine distance with the query face at the
image-embedding level. Stage 2 then re-ranks the top-$k$ shortlisted candidates from Stage 1 using EMD at the patch-embedding level.
distance between $(q_i,g_j)$ and $\boldsymbol{D}=(d_{ij})\in\mathbb{R}^{N\times N}$ be the
ground distance matrix of all pair-wise distances.
We want to find an optimal flow $F$ that minimizes the following cost function, $i.e.$ the sum of weighted pair-wise distances across the two sets of facial features:
(1)
$$\operatorname{COST}(\mathcal{Q},\mathcal{G},\boldsymbol{F})=\sum_{i=1}^N\sum_{j=1}^Nd_{ij}f_{ij}$$
by [60], we also divide an image into a grid but we take the embeddings of the local patches from the last convolutional layers of each network. That is, in FI, face images are aligned and cropped such that the entire face covers most of the image (see Fig. la). Therefore, without facial occlusion, every image patch is supposed to contain useful identity information, which is in contrast to natural photos [66].
Our grid sizes $H\times W$ for ArcFace, FaceNet, and CosFace are respectively, $8\times8,3\times3$, and 6×7, which are the corresponding spatial dimensions of their last convolutional layers (see definitions of these layers in Sec. S1). That is, each feature $q_i$ is an embedding of size 1×1×C where C is the number of channels $(i.e.512,1792$, and 512 for ArcFace, FaceNet, and CosFace, respectively).
Ground distance Like[66,71],we use cosine distance as the ground distance $d_ij$ between the embeddings ($q_i,g_j)$ of two patches:
$$d_{ij}=1-\frac{\langle q_i,g_j\rangle}{\left\|q_i\right\|\left\|g_j\right\|}$$
(2)
$$\begin{aligned}&f_{ij}\geq0\\&\sum_{j=1}^{N}f_{ij}\leq w_{q_{i}},\mathrm{and}\sum_{i=1}^{N}f_{ij}\leq w_{g_{j}},i,j\in[1,N]\\&\sum_{j=1}^{N}\sum_{i=1}^{N}f_{ij}=\operatorname*{min}\left(\sum_{j=1}^{N}w_{g_{j}},\sum_{i=1}^{N}w_{q_{i}}\right).\end{aligned}$$
(3
(4)
(5)
where $\langle.\rangle$ is the dot product between two feature vectors.
## 2.5. Feature weighting
As in [66,71], we normalize the weights of a face such that the total weights of features is 1 $i.e.\sum_{i=1}^Nw_{q_i}=$ $\sum_{j=1}^{N}w_{g_{j}}=1$,which is also the total flow in Eq.(4). Note that EMD is a metric iff two distributions have an equal total weight and the ground distance function is a metric[16].
We use the iterative Sinkhorn algorithm[18] to effciently solve the linear programming problem in Eq.(1), which yields the final EMD between two faces $Q$ and $\mathcal{G}.$
Facial features In image retrieval using EMD, a set of features $\{q_i\}$ can be a collection of dominant colors [45], spatial frequencies [45], or a histogram-like descriptor based on the local patches of reference identities [60]. Inspired
EMD in our FI intuitively is an optimal plan to match all weighted features across two images. Therefore, how to weight features is an important step. Here, we thoroughly explore five different feature-weighting techniques for FI. $\hat{\textbf{Uniform}}$ Zhang et al.[66] found that it is beneficial to assign lower weight to less informative regions (e.g. background or occlusion) and higher weight to discriminative
|
[
[
0.5556640625,
0.0953369140625,
0.58642578125,
0.0953369140625,
0.58642578125,
0.111083984375,
0.5556640625,
0.111083984375
],
[
0.615234375,
0.0953369140625,
0.84228515625,
0.0953369140625,
0.84228515625,
0.10919189453125,
0.615234375,
0.10919189453125
],
[
0.779296875,
0.11932373046875,
0.82275390625,
0.11932373046875,
0.82275390625,
0.1351318359375,
0.779296875,
0.1351318359375
],
[
0.6298828125,
0.1256103515625,
0.76123046875,
0.1275634765625,
0.76123046875,
0.14453125,
0.6298828125,
0.1427001953125
],
[
0.77392578125,
0.1363525390625,
0.830078125,
0.1363525390625,
0.830078125,
0.15087890625,
0.77392578125,
0.15087890625
],
[
0.316162109375,
0.147705078125,
0.373291015625,
0.147705078125,
0.373291015625,
0.1578369140625,
0.316162109375,
0.1578369140625
],
[
0.459228515625,
0.1483154296875,
0.48046875,
0.1483154296875,
0.48046875,
0.1572265625,
0.459228515625,
0.1572265625
],
[
0.56201171875,
0.173583984375,
0.57177734375,
0.173583984375,
0.57177734375,
0.18115234375,
0.56201171875,
0.18115234375
],
[
0.236083984375,
0.1900634765625,
0.26953125,
0.1900634765625,
0.26953125,
0.20703125,
0.236083984375,
0.20703125
],
[
0.303955078125,
0.1932373046875,
0.3349609375,
0.1932373046875,
0.3349609375,
0.2020263671875,
0.303955078125,
0.2020263671875
],
[
0.45751953125,
0.222900390625,
0.48046875,
0.222900390625,
0.48046875,
0.2342529296875,
0.45751953125,
0.2342529296875
],
[
0.336669921875,
0.2310791015625,
0.35693359375,
0.2310791015625,
0.35693359375,
0.23486328125,
0.336669921875,
0.23486328125
],
[
0.57275390625,
0.24560546875,
0.8447265625,
0.246826171875,
0.8447265625,
0.261474609375,
0.57275390625,
0.260009765625
],
[
0.57421875,
0.264404296875,
0.70751953125,
0.264404296875,
0.70751953125,
0.2822265625,
0.57421875,
0.2822265625
],
[
0.2279052734375,
0.285888671875,
0.233642578125,
0.285888671875,
0.233642578125,
0.291748046875,
0.2279052734375,
0.291748046875
],
[
0.313720703125,
0.2822265625,
0.326904296875,
0.2822265625,
0.326904296875,
0.29345703125,
0.313720703125,
0.29345703125
],
[
0.418212890625,
0.2890625,
0.450927734375,
0.2890625,
0.450927734375,
0.3017578125,
0.418212890625,
0.3017578125
],
[
0.26318359375,
0.29736328125,
0.289306640625,
0.29736328125,
0.289306640625,
0.309326171875,
0.26318359375,
0.309326171875
],
[
0.7587890625,
0.294189453125,
0.8388671875,
0.294189453125,
0.8388671875,
0.3212890625,
0.7587890625,
0.3212890625
],
[
0.348876953125,
0.298583984375,
0.38818359375,
0.298583984375,
0.38818359375,
0.309326171875,
0.348876953125,
0.309326171875
],
[
0.1422119140625,
0.298583984375,
0.200927734375,
0.298583984375,
0.200927734375,
0.309326171875,
0.1422119140625,
0.309326171875
],
[
0.3505859375,
0.30810546875,
0.39453125,
0.30810546875,
0.39453125,
0.31494140625,
0.3505859375,
0.31494140625
],
[
0.76708984375,
0.32763671875,
0.7900390625,
0.32763671875,
0.7900390625,
0.339111328125,
0.76708984375,
0.339111328125
],
[
0.65625,
0.337158203125,
0.7060546875,
0.343505859375,
0.70361328125,
0.35595703125,
0.65283203125,
0.349853515625
],
[
0.6044921875,
0.34521484375,
0.63134765625,
0.34521484375,
0.63134765625,
0.349853515625,
0.6044921875,
0.349853515625
],
[
0.31201171875,
0.359130859375,
0.325927734375,
0.359130859375,
0.325927734375,
0.37060546875,
0.31201171875,
0.37060546875
],
[
0.4150390625,
0.363037109375,
0.453369140625,
0.363037109375,
0.453369140625,
0.37890625,
0.4150390625,
0.37890625
],
[
0.6953125,
0.369384765625,
0.70751953125,
0.369384765625,
0.70751953125,
0.375732421875,
0.6953125,
0.375732421875
],
[
0.7138671875,
0.369384765625,
0.72607421875,
0.369384765625,
0.72607421875,
0.37890625,
0.7138671875,
0.37890625
],
[
0.71728515625,
0.3642578125,
0.72802734375,
0.3642578125,
0.72802734375,
0.367431640625,
0.71728515625,
0.367431640625
],
[
0.73779296875,
0.371826171875,
0.74658203125,
0.371826171875,
0.74658203125,
0.375,
0.73779296875,
0.375
],
[
0.77294921875,
0.371826171875,
0.833984375,
0.371826171875,
0.833984375,
0.3818359375,
0.77294921875,
0.3818359375
],
[
0.69921875,
0.37255859375,
0.70751953125,
0.37890625,
0.701171875,
0.3837890625,
0.69287109375,
0.37744140625
],
[
0.2646484375,
0.373779296875,
0.289306640625,
0.373779296875,
0.289306640625,
0.385009765625,
0.2646484375,
0.385009765625
],
[
0.1405029296875,
0.3818359375,
0.21728515625,
0.3818359375,
0.21728515625,
0.392578125,
0.1405029296875,
0.392578125
],
[
0.7158203125,
0.37744140625,
0.7255859375,
0.37744140625,
0.7255859375,
0.3818359375,
0.7158203125,
0.3818359375
],
[
0.74267578125,
0.385009765625,
0.7509765625,
0.385009765625,
0.7509765625,
0.389404296875,
0.74267578125,
0.389404296875
],
[
0.5458984375,
0.395751953125,
0.64892578125,
0.395751953125,
0.64892578125,
0.406005859375,
0.5458984375,
0.406005859375
],
[
0.67724609375,
0.393310546875,
0.736328125,
0.393310546875,
0.736328125,
0.40478515625,
0.67724609375,
0.40478515625
],
[
0.080078125,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.080078125,
0.436279296875,
0.86669921875,
0.436279296875,
0.86669921875,
0.449462890625,
0.080078125,
0.449462890625
],
[
0.07843017578125,
0.458251953125,
0.471435546875,
0.455078125,
0.471435546875,
0.473388671875,
0.07843017578125,
0.4765625
],
[
0.5009765625,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.07843017578125,
0.47607421875,
0.40185546875,
0.474853515625,
0.40185546875,
0.488037109375,
0.07843017578125,
0.4892578125
],
[
0.50341796875,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.098876953125,
0.490478515625,
0.4697265625,
0.490478515625,
0.4697265625,
0.50390625,
0.098876953125,
0.50390625
],
[
0.50244140625,
0.490478515625,
0.89208984375,
0.490478515625,
0.89208984375,
0.50390625,
0.50244140625,
0.50390625
],
[
0.07843017578125,
0.50390625,
0.4697265625,
0.5048828125,
0.4697265625,
0.51953125,
0.07843017578125,
0.51806640625
],
[
0.50244140625,
0.505859375,
0.8935546875,
0.505859375,
0.8935546875,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.52099609375,
0.386474609375,
0.52099609375,
0.386474609375,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.50244140625,
0.52099609375,
0.89208984375,
0.52099609375,
0.89208984375,
0.5341796875,
0.50244140625,
0.5341796875
],
[
0.29736328125,
0.537109375,
0.32275390625,
0.537109375,
0.32275390625,
0.5498046875,
0.29736328125,
0.5498046875
],
[
0.318603515625,
0.537109375,
0.339111328125,
0.537109375,
0.339111328125,
0.5498046875,
0.318603515625,
0.5498046875
],
[
0.5009765625,
0.53515625,
0.8896484375,
0.53515625,
0.8896484375,
0.548828125,
0.5009765625,
0.548828125
],
[
0.28759765625,
0.54443359375,
0.38818359375,
0.54736328125,
0.386474609375,
0.5693359375,
0.285888671875,
0.56640625
],
[
0.16259765625,
0.5517578125,
0.294189453125,
0.5517578125,
0.294189453125,
0.56494140625,
0.16259765625,
0.56494140625
],
[
0.44677734375,
0.55322265625,
0.4697265625,
0.55322265625,
0.4697265625,
0.56494140625,
0.44677734375,
0.56494140625
],
[
0.5009765625,
0.5498046875,
0.8505859375,
0.5498046875,
0.8505859375,
0.56298828125,
0.5009765625,
0.56298828125
],
[
0.294189453125,
0.5693359375,
0.34228515625,
0.5693359375,
0.34228515625,
0.580078125,
0.294189453125,
0.580078125
],
[
0.521484375,
0.56494140625,
0.890625,
0.564453125,
0.890625,
0.5791015625,
0.521484375,
0.580078125
],
[
0.50244140625,
0.58056640625,
0.89208984375,
0.58056640625,
0.89208984375,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.08905029296875,
0.5947265625,
0.11358642578125,
0.5947265625,
0.11358642578125,
0.60693359375,
0.08905029296875,
0.60693359375
],
[
0.1568603515625,
0.5947265625,
0.2115478515625,
0.59228515625,
0.2132568359375,
0.60693359375,
0.158447265625,
0.609375
],
[
0.44677734375,
0.5947265625,
0.4697265625,
0.5947265625,
0.4697265625,
0.60791015625,
0.44677734375,
0.60791015625
],
[
0.5009765625,
0.59521484375,
0.89208984375,
0.5947265625,
0.89208984375,
0.609375,
0.5009765625,
0.60986328125
],
[
0.261474609375,
0.611328125,
0.279296875,
0.611328125,
0.279296875,
0.62548828125,
0.261474609375,
0.62548828125
],
[
0.50244140625,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.50244140625,
0.625
],
[
0.1380615234375,
0.6220703125,
0.15283203125,
0.62744140625,
0.146240234375,
0.63818359375,
0.131591796875,
0.6318359375
],
[
0.136474609375,
0.6123046875,
0.1544189453125,
0.6123046875,
0.1544189453125,
0.62548828125,
0.136474609375,
0.62548828125
],
[
0.28515625,
0.626953125,
0.468017578125,
0.62548828125,
0.468017578125,
0.64013671875,
0.28515625,
0.6416015625
],
[
0.5009765625,
0.62548828125,
0.89208984375,
0.625,
0.89208984375,
0.6396484375,
0.5009765625,
0.64013671875
],
[
0.50244140625,
0.64013671875,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50244140625,
0.65380859375
],
[
0.156005859375,
0.62744140625,
0.273681640625,
0.625,
0.273681640625,
0.6396484375,
0.156005859375,
0.642578125
],
[
0.13232421875,
0.64306640625,
0.158447265625,
0.64306640625,
0.158447265625,
0.65576171875,
0.13232421875,
0.65576171875
],
[
0.261474609375,
0.64453125,
0.281005859375,
0.64453125,
0.281005859375,
0.6513671875,
0.261474609375,
0.6513671875
],
[
0.111083984375,
0.6591796875,
0.156005859375,
0.6591796875,
0.156005859375,
0.673828125,
0.111083984375,
0.673828125
],
[
0.5009765625,
0.65478515625,
0.7861328125,
0.65576171875,
0.7861328125,
0.67041015625,
0.5009765625,
0.66943359375
],
[
0.2467041015625,
0.662109375,
0.26318359375,
0.662109375,
0.26318359375,
0.67041015625,
0.2467041015625,
0.67041015625
],
[
0.30810546875,
0.6611328125,
0.32275390625,
0.6611328125,
0.32275390625,
0.671875,
0.30810546875,
0.671875
],
[
0.1192626953125,
0.66845703125,
0.233642578125,
0.67236328125,
0.2320556640625,
0.6943359375,
0.11767578125,
0.68994140625
],
[
0.50244140625,
0.671875,
0.89208984375,
0.671875,
0.89208984375,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.3095703125,
0.67529296875,
0.314453125,
0.67529296875,
0.314453125,
0.68310546875,
0.3095703125,
0.68310546875
],
[
0.1094970703125,
0.673828125,
0.1258544921875,
0.673828125,
0.1258544921875,
0.689453125,
0.1094970703125,
0.689453125
],
[
0.328369140625,
0.67822265625,
0.348876953125,
0.68310546875,
0.346435546875,
0.68994140625,
0.325927734375,
0.68505859375
],
[
0.4453125,
0.67529296875,
0.4697265625,
0.67529296875,
0.4697265625,
0.68798828125,
0.4453125,
0.68798828125
],
[
0.285888671875,
0.681640625,
0.301513671875,
0.6845703125,
0.2998046875,
0.689453125,
0.28515625,
0.6875
],
[
0.50244140625,
0.6875,
0.8935546875,
0.6875,
0.8935546875,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.271240234375,
0.6796875,
0.28759765625,
0.6796875,
0.28759765625,
0.6875,
0.271240234375,
0.6875
],
[
0.1070556640625,
0.6923828125,
0.158447265625,
0.6923828125,
0.158447265625,
0.70263671875,
0.1070556640625,
0.70263671875
],
[
0.2418212890625,
0.693359375,
0.26708984375,
0.693359375,
0.26708984375,
0.7021484375,
0.2418212890625,
0.7021484375
],
[
0.306396484375,
0.693359375,
0.325927734375,
0.693359375,
0.325927734375,
0.7021484375,
0.306396484375,
0.7021484375
],
[
0.50341796875,
0.70263671875,
0.58837890625,
0.70263671875,
0.58837890625,
0.7138671875,
0.50341796875,
0.7138671875
],
[
0.7060546875,
0.71044921875,
0.75634765625,
0.71044921875,
0.75634765625,
0.724609375,
0.7060546875,
0.724609375
],
[
0.62744140625,
0.71728515625,
0.69287109375,
0.71728515625,
0.69287109375,
0.73291015625,
0.62744140625,
0.73291015625
],
[
0.869140625,
0.71826171875,
0.8935546875,
0.71826171875,
0.8935546875,
0.73193359375,
0.869140625,
0.73193359375
],
[
0.69921875,
0.7265625,
0.7646484375,
0.7265625,
0.7646484375,
0.7412109375,
0.69921875,
0.7412109375
],
[
0.08087158203125,
0.7197265625,
0.46728515625,
0.7197265625,
0.46728515625,
0.72998046875,
0.08087158203125,
0.72998046875
],
[
0.080078125,
0.73193359375,
0.390625,
0.73193359375,
0.390625,
0.74560546875,
0.080078125,
0.74560546875
],
[
0.38232421875,
0.734375,
0.40185546875,
0.73193359375,
0.40185546875,
0.736328125,
0.384033203125,
0.73779296875
],
[
0.406982421875,
0.736328125,
0.4453125,
0.736328125,
0.4453125,
0.74755859375,
0.406982421875,
0.74755859375
],
[
0.44775390625,
0.7392578125,
0.463134765625,
0.7373046875,
0.464111328125,
0.7412109375,
0.449462890625,
0.74365234375
],
[
0.09478759765625,
0.74853515625,
0.4697265625,
0.74755859375,
0.4697265625,
0.76513671875,
0.09478759765625,
0.7666015625
],
[
0.51953125,
0.7470703125,
0.89208984375,
0.74853515625,
0.89208984375,
0.76318359375,
0.51953125,
0.76123046875
],
[
0.080078125,
0.76513671875,
0.468017578125,
0.76513671875,
0.468017578125,
0.77880859375,
0.080078125,
0.77880859375
],
[
0.08251953125,
0.75,
0.1070556640625,
0.75,
0.1070556640625,
0.76123046875,
0.08251953125,
0.76123046875
],
[
0.50244140625,
0.77099609375,
0.6748046875,
0.7734375,
0.6748046875,
0.78857421875,
0.50244140625,
0.78515625
],
[
0.080078125,
0.78076171875,
0.449462890625,
0.78076171875,
0.449462890625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.1004638671875,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.1004638671875,
0.81005859375
],
[
0.521484375,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.521484375,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.82568359375,
0.449462890625,
0.82568359375,
0.449462890625,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.84130859375,
0.8798828125,
0.84130859375,
0.8798828125,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.465576171875,
0.85693359375,
0.465576171875,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.869140625,
0.50341796875,
0.869140625
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.5009765625,
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
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"fa",
"Stage 1: Image similarity-based ranking",
"(fq,fa)",
"distance (fq,fg) = 1",
"Ifalllfall",
"Query image",
"CNN",
"g",
"",
"",
"CNN",
"Tima",
"Stage 2: Patch-wise similarity-based re-ranking",
"",
"",
"qi",
"(wa,)",
"CNN",
"酒香",
"Feature",
"Query image",
"Weiphting",
"flow",
"C",
"",
"9j",
"(wg)",
"0.2",
"0.6",
"",
"",
"EMD distance",
"",
"CNN",
"Top-k candidates",
"O",
"S",
"Ground distance matrix",
"Optimal flow",
"Figure 2. Our 2-stage face identification pipeline. Stage 1 ranks gallery images based on their cosine distance with the query face at the",
"image-embedding level. Stage 2 then re-ranks the top-k shortlisted candidates from Stage 1 using EMD at the patch-embedding level.",
"distance between (qi, gj) and D = (di;) E RNxN be the",
"by [6O], we also divide an image into a grid but we take",
"ground distance matrix of all pair-wise distances.",
"theembeddingsofthelocalpatchesfromthelastconvolu-",
"We want to find an optimal flow F that minimizes the",
"tional layers of each network. That is, in FI, face images are",
"following cost function, i.e.the sum of weighted pair-wise",
"aligned and cropped such that the entire face covers most of",
"distances across the two sets of facial features:",
"the image (seeFig.la).Therefore,withoutfacial occlusion,",
"N",
"N",
"every image patch is supposed to contain useful identity in-",
"difi",
"COST(Q,9,F)=",
"(1)",
"formation, which is in contrast to natural photos [66]",
"z-1j-1",
"Our grid sizes H × W for ArcFace, FaceNet, and Cos-",
"Face are respectively, 8×8, 3×3, and 6x7, which are the",
"s.t.",
"fii≥0",
"(2)",
"corresponding spatial dimensions of their last convolutional",
"N",
"layers (see definitions of these layers in Sec. Si). That is,",
"<",
"N",
"fij≤wgi,i,je[1, N] (3)",
"each feature qi is an embedding of size 1x1xC where C",
"is the number of channels (i.e. 512, 1792, and 512 for Arc-",
"fij≤Wgu,and",
"j=1",
"-1",
"NN",
"Face, FaceNet, and CosFace, respectively).",
"N",
"N",
"fij=min",
"Ground distance Like [66, 71], we use cosine distance as",
"",
">",
"Wa",
"(4)",
"04%",
"the ground distance dij between the embeddings (qi, gj) of",
"w",
"j=1 i=1",
"1",
"i=1",
"two patches:",
"(qi,9i)",
"dij = 1 -",
"(5)",
"Ilqill lg; ll",
"Asin[66.7ll.wenormalizetheweightsofafacesuch",
"that the total weights of features is 1 i.e. ",
"N",
"Wqi",
"",
"n=, wg, = 1, which is also the total flow in Eq. (4). Note",
"where () is the dot product between two feature vectors.",
"that EMD is a metric iff two distributions have an equal total",
"T",
"2.5.Featureweighting",
"weight and the ground distance function is a metric [16]",
"We use the iterative Sinkhorn algorithm [18] to effi-",
"EMD in our FI intuitively is an optimal plan to match",
"ciently solve the linear programming problem in Eq. (1),",
"all weighted features across two images. Therefore, how to",
"which yields the final EMD between two faces Q and g.",
"weight features is an important step. Here, we thoroughly",
"Facial features In image retrieval using EMD, a set of fea-",
"explore five different feature-weighting techniques for FI.",
"tures [qi] can be a collection of dominant colors [45], spa-",
"Uniform Zhang et al. [66] found that it is beneficial to",
"tial frequencies [45], or a histogram-like descriptor based",
"assign lower weight to less informative regions (e.g. back-",
"on the local patches of reference identities [60]. Inspired",
"ground or occlusion) and higher weight to discriminative"
] |
{
"arxiv_id": "2112.04016",
"language": "en",
"page_idx": 2
}
|
||
2112.04016-en-3
|
2112.04016-en-3.jpg
|
areas $(e.g.$ those containing salient objects). Yet, assigning an equal weight to all $N=H\times W$ patches is worth testing given that background noise is often cropped out of the pre-processed face image (Fig. 1):
$$w_{q_i}=w_{g_i}=\dfrac{1}{N},\text{where}1\leq k\leq N$$
(6)
Average Pooling Correlation (APC) Instead of uniformly weighting all patch embeddings, an alternative from [66] would be to weight a given feature $q_i$ proportional to its correlation to the entire other image in consideration. That is, the weight $w_{q_{i}}$ would be the dot product between the feature $q_i$ and the average pooling output of all embeddings $\{g_j\}_1^N$ of the gallery image:
$$w_{q_i}=\max\left(0,\langle q_i,\:\frac{\sum_j^Ng_j}{N}\rangle\right),w_{g_j}=\max\left(0,\langle g_j,\frac{\sum_i^Nq_i}{N}\rangle\right)$$
(7)
where $\max(.)$ keeps the weights always non-negative. APC tends to assign near-zero weight to occluded regions and, interestingly, also minimizes the weight of eyes and mouth in a non-occluded gallery image (see Fig. 3b; blue shades around both the mask and the non-occluded mouth). Cross Correlation (CC) APC[66] is different from CC introduced in [71], which is the same as APC except that CC uses the output vector from the last linear layer (see code) instead of the global average pooling vector in APC.
Spatial Correlation (SC) While both APC and CC “summarize" an entire other gallery image into a vector first, and then compute its correlation with a given patch $q_i$ in the query. In contrast, an alternative, inspired by [53], is to take the sum of the cosine similarity between the query patch $q_i$ and every patch in each gallery image $\{g_j\}_1^N{:}$
$$w_{q_i}=\max\big(0,\sum_j^N\frac{\langle q_i,g_j\rangle}{\|q_i\|\|g_j\|}\big),w_{g_j}=\max\big(0,\sum_i^N\frac{\langle q_i,g_j\rangle}{\|q_i\|\|g_j\|}\big)$$
We observe that SC often assigns a higher weight to oc-
cluded regions $e.g.$, masks and sunglasses (Fig. 3b).
Landmarking (LMK) While the previous three techniques adaptively rely on the image-patch similarity (APC, CC) or patch-wise similarity (SC) to weight a given patch embedding, their considered important points may or may not align with facial landmarks, which are known to be important for many face-related tasks. Here, as a baseline for APC, CC, and SC, we use dlib [26] to predict 68 keypoints in each face image (see Fig. 3c) and weight each patchembedding by the density of the keypoints inside the patch area. Our LMK weight distribution appears Gaussian-like with the peak often right below the nose (Fig. 3c).
(c)LMK
$(a)SC$
$(b)APC$
<FigureHere>
Figure 3. The results of assigning weights to 4×4 patches for ArcFace under three different techniques. Based on per-patch density of detected landmarks (---), LMK (c) often assigns higher weight to the center of a face (regardless of occlusions). In contrast, SC and APC assign higher weight to patches of higher patchwise and patch-vs-image similarity, respectively. APC tends to down-weight a facial feature $(e.g.$ blue regions around sunglasses or mouth) if its corresponding feature is occluded in the other image (b). In contrast, SC is insensitive to occlusions (a). See Fig. S1 & Fig. S2 for more heatmap examples of feature weighting.
## 3. Ablation Studies
We perform three ablation studies to rigorously evaluate the key design choices in our 2-stage FI approach: (1) Which feature-weighting techniques to use (Sec. 3.1)? (2) re-ranking using both EMD and cosine distance (Sec. 3.2); and (3) comparing patches or images in Stage 1 (Sec. 3.3). Experiment For all three experiments, we use ArcFace to perform FI on both LFW [65] and LFW-crop. For LFW, we take all 1,680 people who have $\geq2$ images for a total of 9,164 images. When taking each image as a query, we search in a gallery of the remaining 9,163 images. For the experiments with LFW-crop, we use all 13,233 original LFW images as the gallery. To create a query set of 13,233 cropped images, we clone the gallery and crop each image randomly to its 70% and upsample it back to the original size of 128×128 (see examples in Fig.5d). That is, LFW-crop tests identifying a cropped $(i.e.$ close-up, and misaligned) image given the unchanged LFW gallery. LFW and LFW-crop tests offer contrast insights (ID vs. OOD).
In Stage 2, i.e. re-ranking the top-$k$ candidates, we test different values of $k\in\{100,200,300\}$ and do not find the performance to change substantially. At $k=100$, our 2- stage precision is already close to the maximum precision of 99.88 under a perfect re-ranking (see Tab. la; Max prec.).
3.1. Comparing feature weighting techniques
Here, we evaluate the precision of our 2-stage FI as we sweep across five different feature-weighting techniques and two grid sizes (8×8 and 4×4). In an 8×8 grid, we observe that some facial features such as the eyes are often split in half across two patches (see Fig. S5), which may impair the patch-wise similarity. Therefore, for each weight-
|
[
[
0.53125,
0.08709716796875,
0.57421875,
0.08709716796875,
0.57421875,
0.10162353515625,
0.53125,
0.10162353515625
],
[
0.625,
0.08709716796875,
0.6806640625,
0.08709716796875,
0.6806640625,
0.10162353515625,
0.625,
0.10162353515625
],
[
0.765625,
0.08709716796875,
0.82275390625,
0.08709716796875,
0.82275390625,
0.10162353515625,
0.765625,
0.10162353515625
],
[
0.08087158203125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.11932373046875,
0.08087158203125,
0.11932373046875
],
[
0.07843017578125,
0.091552734375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.106689453125,
0.07843017578125,
0.1060791015625
],
[
0.07843017578125,
0.12249755859375,
0.4697265625,
0.121826171875,
0.4697265625,
0.1363525390625,
0.07843017578125,
0.136962890625
],
[
0.080078125,
0.1383056640625,
0.306396484375,
0.1383056640625,
0.306396484375,
0.15283203125,
0.080078125,
0.15283203125
],
[
0.15283203125,
0.175537109375,
0.2442626953125,
0.177978515625,
0.2442626953125,
0.1925048828125,
0.15283203125,
0.1900634765625
],
[
0.2484130859375,
0.17041015625,
0.259033203125,
0.1685791015625,
0.26318359375,
0.177978515625,
0.25244140625,
0.179931640625
],
[
0.26318359375,
0.175537109375,
0.393798828125,
0.175537109375,
0.393798828125,
0.189453125,
0.26318359375,
0.189453125
],
[
0.4453125,
0.175537109375,
0.4697265625,
0.175537109375,
0.4697265625,
0.189453125,
0.4453125,
0.189453125
],
[
0.2459716796875,
0.186279296875,
0.265625,
0.186279296875,
0.265625,
0.1944580078125,
0.2459716796875,
0.1944580078125
],
[
0.08087158203125,
0.2083740234375,
0.468017578125,
0.2083740234375,
0.468017578125,
0.2215576171875,
0.08087158203125,
0.2215576171875
],
[
0.080078125,
0.2235107421875,
0.468017578125,
0.2235107421875,
0.468017578125,
0.2374267578125,
0.080078125,
0.2374267578125
],
[
0.50244140625,
0.23291015625,
0.8896484375,
0.23291015625,
0.8896484375,
0.246826171875,
0.50244140625,
0.246826171875
],
[
0.080078125,
0.2392578125,
0.4697265625,
0.2392578125,
0.4697265625,
0.25390625,
0.080078125,
0.25390625
],
[
0.50244140625,
0.2474365234375,
0.890625,
0.2474365234375,
0.890625,
0.261474609375,
0.50244140625,
0.261474609375
],
[
0.08087158203125,
0.255615234375,
0.468017578125,
0.255615234375,
0.468017578125,
0.266357421875,
0.08087158203125,
0.266357421875
],
[
0.50244140625,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.50244140625,
0.274658203125
],
[
0.080078125,
0.26953125,
0.4697265625,
0.26953125,
0.4697265625,
0.28271484375,
0.080078125,
0.28271484375
],
[
0.50341796875,
0.276611328125,
0.890625,
0.276611328125,
0.890625,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.080078125,
0.2841796875,
0.468017578125,
0.2841796875,
0.468017578125,
0.29736328125,
0.080078125,
0.29736328125
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.08087158203125,
0.2998046875,
0.2646484375,
0.298583984375,
0.2646484375,
0.313232421875,
0.08087158203125,
0.314453125
],
[
0.5048828125,
0.304931640625,
0.890625,
0.304931640625,
0.890625,
0.31494140625,
0.5048828125,
0.31494140625
],
[
0.50244140625,
0.31640625,
0.8935546875,
0.31640625,
0.8935546875,
0.330078125,
0.50244140625,
0.330078125
],
[
0.2156982421875,
0.330078125,
0.275390625,
0.330078125,
0.275390625,
0.3447265625,
0.2156982421875,
0.3447265625
],
[
0.431396484375,
0.326904296875,
0.44677734375,
0.326904296875,
0.44677734375,
0.333984375,
0.431396484375,
0.333984375
],
[
0.441162109375,
0.330810546875,
0.4697265625,
0.333251953125,
0.468017578125,
0.343505859375,
0.439453125,
0.341552734375
],
[
0.5009765625,
0.330810546875,
0.890625,
0.330810546875,
0.890625,
0.3447265625,
0.5009765625,
0.3447265625
],
[
0.268798828125,
0.333984375,
0.41259765625,
0.3359375,
0.41259765625,
0.354248046875,
0.268798828125,
0.3515625
],
[
0.07843017578125,
0.3359375,
0.2115478515625,
0.3359375,
0.2115478515625,
0.354248046875,
0.07843017578125,
0.354248046875
],
[
0.426513671875,
0.3359375,
0.438720703125,
0.333984375,
0.439453125,
0.339111328125,
0.42724609375,
0.341552734375
],
[
0.2303466796875,
0.347900390625,
0.2459716796875,
0.347900390625,
0.2459716796875,
0.356689453125,
0.2303466796875,
0.356689453125
],
[
0.431396484375,
0.347900390625,
0.44677734375,
0.347900390625,
0.44677734375,
0.35791015625,
0.431396484375,
0.35791015625
],
[
0.5048828125,
0.3466796875,
0.8896484375,
0.3466796875,
0.8896484375,
0.356689453125,
0.5048828125,
0.356689453125
],
[
0.44677734375,
0.35791015625,
0.4697265625,
0.35791015625,
0.4697265625,
0.371826171875,
0.44677734375,
0.371826171875
],
[
0.50244140625,
0.359130859375,
0.859375,
0.359130859375,
0.859375,
0.37255859375,
0.50244140625,
0.37255859375
],
[
0.098876953125,
0.39013671875,
0.468017578125,
0.39013671875,
0.468017578125,
0.404052734375,
0.098876953125,
0.404052734375
],
[
0.08087158203125,
0.40478515625,
0.4697265625,
0.40478515625,
0.4697265625,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50244140625,
0.40087890625,
0.6611328125,
0.40087890625,
0.6611328125,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.434326171875,
0.080078125,
0.434326171875
],
[
0.52294921875,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.52294921875,
0.439453125
],
[
0.080078125,
0.43505859375,
0.4697265625,
0.43505859375,
0.4697265625,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.5009765625,
0.43994140625,
0.89208984375,
0.44140625,
0.89208984375,
0.455810546875,
0.5009765625,
0.455078125
],
[
0.08087158203125,
0.451904296875,
0.465576171875,
0.451904296875,
0.465576171875,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.5009765625,
0.455078125,
0.890625,
0.455810546875,
0.890625,
0.47021484375,
0.5009765625,
0.4697265625
],
[
0.08087158203125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.08087158203125,
0.479736328125
],
[
0.5,
0.4716796875,
0.89208984375,
0.47021484375,
0.89208984375,
0.48486328125,
0.5,
0.486083984375
],
[
0.080078125,
0.481689453125,
0.468017578125,
0.481689453125,
0.468017578125,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.50244140625,
0.4873046875,
0.88623046875,
0.4873046875,
0.88623046875,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.080078125,
0.49755859375,
0.4697265625,
0.49755859375,
0.4697265625,
0.51123046875,
0.080078125,
0.51123046875
],
[
0.50244140625,
0.50244140625,
0.89208984375,
0.50244140625,
0.89208984375,
0.5166015625,
0.50244140625,
0.5166015625
],
[
0.07843017578125,
0.51220703125,
0.423095703125,
0.51123046875,
0.423095703125,
0.52587890625,
0.07843017578125,
0.5263671875
],
[
0.50244140625,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.07843017578125,
0.52783203125,
0.46728515625,
0.5263671875,
0.46728515625,
0.54150390625,
0.07843017578125,
0.54248046875
],
[
0.50341796875,
0.53271484375,
0.89208984375,
0.53271484375,
0.89208984375,
0.546875,
0.50341796875,
0.546875
],
[
0.08087158203125,
0.54541015625,
0.46728515625,
0.54541015625,
0.46728515625,
0.55615234375,
0.08087158203125,
0.55615234375
],
[
0.5009765625,
0.546875,
0.89208984375,
0.54736328125,
0.89208984375,
0.5625,
0.5009765625,
0.56103515625
],
[
0.07843017578125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.5712890625,
0.07843017578125,
0.5712890625
],
[
0.50244140625,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50244140625,
0.5771484375
],
[
0.07843017578125,
0.57470703125,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.07843017578125,
0.58837890625
],
[
0.50244140625,
0.57763671875,
0.8896484375,
0.57763671875,
0.8896484375,
0.59130859375,
0.50244140625,
0.59130859375
],
[
0.07843017578125,
0.58837890625,
0.468017578125,
0.5888671875,
0.468017578125,
0.60400390625,
0.07843017578125,
0.60302734375
],
[
0.5009765625,
0.59130859375,
0.89208984375,
0.59228515625,
0.89208984375,
0.60693359375,
0.5009765625,
0.60595703125
],
[
0.07843017578125,
0.60498046875,
0.3798828125,
0.60498046875,
0.3798828125,
0.619140625,
0.07843017578125,
0.619140625
],
[
0.50341796875,
0.60791015625,
0.89208984375,
0.60791015625,
0.89208984375,
0.6220703125,
0.50341796875,
0.6220703125
],
[
0.50244140625,
0.62353515625,
0.89208984375,
0.62353515625,
0.89208984375,
0.63720703125,
0.50244140625,
0.63720703125
],
[
0.1953125,
0.63330078125,
0.27294921875,
0.63623046875,
0.271240234375,
0.6513671875,
0.193603515625,
0.6484375
],
[
0.39453125,
0.6298828125,
0.410888671875,
0.6298828125,
0.410888671875,
0.6396484375,
0.39453125,
0.6396484375
],
[
0.406982421875,
0.63525390625,
0.4697265625,
0.63525390625,
0.4697265625,
0.6494140625,
0.406982421875,
0.6494140625
],
[
0.080078125,
0.6416015625,
0.200927734375,
0.6416015625,
0.200927734375,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.275390625,
0.6416015625,
0.406982421875,
0.6416015625,
0.406982421875,
0.6591796875,
0.275390625,
0.6591796875
],
[
0.50244140625,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.2017822265625,
0.65283203125,
0.271240234375,
0.65283203125,
0.271240234375,
0.66748046875,
0.2017822265625,
0.66748046875
],
[
0.41259765625,
0.65283203125,
0.476318359375,
0.65283203125,
0.476318359375,
0.66748046875,
0.41259765625,
0.66748046875
],
[
0.50244140625,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.44677734375,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.44677734375,
0.68603515625
],
[
0.50244140625,
0.66845703125,
0.890625,
0.66845703125,
0.890625,
0.681640625,
0.50244140625,
0.681640625
],
[
0.50244140625,
0.6845703125,
0.8759765625,
0.6845703125,
0.8759765625,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.1004638671875,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.1004638671875,
0.71728515625
],
[
0.521484375,
0.69873046875,
0.89208984375,
0.69873046875,
0.89208984375,
0.7119140625,
0.521484375,
0.7119140625
],
[
0.08087158203125,
0.71826171875,
0.416748046875,
0.71826171875,
0.416748046875,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.50244140625,
0.7138671875,
0.89208984375,
0.7138671875,
0.89208984375,
0.72802734375,
0.50244140625,
0.72802734375
],
[
0.080078125,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.5009765625,
0.72900390625,
0.89208984375,
0.72998046875,
0.89208984375,
0.744140625,
0.5009765625,
0.74365234375
],
[
0.5009765625,
0.744140625,
0.89208984375,
0.744140625,
0.89208984375,
0.75830078125,
0.5009765625,
0.75830078125
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.08087158203125,
0.76513671875
],
[
0.5009765625,
0.7587890625,
0.890625,
0.76025390625,
0.890625,
0.7744140625,
0.5009765625,
0.7734375
],
[
0.08087158203125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.080078125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50341796875,
0.787109375,
0.8486328125,
0.787109375,
0.8486328125,
0.8017578125,
0.50341796875,
0.8017578125
],
[
0.080078125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.07843017578125,
0.81103515625,
0.4697265625,
0.81005859375,
0.4697265625,
0.82373046875,
0.07843017578125,
0.82470703125
],
[
0.51953125,
0.8095703125,
0.890625,
0.81005859375,
0.890625,
0.82470703125,
0.51953125,
0.82373046875
],
[
0.07843017578125,
0.82470703125,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.07843017578125,
0.83837890625
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
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
],
[
0.080078125,
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
],
[
0.07843017578125,
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,
0.408447265625,
0.88720703125,
0.408447265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"(a) SC",
"(b) APC",
"(c) LMK",
"an equal weight to all N = H x Wpatches is worth test-",
"areas (e.g. those containing salient objects). Yet, assigning",
"ing given that background noise is often cropped out of the",
"pre-processed face image (Fig. 1):",
"Wqi=Wgi=",
"1",
",where1<k<N",
"(6)",
"N",
"AveragePoolingCorrelation(APC)Instead of uniformly",
"weighting all patch embeddings, an alternative from [66]",
"Figure 3. The results of assigning weights to 4x4 patches for Ar",
"would be to weight a given feature qi proportional to its",
"cFace under three different techniques. Based on per-patch den-",
"correlation to theentireother image in consideration.That",
"sity of detected landmarks (- - -), LMK (c) often assigns higher",
"is, the weight wg, would be the dot product between the",
"weight to the center of a face (regardless of occlusions).In con-",
"feature qi and the average pooling output of all embeddings",
"trast, SC and APC assign higher weight to patches of higher patch-",
"(gil of thegallery image:",
"wise and patch-vs-imagesimilarity,respectively.APC tends to",
"down-weight a facial feature (e.g. blue regions around sunglasses",
"Zj gi)",
"N",
"qi",
"or mouth) if its corresponding feature is occluded in the other im-",
"), wg, = max (0, (gj,",
"Wai = max (0, (qi,",
"",
"N",
"N",
"age (b).In contrast,SCisinsensitiveto occlusions (a).See Fig.Sl",
"(7)",
"& Fig. S2 for more heatmap examples of feature weighting.",
"where max(.) keeps the weights always non-negative.",
"APC tends to assign near-zero weight to occluded regions",
"3.AblationStudies",
"and, interestingly, also minimizes the weight of eyes and",
"We perform three ablation studies to rigorously evalu-",
"mouth in a non-occluded gallery image (see Fig. 3b; blue",
"ate the key design choices in our 2-stage FI approach: (1)",
"shadesaroundboththemaskandthenon-occludedmouth)",
"Which feature-weighting techniques to use (Sec. 3.1)? (2)",
"Cross Correlation (CC)APC [66] is different from CC",
"re-ranking using both EMD and cosine distance (Sec. 3.2);",
"introduced in [71], which is the same as APC except that CC",
"and (3) comparing patches or images in Stage 1 (Sec. 3.3).",
"uses the output vector from the last linear layer (see code)",
"Experiment For all three experiments, we use ArcFace to",
"instead of the global average pooling vector in APC.",
"perform FI on both LFW [65] and LFW-crop. For LFW,",
"Spatial Correlation (SC) While both APC and CC \"sum-",
"we take all 1,680 people who have ≥2 images for a to-",
"marizean entire othergallery image intoa vectorfirst, and",
"tal of 9,164 images. When taking each image as a query,",
"then compute its correlation with a given patch q; in the",
"we search in a gallery of the remaining 9,163 images. For",
"query. In contrast, an alternative, inspired by [53], is to take",
"the experiments with LFW-crop, we use all 13,233 orig",
"the sum of the cosine similarity between the query patch qi",
"inal LFW images as the gallery. To create a query set of",
"and every patch in each gallery image (g,JN :",
"13,233 cropped images, we clone the gallery and crop each",
"image randomly to its 70% and upsample it back to the",
"(qi,gi)",
"N",
"(qi,9j)",
"Wai = max (o, ",
"), wg, = max (o, ",
"original size of 128x128 (see examples in Fig. 5d). That",
"llqalllgill",
"lqilllgill",
"is, LFW-crop tests identifying a cropped (i.e. close-up, and",
"(8)",
"misaligned) image given the unchanged LFW gallery.LFW",
"and LFW-crop tests offer contrast insights (ID vs. OOD)",
"We observe that SC often assigns a higher weight to oc-",
"In Stage 2, i.e. re-ranking the top-k candidates, we test",
"cluded regions e.g., masks and sunglasses (Fig. 3b)",
"different values of k E {100, 200, 300] and do not find the",
"Landmarking (LMK) While the previous three tech-",
"performance to change substantially. At k = 100, our 2-",
"stage precision is already close to the maximum precision",
"niques adaptively rely on the image-patch similarity (APC,",
"of 99.88 under a perfect re-ranking (see Tab. 1a; Max prec.).",
"CC) or patch-wise similarity (SC) to weight a given patch",
"embedding, their considered important points may or may",
"3.1. Comparing feature weighting techniques",
"not align with facial landmarks, which are known to be im-",
"portantfor manyface-relatedtasks.Here,as abaselinefor",
"Here, we evaluate the precision of our 2-stage FI as",
"APC, CC, and SC, we use dlib [26] to predict 68 keypoints",
"we sweep across five different feature-weighting techniques",
"in each face image (see Fig. 3c) and weight each patch-",
"and two grid sizes (8×8 and 4x4). In an 8×8 grid, we ob-",
"embedding by the density of the keypoints inside the patch",
"serve that some facial features such as the eyes are often",
"area. Our LMK weight distribution appears Gaussian-like",
"split in half across two patches (see Fig. S5), which may im-",
"with the peak often right below the nose (Fig. 3c).",
"pair the patch-wise similarity. Therefore, for each weight-"
] |
{
"arxiv_id": "2112.04016",
"language": "en",
"page_idx": 3
}
|
||
2006.14748-en-0
|
2006.14748-en-0.jpg
|
Proper Network Interpretability Helps Adversarial Robustness in Classification
Akhilan Boopathy$^{1}$ Sijia Liu$^{2}$ Gaoyuan Zhang$^{2}$ Cynthia Liu$^{1}$ Pin-Yu Chen$^{2}$ Shiyu Chang$^{2}$ Luca Daniel$^{1}$
$\mathbf{Abstract}$
defensive distillation (Papernot et al., 2016b), training with adversarial examples (Goodfellow et al., 2015), input gradient or curvature regularization (Ross&Doshi-Velez, 2018; Moosavi-Dezfooli et al., 2019), adversarial training via robust optimization (Madry et al., 2018), and TRADES to trade adversarial robustness off against accuracy (Zhang et al., 2019). Different from the aforementioned works, this paper attempts to understand the adversarial robustness of CNNs from the network interpretability perspective, and provides novel insights on when and how interpretability could help robust classification.
Recent works have empirically shown that there exist adversarial examples that can be hidden from neural network interpretability (namely, making network interpretation maps visually similar), or interpretability is itself susceptible to adversarial attacks. In this paper, we theoretically show that with a proper measurement of interpretation, it is actually difficult to prevent prediction-evasion adversarial attacks from causing interpretation discrepancy, as confirmed by experiments on MNIST, CIFAR-10 and Restricted ImageNet. Spurred by that, we develop an interpretability-aware defensive scheme built only on promoting robust interpretation (without the need for resorting to adversarial loss minimization). We show that our defense achieves both robust classification and robust interpretation, outperforming state-of-theart adversarial training methods against attacks of large perturbation in particular.
Having a prediction might not be enough for many realworld machine learning applications. It is crucial to demystify why they make certain decisions. Thus, the problem of network interpretation arises. Various methods have been proposed to understand the mechanism of decision making by CNNs. One category of methods justify a prediction decision by assigning importance values to reflect the influence of individual pixels or image sub-regions on the final classification. Examples include pixel-space sensitivity map methods (Simonyan et al., 2013; Zeiler & Fergus, 2014; Springenberg et al., 2014; Smilkov et al., 2017; Sundararajan et al., 2017) and class-discriminative localization methods (Zhou et al., 2016; Selvaraju et al., 2017; Chattopadhay et al., 2018; Petsiuk et al., 2018), where the former evaluates the sensitivity of a network classification decision to pixel variations at the input, and the latter localizes which parts of an input image were looked at by the network for making a classification decision. We refer readers to Sec.2 for some representative interpretation methods. Besides interpreting CNNs via feature importance maps, some methods peek into the internal response of neural networks. Examples include network dissection (Bau et al.,2017), and learning perceptually-aligned representations from adversarial training (Engstrom et al., 2019).
Some recent works (Xu et al., 2019b;a; Zhang et al., 2018; Subramanya et al., 2018; Ghorbani et al., 2019; Dombrowski et al., 2019; Chen et al., 2019) began to study adversarial robustness by exploring the spectrum between classification accuracy and network interpretability. It was shown in (Xu et al., 2019b;a) that an imperceptible adversarial perturbation to fool classifers can lead to a significant change in a class-specific network interpretation map. Thus, it was
## $1. \textbf{Introduction}$
It has become widely known that convolutional neural networks (CNNs) are vulnerable to adversarial examples, namely, perturbed inputs with the intention to mislead networks' prediction (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot et al., 2016a; Carlini & Wagner, 2017; Chen et al., 2018; Su et al., 2018). The vulnerability of CNNs has spurred extensive research on adversarial attack and defense. To design adversarial attacks, most works have focused on creating either imperceptible input perturbations (Goodfellow et al., 2015; Papernot et al., 2016a; Carlini & Wagner, 2017; Chen et al., 2018) or adversarial patches robust to the physical environment (Eykholt et al., 2018; Brown et al., 2017; Athalye et al., 2017). Many defense methods have also been developed to prevent CNNs from misclassification when facing adversarial attacks. Examples include
argued that such an interpretation discrepancy can be used
$^{1}{\mathrm{Massachusetts~Institute~of~Technology~2MIT-IBM~Watson~AI}}$ Lab, IBM Research. Correspondence to: Akhilan Boopathy, Sijia Liu <akhilan@ mit.edu, lsjxjtu@ gmail.com>.
Proceedings of the 37$^{th}$ International Conference on Machine $Learning$,Vienna, Austria, PMLR 119, 2020. Copyright 2020 by the author(s).
|
[
[
0.09234619140625,
0.1142578125,
0.88232421875,
0.1142578125,
0.88232421875,
0.1287841796875,
0.09234619140625,
0.1287841796875
],
[
0.1004638671875,
0.177978515625,
0.87158203125,
0.177978515625,
0.87158203125,
0.1925048828125,
0.1004638671875,
0.1925048828125
],
[
0.2442626953125,
0.2203369140625,
0.3203125,
0.2203369140625,
0.3203125,
0.23486328125,
0.2442626953125,
0.23486328125
],
[
0.5,
0.2203369140625,
0.8857421875,
0.2215576171875,
0.8857421875,
0.236083984375,
0.5,
0.23486328125
],
[
0.1217041015625,
0.2412109375,
0.441162109375,
0.2412109375,
0.441162109375,
0.255126953125,
0.1217041015625,
0.255126953125
],
[
0.498291015625,
0.238037109375,
0.8857421875,
0.238037109375,
0.8857421875,
0.251953125,
0.498291015625,
0.251953125
],
[
0.1217041015625,
0.2568359375,
0.441162109375,
0.2568359375,
0.441162109375,
0.270751953125,
0.1217041015625,
0.270751953125
],
[
0.5009765625,
0.25244140625,
0.88623046875,
0.25244140625,
0.88623046875,
0.266357421875,
0.5009765625,
0.266357421875
],
[
0.0236968994140625,
0.267578125,
0.058013916015625,
0.267578125,
0.058013916015625,
0.37255859375,
0.0236968994140625,
0.37255859375
],
[
0.12091064453125,
0.272705078125,
0.44287109375,
0.272705078125,
0.44287109375,
0.285888671875,
0.12091064453125,
0.285888671875
],
[
0.5009765625,
0.268310546875,
0.8857421875,
0.268310546875,
0.8857421875,
0.2822265625,
0.5009765625,
0.2822265625
],
[
0.1217041015625,
0.287353515625,
0.44287109375,
0.287353515625,
0.44287109375,
0.300537109375,
0.1217041015625,
0.300537109375
],
[
0.5009765625,
0.28271484375,
0.88623046875,
0.28271484375,
0.88623046875,
0.296630859375,
0.5009765625,
0.296630859375
],
[
0.1217041015625,
0.3017578125,
0.441162109375,
0.3017578125,
0.441162109375,
0.31494140625,
0.1217041015625,
0.31494140625
],
[
0.5009765625,
0.298583984375,
0.8857421875,
0.298583984375,
0.8857421875,
0.3125,
0.5009765625,
0.3125
],
[
0.1217041015625,
0.317626953125,
0.441162109375,
0.317626953125,
0.441162109375,
0.330810546875,
0.1217041015625,
0.330810546875
],
[
0.5009765625,
0.313232421875,
0.8857421875,
0.313232421875,
0.8857421875,
0.326904296875,
0.5009765625,
0.326904296875
],
[
0.12091064453125,
0.333251953125,
0.441162109375,
0.333251953125,
0.441162109375,
0.3466796875,
0.12091064453125,
0.3466796875
],
[
0.50244140625,
0.328857421875,
0.88623046875,
0.328857421875,
0.88623046875,
0.34228515625,
0.50244140625,
0.34228515625
],
[
0.1192626953125,
0.3466796875,
0.441162109375,
0.3466796875,
0.441162109375,
0.35986328125,
0.1192626953125,
0.35986328125
],
[
0.5009765625,
0.343505859375,
0.8857421875,
0.343505859375,
0.8857421875,
0.356689453125,
0.5009765625,
0.356689453125
],
[
0.12091064453125,
0.363037109375,
0.44287109375,
0.363037109375,
0.44287109375,
0.376953125,
0.12091064453125,
0.376953125
],
[
0.5009765625,
0.359130859375,
0.8857421875,
0.359130859375,
0.8857421875,
0.37255859375,
0.5009765625,
0.37255859375
],
[
0.0269622802734375,
0.3662109375,
0.0555419921875,
0.3662109375,
0.0555419921875,
0.395751953125,
0.0269622802734375,
0.395751953125
],
[
0.037567138671875,
0.367431640625,
0.045745849609375,
0.367431640625,
0.045745849609375,
0.37060546875,
0.037567138671875,
0.37060546875
],
[
0.1217041015625,
0.37744140625,
0.44287109375,
0.37744140625,
0.44287109375,
0.391357421875,
0.1217041015625,
0.391357421875
],
[
0.5009765625,
0.375,
0.7099609375,
0.375,
0.7099609375,
0.38818359375,
0.5009765625,
0.38818359375
],
[
0.1217041015625,
0.392578125,
0.441162109375,
0.392578125,
0.441162109375,
0.406005859375,
0.1217041015625,
0.406005859375
],
[
0.5,
0.395751953125,
0.88623046875,
0.395751953125,
0.88623046875,
0.410400390625,
0.5,
0.410400390625
],
[
0.0196075439453125,
0.404052734375,
0.05963134765625,
0.40283203125,
0.062103271484375,
0.488037109375,
0.0228729248046875,
0.4892578125
],
[
0.1217041015625,
0.40771484375,
0.441162109375,
0.40771484375,
0.441162109375,
0.421630859375,
0.1217041015625,
0.421630859375
],
[
0.5,
0.410888671875,
0.88623046875,
0.410400390625,
0.88623046875,
0.4248046875,
0.5,
0.425537109375
],
[
0.1217041015625,
0.4248046875,
0.441162109375,
0.4248046875,
0.441162109375,
0.43505859375,
0.1217041015625,
0.43505859375
],
[
0.5,
0.4267578125,
0.8857421875,
0.4267578125,
0.8857421875,
0.43994140625,
0.5,
0.43994140625
],
[
0.1192626953125,
0.438232421875,
0.44287109375,
0.4375,
0.44287109375,
0.45068359375,
0.1192626953125,
0.451904296875
],
[
0.5009765625,
0.44140625,
0.8857421875,
0.44140625,
0.8857421875,
0.455078125,
0.5009765625,
0.455078125
],
[
0.12091064453125,
0.45263671875,
0.441162109375,
0.45263671875,
0.441162109375,
0.466552734375,
0.12091064453125,
0.466552734375
],
[
0.5009765625,
0.45703125,
0.8857421875,
0.45703125,
0.8857421875,
0.47021484375,
0.5009765625,
0.47021484375
],
[
0.1217041015625,
0.4697265625,
0.439453125,
0.4697265625,
0.439453125,
0.479736328125,
0.1217041015625,
0.479736328125
],
[
0.5,
0.47021484375,
0.88623046875,
0.4716796875,
0.88623046875,
0.486083984375,
0.5,
0.48486328125
],
[
0.1217041015625,
0.48486328125,
0.438720703125,
0.48486328125,
0.438720703125,
0.49560546875,
0.1217041015625,
0.49560546875
],
[
0.5009765625,
0.4873046875,
0.88623046875,
0.4873046875,
0.88623046875,
0.50048828125,
0.5009765625,
0.50048828125
],
[
0.1217041015625,
0.498779296875,
0.443603515625,
0.498779296875,
0.443603515625,
0.51220703125,
0.1217041015625,
0.51220703125
],
[
0.0228729248046875,
0.4892578125,
0.05963134765625,
0.4892578125,
0.05963134765625,
0.7109375,
0.0228729248046875,
0.7109375
],
[
0.50244140625,
0.50390625,
0.8837890625,
0.50390625,
0.8837890625,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.12091064453125,
0.5146484375,
0.326904296875,
0.5146484375,
0.326904296875,
0.52783203125,
0.12091064453125,
0.52783203125
],
[
0.5009765625,
0.517578125,
0.8857421875,
0.517578125,
0.8857421875,
0.53076171875,
0.5009765625,
0.53076171875
],
[
0.50244140625,
0.5341796875,
0.8857421875,
0.5341796875,
0.8857421875,
0.54443359375,
0.50244140625,
0.54443359375
],
[
0.5009765625,
0.54736328125,
0.88623046875,
0.54736328125,
0.88623046875,
0.56103515625,
0.5009765625,
0.56103515625
],
[
0.0906982421875,
0.5546875,
0.2156982421875,
0.5546875,
0.2156982421875,
0.5693359375,
0.0906982421875,
0.5693359375
],
[
0.5,
0.5625,
0.88623046875,
0.5625,
0.88623046875,
0.57568359375,
0.5,
0.57568359375
],
[
0.08905029296875,
0.58056640625,
0.4755859375,
0.58056640625,
0.4755859375,
0.5947265625,
0.08905029296875,
0.5947265625
],
[
0.5,
0.5771484375,
0.8857421875,
0.5771484375,
0.8857421875,
0.59033203125,
0.5,
0.59033203125
],
[
0.08905029296875,
0.5966796875,
0.476318359375,
0.5966796875,
0.476318359375,
0.60986328125,
0.08905029296875,
0.60986328125
],
[
0.5,
0.59228515625,
0.8857421875,
0.59228515625,
0.8857421875,
0.60595703125,
0.5,
0.60595703125
],
[
0.0882568359375,
0.6123046875,
0.4755859375,
0.6123046875,
0.4755859375,
0.62548828125,
0.0882568359375,
0.62548828125
],
[
0.50244140625,
0.609375,
0.8837890625,
0.609375,
0.8837890625,
0.619140625,
0.50244140625,
0.619140625
],
[
0.0882568359375,
0.626953125,
0.4755859375,
0.626953125,
0.4755859375,
0.64013671875,
0.0882568359375,
0.64013671875
],
[
0.5,
0.62353515625,
0.88623046875,
0.62353515625,
0.88623046875,
0.63720703125,
0.5,
0.63720703125
],
[
0.08905029296875,
0.64013671875,
0.473876953125,
0.64013671875,
0.473876953125,
0.65380859375,
0.08905029296875,
0.65380859375
],
[
0.5,
0.63818359375,
0.88623046875,
0.63818359375,
0.88623046875,
0.6513671875,
0.5,
0.6513671875
],
[
0.08660888671875,
0.65576171875,
0.4755859375,
0.65478515625,
0.4755859375,
0.66943359375,
0.08660888671875,
0.67041015625
],
[
0.5,
0.65283203125,
0.88623046875,
0.65283203125,
0.88623046875,
0.66650390625,
0.5,
0.66650390625
],
[
0.0882568359375,
0.671875,
0.4755859375,
0.671875,
0.4755859375,
0.68505859375,
0.0882568359375,
0.68505859375
],
[
0.5,
0.66845703125,
0.8857421875,
0.66845703125,
0.8857421875,
0.681640625,
0.5,
0.681640625
],
[
0.035125732421875,
0.67529296875,
0.0400390625,
0.67529296875,
0.0400390625,
0.6845703125,
0.035125732421875,
0.6845703125
],
[
0.08905029296875,
0.68603515625,
0.473876953125,
0.68603515625,
0.473876953125,
0.69970703125,
0.08905029296875,
0.69970703125
],
[
0.498291015625,
0.681640625,
0.8857421875,
0.68310546875,
0.8857421875,
0.69775390625,
0.498291015625,
0.6962890625
],
[
0.0906982421875,
0.70068359375,
0.47216796875,
0.70068359375,
0.47216796875,
0.7138671875,
0.0906982421875,
0.7138671875
],
[
0.5,
0.69873046875,
0.8857421875,
0.69873046875,
0.8857421875,
0.7119140625,
0.5,
0.7119140625
],
[
0.08905029296875,
0.71728515625,
0.473876953125,
0.71728515625,
0.473876953125,
0.73095703125,
0.08905029296875,
0.73095703125
],
[
0.5009765625,
0.71533203125,
0.8837890625,
0.71533203125,
0.8837890625,
0.72607421875,
0.5009765625,
0.72607421875
],
[
0.0906982421875,
0.73193359375,
0.473876953125,
0.73193359375,
0.473876953125,
0.74560546875,
0.0906982421875,
0.74560546875
],
[
0.5,
0.72900390625,
0.88623046875,
0.72900390625,
0.88623046875,
0.7421875,
0.5,
0.7421875
],
[
0.0906982421875,
0.74755859375,
0.47802734375,
0.74755859375,
0.47802734375,
0.76123046875,
0.0906982421875,
0.76123046875
],
[
0.5,
0.74365234375,
0.71142578125,
0.74365234375,
0.71142578125,
0.7568359375,
0.5,
0.7568359375
],
[
0.08905029296875,
0.76220703125,
0.4755859375,
0.76220703125,
0.4755859375,
0.77587890625,
0.08905029296875,
0.77587890625
],
[
0.5009765625,
0.76513671875,
0.8857421875,
0.76513671875,
0.8857421875,
0.77880859375,
0.5009765625,
0.77880859375
],
[
0.08905029296875,
0.7763671875,
0.4755859375,
0.7763671875,
0.4755859375,
0.79052734375,
0.08905029296875,
0.79052734375
],
[
0.5,
0.78076171875,
0.88623046875,
0.78076171875,
0.88623046875,
0.79443359375,
0.5,
0.79443359375
],
[
0.0882568359375,
0.79248046875,
0.4755859375,
0.79248046875,
0.4755859375,
0.806640625,
0.0882568359375,
0.806640625
],
[
0.5,
0.796875,
0.88623046875,
0.796875,
0.88623046875,
0.81005859375,
0.5,
0.81005859375
],
[
0.1070556640625,
0.814453125,
0.4755859375,
0.81640625,
0.4755859375,
0.83056640625,
0.1070556640625,
0.8291015625
],
[
0.498291015625,
0.8125,
0.8857421875,
0.8125,
0.8857421875,
0.82568359375,
0.498291015625,
0.82568359375
],
[
0.0882568359375,
0.82763671875,
0.473876953125,
0.830078125,
0.473876953125,
0.84326171875,
0.0882568359375,
0.84130859375
],
[
0.5009765625,
0.8271484375,
0.8857421875,
0.8271484375,
0.8857421875,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.0882568359375,
0.84228515625,
0.36279296875,
0.84228515625,
0.36279296875,
0.85595703125,
0.0882568359375,
0.85595703125
],
[
0.498291015625,
0.84033203125,
0.88623046875,
0.84130859375,
0.88623046875,
0.85595703125,
0.498291015625,
0.85498046875
],
[
0.5,
0.85693359375,
0.88623046875,
0.85693359375,
0.88623046875,
0.87060546875,
0.5,
0.87060546875
],
[
0.08660888671875,
0.8642578125,
0.4755859375,
0.8642578125,
0.4755859375,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.5,
0.87255859375,
0.8857421875,
0.87255859375,
0.8857421875,
0.88623046875,
0.5,
0.88623046875
],
[
0.0882568359375,
0.8798828125,
0.473876953125,
0.8798828125,
0.473876953125,
0.89306640625,
0.0882568359375,
0.89306640625
],
[
0.08660888671875,
0.89013671875,
0.1707763671875,
0.892578125,
0.1707763671875,
0.9072265625,
0.08660888671875,
0.90478515625
],
[
0.5,
0.88818359375,
0.88623046875,
0.88818359375,
0.88623046875,
0.9013671875,
0.5,
0.9013671875
]
] |
[
"Proper Network Interpretability Helps Adversarial Robustness in Classification",
"Akhilan Boopathy! Sijia Liu? Gaoyuan Zhang? Cynthia Liu1 Pin-Yu Chen? Shiyu Chang? Luca Daniel!",
"Abstract",
"defensive distillation (Papernot et al., 2016b), training with",
"Recent works have empirically shown that there",
"adversarial examples (Goodfellow et al., 2015), input gradi-",
"exist adversarial examples that can be hidden from",
"ent or curvature regularization (Ross & Doshi-Velez, 2018;",
"2010",
"neural network interpretability (namely, making",
"Moosavi-Dezfooli et al., 2019), adversarial training via ro",
"network interpretation maps visually similar), or",
"bust optimization (Madry et al., 2018), and TRADES to",
"interpretability is itself susceptible to adversarial",
"trade adversarial robustness off against accuracy (Zhang",
"attacks. In this paper, we theoretically show that",
"et al., 2019).Different from the aforementioned works, this",
"with a proper measurement of interpretation, it",
"paper attempts to understand the adversarial robustness of",
"is actually difficult to prevent prediction-evasion",
"CNNs from the network interpretability perspective, and",
"adversarial attacks from causing interpretation dis-",
"provides novel insights on when and how interpretability",
"2",
"",
"crepancy, as confirmed by experiments on MNIST,",
"could help robust classification.",
"CIFAR-10 and Restricted ImageNet. Spurred by",
"Having a prediction might not be enough for many real-",
"[T'S]",
"that, we develop an interpretability-aware defen-",
"world machine learning applications. It is crucial to demys-",
"sivescheme builtonlyonpromotingrobustin-",
"tify why they make certain decisions. Thus, the problem of",
"terpretation (without the need for resorting to ad-",
"network interpretation arises. Various methods have been",
"versarial loss minimization). We show that our",
"proposed to understand the mechanism of decision making",
"defense achieves bothrobust classificationand",
"by CNNs. One category of methods justify a prediction",
"robustinterpretation,outperforming state-of-the",
"decision by assigning importance values to reflect the influ-",
"art adversarial training methods against attacks of",
"8002",
"ence of individual pixels or image sub-regions on the final",
"large perturbation in particular.",
"classification. Examples include pixel-space sensitivity map",
"methods(Simonyanetal..2013:Zeiler&Fergus,2014:",
"Springenberg et al., 2014; Smilkov et al., 2017; Sundarara-",
"1.Introduction",
"jan et al., 2017) and class-discriminative localization meth-",
"It has become widely known that convolutional neural",
"ods (Zhou et al., 2016; Selvaraju et al., 2017; Chattopadhay",
"networks (CNNs) are vulnerable to adversarial examples,",
"et al., 2018; Petsiuk et al., 2018), where the former evaluates",
"namely, perturbed inputs with the intention to mislead net-",
"thesensitivityofanetworkclassificationdecisiontopixel",
"works'prediction (Szegedy et al., 2014; Goodfellow et al.,",
"variations at the input, and the latter localizes which parts of",
"2015; Papernot et al.,2016a; Carlini & Wagner,2017: Chen",
"an input image were looked at by the network for making a",
"et al., 2018; Su et al., 2018). The vulnerability of CNNs has",
"classification decision. We refer readers to Sec.2 for some",
"spurred extensive research on adversarial attack and defense.",
"representative interpretation methods. Besides interpreting",
"",
"To design adversarial attacks, most works have focused on",
"CNNs via feature importance maps, some methods peek",
"creating either imperceptible input perturbations (Goodfel",
"into the internal response of neural networks. Examples",
"low et al.,2015; Papernot et al., 2016a; Carlini & Wagner,",
"includenetwork dissection(Bau etal.,2017),andlearn",
"2017; Chen et al., 2018) or adversarial patches robust to the",
"ing perceptually-aligned representations from adversarial",
"physical environment (Eykholt et al., 2018; Brown et al.,",
"training (Engstrom et al., 2019).",
"2017; Athalye et al., 2017). Many defense methods have",
"Some recent works (Xu et al., 2019b;a; Zhang et al., 2018:",
"also been developed to prevent CNNs from misclassifica-",
"Subramanya et al., 2018: Ghorbani et al., 2019; Dombrowski",
"tion when facing adversarial attacks. Examples include",
"et al., 2019; Chen et al., 2019) began to study adversarial",
"'Massachusetts Institute of Technology °MIT-IBM Watson AI",
"robustness by exploring the spectrum between classification",
"Lab,IBM Research. Correspondence to: Akhilan Boopathy, Sijia",
"accuracy and network interpretability. It was shown in (Xu",
"Liu <akhilan@mit.edu, Isjxjtu@ gmail.com>.",
"et al., 2019b;a) that an imperceptible adversarial perturba-",
"tion to fool classifiers can lead to a significant change in",
"Proceedings of the 37th International Conference on Machine",
"a class-specific network interpretation map. Thus, it was",
"Learning, Vienna, Austria, PMLR 119, 2020. Copyright 2020 by",
"the author(s).",
"argued that such an interpretation discrepancy can be used"
] |
{
"arxiv_id": "2006.14748",
"language": "en",
"page_idx": 0
}
|
||
2006.14748-en-1
|
2006.14748-en-1.jpg
|
as a helpful metric to differentiate adversarial examples from benign inputs. Nevertheless, the work (Zhang et al., 2018; Subramanya et al., 2018) showed that under certain conditions, generating an attack (which we call an interpretability sneaking attack, ISA) that fools the classifier while keeping it stealthy from the coupled interpreter is not significantly more difficult than generating an adversarial input that deceives the classifier only. Here stealthiness refers to keeping the interpretation map of an adversarial example highly similar to that of the corresponding benign example. The existing work had no agreement on the relationship between robust classification and network interpretability. In this work, we will revisit the validity of ISA and propose a solution to improve the adversarial robustness of CNNs by leveraging robust interpretation in a proper way.
The most relevant work to ours is (Chen et al., 2019), which proposed a robust attribution training method with the aid of integrated gradient (IG), an axiomatic attribution map. It showed that the robust attribution training provides a generalization of several commonly-used robust training methods to defend adversarial attacks.
Different from the previous work, our paper contains the
following contributions.
1. By revisiting the validity of ISA, we show that enforc-
ing stealthiness of adversarial examples to a network interpreter could be challenging. Its difficulty relies on how one measures the interpretation discrepancy caused by input perturbations.
2. We propose an $\ell_1$-norm 2-class interpretation discrep-
ancy measure and theoretically show that constraining it helps adversarial robustness. Spurred by that, we develop a principled interpretability-aware robust training method, which provides a means to achieve robust classification by robust interpretation directly.
3. We empirically show that interpretability alone can be
used to defend adversarial attacks for both misclassifcation and misinterpretation. Compared to the IG-based robust attribution training (Chen et al., 2019), our approach is lighter in computation and provides better robustness even when facing a strong adversary.
## $2. \textbf{ Preliminaries and Motivation}$
In this section, we provide a brief background on interpretation methods of CNNs for justifying a classification decision, and motivate the phenomenon of interpretation discrepancy caused by adversarial examples.
To explain what and why CNNs predict, we consider two types of network interpretation methods: a) class activation $map\left(CAM\right)\left(\mathrm{Zhou~et~al.,~2016;~Selvaraju~et~al.,~2017;~Chat-~}\right)$ topadhay et al., 2018) and b) pixel sensitivity map (PSM)
(Simonyan et al., 2013; Springenberg et al., 2014; Smilkov et al., 2017; Sundararajan et al., 2017; Yeh et al., 2019). Let $f(\mathbf{x})\in\mathbb{R}^C$ denote a CNN-based predictor that maps an input $\mathbf{x}\in\mathbb{R}^d$ to a probability vector of $C$ classes. Here $f_c(\mathbf{x})$, the $c$th element of $f(\mathbf{x})$, denotes the classification score (given by logit before the softmax) for class $c.$ Let $I(\mathbf{x},c)$ denote an interpreter (CAM or PSM) that reflects where in x contributes to the classifier's decision on $c.$
Proper Network Interpretability Helps Adversarial Robustness in Classification
CAM-type methods. CAM (Zhou et al., 2016) produces a class-discriminative localization map for CNNs, which performs global averaging pooling over convolutional feature maps prior to the softmax. Let the penultimate layer output $\hat{K}$ feature maps, each of which is denoted by a vector $\begin{array}{l}\text{representation A}_k\in\mathbb{R}^u\text{ for channel }k\in[K].\text{ Here }[K]\\\text{represents the integer set }\{1,2,\ldots,K\}.\text{ The }i\text{th entry of}\\\text{CAM }I_{\mathrm{CAM}}(\mathbf{x},c)\text{ is given by}\end{array}$
(1)
$$[I_{\mathrm{CAM}}(\mathbf{x},c)]_{i}=(1/u)\sum_{k\in[K]}w_{k}^{c}A_{k,i},\:i\in[u],$$
where $w_k^c$ is the linear classification weight that associates the channel $k$ with the class $c$, and $A_k,i$ denotes the $i$the channel $k$ with the class $c$, and $A_k,i$ denotes the $i$th element of $\mathbf{A}_k.$ The rationale behind (1) is that the classification score $f_c(\mathbf{x})$ can be written as the average of CAM values (Zhou et al., 2016), $f_{c}( \mathbf{x} ) = \sum _{i= 1}^{u}[ I_{\mathrm{CAM}}( \mathbf{x} , c) ] _{i}.$ For visual explanation, $I_{\mathrm{CAM}}( \mathbf{x} , c)$ is often up-sampled to the input dimension $d$ using bi-linear interpolation.
GradCAM (Selvaraju et al., 2017) generalizes CAM for CNNs without the architecture`global average pooling $\to$ softmax layer' over the final convolutional maps. Specifically, the weight $w_k^c$ in (1) is given by the gradient of the classification score $f_c(\mathbf{x})$ with respect to (w.r.t.) the feature $\operatorname*{map}\mathbf{A}_{k},w_{k}^{c}=\frac{1}{u}\sum_{i=1}^{u}\frac{\partial f_{c}(\mathbf{x})}{\partial A_{k,i}}.$GradCAM++(Chattopadhay et al., 2018), a generalized formulation of GradCAM utilizes a more involved weighted average of the (positive) pixel-wise gradients but provides a better localization map if an image contains multiple occurrences of the same class In this work, we focus on CAM since it is computationally light and our models used in experiments follow the architecture ‘global average pooling $\to$softmax layer'.
PSM-type methods. PSM assigns importance scores to individual pixels toward explaining the classification decision aboût an input. Êxamples of commonly-used approaches include vanilla gradient (Simonyan et al., 2013), guided backpropogation (Springenberg et al., 2014), SmoothGrad ( Smilkov et al. , 2017) , and integrated gradient (IG) (Sundararajan et al., 2017). In particular, IG satisfes the completeness attribution axiom that PSM ought to obey. Specifically, it averages gradient saliency maps for interpolations between an input x and a baseline image a:
$$[I_{\mathrm{IG}}(\mathbf{x},c)]_{i}=(x_{i}-a_{i})\int_{\alpha=0}^{1}\frac{\partial f_{c}(\mathbf{a}+\alpha(\mathbf{x}-\mathbf{a}))}{\partial x_{i}}d\alpha\\\approx(x_{i}-a_{i})\sum_{j=1}^{m}\frac{\partial f_{c}(\mathbf{a}+\frac{j}{m}(\mathbf{x}-\mathbf{a}))}{\partial x_{i}}\frac{1}{m},\:i\in[d],$$
(2)
|
[
[
0.23779296875,
0.05682373046875,
0.73779296875,
0.05682373046875,
0.73779296875,
0.0706787109375,
0.23779296875,
0.0706787109375
],
[
0.08905029296875,
0.08587646484375,
0.47802734375,
0.08587646484375,
0.47802734375,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.5,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.0882568359375,
0.10162353515625,
0.476318359375,
0.10162353515625,
0.476318359375,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.498291015625,
0.0997314453125,
0.8857421875,
0.10101318359375,
0.8857421875,
0.11553955078125,
0.498291015625,
0.1142578125
],
[
0.0882568359375,
0.11614990234375,
0.4755859375,
0.11614990234375,
0.4755859375,
0.1300048828125,
0.0882568359375,
0.1300048828125
],
[
0.5009765625,
0.11553955078125,
0.8857421875,
0.11553955078125,
0.8857421875,
0.1300048828125,
0.5009765625,
0.1300048828125
],
[
0.0882568359375,
0.1307373046875,
0.473876953125,
0.1307373046875,
0.473876953125,
0.14453125,
0.0882568359375,
0.14453125
],
[
0.5,
0.1307373046875,
0.8857421875,
0.1307373046875,
0.8857421875,
0.1458740234375,
0.5,
0.1458740234375
],
[
0.08660888671875,
0.146484375,
0.4755859375,
0.1458740234375,
0.4755859375,
0.160400390625,
0.08660888671875,
0.1610107421875
],
[
0.5,
0.146484375,
0.88623046875,
0.1458740234375,
0.88623046875,
0.160400390625,
0.5,
0.1610107421875
],
[
0.0882568359375,
0.1622314453125,
0.4755859375,
0.1622314453125,
0.4755859375,
0.175537109375,
0.0882568359375,
0.175537109375
],
[
0.5,
0.1622314453125,
0.88623046875,
0.1622314453125,
0.88623046875,
0.175537109375,
0.5,
0.175537109375
],
[
0.08905029296875,
0.1787109375,
0.47216796875,
0.1787109375,
0.47216796875,
0.189453125,
0.08905029296875,
0.189453125
],
[
0.50244140625,
0.1767578125,
0.88623046875,
0.1767578125,
0.88623046875,
0.1900634765625,
0.50244140625,
0.1900634765625
],
[
0.08905029296875,
0.1932373046875,
0.473876953125,
0.1932373046875,
0.473876953125,
0.203857421875,
0.08905029296875,
0.203857421875
],
[
0.498291015625,
0.1912841796875,
0.85546875,
0.1925048828125,
0.85546875,
0.205810546875,
0.498291015625,
0.2052001953125
],
[
0.0882568359375,
0.20703125,
0.473876953125,
0.20703125,
0.473876953125,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.0882568359375,
0.222900390625,
0.47802734375,
0.222900390625,
0.47802734375,
0.236083984375,
0.0882568359375,
0.236083984375
],
[
0.5009765625,
0.2197265625,
0.8857421875,
0.2197265625,
0.8857421875,
0.23291015625,
0.5009765625,
0.23291015625
],
[
0.08905029296875,
0.2374267578125,
0.4755859375,
0.2374267578125,
0.4755859375,
0.250732421875,
0.08905029296875,
0.250732421875
],
[
0.498291015625,
0.23291015625,
0.88623046875,
0.231689453125,
0.88623046875,
0.24560546875,
0.498291015625,
0.246826171875
],
[
0.08660888671875,
0.250732421875,
0.4755859375,
0.25244140625,
0.4755859375,
0.266357421875,
0.08660888671875,
0.264404296875
],
[
0.5009765625,
0.246826171875,
0.88623046875,
0.24560546875,
0.88623046875,
0.2568359375,
0.5009765625,
0.25830078125
],
[
0.5009765625,
0.25830078125,
0.8857421875,
0.25830078125,
0.8857421875,
0.271484375,
0.5009765625,
0.271484375
],
[
0.0882568359375,
0.267578125,
0.473876953125,
0.267578125,
0.473876953125,
0.281005859375,
0.0882568359375,
0.281005859375
],
[
0.5,
0.270751953125,
0.8857421875,
0.270751953125,
0.8857421875,
0.2841796875,
0.5,
0.2841796875
],
[
0.08905029296875,
0.2841796875,
0.47216796875,
0.2841796875,
0.47216796875,
0.294189453125,
0.08905029296875,
0.294189453125
],
[
0.498291015625,
0.28271484375,
0.88232421875,
0.2822265625,
0.88232421875,
0.296630859375,
0.498291015625,
0.29736328125
],
[
0.0882568359375,
0.298583984375,
0.4248046875,
0.298583984375,
0.4248046875,
0.3125,
0.0882568359375,
0.3125
],
[
0.5,
0.296630859375,
0.88623046875,
0.296630859375,
0.88623046875,
0.311279296875,
0.5,
0.311279296875
],
[
0.5009765625,
0.309326171875,
0.6953125,
0.309326171875,
0.6953125,
0.322509765625,
0.5009765625,
0.322509765625
],
[
0.0882568359375,
0.3193359375,
0.47216796875,
0.3193359375,
0.47216796875,
0.333251953125,
0.0882568359375,
0.333251953125
],
[
0.08905029296875,
0.335205078125,
0.473876953125,
0.335205078125,
0.473876953125,
0.348388671875,
0.08905029296875,
0.348388671875
],
[
0.5537109375,
0.33203125,
0.833984375,
0.33203125,
0.833984375,
0.349853515625,
0.5537109375,
0.349853515625
],
[
0.861328125,
0.335205078125,
0.8857421875,
0.335205078125,
0.8857421875,
0.347900390625,
0.861328125,
0.347900390625
],
[
0.0882568359375,
0.35107421875,
0.473876953125,
0.35107421875,
0.473876953125,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.69384765625,
0.35107421875,
0.7294921875,
0.35107421875,
0.7294921875,
0.3623046875,
0.69384765625,
0.3623046875
],
[
0.0882568359375,
0.365478515625,
0.473876953125,
0.365478515625,
0.473876953125,
0.37890625,
0.0882568359375,
0.37890625
],
[
0.5009765625,
0.37255859375,
0.8857421875,
0.37255859375,
0.8857421875,
0.38330078125,
0.5009765625,
0.38330078125
],
[
0.0882568359375,
0.380615234375,
0.473876953125,
0.380615234375,
0.473876953125,
0.39453125,
0.0882568359375,
0.39453125
],
[
0.5009765625,
0.386474609375,
0.88623046875,
0.386474609375,
0.88623046875,
0.399658203125,
0.5009765625,
0.399658203125
],
[
0.08905029296875,
0.397705078125,
0.339111328125,
0.397705078125,
0.339111328125,
0.40771484375,
0.08905029296875,
0.40771484375
],
[
0.5009765625,
0.401611328125,
0.8857421875,
0.401611328125,
0.8857421875,
0.415283203125,
0.5009765625,
0.415283203125
],
[
0.08905029296875,
0.41845703125,
0.473876953125,
0.41845703125,
0.473876953125,
0.431884765625,
0.08905029296875,
0.431884765625
],
[
0.5,
0.416748046875,
0.8857421875,
0.415283203125,
0.8857421875,
0.429931640625,
0.5,
0.43115234375
],
[
0.0882568359375,
0.434326171875,
0.2484130859375,
0.434326171875,
0.2484130859375,
0.447509765625,
0.0882568359375,
0.447509765625
],
[
0.5009765625,
0.43115234375,
0.88623046875,
0.43115234375,
0.88623046875,
0.44580078125,
0.5009765625,
0.44580078125
],
[
0.5009765625,
0.4462890625,
0.88623046875,
0.4462890625,
0.88623046875,
0.4609375,
0.5009765625,
0.4609375
],
[
0.101318359375,
0.4609375,
0.4755859375,
0.4609375,
0.4755859375,
0.474853515625,
0.101318359375,
0.474853515625
],
[
0.5009765625,
0.462158203125,
0.8349609375,
0.462158203125,
0.8349609375,
0.47607421875,
0.5009765625,
0.47607421875
],
[
0.12091064453125,
0.4765625,
0.473876953125,
0.4765625,
0.473876953125,
0.490478515625,
0.12091064453125,
0.490478515625
],
[
0.5009765625,
0.484130859375,
0.8857421875,
0.484130859375,
0.8857421875,
0.49755859375,
0.5009765625,
0.49755859375
],
[
0.12091064453125,
0.4912109375,
0.473876953125,
0.4912109375,
0.473876953125,
0.505859375,
0.12091064453125,
0.505859375
],
[
0.5,
0.498779296875,
0.8837890625,
0.50048828125,
0.8837890625,
0.5146484375,
0.5,
0.51220703125
],
[
0.1192626953125,
0.5068359375,
0.47216796875,
0.5068359375,
0.47216796875,
0.52099609375,
0.1192626953125,
0.52099609375
],
[
0.5,
0.51513671875,
0.8857421875,
0.51513671875,
0.8857421875,
0.52880859375,
0.5,
0.52880859375
],
[
0.12091064453125,
0.52294921875,
0.32177734375,
0.52294921875,
0.32177734375,
0.5361328125,
0.12091064453125,
0.5361328125
],
[
0.5,
0.53076171875,
0.88623046875,
0.53076171875,
0.88623046875,
0.54443359375,
0.5,
0.54443359375
],
[
0.1004638671875,
0.54345703125,
0.473876953125,
0.54345703125,
0.473876953125,
0.556640625,
0.1004638671875,
0.556640625
],
[
0.5009765625,
0.54541015625,
0.88623046875,
0.54541015625,
0.88623046875,
0.5595703125,
0.5009765625,
0.5595703125
],
[
0.494384765625,
0.55615234375,
0.69287109375,
0.56005859375,
0.69287109375,
0.58203125,
0.494384765625,
0.57763671875
],
[
0.65185546875,
0.5546875,
0.8896484375,
0.56103515625,
0.8896484375,
0.58203125,
0.65185546875,
0.57568359375
],
[
0.1192626953125,
0.57373046875,
0.47216796875,
0.57373046875,
0.47216796875,
0.5869140625,
0.1192626953125,
0.5869140625
],
[
0.12091064453125,
0.56005859375,
0.473876953125,
0.56005859375,
0.473876953125,
0.57373046875,
0.12091064453125,
0.57373046875
],
[
0.5,
0.57763671875,
0.8857421875,
0.5771484375,
0.8857421875,
0.59130859375,
0.5,
0.59228515625
],
[
0.12091064453125,
0.59033203125,
0.473876953125,
0.59033203125,
0.473876953125,
0.60400390625,
0.12091064453125,
0.60400390625
],
[
0.5009765625,
0.59521484375,
0.8857421875,
0.59521484375,
0.8857421875,
0.60595703125,
0.5009765625,
0.60595703125
],
[
0.12091064453125,
0.60400390625,
0.473876953125,
0.60400390625,
0.473876953125,
0.6171875,
0.12091064453125,
0.6171875
],
[
0.5009765625,
0.609375,
0.88623046875,
0.609375,
0.88623046875,
0.62255859375,
0.5009765625,
0.62255859375
],
[
0.12091064453125,
0.62060546875,
0.420654296875,
0.62060546875,
0.420654296875,
0.6337890625,
0.12091064453125,
0.6337890625
],
[
0.5009765625,
0.62548828125,
0.88623046875,
0.62548828125,
0.88623046875,
0.63623046875,
0.5009765625,
0.63623046875
],
[
0.1004638671875,
0.64013671875,
0.473876953125,
0.6416015625,
0.473876953125,
0.65576171875,
0.1004638671875,
0.65478515625
],
[
0.5,
0.6396484375,
0.88818359375,
0.6396484375,
0.88818359375,
0.65283203125,
0.5,
0.65283203125
],
[
0.1217041015625,
0.6591796875,
0.47216796875,
0.6591796875,
0.47216796875,
0.66943359375,
0.1217041015625,
0.66943359375
],
[
0.5,
0.65478515625,
0.88623046875,
0.65478515625,
0.88623046875,
0.66845703125,
0.5,
0.66845703125
],
[
0.12091064453125,
0.671875,
0.473876953125,
0.671875,
0.473876953125,
0.68505859375,
0.12091064453125,
0.68505859375
],
[
0.498291015625,
0.66943359375,
0.859375,
0.66845703125,
0.859375,
0.68310546875,
0.498291015625,
0.6845703125
],
[
0.12091064453125,
0.68798828125,
0.473876953125,
0.68798828125,
0.473876953125,
0.7021484375,
0.12091064453125,
0.7021484375
],
[
0.5,
0.69580078125,
0.88623046875,
0.6962890625,
0.88623046875,
0.7109375,
0.5,
0.71044921875
],
[
0.1192626953125,
0.7041015625,
0.4755859375,
0.70263671875,
0.4755859375,
0.71630859375,
0.1192626953125,
0.71728515625
],
[
0.5009765625,
0.71044921875,
0.88623046875,
0.71044921875,
0.88623046875,
0.7236328125,
0.5009765625,
0.7236328125
],
[
0.11767578125,
0.71728515625,
0.4345703125,
0.71728515625,
0.4345703125,
0.73095703125,
0.11767578125,
0.73095703125
],
[
0.5009765625,
0.72265625,
0.6689453125,
0.72265625,
0.6689453125,
0.736328125,
0.5009765625,
0.736328125
],
[
0.66259765625,
0.72265625,
0.88623046875,
0.72265625,
0.88623046875,
0.736328125,
0.66259765625,
0.736328125
],
[
0.50244140625,
0.73486328125,
0.88623046875,
0.73486328125,
0.88623046875,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.0882568359375,
0.75048828125,
0.35693359375,
0.75244140625,
0.35693359375,
0.767578125,
0.0882568359375,
0.76513671875
],
[
0.50244140625,
0.74755859375,
0.88623046875,
0.74755859375,
0.88623046875,
0.76123046875,
0.50244140625,
0.76123046875
],
[
0.5009765625,
0.76025390625,
0.88623046875,
0.76025390625,
0.88623046875,
0.7734375,
0.5009765625,
0.7734375
],
[
0.08905029296875,
0.77783203125,
0.476318359375,
0.77783203125,
0.476318359375,
0.79248046875,
0.08905029296875,
0.79248046875
],
[
0.5009765625,
0.7734375,
0.88623046875,
0.7734375,
0.88623046875,
0.787109375,
0.5009765625,
0.787109375
],
[
0.5,
0.78515625,
0.88623046875,
0.7861328125,
0.88623046875,
0.7998046875,
0.5,
0.798828125
],
[
0.0882568359375,
0.79345703125,
0.473876953125,
0.79248046875,
0.473876953125,
0.806640625,
0.0882568359375,
0.80810546875
],
[
0.5009765625,
0.798828125,
0.88623046875,
0.798828125,
0.88623046875,
0.8125,
0.5009765625,
0.8125
],
[
0.0882568359375,
0.8095703125,
0.4755859375,
0.8095703125,
0.4755859375,
0.82275390625,
0.0882568359375,
0.82275390625
],
[
0.5009765625,
0.81298828125,
0.833984375,
0.81298828125,
0.833984375,
0.82373046875,
0.5009765625,
0.82373046875
],
[
0.08660888671875,
0.82470703125,
0.381591796875,
0.82470703125,
0.381591796875,
0.83837890625,
0.08660888671875,
0.83837890625
],
[
0.7099609375,
0.82763671875,
0.85693359375,
0.833984375,
0.85546875,
0.85595703125,
0.708984375,
0.849609375
],
[
0.0882568359375,
0.84521484375,
0.473876953125,
0.84521484375,
0.473876953125,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.53173828125,
0.84228515625,
0.65869140625,
0.84228515625,
0.65869140625,
0.85595703125,
0.53173828125,
0.85595703125
],
[
0.76123046875,
0.85107421875,
0.7890625,
0.85107421875,
0.7890625,
0.86279296875,
0.76123046875,
0.86279296875
],
[
0.08660888671875,
0.8623046875,
0.4755859375,
0.861328125,
0.4755859375,
0.875,
0.08660888671875,
0.87548828125
],
[
0.61376953125,
0.8662109375,
0.77783203125,
0.8671875,
0.77783203125,
0.8818359375,
0.61376953125,
0.880859375
],
[
0.0882568359375,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.78173828125,
0.87548828125,
0.83642578125,
0.87548828125,
0.83642578125,
0.89013671875,
0.78173828125,
0.89013671875
],
[
0.86279296875,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.8896484375,
0.86279296875,
0.8896484375
],
[
0.6845703125,
0.88525390625,
0.7138671875,
0.88525390625,
0.7138671875,
0.89794921875,
0.6845703125,
0.89794921875
],
[
0.53515625,
0.87744140625,
0.6259765625,
0.87744140625,
0.6259765625,
0.8896484375,
0.53515625,
0.8896484375
],
[
0.609375,
0.8916015625,
0.63232421875,
0.8916015625,
0.63232421875,
0.90283203125,
0.609375,
0.90283203125
],
[
0.7646484375,
0.88623046875,
0.78515625,
0.88623046875,
0.78515625,
0.896484375,
0.7646484375,
0.896484375
],
[
0.0882568359375,
0.892578125,
0.4755859375,
0.892578125,
0.4755859375,
0.90576171875,
0.0882568359375,
0.90576171875
]
] |
[
"Proper Network Interpretability Helps Adversarial Robustness in Classification",
"as a helpful metric to differentiate adversarial examples",
"(Simonyan et al., 2013; Springenberg et al.,2014; Smilkov",
"from benign inputs. Nevertheless, the work (Zhang et al.,",
"et al., 2017; Sundararajan et al., 2017; Yeh et al., 2019). Let",
"2018; Subramanya et al., 2018) showed that under certain",
"f(x) RC denote a CNN-based predictor that maps an",
"conditions, generating an attack (which we call an inter-",
"input x E IRd to a probability vector of C classes. Here",
"pretability sneaking attack, ISA) that fools the classifier",
"fe(x), the cth element of f(x), denotes the classification",
"while keeping it stealthy from the coupled interpreter is not",
"score (given by logit before the softmax) for class c. Let",
"significantlymore difficultthangeneratinganadversarial in",
"I(x, c) denote an interpreter (CAM or PSM) that reflects",
"putthat deceives theclassifier only.Here stealthinessrefers",
"where in x contributes to the classifier's decision on c.",
"to keeping the interpretation map of an adversarial example",
"highly similar to that of the corresponding benign example.",
"CAM-type methods..CAM (Zhou et al.,2016)produces",
"The existing work had no agreement on the relationship",
"a class-discriminative localization map for CNNs, which",
"between robust classification and network interpretability.",
"performsglobal averaging pooling over convolutionalfea-",
"ture maps prior to the softmax. Let the penultimate layer",
"In this work, we will revisit the validity of ISA and propose",
"output K feature maps, each of which is denoted by a vector",
"asolutiontoimprovetheadversarialrobustnessofCNNs",
"representation A IR\"for channel k E [K]. Here [K",
"by leveraging robust interpretation in a proper way.",
"represents the integer set {1, 2,..., K]. The ith entry of",
"CAM IcAM(x, c) is given by",
"The most relevant work to ours is (Chen et al., 2019), which",
"proposed a robust attribution training method with the aid",
"[IcAM(x,c)]; = (1/u) wAk,i, i [u],",
"(1)",
"of integrated gradient (IG), an axiomatic attribution map",
"ke[K]",
"It showed that the robust attribution training provides a",
"where w is the linear classification weightthat associates",
"generalization of several commonly-used robust training",
"the channel k with the class c, and Aki denotes the ith",
"methods to defend adversarial attacks.",
"element of Ak. The rationale behind (1) is that the classi",
"Different from the previous work, our paper contains the",
"fication score fe(x) can be written as the average of CAM",
"following contributions.",
"values (Zhou et al., 2016), f(x) = u-,[IcAM(x,c)]i",
"For visual explanation, IcAM(x, c) is often up-sampled to",
"1.By revisiting the validity of ISA, we show that enforc-",
"the input dimension d using bi-linear interpolation.",
"ing stealthiness of adversarial examples to a network",
"GradCAM (Selvaraju et al., 2017) generalizes CAM for",
"interpreter could be challenging.Its difficulty relies",
"CNNs without the architecture'global averagepooling-→",
"on how one measures the interpretation discrepancy",
"softmax layer'over the final convolutional maps. Specifi-",
"caused by input perturbations.",
"cally, the weight w in (1) is given by the gradient of the",
"2. We propose an l,-norm 2-class interpretation discrep-",
"classification score f.(x) with respect to (w.r.t.) the feature",
"map Ak, we - - ak",
" fa(x). GradCAM++ (Chattopad-",
"it helps adversarial robustness. Spurred by that, we",
"ancy measure and theoretically show that constraining",
"hay et al., 2018), a generalized formulation of GradCAM",
"develop a principled interpretability-aware robust train-",
"utilizesa more involved weighted average of the (positive)",
"ing method, which provides a means to achieve robust",
"pixel-wise gradients but provides a better localization map",
"classification by robust interpretation directly.",
"if animage contains multiple occurrences of the same class.",
"3.We empirically show that interpretability alone can be",
"In this work, we focus on CAM since it is computation-",
"usedtodefendadversarialattacksforbothmisclassifca",
"ally light and our models used in experiments follow the",
"tion and misinterpretation. Compared to the IG-based",
"architecture global average pooling → softmax layer'.",
"robust attribution training (Chen et al., 2019), our ap-",
"PSM-type methods.PSM assigns importance scores to",
"proach is lighter in computation and provides better",
"individual pixels toward explaining the classification de-",
"robustness even when facing a strong adversary",
"cision about an input.",
"Examplesof commonly-used",
"approaches include vanilla gradient (Simonyan et al.",
"2.Preliminaries and Motivation",
"2013), guided backpropogation (Springenberg et al., 2014),",
"SmoothGrad (Smilkov et al., 2017), and integrated gradient",
"In this section, we provide a brief background on inter-",
"(IG) (Sundararajan et al., 2017). In particular, IG satisfies",
"the completeness attribution axiom that PSM ought to obey",
"pretation methods of CNNs for justifying a classification",
"Specifically, it averages gradient saliency maps for interpo-",
"decision, and motivate the phenomenon of interpretation",
"lations between an input x and a baseline image a:",
"discrepancy caused by adversarial examples.",
"afe(a + a(x - a) da",
"To explain what and why CNNs predict, we consider two",
"[IIG(x,c)]i = (i-",
"Ori",
"types ofnetwork interpretation methods: a) class activation",
"m afe(a+(x-a)) 1",
"map (CAM) (Zhou et al., 2016; Selvaraju et al., 2017; Chat-",
"ie[d],",
"(2)",
"Ori",
"~(ai-a) >",
"j=1",
"m",
"topadhay et al., 2018) and b) pixel sensitivity map (PSM)"
] |
{
"arxiv_id": "2006.14748",
"language": "en",
"page_idx": 1
}
|
||
2006.14748-en-2
|
2006.14748-en-2.jpg
|
Proper Network Interpretability Helps Adversarial Robustness in Classification
<FigureHere>
Figure 1. Interpretation $(I)$ of benign $(\mathbf{x})$ and adversarial $(\mathbf{x}^{\prime})$ image from Restricted ImageNet (Tsipras et al., 2019) with respect to the true label $y=`$monkey’and the target label $y^{\prime}=`$fish'. Here the adversarial example is generated by 10-step PGD attack with perturbation size 0.02 on Wide-Resnet (Madry et al, 2018), and we consider three types of interpretation maps, CAM, GradCAM++ and IG. Given an interpretation method, the first column is $I(\mathbf{x},y)$ versus $I(\mathbf{x}^\prime,y)$, the second column is $I(\mathbf{x},y^{\prime})$ versus $I(\mathbf{x}^\prime,y^{\prime})$, and all maps under each category are normalized w.r.t. their largest value. At the bottom of each column, we quantify the resulting interpretation discrepancy by Kendall's Tau order rank correlation (Selvaraju et al., 2017) between every pair of $I(\mathbf{x},i)$ and $I(\mathbf{x}^{\prime},i)$ for $i=y$ or $y^{\prime}.$
where $m$ is the number of steps in the Riemman approximation of the integral. The completeness axiom (Sundararajan et al., 2017, Proposition 1) states that $\sum _i= 1^d[ I_{\mathrm{IG}}( \mathbf{x} , c) ] _i=$ $f_{c}(\mathbf{x})-f_{c}(\mathbf{a})$, where the baseline image a is often chosen such that $f_{c}(\mathbf{a})\approx0$,e.g., the black image. Note that CAM also satisfies the completeness axiom. PSM is able to highlight fine-grained details in the image, but is computationally intensive and not quite class-discriminative compared to CAM (Selvaraju et al., 2017).
intact. If such adversarial vulnerability exists, it could have serious consequences when classification and interpretation are jointly used in tasks like medical diagnosis (Subramanya et al., 2018), and call into question the faithfulness of interpretation to network classification. It is also suggested from interpretation discrepancy that an interpreter itself could be quite sensitive to input perturbations (even if they were not designed for misclassification). Spurred by that, the robustness of interpretation provides a supplementary robustness metric for CNNs (Ghorbani et al., 2019; Dombrowski et al., 2019; Chen et al., 2019).
## $3. \textbf{ Robustness of Classification vs. Robustness}$ of Interpretation
Interpretation discrepancy caused by adversarial perturbations. Let $\mathbf{x}^\prime=\mathbf{x}+\delta$ represent an adversarial example w.r.t. x, where $\delta$ denotes an adversarial perturbation. By replacing the input image x with $\mathbf{x}^{\prime}$, a CNN will be fooled from the true label y to the target (incorrect) label y'. It was recently shown in (Xu et al., 2019b;a) that the adversary could introduce an evident interpretation discrepancy w.r.t. both the true and the target label in terms of $I(\mathbf{x},y)$ vs. $I(\mathbf{x}^{\prime},y)$,and $I(\mathbf{x},y^{\prime})$ vs. $I( \mathbf{x} ^{\prime }, y^{\prime }) . \textbf{An illustrative example}$ is provided in Figure 1. We see that an adversary suppresses the network interpretation w.r.t. the true label $y$ but $pro-$ motes the interpretation w.r.t. the target label $y^\prime.$ We also observe that compared to IG, CAM and GradCAM++ better localize class-specific discriminative regions.
In this section, we revisit the validity of interpretability sneaking attack (ISA) from the perspective of interpretation discrepancy. We show that it is in fact quite challenging to force an adversarial example to mitigate its associated interpretation discrepancy. Further, we propose a novel measure of interpretation discrepancy, and theoretically show that constraining it prevents the success of adversarial attacks (for misclassification).
Previous work (Zhang et al.,2018; Subramanya et al., 2018) showed that it is $not\textit{difficult to prevent adversarial examples}$ from having large interpretation discrepancy when the latter is measured w.r.t. a single class label (either the true label $y$ or the target label $y^\prime).$ However, we see from Figure 1 that the prediction-evasion adversarial attack alters interpretation maps w.r.t. bothy and $y^\prime.$ This motivates us to rethink whether the single-class interpretation discrepancy measure is proper, and whether ISA is truly easy to bypass
The example in Figure 1 provides two implications on the robustness of classification versus interpretation discrepancy. First, an adversarial example designed for misclassification gives rise to interpretation discrepancy. Spurred by that, the problem of interpretability sneaking attack arises (Zhang et al., 2018; Subramanya et al., 2018): One may wonder whether or not it is easy to generate adversarial examples that mistake classification but keep interpretation
|
[
[
0.236083984375,
0.05682373046875,
0.736328125,
0.05682373046875,
0.736328125,
0.0706787109375,
0.236083984375,
0.0706787109375
],
[
0.31201171875,
0.08331298828125,
0.34716796875,
0.08331298828125,
0.34716796875,
0.0953369140625,
0.31201171875,
0.0953369140625
],
[
0.51708984375,
0.08331298828125,
0.58642578125,
0.08331298828125,
0.58642578125,
0.0947265625,
0.51708984375,
0.0947265625
],
[
0.7607421875,
0.08331298828125,
0.7861328125,
0.08331298828125,
0.7861328125,
0.09661865234375,
0.7607421875,
0.09661865234375
],
[
0.70751953125,
0.0947265625,
0.73388671875,
0.0947265625,
0.73388671875,
0.1060791015625,
0.70751953125,
0.1060791015625
],
[
0.125,
0.083984375,
0.1871337890625,
0.083984375,
0.1871337890625,
0.0953369140625,
0.125,
0.0953369140625
],
[
0.36767578125,
0.09661865234375,
0.396240234375,
0.09661865234375,
0.396240234375,
0.10418701171875,
0.36767578125,
0.10418701171875
],
[
0.81396484375,
0.09661865234375,
0.8388671875,
0.09661865234375,
0.8388671875,
0.10479736328125,
0.81396484375,
0.10479736328125
],
[
0.48779296875,
0.0985107421875,
0.50634765625,
0.0985107421875,
0.50634765625,
0.1029052734375,
0.48779296875,
0.1029052734375
],
[
0.26318359375,
0.09783935546875,
0.281005859375,
0.0947265625,
0.281982421875,
0.10162353515625,
0.2646484375,
0.10479736328125
],
[
0.59228515625,
0.0985107421875,
0.615234375,
0.0985107421875,
0.615234375,
0.1029052734375,
0.59228515625,
0.1029052734375
],
[
0.26708984375,
0.1256103515625,
0.27783203125,
0.1256103515625,
0.27783203125,
0.1287841796875,
0.26708984375,
0.1287841796875
],
[
0.55615234375,
0.271484375,
0.62109375,
0.271484375,
0.62109375,
0.279052734375,
0.55615234375,
0.279052734375
],
[
0.66650390625,
0.271484375,
0.7294921875,
0.271484375,
0.7294921875,
0.279052734375,
0.66650390625,
0.279052734375
],
[
0.77392578125,
0.271484375,
0.86767578125,
0.271484375,
0.86767578125,
0.279052734375,
0.77392578125,
0.279052734375
],
[
0.3349609375,
0.272705078125,
0.374267578125,
0.272705078125,
0.374267578125,
0.276611328125,
0.3349609375,
0.276611328125
],
[
0.2320556640625,
0.272705078125,
0.281005859375,
0.272705078125,
0.281005859375,
0.276611328125,
0.2320556640625,
0.276611328125
],
[
0.449462890625,
0.2734375,
0.461669921875,
0.2734375,
0.461669921875,
0.276611328125,
0.449462890625,
0.276611328125
],
[
0.0882568359375,
0.29541015625,
0.8857421875,
0.29541015625,
0.8857421875,
0.31005859375,
0.0882568359375,
0.31005859375
],
[
0.0882568359375,
0.31005859375,
0.8857421875,
0.31005859375,
0.8857421875,
0.323974609375,
0.0882568359375,
0.323974609375
],
[
0.0882568359375,
0.323974609375,
0.8837890625,
0.324462890625,
0.8837890625,
0.3359375,
0.0882568359375,
0.335205078125
],
[
0.0882568359375,
0.33837890625,
0.8857421875,
0.33837890625,
0.8857421875,
0.3515625,
0.0882568359375,
0.3515625
],
[
0.0882568359375,
0.35107421875,
0.8837890625,
0.35107421875,
0.8837890625,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.0882568359375,
0.365478515625,
0.78173828125,
0.365478515625,
0.78173828125,
0.37890625,
0.0882568359375,
0.37890625
],
[
0.0882568359375,
0.404052734375,
0.47802734375,
0.40478515625,
0.47802734375,
0.419189453125,
0.0882568359375,
0.41845703125
],
[
0.5,
0.40478515625,
0.88623046875,
0.40478515625,
0.88623046875,
0.41845703125,
0.5,
0.41845703125
],
[
0.0882568359375,
0.421630859375,
0.4755859375,
0.421630859375,
0.4755859375,
0.43505859375,
0.0882568359375,
0.43505859375
],
[
0.498291015625,
0.42041015625,
0.88623046875,
0.42041015625,
0.88623046875,
0.434326171875,
0.498291015625,
0.434326171875
],
[
0.08660888671875,
0.43310546875,
0.476318359375,
0.43505859375,
0.476318359375,
0.45263671875,
0.08660888671875,
0.45068359375
],
[
0.5,
0.436279296875,
0.8857421875,
0.436279296875,
0.8857421875,
0.449462890625,
0.5,
0.449462890625
],
[
0.08905029296875,
0.451904296875,
0.473876953125,
0.451904296875,
0.473876953125,
0.46533203125,
0.08905029296875,
0.46533203125
],
[
0.5,
0.45068359375,
0.8857421875,
0.45068359375,
0.8857421875,
0.464111328125,
0.5,
0.464111328125
],
[
0.08905029296875,
0.46728515625,
0.473876953125,
0.46728515625,
0.473876953125,
0.48095703125,
0.08905029296875,
0.48095703125
],
[
0.496826171875,
0.46533203125,
0.8857421875,
0.466552734375,
0.8857421875,
0.479736328125,
0.496826171875,
0.479248046875
],
[
0.0906982421875,
0.484130859375,
0.47216796875,
0.484130859375,
0.47216796875,
0.494384765625,
0.0906982421875,
0.494384765625
],
[
0.498291015625,
0.481689453125,
0.88623046875,
0.48095703125,
0.88623046875,
0.494384765625,
0.498291015625,
0.49560546875
],
[
0.08905029296875,
0.49755859375,
0.4755859375,
0.49755859375,
0.4755859375,
0.51123046875,
0.08905029296875,
0.51123046875
],
[
0.5,
0.496826171875,
0.88623046875,
0.496826171875,
0.88623046875,
0.51025390625,
0.5,
0.51025390625
],
[
0.08905029296875,
0.51318359375,
0.4755859375,
0.51318359375,
0.4755859375,
0.5263671875,
0.08905029296875,
0.5263671875
],
[
0.498291015625,
0.51025390625,
0.8857421875,
0.51123046875,
0.8857421875,
0.52587890625,
0.498291015625,
0.5244140625
],
[
0.0882568359375,
0.52783203125,
0.301513671875,
0.52783203125,
0.301513671875,
0.54150390625,
0.0882568359375,
0.54150390625
],
[
0.5009765625,
0.5263671875,
0.8857421875,
0.5263671875,
0.8857421875,
0.54052734375,
0.5009765625,
0.54052734375
],
[
0.50244140625,
0.54345703125,
0.8857421875,
0.54345703125,
0.8857421875,
0.5537109375,
0.50244140625,
0.5537109375
],
[
0.5009765625,
0.556640625,
0.79345703125,
0.556640625,
0.79345703125,
0.57080078125,
0.5009765625,
0.57080078125
],
[
0.0906982421875,
0.568359375,
0.47216796875,
0.568359375,
0.47216796875,
0.5791015625,
0.0906982421875,
0.5791015625
],
[
0.0906982421875,
0.583984375,
0.47216796875,
0.583984375,
0.47216796875,
0.5947265625,
0.0906982421875,
0.5947265625
],
[
0.5009765625,
0.59033203125,
0.88232421875,
0.59130859375,
0.88232421875,
0.60595703125,
0.5009765625,
0.60498046875
],
[
0.08905029296875,
0.59765625,
0.4755859375,
0.59765625,
0.4755859375,
0.611328125,
0.08905029296875,
0.611328125
],
[
0.0882568359375,
0.6123046875,
0.473876953125,
0.611328125,
0.473876953125,
0.62548828125,
0.0882568359375,
0.626953125
],
[
0.521484375,
0.60986328125,
0.66259765625,
0.60986328125,
0.66259765625,
0.62353515625,
0.521484375,
0.62353515625
],
[
0.08905029296875,
0.62744140625,
0.473876953125,
0.62744140625,
0.473876953125,
0.6416015625,
0.08905029296875,
0.6416015625
],
[
0.5009765625,
0.63623046875,
0.88134765625,
0.63623046875,
0.88134765625,
0.646484375,
0.5009765625,
0.646484375
],
[
0.08905029296875,
0.64306640625,
0.473876953125,
0.64306640625,
0.473876953125,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.5009765625,
0.65087890625,
0.8837890625,
0.65087890625,
0.8837890625,
0.6640625,
0.5009765625,
0.6640625
],
[
0.08660888671875,
0.6572265625,
0.473876953125,
0.65771484375,
0.473876953125,
0.67236328125,
0.08660888671875,
0.671875
],
[
0.5009765625,
0.66552734375,
0.88623046875,
0.66552734375,
0.88623046875,
0.6787109375,
0.5009765625,
0.6787109375
],
[
0.08660888671875,
0.671875,
0.4755859375,
0.67236328125,
0.4755859375,
0.6875,
0.08660888671875,
0.68603515625
],
[
0.498291015625,
0.6787109375,
0.88623046875,
0.6796875,
0.88623046875,
0.6943359375,
0.498291015625,
0.693359375
],
[
0.08905029296875,
0.68798828125,
0.473876953125,
0.68798828125,
0.473876953125,
0.7021484375,
0.08905029296875,
0.7021484375
],
[
0.5,
0.69580078125,
0.88623046875,
0.69580078125,
0.88623046875,
0.708984375,
0.5,
0.708984375
],
[
0.08905029296875,
0.70263671875,
0.4755859375,
0.70263671875,
0.4755859375,
0.71630859375,
0.08905029296875,
0.71630859375
],
[
0.50244140625,
0.7119140625,
0.8857421875,
0.7119140625,
0.8857421875,
0.72265625,
0.50244140625,
0.72265625
],
[
0.08905029296875,
0.71826171875,
0.473876953125,
0.71826171875,
0.473876953125,
0.73193359375,
0.08905029296875,
0.73193359375
],
[
0.5009765625,
0.72607421875,
0.88623046875,
0.72607421875,
0.88623046875,
0.7392578125,
0.5009765625,
0.7392578125
],
[
0.0882568359375,
0.73291015625,
0.473876953125,
0.73291015625,
0.473876953125,
0.7470703125,
0.0882568359375,
0.7470703125
],
[
0.50244140625,
0.7421875,
0.64794921875,
0.7421875,
0.64794921875,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.08905029296875,
0.75,
0.47216796875,
0.75,
0.47216796875,
0.76025390625,
0.08905029296875,
0.76025390625
],
[
0.0882568359375,
0.76318359375,
0.384033203125,
0.76318359375,
0.384033203125,
0.77783203125,
0.0882568359375,
0.77783203125
],
[
0.5009765625,
0.76318359375,
0.8857421875,
0.76318359375,
0.8857421875,
0.7763671875,
0.5009765625,
0.7763671875
],
[
0.50244140625,
0.77978515625,
0.8837890625,
0.77978515625,
0.8837890625,
0.79052734375,
0.50244140625,
0.79052734375
],
[
0.08905029296875,
0.7861328125,
0.47216796875,
0.7861328125,
0.47216796875,
0.7998046875,
0.08905029296875,
0.7998046875
],
[
0.5009765625,
0.79345703125,
0.8857421875,
0.79345703125,
0.8857421875,
0.806640625,
0.5009765625,
0.806640625
],
[
0.08905029296875,
0.8017578125,
0.473876953125,
0.8017578125,
0.473876953125,
0.81494140625,
0.08905029296875,
0.81494140625
],
[
0.5,
0.80810546875,
0.8857421875,
0.80810546875,
0.8857421875,
0.8212890625,
0.5,
0.8212890625
],
[
0.08905029296875,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.830078125,
0.08905029296875,
0.830078125
],
[
0.5,
0.82373046875,
0.8857421875,
0.82373046875,
0.8857421875,
0.8369140625,
0.5,
0.8369140625
],
[
0.0906982421875,
0.83349609375,
0.47216796875,
0.83349609375,
0.47216796875,
0.84326171875,
0.0906982421875,
0.84326171875
],
[
0.498291015625,
0.8369140625,
0.88623046875,
0.83837890625,
0.88623046875,
0.85302734375,
0.498291015625,
0.8515625
],
[
0.0882568359375,
0.8466796875,
0.473876953125,
0.8466796875,
0.473876953125,
0.85986328125,
0.0882568359375,
0.85986328125
],
[
0.498291015625,
0.85400390625,
0.88623046875,
0.85302734375,
0.88623046875,
0.8671875,
0.498291015625,
0.86865234375
],
[
0.0882568359375,
0.861328125,
0.473876953125,
0.861328125,
0.473876953125,
0.875,
0.0882568359375,
0.875
],
[
0.5,
0.87060546875,
0.88232421875,
0.87060546875,
0.88232421875,
0.880859375,
0.5,
0.880859375
],
[
0.0882568359375,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.498291015625,
0.8837890625,
0.8857421875,
0.88330078125,
0.8857421875,
0.89794921875,
0.498291015625,
0.8984375
],
[
0.0882568359375,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90576171875
]
] |
[
"Proper Network Interpretability Helps Adversarial Robustness in Classification",
"CAM",
"GradCAM++",
"IG",
"I(,t)",
"Input image",
"I.y",
"Iy",
".-1",
"1(",
"1-2",
"",
"correlation:",
"correlation:",
"correlation:0.391",
"correlat",
"oreaon",
"",
"Figure 1. Interpretation (I) of benign (x) and adversarial (x) image from Restricted ImageNet (Tsipras et al., 2019) with respect to the",
"true label y-'monkey' and the target label y'=fish'. Here the adversarial example is generated by 10-step PGD attack with perturbation",
"size0.02onWide-Resnet (Madry et al..2018),and we consider three types of interpretationmaps,CAM,GradCAM++andIG.Given an",
"interpretation method, the first column is I(x, y) versus I(x', y), the second column is I(x, y') versus I(x', y'), and all maps under each",
"category are normalized w.r.t. their largest value. At the bottom of each column, we quantify the resulting interpretation discrepancy by",
"Kendall's Tau order rank correlation (Selvaraju et al., 2017) between every pair of I(x, i) and I(x', i) for i = y or y'.",
"where m is the number of steps in the Riemman approxima-",
"intact. If such adversarial vulnerability exists, it could have",
"tion of the integral. The completeness axiom (Sundararajan",
"serious consequences when classification and interpretation",
"et al., 2017, Proposition 1) states that d-,[Iig(x, c)]; =",
"are jointly used in tasks like medical diagnosis (Subra-",
"fe(x) - fe(a), where the baseline image a is often chosen",
"manya et al., 2018), and call into question the faithfulness",
"such that fe(a) ~ O, e.g., the black image. Note that CAM",
"of interpretation to network classification.It is also sug",
"alsosatisfiesthecompletenessaxiom.PSMisabletohigh",
"gested from interpretation discrepancy that an interpreter",
"light fine-grained details in the image, but is computation-",
"itself could be quite sensitive to input perturbations (even",
"ally intensive and not quite class-discriminative compared",
"if they were not designed for misclassification). Spurred",
"to CAM (Selvaraju et al., 2017).",
"by that, the robustness of interpretation provides a supple",
"mentaryrobustnessmetricforCNNs(Ghorbanietal.,2019:",
"Dombrowski et al., 2019; Chen et al., 2019).",
"Interpretation discrepancy caused by adversarial per",
"turbations.Let x' =x + represent an adversarial ex",
"3. Robustness of Classification vs. Robustness",
"ample w.r.t. x, where denotes an adversarial perturbation.",
"By replacing the input image x with x', a CNN will be",
"of Interpretation",
"fooled from the true label y to the target (incorrect) label y",
"In this section, we revisit the validity of interpretability",
"It was recently shown in (Xu et al., 2019b;a) that the adver-",
"sneaking attack (ISA) from the perspective of interpretation",
"sary could introduce an evident interpretation discrepancy",
"discrepancy. We show that it is in fact quite challenging to",
"w.r.t. both the true and the target label in terms of I(x, y) vs.",
"force an adversarial example to mitigate its associated inter-",
"I(x', y), and I(x, y') vs. I(x', y'). An illustrative example",
"pretation discrepancy. Further, we propose a novel measure",
"is provided in Figure 1. We see that an adversary suppresses",
"of interpretationdiscrepancy,andtheoretically showthat",
"the network interpretation w.r.t. the true label y but pro",
"constraining it prevents the success of adversarial attacks",
"motes the interpretation w.r.t. the target label y'. We also",
"(formisclassification).",
"observethatcomparedtoIG,CAMandGradCAM++better",
"localize class-specific discriminative regions.",
"Previous work (Zhang et al., 2018; Subramanya et al., 2018)",
"showedthatitisnotdifficulttopreventadversarial examples",
"The example in Figure 1 provides two implications on the ro",
"from having large interpretation discrepancy when the latter",
"bustness of classification versus interpretation discrepancy",
"is measured w.r.t. a single class label (either the true label",
"First, an adversarial example designed for misclassifica-",
"y or the target label y'). However, we see from Figure 1",
"tion gives rise to interpretation discrepancySpurred by",
"that the prediction-evasion adversarial attack alters inter-",
"that, the problem of interpretability sneaking attack arises",
"pretation maps w.r.t. both y and y'. This motivates us to",
"(Zhang et al., 2018; Subramanya et al., 2018): One may",
"rethinkwhetherthesingle-classinterpretationdiscrepancy",
"wonder whether or not it is easy to generate adversarial",
"measure is proper, and whether ISA is truly easy to bypass",
"examples that mistake classification but keep interpretation"
] |
{
"arxiv_id": "2006.14748",
"language": "en",
"page_idx": 2
}
|
||
2005.14271-en-0
|
2005.14271-en-0.jpg
|
Relation Extraction with Explanation
Hamed Shahbazi, Xiaoli Z. Fern, Reza Ghaeini, Prasad Tadepalli
School of Electrical Engineering and Computer Science, Oregon State University
Corvallis, OR, USA
$\{$shahbazh, xfern, ghaeinim, tadepall$\}$@oregonstate.edu
## Abstract
Recent neural models for relation extraction with distant supervision alleviate the impact of irrelevant sentences in a bag by learning importance weights for the sentences. Efforts thus far have focused on improving extraction accuracy but little is known about their explainability. In this work we annotate a test set with ground-truth sentence-level explanations to evaluate the quality of explanations afforded by the relation extraction models. We demonstrate that replacing the entity mentions in the sentences with their fine-grained entity types not only enhances extraction accuracy but also improves explanation. We also propose to automatically generate“distractor” sentences to augment the bags and train the model to ignore the distractors. Evaluations on the widely used FB-NYT dataset show that our methods achieve new state-of-the-art accuracy while improving model explainability.$^{1}$
## 1 Introduction
Relation extraction with distant supervision associates a pair of entities with a bag of sentences, each containing mentions of both entities. The bag is tagged with relations between the pair in a Knowledge Base (KB), without explicitly indicating which sentence(s) support the relation(s). This method avoids the burden of manual annotations, but presents inherent ambiguity, creating challenges for learning.
To alleviate the impact of the irrelevant sentences many approaches have been proposed including models based on attention (Zeng et al., 2015; Lin et al., 2016; Liu et al., 2017; Luo et al., 2017; Du et al., 2018; Wang et al., 2018; Peng and Denilson, 2019; Bai and Ritter, 2019), approaches that use additional resources (Vashishth et al., 2018; Liu et al., 2018) and methods that utilize supervision
$^{1}$Accepted as a short paper at ACL 2020
data (Pershina et al., 2014; Angeli et al., 2014; Beltagy et al., 2019). These studies primarily focus on improving relation extraction accuracy and little is known about whether the models are making right decision for the right reason or because of some irrelevant biases (Agrawal et al., 2016; Gururangan et al., 2018; Ghaeini et al., 2019).
This paper examines two strong baseline relation extraction models with several explanation mechanisms. We manually annotated a test set from the widely used FB-NYT dataset with ground truth explanations to evaluate the quality of the explanation afforded by these models. We also introduce two different methods for improving relation extraction. First, we demonstrate that replacing the entity mentions with their fine-grained entity types for sentence representation leads to improvement in both the extract accuracy and model explainability. Second, we augment the bags with automatically generated“distractor” sentences (i.e, sentences that contain no supporting information for the relation) and train the model to appropriately ignore the irrelevant information. Our evaluation on the widely used FB-NYT dataset verifies that the proposed methods achieve the new state of the art for the extraction performance along with improved model explainability.
# 2 Problem Setup
Given entity pair $(e_i,e_j)$,we form a bag $B_i,j=$ $\begin{Bmatrix}s_{1},\ldots s_{N_{ij}}\end{Bmatrix}$ with $N_ij$ sentences that contain mentions of both entities and label it by the set of relations between $e_i$ and $e_j$ from the KB. Neural models for relation extraction encode each sentences into a vector representation and a bag $B_{i,j}$ is thus represented by $\{x_1,\ldots x_{N_{ij}}\}$ where $x_i\in\mathbb{R}^d.$
Given a set of bags and the associated labels, the training objective is to learn a model that predicts the probability $P( r$ = $k| B_{i, j})$ that relation
|
[
[
0.310546875,
0.08197021484375,
0.69189453125,
0.08197021484375,
0.69189453125,
0.0968017578125,
0.310546875,
0.0968017578125
],
[
0.2225341796875,
0.1341552734375,
0.78271484375,
0.1353759765625,
0.78271484375,
0.149658203125,
0.2225341796875,
0.1484375
],
[
0.175537109375,
0.152587890625,
0.82861328125,
0.152587890625,
0.82861328125,
0.166259765625,
0.175537109375,
0.166259765625
],
[
0.42236328125,
0.169189453125,
0.5859375,
0.169189453125,
0.5859375,
0.182861328125,
0.42236328125,
0.182861328125
],
[
0.187255859375,
0.1859130859375,
0.82275390625,
0.1859130859375,
0.82275390625,
0.20068359375,
0.187255859375,
0.20068359375
],
[
0.26611328125,
0.26416015625,
0.343505859375,
0.26416015625,
0.343505859375,
0.279052734375,
0.26611328125,
0.279052734375
],
[
0.5146484375,
0.265380859375,
0.88330078125,
0.265380859375,
0.88330078125,
0.279052734375,
0.5146484375,
0.279052734375
],
[
0.51318359375,
0.281982421875,
0.88330078125,
0.28076171875,
0.88330078125,
0.295654296875,
0.51318359375,
0.29638671875
],
[
0.14697265625,
0.292236328125,
0.459228515625,
0.292236328125,
0.459228515625,
0.305908203125,
0.14697265625,
0.305908203125
],
[
0.51318359375,
0.298583984375,
0.88330078125,
0.298583984375,
0.88330078125,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.0235137939453125,
0.30810546875,
0.06463623046875,
0.30810546875,
0.06463623046875,
0.44482421875,
0.0235137939453125,
0.44482421875
],
[
0.14697265625,
0.305908203125,
0.4609375,
0.305908203125,
0.4609375,
0.319580078125,
0.14697265625,
0.319580078125
],
[
0.51123046875,
0.3134765625,
0.88330078125,
0.314208984375,
0.88330078125,
0.328857421875,
0.51123046875,
0.327880859375
],
[
0.14697265625,
0.320556640625,
0.459228515625,
0.32177734375,
0.459228515625,
0.333251953125,
0.14697265625,
0.33203125
],
[
0.1485595703125,
0.3349609375,
0.4609375,
0.3349609375,
0.4609375,
0.3486328125,
0.1485595703125,
0.3486328125
],
[
0.51318359375,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.1485595703125,
0.350830078125,
0.459228515625,
0.350830078125,
0.459228515625,
0.361083984375,
0.1485595703125,
0.361083984375
],
[
0.51318359375,
0.3466796875,
0.88330078125,
0.3466796875,
0.88330078125,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.14697265625,
0.363525390625,
0.462646484375,
0.363525390625,
0.462646484375,
0.377197265625,
0.14697265625,
0.377197265625
],
[
0.51318359375,
0.3623046875,
0.76171875,
0.3623046875,
0.76171875,
0.3759765625,
0.51318359375,
0.3759765625
],
[
0.1485595703125,
0.37890625,
0.459228515625,
0.37890625,
0.459228515625,
0.3896484375,
0.1485595703125,
0.3896484375
],
[
0.53076171875,
0.380126953125,
0.884765625,
0.380126953125,
0.884765625,
0.394287109375,
0.53076171875,
0.394287109375
],
[
0.14697265625,
0.392578125,
0.4609375,
0.392578125,
0.4609375,
0.40625,
0.14697265625,
0.40625
],
[
0.51318359375,
0.396728515625,
0.88671875,
0.396728515625,
0.88671875,
0.410400390625,
0.51318359375,
0.410400390625
],
[
0.14697265625,
0.40625,
0.462646484375,
0.40625,
0.462646484375,
0.419189453125,
0.14697265625,
0.419189453125
],
[
0.5146484375,
0.41455078125,
0.884765625,
0.41455078125,
0.884765625,
0.424560546875,
0.5146484375,
0.424560546875
],
[
0.1485595703125,
0.421630859375,
0.459228515625,
0.421630859375,
0.459228515625,
0.431640625,
0.1485595703125,
0.431640625
],
[
0.51318359375,
0.4287109375,
0.88671875,
0.4287109375,
0.88671875,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.1485595703125,
0.435791015625,
0.459228515625,
0.435791015625,
0.459228515625,
0.446533203125,
0.1485595703125,
0.446533203125
],
[
0.0218353271484375,
0.446533203125,
0.05877685546875,
0.4453125,
0.06378173828125,
0.5166015625,
0.026031494140625,
0.51708984375
],
[
0.1485595703125,
0.45068359375,
0.459228515625,
0.45068359375,
0.459228515625,
0.46142578125,
0.1485595703125,
0.46142578125
],
[
0.5146484375,
0.4453125,
0.884765625,
0.4453125,
0.884765625,
0.458984375,
0.5146484375,
0.458984375
],
[
0.1485595703125,
0.46435546875,
0.45751953125,
0.46435546875,
0.45751953125,
0.474365234375,
0.1485595703125,
0.474365234375
],
[
0.51318359375,
0.460205078125,
0.88330078125,
0.46142578125,
0.88330078125,
0.474365234375,
0.51318359375,
0.473876953125
],
[
0.14697265625,
0.476806640625,
0.4609375,
0.476806640625,
0.4609375,
0.490478515625,
0.14697265625,
0.490478515625
],
[
0.51318359375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.14697265625,
0.4912109375,
0.4609375,
0.4912109375,
0.4609375,
0.5048828125,
0.14697265625,
0.5048828125
],
[
0.51318359375,
0.493408203125,
0.88330078125,
0.493408203125,
0.88330078125,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.145263671875,
0.5048828125,
0.4609375,
0.50439453125,
0.4609375,
0.5185546875,
0.145263671875,
0.51953125
],
[
0.51318359375,
0.5087890625,
0.884765625,
0.5087890625,
0.884765625,
0.5224609375,
0.51318359375,
0.5224609375
],
[
0.0201568603515625,
0.51708984375,
0.060455322265625,
0.51708984375,
0.06378173828125,
0.728515625,
0.0235137939453125,
0.728515625
],
[
0.1502685546875,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53076171875,
0.1502685546875,
0.53076171875
],
[
0.5146484375,
0.525390625,
0.88330078125,
0.525390625,
0.88330078125,
0.5390625,
0.5146484375,
0.5390625
],
[
0.1485595703125,
0.53515625,
0.45751953125,
0.53515625,
0.45751953125,
0.5458984375,
0.1485595703125,
0.5458984375
],
[
0.51123046875,
0.54052734375,
0.884765625,
0.54150390625,
0.884765625,
0.5546875,
0.51123046875,
0.55419921875
],
[
0.14697265625,
0.548828125,
0.4609375,
0.548828125,
0.4609375,
0.5625,
0.14697265625,
0.5625
],
[
0.51318359375,
0.5556640625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57177734375,
0.51318359375,
0.57080078125
],
[
0.1485595703125,
0.56396484375,
0.366943359375,
0.56396484375,
0.366943359375,
0.57470703125,
0.1485595703125,
0.57470703125
],
[
0.51318359375,
0.57373046875,
0.88330078125,
0.572265625,
0.88330078125,
0.58740234375,
0.51318359375,
0.58837890625
],
[
0.1376953125,
0.58740234375,
0.26123046875,
0.58740234375,
0.26123046875,
0.6015625,
0.1376953125,
0.6015625
],
[
0.51318359375,
0.59033203125,
0.884765625,
0.59033203125,
0.884765625,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.1351318359375,
0.59326171875,
0.1427001953125,
0.59326171875,
0.1427001953125,
0.59716796875,
0.1351318359375,
0.59716796875
],
[
0.124267578125,
0.5927734375,
0.130126953125,
0.5927734375,
0.130126953125,
0.5986328125,
0.124267578125,
0.5986328125
],
[
0.51123046875,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.6181640625,
0.51123046875,
0.6181640625
],
[
0.1175537109375,
0.61328125,
0.490234375,
0.6142578125,
0.490234375,
0.62890625,
0.1175537109375,
0.62744140625
],
[
0.51318359375,
0.62158203125,
0.88671875,
0.62158203125,
0.88671875,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.118408203125,
0.63134765625,
0.490234375,
0.63134765625,
0.490234375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.51123046875,
0.6357421875,
0.88671875,
0.63818359375,
0.88671875,
0.65185546875,
0.51123046875,
0.6494140625
],
[
0.118408203125,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.65380859375,
0.884765625,
0.65380859375,
0.884765625,
0.66748046875,
0.51318359375,
0.66748046875
],
[
0.118408203125,
0.6630859375,
0.490234375,
0.6630859375,
0.490234375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.51318359375,
0.67041015625,
0.884765625,
0.67041015625,
0.884765625,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.1175537109375,
0.677734375,
0.489501953125,
0.677734375,
0.489501953125,
0.69140625,
0.1175537109375,
0.69140625
],
[
0.5146484375,
0.68603515625,
0.6708984375,
0.68603515625,
0.6708984375,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.118408203125,
0.69482421875,
0.489501953125,
0.69482421875,
0.489501953125,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.118408203125,
0.71044921875,
0.490234375,
0.71142578125,
0.490234375,
0.7255859375,
0.118408203125,
0.724609375
],
[
0.51318359375,
0.7197265625,
0.67236328125,
0.7197265625,
0.67236328125,
0.73291015625,
0.51318359375,
0.73291015625
],
[
0.118408203125,
0.72705078125,
0.48779296875,
0.72705078125,
0.48779296875,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.1175537109375,
0.7421875,
0.29296875,
0.74365234375,
0.29296875,
0.75830078125,
0.1175537109375,
0.75732421875
],
[
0.51123046875,
0.74658203125,
0.87890625,
0.74658203125,
0.87890625,
0.76416015625,
0.51123046875,
0.76416015625
],
[
0.1368408203125,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.77392578125,
0.1368408203125,
0.77392578125
],
[
0.51123046875,
0.76318359375,
0.88671875,
0.76123046875,
0.88671875,
0.779296875,
0.51123046875,
0.78076171875
],
[
0.1175537109375,
0.775390625,
0.489501953125,
0.77490234375,
0.489501953125,
0.7890625,
0.1175537109375,
0.79052734375
],
[
0.51318359375,
0.78076171875,
0.88330078125,
0.78076171875,
0.88330078125,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.12005615234375,
0.79150390625,
0.490234375,
0.79150390625,
0.490234375,
0.8056640625,
0.12005615234375,
0.8056640625
],
[
0.51123046875,
0.79638671875,
0.88671875,
0.7958984375,
0.88671875,
0.81005859375,
0.51123046875,
0.81103515625
],
[
0.12005615234375,
0.80712890625,
0.490234375,
0.80712890625,
0.490234375,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.51123046875,
0.8134765625,
0.88330078125,
0.8134765625,
0.88330078125,
0.82666015625,
0.51123046875,
0.82666015625
],
[
0.12005615234375,
0.82373046875,
0.491943359375,
0.82373046875,
0.491943359375,
0.83740234375,
0.12005615234375,
0.83740234375
],
[
0.51318359375,
0.8291015625,
0.88330078125,
0.8291015625,
0.88330078125,
0.84326171875,
0.51318359375,
0.84326171875
],
[
0.118408203125,
0.83837890625,
0.489501953125,
0.8388671875,
0.489501953125,
0.85400390625,
0.118408203125,
0.8525390625
],
[
0.509765625,
0.84326171875,
0.8486328125,
0.8427734375,
0.8486328125,
0.86083984375,
0.509765625,
0.8623046875
],
[
0.118408203125,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.118408203125,
0.869140625
],
[
0.53076171875,
0.8623046875,
0.88671875,
0.8623046875,
0.88671875,
0.876953125,
0.53076171875,
0.876953125
],
[
0.118408203125,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.5146484375,
0.87890625,
0.88671875,
0.87890625,
0.88671875,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.139404296875,
0.89453125,
0.387939453125,
0.8935546875,
0.387939453125,
0.90771484375,
0.139404296875,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"Relation Extraction with Explanation",
"Hamed Shahbazi, Xiaoli Z. Fern, Reza Ghaeini, Prasad Tadepalli",
"School of Electrical Engineering and Computer Science, Oregon State University",
"Corvallis, OR, USA",
"[shahbazh, xfern, ghaeinim, tadepall)@oregonstate.edu",
"Abstract",
"data (Pershina et al., 2014; Angeli et al., 2014; Belt-",
"agy et al., 2019). These studies primarily focus on",
"Recent neural models for relation extraction",
"improving relation extraction accuracy and little is",
"22 8",
"with distant supervision alleviate the impact of",
"known about whether the models are making right",
"irrelevant sentences in abag bylearning impor-",
"tance weights for the sentences. Efforts thus",
"decision for the right reason or because of some",
"far havefocused on improving extraction ac",
"irrelevant biases (Agrawal et al., 2016; Gururangan",
"curacy but little is known about their explain-",
"et al., 2018; Ghaeini et al., 2019).",
"ability.In this work we annotate a test set",
"This paper examines two strong baseline relation",
"with ground-truth sentence-level explanations",
"extraction models with several explanation mecha-",
"to evaluate the quality of explanations afforded",
"nisms.Wemanuallyannotatedatestsetfromthe",
"by the relation extraction models. We demon",
"widely used FB-NYT dataset with ground truth ex-",
"strate that replacing the entity mentions in the",
"['s]",
"sentences with their fine-grained entity types",
"planations to evaluate the quality of the explanation",
"not only enhances extraction accuracybut also",
"afforded by these models.We also introduce two",
"improves explanation. We also propose to au-",
"different methods for improving relation extrac",
"tomatically generate \"distractor\" sentences to",
"tion. First, we demonstrate that replacing the entity",
"augment the bags and train the model to ig-",
"mentions with their fine-grained entity types for",
"A",
"norethedistractors.Evaluations onthewidely",
"sentence representation leads to improvement in",
"used FB-NYT dataset show that our methods",
"both the extract accuracy and model explainability.",
"achieve new state-of-the-art accuracy while",
"Second, we augment the bags with automatically",
"improvingmodelexplainability.",
"generated “\"distractor\" sentences (i.e., sentences",
"Introduction",
"that contain no supporting information for the re-",
"",
"",
"lation) and train the model to appropriately ignore",
"Relation extraction with distant supervision asso-",
"the irrelevant information. Our evaluation on the",
"ciates a pair of entities with a bag of sentences,",
"widely used FB-NYT dataset verifies that the pro-",
"each containing mentions of both entities. The",
"posed methods achieve the new state of the art for",
"bag is tagged with relations between the pair in",
"the extraction performance along with improved",
"a Knowledge Base (KB), without explicitly indi-",
"model explainability.",
"cating which sentence(s) support the relation(s).",
"This method avoids the burden of manual anno-",
"2Problem Setup",
"tations, but presents inherent ambiguity, creating",
"challenges for learning.",
"Given entity pair (ei,ej), we form a bag Bi,j ",
"To alleviate the impact of the irrelevant sentences",
"[$1, . SN, with Ni, sentences that contain men-",
"many approaches have been proposed including",
"tions of both entities and label it by the set of rela-",
"models based on attention (Zeng et al., 2015; Lin",
"tions between ei and e; from the KB. Neural mod-",
"et al., 2016; Liu et al., 2017; Luo et al., 2017; Du",
"els for relation extraction encode each sentences",
"et al., 2018; Wang et al., 2018; Peng and Denilson,",
"into a vector representation and a bag Bi,j is thus",
"2019; Bai and Ritter, 2019), approaches that use",
"represented by [ac1,... aN, where c; e Rd.",
"additional resources (Vashishth et al., 2018; Liu",
"Given a set of bags and the associated labels,",
"et al., 2018) and methods that utilize supervision",
"the training objective is to learn a model that pre-",
"'Accepted as a short paper at ACL 2020",
"dicts the probability P(r = kBi.) that relation"
] |
{
"arxiv_id": "2005.14271",
"language": "en",
"page_idx": 0
}
|
||
2005.14271-en-1
|
2005.14271-en-1.jpg
|
$k\in1\ldots K$ and $K$ is the total number of relations in the KB. There are zero to multiple possible relation labels for each bag. Importantly, only some sentences in the bag express any of the relations and the others are irrelevant (provide no information regarding the relations), but such sentences are not labeled.
### 3 $\textbf{Baseline Models}$
We consider two baselines. The first is DirectSup, a recent model achieving the state-of-the-art performance by utilizing auxiliary supervision (Beltagy et al., 2019). The second baseline (CNNs+ATT) revamps the classic attention based method by Lin et al. (2016) but adopts the same sentence encoder as DirectSup for ease of comparisons. In this work, we add a ReLU at the end of the sentence encoder (Beltagy et al., 2019) to produce positive sentence representations. See (Beltagy et al.,2019) for detailed information regarding the sentence encoder. DirectSup. Given a bag of sentences, DirectSup encodes each sentence using CNNs with different filter sizes. The outputs of the CNNs with different filter sizes are concatenated to produce the encoding of the sentence.
Given a bag $B$ and the encoding of its sentences $\{x_1,x_2,...,x_N\}$,DirectSup assigns an importance weight for each sentence based on the output of a binary classifier learned from an additional direct supervision data in a multi-task manner. Given a sentence encoding $x_n$, the binary classifier provides a weight $\alpha_n\in[0,1]$ indicating the likelihood that $x_n$ expresses some form of relations in the KB. As a result, for a bag $B_i,j$, we have importance weights $\{\alpha_1,\ldots,\alpha_N\}.$ It then produces a single bag representation as follows:
$$\bar x=\text{Max-pool}(\{\alpha_1x_1,\dots,\alpha_nx_N\})$$
(1)
and the prediction for relation $k$ is given by:
$$P(r=k|B)=\sigma(\bar{x}\cdotp r_k+b_k)$$
(2)
where $r_k$ is an embedding of relation $k,b_k$ is abias
variable and $\sigma$ is the Sigmoid function.
CNNs+ATT. This model uses the same sentence encoder as DirectSup but differs in the attention mechanism used to decide sentence importance. Specifically, it follows Lin et al. (2016) and computes the importance weights of the sentences in
$k$ exists between $e_i$ and $e_j$ based on $B_{i,j}$, where bag $B$ with encodings $\{x_1,\ldots,x_N\}$ as follows:
(3)
$$\alpha_{k,n}=\frac{\exp(x_nAq_k)}{\sum_{i=1}^N\exp(x_iAq_k)}$$
where $q_k$ is a learned query vector associated with
relation $k$ and $A$ is a diagonal matrix.
Given $\{\alpha_{k,1},...,\alpha_{k,N}\}$,we compute a bag repre-
sentation specific for relation $k$ by:
$$\bar{x}_k=\sum_{n=1}^N\alpha_{k,n}x_n$$
(4)
and the prediction for relation $k$ is given by:
$$P(r=k|B)=\sigma(\bar{x}_k\cdotp r_k+b_k)$$
(5)
where $r_k$ is relation $k$'s embedding and $b_k$ is the
bias.
Entity embedding. Prior work has demonstrated that incorporating entity embeddings into the relation extraction model leads to improved accuracy (Ji et al., 2017; Beltagy et al., 2019). Here we also consider this strategy with the baseline models. Specifically, let $v_i$ and $v_j$ be the entity embedding of $e_i$ and $e_j$, we concatenate the bag representations $\bar{x}$ with $v_i-v_j$ and $v_i\circ v_j$,where $\circ$ is element-wise product. We then apply a linear project layer with ReLU to produce a new bag representation for final prediction with Eq.2 and 5.
For any entity $e_i$ its embedding vector $v_i$ is obtained by concatenating the average of its skipgram (Mikolov et al., 2013) word embeddings and the embeddings produced by Zhang et al. (2019) (produced by using TransE on Wikipedia factual tuples).
Training objective. For all the models in this work we use the binary cross entropy loss function for training:
(6)
$$l=-\sum_{B_{i,j}}\sum_{k=1}^{K}\mathbb{1}_{i,j,k}\log P(r=k|B_{i,j})+\\(1-\mathbb{1}_{i,j,k})\log(1-P(r=k|B_{i,j}))$$
where $1_{i,j,k}$ is an indicator function that takes value
1 if relation $k$ exists for bag $B_i,j.$
## 4 Explanation Mechanisms
The importance weights ( $\alpha’$s, aka attention), generated by the models can be interpreted as explanations. However, recent studies (Ghaeini et al., 2018; Jain et al., 2019; Wiegreffe and Pinter, 2019) have
|
[
[
0.12005615234375,
0.07659912109375,
0.490234375,
0.07659912109375,
0.490234375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07659912109375,
0.869140625,
0.07659912109375,
0.869140625,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.118408203125,
0.0914306640625,
0.48779296875,
0.0914306640625,
0.48779296875,
0.1051025390625,
0.118408203125,
0.1051025390625
],
[
0.1175537109375,
0.10687255859375,
0.489501953125,
0.1092529296875,
0.489501953125,
0.122314453125,
0.1175537109375,
0.12054443359375
],
[
0.67919921875,
0.10272216796875,
0.77490234375,
0.10272216796875,
0.77490234375,
0.1175537109375,
0.67919921875,
0.1175537109375
],
[
0.599609375,
0.114013671875,
0.640625,
0.114013671875,
0.640625,
0.127685546875,
0.599609375,
0.127685546875
],
[
0.8564453125,
0.10986328125,
0.88671875,
0.10986328125,
0.88671875,
0.1270751953125,
0.8564453125,
0.1270751953125
],
[
0.6572265625,
0.1181640625,
0.796875,
0.1181640625,
0.796875,
0.14013671875,
0.6572265625,
0.14013671875
],
[
0.118408203125,
0.12469482421875,
0.489501953125,
0.12469482421875,
0.489501953125,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.118408203125,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.118408203125,
0.155029296875
],
[
0.5146484375,
0.149658203125,
0.88330078125,
0.149658203125,
0.88330078125,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.12005615234375,
0.15673828125,
0.490234375,
0.15673828125,
0.490234375,
0.17041015625,
0.12005615234375,
0.17041015625
],
[
0.51318359375,
0.166259765625,
0.7900390625,
0.166259765625,
0.7900390625,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.12005615234375,
0.174560546875,
0.48779296875,
0.174560546875,
0.48779296875,
0.185302734375,
0.12005615234375,
0.185302734375
],
[
0.53076171875,
0.1817626953125,
0.88330078125,
0.1817626953125,
0.88330078125,
0.196533203125,
0.53076171875,
0.196533203125
],
[
0.1158447265625,
0.188232421875,
0.20654296875,
0.1859130859375,
0.20654296875,
0.20068359375,
0.1158447265625,
0.203125
],
[
0.51123046875,
0.196533203125,
0.771484375,
0.19775390625,
0.771484375,
0.2120361328125,
0.51123046875,
0.21142578125
],
[
0.1175537109375,
0.2197265625,
0.291259765625,
0.2197265625,
0.291259765625,
0.2340087890625,
0.1175537109375,
0.2340087890625
],
[
0.68603515625,
0.2257080078125,
0.7001953125,
0.2257080078125,
0.7001953125,
0.235107421875,
0.68603515625,
0.235107421875
],
[
0.67236328125,
0.2369384765625,
0.76171875,
0.2410888671875,
0.759765625,
0.254638671875,
0.6708984375,
0.25048828125
],
[
0.85791015625,
0.2381591796875,
0.884765625,
0.2381591796875,
0.884765625,
0.251708984375,
0.85791015625,
0.251708984375
],
[
0.1175537109375,
0.2464599609375,
0.489501953125,
0.2476806640625,
0.489501953125,
0.262451171875,
0.1175537109375,
0.26123046875
],
[
0.634765625,
0.240478515625,
0.67333984375,
0.240478515625,
0.67333984375,
0.251708984375,
0.634765625,
0.251708984375
],
[
0.1175537109375,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.68115234375,
0.25830078125,
0.705078125,
0.25830078125,
0.705078125,
0.265380859375,
0.68115234375,
0.265380859375
],
[
0.118408203125,
0.28076171875,
0.489501953125,
0.28076171875,
0.489501953125,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51123046875,
0.276611328125,
0.8388671875,
0.27783203125,
0.8388671875,
0.292236328125,
0.51123046875,
0.29150390625
],
[
0.118408203125,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.58447265625,
0.305908203125,
0.81005859375,
0.305908203125,
0.81005859375,
0.319580078125,
0.58447265625,
0.319580078125
],
[
0.118408203125,
0.312255859375,
0.489501953125,
0.312255859375,
0.489501953125,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.85791015625,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.85791015625,
0.320556640625
],
[
0.12005615234375,
0.328857421875,
0.489501953125,
0.328857421875,
0.489501953125,
0.342529296875,
0.12005615234375,
0.342529296875
],
[
0.5146484375,
0.3349609375,
0.88330078125,
0.3349609375,
0.88330078125,
0.3486328125,
0.5146484375,
0.3486328125
],
[
0.12005615234375,
0.344482421875,
0.490234375,
0.344482421875,
0.490234375,
0.358154296875,
0.12005615234375,
0.358154296875
],
[
0.5146484375,
0.352783203125,
0.5498046875,
0.352783203125,
0.5498046875,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.12005615234375,
0.361083984375,
0.490234375,
0.361083984375,
0.490234375,
0.374755859375,
0.12005615234375,
0.374755859375
],
[
0.51318359375,
0.36767578125,
0.88330078125,
0.36767578125,
0.88330078125,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.1175537109375,
0.377197265625,
0.489501953125,
0.3759765625,
0.489501953125,
0.39013671875,
0.1175537109375,
0.391357421875
],
[
0.51123046875,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.12005615234375,
0.393798828125,
0.489501953125,
0.393798828125,
0.489501953125,
0.40673828125,
0.12005615234375,
0.40673828125
],
[
0.51318359375,
0.3984375,
0.8818359375,
0.399658203125,
0.8818359375,
0.41455078125,
0.51318359375,
0.413330078125
],
[
0.12005615234375,
0.4091796875,
0.489501953125,
0.4091796875,
0.489501953125,
0.4228515625,
0.12005615234375,
0.4228515625
],
[
0.51318359375,
0.416259765625,
0.88330078125,
0.416259765625,
0.88330078125,
0.429931640625,
0.51318359375,
0.429931640625
],
[
0.118408203125,
0.42578125,
0.489501953125,
0.424560546875,
0.489501953125,
0.439453125,
0.118408203125,
0.440673828125
],
[
0.5146484375,
0.432861328125,
0.884765625,
0.432861328125,
0.884765625,
0.446533203125,
0.5146484375,
0.446533203125
],
[
0.12005615234375,
0.4423828125,
0.489501953125,
0.4423828125,
0.489501953125,
0.4560546875,
0.12005615234375,
0.4560546875
],
[
0.51123046875,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.51123046875,
0.4619140625
],
[
0.12005615234375,
0.457763671875,
0.490234375,
0.457763671875,
0.490234375,
0.471435546875,
0.12005615234375,
0.471435546875
],
[
0.51318359375,
0.465576171875,
0.88330078125,
0.465576171875,
0.88330078125,
0.478515625,
0.51318359375,
0.478515625
],
[
0.12005615234375,
0.474365234375,
0.490234375,
0.474365234375,
0.490234375,
0.488037109375,
0.12005615234375,
0.488037109375
],
[
0.51318359375,
0.48095703125,
0.88330078125,
0.48095703125,
0.88330078125,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.1175537109375,
0.490478515625,
0.264404296875,
0.4892578125,
0.264404296875,
0.50439453125,
0.1175537109375,
0.5048828125
],
[
0.51318359375,
0.496337890625,
0.88330078125,
0.496337890625,
0.88330078125,
0.51123046875,
0.51318359375,
0.51123046875
],
[
0.1368408203125,
0.50732421875,
0.490234375,
0.50732421875,
0.490234375,
0.521484375,
0.1368408203125,
0.521484375
],
[
0.51123046875,
0.51123046875,
0.88330078125,
0.5126953125,
0.88330078125,
0.52685546875,
0.51123046875,
0.525390625
],
[
0.118408203125,
0.52392578125,
0.490234375,
0.5224609375,
0.490234375,
0.53759765625,
0.118408203125,
0.5380859375
],
[
0.51123046875,
0.529296875,
0.7197265625,
0.52685546875,
0.7197265625,
0.54150390625,
0.51123046875,
0.54345703125
],
[
0.12005615234375,
0.54150390625,
0.489501953125,
0.54150390625,
0.489501953125,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.53076171875,
0.54443359375,
0.88330078125,
0.54443359375,
0.88330078125,
0.55859375,
0.53076171875,
0.55859375
],
[
0.118408203125,
0.5556640625,
0.489501953125,
0.5556640625,
0.489501953125,
0.5693359375,
0.118408203125,
0.5693359375
],
[
0.51318359375,
0.56103515625,
0.88330078125,
0.56103515625,
0.88330078125,
0.57470703125,
0.51318359375,
0.57470703125
],
[
0.1175537109375,
0.57177734375,
0.490234375,
0.57080078125,
0.490234375,
0.5849609375,
0.1175537109375,
0.5859375
],
[
0.5146484375,
0.57666015625,
0.88330078125,
0.57666015625,
0.88330078125,
0.59033203125,
0.5146484375,
0.59033203125
],
[
0.118408203125,
0.58837890625,
0.490234375,
0.58837890625,
0.490234375,
0.6015625,
0.118408203125,
0.6015625
],
[
0.5146484375,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.5146484375,
0.60693359375
],
[
0.1175537109375,
0.60400390625,
0.489501953125,
0.6025390625,
0.489501953125,
0.61767578125,
0.1175537109375,
0.6181640625
],
[
0.5146484375,
0.609375,
0.88330078125,
0.609375,
0.88330078125,
0.62353515625,
0.5146484375,
0.62353515625
],
[
0.1175537109375,
0.62060546875,
0.489501953125,
0.619140625,
0.489501953125,
0.63427734375,
0.1175537109375,
0.63525390625
],
[
0.5146484375,
0.626953125,
0.5693359375,
0.626953125,
0.5693359375,
0.63818359375,
0.5146484375,
0.63818359375
],
[
0.1175537109375,
0.63525390625,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.1175537109375,
0.6494140625
],
[
0.51318359375,
0.64111328125,
0.8818359375,
0.64111328125,
0.8818359375,
0.65478515625,
0.51318359375,
0.65478515625
],
[
0.118408203125,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.118408203125,
0.6650390625
],
[
0.51318359375,
0.65771484375,
0.88330078125,
0.65771484375,
0.88330078125,
0.67138671875,
0.51318359375,
0.67138671875
],
[
0.1175537109375,
0.66845703125,
0.341796875,
0.66748046875,
0.341796875,
0.681640625,
0.1175537109375,
0.68310546875
],
[
0.51318359375,
0.67333984375,
0.57861328125,
0.67333984375,
0.57861328125,
0.6884765625,
0.51318359375,
0.6884765625
],
[
0.16796875,
0.69873046875,
0.435791015625,
0.70068359375,
0.435791015625,
0.71533203125,
0.16796875,
0.71337890625
],
[
0.46337890625,
0.70068359375,
0.489501953125,
0.70068359375,
0.489501953125,
0.71435546875,
0.46337890625,
0.71435546875
],
[
0.623046875,
0.70166015625,
0.6396484375,
0.70166015625,
0.6396484375,
0.71044921875,
0.623046875,
0.71044921875
],
[
0.61279296875,
0.71142578125,
0.84033203125,
0.71337890625,
0.84033203125,
0.73095703125,
0.61279296875,
0.728515625
],
[
0.587890625,
0.732421875,
0.6455078125,
0.732421875,
0.6455078125,
0.74365234375,
0.587890625,
0.74365234375
],
[
0.85791015625,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.7392578125,
0.85791015625,
0.7392578125
],
[
0.118408203125,
0.73291015625,
0.443359375,
0.73291015625,
0.443359375,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.568359375,
0.7490234375,
0.84033203125,
0.7490234375,
0.84033203125,
0.7666015625,
0.568359375,
0.7666015625
],
[
0.19482421875,
0.76318359375,
0.411376953125,
0.76416015625,
0.411376953125,
0.779296875,
0.19482421875,
0.77783203125
],
[
0.46337890625,
0.765625,
0.489501953125,
0.765625,
0.489501953125,
0.77783203125,
0.46337890625,
0.77783203125
],
[
0.5146484375,
0.77685546875,
0.884765625,
0.77685546875,
0.884765625,
0.79150390625,
0.5146484375,
0.79150390625
],
[
0.1175537109375,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.81103515625,
0.1175537109375,
0.81103515625
],
[
0.51318359375,
0.7919921875,
0.75732421875,
0.79443359375,
0.75732421875,
0.80859375,
0.51318359375,
0.80712890625
],
[
0.1175537109375,
0.8125,
0.405517578125,
0.8134765625,
0.405517578125,
0.82763671875,
0.1175537109375,
0.82666015625
],
[
0.51904296875,
0.82080078125,
0.53564453125,
0.82666015625,
0.52880859375,
0.8349609375,
0.51318359375,
0.8291015625
],
[
0.5341796875,
0.826171875,
0.54150390625,
0.826171875,
0.54150390625,
0.830078125,
0.5341796875,
0.830078125
],
[
0.53662109375,
0.82177734375,
0.76416015625,
0.82177734375,
0.76416015625,
0.8349609375,
0.53662109375,
0.8349609375
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.118408203125,
0.8466796875,
0.491943359375,
0.8466796875,
0.491943359375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51123046875,
0.845703125,
0.88671875,
0.8466796875,
0.88671875,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.118408203125,
0.86328125,
0.495361328125,
0.86328125,
0.495361328125,
0.876953125,
0.118408203125,
0.876953125
],
[
0.51318359375,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.51318359375,
0.876953125
],
[
0.12005615234375,
0.87890625,
0.49365234375,
0.87890625,
0.49365234375,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.51318359375,
0.87890625,
0.884765625,
0.87890625,
0.884765625,
0.892578125,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.8955078125,
0.49365234375,
0.8955078125,
0.49365234375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.89453125,
0.884765625,
0.89453125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"k exists between ei and ej based on Bi.j, where",
"bag B with encodings &i,..., &n as follows:",
"k E 1 . .. K and K is the total number of relations",
"in the KB.There are zero tomultiple possible re-",
"exp(anAqk)",
"Qk,n",
"(3)",
"Erexp(riAqk)",
"lation labels for each bag. Importantly, only some",
"sentences in the bag express any of the relations",
"where qk is a learned query vector associated with",
"and the others are irrelevant (provide no informa-",
"relation k and A is a diagonal matrix.",
"tionregardingtherelations),but such sentences are",
"Given [qQk,1, .., Qk,N ], we compute a bag repre-",
"not labeled.",
"sentation specific for relation k by:",
"3Baseline Models",
"N",
"ak,nan",
"(4)",
"We consider two baselines. The first is DirectSup,",
"Ck=",
"a recent model achieving the state-of-the-art perfor-",
"n=1",
"mance by utilizing auxiliary supervision (Beltagy",
"and the prediction for relation k is given by:",
"et al., 2019). The second baseline (CNNs+ATT)",
"P(r = kB) = o(k:rk+bk)",
"revamps the classic attention based method by Lin",
"(5)",
"et al. (2016) but adopts the same sentence encoder",
"where rk is relation k's embedding and bk is the",
"as DirectSup for ease of comparisons. In this work,",
"bias.",
"we add a ReLU at the end of the sentence encoder",
"Entity embedding. Prior work has demonstrated",
"(Beltagy et al., 2019) to produce positive sentence",
"that incorporating entity embeddings into the rela",
"representations. See (Beltagy et al., 2019) for de-",
"tion extraction model leads to improved accuracy",
"tailed information regarding the sentence encoder.",
"(Ji et al., 2017; Beltagy et al., 2019). Here we",
"DirectSup. Given a bag of sentences, DirectSup",
"also consider this strategy with the baseline models.",
"encodes each sentence using CNNs with different",
"Specifically, let vi and v, be the entity embedding",
"filter sizes. The outputs of the CNNs with different",
"of ei and ej,we concatenate the bag representations",
"filter sizes are concatenated to produce the encod-",
" with Vi - Uj and vi o Vj, where o is element-wise",
"ing of the sentence.",
"product. We then apply a linear project layer with",
"Given a bag B and the encoding of its sentences",
"ReLU to produce a new bag representation for final",
"[i,&2,...,],DirectSup assigns an importance",
"prediction with Eq.2 and 5.",
"weightfor eachsentencebased onthe output ofa",
"For any entity e; its embedding vector w, is ob-",
"binary classifier learned from an additional direct",
"tained by concatenating the average of its skip-",
"supervision data in a multi-task manner. Given a",
"gram (Mikolov et al., 2013) word embeddings and",
"sentence encoding &n,the binary classifier provides",
"the embeddings produced by Zhang et al. (2019)",
"a weight αn E [O, 1] indicating the likelihood that",
"(produced by using TransE on Wikipedia factual",
"&n expresses some form of relations in the KB.",
"tuples).",
"As a result, for a bag Bi.j, we have importance",
"Training objective. For all the models in this work",
"weights fal...., al. It then produces a single",
"we use the binary cross entropy loss function for",
"bag representation as follows:",
"training:",
" =Max-pool({α1r1,...,Qnan))",
"(1)",
"K",
" lij,h log P(r = kBi,j)+",
"Bi. k=1",
"(6)",
"and the prediction for relation k is given by:",
"(1 - li,j,h) log (1 - P(r = k|Bi,5)",
"P(r = k|B) =o(.rk +bk)",
"(2)",
"where Ii.j.k. is an indicator function that takes value",
"where rk is an embedding of relation k, bk is a bias",
"1 if relation k exists for bag Bi,j.",
"variable and is the Sigmoid function.",
"A",
"",
"ExplanationMechanisms",
"CNNs+ATT. This model uses the same sentence",
"encoder as DirectSup but differs in the attention",
"The importance weights (α's, aka attention), gener-",
"mechanism used to decide sentence importance.",
"ated by the models can be interpreted as explana-",
"Specifically, it follows Lin et al. (2016) and com-",
"tions. However, recent studies (Ghaeini et al., 2018;",
"putes the importance weights of the sentences in",
"Jain et al., 2019; Wiegreffe and Pinter, 2019) have"
] |
{
"arxiv_id": "2005.14271",
"language": "en",
"page_idx": 1
}
|
||
2005.14271-en-2
|
2005.14271-en-2.jpg
|
questioned the validity of attention as a faithful explanation of model's behavior. Thus we consider the following additional explanation mechanisms: Saliency. Recent works show that a model's prediction can be explained by examining the input saliency, based on the gradient of the output w.r.t. the inputs (Simonyan et al., 2012; Ross et al., 2017; Ghaeini et al., 2019). We define the saliency of sentence $n$ for relation $k$, denoted by $S_x_{n,k}$, as the L1 norm of the gradient of relation $k$ logit $o_k$ with respect to $x_n.($Appendix. A.1).
Gradient$\times \textbf{input. This is a commonly used mea- }$ sure for input attributions (Shrikumar et al., 2016; Selvaraju et al., 2019). We will refer to this measure as $GI_{x_n,k}$,computed as $\sum_ix_n[i]\times\frac{\partial o_k}{\partial x_n}[i].$ Leave One Out (loo). This measures the sensitivity of $o_k$ to the removal of a sentence. We refer to this measure as $loo_{x_n,k}=(o_k-o_{k,-n})$,where $o_{k,-n}$ is the new logit of relation $k$ after removing sentence $x_n$ from its bag.
# 5 Proposed Methods
We propose two different approaches for improving relation extraction. The first method we propose, introduces a subtle change to the representation of the sentences, which lead to higher performance and better explanation quality. We further propose to automatically generate “distractor” sentences and train the model to appropriately ignore them. Sentence representation. Each sentence in a bag contains entity mentions $m_i$ and $m_j$ for entities $e_i$ and $e_j$ respectively. In prior work $m_i$ and $m_j$ are kept unchanged (Lin et al., 2016; Beltagy et al., 2019). We argue that when entity mentions are used to compute the sentence representation, they provide such rich information that the model may not need to look at the rest of the sentence to deduce a relation. To ensure that our predictions are supported by appropriate sentences, we need to remove this effect. We propose to replace the entity mentions with their Fine-Grained Entity Types (FGET) Ling and Weld (2012) to force the model to identify the relations through the sentences.
Learning from distractors. Prior work studied learning from human provided rationales (Lei et al., 2016; Ross et al., 2017; Bao et al., 2018; Ghaeini et al., 2019) in order to improve model explainability. However, human rationales are expensive to acquire. In this work we propose to learn from automatically generated“distractor”sentences.
Let $B_{i,j}$ be a positive training bag (contains at
least one relation) with entities $(e_i,e_j)$ of FGET $(t_{i},t_{j}).$ Let $R_ij(|R_{ij}|>1)$ be the set of annotated relations for $B_{i,j}.$ For each $k$ in $R_ij$, we sample a “distractor" sentence $s_k^{\prime}$ from the set of sentences in the training set such that 1) it belongs to a bag whose FGET is $(t_i,t_j)$ 2) the bag is not annotated with relation label $k.$ If $s_k^\prime$ is not found this way, we simply choose a random sentence from a random negative bag (bag with no relation). Given $s_k^{\prime}$,we replace its entity mentions with $e_i$ and $e_j$ (or $t_i$ and $t_j$ for FGET-based sentence representation) of a sentence in $B_{i,j}$ and add it to the bag, resulting in an augmented bag $B_{i,j}^{\prime}$ for relation $k.$
To learn from the augmented bags, we feed $B_i,j^{\prime}$ into the model and the goal is to lower the contribution of the distractor sentence in relation to the original sentences in the bag. Specifically, we use $GI$ to measure the sentence-level contribution and define the distractor loss for relation $k$ as follows:
$$\begin{aligned}l_{d,k}^{\prime}=\max(0,\gamma+GI_{x_{k}^{\prime},k}-\max_{x\in B_{i,j}}GI_{x,k})\\+|GI_{x_{k}^{\prime},k}|\end{aligned}$$
(7)
where $x_k^\prime$ is the encoding of distractor sentence $s_k^\prime$ and $\gamma$ is a hyper-parameter for margin. The first term ensures that the contribution of the distractor is lower than the maximum contribution of all the sentences in the original bag and the second term reduces the absolute contribution of the distractor. Although we use $GI$ in Eq.7, other explanation measures such as saliency or the positive portion of the contributions can also be applied here. Moreover a more advanced mechanism for generating distractors will likely lead to a higher performance.
We hence update the loss in Eq. 6 with:
(8)
$$l_m=l+\lambda l_d'$$
where $l_d^{\prime}=\sum_kl_{d,k}^{\prime}$ and $\lambda$ tradeoffs the regular
learning loss with the distractor loss.
### 6 Experiments
In this section, we empirically evaluate our proposed methods both in terms of their relation extraction performance and their explainability.
## 6.1 Dataset and Setup
Dataset. Similar to our baselines and prior work, we use the modified version of FB-NYT dataset. The original FB-NYT dataset was built by Riedel et al. (2010) on New York Times articles which was aligned to Freebase facts. It later was modified
|
[
[
0.12005615234375,
0.07659912109375,
0.490234375,
0.07659912109375,
0.490234375,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.5146484375,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.12005615234375,
0.0926513671875,
0.490234375,
0.0926513671875,
0.490234375,
0.105712890625,
0.12005615234375,
0.105712890625
],
[
0.51123046875,
0.0926513671875,
0.88330078125,
0.0914306640625,
0.88330078125,
0.105712890625,
0.51123046875,
0.10687255859375
],
[
0.12005615234375,
0.10986328125,
0.486083984375,
0.10986328125,
0.486083984375,
0.12054443359375,
0.12005615234375,
0.12054443359375
],
[
0.51123046875,
0.10687255859375,
0.88671875,
0.105712890625,
0.88671875,
0.12353515625,
0.51123046875,
0.12469482421875
],
[
0.118408203125,
0.12469482421875,
0.491943359375,
0.12469482421875,
0.491943359375,
0.1383056640625,
0.118408203125,
0.1383056640625
],
[
0.51318359375,
0.12469482421875,
0.88330078125,
0.12469482421875,
0.88330078125,
0.1383056640625,
0.51318359375,
0.1383056640625
],
[
0.118408203125,
0.14013671875,
0.489501953125,
0.14013671875,
0.489501953125,
0.15380859375,
0.118408203125,
0.15380859375
],
[
0.51123046875,
0.1395263671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.155029296875,
0.51123046875,
0.15380859375
],
[
0.12005615234375,
0.157958984375,
0.491943359375,
0.157958984375,
0.491943359375,
0.171630859375,
0.12005615234375,
0.171630859375
],
[
0.51123046875,
0.15673828125,
0.88330078125,
0.15673828125,
0.88330078125,
0.17041015625,
0.51123046875,
0.17041015625
],
[
0.118408203125,
0.17333984375,
0.49365234375,
0.17333984375,
0.49365234375,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.509765625,
0.171630859375,
0.88330078125,
0.1728515625,
0.88330078125,
0.18701171875,
0.509765625,
0.1859130859375
],
[
0.1175537109375,
0.188232421875,
0.490234375,
0.189453125,
0.490234375,
0.2037353515625,
0.1175537109375,
0.203125
],
[
0.51318359375,
0.189453125,
0.884765625,
0.189453125,
0.884765625,
0.203125,
0.51318359375,
0.203125
],
[
0.1175537109375,
0.204833984375,
0.490234375,
0.2060546875,
0.490234375,
0.2203369140625,
0.1175537109375,
0.2197265625
],
[
0.51318359375,
0.204833984375,
0.884765625,
0.204833984375,
0.884765625,
0.2197265625,
0.51318359375,
0.2197265625
],
[
0.118408203125,
0.221435546875,
0.490234375,
0.221435546875,
0.490234375,
0.236328125,
0.118408203125,
0.236328125
],
[
0.51123046875,
0.2197265625,
0.884765625,
0.2203369140625,
0.884765625,
0.235107421875,
0.51123046875,
0.2340087890625
],
[
0.118408203125,
0.2381591796875,
0.349365234375,
0.2381591796875,
0.349365234375,
0.2529296875,
0.118408203125,
0.2529296875
],
[
0.51123046875,
0.2369384765625,
0.884765625,
0.236328125,
0.884765625,
0.25048828125,
0.51123046875,
0.251708984375
],
[
0.118408203125,
0.254638671875,
0.491943359375,
0.254638671875,
0.491943359375,
0.268310546875,
0.118408203125,
0.268310546875
],
[
0.51318359375,
0.254150390625,
0.884765625,
0.254150390625,
0.884765625,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.118408203125,
0.26953125,
0.489501953125,
0.26953125,
0.489501953125,
0.283203125,
0.118408203125,
0.283203125
],
[
0.51318359375,
0.26953125,
0.791015625,
0.26953125,
0.791015625,
0.283203125,
0.51318359375,
0.283203125
],
[
0.12005615234375,
0.2861328125,
0.490234375,
0.2861328125,
0.490234375,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.53076171875,
0.28515625,
0.87890625,
0.28515625,
0.87890625,
0.2998046875,
0.53076171875,
0.2998046875
],
[
0.1175537109375,
0.2998046875,
0.474365234375,
0.2998046875,
0.474365234375,
0.320556640625,
0.1175537109375,
0.320556640625
],
[
0.51318359375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.118408203125,
0.318359375,
0.490234375,
0.318359375,
0.490234375,
0.33203125,
0.118408203125,
0.33203125
],
[
0.51123046875,
0.31640625,
0.880859375,
0.31640625,
0.880859375,
0.330078125,
0.51123046875,
0.330078125
],
[
0.118408203125,
0.3349609375,
0.490234375,
0.3349609375,
0.490234375,
0.3486328125,
0.118408203125,
0.3486328125
],
[
0.51318359375,
0.334228515625,
0.8818359375,
0.334228515625,
0.8818359375,
0.347412109375,
0.51318359375,
0.347412109375
],
[
0.118408203125,
0.352783203125,
0.489501953125,
0.352783203125,
0.489501953125,
0.366455078125,
0.118408203125,
0.366455078125
],
[
0.5146484375,
0.349853515625,
0.8818359375,
0.349853515625,
0.8818359375,
0.363525390625,
0.5146484375,
0.363525390625
],
[
0.118408203125,
0.3681640625,
0.489501953125,
0.36767578125,
0.489501953125,
0.3818359375,
0.118408203125,
0.383056640625
],
[
0.5146484375,
0.366455078125,
0.87744140625,
0.366455078125,
0.87744140625,
0.380126953125,
0.5146484375,
0.380126953125
],
[
0.118408203125,
0.385498046875,
0.307373046875,
0.385498046875,
0.307373046875,
0.3984375,
0.118408203125,
0.3984375
],
[
0.5341796875,
0.3896484375,
0.7421875,
0.39013671875,
0.7421875,
0.407958984375,
0.5341796875,
0.40673828125
],
[
0.744140625,
0.39013671875,
0.8427734375,
0.39013671875,
0.8427734375,
0.405029296875,
0.744140625,
0.405029296875
],
[
0.74462890625,
0.40380859375,
0.79443359375,
0.40380859375,
0.79443359375,
0.41455078125,
0.74462890625,
0.41455078125
],
[
0.1158447265625,
0.410888671875,
0.313232421875,
0.412109375,
0.313232421875,
0.429931640625,
0.1158447265625,
0.4287109375
],
[
0.85986328125,
0.40673828125,
0.88330078125,
0.40673828125,
0.88330078125,
0.419189453125,
0.85986328125,
0.419189453125
],
[
0.76318359375,
0.41748046875,
0.84619140625,
0.419189453125,
0.84619140625,
0.43701171875,
0.76318359375,
0.435302734375
],
[
0.1175537109375,
0.440673828125,
0.489501953125,
0.441162109375,
0.489501953125,
0.4560546875,
0.1175537109375,
0.454833984375
],
[
0.509765625,
0.4453125,
0.88330078125,
0.44775390625,
0.88330078125,
0.465576171875,
0.509765625,
0.463134765625
],
[
0.1158447265625,
0.457275390625,
0.489501953125,
0.457763671875,
0.489501953125,
0.47265625,
0.1158447265625,
0.471435546875
],
[
0.5146484375,
0.465576171875,
0.88330078125,
0.465576171875,
0.88330078125,
0.478515625,
0.5146484375,
0.478515625
],
[
0.1175537109375,
0.474365234375,
0.490234375,
0.474365234375,
0.490234375,
0.488037109375,
0.1175537109375,
0.488037109375
],
[
0.51318359375,
0.48095703125,
0.884765625,
0.48095703125,
0.884765625,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.1175537109375,
0.490478515625,
0.490234375,
0.4892578125,
0.490234375,
0.50439453125,
0.1175537109375,
0.5048828125
],
[
0.5146484375,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5087890625,
0.5146484375,
0.5087890625
],
[
0.118408203125,
0.50732421875,
0.489501953125,
0.50732421875,
0.489501953125,
0.52099609375,
0.118408203125,
0.52099609375
],
[
0.51123046875,
0.51318359375,
0.8818359375,
0.51318359375,
0.8818359375,
0.52685546875,
0.51123046875,
0.52685546875
],
[
0.118408203125,
0.5224609375,
0.490234375,
0.5224609375,
0.490234375,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.51318359375,
0.529296875,
0.88671875,
0.529296875,
0.88671875,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.118408203125,
0.5390625,
0.48193359375,
0.5390625,
0.48193359375,
0.552734375,
0.118408203125,
0.552734375
],
[
0.51318359375,
0.54443359375,
0.88330078125,
0.5458984375,
0.88330078125,
0.56005859375,
0.51318359375,
0.55859375
],
[
0.1175537109375,
0.55419921875,
0.490234375,
0.5546875,
0.490234375,
0.5693359375,
0.1175537109375,
0.568359375
],
[
0.51123046875,
0.56103515625,
0.88330078125,
0.56103515625,
0.88330078125,
0.57470703125,
0.51123046875,
0.57470703125
],
[
0.118408203125,
0.57177734375,
0.490234375,
0.57177734375,
0.490234375,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.5146484375,
0.57763671875,
0.88330078125,
0.57763671875,
0.88330078125,
0.5927734375,
0.5146484375,
0.5927734375
],
[
0.1175537109375,
0.5859375,
0.489501953125,
0.58740234375,
0.489501953125,
0.6015625,
0.1175537109375,
0.60107421875
],
[
0.51318359375,
0.59423828125,
0.88330078125,
0.59423828125,
0.88330078125,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.118408203125,
0.6025390625,
0.489501953125,
0.6025390625,
0.489501953125,
0.6162109375,
0.118408203125,
0.6162109375
],
[
0.51318359375,
0.60986328125,
0.88671875,
0.60986328125,
0.88671875,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.118408203125,
0.6181640625,
0.489501953125,
0.619140625,
0.489501953125,
0.63427734375,
0.118408203125,
0.6328125
],
[
0.5322265625,
0.6259765625,
0.8251953125,
0.6259765625,
0.8251953125,
0.64013671875,
0.5322265625,
0.64013671875
],
[
0.12005615234375,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.12005615234375,
0.6494140625
],
[
0.12005615234375,
0.65087890625,
0.486083984375,
0.65087890625,
0.486083984375,
0.66455078125,
0.12005615234375,
0.66455078125
],
[
0.646484375,
0.65185546875,
0.75146484375,
0.65185546875,
0.75146484375,
0.66943359375,
0.646484375,
0.66943359375
],
[
0.85791015625,
0.65380859375,
0.884765625,
0.65380859375,
0.884765625,
0.66748046875,
0.85791015625,
0.66748046875
],
[
0.12176513671875,
0.66943359375,
0.486083984375,
0.66943359375,
0.486083984375,
0.6796875,
0.12176513671875,
0.6796875
],
[
0.12005615234375,
0.68408203125,
0.489501953125,
0.68408203125,
0.489501953125,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.509765625,
0.6796875,
0.88818359375,
0.6796875,
0.88818359375,
0.70068359375,
0.509765625,
0.70068359375
],
[
0.12005615234375,
0.70068359375,
0.490234375,
0.70068359375,
0.490234375,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.51318359375,
0.69873046875,
0.78515625,
0.69873046875,
0.78515625,
0.7119140625,
0.51318359375,
0.7119140625
],
[
0.118408203125,
0.71630859375,
0.489501953125,
0.71630859375,
0.489501953125,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.51708984375,
0.72998046875,
0.5263671875,
0.72998046875,
0.5263671875,
0.7353515625,
0.51708984375,
0.7353515625
],
[
0.5439453125,
0.72705078125,
0.65576171875,
0.72705078125,
0.65576171875,
0.73828125,
0.5439453125,
0.73828125
],
[
0.118408203125,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.12005615234375,
0.732421875,
0.489501953125,
0.732421875,
0.489501953125,
0.7451171875,
0.12005615234375,
0.7451171875
],
[
0.51318359375,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.765625,
0.51318359375,
0.765625
],
[
0.118408203125,
0.765625,
0.459228515625,
0.765625,
0.459228515625,
0.779296875,
0.118408203125,
0.779296875
],
[
0.5146484375,
0.76708984375,
0.88330078125,
0.76708984375,
0.88330078125,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.118408203125,
0.78076171875,
0.490234375,
0.78076171875,
0.490234375,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.51318359375,
0.78369140625,
0.84619140625,
0.78369140625,
0.84619140625,
0.79736328125,
0.51318359375,
0.79736328125
],
[
0.118408203125,
0.79736328125,
0.489501953125,
0.79736328125,
0.489501953125,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.12005615234375,
0.8134765625,
0.491943359375,
0.8134765625,
0.491943359375,
0.82666015625,
0.12005615234375,
0.82666015625
],
[
0.51318359375,
0.81005859375,
0.7001953125,
0.81005859375,
0.7001953125,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.118408203125,
0.82763671875,
0.495361328125,
0.8291015625,
0.495361328125,
0.84326171875,
0.118408203125,
0.8427734375
],
[
0.51318359375,
0.83056640625,
0.88671875,
0.83056640625,
0.88671875,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.1175537109375,
0.84423828125,
0.495361328125,
0.845703125,
0.495361328125,
0.85986328125,
0.1175537109375,
0.859375
],
[
0.51123046875,
0.8466796875,
0.884765625,
0.845703125,
0.884765625,
0.85986328125,
0.51123046875,
0.86083984375
],
[
0.118408203125,
0.86083984375,
0.49365234375,
0.86083984375,
0.49365234375,
0.8759765625,
0.118408203125,
0.8759765625
],
[
0.51123046875,
0.8623046875,
0.884765625,
0.86083984375,
0.884765625,
0.8759765625,
0.51123046875,
0.876953125
],
[
0.12005615234375,
0.87744140625,
0.468505859375,
0.87744140625,
0.468505859375,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.51318359375,
0.892578125
],
[
0.1351318359375,
0.8935546875,
0.49365234375,
0.8955078125,
0.49365234375,
0.91015625,
0.1351318359375,
0.90771484375
],
[
0.51318359375,
0.8955078125,
0.884765625,
0.8955078125,
0.884765625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"questioned the validity of attention as a faithful ex-",
"least one relation)with entities (ei,e;)of FGET",
"planation of model's behavior. Thus we consider",
"(ti, t;). Let Ri;(|Riil > 1) be the set of annotated",
"thefollowingadditional explanationmechanisms:",
"relations for Bi.j. For each k in Rij, we sample a",
"Saliency. Recent works show that a model's pre-",
"\"distractor\" sentence s' from the set of sentences",
"diction can be explained by examining the input",
"in the training set such that 1) it belongs to a bag",
"saliency, based on the gradient of the output w.r.t.",
"whose FGET is (ti, t,) 2) the bag is not annotated",
"the inputs (Simonyan et al., 2012; Ross et al., 2017;",
"with relation label k. If s', is not found this way, we",
"Ghaeini et al., 2019). We define the saliency of",
"simply choose a random sentence from a random",
"sentence n for relation k, denoted by San.k, as the",
"negative bag (bag with no relation). Given ss, we",
"L1 norm of the gradient of relation k logit ok with",
"replace its entity mentions with e; and e, (or t; and",
"respect to n.(Appendix. A.1).",
"t, for FGET-based sentence representation) of a",
"Gradient x input. This is a commonly used mea-",
"sentence in Bi.j and add it to the bag, resulting in",
"sure for input attributions (Shrikumar et al., 2016:",
"an augmented bag B', , for relation k.",
"Selvaraju et al., 2019). We will refer to this mea-",
"To learn from the augmented bags, we feed B'",
"sure as GIan,k, computed as E; an[i] × o [i].",
"into the model and the goal is to lower the contri",
"Leave One Out (loo). This measures the sensitiv-",
"bution of the distractor sentence in relation to the",
"ity of ok to the removal of a sentence. We refer",
"original sentences in the bag. Specifically, we use",
"to this measure as loorn.h = (ok - Ok.-n), where",
"GI to measure the sentence-level contribution and",
"Ok,-n is the new logit of relation k after removing",
"define the distractor loss for relation k as follows:",
"sentence &n from its bag.",
"la,h= max(O, +GI, -",
"max GIz,k)",
"TEBij",
"5 Proposed Methods",
"(7)",
"+IGI,l",
"Wepropose two different approachesfor improving",
"where c', is the encoding of distractor sentence sh",
"relation extraction. The first method we propose,",
"and is a hyper-parameter for margin.The first",
"introduces a subtle change to the representation of",
"term ensures that the contribution of the distractor",
"the sentences, which lead to higher performance",
"islowerthanthemaximum contributionof allthe",
"and better explanation quality. We further propose",
"sentences in the original bag and the second term",
"to automatically generate \"distractor\" sentences",
"reduces the absolute contribution of the distractor.",
"and train the model to appropriately ignore them.",
"Although we use GI in Eq.7, other explanation",
"Sentence representation. Each sentence in a bag",
"measures such as saliency or the positive portion of",
"contains entity mentions m, and m, for entities e,",
"the contributions can also be applied here. More-",
"and e; respectively. In prior work m, and m, are",
"over a more advanced mechanism for generating",
"kept unchanged (Lin et al., 2016: Beltagy et al.",
"distractors will likely lead to a higher performance.",
"2019). We argue that when entity mentions are",
"We hence update the loss in Eq. 6 with:",
"used to compute the sentence representation, they",
"provide such rich information that the model may",
"Im = l + 入a",
"(8)",
"not need to look at the rest of the sentence to de",
"duce a relation. To ensure that our predictions are",
"where l'a = Z, 'a, and 入 tradeoffs the regular",
"supported by appropriate sentences, we need to",
"learning loss with the distractor loss.",
"remove this effect. We propose to replace the en-",
"h",
"Experiments",
"(FGET) Ling and Weld (2012) to force the model",
"tity mentions with their Fine-Grained Entity Types",
"In this section, we empirically evaluate our pro-",
"to identify the relations through the sentences.",
"posed methods both in terms of their relation ex-",
"Learning from distractors. Prior work studied",
"traction performance and their explainability.",
"learning from human provided rationales (Lei et al.,",
"2016; Ross et al., 2017; Bao et al., 2018; Ghaeini",
"6.1 Dataset and Setup",
"et al., 2019) in order to improve model explain-",
"Dataset. Similar to our baselines and prior work,",
"ability. However, human rationales are expensive",
"we use the modified version of FB-NYT dataset.",
"to acquire. In this work we propose to learn from",
"The original FB-NYT dataset was built by Riedel",
"automatically generated “distractor\" sentences.",
"et al. (2010) on New York Times articles which",
"Let Bi., be a positive training bag (contains at",
"was aligned to Freebase facts. It later was modified"
] |
{
"arxiv_id": "2005.14271",
"language": "en",
"page_idx": 2
}
|
||
2005.14271-en-4
|
2005.14271-en-4.jpg
|
<table>
<tbody>
<tr>
<th>model</th>
<th>$\overline{{\mathrm{loo}}}$ $\overline{(\mathrm{H})}$</th>
<th>$\overline{\log{0}}$ ${\cdot\overline{(\mathrm{L})}}$</th>
<th>$\overline{S}$ $\overline{(\mathrm{H})}$ $TC$</th>
<th>$\overline{S}$ $\overline{(\mathbf{L})}$ $xr$</th>
<th>$\overline{GI}_{:}$ $\overline{(\mathbf{H})}$</th>
<th>$\overline{GI}_{i}$ $\overline{(\mathrm{L})}$ rr</th>
<th>${\overline{\alpha_{x_{n}}(\mathrm{H})}}$</th>
<th>$\overline{(\mathrm{L})}$ $\alpha_{x_n}$</th>
</tr>
<tr>
<td>$\mathbf{CNNs+ATT}$</td>
<td>0.16</td>
<td>-0.08</td>
<td>0.19</td>
<td>-0.02</td>
<td>0.20</td>
<td>0.04</td>
<td>0.69</td>
<td>0.21</td>
</tr>
<tr>
<td>DirectSup</td>
<td>0.19</td>
<td>0.12</td>
<td>0.08</td>
<td>0.15</td>
<td>0.29</td>
<td>0.19</td>
<td>0.26</td>
<td>-0.12</td>
</tr>
<tr>
<td>CNNs+ ATT+ F</td>
<td>$\overline{0.21}$</td>
<td>$\overline{0.10}$</td>
<td>$\overline{0.36}$</td>
<td>$\overline{0.03}$</td>
<td>$\overline{0.23}$</td>
<td>$\overline{0.00}$</td>
<td>$\overline{0.73}$</td>
<td>$\overline{0.11}$</td>
</tr>
<tr>
<td>DirectSup+ F</td>
<td>0.24</td>
<td>0.15</td>
<td>0.31</td>
<td>-0.19</td>
<td>0.40</td>
<td>-0.17</td>
<td>0.28</td>
<td>0.15</td>
</tr>
<tr>
<td>CNNs+ ATT+ FE</td>
<td>0.01</td>
<td>-0.11</td>
<td>0.21</td>
<td>-0.14</td>
<td>0.20</td>
<td>-0.20</td>
<td>0.24</td>
<td>0.01</td>
</tr>
<tr>
<td>DirectSup+ FE</td>
<td>0.14</td>
<td>-.12</td>
<td>0.19</td>
<td>-0.10</td>
<td>0.29</td>
<td>0.06</td>
<td>0.17</td>
<td>-0.11</td>
</tr>
<tr>
<td>$\mathbf{CNNs+ATT+LD}$</td>
<td>0.18</td>
<td>-0.01</td>
<td>0.22</td>
<td>0.10</td>
<td>0.21</td>
<td>0</td>
<td>0.67</td>
<td>0.11</td>
</tr>
<tr>
<td>CNNs+ ATT+ LD+ F</td>
<td>0.22</td>
<td>-0.11</td>
<td>0.43</td>
<td>0.09</td>
<td>0.28</td>
<td>0.07</td>
<td>0.70</td>
<td>0.12</td>
</tr>
<tr>
<td>DirectSup+ LD+ F</td>
<td>0.23</td>
<td>0.14</td>
<td>0.38</td>
<td>0.01</td>
<td>0.49</td>
<td>0.20</td>
<td>0.45</td>
<td>0.02</td>
</tr>
</tbody>
</table>
L: Low confidence $P(r)\in[0,0.25]$
Table 3: Kendall correlations for top confidence and least confidence range.
which we will examine in the next section. Finally the last three rows of the table show that adding LD to different base models can further improve the AUCs.
Similar to prior work, we observe that incorporating entity embeddings(+E) to the model leads to substantial performance gain across the board. We also observe very similar performance gain when adding FGET and LD to the base models both with and without entity embeddings. Our best model achieved an AUC of 0.341, which improves the previous state-of-the-art by 5.7%.
CNNs+ATT and DirectSup. In contrast, the explanation results presented here show that this comes at the cost of explainability, as demonstrated by the substantially lower correlation scores of CNNs+ATT+FE and DirectSup+FE. This confirms our conjecture that removing entity mentions from the sentence representation leads to more explainable models, possibly by forcing the model to focus on the textual evidence contained in the sentence rather than the word embedding of the mentions.
Finally, we note that adding LD further improves the correlation score on $H$ for $S,GI$ and $\alpha.$ This suggests that learning from distractors is a valuable strategy that not only produces better relation extraction performance, but also enhances the model explanability.
## 6.3 Evaluation of Explanations
## 7 $\textbf{Conclusion}$
In this work we provided an annotated test set with ground-truth sentence-level explanations to evaluate the explanation quality of relation extraction models with distant supervision. Our examination of two baselines show that a model with lower relation extraction accuracy could have higher explanation quality. We proposed methods to improve both the accuracy and explainability. Our proposed methods are based on changing the representation of the sentences and learning from distractor to teach the model to ignore irrelevant information in a bag. Our evaluation on the widely used FBNYT dataset show the effectiveness of our method in achieving state-of-the art performance in both accuracy and explanation quality.
We apply the explanation mechanisms described in Section 4 to produce sentence importance scores for the test set and compute the Kendall Tau correlations for the importance scores using expl-eval.
For each model, to understand its behavior when it predicts correctly versus incorrectly, we consider the subset $H$ (L) of bags/relations that the model outputs high (low) probability, i.e., $p\in[0.76,1]$ ([0,0.25]), for the correct relation. We report the performance on $H$ and $L$ separately in Tab. 3.
Comparing correlation values for $H$ and $L$ in Tab. 3, we observe that when the models are making correct and confident predictions ($H)$,the values of correlation tend to be higher. In contrast, when the model fails to detect the correct relation $(L)$,we see substantially lower correlation scores.
By replacing entity mentions with their FGET in both CNNs+ATT and DirectSup (+F), we observe substantially increased correlation scores for correct predictions (H). The improvement is consistent across all methods that are used to compute the importance scores.
Recall that Tab.2 shows that concatenating FGET with entity mention (+FE) yields improved relation extraction performance for both
# References
Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. 2016. Analyzing the behavior of visual question answering models. $EMNLP.$
|
[
[
0.1435546875,
0.0736083984375,
0.18896484375,
0.0736083984375,
0.18896484375,
0.08489990234375,
0.1435546875,
0.08489990234375
],
[
0.29296875,
0.0736083984375,
0.340087890625,
0.0736083984375,
0.340087890625,
0.08489990234375,
0.29296875,
0.08489990234375
],
[
0.355224609375,
0.0736083984375,
0.403076171875,
0.0736083984375,
0.403076171875,
0.08489990234375,
0.355224609375,
0.08489990234375
],
[
0.51904296875,
0.0736083984375,
0.55517578125,
0.0736083984375,
0.55517578125,
0.08489990234375,
0.51904296875,
0.08489990234375
],
[
0.57080078125,
0.0736083984375,
0.60107421875,
0.0736083984375,
0.60107421875,
0.08489990234375,
0.57080078125,
0.08489990234375
],
[
0.60205078125,
0.0736083984375,
0.640625,
0.0736083984375,
0.640625,
0.08489990234375,
0.60205078125,
0.08489990234375
],
[
0.6572265625,
0.0736083984375,
0.7255859375,
0.0736083984375,
0.7255859375,
0.08489990234375,
0.6572265625,
0.08489990234375
],
[
0.7421875,
0.0736083984375,
0.791015625,
0.0736083984375,
0.791015625,
0.08489990234375,
0.7421875,
0.08489990234375
],
[
0.8076171875,
0.0736083984375,
0.861328125,
0.0736083984375,
0.861328125,
0.08489990234375,
0.8076171875,
0.08489990234375
],
[
0.496337890625,
0.07598876953125,
0.5205078125,
0.07598876953125,
0.5205078125,
0.0843505859375,
0.496337890625,
0.0843505859375
],
[
0.418212890625,
0.0748291015625,
0.478515625,
0.07244873046875,
0.478515625,
0.0843505859375,
0.418212890625,
0.08612060546875
],
[
0.1435546875,
0.08612060546875,
0.2216796875,
0.08612060546875,
0.2216796875,
0.097412109375,
0.1435546875,
0.097412109375
],
[
0.298828125,
0.08612060546875,
0.33154296875,
0.08612060546875,
0.33154296875,
0.09857177734375,
0.298828125,
0.09857177734375
],
[
0.361083984375,
0.08612060546875,
0.39794921875,
0.08612060546875,
0.39794921875,
0.097412109375,
0.361083984375,
0.097412109375
],
[
0.432373046875,
0.08612060546875,
0.465087890625,
0.08612060546875,
0.465087890625,
0.09857177734375,
0.432373046875,
0.09857177734375
],
[
0.5087890625,
0.08612060546875,
0.5439453125,
0.08612060546875,
0.5439453125,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.587890625,
0.08612060546875,
0.62109375,
0.08612060546875,
0.62109375,
0.09857177734375,
0.587890625,
0.09857177734375
],
[
0.67529296875,
0.08612060546875,
0.7060546875,
0.08612060546875,
0.7060546875,
0.09857177734375,
0.67529296875,
0.09857177734375
],
[
0.748046875,
0.08612060546875,
0.7841796875,
0.08612060546875,
0.7841796875,
0.09857177734375,
0.748046875,
0.09857177734375
],
[
0.8134765625,
0.08612060546875,
0.84814453125,
0.08612060546875,
0.84814453125,
0.09857177734375,
0.8134765625,
0.09857177734375
],
[
0.300537109375,
0.097412109375,
0.33154296875,
0.097412109375,
0.33154296875,
0.10986328125,
0.300537109375,
0.10986328125
],
[
0.36279296875,
0.097412109375,
0.3955078125,
0.097412109375,
0.3955078125,
0.10986328125,
0.36279296875,
0.10986328125
],
[
0.432373046875,
0.097412109375,
0.46337890625,
0.097412109375,
0.46337890625,
0.10986328125,
0.432373046875,
0.10986328125
],
[
0.5087890625,
0.097412109375,
0.54150390625,
0.097412109375,
0.54150390625,
0.10986328125,
0.5087890625,
0.10986328125
],
[
0.58837890625,
0.097412109375,
0.62109375,
0.097412109375,
0.62109375,
0.10986328125,
0.58837890625,
0.10986328125
],
[
0.67333984375,
0.0955810546875,
0.7080078125,
0.0955810546875,
0.7080078125,
0.11102294921875,
0.67333984375,
0.11102294921875
],
[
0.75,
0.097412109375,
0.78271484375,
0.097412109375,
0.78271484375,
0.10986328125,
0.75,
0.10986328125
],
[
0.81689453125,
0.097412109375,
0.8505859375,
0.097412109375,
0.8505859375,
0.10986328125,
0.81689453125,
0.10986328125
],
[
0.1427001953125,
0.1092529296875,
0.2418212890625,
0.10986328125,
0.2418212890625,
0.1217041015625,
0.1427001953125,
0.12054443359375
],
[
0.145263671875,
0.09857177734375,
0.210693359375,
0.09857177734375,
0.210693359375,
0.10986328125,
0.145263671875,
0.10986328125
],
[
0.300537109375,
0.10986328125,
0.330078125,
0.10986328125,
0.330078125,
0.1217041015625,
0.300537109375,
0.1217041015625
],
[
0.36279296875,
0.10986328125,
0.393798828125,
0.10986328125,
0.393798828125,
0.1217041015625,
0.36279296875,
0.1217041015625
],
[
0.432373046875,
0.10986328125,
0.46337890625,
0.10986328125,
0.46337890625,
0.122314453125,
0.432373046875,
0.122314453125
],
[
0.5087890625,
0.10986328125,
0.54150390625,
0.10986328125,
0.54150390625,
0.1217041015625,
0.5087890625,
0.1217041015625
],
[
0.59033203125,
0.1092529296875,
0.62060546875,
0.1092529296875,
0.62060546875,
0.1217041015625,
0.59033203125,
0.1217041015625
],
[
0.67529296875,
0.10986328125,
0.7080078125,
0.10986328125,
0.7080078125,
0.1217041015625,
0.67529296875,
0.1217041015625
],
[
0.75,
0.10986328125,
0.78076171875,
0.10986328125,
0.78076171875,
0.1217041015625,
0.75,
0.1217041015625
],
[
0.81689453125,
0.10986328125,
0.84814453125,
0.10986328125,
0.84814453125,
0.1217041015625,
0.81689453125,
0.1217041015625
],
[
0.300537109375,
0.1217041015625,
0.33154296875,
0.1217041015625,
0.33154296875,
0.1341552734375,
0.300537109375,
0.1341552734375
],
[
0.36279296875,
0.1217041015625,
0.393798828125,
0.1217041015625,
0.393798828125,
0.1341552734375,
0.36279296875,
0.1341552734375
],
[
0.432373046875,
0.1217041015625,
0.462646484375,
0.1217041015625,
0.462646484375,
0.1341552734375,
0.432373046875,
0.1341552734375
],
[
0.50732421875,
0.1217041015625,
0.54248046875,
0.1217041015625,
0.54248046875,
0.1341552734375,
0.50732421875,
0.1341552734375
],
[
0.58837890625,
0.1217041015625,
0.62060546875,
0.1217041015625,
0.62060546875,
0.1341552734375,
0.58837890625,
0.1341552734375
],
[
0.67333984375,
0.1217041015625,
0.7080078125,
0.1217041015625,
0.7080078125,
0.1341552734375,
0.67333984375,
0.1341552734375
],
[
0.75,
0.1217041015625,
0.78271484375,
0.1217041015625,
0.78271484375,
0.1341552734375,
0.75,
0.1341552734375
],
[
0.81689453125,
0.1217041015625,
0.84814453125,
0.1217041015625,
0.84814453125,
0.1341552734375,
0.81689453125,
0.1341552734375
],
[
0.300537109375,
0.1312255859375,
0.33154296875,
0.133056640625,
0.330078125,
0.146728515625,
0.298828125,
0.144287109375
],
[
0.14697265625,
0.122314453125,
0.2301025390625,
0.122314453125,
0.2301025390625,
0.1341552734375,
0.14697265625,
0.1341552734375
],
[
0.145263671875,
0.133056640625,
0.252685546875,
0.133056640625,
0.252685546875,
0.144287109375,
0.145263671875,
0.144287109375
],
[
0.432373046875,
0.133056640625,
0.462646484375,
0.133056640625,
0.462646484375,
0.1455078125,
0.432373046875,
0.1455078125
],
[
0.58837890625,
0.133056640625,
0.62060546875,
0.133056640625,
0.62060546875,
0.1455078125,
0.58837890625,
0.1455078125
],
[
0.67236328125,
0.133056640625,
0.7080078125,
0.133056640625,
0.7080078125,
0.1455078125,
0.67236328125,
0.1455078125
],
[
0.75,
0.133056640625,
0.78271484375,
0.133056640625,
0.78271484375,
0.1455078125,
0.75,
0.1455078125
],
[
0.50732421875,
0.1341552734375,
0.54248046875,
0.1341552734375,
0.54248046875,
0.1455078125,
0.50732421875,
0.1455078125
],
[
0.36279296875,
0.1341552734375,
0.397216796875,
0.1341552734375,
0.397216796875,
0.146728515625,
0.36279296875,
0.146728515625
],
[
0.81689453125,
0.1341552734375,
0.84814453125,
0.1341552734375,
0.84814453125,
0.146728515625,
0.81689453125,
0.146728515625
],
[
0.300537109375,
0.1455078125,
0.333251953125,
0.1455078125,
0.333251953125,
0.157958984375,
0.300537109375,
0.157958984375
],
[
0.364501953125,
0.1455078125,
0.3955078125,
0.1455078125,
0.3955078125,
0.157958984375,
0.364501953125,
0.157958984375
],
[
0.432373046875,
0.1455078125,
0.46337890625,
0.1455078125,
0.46337890625,
0.157958984375,
0.432373046875,
0.157958984375
],
[
0.50732421875,
0.1455078125,
0.5439453125,
0.1455078125,
0.5439453125,
0.157958984375,
0.50732421875,
0.157958984375
],
[
0.58837890625,
0.1455078125,
0.62109375,
0.1455078125,
0.62109375,
0.15673828125,
0.58837890625,
0.15673828125
],
[
0.67529296875,
0.1455078125,
0.7060546875,
0.1455078125,
0.7060546875,
0.157958984375,
0.67529296875,
0.157958984375
],
[
0.75,
0.1455078125,
0.78271484375,
0.1455078125,
0.78271484375,
0.157958984375,
0.75,
0.157958984375
],
[
0.8154296875,
0.1455078125,
0.8486328125,
0.1455078125,
0.8486328125,
0.157958984375,
0.8154296875,
0.157958984375
],
[
0.58837890625,
0.15673828125,
0.62060546875,
0.15673828125,
0.62060546875,
0.169189453125,
0.58837890625,
0.169189453125
],
[
0.14697265625,
0.146728515625,
0.240966796875,
0.146728515625,
0.240966796875,
0.157958984375,
0.14697265625,
0.157958984375
],
[
0.67919921875,
0.15673828125,
0.7021484375,
0.15673828125,
0.7021484375,
0.169189453125,
0.67919921875,
0.169189453125
],
[
0.75,
0.15673828125,
0.78271484375,
0.15673828125,
0.78271484375,
0.169189453125,
0.75,
0.169189453125
],
[
0.298828125,
0.157958984375,
0.33154296875,
0.157958984375,
0.33154296875,
0.169189453125,
0.298828125,
0.169189453125
],
[
0.145263671875,
0.157958984375,
0.255126953125,
0.157958984375,
0.255126953125,
0.169189453125,
0.145263671875,
0.169189453125
],
[
0.36279296875,
0.157958984375,
0.397216796875,
0.157958984375,
0.397216796875,
0.169189453125,
0.36279296875,
0.169189453125
],
[
0.432373046875,
0.157958984375,
0.465087890625,
0.157958984375,
0.465087890625,
0.169189453125,
0.432373046875,
0.169189453125
],
[
0.5087890625,
0.157958984375,
0.54150390625,
0.157958984375,
0.54150390625,
0.169189453125,
0.5087890625,
0.169189453125
],
[
0.8154296875,
0.157958984375,
0.84814453125,
0.157958984375,
0.84814453125,
0.169189453125,
0.8154296875,
0.169189453125
],
[
0.1435546875,
0.168701171875,
0.274658203125,
0.16748046875,
0.274658203125,
0.18115234375,
0.1435546875,
0.1817626953125
],
[
0.58837890625,
0.168701171875,
0.62109375,
0.168701171875,
0.62109375,
0.18115234375,
0.58837890625,
0.18115234375
],
[
0.67529296875,
0.16748046875,
0.7080078125,
0.169189453125,
0.7060546875,
0.1817626953125,
0.67333984375,
0.179931640625
],
[
0.361083984375,
0.169189453125,
0.397216796875,
0.169189453125,
0.397216796875,
0.18115234375,
0.361083984375,
0.18115234375
],
[
0.298828125,
0.169189453125,
0.333251953125,
0.169189453125,
0.333251953125,
0.18115234375,
0.298828125,
0.18115234375
],
[
0.432373046875,
0.169189453125,
0.465087890625,
0.169189453125,
0.465087890625,
0.1817626953125,
0.432373046875,
0.1817626953125
],
[
0.5087890625,
0.169189453125,
0.54150390625,
0.169189453125,
0.54150390625,
0.1817626953125,
0.5087890625,
0.1817626953125
],
[
0.75,
0.169189453125,
0.7841796875,
0.169189453125,
0.7841796875,
0.1817626953125,
0.75,
0.1817626953125
],
[
0.8154296875,
0.169189453125,
0.84814453125,
0.169189453125,
0.84814453125,
0.1817626953125,
0.8154296875,
0.1817626953125
],
[
0.298828125,
0.18115234375,
0.333251953125,
0.18115234375,
0.333251953125,
0.193603515625,
0.298828125,
0.193603515625
],
[
0.36279296875,
0.18115234375,
0.397216796875,
0.18115234375,
0.397216796875,
0.193603515625,
0.36279296875,
0.193603515625
],
[
0.432373046875,
0.18115234375,
0.465087890625,
0.18115234375,
0.465087890625,
0.193603515625,
0.432373046875,
0.193603515625
],
[
0.5087890625,
0.18115234375,
0.5400390625,
0.18115234375,
0.5400390625,
0.193603515625,
0.5087890625,
0.193603515625
],
[
0.58837890625,
0.18115234375,
0.62109375,
0.18115234375,
0.62109375,
0.193603515625,
0.58837890625,
0.193603515625
],
[
0.67333984375,
0.1787109375,
0.70947265625,
0.1787109375,
0.70947265625,
0.19482421875,
0.67333984375,
0.19482421875
],
[
0.75,
0.18115234375,
0.78271484375,
0.18115234375,
0.78271484375,
0.193603515625,
0.75,
0.193603515625
],
[
0.8154296875,
0.18115234375,
0.8486328125,
0.18115234375,
0.8486328125,
0.193603515625,
0.8154296875,
0.193603515625
],
[
0.1435546875,
0.1953125,
0.322509765625,
0.1953125,
0.322509765625,
0.2060546875,
0.1435546875,
0.2060546875
],
[
0.145263671875,
0.1817626953125,
0.264404296875,
0.1817626953125,
0.264404296875,
0.1923828125,
0.145263671875,
0.1923828125
],
[
0.1435546875,
0.2060546875,
0.310546875,
0.2060546875,
0.310546875,
0.2197265625,
0.1435546875,
0.2197265625
],
[
0.2418212890625,
0.22802734375,
0.75732421875,
0.2286376953125,
0.75732421875,
0.243408203125,
0.2418212890625,
0.2423095703125
],
[
0.12005615234375,
0.271484375,
0.486083984375,
0.271484375,
0.486083984375,
0.281982421875,
0.12005615234375,
0.281982421875
],
[
0.5146484375,
0.26953125,
0.884765625,
0.26953125,
0.884765625,
0.283203125,
0.5146484375,
0.283203125
],
[
0.118408203125,
0.28515625,
0.490234375,
0.2861328125,
0.490234375,
0.300537109375,
0.118408203125,
0.2998046875
],
[
0.51318359375,
0.2861328125,
0.88330078125,
0.28515625,
0.88330078125,
0.2998046875,
0.51318359375,
0.300537109375
],
[
0.1175537109375,
0.300537109375,
0.48779296875,
0.300537109375,
0.48779296875,
0.314208984375,
0.1175537109375,
0.314208984375
],
[
0.5146484375,
0.303955078125,
0.8818359375,
0.303955078125,
0.8818359375,
0.314208984375,
0.5146484375,
0.314208984375
],
[
0.12005615234375,
0.319580078125,
0.197265625,
0.319580078125,
0.197265625,
0.330810546875,
0.12005615234375,
0.330810546875
],
[
0.515625,
0.320556640625,
0.88330078125,
0.320556640625,
0.88330078125,
0.330810546875,
0.515625,
0.330810546875
],
[
0.1376953125,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.1376953125,
0.349853515625
],
[
0.5146484375,
0.3349609375,
0.88330078125,
0.3349609375,
0.88330078125,
0.3486328125,
0.5146484375,
0.3486328125
],
[
0.118408203125,
0.352783203125,
0.489501953125,
0.352783203125,
0.489501953125,
0.366455078125,
0.118408203125,
0.366455078125
],
[
0.51123046875,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.51123046875,
0.364013671875
],
[
0.118408203125,
0.3681640625,
0.489501953125,
0.3681640625,
0.489501953125,
0.3818359375,
0.118408203125,
0.3818359375
],
[
0.51318359375,
0.36767578125,
0.88330078125,
0.36767578125,
0.88330078125,
0.38134765625,
0.51318359375,
0.38134765625
],
[
0.12176513671875,
0.385986328125,
0.48779296875,
0.385986328125,
0.48779296875,
0.396728515625,
0.12176513671875,
0.396728515625
],
[
0.5146484375,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3955078125,
0.5146484375,
0.3955078125
],
[
0.12005615234375,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.412109375,
0.12005615234375,
0.412109375
],
[
0.51123046875,
0.399658203125,
0.884765625,
0.3984375,
0.884765625,
0.412109375,
0.51123046875,
0.413330078125
],
[
0.1175537109375,
0.416259765625,
0.48779296875,
0.416259765625,
0.48779296875,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.51318359375,
0.4150390625,
0.875,
0.4150390625,
0.875,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.118408203125,
0.431640625,
0.489501953125,
0.431640625,
0.489501953125,
0.4453125,
0.118408203125,
0.4453125
],
[
0.53076171875,
0.43115234375,
0.88330078125,
0.432861328125,
0.88330078125,
0.446533203125,
0.53076171875,
0.44482421875
],
[
0.1175537109375,
0.449462890625,
0.365234375,
0.448974609375,
0.365234375,
0.4619140625,
0.1175537109375,
0.463134765625
],
[
0.51318359375,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.5146484375,
0.465576171875,
0.884765625,
0.465576171875,
0.884765625,
0.478515625,
0.5146484375,
0.478515625
],
[
0.118408203125,
0.479736328125,
0.376220703125,
0.479736328125,
0.376220703125,
0.493408203125,
0.118408203125,
0.493408203125
],
[
0.51318359375,
0.48095703125,
0.884765625,
0.48095703125,
0.884765625,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.51318359375,
0.49755859375,
0.884765625,
0.49755859375,
0.884765625,
0.51123046875,
0.51318359375,
0.51123046875
],
[
0.118408203125,
0.50439453125,
0.489501953125,
0.50439453125,
0.489501953125,
0.51708984375,
0.118408203125,
0.51708984375
],
[
0.118408203125,
0.5185546875,
0.489501953125,
0.51953125,
0.489501953125,
0.53369140625,
0.118408203125,
0.533203125
],
[
0.51318359375,
0.51318359375,
0.615234375,
0.51318359375,
0.615234375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.118408203125,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.54248046875,
0.54443359375,
0.640625,
0.54443359375,
0.640625,
0.5556640625,
0.54248046875,
0.5556640625
],
[
0.118408203125,
0.5517578125,
0.480224609375,
0.5517578125,
0.480224609375,
0.56640625,
0.118408203125,
0.56640625
],
[
0.1368408203125,
0.568359375,
0.489501953125,
0.568359375,
0.489501953125,
0.58203125,
0.1368408203125,
0.58203125
],
[
0.5146484375,
0.57470703125,
0.88330078125,
0.57470703125,
0.88330078125,
0.5849609375,
0.5146484375,
0.5849609375
],
[
0.1175537109375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.51318359375,
0.5888671875,
0.88671875,
0.58837890625,
0.88671875,
0.6025390625,
0.51318359375,
0.60400390625
],
[
0.118408203125,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.6123046875,
0.118408203125,
0.6123046875
],
[
0.51318359375,
0.60546875,
0.884765625,
0.60546875,
0.884765625,
0.619140625,
0.51318359375,
0.619140625
],
[
0.118408203125,
0.6181640625,
0.486083984375,
0.6181640625,
0.486083984375,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51318359375,
0.62158203125,
0.88671875,
0.62158203125,
0.88671875,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.118408203125,
0.63427734375,
0.489501953125,
0.63427734375,
0.489501953125,
0.64794921875,
0.118408203125,
0.64794921875
],
[
0.51318359375,
0.63720703125,
0.88671875,
0.63720703125,
0.88671875,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.1175537109375,
0.6494140625,
0.456787109375,
0.6484375,
0.456787109375,
0.6630859375,
0.1175537109375,
0.66455078125
],
[
0.5146484375,
0.65380859375,
0.88671875,
0.65380859375,
0.88671875,
0.66748046875,
0.5146484375,
0.66748046875
],
[
0.1351318359375,
0.66650390625,
0.489501953125,
0.6650390625,
0.489501953125,
0.6796875,
0.1351318359375,
0.68115234375
],
[
0.5146484375,
0.66943359375,
0.884765625,
0.66943359375,
0.884765625,
0.68310546875,
0.5146484375,
0.68310546875
],
[
0.118408203125,
0.681640625,
0.489501953125,
0.681640625,
0.489501953125,
0.6953125,
0.118408203125,
0.6953125
],
[
0.5146484375,
0.68603515625,
0.88330078125,
0.68603515625,
0.88330078125,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.1175537109375,
0.69970703125,
0.489501953125,
0.69873046875,
0.489501953125,
0.7119140625,
0.1175537109375,
0.71337890625
],
[
0.51318359375,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.118408203125,
0.71533203125,
0.490234375,
0.71533203125,
0.490234375,
0.728515625,
0.118408203125,
0.728515625
],
[
0.51318359375,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.728515625,
0.51318359375,
0.728515625
],
[
0.118408203125,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.74462890625,
0.118408203125,
0.74462890625
],
[
0.51123046875,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.51123046875,
0.74658203125
],
[
0.1175537109375,
0.7490234375,
0.484375,
0.7490234375,
0.484375,
0.7626953125,
0.1175537109375,
0.7626953125
],
[
0.5146484375,
0.751953125,
0.8818359375,
0.751953125,
0.8818359375,
0.7626953125,
0.5146484375,
0.7626953125
],
[
0.1351318359375,
0.76416015625,
0.489501953125,
0.76318359375,
0.489501953125,
0.77783203125,
0.1351318359375,
0.779296875
],
[
0.5146484375,
0.7666015625,
0.884765625,
0.7666015625,
0.884765625,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.1175537109375,
0.77978515625,
0.489501953125,
0.77978515625,
0.489501953125,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.5146484375,
0.78369140625,
0.8818359375,
0.78369140625,
0.8818359375,
0.79443359375,
0.5146484375,
0.79443359375
],
[
0.118408203125,
0.79736328125,
0.490234375,
0.79736328125,
0.490234375,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.51318359375,
0.798828125,
0.759765625,
0.798828125,
0.759765625,
0.8125,
0.51318359375,
0.8125
],
[
0.118408203125,
0.8134765625,
0.490234375,
0.8134765625,
0.490234375,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.118408203125,
0.830078125,
0.489501953125,
0.830078125,
0.489501953125,
0.84326171875,
0.118408203125,
0.84326171875
],
[
0.118408203125,
0.84423828125,
0.285400390625,
0.845703125,
0.285400390625,
0.859375,
0.118408203125,
0.85791015625
],
[
0.51318359375,
0.84326171875,
0.611328125,
0.845703125,
0.611328125,
0.85986328125,
0.51318359375,
0.85791015625
],
[
0.1351318359375,
0.86083984375,
0.489501953125,
0.8623046875,
0.489501953125,
0.876953125,
0.1351318359375,
0.8759765625
],
[
0.51123046875,
0.8681640625,
0.88330078125,
0.86767578125,
0.88330078125,
0.8818359375,
0.51123046875,
0.8828125
],
[
0.1175537109375,
0.87744140625,
0.490234375,
0.87890625,
0.490234375,
0.8935546875,
0.1175537109375,
0.892578125
],
[
0.53076171875,
0.8828125,
0.884765625,
0.8828125,
0.884765625,
0.896484375,
0.53076171875,
0.896484375
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.53076171875,
0.896484375,
0.70947265625,
0.8955078125,
0.70947265625,
0.90673828125,
0.53076171875,
0.90771484375
]
] |
[
"model",
"loo (H)",
"loo (L)",
"(L)",
"GIr",
"(H)",
"GI(L)",
"Qan(H)",
"a(L)",
"Saa",
"SE(H)",
"CNNs+ATT",
"0.16",
"-0.08",
"0.19",
"-0.02",
"0.20",
"0.04",
"0.69",
"0.21",
"0.19",
"0.12",
"0.08",
"0.15",
"0.29",
"0.19",
"0.26",
"-0.12",
"CNNS+ATT+F",
"DirectSup",
"0.21",
"0.10",
"0.36",
"0.03",
"0.23",
"0.00",
"0.73",
"0.11",
"0.24",
"0.15",
"0.31",
"-0.19",
"0.40",
"-0.17",
"0.28",
"0.15",
"0.01",
"DirectSup +F",
"CNNs+ATT +FE",
"0.21",
"0.20",
"-0.20",
"0.24",
"-0.14",
"-0.11",
"0.01",
"0.14",
"-.12",
"0.19",
"-0.10",
"0.29",
"0.06",
"0.17",
"-0.11",
"0.21",
"DirectSup +FE",
"0",
"0.67",
"0.18",
"CNNs+ATT +LD",
"-0.01",
"0.22",
"0.10",
"0.11",
"CNNS+ATT +LD +F",
"0.28",
"0.07",
"-0.11",
"0.22",
"0.43",
"0.09",
"0.70",
"0.12",
"0.23",
"0.14",
"0.38",
"0.01",
"0.49",
"0.20",
"0.45",
"0.02",
"H: High confidence P(r) e [0.76, 1.0]",
"DirectSup +LD +F",
"L: Low confidence P(r) [0, 0.25]",
"Table 3:Kendall correlations for top confidence and least confidence range.",
"whichwewillexaminein thenext section.Finally",
"CNNs+ATT and DirectSup. In contrast, the ex-",
"the last three rows of the table show that adding",
"planation results presented here show that this",
"LD to different base models can further improve",
"comes atthecostof explainability,as demonstrated",
"the AUCs.",
"by the substantiallylower correlation scoresof",
"Similar to prior work, we observe that incorpo-",
"CNNs+ATT+FE and DirectSup+FE. This confirms",
"rating entity embeddings(+E) to the model leads to",
"our conjecture that removing entity mentions from",
"substantial performance gain across the board. We",
"the sentence representation leads to more explain-",
"alsoobserveverysimilarperformancegainwhen",
"ablemodels.possiblybyforcingthemodeltofocus",
"addingFGETandLDtothebasemodelsbothWith",
"on the textual evidence contained in the sentence",
"and without entity embeddings. Our best model",
"rather than the word embedding of the mentions.",
"achieved an AUC of 0.341, which improves the",
"Finally,we note that adding LD further improves",
"previous state-of-the-art by 5.7%.",
"the correlation score on H for S, GI and aα. This",
"suggests that learning from distractors is a valuable",
"6.3Evaluation of Explanations",
"strategy that not only produces better relation ex-",
"traction performance, but also enhances the model",
"We apply the explanation mechanisms described in",
"Section 4 to produce sentence importance scores",
"explanability.",
"for the test set and compute the Kendall Tau corre-",
"Conclusion",
"lations for the importance scores using expl-eval.",
"For each model, to understand its behavior when",
"In this work we provided an annotated test set with",
"it predicts correctly versus incorrectly, we consider",
"ground-truth sentence-level explanations to eval-",
"the subset H (L) of bags/relations that the model",
"uate the explanation quality of relation extraction",
"outputs high (low) probability, i.e., p E [0.76, 1]",
"models with distant supervision. Our examination",
"([0, 0.25]), for the correct relation. We report the",
"of two baselines show that a model with lower rela-",
"performance on H and L separately in Tab. 3.",
"tion extraction accuracy could have higher expla-",
"Comparing correlation values for H and L in",
"nation quality. We proposed methods to improve",
"Tab. 3, we observe that when the models are mak-",
"both the accuracy and explainability. Our proposed",
"ingcorrect and confidentpredictions (H),theval-",
"methods are based on changing the representation",
"ues of correlation tend to be higher. In contrast,",
"ofthe sentences andlearningfrom distractor to",
"when the model fails to detect the correct relation",
"teach the model to ignore irrelevant information",
"(L), we see substantially lower correlation scores.",
"inabag.Ourevaluationonthe widelyused FB.",
"By replacing entity mentions with their FGET",
"NYT dataset show the effectiveness of our method",
"in both CNNs+ATT and DirectSup (+F), we ob-",
"in achieving state-of-the art performanceinboth",
"serve substantially increased correlation scores for",
"accuracy and explanation quality.",
"correct predictions (H). The improvement is con-",
"sistent across all methods that are used to compute",
"the importance scores.",
"References",
"Recall that Tab. 2 shows that concatenating",
"Aishwarya Agrawal, Dhruv Batra, and Devi Parikh.",
"FGET with entity mention (+FE) yields im-",
"2016. Analyzing the behavior of visual question an-",
"proved relation extraction performance for both",
"swering models.EMNLP."
] |
{
"arxiv_id": "2005.14271",
"language": "en",
"page_idx": 4
}
|
||
1511.03575-en-0
|
1511.03575-en-0.jpg
|
# $\textbf{Federated Optimization: }$ Distributed Optimization Beyond the Datacenter
Jakub Konečný* School of Mathematics University of Edinburgh
J.Konecnyêsms.ed.ac.uk
H. Brendan McMahan
Google, Inc.
Seattle, WA 98103
mcmahangoogle.com
Daniel Ramage Google, Inc.
Seattle, WA 98103
dramagegoogle.com
## $\mathbf{Abstract}$
We introduce a new and increasingly relevant setting for distributed optimization in machine learning, where the data defining the optimization are distributed (unevenly) over an extremely large number of nodes, but the goal remains to train a high-quality centralized model. We refer to this setting as Federated Optimization. In this setting, communication effciency is of utmost importance. A motivating example for federated optimization arises when we keep the training data locally on users' mobile devices rather than logging it to a data center for training. Instead, the mobile devices are used as nodes performing computation on their local data in order to update a global model. We suppose that we have an extremely large number of devices in our network, each of which has only a tiny fraction of data available totally; in particular, we expect the number of data points available locally to be much smaller than the number of devices. Additionally, since different users generate data with different patterns, we assume that no device has a representative sample of the overall distribution.
We show that existing algorithms are not suitable for this setting, and propose a new algorithm which shows encouraging experimental results. This work also sets a path for future research needed in the context of federated optimization.
1 $\textbf{Introduction and Problem Formulation}$
The optimization community has seen an explosion of interest in solving problems with finite-sum structure in recent
years. In general, the objective is formulated as
(1)
$$\min\limits_{w\in\mathbb{R}^d}f(w)\quad\text{where}\quad f(w)\stackrel{\text{def}}{=}\frac{1}{n}\sum\limits_{i=1}^nf_i(w).$$
The main source of motivation are problems arising in machine learning. The problem structure (1) covers linear or logistic regressions, support vector machines, but also more complicated algorithms such as conditional random fields or neural networks. The functions $f_i$ are typically specified via a loss function $\ell$ dependent on a pair of input-output data point $\{x_i,y_i\},f_i(w)=\ell(w;x_i,y_i).$
The amount of data that businesses, governments and academic projects collect is rapidly increasing. Consequently, solving the problem (1) is becoming impossible with a single node, as merely storing the whole dataset on a single node becomes infeasible. This gives rise to the distributed optimization setting, where we solve the problem (1) even though no single node has direct access to all the data
In the distributed setting, communication cost is by far the largest bottleneck, as exhibited by a large amount of existing work. Further, many state-of-the-art optimization algorithms are inherently sequential, relying on a large number of
${}^{*}$Work performed while at Google, Inc.
1
|
[
[
0.348876953125,
0.125,
0.650390625,
0.126953125,
0.650390625,
0.147705078125,
0.348876953125,
0.1458740234375
],
[
0.205078125,
0.149658203125,
0.7939453125,
0.15087890625,
0.7939453125,
0.1717529296875,
0.205078125,
0.17041015625
],
[
0.43701171875,
0.2266845703125,
0.59716796875,
0.2266845703125,
0.59716796875,
0.2412109375,
0.43701171875,
0.2412109375
],
[
0.69287109375,
0.2259521484375,
0.8037109375,
0.2279052734375,
0.8037109375,
0.242431640625,
0.69287109375,
0.240478515625
],
[
0.2156982421875,
0.2279052734375,
0.33251953125,
0.2279052734375,
0.33251953125,
0.2412109375,
0.2156982421875,
0.2412109375
],
[
0.0196075439453125,
0.2310791015625,
0.05963134765625,
0.2310791015625,
0.062103271484375,
0.352783203125,
0.0228729248046875,
0.352783203125
],
[
0.191162109375,
0.240478515625,
0.34716796875,
0.2412109375,
0.34716796875,
0.255615234375,
0.191162109375,
0.255126953125
],
[
0.47802734375,
0.24365234375,
0.55810546875,
0.24365234375,
0.55810546875,
0.255126953125,
0.47802734375,
0.255126953125
],
[
0.70751953125,
0.24365234375,
0.78759765625,
0.24365234375,
0.78759765625,
0.255126953125,
0.70751953125,
0.255126953125
],
[
0.1871337890625,
0.25244140625,
0.3505859375,
0.255126953125,
0.3505859375,
0.26953125,
0.1871337890625,
0.267578125
],
[
0.455810546875,
0.255126953125,
0.58251953125,
0.255126953125,
0.58251953125,
0.268310546875,
0.455810546875,
0.268310546875
],
[
0.685546875,
0.25390625,
0.81201171875,
0.25390625,
0.81201171875,
0.268310546875,
0.685546875,
0.268310546875
],
[
0.160888671875,
0.26953125,
0.378173828125,
0.26953125,
0.378173828125,
0.28271484375,
0.160888671875,
0.28271484375
],
[
0.428955078125,
0.268310546875,
0.60693359375,
0.268310546875,
0.60693359375,
0.2822265625,
0.428955078125,
0.2822265625
],
[
0.6611328125,
0.26953125,
0.8388671875,
0.26953125,
0.8388671875,
0.2841796875,
0.6611328125,
0.2841796875
],
[
0.461669921875,
0.318115234375,
0.5390625,
0.318115234375,
0.5390625,
0.3359375,
0.461669921875,
0.3359375
],
[
0.02777099609375,
0.34228515625,
0.053924560546875,
0.34228515625,
0.053924560546875,
0.375,
0.02777099609375,
0.375
],
[
0.1748046875,
0.35107421875,
0.82275390625,
0.35107421875,
0.82275390625,
0.3642578125,
0.1748046875,
0.3642578125
],
[
0.1748046875,
0.365478515625,
0.82421875,
0.365478515625,
0.82421875,
0.37890625,
0.1748046875,
0.37890625
],
[
0.0228729248046875,
0.376953125,
0.062103271484375,
0.376953125,
0.062103271484375,
0.464111328125,
0.0228729248046875,
0.464111328125
],
[
0.176513671875,
0.380615234375,
0.8203125,
0.380615234375,
0.8203125,
0.391357421875,
0.176513671875,
0.391357421875
],
[
0.1748046875,
0.393310546875,
0.814453125,
0.393310546875,
0.814453125,
0.4072265625,
0.1748046875,
0.4072265625
],
[
0.1748046875,
0.4091796875,
0.82421875,
0.4091796875,
0.82421875,
0.42236328125,
0.1748046875,
0.42236328125
],
[
0.1748046875,
0.42236328125,
0.82421875,
0.42236328125,
0.82421875,
0.436279296875,
0.1748046875,
0.436279296875
],
[
0.176513671875,
0.438232421875,
0.82275390625,
0.438232421875,
0.82275390625,
0.448974609375,
0.176513671875,
0.448974609375
],
[
0.1748046875,
0.45068359375,
0.82421875,
0.45068359375,
0.82421875,
0.464111328125,
0.1748046875,
0.464111328125
],
[
0.1748046875,
0.464111328125,
0.826171875,
0.464111328125,
0.826171875,
0.477783203125,
0.1748046875,
0.477783203125
],
[
0.0228729248046875,
0.47021484375,
0.060455322265625,
0.47021484375,
0.060455322265625,
0.5791015625,
0.0228729248046875,
0.5791015625
],
[
0.1732177734375,
0.477783203125,
0.826171875,
0.479248046875,
0.826171875,
0.492431640625,
0.1732177734375,
0.4912109375
],
[
0.1748046875,
0.492431640625,
0.826171875,
0.492431640625,
0.826171875,
0.505859375,
0.1748046875,
0.505859375
],
[
0.1748046875,
0.5068359375,
0.25244140625,
0.5068359375,
0.25244140625,
0.51806640625,
0.1748046875,
0.51806640625
],
[
0.1748046875,
0.521484375,
0.82421875,
0.521484375,
0.82421875,
0.53515625,
0.1748046875,
0.53515625
],
[
0.1748046875,
0.5361328125,
0.82421875,
0.5361328125,
0.82421875,
0.5498046875,
0.1748046875,
0.5498046875
],
[
0.1748046875,
0.55029296875,
0.398681640625,
0.55029296875,
0.398681640625,
0.564453125,
0.1748046875,
0.564453125
],
[
0.0196075439453125,
0.564453125,
0.058013916015625,
0.564453125,
0.060455322265625,
0.6875,
0.0228729248046875,
0.6875
],
[
0.12091064453125,
0.5947265625,
0.125,
0.5947265625,
0.125,
0.599609375,
0.12091064453125,
0.599609375
],
[
0.1422119140625,
0.59033203125,
0.4755859375,
0.59033203125,
0.4755859375,
0.60400390625,
0.1422119140625,
0.60400390625
],
[
0.1168212890625,
0.62060546875,
0.88232421875,
0.62060546875,
0.88232421875,
0.6337890625,
0.1168212890625,
0.6337890625
],
[
0.11517333984375,
0.6337890625,
0.426513671875,
0.6337890625,
0.426513671875,
0.64794921875,
0.11517333984375,
0.64794921875
],
[
0.5107421875,
0.66357421875,
0.59228515625,
0.65478515625,
0.59619140625,
0.67529296875,
0.51318359375,
0.6845703125
],
[
0.5849609375,
0.66552734375,
0.65869140625,
0.66552734375,
0.65869140625,
0.68310546875,
0.5849609375,
0.68310546875
],
[
0.861328125,
0.66650390625,
0.8837890625,
0.66650390625,
0.8837890625,
0.68115234375,
0.861328125,
0.68115234375
],
[
0.441162109375,
0.66748046875,
0.484375,
0.66748046875,
0.484375,
0.6787109375,
0.441162109375,
0.6787109375
],
[
0.340576171875,
0.66650390625,
0.41259765625,
0.66650390625,
0.41259765625,
0.68115234375,
0.340576171875,
0.68115234375
],
[
0.34228515625,
0.6787109375,
0.365234375,
0.68115234375,
0.363525390625,
0.68798828125,
0.340576171875,
0.68603515625
],
[
0.57421875,
0.67529296875,
0.5986328125,
0.67529296875,
0.5986328125,
0.689453125,
0.57421875,
0.689453125
],
[
0.5888671875,
0.6796875,
0.61376953125,
0.68310546875,
0.609375,
0.6962890625,
0.58642578125,
0.6923828125
],
[
0.1168212890625,
0.69970703125,
0.88232421875,
0.69970703125,
0.88232421875,
0.71337890625,
0.1168212890625,
0.71337890625
],
[
0.11517333984375,
0.71337890625,
0.88232421875,
0.71337890625,
0.88232421875,
0.7265625,
0.11517333984375,
0.7265625
],
[
0.11517333984375,
0.72802734375,
0.8837890625,
0.72802734375,
0.8837890625,
0.7412109375,
0.11517333984375,
0.7412109375
],
[
0.11358642578125,
0.74072265625,
0.384033203125,
0.7412109375,
0.384033203125,
0.755859375,
0.11358642578125,
0.7548828125
],
[
0.1168212890625,
0.76220703125,
0.88134765625,
0.76220703125,
0.88134765625,
0.7763671875,
0.1168212890625,
0.7763671875
],
[
0.11517333984375,
0.77587890625,
0.88232421875,
0.7744140625,
0.88232421875,
0.7890625,
0.11517333984375,
0.79052734375
],
[
0.11517333984375,
0.79052734375,
0.88232421875,
0.79052734375,
0.88232421875,
0.8037109375,
0.11517333984375,
0.8037109375
],
[
0.1168212890625,
0.8046875,
0.47216796875,
0.8046875,
0.47216796875,
0.818359375,
0.1168212890625,
0.818359375
],
[
0.1168212890625,
0.82470703125,
0.88134765625,
0.82470703125,
0.88134765625,
0.8388671875,
0.1168212890625,
0.8388671875
],
[
0.1168212890625,
0.8388671875,
0.88232421875,
0.8388671875,
0.88232421875,
0.85400390625,
0.1168212890625,
0.85400390625
],
[
0.1380615234375,
0.861328125,
0.36767578125,
0.861328125,
0.36767578125,
0.875,
0.1380615234375,
0.875
],
[
0.494384765625,
0.8984375,
0.50634765625,
0.8984375,
0.50634765625,
0.912109375,
0.494384765625,
0.912109375
]
] |
[
"Federated Optimization:",
"Distributed Optimization Beyond the Datacenter",
"H. Brendan McMahan",
"Daniel Ramage",
"Jakub Konecny*",
"S0Z AN",
"School of Mathematics",
"Google, Inc.",
"Google, Inc",
"University of Edinburgh",
"Seattle, WA 98103",
"Seattle, WA 98103",
"J.Konecny@sms.ed.ac.uk",
"mcmahan@google.com",
"dramage@google.com",
"Abstract",
"II",
"We introduce a new and increasingly relevant setting for distributed optimization in machine learn-",
"ing, where the data defining the optimization are distributed (unevenly) over an extremely large",
"[T'S]",
"numberofnodes,butthegoalremainstotrainahigh-qualitycentralizedmodel.Werefertothisset",
"ting as Federated Optimization. In this setting, communication efficiency is of utmost importance.",
"A motivating example for federated optimization arises when we keep the training data locally on",
"users' mobile devices rather than logging it to a data center for training. Instead, the mobile devices",
"areusedasnodesperforming computationontheirlocaldata inordertoupdateaglobalmodel.We",
"suppose that we have an extremely large number of devices in our network, each of which has only",
"a tiny fraction of data available totally; in particular, we expect the number of data points available",
"ASLSO'T",
"locally to be much smaller than the number of devices. Additionally, since different users generate",
"data with different patterns, we assume that no device has a representative sample of the overall",
"distribution",
"We show that existing algorithms are not suitable for this setting, and propose a new algorithm which",
"shows encouraging experimental results. This work also sets a path for future research needed in the",
"context of federated optimization.",
"STAIIE",
"-",
"Introduction and Problem Formulation",
"The optimization community has seen an explosion of interest in solving problems with finite-sum structure in recent",
"years. In general, the objective is formulated as",
"f(w) dr 1",
"Z fi(w).",
"(1)",
"where",
"min f(w)",
"WEW",
"n 4",
"i=l",
"The main source of motivation are problems arising in machine learning. The problem structure (1) covers linear or",
"logistic regressions, support vector machines, but also more complicated algorithms such as conditional random fields",
"or neural networks.The functions fi are typically specified via a loss function l dependent on a pair of input-output",
"data point [ri, yi), fi(w) = l(w; ri, yi).",
"The amount of data that businesses, governments and academic projects collect is rapidly increasing. Consequently.",
"solving the problem (1) is becoming impossible with a single node, as merely storing the whole dataset on a single",
"node becomes infeasible. This gives rise to the distributed optimization setting, where we solve the problem (1) even",
"though no single node has direct access to all the data.",
"In the distributed setting, communication cost is by far the largest bottleneck, as exhibited by a large amount of existing",
"work. Further, many state-of-the-art optimization algorithms are inherently sequential, relying on a large number of",
"*Work performed while at Google, Inc.",
"1"
] |
{
"arxiv_id": "1511.03575",
"language": "en",
"page_idx": 0
}
|
||
1511.03575-en-1
|
1511.03575-en-1.jpg
|
very fast iterations. The problem stems from the fact that if one needs to perform a round of communication after each iteration, practical performance drops down dramatically, as the round of communication is much more timeconsuming than a single iteration of the algorithm.
Works including [2,8,3,9] have established basic theoretical results and developed distributed versions of existing methods, but as this is a relatively new area, many open questions remain. Recently, the idea of communication efficient algorithms has gained traction. Such algorithms perform a large amount of computation locally, before each round of communication, ideally balancing the cost of computation and communication[10,11,4,7,14]. There has been an attempt at a thorough understanding of communication lower bounds[1], but the authors note that significant gaps remain in a number of settings.
2 Federated Optimization-The Challenge of Learning from Decentralized Data
With the exception of CoCoA [4,7], the existing work concerning communication efficient optimization [10,14] and/or statistical estimation [13, 12] presupposes, and in fact requires, that the number of nodes is (much) smaller tham the number of datapoints available to each node, and that each node has access to a random sample from the same distribution, and typically that each node has an identical number of data points.
We are hoping to bring to attention a new and increasingly relevant setting for distributed optimization, where typically
none of the above assumption are satisfied, and communication efficiency is of utmost importance. In particular,
algorithms for federated optimization must handle training data with the following characteristics:
$\bullet$ Massively Distributed: Data points are stored across a very large number of nodes $K.$ In particular, the
number of nodes can be much bigger than the average number of training examples stored on a given node
$(n/K).$
$\bullet$ Non-IID: Data on each node may be drawn from a different distribution; that is, the data points available
locally are far from being a representative sample of the overall distribution
$\bullet$ Unbalanced: Different nodes may vary by many orders of magnitude in the number of training examples
they hold.
In this work, we are particularly concerned with sparse data, where some features occur on only a small subset of
nodes or data points. We show that the sparsity structure can be used to develop an effective algorithm for federated
optimization.
We are motivated by the setting where training data lives on users' mobile devices (phones and tablets), and the data may be privacy sensitive. The data $\{x_i,y_i\}_{i=1}^n$ are generated by device usage, e.g. interaction with apps. Examples include predicting the next word a user will type (language modeling for smarter keyboard apps), predicting which photos a user is most likely to share, or predicting which notifications are most important. To train such models using traditional distributed learning algorithms, one would collect the training examples in a centralized location (data center) where it could be shuffled and distributed evenly over compute nodes. Federated optimization provides an alternative model potentially saving significant network bandwidth and providing additional privacy protection. In exchange, users allow some use of their devices'computing power.
Communication constraints arise naturally in the massively distributed setting, as network connectivity may be limited (e.g., we may wish to defer all communication until the mobile device is charging and connected to a wi-fi network). Thus, in realistic scenarios we may be limited to only a single round of communication per day. This implies that, within reasonable bounds, we have access to essentially unlimited local computational power. As a result, the practical objective is solely to minimize the rounds of communication.
Formally, let $K$ be the number of nodes. Let $\{\mathcal{P}_k\}_{k=1}^K$ denote a partition of data point indices $\{1,\ldots,n\}$, so $\mathcal{P}_k$ is the set of data points stored on node $k$, and define $n_k=|\mathcal{P}_k|.$ We can then rephrase the objective(1) as a linear combination of the local empirical objectives $F_k(w)$, defined as
$$f(w)=\sum\limits_{k=1}^K\frac{n_k}{n}F_k(w)\stackrel{\text{def}}{=}\sum\limits_{k=1}^K\frac{n_k}{n}\cdot\frac{1}{n_k}\sum\limits_{i\in\mathcal{P}_k}f_i(w).$$
2
|
[
[
0.11517333984375,
0.0921630859375,
0.88232421875,
0.0921630859375,
0.88232421875,
0.1060791015625,
0.11517333984375,
0.1060791015625
],
[
0.11517333984375,
0.106689453125,
0.8798828125,
0.106689453125,
0.8798828125,
0.12060546875,
0.11517333984375,
0.12060546875
],
[
0.11517333984375,
0.12060546875,
0.449462890625,
0.12060546875,
0.449462890625,
0.1337890625,
0.11517333984375,
0.1337890625
],
[
0.11517333984375,
0.14013671875,
0.88134765625,
0.141357421875,
0.88134765625,
0.1558837890625,
0.11517333984375,
0.1546630859375
],
[
0.11517333984375,
0.1558837890625,
0.88134765625,
0.1558837890625,
0.88134765625,
0.169189453125,
0.11517333984375,
0.169189453125
],
[
0.11517333984375,
0.17041015625,
0.88134765625,
0.17041015625,
0.88134765625,
0.18115234375,
0.11517333984375,
0.18115234375
],
[
0.11517333984375,
0.18310546875,
0.88232421875,
0.18310546875,
0.88232421875,
0.1962890625,
0.11517333984375,
0.1962890625
],
[
0.1168212890625,
0.1988525390625,
0.88134765625,
0.1988525390625,
0.88134765625,
0.208984375,
0.1168212890625,
0.208984375
],
[
0.11358642578125,
0.2108154296875,
0.35693359375,
0.210205078125,
0.35693359375,
0.2247314453125,
0.11358642578125,
0.2259521484375
],
[
0.11767578125,
0.25830078125,
0.13232421875,
0.25830078125,
0.13232421875,
0.266357421875,
0.11767578125,
0.266357421875
],
[
0.12744140625,
0.255126953125,
0.8095703125,
0.255126953125,
0.8095703125,
0.26953125,
0.12744140625,
0.26953125
],
[
0.1168212890625,
0.291748046875,
0.88134765625,
0.291748046875,
0.88134765625,
0.30615234375,
0.1168212890625,
0.30615234375
],
[
0.1168212890625,
0.30615234375,
0.88232421875,
0.30615234375,
0.88232421875,
0.3193359375,
0.1168212890625,
0.3193359375
],
[
0.1168212890625,
0.3193359375,
0.88134765625,
0.3193359375,
0.88134765625,
0.333251953125,
0.1168212890625,
0.333251953125
],
[
0.11358642578125,
0.333984375,
0.6396484375,
0.333984375,
0.6396484375,
0.347900390625,
0.11358642578125,
0.347900390625
],
[
0.1168212890625,
0.354248046875,
0.88134765625,
0.354248046875,
0.88134765625,
0.36865234375,
0.1168212890625,
0.36865234375
],
[
0.11517333984375,
0.36865234375,
0.88232421875,
0.36865234375,
0.88232421875,
0.3818359375,
0.11517333984375,
0.3818359375
],
[
0.11517333984375,
0.3818359375,
0.7587890625,
0.3818359375,
0.7587890625,
0.395751953125,
0.11517333984375,
0.395751953125
],
[
0.1568603515625,
0.412353515625,
0.88134765625,
0.412353515625,
0.88134765625,
0.4267578125,
0.1568603515625,
0.4267578125
],
[
0.1732177734375,
0.427978515625,
0.88232421875,
0.427978515625,
0.88232421875,
0.44140625,
0.1732177734375,
0.44140625
],
[
0.1732177734375,
0.442626953125,
0.226318359375,
0.442626953125,
0.226318359375,
0.45263671875,
0.1732177734375,
0.45263671875
],
[
0.158447265625,
0.453857421875,
0.88232421875,
0.453857421875,
0.88232421875,
0.468505859375,
0.158447265625,
0.468505859375
],
[
0.17236328125,
0.468505859375,
0.6708984375,
0.468505859375,
0.6708984375,
0.481689453125,
0.17236328125,
0.481689453125
],
[
0.158447265625,
0.48095703125,
0.88232421875,
0.481689453125,
0.88232421875,
0.496826171875,
0.158447265625,
0.49560546875
],
[
0.1732177734375,
0.49755859375,
0.24267578125,
0.49755859375,
0.24267578125,
0.5087890625,
0.1732177734375,
0.5087890625
],
[
0.11517333984375,
0.52587890625,
0.88232421875,
0.52587890625,
0.88232421875,
0.54052734375,
0.11517333984375,
0.54052734375
],
[
0.11517333984375,
0.54052734375,
0.8837890625,
0.54052734375,
0.8837890625,
0.5537109375,
0.11517333984375,
0.5537109375
],
[
0.11517333984375,
0.55615234375,
0.20751953125,
0.5537109375,
0.20751953125,
0.56494140625,
0.11517333984375,
0.5673828125
],
[
0.11358642578125,
0.57373046875,
0.88134765625,
0.57470703125,
0.88134765625,
0.5888671875,
0.11358642578125,
0.58837890625
],
[
0.11517333984375,
0.59033203125,
0.88232421875,
0.59033203125,
0.88232421875,
0.60400390625,
0.11517333984375,
0.60400390625
],
[
0.11517333984375,
0.60400390625,
0.88232421875,
0.60400390625,
0.88232421875,
0.6171875,
0.11517333984375,
0.6171875
],
[
0.11517333984375,
0.6171875,
0.88232421875,
0.6171875,
0.88232421875,
0.630859375,
0.11517333984375,
0.630859375
],
[
0.11517333984375,
0.630859375,
0.88232421875,
0.630859375,
0.88232421875,
0.64453125,
0.11517333984375,
0.64453125
],
[
0.1168212890625,
0.646484375,
0.88134765625,
0.646484375,
0.88134765625,
0.6572265625,
0.1168212890625,
0.6572265625
],
[
0.11517333984375,
0.6591796875,
0.88232421875,
0.6591796875,
0.88232421875,
0.67236328125,
0.11517333984375,
0.67236328125
],
[
0.11517333984375,
0.673828125,
0.55810546875,
0.673828125,
0.55810546875,
0.6875,
0.11517333984375,
0.6875
],
[
0.11517333984375,
0.693359375,
0.88232421875,
0.693359375,
0.88232421875,
0.70703125,
0.11517333984375,
0.70703125
],
[
0.11358642578125,
0.70849609375,
0.88134765625,
0.70703125,
0.88134765625,
0.7216796875,
0.11358642578125,
0.72265625
],
[
0.11517333984375,
0.7216796875,
0.88134765625,
0.7216796875,
0.88134765625,
0.73486328125,
0.11517333984375,
0.73486328125
],
[
0.11358642578125,
0.73486328125,
0.88232421875,
0.73486328125,
0.88232421875,
0.74853515625,
0.11358642578125,
0.74853515625
],
[
0.11358642578125,
0.75,
0.5185546875,
0.75,
0.5185546875,
0.76318359375,
0.11358642578125,
0.76318359375
],
[
0.1168212890625,
0.77099609375,
0.88232421875,
0.77099609375,
0.88232421875,
0.7841796875,
0.1168212890625,
0.7841796875
],
[
0.11517333984375,
0.7841796875,
0.8837890625,
0.7841796875,
0.8837890625,
0.798828125,
0.11517333984375,
0.798828125
],
[
0.11358642578125,
0.798828125,
0.53515625,
0.798828125,
0.53515625,
0.8125,
0.11358642578125,
0.8125
],
[
0.396240234375,
0.8291015625,
0.408447265625,
0.8291015625,
0.408447265625,
0.83837890625,
0.396240234375,
0.83837890625
],
[
0.406005859375,
0.8359375,
0.43701171875,
0.8359375,
0.43701171875,
0.8486328125,
0.406005859375,
0.8486328125
],
[
0.5146484375,
0.83203125,
0.521484375,
0.83203125,
0.521484375,
0.83544921875,
0.5146484375,
0.83544921875
],
[
0.423095703125,
0.83837890625,
0.5556640625,
0.8291015625,
0.5595703125,
0.85791015625,
0.42724609375,
0.8671875
],
[
0.56787109375,
0.83837890625,
0.58251953125,
0.83837890625,
0.58251953125,
0.8466796875,
0.56787109375,
0.8466796875
],
[
0.5888671875,
0.84228515625,
0.61669921875,
0.84228515625,
0.61669921875,
0.85498046875,
0.5888671875,
0.85498046875
],
[
0.609375,
0.84228515625,
0.6640625,
0.84228515625,
0.6640625,
0.85693359375,
0.609375,
0.85693359375
],
[
0.564453125,
0.85400390625,
0.58642578125,
0.85400390625,
0.58642578125,
0.8662109375,
0.564453125,
0.8662109375
],
[
0.330078125,
0.8369140625,
0.414306640625,
0.8369140625,
0.414306640625,
0.859375,
0.330078125,
0.859375
],
[
0.410888671875,
0.85791015625,
0.4306640625,
0.85107421875,
0.43701171875,
0.861328125,
0.418212890625,
0.86865234375
],
[
0.496826171875,
0.85693359375,
0.5498046875,
0.84765625,
0.5556640625,
0.8662109375,
0.50244140625,
0.87548828125
],
[
0.58251953125,
0.859375,
0.62109375,
0.859375,
0.62109375,
0.87255859375,
0.58251953125,
0.87255859375
],
[
0.491943359375,
0.896484375,
0.50732421875,
0.896484375,
0.50732421875,
0.9111328125,
0.491943359375,
0.9111328125
],
[
0.390625,
0.859375,
0.418212890625,
0.859375,
0.418212890625,
0.87060546875,
0.390625,
0.87060546875
]
] |
[
"very fast iterations. The problem stems from the fact that if one needs to perform a round of communication after",
"each iteration, practical performance drops down dramatically, as the round of communication is much more time-",
"consuming than a single iteration of the algorithm.",
"Works including [2, 8, 3, 9] have established basic theoretical results and developed distributed versions of existing",
"methods, but as this is a relatively new area, many open questions remain. Recently, the idea of communication",
"efficientalgorithmshasgainedtraction.Suchalgorithmsperformalargeamountof computationlocally,beforeeach",
"round of communication, ideally balancing the cost of computation and communication[10, 11, 4, 7, 14]. There has",
"beenanattemptatathoroughunderstandingof communicationlowerbounds[ll.buttheauthorsnotethat significant",
"gaps remain in a number of settings.",
"2",
"Federated OptimizationThe Challenge of Learning from Decentralized Data",
"With the exception of CoCoA [4, 7], the existing work concerning communication efficient optimization [10, 14]",
"and/or statistical estimation [13, 12] presupposes, and in fact requires, that the number of nodes is (much) smaller than",
"the number of datapoints available to each node, and that each node has access to a random sample from the same",
"distribution, and typically that each node has an identical number of data points.",
"We are hoping to bring to attention a new and increasingly relevant setting for distributed optimization, where typically",
"none of the above assumption are satisfied, and communication efficiency is of utmost importance. In particular,",
"algorithms for federated optimization must handle training data with the following characteristics:",
". Massively Distributed: Data points are stored across a very large number of nodes K. In particular, the",
"number of nodes can be much bigger than the average number of training examples stored on a given node",
"(n/k).",
". Non-IlD: Data on each node may be drawn from a different distribution; that is, the data points available",
"locally are far from being a representative sample of the overall distribution",
".Unbalanced: Different nodes may vary by many orders of magnitude in the number of training examples",
"they hold.",
"In this work, we are particularly concerned with sparse data, where some features occur on only a small subset of",
"nodes or data points. We show that the sparsity structure can be used to develop an effective algorithm for federated",
"optimization.",
"We are motivated by the setting where training data lives on users' mobile devices (phones and tablets), and the data",
"may be privacy sensitive. The data (ci, y:)\"-1 are generated by device usage, e.g. interaction with apps. Examples",
"include predicting the next word a user will type (language modeling for smarter keyboard apps), predicting which",
"photos a user is most likely to share, or predicting which notifications are most important. To train such models",
"using traditional distributed learning algorithms, one would collect the training examples in a centralized location",
"(data center)whereitcouldbeshuffled and distributed evenlyover computenodes.Federated optimizationprovides",
"an alternative model potentially saving significant network bandwidth and providing additional privacy protection. In",
"exchange, users allow some use of their devices' computing power.",
"Communication constraints arise naturally in the massively distributed setting, as network connectivity may be limited",
"(e.g., we may wish to defer all communication until the mobile device is charging and connected to a wi-fi network)",
"Thus, in realistic scenarios we may be limited to only a single round of communication per day. This implies that,",
"within reasonable bounds, we have access to essentially unlimited local computational power. As a result, the practical",
"objective is solely to minimize the rounds of communication.",
"Formally, let K be the number of nodes. Let [P]K-, denote a partition of data point indices [1, ..., n], so Pk is",
"the set of data points stored on node k, and define nk = Pkl. We can then rephrase the objective (1) as a linear",
"combination of the local empirical objectives Fr(w), defined as",
"K",
"nk",
"",
"Fa(u）",
"1",
"",
"fi(w).",
"nk",
"f(w)=",
"n",
"k=in",
"iePh",
"2",
"k=1"
] |
{
"arxiv_id": "1511.03575",
"language": "en",
"page_idx": 1
}
|
||
1511.03575-en-2
|
1511.03575-en-2.jpg
|
3 $\textbf{An Algorithm for Federated Optimization}$
The main motivation for the algorithm we propose comes from a perhaps surprising connection between two algorithms—SVRG[5,6] and DANE[10]. SVRG is a stochastic method with variance reduction for solving the problem (1) on a single node. DANE is an algorithm for distributed optimization which on every iteration solves exactly a new subproblem on each node based on the local data and the gradient of the entire function $f.$
<table>
<tbody>
<tr>
<td>$1ne$</td>
<td> </td>
</tr>
<tr>
<td>$ny$ 0</td>
<td>mizarim algorıthm to produce an anproxımate L.In narticula :if we use SVRG as the local solver. t]</td>
</tr>
<tr>
<td>sque</td>
<td>on of distrihuted SVRC</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<th>$\overline{\mathbf{Alg}}$</th>
<th>$Pf$ version of SVRG</th>
<th> </th>
</tr>
<tr>
<th>1: 2:</th>
<th>parameters: $m=\#$ of stochastic steps per epoch, $h=$stepsize, dat $\textbf{for}s= 0, 1, 2, \ldots \textbf{do}$</th>
<th>artition $\{\mathcal{P}_{k}\}_{k=1}^{K}$, starting point $\tilde{w}_{0}$ - $\triangleright{\mathrm{Overall~iterations}}$</th>
</tr>
<tr>
<td>3: 4:</td>
<td>Compute V$f(\ddot{w}_{s})$ $w_{\mathrm{s}}$ $\textbf{for }k= 1$ to $K\textbf{do in parallel over nodes }k$</td>
<td>$\triangleright$Distributed loop</td>
</tr>
<tr>
<td>$\underline{2.}$</td>
<td>$w_{k}$ = $:w_{s}$</td>
<td> </td>
</tr>
<tr>
<td>7:</td>
<td>Yamnle $\bullet j\in^{\prime}D$ uniformlva at random</td>
<td> </td>
</tr>
<tr>
<td>8:</td>
<td>$-w_{1}-h/Vf.(w)$ $/f.(\eta)+$ $\nabla f(\tilde{w}.)$ 1/11</td>
<td> </td>
</tr>
<tr>
<td>9:</td>
<td>end for</td>
<td> </td>
</tr>
<tr>
<td>10:</td>
<td>end for</td>
<td> </td>
</tr>
<tr>
<td>1 1 T 1 </td>
<td>$\tilde{w}\leftarrow\tilde{w}-$ $-\tilde{w})$</td>
<td>$\triangleright$Aggregate updates from nodes</td>
</tr>
<tr>
<td>12:</td>
<td>$\mathbf{endfor}$</td>
<td> </td>
</tr>
</tbody>
</table>
The algorithm l indeed works very well in many “simple”settings, but fails in the challenging setting of federated
optimization, particularly with sparse data. In order to make the algorithm more robust, we modify it in a number of
ways. The most important changes include
1. Flexible stepsize $h_k-$different for each node, inversely proportional to size of the local data, $n_k$
2. Scaling of stochastic updates by a diagonal matrix $S_k$ (defined below). Step 8 in Algorithm 1 is replaced by
$w_{k}=\tilde{w}_{k}-h_{k}\left(S_{k}\left[\tilde{\nabla f}_{i}(w_{k})-\tilde{\nabla f}_{i}(\tilde{w})\right]+\nabla f(\tilde{w})\right)$
3. The aggregation procedure is adaptive to the data distribution. For some diagonal matrix $A$ (defined below),
the step 11 in Algorithm 1 is replaced by $\tilde{w}=\tilde{w}+A\sum_{k=1}^{K}\frac{n_{k}}{n}(w_{k}-\tilde{w})$
The matrices $S_k$, and $A$ concern sparsity patterns in the data, and are identities in the case of fully dense data. To motivate their use, imagine that the data are distributed in the following way. All of the data points that have non-zero value of a certain feature are stored on a single node $k.$ Consequently, computing stochastic gradient $\nabla f_i(w_k)$ will greatly overestimate the gradient in this direction, because it appears much more frequently on node $k$ than in the entire dataset. Diagonal elements of the matrix $S_k$ are ratios of frequencies of appearance of each feature globally and locally. One can interpret this as scaling the estimates so that they are of the correct magnitude in expectation, taken conditioned on knowledge of the distribution of the sparsity patterns.
In order to motivate use of the matrix $A$, let us assume our problem is separable in the sense that each data point depends only on features belonging to one of a few disjoint clusters of coordinates, and data are distributed according to these clusters. In the case of linear predictors, we could simply solve the problems independently, and add up the results, instead of averaging, to obtain the optimal solution. Although this in not the case in reality, we try to interpolate, and average updates for features that appear on every node, while we take longer steps for features that appear only on a few nodes. Formally, $A_ii=K/\omega_i$, where $\omega_i$ is the number of nodes on which feature $i$ is present. Although this is a heuristic modification, it's omission greatly degrades performance of the algorithm.
## 4 $\textbf{Experiments}$
The dataset presented here was generated based on public posts on a large social network. We randomly picked 10,000 authors that have at least 100 public posts in English, and try to predict whether a post will receive at least one comment (that is, a binary classification task).
3
|
[
[
0.11517333984375,
0.091552734375,
0.496826171875,
0.091552734375,
0.496826171875,
0.10479736328125,
0.11517333984375,
0.10479736328125
],
[
0.1168212890625,
0.121826171875,
0.88134765625,
0.121826171875,
0.88134765625,
0.1351318359375,
0.1168212890625,
0.1351318359375
],
[
0.11517333984375,
0.1363525390625,
0.88232421875,
0.1363525390625,
0.88232421875,
0.149658203125,
0.11517333984375,
0.149658203125
],
[
0.1168212890625,
0.149658203125,
0.88134765625,
0.149658203125,
0.88134765625,
0.1634521484375,
0.1168212890625,
0.1634521484375
],
[
0.11358642578125,
0.1622314453125,
0.705078125,
0.1634521484375,
0.705078125,
0.177978515625,
0.11358642578125,
0.1767578125
],
[
0.11358642578125,
0.18310546875,
0.88232421875,
0.184326171875,
0.88232421875,
0.1988525390625,
0.11358642578125,
0.1976318359375
],
[
0.11517333984375,
0.1988525390625,
0.88232421875,
0.1988525390625,
0.88232421875,
0.212158203125,
0.11517333984375,
0.212158203125
],
[
0.11358642578125,
0.21337890625,
0.6396484375,
0.21337890625,
0.6396484375,
0.2266845703125,
0.11358642578125,
0.2266845703125
],
[
0.1168212890625,
0.2412109375,
0.41015625,
0.2412109375,
0.41015625,
0.255126953125,
0.1168212890625,
0.255126953125
],
[
0.12335205078125,
0.261474609375,
0.82861328125,
0.261474609375,
0.82861328125,
0.27587890625,
0.12335205078125,
0.27587890625
],
[
0.12091064453125,
0.272705078125,
0.289306640625,
0.272705078125,
0.289306640625,
0.285888671875,
0.12091064453125,
0.285888671875
],
[
0.75244140625,
0.274658203125,
0.88134765625,
0.274658203125,
0.88134765625,
0.285400390625,
0.75244140625,
0.285400390625
],
[
0.12335205078125,
0.290283203125,
0.136474609375,
0.290283203125,
0.136474609375,
0.2998046875,
0.12335205078125,
0.2998046875
],
[
0.1666259765625,
0.285400390625,
0.426513671875,
0.285400390625,
0.426513671875,
0.302978515625,
0.1666259765625,
0.302978515625
],
[
0.1650390625,
0.300537109375,
0.453369140625,
0.2998046875,
0.453369140625,
0.314453125,
0.1650390625,
0.31494140625
],
[
0.7607421875,
0.300537109375,
0.8837890625,
0.300537109375,
0.8837890625,
0.31494140625,
0.7607421875,
0.31494140625
],
[
0.191162109375,
0.314453125,
0.3203125,
0.314453125,
0.3203125,
0.32763671875,
0.191162109375,
0.32763671875
],
[
0.12335205078125,
0.302978515625,
0.1380615234375,
0.302978515625,
0.1380615234375,
0.313232421875,
0.12335205078125,
0.313232421875
],
[
0.1217041015625,
0.31494140625,
0.1396484375,
0.31494140625,
0.1396484375,
0.330078125,
0.1217041015625,
0.330078125
],
[
0.1298828125,
0.32568359375,
0.143798828125,
0.3359375,
0.131591796875,
0.3466796875,
0.1168212890625,
0.3359375
],
[
0.19287109375,
0.330810546875,
0.313720703125,
0.330810546875,
0.313720703125,
0.341552734375,
0.19287109375,
0.341552734375
],
[
0.74267578125,
0.328857421875,
0.8837890625,
0.328857421875,
0.8837890625,
0.343505859375,
0.74267578125,
0.343505859375
],
[
0.1217041015625,
0.34228515625,
0.1405029296875,
0.34228515625,
0.1405029296875,
0.35791015625,
0.1217041015625,
0.35791015625
],
[
0.2156982421875,
0.34228515625,
0.45751953125,
0.341552734375,
0.45751953125,
0.35595703125,
0.2156982421875,
0.356689453125
],
[
0.12091064453125,
0.356689453125,
0.1405029296875,
0.356689453125,
0.1405029296875,
0.371826171875,
0.12091064453125,
0.371826171875
],
[
0.2156982421875,
0.356689453125,
0.5390625,
0.356689453125,
0.5390625,
0.371826171875,
0.2156982421875,
0.371826171875
],
[
0.12091064453125,
0.37060546875,
0.1396484375,
0.37060546875,
0.1396484375,
0.3837890625,
0.12091064453125,
0.3837890625
],
[
0.19287109375,
0.371826171875,
0.2442626953125,
0.371826171875,
0.2442626953125,
0.38330078125,
0.19287109375,
0.38330078125
],
[
0.1168212890625,
0.385009765625,
0.1396484375,
0.385009765625,
0.1396484375,
0.3984375,
0.1168212890625,
0.3984375
],
[
0.1683349609375,
0.3837890625,
0.221435546875,
0.3837890625,
0.221435546875,
0.3984375,
0.1683349609375,
0.3984375
],
[
0.11767578125,
0.3984375,
0.1396484375,
0.3984375,
0.1396484375,
0.412353515625,
0.11767578125,
0.412353515625
],
[
0.1641845703125,
0.396484375,
0.36767578125,
0.396484375,
0.36767578125,
0.417236328125,
0.1641845703125,
0.417236328125
],
[
0.66650390625,
0.3984375,
0.8837890625,
0.397705078125,
0.8837890625,
0.412353515625,
0.66650390625,
0.41357421875
],
[
0.11517333984375,
0.412353515625,
0.1968994140625,
0.410400390625,
0.1968994140625,
0.4248046875,
0.11517333984375,
0.4267578125
],
[
0.11517333984375,
0.443115234375,
0.88232421875,
0.443115234375,
0.88232421875,
0.45703125,
0.11517333984375,
0.45703125
],
[
0.11517333984375,
0.458251953125,
0.8837890625,
0.458251953125,
0.8837890625,
0.4716796875,
0.11517333984375,
0.4716796875
],
[
0.11517333984375,
0.4716796875,
0.400390625,
0.4716796875,
0.400390625,
0.48486328125,
0.11517333984375,
0.48486328125
],
[
0.156005859375,
0.49560546875,
0.810546875,
0.49560546875,
0.810546875,
0.51025390625,
0.156005859375,
0.51025390625
],
[
0.1544189453125,
0.51220703125,
0.8798828125,
0.51220703125,
0.8798828125,
0.52294921875,
0.1544189453125,
0.52294921875
],
[
0.1748046875,
0.5244140625,
0.5185546875,
0.5244140625,
0.5185546875,
0.53857421875,
0.1748046875,
0.53857421875
],
[
0.1519775390625,
0.53857421875,
0.88232421875,
0.53857421875,
0.88232421875,
0.5517578125,
0.1519775390625,
0.5517578125
],
[
0.1707763671875,
0.55029296875,
0.65673828125,
0.55322265625,
0.65673828125,
0.57080078125,
0.1707763671875,
0.568359375
],
[
0.1168212890625,
0.5791015625,
0.88232421875,
0.5791015625,
0.88232421875,
0.59326171875,
0.1168212890625,
0.59326171875
],
[
0.1168212890625,
0.59521484375,
0.88134765625,
0.59521484375,
0.88134765625,
0.60595703125,
0.1168212890625,
0.60595703125
],
[
0.1168212890625,
0.60693359375,
0.88134765625,
0.60693359375,
0.88134765625,
0.62060546875,
0.1168212890625,
0.62060546875
],
[
0.11517333984375,
0.6220703125,
0.88134765625,
0.6220703125,
0.88134765625,
0.63623046875,
0.11517333984375,
0.63623046875
],
[
0.11517333984375,
0.6337890625,
0.88134765625,
0.6337890625,
0.88134765625,
0.64794921875,
0.11517333984375,
0.64794921875
],
[
0.11517333984375,
0.6494140625,
0.88232421875,
0.6494140625,
0.88232421875,
0.66357421875,
0.11517333984375,
0.66357421875
],
[
0.11358642578125,
0.66357421875,
0.56884765625,
0.66357421875,
0.56884765625,
0.6767578125,
0.11358642578125,
0.6767578125
],
[
0.11517333984375,
0.6845703125,
0.88232421875,
0.6845703125,
0.88232421875,
0.69775390625,
0.11517333984375,
0.69775390625
],
[
0.1168212890625,
0.69970703125,
0.8798828125,
0.69970703125,
0.8798828125,
0.71044921875,
0.1168212890625,
0.71044921875
],
[
0.1168212890625,
0.7119140625,
0.88232421875,
0.7119140625,
0.88232421875,
0.72607421875,
0.1168212890625,
0.72607421875
],
[
0.11517333984375,
0.72607421875,
0.88232421875,
0.72607421875,
0.88232421875,
0.74072265625,
0.11517333984375,
0.74072265625
],
[
0.11517333984375,
0.7392578125,
0.88134765625,
0.7392578125,
0.88134765625,
0.75244140625,
0.11517333984375,
0.75244140625
],
[
0.11517333984375,
0.75390625,
0.88232421875,
0.75390625,
0.88232421875,
0.767578125,
0.11517333984375,
0.767578125
],
[
0.11517333984375,
0.767578125,
0.78173828125,
0.767578125,
0.78173828125,
0.78076171875,
0.11517333984375,
0.78076171875
],
[
0.11358642578125,
0.8017578125,
0.253173828125,
0.8017578125,
0.253173828125,
0.81640625,
0.11358642578125,
0.81640625
],
[
0.11517333984375,
0.83203125,
0.8837890625,
0.83203125,
0.8837890625,
0.84521484375,
0.11517333984375,
0.84521484375
],
[
0.11517333984375,
0.8466796875,
0.8837890625,
0.8466796875,
0.8837890625,
0.85986328125,
0.11517333984375,
0.85986328125
],
[
0.11517333984375,
0.85986328125,
0.419189453125,
0.85986328125,
0.419189453125,
0.87353515625,
0.11517333984375,
0.87353515625
],
[
0.49267578125,
0.8984375,
0.50634765625,
0.8984375,
0.50634765625,
0.91015625,
0.49267578125,
0.91015625
]
] |
[
"3AnAlgorithmforFederatedOptimization",
"The main motivation for the algorithm we propose comes from a perhaps surprising connection between two algo-",
"rithmsSVRG [5, 6] and DANE [10]. SVRG is a stochastic method with variance reduction for solving the problem",
"(1) on a single node. DANE is an algorithm for distributed optimization which on every iteration solves exactly a new",
"subproblem on each node based on the local data and the gradient of the entire function f.",
"One could modify the DANE algorithm in the following way: instead of solving the DANE subproblem exactly, use",
"any optimization algorithm to produce an approximate solution. In particular, if we use SVRG as the local solver, the",
"sequence of updates is equivalent to the following version of distributed SVRG.",
"Algorithm1Adistributed version of SVRG",
"I: parameters: m = # of stochastic steps per epoch, h = stepsize, data partition [P]k-1, starting point wo",
"2: for s = 0,1,2.... do",
"Overalliterations",
"3:",
"Compute Vf(ws) =Zr- Vfi(ws)",
"for k = 1 to K do in parallel over nodes k",
" Distributed loop",
"Initialize: wk = w.",
"4:",
"5:",
"6.",
"fort=ltomdo",
" Actual update loop",
"7:",
"Sample i e Pk uniformly at random",
"8:",
"wkwk-h(Vf(wk) -Vf;(ws)+Vf(ws))",
"9:",
"end for",
"10:",
"end for",
"11:",
"ww+k-i(wk-)",
">Aggregate updates from nodes",
"12:end for",
"The algorithm 1 indeed works very well in many “simple\" settings, but fails in the challenging setting of federated",
"optimization, particularly with sparse data. In order to make the algorithm more robust, we modify it in a number of",
"ways. The most important changes include",
"1. Flexible stepsize hk different for each node, inversely proportional to size of the local data, nk",
"2.Scalingof stochasticupdatesbyadiagonalmatrixSk(definedbelow).Step8inAlgorithm1isreplacedby",
"wk= Wk-h (S[Vfi(wk)-f:(w)] +Vf(w))",
"3. The aggregation procedure is adaptive to the data distribution. For some diagonal matrix A (defined below),",
"the step 1l in Algorithm 1 is replaced by w = w + Ak=\" (wk -w)",
"The matrices Sk, and A concern sparsity patterns in the data, and are identities in the case of fully dense data. To",
"motivate their use,imagine that the data are distributed in the following way.All of the datapoints that have non-zero",
"value of a certain feature are stored on a single node k. Consequently, computing stochastic gradient f;(w) will",
"greatly overestimate the gradient in this direction, because it appears much more frequently on node k than in the",
"entire dataset. Diagonal elements of the matrix Sk are ratios of frequencies of appearance of each feature globally and",
"locally. One can interpret this as scaling the estimates so that they are of the correct magnitude in expectation, taken",
"conditioned on knowledge of the distribution of the sparsity patterns.",
"In order to motivate use of the matrix A, let us assume our problem is separable in the sense that each data point",
"depends onlyonfeaturesbelongingto one ofafewdisjointclusters of coordinates,and dataaredistributed according",
"to these clusters.In the case of linear predictors, we could simply solve the problems independently, and add up",
"the results, instead of averaging, to obtain the optimal solution. Although this in not the case in reality, we try to",
"interpolate, and average updates for features that appear on every node, while we take longer steps for features that",
"appear only on a few nodes. Formally, Az = K/wi, where w, is the number of nodes on which feature i is present.",
"Although this is a heuristic modification, it's omission greatly degrades performance of the algorithm.",
"4 Experiments",
"The dataset presented here was generated based on public posts on a large social network.We randomly picked",
"10, 000 authors that have at least 100 public posts in English, and try to predict whether a post will receive at least one",
"comment (that is, a binary classification task).",
"3"
] |
{
"arxiv_id": "1511.03575",
"language": "en",
"page_idx": 2
}
|
||
1511.03575-en-3
|
1511.03575-en-3.jpg
|
We split the data chronologically on a per-author basis, taking the earlier $75\%$ for training and the following 25% for testing. The total number of training examples is $n=2,166,693.$ We created a simple bag-of-words language model, based on the 20,000 most frequent words. We then use a logistic regression model to make a prediction based on these features.
We shape the distributed optimization problem as follows. Suppose that each user corresponds to one node, resulting in $\hat{K} = 10, 000. \hat {\text{ The average }n_k}, \hat {\text{number of data points on node }k}, \hat {\text{is thus approximately }216}.$ However, the actua$\tilde{\text{I}}$ numbers $n_k$ range from 75 to 9,000,showing the data distribution is in fact substantially unbalanced. It is natural to expect that different users can exhibit very different patterns in the data generated. This is indeed the case, and hence the distribution to nodes cannot be considered an IID sample from the overall distribution.
<FigureHere>
Figure 1: Rounds of communication vs. objective function (left) and test prediction error (right).
In Figure 1,we compare various optimization algorithms. Two other communication efficient algorithms, DANE[10]
and $\tilde{\mathrm{DiSCO}}[14],\tilde{\mathrm{diverge~in~this~setting,~and~so~are~not~included~in~the~figure.}}$
$\bullet$ The blue squares (OPT) represent the best possible offline value (the optimal value of the optimization task
in the first plot, and the test error corresponding to the optimum in the second plot).
$\bullet$ The teal diamonds (GD) correspond to a simple distributed gradient descent.
$\bullet$ The purple triangles (COCOA) are for the CoCoA algorithm [4].
$\bullet$ The green circles (SVRGfo) give values for our proposed algorithm.
$\bullet$ The red stars (SVRGfoR) correspond to the same algorithm applied to the same problem with randomly
reshuffled data. That is, we keep the unbalanced number of examples per node, but populate each node with
randomly selected examples.
The experiments point at several important realities. First, existing communication efficient algorithms are completely inefficient for federated optimization. The only algorithm that converges, CoCoA, does so at a slower rate than a naive benchmark-distributed gradient descent. Our algorithm, DSVRG, reaches optimality in very few communication rounds, despite the challenging setting of federated optimization. Furthermore, very little performance is lost compared to the setting where the data are randomly reshuffled between nodes, which highlights the robustness of our proposede to the setting are the $a$ are randomly reshuffled between nodes, which highlights the robustness of our proposede ting $a$ method to non-IID data.
We argue that the setting of federated optimization will be increasingly important for practical application, as mobile devices get computationally more powerful and privacy issues ever more pressing. A large number of question remaim open, creating potential for a line of further work. Most importantly, there is no readily available dataset (from usual sources as libsvm or UCI repository) with naturally user-clustered data. Creation of a new, public dataset is crucial to lower the barrier of further engagement of the academic community. More rigorous experiments remain to be done, both on larger datasets and on more challenging problems, such as deep learning. Currently, there is no proper theoretical justification of the method, which would surely drive further improvements. Lastly, in order for this shift to“on-device intelligence”to truly happen, methods from differential privacy need to be incorporated and many parts of practical machine learning pipelines need to be redesigned.
4
|
[
[
0.1168212890625,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.1168212890625,
0.10797119140625,
0.88134765625,
0.10797119140625,
0.88134765625,
0.11871337890625,
0.1168212890625,
0.11871337890625
],
[
0.11517333984375,
0.12060546875,
0.88134765625,
0.12060546875,
0.88134765625,
0.1337890625,
0.11517333984375,
0.1337890625
],
[
0.1168212890625,
0.1363525390625,
0.17236328125,
0.1363525390625,
0.17236328125,
0.147705078125,
0.1168212890625,
0.147705078125
],
[
0.11358642578125,
0.154052734375,
0.88232421875,
0.1546630859375,
0.88232421875,
0.169189453125,
0.11358642578125,
0.1685791015625
],
[
0.1168212890625,
0.17041015625,
0.88134765625,
0.17041015625,
0.88134765625,
0.18115234375,
0.1168212890625,
0.18115234375
],
[
0.11517333984375,
0.18310546875,
0.88232421875,
0.18310546875,
0.88232421875,
0.1962890625,
0.11517333984375,
0.1962890625
],
[
0.11273193359375,
0.1962890625,
0.8837890625,
0.1956787109375,
0.8837890625,
0.210205078125,
0.11273193359375,
0.2108154296875
],
[
0.1168212890625,
0.212158203125,
0.70166015625,
0.212158203125,
0.70166015625,
0.222900390625,
0.1168212890625,
0.222900390625
],
[
0.148681640625,
0.272705078125,
0.16259765625,
0.272705078125,
0.16259765625,
0.281005859375,
0.148681640625,
0.281005859375
],
[
0.1405029296875,
0.323974609375,
0.146240234375,
0.323974609375,
0.146240234375,
0.333251953125,
0.1405029296875,
0.333251953125
],
[
0.1405029296875,
0.333984375,
0.146240234375,
0.333984375,
0.146240234375,
0.34033203125,
0.1405029296875,
0.34033203125
],
[
0.15283203125,
0.3359375,
0.16259765625,
0.3359375,
0.16259765625,
0.339111328125,
0.15283203125,
0.339111328125
],
[
0.78515625,
0.41845703125,
0.7998046875,
0.41845703125,
0.7998046875,
0.4267578125,
0.78515625,
0.4267578125
],
[
0.268798828125,
0.427978515625,
0.30224609375,
0.427978515625,
0.30224609375,
0.43115234375,
0.268798828125,
0.43115234375
],
[
0.30224609375,
0.427978515625,
0.31787109375,
0.427978515625,
0.31787109375,
0.43115234375,
0.30224609375,
0.43115234375
],
[
0.330078125,
0.427978515625,
0.365234375,
0.427978515625,
0.365234375,
0.43115234375,
0.330078125,
0.43115234375
],
[
0.646484375,
0.4248046875,
0.74609375,
0.4248046875,
0.74609375,
0.431884765625,
0.646484375,
0.431884765625
],
[
0.1846923828125,
0.449462890625,
0.814453125,
0.449462890625,
0.814453125,
0.46337890625,
0.1846923828125,
0.46337890625
],
[
0.11767578125,
0.488037109375,
0.56787109375,
0.488037109375,
0.56787109375,
0.498779296875,
0.11767578125,
0.498779296875
],
[
0.56787109375,
0.4873046875,
0.88134765625,
0.4873046875,
0.88134765625,
0.50048828125,
0.56787109375,
0.50048828125
],
[
0.11358642578125,
0.5,
0.615234375,
0.50048828125,
0.615234375,
0.51513671875,
0.11358642578125,
0.5146484375
],
[
0.158447265625,
0.53076171875,
0.8837890625,
0.53076171875,
0.8837890625,
0.54443359375,
0.158447265625,
0.54443359375
],
[
0.1707763671875,
0.54443359375,
0.72119140625,
0.54541015625,
0.72119140625,
0.5595703125,
0.1707763671875,
0.55810546875
],
[
0.158447265625,
0.55810546875,
0.6748046875,
0.55810546875,
0.6748046875,
0.5712890625,
0.158447265625,
0.5712890625
],
[
0.158447265625,
0.57275390625,
0.59716796875,
0.57275390625,
0.59716796875,
0.5859375,
0.158447265625,
0.5859375
],
[
0.16015625,
0.5869140625,
0.619140625,
0.5869140625,
0.619140625,
0.60107421875,
0.16015625,
0.60107421875
],
[
0.158447265625,
0.599609375,
0.8798828125,
0.599609375,
0.8798828125,
0.61279296875,
0.158447265625,
0.61279296875
],
[
0.1748046875,
0.61572265625,
0.88232421875,
0.61572265625,
0.88232421875,
0.62890625,
0.1748046875,
0.62890625
],
[
0.17236328125,
0.62744140625,
0.365234375,
0.62890625,
0.365234375,
0.642578125,
0.17236328125,
0.6416015625
],
[
0.1168212890625,
0.6591796875,
0.88134765625,
0.6591796875,
0.88134765625,
0.67236328125,
0.1168212890625,
0.67236328125
],
[
0.1168212890625,
0.6748046875,
0.88134765625,
0.6748046875,
0.88134765625,
0.68505859375,
0.1168212890625,
0.68505859375
],
[
0.1168212890625,
0.68798828125,
0.193603515625,
0.68798828125,
0.193603515625,
0.69873046875,
0.1168212890625,
0.69873046875
],
[
0.2091064453125,
0.68603515625,
0.88232421875,
0.68603515625,
0.88232421875,
0.69970703125,
0.2091064453125,
0.69970703125
],
[
0.11358642578125,
0.69970703125,
0.88232421875,
0.69970703125,
0.88232421875,
0.71337890625,
0.11358642578125,
0.71337890625
],
[
0.11358642578125,
0.71533203125,
0.8837890625,
0.71533203125,
0.8837890625,
0.72900390625,
0.11358642578125,
0.72900390625
],
[
0.11517333984375,
0.72900390625,
0.275390625,
0.72900390625,
0.275390625,
0.7421875,
0.11517333984375,
0.7421875
],
[
0.1168212890625,
0.75,
0.88232421875,
0.75,
0.88232421875,
0.76318359375,
0.1168212890625,
0.76318359375
],
[
0.11358642578125,
0.76318359375,
0.88232421875,
0.76318359375,
0.88232421875,
0.7763671875,
0.11358642578125,
0.7763671875
],
[
0.11358642578125,
0.77783203125,
0.8837890625,
0.7763671875,
0.8837890625,
0.79052734375,
0.11358642578125,
0.79150390625
],
[
0.11517333984375,
0.79150390625,
0.88232421875,
0.79150390625,
0.88232421875,
0.8046875,
0.11517333984375,
0.8046875
],
[
0.11358642578125,
0.80615234375,
0.8837890625,
0.80615234375,
0.8837890625,
0.8193359375,
0.11358642578125,
0.8193359375
],
[
0.1168212890625,
0.82080078125,
0.88134765625,
0.82080078125,
0.88134765625,
0.83056640625,
0.1168212890625,
0.83056640625
],
[
0.11517333984375,
0.83349609375,
0.88232421875,
0.83349609375,
0.88232421875,
0.8466796875,
0.11517333984375,
0.8466796875
],
[
0.11517333984375,
0.8466796875,
0.8837890625,
0.8466796875,
0.8837890625,
0.85986328125,
0.11517333984375,
0.85986328125
],
[
0.11358642578125,
0.859375,
0.521484375,
0.85986328125,
0.521484375,
0.875,
0.11358642578125,
0.87353515625
],
[
0.494384765625,
0.8994140625,
0.50634765625,
0.8994140625,
0.50634765625,
0.91015625,
0.494384765625,
0.91015625
]
] |
[
"We split the data chronologically on a per-author basis, taking the earlier 75% for training and the following 25% for",
"testing.Thetotalnumberoftrainingexamplesisn=2166,693.Wecreatedasimplebag-of-wordslanguagemodel,",
"based on the 20, 000 most frequent words. We then use a logistic regression model to make a prediction based on these",
"features.",
"We shape the distributed optimization problem as follows. Suppose that each user corresponds to one node, resulting",
"inK =10,000.The average nk,number of datapoints onnodek,is thus approximately216.However,theactual",
"numbers nk range from 75 to 9, 000, showing the data distribution is in fact substantially unbalanced. It is natural to",
"expect that different users can exhibit very different patterns in the data generated. This is indeed the case, and hence",
"thedistributionto nodes cannotbe considered anIID samplefrom theoverall distribution",
"0.68",
"JUC",
"8",
"n",
"25",
"ntunso",
"n",
"incnor",
"RoundsofCommunication",
"Figure I: Rounds of communication vs. objective function (left) and test prediction error (right).",
"InFigurel,wecomparevariousoptimizationalgorithms.Twoother",
"communication efficientalgorithms,DANE [10]",
"and DisCO[14], diverge in this setting, and so are not included in the figure.",
". The blue squares (OPT) represent the best possible offline value (the optimal value of the optimization task",
"in the first plot, and the test error corresponding to the optimum in the second plot).",
". The teal diamonds (GD) correspond to a simple distributed gradient descent.",
". The purple triangles (COCOA) are for the CoCoA algorithm [4]",
"·The green circles (SVRGfo) give values for our proposed algorithm",
". The red stars (SVRGfoR) correspond to the same algorithm applied to the same problem with randomly",
"reshuffled data. That is, we keep the unbalanced number of examples per node, but populate each node with",
"randomlyselected examples.",
"The experiments point at several important realities. First, existing communication efficient algorithms are completely",
"inefficientforfederatedoptimization.Theonlyalgorithmthatconverges,CoCoA,doessoataslowerratethananaive",
"benchmark",
"-distributed gradient descent. Our algorithm, DSVRG, reaches optimality in very few communication",
"rounds, despite the challenging setting of federated optimization. Furthermore, very little performance is lost compared",
"to the setting where the data are randomly reshuffled between nodes, which highlights the robustness of our proposed",
"method to non-IID data.",
"We argue that the setting of federated optimization will be increasingly important for practical application, as mobile",
"devices get computationally more powerful and privacy issues evermore pressing.A large number of question remain",
"open, creating potential for a line of further work. Most importantly, there is no readily available dataset (from usual",
"sources as libsvm or UCI repository) with naturally user-clustered data. Creation of a new, public dataset is crucial",
"to lower the barrier of further engagement of the academic community. More rigorous experiments remain to be",
"done,bothonlargerdatasetsand onmorechallengingproblems,suchasdeeplearning.Currently,thereisnoproper",
"theoretical justification of the method, which would surely drive further improvements. Lastly, in order for this shift",
"to “on-device intelligence\" to truly happen, methods from differential privacy need to be incorporated and many parts",
"of practical machine learningpipelines need to be redesigned.",
"4"
] |
{
"arxiv_id": "1511.03575",
"language": "en",
"page_idx": 3
}
|
||
1511.03575-en-4
|
1511.03575-en-4.jpg
|
# References
[1] Yossi Arjevani and Ohad Shamir. Communication complexity of distributed convex learning and optimization.
$arXivpreprintarXiv:1506.01900,2015.$
[2] Maria-Florina Balcan, Avrim Blum, Shai Fine, and Yishay Mansour. Distributed learning, communication com-
plexity and privacy. $arXi\nu \textit{ preprint arXiv: 1204. 3514, 2012. }$
[3] Olivier Fercoq, Zheng Qu, Peter Richtárik, and Martin Takác. Fast distributed coordinate descent for non-
strongly convex losses. In Machine Learning for Signal Processing $(MLSP),2014$ IEEE International Workshop
$on$, pages 1-6. IEEE, 2014.
[4] Martin Jaggi, Virginia Smith, Martin Takác, Jonathan Terhorst, Sanjay Krishnan, Thomas Hofmann, and
Michael I Jordan. Communication-efficient distributed dual coordinate ascent. In Advances in Neural Infor-
mation Processing Systems, pages 3068-3076,2014.
[5] Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In
$AdvancesinNeuralInformationProcessingSystems$,pages 315-323,2013.
[6] Jakub Konečný and Peter Richtárik. Semi-stochastic gradient descent methods. $arXi\nu preprintarXiv:1312.1666$,
2013.
[7] Chenxin Ma, Virginia Smith, Martin Jaggi, Michael I Jordan, Peter Richtárik, and Martin Takáč. Adding vs.
averaging in distributed primal-dual optimization. $arXi\nu preprint$ arXiv:1502.03508,2015.
[8] Peter Richtárik and Martin Takáč. Distributed coordinate descent method for learning with big data. arXiv
preprint $arXiv: 1310. 2059, 2013.$
[9] Ohad Shamir and Nathan Srebro. Distributed stochastic optimization and learning. In Communication, Control,
and Computing $( Allerton) , 2014$ $52nd$ Annual Allerton Conference $on$, pages 850-857.IEEE, 2014.
[10] Ohad Shamir, Nathan Srebro, and Tong Zhang. Communication efficient distributed optimization using an ap-
proximate newton-type method. $arXi\nu preprintarXiv{:}1312.7853,2013.$
[11] Tianbao Yang. Trading computation for communication: Distributed stochastic dual coordinate ascent. In $Ad-$
vances in Neural Information Processing Systems, pages 629-637, 2013.
[12] Yuchen Zhang, John Duchi, Michael I Jordan, and Martin J Wainwright. Information-theoretic lower bounds for
distributed statistical estimation with communication constraints. In Advancesin Neural Information Processing
$Systems$, pages 2328-2336,2013.
[13] Yuchen Zhang, Martin J Wainwright, and John C Duchi. Communication-efficient algorithms for statistical
optimization. In AdvancesinNeural Information Processing Systems, pages 1502-1510,2012.
[14] Yuchen Zhang and Lin Xiao. Communication-efficient distributed optimization of self-concordant empirical
loss. arXiv preprint $arXiv: 1501. 00263, 2015.$
5
|
[
[
0.11767578125,
0.0921630859375,
0.2091064453125,
0.0921630859375,
0.2091064453125,
0.106689453125,
0.11767578125,
0.106689453125
],
[
0.12335205078125,
0.117431640625,
0.8798828125,
0.11614990234375,
0.8798828125,
0.1307373046875,
0.12335205078125,
0.1319580078125
],
[
0.15283203125,
0.1319580078125,
0.41259765625,
0.1319580078125,
0.41259765625,
0.1458740234375,
0.15283203125,
0.1458740234375
],
[
0.125,
0.15087890625,
0.88134765625,
0.15087890625,
0.88134765625,
0.1641845703125,
0.125,
0.1641845703125
],
[
0.1519775390625,
0.1654052734375,
0.53759765625,
0.1654052734375,
0.53759765625,
0.1787109375,
0.1519775390625,
0.1787109375
],
[
0.125,
0.18310546875,
0.88134765625,
0.18310546875,
0.88134765625,
0.1976318359375,
0.125,
0.1976318359375
],
[
0.15283203125,
0.1976318359375,
0.88134765625,
0.1976318359375,
0.88134765625,
0.2108154296875,
0.15283203125,
0.2108154296875
],
[
0.1502685546875,
0.2108154296875,
0.330810546875,
0.210205078125,
0.330810546875,
0.2247314453125,
0.1502685546875,
0.2259521484375
],
[
0.125,
0.2298583984375,
0.88232421875,
0.2298583984375,
0.88232421875,
0.24365234375,
0.125,
0.24365234375
],
[
0.1502685546875,
0.242431640625,
0.88134765625,
0.24365234375,
0.88134765625,
0.25830078125,
0.1502685546875,
0.2568359375
],
[
0.1502685546875,
0.25830078125,
0.498291015625,
0.2568359375,
0.498291015625,
0.271484375,
0.1502685546875,
0.272705078125
],
[
0.125,
0.27587890625,
0.88232421875,
0.27587890625,
0.88232421875,
0.290283203125,
0.125,
0.290283203125
],
[
0.1502685546875,
0.290283203125,
0.64794921875,
0.2890625,
0.64794921875,
0.3037109375,
0.1502685546875,
0.304931640625
],
[
0.125,
0.309326171875,
0.88134765625,
0.309326171875,
0.88134765625,
0.322509765625,
0.125,
0.322509765625
],
[
0.15283203125,
0.323974609375,
0.188720703125,
0.323974609375,
0.188720703125,
0.335205078125,
0.15283203125,
0.335205078125
],
[
0.125,
0.341552734375,
0.88134765625,
0.341552734375,
0.88134765625,
0.35595703125,
0.125,
0.35595703125
],
[
0.1519775390625,
0.35595703125,
0.74853515625,
0.354736328125,
0.74853515625,
0.369384765625,
0.1519775390625,
0.37060546875
],
[
0.12335205078125,
0.373779296875,
0.88134765625,
0.375,
0.88134765625,
0.389404296875,
0.12335205078125,
0.38818359375
],
[
0.1502685546875,
0.389404296875,
0.366943359375,
0.38818359375,
0.366943359375,
0.401611328125,
0.1502685546875,
0.40283203125
],
[
0.125,
0.4072265625,
0.88134765625,
0.4072265625,
0.88134765625,
0.42041015625,
0.125,
0.42041015625
],
[
0.1502685546875,
0.421630859375,
0.80224609375,
0.42041015625,
0.80224609375,
0.43505859375,
0.1502685546875,
0.436279296875
],
[
0.11358642578125,
0.436279296875,
0.88232421875,
0.438232421875,
0.88232421875,
0.455810546875,
0.11358642578125,
0.453857421875
],
[
0.1502685546875,
0.453857421875,
0.62353515625,
0.45263671875,
0.62353515625,
0.46728515625,
0.1502685546875,
0.468505859375
],
[
0.1168212890625,
0.472900390625,
0.88134765625,
0.472900390625,
0.88134765625,
0.486083984375,
0.1168212890625,
0.486083984375
],
[
0.1519775390625,
0.486083984375,
0.6298828125,
0.48486328125,
0.6298828125,
0.5,
0.1519775390625,
0.50048828125
],
[
0.1168212890625,
0.5048828125,
0.88134765625,
0.5048828125,
0.88134765625,
0.51806640625,
0.1168212890625,
0.51806640625
],
[
0.1502685546875,
0.51806640625,
0.88134765625,
0.51953125,
0.88134765625,
0.5341796875,
0.1502685546875,
0.53271484375
],
[
0.1519775390625,
0.5341796875,
0.371826171875,
0.5341796875,
0.371826171875,
0.54736328125,
0.1519775390625,
0.54736328125
],
[
0.1168212890625,
0.5517578125,
0.88134765625,
0.5517578125,
0.88134765625,
0.56494140625,
0.1168212890625,
0.56494140625
],
[
0.15283203125,
0.56640625,
0.771484375,
0.56640625,
0.771484375,
0.58056640625,
0.15283203125,
0.58056640625
],
[
0.11767578125,
0.583984375,
0.88232421875,
0.583984375,
0.88232421875,
0.59765625,
0.11767578125,
0.59765625
],
[
0.1502685546875,
0.5986328125,
0.449462890625,
0.5986328125,
0.449462890625,
0.6123046875,
0.1502685546875,
0.6123046875
],
[
0.494384765625,
0.8984375,
0.5048828125,
0.8984375,
0.5048828125,
0.912109375,
0.494384765625,
0.912109375
]
] |
[
"References",
"[1] Yossi Arjevani and Ohad Shamir. Communication complexity of distributed convex learning and optimization",
"arXivpreprintarXiv:1506.01900,2015",
"[2] Maria-Florina Balcan, Avrim Blum, Shai Fine, and Yishay Mansour. Distributed learning, communication com-",
"plexity and privacy.arXiv preprint arXiv:1204.3514,2012.",
"[3] Olivier Fercoq, Zheng Qu, Peter Richtarik, and Martin Takac. Fast distributed coordinate descent for non-",
"strongly convexlosses.In MachineLearningfor Signal Processing (MLSP),2014IEEEInternational Workshop",
"on, pages 1-6.IEEE, 2014.",
"[4] Martin Jaggi, Virginia Smith, Martin Takac, Jonathan Terhorst, Sanjay Krishnan, Thomas Hofmann, and",
"Michael I Jordan. Communication-efficient distributed dual coordinate ascent. In Advances in Neural Infor",
"mation Processing Systems, pages 3068-3076,2014.",
"[5] Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In",
"AdvancesinNeural InformationProcessingSystems,pages315-323,2013.",
"[6] Jakub Konecny and Peter Richtarik. Semi-stochastic gradient descent methods. arXiv preprint arXiv:1312.1666",
"2013.",
"[7] Chenxin Ma, Virginia Smith, Martin Jaggi, Michael I Jordan, Peter Richtarik, and Martin Takac. Adding vs.",
"averaging in distributed primal-dual optimization.arXiv preprint arXiv:1502.03508,2015.",
"[8] Peter Richtarik and Martin Taka. Distributed coordinate descent method for learning with big data. arXiv",
"preprint arXiv:1310.2059, 2013.",
"[9] Ohad Shamir and Nathan Srebro.Distributed stochastic optimization and learning.In Communication, Control",
"and Computing (Allerton), 2014 52nd Annual Allerton Conference on, pages 850-857.IEEE, 2014.",
"[10] Ohad Shamir, Nathan Srebro, and Tong Zhang. Communication efficient distributed optimization using an ap-",
"proximate newton-type method. arXiv preprint arXiv:1312.7853, 2013.",
"[1l] Tianbao Yang. Trading computation for communication: Distributed stochastic dual coordinate ascent. In Ad-",
"vances in Neural Information Processing Systems,pages 629-637,2013.",
"[12] Yuchen Zhang, John Duchi, Michael I Jordan, and Martin J Wainwright. Information-theoretic lower bounds for",
"distributed statistical estimation with communication constraints. In Advances in Neural Information Processing",
"Systems, pages 2328-2336, 2013.",
"[13] Yuchen Zhang, Martin J Wainwright, and John C Duchi. Communication-efficient algorithms for statistical",
"optimization. In Advances in Neural Information Processing Systems, pages 1502-1510,2012.",
"[14] Yuchen Zhang and Lin Xiao. Communication-efficient distributed optimization of self-concordant empirical",
"loss.arXivpreprintarXiv:1501.00263,2015.",
"5"
] |
{
"arxiv_id": "1511.03575",
"language": "en",
"page_idx": 4
}
|
||
1906.02979-en-0
|
1906.02979-en-0.jpg
|
# A Wind of Change: Detecting and Evaluating Lexical Semantic Change across Times and Domains
Dominik Schlechtweg$^1,\text{Anna Hätty}^{1,2}$,Marco del Tredici$^3$,Sabine Schulte im Walde$^{1}$
$^{1}$Institute for Natural Language Processing, University of Stuttgart
$^{2}$Robert Bosch GmbH, Corporate Research
$^{3}$Institute for Logic, Language and Computation, University of Amsterdam
$\{$schlecdk, schulte$\}$@ims.uni-stuttgart.de,
anna.haetty@de.bosch.com, m.deltredici@uva.nl
### $\mathbf{Abstract}$
We perform an interdisciplinary large-scale evaluation for detecting lexical semantic divergences in a diachronic and in a synchronic task: semantic sense changes across time, and semantic sense changes across domains. Our work addresses the superficialness and lack of comparison in assessing models of diachronic lexical change, by bringing together and extending benchmark models on a common state-of-the-art evaluation task. In addition, we demonstrate that the same evaluation task and modelling approaches can successfully be utilised for the synchronic detection of domain-specific sense divergences in the field of term extraction.
## 1 Introduction
Relying on an existing German LSC dataset we compare models regarding different combinations of semantic representations, alignment techniques and detection measures, while exploring various pre-processing and parameter settings. Furthermore, we introduce Word Injection to LSC, a modeling idea drawn from term extraction, that overcomes the problem of vector space alignment. Our comparison of state-of-the-art approaches identifies best models and optimal parameter settings, and it suggests modifications to existing models which consistently show superior performance.
Meanwhile, the detection of lexical sense divergences across time-specific corpora is not the only possible application of LSC detection models.
In more general terms, they have
the potential to detect sense divergences between corpora of any type, not necessarily timespecific ones. We acknowledge this observation and further explore a synchronic LSC detection task: identifying domain-specific changes of word senses in comparison to general-language usage, which is addressed, e.g., in term identification and automatic term extraction (Drouin, 2004; Pérez, 2016; Hätty and Schulte im Walde, 2018), and in determining social and dialectal language variations (Del Tredici and Fernández, 2017; Hovy and Purschke, 2018).$^{2}$
For addressing the synchronic LSC task, we present a recent sense-specific term dataset (Hätty et al., 2019) that we created analogously to the existing diachronic dataset, and we show that the diachronic models can be successfully applied to the synchronic task as well. This two-fold evaluation assures robustness and reproducibility of our model comparisons under various conditions.
Diachronic Lexical Semantic Change (LSC) detection, i.e., the automatic detection of word sense changes over time, is a flourishing new feld within NLP (Frermann and Lapata, 2016; Hamilton et al., 2016b; Schlechtweg et al., 2017, i.a.).$^1$ Yet, it is hard to compare the performances of the various models, and optimal parameter choices remain unclear, because up to now most models have been compared on different evaluation tasks and data. Presently, we do not know which model performs best under which conditions, and if more complex model architectures gain performance benefits over simpler models. This situation hinders advances in the field and favors unfelicitous drawings of statistical laws of diachronic LSC (Dubossarsky et al., 2017).
In this study, we provide the first large-scale evaluation of an extensive number of approaches.
$^{1}$An example for diachronic LSC is the German noun Vorwort (Paul, 2002), which was mainly used in the meaning of $`$preposition' before$\approx1800.$ Then Vorwort rapidly acquired a new meaning ‘preface’, which after 1850 has nearly exclusively been used.
$^{2}$An example for domain-specific synchronic LSC is the German noun Form. In general-language use, Form means shape'/ form', while in the cooking domain the predominant meaning is the domain-specific baking tin'.
|
[
[
0.3984375,
0.0736083984375,
0.6025390625,
0.07598876953125,
0.6025390625,
0.09320068359375,
0.3984375,
0.0914306640625
],
[
0.236083984375,
0.09320068359375,
0.76708984375,
0.0955810546875,
0.76708984375,
0.1134033203125,
0.236083984375,
0.11102294921875
],
[
0.361328125,
0.11517333984375,
0.64111328125,
0.11517333984375,
0.64111328125,
0.1300048828125,
0.361328125,
0.1300048828125
],
[
0.1419677734375,
0.1436767578125,
0.86376953125,
0.1436767578125,
0.86376953125,
0.157958984375,
0.1419677734375,
0.157958984375
],
[
0.233642578125,
0.157958984375,
0.77294921875,
0.1591796875,
0.77294921875,
0.177001953125,
0.233642578125,
0.17578125
],
[
0.3251953125,
0.174560546875,
0.681640625,
0.17578125,
0.681640625,
0.193603515625,
0.3251953125,
0.1923828125
],
[
0.2008056640625,
0.190673828125,
0.8076171875,
0.1923828125,
0.8076171875,
0.21142578125,
0.2008056640625,
0.208984375
],
[
0.268798828125,
0.21142578125,
0.740234375,
0.21142578125,
0.740234375,
0.2257080078125,
0.268798828125,
0.2257080078125
],
[
0.233642578125,
0.22802734375,
0.77490234375,
0.22802734375,
0.77490234375,
0.2423095703125,
0.233642578125,
0.2423095703125
],
[
0.265625,
0.265380859375,
0.34375,
0.265380859375,
0.34375,
0.27978515625,
0.265625,
0.27978515625
],
[
0.5126953125,
0.265380859375,
0.884765625,
0.26416015625,
0.884765625,
0.279052734375,
0.5126953125,
0.27978515625
],
[
0.5126953125,
0.281982421875,
0.88330078125,
0.28076171875,
0.88330078125,
0.295654296875,
0.5126953125,
0.29638671875
],
[
0.1470947265625,
0.29443359375,
0.46044921875,
0.29443359375,
0.46044921875,
0.30810546875,
0.1470947265625,
0.30810546875
],
[
0.51171875,
0.297607421875,
0.88330078125,
0.297607421875,
0.88330078125,
0.311279296875,
0.51171875,
0.311279296875
],
[
0.023529052734375,
0.3046875,
0.060516357421875,
0.3046875,
0.060516357421875,
0.410400390625,
0.023529052734375,
0.410400390625
],
[
0.0294189453125,
0.298583984375,
0.05462646484375,
0.298583984375,
0.05462646484375,
0.3134765625,
0.0294189453125,
0.3134765625
],
[
0.1470947265625,
0.308837890625,
0.459716796875,
0.308837890625,
0.459716796875,
0.322509765625,
0.1470947265625,
0.322509765625
],
[
0.51416015625,
0.3154296875,
0.884765625,
0.3154296875,
0.884765625,
0.328857421875,
0.51416015625,
0.328857421875
],
[
0.148681640625,
0.32470703125,
0.459716796875,
0.32470703125,
0.459716796875,
0.3349609375,
0.148681640625,
0.3349609375
],
[
0.51416015625,
0.330810546875,
0.88330078125,
0.330810546875,
0.88330078125,
0.344482421875,
0.51416015625,
0.344482421875
],
[
0.1470947265625,
0.33740234375,
0.459716796875,
0.33740234375,
0.459716796875,
0.350830078125,
0.1470947265625,
0.350830078125
],
[
0.1470947265625,
0.3515625,
0.459716796875,
0.3515625,
0.459716796875,
0.365234375,
0.1470947265625,
0.365234375
],
[
0.5126953125,
0.3466796875,
0.88330078125,
0.345703125,
0.88330078125,
0.35986328125,
0.5126953125,
0.361083984375
],
[
0.1470947265625,
0.365234375,
0.46044921875,
0.365234375,
0.46044921875,
0.37890625,
0.1470947265625,
0.37890625
],
[
0.51416015625,
0.363525390625,
0.884765625,
0.363525390625,
0.884765625,
0.377197265625,
0.51416015625,
0.377197265625
],
[
0.1470947265625,
0.380126953125,
0.46044921875,
0.380126953125,
0.46044921875,
0.393798828125,
0.1470947265625,
0.393798828125
],
[
0.5126953125,
0.377685546875,
0.884765625,
0.37890625,
0.884765625,
0.393798828125,
0.5126953125,
0.392578125
],
[
0.1470947265625,
0.394287109375,
0.46044921875,
0.394287109375,
0.46044921875,
0.407958984375,
0.1470947265625,
0.407958984375
],
[
0.51416015625,
0.3955078125,
0.88232421875,
0.3955078125,
0.88232421875,
0.4091796875,
0.51416015625,
0.4091796875
],
[
0.02520751953125,
0.4091796875,
0.06219482421875,
0.4091796875,
0.06219482421875,
0.490478515625,
0.02520751953125,
0.490478515625
],
[
0.148681640625,
0.410400390625,
0.4580078125,
0.410400390625,
0.4580078125,
0.42041015625,
0.148681640625,
0.42041015625
],
[
0.51416015625,
0.410888671875,
0.884765625,
0.410888671875,
0.884765625,
0.424560546875,
0.51416015625,
0.424560546875
],
[
0.1470947265625,
0.4228515625,
0.46044921875,
0.4228515625,
0.46044921875,
0.435791015625,
0.1470947265625,
0.435791015625
],
[
0.51416015625,
0.427001953125,
0.884765625,
0.427001953125,
0.884765625,
0.441162109375,
0.51416015625,
0.441162109375
],
[
0.148681640625,
0.438232421875,
0.46044921875,
0.438232421875,
0.46044921875,
0.448974609375,
0.148681640625,
0.448974609375
],
[
0.51416015625,
0.443603515625,
0.85986328125,
0.443603515625,
0.85986328125,
0.457275390625,
0.51416015625,
0.457275390625
],
[
0.1470947265625,
0.45068359375,
0.459716796875,
0.45068359375,
0.459716796875,
0.46435546875,
0.1470947265625,
0.46435546875
],
[
0.53125,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.473876953125,
0.53125,
0.473876953125
],
[
0.148681640625,
0.466064453125,
0.46044921875,
0.466064453125,
0.46044921875,
0.476806640625,
0.148681640625,
0.476806640625
],
[
0.1453857421875,
0.478515625,
0.462158203125,
0.47802734375,
0.462158203125,
0.4921875,
0.1453857421875,
0.493408203125
],
[
0.51416015625,
0.476806640625,
0.884765625,
0.476806640625,
0.884765625,
0.490478515625,
0.51416015625,
0.490478515625
],
[
0.0218505859375,
0.485107421875,
0.06219482421875,
0.485107421875,
0.06219482421875,
0.70166015625,
0.0218505859375,
0.70166015625
],
[
0.1470947265625,
0.493408203125,
0.27490234375,
0.493408203125,
0.27490234375,
0.50732421875,
0.1470947265625,
0.50732421875
],
[
0.5126953125,
0.4921875,
0.88330078125,
0.4921875,
0.88330078125,
0.50732421875,
0.5126953125,
0.50732421875
],
[
0.51611328125,
0.51025390625,
0.578125,
0.51025390625,
0.578125,
0.521484375,
0.51611328125,
0.521484375
],
[
0.591796875,
0.5087890625,
0.884765625,
0.5087890625,
0.884765625,
0.5224609375,
0.591796875,
0.5224609375
],
[
0.1395263671875,
0.5224609375,
0.26123046875,
0.5224609375,
0.26123046875,
0.5361328125,
0.1395263671875,
0.5361328125
],
[
0.5126953125,
0.52490234375,
0.88330078125,
0.52392578125,
0.88330078125,
0.5380859375,
0.5126953125,
0.5390625
],
[
0.51416015625,
0.54150390625,
0.88232421875,
0.54150390625,
0.88232421875,
0.5546875,
0.51416015625,
0.5546875
],
[
0.13525390625,
0.52783203125,
0.1419677734375,
0.52783203125,
0.1419677734375,
0.53076171875,
0.13525390625,
0.53076171875
],
[
0.11846923828125,
0.54736328125,
0.4873046875,
0.5498046875,
0.4873046875,
0.5634765625,
0.11846923828125,
0.56103515625
],
[
0.5126953125,
0.55712890625,
0.63037109375,
0.5556640625,
0.63037109375,
0.57080078125,
0.5126953125,
0.57177734375
],
[
0.6376953125,
0.55712890625,
0.88232421875,
0.55712890625,
0.88232421875,
0.57080078125,
0.6376953125,
0.57080078125
],
[
0.12017822265625,
0.5654296875,
0.489013671875,
0.5654296875,
0.489013671875,
0.5791015625,
0.12017822265625,
0.5791015625
],
[
0.51171875,
0.572265625,
0.88330078125,
0.57177734375,
0.88330078125,
0.5859375,
0.51171875,
0.58740234375
],
[
0.11846923828125,
0.58203125,
0.489013671875,
0.58203125,
0.489013671875,
0.595703125,
0.11846923828125,
0.595703125
],
[
0.51416015625,
0.5888671875,
0.884765625,
0.5888671875,
0.884765625,
0.6025390625,
0.51416015625,
0.6025390625
],
[
0.12017822265625,
0.59716796875,
0.49072265625,
0.59716796875,
0.49072265625,
0.61083984375,
0.12017822265625,
0.61083984375
],
[
0.5126953125,
0.60546875,
0.88330078125,
0.60546875,
0.88330078125,
0.619140625,
0.5126953125,
0.619140625
],
[
0.12017822265625,
0.61328125,
0.49072265625,
0.61328125,
0.49072265625,
0.62744140625,
0.12017822265625,
0.62744140625
],
[
0.5126953125,
0.62158203125,
0.88330078125,
0.62158203125,
0.88330078125,
0.6357421875,
0.5126953125,
0.6357421875
],
[
0.11846923828125,
0.63134765625,
0.1630859375,
0.62744140625,
0.1646728515625,
0.64013671875,
0.12017822265625,
0.6435546875
],
[
0.1739501953125,
0.62744140625,
0.489013671875,
0.62890625,
0.489013671875,
0.6435546875,
0.1739501953125,
0.642578125
],
[
0.51416015625,
0.63818359375,
0.880859375,
0.63818359375,
0.880859375,
0.6484375,
0.51416015625,
0.6484375
],
[
0.1168212890625,
0.6455078125,
0.489013671875,
0.64453125,
0.489013671875,
0.6591796875,
0.1168212890625,
0.66015625
],
[
0.51416015625,
0.65380859375,
0.88330078125,
0.65380859375,
0.88330078125,
0.66748046875,
0.51416015625,
0.66748046875
],
[
0.1168212890625,
0.6611328125,
0.4873046875,
0.6611328125,
0.4873046875,
0.6748046875,
0.1168212890625,
0.6748046875
],
[
0.51416015625,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.51416015625,
0.68310546875
],
[
0.1168212890625,
0.677734375,
0.489013671875,
0.6767578125,
0.489013671875,
0.69140625,
0.1168212890625,
0.6923828125
],
[
0.51416015625,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69970703125,
0.51416015625,
0.69970703125
],
[
0.12017822265625,
0.69482421875,
0.49072265625,
0.69482421875,
0.49072265625,
0.7080078125,
0.12017822265625,
0.7080078125
],
[
0.51416015625,
0.70166015625,
0.7724609375,
0.70166015625,
0.7724609375,
0.71533203125,
0.51416015625,
0.71533203125
],
[
0.11846923828125,
0.71044921875,
0.489013671875,
0.71044921875,
0.489013671875,
0.7236328125,
0.11846923828125,
0.7236328125
],
[
0.53125,
0.71875,
0.88232421875,
0.71875,
0.88232421875,
0.73291015625,
0.53125,
0.73291015625
],
[
0.11846923828125,
0.72705078125,
0.4873046875,
0.72705078125,
0.4873046875,
0.74072265625,
0.11846923828125,
0.74072265625
],
[
0.5126953125,
0.73388671875,
0.88232421875,
0.73388671875,
0.88232421875,
0.74755859375,
0.5126953125,
0.74755859375
],
[
0.11846923828125,
0.7421875,
0.4873046875,
0.7421875,
0.4873046875,
0.755859375,
0.11846923828125,
0.755859375
],
[
0.51611328125,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76416015625,
0.51611328125,
0.76416015625
],
[
0.121826171875,
0.76025390625,
0.4873046875,
0.76025390625,
0.4873046875,
0.77099609375,
0.121826171875,
0.77099609375
],
[
0.51416015625,
0.7666015625,
0.88232421875,
0.7666015625,
0.88232421875,
0.77978515625,
0.51416015625,
0.77978515625
],
[
0.11846923828125,
0.77490234375,
0.49072265625,
0.77490234375,
0.49072265625,
0.7880859375,
0.11846923828125,
0.7880859375
],
[
0.11846923828125,
0.79052734375,
0.4580078125,
0.79052734375,
0.4580078125,
0.80419921875,
0.11846923828125,
0.80419921875
],
[
0.51611328125,
0.78515625,
0.88232421875,
0.78515625,
0.88232421875,
0.7958984375,
0.51611328125,
0.7958984375
],
[
0.51416015625,
0.7998046875,
0.88232421875,
0.7998046875,
0.88232421875,
0.8134765625,
0.51416015625,
0.8134765625
],
[
0.1361083984375,
0.80712890625,
0.489013671875,
0.80712890625,
0.489013671875,
0.82080078125,
0.1361083984375,
0.82080078125
],
[
0.51611328125,
0.81787109375,
0.88330078125,
0.81787109375,
0.88330078125,
0.82763671875,
0.51611328125,
0.82763671875
],
[
0.11846923828125,
0.82373046875,
0.489013671875,
0.82373046875,
0.489013671875,
0.83740234375,
0.11846923828125,
0.83740234375
],
[
0.5126953125,
0.8330078125,
0.87744140625,
0.8330078125,
0.87744140625,
0.8466796875,
0.5126953125,
0.8466796875
],
[
0.1395263671875,
0.84716796875,
0.489013671875,
0.84716796875,
0.489013671875,
0.86083984375,
0.1395263671875,
0.86083984375
],
[
0.11846923828125,
0.85986328125,
0.49072265625,
0.85986328125,
0.49072265625,
0.87353515625,
0.11846923828125,
0.87353515625
],
[
0.53173828125,
0.85986328125,
0.884765625,
0.85986328125,
0.884765625,
0.87353515625,
0.53173828125,
0.87353515625
],
[
0.11846923828125,
0.87158203125,
0.4931640625,
0.87158203125,
0.4931640625,
0.88525390625,
0.11846923828125,
0.88525390625
],
[
0.51171875,
0.87060546875,
0.88671875,
0.87158203125,
0.88671875,
0.8857421875,
0.51171875,
0.88525390625
],
[
0.1168212890625,
0.88427734375,
0.489013671875,
0.88427734375,
0.489013671875,
0.89794921875,
0.1168212890625,
0.89794921875
],
[
0.51416015625,
0.88427734375,
0.88671875,
0.88427734375,
0.88671875,
0.89794921875,
0.51416015625,
0.89794921875
],
[
0.12017822265625,
0.89794921875,
0.224365234375,
0.89794921875,
0.224365234375,
0.90771484375,
0.12017822265625,
0.90771484375
],
[
0.51416015625,
0.896484375,
0.7841796875,
0.896484375,
0.7841796875,
0.91015625,
0.51416015625,
0.91015625
]
] |
[
"A Wind of Change:",
"Detecting and Evaluating Lexical Semantic Change",
"across Times and Domains",
"Dominik Schlechtweg', Anna Hattyl,, Marco del Tredici, Sabine Schulte im Walde",
"'Institute for Natural Language Processing, University of Stuttgart",
"2Robert Bosch GmbH, Corporate Research",
"3Institute for Logic, Language and Computation, University of Amsterdam",
"{schlecdk,schulte}@ims.uni-stuttgart.de,",
"anna.haetty@de.bosch.com, m.deltredici@uva.nl",
"Abstract",
"Relying on an existing German LSC dataset we",
"compare models regarding different combinations",
"We perform an interdisciplinary large-scale",
" of semantic representations, alignment techniques",
"50",
"9",
"evaluation for detecting lexical semantic di-",
"and detection measures, while exploring various",
"vergences in adiachronicandina synchronic",
"pre-processing and parameter settings. Further-",
"task: semantic sense changes across time,",
"and semantic sense changes across domains.",
"more, we introduce Word Injection to LSC, a mod-",
"Our work addresses the superficialness and",
"eling idea drawn from term extraction, that over-",
"lack of comparison in assessing models of di-",
"comes the problem of vector space alignment. Our",
"achronic lexical change, by bringing together",
"comparison of state-of-the-art approaches identi-",
"['S]",
"andextendingbenchmarkmodelsonacom-",
"fies best models and optimal parameter settings,",
"mon state-of-the-art evaluation task. In addi-",
"and it suggests modifications to existing models",
"tion, we demonstrate that the same evaluation",
"which consistently show superior performance.",
"task and modelling approaches can success-",
"Meanwhile, the detection of lexical sense di-",
"fullybeutilisedforthesynchronicdetectionof",
"domain-specific sense divergences in the field",
"vergences across time-specific corpora is not",
"1A62620906111",
"of term extraction.",
"the only possible application of LSC detection",
"models..",
"In more general terms,they have",
"Introduction",
"the potential to detect sense divergences be-",
"tween corpora of any type, not necessarily time-",
"",
"Diachronic Lexical Semantic Change (LSC) de-",
"specific ones.",
"We acknowledge this observa-",
"tection, i.e., the automatic detection of word",
"tion and further explore a synchronic LSC detec-",
"sense changes over time, is a flourishing new",
"tion task: identifying domain-specific changes of",
"field within NLP (Frermann and Lapata, 2016;",
"word senses in comparison to general-language",
"Hamilton et al., 2016b; Schlechtweg et al., 2017,",
"usage, which is addressed, e.g., in term identi-",
"i.a.),1",
"Yet, it is hard to compare the perfor-",
"fication and automatic term extraction (Drouin",
"mances of the various models, and optimal pa-",
"2004; Péerez, 2016; Hatty and Schulte im Walde,",
"rameter choices remain unclear, because up to",
"2018), and in determining social and dialectal",
"now most models have been compared on differ-",
"language variations (Del Tredici and Fernandez,",
"ent evaluation tasks and data.Presently, we do",
"2017; Hovy and Purschke, 2018).2",
"not know which model performs best under which",
"For addressing the synchronic LSC task,",
"conditions, and if more complex model architec-",
"we present a recent sense-specific term dataset",
"tures gain performance benefits over simpler mod-",
"(Hatty et al., 2019) that we created analogously to",
"els.This situation hinders advances in the field",
"the existing diachronic dataset, and we show that",
"and favors unfelicitous drawings of statistical laws",
"of diachronic LSC (Dubossarsky et al., 2017).",
"thediachronicmodelscanbesuccessfullyapplied",
"to the synchronic task as well. This two-fold eval-",
"In this study, we provide the first large-scale",
"uation assures robustness and reproducibility of",
"evaluation of an extensive number of approaches.",
"our model comparisons under various conditions.",
"I An example for diachronic LSC is the German noun Vor-",
"wort (Paul, 2002), which was mainly used in the meaning of",
"2An example for domain-specific synchronic LSC is the",
"'preposition' before ~1800. Then Vorwort rapidly acquired",
"German noun Form. In general-language use, Form means",
"a new meaning'preface', which after 1850 has nearly exclu-",
"'shape'\"form', while in the cooking domain the predominant",
"sively been used.",
"meaning is the domain-specific baking tin'."
] |
{
"arxiv_id": "1906.02979",
"language": "en",
"page_idx": 0
}
|
||
1906.02979-en-1
|
1906.02979-en-1.jpg
|
### 2 $\textbf{Related Work}$
Diachronic LSC Detection. Existing approaches for diachronic LSC detection are mainly based on three types of meaning representations: (i) semantic vector spaces, (ii) topic distributions, and (iii) sense clusters. In (i), semantic vector spaces, each word is represented as two vectors reflecting its co-occurrence statistics at different periods of time (Gulordava and Baroni, 2011; Kimet al., 2014; Xu and Kemp, 2015; Eger and Mehler, 2016; Hamilton et al., 2016a,b; Hellrich and Hahn, 2016; Rosenfeld and Erk, 2018). LSC is typically measured by the cosine distance (or some alternative metric) between the two vectors, or by differences in contextual dispersion between the two vectors (Kisselew et al., 2016; Schlechtweg et al., 2017). (ii) Diachronic topic models infer a probability distribution for each word over different word senses (or topics), which are in turn modeled as a distribution over words (Wang and McCallum, 2006; Bamman and Crane, 2011; Wijaya and Yeniterzi, 2011; Lau et al., 2012; Mihalcea and Nastase, 2012; Cook et al., 2014; Frermann and Lapata, 2016). LSC of a word is measured by calculating a novelty score for its senses based on their frequency of use. (iii) Clustering models assign all uses of a word into sense clusters based on some contextual property (Mitra et al., 2015). Word sense clustering models are similar to topic models in that they map uses to senses. Accordingly, LSC of a word is measured similarly as in (ii). For an overview on diachronic LSC detection, see Tahmasebi et al. (2018).
Synchronic LSC Detection. We use the term synchronic LSC to refer to NLP research areas with a focus on how the meanings of words vary across domains or communities of speakers.
Synchronic LSC per se is not
widely researched; for meaning shifts across domains, there is strongly related research which is concerned with domain-specific word sense disambiguation (Maynard and Ananiadou, 1998; Chen and Al-Mubaid, 2006; Taghipour and Ng, 2015; Daille et al., 2016) or term ambiguity detection (Baldwin et al., 2013; Wang et al., 2013). The only notable work for explicitly measuring across domain meaning shifts is Ferrari et al. (2017), which is based on semantic vector spaces and cosine distance. Synchronic LSC across commu-
nities has been investigated as meaning variation in online communities, leveraging the large-scale data which has become available thanks to online social platforms (Del Tredici and Fernández, 2017; Rotabi et al., 2017).
Evaluation. Existing evaluation procedures for LSC detection can be distinguished into evaluation on (i) empirically observed data, and (ii) synthetic data or related tasks. (i) includes case studies of individual words (Sagi et al., 2009; Jatowt and Duh, 2014; Hamilton et al., 2016a),
stand-alone
comparison of a few hand-selected words (Wijaya and Yeniterzi, 2011; Hamilton et al., 2016b;
2017),
Del Tredici and Fernández,
comparison of hand-selected changing vs. semantically stable words (Lau et al., 2012; Cook et al., 2014), and post-hoc evaluation of the predictions of the presented models (Cook and Stevenson, 2010; Kulkarni et al., 2015; Del Tredici et al., 2016; Eger and Mehler, 2016; Ferrari et al. 2017).
Schlechtweg et al. (2017) propose a
small-scale annotation of diachronic metaphoric
change.
(ii)
procedures
Synthetic
evaluation
include
simulate
$LSC$
studies
that
Kulkarni et al.,
(Cook and Stevenson,
2010;
2015; Rosenfeld and Erk, 2018), evaluate sense assignments in WordNet (Mitra et al., 2015; Frermann and Lapata, 2016), identify text creation dates, (Mihalcea and Nastase, 2012; Frermann and Lapata, 2016), or predict the loglikelihood of textual data (Frermann and Lapata, 2016).
Overall, the various studies use different evaluation tasks and data, with little overlap. Most evaluation data has not been annotated. Models were rarely compared to previously suggested ones, especially if the models differed in meaning representations. Moreover, for the diachronic task, synthetic datasets are used which do not reflect actual diachronic changes.
## 3 $\textbf{Task and Data}$
Our study makes use of the evaluation framework proposed in Schlechtweg et al.(2018), where diachronic LSC detection is defined as a comparison between word uses in two time-specific corpora. We further applied the framework to create an analogous synchronic LSC dataset that compares word uses across general-language and domainspecific corpora. The common, meta-level task in
|
[
[
0.12017822265625,
0.0748291015625,
0.272216796875,
0.0748291015625,
0.272216796875,
0.08905029296875,
0.12017822265625,
0.08905029296875
],
[
0.51611328125,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.51611328125,
0.09027099609375
],
[
0.51416015625,
0.09320068359375,
0.88330078125,
0.09320068359375,
0.88330078125,
0.10687255859375,
0.51416015625,
0.10687255859375
],
[
0.11846923828125,
0.105712890625,
0.44873046875,
0.105712890625,
0.44873046875,
0.119384765625,
0.11846923828125,
0.119384765625
],
[
0.441162109375,
0.10809326171875,
0.489013671875,
0.10809326171875,
0.489013671875,
0.119384765625,
0.441162109375,
0.119384765625
],
[
0.5126953125,
0.10809326171875,
0.880859375,
0.10809326171875,
0.880859375,
0.1217041015625,
0.5126953125,
0.1217041015625
],
[
0.11846923828125,
0.122314453125,
0.489013671875,
0.1217041015625,
0.489013671875,
0.135986328125,
0.11846923828125,
0.13720703125
],
[
0.5126953125,
0.12469482421875,
0.88232421875,
0.12469482421875,
0.88232421875,
0.1383056640625,
0.5126953125,
0.1383056640625
],
[
0.11846923828125,
0.1395263671875,
0.49072265625,
0.1395263671875,
0.49072265625,
0.152587890625,
0.11846923828125,
0.152587890625
],
[
0.5126953125,
0.1395263671875,
0.70654296875,
0.1395263671875,
0.70654296875,
0.152587890625,
0.5126953125,
0.152587890625
],
[
0.1168212890625,
0.15380859375,
0.489013671875,
0.155029296875,
0.489013671875,
0.169189453125,
0.1168212890625,
0.168701171875
],
[
0.5126953125,
0.1644287109375,
0.884765625,
0.1650390625,
0.884765625,
0.179931640625,
0.5126953125,
0.1787109375
],
[
0.11846923828125,
0.171630859375,
0.49072265625,
0.171630859375,
0.49072265625,
0.185302734375,
0.11846923828125,
0.185302734375
],
[
0.51171875,
0.18115234375,
0.88330078125,
0.18115234375,
0.88330078125,
0.19482421875,
0.51171875,
0.19482421875
],
[
0.11846923828125,
0.18701171875,
0.49072265625,
0.18701171875,
0.49072265625,
0.20068359375,
0.11846923828125,
0.20068359375
],
[
0.51416015625,
0.19775390625,
0.88330078125,
0.19775390625,
0.88330078125,
0.21142578125,
0.51416015625,
0.21142578125
],
[
0.121826171875,
0.204833984375,
0.489013671875,
0.204833984375,
0.489013671875,
0.215576171875,
0.121826171875,
0.215576171875
],
[
0.51171875,
0.2120361328125,
0.88330078125,
0.213134765625,
0.88330078125,
0.22802734375,
0.51171875,
0.226806640625
],
[
0.12017822265625,
0.218505859375,
0.4873046875,
0.218505859375,
0.4873046875,
0.232177734375,
0.12017822265625,
0.232177734375
],
[
0.12017822265625,
0.235107421875,
0.49072265625,
0.235107421875,
0.49072265625,
0.248779296875,
0.12017822265625,
0.248779296875
],
[
0.51416015625,
0.2298583984375,
0.88330078125,
0.2298583984375,
0.88330078125,
0.24462890625,
0.51416015625,
0.24462890625
],
[
0.11846923828125,
0.25048828125,
0.4873046875,
0.25048828125,
0.4873046875,
0.26416015625,
0.11846923828125,
0.26416015625
],
[
0.5126953125,
0.2464599609375,
0.88330078125,
0.2464599609375,
0.88330078125,
0.260009765625,
0.5126953125,
0.260009765625
],
[
0.11846923828125,
0.2666015625,
0.489013671875,
0.265380859375,
0.489013671875,
0.27978515625,
0.11846923828125,
0.28076171875
],
[
0.51416015625,
0.262451171875,
0.884765625,
0.262451171875,
0.884765625,
0.275634765625,
0.51416015625,
0.275634765625
],
[
0.51416015625,
0.27783203125,
0.88330078125,
0.276611328125,
0.88330078125,
0.29150390625,
0.51416015625,
0.292236328125
],
[
0.12017822265625,
0.283935546875,
0.489013671875,
0.283935546875,
0.489013671875,
0.297607421875,
0.12017822265625,
0.297607421875
],
[
0.51611328125,
0.295654296875,
0.57373046875,
0.295654296875,
0.57373046875,
0.30712890625,
0.51611328125,
0.30712890625
],
[
0.59765625,
0.29443359375,
0.80908203125,
0.29443359375,
0.80908203125,
0.30810546875,
0.59765625,
0.30810546875
],
[
0.83447265625,
0.295654296875,
0.88232421875,
0.295654296875,
0.88232421875,
0.30712890625,
0.83447265625,
0.30712890625
],
[
0.121826171875,
0.3017578125,
0.4873046875,
0.3017578125,
0.4873046875,
0.312255859375,
0.121826171875,
0.312255859375
],
[
0.51416015625,
0.31005859375,
0.88232421875,
0.31005859375,
0.88232421875,
0.32373046875,
0.51416015625,
0.32373046875
],
[
0.12017822265625,
0.317626953125,
0.4873046875,
0.317626953125,
0.4873046875,
0.327880859375,
0.12017822265625,
0.327880859375
],
[
0.5126953125,
0.325927734375,
0.88232421875,
0.325927734375,
0.88232421875,
0.339111328125,
0.5126953125,
0.339111328125
],
[
0.12017822265625,
0.33203125,
0.489013671875,
0.33203125,
0.489013671875,
0.345703125,
0.12017822265625,
0.345703125
],
[
0.11846923828125,
0.3466796875,
0.4873046875,
0.3466796875,
0.4873046875,
0.35986328125,
0.11846923828125,
0.35986328125
],
[
0.51416015625,
0.342529296875,
0.88330078125,
0.342529296875,
0.88330078125,
0.355712890625,
0.51416015625,
0.355712890625
],
[
0.5126953125,
0.35693359375,
0.88232421875,
0.35693359375,
0.88232421875,
0.371826171875,
0.5126953125,
0.371826171875
],
[
0.1168212890625,
0.363525390625,
0.489013671875,
0.363525390625,
0.489013671875,
0.377197265625,
0.1168212890625,
0.377197265625
],
[
0.5126953125,
0.372314453125,
0.88330078125,
0.37353515625,
0.88330078125,
0.388427734375,
0.5126953125,
0.38720703125
],
[
0.11846923828125,
0.380126953125,
0.49072265625,
0.380126953125,
0.49072265625,
0.393798828125,
0.11846923828125,
0.393798828125
],
[
0.51416015625,
0.388427734375,
0.56640625,
0.39013671875,
0.56494140625,
0.405029296875,
0.5126953125,
0.402587890625
],
[
0.572265625,
0.39013671875,
0.724609375,
0.3896484375,
0.724609375,
0.40380859375,
0.572265625,
0.405029296875
],
[
0.71533203125,
0.39013671875,
0.88330078125,
0.39013671875,
0.88330078125,
0.40380859375,
0.71533203125,
0.40380859375
],
[
0.51416015625,
0.40625,
0.56494140625,
0.40625,
0.56494140625,
0.42041015625,
0.51416015625,
0.42041015625
],
[
0.1168212890625,
0.3955078125,
0.4873046875,
0.394287109375,
0.4873046875,
0.4091796875,
0.1168212890625,
0.410400390625
],
[
0.5908203125,
0.40625,
0.88671875,
0.40673828125,
0.88671875,
0.421630859375,
0.5908203125,
0.42041015625
],
[
0.11846923828125,
0.412109375,
0.489013671875,
0.412109375,
0.489013671875,
0.42578125,
0.11846923828125,
0.42578125
],
[
0.12017822265625,
0.427490234375,
0.4873046875,
0.427490234375,
0.4873046875,
0.441162109375,
0.12017822265625,
0.441162109375
],
[
0.51416015625,
0.42333984375,
0.884765625,
0.42333984375,
0.884765625,
0.43701171875,
0.51416015625,
0.43701171875
],
[
0.11846923828125,
0.4423828125,
0.489013671875,
0.443603515625,
0.489013671875,
0.457763671875,
0.11846923828125,
0.457275390625
],
[
0.5126953125,
0.439453125,
0.57373046875,
0.439453125,
0.57373046875,
0.45361328125,
0.5126953125,
0.45361328125
],
[
0.53125,
0.454833984375,
0.60693359375,
0.454833984375,
0.60693359375,
0.4697265625,
0.53125,
0.4697265625
],
[
0.6337890625,
0.454833984375,
0.716796875,
0.454833984375,
0.716796875,
0.4697265625,
0.6337890625,
0.4697265625
],
[
0.74365234375,
0.454833984375,
0.82861328125,
0.454833984375,
0.82861328125,
0.4697265625,
0.74365234375,
0.4697265625
],
[
0.85791015625,
0.4560546875,
0.884765625,
0.4560546875,
0.884765625,
0.4697265625,
0.85791015625,
0.4697265625
],
[
0.84619140625,
0.471435546875,
0.884765625,
0.471435546875,
0.884765625,
0.48388671875,
0.84619140625,
0.48388671875
],
[
0.11846923828125,
0.458984375,
0.4873046875,
0.460205078125,
0.4873046875,
0.474365234375,
0.11846923828125,
0.473876953125
],
[
0.51416015625,
0.47265625,
0.57568359375,
0.47265625,
0.57568359375,
0.48388671875,
0.51416015625,
0.48388671875
],
[
0.59326171875,
0.47265625,
0.681640625,
0.47265625,
0.681640625,
0.48388671875,
0.59326171875,
0.48388671875
],
[
0.681640625,
0.47265625,
0.72021484375,
0.47265625,
0.72021484375,
0.48388671875,
0.681640625,
0.48388671875
],
[
0.74951171875,
0.47265625,
0.81787109375,
0.47265625,
0.81787109375,
0.48388671875,
0.74951171875,
0.48388671875
],
[
0.51611328125,
0.48681640625,
0.70068359375,
0.48681640625,
0.70068359375,
0.50048828125,
0.51611328125,
0.50048828125
],
[
0.11846923828125,
0.476806640625,
0.486572265625,
0.476806640625,
0.486572265625,
0.490478515625,
0.11846923828125,
0.490478515625
],
[
0.69921875,
0.488037109375,
0.7705078125,
0.488037109375,
0.7705078125,
0.49951171875,
0.69921875,
0.49951171875
],
[
0.7646484375,
0.48681640625,
0.88671875,
0.48681640625,
0.88671875,
0.50048828125,
0.7646484375,
0.50048828125
],
[
0.11846923828125,
0.493408203125,
0.49072265625,
0.493408203125,
0.49072265625,
0.50732421875,
0.11846923828125,
0.50732421875
],
[
0.69580078125,
0.4912109375,
0.70166015625,
0.4912109375,
0.70166015625,
0.496337890625,
0.69580078125,
0.496337890625
],
[
0.5126953125,
0.501953125,
0.88671875,
0.5029296875,
0.88671875,
0.51708984375,
0.5126953125,
0.5166015625
],
[
0.12017822265625,
0.5087890625,
0.489013671875,
0.5087890625,
0.489013671875,
0.5224609375,
0.12017822265625,
0.5224609375
],
[
0.12017822265625,
0.52490234375,
0.489013671875,
0.52490234375,
0.489013671875,
0.5380859375,
0.12017822265625,
0.5380859375
],
[
0.5126953125,
0.51953125,
0.884765625,
0.5185546875,
0.884765625,
0.533203125,
0.5126953125,
0.53369140625
],
[
0.51416015625,
0.5361328125,
0.8359375,
0.5361328125,
0.8359375,
0.5498046875,
0.51416015625,
0.5498046875
],
[
0.83447265625,
0.53759765625,
0.884765625,
0.53759765625,
0.884765625,
0.548828125,
0.83447265625,
0.548828125
],
[
0.51416015625,
0.5517578125,
0.884765625,
0.5517578125,
0.884765625,
0.5654296875,
0.51416015625,
0.5654296875
],
[
0.11846923828125,
0.54052734375,
0.489013671875,
0.5390625,
0.489013671875,
0.55419921875,
0.11846923828125,
0.5546875
],
[
0.11846923828125,
0.5556640625,
0.489013671875,
0.5556640625,
0.489013671875,
0.5693359375,
0.11846923828125,
0.5693359375
],
[
0.5126953125,
0.56787109375,
0.880859375,
0.56787109375,
0.880859375,
0.58056640625,
0.5126953125,
0.58056640625
],
[
0.11846923828125,
0.57373046875,
0.4873046875,
0.57373046875,
0.4873046875,
0.58740234375,
0.11846923828125,
0.58740234375
],
[
0.5126953125,
0.5830078125,
0.88330078125,
0.58447265625,
0.88330078125,
0.5986328125,
0.5126953125,
0.59716796875
],
[
0.12017822265625,
0.5888671875,
0.489013671875,
0.5888671875,
0.489013671875,
0.6025390625,
0.12017822265625,
0.6025390625
],
[
0.5126953125,
0.59716796875,
0.56494140625,
0.599609375,
0.56396484375,
0.6142578125,
0.51171875,
0.6123046875
],
[
0.11846923828125,
0.60498046875,
0.489013671875,
0.60498046875,
0.489013671875,
0.6181640625,
0.11846923828125,
0.6181640625
],
[
0.11846923828125,
0.62158203125,
0.402587890625,
0.62158203125,
0.402587890625,
0.6357421875,
0.11846923828125,
0.6357421875
],
[
0.53369140625,
0.6162109375,
0.88232421875,
0.6162109375,
0.88232421875,
0.6298828125,
0.53369140625,
0.6298828125
],
[
0.5126953125,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.64453125,
0.5126953125,
0.64453125
],
[
0.11846923828125,
0.65283203125,
0.489013671875,
0.65283203125,
0.489013671875,
0.66748046875,
0.11846923828125,
0.66748046875
],
[
0.51611328125,
0.6484375,
0.88330078125,
0.6484375,
0.88330078125,
0.662109375,
0.51611328125,
0.662109375
],
[
0.1168212890625,
0.66943359375,
0.4873046875,
0.66845703125,
0.4873046875,
0.68310546875,
0.1168212890625,
0.68408203125
],
[
0.51416015625,
0.6650390625,
0.88232421875,
0.6650390625,
0.88232421875,
0.6787109375,
0.51416015625,
0.6787109375
],
[
0.11846923828125,
0.68505859375,
0.489013671875,
0.68505859375,
0.489013671875,
0.69873046875,
0.11846923828125,
0.69873046875
],
[
0.51416015625,
0.68115234375,
0.88232421875,
0.68115234375,
0.88232421875,
0.69482421875,
0.51416015625,
0.69482421875
],
[
0.5126953125,
0.6953125,
0.88330078125,
0.69677734375,
0.88330078125,
0.71142578125,
0.5126953125,
0.71044921875
],
[
0.11846923828125,
0.70166015625,
0.489013671875,
0.70068359375,
0.489013671875,
0.71533203125,
0.11846923828125,
0.71630859375
],
[
0.216796875,
0.71630859375,
0.49072265625,
0.71875,
0.49072265625,
0.732421875,
0.216796875,
0.72998046875
],
[
0.51416015625,
0.71337890625,
0.884765625,
0.71337890625,
0.884765625,
0.72705078125,
0.51416015625,
0.72705078125
],
[
0.121826171875,
0.7197265625,
0.191650390625,
0.7197265625,
0.191650390625,
0.73095703125,
0.121826171875,
0.73095703125
],
[
0.5126953125,
0.72802734375,
0.662109375,
0.728515625,
0.662109375,
0.74365234375,
0.5126953125,
0.7421875
],
[
0.12017822265625,
0.73388671875,
0.489013671875,
0.73388671875,
0.489013671875,
0.74755859375,
0.12017822265625,
0.74755859375
],
[
0.12017822265625,
0.75048828125,
0.489013671875,
0.75048828125,
0.489013671875,
0.76416015625,
0.12017822265625,
0.76416015625
],
[
0.54296875,
0.75732421875,
0.67236328125,
0.75732421875,
0.67236328125,
0.771484375,
0.54296875,
0.771484375
],
[
0.11846923828125,
0.7666015625,
0.489013671875,
0.7666015625,
0.489013671875,
0.77978515625,
0.11846923828125,
0.77978515625
],
[
0.51611328125,
0.76318359375,
0.5234375,
0.75830078125,
0.529296875,
0.7626953125,
0.52197265625,
0.76708984375
],
[
0.11846923828125,
0.7822265625,
0.489013671875,
0.78076171875,
0.489013671875,
0.7958984375,
0.11846923828125,
0.79638671875
],
[
0.51416015625,
0.783203125,
0.884765625,
0.783203125,
0.884765625,
0.79638671875,
0.51416015625,
0.79638671875
],
[
0.11846923828125,
0.79736328125,
0.489013671875,
0.798828125,
0.489013671875,
0.8134765625,
0.11846923828125,
0.8125
],
[
0.5126953125,
0.798828125,
0.88232421875,
0.79736328125,
0.88232421875,
0.8125,
0.5126953125,
0.8134765625
],
[
0.12017822265625,
0.81396484375,
0.49072265625,
0.81396484375,
0.49072265625,
0.82763671875,
0.12017822265625,
0.82763671875
],
[
0.51611328125,
0.81640625,
0.88330078125,
0.81640625,
0.88330078125,
0.82666015625,
0.51611328125,
0.82666015625
],
[
0.1168212890625,
0.83056640625,
0.489013671875,
0.83056640625,
0.489013671875,
0.84423828125,
0.1168212890625,
0.84423828125
],
[
0.5126953125,
0.830078125,
0.88232421875,
0.83056640625,
0.88232421875,
0.845703125,
0.5126953125,
0.84423828125
],
[
0.11846923828125,
0.84716796875,
0.49072265625,
0.84716796875,
0.49072265625,
0.86083984375,
0.11846923828125,
0.86083984375
],
[
0.5126953125,
0.845703125,
0.884765625,
0.84716796875,
0.884765625,
0.8623046875,
0.5126953125,
0.85986328125
],
[
0.11846923828125,
0.86328125,
0.49072265625,
0.86328125,
0.49072265625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.5126953125,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.5126953125,
0.876953125
],
[
0.11846923828125,
0.87744140625,
0.492431640625,
0.87890625,
0.492431640625,
0.8935546875,
0.11846923828125,
0.892578125
],
[
0.5126953125,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.5126953125,
0.8935546875
],
[
0.12017822265625,
0.8955078125,
0.492431640625,
0.8955078125,
0.492431640625,
0.9091796875,
0.12017822265625,
0.9091796875
],
[
0.51416015625,
0.8955078125,
0.88330078125,
0.8955078125,
0.88330078125,
0.9091796875,
0.51416015625,
0.9091796875
]
] |
[
"2 Related Work",
"nities has been investigated as meaning variation",
"in online communities, leveraging the large-scale",
"DiachronicLSC Detection. Existing",
"ap-",
"data which has become available thanks to on",
"proaches for diachronic LSC detection are mainly",
"line social platforms (Del Tredici and Fernandez,",
"based on three types of meaning representations:",
"2017: Rotabi et al., 2017)",
"(i) semantic vector spaces, (ii) topic distribu-",
"Evaluation.Existing evaluation procedures for",
"tions, and (ii) sense clusters. In (i), semantic",
"LSC detection can be distinguished into evaluation",
"vector spaces, each word is represented as two",
"on (i) empirically observed data, and (i) synthetic",
"vectors reflecting its co-occurrence statistics at",
"data or related tasks. (i) includes case studies of in-",
"different periods of time (Gulordava and Baroni,",
"2011; Kim et al., 2014; Xuand Kemp, 2015;",
"dividual words (Sagi et al., 2009; Jatowt and Duh,",
"Eger and Mehler, 2016; Hamilton et al., 2016a,b:",
"2014;Hamilton etal.,2016a),stand-alone",
"Hellrich and Hahn, 2016;Rosenfeld and Erk,",
"comparison of a few hand-selected words",
"(Wijaya and Yeniterzi, 2011; Hamilton et al.,",
"2018). LSC is typically measured by the cosine",
"2016b;",
"Del Tredici and Fernandez,",
"2017)",
"distance(orsomealternativemetric)betweenthe",
"comparison of hand-selected changing vs. seman-",
"two vectors, or by differences in contextual dis-",
"tically stable words (Lau et al., 2012: Cook et al..",
"persion between the two vectors (Kisselew et al.,",
"2016: Schlechtweg et al., 2017). (ii) Diachronic",
"2014), and post-hoc evaluation of the predictions",
"of the presented models (Cook and Stevenson,",
"topic models infer a probability distribution",
"2010: Kulkarni et al., 2015; Del Tredici et al.,",
"for each word over different word senses (or",
"2016;",
"Eger and Mehler,",
"2016;Ferrari et al..",
"2017).",
"topics), which are in turn modeled as a distri-",
"Schlechtweg et al. (2017) propose a",
"bution over words (Wang and McCallum, 2006;",
"Bamman and Crane, 2011; Wijaya and Yeniterzi,",
"small-scale annotation of diachronic metaphoric",
"2011; Lau et al., 2012; Mihalcea and Nastase,",
"change.",
"Synthetic",
"evaluation",
"procedures",
"(ii)",
"LSC",
"2012; Cook et al.,2014; Frermann and Lapata,",
"include",
"studies",
"that",
"simulate",
"(Cook and Stevenson,",
"2016). LSC of a word is measured by calculating",
"2010:",
"Kulkarni et al.,",
"a novelty score for its senses based on their",
"",
"2015;Rosenfeld and Erk,2018),evaluate sense",
"frequency of use. (i) Clustering models assign",
"all uses of a word into sense clusters based on",
"assignments in WordNet (Mitra et al., 2015;",
"Frermann and Lapata,2016),identify",
"text",
"creation dates,(Mihalcea and Nastase, 2012;",
"some contextual property (Mitra et al., 2015).",
"Word sense clustering models are similar to",
"Frermann and Lapata, 2016), or predict the log-",
"topic models in that they map uses to senses.",
"likelihood of textual data (Frermann and Lapata,",
"Accordingly, LSC of a word is measured similarly",
"2016).",
"as in (i).For an overview on diachronic LSC",
"detection, see Tahmasebi et al. (2018).",
"Overall, the various studies use different evalu-",
"ation tasks and data. with little overlap. Most eval-",
"Synchronic LSC Detection. We use the term",
"uation data has not been annotated. Models were",
"synchronic LSC to refer to NLP research ar-",
"rarely compared to previously suggested ones, es-",
"eas with a focus on how the meanings of",
"pecially if the models differed in meaning repre-",
"sentations. Moreover, for the diachronic task, syn-",
"words vary across domains or communities of",
"SynchronicLSCperseis not",
"thetic datasets are used which do not reflect actual",
"speakers.",
"diachronic changes.",
"widely researched; for meaning shifts across do-",
"mains, there is strongly related research which",
"Task and Data",
"is concerned with domain-specific word sense",
"Y",
"disambiguation (Maynard and Ananiadou, 1998;",
"Our study makes use of the evaluation framework",
"Chen and Al-Mubaid, 2006; Taghipour and Ng,",
"proposed in Schlechtweg et al. (2018), where di-",
"2015; Daille et al., 2016) or term ambiguity detec-",
"achronicLSC detectionisdefined as a comparison",
"tion (Baldwin et al., 2013; Wang et al., 2013). The",
"between word uses in two time-specific corpora.",
"only notable work for explicitly measuring across",
"Wefurther applied theframeworkto create an",
"domain meaning shifts is Ferrari et al. (2017),",
" analogous synchronic LSC dataset that compares",
"which is based on semantic vector spaces and co-",
"word uses across general-language and domain-",
"sine distance. Synchronic LSC across commu-",
"specific corpora. The common, meta-level task in"
] |
{
"arxiv_id": "1906.02979",
"language": "en",
"page_idx": 1
}
|
||
1906.02979-en-3
|
1906.02979-en-3.jpg
|
each row vector represents a word $w$ in the vocabulary $V$ reflecting its co-occurrence statistics (Turney and Pantel, 2010). We compare two stateof-the-art approaches to learn these vectors from co-occurrence data, (i) counting and (ii) predicting, and construct vector spaces for each time period and domain.
## 4.1.1 Count-based Vector Spaces
In a count-based semantic vector space the matrix $M$ is high-dimensional and sparse. The value of each matrix cell $M_{i,j}$ represents the number of cooccurrences of the word $w_i$ and the context $c_j$, $\# ( w_i, c_j) .$ In line with Hamilton et al. (2016b) we apply a number of transformations to these raw co-occurrence matrices, as previous work has shown that this improves results on different tasks (Bullinaria and Levy, 2012; Levy et al., 2015).
Positive Pointwise Mutual Information (PPMI).
In PPMI representations the co-occurrence counts in each matrix cell $M_{i,j}$ are weighted by the positive mutual information of target $w_i$ and context $c_j$ reflecting their degree of association. The values of the transformed matrix are
where $k$ > 1 is a prior on the probability of observing an actual occurrence of $(w_i,c_j)$ and $0<\alpha<1$ is a smoothing parameter reducing PPMI's bias towards rare words (Levy and Goldberg, 2014; Levy et al., 2015).
Singular Value Decomposition (SVD). Truncated SVD finds the optimal rank $d$ factorization of matrix $M$ with respect to L2 loss (Eckart and Young, 1936). We use truncated SVD to obtain low-dimensional approximations of the PPMI representations by factorizing $M^\mathrm{PPMI}$ into the product of the three matrices $U\Sigma V^\top.$ We keep only the top $d$ elements of $\Sigma$ and obtain
$$M^{\mathrm{SVD}}=U_d\Sigma_d^p,$$
where $p$ is an eigenvalue weighting parameter (Levy et al., 2015). The $i$th row of $M^\mathrm{SVD}$ corresponds to $w_i’$s $d$-dimensional representation.
Random Indexing (RI). RI is a dimensionality reduction technique based on the JohnsonLindenstrauss lemma according to which points in a vector space can be mapped into a randomly
selected subspace under approximate preservation of the distances between points, if the subspace has a sufficiently high dimensionality (Johnson and Lindenstrauss, 1984; Sahlgren, 2004). We reduce the dimensionality of a countbased matrix $M$ by multiplying it with a random matrix $R:$
$$M^{\mathrm{RI}}=MR^{|\mathcal{V}|\times d},$$
where the $i$th row of $M^\mathrm{RI}$ corresponds to $w_i’$s $d$- dimensional semantic representation. The choice of the random vectors corresponding to the rows in $R$ is important for RI. We follow previous work (Basile et al., 2015) and use sparse ternary random vectors with a small number $s$ of randomly distributed -1s and +1s, all other elements set to 0, and we apply subsampling with a threshold $t.$
## 4.1.2 Predictive Vector Spaces
Skip-Gram with Negative Sampling (SGNS) $\tilde{\text{differs from count-based techniques in that it di-}}$ rectly represents each word $w\in V$ and each context $c\in V$ as a $d$-dimensional vector by implicitly factorizing $M=WC^\top$ when solving
$$M_{i,j}^{\mathrm{PPM}}=\max\left\{\log\left(\frac{\#(w_{i},c_{j})\sum_{c}\#(c)^{\alpha}}{\#(w_{i})\#(c_{j})^{\alpha}}\right)-\log(k),0\right\},\quad\arg\max_{\theta}\sum_{(w,c)\in D}\log\sigma(v_{c}\cdot v_{w})+\sum_{(w,c)\in D^{\prime}}\log\sigma(-v_{c}\cdot v_{w}),$$
where $\sigma(x)=\frac1{1+e^{-x}},D$ is the set of all observed word-context pairs and $D^{\prime}$ is the set of randomly generated negative samples (Mikolov et al., 2013a,b; Goldberg and Levy, 2014). The optimized parameters $\theta$ are $v_{c_i}=C_{i*}$ and $v_{w_i}=W_{i*}$ for $w,c\in V,i\in1,...,d.D^{\prime}$ is obtained by drawing $k$ contexts from the empirical unigram distribution $P(c)=\frac{\#(c)}{|D|}$ for each observation of (w,c), cf. Levy et al. (2015). SGNS and PPMI representations are highly related in that the cells of the implicitly factorized matrix $M$ are PPMI values shifted by the constant $k$ (Levy and Goldberg, 2014). Hence, SGNS and PPMI share the hyperparameter $k.$ The final SGNS matrix is given by
$$M^{\mathrm{SGNS}}=W,$$
where the $i$th row of $M^\mathrm{SGNS}$ corresponds to $w_i’$s $d$-dimensional semantic representation. As in RI we apply subsampling with a threshold $t.$ SGNS with particular parameter configurations has shown to outperform transformed count-based techniques on a variety of tasks (Baroni et al., 2014;Levy et al., 2015).
|
[
[
0.12017822265625,
0.07659912109375,
0.49072265625,
0.07659912109375,
0.49072265625,
0.09027099609375,
0.12017822265625,
0.09027099609375
],
[
0.51611328125,
0.07659912109375,
0.88232421875,
0.07659912109375,
0.88232421875,
0.09027099609375,
0.51611328125,
0.09027099609375
],
[
0.12017822265625,
0.0926513671875,
0.49072265625,
0.0926513671875,
0.49072265625,
0.105712890625,
0.12017822265625,
0.105712890625
],
[
0.5126953125,
0.0914306640625,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.5126953125,
0.105712890625
],
[
0.11846923828125,
0.10687255859375,
0.489013671875,
0.10809326171875,
0.489013671875,
0.122314453125,
0.11846923828125,
0.1217041015625
],
[
0.51416015625,
0.10809326171875,
0.88232421875,
0.10687255859375,
0.88232421875,
0.1217041015625,
0.51416015625,
0.122314453125
],
[
0.12017822265625,
0.12469482421875,
0.49072265625,
0.12469482421875,
0.49072265625,
0.1395263671875,
0.12017822265625,
0.1395263671875
],
[
0.51416015625,
0.12353515625,
0.88232421875,
0.12469482421875,
0.88232421875,
0.1395263671875,
0.51416015625,
0.1383056640625
],
[
0.121826171875,
0.142578125,
0.4873046875,
0.142578125,
0.4873046875,
0.152587890625,
0.121826171875,
0.152587890625
],
[
0.51611328125,
0.14013671875,
0.88232421875,
0.14013671875,
0.88232421875,
0.15380859375,
0.51611328125,
0.15380859375
],
[
0.11846923828125,
0.15673828125,
0.489013671875,
0.15673828125,
0.489013671875,
0.17041015625,
0.11846923828125,
0.17041015625
],
[
0.51416015625,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.171630859375,
0.51416015625,
0.171630859375
],
[
0.1168212890625,
0.1728515625,
0.2479248046875,
0.1728515625,
0.2479248046875,
0.1859130859375,
0.1168212890625,
0.1859130859375
],
[
0.51416015625,
0.1728515625,
0.591796875,
0.1728515625,
0.591796875,
0.18701171875,
0.51416015625,
0.18701171875
],
[
0.11846923828125,
0.19775390625,
0.38916015625,
0.198974609375,
0.38916015625,
0.213134765625,
0.11846923828125,
0.2120361328125
],
[
0.62939453125,
0.1953125,
0.76904296875,
0.19482421875,
0.76904296875,
0.213134765625,
0.62939453125,
0.21435546875
],
[
0.12017822265625,
0.218505859375,
0.4873046875,
0.218505859375,
0.4873046875,
0.2327880859375,
0.12017822265625,
0.2327880859375
],
[
0.51611328125,
0.2244873046875,
0.88330078125,
0.2244873046875,
0.88330078125,
0.2392578125,
0.51611328125,
0.2392578125
],
[
0.12017822265625,
0.235107421875,
0.489013671875,
0.235107421875,
0.489013671875,
0.248779296875,
0.12017822265625,
0.248779296875
],
[
0.51416015625,
0.2423095703125,
0.884765625,
0.2423095703125,
0.884765625,
0.255859375,
0.51416015625,
0.255859375
],
[
0.11846923828125,
0.25048828125,
0.4873046875,
0.2493896484375,
0.4873046875,
0.26416015625,
0.11846923828125,
0.265380859375
],
[
0.51416015625,
0.260009765625,
0.88232421875,
0.260009765625,
0.88232421875,
0.270751953125,
0.51416015625,
0.270751953125
],
[
0.1168212890625,
0.2666015625,
0.4873046875,
0.267333984375,
0.4873046875,
0.281982421875,
0.1168212890625,
0.28076171875
],
[
0.51416015625,
0.273681640625,
0.88330078125,
0.273681640625,
0.88330078125,
0.2880859375,
0.51416015625,
0.2880859375
],
[
0.121826171875,
0.283935546875,
0.489013671875,
0.283935546875,
0.489013671875,
0.297607421875,
0.121826171875,
0.297607421875
],
[
0.51416015625,
0.289306640625,
0.884765625,
0.290283203125,
0.884765625,
0.3046875,
0.51416015625,
0.303955078125
],
[
0.12017822265625,
0.300537109375,
0.489013671875,
0.300537109375,
0.489013671875,
0.314208984375,
0.12017822265625,
0.314208984375
],
[
0.51611328125,
0.30810546875,
0.88232421875,
0.30810546875,
0.88232421875,
0.318359375,
0.51611328125,
0.318359375
],
[
0.12017822265625,
0.31640625,
0.49072265625,
0.31640625,
0.49072265625,
0.330078125,
0.12017822265625,
0.330078125
],
[
0.5126953125,
0.32177734375,
0.88330078125,
0.322509765625,
0.88330078125,
0.33740234375,
0.5126953125,
0.336181640625
],
[
0.11846923828125,
0.33203125,
0.489013671875,
0.33203125,
0.489013671875,
0.345703125,
0.11846923828125,
0.345703125
],
[
0.51416015625,
0.339111328125,
0.85205078125,
0.339111328125,
0.85205078125,
0.35400390625,
0.51416015625,
0.35400390625
],
[
0.11846923828125,
0.347412109375,
0.4638671875,
0.347412109375,
0.4638671875,
0.3623046875,
0.11846923828125,
0.3623046875
],
[
0.51416015625,
0.3681640625,
0.765625,
0.3681640625,
0.765625,
0.3818359375,
0.51416015625,
0.3818359375
],
[
0.1168212890625,
0.377197265625,
0.4873046875,
0.3759765625,
0.4873046875,
0.39013671875,
0.1168212890625,
0.391357421875
],
[
0.11846923828125,
0.393798828125,
0.49072265625,
0.393798828125,
0.49072265625,
0.40673828125,
0.11846923828125,
0.40673828125
],
[
0.51416015625,
0.392578125,
0.88330078125,
0.392578125,
0.88330078125,
0.40625,
0.51416015625,
0.40625
],
[
0.11846923828125,
0.410400390625,
0.489013671875,
0.410400390625,
0.489013671875,
0.42333984375,
0.11846923828125,
0.42333984375
],
[
0.51416015625,
0.4091796875,
0.88232421875,
0.4091796875,
0.88232421875,
0.4228515625,
0.51416015625,
0.4228515625
],
[
0.12017822265625,
0.42578125,
0.489013671875,
0.42578125,
0.489013671875,
0.440673828125,
0.12017822265625,
0.440673828125
],
[
0.51416015625,
0.424560546875,
0.88232421875,
0.424560546875,
0.88232421875,
0.438232421875,
0.51416015625,
0.438232421875
],
[
0.11846923828125,
0.4423828125,
0.49072265625,
0.4423828125,
0.49072265625,
0.4560546875,
0.11846923828125,
0.4560546875
],
[
0.5126953125,
0.439453125,
0.88232421875,
0.440673828125,
0.88232421875,
0.454833984375,
0.5126953125,
0.45361328125
],
[
0.1168212890625,
0.457275390625,
0.3369140625,
0.457763671875,
0.3369140625,
0.47265625,
0.1168212890625,
0.471435546875
],
[
0.51171875,
0.45361328125,
0.79736328125,
0.454833984375,
0.79736328125,
0.47265625,
0.51171875,
0.471435546875
],
[
0.257080078125,
0.485107421875,
0.396728515625,
0.485107421875,
0.396728515625,
0.49951171875,
0.257080078125,
0.49951171875
],
[
0.57958984375,
0.486328125,
0.73193359375,
0.486328125,
0.73193359375,
0.50048828125,
0.57958984375,
0.50048828125
],
[
0.765625,
0.485107421875,
0.88330078125,
0.486328125,
0.88330078125,
0.50048828125,
0.765625,
0.49951171875
],
[
0.74365234375,
0.490478515625,
0.75390625,
0.490478515625,
0.75390625,
0.49462890625,
0.74365234375,
0.49462890625
],
[
0.1226806640625,
0.4912109375,
0.1646728515625,
0.4912109375,
0.1646728515625,
0.5048828125,
0.1226806640625,
0.5048828125
],
[
0.51611328125,
0.488037109375,
0.57568359375,
0.488037109375,
0.57568359375,
0.49951171875,
0.51611328125,
0.49951171875
],
[
0.57568359375,
0.4912109375,
0.5830078125,
0.4912109375,
0.5830078125,
0.495361328125,
0.57568359375,
0.495361328125
],
[
0.28076171875,
0.501953125,
0.3671875,
0.501953125,
0.3671875,
0.51318359375,
0.28076171875,
0.51318359375
],
[
0.41943359375,
0.493408203125,
0.46044921875,
0.493408203125,
0.46044921875,
0.50439453125,
0.41943359375,
0.50439453125
],
[
0.58740234375,
0.50732421875,
0.603515625,
0.50732421875,
0.603515625,
0.51025390625,
0.58740234375,
0.51025390625
],
[
0.72705078125,
0.5029296875,
0.77490234375,
0.501953125,
0.77490234375,
0.51025390625,
0.72705078125,
0.51123046875
],
[
0.11846923828125,
0.525390625,
0.486572265625,
0.525390625,
0.486572265625,
0.5390625,
0.11846923828125,
0.5390625
],
[
0.51416015625,
0.529296875,
0.6435546875,
0.529296875,
0.6435546875,
0.5419921875,
0.51416015625,
0.5419921875
],
[
0.6796875,
0.52783203125,
0.88232421875,
0.52783203125,
0.88232421875,
0.54150390625,
0.6796875,
0.54150390625
],
[
0.1168212890625,
0.54052734375,
0.489013671875,
0.5419921875,
0.489013671875,
0.55712890625,
0.1168212890625,
0.5546875
],
[
0.51416015625,
0.54443359375,
0.88232421875,
0.54443359375,
0.88232421875,
0.55859375,
0.51416015625,
0.55859375
],
[
0.1168212890625,
0.55712890625,
0.489013671875,
0.55810546875,
0.489013671875,
0.572265625,
0.1168212890625,
0.57177734375
],
[
0.51611328125,
0.56005859375,
0.88232421875,
0.56005859375,
0.88232421875,
0.57373046875,
0.51611328125,
0.57373046875
],
[
0.12017822265625,
0.57373046875,
0.49072265625,
0.57373046875,
0.49072265625,
0.58740234375,
0.12017822265625,
0.58740234375
],
[
0.51416015625,
0.57666015625,
0.88232421875,
0.57666015625,
0.88232421875,
0.59033203125,
0.51416015625,
0.59033203125
],
[
0.12017822265625,
0.59033203125,
0.4580078125,
0.59033203125,
0.4580078125,
0.60498046875,
0.12017822265625,
0.60498046875
],
[
0.5126953125,
0.5927734375,
0.88330078125,
0.59130859375,
0.88330078125,
0.60546875,
0.5126953125,
0.60693359375
],
[
0.5126953125,
0.609375,
0.88232421875,
0.60693359375,
0.88232421875,
0.62060546875,
0.5126953125,
0.62255859375
],
[
0.12017822265625,
0.619140625,
0.49072265625,
0.619140625,
0.49072265625,
0.6328125,
0.12017822265625,
0.6328125
],
[
0.51611328125,
0.62451171875,
0.88232421875,
0.62451171875,
0.88232421875,
0.6396484375,
0.51611328125,
0.6396484375
],
[
0.11846923828125,
0.6357421875,
0.4873046875,
0.6357421875,
0.4873046875,
0.6494140625,
0.11846923828125,
0.6494140625
],
[
0.51171875,
0.63720703125,
0.88671875,
0.63720703125,
0.88671875,
0.65771484375,
0.51171875,
0.65771484375
],
[
0.11846923828125,
0.65087890625,
0.486572265625,
0.65087890625,
0.486572265625,
0.66455078125,
0.11846923828125,
0.66455078125
],
[
0.5126953125,
0.65673828125,
0.880859375,
0.65673828125,
0.880859375,
0.67041015625,
0.5126953125,
0.67041015625
],
[
0.11846923828125,
0.66748046875,
0.489013671875,
0.6650390625,
0.489013671875,
0.6796875,
0.11846923828125,
0.681640625
],
[
0.51611328125,
0.6748046875,
0.884765625,
0.6748046875,
0.884765625,
0.6884765625,
0.51611328125,
0.6884765625
],
[
0.12017822265625,
0.68408203125,
0.489013671875,
0.68408203125,
0.489013671875,
0.69775390625,
0.12017822265625,
0.69775390625
],
[
0.51416015625,
0.689453125,
0.880859375,
0.689453125,
0.880859375,
0.703125,
0.51416015625,
0.703125
],
[
0.1168212890625,
0.69775390625,
0.49072265625,
0.6953125,
0.49072265625,
0.71337890625,
0.1168212890625,
0.71533203125
],
[
0.51611328125,
0.7060546875,
0.88232421875,
0.7060546875,
0.88232421875,
0.72021484375,
0.51611328125,
0.72021484375
],
[
0.12017822265625,
0.71435546875,
0.489013671875,
0.71435546875,
0.489013671875,
0.728515625,
0.12017822265625,
0.728515625
],
[
0.5126953125,
0.72021484375,
0.88232421875,
0.72265625,
0.88232421875,
0.73681640625,
0.5126953125,
0.7353515625
],
[
0.12017822265625,
0.73291015625,
0.413330078125,
0.73291015625,
0.413330078125,
0.74658203125,
0.12017822265625,
0.74658203125
],
[
0.5126953125,
0.73828125,
0.869140625,
0.73828125,
0.869140625,
0.7529296875,
0.5126953125,
0.7529296875
],
[
0.239501953125,
0.751953125,
0.365478515625,
0.7548828125,
0.36376953125,
0.77685546875,
0.23779296875,
0.77392578125
],
[
0.6435546875,
0.76025390625,
0.7529296875,
0.7626953125,
0.7529296875,
0.78076171875,
0.6435546875,
0.779296875
],
[
0.12017822265625,
0.78369140625,
0.4873046875,
0.78369140625,
0.4873046875,
0.79736328125,
0.12017822265625,
0.79736328125
],
[
0.51171875,
0.7861328125,
0.88671875,
0.78759765625,
0.88671875,
0.80810546875,
0.51171875,
0.80712890625
],
[
0.115966796875,
0.7958984375,
0.49072265625,
0.79638671875,
0.49072265625,
0.81787109375,
0.115966796875,
0.81640625
],
[
0.51611328125,
0.80859375,
0.884765625,
0.80859375,
0.884765625,
0.822265625,
0.51611328125,
0.822265625
],
[
0.12017822265625,
0.81787109375,
0.44873046875,
0.81787109375,
0.44873046875,
0.83056640625,
0.12017822265625,
0.83056640625
],
[
0.5126953125,
0.822265625,
0.88330078125,
0.82373046875,
0.88330078125,
0.83837890625,
0.5126953125,
0.83740234375
],
[
0.51416015625,
0.8388671875,
0.88330078125,
0.84033203125,
0.88330078125,
0.85498046875,
0.51416015625,
0.85400390625
],
[
0.11846923828125,
0.8466796875,
0.4873046875,
0.8466796875,
0.4873046875,
0.85986328125,
0.11846923828125,
0.85986328125
],
[
0.51416015625,
0.85693359375,
0.884765625,
0.85693359375,
0.884765625,
0.87060546875,
0.51416015625,
0.87060546875
],
[
0.11846923828125,
0.86328125,
0.49072265625,
0.86328125,
0.49072265625,
0.876953125,
0.11846923828125,
0.876953125
],
[
0.1168212890625,
0.87744140625,
0.492431640625,
0.87890625,
0.492431640625,
0.8935546875,
0.1168212890625,
0.892578125
],
[
0.51416015625,
0.87353515625,
0.884765625,
0.87353515625,
0.884765625,
0.88720703125,
0.51416015625,
0.88720703125
],
[
0.51416015625,
0.8896484375,
0.697265625,
0.8896484375,
0.697265625,
0.90283203125,
0.51416015625,
0.90283203125
],
[
0.1168212890625,
0.8955078125,
0.486572265625,
0.8955078125,
0.486572265625,
0.9091796875,
0.1168212890625,
0.9091796875
]
] |
[
"each row vector represents a word w in the vo-",
"selected subspace under approximate preserva-",
"cabulary V reflecting its co-occurrence statistics",
"tion of the distances between points, if the",
"(Turney and Pantel, 2010). We compare two state-",
"subspace has a sufficiently high dimensional-",
"of-the-art approaches to learn these vectors from",
"ity (Johnson and Lindenstrauss, 1984; Sahlgren,",
"co-occurrence data(i) counting and (ii)predict-",
"2004). We reduce the dimensionality of a count-",
"ing, and construct vector spaces for each time pe-",
"based matrix M by multiplying it with a random",
"riod and domain.",
"matrix R:",
"4.1.1 Count-based Vector Spaces",
"MRI= MRIVIxd,",
"In a count-based semantic vector space the matrix",
"where the ith row of MRI corresponds to w;'s d-",
"M is high-dimensional and sparse. The value of",
"dimensional semantic representation. The choice",
"each matrix cell Mi.j represents the number of co-",
"of the random vectors corresponding to the rows",
"occurrences of the word wi and the context Cj,",
"in R is important for RI. We follow previous work",
"#(wi,c,). In line with Hamilton et al. (2016b)",
"(Basile et al., 2015) and use sparse ternary random",
"we apply a number of transformations to these",
"vectors with a small number s of randomly dis-",
"raw co-occurrence matrices, as previous work has",
"tributed -1s and +1s, all other elements set to O,",
"shown that this improves results on different tasks",
"and we apply subsampling with a threshold t.",
"(Bullinaria and Levy, 2012; Levy et al., 2015).",
"4.1.2Predictive Vector Spaces",
"Positive Pointwise Mutual Information (PPMI).",
"In PPMI representations the co-occurrence counts",
"Skip-Gram with Negative Sampling (SGNS)",
"in each matrix cell Mi. are weighted by the posi-",
"differs from count-based techniques in that it di-",
"tive mutual information of target w; and context cj",
"rectly represents each word w E V and each con-",
"reflecting their degree of association. The values",
"text c e V as a d-dimensional vector by implicitly",
"of the transformed matrix are",
"factorizing M = WCT when solving",
"#(wi,c) 。#(c)",
" loga(ve-vw)+",
"log a(-ve*ww),",
"",
"MPPMI",
"argmax",
"",
"#(wi)#(ci)",
"log(k",
"",
"(W,c)ED",
"where k > 1 is a prior on the probability",
"where α(α) =",
",D is the set of all ob-",
"of observing an actual occurrence of (wi,c,)",
"served word-context pairs and D' is the set of ran-",
"and 0 < α< 1 is a smoothing param-",
"domly generated negative samples (Mikolov et al.,",
"eter reducing PPMI's bias towards rare words",
"2013a,b; Goldberg and Levy, 2014). The opti-",
"(Levy and Goldberg, 2014; Levy et al., 2015).",
"mized parameters are Uc, = Ci* and Vw - Wi",
"for w,c eV, i el,....d.D' is obtained by draw-",
"Singular Value Decomposition (SVD). Trun-",
"ing k contexts from the empirical unigram distri-",
"cated SVD finds the optimal rank d factor-",
"bution P(c) = # for each observation of (w,c),",
"ization of matrix M with respect to L2 loss",
"cf. Levy et al. (2015).SGNS and PPMI repre-",
"(Eckart and Young, 1936). We use truncated SVD",
"sentations are highly related in that the cells of",
"to obtain low-dimensional approximations of the",
"the implicitly factorized matrix M are PPMI val-",
"PPMI representations by factorizing MPPMI into",
"ues shifted by the constant k (Levy and Goldberg,",
"the product of the three matrices UVT : We keep",
"2014). Hence, SGNS and PPMI share the hyper-",
"only the top d elements of and obtain",
"parameter k. The final SGNS matrix is given by",
"MSVD = UaZa,",
"MSGNS = W,",
"where p is an eigenvalue weighting parameter",
"where the ith row of MSGNS corresponds to w's",
"(Levy et al., 2015). The ith row of MsvD corre-",
"d-dimensional semantic representation.As in",
"sponds to wi's d-dimensional representation.",
"RI we apply subsampling with a threshold t.",
"SGNS with particular parameter configurations",
"Random Indexing (RD). RI is a dimensional-",
"has shown to outperform transformed count-based",
"ity reduction technique based on the Johnson-",
"Lindenstrauss lemma according to which points",
"techniques on a variety of tasks (Baroni et al.,",
"2014; Levy et al., 2015).",
"in a vector space can be mapped into a randomly"
] |
{
"arxiv_id": "1906.02979",
"language": "en",
"page_idx": 3
}
|
||
1906.02979-en-4
|
1906.02979-en-4.jpg
|
## 4.1.3 Alignment
$\textbf{Column Intersection ( CI) . }$ In order to make the matrices $A$ and $B$ from time periods $a<b$ (or domains $a$ and $b)$ comparable, they have to be aligned via a common coordinate axis. This is rather straightforward for count and PPMI representations, because their columns correspond to context words which often occur in both $A$ and $B$ (Hamilton et al., 2016b). In this case, the alignment for $A$ and $B$ is
$$\begin{aligned}A_{*j}^{\text{CI}}&=A_{*j}\quad\text{for all}\:c_j\in V_a\cap V_b,\\B_{*j}^{\text{CI}}&=B_{*j}\quad\text{for all}\:c_j\in V_a\cap V_b,\end{aligned}$$
where $X_*j$ denotes the $j$th column of $X.$
Shared Random Vectors (SRV). RI offers an elegant way to align count-based vector spaces and reduce their dimensionality at the same time (Basile et al., 2015). Instead of multiplying count matrices $A$ and $B$ each by a separate random matrix $R_A$ and $R_B$ they may be multiplied both by the same random matrix $R$ representing them in the same low-dimensional random space. Hence, $A$ and $B$ are aligned by
$$\begin{aligned}A^{\mathrm{SVR}}&=AR,\\B^{\mathrm{SVR}}&=BR.\end{aligned}$$
We follow Basile et al. and adopt a slight variation of this procedure: instead of multiplying both matrices by exactly the same random matrix (corresponding to an intersection of their columns) we first construct a shared random matrix and then multiply $A$ and $B$ by the respective sub-matrix.
Orthogonal Procrustes (OP). In the lowdimensional vector spaces produced by SVD, RI and SGNS the columns may represent different coordinate axes (orthogonal variants) and thus cannot directly be aligned to each other. Following Hamilton et al. (2016b) we apply OP analysis to solve this problem. We represent the dictionary as a binary matrix D, so that $D_{i,j}=1$ if $w_i\in V_b$ (the $i$th word in the vocabulary at time $b)$ corresponds to $w_j\in V_a.$ The goal is then to find the optimal mapping matrix $W^*$ such that the sum of squared Euclidean distances between $B’$s mapping $B_{i*}W$ and $A_{j*}$ for the dictionary entries $D_{i,j}$ is minimized:
$$W^*=\arg\min_W\sum_i\sum_jD_{i,j}\|B_{i*}W-A_{j*}\|^2.$$
Following standard practice we length-normalize and mean-center $A$ and $B$ in a pre-processing step (Artetxe et al., 2017), and constrain $W$ to be orthogonal, which preserves distances within each time period. Under this constraint, minimizing the squared Euclidean distance becomes equivalent to maximizing the dot product when finding the optimal rotational alignment (Hamilton et al., 2016b; Artetxe et al., 2017). The optimal solution for this problem is then given by $W^*=\underline{U}V^\top$, where $B^{\top}DA=U\Sigma V^{\top}$is the SVD of $B^{\top}DA.$ Hence, $A$ and $B$ are aligned by
$$\begin{aligned}A^{\mathrm{OP}}&=A,\\B^{\mathrm{OP}}&=BW^{*},\end{aligned}$$
where $A$ and $B$ correspond to their preprocessed versions. We also experiment with two variants: OP- omits mean-centering (Hamilton et al., 2016b), which is potentially harmful as a better solution may be found after mean-centering. OP+ corresponds to OP with additional pre- and postprocessing steps and has been shown to improve performance in research on bilingual lexicon induction (Artetxe et al., 2018a,b). We apply all OP variants only to the low-dimensional matrices.
Vector Initialization (VI). In VI we first learn $A^\mathrm{VI}$ using standard SGNS and then initialize the SGNS model for learning $B^\mathrm{VI}$ on $A^\mathrm{VI}$ (Kim et al., 2014). The idea is that if a word is used in similar contexts in $a$ and $b$, its vector will be updated only slightly, while more different contexts lead to a stronger update.
Word Injection (WI). Finally, we use the word injection approach by Ferrari et al. (2017) where target words are substituted by a placeholder in one corpus before learning semantic representations, and a single matrix $\tilde{M}^\mathrm{WI}$ is constructed for both corpora after mixing their sentences. The advantage of this approach is that all vector learning methods described above can be directly applied to the mixed corpus, and target vectors are constructed directly in the same space, so no post-hoc alignment is necessary.
### 4.2 $\textbf{Topic Distributions}$
Sense ChANge (SCAN). SCAN models LSC of word senses via smooth and gradual changes in associated topics (Frermann and Lapata, 2016).
|
[
[
0.12017822265625,
0.07598876953125,
0.259765625,
0.07598876953125,
0.259765625,
0.09027099609375,
0.12017822265625,
0.09027099609375
],
[
0.51416015625,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.09027099609375,
0.51416015625,
0.09027099609375
],
[
0.51416015625,
0.0914306640625,
0.88330078125,
0.0926513671875,
0.88330078125,
0.10687255859375,
0.51416015625,
0.105712890625
],
[
0.12017822265625,
0.09857177734375,
0.49072265625,
0.09857177734375,
0.49072265625,
0.11224365234375,
0.12017822265625,
0.11224365234375
],
[
0.51416015625,
0.10809326171875,
0.880859375,
0.10809326171875,
0.880859375,
0.1217041015625,
0.51416015625,
0.1217041015625
],
[
0.11846923828125,
0.11517333984375,
0.49072265625,
0.11517333984375,
0.49072265625,
0.12890625,
0.11846923828125,
0.12890625
],
[
0.5126953125,
0.12469482421875,
0.88232421875,
0.12469482421875,
0.88232421875,
0.1383056640625,
0.5126953125,
0.1383056640625
],
[
0.11846923828125,
0.1312255859375,
0.49072265625,
0.1312255859375,
0.49072265625,
0.144287109375,
0.11846923828125,
0.144287109375
],
[
0.51416015625,
0.14013671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.15380859375,
0.51416015625,
0.15380859375
],
[
0.12017822265625,
0.1478271484375,
0.49072265625,
0.1478271484375,
0.49072265625,
0.160888671875,
0.12017822265625,
0.160888671875
],
[
0.11846923828125,
0.162109375,
0.489013671875,
0.162109375,
0.489013671875,
0.17578125,
0.11846923828125,
0.17578125
],
[
0.51416015625,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.17041015625,
0.51416015625,
0.17041015625
],
[
0.11846923828125,
0.1776123046875,
0.489013671875,
0.1776123046875,
0.489013671875,
0.191162109375,
0.11846923828125,
0.191162109375
],
[
0.51611328125,
0.174560546875,
0.88232421875,
0.174560546875,
0.88232421875,
0.185302734375,
0.51611328125,
0.185302734375
],
[
0.5126953125,
0.188232421875,
0.88671875,
0.18701171875,
0.88671875,
0.201904296875,
0.5126953125,
0.203125
],
[
0.12017822265625,
0.1953125,
0.489013671875,
0.1953125,
0.489013671875,
0.208984375,
0.12017822265625,
0.208984375
],
[
0.11846923828125,
0.210205078125,
0.4873046875,
0.21142578125,
0.4873046875,
0.2257080078125,
0.11846923828125,
0.2244873046875
],
[
0.51611328125,
0.204833984375,
0.88671875,
0.204833984375,
0.88671875,
0.218505859375,
0.51611328125,
0.218505859375
],
[
0.51416015625,
0.2203369140625,
0.88330078125,
0.2203369140625,
0.88330078125,
0.235107421875,
0.51416015625,
0.235107421875
],
[
0.1168212890625,
0.226806640625,
0.272216796875,
0.2257080078125,
0.272216796875,
0.240478515625,
0.1168212890625,
0.2410888671875
],
[
0.51416015625,
0.235107421875,
0.88330078125,
0.2369384765625,
0.88330078125,
0.251708984375,
0.51416015625,
0.2493896484375
],
[
0.517578125,
0.254150390625,
0.68994140625,
0.254150390625,
0.68994140625,
0.267333984375,
0.517578125,
0.267333984375
],
[
0.1873779296875,
0.26318359375,
0.43115234375,
0.26318359375,
0.43115234375,
0.28076171875,
0.1873779296875,
0.28076171875
],
[
0.1781005859375,
0.283203125,
0.427001953125,
0.283203125,
0.427001953125,
0.303955078125,
0.1781005859375,
0.303955078125
],
[
0.64208984375,
0.27978515625,
0.72705078125,
0.281982421875,
0.72705078125,
0.2998046875,
0.64208984375,
0.297607421875
],
[
0.64208984375,
0.300537109375,
0.75390625,
0.302734375,
0.75390625,
0.320556640625,
0.64208984375,
0.318359375
],
[
0.12017822265625,
0.311279296875,
0.41943359375,
0.311279296875,
0.41943359375,
0.325927734375,
0.12017822265625,
0.325927734375
],
[
0.1168212890625,
0.33740234375,
0.4931640625,
0.339111328125,
0.4931640625,
0.35400390625,
0.1168212890625,
0.3515625
],
[
0.51416015625,
0.336181640625,
0.88232421875,
0.336181640625,
0.88232421875,
0.349853515625,
0.51416015625,
0.349853515625
],
[
0.1168212890625,
0.35400390625,
0.492431640625,
0.355224609375,
0.492431640625,
0.369384765625,
0.1168212890625,
0.3681640625
],
[
0.51416015625,
0.355224609375,
0.603515625,
0.355224609375,
0.603515625,
0.366455078125,
0.51416015625,
0.366455078125
],
[
0.5966796875,
0.35400390625,
0.88232421875,
0.35400390625,
0.88232421875,
0.36767578125,
0.5966796875,
0.36767578125
],
[
0.12017822265625,
0.372314453125,
0.49072265625,
0.372314453125,
0.49072265625,
0.383056640625,
0.12017822265625,
0.383056640625
],
[
0.51416015625,
0.369384765625,
0.88330078125,
0.369384765625,
0.88330078125,
0.383056640625,
0.51416015625,
0.383056640625
],
[
0.11846923828125,
0.38720703125,
0.49072265625,
0.38720703125,
0.49072265625,
0.40087890625,
0.11846923828125,
0.40087890625
],
[
0.51416015625,
0.385986328125,
0.884765625,
0.385986328125,
0.884765625,
0.399658203125,
0.51416015625,
0.399658203125
],
[
0.11846923828125,
0.402099609375,
0.49072265625,
0.402587890625,
0.49072265625,
0.41748046875,
0.11846923828125,
0.416259765625
],
[
0.517578125,
0.40380859375,
0.880859375,
0.40380859375,
0.880859375,
0.41455078125,
0.517578125,
0.41455078125
],
[
0.12017822265625,
0.419189453125,
0.489013671875,
0.419189453125,
0.489013671875,
0.432861328125,
0.12017822265625,
0.432861328125
],
[
0.51416015625,
0.418701171875,
0.88232421875,
0.418701171875,
0.88232421875,
0.431640625,
0.51416015625,
0.431640625
],
[
0.12017822265625,
0.435791015625,
0.49072265625,
0.435791015625,
0.49072265625,
0.449462890625,
0.12017822265625,
0.449462890625
],
[
0.5126953125,
0.43408203125,
0.88330078125,
0.432861328125,
0.88330078125,
0.44775390625,
0.5126953125,
0.448974609375
],
[
0.12017822265625,
0.451904296875,
0.49072265625,
0.451904296875,
0.49072265625,
0.466064453125,
0.12017822265625,
0.466064453125
],
[
0.5126953125,
0.45068359375,
0.88232421875,
0.448974609375,
0.88232421875,
0.4619140625,
0.5126953125,
0.46435546875
],
[
0.11846923828125,
0.466064453125,
0.294189453125,
0.468505859375,
0.294189453125,
0.482177734375,
0.11846923828125,
0.479736328125
],
[
0.51611328125,
0.466064453125,
0.88330078125,
0.466064453125,
0.88330078125,
0.479736328125,
0.51611328125,
0.479736328125
],
[
0.2529296875,
0.486328125,
0.35791015625,
0.488037109375,
0.35791015625,
0.50732421875,
0.2529296875,
0.5048828125
],
[
0.51416015625,
0.48388671875,
0.85205078125,
0.48388671875,
0.85205078125,
0.49462890625,
0.51416015625,
0.49462890625
],
[
0.251220703125,
0.50732421875,
0.356201171875,
0.51025390625,
0.35546875,
0.529296875,
0.2496337890625,
0.525390625
],
[
0.51416015625,
0.5126953125,
0.884765625,
0.51318359375,
0.884765625,
0.52783203125,
0.51416015625,
0.52685546875
],
[
0.51416015625,
0.52685546875,
0.884765625,
0.529296875,
0.884765625,
0.54638671875,
0.51416015625,
0.54443359375
],
[
0.12017822265625,
0.53759765625,
0.489013671875,
0.53759765625,
0.489013671875,
0.55029296875,
0.12017822265625,
0.55029296875
],
[
0.51611328125,
0.5458984375,
0.88330078125,
0.5458984375,
0.88330078125,
0.56005859375,
0.51611328125,
0.56005859375
],
[
0.11846923828125,
0.55419921875,
0.489013671875,
0.55419921875,
0.489013671875,
0.568359375,
0.11846923828125,
0.568359375
],
[
0.51611328125,
0.5625,
0.88330078125,
0.5625,
0.88330078125,
0.576171875,
0.51611328125,
0.576171875
],
[
0.11846923828125,
0.57080078125,
0.4873046875,
0.57080078125,
0.4873046875,
0.58447265625,
0.11846923828125,
0.58447265625
],
[
0.51416015625,
0.5791015625,
0.884765625,
0.5791015625,
0.884765625,
0.5927734375,
0.51416015625,
0.5927734375
],
[
0.12017822265625,
0.5859375,
0.489013671875,
0.5859375,
0.489013671875,
0.599609375,
0.12017822265625,
0.599609375
],
[
0.5126953125,
0.59423828125,
0.884765625,
0.59326171875,
0.884765625,
0.60791015625,
0.5126953125,
0.609375
],
[
0.12017822265625,
0.60400390625,
0.4873046875,
0.60400390625,
0.4873046875,
0.6142578125,
0.12017822265625,
0.6142578125
],
[
0.51416015625,
0.61083984375,
0.64892578125,
0.61083984375,
0.64892578125,
0.62451171875,
0.51416015625,
0.62451171875
],
[
0.11846923828125,
0.6181640625,
0.46728515625,
0.6181640625,
0.46728515625,
0.6318359375,
0.11846923828125,
0.6318359375
],
[
0.12017822265625,
0.6455078125,
0.489013671875,
0.6455078125,
0.489013671875,
0.6591796875,
0.12017822265625,
0.6591796875
],
[
0.51611328125,
0.64111328125,
0.884765625,
0.64111328125,
0.884765625,
0.65625,
0.51611328125,
0.65625
],
[
0.11846923828125,
0.662109375,
0.489013671875,
0.662109375,
0.489013671875,
0.6767578125,
0.11846923828125,
0.6767578125
],
[
0.5126953125,
0.65771484375,
0.88330078125,
0.65673828125,
0.88330078125,
0.67138671875,
0.5126953125,
0.6728515625
],
[
0.11846923828125,
0.677734375,
0.49072265625,
0.677734375,
0.49072265625,
0.6923828125,
0.11846923828125,
0.6923828125
],
[
0.51416015625,
0.6748046875,
0.884765625,
0.6748046875,
0.884765625,
0.6884765625,
0.51416015625,
0.6884765625
],
[
0.11846923828125,
0.69482421875,
0.492431640625,
0.69482421875,
0.492431640625,
0.7080078125,
0.11846923828125,
0.7080078125
],
[
0.51416015625,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.705078125,
0.51416015625,
0.705078125
],
[
0.51171875,
0.70361328125,
0.88671875,
0.703125,
0.88671875,
0.72021484375,
0.51171875,
0.7216796875
],
[
0.12017822265625,
0.71044921875,
0.489013671875,
0.71044921875,
0.489013671875,
0.7236328125,
0.12017822265625,
0.7236328125
],
[
0.1168212890625,
0.724609375,
0.49072265625,
0.7255859375,
0.49072265625,
0.74072265625,
0.1168212890625,
0.7392578125
],
[
0.51416015625,
0.72265625,
0.88330078125,
0.72265625,
0.88330078125,
0.736328125,
0.51416015625,
0.736328125
],
[
0.11846923828125,
0.7421875,
0.489013671875,
0.7421875,
0.489013671875,
0.755859375,
0.11846923828125,
0.755859375
],
[
0.51416015625,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.51416015625,
0.7529296875
],
[
0.11846923828125,
0.75830078125,
0.49072265625,
0.75830078125,
0.49072265625,
0.7724609375,
0.11846923828125,
0.7724609375
],
[
0.5126953125,
0.75439453125,
0.88330078125,
0.75439453125,
0.88330078125,
0.76708984375,
0.5126953125,
0.76708984375
],
[
0.12017822265625,
0.77490234375,
0.49072265625,
0.77490234375,
0.49072265625,
0.7880859375,
0.12017822265625,
0.7880859375
],
[
0.5126953125,
0.76953125,
0.88330078125,
0.77099609375,
0.88330078125,
0.78515625,
0.5126953125,
0.78369140625
],
[
0.11846923828125,
0.79052734375,
0.49072265625,
0.7890625,
0.49072265625,
0.80419921875,
0.11846923828125,
0.8046875
],
[
0.51611328125,
0.78759765625,
0.884765625,
0.78759765625,
0.884765625,
0.80029296875,
0.51611328125,
0.80029296875
],
[
0.12017822265625,
0.80712890625,
0.49072265625,
0.80712890625,
0.49072265625,
0.82177734375,
0.12017822265625,
0.82177734375
],
[
0.5126953125,
0.802734375,
0.685546875,
0.80419921875,
0.685546875,
0.818359375,
0.5126953125,
0.81787109375
],
[
0.12017822265625,
0.82470703125,
0.489013671875,
0.82470703125,
0.489013671875,
0.8349609375,
0.12017822265625,
0.8349609375
],
[
0.11846923828125,
0.8388671875,
0.49072265625,
0.83837890625,
0.49072265625,
0.8525390625,
0.11846923828125,
0.85400390625
],
[
0.51416015625,
0.8359375,
0.70947265625,
0.8359375,
0.70947265625,
0.85107421875,
0.51416015625,
0.85107421875
],
[
0.12017822265625,
0.85498046875,
0.259765625,
0.85400390625,
0.259765625,
0.8681640625,
0.12017822265625,
0.869140625
],
[
0.51416015625,
0.8623046875,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.51416015625,
0.876953125
],
[
0.13525390625,
0.88134765625,
0.265625,
0.88427734375,
0.265625,
0.8984375,
0.13525390625,
0.8955078125
],
[
0.288330078125,
0.88134765625,
0.471435546875,
0.88134765625,
0.471435546875,
0.8984375,
0.288330078125,
0.8984375
],
[
0.5126953125,
0.87744140625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.5126953125,
0.892578125
],
[
0.51611328125,
0.8955078125,
0.88232421875,
0.8955078125,
0.88232421875,
0.9091796875,
0.51611328125,
0.9091796875
],
[
0.257080078125,
0.8828125,
0.267333984375,
0.8828125,
0.267333984375,
0.88720703125,
0.257080078125,
0.88720703125
]
] |
[
"4.1.3Alignment",
"Following standard practice we length-normalize",
"and mean-center A and B in a pre-processing step",
"Column Intersection (CI). In order to make the",
"(Artetxe et al., 2017), and constrain W to be or-",
"matrices A and B from time periods a < b (or",
"thogonal, which preserves distances within each",
"domains a and b) comparable, they have to be",
"time period. Under this constraint, minimizing the",
"aligned via a common coordinate axis.This is",
"rather straightforward for count and PPMI repre-",
"squared Euclidean distance becomes equivalent to",
"sentations. because their columns correspond to",
"maximizing the dotproduct when finding the opti-",
"mal rotational alignment (Hamilton et al., 2016b;",
"context words which often occur in both A and",
"B (Hamilton et al., 2016b). In this case, the align-",
"Artetxe et al., 2017). The optimal solution for this",
"problem is then given by W* = UVT, where",
"ment for A and B is",
"BTDA = UEVT is the SVD of BTDA. Hence,",
"A and B are aligned by",
"A, = Axj for all cj e Va n Vb,",
"B = Bj for all cj E Va n Vb,",
"AOP= A,",
"BOP = BW*,",
"where X*j denotes the jth column of X.",
"Shared Random Vectors (SRV). RI offers an",
"where A and B correspond to their preprocessed",
"elegant way to align count-based vector spaces",
"versions.",
" We also experiment with two vari-",
"andreduce their dimensionality atthe same time",
"ants: OP_ omits mean-centering (Hamilton et al.",
"(Basile et al., 2015). Instead of multiplying count",
"2016b), which is potentially harmful as a better",
"matrices A and B each by a separate random ma-",
"solutionmaybefoundaftermean-centering.OP+",
"trix RA and RB they may be multiplied both by",
"corresponds to OP with additional pre- and post-",
"the same random matrix R representing them in",
"processing steps and has been shown to improve",
"the same low-dimensional random space. Hence,",
"performance in research on bilingual lexicon in-",
"A and B are aligned by",
"duction (Artetxe et al., 2018a,b). We apply all OP",
"ASVR = AR,",
"variants onlytothe low-dimensional matrices",
"BSVR = BR.",
"Vector Initialization (V).In VI we first learn",
"AVI using standard SGNS and then initialize the",
"We follow Basile et al. and adopt a slight varia-",
"SGNS model for learning BVI on AVI (Kim et al.,",
"tion of this procedure: instead of multiplying both",
"2014). The idea is that if a word is used in simi-",
"matrices by exactly the same random matrix (cor-",
"lar contexts in a and b, its vector will be updated",
"responding to an intersection of their columns) we",
"only slightly, while more different contexts lead to",
"first construct a shared random matrix and then",
"a stronger update.",
"multiply A and B by the respective sub-matrix.",
"Orthogonal Procrustes (OP). In the low-",
"Word Injection (WI). Finally, we use the word",
"dimensional vector spaces produced by SVD, RI",
"injection approach by Ferrari et al. (2017) where",
"and SGNS the columns may represent different",
"target words are substituted by a placeholder in",
"coordinate axes (orthogonal variants) and thus",
"one corpus before learning semantic representa-",
"tions, and a single matrix MWI is constructed for",
"cannot directly be aligned to each other. Follow-",
"ing Hamilton et al. (2016b) we apply OP analysis",
"both corpora after mixing their sentences. The ad-",
"to solve this problem.We represent the dictio-",
"vantage of this approach is that all vector learning",
"nary as a binary matrix D, so that Di.i = 1 if",
"methods described above can be directly applied",
"Wi E Ve (the ith word in the vocabulary at time",
"to the mixed corpus, and target vectors are con-",
"b) corresponds to wi E Va. The goal is then to",
"structed directly in the same space, so no post-hoc",
"find the optimal mapping matrix W* such that the",
"alignment is necessary.",
"sum of sguared Euclidean distances between B's",
"mapping BiW and Aj for the dictionary entries",
"4.2Topic Distributions",
"Di, is minimized:",
"Sense ChANge (SCAN). SCAN models LSC",
"W* = arg min",
"DiillBiW-Ajl?",
"of word senses via smooth and gradual changes",
"in associated topics (Frermann and Lapata, 2016).",
""
] |
{
"arxiv_id": "1906.02979",
"language": "en",
"page_idx": 4
}
|
||
1909.11059-en-2
|
1909.11059-en-2.jpg
|
ities. ViLBERT is tested on a variety of downstream tasks including VQA, referring expression, and image-to-text retrieval. LXMERT only focuses on a particular problem space $(i.e.$,VQA and visual reasoning) and the generalization ability further compromises when the datasets from the downstream tasks are also exploited in the pre-training stage. The most similar work to ours is VideoBERT (Sun et al. 2019b), which addresses generation-based tasks $(e.g.$, video captioning) and understanding-based tasks $(e.g.$, action classification). However, it separates the visual encoder and the language decoder and performs pre-training only on the encoder, leaving decoder uninitialized. In contrast, we propose a unified model for both encoding and decoding and fully leverage the benefit of pre-training.
Image Captioning & VQA. Most of the recent works on image captioning are built upon (Anderson et al.2018), where a language model gets clues for sentence generation through dynamically attending on object regions in the image extracted from pre-trained object detectors. Follow-up works further capture the relationships among object regions by using Graph Convolutional Networks (GCNs) (Yao et al. 2018), incorporating language inductive bias (Yang et al. 2019), or enforcing region grounding between image and text (Lu et al. 2018; Zhou et al. 2019). VQA is another prevalent research area in vision and language. Since its initial proposal (Antol et al. 2015), there has been a significant amount of works proposing model architectures to fuse question and image representations (Kim, Jun, and Zhang 2018; Anderson et al. 2018; Gao et al. 2019), new datasets or models to reduce the dataset bias (Zhang et al.2016; Goyal et al. 2017; Agrawal et al. 2017) and ground the answer in the question (Lewis and Fan 2019). We use our base architecture to perform both image captioning and VQA with minor model structure differences.
# Vision-Language Pre-training
We denote the input image as $I$ and the associated/target sentence description (words) as $S.$ We extract a fxed number $N$ of object regions from the image using an off-the-shelf ob- $\begin{array}{l}{\text{ject detector, denoted as}}\left\{r_{1},\ldots,r_{N}\right\}\text{and the correspond-}\\{\text{ing region features as }R=[R_{1},\ldots,R_{N}]\in\mathbb{R}^{d\times N},\text{region}}\\{\text{object labels (probabilities) as }C=[C_{1},\ldots,C_{N}]\in\mathbb{R}^{l\times N},}\end{array}$ and region geometric information as $G=[G_1,\ldots,G_N]\in$ $\mathbb{R}^{o\times N}$,where $d$ is the embedding size, $l$ indicates the number of the object classes of the object detector, and $o=5$ consists of four values for top left and bottom right corner coordinates of the region bounding box (normalized between 0 and 1) and one value for its relative area $(i.e.$, ratio of the bounding box area to the image area, also between 0 and 1). The words in $S$ are represented as one-hot vectors which are further encoded to word embeddings with embedding size $e:$ $y_t\in\mathbb{R}^e$ where $t\in\{1,2,\ldots,T\}$ and $T$ indicates the length of the sentence.
Vision-Language Transformer Network
Our vision-language Transformer network, which unifies the Transformer encoder and decoder into a single model, is depicted in Fig. 2 (left). The model input consists of the class-
aware region embedding, word embedding and three special
tokens. The region embedding is defined as:
$r_{i}=W_{r}R_{i}+W_{p}[$LayerNorm$(W_cC_{i})|$LayerNorm$(W_gG_{i})]$
## (1)
where $[\cdot|\cdot]$ indicates the concatenation on the feature dimen-
sion, LayerNorm represents Layer Normalization. The second term mimics the positional embedding in BERT, but adding extra region class information, and $\tilde{W}_{r},W_{p},W_{c},W_{g}$ are the embedding weights (the bias term and the nonlinearity term are omitted). Note that here we overload the notation of $r_{i}\in \mathbb{R} ^{d}\left ( i\in \{ 1, 2, . . . , N\} \right )$to also represent classaware region embeddings. In addition, we add segment embeddings to $r_i$ as in BERT where all the regions share the same segment embedding where the values depend on the objectives $(i.e.$, seq2seq and bidirectional, see the following section).
The word embeddings are similarly defined as in (Devlin et al. 2018), adding up $y_{t}$ with positional embeddings and segment embeddings, which is again overloaded as $y_t.$ We define three special tokens [CLS], [SEP], [STOP], where $[\hat{\mathrm{CLS}}]$ indicates the start of the visual input, [SEP] marks the boundary between the visual input and the sentence input, and [STOP] determines the end of the sentence. The [MASK] tokens indicate the masked words which will be explained in the next section.
## Pre-training Objectives In the BERT masked language modeling objective, 15% of
the input text tokens are first replaced with either a special [MASK] token, a random token or the original token, at random with chances equal to $80\%,10\%$, and $10\%$, respectively. Then, at the model output, the hidden state from the last Transformer block is projected to word likelihoods where the masked tokens are predicted in the form of a classification problem. Through this reconstruction, the model learns the dependencies in the context and forms a language model. We follow the same scheme and consider two specific objectives: the bidirectional objective (bidirectional) as in BERT and the sequence to sequence objective (seq2seq), inspired by (Dong et al. 2019).
As shown in Fig. 2 (right), the only difference between the two objectives lie in the self-attention mask. The mask used for the bidirectional objective allows unrestricted message passing between the visual modality and the language modality while in seq2seq, the to-bepredicted word cannot attend to the words in the future, $i.e.$, it satisfies the auto-regressive property. More formally, we define the input to the first Transformer block $\begin{array}{rcl}{\mathrm{as~}H^{0}}&{=\:[r_{[\mathrm{CLS]}},r_{1},\ldots,r_{N},y_{[\mathrm{SEP]}},y_{1},\ldots,y_{T},y_{[\mathrm{STOP]}}]\:\in}\\{\mathrm{m}_{d\times U}}&{=\:[r_{[\mathrm{CLS]}},r_{1},\ldots,r_{N},y_{[\mathrm{SEP]}},y_{1},\ldots,y_{T},y_{[\mathrm{STOP]}}]\:\in}\end{array}$ $\mathbb{R}^{d\times U}$ where$U=N+T+3$,and then the encoding at different levels of Transformer as $H^l=$Transformer$(H^l-1),l\in$ $[1,L].$ We further defne a self-attention mask as $M\in$ $\mathbb{R}^{[-,-]}{\mathbb{R}^{U\times U)},\mathrm{~where}}$
$$M_{jk}=\left\{\begin{matrix}0,&\text{allow to attend}\\-\infty,&\text{prevent from attending}\end{matrix}\right.\quad j,k=1,\ldots,U.$$
(2)
For simplicity, we assume a single attention head in the self-
attention module. Then, the self-attention output on $H^{l-1}$
|
[
[
0.08660888671875,
0.0694580078125,
0.47802734375,
0.0694580078125,
0.47802734375,
0.08331298828125,
0.08660888671875,
0.08331298828125
],
[
0.51953125,
0.0694580078125,
0.91259765625,
0.0694580078125,
0.91259765625,
0.08331298828125,
0.51953125,
0.08331298828125
],
[
0.0849609375,
0.08209228515625,
0.476318359375,
0.08331298828125,
0.476318359375,
0.09783935546875,
0.0849609375,
0.09661865234375
],
[
0.51953125,
0.083984375,
0.8095703125,
0.083984375,
0.8095703125,
0.09783935546875,
0.51953125,
0.09783935546875
],
[
0.08660888671875,
0.09783935546875,
0.47802734375,
0.09783935546875,
0.47802734375,
0.111083984375,
0.08660888671875,
0.111083984375
],
[
0.521484375,
0.10101318359375,
0.90869140625,
0.10101318359375,
0.90869140625,
0.11553955078125,
0.521484375,
0.11553955078125
],
[
0.08905029296875,
0.11297607421875,
0.476318359375,
0.11297607421875,
0.476318359375,
0.12371826171875,
0.08905029296875,
0.12371826171875
],
[
0.89208984375,
0.11553955078125,
0.9111328125,
0.11553955078125,
0.9111328125,
0.1275634765625,
0.89208984375,
0.1275634765625
],
[
0.08660888671875,
0.126953125,
0.476318359375,
0.126953125,
0.476318359375,
0.136962890625,
0.08660888671875,
0.136962890625
],
[
0.521484375,
0.1287841796875,
0.9111328125,
0.1287841796875,
0.9111328125,
0.1427001953125,
0.521484375,
0.1427001953125
],
[
0.0849609375,
0.1383056640625,
0.476318359375,
0.1383056640625,
0.476318359375,
0.1514892578125,
0.0849609375,
0.1514892578125
],
[
0.521484375,
0.143310546875,
0.91015625,
0.143310546875,
0.91015625,
0.1572265625,
0.521484375,
0.1572265625
],
[
0.08660888671875,
0.154052734375,
0.476318359375,
0.154052734375,
0.476318359375,
0.1641845703125,
0.08660888671875,
0.1641845703125
],
[
0.51953125,
0.1572265625,
0.91259765625,
0.1572265625,
0.91259765625,
0.17041015625,
0.51953125,
0.17041015625
],
[
0.08660888671875,
0.1673583984375,
0.47802734375,
0.1673583984375,
0.47802734375,
0.18115234375,
0.08660888671875,
0.18115234375
],
[
0.5185546875,
0.17041015625,
0.91259765625,
0.1717529296875,
0.91259765625,
0.1849365234375,
0.5185546875,
0.184326171875
],
[
0.08660888671875,
0.18115234375,
0.476318359375,
0.18115234375,
0.476318359375,
0.1944580078125,
0.08660888671875,
0.1944580078125
],
[
0.521484375,
0.1849365234375,
0.91015625,
0.1849365234375,
0.91015625,
0.1988525390625,
0.521484375,
0.1988525390625
],
[
0.0849609375,
0.1944580078125,
0.476318359375,
0.1944580078125,
0.476318359375,
0.2083740234375,
0.0849609375,
0.2083740234375
],
[
0.52294921875,
0.199462890625,
0.90869140625,
0.199462890625,
0.90869140625,
0.210205078125,
0.52294921875,
0.210205078125
],
[
0.0882568359375,
0.210205078125,
0.476318359375,
0.210205078125,
0.476318359375,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.5185546875,
0.2108154296875,
0.9111328125,
0.212158203125,
0.9111328125,
0.2259521484375,
0.5185546875,
0.2247314453125
],
[
0.08660888671875,
0.222900390625,
0.47802734375,
0.222900390625,
0.47802734375,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.521484375,
0.2259521484375,
0.91259765625,
0.2259521484375,
0.91259765625,
0.2392578125,
0.521484375,
0.2392578125
],
[
0.08660888671875,
0.236083984375,
0.476318359375,
0.236083984375,
0.476318359375,
0.25,
0.08660888671875,
0.25
],
[
0.51953125,
0.240478515625,
0.9140625,
0.240478515625,
0.9140625,
0.25390625,
0.51953125,
0.25390625
],
[
0.0882568359375,
0.251953125,
0.316162109375,
0.251953125,
0.316162109375,
0.261962890625,
0.0882568359375,
0.261962890625
],
[
0.52294921875,
0.25390625,
0.9111328125,
0.25244140625,
0.9111328125,
0.266357421875,
0.52294921875,
0.267578125
],
[
0.08660888671875,
0.264404296875,
0.4794921875,
0.264404296875,
0.4794921875,
0.27783203125,
0.08660888671875,
0.27783203125
],
[
0.521484375,
0.267578125,
0.9140625,
0.267578125,
0.9140625,
0.2822265625,
0.521484375,
0.2822265625
],
[
0.08416748046875,
0.27783203125,
0.47802734375,
0.276611328125,
0.47802734375,
0.291748046875,
0.08416748046875,
0.292236328125
],
[
0.521484375,
0.28271484375,
0.57861328125,
0.28271484375,
0.57861328125,
0.294189453125,
0.521484375,
0.294189453125
],
[
0.08660888671875,
0.292236328125,
0.4794921875,
0.292236328125,
0.4794921875,
0.30615234375,
0.08660888671875,
0.30615234375
],
[
0.53759765625,
0.294189453125,
0.91259765625,
0.294189453125,
0.91259765625,
0.30810546875,
0.53759765625,
0.30810546875
],
[
0.08660888671875,
0.30615234375,
0.47802734375,
0.30615234375,
0.47802734375,
0.3193359375,
0.08660888671875,
0.3193359375
],
[
0.521484375,
0.309326171875,
0.91259765625,
0.309326171875,
0.91259765625,
0.323974609375,
0.521484375,
0.323974609375
],
[
0.08660888671875,
0.3212890625,
0.47802734375,
0.3212890625,
0.47802734375,
0.33203125,
0.08660888671875,
0.33203125
],
[
0.521484375,
0.323974609375,
0.91259765625,
0.323974609375,
0.91259765625,
0.337158203125,
0.521484375,
0.337158203125
],
[
0.08905029296875,
0.335205078125,
0.476318359375,
0.335205078125,
0.476318359375,
0.34521484375,
0.08905029296875,
0.34521484375
],
[
0.521484375,
0.337158203125,
0.91259765625,
0.337158203125,
0.91259765625,
0.35107421875,
0.521484375,
0.35107421875
],
[
0.0849609375,
0.3466796875,
0.476318359375,
0.34521484375,
0.476318359375,
0.35986328125,
0.0849609375,
0.361083984375
],
[
0.52294921875,
0.35107421875,
0.91259765625,
0.35107421875,
0.91259765625,
0.3642578125,
0.52294921875,
0.3642578125
],
[
0.08660888671875,
0.361083984375,
0.476318359375,
0.361083984375,
0.476318359375,
0.375,
0.08660888671875,
0.375
],
[
0.51953125,
0.3642578125,
0.9111328125,
0.3642578125,
0.9111328125,
0.37744140625,
0.51953125,
0.37744140625
],
[
0.0882568359375,
0.375,
0.47802734375,
0.375,
0.47802734375,
0.38818359375,
0.0882568359375,
0.38818359375
],
[
0.52294921875,
0.37744140625,
0.91259765625,
0.37744140625,
0.91259765625,
0.391357421875,
0.52294921875,
0.391357421875
],
[
0.08660888671875,
0.38818359375,
0.4794921875,
0.38818359375,
0.4794921875,
0.401611328125,
0.08660888671875,
0.401611328125
],
[
0.52294921875,
0.391357421875,
0.9140625,
0.391357421875,
0.9140625,
0.40478515625,
0.52294921875,
0.40478515625
],
[
0.0882568359375,
0.40283203125,
0.473876953125,
0.401611328125,
0.473876953125,
0.415283203125,
0.0882568359375,
0.416748046875
],
[
0.51953125,
0.406005859375,
0.71337890625,
0.406005859375,
0.71337890625,
0.419189453125,
0.51953125,
0.419189453125
],
[
0.08660888671875,
0.416748046875,
0.476318359375,
0.416748046875,
0.476318359375,
0.429931640625,
0.08660888671875,
0.429931640625
],
[
0.0849609375,
0.43115234375,
0.4794921875,
0.43115234375,
0.4794921875,
0.4443359375,
0.0849609375,
0.4443359375
],
[
0.521484375,
0.43115234375,
0.701171875,
0.43115234375,
0.701171875,
0.4443359375,
0.521484375,
0.4443359375
],
[
0.0882568359375,
0.4443359375,
0.47802734375,
0.4443359375,
0.47802734375,
0.458251953125,
0.0882568359375,
0.458251953125
],
[
0.51953125,
0.447509765625,
0.9140625,
0.447509765625,
0.9140625,
0.4609375,
0.51953125,
0.4609375
],
[
0.08660888671875,
0.458251953125,
0.476318359375,
0.458251953125,
0.476318359375,
0.4716796875,
0.08660888671875,
0.4716796875
],
[
0.521484375,
0.46337890625,
0.91015625,
0.46337890625,
0.91015625,
0.473388671875,
0.521484375,
0.473388671875
],
[
0.08660888671875,
0.472900390625,
0.47802734375,
0.472900390625,
0.47802734375,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.521484375,
0.47607421875,
0.9111328125,
0.47607421875,
0.9111328125,
0.4892578125,
0.521484375,
0.4892578125
],
[
0.0882568359375,
0.4873046875,
0.473876953125,
0.4873046875,
0.473876953125,
0.49755859375,
0.0882568359375,
0.49755859375
],
[
0.52294921875,
0.4912109375,
0.91015625,
0.4912109375,
0.91015625,
0.501953125,
0.52294921875,
0.501953125
],
[
0.08416748046875,
0.49755859375,
0.47802734375,
0.498779296875,
0.47802734375,
0.51318359375,
0.08416748046875,
0.51220703125
],
[
0.51953125,
0.50390625,
0.91259765625,
0.50244140625,
0.91259765625,
0.5166015625,
0.51953125,
0.517578125
],
[
0.0849609375,
0.51318359375,
0.47802734375,
0.51318359375,
0.47802734375,
0.5263671875,
0.0849609375,
0.5263671875
],
[
0.521484375,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.521484375,
0.53076171875
],
[
0.08660888671875,
0.5263671875,
0.346435546875,
0.5263671875,
0.346435546875,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.52294921875,
0.53271484375,
0.90673828125,
0.53271484375,
0.90673828125,
0.54345703125,
0.52294921875,
0.54345703125
],
[
0.52392578125,
0.546875,
0.9111328125,
0.546875,
0.9111328125,
0.556640625,
0.52392578125,
0.556640625
],
[
0.1544189453125,
0.5537109375,
0.41015625,
0.55615234375,
0.41015625,
0.57373046875,
0.1544189453125,
0.5712890625
],
[
0.51953125,
0.55810546875,
0.91259765625,
0.5595703125,
0.91259765625,
0.57275390625,
0.51953125,
0.5712890625
],
[
0.51953125,
0.5712890625,
0.9111328125,
0.57275390625,
0.9111328125,
0.5859375,
0.51953125,
0.58544921875
],
[
0.0882568359375,
0.57763671875,
0.476318359375,
0.57763671875,
0.476318359375,
0.59130859375,
0.0882568359375,
0.59130859375
],
[
0.0882568359375,
0.59326171875,
0.4755859375,
0.59326171875,
0.4755859375,
0.60400390625,
0.0882568359375,
0.60400390625
],
[
0.52294921875,
0.58837890625,
0.91259765625,
0.58837890625,
0.91259765625,
0.5986328125,
0.52294921875,
0.5986328125
],
[
0.51953125,
0.5986328125,
0.91259765625,
0.599609375,
0.91259765625,
0.6142578125,
0.51953125,
0.61279296875
],
[
0.08660888671875,
0.60595703125,
0.4755859375,
0.60595703125,
0.4755859375,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.5185546875,
0.61279296875,
0.72119140625,
0.6123046875,
0.72119140625,
0.626953125,
0.5185546875,
0.62744140625
],
[
0.08660888671875,
0.619140625,
0.476318359375,
0.619140625,
0.476318359375,
0.6337890625,
0.08660888671875,
0.6337890625
],
[
0.31201171875,
0.62890625,
0.48046875,
0.626953125,
0.48046875,
0.6484375,
0.31201171875,
0.65087890625
],
[
0.5361328125,
0.62744140625,
0.91015625,
0.62744140625,
0.91015625,
0.6416015625,
0.5361328125,
0.6416015625
],
[
0.08905029296875,
0.63525390625,
0.31201171875,
0.63525390625,
0.31201171875,
0.64599609375,
0.08905029296875,
0.64599609375
],
[
0.08660888671875,
0.646484375,
0.3505859375,
0.646484375,
0.3505859375,
0.66015625,
0.08660888671875,
0.66015625
],
[
0.346435546875,
0.6484375,
0.471435546875,
0.64306640625,
0.47216796875,
0.65771484375,
0.34716796875,
0.66357421875
],
[
0.521484375,
0.64306640625,
0.90869140625,
0.64306640625,
0.90869140625,
0.65380859375,
0.521484375,
0.65380859375
],
[
0.51953125,
0.65380859375,
0.9111328125,
0.65576171875,
0.9111328125,
0.66943359375,
0.51953125,
0.66748046875
],
[
0.0849609375,
0.66015625,
0.47802734375,
0.6591796875,
0.47802734375,
0.673828125,
0.0849609375,
0.6748046875
],
[
0.0849609375,
0.67236328125,
0.47802734375,
0.6748046875,
0.47802734375,
0.68798828125,
0.0849609375,
0.68603515625
],
[
0.51953125,
0.66943359375,
0.91259765625,
0.66943359375,
0.91259765625,
0.68310546875,
0.51953125,
0.68310546875
],
[
0.0849609375,
0.68798828125,
0.47802734375,
0.68798828125,
0.47802734375,
0.7021484375,
0.0849609375,
0.7021484375
],
[
0.51953125,
0.6845703125,
0.91015625,
0.6845703125,
0.91015625,
0.69775390625,
0.51953125,
0.69775390625
],
[
0.521484375,
0.69775390625,
0.91259765625,
0.69775390625,
0.91259765625,
0.7109375,
0.521484375,
0.7109375
],
[
0.0882568359375,
0.7041015625,
0.473876953125,
0.7041015625,
0.473876953125,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.08660888671875,
0.71630859375,
0.47802734375,
0.71630859375,
0.47802734375,
0.72998046875,
0.08660888671875,
0.72998046875
],
[
0.521484375,
0.7109375,
0.91015625,
0.7109375,
0.91015625,
0.724609375,
0.521484375,
0.724609375
],
[
0.08660888671875,
0.72998046875,
0.47802734375,
0.72998046875,
0.47802734375,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.51953125,
0.724609375,
0.91259765625,
0.724609375,
0.91259765625,
0.73779296875,
0.51953125,
0.73779296875
],
[
0.51708984375,
0.736328125,
0.9150390625,
0.73779296875,
0.9150390625,
0.755859375,
0.51708984375,
0.75390625
],
[
0.0882568359375,
0.74560546875,
0.4755859375,
0.74560546875,
0.4755859375,
0.755859375,
0.0882568359375,
0.755859375
],
[
0.51708984375,
0.75,
0.91259765625,
0.75244140625,
0.91259765625,
0.77099609375,
0.51708984375,
0.767578125
],
[
0.0882568359375,
0.75830078125,
0.47802734375,
0.75830078125,
0.47802734375,
0.771484375,
0.0882568359375,
0.771484375
],
[
0.5185546875,
0.76513671875,
0.9150390625,
0.7646484375,
0.9150390625,
0.7822265625,
0.5185546875,
0.78271484375
],
[
0.08660888671875,
0.771484375,
0.47802734375,
0.771484375,
0.47802734375,
0.78515625,
0.08660888671875,
0.78515625
],
[
0.52294921875,
0.7822265625,
0.8935546875,
0.7822265625,
0.8935546875,
0.79541015625,
0.52294921875,
0.79541015625
],
[
0.8935546875,
0.78271484375,
0.9111328125,
0.7861328125,
0.91015625,
0.79345703125,
0.890625,
0.79052734375
],
[
0.521484375,
0.79345703125,
0.62109375,
0.79736328125,
0.619140625,
0.8125,
0.51953125,
0.80810546875
],
[
0.08660888671875,
0.78515625,
0.47802734375,
0.78515625,
0.47802734375,
0.7998046875,
0.08660888671875,
0.7998046875
],
[
0.08660888671875,
0.80029296875,
0.191162109375,
0.80029296875,
0.191162109375,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.58251953125,
0.81494140625,
0.71337890625,
0.81494140625,
0.71337890625,
0.830078125,
0.58251953125,
0.830078125
],
[
0.525390625,
0.82275390625,
0.55810546875,
0.8271484375,
0.5556640625,
0.8388671875,
0.52294921875,
0.83544921875
],
[
0.79833984375,
0.82373046875,
0.90869140625,
0.82275390625,
0.90869140625,
0.8369140625,
0.79833984375,
0.83837890625
],
[
0.60107421875,
0.83349609375,
0.78515625,
0.83203125,
0.78515625,
0.8466796875,
0.60107421875,
0.84765625
],
[
0.0882568359375,
0.8271484375,
0.390625,
0.8271484375,
0.390625,
0.84130859375,
0.0882568359375,
0.84130859375
],
[
0.08660888671875,
0.84765625,
0.4794921875,
0.84765625,
0.4794921875,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.890625,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.890625,
0.861328125
],
[
0.08905029296875,
0.86279296875,
0.4755859375,
0.86279296875,
0.4755859375,
0.87353515625,
0.08905029296875,
0.87353515625
],
[
0.521484375,
0.861328125,
0.91015625,
0.861328125,
0.91015625,
0.875,
0.521484375,
0.875
],
[
0.08660888671875,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.51708984375,
0.87353515625,
0.91259765625,
0.8720703125,
0.91259765625,
0.8896484375,
0.51708984375,
0.8916015625
]
] |
[
"ities. ViLBERT is tested on a variety of downstream tasks",
"aware region embedding, word embedding and three special",
"including VQA, referring expression, and image-to-text re-",
"tokens. The region embedding is defined as:",
"trieval. LXMERT only focuses on a particular problem space",
"ri = W.R; + W,[LayerNorm(W.Ci)|LayerNorm(W.G,)",
"(i.e.,VQAandvisualreasoning)andthegeneralizationabil-",
"(1)",
"ityfurthercompromiseswhenthedatasetsfromthedown-",
"where [J-l indicates the concatenation on the feature dimen-",
" stream tasks are also exploited in the pre-training stage. The",
"sion, LayerNorm represents Layer Normalization. The sec-",
"mostsimilarworktooursisVideoBERT(Sunetal.2019b)",
"ond term mimics the positional embedding in BERT, but",
"which addresses generation-based tasks (e.g., video caption-",
"adding extra region class information,and W,Wp,We,Wg",
"ing) and understanding-based tasks (e.g., action classifica-",
"are the embedding weights (the bias term and the nonlinear-",
"tion). However, it separates the visual encoder and the lan-",
"ityterm are omitted).Note that here we overload the nota",
"guage decoder andperformspre-training onlyontheen-",
"tion of ri e Rd (i e {1, 2,.., N) to also represent class-",
"coder, leaving decoder uninitialized. In contrast, we propose",
"aware region embeddings. In addition, we add segment em-",
"a unified model for both encoding and decoding and fully",
"beddings to r; as in BERT where all the regions share the",
"leveragethebenefitof pre-training.",
"same segment embedding where the values depend on the",
"Image Captioning & VQA. Most of the recent works on",
"objectives (i.e., seq2seq and bidirectional, see the following",
"image captioning are built upon (Anderson et al. 2018),",
"section).",
"where a language model gets clues for sentence generation",
"The word embeddings are similarly defined as in (Devlin",
"through dynamically attending on object regions in the im-",
"et al. 2018), adding up yt with positional embeddings and",
"age extracted from pre-trained object detectors.Follow-up",
"segment embeddings, which is again overloaded as Yt. We",
"worksfurthercapturetherelationshipsamongobjectregions",
"define three special tokens [CLS], [SEP], [STOP], where",
"by using Graph Convolutional Networks (GCNs) (Yao et al.",
"[CLS] indicates the start of the visual input, [SEP] marks",
"2018), incorporating language inductive bias (Yang et al.",
"the boundary between the visual input and the sentence in-",
"2019), or enforcing region grounding between image and",
"put, and [STOP] determines the end of the sentence. The",
"text (Lu et al. 2018; Zhou et al. 2019). VQA is another",
"[MASK] tokens indicate the masked words which will be",
"prevalent research area in vision and language. Since its ini",
"explained in the next section.",
"tial proposal (Antol et al. 2015), there has been a signifi-",
"cant amount of works proposing model architectures to fuse",
"Pre-training Objectives",
"question and image representations (Kim, Jun, and Zhang",
"In the BERT masked language modeling objective, 15% of",
"2018; Anderson et al. 2018; Gao et al. 2019), new datasets",
"the input text tokens are first replaced with either a spe-",
"or models to reduce the dataset bias (Zhang et al. 2016;",
"cial [MASK] token, a random token or the original token,",
"Goyaletal.2017:Agrawaletal.2017)andgroundthean",
"atrandomwithchancesequalto80%,10%,and10%,re-",
" swer in the question (Lewis and Fan 2019). We use our base",
"spectively. Then, at the model output, the hidden state from",
"architecture to perform both image captioning and VQA",
"the last Transformer block is projected to word likelihoods",
"with minor model structure differences.",
"wherethemaskedtokensarepredictedintheformof aclas",
"sificationproblem.Throughthisreconstruction,themodel",
"Vision-Language Pre-training",
"learns the dependencies in the context and forms a language",
"model.Wefollow the same scheme and consider two spe-",
"We denote the input image as I and the associated/target sen-",
"tencedescription(words)asS.Weextractafixed numberN",
"cificobjectives:thebidirectionalobjective(bidirectional)as",
"in BERT and the sequence to sequence objective (seq2seq),",
"of object regions from the image using an off-the-shelf ob-",
"inspired by (Dong et al. 2019).",
"ject detector, denoted as (ri,..., rn ] and the correspond-",
".,RNJ ERdxN,region",
"As shown in Fig. 2 (right), the only difference be-",
"ing regionfeaturesas R =Ri..",
"object labels (probabilities) as C =[Ci",
"....,CNJ ERIXN",
"tweenthe two objectives liein the self-attention mask",
"The mask used for the bidirectional objective allows un-",
"and region geometric information as G [Gi,..., Gn] E",
"Rox N.wheredis the embedding size,I indicates the number",
"restricted message passing between the visual modality",
"of the object classes of the object detector, and o = 5 con-",
"and the language modality while in seq2seq, the to-be-",
"predicted word cannot attend to the words in the fu-",
"sistsoffourvaluesfortopleftandbottomrightcornerco",
"ordinates of the region bounding box (normalized between",
"ture, i.e., it satisfies the auto-regressive property. More for-",
"O and 1) and one value for its relative area (i.e., ratio of the",
"mally, we define the input to the first Transformer block",
"as HO = [r[CLS], ri,..., rN, J[SEP], Yi,..., YT, J[STOP] E",
"boundingboxareatotheimagearea.alsobetweenOandl)",
"RdxU where U = N+T+3, and then the encoding at differ-",
"The words in S are represented as one-hot vectors which are",
"ent levels of Transformer as Hl = Transformer(Hi-1), 1 e",
"further encoded to word embeddings with embedding size e:",
"[1, L]. We further define a self-attention mask as M",
"E",
"RUxU),where",
"Yt e Re where t e [1, 2, ..., T and T indicates the length",
"of the sentence.",
"(0, allow to attend",
"Mik",
"j,k= 1,...,U.",
"0o, prevent from attending",
"Vision-Language Transformer Network",
"Our vision-language Transformer network, which unifies the",
"(2)",
"Transformer encoder and decoder into a singlemodel,is de-",
"For simplicity, we assume a single attention head in the self-",
"picted in Fig. 2 (left). The model input consists of the class-",
"attention module. Then, the seif-attention output on Hl-1"
] |
{
"arxiv_id": "1909.11059",
"language": "en",
"page_idx": 2
}
|
||
2105.05987-en-0
|
2105.05987-en-0.jpg
|
Two Influence Maximization Games on Graphs Made Temporal'
Niclas Boehmer; Vincent Froese, Julia Henkel, Yvonne Lasars, Rolf Niedermeier, and
Malte Renke$n^{\ddagger }$
Technische Universität Berlin, Faculty IV, Algorithmics and Computational Complexity
$\{$niclas.boehmer, vincent.froese$\}\textcircled{\mathrm{O}}$tu-berlin.de,
$\{$henkel, y.lasars$\}\textcircled{\mathrm{Qcampus.tu-berlin.de,}}$
${\mathrm{m.renken@tu-berlin.de}}$
February 22, 2023
$\mathbf{Abstract}$
To address the dynamic nature of real-world networks, we generalize competitive diffusion games and Voronoi games from static to temporal graphs, where edges may appear or disappear over time. This establishes a new direction of studies in the area of graph games, motivated by applications such as influence spreading. As a first step, we investigate the existence of Nash equilibria in 2-player competitive diffusion and Voronoi games on different temporal graph classes. Even when restricting our studies to temporal trees and cycles, this turns out to be a challenging undertaking, revealing significant differences between the two games in the temporal setting. Notably, both games are equivalent on static trees and cycles. Our two main technical results are (algorithmic) proofs for the existence of Nash equilibria in 2-player competitive diffusion and temporal Voronoi games when the edges are restricted not to disappear over time.
$\begin{array}{c}\text{Keywords: Temporal Graph Games, Competitive Diffusion Games, Voronoi Games, Nash}\\\\\text{Keywords: Temporal Graph Games, Competitive Diffusion Games, Voronoi Games, Nash}\end{array}$
Equilibria, Temporal Graph Classes
## 1 Introduction
As graph games help us to reason about a networked world, playing games on graphs is an intensively researched topic since decades. In this work, we focus on competitive games on undirected graphs. Here, some external parties influence a small subset of agents who then spread some information through the network. Typical application scenarios for these occur, for instance, when political parties want to gain influence in a social network or in the context of viral marketing.
Looking at two prominent, somewhat similar representatives, namely competitive diffusion games and Voronoi games, we put forward to model network dynamics more realistically. Specifically, while to the best of our knowledge almost all work on graph games focused on static graphs, we initiate the study of these two games on temporal graphs. Roughly speaking, in a temporal graph, the edge set may evolve over discrete time steps, while the vertex set remains unchanged,
$^{*}$An extended abstract of this article appeared in the proceedings of the Thirtieth International Joint Confer-
ence on Artificial Intelligence (IJCAI 2021), pages 45-51, 2021.
Supported by DFG projects MaMu (NI 369/19) and ComSoc-MPMS (NI 369/22).
${}^{\ddagger}$Supported by DFG project MATE (NI 369/17).
1
|
[
[
0.12939453125,
0.1341552734375,
0.87158203125,
0.1341552734375,
0.87158203125,
0.1519775390625,
0.12939453125,
0.1519775390625
],
[
0.1319580078125,
0.1728515625,
0.86376953125,
0.1728515625,
0.86376953125,
0.18701171875,
0.1319580078125,
0.18701171875
],
[
0.435302734375,
0.191162109375,
0.56982421875,
0.191162109375,
0.56982421875,
0.204833984375,
0.435302734375,
0.204833984375
],
[
0.176513671875,
0.2197265625,
0.82275390625,
0.2197265625,
0.82275390625,
0.2327880859375,
0.176513671875,
0.2327880859375
],
[
0.3310546875,
0.2340087890625,
0.67236328125,
0.2340087890625,
0.67236328125,
0.2476806640625,
0.3310546875,
0.2476806640625
],
[
0.356201171875,
0.248779296875,
0.64892578125,
0.248779296875,
0.64892578125,
0.262451171875,
0.356201171875,
0.262451171875
],
[
0.417724609375,
0.26416015625,
0.5830078125,
0.26416015625,
0.5830078125,
0.27490234375,
0.417724609375,
0.27490234375
],
[
0.02520751953125,
0.283935546875,
0.06219482421875,
0.283935546875,
0.058837890625,
0.3896484375,
0.0218505859375,
0.3896484375
],
[
0.41943359375,
0.300537109375,
0.57958984375,
0.3017578125,
0.57958984375,
0.319580078125,
0.41943359375,
0.318359375
],
[
0.4638671875,
0.366455078125,
0.53857421875,
0.366455078125,
0.53857421875,
0.38134765625,
0.4638671875,
0.38134765625
],
[
0.0277252197265625,
0.380126953125,
0.0579833984375,
0.380126953125,
0.0579833984375,
0.410888671875,
0.0277252197265625,
0.410888671875
],
[
0.18994140625,
0.388427734375,
0.83447265625,
0.388427734375,
0.83447265625,
0.402099609375,
0.18994140625,
0.402099609375
],
[
0.1663818359375,
0.40380859375,
0.8359375,
0.40380859375,
0.8359375,
0.41748046875,
0.1663818359375,
0.41748046875
],
[
0.02520751953125,
0.413330078125,
0.06219482421875,
0.413330078125,
0.06219482421875,
0.493408203125,
0.02520751953125,
0.493408203125
],
[
0.1630859375,
0.416259765625,
0.8359375,
0.41748046875,
0.8359375,
0.431640625,
0.1630859375,
0.43115234375
],
[
0.1646728515625,
0.431640625,
0.83447265625,
0.431640625,
0.83447265625,
0.4453125,
0.1646728515625,
0.4453125
],
[
0.1663818359375,
0.44775390625,
0.8330078125,
0.44775390625,
0.8330078125,
0.457763671875,
0.1663818359375,
0.457763671875
],
[
0.1646728515625,
0.460205078125,
0.8330078125,
0.460205078125,
0.8330078125,
0.473876953125,
0.1646728515625,
0.473876953125
],
[
0.1646728515625,
0.474365234375,
0.83447265625,
0.474365234375,
0.83447265625,
0.488037109375,
0.1646728515625,
0.488037109375
],
[
0.1680908203125,
0.490478515625,
0.83447265625,
0.490478515625,
0.83447265625,
0.50048828125,
0.1680908203125,
0.50048828125
],
[
0.0193328857421875,
0.496337890625,
0.060516357421875,
0.496337890625,
0.0638427734375,
0.7060546875,
0.0218505859375,
0.7060546875
],
[
0.1663818359375,
0.5029296875,
0.83447265625,
0.5029296875,
0.83447265625,
0.5166015625,
0.1663818359375,
0.5166015625
],
[
0.1663818359375,
0.5185546875,
0.8330078125,
0.5185546875,
0.8330078125,
0.529296875,
0.1663818359375,
0.529296875
],
[
0.1646728515625,
0.53076171875,
0.363037109375,
0.53076171875,
0.363037109375,
0.54443359375,
0.1646728515625,
0.54443359375
],
[
0.1646728515625,
0.5517578125,
0.8330078125,
0.5517578125,
0.8330078125,
0.5654296875,
0.1646728515625,
0.5654296875
],
[
0.1646728515625,
0.5654296875,
0.429443359375,
0.5654296875,
0.429443359375,
0.5791015625,
0.1646728515625,
0.5791015625
],
[
0.1260986328125,
0.6025390625,
0.1395263671875,
0.6123046875,
0.127685546875,
0.62060546875,
0.1142578125,
0.61083984375
],
[
0.1605224609375,
0.60498046875,
0.31103515625,
0.60498046875,
0.31103515625,
0.619140625,
0.1605224609375,
0.619140625
],
[
0.12017822265625,
0.6357421875,
0.87890625,
0.6357421875,
0.87890625,
0.65087890625,
0.12017822265625,
0.65087890625
],
[
0.12017822265625,
0.65380859375,
0.87890625,
0.65380859375,
0.87890625,
0.66455078125,
0.12017822265625,
0.66455078125
],
[
0.12017822265625,
0.66748046875,
0.87744140625,
0.66748046875,
0.87744140625,
0.68115234375,
0.12017822265625,
0.68115234375
],
[
0.12017822265625,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69873046875,
0.12017822265625,
0.69873046875
],
[
0.12017822265625,
0.70068359375,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.12017822265625,
0.71435546875
],
[
0.11846923828125,
0.71533203125,
0.272216796875,
0.71630859375,
0.272216796875,
0.73095703125,
0.11846923828125,
0.72998046875
],
[
0.148681640625,
0.732421875,
0.87890625,
0.732421875,
0.87890625,
0.7451171875,
0.148681640625,
0.7451171875
],
[
0.12017822265625,
0.74755859375,
0.880859375,
0.74755859375,
0.880859375,
0.76123046875,
0.12017822265625,
0.76123046875
],
[
0.12017822265625,
0.765625,
0.880859375,
0.765625,
0.880859375,
0.779296875,
0.12017822265625,
0.779296875
],
[
0.12017822265625,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.79443359375,
0.12017822265625,
0.79443359375
],
[
0.11846923828125,
0.79736328125,
0.88232421875,
0.79638671875,
0.88232421875,
0.81103515625,
0.11846923828125,
0.8125
],
[
0.1419677734375,
0.82080078125,
0.88232421875,
0.82080078125,
0.88232421875,
0.83447265625,
0.1419677734375,
0.83447265625
],
[
0.11846923828125,
0.83447265625,
0.546875,
0.83447265625,
0.546875,
0.8486328125,
0.11846923828125,
0.8486328125
],
[
0.1436767578125,
0.84716796875,
0.70751953125,
0.84716796875,
0.70751953125,
0.86083984375,
0.1436767578125,
0.86083984375
],
[
0.1419677734375,
0.86083984375,
0.473876953125,
0.86083984375,
0.473876953125,
0.87451171875,
0.1419677734375,
0.87451171875
],
[
0.49072265625,
0.90283203125,
0.50830078125,
0.90283203125,
0.50830078125,
0.91748046875,
0.49072265625,
0.91748046875
]
] |
[
"Two Influence Maximization Games on Graphs Made Temporal",
"Niclas Boehmer! Vincent Froese. Julia Henkel. Yvonne Lasars, Rolf Niedermeier, ano",
"Malte Renkent",
"Technische Universitat Berlin, Faculty IV, Algorithmics and Computational Complexity",
"[niclas.boehmer, vincent.froese}@tu-berlin.de,",
"[henkel, y.lasars]@campus.tu-berlin.de,",
"m.renken@tu-berlin.de",
"",
"February 22, 2023",
"Abstract",
"2",
"To address the dynamic nature of real-world networks, we generalize competitive diffusion",
"games and Voronoi games from static to temporal graphs, where edges may appear or",
"[D'S]",
"disappear over time. This establishes a new direction of studies in the area of graph games,",
"motivated by applications such as influence spreading. As a first step, we investigate the",
"existenceofNashequilibriain2-playercompetitivediffusionandVoronoigamesondifferent",
"temporal graph classes. Even when restricting our studies to temporal trees and cycles, this",
"turns out to be a challenging undertaking, revealing significant differences between the two",
"gamesinthetemporalsetting.Notably,bothgamesareequivalentonstatictreesand cycles.",
"CA8650S01ZA1XI1",
"Our two main technical results are (algorithmic) proofs for the existence of Nash equilibria",
"in 2-player competitivediffusion and temporal Voronoigames when theedges arerestricted",
"not to disappear over time",
"Keywords: Temporal Graph Games, Competitive Diffusion Games, Voronoi Games, Nash",
"Equilibria, Temporal Graph Classes",
"",
"Introduction",
"As graph games help us to reason about a networked world, playing games on graphs is an",
"intensively researched topic since decades.In this work.we focus on competitive games on",
"undirected graphs. Here, some external parties influence a small subset of agents who then",
"spread some information through the network. Typical application scenarios for these occur,",
"for instance, when political parties want to gain influence in a social network or in the context",
"of viral marketing.",
"Looking at twoprominent, somewhat similar representatives,namely competitive diffusion",
"games and Voronoi games, we put forward to model network dynamics more realistically. Specifi-",
"cally, while to the best of our knowledge almost all work on graph games focused on static graphs,",
"we initiate the study of these two games on temporal graphs. Roughly speaking, in a temporal",
"graph, the edge set may evolve over discrete time steps, while the vertex set remains unchanged,",
"*An extended abstract of this article appeared in the proceedings of the Thirtieth International Joint Confer-",
"ence on Artificial Intelligence (IJCAI 2021), pages 45-51, 2021",
"'Supported by DFG projects MaMu (NI 369/19) and ComSoc-MPMS (NI 369/22).",
"Supported by DFG project MATE (NI 369/17).",
"1"
] |
{
"arxiv_id": "2105.05987",
"language": "en",
"page_idx": 0
}
|
||
2105.05987-en-1
|
2105.05987-en-1.jpg
|
yielding time-ordered graph layers with different edge sets. Moving to the temporal setting has dramatic consequences. For example, while competitive diffusion games and Voronoi games are equivalent on static trees and cycles [26] and we understand well their properties in these simple but important special cases, they are no longer equivalent in the temporal case and their properties are much more challenging to analyze.
Our study takes a first step towards understanding both games on temporal paths, trees, and cycles. It turns out that these backbone structures of graphs already confront us with several technically challenging questions when looking for the existence (and computation) of Nash equilibria—one of the most fundamental game-theoretic concepts. We refer to the next section for all formal definitions and examples of the two (temporal) games. Intuitively speaking, in both games one can think of each player having a color and trying to color as many vertices as possible by her own color; the coloring process starts in a vertex freely chosen by each player and acts through the neighborhood relation of the graph. Herein, the distance of a vertex to the start vertices plays a central role. In both games, a player colors all vertices that are closest to her start vertex. Moreover, while in competitive diffusion games a vertex that is at the same distance to two start vertices of competing players can still get one of the two colors, this is not the case for Voronoi games.
$\textbf{Related work. Competitive diffusion games were introduced by Alon et al. [ 2] . }$Research on competitive diffusion games so far mainly focused on the existence of Nash equilibria on a variety of graph classes for different numbers of players [2,6,17,23,24,25,27]. Also, the (parameterized) computational complexity of deciding the existence of a Nash equilibrium has been studied [14,18].
Voronoi games have been originally studied for a one-dimensional or two-dimensional continuous space [1,4,8,9,15]. There, it is typically assumed that players choose their initial sets of points sequentially and that a player wins the game if a certain fraction of all points is closest to her. Voronoi games on graphs have also been studied on different graph classes for various numbers of players [3,10,16,20,26,28]. Again a focus lies on determining for which graphs a Nash equilibrium exists and how to compute one.
From a broader perspective, analyzing games played on graphs from a game-theoretic perspective is an intensively researched topic; recent examples include Schelling games [7,12], $b$-matching games [19],or network creation games [11].
As mentioned before, we know of basically no work systematically studying graph games in a temporal setting. The only exception we are aware of is in the context of pursuit-evasion games: Erlebach and Spooner [13] studied the pursuit-evasion game of cops and robbers on some specific temporal graphs, namely so-called edge-periodic graphs. Also different from our studies, their focus was on computing winning strategies for the players. Morawietz et al. [22] and Morawietz and Wolf [21] extended this study, also answering an open question of Erlebach and Spooner [13].
Our contributions. We put forward the study of game-theoretic models on temporal graphs We do so by generalizing two well-studied (static) graph games to temporal graphs, namely competitive diffusion games and Voronoi games. For the two resulting temporal graph games we analyze the (constructive) existence of Nash equilibria on different temporal graph classes, focusing on different types of temporal paths, trees, and cycles (see Table 1 for an overview of our results). We observe that, in contrast to the static case where both games are equivalent and a Nash equilibrium is guaranteed to exist [23], the games exhibit far more complex dynamics and a quite different behavior on temporal trees and cycles. Our main results are two involved proofs of guaranteed existence of Nash equilibria, namely in temporal diffusion games and
2
|
[
[
0.12017822265625,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.12017822265625,
0.1009521484375
],
[
0.12017822265625,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.1175537109375,
0.12017822265625,
0.1175537109375
],
[
0.12017822265625,
0.119384765625,
0.87890625,
0.119384765625,
0.87890625,
0.133056640625,
0.12017822265625,
0.133056640625
],
[
0.12017822265625,
0.1341552734375,
0.87890625,
0.1341552734375,
0.87890625,
0.1478271484375,
0.12017822265625,
0.1478271484375
],
[
0.11846923828125,
0.152587890625,
0.50830078125,
0.1519775390625,
0.50830078125,
0.1650390625,
0.11846923828125,
0.166259765625
],
[
0.1470947265625,
0.166259765625,
0.880859375,
0.16748046875,
0.880859375,
0.1817626953125,
0.1470947265625,
0.18115234375
],
[
0.1168212890625,
0.182861328125,
0.87890625,
0.18408203125,
0.87890625,
0.198974609375,
0.1168212890625,
0.19775390625
],
[
0.11846923828125,
0.20068359375,
0.880859375,
0.20068359375,
0.880859375,
0.21435546875,
0.11846923828125,
0.21435546875
],
[
0.12017822265625,
0.2161865234375,
0.87890625,
0.2161865234375,
0.87890625,
0.2298583984375,
0.12017822265625,
0.2298583984375
],
[
0.11846923828125,
0.23095703125,
0.880859375,
0.232177734375,
0.880859375,
0.2464599609375,
0.11846923828125,
0.2452392578125
],
[
0.12017822265625,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.262451171875,
0.12017822265625,
0.262451171875
],
[
0.11846923828125,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.279052734375,
0.11846923828125,
0.279052734375
],
[
0.12017822265625,
0.279052734375,
0.87890625,
0.279052734375,
0.87890625,
0.29345703125,
0.12017822265625,
0.29345703125
],
[
0.11846923828125,
0.29638671875,
0.880859375,
0.29638671875,
0.880859375,
0.311279296875,
0.11846923828125,
0.311279296875
],
[
0.1168212890625,
0.311279296875,
0.880859375,
0.3134765625,
0.880859375,
0.32666015625,
0.1168212890625,
0.32470703125
],
[
0.1168212890625,
0.32666015625,
0.88330078125,
0.327880859375,
0.88330078125,
0.345703125,
0.1168212890625,
0.344482421875
],
[
0.11846923828125,
0.345703125,
0.34375,
0.345703125,
0.34375,
0.359375,
0.11846923828125,
0.359375
],
[
0.12017822265625,
0.37890625,
0.880859375,
0.37890625,
0.880859375,
0.393798828125,
0.12017822265625,
0.393798828125
],
[
0.12017822265625,
0.394287109375,
0.880859375,
0.394287109375,
0.880859375,
0.4091796875,
0.12017822265625,
0.4091796875
],
[
0.12017822265625,
0.410888671875,
0.880859375,
0.410888671875,
0.880859375,
0.42578125,
0.12017822265625,
0.42578125
],
[
0.11846923828125,
0.427490234375,
0.87890625,
0.427490234375,
0.87890625,
0.441162109375,
0.11846923828125,
0.441162109375
],
[
0.12017822265625,
0.443603515625,
0.29150390625,
0.443603515625,
0.29150390625,
0.457275390625,
0.12017822265625,
0.457275390625
],
[
0.148681640625,
0.458984375,
0.88232421875,
0.458984375,
0.88232421875,
0.473876953125,
0.148681640625,
0.473876953125
],
[
0.12017822265625,
0.4755859375,
0.880859375,
0.4755859375,
0.880859375,
0.490478515625,
0.12017822265625,
0.490478515625
],
[
0.11846923828125,
0.4921875,
0.880859375,
0.4921875,
0.880859375,
0.505859375,
0.11846923828125,
0.505859375
],
[
0.11846923828125,
0.50830078125,
0.88232421875,
0.50830078125,
0.88232421875,
0.521484375,
0.11846923828125,
0.521484375
],
[
0.11846923828125,
0.5224609375,
0.880859375,
0.52392578125,
0.880859375,
0.5380859375,
0.11846923828125,
0.53759765625
],
[
0.1168212890625,
0.5380859375,
0.52001953125,
0.5390625,
0.52001953125,
0.55419921875,
0.1168212890625,
0.552734375
],
[
0.1470947265625,
0.5556640625,
0.880859375,
0.5556640625,
0.880859375,
0.57080078125,
0.1470947265625,
0.57080078125
],
[
0.12017822265625,
0.572265625,
0.88232421875,
0.572265625,
0.88232421875,
0.58740234375,
0.12017822265625,
0.58740234375
],
[
0.11846923828125,
0.58740234375,
0.56201171875,
0.58740234375,
0.56201171875,
0.6015625,
0.11846923828125,
0.6015625
],
[
0.1470947265625,
0.60498046875,
0.880859375,
0.60498046875,
0.880859375,
0.619140625,
0.1470947265625,
0.619140625
],
[
0.1168212890625,
0.619140625,
0.88232421875,
0.62060546875,
0.88232421875,
0.63525390625,
0.1168212890625,
0.63427734375
],
[
0.12017822265625,
0.6357421875,
0.880859375,
0.6357421875,
0.880859375,
0.65087890625,
0.12017822265625,
0.65087890625
],
[
0.12017822265625,
0.65283203125,
0.87744140625,
0.65283203125,
0.87744140625,
0.66650390625,
0.12017822265625,
0.66650390625
],
[
0.11846923828125,
0.66845703125,
0.87890625,
0.66845703125,
0.87890625,
0.68310546875,
0.11846923828125,
0.68310546875
],
[
0.12017822265625,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69873046875,
0.12017822265625,
0.69873046875
],
[
0.12017822265625,
0.70068359375,
0.26123046875,
0.70068359375,
0.26123046875,
0.71533203125,
0.12017822265625,
0.71533203125
],
[
0.12017822265625,
0.7353515625,
0.88232421875,
0.7353515625,
0.88232421875,
0.7490234375,
0.12017822265625,
0.7490234375
],
[
0.11846923828125,
0.75,
0.87890625,
0.75048828125,
0.87890625,
0.765625,
0.11846923828125,
0.76416015625
],
[
0.11846923828125,
0.7666015625,
0.87890625,
0.76708984375,
0.87890625,
0.7822265625,
0.11846923828125,
0.78076171875
],
[
0.121826171875,
0.783203125,
0.87646484375,
0.783203125,
0.87646484375,
0.79638671875,
0.121826171875,
0.79638671875
],
[
0.12017822265625,
0.7998046875,
0.880859375,
0.7998046875,
0.880859375,
0.8134765625,
0.12017822265625,
0.8134765625
],
[
0.12017822265625,
0.8154296875,
0.880859375,
0.8154296875,
0.880859375,
0.8291015625,
0.12017822265625,
0.8291015625
],
[
0.12017822265625,
0.83203125,
0.880859375,
0.83203125,
0.880859375,
0.845703125,
0.12017822265625,
0.845703125
],
[
0.12017822265625,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.12017822265625,
0.86083984375
],
[
0.11846923828125,
0.8642578125,
0.880859375,
0.8642578125,
0.880859375,
0.87744140625,
0.11846923828125,
0.87744140625
],
[
0.49072265625,
0.90185546875,
0.51025390625,
0.90185546875,
0.51025390625,
0.91748046875,
0.49072265625,
0.91748046875
]
] |
[
"yielding time-ordered graph layers with different edge sets. Moving to the temporal setting has",
"dramatic consequences. For example, while competitive diffusion games and Voronoi games",
"are equivalent on static trees and cycles [26] and we understand well their properties in these",
"simple but important special cases, they are no longer equivalent in the temporal case and their",
"properties are much more challenging to analyze",
"Our study takes a first step towards understanding both games on temporal paths, trees, and",
"cycles. It turns out that these backbone structures of graphs already confront us with several",
"technically challenging questions when looking for the existence (and computation) of Nash",
"equilibriaone of the most fundamental game-theoretic concepts. We refer to the next section",
"for all formal definitions and examples of the two (temporal) games. Intuitively speaking, in",
"both games one can think of each player having a color and trying to color as many vertices as",
"possible by her own color; the coloring process starts in a vertex freely chosen by each player",
"and acts through the neighborhood relation of the graph. Herein, the distance of a vertex to",
"the start vertices plays a central role. In both games, a player colors all vertices that are closest",
"to her start vertex. Moreover, while in competitive diffusion games a vertex that is at the same",
"distance to two start vertices of competing players can still get one of the two colors, this is not",
"the case for Voronoi games.",
"Related work. Competitive diffusion games were introduced by Alon et al. [2]. Research",
"on competitive diffusion games so far mainly focused on the existence of Nash equilibria on",
"a variety of graph classes for different numbers of players [2, 6, 17, 23, 24, 25, 27]. Also, the",
"(parameterized) computational complexity of deciding the existence of a Nash equilibrium has",
"been studied [14, 18].",
"Voronoi games have been originally studied for a one-dimensional or two-dimensional con-",
"tinuous space [1, 4, 8, 9, 15]. There, it is typically assumed that players choose their initial sets",
"of points sequentially and that a player wins the game if a certain fraction of all points is closest",
"to her. Voronoi games on graphs have also been studied on different graph classes for various",
"numbers of players [3, 10, 16, 20, 26, 28]. Again a focus lies on determining for which graphs a",
"Nash equilibrium exists and how to compute one.",
"From a broader perspective, analyzing games played on graphs from a game-theoretic per-",
"spective is an intensively researched topic; recent examples include Schelling games [7, 12],",
"b-matching games [19], or network creation games [11].",
"As mentioned before, we know of basically no work systematically studying graph games",
"in a temporal setting. The only exception we are aware of is in the context of pursuit-evasion",
"games: Erlebach and Spooner [13] studied the pursuit-evasion game of cops and robbers on",
"some specific temporal graphs, namely so-called edge-periodic graphs. Also different from our",
"studies, their focus was on computing winning strategies for the players. Morawietz et al. [22]",
"and Morawietz and Wolf [21] extended this study, also answering an open question of Erlebach",
"and Spooner [13].",
"Our contributions.We put forward the study of game-theoretic models on temporal graphs",
"We do so by generalizing two well-studied (static) graph games to temporal graphs, namely",
"competitive diffusion games and Voronoi games. For the two resulting temporal graph games",
"we analyze the (constructive) existence of Nash equilibria on different temporal graph classes",
"focusing on different types of temporal paths, trees, and cycles (see Table 1 for an overview of",
"our results). We observe that, in contrast to the static case where both games are equivalent and",
"a Nash equilibrium is guaranteed to exist [23]l, the games exhibit far more complex dynamics",
"and a quite different behavior on temporal trees and cycles. Our main results are two involved",
"proofs of guaranteed existence of Nash equilibria, namely in temporal diffusion games and",
"2"
] |
{
"arxiv_id": "2105.05987",
"language": "en",
"page_idx": 1
}
|
||
2105.05987-en-2
|
2105.05987-en-2.jpg
|
Table 1: Overview of our results.“X” means that a Nash equilibrium is not guaranteed to exist. $“\checkmark” means that a Nash equilibrium always exists. Entries in parentheses are implied by other table$ cells. See Section 2 for formal definitions. Voronoi games and diffusion games on static trees and static cycles are guaranteed to admit a Nash equilibrium.
<table>
<tbody>
<tr>
<th> </th>
<th>Temporally connected</th>
<th>Monotonically growing</th>
<th>Monotonically shrinking</th>
</tr>
<tr>
<td>${\mathrm{Diffusion}}$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Temporal $\textbf{I Paths}$</td>
<td>$\langle V\rangle$</td>
<td>$^{\prime}V$</td>
<td>$X$ (Theorem 4</td>
</tr>
<tr>
<td>Temporal Trees</td>
<td>(Theorem 3</td>
<td>$^{\prime}V$</td>
<td>$(X)$</td>
</tr>
<tr>
<td>$\operatorname{Temporal}$ Cycles</td>
<td>$X$ (Theorem 5)</td>
<td>(Theorem 7)</td>
<td>$X$ (Theorem 6</td>
</tr>
<tr>
<td>$\mathbf{Voronoi}$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Temporal $\textbf{l}$Paths</td>
<td>$X$ (Theorem 32)</td>
<td>$^{\prime}V$</td>
<td>$X$ (Corollary 731</td>
</tr>
<tr>
<td>Temporal Trees</td>
<td>$(X)$</td>
<td>(Theorem 33)</td>
<td>$(X)$</td>
</tr>
<tr>
<td>Temporal Cycles</td>
<td>$X$ (Theorem 32)</td>
<td>(Theorem 55</td>
<td>$X$ (Corollary $\cdot31$</td>
</tr>
</tbody>
</table>
temporal Voronoi games on so-called monotonically growing temporal cycles. On a high level, one conclusion from our work is that temporal Voronoi games seem to exhibit a more intricate behavior than temporal diffusion games.
Organization of the paper. In Section 2, we provide some background on temporal graphs and define temporal competitive diffusion and temporal Voronoi games. In Section 3, we analyze temporal competitive diffusion games. We first consider these games on temporal trees (Section 3.1) and afterwards on temporal cycles (Section 3.2). In Section 3.2.3, we present our technically most involved result for diffusion games, i.e., we show that every temporal competitive diffusion game on a monotonically growing temporal cycle admits a Nash equilibrium. After that, in Section 4, we turn to temporal Voronoi games. Here, the structure is a bit different. We first show negative results on monotonically shrinking paths and cycles (Section 4.1) and then on temporally connected trees and cycles (Section 4.2). Lastly, in Section 4.3, we study monotonically growing trees and in Section 4.4.1, present our most complicated result showing that a Nash equilibrium exists in all temporal Voronoi games on monotonically growing cycles.
## 2 Preliminaries
$\operatorname{For}a\leq b\in\mathbb{N},\operatorname*{let}\left[a,b\right]:=\{a,a+1,\ldots,b\},\left[a,b[:=\{a,a+1,\ldots,b-1\},\operatorname*{and}\left[a,b\right]:=\{a+1,\ldots,b\}.\right.$ Further, let $[n]:=[1,n]$ for $n\in\mathbb{N}.$ For any proposition $P$, the Iverson bracket $[P]$ evaluates to 1 if $P$ is true and to 0 if $P$ is false.
# 2.1 Temporal graphs
A $temporal\textit{ graph G }= ( V, ( E_t) _{t= 1}^\infty )$ consists of a set of vertices $V$ and a sequence of edge sets $(\dot{E}_{t})_{t=1}^{\infty}$ with $E_t\subseteq\binom V2.$ We call the graph $G_t=(V,E_t)$ the $t$-$th\textit{layer of}\tilde{\mathcal{G} }$ and $\mathcal{G}_\downarrow=(V,E_\downarrow)$ $\operatorname { \mathrm{with}} E_{\downarrow }: = \bigcup _{t= 1}^{\infty }E_{t}\operatorname { \mathrm{the~}} underlying\operatorname { graph} \operatorname { \mathrm{of}} \mathcal{G} .$
If there is an integer $\tau(\mathcal{G})$ such that $E_t=E_\tau$ for all $t\geq\tau(\mathcal{G})$, then the minimum such integer $\tau(\mathcal{G})$ is called the lifetime of $\mathcal{G}.$ Intuitively, this means that the graph stops changing after time $\tau(\mathcal{G}).$ Since the temporal games we consider (which will be defined in Section 2.2) always end after some fnite amount of time (depending on the temporal graph), it does not
3
|
[
[
0.11846923828125,
0.0843505859375,
0.88232421875,
0.0843505859375,
0.88232421875,
0.097412109375,
0.11846923828125,
0.097412109375
],
[
0.1168212890625,
0.097412109375,
0.880859375,
0.09857177734375,
0.880859375,
0.11224365234375,
0.1168212890625,
0.11102294921875
],
[
0.11846923828125,
0.1134033203125,
0.88232421875,
0.1134033203125,
0.88232421875,
0.1270751953125,
0.11846923828125,
0.1270751953125
],
[
0.11846923828125,
0.127685546875,
0.49658203125,
0.127685546875,
0.49658203125,
0.141357421875,
0.11846923828125,
0.141357421875
],
[
0.373046875,
0.1591796875,
0.45458984375,
0.1591796875,
0.45458984375,
0.17333984375,
0.373046875,
0.17333984375
],
[
0.50244140625,
0.1591796875,
0.603515625,
0.1591796875,
0.603515625,
0.17333984375,
0.50244140625,
0.17333984375
],
[
0.63623046875,
0.1591796875,
0.736328125,
0.1591796875,
0.736328125,
0.17333984375,
0.63623046875,
0.17333984375
],
[
0.373046875,
0.177001953125,
0.4462890625,
0.177001953125,
0.4462890625,
0.188232421875,
0.373046875,
0.188232421875
],
[
0.5009765625,
0.1776123046875,
0.56396484375,
0.1776123046875,
0.56396484375,
0.1923828125,
0.5009765625,
0.1923828125
],
[
0.63623046875,
0.17333984375,
0.70654296875,
0.177001953125,
0.705078125,
0.1923828125,
0.63525390625,
0.189453125
],
[
0.2403564453125,
0.1995849609375,
0.3134765625,
0.1995849609375,
0.3134765625,
0.21435546875,
0.2403564453125,
0.21435546875
],
[
0.365478515625,
0.2120361328125,
0.3984375,
0.2161865234375,
0.39404296875,
0.2327880859375,
0.361328125,
0.2286376953125
],
[
0.49658203125,
0.215576171875,
0.529296875,
0.215576171875,
0.529296875,
0.2327880859375,
0.49658203125,
0.2327880859375
],
[
0.63623046875,
0.2161865234375,
0.7421875,
0.2161865234375,
0.7421875,
0.23095703125,
0.63623046875,
0.23095703125
],
[
0.6318359375,
0.2327880859375,
0.6572265625,
0.2327880859375,
0.6572265625,
0.2476806640625,
0.6318359375,
0.2476806640625
],
[
0.2420654296875,
0.218505859375,
0.356201171875,
0.218505859375,
0.356201171875,
0.23095703125,
0.2420654296875,
0.23095703125
],
[
0.2403564453125,
0.2340087890625,
0.35205078125,
0.2340087890625,
0.35205078125,
0.248779296875,
0.2403564453125,
0.248779296875
],
[
0.374755859375,
0.2340087890625,
0.486572265625,
0.2340087890625,
0.486572265625,
0.248779296875,
0.374755859375,
0.248779296875
],
[
0.498291015625,
0.2340087890625,
0.525390625,
0.2340087890625,
0.525390625,
0.2476806640625,
0.498291015625,
0.2476806640625
],
[
0.239501953125,
0.25048828125,
0.35546875,
0.25048828125,
0.35546875,
0.265380859375,
0.239501953125,
0.265380859375
],
[
0.371337890625,
0.25048828125,
0.47900390625,
0.25048828125,
0.47900390625,
0.265380859375,
0.371337890625,
0.265380859375
],
[
0.50439453125,
0.25048828125,
0.6142578125,
0.25048828125,
0.6142578125,
0.265380859375,
0.50439453125,
0.265380859375
],
[
0.63525390625,
0.25048828125,
0.7421875,
0.25048828125,
0.7421875,
0.265380859375,
0.63525390625,
0.265380859375
],
[
0.2403564453125,
0.276611328125,
0.3076171875,
0.276611328125,
0.3076171875,
0.29150390625,
0.2403564453125,
0.29150390625
],
[
0.374755859375,
0.29345703125,
0.4931640625,
0.29345703125,
0.4931640625,
0.30810546875,
0.374755859375,
0.30810546875
],
[
0.63623046875,
0.29345703125,
0.75390625,
0.29345703125,
0.75390625,
0.30810546875,
0.63623046875,
0.30810546875
],
[
0.489013671875,
0.295654296875,
0.525390625,
0.295654296875,
0.525390625,
0.30712890625,
0.489013671875,
0.30712890625
],
[
0.2420654296875,
0.295654296875,
0.353759765625,
0.295654296875,
0.353759765625,
0.30712890625,
0.2420654296875,
0.30712890625
],
[
0.36962890625,
0.311279296875,
0.392333984375,
0.311279296875,
0.392333984375,
0.32470703125,
0.36962890625,
0.32470703125
],
[
0.50439453125,
0.308837890625,
0.65869140625,
0.308837890625,
0.65869140625,
0.32666015625,
0.50439453125,
0.32666015625
],
[
0.371337890625,
0.32666015625,
0.489013671875,
0.327880859375,
0.489013671875,
0.342529296875,
0.371337890625,
0.341552734375
],
[
0.2420654296875,
0.3134765625,
0.353759765625,
0.3134765625,
0.353759765625,
0.32470703125,
0.2420654296875,
0.32470703125
],
[
0.2403564453125,
0.328857421875,
0.359619140625,
0.328857421875,
0.359619140625,
0.342529296875,
0.2403564453125,
0.342529296875
],
[
0.50244140625,
0.328857421875,
0.63525390625,
0.328857421875,
0.63525390625,
0.342529296875,
0.50244140625,
0.342529296875
],
[
0.6318359375,
0.32666015625,
0.75537109375,
0.327880859375,
0.75537109375,
0.342529296875,
0.6318359375,
0.341552734375
],
[
0.12017822265625,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.385986328125,
0.12017822265625,
0.385986328125
],
[
0.12017822265625,
0.3896484375,
0.880859375,
0.3896484375,
0.880859375,
0.402587890625,
0.12017822265625,
0.402587890625
],
[
0.11846923828125,
0.405029296875,
0.4462890625,
0.405029296875,
0.4462890625,
0.419189453125,
0.11846923828125,
0.419189453125
],
[
0.12017822265625,
0.439453125,
0.880859375,
0.439453125,
0.880859375,
0.45361328125,
0.12017822265625,
0.45361328125
],
[
0.12017822265625,
0.4560546875,
0.88232421875,
0.4560546875,
0.88232421875,
0.4697265625,
0.12017822265625,
0.4697265625
],
[
0.12017822265625,
0.471435546875,
0.88232421875,
0.471435546875,
0.88232421875,
0.486328125,
0.12017822265625,
0.486328125
],
[
0.1168212890625,
0.486328125,
0.88232421875,
0.48681640625,
0.88232421875,
0.501953125,
0.1168212890625,
0.50048828125
],
[
0.12017822265625,
0.50439453125,
0.88232421875,
0.50439453125,
0.88232421875,
0.51708984375,
0.12017822265625,
0.51708984375
],
[
0.12017822265625,
0.51953125,
0.88232421875,
0.51953125,
0.88232421875,
0.533203125,
0.12017822265625,
0.533203125
],
[
0.12017822265625,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.12017822265625,
0.5498046875
],
[
0.12017822265625,
0.55029296875,
0.87890625,
0.55029296875,
0.87890625,
0.56396484375,
0.12017822265625,
0.56396484375
],
[
0.12017822265625,
0.568359375,
0.87744140625,
0.568359375,
0.87744140625,
0.58203125,
0.12017822265625,
0.58203125
],
[
0.11846923828125,
0.5830078125,
0.880859375,
0.58447265625,
0.880859375,
0.5986328125,
0.11846923828125,
0.59716796875
],
[
0.11846923828125,
0.60107421875,
0.88232421875,
0.60107421875,
0.88232421875,
0.6142578125,
0.11846923828125,
0.6142578125
],
[
0.11846923828125,
0.63818359375,
0.1395263671875,
0.63818359375,
0.1395263671875,
0.65283203125,
0.11846923828125,
0.65283203125
],
[
0.1546630859375,
0.6357421875,
0.3193359375,
0.63720703125,
0.3193359375,
0.65478515625,
0.1546630859375,
0.65380859375
],
[
0.1168212890625,
0.66845703125,
0.59912109375,
0.66943359375,
0.59912109375,
0.68408203125,
0.1168212890625,
0.68310546875
],
[
0.59765625,
0.66943359375,
0.880859375,
0.66943359375,
0.880859375,
0.68408203125,
0.59765625,
0.68408203125
],
[
0.11846923828125,
0.68603515625,
0.880859375,
0.68603515625,
0.880859375,
0.70068359375,
0.11846923828125,
0.70068359375
],
[
0.11846923828125,
0.70068359375,
0.421875,
0.70068359375,
0.421875,
0.71435546875,
0.11846923828125,
0.71435546875
],
[
0.12017822265625,
0.73681640625,
0.1571044921875,
0.73681640625,
0.1571044921875,
0.75,
0.12017822265625,
0.75
],
[
0.1646728515625,
0.7353515625,
0.342041015625,
0.73388671875,
0.342041015625,
0.751953125,
0.1646728515625,
0.7529296875
],
[
0.1588134765625,
0.7412109375,
0.1663818359375,
0.7412109375,
0.1663818359375,
0.7451171875,
0.1588134765625,
0.7451171875
],
[
0.115966796875,
0.7587890625,
0.880859375,
0.75830078125,
0.880859375,
0.775390625,
0.115966796875,
0.77685546875
],
[
0.1168212890625,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7919921875,
0.1168212890625,
0.7919921875
],
[
0.11846923828125,
0.79345703125,
0.492431640625,
0.79345703125,
0.492431640625,
0.80810546875,
0.11846923828125,
0.80810546875
],
[
0.1470947265625,
0.80859375,
0.880859375,
0.80859375,
0.880859375,
0.82373046875,
0.1470947265625,
0.82373046875
],
[
0.11846923828125,
0.82470703125,
0.880859375,
0.826171875,
0.880859375,
0.84033203125,
0.11846923828125,
0.8388671875
],
[
0.11846923828125,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.11846923828125,
0.85498046875
],
[
0.12017822265625,
0.85791015625,
0.87890625,
0.85791015625,
0.87890625,
0.87158203125,
0.12017822265625,
0.87158203125
],
[
0.49072265625,
0.90185546875,
0.51025390625,
0.90185546875,
0.51025390625,
0.91748046875,
0.49072265625,
0.91748046875
]
] |
[
"Table l: Overview of our results. \"x\" means that a Nash equilibrium is not guaranteed to exist.",
"\"\" means that a Nash equilibrium always exists. Entries in parentheses are implied by other table",
"cells. See Section 2 for formal definitions. Voronoi games and diffusion games on static trees and static",
"cycles are guaranteed to admit a Nash equilibrium.",
"Temporally",
"Monotonically",
"Monotonically",
"connected",
"growing",
"shrinking",
"Diffusion",
"()",
"()",
"X (Theorem 4)",
"(X)",
"Temporal Paths",
"Temporal Trees",
"/ (Theorem 3)",
"()",
"Temporal Cycles",
"X (Theorem 5)",
"/ (Theorem 7)",
"X (Theorem 6)",
"Voronoi",
"X (Theorem 32)",
"X (Corollary 31)",
")",
"Temporal Paths",
"(X)",
"V (Theorem 33) (X)",
"X (Theorem 32)",
"Temporal Trees",
"Temporal Cycles",
"(Theorem 55)",
"× (Corollary 31)",
"temporal Voronoi games on so-called monotonically growing temporal cycles. On a high level.",
"one conclusion from our work is that temporal Voronoi games seem to exhibit a more intricate",
"behavior than temporal diffusion games.",
"Organization of the paper. In Section 2, we provide some background on temporal graphs",
"and define temporal competitive diffusion and temporal Voronoi games. In Section 3, we an-",
"alyze temporal competitive diffusion games. We first consider these games on temporal trees",
"(Section 3.1) and afterwards on temporal cycles (Section 3.2). In Section 3.2.3, we present our",
"technically most involved result for diffusion games, i.e., we show that every temporal competi-",
"tive diffusion game on a monotonically growing temporal cycle admits a Nash equilibrium. After",
"that, in Section 4, we turn to temporal Voronoi games. Here, the structure is a bit different",
"We first show negative results on monotonically shrinking paths and cycles (Section 4.1) and",
"then on temporally connected trees and cycles (Section 4.2). Lastly, in Section 4.3, we study",
"monotonically growing trees and in Section 4.4.1, present our most complicated result showing",
"that a Nash equilibrium exists in all temporal Voronoi games on monotonically growing cycles.",
"2",
"Preliminaries",
"For a < b e N, let [a,b] := {a,a+1,...,b], [a,b[;= [a,a+1, .",
"..,b-1], and ]a,b] := [a+1,...,b]",
"Further, let [n] := [1,n] for n E N. For any proposition P, the Iverson bracket [P] evaluates",
"to 1 if P is true and to 0 if P is false.",
"2.1",
" Temporal graphs",
"",
"A temporal graph g = (V,(Et)1) consists of a set of vertices V and a sequence of edge",
"sets (Et)-i with Et (). We call the graph Gt = (V, Et) the t-th layer of g and Gj = (V, E))",
"with Ej := Ut-i Et the underlying graph of g.",
"If there is an integer (9) such that Et = E, for all t ≥ (G), then the minimum such",
"integer r() is called the lifetime of g. Intuitively, this means that the graph stops changing",
"after time T(g). Since the temporal games we consider (which will be defined in Section 2.2)",
"always end after some finite amount of time (depending on the temporal graph), it does not",
"3"
] |
{
"arxiv_id": "2105.05987",
"language": "en",
"page_idx": 2
}
|
||
2105.05987-en-3
|
2105.05987-en-3.jpg
|
constitute a loss of generality to assume that all our temporal graphs have finite lifetime $\tau.$ We
will then simply omit specifying $E_i$ for $i>\tau$ (implying that $E_i=E_\tau$ for all $i>\tau).$
A temporalforest (resp. $tree,path,cycle)$ is a temporal graph whose underlying graph is a forest (resp. path, cycle). In the case of paths, we will use the convention that the path runs from left to right and that its vertices are denoted $1,\ldots,n$ in that order. Similarly, we will denote the vertices of a cycle by either $1,\ldots,n$ or $0,\ldots,(n-1)$ in counterclockwise order.
For the defnition of temporal Voronoi games, we require a notion of temporal distance between two vertices. In a temporal graph $\mathcal{G}=(V,(E_{t})_{t=1}^{\infty})$, we define a temporalwalk from a vertex $v_0$ to a vertex $v_d$ as a sequence of tuples $(\{v_0,v_1\},t_1),(\{v_1,v_2\},t_2),\ldots,(\{v_{d-1},v_d\},t_d)$ such that the following properties hold:
$\bullet$ $t_{i}< t_{i+ 1}$ for all $i\in[d-1]$,
$\bullet$ $\{ v_{i- 1}, v_{i}\} \in E_{t_{i}}$ for all $i\in[d].$
We refer to $t_d$ as the $arrival\textit{time of the temporal walk. Moreover, we call a temporal walk}$ from $v_0$ to $v_d\textit{foremost if there is no temporal walk from }v_0$ to $v_d$ with a smaller arrival time. We now define the temporal distance td$(u,v)$ from $u$ to $v$ as the arrival time of a foremost walk$^1$ from $u$ to $v.$ If there is no such walk, then set td$(u,v)=\infty.$ Notably, in contrast to the static case, temporal distances are not necessarily symmetric, that is, td$(u,v)\neq$td$(v,u)$ is possible. By convention, we set td$(v,v)=0$ for any vertex $v.$ For two vertices $u$ and $v$, we say that $u\textit{reaches}$ $v$ until step $\ell$ if td$(u,v)\leq\ell$ and that $u\textit{reaches }vin$ step $( or$ $at$ $time)$ $\ell$ if td$(u,v)=\ell.$ The set of all vertices reachable from $v$ until time $\ell$ in $\mathcal{G}$ is denoted by $\Omega_\mathcal{G}^\ell(v):=\{w\mid$td$(v,w)\leq\ell\}$ where the subscript $\mathcal{G}$ is omitted if it is clear from context.
A temporal graph $\mathcal{G}$ is $temporally\textit{connected if td}( u, v) < \infty$ for all vertex pairs $u,v.$ Further, we call $\mathcal{G} \textit{monotonically growing if no edge disappears over time, i. e. , }E_{t}\subseteq E_{t+ 1}$ for all $t.$ Symmetrically, $\mathcal{G}$ is $monotonically\textit{ shrinking if edges do not appear over time, i. e. , }E_{t+ 1}\subseteq E_t$ for all $t.$
2.2 Games on temporal graphs
We focus on games with two players. Nevertheless, to highlight the nature of our definitions, we directly introduce both games for an arbitrary number of players. Since the games are somewhat similar, we start by making some general definitions for both temporal games before describing the specifics. For a temporal graph $\mathcal{G}=(V,(E_t)_{t=1}^\infty)$ and a number $k\in\mathbb{N}$ of players, Diff$(\mathcal{G},k)$ (Vor$(\mathcal{G},k))$ denotes the $k$-player $temporal\textit{diffusion game }( temporal\textit{ Voronoi game})$ on the temporal graph $\mathcal{G}$, where each player has a distinct color in $[k].$ Moreover, we use the color 0 to which we refer as $gray.$ The $strategy\textit{space of each player }i\in [ k]$ is the vertex set $V$, that is, each player $i$ selects a single vertex $p_i\in V$, which is then immediately colored by her color $i.$ If two players pick the same vertex, then it is colored gray. A $strategy\textit{profle of the}$ game is a tuple $(p_1,\ldots,p_k)\in V^k$ containing the initially chosen vertex of each player. We also use the term position to refer to the vertex $p_i$ chosen by player $i.$
Now, for both games the strategy profile $(p_1,\ldots,p_k)$ determines an initial (partial) coloring of the graph at time $t=0.$ We use $U_i^t(p_1,\ldots,p_k)$ to denote the set of vertices having color $i$ at time $t.$ Then $U_i^0(p_1,\ldots,p_k)=\{p_i\}$, unless some other player chose the same vertex, in which case $U_i^0(p_1,\ldots,p_k)=\emptyset.$ In each of the two games, the coloring at time $t$ then determines the coloring at time $t+1$ by the specific rules described below.
Let further $u_i^t(p_1,\ldots,p_k):=|U_i^t(p_1,\ldots,p_k)|$ be the number of vertices having color $i$ at time $t.$ As vertices will never change their color once they have been colored, the sequence
$^{1}$We consider foremost walks since earliest arrival seems more natural than other concepts such as fastest or shortest [5] in the context of influence spreading. Moreover, we require strict inequality between $t_i$ and $t_i+1$ since this follows the static case and the diffusion process more closely.
4
|
[
[
0.12017822265625,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.12017822265625,
0.1009521484375
],
[
0.12017822265625,
0.10394287109375,
0.7900390625,
0.10394287109375,
0.7900390625,
0.1175537109375,
0.12017822265625,
0.1175537109375
],
[
0.148681640625,
0.119384765625,
0.880859375,
0.119384765625,
0.880859375,
0.1341552734375,
0.148681640625,
0.1341552734375
],
[
0.11846923828125,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.149658203125,
0.11846923828125,
0.149658203125
],
[
0.12017822265625,
0.1519775390625,
0.55810546875,
0.1519775390625,
0.55810546875,
0.1650390625,
0.12017822265625,
0.1650390625
],
[
0.42529296875,
0.1561279296875,
0.43701171875,
0.1561279296875,
0.43701171875,
0.1602783203125,
0.42529296875,
0.1602783203125
],
[
0.5830078125,
0.15087890625,
0.880859375,
0.15087890625,
0.880859375,
0.1644287109375,
0.5830078125,
0.1644287109375
],
[
0.11846923828125,
0.16748046875,
0.4404296875,
0.16748046875,
0.4404296875,
0.18115234375,
0.11846923828125,
0.18115234375
],
[
0.49072265625,
0.171630859375,
0.54443359375,
0.171630859375,
0.54443359375,
0.1787109375,
0.49072265625,
0.1787109375
],
[
0.61083984375,
0.168701171875,
0.84619140625,
0.168701171875,
0.84619140625,
0.1817626953125,
0.61083984375,
0.1817626953125
],
[
0.572265625,
0.174560546875,
0.591796875,
0.174560546875,
0.591796875,
0.1776123046875,
0.572265625,
0.1776123046875
],
[
0.1470947265625,
0.18408203125,
0.462158203125,
0.18408203125,
0.462158203125,
0.19775390625,
0.1470947265625,
0.19775390625
],
[
0.57080078125,
0.18408203125,
0.87890625,
0.18408203125,
0.87890625,
0.19775390625,
0.57080078125,
0.19775390625
],
[
0.1168212890625,
0.198974609375,
0.498291015625,
0.20068359375,
0.498291015625,
0.21435546875,
0.1168212890625,
0.2120361328125
],
[
0.50830078125,
0.20068359375,
0.87890625,
0.20068359375,
0.87890625,
0.21435546875,
0.50830078125,
0.21435546875
],
[
0.1168212890625,
0.215576171875,
0.87890625,
0.2161865234375,
0.87890625,
0.23095703125,
0.1168212890625,
0.2298583984375
],
[
0.11846923828125,
0.2327880859375,
0.438720703125,
0.2327880859375,
0.438720703125,
0.2464599609375,
0.11846923828125,
0.2464599609375
],
[
0.148681640625,
0.2476806640625,
0.37890625,
0.2476806640625,
0.37890625,
0.262451171875,
0.148681640625,
0.262451171875
],
[
0.148681640625,
0.26416015625,
0.39990234375,
0.26318359375,
0.39990234375,
0.27783203125,
0.148681640625,
0.279052734375
],
[
0.1470947265625,
0.27978515625,
0.88232421875,
0.27978515625,
0.88232421875,
0.29443359375,
0.1470947265625,
0.29443359375
],
[
0.12017822265625,
0.29638671875,
0.88232421875,
0.29638671875,
0.88232421875,
0.311279296875,
0.12017822265625,
0.311279296875
],
[
0.11846923828125,
0.312255859375,
0.88232421875,
0.312255859375,
0.88232421875,
0.32666015625,
0.11846923828125,
0.32666015625
],
[
0.12017822265625,
0.328857421875,
0.88232421875,
0.328857421875,
0.88232421875,
0.34326171875,
0.12017822265625,
0.34326171875
],
[
0.12017822265625,
0.344482421875,
0.87890625,
0.344482421875,
0.87890625,
0.358154296875,
0.12017822265625,
0.358154296875
],
[
0.12017822265625,
0.361083984375,
0.880859375,
0.361083984375,
0.880859375,
0.374755859375,
0.12017822265625,
0.374755859375
],
[
0.12017822265625,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.391357421875,
0.12017822265625,
0.391357421875
],
[
0.12017822265625,
0.393798828125,
0.87890625,
0.393798828125,
0.87890625,
0.407958984375,
0.12017822265625,
0.407958984375
],
[
0.11846923828125,
0.407958984375,
0.59326171875,
0.4091796875,
0.59326171875,
0.42333984375,
0.11846923828125,
0.4228515625
],
[
0.148681640625,
0.424560546875,
0.88232421875,
0.424560546875,
0.88232421875,
0.439453125,
0.148681640625,
0.439453125
],
[
0.12017822265625,
0.441162109375,
0.88330078125,
0.441162109375,
0.88330078125,
0.4560546875,
0.12017822265625,
0.4560546875
],
[
0.11846923828125,
0.4560546875,
0.880859375,
0.457275390625,
0.880859375,
0.471435546875,
0.11846923828125,
0.47021484375
],
[
0.11846923828125,
0.47265625,
0.18994140625,
0.47265625,
0.18994140625,
0.48681640625,
0.11846923828125,
0.48681640625
],
[
0.12017822265625,
0.50732421875,
0.4404296875,
0.50732421875,
0.4404296875,
0.521484375,
0.12017822265625,
0.521484375
],
[
0.12017822265625,
0.5322265625,
0.88232421875,
0.5322265625,
0.88232421875,
0.54638671875,
0.12017822265625,
0.54638671875
],
[
0.12017822265625,
0.548828125,
0.880859375,
0.548828125,
0.880859375,
0.5625,
0.12017822265625,
0.5625
],
[
0.11846923828125,
0.5634765625,
0.880859375,
0.56396484375,
0.880859375,
0.5791015625,
0.11846923828125,
0.57763671875
],
[
0.12017822265625,
0.58056640625,
0.880859375,
0.58056640625,
0.880859375,
0.595703125,
0.12017822265625,
0.595703125
],
[
0.12017822265625,
0.5966796875,
0.87890625,
0.5966796875,
0.87890625,
0.60986328125,
0.12017822265625,
0.60986328125
],
[
0.11846923828125,
0.61328125,
0.880859375,
0.61328125,
0.880859375,
0.62744140625,
0.11846923828125,
0.62744140625
],
[
0.11846923828125,
0.6298828125,
0.88232421875,
0.6298828125,
0.88232421875,
0.6435546875,
0.11846923828125,
0.6435546875
],
[
0.11846923828125,
0.6455078125,
0.88232421875,
0.6455078125,
0.88232421875,
0.66015625,
0.11846923828125,
0.66015625
],
[
0.11846923828125,
0.6611328125,
0.880859375,
0.6611328125,
0.880859375,
0.67578125,
0.11846923828125,
0.67578125
],
[
0.1168212890625,
0.67578125,
0.88330078125,
0.6748046875,
0.88330078125,
0.6923828125,
0.1168212890625,
0.693359375
],
[
0.12017822265625,
0.693359375,
0.6455078125,
0.693359375,
0.6455078125,
0.70703125,
0.12017822265625,
0.70703125
],
[
0.1470947265625,
0.708984375,
0.880859375,
0.708984375,
0.880859375,
0.7236328125,
0.1470947265625,
0.7236328125
],
[
0.11846923828125,
0.724609375,
0.88232421875,
0.724609375,
0.88232421875,
0.7421875,
0.11846923828125,
0.7421875
],
[
0.11846923828125,
0.74072265625,
0.88232421875,
0.74072265625,
0.88232421875,
0.75830078125,
0.11846923828125,
0.75830078125
],
[
0.11846923828125,
0.75732421875,
0.88232421875,
0.75732421875,
0.88232421875,
0.77490234375,
0.11846923828125,
0.77490234375
],
[
0.12017822265625,
0.77490234375,
0.5966796875,
0.77490234375,
0.5966796875,
0.7880859375,
0.12017822265625,
0.7880859375
],
[
0.1419677734375,
0.7880859375,
0.88232421875,
0.78759765625,
0.88232421875,
0.8046875,
0.1419677734375,
0.8056640625
],
[
0.1168212890625,
0.8046875,
0.880859375,
0.8056640625,
0.880859375,
0.82080078125,
0.1168212890625,
0.8193359375
],
[
0.1412353515625,
0.830078125,
0.88232421875,
0.830078125,
0.88232421875,
0.84326171875,
0.1412353515625,
0.84326171875
],
[
0.11846923828125,
0.84326171875,
0.880859375,
0.84326171875,
0.880859375,
0.85693359375,
0.11846923828125,
0.85693359375
],
[
0.1168212890625,
0.85498046875,
0.56494140625,
0.85546875,
0.56494140625,
0.87060546875,
0.1168212890625,
0.869140625
],
[
0.49072265625,
0.90185546875,
0.51171875,
0.90185546875,
0.51171875,
0.91748046875,
0.49072265625,
0.91748046875
]
] |
[
"constitute a loss of generality to assume that all our temporal graphs have finite lifetime T. We",
"will then simply omit specifying E, for i > T (implying that E, = E, for all i > T).",
"A temporal forest (resp. tree, path, cycle) is a temporal graph whose underlying graph is a",
"forest (resp. path, cycle). In the case of paths, we will use the convention that the path runs",
"from left to right and that its vertices are denoted 1.",
"",
", n in that order. Similarly, we will",
"denote the vertices of a cycle by either",
"Qor.",
"1)in counterclockwise order.",
"",
"For the definition of temporal Vorono",
"require a notion of temporal distance",
"between two vertices.In atemporalgraphg",
"(V,(Et)-), we define a temporal walk from",
"a vertex vo to a vertex Va as a sequence of tuples ({vo, vi),ti), ({vi,v2),t2),...,({vd-1, vd),ta)",
"such that the following properties hold:",
"· ti <ti+i for all ie [d - 1],",
". [vi-1, vi] E Et, for all i e [d].",
"We refer to ta as the arrival time of the temporal walk. Moreover, we call a temporal walk",
"from vo to a foremost if there is no temporal walk from vo to va with a smaller arrival time. We",
"now define the temporal distance td(u, ) from u to v as the arrival time of a foremost walki from",
"u to u. If there is no such walk, then set td(u, v) = oo. Notably, in contrast to the static case,",
"temporal distances are not necessarily symmetric, that is, td(u, u) td(, u) is possible. By",
"convention, we set td(, v) = o for any vertex v. For two vertices u and y, we say that u reaches",
"w until step l if td(u, v) ≤ l and that u reaches w in step (or at time) l if td(u, v) = l. The",
"set of all vertices reachable from u until time l in G is denoted by 2(u) := [w I td(u, w) ≤ l)",
"where the subscript g is omitted if it is clear from context.",
"A temporal graph g is temporally connected if td(u, u) < o for all vertex pairs u, u. Further,",
"we call g monotonically growing if no edge disappears over time, i.e., Et Et+1 for all t.",
"Symmetrically, g is monotonically shrinking if edges do not appear over time, i.e., Et+i C Et",
"for all t.",
"2.2 Games on temporal graphs",
"We focus on games with two players. Nevertheless, to highlight the nature of our definitions,",
"we directly introduce both games for an arbitrary number of players. Since the games are",
"somewhat similar, we start by making some general definitions for both temporal games before",
"describing the specifics: For a temporal graph G = (V,(Et)-i) and a number k E N of players,",
"Diff(G, k) (Vor(G,k)) denotes the k-player temporal diffusion game (temporal Voronoi game)",
"on the temporal graph G, where each player has a distinct color in [k]. Moreover, we use the",
"color O to which we refer as gray. The strategy space of each player i e [k] is the vertex set V,",
"that is, each player i selects a single vertex pi e V, which is then immediately colored by her",
"color i. If two players pick the same vertex, then it is colored gray. A strategy profile of the",
"game is a tuple (pi, .. , pk) e Vk containing the initially chosen vertex of each player. We also",
"use the term position to refer to the vertex pi chosen by player i.",
"Now, for both games the strategy profile (pi, .. . , Pk) determines an initial (partial) coloring",
"of the graph at time t = 0. We use Ut(pi,..., Pk) to denote the set of vertices having color i at",
"time t. Then U?(pi, ..., Pk) = (pi), unless some other player chose the same vertex, in which",
"case U?(pi,..., pk) = 0. In each of the two games, the coloring at time t then determines the",
"coloring at time t+ 1 by the specific rules described below.",
"Let further ut(pi,.., pk) :- JUt(pi,..., pk)I be the number of vertices having color i at",
"time t. As vertices will never change their color once they have been colored, the sequence",
"IWe consider foremost walks since earliest arrival seems more natural than other concepts such as fastest or",
"shortest [5] in the context of influence spreading. Moreover, we require strict inequality between t, and ti+1 since",
"this follows the static case and the diffusion process more closely.",
"4"
] |
{
"arxiv_id": "2105.05987",
"language": "en",
"page_idx": 3
}
|
||
2105.05987-en-4
|
2105.05987-en-4.jpg
|
0000000
500000
6
1 $p_{1}$ $p_{2}$
$\operatorname{vertex}$
0000000500000
6
1 $p_1$ $p_2$
vertex
(b) Temporal Voronoi game.
( a) Temporal diffusion game.
Figure 1: Example of our two games on the temporal path $\mathcal{G}=([6],E_1,\ldots,E_5)$ with $E_t=\{\{t,t+1\}\}$ for $t\in[5]$,where player 1 selects vertex $p_1=2$ and player 2 selects vertex $p_2=3.$ In the diffusion game, the color of a player cannot“pass” a vertex colored by the other player. Thus, the two players split the vertices, that is, player 1 colors the vertices in [1,2], while player 2 colors the vertices in [3,6]. In contrast to this, in the Voronoi game, player 1 is able to“catch up” with player 2: They both reach the vertices 4,5, and 6 at the same time. The displayed strategy profile is a Nash equilibrium in Vor$(\mathcal{G},2)$ but not in Diff$(\mathcal{G},2).$
$(U_{i}^{t}(p_{1},\ldots,p_{k}))_{t=0}^{\infty}$ must eventually become constant. We denote its limit by $U_i(p_1,\ldots,p_k).$ The payoff or outcome for player $i$ is then $u_i(p_1,\ldots,p_k):=|U_i(p_1,\ldots,p_k)|$, and the players aim to maximize this value. Consequently, we say that a player $i\in[k]$ plays a $best\textit{re- }$ sponse to the other players in the strategy profile $(p_1,\ldots,p_k)$ if for all vertices $p^\prime\in V$ it holds that $u_i(p_1,\ldots,p_{i-1},p^{\prime},p_{i+1},\ldots,p_k)\leq u_i(p_1,\ldots,p_k).$ A strategy profile $(p_1,\ldots,p_k)$ is a Nash equilibrium if every player $i\in[k]$ plays a best response to the other players. Any strategy profile of the form $(p,p,\ldots,p)$ in which all players choose the same vertex is called $trivial.$
It remains to specify how the strategy profile $(p_1,\ldots,p_k)$ determines the coloring of the
vertices $V$ in the two games.
Temporal diffusion games. In a $temporal\textit{diffusion game, the temporal graph }\mathcal{G}$ is colored by the following propagation process over time. We call a vertex uncolored if no color has been assigned to it (so far). In step $t$,we consider the layer $G_t.$ We color a so far uncolored vertex $v$ with color $i\in[k]$ if $v$ has at least one neighbor in $G_t$ that is colored with color $i\in[k]$ and no neighbor in $G_t$ that is colored with any other color $j\in[k]\setminus\{i\}.$ Every uncolored vertex with at least two neighbors in $G_t$ colored by two different colors $i,j\in[k]$ is colored gray.
Temporal Voronoi games. In a $temporal\textit{ Voronoi game, at time }t$ a vertex $v$ is colored with color $i\in[k]$ if $p_i$ is the only position of a player that can reach $v$ until time $t$ by a temporal walk. (If an uncolored vertex is simultaneously reached by two or more players, then it is colored gray.) In effect, $v\in U_i(p_1,\ldots,p_k)$ if and only if td$(p_i,v)<$td$(p_j,v)$ holds for all $j\neq i.$
Note that we defined temporal diffusion games and temporal Voronoi games such that both temporal games played on a temporal graph $\mathcal{G}$ with lifetime 1 are equivalent to the (nontemporal) game played on the static graph $\mathcal{G}_\downarrow.$ An example of a temporal diffusion game and a temporal Voronoi game is shown in Figure 1. Notably, in the displayed temporal diffusion game each player colors a superset of the vertices they color in the temporal Voronoi game. In fact, by definition of the two games, this holds for every temporal graph (as in the static case).
Difference games. We also define difference versions of both temporal diffusion and Voronoi games, as they will prove useful in analyzing the classical versions. We only define these for $k=2$ players, since there is no single canonical generalization to more players. The key difference in
5
|
[
[
0.283935546875,
0.0885009765625,
0.336181640625,
0.0885009765625,
0.336181640625,
0.09857177734375,
0.283935546875,
0.09857177734375
],
[
0.37890625,
0.0885009765625,
0.4404296875,
0.0885009765625,
0.4404296875,
0.09857177734375,
0.37890625,
0.09857177734375
],
[
0.58740234375,
0.0885009765625,
0.6513671875,
0.0885009765625,
0.6513671875,
0.09857177734375,
0.58740234375,
0.09857177734375
],
[
0.6552734375,
0.0914306640625,
0.67236328125,
0.0914306640625,
0.67236328125,
0.0955810546875,
0.6552734375,
0.0955810546875
],
[
0.673828125,
0.0885009765625,
0.7421875,
0.0885009765625,
0.7421875,
0.09857177734375,
0.673828125,
0.09857177734375
],
[
0.3984375,
0.105712890625,
0.417724609375,
0.105712890625,
0.417724609375,
0.1134033203125,
0.3984375,
0.1134033203125
],
[
0.41943359375,
0.105712890625,
0.43701171875,
0.105712890625,
0.43701171875,
0.114013671875,
0.41943359375,
0.114013671875
],
[
0.673828125,
0.1051025390625,
0.72265625,
0.1051025390625,
0.72265625,
0.11517333984375,
0.673828125,
0.11517333984375
],
[
0.72265625,
0.105712890625,
0.740234375,
0.105712890625,
0.740234375,
0.114013671875,
0.72265625,
0.114013671875
],
[
0.72607421875,
0.122314453125,
0.740234375,
0.122314453125,
0.740234375,
0.1312255859375,
0.72607421875,
0.1312255859375
],
[
0.65869140625,
0.10809326171875,
0.67236328125,
0.10809326171875,
0.67236328125,
0.11224365234375,
0.65869140625,
0.11224365234375
],
[
0.427001953125,
0.12353515625,
0.4345703125,
0.12353515625,
0.4345703125,
0.127685546875,
0.427001953125,
0.127685546875
],
[
0.69580078125,
0.12353515625,
0.72119140625,
0.12353515625,
0.72119140625,
0.1312255859375,
0.69580078125,
0.1312255859375
],
[
0.6796875,
0.12469482421875,
0.69580078125,
0.12469482421875,
0.69580078125,
0.12890625,
0.6796875,
0.12890625
],
[
0.402587890625,
0.12469482421875,
0.411865234375,
0.12469482421875,
0.411865234375,
0.12890625,
0.402587890625,
0.12890625
],
[
0.56982421875,
0.135986328125,
0.5771484375,
0.14013671875,
0.56982421875,
0.146728515625,
0.56201171875,
0.1436767578125
],
[
0.72265625,
0.1395263671875,
0.740234375,
0.1395263671875,
0.740234375,
0.1478271484375,
0.72265625,
0.1478271484375
],
[
0.69580078125,
0.14013671875,
0.72265625,
0.14013671875,
0.72265625,
0.1478271484375,
0.69580078125,
0.1478271484375
],
[
0.427001953125,
0.14013671875,
0.432861328125,
0.14013671875,
0.432861328125,
0.146728515625,
0.427001953125,
0.146728515625
],
[
0.705078125,
0.1519775390625,
0.74365234375,
0.1561279296875,
0.740234375,
0.166259765625,
0.70166015625,
0.162109375
],
[
0.427001953125,
0.157958984375,
0.432861328125,
0.157958984375,
0.432861328125,
0.163330078125,
0.427001953125,
0.163330078125
],
[
0.6748046875,
0.17333984375,
0.70068359375,
0.17333984375,
0.70068359375,
0.18115234375,
0.6748046875,
0.18115234375
],
[
0.69921875,
0.17333984375,
0.72119140625,
0.17333984375,
0.72119140625,
0.18115234375,
0.69921875,
0.18115234375
],
[
0.72705078125,
0.17333984375,
0.736328125,
0.17333984375,
0.736328125,
0.1787109375,
0.72705078125,
0.1787109375
],
[
0.65869140625,
0.177001953125,
0.6748046875,
0.177001953125,
0.6748046875,
0.179931640625,
0.65869140625,
0.179931640625
],
[
0.28662109375,
0.17578125,
0.29736328125,
0.17578125,
0.29736328125,
0.179931640625,
0.28662109375,
0.179931640625
],
[
0.42529296875,
0.188232421875,
0.4404296875,
0.188232421875,
0.4404296875,
0.198974609375,
0.42529296875,
0.198974609375
],
[
0.6142578125,
0.188232421875,
0.67626953125,
0.188232421875,
0.67626953125,
0.203125,
0.6142578125,
0.203125
],
[
0.72705078125,
0.1859130859375,
0.74560546875,
0.1859130859375,
0.74560546875,
0.20068359375,
0.72705078125,
0.20068359375
],
[
0.34619140625,
0.203125,
0.3984375,
0.203125,
0.3984375,
0.21435546875,
0.34619140625,
0.21435546875
],
[
0.317626953125,
0.190673828125,
0.371337890625,
0.190673828125,
0.371337890625,
0.201904296875,
0.317626953125,
0.201904296875
],
[
0.6494140625,
0.203125,
0.70166015625,
0.203125,
0.70166015625,
0.21435546875,
0.6494140625,
0.21435546875
],
[
0.2496337890625,
0.221435546875,
0.4462890625,
0.221435546875,
0.4462890625,
0.235107421875,
0.2496337890625,
0.235107421875
],
[
0.552734375,
0.221435546875,
0.74609375,
0.221435546875,
0.74609375,
0.235107421875,
0.552734375,
0.235107421875
],
[
0.12017822265625,
0.2464599609375,
0.87744140625,
0.2464599609375,
0.87744140625,
0.26123046875,
0.12017822265625,
0.26123046875
],
[
0.12017822265625,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.275634765625,
0.12017822265625,
0.275634765625
],
[
0.12017822265625,
0.275634765625,
0.880859375,
0.275634765625,
0.880859375,
0.289306640625,
0.12017822265625,
0.289306640625
],
[
0.121826171875,
0.290283203125,
0.87744140625,
0.290283203125,
0.87744140625,
0.300537109375,
0.121826171875,
0.300537109375
],
[
0.11846923828125,
0.303955078125,
0.87890625,
0.303955078125,
0.87890625,
0.317626953125,
0.11846923828125,
0.317626953125
],
[
0.12017822265625,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.33203125,
0.12017822265625,
0.33203125
],
[
0.11846923828125,
0.333251953125,
0.2763671875,
0.333251953125,
0.2763671875,
0.3466796875,
0.11846923828125,
0.3466796875
],
[
0.11846923828125,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.38720703125,
0.11846923828125,
0.38720703125
],
[
0.115966796875,
0.385986328125,
0.88330078125,
0.38720703125,
0.88330078125,
0.405029296875,
0.115966796875,
0.40380859375
],
[
0.11846923828125,
0.405029296875,
0.88232421875,
0.405029296875,
0.88232421875,
0.419189453125,
0.11846923828125,
0.419189453125
],
[
0.1168212890625,
0.419189453125,
0.88330078125,
0.419189453125,
0.88330078125,
0.43701171875,
0.1168212890625,
0.43701171875
],
[
0.1168212890625,
0.435791015625,
0.88232421875,
0.435791015625,
0.88232421875,
0.45361328125,
0.1168212890625,
0.45361328125
],
[
0.11846923828125,
0.45361328125,
0.880859375,
0.45361328125,
0.880859375,
0.468505859375,
0.11846923828125,
0.468505859375
],
[
0.11846923828125,
0.4697265625,
0.7900390625,
0.4697265625,
0.7900390625,
0.48388671875,
0.11846923828125,
0.48388671875
],
[
0.1470947265625,
0.486328125,
0.880859375,
0.486328125,
0.880859375,
0.50048828125,
0.1470947265625,
0.50048828125
],
[
0.1168212890625,
0.50048828125,
0.35205078125,
0.501953125,
0.35205078125,
0.5166015625,
0.1168212890625,
0.515625
],
[
0.12017822265625,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.12017822265625,
0.5498046875
],
[
0.1168212890625,
0.5517578125,
0.880859375,
0.55029296875,
0.880859375,
0.5654296875,
0.1168212890625,
0.56640625
],
[
0.12017822265625,
0.568359375,
0.87890625,
0.568359375,
0.87890625,
0.58203125,
0.12017822265625,
0.58203125
],
[
0.11846923828125,
0.5830078125,
0.880859375,
0.58447265625,
0.880859375,
0.5986328125,
0.11846923828125,
0.59716796875
],
[
0.11846923828125,
0.60107421875,
0.880859375,
0.60107421875,
0.880859375,
0.615234375,
0.11846923828125,
0.615234375
],
[
0.11846923828125,
0.615234375,
0.7900390625,
0.615234375,
0.7900390625,
0.6298828125,
0.11846923828125,
0.6298828125
],
[
0.12017822265625,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.66455078125,
0.12017822265625,
0.66455078125
],
[
0.11846923828125,
0.66650390625,
0.880859375,
0.6650390625,
0.880859375,
0.6796875,
0.11846923828125,
0.68115234375
],
[
0.11846923828125,
0.68310546875,
0.880859375,
0.68310546875,
0.880859375,
0.69677734375,
0.11846923828125,
0.69677734375
],
[
0.11846923828125,
0.69970703125,
0.76123046875,
0.69970703125,
0.76123046875,
0.71337890625,
0.11846923828125,
0.71337890625
],
[
0.1470947265625,
0.71875,
0.473876953125,
0.71875,
0.473876953125,
0.732421875,
0.1470947265625,
0.732421875
],
[
0.52001953125,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.52001953125,
0.732421875
],
[
0.11846923828125,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.11846923828125,
0.7490234375
],
[
0.12017822265625,
0.75048828125,
0.87890625,
0.75048828125,
0.87890625,
0.76416015625,
0.12017822265625,
0.76416015625
],
[
0.1168212890625,
0.765625,
0.880859375,
0.7666015625,
0.880859375,
0.78076171875,
0.1168212890625,
0.77978515625
],
[
0.12017822265625,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79638671875,
0.12017822265625,
0.79638671875
],
[
0.11846923828125,
0.7998046875,
0.88232421875,
0.7998046875,
0.88232421875,
0.8134765625,
0.11846923828125,
0.8134765625
],
[
0.12017822265625,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.84716796875,
0.12017822265625,
0.84716796875
],
[
0.12017822265625,
0.849609375,
0.880859375,
0.849609375,
0.880859375,
0.86328125,
0.12017822265625,
0.86328125
],
[
0.1168212890625,
0.865234375,
0.880859375,
0.8642578125,
0.880859375,
0.87890625,
0.1168212890625,
0.8798828125
],
[
0.492431640625,
0.90283203125,
0.51025390625,
0.90283203125,
0.51025390625,
0.91748046875,
0.492431640625,
0.91748046875
]
] |
[
"00",
"000",
"00.",
"",
"00",
"0",
"O",
"00",
"O",
"O",
"",
"C",
"0",
"",
"",
"2",
"O",
"O",
"C",
"00",
"C",
"0",
"O",
"",
"",
"",
"6",
"1 pip2",
"6",
"vertex",
"pip2",
"vertex",
"(a) Temporal diffusion game.",
"(b) Temporal Voronoi game.",
"Figure l: Example of our two games on the temporal path g = ([6], Ei, . .., Es) with Et = [(t,t + 1]]",
"for t E [5], where player 1 selects vertex pi = 2 and player 2 selects vertex p2 = 3. In the diffusion game,",
"the color of a player cannot “pass\" a vertex colored by the other player. Thus, the two players split",
"the vertices. that is. plaver 1 colors the vertices in [1.2l. while plaver 2 colors the vertices in J3.6l. In",
"contrast to this, in the Voronoi game, player 1 is able to “catch up\" with player 2: They both reach the",
"vertices 4, 5, and 6 at the same time. The displayed strategy profile is a Nash equilibrium in Vor(G,2)",
"but not in Diff(G,2).",
"(Ut(pi,...,pk))-o must eventually become constant. We denote its limit by U;(pi,...,Pk)",
"The payoff or outcome for player i is then ui(p1,...,pk) := [U;(p1,...,pk)l, and the play-",
"ers aim to maximize this value. Consequently, we say that a player i E [k] plays a best re-",
"sponse to the other players in the strategy profile (pi,...,Pk) if for all vertices p' e V it holds",
"that u;(p1,..., Pi-1,p',pi+1,..., Pk) ≤ui(pi,...,Pk). A strategy profile (p1,...,Pk) is a Nash",
"equilibrium if every player i E [k] plays a best response to the other players. Any strategy profile",
"of the form (p, P, .: : ,p) in which all players choose the same vertex is called trivial.",
"It remains to specify how the strategy profile (pi,...,pk) determines the coloring of the",
"vertices V in the two games.",
"Temporal diffusion games. In a temporal diffusion game, the temporal graph g is colored",
"by the following propagation process over time. We call a vertex uncolored if no color has been",
"assigned to it (so far). In step t, we consider the layer Gt. We color a so far uncolored vertex u",
"with color i e [k] if has at least one neighbor in Gt that is colored with color i e [k] and no",
"neighbor in Gt that is colored with any other color j E [k] / [i]. Every uncolored vertex with",
"at least two neighbors in Gt colored by two different colors i, j e [k] is colored gray.",
"Temporal Voronoi games. In a temporal Voronoi game, at time t a vertex v is colored with",
"color i e [k] if pi is the only position of a player that can reach w until time t by a temporal walk.",
"(If an uncolored vertex is simultaneously reached by two or more players, then it is colored gray.)",
"In effect, u E U;(pi,..., pk) if and only if td(pi, u) < td(pi, v) holds for all j + i.",
"Note that we defined temporal diffusion",
" and temporal Voronoi games such that both",
"temporal games played on a temporal graph g with lifetime 1 are equivalent to the (non-",
"temporal) game played on the static graph Gr: An example of a temporal diffusion game and",
"a temporal Voronoi game is shown in Figure 1. Notably, in the displayed temporal diffusion",
"game each player colors a superset of the vertices they color in the temporal Voronoi game. In",
"fact, by definition of the two games, this holds for every temporal graph (as in the static case).",
"Difference games. We also define difference versions of both temporal diffusion and Voronoi",
"games, as they will prove useful in analyzing the classical versions. We only define these for k = 2",
"players, since there is no single canonical generalization to more players. The key difference in",
"5"
] |
{
"arxiv_id": "2105.05987",
"language": "en",
"page_idx": 4
}
|
||
1312.5548-en-0
|
1312.5548-en-0.jpg
|
# $\textbf{My First Deep Learning System of 1991}$ +Deep Learning Timeline 1962-2013
Jürgen Schmidhuber
The Swiss AI Lab IDSIA, Galleria 2,6928 Manno-Lugano
University of Lugano & SUPSI, Switzerland
20 December 2013
Abstract
Deep Learning has attracted significant attention in recent years. Here I present a brief overview of
my first Deep Learner of 1991, and its historic context, with a timeline of Deep Learning highlights.
$Note:$ As a machine learning researcher I am obsessed with proper credit assignment. This draft is the result of an experiment in rapid massive open online peer review. Since 20 September 2013, subsequent revisions published under www.deeplearning.me have absorbed many suggestions for improvements by experts. The abbreviation“TL”is used to refer to subsections of the timeline section.
<FigureHere>
Figure 1: My first Deep Learning system of 1991 used a deep stack of recurrent neural networks (a Neural Hierarchical Temporal Memory) pre-trained in unsupervised fashion to accelerate subsequent supervised learning [79,81,82].
|
[
[
0.265625,
0.1875,
0.71728515625,
0.1875,
0.71728515625,
0.2052001953125,
0.265625,
0.2052001953125
],
[
0.275390625,
0.21337890625,
0.708984375,
0.212158203125,
0.708984375,
0.23291015625,
0.275390625,
0.2342529296875
],
[
0.41015625,
0.255126953125,
0.57421875,
0.255126953125,
0.57421875,
0.26953125,
0.41015625,
0.26953125
],
[
0.0228729248046875,
0.268310546875,
0.056365966796875,
0.268310546875,
0.05963134765625,
0.41845703125,
0.02532958984375,
0.41845703125
],
[
0.261474609375,
0.271484375,
0.72119140625,
0.271484375,
0.72119140625,
0.285888671875,
0.261474609375,
0.285888671875
],
[
0.31787109375,
0.290283203125,
0.66650390625,
0.290283203125,
0.66650390625,
0.304931640625,
0.31787109375,
0.304931640625
],
[
0.418212890625,
0.3193359375,
0.56787109375,
0.3193359375,
0.56787109375,
0.333251953125,
0.418212890625,
0.333251953125
],
[
0.464111328125,
0.367431640625,
0.521484375,
0.367431640625,
0.521484375,
0.37890625,
0.464111328125,
0.37890625
],
[
0.2059326171875,
0.386474609375,
0.7998046875,
0.386962890625,
0.7998046875,
0.40087890625,
0.2059326171875,
0.399658203125
],
[
0.1846923828125,
0.401611328125,
0.77294921875,
0.401611328125,
0.77294921875,
0.415283203125,
0.1846923828125,
0.415283203125
],
[
0.0228729248046875,
0.4091796875,
0.056365966796875,
0.40771484375,
0.05963134765625,
0.49365234375,
0.02532958984375,
0.494384765625
],
[
0.20751953125,
0.415283203125,
0.79833984375,
0.415283203125,
0.79833984375,
0.4287109375,
0.20751953125,
0.4287109375
],
[
0.1871337890625,
0.43115234375,
0.79736328125,
0.43115234375,
0.79736328125,
0.44140625,
0.1871337890625,
0.44140625
],
[
0.1854248046875,
0.443115234375,
0.79736328125,
0.443115234375,
0.79736328125,
0.45703125,
0.1854248046875,
0.45703125
],
[
0.1854248046875,
0.45703125,
0.68701171875,
0.45703125,
0.68701171875,
0.47021484375,
0.1854248046875,
0.47021484375
],
[
0.0228729248046875,
0.50048828125,
0.0555419921875,
0.50048828125,
0.058013916015625,
0.7021484375,
0.02532958984375,
0.7021484375
],
[
0.035125732421875,
0.49560546875,
0.0482177734375,
0.49560546875,
0.0482177734375,
0.5048828125,
0.035125732421875,
0.5048828125
],
[
0.1829833984375,
0.73095703125,
0.2059326171875,
0.75048828125,
0.191162109375,
0.76123046875,
0.1683349609375,
0.7412109375
],
[
0.158447265625,
0.7587890625,
0.1732177734375,
0.7587890625,
0.1732177734375,
0.767578125,
0.158447265625,
0.767578125
],
[
0.193603515625,
0.7548828125,
0.2197265625,
0.7470703125,
0.240234375,
0.7822265625,
0.214111328125,
0.79052734375
],
[
0.176513671875,
0.7734375,
0.1968994140625,
0.76953125,
0.2059326171875,
0.79541015625,
0.1854248046875,
0.7998046875
],
[
0.146240234375,
0.8515625,
0.8388671875,
0.8515625,
0.8388671875,
0.86572265625,
0.146240234375,
0.86572265625
],
[
0.146240234375,
0.8671875,
0.8388671875,
0.8671875,
0.8388671875,
0.880859375,
0.146240234375,
0.880859375
],
[
0.1446533203125,
0.8818359375,
0.28515625,
0.8818359375,
0.28515625,
0.89599609375,
0.1446533203125,
0.89599609375
]
] |
[
"MyFirst DeepLearning Systemof 1991",
"+ Deep Learning Timeline 1962-2013",
"Juirgen Schmidhuber",
"0",
"The Swiss AI Lab IDSIA, Galleria 2, 6928 Manno-Lugano",
"University of Lugano & SUPSI, Switzerland",
"20December2013",
"Abstract",
"DeepLearning has attracted significant attention in recent years.HereI present a brief overview of",
"my first Deep Learner of 1991, and its historic context, with a timeline of Deep Learning highlights",
"[N'S]",
"Note: As a machine learning researcher I am obsessed with proper credit assignment. This draft is the",
"resultof anexperimentinrapidmassiveopenonlinepeerreview.Since20September2013,subsequent",
"revisions published under www.deeplearning.me have absorbed many suggestions for improvements by",
"experts. The abbreviation \"TL\" is used to refer to subsections of the timeline section.",
"",
"",
"P",
"N",
"",
"",
"Figure l: My first Deep Learning system of 1991 used a deep stack of recurrent neural networks (a Neural",
"Hierarchical Temporal Memory) pre-trained in unsupervised fashion to accelerate subsequent supervised",
"learning [79, 81, 82]."
] |
{
"arxiv_id": "1312.5548",
"language": "en",
"page_idx": 0
}
|
||
1312.5548-en-1
|
1312.5548-en-1.jpg
|
In 2009, our Deep Learning Artificial Neural Networks became the first Deep Learners to win official international pattern recognition competitions [40, 83] (with secret test sets known only to the organisers); by 2012 they had won eight of them (TL 1.13), including the first contests on object detection in large images (ICPR 2012)[2,16] and image segmentation (ISBI 2012)[3,15]. In 2011, they achieved the world's first superhuman visual pattern recognition results [20,19]. Others have implemented very similar techniques, e.g,[51], and won additional contests or set benchmark records since 2012, e.g.,(TL 1.13, TL 1.14). The field of Deep Learning research is far older though—compare the timeline (TL) further down.
My first Deep Learner dates back to 1991 [79,81,82](TL 1.7). It can perform credit assignment across hundreds of nonlinear operators or neural layers, by using unsupervised pre-training for a stack of recurrent neural networks (RNN) (deep by nature) as in Figure 1. Such RNN are general computers more powerful than normal feedforward NN, and can encode entire sequences of input vectors.
The basic idea is still relevant today. Each RNN is trained for a while in unsupervised fashion to predict its next input. From then on, only unexpected inputs (errors) convey new information and get fed to the next higher RNN which thus ticks on a slower, self-organising time scale. It can easily be shown that no information gets lost. It just gets compressed (much of machine learning is essentially about compression). We get less and less redundant input sequence encodings in deeper and deeper levels of this hierarchical temporal memory, which compresses data in both space (like feedforward NN) and time. There also is a continuous variant [84].
One ancient illustrative Deep Learning experiment of 1993 [82] required credit assignment across 1200 time steps, or through 1200 subsequent nonlinear virtual layers. The top level code of the initially unsupervised RNN stack, however, got so compact that (previously infeasible) sequence classification through additional supervised learning became possible.
There is a way of compressing higher levels down into lower levels, thus partially collapsing the hierarchical temporal memory. The trick is to retrain lower-level RNN to continually imitate (predict) the hidden units of already trained, slower, higher-level RNN, through additional predictive output neurons [81, 79, 82]. This helps the lower RNN to develop appropriate, rarely changing memories that may bridge very long time lags.
The Deep Learner of 1991 was a first way of overcoming the Fundamental Deep Learning Problem identified and analysed in 1991 by my very first student (now professor) Sepp Hochreiter (TL 1.6): the problem of vanishing or exploding gradients [46,47,10]. The latter motivated all our subsequent Deep Learning research of the 1990s and 2000s.
Through supervised LSTM RNN (1997) (e.g.,[48,32,39,36,37,40,38,83], TL 1.8) we could eventually perform similar feats as with the 1991 system [81,82], overcoming the Fundamental Deep Learning Problem without any unsupervised pre-training. Moreover, LSTM could also learn tasks unlearnable by the partially unsupervised 1991 chunker [81,82].
Particularly successful are stacks of LSTM RNN [40] trained by Connectionist Temporal Classification (CTC)[36]. On faster computers of 2009, this became the first RNN system ever to win an official international pattern recognition competition [40,83], through the work of my PhD student and postdoc Alex Graves, e.g.,[40]. To my knowledge, this also was the first Deep Learning system ever (recurrent or not) to win such a contest (TL 1.10). (In fact, it won three different ICDAR 2009 contests on connected handwriting in three different languages, e.g.,[83,40], TL 1.10.) A while ago, Alex moved on to Geoffrey Hinton's lab (Univ. Toronto), where a stack[40] of our bidirectional LSTM RNN [39] also broke a famous TIMIT speech recognition record [38] (TL 1.14), despite thousands of man years previously spent on HMM-based speech recognition research. CTC-LSTM also helped to score first at NIST's OpenHaRT2013 evaluation [11].
Recently, well-known entrepreneurs also got interested[43,52] in such hierarchical temporal memories
[81,82](TL 1.7).
The expression Deep Learning actually got coined relatively late, around 2006, in the context of unsupervised pre-training for less general feedforward networks [44] (TL 1.9). Such a system reached $1.2\%$ error rate[44] on the MNIST handwritten digits[54,55], perhaps the most famous benchmark of Machine Learning. Our team first showed that good old backpropagation (TL 1.2) on GPUs (with training pattern distortions [6, 86] but without any unsupervised pre-training) can actually achieve a three times better result of $0.35\%[17]$- back then, a world record (a previous standard net achieved $0.7\%[86];$ a backprop-trained [54, 55] Convolutional NN (CNN or convnet) [29, 30, 54, 55] got $0.39\%$ [70](TL 1.9);
2
|
[
[
0.1666259765625,
0.136962890625,
0.8388671875,
0.1383056640625,
0.8388671875,
0.15283203125,
0.1666259765625,
0.1514892578125
],
[
0.143798828125,
0.154052734375,
0.8408203125,
0.1546630859375,
0.8408203125,
0.1685791015625,
0.143798828125,
0.1673583984375
],
[
0.146240234375,
0.17041015625,
0.83837890625,
0.17041015625,
0.83837890625,
0.18115234375,
0.146240234375,
0.18115234375
],
[
0.1446533203125,
0.18310546875,
0.83837890625,
0.18310546875,
0.83837890625,
0.1962890625,
0.1446533203125,
0.1962890625
],
[
0.146240234375,
0.199462890625,
0.8388671875,
0.199462890625,
0.8388671875,
0.21337890625,
0.146240234375,
0.21337890625
],
[
0.1446533203125,
0.2139892578125,
0.83837890625,
0.21337890625,
0.83837890625,
0.2279052734375,
0.1446533203125,
0.2291259765625
],
[
0.1446533203125,
0.2279052734375,
0.833984375,
0.2298583984375,
0.833984375,
0.24365234375,
0.1446533203125,
0.2412109375
],
[
0.1690673828125,
0.2442626953125,
0.8388671875,
0.2442626953125,
0.8388671875,
0.25830078125,
0.1690673828125,
0.25830078125
],
[
0.1446533203125,
0.260009765625,
0.8388671875,
0.260009765625,
0.8388671875,
0.2734375,
0.1446533203125,
0.2734375
],
[
0.143798828125,
0.2734375,
0.8388671875,
0.274658203125,
0.8388671875,
0.2890625,
0.143798828125,
0.28857421875
],
[
0.1446533203125,
0.290283203125,
0.66845703125,
0.290283203125,
0.66845703125,
0.3037109375,
0.1446533203125,
0.3037109375
],
[
0.1683349609375,
0.304931640625,
0.8388671875,
0.304931640625,
0.8388671875,
0.3193359375,
0.1683349609375,
0.3193359375
],
[
0.1446533203125,
0.32080078125,
0.8388671875,
0.32080078125,
0.8388671875,
0.333984375,
0.1446533203125,
0.333984375
],
[
0.1446533203125,
0.335205078125,
0.8388671875,
0.335205078125,
0.8388671875,
0.348388671875,
0.1446533203125,
0.348388671875
],
[
0.146240234375,
0.349853515625,
0.8349609375,
0.349853515625,
0.8349609375,
0.363037109375,
0.146240234375,
0.363037109375
],
[
0.146240234375,
0.365478515625,
0.8388671875,
0.365478515625,
0.8388671875,
0.37890625,
0.146240234375,
0.37890625
],
[
0.1446533203125,
0.380615234375,
0.8408203125,
0.380615234375,
0.8408203125,
0.39453125,
0.1446533203125,
0.39453125
],
[
0.1446533203125,
0.395751953125,
0.303955078125,
0.395751953125,
0.303955078125,
0.4091796875,
0.1446533203125,
0.4091796875
],
[
0.1707763671875,
0.410400390625,
0.83837890625,
0.410400390625,
0.83837890625,
0.423583984375,
0.1707763671875,
0.423583984375
],
[
0.1446533203125,
0.425537109375,
0.83837890625,
0.425537109375,
0.83837890625,
0.43994140625,
0.1446533203125,
0.43994140625
],
[
0.1446533203125,
0.44140625,
0.83837890625,
0.44140625,
0.83837890625,
0.455078125,
0.1446533203125,
0.455078125
],
[
0.143798828125,
0.455078125,
0.4599609375,
0.455810546875,
0.4599609375,
0.47021484375,
0.143798828125,
0.4697265625
],
[
0.1690673828125,
0.47021484375,
0.83837890625,
0.47021484375,
0.83837890625,
0.48486328125,
0.1690673828125,
0.48486328125
],
[
0.146240234375,
0.486083984375,
0.83642578125,
0.486083984375,
0.83642578125,
0.5,
0.146240234375,
0.5
],
[
0.143798828125,
0.5,
0.83837890625,
0.50048828125,
0.83837890625,
0.51513671875,
0.143798828125,
0.5146484375
],
[
0.1446533203125,
0.5166015625,
0.83642578125,
0.5166015625,
0.83642578125,
0.52978515625,
0.1446533203125,
0.52978515625
],
[
0.143798828125,
0.52978515625,
0.32177734375,
0.53076171875,
0.32177734375,
0.54541015625,
0.143798828125,
0.54443359375
],
[
0.1690673828125,
0.54541015625,
0.8388671875,
0.54541015625,
0.8388671875,
0.56005859375,
0.1690673828125,
0.56005859375
],
[
0.1446533203125,
0.56005859375,
0.8388671875,
0.56103515625,
0.8388671875,
0.57568359375,
0.1446533203125,
0.57470703125
],
[
0.143798828125,
0.57568359375,
0.8388671875,
0.5771484375,
0.8388671875,
0.59130859375,
0.143798828125,
0.59033203125
],
[
0.1446533203125,
0.59033203125,
0.426513671875,
0.59130859375,
0.426513671875,
0.60595703125,
0.1446533203125,
0.60498046875
],
[
0.1683349609375,
0.60595703125,
0.83642578125,
0.60595703125,
0.83642578125,
0.619140625,
0.1683349609375,
0.619140625
],
[
0.146240234375,
0.6220703125,
0.83837890625,
0.6220703125,
0.83837890625,
0.63525390625,
0.146240234375,
0.63525390625
],
[
0.1446533203125,
0.63623046875,
0.83837890625,
0.63720703125,
0.83837890625,
0.6513671875,
0.1446533203125,
0.65087890625
],
[
0.146240234375,
0.65380859375,
0.464111328125,
0.65380859375,
0.464111328125,
0.6640625,
0.146240234375,
0.6640625
],
[
0.1707763671875,
0.66748046875,
0.8388671875,
0.66748046875,
0.8388671875,
0.68115234375,
0.1707763671875,
0.68115234375
],
[
0.146240234375,
0.681640625,
0.83837890625,
0.681640625,
0.83837890625,
0.69580078125,
0.146240234375,
0.69580078125
],
[
0.146240234375,
0.69775390625,
0.8388671875,
0.69775390625,
0.8388671875,
0.7109375,
0.146240234375,
0.7109375
],
[
0.1446533203125,
0.71337890625,
0.8388671875,
0.71337890625,
0.8388671875,
0.7265625,
0.1446533203125,
0.7265625
],
[
0.1478271484375,
0.72802734375,
0.83837890625,
0.72802734375,
0.83837890625,
0.7412109375,
0.1478271484375,
0.7412109375
],
[
0.1446533203125,
0.7421875,
0.8388671875,
0.7421875,
0.8388671875,
0.7568359375,
0.1446533203125,
0.7568359375
],
[
0.1446533203125,
0.75830078125,
0.83837890625,
0.75830078125,
0.83837890625,
0.771484375,
0.1446533203125,
0.771484375
],
[
0.146240234375,
0.7734375,
0.8388671875,
0.7734375,
0.8388671875,
0.787109375,
0.146240234375,
0.787109375
],
[
0.1446533203125,
0.78857421875,
0.8388671875,
0.78857421875,
0.8388671875,
0.8017578125,
0.1446533203125,
0.8017578125
],
[
0.146240234375,
0.8046875,
0.178955078125,
0.8046875,
0.178955078125,
0.81640625,
0.146240234375,
0.81640625
],
[
0.1683349609375,
0.818359375,
0.83837890625,
0.818359375,
0.83837890625,
0.83203125,
0.1683349609375,
0.83203125
],
[
0.146240234375,
0.833984375,
0.261474609375,
0.833984375,
0.261474609375,
0.84765625,
0.146240234375,
0.84765625
],
[
0.1690673828125,
0.84765625,
0.83642578125,
0.84765625,
0.83642578125,
0.8623046875,
0.1690673828125,
0.8623046875
],
[
0.1446533203125,
0.86279296875,
0.83837890625,
0.861328125,
0.83837890625,
0.87548828125,
0.1446533203125,
0.87744140625
],
[
0.1446533203125,
0.87890625,
0.8388671875,
0.87890625,
0.8388671875,
0.892578125,
0.1446533203125,
0.892578125
],
[
0.1446533203125,
0.89306640625,
0.8388671875,
0.89306640625,
0.8388671875,
0.90771484375,
0.1446533203125,
0.90771484375
],
[
0.146240234375,
0.9091796875,
0.8388671875,
0.9091796875,
0.8388671875,
0.92236328125,
0.146240234375,
0.92236328125
],
[
0.143798828125,
0.92236328125,
0.8408203125,
0.923828125,
0.8408203125,
0.93798828125,
0.143798828125,
0.9375
],
[
0.1446533203125,
0.939453125,
0.8388671875,
0.939453125,
0.8388671875,
0.95263671875,
0.1446533203125,
0.95263671875
],
[
0.491943359375,
0.974609375,
0.50244140625,
0.98193359375,
0.491943359375,
0.9921875,
0.48046875,
0.98486328125
]
] |
[
"In 2009, our Deep Learning Artificial Neural Networks became the first Deep Learners to win official",
"international pattern recognition competitions [40, 83] (with secret test setsknown only to the organisers);",
"by 2012theyhad won eight of them (TL 1.13),includingthe first contests on object detection in large",
"images (ICPR 2012) [2, 16] and image segmentation (ISB1 2012) [3, 15]. In 2011, they achieved the",
"world's first superhuman visual pattern recognition results [20, 19]. Others have implemented very similar",
"techniques, e.g., [51], and won additional contests or set benchmark records since 2012, e.g., (TL 1.13, TL",
"1.14).Thefield ofDeepLearningresearchisfarolderthoughcomparethetimeline(TL)furtherdown.",
"My first Deep Learner dates back to 1991 [79, 81, 82] (TL 1.7). It can perform credit assignment across",
"hundreds of nonlinear operators or neural layers, by using unsupervised pre-training for a stack of recurrent",
"neural networks (RNN) (deep by nature) as in Figure 1. Such RNN are general computers more powerful",
"than normal feedforward NN, and can encode entire sequences of input vectors.",
"The basic idea is still relevant today. Each RNN is trained for a while in unsupervised fashion to predict",
"its next input. From then on, only unexpected inputs (errors) convey new information and get fed to the",
"next higher RNN which thus ticks on a slower, self-organising time scale. It can easily be shown that no",
"information gets lost. It just gets compressed (much of machine learning is essentially about compression)",
"We get less and less redundant input sequence encodings in deeper and deeper levels of this hierarchical",
"temporal memory, which compresses data in both space (like feedforward NN) and time. There also is a",
"continuous variant [84].",
"One ancient illustrative Deep Learning experiment of 1993[82] required credit assignment across 1200",
"time steps, or through 1200 subsequent nonlinear virtual layers. The top level code of the initially unsu-",
"pervised RNN stack, however, got so compact that (previously infeasible) sequence classification through",
"additional supervisedlearningbecamepossible.",
"There is a way of compressing higher levels down into lower levels, thus partially collapsing the",
"hierarchical temporal memory. The trick is to retrain lower-level RNN to continually imitate (predict)",
"the hidden units of already trained, slower, higher-level RNN, through additional predictive output neu-",
"rons [81, 79, 82]. This helps the lower RNN to develop appropriate, rarely changing memories that may",
"bridge very long time lags.",
"The Deep Learner of 1991 was a first way of overcoming the Fundamental Deep Learning Problem",
"identified and analysed in 1991 by my very first student (now professor) Sepp Hochreiter (TL 1.6): the",
"problem of vanishing or exploding gradients [46, 47, 10]. The latter motivated all our subsequent Deep",
"Learning research of the 1990s and 2000s.",
"Through supervised LSTM RNN (1997) (e.g., [48, 32, 39, 36, 37, 40, 38, 83], TL 1.8) we could even-",
"tually perform similar feats as with the 1991 system [81, 82], overcoming the Fundamental Deep Learning",
"Problem without any unsupervised pre-training. Moreover, LSTM could also learn tasks unlearnable by",
"thepartiallyunsupervised1991chunker[81,82]",
"Particularly successful are stacks of LSTM RNN [4O] trained by Connectionist Temporal Classification",
"(CTC) [36]. On faster computers of 2009, this became the first RNN system ever to win an official inter-",
"national pattern recognition competition [4o, 83], through the work of my PhD student and postdoc Alex",
"Graves, e.g., [4o]. To my knowledge, this also was the first Deep Learning system ever (recurrent or not) to",
"win such a contest (TL 1.10). (In fact, it won three different ICDAR 2009 contests on connected handwrit-",
"ing in three different languages, e.g., [83, 40], TL 1.10.) A while ago, Alex moved on to Geoffrey Hinton's",
"lab (Univ. Toronto), where a stack [4O] of our bidirectional LSTM RNN [39] also broke a famous TIMIT",
"speech recognition record [38] (TL 1.14), despite thousands of man years previously spent on HMM-based",
"speech recognition research. CTC-LSTM also helped to score first at NIST's OpenHaRT2013 evaluation",
"[11].",
"Recently, well-known entrepreneurs also got interested [43, 52] in such hierarchical temporal memories",
"[81, 82]1 (TL 1.7).",
"The expression Deep Learning actually got coined relatively late, around 2006, in the context of un-",
"supervised pre-training for less general feedforward networks [44] (TL 1.9). Such a system reached 1.2%",
"error rate [44] on the MNIST handwritten digits [54, 55], perhaps the most famous benchmark of Ma-",
"chine Learning. Our team first showed that good old backpropagation (TL 1.2) on GPUs (with training",
"pattern distortions [6, 86] but without any unsupervised pre-training) can actually achieve a three times",
"better result of 0.35% [17] - back then, a world record (a previous standard net achieved 0.7% [86]; a",
"backprop-trained [54, 55] Convolutional NN (CNN or convnet) [29, 30, 54, 55] got 0.39% [70](TL 1.9);",
"2"
] |
{
"arxiv_id": "1312.5548",
"language": "en",
"page_idx": 1
}
|
||
1312.5548-en-2
|
1312.5548-en-2.jpg
|
plain backprop without distortions except for small saccadic eye movement-like translations already got $0.95\%).$ Then we replaced our standard net by a biologically rather plausible architecture inspired by early neuroscience-related work [29,49,30,54]: Deep and Wide GPU-based Multi-Column Max-Pooling $CNN\left(MCMPCNN\right)\left[18,21,22\right]$with alternating backprop-based [54,55,71] weight-sharing convolutional layers [30,54,56,8] and winner-take-all [29,30] max-pooling [72,71,77] layers (see[14] for early GPU-based CNN). MCMPCNN are committees of MPCNN [19] with simple democratic output averaging (compare earlier more sophisticated ensemble methods [76]). Object detection[16] and image segmentation [15](TL 1.13) profit from fast MPCNN-based image scans [63,62]. Our supervised GPU-MCMPCNN was the first system to achieve superhuman performance in an official international competition (with deadline and secret test set known only to the organisers) [20, 19] (TL 1.12) (compare [85]), and the first with human-competitive performance (around $0.2\%)$ on MNIST [22]. Since 2011, it has won numerous additional competitions on a routine basis, e.g., (TL 1.13, TL 1.14).
Our (multi-column [19]) GPU-MPCNN [18] (TL 1.12) were adopted by the groups of Univ. Toronto $/Stanford/Google$,e.g.,[51,25](TL 1.13,TL 1.14).Apple Inc, the most profitable smartphone maker, hired Ueli Meier, member of our Deep Learning team that won the ICDAR 2011 Chinese handwriting contest [22]. ArcelorMittal, the world's top steel producer, is using our methods for material defect detection and classification, e.g.,[63]. Other users include a leading automotive supplier, recent start-ups such as deepmind (which hired four of my former PhD students/postdocs), and many other companies and leading research labs. One of the most important applications of our techniques is biomedical imaging [16](TL 1.13, TL 1.14), e.g., for cancer prognosis or plaque detection in CT heart scans.
Remarkably, the most successful Deep Learning algorithms in most international contests since 2009 (TL 1.10-1.14) are adaptations and extensions of an over 40-year-old algorithm, namely, supervised efficient backprop [59, 90] (TL 1.2) (compare [60,53,75,12,67]) or BPTT/RTRL for RNN, e.g.,[94,73,91, 68,80,61](exceptions include two 2011 contests specialised on transfer learning [35]—but compare [23]). In particular, as of 2013, state-of-the-art feedforward nets (TL 1.12-1.13) are GPU-based [18] multicolumn [22] combinations of two ancient concepts: Backpropagation (TL 1.2) applied [55] to Neocognitronlike convolutional architectures (TL 1.3) (with max-pooling layers [72,71,77] instead of alternative [29, 30,31,78] local winner-take-all methods). (Plus additional tricks from the 1990s and 2000s,e.g.,[65,64].) In the quite different deep recurrent case, supervised systems also dominate, e.g.,[48,36,40,37,61,38] (TL 1.10, TL 1.14).
In particular, most competition-winning or benchmark record-setting Deep Learners (TL 1.10- TL 1.14) use one of two supervised techniques developed in my lab: (1) recurrent LSTM (1997) (TL 1.8) trained by CTC (2006) [36], or (2) feedforward GPU-MPCNN (2011) [18] (TL 1.12) (building on earlier work since the 1960s mentioned in the text above). Nevertheless, in many applications it can still be advantageous to combine the best of both worlds- supervised learning and unsupervised pre-training, like in my 1991 system described above [79,81,82].
3
|
[
[
0.1446533203125,
0.1383056640625,
0.8388671875,
0.1395263671875,
0.8388671875,
0.154052734375,
0.1446533203125,
0.15283203125
],
[
0.1446533203125,
0.154052734375,
0.83642578125,
0.154052734375,
0.83642578125,
0.1673583984375,
0.1446533203125,
0.1673583984375
],
[
0.146240234375,
0.1685791015625,
0.83642578125,
0.1685791015625,
0.83642578125,
0.1817626953125,
0.146240234375,
0.1817626953125
],
[
0.1446533203125,
0.18310546875,
0.8388671875,
0.184326171875,
0.8388671875,
0.1988525390625,
0.1446533203125,
0.1976318359375
],
[
0.143798828125,
0.1988525390625,
0.83837890625,
0.199462890625,
0.83837890625,
0.2139892578125,
0.143798828125,
0.21337890625
],
[
0.1446533203125,
0.21337890625,
0.83837890625,
0.2139892578125,
0.83837890625,
0.2291259765625,
0.1446533203125,
0.2279052734375
],
[
0.146240234375,
0.2298583984375,
0.83837890625,
0.2298583984375,
0.83837890625,
0.24365234375,
0.146240234375,
0.24365234375
],
[
0.1446533203125,
0.24560546875,
0.83837890625,
0.24560546875,
0.83837890625,
0.2587890625,
0.1446533203125,
0.2587890625
],
[
0.146240234375,
0.261962890625,
0.83642578125,
0.261962890625,
0.83642578125,
0.272705078125,
0.146240234375,
0.272705078125
],
[
0.1446533203125,
0.27587890625,
0.8388671875,
0.27587890625,
0.8388671875,
0.2890625,
0.1446533203125,
0.2890625
],
[
0.1446533203125,
0.2890625,
0.83642578125,
0.2890625,
0.83642578125,
0.302978515625,
0.1446533203125,
0.302978515625
],
[
0.1446533203125,
0.30615234375,
0.5595703125,
0.30615234375,
0.5595703125,
0.3193359375,
0.1446533203125,
0.3193359375
],
[
0.1707763671875,
0.32080078125,
0.8388671875,
0.32080078125,
0.8388671875,
0.333984375,
0.1707763671875,
0.333984375
],
[
0.1446533203125,
0.333984375,
0.8388671875,
0.335205078125,
0.8388671875,
0.349853515625,
0.1446533203125,
0.348388671875
],
[
0.1446533203125,
0.35107421875,
0.83837890625,
0.35107421875,
0.83837890625,
0.3642578125,
0.1446533203125,
0.3642578125
],
[
0.1446533203125,
0.3662109375,
0.83837890625,
0.3662109375,
0.83837890625,
0.380126953125,
0.1446533203125,
0.380126953125
],
[
0.146240234375,
0.380615234375,
0.8388671875,
0.380615234375,
0.8388671875,
0.395751953125,
0.146240234375,
0.395751953125
],
[
0.1446533203125,
0.396484375,
0.83642578125,
0.396484375,
0.83642578125,
0.410400390625,
0.1446533203125,
0.410400390625
],
[
0.1446533203125,
0.410888671875,
0.8388671875,
0.410888671875,
0.8388671875,
0.4248046875,
0.1446533203125,
0.4248046875
],
[
0.146240234375,
0.4267578125,
0.6650390625,
0.4267578125,
0.6650390625,
0.43994140625,
0.146240234375,
0.43994140625
],
[
0.1683349609375,
0.44140625,
0.83837890625,
0.44140625,
0.83837890625,
0.455078125,
0.1683349609375,
0.455078125
],
[
0.146240234375,
0.45703125,
0.83642578125,
0.45703125,
0.83642578125,
0.47021484375,
0.146240234375,
0.47021484375
],
[
0.146240234375,
0.4716796875,
0.8388671875,
0.4716796875,
0.8388671875,
0.486083984375,
0.146240234375,
0.486083984375
],
[
0.146240234375,
0.4873046875,
0.83837890625,
0.4873046875,
0.83837890625,
0.50048828125,
0.146240234375,
0.50048828125
],
[
0.1446533203125,
0.501953125,
0.83837890625,
0.501953125,
0.83837890625,
0.51513671875,
0.1446533203125,
0.51513671875
],
[
0.146240234375,
0.517578125,
0.84326171875,
0.517578125,
0.84326171875,
0.53076171875,
0.146240234375,
0.53076171875
],
[
0.1446533203125,
0.5322265625,
0.83837890625,
0.5322265625,
0.83837890625,
0.54541015625,
0.1446533203125,
0.54541015625
],
[
0.1446533203125,
0.54541015625,
0.83642578125,
0.54541015625,
0.83642578125,
0.5595703125,
0.1446533203125,
0.5595703125
],
[
0.1446533203125,
0.5625,
0.8388671875,
0.5625,
0.8388671875,
0.57568359375,
0.1446533203125,
0.57568359375
],
[
0.146240234375,
0.5771484375,
0.273681640625,
0.5771484375,
0.273681640625,
0.59033203125,
0.146240234375,
0.59033203125
],
[
0.1683349609375,
0.59228515625,
0.8388671875,
0.59228515625,
0.8388671875,
0.60693359375,
0.1683349609375,
0.60693359375
],
[
0.1446533203125,
0.60595703125,
0.8388671875,
0.60693359375,
0.8388671875,
0.6220703125,
0.1446533203125,
0.62060546875
],
[
0.1446533203125,
0.62255859375,
0.8388671875,
0.62255859375,
0.8388671875,
0.63623046875,
0.1446533203125,
0.63623046875
],
[
0.1446533203125,
0.63818359375,
0.8388671875,
0.63818359375,
0.8388671875,
0.65283203125,
0.1446533203125,
0.65283203125
],
[
0.146240234375,
0.65283203125,
0.8388671875,
0.65283203125,
0.8388671875,
0.66650390625,
0.146240234375,
0.66650390625
],
[
0.1446533203125,
0.66845703125,
0.463134765625,
0.66845703125,
0.463134765625,
0.681640625,
0.1446533203125,
0.681640625
],
[
0.486083984375,
0.97802734375,
0.498291015625,
0.97802734375,
0.498291015625,
0.99072265625,
0.486083984375,
0.99072265625
]
] |
[
"plain backprop without distortions except for small saccadic eye movement-like translations already got",
"0.95%).Then we replaced our standard net by a biologically rather plausible architecture inspired by",
"early neuroscience-related work [29, 49, 30, 54]:Deep and Wide GPU-based Multi-Column Max-Pooling",
"CNN (MCMPCNN) [18, 21, 22] with alternating backprop-based [54, 55, 71] weight-sharing convolutional",
"layers [30, 54, 56, 8] and winner-take-all [29, 30] max-pooling [72, 71, 77] layers (see [14] for early",
"GPU-based CNN).MCMPCNN are committees of MPCNN [19] with simple democratic output averaging",
"(compare earlier more sophisticated ensemble methods [76]). Object detection [16] and image segmenta-",
"tion [15] (TL 1.13) profit from fast MPCNN-based image scans [63, 62]. Our supervised GPU-MCMPCNN",
"wasthefirst systemtoachievesuperhumanperformanceinanofficialinternationalcompetition(withdead-",
"line and secret test set known only to the organisers) [20, 19] (TL 1.12) (compare [85]), and the first with",
"human-competitive performance (around 0.2%) on MNIST [22]. Since 2011, it has won numerous addi-",
"tional competitions on a routine basis, e.g., (TL 1.13, TL 1.14).",
"Our (multi-column [19]) GPU-MPCNN [18] (TL 1.12) were adopted by the groups of Univ. Toronto",
"/ Stanford / Google, e.g, [51, 25] (TL 1.13, TL 1.14). Apple Inc., the most profitable smartphone maker,",
"hired Ueli Meier, member of our Deep Learning team that won the ICDAR 2011 Chinese handwriting con-",
"test [22]. ArcelorMittal, the world's top steel producer, is using our methods for material defect detection",
"and classification, e.g., [63]. Other users include a leading automotive supplier, recent start-ups such as",
"deepmind (which hired four of myformer PhD students/postdocs), and many other companies and leading",
"research labs. One of the most important applications of our techniques is biomedical imaging [16] (TL",
"1.13, TL 1.14), e.g., for cancer prognosis or plaque detection in CT heart scans.",
"Remarkably, the most successful Deep Learning algorithms in most international contests since 2009",
"(TL 1.10-1.14) are adaptations and extensions of an over 40-year-old algorithm, namely, supervised effi-",
"cient backprop [59, 90] (TL 1.2) (compare [60, 53, 75, 12, 67]) or BPTT/RTRL for RNN, e.g., [94, 73, 91,",
"68, 80, 61] (exceptions include two 2011 contests specialised on transfer learning [35]—but compare [23]).",
"In particular, as of 2013, state-of-the-art feedforward nets (TL 1.12-1.13) are GPU-based [18] multi-",
"column[22] combinations of two ancient concepts:Backpropagation (TL 1.2) applied [55] to Neocognitron-",
"like convolutional architectures (TL 1.3) (with max-pooling layers [72, 71, 77] instead of alternative [29,",
"30, 31, 78] local winner-take-all methods). (Plus additional tricks from the 1990s and 2000s, e.g. [65, 64].)",
"In the quite different deep recurrent case, supervised systems also dominate, e.g., [48, 36, 40, 37, 61, 38]",
"(TL 1.10, TL 1.14),",
"In particular, most competition-winning or benchmark record-setting Deep Learners (TL 1.10 - TL",
"1.14) use one of two supervised techniques developed in my lab: (1) recurrent LSTM (1997) (TL 1.8)",
"trained by CTC (2006) [36], or (2) feedforward GPU-MPCNN (2011) [18] (TL 1.12) (building on earlier",
"work since the 1960s mentioned in the text above). Nevertheless, in many applications it can still be",
"advantageous to combine the best of both worlds - supervised learning and unsupervised pre-training, like",
"in my 1991 system described above [79, 81, 82].",
"3"
] |
{
"arxiv_id": "1312.5548",
"language": "en",
"page_idx": 2
}
|
||
1312.5548-en-3
|
1312.5548-en-3.jpg
|
1 $\textbf{Timeline of Deep Learning Highlights}$
1.1 1962{:} Neurobiological Inspiration Through Simple Cells and Complex Cells
Hubel and Wiesel described simple cells and complex cells in the visual cortex [49]. This inspired later deep artificial neural network (NN) architectures (TL 1.3) used in certain modern award-winning Deep Learners (TL 1.12-1.14). (The author of the present paper was conceived in 1962.)
1.2 $1970\pm$ a Decade or so: Backpropagation
Error functions and their gradients for complex, nonlinear, multi-stage, differentiable, NN-related systems have been discussed at least since the early 1960s, e.g..[41,50,13,27,12,93,4,26]. Gradient descent[42] in such systems can be performed [13,50,12] by iterating the ancient chain rule [57,58] in dynamic programming style[9] (compare simplified derivation using chain rule only [27]). However, efficient error backpropagation (BP) in arbitrary, possibly sparse, NN-like networks apparently was fırst described by Linnainmaa in 1970 [59, 60] (he did not refer to NN though). BP is also known as the reverse mode of automatic differentiation [41], where the costs of forward activation spreading essentially equal the costs of backward derivative calculation. See early FORTRAN code[59], and compare[66]. Compare the concept of ordered derivative [89] and related work [28], with NN-specific discussion [89] (section 5.5.1), and the first NN-specific efficient BP of 1981 by Werbos [90,92]. Compare [53,75,67] and generalisations for sequence-processing recurrent NN,e.g.,[94,73,91,68,80,61]. See also natural gradients[5]. As of 2013, BP is still the central Deep Learning algorithm.
1.3 $1979: \textbf{ Deep Neocognitron, Weight Sharing, Convolution}$
Fukushima’s deep Neocognitron architecture [29,30,31] incorporated neurophysiological insights (TL 1.1) [49]. It introduced weight-sharing Convolutional Neural Networks (CNN) as well as winner-take-all layers. It is very similar to the architecture of modern, competition-winning, purely $supervised$, feedforward, gradient-based Deep Learners (TL 1.12-1.14). Fukushima, however, used local $un$supervised learning rules instead.
1.4 1987{:} Autoencoder Hierarchies
In 1987, Ballard published ideas on unsupervised autoencoder hierarchies [7], related to post-2000 feedforward Deep Learners (TL 1.9) based on unsupervised pre-training, e.g.,[44]; compare survey [45] and somewhat related RAAMs [69].
1.5 1989: Backpropagation for CNN
LeCun et al. [54,55] applied backpropagation (TL 1.2) to Fukushima's weight-sharing convolutional neural layers (TL 1.3) [29, 30, 54]. This combination has become an essential ingredient of many modern, competition-winning, feedforward, visual Deep Learners (TL 1.12-1.13).
1.6 1991: Fundamental Deep Learning Problem
By the early 1990s, experiments had shown that deep feedforward or recurrent networks are hard to train by backpropagation (TL 1.2). My student Hochreiter[46] discovered and analyzed the reason, namely, the Fundamental Deep Learning Problem due to vanishing or exploding gradients. Compare [47].
1.7 1991{:} Deep Hierarchy of Recurrent NN
My first recurrent Deep Learning system (present paper) partially overcame the fundamental problem (TL 1.6) through a deep RNN stack pre-trained in unsupervised fashion [79, 81, 82] to accelerate subsequent supervised learning. This was a working Deep Learner in the modern post-2000 sense, and also the first Neural Hierarchical Temporal Memory.
4
|
[
[
0.1478271484375,
0.1395263671875,
0.158447265625,
0.1395263671875,
0.158447265625,
0.1483154296875,
0.1478271484375,
0.1483154296875
],
[
0.176513671875,
0.1351318359375,
0.560546875,
0.1363525390625,
0.560546875,
0.154052734375,
0.176513671875,
0.15283203125
],
[
0.1446533203125,
0.1685791015625,
0.82177734375,
0.1685791015625,
0.82177734375,
0.18310546875,
0.1446533203125,
0.18310546875
],
[
0.1446533203125,
0.1932373046875,
0.83837890625,
0.1932373046875,
0.83837890625,
0.20703125,
0.1446533203125,
0.20703125
],
[
0.1446533203125,
0.208984375,
0.83837890625,
0.208984375,
0.83837890625,
0.222900390625,
0.1446533203125,
0.222900390625
],
[
0.1446533203125,
0.2235107421875,
0.68701171875,
0.2235107421875,
0.68701171875,
0.238037109375,
0.1446533203125,
0.238037109375
],
[
0.143798828125,
0.255126953125,
0.53369140625,
0.255615234375,
0.53369140625,
0.2734375,
0.143798828125,
0.272705078125
],
[
0.146240234375,
0.2822265625,
0.83642578125,
0.2822265625,
0.83642578125,
0.29541015625,
0.146240234375,
0.29541015625
],
[
0.1446533203125,
0.29736328125,
0.83837890625,
0.29736328125,
0.83837890625,
0.3125,
0.1446533203125,
0.3125
],
[
0.146240234375,
0.313232421875,
0.8388671875,
0.313232421875,
0.8388671875,
0.326904296875,
0.146240234375,
0.326904296875
],
[
0.1446533203125,
0.32763671875,
0.83837890625,
0.326904296875,
0.83837890625,
0.341552734375,
0.1446533203125,
0.34228515625
],
[
0.1446533203125,
0.343505859375,
0.83837890625,
0.343505859375,
0.83837890625,
0.356689453125,
0.1446533203125,
0.356689453125
],
[
0.146240234375,
0.356689453125,
0.83837890625,
0.356689453125,
0.83837890625,
0.37060546875,
0.146240234375,
0.37060546875
],
[
0.146240234375,
0.373779296875,
0.8388671875,
0.373779296875,
0.8388671875,
0.386962890625,
0.146240234375,
0.386962890625
],
[
0.1446533203125,
0.386962890625,
0.83837890625,
0.386962890625,
0.83837890625,
0.40087890625,
0.1446533203125,
0.40087890625
],
[
0.146240234375,
0.404052734375,
0.8388671875,
0.404052734375,
0.8388671875,
0.417236328125,
0.146240234375,
0.417236328125
],
[
0.1446533203125,
0.41845703125,
0.83837890625,
0.41845703125,
0.83837890625,
0.431884765625,
0.1446533203125,
0.431884765625
],
[
0.146240234375,
0.434326171875,
0.8388671875,
0.434326171875,
0.8388671875,
0.447509765625,
0.146240234375,
0.447509765625
],
[
0.1446533203125,
0.449462890625,
0.45751953125,
0.449462890625,
0.45751953125,
0.46337890625,
0.1446533203125,
0.46337890625
],
[
0.146240234375,
0.48291015625,
0.654296875,
0.48291015625,
0.654296875,
0.49755859375,
0.146240234375,
0.49755859375
],
[
0.146240234375,
0.50830078125,
0.83837890625,
0.50830078125,
0.83837890625,
0.521484375,
0.146240234375,
0.521484375
],
[
0.1478271484375,
0.5244140625,
0.833984375,
0.5244140625,
0.833984375,
0.53515625,
0.1478271484375,
0.53515625
],
[
0.146240234375,
0.53857421875,
0.83837890625,
0.53857421875,
0.83837890625,
0.5517578125,
0.146240234375,
0.5517578125
],
[
0.146240234375,
0.5537109375,
0.8388671875,
0.5537109375,
0.8388671875,
0.5673828125,
0.146240234375,
0.5673828125
],
[
0.1446533203125,
0.5693359375,
0.19775390625,
0.5693359375,
0.19775390625,
0.58056640625,
0.1446533203125,
0.58056640625
],
[
0.146240234375,
0.60107421875,
0.450927734375,
0.60107421875,
0.450927734375,
0.61572265625,
0.146240234375,
0.61572265625
],
[
0.143798828125,
0.62548828125,
0.83837890625,
0.625,
0.83837890625,
0.6396484375,
0.143798828125,
0.64013671875
],
[
0.1446533203125,
0.642578125,
0.8388671875,
0.642578125,
0.8388671875,
0.6572265625,
0.1446533203125,
0.6572265625
],
[
0.1446533203125,
0.65771484375,
0.35693359375,
0.65771484375,
0.35693359375,
0.671875,
0.1446533203125,
0.671875
],
[
0.143798828125,
0.68798828125,
0.464111328125,
0.689453125,
0.464111328125,
0.70703125,
0.143798828125,
0.70556640625
],
[
0.1446533203125,
0.71533203125,
0.8388671875,
0.7138671875,
0.8388671875,
0.72900390625,
0.1446533203125,
0.72998046875
],
[
0.1446533203125,
0.73095703125,
0.83837890625,
0.73095703125,
0.83837890625,
0.744140625,
0.1446533203125,
0.744140625
],
[
0.1446533203125,
0.7470703125,
0.62353515625,
0.7470703125,
0.62353515625,
0.76025390625,
0.1446533203125,
0.76025390625
],
[
0.143798828125,
0.77783203125,
0.560546875,
0.77880859375,
0.560546875,
0.796875,
0.143798828125,
0.79541015625
],
[
0.146240234375,
0.8046875,
0.83837890625,
0.8046875,
0.83837890625,
0.818359375,
0.146240234375,
0.818359375
],
[
0.1446533203125,
0.82080078125,
0.8388671875,
0.82080078125,
0.8388671875,
0.833984375,
0.1446533203125,
0.833984375
],
[
0.1446533203125,
0.8359375,
0.7607421875,
0.8359375,
0.7607421875,
0.849609375,
0.1446533203125,
0.849609375
],
[
0.1422119140625,
0.8671875,
0.51953125,
0.8662109375,
0.51953125,
0.8837890625,
0.1422119140625,
0.88525390625
],
[
0.1446533203125,
0.89306640625,
0.83837890625,
0.89306640625,
0.83837890625,
0.90771484375,
0.1446533203125,
0.90771484375
],
[
0.146240234375,
0.9091796875,
0.83837890625,
0.9091796875,
0.83837890625,
0.92236328125,
0.146240234375,
0.92236328125
],
[
0.146240234375,
0.9248046875,
0.8388671875,
0.9248046875,
0.8388671875,
0.93798828125,
0.146240234375,
0.93798828125
],
[
0.143798828125,
0.93798828125,
0.408447265625,
0.939453125,
0.408447265625,
0.9541015625,
0.143798828125,
0.95263671875
],
[
0.491943359375,
0.974609375,
0.50341796875,
0.984375,
0.491943359375,
0.99365234375,
0.4794921875,
0.984375
]
] |
[
"1",
" Timeline of Deep Learning Highlights",
"1.1 1962: Neurobiological Inspiration Through Simple Cells and Complex Cells",
"Hubel and Wiesel described simple cells and complex cells in the visual cortex [49]. This inspired later",
"deep artificial neural network (NN) architectures (TL 1.3) used in certain modern award-winning Deep",
"Learners (TL 1.12-1.14). (The author of the present paper was conceived in 1962.)",
"1.2 1970 ± a Decade or so: Backpropagation",
"Error functions and their gradients for complex, nonlinear, multi-stage, differentiable, NN-related systems",
"have been discussed at least since the early 1960s, e.g., [41, 50, 13, 27, 12, 93, 4, 26]. Gradient descent [42]",
"in such systems can be performed [13, 50, 12] by iterating the ancient chain rule [57, 58] in dynamic",
"programming style [9] (compare simplified derivation using chain rule only [27]). However, efficient error",
"backpropagation (BP) in arbitrary, possibly sparse, NN-like networks apparently was first described by",
"Linnainmaa in 1970 [59, 60] (he did not refer to NN though). BP is also known as the reverse mode of",
"automatic differentiation [41], where the costs of forward activation spreading essentially equal the costs of",
"backward derivative calculation. See early FORTRAN code [59], and compare [66]. Compare the concept",
"of ordered derivative [89] and related work [28], with NN-specific discussion [89] (section 5.5.1), and the",
"first NN-specific efficient BP of 1981 by Werbos [90, 92]. Compare [53, 75, 67] and generalisations for",
"sequence-processing recurrent NN, e.g., [94, 73, 91, 68, 80, 61]. See also natural gradients [5]. As of 2013,",
"BP is still the central Deep Learning algorithm.",
"1.31979: Deep Neocognitron, Weight Sharing, Convolution",
"Fukushima's deep Neocognitron architecture [29, 30,31]incorporated neurophysiological insights (TL 1.1)",
"[49].Itintroduced weight-sharing Convolutional Neural Networks (CNN) as well as winner-take-all lay",
"ers. It is very similar to the architecture of modern, competition-winning,purely supervised, feedforward,",
"gradient-based Deep Learners (TL 1.12-1.14). Fukushima, however, used local unsupervised learning rules",
"instead.",
"1.41987:Autoencoder Hierarchies",
"In 1987, Ballard published ideas on unsupervised autoencoder hierarchies [7], related to post-2000 feed-",
"forward Deep Learners (TL 1.9) based on unsupervised pre-training, e.g., [44]; compare survey [45] and",
"somewhat related RAAMs [69].",
"1.51989: Backpropagation for CNN",
"LeCun et al. [54, 55] applied backpropagation (TL 1.2) to Fukushima's weight-sharing convolutional",
"neural layers (TL 1.3) [29, 30, 54]. This combination has become an essential ingredient of many modern,",
"competition-winning, feedforward, visual Deep Learners (TL 1.12-1.13).",
"1.61991: Fundamental Deep Learning Problem",
"By the early 199Os, experiments had shown that deep feedforward or recurrent networks are hard to train",
"by backpropagation (TL 1.2). My student Hochreiter [46] discovered and analyzed the reason, namely, the",
"Fundamental Deep Learning Problem due to vanishing or exploding gradients. Compare [47].",
"1.71991: Deep Hierarchy of Recurrent NN",
"My first recurrent Deep Learning system (present paper) partially overcame the fundamental problem (TL",
"1.6) through a deep RNN stack pre-trained in unsupervised fashion [79, 81, 82] to accelerate subsequent",
"supervised learning. This was a working Deep Learner in the modern post-200O sense, and also the first",
"Neural Hierarchical Temporal Memory.",
"X"
] |
{
"arxiv_id": "1312.5548",
"language": "en",
"page_idx": 3
}
|
||
1312.5548-en-4
|
1312.5548-en-4.jpg
|
1.8 1997{:} Supervised Deep Learner (LSTM)
Long Short-Term Memory (LSTM) recurrent neural networks (RNN) became the first purely supervised Deep Learners, e.g., [48, 33, 39, 36, 37, 40, 38]. LSTM RNN were able to learn solutions to many previously unlearnable problems (see also TL 1.10, TL 1.14).
1.9 2006: Deep Belief Networks / CNN Results
A paper by Hinton and Salakhutdinov [44] focused on unsupervised pre-training of feedforward NN to accelerate subsequent supervised learning (compare TL 1.7). This helped to arouse interest in deep NN (keywords: restricted Boltzmann machines; Deep Belief Networks). In the same year, a BP-trained CNN (TL 1.3, TL 1.5) by Ranzato et al. [70] set a new record on the famous MNIST handwritten digit recognition benchmark[54], using training pattern deformations [6,86].
1.10 2009: First Competitions Won by Deep Learning
2009 saw the first Deep Learning systems to win official international pattern recognition contests (with secret test sets known only to the organisers): three connected handwriting competitions at ICDAR 2009 were won by deep LSTM RNN [40, 83] performing simultaneous segmentation and recognition
1.11 2010{:} Plain Backpropagation on GPUs Yields Excellent Results
In 2010, a new MNIST record was set by good old backpropagation (TL 1.2) in deep but otherwise standard NN, without unsupervised pre-training, and without convolution (but with training pattern deformations). This was made possible mainly by boosting computing power through a fast GPU implementation [17]. (A year later, first human-competitive performance on MNIST was achieved by a deep MCMPCNN (TL 1.12) [22].)
1.12 2011: MPCNN on GPU / First Superhuman Visual Pattern Recognition
In 2011, Ciresan $etal.$ introduced supervised $GPU-based$ Max-Pooling CNN or Convnets (MPCNN)[18], today used by most if not all feedforward competition-winning deep NN (TL 1.13, TL 1.14). The first superhuman visual pattern recognition in a controlled competition (traffic signs [87]) was achieved [20,19] (twice better than humans, three times better than the closest artificial NN competitor, six times better than the best non-neural method), through deep and wide Multi-Column (MC) GPU-MPCNN[18,19], the current gold standard for deep feedforward NN.
1.13 2012{:} First Contests Won on Object Detection and Image Segmentation
2012 saw the first Deep learning system (a GPU-MCMPCNN [18,19], TL 1.12) to win a contest on visual object detection in large images (as opposed to mere recognition/classification): the ICPR 2012 Contest on Mitosis Detection in Breast Cancer Histological Images [2,74,16]. An MC (TL 1.12) variant of a GPU-MPCNN also achieved best results on the ImageNet classification benchmark[51]. 2012 also saw the first pure image segmentation contest won by Deep Learning (again through a GPU-MCMPCNN), namely, the ISBI 2012 Challenge on segmenting neuronal structures [3,15]. This was the 8th international pattern recognition contest won by my team since 2009 [1].
1.14 2013: More Contests and Benchmark Records
In 2013, a new TIMITphonemerecognition record was set by deep LSTM RNN [38](TL 1.8, TL 1.10). A new record[24] on the $ICDAR$ Chinese handwriting recognition benchmark (over 3700 classes) was set on a desktop machine by a GPU-MCMPCNN with almost human performance. The $MICCAI$ 2013 Grand $Challenge\textit{ on Mitosis Detection was won by a GPU- MCMPCNN [ 88, 16] . Deep GPU- MPCNN[ 18] also}$ helped to achieve new best results on ImageNet classification [95] and PASCAL object detection [34].
5
|
[
[
0.146240234375,
0.136962890625,
0.525390625,
0.136962890625,
0.525390625,
0.1514892578125,
0.146240234375,
0.1514892578125
],
[
0.146240234375,
0.1622314453125,
0.8388671875,
0.1622314453125,
0.8388671875,
0.1767578125,
0.146240234375,
0.1767578125
],
[
0.146240234375,
0.177978515625,
0.83837890625,
0.177978515625,
0.83837890625,
0.1925048828125,
0.146240234375,
0.1925048828125
],
[
0.1446533203125,
0.1932373046875,
0.51318359375,
0.1932373046875,
0.51318359375,
0.20703125,
0.1446533203125,
0.20703125
],
[
0.146240234375,
0.2279052734375,
0.5439453125,
0.2279052734375,
0.5439453125,
0.242431640625,
0.146240234375,
0.242431640625
],
[
0.1446533203125,
0.251953125,
0.83837890625,
0.250732421875,
0.83837890625,
0.26513671875,
0.1446533203125,
0.266357421875
],
[
0.1446533203125,
0.268310546875,
0.83837890625,
0.268310546875,
0.83837890625,
0.2822265625,
0.1446533203125,
0.2822265625
],
[
0.146240234375,
0.28271484375,
0.83642578125,
0.28271484375,
0.83642578125,
0.296630859375,
0.146240234375,
0.296630859375
],
[
0.146240234375,
0.298583984375,
0.83837890625,
0.298583984375,
0.83837890625,
0.3125,
0.146240234375,
0.3125
],
[
0.1446533203125,
0.313232421875,
0.5400390625,
0.313232421875,
0.5400390625,
0.32763671875,
0.1446533203125,
0.32763671875
],
[
0.1446533203125,
0.34521484375,
0.609375,
0.3466796875,
0.609375,
0.3642578125,
0.1446533203125,
0.363037109375
],
[
0.1446533203125,
0.37255859375,
0.8388671875,
0.37255859375,
0.8388671875,
0.386474609375,
0.1446533203125,
0.386474609375
],
[
0.146240234375,
0.38818359375,
0.8388671875,
0.38818359375,
0.8388671875,
0.401611328125,
0.146240234375,
0.401611328125
],
[
0.1446533203125,
0.40283203125,
0.77392578125,
0.40283203125,
0.77392578125,
0.416748046875,
0.1446533203125,
0.416748046875
],
[
0.146240234375,
0.4375,
0.72802734375,
0.4375,
0.72802734375,
0.451904296875,
0.146240234375,
0.451904296875
],
[
0.143798828125,
0.460205078125,
0.8388671875,
0.4609375,
0.8388671875,
0.47607421875,
0.143798828125,
0.474853515625
],
[
0.1446533203125,
0.477783203125,
0.83642578125,
0.477783203125,
0.83642578125,
0.4912109375,
0.1446533203125,
0.4912109375
],
[
0.146240234375,
0.492431640625,
0.83837890625,
0.492431640625,
0.83837890625,
0.505859375,
0.146240234375,
0.505859375
],
[
0.1446533203125,
0.50830078125,
0.83837890625,
0.50830078125,
0.83837890625,
0.521484375,
0.1446533203125,
0.521484375
],
[
0.1446533203125,
0.52294921875,
0.1854248046875,
0.52294921875,
0.1854248046875,
0.537109375,
0.1446533203125,
0.537109375
],
[
0.146240234375,
0.556640625,
0.7939453125,
0.556640625,
0.7939453125,
0.5712890625,
0.146240234375,
0.5712890625
],
[
0.146240234375,
0.58056640625,
0.83642578125,
0.58056640625,
0.83642578125,
0.59521484375,
0.146240234375,
0.59521484375
],
[
0.1446533203125,
0.59765625,
0.83837890625,
0.59765625,
0.83837890625,
0.611328125,
0.1446533203125,
0.611328125
],
[
0.146240234375,
0.6123046875,
0.83837890625,
0.611328125,
0.83837890625,
0.62548828125,
0.146240234375,
0.626953125
],
[
0.146240234375,
0.62744140625,
0.83837890625,
0.62744140625,
0.83837890625,
0.6416015625,
0.146240234375,
0.6416015625
],
[
0.143798828125,
0.642578125,
0.8388671875,
0.6416015625,
0.8388671875,
0.65576171875,
0.143798828125,
0.6572265625
],
[
0.1446533203125,
0.6572265625,
0.4599609375,
0.6572265625,
0.4599609375,
0.67041015625,
0.1446533203125,
0.67041015625
],
[
0.146240234375,
0.6923828125,
0.7939453125,
0.6923828125,
0.7939453125,
0.70703125,
0.146240234375,
0.70703125
],
[
0.146240234375,
0.71630859375,
0.8388671875,
0.71630859375,
0.8388671875,
0.73095703125,
0.146240234375,
0.73095703125
],
[
0.146240234375,
0.73291015625,
0.83837890625,
0.73291015625,
0.83837890625,
0.7470703125,
0.146240234375,
0.7470703125
],
[
0.146240234375,
0.7470703125,
0.83837890625,
0.7470703125,
0.83837890625,
0.76025390625,
0.146240234375,
0.76025390625
],
[
0.146240234375,
0.76318359375,
0.83837890625,
0.76318359375,
0.83837890625,
0.7763671875,
0.146240234375,
0.7763671875
],
[
0.1446533203125,
0.7763671875,
0.83837890625,
0.77783203125,
0.83837890625,
0.79248046875,
0.1446533203125,
0.79150390625
],
[
0.1446533203125,
0.79248046875,
0.83837890625,
0.79248046875,
0.83837890625,
0.80615234375,
0.1446533203125,
0.80615234375
],
[
0.1446533203125,
0.80810546875,
0.48779296875,
0.80810546875,
0.48779296875,
0.82275390625,
0.1446533203125,
0.82275390625
],
[
0.146240234375,
0.84228515625,
0.5849609375,
0.84228515625,
0.5849609375,
0.85693359375,
0.146240234375,
0.85693359375
],
[
0.1446533203125,
0.8671875,
0.83642578125,
0.8671875,
0.83642578125,
0.880859375,
0.1446533203125,
0.880859375
],
[
0.146240234375,
0.88330078125,
0.83837890625,
0.88330078125,
0.83837890625,
0.896484375,
0.146240234375,
0.896484375
],
[
0.146240234375,
0.89794921875,
0.83837890625,
0.89794921875,
0.83837890625,
0.9111328125,
0.146240234375,
0.9111328125
],
[
0.146240234375,
0.91357421875,
0.83837890625,
0.91357421875,
0.83837890625,
0.9267578125,
0.146240234375,
0.9267578125
],
[
0.143798828125,
0.92822265625,
0.83837890625,
0.92822265625,
0.83837890625,
0.94140625,
0.143798828125,
0.94140625
],
[
0.483642578125,
0.97607421875,
0.5009765625,
0.97607421875,
0.5009765625,
0.99072265625,
0.483642578125,
0.99072265625
]
] |
[
"1.8 1997: Supervised Deep Learner (LSTM)",
"Long Short-Term Memory (LSTM) recurrent neural networks (RNN) became the first purely supervised",
"Deep Learners, e.g., [48, 33, 39, 36, 37, 40, 38]. LSTM RNN were able to learn solutions to many previ-",
"ously unlearnable problems (see also TL 1.10, TL 1.14).",
"1.92006:DeepBelief Networks/CNNResults",
"A paper by Hinton and Salakhutdinov [44] focused on unsupervised pre-training of feedforward NN to",
"accelerate subsequent supervised learning (compare TL 1.7). This helped to arouse interest in deep NN",
"(keywords: restricted Boltzmann machines; Deep Belief Networks). In the same year, a BP-trained CNN",
"(TL 1.3, TL 1.5) by Ranzato et al. [70] set a new record on the famous MNIST handwritten digit recognition",
"benchmark [54], using training pattern deformations [6, 86].",
"1.10 2009: First Competitions Won by Deep Learning",
"2009 saw the first Deep Learning systems to win official international pattern recognition contests (with",
"secret test sets known only to the organisers): three connected handwriting competitions at ICDAR 2009",
"were won by deep LSTM RNN [40, 83] performing simultaneous segmentation and recognition.",
"1.112010: Plain Backpropagation on GPUs Yields Excellent Results",
"In 2010,a newMNIST record was set bygood old backpropagation (TL1.2)in deep but otherwise standard",
"NN, without unsupervised pre-training, and without convolution (but with training pattern deformations)",
"This was made possible mainly by boosting computing power through a fast GPU implementation [17]. (A",
"year later, first human-competitive performance on MNIST was achieved by a deep MCMPCNN (TL 1.12)",
"[22].]",
"1.122011:MPCNN on GPU/First Superhuman Visual Pattern Recognition",
"In 2011, Ciresan et al. introduced supervised GPU-based Max-Pooling CNN or Convnets (MPCNN) [18].",
"today used by most if not all feedforward competition-winning deep NN (TL 1.13, TL 1.14). The first",
"superhuman visual pattern recognition in a controlled competition (traffic signs [87D) was achieved [20, 19]",
"(twice better than humans, three times better than the closest artificial NN competitor, six times better than",
"the best non-neural method), through deep and wide Multi-Column (MC) GPU-MPCNN [18, 19], the",
"current gold standard for deep feedforward NN.",
"1.132012: First Contests Won on Object Detection and Image Segmentation",
"2012 saw the first Deep learning system (a GPU-MCMPCNN [18, 19], TL 1.12) to win a contest on visual",
"object detection in large images (as opposed to mere recognition/classification): the ICPR 2012 Contest",
"on Mitosis Detection in Breast Cancer Histological Images [2, 74, 16]. An MC (TL 1.12) variant of a",
"GPU-MPCNN also achieved best results on the ImageNet classification benchmark [51]. 2012 also saw the",
"first pure image segmentation contest won by Deep Learning (again through a GPU-MCMPCNN), namely,",
"the ISB1 2012 Challenge on segmenting neuronal structures [3, 15]. This was the 8th international pattern",
"recognition contest won by my team since 2009 [1].",
"1.142013: More Contests and Benchmark Records",
"In 2013, a new TIMIT phoneme recognition record was set by deep LSTM RNN [38] (TL 1.8, TL 1.10),",
"A new record [24] on the ICDAR Chinese handwriting recognition benchmark (over 3700 classes) was set",
"on a desktop machine by a GPU-MCMPCNN with almost human performance. The MICCAI 2013 Grand",
"Challenge on Mitosis Detection was won by a GPU-MCMPCNN [88, 16]. Deep GPU-MPCNN [18] also",
"helped to achieve new best results on ImageNet classification [95] and PASCAL object detection [34].",
"5"
] |
{
"arxiv_id": "1312.5548",
"language": "en",
"page_idx": 4
}
|
||
1606.01933-en-0
|
1606.01933-en-0.jpg
|
A Decomposable Attention Model for Natural Language Inference
$\begin{aligned}&\textbf{nkur P. Parikh}\:&\textbf{Oscar Täckström}\:&\textbf{Dipanjan Das}\:&\text{Jakob Uszkoreit}\\&\text{Google}\:&\text{Google}\:&\text{Google}\:&\text{Google}\\&\text{Vew York, NY}\:&&\text{New York, NY}\:&&\text{Mountain View, CA}\end{aligned}$
$\{$aparikh,oscart,dipanjand,uszkoreit$\}$google.com
### $\mathbf{Abstract}$
We propose a simple neural architecture for natural language inference. Our approach uses attention to decompose the problem into subproblems that can be solved separately, thus making it trivially parallelizable. On the Stanford Natural Language Inference (SNLI) dataset, we obtain state-of-the-art results with almost an order of magnitude fewer parameters than previous work and without relying on any word-order information. Adding intra-sentence attention that takes a minimum amount of order into account yields further improvements.
## 1 Introduction
Natural language inference (NLI) refers to the prob-
lem of determining entailment and contradiction re- is challenging to construct a compact representation lationships between a premise and a hypothesis. NLI that expresses its entire meaning. However, it is fairly is a central problem in language understanding (Katz, easy to conclude that the second sentence follows 1972; Bos and Markert, 2005; van Benthem, 2008; from the first one, by simply aligning Bob with Bob MacCartney and Manning, 2009) and recently the and cannot sleep with awake and recognizing that large SNLI corpus of 570K sentence pairs was cre- these are synonyms. Similarly, one can conclude ated for this task (Bowman et al., 2015). We present that $It$ is sunny outside contradicts the first sentence a new model for NLI and leverage this corpus for
comparison with prior work.
A large body of work based on neural networks
for text similarity tasks including NLI has been pub- more lightweight approach to NLI within a neural lished in recent years (Hu et al., 2014; Rocktäschel framework; with considerably fewer parameters, our et al., 2016; Wang and Jiang, 2016; Yin et al., 2016, model outperforms more complex existing neural ar- $interalia).$ The dominating trend in these models is chitectures. In contrast to existing approaches, our to build complex, deep text representation models, approach only relies on alignment and is fully comfor example, with convolutional networks (LeCun et putationally decomposable with respect to the input al., 1990, CNNs henceforth) or long short-term mem- text. An overview of our approach is given in Figory networks (Hochreiter and Schmidhuber, 1997, ure l. Given two sentences, where each word is repre-
LSTMs henceforth) with the goal of deeper sentence comprehension. While these approaches have yielded impressive results, they are often computationally very expensive, and result in models having millions of parameters (excluding embeddings).
Here, we take a different approach, arguing that for natural language inference it can often suffice to simply align bits of local text substructure and then aggregate this information. For example, consider the following sentences:
$\bullet \textit{Bob is in his room, but because of the thunder}$
and lightning outside, he cannot sleep.
$\bullet \textit{Bob is awake. }$
$\bullet \textit{It is sunny outside. }$
The first sentence is complex in structure and it by aligning thunder and lightning with sunny and recognizing that these are most likely incompatible.
We leverage this intuition to build a simpler and
|
[
[
0.1683349609375,
0.0997314453125,
0.8310546875,
0.0997314453125,
0.8310546875,
0.117431640625,
0.1683349609375,
0.117431640625
],
[
0.1298828125,
0.154052734375,
0.27294921875,
0.1558837890625,
0.27294921875,
0.173583984375,
0.1298828125,
0.1717529296875
],
[
0.325927734375,
0.1578369140625,
0.473876953125,
0.1578369140625,
0.473876953125,
0.17236328125,
0.325927734375,
0.17236328125
],
[
0.54150390625,
0.1578369140625,
0.65869140625,
0.1578369140625,
0.65869140625,
0.17236328125,
0.54150390625,
0.17236328125
],
[
0.73046875,
0.1578369140625,
0.869140625,
0.1578369140625,
0.869140625,
0.17236328125,
0.73046875,
0.17236328125
],
[
0.17236328125,
0.175537109375,
0.233642578125,
0.175537109375,
0.233642578125,
0.1912841796875,
0.17236328125,
0.1912841796875
],
[
0.369384765625,
0.175537109375,
0.4306640625,
0.175537109375,
0.4306640625,
0.1900634765625,
0.369384765625,
0.1900634765625
],
[
0.56884765625,
0.175537109375,
0.63134765625,
0.175537109375,
0.63134765625,
0.1900634765625,
0.56884765625,
0.1900634765625
],
[
0.76953125,
0.175537109375,
0.830078125,
0.175537109375,
0.830078125,
0.1912841796875,
0.76953125,
0.1912841796875
],
[
0.1405029296875,
0.1932373046875,
0.2607421875,
0.1932373046875,
0.2607421875,
0.2083740234375,
0.1405029296875,
0.2083740234375
],
[
0.339111328125,
0.1932373046875,
0.459228515625,
0.1932373046875,
0.459228515625,
0.2083740234375,
0.339111328125,
0.2083740234375
],
[
0.5390625,
0.1932373046875,
0.65869140625,
0.1932373046875,
0.65869140625,
0.2083740234375,
0.5390625,
0.2083740234375
],
[
0.7197265625,
0.1932373046875,
0.87841796875,
0.1932373046875,
0.87841796875,
0.2083740234375,
0.7197265625,
0.2083740234375
],
[
0.2279052734375,
0.23291015625,
0.77294921875,
0.23291015625,
0.77294921875,
0.2474365234375,
0.2279052734375,
0.2474365234375
],
[
0.021240234375,
0.261962890625,
0.06451416015625,
0.261962890625,
0.06451416015625,
0.40283203125,
0.021240234375,
0.40283203125
],
[
0.2646484375,
0.317626953125,
0.34228515625,
0.317626953125,
0.34228515625,
0.335205078125,
0.2646484375,
0.335205078125
],
[
0.5107421875,
0.32080078125,
0.8837890625,
0.32080078125,
0.8837890625,
0.333984375,
0.5107421875,
0.333984375
],
[
0.5087890625,
0.337158203125,
0.88232421875,
0.337158203125,
0.88232421875,
0.3515625,
0.5087890625,
0.3515625
],
[
0.1519775390625,
0.352783203125,
0.453369140625,
0.352783203125,
0.453369140625,
0.3662109375,
0.1519775390625,
0.3662109375
],
[
0.5087890625,
0.354736328125,
0.88232421875,
0.354736328125,
0.88232421875,
0.369384765625,
0.5087890625,
0.369384765625
],
[
0.1519775390625,
0.36865234375,
0.455078125,
0.36865234375,
0.455078125,
0.3818359375,
0.1519775390625,
0.3818359375
],
[
0.5107421875,
0.371826171875,
0.88232421875,
0.371826171875,
0.88232421875,
0.386474609375,
0.5107421875,
0.386474609375
],
[
0.1519775390625,
0.3837890625,
0.455078125,
0.3837890625,
0.455078125,
0.397705078125,
0.1519775390625,
0.397705078125
],
[
0.5087890625,
0.389404296875,
0.85302734375,
0.389404296875,
0.85302734375,
0.40283203125,
0.5087890625,
0.40283203125
],
[
0.1502685546875,
0.3984375,
0.453369140625,
0.3984375,
0.453369140625,
0.412353515625,
0.1502685546875,
0.412353515625
],
[
0.02532958984375,
0.40478515625,
0.05963134765625,
0.40478515625,
0.05963134765625,
0.48486328125,
0.02532958984375,
0.48486328125
],
[
0.52685546875,
0.406005859375,
0.88134765625,
0.406005859375,
0.88134765625,
0.42041015625,
0.52685546875,
0.42041015625
],
[
0.1519775390625,
0.41357421875,
0.453369140625,
0.41357421875,
0.453369140625,
0.4267578125,
0.1519775390625,
0.4267578125
],
[
0.15283203125,
0.427978515625,
0.451904296875,
0.427978515625,
0.451904296875,
0.442626953125,
0.15283203125,
0.442626953125
],
[
0.5107421875,
0.423583984375,
0.88134765625,
0.423583984375,
0.88134765625,
0.4375,
0.5107421875,
0.4375
],
[
0.1519775390625,
0.443115234375,
0.453369140625,
0.443115234375,
0.453369140625,
0.45703125,
0.1519775390625,
0.45703125
],
[
0.50732421875,
0.43994140625,
0.88232421875,
0.439453125,
0.88232421875,
0.453857421875,
0.50732421875,
0.455078125
],
[
0.1519775390625,
0.458984375,
0.451904296875,
0.458984375,
0.451904296875,
0.472900390625,
0.1519775390625,
0.472900390625
],
[
0.5087890625,
0.458251953125,
0.88232421875,
0.45703125,
0.88232421875,
0.4716796875,
0.5087890625,
0.472900390625
],
[
0.1502685546875,
0.472900390625,
0.455078125,
0.473388671875,
0.455078125,
0.488037109375,
0.1502685546875,
0.4873046875
],
[
0.5087890625,
0.473388671875,
0.685546875,
0.474853515625,
0.685546875,
0.4892578125,
0.5087890625,
0.488037109375
],
[
0.0196075439453125,
0.488037109375,
0.058013916015625,
0.488037109375,
0.060455322265625,
0.7119140625,
0.0228729248046875,
0.7119140625
],
[
0.1502685546875,
0.4892578125,
0.455078125,
0.4892578125,
0.455078125,
0.50244140625,
0.1502685546875,
0.50244140625
],
[
0.1502685546875,
0.5048828125,
0.455078125,
0.5048828125,
0.455078125,
0.51806640625,
0.1502685546875,
0.51806640625
],
[
0.529296875,
0.501953125,
0.88134765625,
0.501953125,
0.88134765625,
0.51513671875,
0.529296875,
0.51513671875
],
[
0.1502685546875,
0.51953125,
0.340576171875,
0.51953125,
0.340576171875,
0.53271484375,
0.1502685546875,
0.53271484375
],
[
0.5439453125,
0.51953125,
0.82177734375,
0.51953125,
0.82177734375,
0.53271484375,
0.5439453125,
0.53271484375
],
[
0.52783203125,
0.54052734375,
0.646484375,
0.54052734375,
0.646484375,
0.5546875,
0.52783203125,
0.5546875
],
[
0.11517333984375,
0.564453125,
0.25244140625,
0.564453125,
0.25244140625,
0.5791015625,
0.11517333984375,
0.5791015625
],
[
0.529296875,
0.5625,
0.6787109375,
0.5625,
0.6787109375,
0.5771484375,
0.529296875,
0.5771484375
],
[
0.1168212890625,
0.5888671875,
0.490234375,
0.5888671875,
0.490234375,
0.60302734375,
0.1168212890625,
0.60302734375
],
[
0.52685546875,
0.5888671875,
0.88232421875,
0.5888671875,
0.88232421875,
0.60302734375,
0.52685546875,
0.60302734375
],
[
0.11517333984375,
0.60693359375,
0.494384765625,
0.60693359375,
0.494384765625,
0.62060546875,
0.11517333984375,
0.62060546875
],
[
0.5087890625,
0.60693359375,
0.88232421875,
0.60693359375,
0.88232421875,
0.62060546875,
0.5087890625,
0.62060546875
],
[
0.11517333984375,
0.62353515625,
0.491943359375,
0.62353515625,
0.491943359375,
0.63818359375,
0.11517333984375,
0.63818359375
],
[
0.5087890625,
0.62353515625,
0.88134765625,
0.62353515625,
0.88134765625,
0.63720703125,
0.5087890625,
0.63720703125
],
[
0.1168212890625,
0.64013671875,
0.49267578125,
0.64013671875,
0.49267578125,
0.65478515625,
0.1168212890625,
0.65478515625
],
[
0.50732421875,
0.6416015625,
0.88134765625,
0.64013671875,
0.88134765625,
0.65478515625,
0.50732421875,
0.65576171875
],
[
0.11767578125,
0.6572265625,
0.491943359375,
0.6572265625,
0.491943359375,
0.671875,
0.11767578125,
0.671875
],
[
0.50732421875,
0.65771484375,
0.88232421875,
0.6572265625,
0.88232421875,
0.671875,
0.50732421875,
0.67236328125
],
[
0.1168212890625,
0.6748046875,
0.491943359375,
0.6748046875,
0.491943359375,
0.68798828125,
0.1168212890625,
0.68798828125
],
[
0.50732421875,
0.67529296875,
0.88232421875,
0.67529296875,
0.88232421875,
0.689453125,
0.50732421875,
0.689453125
],
[
0.11517333984375,
0.68994140625,
0.49267578125,
0.69140625,
0.49267578125,
0.70556640625,
0.11517333984375,
0.705078125
],
[
0.5087890625,
0.6923828125,
0.88134765625,
0.6923828125,
0.88134765625,
0.70703125,
0.5087890625,
0.70703125
],
[
0.11517333984375,
0.708984375,
0.490234375,
0.708984375,
0.490234375,
0.72265625,
0.11517333984375,
0.72265625
],
[
0.50732421875,
0.70849609375,
0.88134765625,
0.70849609375,
0.88134765625,
0.72265625,
0.50732421875,
0.72265625
],
[
0.11517333984375,
0.7265625,
0.491943359375,
0.7265625,
0.491943359375,
0.74072265625,
0.11517333984375,
0.74072265625
],
[
0.5087890625,
0.7265625,
0.88134765625,
0.7265625,
0.88134765625,
0.74072265625,
0.5087890625,
0.74072265625
],
[
0.11358642578125,
0.74365234375,
0.3203125,
0.7421875,
0.3203125,
0.7568359375,
0.11358642578125,
0.75830078125
],
[
0.5087890625,
0.74365234375,
0.87744140625,
0.74365234375,
0.87744140625,
0.7568359375,
0.5087890625,
0.7568359375
],
[
0.134033203125,
0.76025390625,
0.490234375,
0.76025390625,
0.490234375,
0.7744140625,
0.134033203125,
0.7744140625
],
[
0.52783203125,
0.7587890625,
0.88134765625,
0.7587890625,
0.88134765625,
0.7734375,
0.52783203125,
0.7734375
],
[
0.1168212890625,
0.77783203125,
0.491943359375,
0.77783203125,
0.491943359375,
0.79150390625,
0.1168212890625,
0.79150390625
],
[
0.5087890625,
0.77783203125,
0.88134765625,
0.77783203125,
0.88134765625,
0.79248046875,
0.5087890625,
0.79248046875
],
[
0.1168212890625,
0.79443359375,
0.491943359375,
0.79443359375,
0.491943359375,
0.8095703125,
0.1168212890625,
0.8095703125
],
[
0.50732421875,
0.79345703125,
0.88232421875,
0.79541015625,
0.88232421875,
0.8095703125,
0.50732421875,
0.806640625
],
[
0.1168212890625,
0.8125,
0.494384765625,
0.8125,
0.494384765625,
0.8271484375,
0.1168212890625,
0.8271484375
],
[
0.50634765625,
0.81298828125,
0.8857421875,
0.81298828125,
0.8857421875,
0.8271484375,
0.50634765625,
0.8271484375
],
[
0.11517333984375,
0.8291015625,
0.491943359375,
0.8291015625,
0.491943359375,
0.84326171875,
0.11517333984375,
0.84326171875
],
[
0.5087890625,
0.830078125,
0.8837890625,
0.830078125,
0.8837890625,
0.8447265625,
0.5087890625,
0.8447265625
],
[
0.11358642578125,
0.84521484375,
0.494384765625,
0.8466796875,
0.494384765625,
0.861328125,
0.11358642578125,
0.85986328125
],
[
0.50634765625,
0.84521484375,
0.8837890625,
0.84326171875,
0.8837890625,
0.861328125,
0.50634765625,
0.86279296875
],
[
0.11517333984375,
0.86279296875,
0.490234375,
0.86279296875,
0.490234375,
0.87744140625,
0.11517333984375,
0.87744140625
],
[
0.5087890625,
0.8642578125,
0.88232421875,
0.8642578125,
0.88232421875,
0.87744140625,
0.5087890625,
0.87744140625
],
[
0.11517333984375,
0.8798828125,
0.494384765625,
0.8798828125,
0.494384765625,
0.89453125,
0.11517333984375,
0.89453125
],
[
0.50732421875,
0.880859375,
0.88232421875,
0.880859375,
0.88232421875,
0.89599609375,
0.50732421875,
0.89599609375
],
[
0.11358642578125,
0.89794921875,
0.494384765625,
0.896484375,
0.494384765625,
0.9111328125,
0.11358642578125,
0.912109375
],
[
0.50732421875,
0.89794921875,
0.8837890625,
0.89794921875,
0.8837890625,
0.912109375,
0.50732421875,
0.912109375
]
] |
[
"A Decomposable Attention Model for Natural Language Inference",
"Ankur P. Parikh",
"Oscar Tackstrom",
"Dipanjan Das",
"Jakob Uszkoreit",
"Google",
"Google",
"Google",
"Google",
"New York, NY",
"New York, NY",
"New York, NY",
"Mountain View, CA",
"faparikh,oscart,dipanjand,uszkoreitl@google.com",
"910z des 5",
"Abstract",
"LSTMs henceforth) with the goal of deeper sen-",
"tence comprehension. While these approaches have",
"We propose a simple neural architecture for nat-",
"yielded impressive results, they are often computa-",
"ural language inference. Our approach uses at-",
"tionally very expensive, and result in models having",
"tention to decompose the problem into subprob-",
"millions of parameters (excluding embeddings).",
"lems that can be solved separately, thus making",
"['S]",
"Here, we take a different approach, arguing that",
"it trivially parallelizable. On the Stanford Natu-",
"ral Language Inference (SNLI) dataset, we ob.",
"for natural language inference it can often suffice to",
"tain state-of-the-art results with almost an order",
"simply align bits of local text substructure and then",
"of magnitude fewer parameters than previous",
"aggregate this information. For example, consider",
"work and without relying on any word-order in-",
"the following sentences:",
"00609091",
"formation. Adding intra-sentence attention that",
"takes a minimum amount of order into account",
" Bob is in his room, but because of the thunder",
"yields further improvements.",
"and lightning outside, he cannot sleep.",
". Bob is awake.",
"1 Introduction",
"It is sunny outside.",
"Natural language inference (NLD) refers to the prob-",
"The first sentence is complex in structure and it",
"lem of determining entailment and contradiction re-",
"is challenging to construct a compact representation",
"lationships between a premise and a hypothesis. NLI",
"that expresses its entire meaning. However, it is fairly",
"is a central problem in language understanding (Katz,",
"easy to conclude that the second sentence follows",
"1972; Bos and Markert, 2005; van Benthem, 2008;",
"from the first one, by simply aligning Bob with Bob",
"MacCartney and Manning, 2009) and recently the",
"and cannot sleep with awake and recognizing that",
"large SNLI corpus of 570K sentence pairs was cre-",
"these are synonyms. Similarly, one can conclude",
"ated for this task (Bowman et al., 2015). We present",
"that It is sunny outside contradicts the first sentence",
"a new model for NLI and leverage this corpus for",
"by aligning thunder and lightning with sunny and",
"comparisonwithpriorwork.",
"recognizing that these are most likely incompatible",
"A large body of work based on neural networks",
"We leverage this intuition to build a simpler and",
"for text similarity tasks including NLI has been pub-",
"more lightweight approach to NLI within a neural",
"lished in recent years (Hu et al., 2014; Rocktaschel",
"framework;withconsiderablyfewerparameters,our",
"et al., 2016; Wang and Jiang, 2016; Yin et al., 2016,",
"model outperforms more complex existing neural ar-",
"inter alia). The dominating trend in these models is",
"chitectures. In contrast to existing approaches, our",
"to build complex, deep text representation models,",
"approach only relies on alignment and is fully com-",
"for example, with convolutional networks (LeCun et",
"putationally decomposable withrespect totheinput",
"al., 1990, CNNs henceforth) or long short-term mem-",
"text. An overview of our approach is given in Fig",
"ory networks (Hochreiter and Schmidhuber, 1997,",
"ure 1. Given two sentences, where each word is repre-"
] |
{
"arxiv_id": "1606.01933",
"language": "en",
"page_idx": 0
}
|
||
1606.01933-en-1
|
1606.01933-en-1.jpg
|
sented by an embedding vector, we first create a soft alignment matrix using neural attention (Bahdanau et al., 2015). We then use the (soft) alignment to decompose the task into subproblems that are solved separately. Finally, the results of these subproblems are merged to produce the final classification. In addition, we optionally apply intra-sentence attention (Cheng et al., 2016) to endow the model with a richer encoding of substructures prior to the alignment step. Asymptotically our approach does the same total work as a vanilla LSTM encoder, while being trivially parallelizable across sentence length, which can allow for considerable speedups in low-latency settings. Empirical results on the SNLI corpus show that our approach achieves state-of-the-art results, while using almost an order of magnitude fewer parameters compared to complex LSTM-based approaches.
## 2 Related Work
Our method is motivated by the central role played by alignment in machine translation (Koehn, 2009) and previous approaches to sentence similarity modeling (Haghighi et al., 2005; Das and Smith, 2009; Chang et al., 2010; Fader et al., 2013), natural language inference (Marsi and Krahmer, 2005; MacCartney et al., 2006; Hickl and Bensley, 2007; MacCartney et al., 2008), and semantic parsing (Andreas et al., 2013). The neural counterpart to alignment, $atten-$ tion (Bahdanau et al., 2015), which is a key part of our approach, was originally proposed and has been predominantly used in conjunction with LSTMs (Rocktäschel et al., 2016; Wang and Jiang, 2016) and to a lesser extent with CNNs (Yin et al., 2016). In contrast, our use of attention is purely based on word Aggregate. Finally, aggregate the sets $\{\mathbf{v}_{1,i}\}_{i=1}^{\ell_a}$
embeddings and our method essentially consists of feed-forward networks that operate largely independently of word order.
# 3 Approach
Let a $=(a_1,\ldots,a_{\ell_{a}})$ and b $=(b_1,\ldots,b_{\ell_{b}})$ be the two input sentences of length $\ell_a$ and $\ell_b$, respectively. We assume that each $a_i,b_j\in\mathbb{R}^d$ is a word embedding vector of dimension $d$ and that each sentence is prepended with a “NULL” This decomposition avoids the quadratic complexity token. Our training data comes in the form of that would be associated with separately applying $F^{\prime}$ labeled pairs $\{ \mathbf{a} ^{( n) }, \mathbf{b} ^{( n) }, \mathbf{y} ^{( n) }\} _{n= 1}^N, \mathrm{~where~y}^{( n) }=$ $\ell _a\times \ell _b$ times. Instead, only $\ell _a+ \ell _b$ applications of $(y_1^{(n)},\ldots,y_C^{(n)})$ is an indicator vector encoding the $F$ are needed. We take $F$ to be a feed-forward neural label and $C$ is the number of output classes. At test network with ReLU activations (Glorot et al., 2011).
<FigureHere>
Figure 1: Pictoral overview of the approach, showing the Attend
(left), Compare (center) and Aggregate (right) steps.
time, we receive a pair of sentences (a,b) and our
goal is to predict the correct label y.
$\textbf{Input representation. }$ Let $\bar{\mathbf{a}}=(\bar{a}_1,\ldots,\bar{a}_{\ell_a})$ and $\bar{\mathbf{b}}=(\bar{b}_{1},\ldots,\bar{b}_{\ell_{b}})$ denote the input representation of each fragment that is fed to subsequent steps of the algorithm. The vanilla version of our model simply defines $\bar{\mathbf{a}}:=\mathbf{a}$ and $\bar{\mathbf{b}}:=\mathbf{b}.$ With this input representation, our model does not make use of word order. However, we discuss an extension using intrasentence attention in Section 3.4 that uses a minimal amount of sequence information.
The core model consists of the following three components (see Figure 1), which are trained jointly: Attend. First, soft-align the elements of ā and b using a variant of neural attention (Bahdanau et al., 2015) and decompose the problem into the comparison of aligned subphrases.
Compare. Second, separately compare each aligned subphrase to produce a set of vectors
$\left\{\mathbf{v}_{1,i}\right\}_{i=1}^{\ell_{a}}$for a and$\left\{\mathbf{v}_{2,j}\right\}_{j=1}^{\ell_{b}}$for b.Each v$_1,i$ is a nonlinear combination of $a_i$ and its (softly) aligned subphrase in b (and analogously for v$_2,j).$
and $\{\mathbf{v}_{2,j}\}_{j=1}^{\ell_{b}}$ from the previous step and use the
result to predict the label $\hat{y}.$
## 3.1 Attend
We first obtain unnormalized attention weights $e_{ij}$, computed by a function $F^{\prime}$, which decomposes as:
$$e_{ij}:=F'(\bar a_i,\bar b_j):=F(\bar a_i)^TF(\bar b_j)\:.$$
(1)
|
[
[
0.1168212890625,
0.091552734375,
0.48779296875,
0.091552734375,
0.48779296875,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.59326171875,
0.0997314453125,
0.61669921875,
0.0997314453125,
0.61669921875,
0.111083984375,
0.59326171875,
0.111083984375
],
[
0.65625,
0.0985107421875,
0.74658203125,
0.0985107421875,
0.74658203125,
0.121826171875,
0.65625,
0.121826171875
],
[
0.1168212890625,
0.111083984375,
0.488525390625,
0.111083984375,
0.488525390625,
0.125,
0.1168212890625,
0.125
],
[
0.1168212890625,
0.1275634765625,
0.488525390625,
0.1275634765625,
0.488525390625,
0.141357421875,
0.1168212890625,
0.141357421875
],
[
0.65625,
0.1331787109375,
0.88134765625,
0.1331787109375,
0.88134765625,
0.15087890625,
0.65625,
0.15087890625
],
[
0.1168212890625,
0.14453125,
0.490234375,
0.14453125,
0.490234375,
0.1578369140625,
0.1168212890625,
0.1578369140625
],
[
0.11358642578125,
0.1610107421875,
0.488525390625,
0.160400390625,
0.488525390625,
0.1749267578125,
0.11358642578125,
0.175537109375
],
[
0.6357421875,
0.1673583984375,
0.74609375,
0.1673583984375,
0.74609375,
0.189453125,
0.6357421875,
0.189453125
],
[
0.11517333984375,
0.1787109375,
0.490234375,
0.1787109375,
0.490234375,
0.1925048828125,
0.11517333984375,
0.1925048828125
],
[
0.62353515625,
0.1787109375,
0.6298828125,
0.1787109375,
0.6298828125,
0.18310546875,
0.62353515625,
0.18310546875
],
[
0.1168212890625,
0.1962890625,
0.490234375,
0.1962890625,
0.490234375,
0.210205078125,
0.1168212890625,
0.210205078125
],
[
0.5107421875,
0.199462890625,
0.8837890625,
0.2008056640625,
0.8837890625,
0.21533203125,
0.5107421875,
0.2139892578125
],
[
0.1168212890625,
0.21337890625,
0.490234375,
0.21337890625,
0.490234375,
0.2266845703125,
0.1168212890625,
0.2266845703125
],
[
0.5087890625,
0.2183837890625,
0.81787109375,
0.2183837890625,
0.81787109375,
0.231689453125,
0.5087890625,
0.231689453125
],
[
0.11358642578125,
0.2298583984375,
0.490234375,
0.2298583984375,
0.490234375,
0.2442626953125,
0.11358642578125,
0.2442626953125
],
[
0.134033203125,
0.2474365234375,
0.490234375,
0.2474365234375,
0.490234375,
0.261962890625,
0.134033203125,
0.261962890625
],
[
0.5107421875,
0.2474365234375,
0.88134765625,
0.2474365234375,
0.88134765625,
0.261962890625,
0.5107421875,
0.261962890625
],
[
0.11517333984375,
0.264404296875,
0.49267578125,
0.264404296875,
0.49267578125,
0.27783203125,
0.11517333984375,
0.27783203125
],
[
0.5087890625,
0.26513671875,
0.771484375,
0.26513671875,
0.771484375,
0.27978515625,
0.5087890625,
0.27978515625
],
[
0.11358642578125,
0.281005859375,
0.490234375,
0.27978515625,
0.490234375,
0.294189453125,
0.11358642578125,
0.29541015625
],
[
0.5087890625,
0.2890625,
0.88232421875,
0.290283203125,
0.88232421875,
0.30810546875,
0.5087890625,
0.306884765625
],
[
0.11517333984375,
0.298583984375,
0.490234375,
0.298583984375,
0.490234375,
0.3125,
0.11517333984375,
0.3125
],
[
0.1168212890625,
0.31494140625,
0.490234375,
0.31494140625,
0.490234375,
0.328857421875,
0.1168212890625,
0.328857421875
],
[
0.5107421875,
0.309326171875,
0.88232421875,
0.309326171875,
0.88232421875,
0.323974609375,
0.5107421875,
0.323974609375
],
[
0.5107421875,
0.326904296875,
0.88134765625,
0.326904296875,
0.88134765625,
0.34033203125,
0.5107421875,
0.34033203125
],
[
0.11358642578125,
0.333251953125,
0.490234375,
0.333251953125,
0.490234375,
0.3466796875,
0.11358642578125,
0.3466796875
],
[
0.51123046875,
0.343505859375,
0.88134765625,
0.343505859375,
0.88134765625,
0.356689453125,
0.51123046875,
0.356689453125
],
[
0.11517333984375,
0.35107421875,
0.490234375,
0.35107421875,
0.490234375,
0.3642578125,
0.11517333984375,
0.3642578125
],
[
0.5107421875,
0.359130859375,
0.88232421875,
0.35986328125,
0.88232421875,
0.375,
0.5107421875,
0.373779296875
],
[
0.11273193359375,
0.3662109375,
0.45751953125,
0.367431640625,
0.45751953125,
0.3818359375,
0.11273193359375,
0.380615234375
],
[
0.5107421875,
0.37744140625,
0.88232421875,
0.37744140625,
0.88232421875,
0.392578125,
0.5107421875,
0.392578125
],
[
0.11517333984375,
0.392578125,
0.2646484375,
0.392578125,
0.2646484375,
0.4072265625,
0.11517333984375,
0.4072265625
],
[
0.5107421875,
0.395751953125,
0.8837890625,
0.395751953125,
0.8837890625,
0.4091796875,
0.5107421875,
0.4091796875
],
[
0.11517333984375,
0.417236328125,
0.488525390625,
0.417236328125,
0.488525390625,
0.43115234375,
0.11517333984375,
0.43115234375
],
[
0.5107421875,
0.412353515625,
0.88232421875,
0.412353515625,
0.88232421875,
0.425537109375,
0.5107421875,
0.425537109375
],
[
0.11517333984375,
0.434326171875,
0.488525390625,
0.43310546875,
0.488525390625,
0.447509765625,
0.11517333984375,
0.448974609375
],
[
0.5087890625,
0.429931640625,
0.74853515625,
0.429931640625,
0.74853515625,
0.443115234375,
0.5087890625,
0.443115234375
],
[
0.52783203125,
0.44580078125,
0.88134765625,
0.44580078125,
0.88134765625,
0.460205078125,
0.52783203125,
0.460205078125
],
[
0.1168212890625,
0.451904296875,
0.48779296875,
0.451904296875,
0.48779296875,
0.46533203125,
0.1168212890625,
0.46533203125
],
[
0.1168212890625,
0.468505859375,
0.488525390625,
0.468505859375,
0.488525390625,
0.481689453125,
0.1168212890625,
0.481689453125
],
[
0.5087890625,
0.46337890625,
0.88232421875,
0.46337890625,
0.88232421875,
0.477783203125,
0.5087890625,
0.477783203125
],
[
0.11358642578125,
0.484130859375,
0.490234375,
0.48486328125,
0.490234375,
0.5,
0.11358642578125,
0.498779296875
],
[
0.5107421875,
0.4892578125,
0.88232421875,
0.4892578125,
0.88232421875,
0.50244140625,
0.5107421875,
0.50244140625
],
[
0.11358642578125,
0.501953125,
0.491943359375,
0.50244140625,
0.491943359375,
0.517578125,
0.11358642578125,
0.5166015625
],
[
0.5087890625,
0.5068359375,
0.8837890625,
0.5068359375,
0.8837890625,
0.52099609375,
0.5087890625,
0.52099609375
],
[
0.11517333984375,
0.51953125,
0.491943359375,
0.51953125,
0.491943359375,
0.5341796875,
0.11517333984375,
0.5341796875
],
[
0.50732421875,
0.52294921875,
0.8857421875,
0.5234375,
0.8857421875,
0.53857421875,
0.50732421875,
0.537109375
],
[
0.11517333984375,
0.537109375,
0.49267578125,
0.537109375,
0.49267578125,
0.5517578125,
0.11517333984375,
0.5517578125
],
[
0.50634765625,
0.54052734375,
0.701171875,
0.54150390625,
0.701171875,
0.55615234375,
0.50634765625,
0.5546875
],
[
0.11358642578125,
0.55322265625,
0.49267578125,
0.5537109375,
0.49267578125,
0.568359375,
0.11358642578125,
0.5673828125
],
[
0.11517333984375,
0.5712890625,
0.494384765625,
0.5712890625,
0.494384765625,
0.58544921875,
0.11517333984375,
0.58544921875
],
[
0.5087890625,
0.5673828125,
0.88134765625,
0.5673828125,
0.88134765625,
0.58203125,
0.5087890625,
0.58203125
],
[
0.11517333984375,
0.5869140625,
0.491943359375,
0.5869140625,
0.491943359375,
0.6015625,
0.11517333984375,
0.6015625
],
[
0.5087890625,
0.58544921875,
0.88134765625,
0.58544921875,
0.88134765625,
0.5986328125,
0.5087890625,
0.5986328125
],
[
0.50634765625,
0.5986328125,
0.8837890625,
0.5986328125,
0.8837890625,
0.619140625,
0.50634765625,
0.619140625
],
[
0.11358642578125,
0.60498046875,
0.49267578125,
0.60400390625,
0.49267578125,
0.61865234375,
0.11358642578125,
0.619140625
],
[
0.1168212890625,
0.62255859375,
0.491943359375,
0.62255859375,
0.491943359375,
0.63623046875,
0.1168212890625,
0.63623046875
],
[
0.50634765625,
0.619140625,
0.88232421875,
0.619140625,
0.88232421875,
0.63330078125,
0.50634765625,
0.63330078125
],
[
0.11517333984375,
0.6396484375,
0.494384765625,
0.6396484375,
0.494384765625,
0.65283203125,
0.11517333984375,
0.65283203125
],
[
0.50732421875,
0.63525390625,
0.81201171875,
0.63720703125,
0.81201171875,
0.6513671875,
0.50732421875,
0.6494140625
],
[
0.11517333984375,
0.6572265625,
0.490234375,
0.6572265625,
0.490234375,
0.67041015625,
0.11517333984375,
0.67041015625
],
[
0.50732421875,
0.6591796875,
0.87353515625,
0.66015625,
0.87353515625,
0.68115234375,
0.50732421875,
0.6796875
],
[
0.11358642578125,
0.673828125,
0.491943359375,
0.67236328125,
0.491943359375,
0.6875,
0.11358642578125,
0.68798828125
],
[
0.5048828125,
0.6787109375,
0.8857421875,
0.6767578125,
0.8857421875,
0.69775390625,
0.5048828125,
0.69970703125
],
[
0.1168212890625,
0.69140625,
0.49267578125,
0.69140625,
0.49267578125,
0.70556640625,
0.1168212890625,
0.70556640625
],
[
0.50732421875,
0.69873046875,
0.70751953125,
0.69873046875,
0.70751953125,
0.71337890625,
0.50732421875,
0.71337890625
],
[
0.11517333984375,
0.70849609375,
0.26953125,
0.70849609375,
0.26953125,
0.72265625,
0.11517333984375,
0.72265625
],
[
0.111083984375,
0.73095703125,
0.2320556640625,
0.73291015625,
0.2320556640625,
0.75048828125,
0.111083984375,
0.74853515625
],
[
0.5087890625,
0.7265625,
0.60693359375,
0.7265625,
0.60693359375,
0.7412109375,
0.5087890625,
0.7412109375
],
[
0.50634765625,
0.7470703125,
0.8837890625,
0.74853515625,
0.8837890625,
0.7666015625,
0.50634765625,
0.7646484375
],
[
0.111083984375,
0.755859375,
0.49267578125,
0.75390625,
0.49267578125,
0.7734375,
0.111083984375,
0.77587890625
],
[
0.5087890625,
0.7666015625,
0.875,
0.7666015625,
0.875,
0.78076171875,
0.5087890625,
0.78076171875
],
[
0.11517333984375,
0.7744140625,
0.490234375,
0.7744140625,
0.490234375,
0.7890625,
0.11517333984375,
0.7890625
],
[
0.11273193359375,
0.7890625,
0.491943359375,
0.787109375,
0.491943359375,
0.80810546875,
0.11273193359375,
0.81005859375
],
[
0.56787109375,
0.79541015625,
0.83642578125,
0.79443359375,
0.83642578125,
0.81298828125,
0.56787109375,
0.814453125
],
[
0.8505859375,
0.798828125,
0.88134765625,
0.798828125,
0.88134765625,
0.81103515625,
0.8505859375,
0.81103515625
],
[
0.11517333984375,
0.8095703125,
0.490234375,
0.8095703125,
0.490234375,
0.82373046875,
0.11517333984375,
0.82373046875
],
[
0.11358642578125,
0.82568359375,
0.491943359375,
0.82470703125,
0.491943359375,
0.8388671875,
0.11358642578125,
0.84033203125
],
[
0.5087890625,
0.8291015625,
0.88134765625,
0.8291015625,
0.88134765625,
0.84326171875,
0.5087890625,
0.84326171875
],
[
0.11517333984375,
0.84326171875,
0.491943359375,
0.84326171875,
0.491943359375,
0.85693359375,
0.11517333984375,
0.85693359375
],
[
0.5087890625,
0.8466796875,
0.8798828125,
0.8466796875,
0.8798828125,
0.861328125,
0.5087890625,
0.861328125
],
[
0.111083984375,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.880859375,
0.111083984375,
0.880859375
],
[
0.5087890625,
0.86279296875,
0.8837890625,
0.8623046875,
0.8837890625,
0.876953125,
0.5087890625,
0.87744140625
],
[
0.11767578125,
0.880859375,
0.1396484375,
0.880859375,
0.1396484375,
0.89453125,
0.11767578125,
0.89453125
],
[
0.15283203125,
0.875,
0.494384765625,
0.87255859375,
0.494384765625,
0.89794921875,
0.15283203125,
0.8994140625
],
[
0.5087890625,
0.880859375,
0.88232421875,
0.880859375,
0.88232421875,
0.89453125,
0.5087890625,
0.89453125
],
[
0.2034912109375,
0.88720703125,
0.2115478515625,
0.88330078125,
0.2156982421875,
0.88720703125,
0.20751953125,
0.8916015625
],
[
0.11358642578125,
0.896484375,
0.494384765625,
0.89794921875,
0.494384765625,
0.912109375,
0.11358642578125,
0.9111328125
],
[
0.50634765625,
0.89794921875,
0.8857421875,
0.89794921875,
0.8857421875,
0.9111328125,
0.50634765625,
0.9111328125
]
] |
[
"sented by an embedding vector, we first create a soft",
"",
"I=G(I)",
"alignment matrix using neural attention (Bahdanau",
"et al., 2015). We then use the (soft) alignment to",
"I=G( ) 9=H( +I+.+)",
"decompose the task into subproblems that are solved",
"separately. Finally, the results of these subproblems",
"DI=G(I)",
"are merged to produce the final classification. In ad-",
"",
"dition, we optionally apply intra-sentence attention",
"Figure 1: Pictoral overview of the approach, showing the Attend",
"(Cheng et al., 2016) to endow the model with a richer",
"(left), Compare (center) and Aggregate (right) steps.",
"encoding of substructures prior to the alignment step",
"Asymptotically our approach does the same total",
"time, we receive a pair of sentences (a, b) and our",
"work as a vanilla LSTM encoder, while being triv-",
"goal is to predict the correct label y.",
"ially parallelizable across sentence length, which can",
"Input representation. Let a -- (ai,..., ag.) and",
"allow for considerable speedups in low-latency set-",
"tings. Empirical results on the SNLI corpus show that",
"b = (bi,..., ben) denote the input representation of",
"each fragment that is fed to subsequent steps of the",
"our approach achieves state-of-the-art results, while",
"algorithm. The vanilla version of our model simply",
"using almost an order of magnitude fewer parameters",
"defines a := a and b := b. With this input rep",
"compared to complex LSTM-based approaches.",
"resentation, our model does not make use of word",
"2 Related Work",
"order. However, we discuss an extension using intra-",
"Our method is motivated by the central role played by",
"sentence attention in Section 3.4 that uses a minimal",
"alignment in machine translation (Koehn, 2009) and",
"amount of sequence information.",
"The core model consists of the following three",
"previous approaches to sentence similarity modeling",
"(Haghighi et al., 2005; Das and Smith, 2009; Chang",
"components (see Figure 1), which are trained jointly:",
"et al., 2010; Fader et al., 2013), natural language",
"Attend. First, soft-align the elements of a and b",
"inference (Marsi and Krahmer,2005; MacCartney",
"using a variant of neural attention (Bahdanau et al.",
"et al., 2006; Hickl and Bensley, 2007; MacCartney",
"2015)anddecompose theprobleminto the compari-",
"et al., 2008), and semantic parsing (Andreas et al.,",
"son of aligned subphrases.",
"2013). The neural counterpart to alignment, atten-",
"tion (Bahdanau et al., 2015), which is a key part",
"Compare. Second, separately compare each",
"of our approach,was originally proposed and has",
"aligned subphrase to produce a set of vectors",
"{v1,;}'=} for a and [v2,j'-1 for b. Each v1, is",
"been predominantly used in conjunction with LSTMs",
"(Rocktaschel et al., 2016; Wang and Jiang, 2016) and",
"a nonlinear combination of ai and its (softly) aligned",
"to a lesser extent with CNNs (Yin et al., 2016). In",
"subphrase in b (and analogously for V2,j).",
"contrast, our use of attention is purely based on word",
"Aggregate.Finally, aggregate the sets [vi,l'",
"embeddings and our method essentially consists of",
"and (v2,jl'= from the previous step and use the",
"feed-forward networks that operate largely indepen-",
"result to predict the label g.",
"dently of word order.",
"3Approach",
"3.1 Attend",
"We first obtain unnormalized attention weights eij,",
"Let a = (ai,...,aa) and b = (b1,...,be.) be",
"computed by a function F', which decomposes as:",
"the two input sentences of length la and b, re-",
"spectively. We assume that each ai, bj Rd",
"eij := F'(ai,b,) := F(ai)T F(b).",
"(1)",
"is a word embedding vector of dimension d and",
"that each sentence is prepended with a “NULL\"",
"This decomposition avoids the quadratic complexity",
"token.Our training data comes in the form of",
"that would be associated with separately applying F",
"labeled pairs (a(n),b(n), y(n)M=1, where y(n)",
"la × lp times.Instead, only la + lbapplications of",
"(yi",
"....ye) is an indicator vector encoding the",
"F are needed. We take F to be a feed-forward neural",
"",
"label and C is the number of output classes. At test",
"network with ReLU activations (Glorot et al., 2011)."
] |
{
"arxiv_id": "1606.01933",
"language": "en",
"page_idx": 1
}
|
||
1606.01933-en-2
|
1606.01933-en-2.jpg
|
These attention weights are normalized as follows: 3.4 Intra-Sentence Attention (Optional)
$$\beta_{i}:=\sum_{j=1}^{\ell_{b}}\frac{\exp(e_{ij})}{\sum_{k=1}^{\ell_{b}}\exp(e_{ik})}\bar{b}_{j}\:,\\\alpha_{j}:=\sum_{i=1}^{\ell_{a}}\frac{\exp(e_{ij})}{\sum_{k=1}^{\ell_{a}}\exp(e_{kj})}\bar{a}_{i}\:.$$
(2)
Here $\beta_i$ is the subphrase in b that is (softly) aligned
to $\bar{a}_i$ and vice versa for $\alpha_j.$
# 3.2 Compare
Next, we separately compare the aligned phrases $\left\{(\bar{a}_{i},\beta_{i})\right\}_{i=1}^{\ell_{a}}$ and$\left\{(\bar{b}_{j},\alpha_{j})\right\}_{j=1}^{\ell_{b}}$ using a function $G$, which in this work is again a feed-forward network:
$$\begin{aligned}\mathbf{v}_{1,i}&:=G([\bar{a}_i,\beta_i])&\forall i\in[1,\dots,\ell_a]\:,\\\mathbf{v}_{2,j}&:=G([\bar{b}_j,\alpha_j])&\forall j\in[1,\dots,\ell_b]\:.\end{aligned}$$
(3)
where the brackets $[\cdot,\cdot]$ denote concatenation. Note that since there are only a linear number of terms in this case, we do not need to apply a decomposition as was done in the previous step. Thus $G$ can jointly take into account both $\bar{a}_i$, and $\beta_i.$
## 3.3 Aggregate
We now have two sets of comparison vectors $\begin{Bmatrix}\mathbf{v}_{1,i}\end{Bmatrix}_{i=1}^{\ell_a}$ and $\{\mathbf{v}_{2,j}\}_{j=1}^{\ell_b}.$ We first aggregate over each set by summation:
(4)
$$\mathbf{v}_1=\sum_{i=1}^{\ell_a}\mathbf{v}_{1,i}\quad,\quad\mathbf{v}_2=\sum_{j=1}^{\ell_b}\mathbf{v}_{2,j}\:.$$
and feed the result through a final classifier $H$, that is a feed forward network followed by a linear layer:
(5)
$$\hat{\mathbf{y}}=H([\mathbf{v}_1,\mathbf{v}_2])\:,$$
where $\hat{\mathbf{y}}\in\mathbb{R}^{C}$ represents the predicted (unnormalized) scores for each class and consequently the predicted class is given by $\hat{y}=\operatorname{argmax}_i\hat{\mathbf{y}}_i.$
For training, we use multi-class cross-entropy loss with dropout regularization (Srivastava et al., 2014):
$$\begin{aligned}L(\theta_F,\theta_G,\theta_H)=\frac{1}{N}\sum_{n=1}^{N}\sum_{c=1}^{C}y_{c}^{(n)}\log\frac{\exp(\hat{y}_{c})}{\sum_{c^{\prime}=1}^{C}\exp(\hat{y}_{c^{\prime}})}\:.\end{aligned}$$
Here $\theta_F,\theta_G,\theta_{H}$ denote the learnable parameters of
the functions F, G and H, respectively.
In the above model, the input representations are simple word embeddings. However, we can augment this input representation with intra-sentence attention to encode compositional relationships between words within each sentence, as proposed by Cheng et al. (2016). Similar to Eqs. 1 and 2, we define
$$f_{ij}:=F_{\text{intra}}(a_i)^TF_{\text{intra}}(a_j)\:,$$
(6)
where $F_{\mathrm{intra}}$ is a feed-forward network. We then cre-
ate the self-aligned phrases
(7)
$$\begin{aligned}a_i'&:=\sum_{j=1}^{\ell_a}\frac{\exp(f_{ij}+d_{i-j})}{\sum_{k=1}^{\ell_a}\exp(f_{ik}+d_{i-k})}a_j\:.\end{aligned}$$
The distance-sensitive bias terms $d_{i-j}\in\mathbb{R}$ provides the model with a minimal amount of sequence information, while remaining parallelizable. These terms are bucketed such that all distances greater than 10 words share the same bias. The input representation for subsequent steps is then defined as $\bar{a}_i:=[a_i,a_i^{\prime}]$ and analogously $\overline{b}_i:=[b_i,b_i^{\prime}].$
## 4 Computational Complexity
We now discuss the asymptotic complexity of our approach and how it offers a higher degree of parallelism than LSTM-based approaches. Recall that $d$ denotes embedding dimension and $\ell$ means sentence length. For simplicity we assume that all hidden dimensions are $d$ and that the complexity of matrix$( d\times d)$-vector$(d\times1)$ multiplication is $O(d^{2}).$
A key assumption of our analysis is that $\ell<d$, which we believe is reasonable and is true of the SNLI dataset (Bowman et al., 2015) where $\ell<80$, whereas recent LSTM-based approaches have used $d\geq300.$ This assumption allows us to bound the complexity of computing the $\ell^{2}$ attention weights.
Complexity of LSTMs. The complexity of an LSTM cell is $O(d^{2})$, resulting in a complexity of $O(\ell d^2)$ to encode the sentence. Adding attention as in Rocktäschel et al. (2016) increases this complexity to $O(\ell d^{2}+\ell^{2}d).$
Complexity of our Approach. Application of a feed-forward network requires $O(d^{2})$ steps. Thus, the Compare and Aggregate steps have complexity $O(\ell d^{2})$ and $O(d^{2})$ respectively. For the Attend step,
|
[
[
0.1356201171875,
0.091552734375,
0.82421875,
0.091552734375,
0.82421875,
0.1060791015625,
0.1356201171875,
0.1060791015625
],
[
0.5107421875,
0.11553955078125,
0.88232421875,
0.11553955078125,
0.88232421875,
0.1300048828125,
0.5107421875,
0.1300048828125
],
[
0.2484130859375,
0.12249755859375,
0.26318359375,
0.12249755859375,
0.26318359375,
0.1331787109375,
0.2484130859375,
0.1331787109375
],
[
0.293212890625,
0.1256103515625,
0.358642578125,
0.1256103515625,
0.358642578125,
0.141357421875,
0.293212890625,
0.141357421875
],
[
0.5087890625,
0.1319580078125,
0.88232421875,
0.1307373046875,
0.88232421875,
0.1458740234375,
0.5087890625,
0.146484375
],
[
0.1993408203125,
0.1383056640625,
0.2296142578125,
0.1383056640625,
0.2296142578125,
0.149658203125,
0.1993408203125,
0.149658203125
],
[
0.29736328125,
0.147705078125,
0.381591796875,
0.1458740234375,
0.381591796875,
0.1634521484375,
0.29736328125,
0.1654052734375
],
[
0.5107421875,
0.149658203125,
0.88232421875,
0.149658203125,
0.88232421875,
0.1634521484375,
0.5107421875,
0.1634521484375
],
[
0.5107421875,
0.1666259765625,
0.88134765625,
0.1666259765625,
0.88134765625,
0.179931640625,
0.5107421875,
0.179931640625
],
[
0.250732421875,
0.175537109375,
0.2548828125,
0.175537109375,
0.2548828125,
0.1817626953125,
0.250732421875,
0.1817626953125
],
[
0.293212890625,
0.1767578125,
0.359375,
0.1767578125,
0.359375,
0.1925048828125,
0.293212890625,
0.1925048828125
],
[
0.46728515625,
0.1875,
0.491943359375,
0.1875,
0.491943359375,
0.2008056640625,
0.46728515625,
0.2008056640625
],
[
0.5087890625,
0.18310546875,
0.8857421875,
0.18310546875,
0.8857421875,
0.1962890625,
0.5087890625,
0.1962890625
],
[
0.19775390625,
0.1925048828125,
0.2296142578125,
0.1925048828125,
0.2296142578125,
0.199462890625,
0.19775390625,
0.199462890625
],
[
0.240234375,
0.199462890625,
0.386474609375,
0.1956787109375,
0.386474609375,
0.21337890625,
0.240234375,
0.2171630859375
],
[
0.51123046875,
0.2008056640625,
0.810546875,
0.2008056640625,
0.810546875,
0.2139892578125,
0.51123046875,
0.2139892578125
],
[
0.1168212890625,
0.2298583984375,
0.490234375,
0.2298583984375,
0.490234375,
0.2442626953125,
0.1168212890625,
0.2442626953125
],
[
0.59619140625,
0.2298583984375,
0.79736328125,
0.2298583984375,
0.79736328125,
0.2474365234375,
0.59619140625,
0.2474365234375
],
[
0.859375,
0.231689453125,
0.8837890625,
0.231689453125,
0.8837890625,
0.24560546875,
0.859375,
0.24560546875
],
[
0.11273193359375,
0.2442626953125,
0.310546875,
0.246826171875,
0.310546875,
0.264404296875,
0.11273193359375,
0.261962890625
],
[
0.50732421875,
0.260009765625,
0.88623046875,
0.2587890625,
0.88623046875,
0.276611328125,
0.50732421875,
0.27783203125
],
[
0.11358642578125,
0.272705078125,
0.2296142578125,
0.274658203125,
0.2296142578125,
0.291748046875,
0.11358642578125,
0.2890625
],
[
0.5087890625,
0.27978515625,
0.7099609375,
0.27978515625,
0.7099609375,
0.29345703125,
0.5087890625,
0.29345703125
],
[
0.1168212890625,
0.29541015625,
0.48779296875,
0.29541015625,
0.48779296875,
0.31005859375,
0.1168212890625,
0.31005859375
],
[
0.6044921875,
0.3125,
0.62158203125,
0.3037109375,
0.63232421875,
0.318115234375,
0.615234375,
0.32568359375
],
[
0.654296875,
0.311279296875,
0.779296875,
0.313232421875,
0.779296875,
0.330810546875,
0.654296875,
0.328857421875
],
[
0.560546875,
0.322509765625,
0.62744140625,
0.3193359375,
0.62841796875,
0.335205078125,
0.56201171875,
0.33837890625
],
[
0.1168212890625,
0.313232421875,
0.491943359375,
0.313232421875,
0.491943359375,
0.330810546875,
0.1168212890625,
0.330810546875
],
[
0.859375,
0.323974609375,
0.88232421875,
0.323974609375,
0.88232421875,
0.3359375,
0.859375,
0.3359375
],
[
0.11517333984375,
0.33203125,
0.488525390625,
0.33203125,
0.488525390625,
0.34521484375,
0.11517333984375,
0.34521484375
],
[
0.60302734375,
0.33203125,
0.810546875,
0.32568359375,
0.81201171875,
0.35107421875,
0.6044921875,
0.356689453125
],
[
0.1650390625,
0.35986328125,
0.438720703125,
0.35986328125,
0.438720703125,
0.375,
0.1650390625,
0.375
],
[
0.5107421875,
0.3662109375,
0.88134765625,
0.367431640625,
0.88134765625,
0.3818359375,
0.5107421875,
0.380615234375
],
[
0.1650390625,
0.38330078125,
0.453369140625,
0.38330078125,
0.453369140625,
0.397705078125,
0.1650390625,
0.397705078125
],
[
0.45751953125,
0.3837890625,
0.488525390625,
0.3837890625,
0.488525390625,
0.395751953125,
0.45751953125,
0.395751953125
],
[
0.50732421875,
0.3837890625,
0.8798828125,
0.3837890625,
0.8798828125,
0.397705078125,
0.50732421875,
0.397705078125
],
[
0.5087890625,
0.401611328125,
0.88134765625,
0.401611328125,
0.88134765625,
0.415283203125,
0.5087890625,
0.415283203125
],
[
0.11517333984375,
0.410888671875,
0.488525390625,
0.410400390625,
0.488525390625,
0.4248046875,
0.11517333984375,
0.425537109375
],
[
0.5087890625,
0.419189453125,
0.88134765625,
0.419189453125,
0.88134765625,
0.43310546875,
0.5087890625,
0.43310546875
],
[
0.1168212890625,
0.4287109375,
0.491943359375,
0.4287109375,
0.491943359375,
0.443115234375,
0.1168212890625,
0.443115234375
],
[
0.50732421875,
0.436279296875,
0.8798828125,
0.436279296875,
0.8798828125,
0.449462890625,
0.50732421875,
0.449462890625
],
[
0.1168212890625,
0.4462890625,
0.491943359375,
0.4462890625,
0.491943359375,
0.4609375,
0.1168212890625,
0.4609375
],
[
0.5087890625,
0.45263671875,
0.88134765625,
0.45263671875,
0.88134765625,
0.46728515625,
0.5087890625,
0.46728515625
],
[
0.11517333984375,
0.464111328125,
0.488525390625,
0.464111328125,
0.488525390625,
0.477783203125,
0.11517333984375,
0.477783203125
],
[
0.50634765625,
0.46728515625,
0.7236328125,
0.4697265625,
0.7236328125,
0.4873046875,
0.50634765625,
0.48486328125
],
[
0.11517333984375,
0.48095703125,
0.3544921875,
0.48095703125,
0.3544921875,
0.494384765625,
0.11517333984375,
0.494384765625
],
[
0.5107421875,
0.49560546875,
0.76708984375,
0.49560546875,
0.76708984375,
0.51025390625,
0.5107421875,
0.51025390625
],
[
0.11273193359375,
0.505859375,
0.23779296875,
0.50830078125,
0.23779296875,
0.52587890625,
0.11273193359375,
0.5234375
],
[
0.5087890625,
0.51806640625,
0.88232421875,
0.51953125,
0.88232421875,
0.5341796875,
0.5087890625,
0.53271484375
],
[
0.1168212890625,
0.52978515625,
0.488525390625,
0.52978515625,
0.488525390625,
0.54443359375,
0.1168212890625,
0.54443359375
],
[
0.5087890625,
0.537109375,
0.8837890625,
0.5361328125,
0.8837890625,
0.55029296875,
0.5087890625,
0.5517578125
],
[
0.11517333984375,
0.54541015625,
0.494384765625,
0.54541015625,
0.494384765625,
0.56640625,
0.11517333984375,
0.56640625
],
[
0.5107421875,
0.5537109375,
0.88232421875,
0.5537109375,
0.88232421875,
0.568359375,
0.5107421875,
0.568359375
],
[
0.11358642578125,
0.56494140625,
0.28759765625,
0.56494140625,
0.28759765625,
0.5791015625,
0.11358642578125,
0.5791015625
],
[
0.5107421875,
0.5712890625,
0.88232421875,
0.5712890625,
0.88232421875,
0.58544921875,
0.5107421875,
0.58544921875
],
[
0.373291015625,
0.59130859375,
0.393798828125,
0.59130859375,
0.393798828125,
0.60498046875,
0.373291015625,
0.60498046875
],
[
0.5107421875,
0.5869140625,
0.8837890625,
0.5869140625,
0.8837890625,
0.60107421875,
0.5107421875,
0.60107421875
],
[
0.191162109375,
0.5966796875,
0.20751953125,
0.59033203125,
0.2156982421875,
0.6015625,
0.1993408203125,
0.60791015625
],
[
0.36767578125,
0.60400390625,
0.4345703125,
0.60693359375,
0.43310546875,
0.62548828125,
0.366943359375,
0.62255859375
],
[
0.5087890625,
0.60400390625,
0.8857421875,
0.60498046875,
0.8857421875,
0.619140625,
0.5087890625,
0.61865234375
],
[
0.4697265625,
0.60986328125,
0.490234375,
0.60986328125,
0.490234375,
0.62060546875,
0.4697265625,
0.62060546875
],
[
0.1502685546875,
0.61279296875,
0.17236328125,
0.61279296875,
0.17236328125,
0.62060546875,
0.1502685546875,
0.62060546875
],
[
0.1953125,
0.60791015625,
0.205078125,
0.60791015625,
0.205078125,
0.6123046875,
0.1953125,
0.6123046875
],
[
0.2156982421875,
0.6142578125,
0.24267578125,
0.6142578125,
0.24267578125,
0.6220703125,
0.2156982421875,
0.6220703125
],
[
0.3349609375,
0.6142578125,
0.35693359375,
0.6142578125,
0.35693359375,
0.61865234375,
0.3349609375,
0.61865234375
],
[
0.370849609375,
0.62548828125,
0.396240234375,
0.62548828125,
0.396240234375,
0.6396484375,
0.370849609375,
0.6396484375
],
[
0.5087890625,
0.62255859375,
0.8837890625,
0.62060546875,
0.8837890625,
0.63525390625,
0.5087890625,
0.63720703125
],
[
0.52685546875,
0.6396484375,
0.8857421875,
0.6396484375,
0.8857421875,
0.65283203125,
0.52685546875,
0.65283203125
],
[
0.193603515625,
0.626953125,
0.214111328125,
0.626953125,
0.214111328125,
0.63720703125,
0.193603515625,
0.63720703125
],
[
0.11517333984375,
0.6494140625,
0.490234375,
0.65087890625,
0.490234375,
0.66552734375,
0.11517333984375,
0.6640625
],
[
0.5087890625,
0.65576171875,
0.88232421875,
0.65576171875,
0.88232421875,
0.66943359375,
0.5087890625,
0.66943359375
],
[
0.11517333984375,
0.66845703125,
0.494384765625,
0.66845703125,
0.494384765625,
0.681640625,
0.11517333984375,
0.681640625
],
[
0.5087890625,
0.671875,
0.8837890625,
0.67236328125,
0.8837890625,
0.6875,
0.5087890625,
0.68603515625
],
[
0.5087890625,
0.689453125,
0.88232421875,
0.68994140625,
0.88232421875,
0.705078125,
0.5087890625,
0.7041015625
],
[
0.233642578125,
0.69775390625,
0.363525390625,
0.69775390625,
0.363525390625,
0.7119140625,
0.233642578125,
0.7119140625
],
[
0.465576171875,
0.69873046875,
0.491943359375,
0.69873046875,
0.491943359375,
0.7119140625,
0.465576171875,
0.7119140625
],
[
0.5107421875,
0.70849609375,
0.88232421875,
0.70849609375,
0.88232421875,
0.7216796875,
0.5107421875,
0.7216796875
],
[
0.11517333984375,
0.72802734375,
0.495849609375,
0.72802734375,
0.495849609375,
0.7421875,
0.11517333984375,
0.7421875
],
[
0.5087890625,
0.724609375,
0.869140625,
0.724609375,
0.869140625,
0.7392578125,
0.5087890625,
0.7392578125
],
[
0.11358642578125,
0.744140625,
0.49267578125,
0.74560546875,
0.49267578125,
0.76025390625,
0.11358642578125,
0.7587890625
],
[
0.5087890625,
0.75,
0.88134765625,
0.75048828125,
0.88134765625,
0.76513671875,
0.5087890625,
0.7646484375
],
[
0.11358642578125,
0.76123046875,
0.398681640625,
0.76318359375,
0.398681640625,
0.77783203125,
0.11358642578125,
0.77587890625
],
[
0.5087890625,
0.767578125,
0.88232421875,
0.76806640625,
0.88232421875,
0.78271484375,
0.5087890625,
0.7822265625
],
[
0.13232421875,
0.77978515625,
0.490234375,
0.77978515625,
0.490234375,
0.79443359375,
0.13232421875,
0.79443359375
],
[
0.5107421875,
0.7841796875,
0.88232421875,
0.78515625,
0.88232421875,
0.7998046875,
0.5107421875,
0.798828125
],
[
0.11517333984375,
0.796875,
0.490234375,
0.796875,
0.490234375,
0.81103515625,
0.11517333984375,
0.81103515625
],
[
0.50732421875,
0.8017578125,
0.88134765625,
0.80322265625,
0.88134765625,
0.8173828125,
0.50732421875,
0.81640625
],
[
0.273681640625,
0.82470703125,
0.291748046875,
0.82470703125,
0.291748046875,
0.83349609375,
0.273681640625,
0.83349609375
],
[
0.5107421875,
0.8193359375,
0.6337890625,
0.8193359375,
0.6337890625,
0.833984375,
0.5107421875,
0.833984375
],
[
0.298095703125,
0.8271484375,
0.313720703125,
0.8271484375,
0.313720703125,
0.83056640625,
0.298095703125,
0.83056640625
],
[
0.25244140625,
0.83349609375,
0.261474609375,
0.83349609375,
0.261474609375,
0.84033203125,
0.25244140625,
0.84033203125
],
[
0.410888671875,
0.8291015625,
0.468017578125,
0.8291015625,
0.468017578125,
0.8447265625,
0.410888671875,
0.8447265625
],
[
0.1168212890625,
0.8359375,
0.2303466796875,
0.83837890625,
0.2303466796875,
0.85400390625,
0.1168212890625,
0.8515625
],
[
0.236083984375,
0.8359375,
0.386474609375,
0.8271484375,
0.390625,
0.85693359375,
0.240234375,
0.8662109375
],
[
0.400390625,
0.849609375,
0.41259765625,
0.849609375,
0.41259765625,
0.85400390625,
0.400390625,
0.85400390625
],
[
0.50634765625,
0.8466796875,
0.88134765625,
0.8466796875,
0.88134765625,
0.861328125,
0.50634765625,
0.861328125
],
[
0.384033203125,
0.85302734375,
0.5,
0.849609375,
0.5,
0.8642578125,
0.384033203125,
0.8671875
],
[
0.281005859375,
0.859375,
0.310546875,
0.859375,
0.310546875,
0.8662109375,
0.281005859375,
0.8662109375
],
[
0.5107421875,
0.86279296875,
0.88232421875,
0.86279296875,
0.88232421875,
0.87744140625,
0.5107421875,
0.87744140625
],
[
0.1168212890625,
0.8798828125,
0.495849609375,
0.8798828125,
0.495849609375,
0.89453125,
0.1168212890625,
0.89453125
],
[
0.50732421875,
0.87890625,
0.88134765625,
0.880859375,
0.88134765625,
0.89599609375,
0.50732421875,
0.89306640625
],
[
0.11358642578125,
0.896484375,
0.3896484375,
0.89794921875,
0.3896484375,
0.912109375,
0.11358642578125,
0.9111328125
],
[
0.51123046875,
0.89794921875,
0.88232421875,
0.89794921875,
0.88232421875,
0.912109375,
0.51123046875,
0.912109375
]
] |
[
"These attention weights are normalized as follows:3.4 Intra-Sentence Attention (Optional)",
"In the above model, the input representations are",
"lb",
"exp(eii)",
"simple word embeddings.However,we can augment",
"Bi:3",
"-1exp(eik)",
"this input representation with intra-sentence attention",
"to encode compositional relationships between words",
"",
"exp(ei)",
"(2)",
"within each sentence, as proposed by Cheng et al.",
"0:E",
"=i k-1 exp(eki)",
"(2016). Similar to Eqs. 1 and 2, we define",
"Here β; is the subphrase in b that is (softly) aligned",
"fij := Fintra(ai)T Fintra(aj),",
"(6)",
"to ai and vice versa for Qj.",
"where Fintra is a feed-forward network. We then cre-",
"3.2Compare",
"ate the self-aligned phrases",
"Next, we separately compare the aligned phrases",
"",
"exp(fij +di-s)",
"a, :=/",
"[(ai, β:))f-1 and [(bj, Qj)- using a function G,",
"(7)",
"which in this work is again a feed-forward network:",
"台k=1 exp(fis + d-k)",
"vi..:=G([ai, BiD) Vi [1,..., lal,",
"The distance-sensitive bias terms di-i e R provides",
"V2,j := G([bj,al) Vj e[1,..., lb] .",
"(3)",
"the model with a minimal amount of sequence infor",
"mation, while remaining parallelizable. These terms",
"where the brackets [-,-| denote concatenation. Note",
"are bucketed such that all distances greater than 10",
"that since there are only a linear number of terms in",
"words share the same bias. The input representation",
"this case, we do not need to apply a decomposition",
"for subsequent steps is then defined as a, :- [ai, a']",
"as was done in the previous step. Thus G can jointly",
"and analogously bi := [bi, b)].",
"take into account both ai, and βi.",
"4ComputationalComplexity",
"3.3 Aggregate",
"We now discuss the asymptotic complexity of our",
"We now have two sets of comparison vectors",
"approach and how it offers a higher degree of par-",
"[vi,}\" and [v2,i'-1. We first aggregate over",
"allelism than LSTM-based approaches. Recall that",
"each set by summation:",
"d denotes embedding dimension and l means sen-",
"lb",
"tence length. For simplicity we assume that all hid-",
"e",
"v2,j",
"den dimensions are d and that the complexity of",
"(4)",
"V1",
"",
"V1.",
"10",
"j-1",
"matrix(d × d)-vector(d × 1) multiplication is O(d2).",
"A key assumption of our analysis is that l < d,",
"=1",
"and feed the result through a final classifier H, that",
"which we believe is reasonable and is true of the",
"is a feed forward network followed by a linear layer:",
"SNLI dataset (Bowman et al., 2015) where l < 80,",
"whereas recent LSTM-based approaches have used",
"y = H([vi, V2l)",
"(5)",
"d ≥ 300. This assumption allows us to bound the",
"where y e RC represents the predicted (unnormal-",
"complexity of computing the g? attention weights.",
"ized) scores for each class and consequently the pre-",
"Complexity of LSTMs. The complexity of an",
"dicted class is given by y = argmax,yi.",
"LSTM cell is O(d?), resulting in a complexity of",
"For training, we use multi-class cross-entropy loss",
"O(ld?) to encode the sentence. Adding attention as",
"with dropout regularization (Srivastava et al., 2014):",
"in Rocktaschel et al. (2016) increases this complexity",
"N",
"to O(ld2 + l2d).",
"",
"",
"exp(yc)",
"L(OF,OG,OH) :",
"1og",
"",
"Complexity of our Approach. Application of a",
"Zc=1 exp(yc)",
"1C=",
"feed-forward network requires O(d?) steps. Thus.",
"Here F, g, H denote the learnable parameters of",
"the Compare and Aggregate steps have complexity",
"the functions F, G and H, respectively.",
"O(ld2) and O(d2) respectively. For the Attend step"
] |
{
"arxiv_id": "1606.01933",
"language": "en",
"page_idx": 2
}
|
||
1606.01933-en-3
|
1606.01933-en-3.jpg
|
<table>
<tbody>
<tr>
<th>Method</th>
<th>Train Acc</th>
<th>$\operatorname{Test}\operatorname{Acc}$</th>
<th>$\#$Parameters</th>
</tr>
<tr>
<td>Lexicalized Classifier (Bowman et al., 2015)</td>
<td>99.7</td>
<td>78.2</td>
<td> </td>
</tr>
<tr>
<td>${\mathrm{ul..2016}}$ 3000 $a$</td>
<td>83.9</td>
<td>80.6</td>
<td>$3.0\mathbf{M}$</td>
</tr>
<tr>
<td>$||||\gamma||$ 1(ikl) Vendrov et al 2015 ancoders</td>
<td>98.8</td>
<td>81.4</td>
<td>$15.0\mathbf{M}$</td>
</tr>
<tr>
<td>$\left\{(00)\right\}$ 1(CNN encoders (Mou et al..2015)</td>
<td>83.3</td>
<td>82.1</td>
<td>3.5M</td>
</tr>
<tr>
<td>300D SPINN-P ${\mathrm{al.,2016}}$</td>
<td>89.2</td>
<td>83.2</td>
<td>$3.7\mathbf{M}$</td>
</tr>
<tr>
<td>100D al 2016)</td>
<td>85.3</td>
<td>83.5</td>
<td>252K</td>
</tr>
<tr>
<td>3(00) $\mathbf{g},2016)$</td>
<td>92.0</td>
<td>86.1</td>
<td>$1.9\mathbf{M}$</td>
</tr>
<tr>
<td>450D LSTMN with deep g et al., 2016) 112 tuls10 (Cheng</td>
<td>88.5</td>
<td>86.3</td>
<td>$3.4\mathbf{M}$</td>
</tr>
<tr>
<td>${\mathrm{n}}\left({\mathrm{vanilla}}\right)$ Our approach</td>
<td>89.5</td>
<td>86.3</td>
<td>$382K$</td>
</tr>
<tr>
<td>attention</td>
<td>90.5</td>
<td>86.8</td>
<td>$582K$</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<th>Method</th>
<th>$N$</th>
<th>$E$</th>
<th>$\mathbf{C}$</th>
</tr>
<tr>
<td>Bowman et al. (2016)</td>
<td>80.6</td>
<td>88.2</td>
<td>85.5 $\frac{1}{2}=\frac{1}{2}=\frac{1}{2}$</td>
</tr>
<tr>
<td>Wang and Jiang ( 2016) </td>
<td>81.6</td>
<td>91.6</td>
<td>87.4 </td>
</tr>
<tr>
<td>Our a approach (vanilla)</td>
<td>83.6</td>
<td>91.3</td>
<td>85.8</td>
</tr>
<tr>
<td>Our approach w/ intra att.</td>
<td>83.7</td>
<td>92.1</td>
<td>86.7</td>
</tr>
</tbody>
</table>
Table 2: Breakdown of accuracy with respect to classes on SNLI
development set. $\mathbf{N=}$neutral, E=entailment, C=contradiction.
$F$ is evaluated $O(\ell)$ times, giving a complexity of $O(\ell d^{2}).$ Each attention weight $e_ij$ requires one dot product, resulting in a complexity of $O(\ell^2d).$
Thus the total complexity of the model is $O( \ell d^{2}+$ contained examples of similar length.
$\ell^2d)$,which is equal to that of an LSTM with attention. However, note that with the assumption that $\ell<d$, this becomes $O(\ell d^2)$ which is the same complexity as a regular LSTM. Moreover, unlike the LSTM, our approach has the advantage of being parallelizable over $\ell$, which can be useful at test time.
Table 1: Train/test accuracies on the SNLI dataset and number of parameters (excluding embeddings) for each approach
for training, 9,842 for development, and 9,824 for testing. We use the tokenized sentences from the non-binary parse provided in the dataset and prepend each sentence with a “NULL” token. During training, each sentence was padded up to the maximum length of the batch for efficient training (the padding was explicitly masked out so as not to affect the objective/gradients). For efficient batching in TensorFlow, we semi-sorted the training data to first contain examples where both sentences had length less than 20, followed by those with length less than 50, and then the rest. This ensured that most training batches
$\mathbf{Embeddings:}$ We use 300 dimensional GloVe embeddings (Pennington et al., 2014) to represent words. Each embedding vector was normalized to have $\ell_2$ norm of 1 and projected down to 200 dimensions, a number determined via hyperparameter tuning. Out-of-vocabulary (OOV) words are hashed to one of 100 random embeddings each initialized to mean 0 and standard deviation 1. All embeddings remain fxed during training, but the projection matrix is trained. All other parameter weights (hidden layers etc.) were initialized from random Gaussians with mean 0 and standard deviation 0.01.
Each hyperparameter setting was run on a single machine with 10 asynchronous gradient-update threads, using Adagrad (Duchi et al., 2011) for optimization with the default initial accumulator value of 0.1. Dropout regularization (Srivastava et al., 2014) was used for all ReLU layers, but not for the final linear layer. We additionally tuned the following hyperparameters and present their chosen values in
## 5 Experiments
We evaluate our approach on the Stanford Natural Language Inference (SNLI) dataset (Bowman et al., 2015). Given a sentences pair $(\mathbf{a},\mathbf{b})$, the task is to predict whether b is entailed by a, b contradicts a, or whether their relationship is neutral.
### 5.1 Implementation Details
The method was implemented in TensorFlow (Abadi
et al., 2015).
Data preprocessing: Following Bowman et al. (2015), we remove examples labeled“-”(no gold label) from the dataset, which leaves 549,367 pairs
|
[
[
0.59619140625,
0.09344482421875,
0.67333984375,
0.0953369140625,
0.67333984375,
0.111083984375,
0.59619140625,
0.10919189453125
],
[
0.7685546875,
0.0947265625,
0.86376953125,
0.09661865234375,
0.86376953125,
0.111083984375,
0.7685546875,
0.10919189453125
],
[
0.68896484375,
0.09783935546875,
0.75634765625,
0.09783935546875,
0.75634765625,
0.10919189453125,
0.68896484375,
0.10919189453125
],
[
0.1356201171875,
0.0953369140625,
0.1968994140625,
0.0953369140625,
0.1968994140625,
0.10986328125,
0.1356201171875,
0.10986328125
],
[
0.61669921875,
0.11932373046875,
0.654296875,
0.11932373046875,
0.654296875,
0.1351318359375,
0.61669921875,
0.1351318359375
],
[
0.70361328125,
0.11932373046875,
0.74169921875,
0.11932373046875,
0.74169921875,
0.1351318359375,
0.70361328125,
0.1351318359375
],
[
0.61669921875,
0.1427001953125,
0.654296875,
0.1427001953125,
0.654296875,
0.1590576171875,
0.61669921875,
0.1590576171875
],
[
0.136474609375,
0.12060546875,
0.45751953125,
0.12060546875,
0.45751953125,
0.1337890625,
0.136474609375,
0.1337890625
],
[
0.70361328125,
0.1427001953125,
0.74169921875,
0.1427001953125,
0.74169921875,
0.1590576171875,
0.70361328125,
0.1590576171875
],
[
0.81884765625,
0.1427001953125,
0.865234375,
0.1427001953125,
0.865234375,
0.1578369140625,
0.81884765625,
0.1578369140625
],
[
0.61669921875,
0.160400390625,
0.654296875,
0.160400390625,
0.654296875,
0.175537109375,
0.61669921875,
0.175537109375
],
[
0.136474609375,
0.143310546875,
0.5048828125,
0.143310546875,
0.5048828125,
0.1578369140625,
0.136474609375,
0.1578369140625
],
[
0.70361328125,
0.160400390625,
0.74169921875,
0.160400390625,
0.74169921875,
0.175537109375,
0.70361328125,
0.175537109375
],
[
0.810546875,
0.160400390625,
0.86376953125,
0.160400390625,
0.86376953125,
0.175537109375,
0.810546875,
0.175537109375
],
[
0.136474609375,
0.1767578125,
0.5,
0.177978515625,
0.5,
0.1925048828125,
0.136474609375,
0.1912841796875
],
[
0.1380615234375,
0.1610107421875,
0.53369140625,
0.1610107421875,
0.53369140625,
0.175537109375,
0.1380615234375,
0.175537109375
],
[
0.61669921875,
0.1767578125,
0.654296875,
0.1767578125,
0.654296875,
0.1932373046875,
0.61669921875,
0.1932373046875
],
[
0.70361328125,
0.1767578125,
0.740234375,
0.1767578125,
0.740234375,
0.1932373046875,
0.70361328125,
0.1932373046875
],
[
0.81884765625,
0.1767578125,
0.86376953125,
0.1767578125,
0.86376953125,
0.1925048828125,
0.81884765625,
0.1925048828125
],
[
0.61669921875,
0.1932373046875,
0.654296875,
0.1932373046875,
0.654296875,
0.210205078125,
0.61669921875,
0.210205078125
],
[
0.70361328125,
0.1932373046875,
0.74169921875,
0.1932373046875,
0.74169921875,
0.210205078125,
0.70361328125,
0.210205078125
],
[
0.81884765625,
0.1944580078125,
0.86376953125,
0.1944580078125,
0.86376953125,
0.210205078125,
0.81884765625,
0.210205078125
],
[
0.1356201171875,
0.2171630859375,
0.521484375,
0.2183837890625,
0.521484375,
0.23291015625,
0.1356201171875,
0.231689453125
],
[
0.1380615234375,
0.1944580078125,
0.48779296875,
0.1944580078125,
0.48779296875,
0.2083740234375,
0.1380615234375,
0.2083740234375
],
[
0.61669921875,
0.2183837890625,
0.654296875,
0.2183837890625,
0.654296875,
0.2342529296875,
0.61669921875,
0.2342529296875
],
[
0.70361328125,
0.2183837890625,
0.74169921875,
0.2183837890625,
0.74169921875,
0.2342529296875,
0.70361328125,
0.2342529296875
],
[
0.81787109375,
0.2183837890625,
0.86376953125,
0.2183837890625,
0.86376953125,
0.2342529296875,
0.81787109375,
0.2342529296875
],
[
0.61669921875,
0.23486328125,
0.654296875,
0.23486328125,
0.654296875,
0.250732421875,
0.61669921875,
0.250732421875
],
[
0.70361328125,
0.23486328125,
0.740234375,
0.23486328125,
0.740234375,
0.250732421875,
0.70361328125,
0.250732421875
],
[
0.8203125,
0.23486328125,
0.86376953125,
0.23486328125,
0.86376953125,
0.250732421875,
0.8203125,
0.250732421875
],
[
0.81884765625,
0.250732421875,
0.86376953125,
0.25244140625,
0.86279296875,
0.268310546875,
0.81787109375,
0.266357421875
],
[
0.136474609375,
0.236083984375,
0.419189453125,
0.236083984375,
0.419189453125,
0.250732421875,
0.136474609375,
0.250732421875
],
[
0.61669921875,
0.251953125,
0.654296875,
0.251953125,
0.654296875,
0.268310546875,
0.61669921875,
0.268310546875
],
[
0.70361328125,
0.251953125,
0.74169921875,
0.251953125,
0.74169921875,
0.268310546875,
0.70361328125,
0.268310546875
],
[
0.61669921875,
0.276611328125,
0.654296875,
0.276611328125,
0.654296875,
0.292236328125,
0.61669921875,
0.292236328125
],
[
0.1380615234375,
0.25244140625,
0.580078125,
0.25244140625,
0.580078125,
0.267578125,
0.1380615234375,
0.267578125
],
[
0.70361328125,
0.276611328125,
0.74169921875,
0.276611328125,
0.74169921875,
0.292236328125,
0.70361328125,
0.292236328125
],
[
0.81787109375,
0.276611328125,
0.86376953125,
0.276611328125,
0.86376953125,
0.292236328125,
0.81787109375,
0.292236328125
],
[
0.61669921875,
0.29345703125,
0.654296875,
0.29345703125,
0.654296875,
0.309326171875,
0.61669921875,
0.309326171875
],
[
0.136474609375,
0.27783203125,
0.30224609375,
0.27783203125,
0.30224609375,
0.292236328125,
0.136474609375,
0.292236328125
],
[
0.70361328125,
0.29345703125,
0.74169921875,
0.29345703125,
0.74169921875,
0.309326171875,
0.70361328125,
0.309326171875
],
[
0.81787109375,
0.294189453125,
0.86376953125,
0.294189453125,
0.86376953125,
0.31005859375,
0.81787109375,
0.31005859375
],
[
0.1422119140625,
0.3212890625,
0.8544921875,
0.322509765625,
0.8544921875,
0.337158203125,
0.1422119140625,
0.3359375
],
[
0.1380615234375,
0.29541015625,
0.4453125,
0.29541015625,
0.4453125,
0.31005859375,
0.1380615234375,
0.31005859375
],
[
0.435546875,
0.349853515625,
0.4599609375,
0.349853515625,
0.4599609375,
0.36865234375,
0.435546875,
0.36865234375
],
[
0.5107421875,
0.349853515625,
0.88232421875,
0.349853515625,
0.88232421875,
0.3642578125,
0.5107421875,
0.3642578125
],
[
0.3857421875,
0.35107421875,
0.41015625,
0.35107421875,
0.41015625,
0.36865234375,
0.3857421875,
0.36865234375
],
[
0.12335205078125,
0.3515625,
0.1854248046875,
0.3515625,
0.1854248046875,
0.3662109375,
0.12335205078125,
0.3662109375
],
[
0.3349609375,
0.3515625,
0.35546875,
0.3515625,
0.35546875,
0.367431640625,
0.3349609375,
0.367431640625
],
[
0.5107421875,
0.367431640625,
0.88134765625,
0.367431640625,
0.88134765625,
0.380615234375,
0.5107421875,
0.380615234375
],
[
0.125,
0.375732421875,
0.281982421875,
0.375732421875,
0.281982421875,
0.39013671875,
0.125,
0.39013671875
],
[
0.326904296875,
0.375732421875,
0.365234375,
0.375732421875,
0.365234375,
0.391357421875,
0.326904296875,
0.391357421875
],
[
0.375732421875,
0.375732421875,
0.4150390625,
0.375732421875,
0.4150390625,
0.391357421875,
0.375732421875,
0.391357421875
],
[
0.428955078125,
0.376953125,
0.46728515625,
0.376953125,
0.46728515625,
0.39013671875,
0.428955078125,
0.39013671875
],
[
0.5087890625,
0.3837890625,
0.88134765625,
0.3837890625,
0.88134765625,
0.3984375,
0.5087890625,
0.3984375
],
[
0.328369140625,
0.393310546875,
0.366943359375,
0.393310546875,
0.366943359375,
0.4091796875,
0.328369140625,
0.4091796875
],
[
0.374267578125,
0.391357421875,
0.4150390625,
0.393310546875,
0.414306640625,
0.410400390625,
0.373291015625,
0.40771484375
],
[
0.426513671875,
0.393310546875,
0.46728515625,
0.393310546875,
0.46728515625,
0.4091796875,
0.426513671875,
0.4091796875
],
[
0.5107421875,
0.40087890625,
0.88232421875,
0.40087890625,
0.88232421875,
0.415283203125,
0.5107421875,
0.415283203125
],
[
0.1258544921875,
0.39453125,
0.293212890625,
0.39453125,
0.293212890625,
0.4091796875,
0.1258544921875,
0.4091796875
],
[
0.125,
0.410400390625,
0.291748046875,
0.410400390625,
0.291748046875,
0.427978515625,
0.125,
0.427978515625
],
[
0.328369140625,
0.410400390625,
0.366943359375,
0.410400390625,
0.366943359375,
0.425537109375,
0.328369140625,
0.425537109375
],
[
0.375732421875,
0.410400390625,
0.4150390625,
0.410400390625,
0.4150390625,
0.425537109375,
0.375732421875,
0.425537109375
],
[
0.42724609375,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.425537109375,
0.42724609375,
0.425537109375
],
[
0.5107421875,
0.41845703125,
0.88134765625,
0.41845703125,
0.88134765625,
0.431884765625,
0.5107421875,
0.431884765625
],
[
0.326904296875,
0.4267578125,
0.365234375,
0.4267578125,
0.365234375,
0.442626953125,
0.326904296875,
0.442626953125
],
[
0.37744140625,
0.4267578125,
0.414306640625,
0.4267578125,
0.414306640625,
0.442626953125,
0.37744140625,
0.442626953125
],
[
0.42724609375,
0.427978515625,
0.465576171875,
0.427978515625,
0.465576171875,
0.443115234375,
0.42724609375,
0.443115234375
],
[
0.5107421875,
0.43505859375,
0.88232421875,
0.43505859375,
0.88232421875,
0.448974609375,
0.5107421875,
0.448974609375
],
[
0.1258544921875,
0.4287109375,
0.31201171875,
0.4287109375,
0.31201171875,
0.443115234375,
0.1258544921875,
0.443115234375
],
[
0.11517333984375,
0.448974609375,
0.491943359375,
0.449462890625,
0.491943359375,
0.464111328125,
0.11517333984375,
0.46337890625
],
[
0.5107421875,
0.45263671875,
0.88232421875,
0.45263671875,
0.88232421875,
0.466552734375,
0.5107421875,
0.466552734375
],
[
0.11358642578125,
0.46728515625,
0.473876953125,
0.46728515625,
0.473876953125,
0.48095703125,
0.11358642578125,
0.48095703125
],
[
0.5087890625,
0.4697265625,
0.8857421875,
0.468505859375,
0.8857421875,
0.48291015625,
0.5087890625,
0.484130859375
],
[
0.5087890625,
0.486083984375,
0.8857421875,
0.486083984375,
0.8857421875,
0.50048828125,
0.5087890625,
0.50048828125
],
[
0.1168212890625,
0.494384765625,
0.490234375,
0.494384765625,
0.490234375,
0.5087890625,
0.1168212890625,
0.5087890625
],
[
0.5087890625,
0.50390625,
0.88232421875,
0.50390625,
0.88232421875,
0.517578125,
0.5087890625,
0.517578125
],
[
0.1168212890625,
0.51318359375,
0.490234375,
0.51318359375,
0.490234375,
0.52783203125,
0.1168212890625,
0.52783203125
],
[
0.5087890625,
0.521484375,
0.88232421875,
0.521484375,
0.88232421875,
0.53515625,
0.5087890625,
0.53515625
],
[
0.11358642578125,
0.52978515625,
0.438720703125,
0.52880859375,
0.438720703125,
0.54345703125,
0.11358642578125,
0.54443359375
],
[
0.50732421875,
0.537109375,
0.8837890625,
0.53857421875,
0.8837890625,
0.55322265625,
0.50732421875,
0.5517578125
],
[
0.134033203125,
0.546875,
0.488525390625,
0.54541015625,
0.488525390625,
0.56005859375,
0.134033203125,
0.56103515625
],
[
0.5087890625,
0.5546875,
0.779296875,
0.5546875,
0.779296875,
0.5693359375,
0.5087890625,
0.5693359375
],
[
0.11767578125,
0.564453125,
0.490234375,
0.564453125,
0.490234375,
0.5791015625,
0.11767578125,
0.5791015625
],
[
0.52685546875,
0.5712890625,
0.88232421875,
0.57080078125,
0.88232421875,
0.58544921875,
0.52685546875,
0.5859375
],
[
0.1168212890625,
0.58203125,
0.490234375,
0.58203125,
0.490234375,
0.59521484375,
0.1168212890625,
0.59521484375
],
[
0.5087890625,
0.5888671875,
0.88232421875,
0.5888671875,
0.88232421875,
0.60400390625,
0.5087890625,
0.60400390625
],
[
0.1168212890625,
0.5986328125,
0.491943359375,
0.5986328125,
0.491943359375,
0.6123046875,
0.1168212890625,
0.6123046875
],
[
0.50732421875,
0.60595703125,
0.88232421875,
0.60595703125,
0.88232421875,
0.62060546875,
0.50732421875,
0.62060546875
],
[
0.11517333984375,
0.6162109375,
0.490234375,
0.61572265625,
0.490234375,
0.6298828125,
0.11517333984375,
0.630859375
],
[
0.50732421875,
0.62353515625,
0.8837890625,
0.62353515625,
0.8837890625,
0.63818359375,
0.50732421875,
0.63818359375
],
[
0.1168212890625,
0.6337890625,
0.490234375,
0.6337890625,
0.490234375,
0.64794921875,
0.1168212890625,
0.64794921875
],
[
0.5087890625,
0.6416015625,
0.8837890625,
0.6416015625,
0.8837890625,
0.65478515625,
0.5087890625,
0.65478515625
],
[
0.11517333984375,
0.65087890625,
0.473876953125,
0.65087890625,
0.473876953125,
0.6640625,
0.11517333984375,
0.6640625
],
[
0.5087890625,
0.65771484375,
0.8837890625,
0.65771484375,
0.8837890625,
0.671875,
0.5087890625,
0.671875
],
[
0.11517333984375,
0.6787109375,
0.2548828125,
0.6787109375,
0.2548828125,
0.693359375,
0.11517333984375,
0.693359375
],
[
0.50732421875,
0.6748046875,
0.88134765625,
0.6748046875,
0.88134765625,
0.68798828125,
0.50732421875,
0.68798828125
],
[
0.5087890625,
0.6923828125,
0.88134765625,
0.6923828125,
0.88134765625,
0.70556640625,
0.5087890625,
0.70556640625
],
[
0.11517333984375,
0.705078125,
0.490234375,
0.705078125,
0.490234375,
0.7197265625,
0.11517333984375,
0.7197265625
],
[
0.50634765625,
0.70849609375,
0.8837890625,
0.708984375,
0.8837890625,
0.7236328125,
0.50634765625,
0.72265625
],
[
0.11517333984375,
0.7216796875,
0.494384765625,
0.7216796875,
0.494384765625,
0.736328125,
0.11517333984375,
0.736328125
],
[
0.50732421875,
0.7265625,
0.88134765625,
0.7265625,
0.88134765625,
0.74072265625,
0.50732421875,
0.74072265625
],
[
0.11517333984375,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.75244140625,
0.11517333984375,
0.75244140625
],
[
0.5087890625,
0.74365234375,
0.88232421875,
0.74365234375,
0.88232421875,
0.7568359375,
0.5087890625,
0.7568359375
],
[
0.11517333984375,
0.755859375,
0.491943359375,
0.755859375,
0.491943359375,
0.77099609375,
0.11517333984375,
0.77099609375
],
[
0.50732421875,
0.76025390625,
0.8037109375,
0.76025390625,
0.8037109375,
0.7734375,
0.50732421875,
0.7734375
],
[
0.11358642578125,
0.7734375,
0.39794921875,
0.7734375,
0.39794921875,
0.787109375,
0.11358642578125,
0.787109375
],
[
0.529296875,
0.77783203125,
0.88232421875,
0.77783203125,
0.88232421875,
0.79248046875,
0.529296875,
0.79248046875
],
[
0.5107421875,
0.79443359375,
0.88232421875,
0.79443359375,
0.88232421875,
0.8095703125,
0.5107421875,
0.8095703125
],
[
0.11517333984375,
0.8037109375,
0.336669921875,
0.8037109375,
0.336669921875,
0.818359375,
0.11517333984375,
0.818359375
],
[
0.5107421875,
0.8125,
0.8857421875,
0.8125,
0.8857421875,
0.8271484375,
0.5107421875,
0.8271484375
],
[
0.11517333984375,
0.82763671875,
0.490234375,
0.82763671875,
0.490234375,
0.84228515625,
0.11517333984375,
0.84228515625
],
[
0.50634765625,
0.830078125,
0.8837890625,
0.830078125,
0.8837890625,
0.84326171875,
0.50634765625,
0.84326171875
],
[
0.11517333984375,
0.8447265625,
0.20751953125,
0.8447265625,
0.20751953125,
0.859375,
0.11517333984375,
0.859375
],
[
0.5087890625,
0.8466796875,
0.8837890625,
0.8466796875,
0.8837890625,
0.861328125,
0.5087890625,
0.861328125
],
[
0.13232421875,
0.86279296875,
0.494384765625,
0.86279296875,
0.494384765625,
0.87744140625,
0.13232421875,
0.87744140625
],
[
0.50634765625,
0.86279296875,
0.88232421875,
0.86279296875,
0.88232421875,
0.87744140625,
0.50634765625,
0.87744140625
],
[
0.1168212890625,
0.8798828125,
0.491943359375,
0.8798828125,
0.491943359375,
0.89453125,
0.1168212890625,
0.89453125
],
[
0.5048828125,
0.8798828125,
0.88232421875,
0.880859375,
0.88232421875,
0.89599609375,
0.5048828125,
0.89453125
],
[
0.11358642578125,
0.896484375,
0.491943359375,
0.89794921875,
0.491943359375,
0.912109375,
0.11358642578125,
0.9111328125
],
[
0.50634765625,
0.8984375,
0.88232421875,
0.8984375,
0.88232421875,
0.912109375,
0.50634765625,
0.912109375
]
] |
[
"Train Acc",
"#Parameters",
"Test Acc",
"Method",
"99.7",
"78.2",
"83.9",
"Lexicalized Classifier (Bowman et al., 2015)",
"80.6",
"3.0M",
"98.8",
"300D LSTM RNN encoders (Bowman et al., 2016)",
"81.4",
"15.0M",
"300D tree-based CNN encoders (Mou et al., 2015)",
"1024D pretrained GRU encoders (Vendrov et al., 2015)",
"83.3",
"82.1",
"3.5M",
"89.2",
"83.2",
"3.7M",
"100D LSTM with attention (Rocktaschel et al., 2016)",
"300D SPINN-PI encoders (Bowman et al., 2016)",
"85.3",
"83.5",
"252K",
"92.0",
"86.1",
"1.9M",
"3.4M",
"300D mLSTM (Wang and Jiang, 2016)",
"88.5",
"86.3",
"89.5",
"450D LSTMN with deep attention fusion (Cheng et al., 2016)",
"86.3",
"382K",
"90.5",
"Our approach (vanilla)",
"86.8",
"582K",
"Table 1: Train/test accuracies on the SNLI dataset and number of parameters (excluding embeddings) for each approach",
"Our approach with intra-sentence attention",
"c",
"for training, 9,842 for development, and 9,824 for",
"E",
"Method",
"N",
"testing. We use the tokenized sentences from the",
"Bowman et al. (2016)",
"80.6",
"88.2",
"85.5",
"non-binary parse provided in the dataset and prepend",
"81.6",
"91.6",
"87.4",
"each sentence with a \"NULL\" token. During training.",
"Wang and Jiang (2016)",
"Our approach (vanilla)",
"83.6",
"91.3",
"85.8",
"each sentence was padded up to the maximum length",
"83.7",
"92.1",
"86.7",
"of the batch for efficient training (the paddingwas",
"Our approach w/ intra att.",
"Table 2: Breakdown of accuracy with respect to classes on SNLI",
"explicitly masked out so as not to affect the objec-",
"development set. N=neutral, E=entailment, C=contradiction",
"tive/gradients).For efficient batching in TensorFlow,",
"we semi-sorted the training data to first contain ex-",
"F is evaluated O(e) times, giving a complexity of",
"amples where both sentences had length less than",
"O(ld2). Each attention weight eij requires one dot",
"20, followed by those with length less than 50, and",
"product, resulting in a complexity of O(?d).",
"then the rest. This ensured that most training batches",
"Thus the total complexity of the model is O(ed2 +",
"contained examples of similar length.",
"e? d), which is equal to that of an LSTM with atten-",
"Embeddings:We use 300 dimensional GloVe",
"tion. However, note that with the assumption that",
"embeddings (Pennington et al., 2014) to represent",
" < d, this becomes O(ld?) which is the same com-",
"words. Each embedding vector was normalized to",
"plexity as a regular LSTM. Moreover, unlike the",
"have l2 norm of 1 and projected down to 200 di-",
"LSTM, our approach has the advantage of being par-",
"mensions, a number determined via hyperparameter",
"allelizable over l, which can be useful at test time",
"tuning. Out-of-vocabulary (OOV) words are hashed",
"5Experiments",
"to one of 100 random embeddings each initialized",
"to mean O and standard deviation 1. All embeddings",
"We evaluate our approach on the Stanford Natural",
"remain fixed during training, but the projection ma-",
"Language Inference (SNLI) dataset (Bowman et al.,",
"trix is trained. All other parameter weights (hidden",
"2015). Given a sentences pair (a, b), the task is to",
"layers etc.)were initialized from random Gaussians",
"predict whether b is entailed by a, b contradicts a,",
"with mean 0 and standard deviation 0.01.",
"or whether their relationship is neutral.",
"Each hyperparameter setting was run on a sin-",
"gle machine with 10 asynchronous gradient-update",
"5.1ImplementationDetails",
"threads, using Adagrad (Duchi et al., 2011) for opti-",
"The method was implemented in TensorFlow (Abadi",
"mization with the default initial accumulator value of",
"et al., 2015).",
"0.1. Dropout regularization (Srivastava et al., 2014)",
"Data preprocessing: Following Bowman et al.",
"was used for all ReLU layers, but not for the final",
"(2015), we remove examples labeled _\" (no gold",
"linear layer.We additionally tuned the following",
"label) from the dataset, which leaves 549,367 pairs",
"hyperparameters and present their chosen values in"
] |
{
"arxiv_id": "1606.01933",
"language": "en",
"page_idx": 3
}
|
||
1606.01933-en-4
|
1606.01933-en-4.jpg
|
<table>
<tbody>
<tr>
<th>ID</th>
<th>Sentence l</th>
<th>Sentence 2</th>
<th>DA$\left ( \mathrm{vanilla}\right )$</th>
<th>$DA$ (intra att.) 1</th>
<th>SPINN- PI</th>
<th>mLSTM</th>
<th>Gold</th>
</tr>
<tr>
<td>A</td>
<td>Two kids are standing in the ocean hugging each other.</td>
<td>Two kids enjoy their day at the beach.</td>
<td>$N$</td>
<td>$N$</td>
<td>$E$</td>
<td>$E$</td>
<td>$N$</td>
</tr>
<tr>
<td>$B$</td>
<td>A dancer in costumer performs on stage while a man watches.</td>
<td>the man is captivated</td>
<td>$N$</td>
<td>$N$</td>
<td>$E$</td>
<td>$E$</td>
<td>$N$</td>
</tr>
<tr>
<td>$C$</td>
<td>They are sitting on the edge of a fountain</td>
<td>The fountain is splashing the persons seated.</td>
<td>$N$</td>
<td>$N$</td>
<td>C</td>
<td>C</td>
<td>$N$</td>
</tr>
<tr>
<td>$D$</td>
<td>Two dogs play with tennis ball in field</td>
<td>Dogs are watching a tennis match.</td>
<td>$N$</td>
<td>C</td>
<td>C</td>
<td>C</td>
<td>C</td>
</tr>
<tr>
<td>E</td>
<td>Two kids begin to make a snowman on a sunny winter day.</td>
<td>Two penguins making a snowman.</td>
<td>$N$</td>
<td>C</td>
<td>C</td>
<td>$C$</td>
<td>C</td>
</tr>
<tr>
<td>$F$</td>
<td>The horses pull the carriage, holding people and a dog, through the rain.</td>
<td>Horses ride in a carriage pulled by a dog.</td>
<td>$E$</td>
<td>$E$</td>
<td>C</td>
<td>$C$</td>
<td>C</td>
</tr>
<tr>
<td>$G$</td>
<td>A woman closes her eyes as she plays her cello.</td>
<td>The woman has her eyes open.</td>
<td>$E$</td>
<td>$E$</td>
<td>$E$</td>
<td>E</td>
<td>$C$</td>
</tr>
<tr>
<td>$H$</td>
<td>Two women having drinks and smoking cigarettes at the bar.</td>
<td>Three women are at a bar.</td>
<td>$E$</td>
<td>$E$</td>
<td>$E$</td>
<td>$E$</td>
<td>C</td>
</tr>
<tr>
<td>I</td>
<td>hand</td>
<td>A band watches the fans play</td>
<td>$E$</td>
<td>$E$</td>
<td>$E$</td>
<td>$E$</td>
<td>C</td>
</tr>
</tbody>
</table>
Table 3: Example wins and losses compared to other approaches. DA (Decomposable Attention) refers to our approach while
SPINN-PI and mLSTM are previously developed methods (see Table 1).
parentheses: network size(2-layers, each with 200 neurons), batch size (4), 1 dropout ratio (0.2) and learning rate (0.05-vanilla, 0.025-intra-attention). All settings were run for 50 million steps (each step indicates one batch) but model parameters were saved frequently as training progressed and we chose the model that did best on the development set.
## 5.2 Results
our vanilla approach is incorrect but mLSTM and SPINN-PI are correct. Example F requires sequential information and neither variant of our approach can predict the correct class. Examples D-E are interesting however, since they don't require word order information, yet intra-attention seems to help. We suspect this may be because the word embeddings are not fine-grained enough for the algorithm to conclude that “play/watch”is a contradiction, but intra-attention, by adding an extra layer of composition/nonlinearity to incorporate context, compensates for this.
Finally, Examples G-I are cases that all methods get wrong. The first is actually representative of many examples in this category where there is one critical word that separates the two sentences (close vs open in this case) and goes unnoticed by the algorithms. Examples H requires inference about numbers and Example I needs sequence information.
## 6 Conclusion
Results in terms of 3-class accuracy are shown in Table 1. Our vanilla approach achieves state-of-theart results with almost an order of magnitude fewer parameters than the LSTMN of Cheng et al. (2016). Adding intra-sentence attention gives a considerable improvement of 0.5 percentage points over the existing state of the art. Table 2 gives a breakdown of accuracy on the development set showing that most of our gains stem from $neutral$,while most losses come from contradiction pairs.
Table 3 shows some wins and losses. Examples AC are cases where both variants of our approach are correct while both SPINN-PI (Bowman et al., 2016) and the mLSTM (Wang and Jiang, 2016) are incorrect. In the first two cases, both sentences contain phrases that are either identical or highly lexically related (e.g.“Two kids”and“ocean/beach”) and our approach correctly favors neutral in these cases. In pairwise comparisons are relatively more important
Example C, it is possible that relying on word-order may confuse SPINN-PI and the mLSTM due to how “fountain" is the object of a preposition in the first sentence but the subject of the second.
The second set of examples (D-F) are cases where
We presented a simple attention-based approach to natural language inference that is trivially parallelizable. The approach outperforms considerably more complex neural methods aiming for text understanding. Our results suggest that, at least for this task, than global sentence-level representations.
## Acknowledgements
We thank Slav Petrov, Tom Kwiatkowski, Yoon Kim, Erick Fonseca, Mark Neumann for useful discussion and Sam Bowman and Shuohang Wang for providing us their model outputs for error analysis.
$^{1}$16 or 32 also work well and are a bit more stable.
|
[
[
0.59228515625,
0.0921630859375,
0.65283203125,
0.0921630859375,
0.65283203125,
0.10418701171875,
0.59228515625,
0.10418701171875
],
[
0.66015625,
0.091552734375,
0.73193359375,
0.091552734375,
0.73193359375,
0.1029052734375,
0.66015625,
0.1029052734375
],
[
0.736328125,
0.0921630859375,
0.7890625,
0.0921630859375,
0.7890625,
0.10418701171875,
0.736328125,
0.10418701171875
],
[
0.7998046875,
0.0921630859375,
0.84228515625,
0.0921630859375,
0.84228515625,
0.10418701171875,
0.7998046875,
0.10418701171875
],
[
0.8486328125,
0.0921630859375,
0.8798828125,
0.0921630859375,
0.8798828125,
0.10418701171875,
0.8486328125,
0.10418701171875
],
[
0.148681640625,
0.09344482421875,
0.200927734375,
0.09344482421875,
0.200927734375,
0.10418701171875,
0.148681640625,
0.10418701171875
],
[
0.125,
0.09344482421875,
0.1405029296875,
0.09344482421875,
0.1405029296875,
0.1029052734375,
0.125,
0.1029052734375
],
[
0.370849609375,
0.09344482421875,
0.423095703125,
0.09344482421875,
0.423095703125,
0.10418701171875,
0.370849609375,
0.10418701171875
],
[
0.1478271484375,
0.1060791015625,
0.366943359375,
0.106689453125,
0.366943359375,
0.121826171875,
0.1478271484375,
0.12060546875
],
[
0.35693359375,
0.10986328125,
0.54833984375,
0.11236572265625,
0.54833984375,
0.126953125,
0.35693359375,
0.125
],
[
0.68701171875,
0.11236572265625,
0.705078125,
0.11236572265625,
0.705078125,
0.1275634765625,
0.68701171875,
0.1275634765625
],
[
0.7548828125,
0.11297607421875,
0.771484375,
0.11297607421875,
0.771484375,
0.1275634765625,
0.7548828125,
0.1275634765625
],
[
0.81201171875,
0.11297607421875,
0.82861328125,
0.11297607421875,
0.82861328125,
0.1275634765625,
0.81201171875,
0.1275634765625
],
[
0.85693359375,
0.11297607421875,
0.87353515625,
0.11297607421875,
0.87353515625,
0.1275634765625,
0.85693359375,
0.1275634765625
],
[
0.615234375,
0.1142578125,
0.62841796875,
0.1142578125,
0.62841796875,
0.1256103515625,
0.615234375,
0.1256103515625
],
[
0.1298828125,
0.11614990234375,
0.1356201171875,
0.11614990234375,
0.1356201171875,
0.121826171875,
0.1298828125,
0.121826171875
],
[
0.146240234375,
0.125,
0.473876953125,
0.1319580078125,
0.473876953125,
0.149658203125,
0.146240234375,
0.1427001953125
],
[
0.1502685546875,
0.11932373046875,
0.2034912109375,
0.11932373046875,
0.2034912109375,
0.1300048828125,
0.1502685546875,
0.1300048828125
],
[
0.61279296875,
0.1337890625,
0.6298828125,
0.1337890625,
0.6298828125,
0.1483154296875,
0.61279296875,
0.1483154296875
],
[
0.68701171875,
0.1337890625,
0.705078125,
0.1337890625,
0.705078125,
0.1483154296875,
0.68701171875,
0.1483154296875
],
[
0.7548828125,
0.1337890625,
0.771484375,
0.1337890625,
0.771484375,
0.1483154296875,
0.7548828125,
0.1483154296875
],
[
0.81201171875,
0.1337890625,
0.830078125,
0.1337890625,
0.830078125,
0.1483154296875,
0.81201171875,
0.1483154296875
],
[
0.85546875,
0.1337890625,
0.87353515625,
0.1337890625,
0.87353515625,
0.1483154296875,
0.85546875,
0.1483154296875
],
[
0.15283203125,
0.143310546875,
0.2181396484375,
0.143310546875,
0.2181396484375,
0.15087890625,
0.15283203125,
0.15087890625
],
[
0.61376953125,
0.149658203125,
0.6298828125,
0.149658203125,
0.6298828125,
0.1654052734375,
0.61376953125,
0.1654052734375
],
[
0.1282958984375,
0.1383056640625,
0.1356201171875,
0.1383056640625,
0.1356201171875,
0.143310546875,
0.1282958984375,
0.143310546875
],
[
0.68701171875,
0.15087890625,
0.705078125,
0.15087890625,
0.705078125,
0.1654052734375,
0.68701171875,
0.1654052734375
],
[
0.75732421875,
0.15087890625,
0.76953125,
0.15087890625,
0.76953125,
0.1641845703125,
0.75732421875,
0.1641845703125
],
[
0.81201171875,
0.149658203125,
0.8310546875,
0.149658203125,
0.8310546875,
0.1641845703125,
0.81201171875,
0.1641845703125
],
[
0.85546875,
0.15087890625,
0.875,
0.15087890625,
0.875,
0.1654052734375,
0.85546875,
0.1654052734375
],
[
0.369384765625,
0.1514892578125,
0.57421875,
0.1514892578125,
0.57421875,
0.1622314453125,
0.369384765625,
0.1622314453125
],
[
0.1282958984375,
0.1546630859375,
0.1356201171875,
0.1546630859375,
0.1356201171875,
0.160400390625,
0.1282958984375,
0.160400390625
],
[
0.148681640625,
0.1514892578125,
0.34228515625,
0.1514892578125,
0.34228515625,
0.1654052734375,
0.148681640625,
0.1654052734375
],
[
0.689453125,
0.1641845703125,
0.70751953125,
0.1673583984375,
0.70166015625,
0.18310546875,
0.6845703125,
0.179931640625
],
[
0.75634765625,
0.1654052734375,
0.77294921875,
0.1654052734375,
0.77294921875,
0.1817626953125,
0.75634765625,
0.1817626953125
],
[
0.814453125,
0.1641845703125,
0.8310546875,
0.1673583984375,
0.826171875,
0.18310546875,
0.8095703125,
0.179931640625
],
[
0.61376953125,
0.1666259765625,
0.6298828125,
0.1666259765625,
0.6298828125,
0.1817626953125,
0.61376953125,
0.1817626953125
],
[
0.369384765625,
0.1666259765625,
0.53173828125,
0.1666259765625,
0.53173828125,
0.179931640625,
0.369384765625,
0.179931640625
],
[
0.85693359375,
0.1666259765625,
0.87353515625,
0.1666259765625,
0.87353515625,
0.18115234375,
0.85693359375,
0.18115234375
],
[
0.148681640625,
0.1685791015625,
0.328369140625,
0.1685791015625,
0.328369140625,
0.1787109375,
0.148681640625,
0.1787109375
],
[
0.1298828125,
0.17041015625,
0.1356201171875,
0.17041015625,
0.1356201171875,
0.175537109375,
0.1298828125,
0.175537109375
],
[
0.148681640625,
0.177978515625,
0.361083984375,
0.1787109375,
0.361083984375,
0.1900634765625,
0.148681640625,
0.189453125
],
[
0.62158203125,
0.1787109375,
0.6357421875,
0.189453125,
0.62158203125,
0.199462890625,
0.60888671875,
0.189453125
],
[
0.68896484375,
0.1817626953125,
0.705078125,
0.1817626953125,
0.705078125,
0.1988525390625,
0.68896484375,
0.1988525390625
],
[
0.75634765625,
0.1817626953125,
0.77392578125,
0.1817626953125,
0.77392578125,
0.1976318359375,
0.75634765625,
0.1976318359375
],
[
0.81201171875,
0.1817626953125,
0.830078125,
0.1817626953125,
0.830078125,
0.1976318359375,
0.81201171875,
0.1976318359375
],
[
0.85888671875,
0.184326171875,
0.87109375,
0.184326171875,
0.87109375,
0.1962890625,
0.85888671875,
0.1962890625
],
[
0.36767578125,
0.1849365234375,
0.53125,
0.1849365234375,
0.53125,
0.1956787109375,
0.36767578125,
0.1956787109375
],
[
0.12744140625,
0.1849365234375,
0.1380615234375,
0.1849365234375,
0.1380615234375,
0.1944580078125,
0.12744140625,
0.1944580078125
],
[
0.1446533203125,
0.1932373046875,
0.564453125,
0.2020263671875,
0.564453125,
0.222900390625,
0.1446533203125,
0.2139892578125
],
[
0.1502685546875,
0.1900634765625,
0.2034912109375,
0.1900634765625,
0.2034912109375,
0.2020263671875,
0.1502685546875,
0.2020263671875
],
[
0.61376953125,
0.2052001953125,
0.6298828125,
0.2052001953125,
0.6298828125,
0.2197265625,
0.61376953125,
0.2197265625
],
[
0.68701171875,
0.2052001953125,
0.70361328125,
0.2052001953125,
0.70361328125,
0.2197265625,
0.68701171875,
0.2197265625
],
[
0.75634765625,
0.2052001953125,
0.77294921875,
0.2052001953125,
0.77294921875,
0.2197265625,
0.75634765625,
0.2197265625
],
[
0.81201171875,
0.2052001953125,
0.830078125,
0.2052001953125,
0.830078125,
0.2197265625,
0.81201171875,
0.2197265625
],
[
0.85693359375,
0.2052001953125,
0.87353515625,
0.2052001953125,
0.87353515625,
0.2197265625,
0.85693359375,
0.2197265625
],
[
0.1519775390625,
0.212158203125,
0.27685546875,
0.212158203125,
0.27685546875,
0.222900390625,
0.1519775390625,
0.222900390625
],
[
0.1446533203125,
0.2197265625,
0.515625,
0.2279052734375,
0.515625,
0.248779296875,
0.1446533203125,
0.240478515625
],
[
0.61376953125,
0.2310791015625,
0.6298828125,
0.2310791015625,
0.6298828125,
0.24560546875,
0.61376953125,
0.24560546875
],
[
0.68701171875,
0.2310791015625,
0.705078125,
0.2310791015625,
0.705078125,
0.24560546875,
0.68701171875,
0.24560546875
],
[
0.7548828125,
0.2310791015625,
0.77294921875,
0.2310791015625,
0.77294921875,
0.24560546875,
0.7548828125,
0.24560546875
],
[
0.81201171875,
0.2310791015625,
0.82861328125,
0.2310791015625,
0.82861328125,
0.24560546875,
0.81201171875,
0.24560546875
],
[
0.85693359375,
0.2298583984375,
0.87353515625,
0.2298583984375,
0.87353515625,
0.24560546875,
0.85693359375,
0.24560546875
],
[
0.1519775390625,
0.240478515625,
0.1748046875,
0.240478515625,
0.1748046875,
0.248779296875,
0.1519775390625,
0.248779296875
],
[
0.12744140625,
0.23291015625,
0.1396484375,
0.23291015625,
0.1396484375,
0.2412109375,
0.12744140625,
0.2412109375
],
[
0.143798828125,
0.2412109375,
0.496826171875,
0.25,
0.496826171875,
0.270751953125,
0.143798828125,
0.261962890625
],
[
0.62158203125,
0.250732421875,
0.6357421875,
0.261474609375,
0.62353515625,
0.271484375,
0.60888671875,
0.260009765625
],
[
0.68701171875,
0.25390625,
0.705078125,
0.25390625,
0.705078125,
0.268310546875,
0.68701171875,
0.268310546875
],
[
0.7548828125,
0.25244140625,
0.77294921875,
0.25244140625,
0.77294921875,
0.267578125,
0.7548828125,
0.267578125
],
[
0.81201171875,
0.25390625,
0.830078125,
0.25390625,
0.830078125,
0.268310546875,
0.81201171875,
0.268310546875
],
[
0.85546875,
0.25390625,
0.87353515625,
0.25390625,
0.87353515625,
0.267578125,
0.85546875,
0.267578125
],
[
0.148681640625,
0.261474609375,
0.2459716796875,
0.261474609375,
0.2459716796875,
0.271484375,
0.148681640625,
0.271484375
],
[
0.61376953125,
0.26953125,
0.6298828125,
0.26953125,
0.6298828125,
0.2841796875,
0.61376953125,
0.2841796875
],
[
0.1298828125,
0.2568359375,
0.136474609375,
0.2568359375,
0.136474609375,
0.26318359375,
0.1298828125,
0.26318359375
],
[
0.148681640625,
0.270751953125,
0.316162109375,
0.270751953125,
0.316162109375,
0.2841796875,
0.148681640625,
0.2841796875
],
[
0.370849609375,
0.270751953125,
0.50732421875,
0.271484375,
0.50732421875,
0.28271484375,
0.370849609375,
0.2822265625
],
[
0.689453125,
0.270751953125,
0.70166015625,
0.270751953125,
0.70166015625,
0.2822265625,
0.689453125,
0.2822265625
],
[
0.7548828125,
0.26953125,
0.77294921875,
0.26953125,
0.77294921875,
0.28271484375,
0.7548828125,
0.28271484375
],
[
0.814453125,
0.270751953125,
0.82666015625,
0.270751953125,
0.82666015625,
0.28271484375,
0.814453125,
0.28271484375
],
[
0.85888671875,
0.270751953125,
0.87109375,
0.270751953125,
0.87109375,
0.28271484375,
0.85888671875,
0.28271484375
],
[
0.11358642578125,
0.292236328125,
0.8837890625,
0.29345703125,
0.8837890625,
0.30810546875,
0.11358642578125,
0.306884765625
],
[
0.11517333984375,
0.311279296875,
0.54345703125,
0.311279296875,
0.54345703125,
0.324462890625,
0.11517333984375,
0.324462890625
],
[
0.11517333984375,
0.3359375,
0.488525390625,
0.3359375,
0.488525390625,
0.35107421875,
0.11517333984375,
0.35107421875
],
[
0.50634765625,
0.337158203125,
0.8837890625,
0.3359375,
0.8837890625,
0.35107421875,
0.50634765625,
0.3515625
],
[
0.11517333984375,
0.354736328125,
0.491943359375,
0.354736328125,
0.491943359375,
0.36865234375,
0.11517333984375,
0.36865234375
],
[
0.5087890625,
0.354736328125,
0.8837890625,
0.354736328125,
0.8837890625,
0.369384765625,
0.5087890625,
0.369384765625
],
[
0.11517333984375,
0.371826171875,
0.491943359375,
0.371826171875,
0.491943359375,
0.386474609375,
0.11517333984375,
0.386474609375
],
[
0.5087890625,
0.37255859375,
0.8837890625,
0.37255859375,
0.8837890625,
0.386474609375,
0.5087890625,
0.386474609375
],
[
0.1168212890625,
0.389404296875,
0.490234375,
0.389404296875,
0.490234375,
0.40283203125,
0.1168212890625,
0.40283203125
],
[
0.50634765625,
0.389404296875,
0.8857421875,
0.38818359375,
0.8857421875,
0.40283203125,
0.50634765625,
0.404052734375
],
[
0.11517333984375,
0.4072265625,
0.491943359375,
0.4072265625,
0.491943359375,
0.42041015625,
0.11517333984375,
0.42041015625
],
[
0.50732421875,
0.4072265625,
0.8857421875,
0.4072265625,
0.8857421875,
0.42041015625,
0.50732421875,
0.42041015625
],
[
0.1168212890625,
0.423583984375,
0.491943359375,
0.423583984375,
0.491943359375,
0.438232421875,
0.1168212890625,
0.438232421875
],
[
0.50732421875,
0.423583984375,
0.8837890625,
0.423583984375,
0.8837890625,
0.438232421875,
0.50732421875,
0.438232421875
],
[
0.11517333984375,
0.44140625,
0.42724609375,
0.44140625,
0.42724609375,
0.455078125,
0.11517333984375,
0.455078125
],
[
0.5087890625,
0.43994140625,
0.8857421875,
0.439453125,
0.8857421875,
0.453857421875,
0.5087890625,
0.455078125
],
[
0.5107421875,
0.458251953125,
0.88232421875,
0.458251953125,
0.88232421875,
0.4716796875,
0.5107421875,
0.4716796875
],
[
0.11517333984375,
0.4697265625,
0.2156982421875,
0.4697265625,
0.2156982421875,
0.484130859375,
0.11517333984375,
0.484130859375
],
[
0.5087890625,
0.474853515625,
0.88232421875,
0.474853515625,
0.88232421875,
0.488037109375,
0.5087890625,
0.488037109375
],
[
0.11517333984375,
0.492431640625,
0.490234375,
0.49365234375,
0.490234375,
0.50830078125,
0.11517333984375,
0.5068359375
],
[
0.5087890625,
0.49365234375,
0.8837890625,
0.49365234375,
0.8837890625,
0.50390625,
0.5087890625,
0.50390625
],
[
0.11517333984375,
0.51123046875,
0.49267578125,
0.51123046875,
0.49267578125,
0.5244140625,
0.11517333984375,
0.5244140625
],
[
0.50732421875,
0.5087890625,
0.88232421875,
0.5087890625,
0.88232421875,
0.52294921875,
0.50732421875,
0.52294921875
],
[
0.11517333984375,
0.52783203125,
0.491943359375,
0.52783203125,
0.491943359375,
0.54150390625,
0.11517333984375,
0.54150390625
],
[
0.5087890625,
0.52587890625,
0.56787109375,
0.52587890625,
0.56787109375,
0.54052734375,
0.5087890625,
0.54052734375
],
[
0.1168212890625,
0.54345703125,
0.488525390625,
0.54345703125,
0.488525390625,
0.55810546875,
0.1168212890625,
0.55810546875
],
[
0.52685546875,
0.54248046875,
0.88232421875,
0.54248046875,
0.88232421875,
0.556640625,
0.52685546875,
0.556640625
],
[
0.1168212890625,
0.56298828125,
0.491943359375,
0.56298828125,
0.491943359375,
0.5771484375,
0.1168212890625,
0.5771484375
],
[
0.50732421875,
0.56005859375,
0.88134765625,
0.5595703125,
0.88134765625,
0.57373046875,
0.50732421875,
0.57470703125
],
[
0.1168212890625,
0.580078125,
0.495849609375,
0.580078125,
0.495849609375,
0.5947265625,
0.1168212890625,
0.5947265625
],
[
0.50732421875,
0.5771484375,
0.8798828125,
0.5771484375,
0.8798828125,
0.59130859375,
0.50732421875,
0.59130859375
],
[
0.11517333984375,
0.5966796875,
0.494384765625,
0.5966796875,
0.494384765625,
0.611328125,
0.11517333984375,
0.611328125
],
[
0.50732421875,
0.59521484375,
0.8837890625,
0.59521484375,
0.8837890625,
0.609375,
0.50732421875,
0.609375
],
[
0.11517333984375,
0.6142578125,
0.49267578125,
0.6142578125,
0.49267578125,
0.62744140625,
0.11517333984375,
0.62744140625
],
[
0.50732421875,
0.6123046875,
0.8857421875,
0.6123046875,
0.8857421875,
0.62548828125,
0.50732421875,
0.62548828125
],
[
0.11517333984375,
0.630859375,
0.491943359375,
0.630859375,
0.491943359375,
0.64453125,
0.11517333984375,
0.64453125
],
[
0.5107421875,
0.62890625,
0.8837890625,
0.62890625,
0.8837890625,
0.64306640625,
0.5107421875,
0.64306640625
],
[
0.11517333984375,
0.64794921875,
0.340576171875,
0.64794921875,
0.340576171875,
0.662109375,
0.11517333984375,
0.662109375
],
[
0.5107421875,
0.646484375,
0.7939453125,
0.646484375,
0.7939453125,
0.66015625,
0.5107421875,
0.66015625
],
[
0.13232421875,
0.6640625,
0.491943359375,
0.66552734375,
0.491943359375,
0.6796875,
0.13232421875,
0.6787109375
],
[
0.51123046875,
0.673828125,
0.5390625,
0.673828125,
0.5390625,
0.681640625,
0.51123046875,
0.681640625
],
[
0.5361328125,
0.671875,
0.6337890625,
0.671875,
0.6337890625,
0.68603515625,
0.5361328125,
0.68603515625
],
[
0.11517333984375,
0.68310546875,
0.490234375,
0.68310546875,
0.490234375,
0.6962890625,
0.11517333984375,
0.6962890625
],
[
0.11517333984375,
0.69970703125,
0.491943359375,
0.69970703125,
0.491943359375,
0.71337890625,
0.11517333984375,
0.71337890625
],
[
0.50732421875,
0.6943359375,
0.88134765625,
0.6943359375,
0.88134765625,
0.70849609375,
0.50732421875,
0.70849609375
],
[
0.11358642578125,
0.71630859375,
0.494384765625,
0.71630859375,
0.494384765625,
0.72998046875,
0.11358642578125,
0.72998046875
],
[
0.50732421875,
0.7119140625,
0.8857421875,
0.7119140625,
0.8857421875,
0.7265625,
0.50732421875,
0.7265625
],
[
0.1168212890625,
0.73486328125,
0.488525390625,
0.73486328125,
0.488525390625,
0.74560546875,
0.1168212890625,
0.74560546875
],
[
0.50732421875,
0.72998046875,
0.8837890625,
0.72998046875,
0.8837890625,
0.74365234375,
0.50732421875,
0.74365234375
],
[
0.11517333984375,
0.75048828125,
0.488525390625,
0.75,
0.488525390625,
0.7646484375,
0.11517333984375,
0.76513671875
],
[
0.50732421875,
0.74755859375,
0.8857421875,
0.74755859375,
0.8857421875,
0.76123046875,
0.50732421875,
0.76123046875
],
[
0.11517333984375,
0.76806640625,
0.490234375,
0.76806640625,
0.490234375,
0.7822265625,
0.11517333984375,
0.7822265625
],
[
0.50732421875,
0.7646484375,
0.8857421875,
0.7646484375,
0.8857421875,
0.77880859375,
0.50732421875,
0.77880859375
],
[
0.50634765625,
0.77978515625,
0.88134765625,
0.77978515625,
0.88134765625,
0.79443359375,
0.50634765625,
0.79443359375
],
[
0.1168212890625,
0.7861328125,
0.490234375,
0.7861328125,
0.490234375,
0.7998046875,
0.1168212890625,
0.7998046875
],
[
0.1168212890625,
0.80322265625,
0.491943359375,
0.80322265625,
0.491943359375,
0.81640625,
0.1168212890625,
0.81640625
],
[
0.50732421875,
0.798828125,
0.814453125,
0.798828125,
0.814453125,
0.8125,
0.50732421875,
0.8125
],
[
0.1168212890625,
0.8193359375,
0.488525390625,
0.8193359375,
0.488525390625,
0.83349609375,
0.1168212890625,
0.83349609375
],
[
0.51123046875,
0.82763671875,
0.6748046875,
0.82763671875,
0.6748046875,
0.84228515625,
0.51123046875,
0.84228515625
],
[
0.11273193359375,
0.83544921875,
0.490234375,
0.8359375,
0.490234375,
0.85107421875,
0.11273193359375,
0.849609375
],
[
0.11517333984375,
0.85400390625,
0.39208984375,
0.85400390625,
0.39208984375,
0.8671875,
0.11517333984375,
0.8671875
],
[
0.5087890625,
0.8515625,
0.88232421875,
0.8515625,
0.88232421875,
0.86572265625,
0.5087890625,
0.86572265625
],
[
0.134033203125,
0.8720703125,
0.490234375,
0.8720703125,
0.490234375,
0.88525390625,
0.134033203125,
0.88525390625
],
[
0.5087890625,
0.8671875,
0.88232421875,
0.8671875,
0.88232421875,
0.880859375,
0.5087890625,
0.880859375
],
[
0.5087890625,
0.88330078125,
0.88232421875,
0.88330078125,
0.88232421875,
0.896484375,
0.5087890625,
0.896484375
],
[
0.134033203125,
0.896484375,
0.4345703125,
0.89794921875,
0.4345703125,
0.912109375,
0.134033203125,
0.9111328125
],
[
0.5087890625,
0.8984375,
0.80126953125,
0.8984375,
0.80126953125,
0.912109375,
0.5087890625,
0.912109375
]
] |
[
"DA (vanilla)",
"DA (intra att.)",
"SPINN-PI",
"mLSTM",
"Gold",
"Sentence 1",
"ID",
"Sentence 2",
"Two kids are standing in the ocean hugging",
" Two kids enjoy their day at the beach.",
"N",
"E",
"E",
"N",
"N",
"A",
"Adancerincostumerperformsonstagewhilehemaniscaptivated",
"cach other.",
"N",
"N",
"E",
"E",
"N",
"amanwatches",
"N",
"B",
"N",
"c",
"c",
"N",
"The fountain is splashing the persons seated",
"C",
"They are stting on the edge of a fountain",
"c",
"c",
"c",
"N",
"Dogs are watching a tennis match.",
"c",
"Two dogs play with tennis ball in field",
"D",
"Two kids begin to make a snowman on a sunny",
"1",
"c",
"c",
"c",
"c",
"Two penguins making a snowman",
"E",
"urdinsrrule",
"winter day.",
"E",
"E",
"c",
"c",
"c",
"and a dog, through the rain",
"Awomancloseshereyesasshelysherhewomanhashereyesopen",
"E",
"E",
"E",
"E",
"c",
"cello",
"G",
"wo women having drinks and smoking Thre womenareata bar.",
"A",
"E",
"E",
"E",
"c",
"cigarettes at the bar.",
"E",
"H",
"A band playing with fans watching.",
"A band watches the fans play",
"E",
"E",
"E",
"c",
"Table 3: Example wins and losses compared to other approaches. DA (Decomposable Attention) refers to our approach while",
"SPINN-PI and mLSTM are previously developed methods (see Table 1).",
"parentheses: network size (2-layers, each with 200",
"our vanilla approach is incorrect but mLSTM and",
"neurons), batch size (4), dropout ratio (0.2) and",
"SPINN-PI are correct. Example F requires sequen-",
"learning rate (0.05-vanilla, 0.025-intra-attention).",
"tial information and neither variant of our approach",
"All settings were run for 50 million steps (each step",
"can predict the correct class. Examples D-E are in-",
"indicates one batch) but model parameters were saved",
"teresting however, since they don't require word or-",
"frequently as training progressed and we chose the",
"der information, yet intra-attention seems to help",
"model that did best on the development set.",
"We suspect this may be because the word embed-",
"dings are not fine-grained enough for the algorithm",
"5.2 Results",
"to conclude that “play/watch\"is a contradiction, but",
"Results in terms of 3-class accuracy are shown in",
"intra-attention,byaddinganextralayerofcomposi-",
"Table 1. Our vanilla approach achieves state-of-the-",
"tion/nonlinearity to incorporate context, compensates",
"art results with almost an order of magnitude fewer",
"for this.",
"parameters than the LSTMN of Cheng et al. (2016)",
"Finally, Examples G-I are cases that all methods",
"Adding intra-sentence attention gives a considerable",
"get wrong. The first is actually representative of many",
"improvement of 0.5 percentage points over the ex-",
"examples in this category where there is one critical",
"isting state of the art. Table 2 gives a breakdown of",
"word that separates the two sentences (close vs open",
"accuracy on the development set showing that most",
"in this case) and goes unnoticed by the algorithms.",
"of our gains stem from neutral, while most losses",
"Examples H requires inference about numbers and",
"come from contradiction pairs.",
"Example I needs sequence information.",
"Table 3 shows some wins and losses. Examples A-",
"6",
"Conclusion",
"C are cases where both variants of our approach are",
"correct while both SPINN-PI (Bowman et al., 2016)",
"We presented a simple attention-based approach to",
"and the mLSTM (Wang and Jiang, 2016) are incor-",
"natural language inference that is trivially paralleliz-",
"rect.Inthefirsttwocases,bothsentencescontain",
"able. The approach outperforms considerably more",
"phrases that are either identical or highly lexically",
"complex neural methods aiming for text understand-",
"related (e.g.\"\"Two kids\"and “ocean /beach\") and our",
"ing. Our results suggest that, at least for this task,",
" pairwise comparisons are relatively more important",
"approach correctly favors neutral in these cases. In",
"Example C, it is possible that relying on word-order",
"than global sentence-level representations.",
"may confuse SPINN-PI and the mLSTM due to how",
"Acknowledgements",
"“\"fountain\" is the object of a preposition in the first",
"sentence but the subject of the second.",
"We thank Slav Petrov, Tom Kwiatkowski, Yoon Kim.",
"The second set of examples (D-F) are cases where",
"Erick Fonseca, Mark Neumann for useful discussion",
"and Sam Bowman and Shuohang Wang for providing",
"'16 or 32 also work well and are a bit more stable.",
"us their model outputs for error analysis."
] |
{
"arxiv_id": "1606.01933",
"language": "en",
"page_idx": 4
}
|
||
2007.03630-en-0
|
2007.03630-en-0.jpg
|
The CMS monitoring infrastructure and applications
$Christian\text{ Ariza-Porras}^{1,*},Valentin\text{ Kuznetsov}^{2,**}$, and Federica Legger$^{3,***}$
$^{1}$CERN, Geneva, Switzerland
$^{2}$Cornell University, Ithaca, NY, USA 14850
$^{3}$Istituto Nazionale di Fisica Nucleare, via Pietro Giuria 1, Torino, Italy
Abstract. The globally distributed computing infrastructure required to cope with the multi-petabytes datasets produced by the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider (LHC) at CERN comprises several subsystems, such as workload management, data management, data transfers, and submission of users' and centrally managed production requests The performance and status of all subsystems must be constantly monitored to guarantee the efficient operation of the whole infrastructure. Moreover, key metrics need to be tracked to evaluate and study the system performance over time. The CMS monitoring architecture allows both real-time and historical monitoring of a variety of data sources and is based on scalable and open source solutions tailored to satisfy the experiment's monitoring needs. We present the monitoring data flow and software architecture for the CMS distributed computing applications. We discuss the challenges, components, current achievements, and future developments of the CMS monitoring infrastructure.
### 1 Introduction
Data from the CMS experiment[1] at the LHC are stored and processed in a tiered distributed computing infrastructure. More than one hundred computing centers worldwide, connected through a set of grid services responsible for storage, computing, and connectivity are used to process LHC data and produce Monte Carlo simulated events of relevant physics processes A recent overview of the computing model of the LHC experiments can be found in [2].
The CMS distributed computing infrastructure includes central services and middleware that handle authentication, workload management, and data management. The workload management system executes payloads in compute nodes provisioned through GlideinWMS [3] and made available as execution slots in a Vanilla Universe HTCondor pool [4]. HTCondor jobs are submitted via specific workload management tools: WMAgent for central data processing and Monte Carlo production jobs, and CRAB for user jobs [5]. The data management system is modular, and includes several components: PhEDEx, the data transfer and location system; DBS, the Data Bookkeeping Service, a metadata catalog; and DAS, the Data Aggregation Service designed to aggregate views and provide them to users and services[6]. Data from these services are available to CMS collaborators through a web suite of services known as CMSWEB
$^*$e-mail: christian.ariza.porras@cern.ch
$^{**}$e-mail: vkuznet@protonmail.com $^{***}$e-mail: federica.legger@cern.ch
|
[
[
0.11932373046875,
0.184814453125,
0.79443359375,
0.184814453125,
0.79443359375,
0.203125,
0.11932373046875,
0.203125
],
[
0.029052734375,
0.1995849609375,
0.07470703125,
0.1995849609375,
0.07470703125,
0.269287109375,
0.029052734375,
0.269287109375
],
[
0.11517333984375,
0.2305908203125,
0.74169921875,
0.227783203125,
0.74169921875,
0.2489013671875,
0.11517333984375,
0.251708984375
],
[
0.029052734375,
0.260986328125,
0.07470703125,
0.260986328125,
0.07470703125,
0.330078125,
0.029052734375,
0.330078125
],
[
0.1182861328125,
0.254638671875,
0.34130859375,
0.256103515625,
0.34130859375,
0.270751953125,
0.1182861328125,
0.269287109375
],
[
0.0435791015625,
0.262939453125,
0.05291748046875,
0.262939453125,
0.05291748046875,
0.2666015625,
0.0435791015625,
0.2666015625
],
[
0.1162109375,
0.271484375,
0.44921875,
0.271484375,
0.44921875,
0.286376953125,
0.1162109375,
0.286376953125
],
[
0.1182861328125,
0.287109375,
0.64501953125,
0.287109375,
0.64501953125,
0.3017578125,
0.1182861328125,
0.3017578125
],
[
0.204345703125,
0.316650390625,
0.79443359375,
0.316650390625,
0.79443359375,
0.330078125,
0.204345703125,
0.330078125
],
[
0.02801513671875,
0.332763671875,
0.0726318359375,
0.332275390625,
0.07574462890625,
0.418212890625,
0.0311279296875,
0.419677734375
],
[
0.2054443359375,
0.332275390625,
0.79443359375,
0.332275390625,
0.79443359375,
0.345458984375,
0.2054443359375,
0.345458984375
],
[
0.204345703125,
0.34765625,
0.79443359375,
0.34765625,
0.79443359375,
0.361083984375,
0.204345703125,
0.361083984375
],
[
0.204345703125,
0.36328125,
0.79443359375,
0.36328125,
0.79443359375,
0.37646484375,
0.204345703125,
0.37646484375
],
[
0.204345703125,
0.378662109375,
0.79345703125,
0.378662109375,
0.79345703125,
0.39208984375,
0.204345703125,
0.39208984375
],
[
0.204345703125,
0.394287109375,
0.79443359375,
0.394287109375,
0.79443359375,
0.408203125,
0.204345703125,
0.408203125
],
[
0.204345703125,
0.40966796875,
0.79345703125,
0.40966796875,
0.79345703125,
0.423828125,
0.204345703125,
0.423828125
],
[
0.02490234375,
0.4267578125,
0.0726318359375,
0.4267578125,
0.07781982421875,
0.67138671875,
0.029052734375,
0.67138671875
],
[
0.204345703125,
0.424560546875,
0.79443359375,
0.424560546875,
0.79443359375,
0.43798828125,
0.204345703125,
0.43798828125
],
[
0.204345703125,
0.43994140625,
0.79443359375,
0.43994140625,
0.79443359375,
0.4541015625,
0.204345703125,
0.4541015625
],
[
0.204345703125,
0.45556640625,
0.79443359375,
0.45556640625,
0.79443359375,
0.4697265625,
0.204345703125,
0.4697265625
],
[
0.204345703125,
0.4716796875,
0.79443359375,
0.4716796875,
0.79443359375,
0.485107421875,
0.204345703125,
0.485107421875
],
[
0.2022705078125,
0.48583984375,
0.79345703125,
0.4873046875,
0.79345703125,
0.50048828125,
0.2022705078125,
0.5
],
[
0.204345703125,
0.5029296875,
0.79345703125,
0.5029296875,
0.79345703125,
0.51611328125,
0.204345703125,
0.51611328125
],
[
0.2022705078125,
0.517578125,
0.67724609375,
0.517578125,
0.67724609375,
0.53125,
0.2022705078125,
0.53125
],
[
0.11517333984375,
0.5791015625,
0.279052734375,
0.5791015625,
0.279052734375,
0.59326171875,
0.11517333984375,
0.59326171875
],
[
0.1162109375,
0.60986328125,
0.8828125,
0.60986328125,
0.8828125,
0.625,
0.1162109375,
0.625
],
[
0.1182861328125,
0.62744140625,
0.8828125,
0.62744140625,
0.8828125,
0.642578125,
0.1182861328125,
0.642578125
],
[
0.1162109375,
0.64453125,
0.8828125,
0.64453125,
0.8828125,
0.65869140625,
0.1162109375,
0.65869140625
],
[
0.1162109375,
0.662109375,
0.8828125,
0.66064453125,
0.8828125,
0.67578125,
0.1162109375,
0.67724609375
],
[
0.1182861328125,
0.67919921875,
0.84423828125,
0.67919921875,
0.84423828125,
0.6923828125,
0.1182861328125,
0.6923828125
],
[
0.1483154296875,
0.6953125,
0.8828125,
0.6953125,
0.8828125,
0.7099609375,
0.1483154296875,
0.7099609375
],
[
0.1162109375,
0.71240234375,
0.8837890625,
0.71240234375,
0.8837890625,
0.72705078125,
0.1162109375,
0.72705078125
],
[
0.1162109375,
0.728515625,
0.8837890625,
0.728515625,
0.8837890625,
0.7431640625,
0.1162109375,
0.7431640625
],
[
0.1162109375,
0.74560546875,
0.8837890625,
0.74560546875,
0.8837890625,
0.76025390625,
0.1162109375,
0.76025390625
],
[
0.1162109375,
0.76318359375,
0.8828125,
0.76318359375,
0.8828125,
0.77783203125,
0.1162109375,
0.77783203125
],
[
0.1162109375,
0.77978515625,
0.880859375,
0.77978515625,
0.880859375,
0.794921875,
0.1162109375,
0.794921875
],
[
0.1162109375,
0.79638671875,
0.8837890625,
0.79638671875,
0.8837890625,
0.81005859375,
0.1162109375,
0.81005859375
],
[
0.1162109375,
0.81298828125,
0.8828125,
0.81298828125,
0.8828125,
0.828125,
0.1162109375,
0.828125
],
[
0.1162109375,
0.82861328125,
0.8837890625,
0.82861328125,
0.8837890625,
0.84619140625,
0.1162109375,
0.84619140625
],
[
0.1162109375,
0.84765625,
0.8837890625,
0.84765625,
0.8837890625,
0.86083984375,
0.1162109375,
0.86083984375
],
[
0.11517333984375,
0.86328125,
0.296630859375,
0.86328125,
0.296630859375,
0.8779296875,
0.11517333984375,
0.8779296875
],
[
0.1358642578125,
0.88720703125,
0.402587890625,
0.88720703125,
0.402587890625,
0.900390625,
0.1358642578125,
0.900390625
],
[
0.1307373046875,
0.900390625,
0.372314453125,
0.900390625,
0.372314453125,
0.9140625,
0.1307373046875,
0.9140625
],
[
0.12139892578125,
0.9111328125,
0.3662109375,
0.91162109375,
0.3662109375,
0.93017578125,
0.12139892578125,
0.9287109375
]
] |
[
"The CMS monitoringinfrastructureand applications",
"000",
"Christian Ariza-Porras'*, Valentin Kuznetsov2.*, and Federica Legger3,*",
"[n L",
"ICERN,Geneva,Switzerland",
"",
"2Cornell University, Ithaca, NY, USA 14850",
"3Istituto Nazionale di Fisica Nucleare, via Pietro Giuria 1, Torino, Italy",
"Abstract.The globally distributed computing infrastructure required to cope",
"[S'S]",
"with the multi-petabytes datasets produced by the Compact Muon Solenoid",
"(CMS) experiment at the Large Hadron Collider (LHC) at CERN comprises",
"several subsystems, such as workload management, data management, data",
"transfers, and submission of users'and centrally managed production requests.",
"The performance and status of all subsystems must be constantly monitored",
"to guarantee the efficient operation of the whole infrastructure. Moreover, key",
"A59000A",
"metrics need to be tracked to evaluate and study the system performance over",
"time. The CMS monitoring architecture allows both real-time and historical",
"monitoring of a variety of data sources and is based on scalable and open source",
"solutions tailored to satisfy the experiment's monitoring needs. We present the",
"monitoringdata flowand software architectureforthe CMSdistributed comput-",
"ing applications. We discuss the challenges, components, current achievements,",
"and future developments of the CMS monitoring infrastructure.",
"1 Introduction",
"Data from the CMS experiment [1] at the LHC are stored and processed in a tiered distributed",
"computing infrastructure.More than one hundred computing centers worldwide, connected",
"through a set of grid services responsible for storage, computing, and connectivity are used to",
"process LHC data and produce Monte Carlo simulated events of relevant physics processes.",
"Arecent overview of the computing model of the LHC experiments can befound in [2].",
"The CMS distributed computing infrastructure includes central services and middleware",
"that handle authentication, workload management, and data management. The workload",
"management system executes payloads in compute nodes provisioned through Glidein WMS",
"[3] and made available as execution slots in a Vanilla Universe HTCondor pool [4]. HTCon-",
"dor jobs are submitted via specific workload management tools: WMAgent for central data",
"processing and Monte Carlo production jobs, and CRAB for user jobs [5]. The data man-",
"agement system is modular, and includes several components:PhEDEx, the data transfer and",
"location system; DBS, the Data Bookkeeping Service, a metadata catalog; and DAS, the Data",
"Aggregation Service designed to aggregate views and provide them to users and services [6].",
"Data from these services are available to CMS collaborators through a web suite of services",
"known as CMSWEB.",
"*e-mail: christian.ariza.porras@cern.ch",
"**e-mail: vkuznet@protonmail.com",
"***e-mail: federica.legger@cern.ch"
] |
{
"arxiv_id": "2007.03630",
"language": "en",
"page_idx": 0
}
|
||
2007.03630-en-1
|
2007.03630-en-1.jpg
|
<FigureHere>
Figure 1. The data flow in the CERN MONIT infrastructure. The various steps of a typical monitoring
workflow (data injection, transport, processing, storage, and access) are shown from left to right.
Until recently, most services were monitored through custom tools and web applications, and logging information was scattered over several sources and mostly accessible only by experts. The maintenance and operation of the monitoring applications were becoming increasingly time-consuming and complicated due to the high turnover of experts in the computing community.
Nowadays, several solutions are available to gather, store, and process large amounts of data such as those produced by monitoring and logging services of computing applications Many technologies are available under an open-source licence, and are developed and sup ported by large software companies and communities.
During the last two years, the CMS computing community has been gradually abandoning in-house solutions in favour of the adoption of widely-used technologies based on opensource, scalable, and non-SQL tools, such as Hadoop [7], InfluxDB [8], and ElasticSearch [9]. These services are available through MONIT [15], a central monitoring infrastructure provided by the CERN IT department, and allow for the easy deployment of monitoring and accounting applications using visualisation tools such as Kibana[10] and Grafana[11]. Grafana allows users to automatically raise alarms and send notifications when anomalous conditions in the monitoring data are met. Data sources from different subsystems can be used to build complex monitoring workflows and predictive analytics, and performance studies. To complement the monitoring infrastructure provided by the CERN IT department, the CMS monitoring group set up and runs additional monitoring applications based on technologies such as VictoriaMetrics [12], NATS [13], and Prometheus [14].
In the following sections, we describe MONIT, the monitoring infrastructure at CERN, the organisation of CMS monitoring applications based on MONIT and CMS own infrastructure, and future developments.
## 2 The CERN MONIT infrastructure
The CERN IT department offers a variety of monitoring services through the MONIT infrastructure. A typical monitoring workflow consists of the following steps (see also Fig 1):
$\bullet$ Data injection: data is pushed into the MONIT infrastructure through ActiveMQ[16] messaging. As an alternative, data providers inside the CERN network boundaries may use an HTTP endpoint. Log files may be directly injected using Logstash [17]
$\bullet$ Data transport and processing: Data are streamed into MONIT using Apache Kafka[18] Apache Spark[19] and Apache Flume [20] are used for further data processing. Data may
|
[
[
0.161865234375,
0.08746337890625,
0.2427978515625,
0.08953857421875,
0.2427978515625,
0.1036376953125,
0.161865234375,
0.1015625
],
[
0.365234375,
0.08953857421875,
0.4326171875,
0.08953857421875,
0.4326171875,
0.09661865234375,
0.365234375,
0.09661865234375
],
[
0.5654296875,
0.08746337890625,
0.6298828125,
0.08953857421875,
0.62744140625,
0.1036376953125,
0.5634765625,
0.1015625
],
[
0.363037109375,
0.0972900390625,
0.44189453125,
0.0972900390625,
0.44189453125,
0.10791015625,
0.363037109375,
0.10791015625
],
[
0.765625,
0.0924072265625,
0.80078125,
0.0924072265625,
0.80078125,
0.100830078125,
0.765625,
0.100830078125
],
[
0.1763916015625,
0.1099853515625,
0.2271728515625,
0.1099853515625,
0.2271728515625,
0.12127685546875,
0.1763916015625,
0.12127685546875
],
[
0.287353515625,
0.10577392578125,
0.319580078125,
0.10577392578125,
0.319580078125,
0.1170654296875,
0.287353515625,
0.1170654296875
],
[
0.5693359375,
0.1114501953125,
0.63671875,
0.1114501953125,
0.63671875,
0.12127685546875,
0.5693359375,
0.12127685546875
],
[
0.77783203125,
0.113525390625,
0.80712890625,
0.113525390625,
0.80712890625,
0.12060546875,
0.77783203125,
0.12060546875
],
[
0.56201171875,
0.11492919921875,
0.57177734375,
0.11492919921875,
0.57177734375,
0.11846923828125,
0.56201171875,
0.11846923828125
],
[
0.278076171875,
0.11492919921875,
0.32568359375,
0.11492919921875,
0.32568359375,
0.12548828125,
0.278076171875,
0.12548828125
],
[
0.35986328125,
0.1163330078125,
0.420166015625,
0.1142578125,
0.421142578125,
0.126953125,
0.362060546875,
0.1290283203125
],
[
0.179443359375,
0.13818359375,
0.22412109375,
0.13818359375,
0.22412109375,
0.1488037109375,
0.179443359375,
0.1488037109375
],
[
0.55615234375,
0.1431884765625,
0.62255859375,
0.1431884765625,
0.62255859375,
0.157958984375,
0.55615234375,
0.157958984375
],
[
0.75732421875,
0.146728515625,
0.8134765625,
0.145263671875,
0.8134765625,
0.1536865234375,
0.75732421875,
0.1551513671875
],
[
0.27587890625,
0.1650390625,
0.327880859375,
0.1650390625,
0.327880859375,
0.1756591796875,
0.27587890625,
0.1756591796875
],
[
0.35986328125,
0.1629638671875,
0.42333984375,
0.1629638671875,
0.42333984375,
0.177734375,
0.35986328125,
0.177734375
],
[
0.2841796875,
0.1749267578125,
0.319580078125,
0.1749267578125,
0.319580078125,
0.1826171875,
0.2841796875,
0.1826171875
],
[
0.1888427734375,
0.1663818359375,
0.219970703125,
0.1663818359375,
0.219970703125,
0.1749267578125,
0.1888427734375,
0.1749267578125
],
[
0.583984375,
0.177001953125,
0.615234375,
0.1820068359375,
0.61181640625,
0.191162109375,
0.5810546875,
0.1861572265625
],
[
0.75634765625,
0.1820068359375,
0.77783203125,
0.1798095703125,
0.78125,
0.188232421875,
0.75732421875,
0.1904296875
],
[
0.77587890625,
0.1812744140625,
0.8154296875,
0.1812744140625,
0.8154296875,
0.1904296875,
0.77587890625,
0.1904296875
],
[
0.11309814453125,
0.22705078125,
0.8857421875,
0.227783203125,
0.8857421875,
0.245361328125,
0.11309814453125,
0.2447509765625
],
[
0.281005859375,
0.18408203125,
0.32568359375,
0.18408203125,
0.32568359375,
0.191162109375,
0.281005859375,
0.191162109375
],
[
0.1162109375,
0.246826171875,
0.8369140625,
0.246826171875,
0.8369140625,
0.26025390625,
0.1162109375,
0.26025390625
],
[
0.1483154296875,
0.3046875,
0.880859375,
0.3046875,
0.880859375,
0.319580078125,
0.1483154296875,
0.319580078125
],
[
0.11517333984375,
0.321533203125,
0.880859375,
0.322265625,
0.880859375,
0.337158203125,
0.11517333984375,
0.33642578125
],
[
0.1162109375,
0.339111328125,
0.87744140625,
0.339111328125,
0.87744140625,
0.35400390625,
0.1162109375,
0.35400390625
],
[
0.1162109375,
0.35693359375,
0.8828125,
0.35693359375,
0.8828125,
0.37158203125,
0.1162109375,
0.37158203125
],
[
0.1162109375,
0.373779296875,
0.2147216796875,
0.373779296875,
0.2147216796875,
0.387939453125,
0.1162109375,
0.387939453125
],
[
0.146240234375,
0.391357421875,
0.8837890625,
0.391357421875,
0.8837890625,
0.40625,
0.146240234375,
0.40625
],
[
0.1182861328125,
0.408203125,
0.8828125,
0.408203125,
0.8828125,
0.42236328125,
0.1182861328125,
0.42236328125
],
[
0.11517333984375,
0.423828125,
0.880859375,
0.424560546875,
0.880859375,
0.439453125,
0.11517333984375,
0.43798828125
],
[
0.1162109375,
0.442138671875,
0.5654296875,
0.442138671875,
0.5654296875,
0.45556640625,
0.1162109375,
0.45556640625
],
[
0.146240234375,
0.458984375,
0.8828125,
0.458984375,
0.8828125,
0.473876953125,
0.146240234375,
0.473876953125
],
[
0.1162109375,
0.476806640625,
0.8798828125,
0.476806640625,
0.8798828125,
0.490234375,
0.1162109375,
0.490234375
],
[
0.1162109375,
0.49365234375,
0.8828125,
0.49365234375,
0.8828125,
0.5068359375,
0.1162109375,
0.5068359375
],
[
0.1162109375,
0.509765625,
0.8828125,
0.509765625,
0.8828125,
0.5234375,
0.1162109375,
0.5234375
],
[
0.1162109375,
0.52685546875,
0.8828125,
0.52685546875,
0.8828125,
0.54150390625,
0.1162109375,
0.54150390625
],
[
0.11517333984375,
0.5439453125,
0.880859375,
0.54248046875,
0.880859375,
0.55712890625,
0.11517333984375,
0.5576171875
],
[
0.1162109375,
0.56005859375,
0.8828125,
0.56005859375,
0.8828125,
0.5732421875,
0.1162109375,
0.5732421875
],
[
0.1162109375,
0.57763671875,
0.8828125,
0.57763671875,
0.8828125,
0.5908203125,
0.1162109375,
0.5908203125
],
[
0.1162109375,
0.5947265625,
0.8828125,
0.5947265625,
0.8828125,
0.60791015625,
0.1162109375,
0.60791015625
],
[
0.1162109375,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.1162109375,
0.62353515625
],
[
0.11309814453125,
0.626953125,
0.8837890625,
0.62548828125,
0.8837890625,
0.64306640625,
0.11309814453125,
0.64453125
],
[
0.1162109375,
0.6455078125,
0.64013671875,
0.6455078125,
0.64013671875,
0.65869140625,
0.1162109375,
0.65869140625
],
[
0.146240234375,
0.662109375,
0.8828125,
0.662109375,
0.8828125,
0.67724609375,
0.146240234375,
0.67724609375
],
[
0.1162109375,
0.6796875,
0.8828125,
0.6796875,
0.8828125,
0.69482421875,
0.1162109375,
0.69482421875
],
[
0.11517333984375,
0.697265625,
0.369384765625,
0.697265625,
0.369384765625,
0.7109375,
0.11517333984375,
0.7109375
],
[
0.1162109375,
0.7431640625,
0.5,
0.7431640625,
0.5,
0.75830078125,
0.1162109375,
0.75830078125
],
[
0.1182861328125,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79248046875,
0.1182861328125,
0.79248046875
],
[
0.1162109375,
0.79541015625,
0.880859375,
0.79541015625,
0.880859375,
0.81005859375,
0.1162109375,
0.81005859375
],
[
0.11517333984375,
0.8125,
0.14208984375,
0.8125,
0.14208984375,
0.82666015625,
0.11517333984375,
0.82666015625
],
[
0.1182861328125,
0.8369140625,
0.8828125,
0.8369140625,
0.8828125,
0.8505859375,
0.1182861328125,
0.8505859375
],
[
0.1328125,
0.8544921875,
0.8828125,
0.8544921875,
0.8828125,
0.86962890625,
0.1328125,
0.86962890625
],
[
0.1328125,
0.87158203125,
0.78515625,
0.87158203125,
0.78515625,
0.8857421875,
0.1328125,
0.8857421875
],
[
0.1182861328125,
0.8955078125,
0.880859375,
0.8955078125,
0.880859375,
0.90966796875,
0.1182861328125,
0.90966796875
],
[
0.1358642578125,
0.9130859375,
0.8828125,
0.9130859375,
0.8828125,
0.92822265625,
0.1358642578125,
0.92822265625
]
] |
[
"Data Providers",
"Datatranspor",
"Data storage",
"andprocessing",
"Dataac",
"HTCondor",
"Data",
"ElasticSearch",
"Kiban",
"",
"injection",
"&eKafka",
"WMAgent",
"InfuxDe",
"OGrafana",
"ActiveMQ",
"MSpark",
"HTTP",
"CRAB",
"HDFS",
"",
"SWAN",
"Figure 1. The data flow in the CERN MONIT infrastructure. The various steps of a typical monitoring",
"Logstash",
"workflow (data injection, transport, processing, storage, and access) are shown from left to right.",
"Until recently, most services were monitored through custom tools and web applications.",
"and logging information was scattered over several sources and mostly accessible only by ex-",
"perts. The maintenance and operation of the monitoring applications were becoming increas",
"ingly time-consuming and complicated due to the high turnover of experts in the computing",
"community.",
"Nowadays, several solutions are available to gather, store, and process large amounts of",
"data such as those produced by monitoring and logging services of computing applications.",
"Manytechnologies areavailableunderan open-sourcelicence,andaredeveloped and sup-",
"portedbylargesoftwarecompaniesandcommunities.",
"During the last two years, the CMS computing community has been gradually abandoning",
"in-house solutions in favour of the adoption of widely-used technologies based on open-",
"source, scalable,and non-SQL tools,such as Hadoop[7],InfluxDB [8],and ElasticSearch",
"[9]l.These services areavailable throughMONIT [15],a central monitoring infrastructure",
"provided by the CERN IT department, and allow for the easy deployment of monitoring",
"and accountingapplicationsusingvisualisation tools such as Kibana[1o]and Grafana[11]",
"Grafana allows users to automatically raise alarms and send notifications when anomalous",
"conditions in the monitoring data aremet.Data sources from different subsystems can beused",
"tobuild complexmonitoring workflows and predictive analytics,andperformance studies.To",
"complement themonitoring infrastructureprovided by the CERNIT department.theCMS",
"monitoring group set up and runs additional monitoring applications based on technologies",
"such as VictoriaMetrics [12],NATS [13],and Prometheus[14].",
"In the following sections, we describe MONIT, the monitoring infrastructure at CERN,",
"the organisation of CMS monitoring applications based on MONIT and CMS own infrastruc-",
"ture,and futuredevelopments.",
"2TheCERNMONITinfrastructure",
"The CERN IT department offers a variety of monitoring services through the MONIT in-",
"frastructure. A typical monitoring workflow consists of the following steps (see also Fig.",
"1):",
"·Data injection: data is pushed into the MONIT infrastructure throughActiveMQ[16]",
"messaging. As an alternative, data providers inside the CERN network boundaries may",
"use an HTTP endpoint. Log files may be directly injected using Logstash [17].",
".Data transport and processing: Data are streamed into MONIT using Apache Kafka [18].",
"Apache Spark [19] and Apache Flume [20] are used for further data processing. Data may"
] |
{
"arxiv_id": "2007.03630",
"language": "en",
"page_idx": 1
}
|
||
2007.03630-en-2
|
2007.03630-en-2.jpg
|
be enriched with additional information or data from several sources can be aggregated as
needed.
$\bullet$ Data storage: Three storage technologies are available, depending on the needs of the particular use case: data volume, schema, and retention policy. ElasticSearch (ES) is used to store semi-structured data for a short period of time (up to one month). InfluxDB is used to store time-series structured, aggregated data for a longer period of time (up to five years, depending on the granularity). The Hadoop Distributed File System (HDFS) is used for long-term (currently unlimited in time) storage of data in a variety of formats (Apache Avro [21], Apache Parquet [22], JSON [23], plain text).
$\bullet$ Data access: For data access we rely on the stack ES/Kibana, Grafana, and the SWAN
[24] service at CERN. Kibana is used for data exploration and visualisation of ES data sources. Grafana is a visualisation tool that can read data from several sources such as ES, InfluxDB, Prometheus, Graphite[25], Open TSDB [26], and MySQL[27]. Data sources can also be accessed through the Grafana proxy. The Apache Spark framework is used to process data on HDFS. The SWAN service provides access to the CERN Hadoop clusters through a JupyterHub interface [28].
The ActiveMQ servers and the SWAN service are not strictly part of the MONIT infras-
tructure, but are offered and managed by the CERN IT department.
# 3 The CMS monitoring infrastructure
In this section we provide a comprehensive overview of various components of the CMS monitoring infrastructure. In Section 3.1 we describe how CMS is building monitoring applications based on the MONIT infrastructure. In Section 3.2 we provide details of the CMS specific monitoring infrastructure, and we discuss the CMS monitoring Kubernetes (k8s) [29] clusters in Section 3.3.
3.1 CMS usage of the MONIT infrastructure
A variety of CMS data producers regularly inject data into the MONIT infrastructure. There are currently more than twenty-five active CMS monitoring workflows, covering a large variety of systems: HTCondor job parameters, GlideinWMS submission infrastructure, data transfers and access patterns, CRAB and WMAgent tasks, CMSWEB services. With reference to the data flow outlined in Section 2, we describe in the following paragraphs how CMS is using the MONIT infrastructure.
## 3.1.1 Data injection
Most of CMS monitoring data producers use the ActiveMQ endpoints—with a few exceptions for producers inside the CERN internal network which use an HTTP REST endpoint - to inject data in JSON format. The CERN MONIT infrastructure consumes the data via a Kafka pipeline and redirects them to the ES, InfluxDB, and HDFS data sinks. We do not impose a specific schema on injected documents, but the schema should be consistent over time. The document schema and daily data volume are agreed a priory for each CMS data producer with the CERN MONIT team, which then allocates and controls the resources to consume the data in their infrastructure.
|
[
[
0.1307373046875,
0.0853271484375,
0.8828125,
0.0860595703125,
0.8828125,
0.100830078125,
0.1307373046875,
0.09942626953125
],
[
0.1337890625,
0.1051025390625,
0.19921875,
0.1051025390625,
0.19921875,
0.1163330078125,
0.1337890625,
0.1163330078125
],
[
0.1182861328125,
0.127685546875,
0.8828125,
0.127685546875,
0.8828125,
0.1424560546875,
0.1182861328125,
0.1424560546875
],
[
0.1328125,
0.145263671875,
0.8837890625,
0.14453125,
0.8837890625,
0.15869140625,
0.1328125,
0.1600341796875
],
[
0.1328125,
0.1622314453125,
0.8837890625,
0.1622314453125,
0.8837890625,
0.177001953125,
0.1328125,
0.177001953125
],
[
0.1337890625,
0.1790771484375,
0.8837890625,
0.1790771484375,
0.8837890625,
0.1932373046875,
0.1337890625,
0.1932373046875
],
[
0.1358642578125,
0.1953125,
0.8828125,
0.1953125,
0.8828125,
0.210205078125,
0.1358642578125,
0.210205078125
],
[
0.1337890625,
0.2122802734375,
0.8828125,
0.2122802734375,
0.8828125,
0.22705078125,
0.1337890625,
0.22705078125
],
[
0.1337890625,
0.2298583984375,
0.59765625,
0.2298583984375,
0.59765625,
0.2432861328125,
0.1337890625,
0.2432861328125
],
[
0.1182861328125,
0.25390625,
0.8828125,
0.25390625,
0.8828125,
0.267333984375,
0.1182861328125,
0.267333984375
],
[
0.1337890625,
0.270751953125,
0.8837890625,
0.269287109375,
0.8837890625,
0.2841796875,
0.1337890625,
0.284912109375
],
[
0.1337890625,
0.288330078125,
0.8828125,
0.288330078125,
0.8828125,
0.3017578125,
0.1337890625,
0.3017578125
],
[
0.1337890625,
0.303955078125,
0.8828125,
0.303955078125,
0.8828125,
0.3173828125,
0.1337890625,
0.3173828125
],
[
0.1337890625,
0.320068359375,
0.8828125,
0.320068359375,
0.8828125,
0.3349609375,
0.1337890625,
0.3349609375
],
[
0.1337890625,
0.338623046875,
0.8828125,
0.338623046875,
0.8828125,
0.3525390625,
0.1337890625,
0.3525390625
],
[
0.1337890625,
0.354736328125,
0.43994140625,
0.354736328125,
0.43994140625,
0.36962890625,
0.1337890625,
0.36962890625
],
[
0.146240234375,
0.380126953125,
0.8828125,
0.380126953125,
0.8828125,
0.394287109375,
0.146240234375,
0.394287109375
],
[
0.11517333984375,
0.397705078125,
0.669921875,
0.397705078125,
0.669921875,
0.411865234375,
0.11517333984375,
0.411865234375
],
[
0.11309814453125,
0.443603515625,
0.5341796875,
0.4443359375,
0.5341796875,
0.4619140625,
0.11309814453125,
0.461181640625
],
[
0.11517333984375,
0.48095703125,
0.8828125,
0.48095703125,
0.8828125,
0.495849609375,
0.11517333984375,
0.495849609375
],
[
0.11517333984375,
0.497802734375,
0.8828125,
0.49853515625,
0.8828125,
0.51318359375,
0.11517333984375,
0.51220703125
],
[
0.11517333984375,
0.515625,
0.8828125,
0.51416015625,
0.8828125,
0.52880859375,
0.11517333984375,
0.52978515625
],
[
0.1162109375,
0.53173828125,
0.880859375,
0.53173828125,
0.880859375,
0.544921875,
0.1162109375,
0.544921875
],
[
0.11517333984375,
0.548828125,
0.30712890625,
0.548828125,
0.30712890625,
0.5634765625,
0.11517333984375,
0.5634765625
],
[
0.1162109375,
0.5947265625,
0.52294921875,
0.5947265625,
0.52294921875,
0.609375,
0.1162109375,
0.609375
],
[
0.1162109375,
0.62890625,
0.8828125,
0.62890625,
0.8828125,
0.64306640625,
0.1162109375,
0.64306640625
],
[
0.1162109375,
0.646484375,
0.8828125,
0.646484375,
0.8828125,
0.66064453125,
0.1162109375,
0.66064453125
],
[
0.1162109375,
0.6630859375,
0.8828125,
0.6630859375,
0.8828125,
0.677734375,
0.1162109375,
0.677734375
],
[
0.11517333984375,
0.6796875,
0.8828125,
0.6796875,
0.8828125,
0.69482421875,
0.11517333984375,
0.69482421875
],
[
0.1162109375,
0.697265625,
0.8828125,
0.697265625,
0.8828125,
0.71240234375,
0.1162109375,
0.71240234375
],
[
0.11517333984375,
0.71435546875,
0.40771484375,
0.71435546875,
0.40771484375,
0.728515625,
0.11517333984375,
0.728515625
],
[
0.1162109375,
0.7587890625,
0.294677734375,
0.7587890625,
0.294677734375,
0.7734375,
0.1162109375,
0.7734375
],
[
0.11517333984375,
0.79345703125,
0.8828125,
0.794921875,
0.8828125,
0.80908203125,
0.11517333984375,
0.80810546875
],
[
0.1162109375,
0.81103515625,
0.8837890625,
0.81103515625,
0.8837890625,
0.82568359375,
0.1162109375,
0.82568359375
],
[
0.12451171875,
0.82861328125,
0.8828125,
0.82861328125,
0.8828125,
0.841796875,
0.12451171875,
0.841796875
],
[
0.1162109375,
0.845703125,
0.8837890625,
0.845703125,
0.8837890625,
0.8603515625,
0.1162109375,
0.8603515625
],
[
0.11309814453125,
0.8623046875,
0.8837890625,
0.86083984375,
0.8837890625,
0.8759765625,
0.11309814453125,
0.87646484375
],
[
0.11517333984375,
0.8798828125,
0.8828125,
0.8798828125,
0.8828125,
0.8935546875,
0.11517333984375,
0.8935546875
],
[
0.1162109375,
0.896484375,
0.8837890625,
0.896484375,
0.8837890625,
0.90966796875,
0.1162109375,
0.90966796875
],
[
0.11309814453125,
0.91162109375,
0.44921875,
0.9130859375,
0.44921875,
0.92822265625,
0.11309814453125,
0.9267578125
]
] |
[
"beenrichedwithadditional information ordatafrom several sourcescanbeaggregatedas",
"needed.",
". Data storage: Three storage technologies are available, depending on the needs of the",
"particular use case:data volume, schema, and retention policy.ElasticSearch (ES)is used",
"to store semi-structured data for a short period of time (up to one month). InfluxDB is",
"used to store time-series structured, aggregated data for a longer period of time (up to five",
"years, depending on the granularity). The Hadoop Distributed File System (HDFS) is used",
"for long-term (currently unlimited in time) storage of data in a variety of formats (Apache",
"Avro [21], ApacheParquet [22],JSON[23],plain text).",
".Data access: For data access we rely on the stack ES/Kibana, Grafana, and the SWAN",
"[24] service at CERN.Kibana is used for data exploration and visualisation of ES data",
"sources. Grafana is a visualisation tool that can read data from several sources such as ES,",
"InfluxDB,Prometheus, Graphite [25],Open TSDB [26], and MySQL[27]. Data sources",
"can also be accessed through the Grafana proxy. The Apache Spark framework is used to",
"process data on HDFS. The SWAN service provides access to the CERN Hadoop clusters",
"through a JupyterHub interface [28].",
"The ActiveMQ servers and the SWAN service are not strictly part of the MONIT infras-",
"tructure, but are offered and managed by the CERNIT department.",
"3TheCMSmonitoringinfrastructure",
"In this section we provide a comprehensive overview of various components of the CMS",
"monitoring infrastructure.InSection3.1wedescribehowCMS is buildingmonitoringap-",
"plicationsbased ontheMONIT infrastructure.In Section3.2weprovidedetails of theCMS",
"specificmonitoring infrastructure,andwediscusstheCMS monitoringKubernetes (k8s)[29]",
"clusters in Section 3.3.",
"3.1CMSusageoftheMONITinfrastructure",
"A variety of CMS data producers regularly inject data into the MONIT infrastructure. There",
"are currently more than twenty-five active CMS monitoring workflows, covering a large va-",
"riety of systems: HTCondor job parameters, GlideinWMS submission infrastructure, data",
"transfers and access patterns, CRAB and WMAgent tasks, CMSWEB services. With refer-",
"ence to the data flow outlined in Section 2, we describe in the following paragraphs how CMS",
"is using the MONIT infrastructure.",
"3.1.1 Data injection",
"Most of CMS monitoringdata producers use the ActiveMQ endpoints-witha few excep-",
"tions for producers inside the CERN internal network which use an HTTP REST endpoint",
"-to inject data in JSON format. The CERN MONIT infrastructure consumes the data via",
"a Kafka pipeline and redirects them to the ES, InfluxDB, and HDFS data sinks. We do not",
"impose a specific schema on injected documents, but the schema should be consistent over",
"time. The document schema and daily data volume are agreed a priory for each CMS data",
"producer with the CERN MONIT team, which then allocates and controls the resources to",
"consume the data in their infrastructure."
] |
{
"arxiv_id": "2007.03630",
"language": "en",
"page_idx": 2
}
|
||
2007.03630-en-3
|
2007.03630-en-3.jpg
|
# 3.1.2 Data storage
The CMS data producers exploit all the provided storage options depending on their particular use case. Data in ES are stored for the last 30-40 days depending on the data source. In addition to the raw data indexes, we take advantage of the document indexing features of ES to create purpose-specific indexes which incorporate some additional logic. For example, we compute and aggregate over HTCondor jobs that have been retried several times, or over temporary job statuses. InfluxDB is used to store aggregated structured data as time series for a limited set of tags and felds. This is used to build historical views of key performance metrics.
## 3.1.3 Data access
Access to the data stored in MONIT is provided through visualization tools such as Grafana and Kibana for data in InfluxDB and ES, and Apache spark jobs (either through the SWAN service or standalone scripts) for data in HDFS. Kibana is mainly used for interactive data exploration of the ES data sources. The search capabilities of ES enable our users to quickly create ad-hoc queries and visualizations. Most CMS official dashboards are implemented in Grafana. Additionally, users can create their own personalised views.
Data stored in HDFS are typically accessed using Spark-based workflows, and are used to create dedicated views over long time periods (for example, to regularly produce yearly data popularity plots), or for complex queries and visualizations that can not be implemented in Grafana or Kibana.
3.1.4 Experiences with the MONIT infrastructure
In the past two years several custom CMS monitoring applications have been ported to the MONIT infrastructure. The first tests with the new infrastructure started in 2017, and the migration process formally began in October 2018. Since then the number of data sources, stored data volume, and usage has steadily grown (see Fig 2).
HTCondor job monitoring is one of the most important views, since it provides an overview of the performances of the CMS distributed computing infrastructure, and was one of the first to be migrated. Data from the HTCondor pool are collected, processed, and injected into MONIT at regular intervals (every twelve minutes) by a service, implemented in python, called spider. The spider sends data to MONIT through the ActiveMQ endpoint. The HTCondor data are finally stored in ES, InfluxDB, and HDFS.
The complete migration took about one year through various steps: first an evaluation phase, where all use cases were gathered and requirements for the new application drafted, then the actual development phase, where the complete data flow was implemented, then a final evaluation phase, where the new data flow was validated against the old one, first from CMS monitoring experts, and then from the wider CMS community. The migration required the coordination and interplay of several teams: the MONIT team, the developers of the CMS monitoring group, and the CMS stakeholders (the data providers and consumers). In particular, it was crucial to define the schema of the data in a way that all user requirements can be satisfied using the visualization options of Grafana or Kibana.
A major problem we encountered was the lack of schema validation in the MONIT infrastructure which may lead to some documents being rejected by ES, if the same field has different types in different documents. To avoid this problem, we introduced schema validation of the injected documents through the CMS client tools, and we are gradually enforcing
|
[
[
0.11309814453125,
0.08392333984375,
0.287353515625,
0.0860595703125,
0.287353515625,
0.100830078125,
0.11309814453125,
0.0987548828125
],
[
0.1162109375,
0.1170654296875,
0.8828125,
0.1170654296875,
0.8828125,
0.1318359375,
0.1162109375,
0.1318359375
],
[
0.11517333984375,
0.1353759765625,
0.8828125,
0.1353759765625,
0.8828125,
0.1488037109375,
0.11517333984375,
0.1488037109375
],
[
0.1162109375,
0.151611328125,
0.8837890625,
0.151611328125,
0.8837890625,
0.1650390625,
0.1162109375,
0.1650390625
],
[
0.1162109375,
0.1693115234375,
0.8828125,
0.1693115234375,
0.8828125,
0.1826171875,
0.1162109375,
0.1826171875
],
[
0.11517333984375,
0.1861572265625,
0.8828125,
0.184814453125,
0.8828125,
0.1995849609375,
0.11517333984375,
0.200927734375
],
[
0.1162109375,
0.203125,
0.880859375,
0.203125,
0.880859375,
0.216552734375,
0.1162109375,
0.216552734375
],
[
0.11309814453125,
0.2186279296875,
0.8828125,
0.2193603515625,
0.8828125,
0.234130859375,
0.11309814453125,
0.232666015625
],
[
0.11517333984375,
0.2362060546875,
0.18359375,
0.2362060546875,
0.18359375,
0.250244140625,
0.11517333984375,
0.250244140625
],
[
0.11517333984375,
0.280029296875,
0.2841796875,
0.280029296875,
0.2841796875,
0.294677734375,
0.11517333984375,
0.294677734375
],
[
0.1162109375,
0.31103515625,
0.8837890625,
0.31103515625,
0.8837890625,
0.325927734375,
0.1162109375,
0.325927734375
],
[
0.11517333984375,
0.32861328125,
0.8828125,
0.32861328125,
0.8828125,
0.343505859375,
0.11517333984375,
0.343505859375
],
[
0.1162109375,
0.34423828125,
0.880859375,
0.34423828125,
0.880859375,
0.35888671875,
0.1162109375,
0.35888671875
],
[
0.11517333984375,
0.361083984375,
0.8828125,
0.362548828125,
0.8828125,
0.37646484375,
0.11517333984375,
0.3759765625
],
[
0.1162109375,
0.378662109375,
0.8828125,
0.378662109375,
0.8828125,
0.3935546875,
0.1162109375,
0.3935546875
],
[
0.1162109375,
0.396240234375,
0.69189453125,
0.396240234375,
0.69189453125,
0.40966796875,
0.1162109375,
0.40966796875
],
[
0.146240234375,
0.413330078125,
0.8837890625,
0.413330078125,
0.8837890625,
0.427978515625,
0.146240234375,
0.427978515625
],
[
0.1162109375,
0.430908203125,
0.8828125,
0.430908203125,
0.8828125,
0.44580078125,
0.1162109375,
0.44580078125
],
[
0.1162109375,
0.44775390625,
0.8837890625,
0.44775390625,
0.8837890625,
0.4619140625,
0.1162109375,
0.4619140625
],
[
0.11309814453125,
0.46337890625,
0.297607421875,
0.464111328125,
0.297607421875,
0.478759765625,
0.11309814453125,
0.4775390625
],
[
0.1162109375,
0.5068359375,
0.54443359375,
0.5068359375,
0.54443359375,
0.52197265625,
0.1162109375,
0.52197265625
],
[
0.1162109375,
0.53955078125,
0.8837890625,
0.53955078125,
0.8837890625,
0.5537109375,
0.1162109375,
0.5537109375
],
[
0.11517333984375,
0.55712890625,
0.8837890625,
0.55712890625,
0.8837890625,
0.5703125,
0.11517333984375,
0.5703125
],
[
0.1162109375,
0.5732421875,
0.8828125,
0.5732421875,
0.8828125,
0.58837890625,
0.1162109375,
0.58837890625
],
[
0.11517333984375,
0.5908203125,
0.62451171875,
0.5908203125,
0.62451171875,
0.60498046875,
0.11517333984375,
0.60498046875
],
[
0.146240234375,
0.60791015625,
0.8837890625,
0.60791015625,
0.8837890625,
0.62255859375,
0.146240234375,
0.62255859375
],
[
0.1162109375,
0.625,
0.8857421875,
0.625,
0.8857421875,
0.63818359375,
0.1162109375,
0.63818359375
],
[
0.1162109375,
0.64111328125,
0.8837890625,
0.64111328125,
0.8837890625,
0.65576171875,
0.1162109375,
0.65576171875
],
[
0.11517333984375,
0.6572265625,
0.8828125,
0.6572265625,
0.8828125,
0.67138671875,
0.11517333984375,
0.67138671875
],
[
0.1162109375,
0.6748046875,
0.8837890625,
0.6748046875,
0.8837890625,
0.68896484375,
0.1162109375,
0.68896484375
],
[
0.1162109375,
0.6923828125,
0.6328125,
0.6923828125,
0.6328125,
0.7060546875,
0.1162109375,
0.7060546875
],
[
0.1483154296875,
0.7099609375,
0.8828125,
0.7099609375,
0.8828125,
0.7236328125,
0.1483154296875,
0.7236328125
],
[
0.1162109375,
0.72705078125,
0.8828125,
0.72705078125,
0.8828125,
0.74072265625,
0.1162109375,
0.74072265625
],
[
0.1162109375,
0.7431640625,
0.8837890625,
0.7431640625,
0.8837890625,
0.7568359375,
0.1162109375,
0.7568359375
],
[
0.1162109375,
0.7607421875,
0.8828125,
0.7607421875,
0.8828125,
0.7744140625,
0.1162109375,
0.7744140625
],
[
0.11517333984375,
0.7763671875,
0.8837890625,
0.77587890625,
0.8837890625,
0.79052734375,
0.11517333984375,
0.791015625
],
[
0.11517333984375,
0.79248046875,
0.8828125,
0.79345703125,
0.8828125,
0.80810546875,
0.11517333984375,
0.806640625
],
[
0.1162109375,
0.81103515625,
0.8828125,
0.81103515625,
0.8828125,
0.82421875,
0.1162109375,
0.82421875
],
[
0.1162109375,
0.828125,
0.8837890625,
0.828125,
0.8837890625,
0.841796875,
0.1162109375,
0.841796875
],
[
0.11517333984375,
0.84423828125,
0.68896484375,
0.84423828125,
0.68896484375,
0.8583984375,
0.11517333984375,
0.8583984375
],
[
0.146240234375,
0.8623046875,
0.8828125,
0.86083984375,
0.8828125,
0.8759765625,
0.146240234375,
0.87646484375
],
[
0.1162109375,
0.87890625,
0.8837890625,
0.87890625,
0.8837890625,
0.8935546875,
0.1162109375,
0.8935546875
],
[
0.1162109375,
0.896484375,
0.8828125,
0.896484375,
0.8828125,
0.90966796875,
0.1162109375,
0.90966796875
],
[
0.11517333984375,
0.9130859375,
0.8828125,
0.9130859375,
0.8828125,
0.92822265625,
0.11517333984375,
0.92822265625
]
] |
[
"3.1.2Datastorage",
"The CMS data producers exploit all the provided storage options depending on their particular",
"use case. Data in ES are stored for the last 30-40 days depending on the data source. In",
"addition to the raw data indexes, we take advantage of the document indexing features of",
"ES to create purpose-specific indexes which incorporate some additional logic.For example,",
"we compute and aggregate over HTCondor jobs that have been retried several times, or over",
"temporary job statuses. InfluxDB is used to store aggregated structured data as time series",
"for a limited set of tags and fields.This is used to build historical views ofkey performance",
"metrics.",
"3.1.3Dataaccess",
"Access to the data stored in MONIT is provided through visualization tools such as Grafana",
"and Kibana for data in InfluxDB and ES, and Apache spark jobs (either through the SWAN",
"service or standalone scripts) for data in HDFS. Kibana is mainly used for interactive data",
"exploration of the ES data sources.The searchcapabilities of ES enable ourusers toquickly",
"create ad-hoc queries and visualizations. Most CMS official dashboards are implemented in",
"Grafana.Additionally,users cancreatetheirownpersonalisedviews.",
"Data stored in HDFS are typically accessed using Spark-based workflows, and are used",
"to create dedicated views over long time periods (for example, to regularly produce yearly",
"data popularity plots), or for complex queries and visualizations that can not be implemented",
"inGrafanaorKibana.",
"3.1.4 Experiences withtheMONIT infrastructure",
"In the past two years several custom CMS monitoring applications have been ported to the",
"MONIT infrastructure. The first tests with the new infrastructure started in 2017, and the",
"migration process formally began in October 2018. Since then the number of data sources,",
"stored data volume, and usage has steadily grown (see Fig 2).",
"HTCondor job monitoring is one of the most important views, since it provides an",
"overview of the performances of the CMS distributed computing infrastructure,and was one",
"of the first to be migrated. Data from the HTCondor pool are collected, processed, and in-",
"jected into MONIT at regular intervals (every twelve minutes) by a service, implemented in",
"python, called spider. The spider sends data to MONIT through the ActiveMQ endpoint. The",
"HTCondor data are finally stored in ES, InfluxDB, and HDFS.",
"The complete migration took about one year through various steps: first an evaluation",
"phase, where all use cases were gathered and requirements for the new application drafted,",
"then the actual development phase, where the complete data flow was implemented, then a",
"final evaluation phase, where the new data flow was validated against the old one, first from",
"CMS monitoring experts, and then from the wider CMS community.The migration required",
"the coordination and interplay of several teams:the MONIT team, thedevelopers of the",
"CMS monitoring group, and the CMS stakeholders (the data providers and consumers). In",
"particular, it was crucial to define the schema of the data in a way that all user requirements",
"can be satisfied using the visualization options of Grafana or Kibana.",
"Amajorproblem we encountered was thelack of schema validation in the MONIT in-",
"frastructure which may lead to some documents being rejected by ES, if the same field has",
"different types in different documents. To avoid this problem, we introduced schema valida-",
"tion of the injected documents through the CMS client tools, and we are gradually enforcing"
] |
{
"arxiv_id": "2007.03630",
"language": "en",
"page_idx": 3
}
|
||
2007.03630-en-4
|
2007.03630-en-4.jpg
|
<FigureHere>
Figure 2. Time evolution distributions showing the development and adoption of the CMS Monitoring infrastructure. a) From top to bottom, the evolution of the total number of CMS data producers, the data stored in HDFS, and the monthly data volume. b) The number of Grafana dashboards using ES, InfluxDB, and Prometheus as data sources, ordered by date of creation.
it for all CMS data providers. We also noticed that some attributes are reserved by ES (such
as version, timestamp, uuid), and those should be avoided.
In InfluxDB a limited set of aggregated tags are stored. The aggregations are computed every twelve minutes and stored for one week. Additional aggregations, e.g. one-hour, oneday, seven-days and thirty-days bins are stored for five years. We currently store eighteen
|
[
[
0.1556396484375,
0.100830078125,
0.369384765625,
0.1015625,
0.369384765625,
0.11492919921875,
0.1556396484375,
0.1142578125
],
[
0.371337890625,
0.12481689453125,
0.460693359375,
0.126953125,
0.460693359375,
0.134033203125,
0.371337890625,
0.1318359375
],
[
0.5693359375,
0.1226806640625,
0.65234375,
0.12481689453125,
0.65234375,
0.1353759765625,
0.5693359375,
0.13330078125
],
[
0.3662109375,
0.12548828125,
0.37451171875,
0.1290283203125,
0.371337890625,
0.13330078125,
0.363037109375,
0.1297607421875
],
[
0.15771484375,
0.12548828125,
0.2354736328125,
0.12548828125,
0.2354736328125,
0.13330078125,
0.15771484375,
0.13330078125
],
[
0.365234375,
0.1361083984375,
0.436767578125,
0.1361083984375,
0.436767578125,
0.1431884765625,
0.365234375,
0.1431884765625
],
[
0.57275390625,
0.1361083984375,
0.62744140625,
0.1361083984375,
0.62744140625,
0.146728515625,
0.57275390625,
0.146728515625
],
[
0.337158203125,
0.170654296875,
0.35888671875,
0.170654296875,
0.35888671875,
0.1749267578125,
0.337158203125,
0.1749267578125
],
[
0.15771484375,
0.1368408203125,
0.2841796875,
0.1368408203125,
0.2841796875,
0.14453125,
0.15771484375,
0.14453125
],
[
0.25732421875,
0.228515625,
0.365234375,
0.228515625,
0.365234375,
0.2362060546875,
0.25732421875,
0.2362060546875
],
[
0.229248046875,
0.26513671875,
0.2427978515625,
0.26513671875,
0.2427978515625,
0.268798828125,
0.229248046875,
0.268798828125
],
[
0.26025390625,
0.2890625,
0.29345703125,
0.2890625,
0.29345703125,
0.292724609375,
0.26025390625,
0.292724609375
],
[
0.161865234375,
0.36328125,
0.3154296875,
0.364013671875,
0.3154296875,
0.3779296875,
0.161865234375,
0.37646484375
],
[
0.1556396484375,
0.408935546875,
0.70556640625,
0.408935546875,
0.70556640625,
0.42236328125,
0.1556396484375,
0.42236328125
],
[
0.5498046875,
0.430908203125,
0.60888671875,
0.430908203125,
0.60888671875,
0.43798828125,
0.5498046875,
0.43798828125
],
[
0.281005859375,
0.44580078125,
0.337158203125,
0.44580078125,
0.337158203125,
0.45263671875,
0.281005859375,
0.45263671875
],
[
0.33203125,
0.443603515625,
0.43359375,
0.443603515625,
0.43359375,
0.4541015625,
0.33203125,
0.4541015625
],
[
0.1607666015625,
0.4619140625,
0.186767578125,
0.4619140625,
0.186767578125,
0.466064453125,
0.1607666015625,
0.466064453125
],
[
0.19189453125,
0.4619140625,
0.2105712890625,
0.4619140625,
0.2105712890625,
0.466064453125,
0.19189453125,
0.466064453125
],
[
0.57470703125,
0.50341796875,
0.603515625,
0.50341796875,
0.603515625,
0.5078125,
0.57470703125,
0.5078125
],
[
0.15869140625,
0.533203125,
0.1981201171875,
0.533203125,
0.1981201171875,
0.54248046875,
0.15869140625,
0.54248046875
],
[
0.30078125,
0.53369140625,
0.3681640625,
0.53369140625,
0.3681640625,
0.54150390625,
0.30078125,
0.54150390625
],
[
0.362060546875,
0.53125,
0.5029296875,
0.533203125,
0.5029296875,
0.5439453125,
0.362060546875,
0.54150390625
],
[
0.35693359375,
0.544921875,
0.445068359375,
0.544921875,
0.445068359375,
0.552734375,
0.35693359375,
0.552734375
],
[
0.6494140625,
0.57666015625,
0.6669921875,
0.57666015625,
0.6669921875,
0.5810546875,
0.6494140625,
0.5810546875
],
[
0.3349609375,
0.5986328125,
0.4150390625,
0.5986328125,
0.4150390625,
0.609375,
0.3349609375,
0.609375
],
[
0.427490234375,
0.60107421875,
0.50927734375,
0.60107421875,
0.50927734375,
0.60791015625,
0.427490234375,
0.60791015625
],
[
0.256103515625,
0.64892578125,
0.28515625,
0.64892578125,
0.28515625,
0.66015625,
0.256103515625,
0.66015625
],
[
0.42333984375,
0.64892578125,
0.454345703125,
0.64892578125,
0.454345703125,
0.66015625,
0.42333984375,
0.66015625
],
[
0.603515625,
0.64892578125,
0.6337890625,
0.64892578125,
0.6337890625,
0.66015625,
0.603515625,
0.66015625
],
[
0.79052734375,
0.6494140625,
0.80908203125,
0.6494140625,
0.80908203125,
0.658203125,
0.79052734375,
0.658203125
],
[
0.79052734375,
0.6708984375,
0.80908203125,
0.6708984375,
0.80908203125,
0.677734375,
0.79052734375,
0.677734375
],
[
0.2012939453125,
0.68408203125,
0.22412109375,
0.68408203125,
0.22412109375,
0.6884765625,
0.2012939453125,
0.6884765625
],
[
0.5029296875,
0.68115234375,
0.5185546875,
0.68115234375,
0.5185546875,
0.69091796875,
0.5029296875,
0.69091796875
],
[
0.179443359375,
0.69970703125,
0.19921875,
0.69970703125,
0.19921875,
0.70361328125,
0.179443359375,
0.70361328125
],
[
0.46484375,
0.6953125,
0.482421875,
0.6953125,
0.482421875,
0.70458984375,
0.46484375,
0.70458984375
],
[
0.11517333984375,
0.736328125,
0.8828125,
0.73828125,
0.8828125,
0.75244140625,
0.11517333984375,
0.75048828125
],
[
0.1162109375,
0.75439453125,
0.8828125,
0.75439453125,
0.8828125,
0.7685546875,
0.1162109375,
0.7685546875
],
[
0.1162109375,
0.76953125,
0.880859375,
0.76953125,
0.880859375,
0.78271484375,
0.1162109375,
0.78271484375
],
[
0.1162109375,
0.7841796875,
0.64501953125,
0.7841796875,
0.64501953125,
0.79736328125,
0.1162109375,
0.79736328125
],
[
0.11517333984375,
0.84423828125,
0.8828125,
0.84423828125,
0.8828125,
0.85888671875,
0.11517333984375,
0.85888671875
],
[
0.11517333984375,
0.8623046875,
0.599609375,
0.8623046875,
0.599609375,
0.8759765625,
0.11517333984375,
0.8759765625
],
[
0.146240234375,
0.87890625,
0.8828125,
0.87890625,
0.8828125,
0.8935546875,
0.146240234375,
0.8935546875
],
[
0.1162109375,
0.896484375,
0.8828125,
0.896484375,
0.8828125,
0.9111328125,
0.1162109375,
0.9111328125
],
[
0.1162109375,
0.9130859375,
0.8828125,
0.9130859375,
0.8828125,
0.92822265625,
0.1162109375,
0.92822265625
]
] |
[
"a)EvolutionofHDFSusage",
"DataManaqement",
"ServiceStatus",
"",
"HTCondorJobs",
"WehService",
"Training",
"e.11",
"DistributedInfrastructure",
"Accumulatedsize:25TE",
"",
"uonth",
"HTCondorJobs(22,934.91GB)",
"b)Usageof Grafanadata-sourcetypesindashboardsbycreationdate",
"Dctober2018",
"Firstexplor",
"brationwithGrafana",
"slacti",
"0",
"ar201",
"influxdb",
"Thedevelopn",
"mentofthe JobMonitoring",
"Dashboard starts",
"20",
"Proofofconcep",
"vithPrometheus",
"2017",
"2018",
"2019",
"2020",
"152",
"",
"73",
"ron",
"63",
"Figure 2. Time evolution distributions showing the development and adoption of the CMS Monitoring",
"infrastructure. a) From top to bottom, the evolution of the total number of CMS data producers, the",
"data stored in HDFS, and the monthly data volume. b) The number of Grafana dashboards using ES.",
"InfluxDB,and Prometheus as data sources, ordered by date of creation",
"it for all CMS data providers. We also noticed that some attributes are reserved by ES (such",
"as version, timestamp, uuid), and those should be avoided.",
"In InfluxDB a limited set of aggregated tags are stored. The aggregations are computed",
"every twelve minutes and stored for one week. Additional aggregations, e.g. one-hour, one-",
"day, seven-days and thirty-days bins are stored for five years. We currently store eighteen"
] |
{
"arxiv_id": "2007.03630",
"language": "en",
"page_idx": 4
}
|
||
1810.11344-en-0
|
1810.11344-en-0.jpg
|
Benefits of over-parameterization with EM
Daniel Hsu
Columbia University
djhsu@cs.columbia.edu
Arian Maleki
Columbia University
arian@stat.columbia.edu
Ji Xu
Columbia University
jixu@cs.columbia.edu
### Abstract
Expectation Maximization (EM) is among the most popular algorithms for maximum likelihood estimation, but it is generally only guaranteed to find its stationary points of the log-likelihood objective. The goal of this article is to present theoretical and empirical evidence that over-parameterization can help EM avoid spurious $\hat{\text{local optima in the log-likelihood. }}\hat{\text{We consider the problem of estimating the}}$ $\hat{\text{mean vectors of a Gaussian mixture model in a scenario where the mixing weights}}$ are known. Our study shows that the global behavior of EM, when one uses an over-parameterized model in which the mixing weights are treated as unknown, is better than that when one uses the (correct) model with the mixing weights fixed to the known values. For symmetric Gaussians mixtures with two components, we prove that introducing the (statistically redundant) weight parameters enables EM $\hat{\text{to find the global maximizer of the log-likelihood starting from almost any initial}}$ mean parameters, whereas EM without this over-parameterization may very often fail. For other Gaussian mixtures, we provide empirical evidence that shows similar behavior. Our results corroborate the value of over-parameterization in solving non-convex optimization problems, previously observed in other domains.
## 1 Introduction
In a Gaussian mixture model (GMM), the observed data $\mathcal{Y}=\{y_1,y_2,\ldots,y_n\}\subset\mathbb{R}^d$ comprise an
i.i.d. sample from a mixture of $k$ Gaussians:
(1)
$$\boldsymbol{y}_{1},\ldots,\boldsymbol{y}_{n}\stackrel{\mathrm{i.i.d.}}{\sim}\sum_{i=1}^{k}w_{i}^{*}\:\mathcal{N}(\boldsymbol{\theta}_{i}^{*},\boldsymbol{\Sigma}_{i}^{*})$$
where $(w_i^*,\boldsymbol{\theta}_i^*,\boldsymbol{\Sigma}_i^*)$ denote the weight, mean, and covariance matrix of the $i^\mathrm{th}$ mixture component. Parameters of the GMM are often estimated using the Expectation Maximization (EM) algorithm, which aims to find the maximizer of the log-likelihood objective. However, the log-likelihood function is not concave, so EM is only guaranteed to find its stationary points. This leads to the following natural and fundamental question in the study of EM and non-convex optimization: How can EM escape spurious local maxima and saddle points to reach the maximum likelihood estimate (MLE)? In this work, we give theoretical and empirical evidence that over-parameterizing the mixture model can help EM achieve this objective.
Our evidence is based on models in (1) where the mixture components share a known, common covariance, i.e., we fix $\Sigma_i^*=\Sigma^*$ for all $i.$ First, we assume that the mixing weights $w_i$ are also fixed to known values. Under this model, which we call $Modell$,EM finds a stationary point of the log-likelihood function in the parameter space of component means $(\boldsymbol\theta_1,\ldots,\boldsymbol\theta_k).$ Next, we over-parameterize Model l as follows. Despite the fact that the weights fixed in Model l, we now pretend that they are not fxed. This gives a second model, which we call Model 2. Parameter estimation for Model 2 requires EM to estimate the mixing weights in addition to the component means. Finding the global maximizer of the log-likelihood over this enlarged parameter space is 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada
|
[
[
0.24267578125,
0.126953125,
0.75732421875,
0.1256103515625,
0.75732421875,
0.143310546875,
0.24267578125,
0.14453125
],
[
0.261474609375,
0.20263671875,
0.3056640625,
0.20263671875,
0.3056640625,
0.2171630859375,
0.261474609375,
0.2171630859375
],
[
0.44775390625,
0.203857421875,
0.525390625,
0.203857421875,
0.525390625,
0.21533203125,
0.44775390625,
0.21533203125
],
[
0.65869140625,
0.203857421875,
0.75244140625,
0.203857421875,
0.75244140625,
0.21533203125,
0.65869140625,
0.21533203125
],
[
0.2132568359375,
0.216552734375,
0.3505859375,
0.216552734375,
0.3505859375,
0.2310791015625,
0.2132568359375,
0.2310791015625
],
[
0.416748046875,
0.216552734375,
0.5556640625,
0.216552734375,
0.5556640625,
0.2310791015625,
0.416748046875,
0.2310791015625
],
[
0.6357421875,
0.216552734375,
0.77294921875,
0.216552734375,
0.77294921875,
0.2310791015625,
0.6357421875,
0.2310791015625
],
[
0.1953125,
0.231689453125,
0.369384765625,
0.231689453125,
0.369384765625,
0.242431640625,
0.1953125,
0.242431640625
],
[
0.396240234375,
0.231689453125,
0.580078125,
0.231689453125,
0.580078125,
0.242431640625,
0.396240234375,
0.242431640625
],
[
0.60546875,
0.231689453125,
0.8037109375,
0.231689453125,
0.8037109375,
0.242431640625,
0.60546875,
0.242431640625
],
[
0.0228729248046875,
0.26513671875,
0.05963134765625,
0.26513671875,
0.05963134765625,
0.401611328125,
0.0228729248046875,
0.401611328125
],
[
0.4599609375,
0.27783203125,
0.5390625,
0.27978515625,
0.5390625,
0.29736328125,
0.4599609375,
0.29541015625
],
[
0.233642578125,
0.311279296875,
0.76708984375,
0.311279296875,
0.76708984375,
0.32568359375,
0.233642578125,
0.32568359375
],
[
0.233642578125,
0.32568359375,
0.765625,
0.32568359375,
0.765625,
0.34033203125,
0.233642578125,
0.34033203125
],
[
0.2320556640625,
0.34033203125,
0.76708984375,
0.339111328125,
0.76708984375,
0.352783203125,
0.2320556640625,
0.354248046875
],
[
0.2320556640625,
0.352783203125,
0.7646484375,
0.352783203125,
0.7646484375,
0.3662109375,
0.2320556640625,
0.3662109375
],
[
0.233642578125,
0.367431640625,
0.765625,
0.367431640625,
0.765625,
0.380615234375,
0.233642578125,
0.380615234375
],
[
0.2320556640625,
0.380615234375,
0.7646484375,
0.380615234375,
0.7646484375,
0.39453125,
0.2320556640625,
0.39453125
],
[
0.0228729248046875,
0.401611328125,
0.062103271484375,
0.401611328125,
0.062103271484375,
0.4892578125,
0.0228729248046875,
0.4892578125
],
[
0.2344970703125,
0.396484375,
0.7646484375,
0.396484375,
0.7646484375,
0.4072265625,
0.2344970703125,
0.4072265625
],
[
0.2320556640625,
0.4091796875,
0.7646484375,
0.4091796875,
0.7646484375,
0.42236328125,
0.2320556640625,
0.42236328125
],
[
0.233642578125,
0.42236328125,
0.765625,
0.42236328125,
0.765625,
0.436279296875,
0.233642578125,
0.436279296875
],
[
0.2344970703125,
0.438232421875,
0.7646484375,
0.438232421875,
0.7646484375,
0.448974609375,
0.2344970703125,
0.448974609375
],
[
0.233642578125,
0.45068359375,
0.76708984375,
0.45068359375,
0.76708984375,
0.464111328125,
0.233642578125,
0.464111328125
],
[
0.2344970703125,
0.46533203125,
0.765625,
0.46533203125,
0.765625,
0.47607421875,
0.2344970703125,
0.47607421875
],
[
0.233642578125,
0.477783203125,
0.765625,
0.477783203125,
0.765625,
0.4912109375,
0.233642578125,
0.4912109375
],
[
0.021240234375,
0.4912109375,
0.05963134765625,
0.4912109375,
0.05963134765625,
0.7119140625,
0.021240234375,
0.7119140625
],
[
0.233642578125,
0.4912109375,
0.765625,
0.4912109375,
0.765625,
0.5048828125,
0.233642578125,
0.5048828125
],
[
0.2320556640625,
0.50390625,
0.765625,
0.5048828125,
0.765625,
0.51953125,
0.2320556640625,
0.51806640625
],
[
0.2320556640625,
0.51953125,
0.71826171875,
0.51953125,
0.71826171875,
0.53271484375,
0.2320556640625,
0.53271484375
],
[
0.178955078125,
0.5625,
0.1854248046875,
0.5625,
0.1854248046875,
0.5693359375,
0.178955078125,
0.5693359375
],
[
0.1993408203125,
0.5595703125,
0.31201171875,
0.5595703125,
0.31201171875,
0.57373046875,
0.1993408203125,
0.57373046875
],
[
0.1707763671875,
0.5869140625,
0.826171875,
0.5859375,
0.826171875,
0.60400390625,
0.1707763671875,
0.60498046875
],
[
0.1748046875,
0.60498046875,
0.464111328125,
0.60498046875,
0.464111328125,
0.61572265625,
0.1748046875,
0.61572265625
],
[
0.5087890625,
0.62744140625,
0.51708984375,
0.62744140625,
0.51708984375,
0.63330078125,
0.5087890625,
0.63330078125
],
[
0.463134765625,
0.63525390625,
0.484375,
0.63525390625,
0.484375,
0.64794921875,
0.463134765625,
0.64794921875
],
[
0.5,
0.63623046875,
0.6259765625,
0.63623046875,
0.6259765625,
0.65380859375,
0.5,
0.65380859375
],
[
0.80224609375,
0.63818359375,
0.826171875,
0.63818359375,
0.826171875,
0.65283203125,
0.80224609375,
0.65283203125
],
[
0.5009765625,
0.65478515625,
0.52392578125,
0.65478515625,
0.52392578125,
0.66748046875,
0.5009765625,
0.66748046875
],
[
0.371826171875,
0.64013671875,
0.453369140625,
0.64013671875,
0.453369140625,
0.65478515625,
0.371826171875,
0.65478515625
],
[
0.1732177734375,
0.673828125,
0.826171875,
0.6748046875,
0.826171875,
0.689453125,
0.1732177734375,
0.68798828125
],
[
0.1748046875,
0.68798828125,
0.826171875,
0.68798828125,
0.826171875,
0.70263671875,
0.1748046875,
0.70263671875
],
[
0.1732177734375,
0.70263671875,
0.82177734375,
0.70263671875,
0.82177734375,
0.71337890625,
0.1732177734375,
0.71337890625
],
[
0.1732177734375,
0.71728515625,
0.82421875,
0.71728515625,
0.82421875,
0.73095703125,
0.1732177734375,
0.73095703125
],
[
0.1748046875,
0.72998046875,
0.82421875,
0.72998046875,
0.82421875,
0.74365234375,
0.1748046875,
0.74365234375
],
[
0.1748046875,
0.744140625,
0.826171875,
0.744140625,
0.826171875,
0.75830078125,
0.1748046875,
0.75830078125
],
[
0.1748046875,
0.76025390625,
0.82275390625,
0.76025390625,
0.82275390625,
0.77099609375,
0.1748046875,
0.77099609375
],
[
0.1748046875,
0.771484375,
0.408447265625,
0.771484375,
0.408447265625,
0.78515625,
0.1748046875,
0.78515625
],
[
0.1748046875,
0.79248046875,
0.82421875,
0.79248046875,
0.82421875,
0.80615234375,
0.1748046875,
0.80615234375
],
[
0.1748046875,
0.80615234375,
0.82421875,
0.80615234375,
0.82421875,
0.82080078125,
0.1748046875,
0.82080078125
],
[
0.1748046875,
0.82080078125,
0.826171875,
0.82080078125,
0.826171875,
0.833984375,
0.1748046875,
0.833984375
],
[
0.1732177734375,
0.833984375,
0.82275390625,
0.833984375,
0.82275390625,
0.8486328125,
0.1732177734375,
0.8486328125
],
[
0.1732177734375,
0.84765625,
0.82275390625,
0.84765625,
0.82275390625,
0.861328125,
0.1732177734375,
0.861328125
],
[
0.1748046875,
0.861328125,
0.82421875,
0.861328125,
0.82421875,
0.875,
0.1748046875,
0.875
],
[
0.1748046875,
0.87548828125,
0.82421875,
0.87548828125,
0.82421875,
0.8896484375,
0.1748046875,
0.8896484375
],
[
0.1748046875,
0.88818359375,
0.826171875,
0.88818359375,
0.826171875,
0.90283203125,
0.1748046875,
0.90283203125
],
[
0.1732177734375,
0.92236328125,
0.7255859375,
0.923828125,
0.7255859375,
0.93798828125,
0.1732177734375,
0.9375
]
] |
[
"Benefits of over-parameterization with EM",
"JiXu",
"Daniel Hsu",
"Arian Maleki",
"Columbia University",
"Columbia University",
"Columbia University",
"jixu@cs.columbia.edu",
"djhsu@cs.columbia.edu",
"arian@stat.columbia.edu",
"81021009",
"Abstract",
"Expectation Maximization (EM) is among the most popular algorithms for maxi-",
"mum likelihood estimation, but it is generally only guaranteed to find its stationary",
"points of the log-likelihood objective. The goal of this article is to present theoreti-",
"cal and empirical evidence that over-parameterization can help EM avoid spurious",
"local optima in the log-likelihood. We consider the problem of estimating the",
"mean vectors of a Gaussian mixture model in a scenario where the mixing weights",
"[T'S]",
"areknown.OurstudyshowsthattheglobalbehaviorofEM,whenoneusesan",
"over-parameterized model in which the mixing weights are treated as unknown, is",
"better than that when one uses the (correct) model with the mixing weights fixed to",
"theknownvalues.ForsymmetricGaussiansmixtureswithtwocomponents,we",
"prove that introducing the (statistically redundant) weight parameters enables EM",
"tofindtheglobalmaximizerofthelog-likelihoodstartingfromalmostanyinitial",
"mean parameters, whereas EM without this over-parameterization may very often",
"11011",
"fail. For other Gaussian mixtures, we provide empirical evidence that shows similar",
"behavior. Our results corroborate the value of over-parameterization in solving",
"non-convex optimization problems, previously observed in other domains.",
"1",
"Introduction",
"In a Gaussian mixture model (GMM), the observed data ) = (yi, y2, ..., yn) C Rd comprise an",
"i.i.d. sample from a mixture of k Gaussians:",
"",
"iid.",
"Ew N(Ot,)",
"(1)",
"i=1",
"Yi,...,Yn",
"where (w,, ,, ) denote the weight, mean, and covariance matrix of the ith mixture component.",
"Parameters of the GMM are often estimated using the Expectation Maximization (EM) algorithm,",
"whichaimstofindthemaximizerofthelog-likelihood objective.However,thelog-likelihoodfunction",
"is not concave, so EM is only guaranteed to find its stationary points. This leads to the following",
"natural and fundamental question in the study of EM and non-convex optimization: How can EM",
"escape spurious local maxima and saddle points to reach the maximum likelihood estimate (MLE)?",
"Inthiswork,wegivetheoretical andempirical evidencethatover-parameterizingthemixturemodel",
"can help EM achieve this objective.",
"Our evidence is based on models in (Il) where the mixture components share a known, common",
"covariance, i.e., we fix E, = E* for all i. First, we assume that the mixing weights w; are also",
"fixed to known values. Under this model, which we call Model I, EM finds a stationary point of",
"the log-likelihood function in the parameter space of component means (1,:::, k). Next, we",
"over-parameterize Model 1 as follows. Despite the fact that the weights fixed in Model 1, we now",
"pretend that they are not fixed. This gives a second model, which we call Model 2. Parameter",
"estimation for Model 2 requires EM to estimate the mixing weights in addition to the component",
"means. Finding the global maximizer of the log-likelihood over this enlarged parameter space is",
"32nd Conference on Neural Information Processing Systems (NIPS 2018), Montreal, Canada."
] |
{
"arxiv_id": "1810.11344",
"language": "en",
"page_idx": 0
}
|
||
1810.11344-en-1
|
1810.11344-en-1.jpg
|
seemingly more difficult for Model 2 than it is for Model 1, and perhaps needlessly so. However, in
this paper we present theoretical and empirical evidence to the contrary.
1. For mixtures of two symmetric Gaussians (i.e., $k=2$ and $\boldsymbol{\theta}_1^*=-\boldsymbol{\theta}_2^*)$, we prove that EM for
Model 2 converges to the global maximizer of the log-likelihood objective with almost any initialization of the mean parameters, while EM for Model 1 will fail to do so for many choices of $(w_1^*,w_2^*).$ These results are established for idealized executions of EM in an infinite sample size limit, which we complement with finite sample results.
2. We prove that the spurious local maxima in the (population) log-likelihood objective for Model
1 are eliminated in the objective for Model 2.
3. We present an empirical study to show that for more general mixtures of Gaussians, with a
variety of model parameters and sample sizes, EM for Model 2 has higher probability to find the
MLE than Model 1 under random initializations.
Related work. Since Dempster's 1977 paper [Dempster et al., 1977],the EM algorithm has become one of the most popular algorithms to find the MLE for mixture models. Due to its popularity, the convergence analysis of EM has attracted researchers' attention for years. Local convergence of EM has been shown by Wu [1983], Xu and Jordan [1996], Tseng [2004], Chrétien and Hero [2008]. Further, for certain models and under various assumptions about the initialization, EM has been shown to converge to the MLE [Redner and Walker, 1984, Balakrishnan et al., 2017, Klusowski and Brinda, 2016, Yan et al., 2017]. Typically, the initialization is required to be sufficiently close to the true parameter values of the data-generating distribution. Much less is known about global convergence of EM, as the landscape of the log-likelihood function has not been well-studied. For GMMs, Xu et al.[2016] and [Daskalakis et al., 2017] study mixtures of two Gaussians with equal weights and show that the log-likelihood objective has only two global maxima and one saddle point; and if EM is randomly initialized in a natural way, the probability that EM converges to this saddle point is zero. (Our Theorem 2 generalizes these results.) It is known that for mixtures of three or more Gaussians, global convergence is not generally possible[Jin et al.,2016].
The value of over-parameterization for local or greedy search algorithms that aim to find a global minimizer of non-convex objectives has been rigorously established in other domains. Matrix completion is a concrete example: the goal is to recover of a rank $r\ll n$ matrix $M\in\mathbb{R}^n\times n$ from observations of randomly chosen entries [Candès and Recht, 2009]. A direct approach to this problem is to find the matrix $X\in\mathbb{R}^{n\times n}$ of minimum rank that is consistent with the observed entries of $M.$ However, this optimization problem is NP-hard in general, despite the fact that there are only $2nr-r^2\ll n^2\hat{\text{ degrees-of-freedom. An indirect approach to this matrix completion problem is to find}}$ a matrix $X$ of smallest nuclear norm, subject to the same constraints; this is a convex relaxation of the rank minimization problem. By considering all $n^2$ degrees-of-freedom, Candès and Tao[2010] show that the matrix $M$ is exactly recovered via nuclear norm minimization as soon as $\Omega(nr\log^6n)$ entries are observed (with high probability). Notably, this combination of over-parameterization with convex relaxation works well in many other research problems such as sparse-PCA [d'Aspremont et al., 2005] and compressive sensing [Donoho, 2006]. However, many problems (like ours) do not have a straightforward convex relaxation. Therefore, it is important to understand how over-parameterization can help one solve a non-convex problem other than convex relaxation.
Another line of work in which the value of over-parameterization is observed is in deep learning. It is conjectured that the use of over-parameterization is the main reason for the success of local search algorithms in learning good parameters for neural nets [Livni et al., 2014, Safran and Shamir, 2017]. Recently, Haeffele and Vidal [2015], Nguyen and Hein [2017,2018], Soltani and Hegde [2018], Du and Lee[2018] confırm this observation for many neural networks such as feedforward and convolutional neural networks.
2 Theoretical results
In this section, we present our main theoretical results concerning EM and two-component Gaussian
mixture models.
2
|
[
[
0.1732177734375,
0.0921630859375,
0.82421875,
0.091552734375,
0.82421875,
0.1060791015625,
0.1732177734375,
0.106689453125
],
[
0.1732177734375,
0.106689453125,
0.64208984375,
0.106689453125,
0.64208984375,
0.12060546875,
0.1732177734375,
0.12060546875
],
[
0.1871337890625,
0.136962890625,
0.82275390625,
0.1363525390625,
0.82275390625,
0.15087890625,
0.1871337890625,
0.1514892578125
],
[
0.2059326171875,
0.149658203125,
0.82275390625,
0.15087890625,
0.82275390625,
0.1654052734375,
0.2059326171875,
0.1641845703125
],
[
0.20751953125,
0.1654052734375,
0.82275390625,
0.1654052734375,
0.82275390625,
0.1787109375,
0.20751953125,
0.1787109375
],
[
0.20751953125,
0.179931640625,
0.82421875,
0.179931640625,
0.82421875,
0.1932373046875,
0.20751953125,
0.1932373046875
],
[
0.20751953125,
0.1932373046875,
0.59619140625,
0.1932373046875,
0.59619140625,
0.20703125,
0.20751953125,
0.20703125
],
[
0.1846923828125,
0.212158203125,
0.82421875,
0.21337890625,
0.82421875,
0.2279052734375,
0.1846923828125,
0.2266845703125
],
[
0.2059326171875,
0.2291259765625,
0.50244140625,
0.2291259765625,
0.50244140625,
0.242431640625,
0.2059326171875,
0.242431640625
],
[
0.1846923828125,
0.25,
0.82275390625,
0.25,
0.82275390625,
0.26318359375,
0.1846923828125,
0.26318359375
],
[
0.20751953125,
0.26318359375,
0.82421875,
0.26318359375,
0.82421875,
0.276611328125,
0.20751953125,
0.276611328125
],
[
0.20751953125,
0.276611328125,
0.52685546875,
0.276611328125,
0.52685546875,
0.290283203125,
0.20751953125,
0.290283203125
],
[
0.1748046875,
0.317626953125,
0.82421875,
0.317626953125,
0.82421875,
0.330810546875,
0.1748046875,
0.330810546875
],
[
0.1748046875,
0.33203125,
0.82421875,
0.33203125,
0.82421875,
0.34521484375,
0.1748046875,
0.34521484375
],
[
0.1732177734375,
0.34521484375,
0.826171875,
0.3447265625,
0.826171875,
0.35791015625,
0.1732177734375,
0.359130859375
],
[
0.1748046875,
0.359130859375,
0.826171875,
0.359130859375,
0.826171875,
0.37255859375,
0.1748046875,
0.37255859375
],
[
0.1748046875,
0.37255859375,
0.82421875,
0.37255859375,
0.82421875,
0.386474609375,
0.1748046875,
0.386474609375
],
[
0.1748046875,
0.386474609375,
0.82421875,
0.386474609375,
0.82421875,
0.399658203125,
0.1748046875,
0.399658203125
],
[
0.1732177734375,
0.40087890625,
0.82275390625,
0.40087890625,
0.82275390625,
0.4140625,
0.1732177734375,
0.4140625
],
[
0.1748046875,
0.41357421875,
0.82421875,
0.41357421875,
0.82421875,
0.427978515625,
0.1748046875,
0.427978515625
],
[
0.1748046875,
0.427978515625,
0.82421875,
0.427978515625,
0.82421875,
0.44140625,
0.1748046875,
0.44140625
],
[
0.1748046875,
0.44140625,
0.82421875,
0.44140625,
0.82421875,
0.455078125,
0.1748046875,
0.455078125
],
[
0.1748046875,
0.455810546875,
0.826171875,
0.455810546875,
0.826171875,
0.4697265625,
0.1748046875,
0.4697265625
],
[
0.1748046875,
0.4697265625,
0.82421875,
0.4697265625,
0.82421875,
0.48291015625,
0.1748046875,
0.48291015625
],
[
0.1748046875,
0.484130859375,
0.82421875,
0.484130859375,
0.82421875,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.1748046875,
0.496826171875,
0.689453125,
0.496826171875,
0.689453125,
0.51025390625,
0.1748046875,
0.51025390625
],
[
0.1732177734375,
0.517578125,
0.82421875,
0.517578125,
0.82421875,
0.53076171875,
0.1732177734375,
0.53076171875
],
[
0.1748046875,
0.53076171875,
0.77294921875,
0.53076171875,
0.77294921875,
0.54443359375,
0.1748046875,
0.54443359375
],
[
0.76708984375,
0.5322265625,
0.82275390625,
0.5322265625,
0.82275390625,
0.54345703125,
0.76708984375,
0.54345703125
],
[
0.1707763671875,
0.54345703125,
0.826171875,
0.54150390625,
0.826171875,
0.5595703125,
0.1707763671875,
0.56103515625
],
[
0.1748046875,
0.5595703125,
0.82421875,
0.5595703125,
0.82421875,
0.57275390625,
0.1748046875,
0.57275390625
],
[
0.1748046875,
0.5712890625,
0.826171875,
0.5712890625,
0.826171875,
0.5859375,
0.1748046875,
0.5859375
],
[
0.1748046875,
0.5859375,
0.82421875,
0.5859375,
0.82421875,
0.599609375,
0.1748046875,
0.599609375
],
[
0.17236328125,
0.5986328125,
0.82421875,
0.599609375,
0.82421875,
0.6142578125,
0.17236328125,
0.61279296875
],
[
0.1732177734375,
0.6142578125,
0.82421875,
0.6142578125,
0.82421875,
0.62744140625,
0.1732177734375,
0.62744140625
],
[
0.1748046875,
0.62744140625,
0.82421875,
0.62744140625,
0.82421875,
0.6416015625,
0.1748046875,
0.6416015625
],
[
0.1732177734375,
0.642578125,
0.82421875,
0.642578125,
0.82421875,
0.65576171875,
0.1732177734375,
0.65576171875
],
[
0.1748046875,
0.6572265625,
0.82421875,
0.6572265625,
0.82421875,
0.67041015625,
0.1748046875,
0.67041015625
],
[
0.176513671875,
0.67236328125,
0.82421875,
0.67236328125,
0.82421875,
0.68310546875,
0.176513671875,
0.68310546875
],
[
0.1748046875,
0.6845703125,
0.826171875,
0.6845703125,
0.826171875,
0.69775390625,
0.1748046875,
0.69775390625
],
[
0.1748046875,
0.69873046875,
0.82421875,
0.69873046875,
0.82421875,
0.7119140625,
0.1748046875,
0.7119140625
],
[
0.1748046875,
0.7119140625,
0.6396484375,
0.7119140625,
0.6396484375,
0.72607421875,
0.1748046875,
0.72607421875
],
[
0.1732177734375,
0.73095703125,
0.826171875,
0.73193359375,
0.826171875,
0.7470703125,
0.1732177734375,
0.74560546875
],
[
0.1748046875,
0.7470703125,
0.82421875,
0.7470703125,
0.82421875,
0.76025390625,
0.1748046875,
0.76025390625
],
[
0.1748046875,
0.76025390625,
0.826171875,
0.76025390625,
0.826171875,
0.7734375,
0.1748046875,
0.7734375
],
[
0.1748046875,
0.7734375,
0.82421875,
0.7734375,
0.82421875,
0.787109375,
0.1748046875,
0.787109375
],
[
0.176513671875,
0.787109375,
0.82421875,
0.787109375,
0.82421875,
0.80029296875,
0.176513671875,
0.80029296875
],
[
0.1748046875,
0.8017578125,
0.404296875,
0.8017578125,
0.404296875,
0.81494140625,
0.1748046875,
0.81494140625
],
[
0.176513671875,
0.8486328125,
0.1895751953125,
0.8486328125,
0.1895751953125,
0.859375,
0.176513671875,
0.859375
],
[
0.1993408203125,
0.8466796875,
0.361083984375,
0.8466796875,
0.361083984375,
0.861328125,
0.1993408203125,
0.861328125
],
[
0.17236328125,
0.88330078125,
0.82421875,
0.8837890625,
0.82421875,
0.8984375,
0.17236328125,
0.89794921875
],
[
0.1732177734375,
0.8984375,
0.281982421875,
0.8984375,
0.281982421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.491943359375,
0.9345703125,
0.50732421875,
0.9345703125,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"seemingly more difficult for Model 2 than it is for Model 1, and perhaps needlessly so. However, in",
"this paper we present theoretical and empirical evidence to the contrary",
"1. For mixtures of two symmetric Gaussians (i.e., k = 2 and o =-o), we prove that EM for",
"Model 2 converges to the global maximizer of the log-likelihood objective with almost any",
"initialization of the mean parameters, while EM for Model I will fail to do so for many choices",
"of (w,w).These results are established for idealized executions of EM in an infinite sample",
"size limit, which we complement with finite sample results.",
"2. We prove that the spurious local maxima in the (population) log-likelihood objective for Model",
"I are eliminated in the objective for Model 2.",
"3. We present an empirical study to show that for more general mixtures of Gaussians, with a",
"variety of model parameters and sample sizes, EM for Model 2 has higher probability to find the",
"MLE than Model 1 under random initializations.",
"Related work.Since Dempster's 1977 paper [Dempster et al., 1977], the EM algorithm has become",
"one of the most popular algorithms to find the MLE for mixture models. Due to its popularity, the",
"convergence analysis of EM has attracted researchers'attention for years.Local convergence of",
"EM has been shown by Wu [1983], Xu and Jordan [1996], Tseng [2004], Chretien and Hero [2008].",
"Further, for certain models and under various assumptions about the initialization, EM has been",
"shown to converge to the MLE [Redner and Walker, 1984, Balakrishnan et al., 2017, Klusowski",
"and Brinda, 2016, Yan et al., 2017]. Typically, the initialization is required to be sufficiently close",
"to the true parameter values of the data-generating distribution. Much less is known about global",
"convergence of EM, as the landscape of the log-likelihood function has not been well-studied. For",
"GMMs, Xu et al. [2016] and [Daskalakis et al., 2017] study mixtures of two Gaussians with equal",
"weights and show that the log-likelihood objective has only two global maxima and one saddle point;",
"and if EM is randomly initialized in a natural way, the probability that EM converges to this saddle",
"point is zero. (Our Theorem 2 generalizes these results.) It is known that for mixtures of three or",
"more Gaussians, global convergence is not generally possible [Jin et al., 2016].",
"The value of over-parameterization for local or greedy search algorithms that aim to find a global",
"minimizer of non-convex objectives has been rigorously established in other domains.",
"Matrix",
"completion is a concrete example: the goal is to recover of a rank r < n matrix M e Rnxn from",
"observations of randomly chosen entries [Candes and Recht, 2009]. A direct approach to this problem",
"is to find the matrix X e Rnxn of minimum rank that is consistent with the observed entries of",
"M. However, this optimization problem is NP-hard in general, despite the fact that there are only",
"2nr-r2 < n? degrees-of-freedom. An indirect approach to this matrix completion problem is to find",
"a matrix X of smallest nuclear norm, subject to the same constraints; this is a convex relaxation of the",
"rank minimization problem. By considering all n? degrees-of-freedom, Candes and Tao [2010] show",
"that the matrix M is exactlyrecovered via nuclear norm minimization as soon as 2(nr logn) entries",
"are observed (with high probability).Notably, this combination of over-parameterization with convex",
"relaxationworks wellin many otherresearchproblems such as sparse-PCA[d'Aspremont etal..",
"2005] and compressive sensing [Donoho, 2006]. However, many problems (like ours) do not have a",
"straightforward convex relaxation.Therefore, it is important to understand how over-parameterization",
"can help one solve a non-convex problem other than convex relaxation.",
"Another line of work in which the value of over-parameterization is observed is in deep learning.",
"It is conjectured that the use of over-parameterization is the main reason for the success of local",
"search algorithms in learning good parameters for neural nets [Livni et al., 2014, Safran and Shamir,",
"2017]. Recently, Haeffele and Vidal [2015], Nguyen and Hein [2017, 2018], Soltani and Hegde",
"[2018], Du and Lee [2018] confirm this observation for many neural networks such as feedforward",
"and convolutional neural networks.",
"2",
"Theoretical results",
"In this section, we present our main theoretical results concerning EM and two-component Gaussian",
"mixture models.",
"2"
] |
{
"arxiv_id": "1810.11344",
"language": "en",
"page_idx": 1
}
|
||
1810.11344-en-2
|
1810.11344-en-2.jpg
|
2.1 Sample-based EM and Population EM
Without loss of generality, we assume $\Sigma^*=I$. We consider the following Gaussian mixture model:
$$y_{1},\ldots,y_{n}\stackrel{\mathrm{i.i.d.}}{\sim}w_{1}^{*}\mathcal{N}(\boldsymbol{\theta}^{*},\boldsymbol{I})+w_{2}^{*}\mathcal{N}(-\boldsymbol{\theta}^{*},\boldsymbol{I}).$$
(2)
The mixing weights $w_1^*$ and $w_2^*$ are fxed (i.e., assumed to be known). Without loss of generality, we also assume that $w_1^*\geq w_2^*>\bar{0}($and, of course, $w_1^*+w_2^*=1).$ The only parameter to estimate is the mean vector $\theta^*.\bar{\text{TheEMalgorithmforthismodelusesthefollowingiterations:}}$
$$\hat{\boldsymbol{\theta}}^{\langle t+1\rangle}\:=\:\frac{1}{n}\sum_{i=1}^{n}\left[\frac{w_{1}^{*}e^{\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}-w_{2}^{*}e^{-\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}{w_{1}^{*}e^{\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}+w_{2}^{*}e^{-\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}\boldsymbol{y}_{i}\right].$$
(3)
We refer to this algorithm as Sample-based EM$_1:$ it is the EM algorithm one would normally use when the mixing weights are known. In spite of this, we also consider an EM algorithm that pretends that the weights are not known, and estimates them alongside the mean parameters. We refer to this algorithm as Sample-based EM$_2$, which uses the following iterations:
$$\begin{aligned}&\hat{w}_{1}^{\langle t+1\rangle}&&=\quad\frac{1}{n}\sum_{i=1}^{n}\left[\frac{\hat{w}_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}{\hat{w}_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}+\hat{w}_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}\right]\:=\:1-\hat{w}_{2}^{\langle t+1\rangle}.\\&\hat{\boldsymbol{\theta}}^{\langle t+1\rangle}&&=\quad\frac{1}{n}\sum_{i=1}^{n}\left[\frac{\hat{w}_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}-\hat{w}_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}{\hat{w}_{1}^{\langle t\rangle}e\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle+\hat{w}_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y}_{i},\hat{\boldsymbol{\theta}}^{\langle t\rangle}\rangle}}\boldsymbol{y}_{i}\right].\\\end{aligned}$$
(4)
This is the EM algorithm for a different Gaussian mixture model in which the weights $w_1^*$ and $w_2^*$ are not fixed (i.e., unknown), and hence must be estimated. Our goal is to study the global convergence properties of the above two EM algorithms on data from the first model, where the mixing weights are, in fact, known.
We study idealized executions of the EM algorithms in the large sample limit, where the algorithms are modified to be computed over an infinitely large i.i.d. sample drawn from the mixture distribution in (2). Specifically, we replace the empirical averages in (3) and (4) with the expectations with respect to the mixture distribution. We obtain the following two modified EM algorithms, which we refer to as Population EM$_1$ and Population EM$_2:$
$\bullet$ Population EM$_1:$
$$\boldsymbol{\theta}^{\langle t+1\rangle}\quad=\quad\mathbb{E}_{\boldsymbol{y}\sim f^*}\left[\frac{w_1^*e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}-w_2^*e^{-\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}}{w_1^*e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}+w_2^*e^{-\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}}\boldsymbol{y}\right]\:=:\:H(\boldsymbol{\theta}^{\langle t\rangle};\boldsymbol{\theta}^*,w_1^*),$$
(5)
where $f^*=f^*(\boldsymbol{\theta}^*,w_1^*)$ here denotes the true distribution of $y_i$ given in (2).
$\bullet$ Population EM$_2:$ Set $w_1^{\langle0\rangle}=w_2^{\langle0\rangle}=0.5^1$, and run
(6)
$$\begin{aligned}&w_{1}^{\langle t+1\rangle}&&=\quad\mathbb{E}_{\boldsymbol{y}\sim f^{*}}\left[\frac{w_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}}{w_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}+w_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}}\right]\:=:\:G_{w}(\boldsymbol{\theta}^{\langle t\rangle},w^{\langle t\rangle};\boldsymbol{\theta}^{*},w_{1}^{*})\\&&&=\quad1-w_{2}^{\langle t+1\rangle}.\\&\boldsymbol{\theta}^{\langle t+1\rangle}&&=\quad\mathbb{E}_{\boldsymbol{y}\sim f^{*}}\left[\frac{w_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}-w_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y},\boldsymbol{\theta}^{\langle t\rangle}\rangle}}{w_{1}^{\langle t\rangle}e^{\langle\boldsymbol{y},\boldsymbol{\theta}^{(t)}\rangle}+w_{2}^{\langle t\rangle}e^{-\langle\boldsymbol{y},\boldsymbol{\theta}^{(t)}\rangle}}\boldsymbol{y}\right]\:=:\:G_{\theta}(\boldsymbol{\theta}^{\langle t\rangle},w^{\langle t\rangle};\boldsymbol{\theta}^{*},w_{1}^{*}).\end{aligned}$$
As $n\to\infty$, we can show the performance of Sample-based EM$_\star$ converges to that of the Population EM$_\star$ in probability. This argument has been used rigorously in many previous works on EM [Balakrishnan et al., 2017, Xu et al., 2016, Klusowski and Brinda, 2016, Daskalakis et al. 2017]. The main goal of this section, however, is to study the dynamics of Population EM$_1$ and Population EM$_2$, and the landscape of the log-likelihood objectives of the two models
l Using equal initial weights is a natural way to initialize EM when the weights are unknown.
3
|
[
[
0.1732177734375,
0.09027099609375,
0.48046875,
0.091552734375,
0.48046875,
0.1060791015625,
0.1732177734375,
0.10479736328125
],
[
0.1748046875,
0.117431640625,
0.826171875,
0.117431640625,
0.826171875,
0.1307373046875,
0.1748046875,
0.1307373046875
],
[
0.419189453125,
0.1331787109375,
0.6640625,
0.1383056640625,
0.6640625,
0.160400390625,
0.419189453125,
0.1546630859375
],
[
0.334228515625,
0.1427001953125,
0.416748046875,
0.1427001953125,
0.416748046875,
0.1572265625,
0.334228515625,
0.1572265625
],
[
0.8037109375,
0.141357421875,
0.82421875,
0.141357421875,
0.82421875,
0.1546630859375,
0.8037109375,
0.1546630859375
],
[
0.1732177734375,
0.1590576171875,
0.82275390625,
0.1590576171875,
0.82275390625,
0.17236328125,
0.1732177734375,
0.17236328125
],
[
0.1748046875,
0.173583984375,
0.82421875,
0.173583984375,
0.82421875,
0.1875,
0.1748046875,
0.1875
],
[
0.1732177734375,
0.1881103515625,
0.69921875,
0.1881103515625,
0.69921875,
0.2020263671875,
0.1732177734375,
0.2020263671875
],
[
0.455078125,
0.21337890625,
0.54345703125,
0.2052001953125,
0.5458984375,
0.222900390625,
0.45751953125,
0.2310791015625
],
[
0.53515625,
0.210205078125,
0.64892578125,
0.205810546875,
0.650390625,
0.2247314453125,
0.5361328125,
0.2291259765625
],
[
0.306396484375,
0.2197265625,
0.358642578125,
0.20703125,
0.366943359375,
0.2259521484375,
0.314453125,
0.238037109375
],
[
0.41015625,
0.2197265625,
0.418212890625,
0.2197265625,
0.418212890625,
0.2235107421875,
0.41015625,
0.2235107421875
],
[
0.419189453125,
0.2259521484375,
0.431396484375,
0.222900390625,
0.4345703125,
0.2310791015625,
0.42236328125,
0.2342529296875
],
[
0.80224609375,
0.2235107421875,
0.826171875,
0.2235107421875,
0.826171875,
0.2374267578125,
0.80224609375,
0.2374267578125
],
[
0.455078125,
0.2310791015625,
0.66259765625,
0.2259521484375,
0.6640625,
0.2474365234375,
0.455810546875,
0.25244140625
],
[
0.1732177734375,
0.2568359375,
0.82275390625,
0.2568359375,
0.82275390625,
0.270751953125,
0.1732177734375,
0.270751953125
],
[
0.1748046875,
0.271484375,
0.82421875,
0.271484375,
0.82421875,
0.285400390625,
0.1748046875,
0.285400390625
],
[
0.17236328125,
0.2841796875,
0.82421875,
0.2841796875,
0.82421875,
0.29736328125,
0.17236328125,
0.29736328125
],
[
0.1732177734375,
0.2998046875,
0.62841796875,
0.2998046875,
0.62841796875,
0.313232421875,
0.1732177734375,
0.313232421875
],
[
0.46728515625,
0.32080078125,
0.55810546875,
0.318115234375,
0.55810546875,
0.3359375,
0.46728515625,
0.33837890625
],
[
0.25732421875,
0.335205078125,
0.3056640625,
0.322509765625,
0.313720703125,
0.339111328125,
0.265625,
0.3515625
],
[
0.68115234375,
0.3359375,
0.69921875,
0.3359375,
0.69921875,
0.3447265625,
0.68115234375,
0.3447265625
],
[
0.701171875,
0.33203125,
0.7255859375,
0.33203125,
0.7255859375,
0.339111328125,
0.701171875,
0.339111328125
],
[
0.423095703125,
0.341552734375,
0.54345703125,
0.341552734375,
0.54345703125,
0.359130859375,
0.423095703125,
0.359130859375
],
[
0.52685546875,
0.3447265625,
0.5986328125,
0.341552734375,
0.6005859375,
0.35595703125,
0.52783203125,
0.359130859375
],
[
0.428955078125,
0.37255859375,
0.5009765625,
0.369384765625,
0.50244140625,
0.38330078125,
0.4306640625,
0.386474609375
],
[
0.53125,
0.373779296875,
0.611328125,
0.36865234375,
0.61279296875,
0.38330078125,
0.53173828125,
0.38818359375
],
[
0.26318359375,
0.3818359375,
0.303955078125,
0.37255859375,
0.3095703125,
0.386962890625,
0.268798828125,
0.396484375
],
[
0.80224609375,
0.3837890625,
0.826171875,
0.3837890625,
0.826171875,
0.397705078125,
0.80224609375,
0.397705078125
],
[
0.428955078125,
0.39453125,
0.491943359375,
0.391357421875,
0.49267578125,
0.406005859375,
0.4306640625,
0.4091796875
],
[
0.5458984375,
0.39453125,
0.61767578125,
0.386962890625,
0.62109375,
0.40283203125,
0.54833984375,
0.410400390625
],
[
0.1732177734375,
0.41845703125,
0.82177734375,
0.41845703125,
0.82177734375,
0.431884765625,
0.1732177734375,
0.431884765625
],
[
0.1748046875,
0.43310546875,
0.39453125,
0.43310546875,
0.39453125,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.408447265625,
0.43310546875,
0.82177734375,
0.43310546875,
0.82177734375,
0.4462890625,
0.408447265625,
0.4462890625
],
[
0.1732177734375,
0.4462890625,
0.82421875,
0.4462890625,
0.82421875,
0.460205078125,
0.1732177734375,
0.460205078125
],
[
0.1748046875,
0.460205078125,
0.30224609375,
0.460205078125,
0.30224609375,
0.473388671875,
0.1748046875,
0.473388671875
],
[
0.17236328125,
0.479248046875,
0.82275390625,
0.479736328125,
0.82275390625,
0.494384765625,
0.17236328125,
0.49365234375
],
[
0.176513671875,
0.496826171875,
0.82177734375,
0.496826171875,
0.82177734375,
0.5068359375,
0.176513671875,
0.5068359375
],
[
0.1732177734375,
0.5087890625,
0.82177734375,
0.5087890625,
0.82177734375,
0.52294921875,
0.1732177734375,
0.52294921875
],
[
0.17236328125,
0.52099609375,
0.826171875,
0.521484375,
0.826171875,
0.5361328125,
0.17236328125,
0.53515625
],
[
0.1732177734375,
0.5361328125,
0.443603515625,
0.5361328125,
0.443603515625,
0.5498046875,
0.1732177734375,
0.5498046875
],
[
0.191162109375,
0.56005859375,
0.32177734375,
0.56005859375,
0.32177734375,
0.57373046875,
0.191162109375,
0.57373046875
],
[
0.4150390625,
0.59228515625,
0.5986328125,
0.59033203125,
0.5986328125,
0.609375,
0.4150390625,
0.611328125
],
[
0.268798828125,
0.60107421875,
0.334228515625,
0.5986328125,
0.3349609375,
0.6142578125,
0.26953125,
0.6162109375
],
[
0.62158203125,
0.599609375,
0.76123046875,
0.6015625,
0.76123046875,
0.619140625,
0.62158203125,
0.6171875
],
[
0.328369140625,
0.60595703125,
0.39208984375,
0.60595703125,
0.39208984375,
0.6171875,
0.328369140625,
0.6171875
],
[
0.41259765625,
0.60986328125,
0.611328125,
0.60400390625,
0.61279296875,
0.62890625,
0.414306640625,
0.63525390625
],
[
0.80224609375,
0.60498046875,
0.826171875,
0.60498046875,
0.826171875,
0.61865234375,
0.80224609375,
0.61865234375
],
[
0.205078125,
0.6337890625,
0.70361328125,
0.63525390625,
0.70361328125,
0.65283203125,
0.205078125,
0.6513671875
],
[
0.191162109375,
0.65478515625,
0.5390625,
0.65478515625,
0.5390625,
0.66943359375,
0.191162109375,
0.66943359375
],
[
0.44677734375,
0.689453125,
0.53369140625,
0.68310546875,
0.5361328125,
0.70068359375,
0.449462890625,
0.70703125
],
[
0.3056640625,
0.6962890625,
0.378173828125,
0.7021484375,
0.37744140625,
0.71728515625,
0.303955078125,
0.7119140625
],
[
0.62353515625,
0.69775390625,
0.78759765625,
0.69873046875,
0.78759765625,
0.71630859375,
0.62353515625,
0.71533203125
],
[
0.79833984375,
0.70263671875,
0.82177734375,
0.69873046875,
0.826171875,
0.7119140625,
0.80224609375,
0.71630859375
],
[
0.38818359375,
0.708984375,
0.5849609375,
0.70703125,
0.5849609375,
0.72607421875,
0.38818359375,
0.72802734375
],
[
0.2418212890625,
0.7021484375,
0.291748046875,
0.6923828125,
0.29736328125,
0.70849609375,
0.2467041015625,
0.71728515625
],
[
0.3056640625,
0.734375,
0.41259765625,
0.72900390625,
0.414306640625,
0.74755859375,
0.306396484375,
0.75244140625
],
[
0.39208984375,
0.77294921875,
0.48046875,
0.77294921875,
0.48046875,
0.787109375,
0.39208984375,
0.787109375
],
[
0.498291015625,
0.7734375,
0.5849609375,
0.76953125,
0.58642578125,
0.7841796875,
0.5,
0.78857421875
],
[
0.314453125,
0.7822265625,
0.369384765625,
0.787109375,
0.366943359375,
0.80029296875,
0.31201171875,
0.79541015625
],
[
0.625,
0.78076171875,
0.826171875,
0.78076171875,
0.826171875,
0.798828125,
0.625,
0.798828125
],
[
0.51953125,
0.79345703125,
0.583984375,
0.79052734375,
0.5849609375,
0.8046875,
0.521484375,
0.80810546875
],
[
0.23779296875,
0.78271484375,
0.2900390625,
0.77783203125,
0.293212890625,
0.79150390625,
0.240234375,
0.796875
],
[
0.39453125,
0.79443359375,
0.51123046875,
0.79443359375,
0.51123046875,
0.8095703125,
0.39453125,
0.8095703125
],
[
0.178955078125,
0.82373046875,
0.2296142578125,
0.82373046875,
0.2296142578125,
0.83056640625,
0.178955078125,
0.83056640625
],
[
0.223876953125,
0.82080078125,
0.82275390625,
0.82080078125,
0.82275390625,
0.833984375,
0.223876953125,
0.833984375
],
[
0.1732177734375,
0.83349609375,
0.82421875,
0.833984375,
0.82421875,
0.8486328125,
0.1732177734375,
0.84765625
],
[
0.176513671875,
0.849609375,
0.82177734375,
0.849609375,
0.82177734375,
0.85986328125,
0.176513671875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.826171875,
0.861328125,
0.826171875,
0.87548828125,
0.1748046875,
0.875
],
[
0.1748046875,
0.87548828125,
0.73876953125,
0.87548828125,
0.73876953125,
0.8896484375,
0.1748046875,
0.8896484375
],
[
0.1953125,
0.896484375,
0.740234375,
0.8984375,
0.740234375,
0.91357421875,
0.1953125,
0.9111328125
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"2.1Sample-basedEMandPopulationEM",
"Without loss of generality,we assume *=I.We consider the following Gaussian mixture model:",
"iidwN(o*, I) +wN(-0*, I),",
"Yi,...,Yn",
"(2)",
"The mixing weights w* and w, are fixed (i.e., assumed to be known). Without loss of generality, we",
"also assume that w* ≥ w > 0 (and, of course, w* + w* = 1). The only parameter to estimate is the",
"mean vector *. The EM algorithm for this model uses the following iterations:",
"wie(y,e(t)",
"- wse-(y,o(t)",
"(t+1)",
"",
"1",
"(3)",
"wie(y.e)+we-(y.(j",
"We refer to this algorithm as Sample-based EMi: it is the EM algorithm one would normally use",
"when the mixing weights are known. In spite of this, we also consider an EM algorithm that pretends",
"that the weights are not known, and estimates them alongside the mean parameters. We refer to this",
"algorithm as Sample-based EM2, which uses the following iterations:",
"w(t)e(y.(t)",
"w,(t+1)",
"w",
"t+1",
"(t)e(ys,e() +wst)",
"t)e-(y.,o",
"t)e(y,e(t))",
"(t\"e-(y,o(t))",
"o(t+1)",
"(4)",
"te(yi,(t)",
"e-(y,o(),",
"This is the EM algorithm for a different Gaussian mixture model in which the weights w* and w, are",
"not fixed (i.e., unknown), and hen",
" must be estimated. Our goal is to study the global convergence",
"properties of the above two EM algorithms on data from the first model, where the mixing weights",
"are, in fact, known.",
"We study idealized executions of the EM algorithms in the large sample limit, where the algorithms",
"aremodifiedtobecomputedoveraninfinitelylargei.i.d.sampledrawnfromthemixturedistribution",
"in (2). Specifically, we replace the empirical averages in (3) and (4) with the expectations with respect",
"to the mixture distribution. We obtain the following two modified EM algorithms, which we refer to",
"as Population EMi and Population EM2:",
". Population EMi:",
"wie(y,o(t)-wse-(y,o(t\")",
"0(t+1)",
"=: H(o(t);0*, w),",
"=Eu~",
"wie(u,o)+we-(u,o)",
"(5)",
"where f* = f*(0*, w) here denotes the true distribution of yi given in (2).",
". Population EM2: Set w(0)= w,0) = 0.5l, and run",
"w$t)e(y,o(t)",
"=Eu~f",
": Gw(o(t),w(t);0*,w1)",
"(6)",
"[wt)e(y,(t) +wst)e-(y,o(t)",
"w(t+1)",
"= 1- w++1)",
"w(t)e(y.o(t)).",
"(t)e-(y,o(t)",
"Ey~f",
": Ge(o(t),w(t);0*,w). (7)",
"e-(y,o(t))",
"(t+1)",
"w(t)e(y,(t))+w,",
"Asn-",
"→ oo, we can show the performance of Sample-based EM converges to that of the",
"Population EM in probability. This argument has been used rigorously in many previous works",
"onEM[Balakrishnan et al.,2017,Xuet al.,2016.KlusowskiandBrinda,2016.Daskalakisetal.",
"2017]. The main goal of this section, however, is to study the dynamics of Population EMi and",
"Population EM2, and the landscape of the log-likelihood objectives of the two models.",
"'Using equal initial weights is a natural way to initialize EM when the weights are unknown.",
"3"
] |
{
"arxiv_id": "1810.11344",
"language": "en",
"page_idx": 2
}
|
||
1810.11344-en-3
|
1810.11344-en-3.jpg
|
## 2.2 Main theoretical results
Let us first consider the special case $w_1^*=w_2^*=0.5$ Then, it is straightforward to show that $w_1^{\langle t\rangle}=$ $w_2^{\langle t\rangle}=0.5$ for all $t\text{in Population EM}_2.$ Hence, Population EM$_2$ is equivalent to Population EM$_1.$ Global convergence of $\theta^{\langle t\rangle}$ to $\theta^*$ for this case was recently established by Xu et al.[2016, Theorem 1] for almost all initial $\theta^{\langle0\rangle}$ (see also [Daskalakis et al., 2017]).
We first show that the same global convergence may $not\text{holdforPopulationEM}_1$ when $w_1^*\neq w_2^*.$ Theorem $1.ConsiderPopulationEM_{1}$ in dimension one $(i.e,\theta^*\in\mathbb{R}).For$ any $\theta^*>0,there$ $exists\delta > 0, such$ that given $w_{1}^{* }\in ( 0. 5, 0. 5+ \delta )$ and initialization $\theta ^{\langle 0\rangle }\leq - \theta ^{* }, the$ Population $EM_{1}$ estimate $\theta^{\langle t\rangle}$ converges to a fixed point $\theta_\mathrm{wrong}$ inside $(-\theta^*,0).$
This theorem, which is proved in Appendix A, implies that if we use random initialization, Population EM$_1$ may converge to the wrong fixed point with constant probability. We illustrate this in Figure 1. The iterates of Population EM$_1$ converge to a fixed point of the function $\theta\mapsto H(\theta;\theta^*,w_1^*)$ defined in (5). We have plotted this function for several different values of $w_1^*$ in the left panel of Figure 1. When $w_1^*$ is close to $1,H(\theta;\theta^*,w_1^*)$ has only one fixed point and that is at $\theta=\theta^*.$ Hence in this case, the estimates produced by Population EM$_1$ converge to the true $\theta^*.$ However, when we decrease the value of $w_1^*$ below a certain threshold (which is numerically found to be approximately 0.77 for $\theta ^* = 1) $,two other fixed points of $H( \theta ; \theta ^* , w_1^* )$ emerge. These new fixed points are foils for $\mathrm{Population~EM}_{1}.$
From the failure of Population EM$_1$, one may expect the over-parameterized Population EM$_2$ to fail as well. Yet, surprisingly, our second theorem proves the opposite is true: Population EM$_2$ has global convergence even when $w_1^*\neq w_2^*.$
Theorem 2. For any $w_1^*\in[0.5,1)$, the Population EM$_2$ estimate $(\boldsymbol{\theta}^t,w^{\langle t\rangle})$ converges to either $(\boldsymbol{\theta}^*,w_1^*)or\left(-\boldsymbol{\theta}^*,w_2^*\right)withanyinitialization\boldsymbol{\theta}^{\langle0\rangle}exceptonthehyperplane\left\langle\boldsymbol{\theta}^{\langle0\rangle},\boldsymbol{\theta}^*\right\rangle=0.Further$- more, the convergence speed is geometric after some finite number of iterations, i.e., there exists a finite number $T$ and constant $\rho \in ( 0, 1)$ such that the following $hold.$
$\bullet$ $If\langle \boldsymbol{\theta }^{\langle 0\rangle }, \boldsymbol{\theta }^{* }\rangle > 0, then$ for all $t> T$,
$$\|\boldsymbol{\theta}^{\langle t+1\rangle}-\boldsymbol{\theta}^*\|^2+|w_1^{\langle t+1\rangle}-w_1^*|^2\leq\rho^{t-T}\left(\|\boldsymbol{\theta}^{\langle T\rangle}-\boldsymbol{\theta}^*\|^2+(w_1^{\langle T\rangle}-w_1^*)^2\right).$$
$\bullet$ $If\langle \boldsymbol{\theta }^{\langle 0\rangle }, \boldsymbol{\theta }^{* }\rangle < 0, then$ for all $t> T$,
$$\|\boldsymbol{\theta}^{\langle t+1\rangle}+\boldsymbol{\theta}^*\|^2+|w_1^{\langle t+1\rangle}-w_2^*|^2\leq\rho^{t-T}\left(\|\boldsymbol{\theta}^{\langle T\rangle}+\boldsymbol{\theta}^*\|^2+(w_1^{\langle T\rangle}-w_2^*)^2\right).$$
Theorem 2 implies that if we use random initialization for $\theta^{\langle0\rangle}$, with probability one, the
Population EM$_2$ estimates converge to the true parameters.
The failure of Population EM$_1$ and success of Population EM$_2$ can be explained intuitively. Let $C_1$ and $C_2$, respectively, denote the true mixture components with parameters $(w_1^*,\theta^*)$ and $(w_2^*,-\theta^*).$ Due to the symmetry in Population EM$_1$, we are assured that among the two estimated mixture components, one will have a positive mean, and the other will have a negative mean: call these $\hat{C}_+$ and $\hat{C}_-$, respectively. Assume $\theta^*>0$ and $w_1^*>0.5$, and consider initializing the Population EM$_1$ with $\theta^{\langle0\rangle}:=-\theta^*.$ This initialization incorrectly associates $\hat{C}_-$ with the larger weight $w_1^*$ instead of the smaller weight $w_2^*.$ This causes, in the E-step of EM, the component $\hat{C}_-$ to become“responsible” for an overly large share of the overall probability mass, and in particular an overly large share of the mass from $\tilde{C}_1$ (which has a positive mean). Thus, in the M-step of EM, when the mean of the estimated component $\hat{C}_-$ is updated, it is pulled rightward towards $+\infty.$ It is possible that this rightward pull would cause the estimated mean of $\hat{C}_-$ to become positive\_in which case the roles of $\hat{C}_{+}$ and $\hat{C}_{-}$ would switch—but this will not happen as long as $w_1^*$ is sufficiently bounded away from 1 (but still> $0.5)^2$ The result is a bias in the estimation of $\theta^*,\text{thus explaining why the Population EM}_1$ estimate converges to some $\theta_\mathrm{wrong}\in(-\theta^*,0)$ when $w_1^*$ is not too large.
$^{2}$When $w_1^*$ is indeed very close to 1,then almost all of the probability mass of the true distribution comes from $C_1$, which has positive mean. So, in the M-step discussed above, the rightward pull of the mean of $\hat{C}_-$ may
4
|
[
[
0.1732177734375,
0.09027099609375,
0.3798828125,
0.091552734375,
0.3798828125,
0.1060791015625,
0.1732177734375,
0.10479736328125
],
[
0.17236328125,
0.11553955078125,
0.81396484375,
0.1142578125,
0.81396484375,
0.1319580078125,
0.17236328125,
0.1331787109375
],
[
0.1732177734375,
0.1331787109375,
0.82275390625,
0.1337890625,
0.82275390625,
0.1483154296875,
0.1732177734375,
0.147705078125
],
[
0.1748046875,
0.15087890625,
0.826171875,
0.15087890625,
0.826171875,
0.1654052734375,
0.1748046875,
0.1654052734375
],
[
0.1732177734375,
0.1654052734375,
0.57177734375,
0.1654052734375,
0.57177734375,
0.18310546875,
0.1732177734375,
0.18310546875
],
[
0.1748046875,
0.189453125,
0.81640625,
0.189453125,
0.81640625,
0.20263671875,
0.1748046875,
0.20263671875
],
[
0.1732177734375,
0.2052001953125,
0.82275390625,
0.205810546875,
0.82275390625,
0.2203369140625,
0.1732177734375,
0.2197265625
],
[
0.17236328125,
0.2183837890625,
0.826171875,
0.2183837890625,
0.826171875,
0.236083984375,
0.17236328125,
0.236083984375
],
[
0.1707763671875,
0.23291015625,
0.5849609375,
0.23486328125,
0.5849609375,
0.25244140625,
0.1707763671875,
0.250732421875
],
[
0.1748046875,
0.261474609375,
0.82421875,
0.261474609375,
0.82421875,
0.274658203125,
0.1748046875,
0.274658203125
],
[
0.1748046875,
0.274658203125,
0.82275390625,
0.274658203125,
0.82275390625,
0.2890625,
0.1748046875,
0.2890625
],
[
0.176513671875,
0.28857421875,
0.82421875,
0.28857421875,
0.82421875,
0.3017578125,
0.176513671875,
0.3017578125
],
[
0.176513671875,
0.3037109375,
0.82275390625,
0.3037109375,
0.82275390625,
0.314453125,
0.176513671875,
0.314453125
],
[
0.176513671875,
0.31640625,
0.82421875,
0.31640625,
0.82421875,
0.330078125,
0.176513671875,
0.330078125
],
[
0.1748046875,
0.330078125,
0.82421875,
0.330078125,
0.82421875,
0.343505859375,
0.1748046875,
0.343505859375
],
[
0.1748046875,
0.343505859375,
0.82275390625,
0.343505859375,
0.82275390625,
0.356689453125,
0.1748046875,
0.356689453125
],
[
0.1732177734375,
0.35595703125,
0.82275390625,
0.356689453125,
0.82275390625,
0.371826171875,
0.1732177734375,
0.37060546875
],
[
0.1748046875,
0.371826171875,
0.289306640625,
0.371826171875,
0.289306640625,
0.385009765625,
0.1748046875,
0.385009765625
],
[
0.1748046875,
0.391357421875,
0.82275390625,
0.391357421875,
0.82275390625,
0.406005859375,
0.1748046875,
0.406005859375
],
[
0.1748046875,
0.406005859375,
0.82275390625,
0.406005859375,
0.82275390625,
0.419189453125,
0.1748046875,
0.419189453125
],
[
0.1748046875,
0.42041015625,
0.40283203125,
0.42041015625,
0.40283203125,
0.434326171875,
0.1748046875,
0.434326171875
],
[
0.1732177734375,
0.436279296875,
0.82421875,
0.436279296875,
0.82421875,
0.453857421875,
0.1732177734375,
0.453857421875
],
[
0.1732177734375,
0.45263671875,
0.82666015625,
0.45263671875,
0.82666015625,
0.47021484375,
0.1732177734375,
0.47021484375
],
[
0.176513671875,
0.4716796875,
0.82421875,
0.4716796875,
0.82421875,
0.481689453125,
0.176513671875,
0.481689453125
],
[
0.1732177734375,
0.484130859375,
0.62841796875,
0.484130859375,
0.62841796875,
0.49755859375,
0.1732177734375,
0.49755859375
],
[
0.191162109375,
0.5087890625,
0.4453125,
0.5087890625,
0.4453125,
0.5234375,
0.191162109375,
0.5234375
],
[
0.25732421875,
0.53076171875,
0.76123046875,
0.53076171875,
0.76123046875,
0.5517578125,
0.25732421875,
0.5517578125
],
[
0.191162109375,
0.5673828125,
0.44677734375,
0.5673828125,
0.44677734375,
0.58544921875,
0.191162109375,
0.58544921875
],
[
0.256591796875,
0.5888671875,
0.76123046875,
0.59033203125,
0.76123046875,
0.6123046875,
0.256591796875,
0.611328125
],
[
0.1732177734375,
0.626953125,
0.82421875,
0.626953125,
0.82421875,
0.64453125,
0.1732177734375,
0.64453125
],
[
0.1748046875,
0.64306640625,
0.5595703125,
0.64453125,
0.5595703125,
0.65771484375,
0.1748046875,
0.6572265625
],
[
0.1748046875,
0.6640625,
0.82177734375,
0.6640625,
0.82177734375,
0.67822265625,
0.1748046875,
0.67822265625
],
[
0.1732177734375,
0.67822265625,
0.82177734375,
0.67822265625,
0.82177734375,
0.69140625,
0.1732177734375,
0.69140625
],
[
0.1732177734375,
0.69140625,
0.82177734375,
0.69140625,
0.82177734375,
0.705078125,
0.1732177734375,
0.705078125
],
[
0.1732177734375,
0.70849609375,
0.82177734375,
0.70703125,
0.82177734375,
0.7216796875,
0.1732177734375,
0.72265625
],
[
0.1748046875,
0.7236328125,
0.82177734375,
0.7236328125,
0.82177734375,
0.73779296875,
0.1748046875,
0.73779296875
],
[
0.1748046875,
0.7392578125,
0.826171875,
0.7392578125,
0.826171875,
0.75390625,
0.1748046875,
0.75390625
],
[
0.1732177734375,
0.755859375,
0.826171875,
0.7548828125,
0.826171875,
0.76953125,
0.1732177734375,
0.77099609375
],
[
0.1748046875,
0.76953125,
0.82421875,
0.76953125,
0.82421875,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1748046875,
0.7841796875,
0.82275390625,
0.7841796875,
0.82275390625,
0.79736328125,
0.1748046875,
0.79736328125
],
[
0.1732177734375,
0.80029296875,
0.82421875,
0.80029296875,
0.82421875,
0.814453125,
0.1732177734375,
0.814453125
],
[
0.1732177734375,
0.81640625,
0.82421875,
0.81640625,
0.82421875,
0.830078125,
0.1732177734375,
0.830078125
],
[
0.1732177734375,
0.83056640625,
0.826171875,
0.83203125,
0.826171875,
0.8466796875,
0.1732177734375,
0.84521484375
],
[
0.1748046875,
0.8466796875,
0.82275390625,
0.8466796875,
0.82275390625,
0.85986328125,
0.1748046875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.640625,
0.85986328125,
0.640625,
0.875,
0.1748046875,
0.875
],
[
0.193603515625,
0.8818359375,
0.82421875,
0.8837890625,
0.82421875,
0.8984375,
0.193603515625,
0.896484375
],
[
0.1732177734375,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1732177734375,
0.912109375
],
[
0.490234375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"2.2 Main theoretical results",
"Let us first consider the special case w = wz = 0.5. Then, it is straightforward to show that wst)",
"wit) = 0.5 for all t in Population EM2. Hence, Population EM2 is equivalent to Population EMi",
"Global convergence of (t) to * for this case was recently established by Xu et al. [2016, Theorem 1]",
"for almost all initial (0) (see also [Daskalakis et al., 2017).",
"We first show that the same global convergence may not hold for Population EMi when w* + w,.",
"Theorem 1. Consider Population EMi in dimension one (i.e., 0* E R). For any * > O, there",
"exists d > 0, such that given w e (0.5, 0.5 + §) and initialization 0(0) ≤ -0*, the Population EMi",
"estimate (t) converges to a fixed point Owrong inside (-0*, O).",
"This theorem, which is proved in Appendix A, implies that if we use random initialization,",
"Population EMi may converge to the wrong fixed point with constant probability. We illustrate this in",
"Figure 1. The iterates of Population EMi converge to a fixed point of the function - H(; 0*, w*)",
"definedin(5).Wehaveplottedthisfunctionforseveral differentvalues of w*intheleftpanel of",
"Figure 1. When w* is close to 1, H(0; 0*, w*) has only one fixed point and that is at = *. Hence,",
"in this case, the estimates produced by Population EMi converge to the true O*. However, when we",
"decrease the value of w* below a certain threshold (which is numerically found to be approximately",
"0.77 for * = 1), two other fixed points of H(0; 0*, w) emerge. These new fixed points are foils for",
"Population EMi.",
"From the failure of Population EMi, one may expect the over-parameterized Population EM2 to fail",
"as well. Yet, surprisingly, our second theorem proves the opposite is true: Population EM2 has global",
"convergence even when w + w2.",
"Theorem 2. For any wi E [0.5, 1), the Population EM2 estimate (ot, w(t) converges to either",
"(0*, wi) or (-0*, w2) with any initialization o(0) except on the hyperplane (0(0), o*) = 0. Further-",
"more,theconvergencespeedisgeometricaftersomefinitenumberofiterations,i.e.,thereexistsa",
"finite number T and constant p E (O, 1) such that the following hold.",
". If (o(0), o*) > 0, then for all t > T,",
"I10(++1) - 0*I? + [w++1) w/? ≤pt-T (10(T) - 0*I? + (w(T) -w)2)",
". If (o(0),0*) < 0, then for all t > T,",
"I10(++1) + 0* + [t+1) w2)2 ≤pt-T (I(T) + 0*2 + (T) 2)2)",
"Theorem 2 implies that if we use random initialization for o(o), with probability one, the",
"Population EM2 estimates converge to the true parameters.",
"The failure of Population EMi and success of Population EM2 can be explained intuitively. Let Ci",
"and C2,respectively,denote thetrue mixture components with parameters (w*,*)and (w,,-*)",
"Due to the symmetry in Population EMi, we are assured that among the two estimated mixture",
"components, one will have a positive mean, and the other will have a negative mean: call these C+",
"and C-, respectively. Assume 0* > 0 and w’ > 0.5, and consider initializing the Population EMi",
"with o(o) :=-*. This initialization incorrectly associates C_ with the larger weight w* instead of",
"the smaller weight w,. This causes, in the E-step of EM, the component C_ to become “responsible\"",
"for an overly large share of the overall probability mass, and in particular an overly large share of",
"the mass from Ci (which has a positive mean). Thus, in the M-step of EM, when the mean of the",
"estimated component C_ is updated, it is pulled rightward towards +oo. It is possible that this",
"rightward pull would cause the estimated mean of C_ to become positivein which case the roles of",
"C+ and C_ would switchbut this will not happen as long as w* is sufficiently bounded away from 1",
"(but still > 0.5).2 The result is a bias in the estimation of *, thus explaining why the Population EM",
"estimate converges to some Owrong E (-*, O) when w* is not too large.",
"2when wi is indeed very close to 1, then almost all of the probability mass of the true distribution comes",
"from Ci, which has positive mean. So, in the M-step discussed above, the rightward pull of the mean of C. may",
"4"
] |
{
"arxiv_id": "1810.11344",
"language": "en",
"page_idx": 3
}
|
||
1810.11344-en-4
|
1810.11344-en-4.jpg
|
Our discussion confırms that one way Population EM$_{1}$ may fail (in dimension one) is if it is initialized with $\theta^{\langle0\rangle}$ having the “incorrect”sign (e.g, $\theta^{\langle0\rangle}=-\theta^*).$ On the other hand, the performance of Population EM$_2$ does not depend on the sign of the initial $\theta^{\langle0\rangle}.$ Recall that $\hat{\text{the estimates of Population EM}}_2$ converge to the fixed points of the mapping $\mathcal{M}:(\boldsymbol{\theta},w_1)\mapsto$ $(G_{\theta}(\boldsymbol{\theta},w_{1};\boldsymbol{\theta}^{*},w_{1}^{*}),\overline{G}_{w}(\boldsymbol{\theta},w_{1};\boldsymbol{\theta}^{*},w_{1}^{*}))$, as defined in (6) and (7). One can check that for all ${\boldsymbol{\theta}},w_{1},\boldsymbol{\theta}^{*},w_{1}^{*}$,we have
(8)
$$G_{\theta}(\boldsymbol{\theta},w_{1};\boldsymbol{\theta}^{*},w_{1}^{*})+G_{\theta}(-\boldsymbol{\theta},w_{2};\boldsymbol{\theta}^{*},w_{1}^{*})=0,\\G_{w}(\boldsymbol{\theta},w_{1};\boldsymbol{\theta}^{*},w_{1}^{*})+G_{w}(-\boldsymbol{\theta},w_{2};\boldsymbol{\theta}^{*},w_{1}^{*})=1.$$
Hence, $(\boldsymbol{\theta},w_1)$ is a fixed point of $\mathcal{M}$ if and only if $(-\boldsymbol{\theta},w_2)$ is a fixed point of $\mathcal{M}$ as well. Therefore, Population EM$_2$ is insensitive to the sign of the initial $\theta^{\langle0\rangle}.$ This property can be extended to mixtures of $k>2$ Gaussians as well. In these cases, the performance of EM for Model 2 is insensitive to permutations of the component parameters. Hence, because of this nice property, as we will confirm $\hat{\text{in our simulations, when the mixture components are well-separated, EM for Model 2 performs well}}$ for most of the initializations, while EM for Model 1 fails in many cases.
One limitation of our permutation-free explanation is that the argument only holds when the weights $i$n Population EM$_2$ are initialized to be uniform. However, the benefits of over-parameterization are not limited to this case. Indeed, when we compare the landscapes of the log-likelihood objective for (the mixture models corresponding to) Population EM$_1$ and Population EM$_2$,we find that overparameterization eliminates spurious local maxima that were obstacles for Population EM$_1.$
Theorem 3. For all $w_1^*\neq0.5$, the log-likelihood objective optimized by Population EM{2} has only one saddle point $(\boldsymbol{\theta},\dot{w_1})=(\boldsymbol{0},1/2)$ and no local maximizers besides the two global maximizers $( \boldsymbol{\theta }, w_{1}) = ( \dot{\boldsymbol{\theta }} ^{* }, w_{1}^{* })$ and $( \boldsymbol{\theta }, w_{1}) = ( - \boldsymbol{\theta }^{* }, w_{2}^{* }) .$
The proof of this theorem is presented in Appendix C.
Remark 1. Consider the landscape of the log-likelihood objective for Population EM $_2$ and the point $( \theta _{\mathrm{wrong}}, w_{1}^{* }) , where$ $\theta _{\mathrm{wrong}}is$ the local maximizer suggested $by$ Theorem $1. Theorem$ 3 implies that we can still easily escape this point due to the non-zero gradient in the direction of w and thus $( \theta _{\mathrm{wrong}}, w_{1}^{* }) is\textit{not even a saddle point. We emphasize that this is exactly the mechanism that we have}$ hopedfor the purpose and benefit $of$ $over- parameterization.$
$\textbf{Remark 2. Note that although }( \boldsymbol{\theta }, w_{1}) = ( ( w_{1}^{* }- w_{2}^{* }) \boldsymbol{\theta }^{* }, 1)$ $or\left ( ( w_{2}^{* }- w_{1}^{* }) \boldsymbol{\theta }^{* }, 0\right )$ are the two fixed $pointsforPopulationEM_{2}$ $as$ $well, they$ are not the first order stationary points $of$ the $log- likelihood$ objective if $w_1^*\neq0.5.$
Finally, to complete the analysis of EM for the mixtures of two Gaussians, we present the following
result that applies to Sample-based EM$_2.$
Theorem 4. Let $(\hat{\boldsymbol{\theta}}^{\langle t\rangle},\hat{w}_{1}^{\langle t\rangle})$ be the estimates of Sample-based EM $_2.$ Suppose $\hat{\boldsymbol{\theta}}^{\langle0\rangle}=\boldsymbol{\theta}^{\langle0\rangle},\hat{w}_{1}^{\langle0\rangle}=$
$w_1^{\langle 0\rangle }= \frac 12$ and $\langle \boldsymbol{\theta }^{\langle 0\rangle }, \boldsymbol{\theta }^* \rangle \neq 0.$ Then $we$ have
$$\lim_{t\to\infty}\sup\|\hat{\boldsymbol{\theta}}^{\langle t\rangle}-\boldsymbol{\theta}^{\langle t\rangle}\|\:\to\:0\quad and\quad\lim_{t\to\infty}\sup|\hat{w}_{1}^{\langle t\rangle}-w_{1}^{\langle t\rangle}|\:\to\:0\quad as\:n\to\infty\:,$$
where convergence is in probability
The proof of this theorem uses the same approach as Xu et al.[2016] and is presented in Appendix D.
2.3 Roadmap of the proof for Theorem 2
Our first lemma, proved in Appendix B.1, confirms that if $\langle\boldsymbol{\theta}^{\langle0\rangle},\boldsymbol{\theta}^*\rangle>0$, then $\langle\boldsymbol{\theta}^{\langle t\rangle},\boldsymbol{\theta}^*\rangle>0$ for every $t$ and $w_1^{\langle t\rangle}\in(0.5,1).$ In other words, the estimates of the Population EM$_2$ remain in the correct hyperplane, and the weight moves in the right direction, too.
be so strong that the updated mean estimate becomes positive. Since the model enforces that the mean estimates of $\hat{C}_+$ and $\hat{C}_-$ be negations of each other, the roles of $\hat{C}_+$ and $\hat{C}_-$ switch, and now it is $\hat{C}_+$ that becomes associated with the larger mixing weight $w_1^*.$ In this case, owing to the symmetry assumption, Population EM$_1$ may be able to successfully converge to $\theta^*.$ We revisit this issue in the numerical study, where the symmetry assumption is removed.
5
|
[
[
0.176513671875,
0.0921630859375,
0.82421875,
0.0921630859375,
0.82421875,
0.1060791015625,
0.176513671875,
0.1060791015625
],
[
0.1748046875,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.121826171875,
0.1748046875,
0.121826171875
],
[
0.1732177734375,
0.12249755859375,
0.82421875,
0.121826171875,
0.82421875,
0.1363525390625,
0.1732177734375,
0.136962890625
],
[
0.176513671875,
0.136962890625,
0.82177734375,
0.136962890625,
0.82177734375,
0.15087890625,
0.176513671875,
0.15087890625
],
[
0.176513671875,
0.15087890625,
0.82421875,
0.15087890625,
0.82421875,
0.1654052734375,
0.176513671875,
0.1654052734375
],
[
0.17236328125,
0.1634521484375,
0.325927734375,
0.1641845703125,
0.325927734375,
0.1787109375,
0.17236328125,
0.177978515625
],
[
0.348876953125,
0.186279296875,
0.65869140625,
0.186279296875,
0.65869140625,
0.2008056640625,
0.348876953125,
0.2008056640625
],
[
0.80224609375,
0.1962890625,
0.826171875,
0.1962890625,
0.826171875,
0.2108154296875,
0.80224609375,
0.2108154296875
],
[
0.343017578125,
0.205810546875,
0.65673828125,
0.205810546875,
0.65673828125,
0.2197265625,
0.343017578125,
0.2197265625
],
[
0.1748046875,
0.2291259765625,
0.826171875,
0.2291259765625,
0.826171875,
0.242431640625,
0.1748046875,
0.242431640625
],
[
0.17236328125,
0.2412109375,
0.826171875,
0.242431640625,
0.826171875,
0.260009765625,
0.17236328125,
0.2587890625
],
[
0.1748046875,
0.2587890625,
0.82275390625,
0.2587890625,
0.82275390625,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.176513671875,
0.272705078125,
0.82177734375,
0.272705078125,
0.82177734375,
0.285888671875,
0.176513671875,
0.285888671875
],
[
0.1732177734375,
0.285400390625,
0.82275390625,
0.285400390625,
0.82275390625,
0.298583984375,
0.1732177734375,
0.298583984375
],
[
0.1732177734375,
0.2998046875,
0.65283203125,
0.2998046875,
0.65283203125,
0.313232421875,
0.1732177734375,
0.313232421875
],
[
0.1748046875,
0.32080078125,
0.82275390625,
0.32080078125,
0.82275390625,
0.333984375,
0.1748046875,
0.333984375
],
[
0.1748046875,
0.335205078125,
0.82275390625,
0.335205078125,
0.82275390625,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.1748046875,
0.348388671875,
0.82275390625,
0.348388671875,
0.82275390625,
0.3623046875,
0.1748046875,
0.3623046875
],
[
0.1748046875,
0.3623046875,
0.82421875,
0.3623046875,
0.82421875,
0.375732421875,
0.1748046875,
0.375732421875
],
[
0.1748046875,
0.375732421875,
0.77294921875,
0.375732421875,
0.77294921875,
0.39013671875,
0.1748046875,
0.39013671875
],
[
0.1748046875,
0.39453125,
0.82275390625,
0.39453125,
0.82275390625,
0.40771484375,
0.1748046875,
0.40771484375
],
[
0.176513671875,
0.4091796875,
0.82275390625,
0.4091796875,
0.82275390625,
0.42236328125,
0.176513671875,
0.42236328125
],
[
0.1748046875,
0.421630859375,
0.476318359375,
0.421630859375,
0.476318359375,
0.436279296875,
0.1748046875,
0.436279296875
],
[
0.1748046875,
0.447509765625,
0.52685546875,
0.447509765625,
0.52685546875,
0.4609375,
0.1748046875,
0.4609375
],
[
0.176513671875,
0.466552734375,
0.82275390625,
0.466552734375,
0.82275390625,
0.479736328125,
0.176513671875,
0.479736328125
],
[
0.17236328125,
0.479248046875,
0.82275390625,
0.479248046875,
0.82275390625,
0.496826171875,
0.17236328125,
0.496826171875
],
[
0.17724609375,
0.49560546875,
0.82177734375,
0.49560546875,
0.82177734375,
0.505859375,
0.17724609375,
0.505859375
],
[
0.1748046875,
0.50830078125,
0.82275390625,
0.50830078125,
0.82275390625,
0.52294921875,
0.1748046875,
0.52294921875
],
[
0.1748046875,
0.521484375,
0.56884765625,
0.521484375,
0.56884765625,
0.53515625,
0.1748046875,
0.53515625
],
[
0.1748046875,
0.54052734375,
0.82421875,
0.54052734375,
0.82421875,
0.5537109375,
0.1748046875,
0.5537109375
],
[
0.1748046875,
0.5537109375,
0.82421875,
0.5537109375,
0.82421875,
0.5673828125,
0.1748046875,
0.5673828125
],
[
0.1732177734375,
0.5673828125,
0.318603515625,
0.5673828125,
0.318603515625,
0.58203125,
0.1732177734375,
0.58203125
],
[
0.1748046875,
0.59228515625,
0.82275390625,
0.59228515625,
0.82275390625,
0.60693359375,
0.1748046875,
0.60693359375
],
[
0.1732177734375,
0.60693359375,
0.443603515625,
0.60693359375,
0.443603515625,
0.6220703125,
0.1732177734375,
0.6220703125
],
[
0.1690673828125,
0.62255859375,
0.7255859375,
0.62353515625,
0.7255859375,
0.6484375,
0.1690673828125,
0.64794921875
],
[
0.7158203125,
0.626953125,
0.80810546875,
0.62890625,
0.80810546875,
0.64306640625,
0.7158203125,
0.6416015625
],
[
0.1732177734375,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.6640625,
0.1732177734375,
0.6640625
],
[
0.2418212890625,
0.6787109375,
0.32177734375,
0.6787109375,
0.32177734375,
0.693359375,
0.2418212890625,
0.693359375
],
[
0.318603515625,
0.6748046875,
0.75732421875,
0.6767578125,
0.75732421875,
0.69580078125,
0.318603515625,
0.693359375
],
[
0.1732177734375,
0.70703125,
0.408447265625,
0.70703125,
0.408447265625,
0.7216796875,
0.1732177734375,
0.7216796875
],
[
0.1732177734375,
0.73291015625,
0.82177734375,
0.73291015625,
0.82177734375,
0.7470703125,
0.1732177734375,
0.7470703125
],
[
0.17236328125,
0.76318359375,
0.4697265625,
0.76318359375,
0.4697265625,
0.77783203125,
0.17236328125,
0.77783203125
],
[
0.1732177734375,
0.79150390625,
0.82177734375,
0.79052734375,
0.82177734375,
0.8046875,
0.1732177734375,
0.80615234375
],
[
0.1748046875,
0.8095703125,
0.82275390625,
0.8095703125,
0.82275390625,
0.82373046875,
0.1748046875,
0.82373046875
],
[
0.1748046875,
0.82373046875,
0.56884765625,
0.82373046875,
0.56884765625,
0.83837890625,
0.1748046875,
0.83837890625
],
[
0.1748046875,
0.84521484375,
0.82421875,
0.84521484375,
0.82421875,
0.85986328125,
0.1748046875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.82275390625,
0.85986328125,
0.82275390625,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.1748046875,
0.87353515625,
0.82275390625,
0.87353515625,
0.82275390625,
0.88720703125,
0.1748046875,
0.88720703125
],
[
0.1748046875,
0.88818359375,
0.82177734375,
0.88818359375,
0.82177734375,
0.8984375,
0.1748046875,
0.8984375
],
[
0.1748046875,
0.90087890625,
0.313720703125,
0.90087890625,
0.313720703125,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94873046875,
0.494384765625,
0.94873046875
]
] |
[
"Our discussion confirms that one way Population EMi may fail (in dimension one) is if it is",
"initialized with o(o) having the “incorrect\"sign (e.g., o(o) = -0*). On the other hand, the",
"performance of Population EM2 does not depend on the sign of the initial o(o).Recall that",
"the estimates of Population EM2 converge to the fixed points of the mapping M : (0, wi) →",
"(Ge(0, wr;0*, w),Gw(0, wi;0*, w)), as defined in (6) and (7). One can check that for all",
"O, wi, o*, w', we have",
"Ge(0, W1;0*, w) +Ge(-0,W2;0*, w*) = 0,",
"(8)",
"Gw(0,W1; 0*,w*) +Gw(-0, w2;0*,w*) = 1.",
"Hence, (0, wi) is a fixed point of M if and only if (-0, w2) is a fixed point of M as well. Therefore,",
"Population EM2 is insensitive to the sign of the initial (o). This property can be extended to mixtures",
"of k > 2 Gaussians as well. In these cases, the performance of EM for Model 2 is insensitive to",
"permutations of the component parameters. Hence, because of this nice property, as we will confirm",
"in our simulations, when the mixture components are well-separated, EM for Model 2 performs well",
"for most of the initializations, while EM for Model 1 fails in many cases.",
"One limitation of our permutation-free explanation is that the argument only holds when the weights",
"in Population EM2 are initialized to be uniform. However, the benefits of over-parameterization are",
"not limited to this case. Indeed, when we compare the landscapes of the log-likelihood objective",
"for (the mixture models corresponding to) Population EMi and Population EM2, we find that over-",
"parameterization eliminates spurious local maxima that were obstacles for Population EMi.",
"Theorem 3. For all w* 0.5, the log-likelihood objective optimized by Population EM2 has only",
"one saddle point (0, w1) = (, 1/2) and no local maximizers besides the two global maximizers",
"(0,w1) = (0*,w)and (0, w1) = (-0*,w)",
"The proof of this theorem is presented in Appendix C",
"Remark 1.Consider the landscape of the log-likelihood objective for Population EM2 and the point",
"(wrong, wi), where Qwrong is the local maximizer suggested by Theorem 1. Theorem 3 implies that",
"we can still easily escape this point due tothe non-zero gradientin the directionof wr andthus",
"(Owrong, w) is not even a saddle point. We emphasize that this is exactly the mechanism that we have",
"hoped for the purpose and benefit of over-parameterization.",
"Remark 2. Note that although (0, wi) = ((w* -w)o*, 1) or ((w, - w*)o*, O) are the two fixed",
"points for Population EM2 as well, they are not the first order stationary points of the log-likelihood",
"objective if w + 0.5.",
"Finally, to complete the analysis of EM for the mixtures of two Gaussians, we present the following",
"result that applies to Sample-based EM2.",
"Theorem 4. Let (e\", t\") be the estimates of Sample-based EM2. Suppose (0)",
"= 0(0) W(0)",
"wi0) =号 and (o(0), 0*) 0. Then we have",
"lim sup [o",
"t)-(t)l→0 and limsup|t-wt}→0 asn→,",
"where convergence is in probability",
"The proof of this theorem uses the same approach as Xu et al. [2016] and is presented in Appendix D",
"2.3Roadmap of the proof for Theorem 2",
"Our first lemma, proved in Appendix B.1, confirms that if (o(0), 0*) > 0, then (o(t), 0*) > 0 fo",
"every t and w(t) e (0.5, 1). In other words, the estimates of the Population EM2 remain in the correct",
"hyperplane, and the weight moves in the right direction, too.",
"be so strong that the updated mean estimate becomes positive. Since the model enforces that the mean estimates",
"of C+ and C_ be negations of each other, the roles of C+ and C- switch, and now it is C+ that becomes",
"associated with the larger mixing weight wi. In this case, owing to the symmetry assumption, Population EMi",
"may be ableto successfully convergeto *We revisit this issuein the numerical study,where the symmetry",
"assumptionisremoved.",
"5"
] |
{
"arxiv_id": "1810.11344",
"language": "en",
"page_idx": 4
}
|
||
2007.02454-en-0
|
2007.02454-en-0.jpg
|
# Self-Challenging Improves Cross-Domain Generalization
Zeyi Huang$^\star$, Haohan Wang$^\star$, Eric P. Xing, and Dong Huang
School of Computer Science, Carnegie Mellon University
$\{$zeyih@andrew, haohanw@cs, epxing@cs, donghuang}.cmu.edu
$\mathbf{Abstract.~Convolutional~Neural~Networks~(CNN)~conduct~image~clas-}$ sification by activating dominant features that correlated with labels. When the training and testing data are under similar distributions, their dominant features are similar, leading to decent test performance. The performance is nonetheless unmet when tested with different distributions, leading to the challenges in cross-domain image classification. We introduce a simple training heuristic, Representation Self-Challenging (RSC),that significantly improves the generalization of CNN to the outof-domain data. RSC iteratively challenges (discards) the dominant features activated on the training data, and forces the network to activate remaining features that correlates with labels. This process appears to activate feature representations applicable to out-of-domain data without prior knowledge of new domain and without learning extra network parameters. We present theoretical properties and conditions of RSC for improving cross-domain generalization. The experiments endorse the simple, effective, and architecture-agnostic nature of our RSC method. Keywords: cross-domain generalization, robustness
## 1 Introduction
Imagine teaching a child to visually differentiate “dog” from“cat": when presented with a collection of illustrations from her picture books, she may immediately answer that“cats tend to have chubby faces” and end the learning. However, if we continue to ask for more differences, she may start to notice other features like ears or body-size. We conjecture this follow-up challenge question plays a significant role in helping human reach the remarkable generalization ability. Most people should be able to differentiate “cat” from“dog”visually even when the images are presented in irregular qualities. After all, we did not stop learning after we picked up the first clue when we were children, even the first clue was good enough to help us recognize all the images in our textbook.
Nowadays, deep neural networks have exhibited remarkable empirical results over various computer vision tasks, yet these impressive performances seem unmet when the models are tested with the samples in irregular qualities $(i.e.$,
$^{\star}$ equal contribution
|
[
[
0.1473388671875,
0.052001953125,
0.8466796875,
0.05279541015625,
0.8466796875,
0.07470703125,
0.1473388671875,
0.07391357421875
],
[
0.0338134765625,
0.07232666015625,
0.0821533203125,
0.07232666015625,
0.08575439453125,
0.223388671875,
0.037445068359375,
0.223388671875
],
[
0.37451171875,
0.08203125,
0.62451171875,
0.0828857421875,
0.62451171875,
0.10400390625,
0.37451171875,
0.10235595703125
],
[
0.1739501953125,
0.141357421875,
0.82470703125,
0.1429443359375,
0.82470703125,
0.160888671875,
0.1739501953125,
0.1591796875
],
[
0.2210693359375,
0.1754150390625,
0.775390625,
0.17626953125,
0.775390625,
0.194091796875,
0.2210693359375,
0.193359375
],
[
0.182373046875,
0.1973876953125,
0.81640625,
0.1973876953125,
0.81640625,
0.2120361328125,
0.182373046875,
0.2120361328125
],
[
0.0338134765625,
0.223388671875,
0.08331298828125,
0.2225341796875,
0.088134765625,
0.329833984375,
0.03863525390625,
0.33056640625
],
[
0.059173583984375,
0.216064453125,
0.07244873046875,
0.216064453125,
0.07244873046875,
0.2298583984375,
0.059173583984375,
0.2298583984375
],
[
0.1522216796875,
0.256591796875,
0.8466796875,
0.256591796875,
0.8466796875,
0.271240234375,
0.1522216796875,
0.271240234375
],
[
0.1497802734375,
0.275390625,
0.84765625,
0.275390625,
0.84765625,
0.2900390625,
0.1497802734375,
0.2900390625
],
[
0.1510009765625,
0.293212890625,
0.84765625,
0.293212890625,
0.84765625,
0.307861328125,
0.1510009765625,
0.307861328125
],
[
0.1497802734375,
0.31201171875,
0.84912109375,
0.31201171875,
0.84912109375,
0.32568359375,
0.1497802734375,
0.32568359375
],
[
0.05194091796875,
0.337158203125,
0.07244873046875,
0.337158203125,
0.07244873046875,
0.351806640625,
0.05194091796875,
0.351806640625
],
[
0.1473388671875,
0.329833984375,
0.8466796875,
0.3291015625,
0.8466796875,
0.343505859375,
0.1473388671875,
0.344482421875
],
[
0.0338134765625,
0.341064453125,
0.0821533203125,
0.341064453125,
0.088134765625,
0.62158203125,
0.03985595703125,
0.62158203125
],
[
0.1497802734375,
0.34765625,
0.84912109375,
0.34765625,
0.84912109375,
0.360595703125,
0.1497802734375,
0.360595703125
],
[
0.1497802734375,
0.364013671875,
0.84912109375,
0.364013671875,
0.84912109375,
0.378662109375,
0.1497802734375,
0.378662109375
],
[
0.1510009765625,
0.3818359375,
0.8466796875,
0.3818359375,
0.8466796875,
0.396484375,
0.1510009765625,
0.396484375
],
[
0.1497802734375,
0.399658203125,
0.8466796875,
0.399658203125,
0.8466796875,
0.414306640625,
0.1497802734375,
0.414306640625
],
[
0.1497802734375,
0.41748046875,
0.8466796875,
0.41748046875,
0.8466796875,
0.43212890625,
0.1497802734375,
0.43212890625
],
[
0.1473388671875,
0.4345703125,
0.84912109375,
0.436279296875,
0.84912109375,
0.450927734375,
0.1473388671875,
0.44921875
],
[
0.1497802734375,
0.4541015625,
0.8466796875,
0.4541015625,
0.8466796875,
0.467041015625,
0.1497802734375,
0.467041015625
],
[
0.1510009765625,
0.47119140625,
0.8515625,
0.47119140625,
0.8515625,
0.48583984375,
0.1510009765625,
0.48583984375
],
[
0.1497802734375,
0.489013671875,
0.84765625,
0.489013671875,
0.84765625,
0.50341796875,
0.1497802734375,
0.50341796875
],
[
0.1497802734375,
0.5078125,
0.84765625,
0.5078125,
0.84765625,
0.521484375,
0.1497802734375,
0.521484375
],
[
0.1497802734375,
0.525390625,
0.8408203125,
0.525390625,
0.8408203125,
0.5400390625,
0.1497802734375,
0.5400390625
],
[
0.1473388671875,
0.55712890625,
0.66650390625,
0.55810546875,
0.66650390625,
0.576171875,
0.1473388671875,
0.57421875
],
[
0.0821533203125,
0.61181640625,
0.09417724609375,
0.61181640625,
0.09417724609375,
0.6220703125,
0.0821533203125,
0.6220703125
],
[
0.1292724609375,
0.609375,
0.310302734375,
0.609375,
0.310302734375,
0.62646484375,
0.1292724609375,
0.62646484375
],
[
0.11834716796875,
0.6669921875,
0.91552734375,
0.6669921875,
0.91552734375,
0.681640625,
0.11834716796875,
0.681640625
],
[
0.0821533203125,
0.68798828125,
0.9130859375,
0.68798828125,
0.9130859375,
0.70263671875,
0.0821533203125,
0.70263671875
],
[
0.0784912109375,
0.705078125,
0.91552734375,
0.7060546875,
0.91552734375,
0.7236328125,
0.0784912109375,
0.72216796875
],
[
0.0821533203125,
0.72705078125,
0.91650390625,
0.72705078125,
0.91650390625,
0.74072265625,
0.0821533203125,
0.74072265625
],
[
0.0809326171875,
0.74658203125,
0.91650390625,
0.74658203125,
0.91650390625,
0.76123046875,
0.0809326171875,
0.76123046875
],
[
0.0809326171875,
0.76513671875,
0.91650390625,
0.76513671875,
0.91650390625,
0.77978515625,
0.0809326171875,
0.77978515625
],
[
0.0821533203125,
0.78466796875,
0.91552734375,
0.78466796875,
0.91552734375,
0.79931640625,
0.0821533203125,
0.79931640625
],
[
0.0809326171875,
0.80517578125,
0.91650390625,
0.80517578125,
0.91650390625,
0.81982421875,
0.0809326171875,
0.81982421875
],
[
0.0772705078125,
0.82275390625,
0.91650390625,
0.8212890625,
0.91650390625,
0.83935546875,
0.0772705078125,
0.83984375
],
[
0.0809326171875,
0.84326171875,
0.908203125,
0.84326171875,
0.908203125,
0.85791015625,
0.0809326171875,
0.85791015625
],
[
0.11474609375,
0.861328125,
0.92041015625,
0.861328125,
0.92041015625,
0.87890625,
0.11474609375,
0.87890625
],
[
0.0809326171875,
0.8828125,
0.91552734375,
0.8828125,
0.91552734375,
0.8974609375,
0.0809326171875,
0.8974609375
],
[
0.0772705078125,
0.8984375,
0.91650390625,
0.90087890625,
0.91650390625,
0.91796875,
0.0772705078125,
0.91650390625
],
[
0.1051025390625,
0.931640625,
0.302001953125,
0.931640625,
0.302001953125,
0.9462890625,
0.1051025390625,
0.9462890625
]
] |
[
"Self-Challenging Improves Cross-Domain",
"",
"Generalization",
"Zeyi Huang*,Haohan Wang*, Eric P.Xing, and Dong Huang",
"School of Computer Science, Carnegie Mellon University",
"[zeyih@andrew,haohanw@cs,epxing@cs,donghuang].cmu.edu",
"[A'S]",
"",
"Abstract.Convolutional Neural Networks (CNN)conduct image clas-",
"sification by activating dominant features that correlated with labels.",
"When the training and testing data are under similar distributions, their",
"dominant features are similar, leading to decent test performance. The",
"",
"performance is nonetheless unmet when tested with different distribu-",
"",
"tions,leading to thechallenges in cross-domain imageclassification.We",
"introduce a simple training heuristic, Representation Self-Challenging",
"(RSC), that significantly improves the generalization of CNN to the out-",
"of-domain data.RSC iteratively challenges (discards)the dominant fea-",
"tures activated on the training data, and forces the network to activate",
"remaining features that correlates with labels.This process appears to",
"activate feature representations applicable to out-of-domain data with-",
"out prior knowledge of new domain and withoutlearning extra network",
"parameters. We present theoretical properties and conditions of RsC",
"forimproving cross-domain generalization.Theexperiments endorse the",
"simple, effective, and architecture-agnostic nature of our RSC method.",
"Keywords: cross-domain generalization, robustness",
"1",
"Introduction",
"Imagine teaching a child to visually differentiate “dog\" from “\"cat\":when",
"presented with a collection of illustrations from her picture books, she may",
"immediately answer that“cats tend to have chubbyfaces\"and end the learning.",
"However,if we continue to ask for more differences, she may start tonotice other",
"features like ears or body-size. We conjecture this follow-up challenge question",
"plays a significant role in helping human reach the remarkable generalization",
"ability. Most people should be able to differentiate“cat\" from“dog\"visually",
"even when the images are presented in irregular qualities.After all, we did not",
"stop learning after we picked up the first clue when we were children, even the",
"first clue was good enough to help us recognize all the images in our textbook",
"Nowadays, deep neural networks have exhibited remarkable empirical results",
"over various computer vision tasks, yet these impressive performances seem un-",
"met when the models are tested with the samples in irregular qualities (i.e.,",
"*equalcontribution"
] |
{
"arxiv_id": "2007.02454",
"language": "en",
"page_idx": 0
}
|
||
2007.02454-en-1
|
2007.02454-en-1.jpg
|
2
${\mathrm{Huang~et~al.}}$
<FigureHere>
Fig.1. The essence of our Representation Self-Challenging (RSC) training method: top two panels: the algorithm mutes the feature representations associated with the highest gradient, such that the network is forced to predict the labels through other features; bottom panel: after training, the model is expected to leverage more features for prediction in comparison to models trained convetionally.
out-of-domain data, samples collected from the distributions that are similar to, but different from the distributions of the training samples). To account for this discrepancy, technologies have been invented under the domain adaptation regime [2,3], where the goal is to train a model invariant to the distributional differences between the source domain ( i. e. , the distribution of the training sam- ples) and the target domain $(i.e.$, the distribution of the testing samples) [5,32].
As the influence of machine learning increases, the industry starts to demand the models that can be applied to the domains that are not seen during the training phase. Domain generalization [18],as an extension of domain adaptation, has been studied as a response. The central goal is to train a model that can align the signals from multiple source domains.
Further, Wang $et\textit{al. extend the problem to ask how to train a model that}$ generalizes to an arbitrary domain with only the training samples, but not the corresponding domain information, as these domain information may not be available in the real world [31]. Our paper builds upon this set-up and aims to offer a solution that allows the model to be robustly trained without domain information and to empirically perform well on unseen domains.
In this paper, we introduce a simple training heuristic that improves crossdomain generalization. This approach discards the representations associated with the higher gradients at each epoch, and forces the model to predict with remaining information. Intuitively, in a image classification problem, our heuristic works like a“self-challenging”mechanism as it prevents the fully-connected layers to predict with the most predictive subsets of features, such as the most frequent color, edges, or shapes in the training data. We name our method Representation Self Challenging (RSC) and illustrate its main idea in Figure 1.
|
[
[
0.0784912109375,
0.01300048828125,
0.1026611328125,
0.0162506103515625,
0.09661865234375,
0.03411865234375,
0.07244873046875,
0.0308685302734375
],
[
0.15576171875,
0.0138092041015625,
0.283935546875,
0.0162506103515625,
0.283935546875,
0.03411865234375,
0.15576171875,
0.03167724609375
],
[
0.48193359375,
0.0780029296875,
0.5107421875,
0.0780029296875,
0.5107421875,
0.09014892578125,
0.48193359375,
0.09014892578125
],
[
0.6728515625,
0.07958984375,
0.7197265625,
0.08612060546875,
0.7138671875,
0.10723876953125,
0.66650390625,
0.1007080078125
],
[
0.47705078125,
0.08935546875,
0.5009765625,
0.08935546875,
0.5009765625,
0.0966796875,
0.47705078125,
0.0966796875
],
[
0.5712890625,
0.10882568359375,
0.6015625,
0.10882568359375,
0.6015625,
0.11700439453125,
0.5712890625,
0.11700439453125
],
[
0.51806640625,
0.1226806640625,
0.6025390625,
0.1226806640625,
0.6025390625,
0.1331787109375,
0.51806640625,
0.1331787109375
],
[
0.572265625,
0.1697998046875,
0.59765625,
0.17138671875,
0.5966796875,
0.1795654296875,
0.5712890625,
0.1778564453125
],
[
0.57861328125,
0.1851806640625,
0.5927734375,
0.1851806640625,
0.5927734375,
0.19091796875,
0.57861328125,
0.19091796875
],
[
0.56982421875,
0.1982421875,
0.6015625,
0.1982421875,
0.6015625,
0.2071533203125,
0.56982421875,
0.2071533203125
],
[
0.5400390625,
0.2120361328125,
0.6025390625,
0.2120361328125,
0.6025390625,
0.2225341796875,
0.5400390625,
0.2225341796875
],
[
0.496337890625,
0.24853515625,
0.50732421875,
0.245361328125,
0.5107421875,
0.2509765625,
0.5,
0.25341796875
],
[
0.576171875,
0.2445068359375,
0.59423828125,
0.247802734375,
0.591796875,
0.255859375,
0.572265625,
0.252685546875
],
[
0.488037109375,
0.256591796875,
0.5107421875,
0.256591796875,
0.5107421875,
0.269775390625,
0.488037109375,
0.269775390625
],
[
0.56982421875,
0.25830078125,
0.59423828125,
0.25830078125,
0.59423828125,
0.2705078125,
0.56982421875,
0.2705078125
],
[
0.56982421875,
0.272216796875,
0.6025390625,
0.272216796875,
0.6025390625,
0.284423828125,
0.56982421875,
0.284423828125
],
[
0.56982421875,
0.284423828125,
0.6015625,
0.286865234375,
0.59912109375,
0.29736328125,
0.56787109375,
0.295654296875
],
[
0.50830078125,
0.30126953125,
0.6025390625,
0.30126953125,
0.6025390625,
0.31201171875,
0.50830078125,
0.31201171875
],
[
0.0821533203125,
0.339599609375,
0.91650390625,
0.339599609375,
0.91650390625,
0.353271484375,
0.0821533203125,
0.353271484375
],
[
0.0809326171875,
0.358154296875,
0.91650390625,
0.358154296875,
0.91650390625,
0.371337890625,
0.0809326171875,
0.371337890625
],
[
0.0784912109375,
0.37451171875,
0.91650390625,
0.373779296875,
0.91650390625,
0.38818359375,
0.0784912109375,
0.38916015625
],
[
0.0809326171875,
0.392333984375,
0.91796875,
0.392333984375,
0.91796875,
0.406982421875,
0.0809326171875,
0.406982421875
],
[
0.0821533203125,
0.41015625,
0.67626953125,
0.41015625,
0.67626953125,
0.4248046875,
0.0821533203125,
0.4248046875
],
[
0.0821533203125,
0.45166015625,
0.91650390625,
0.45166015625,
0.91650390625,
0.46630859375,
0.0821533203125,
0.46630859375
],
[
0.0809326171875,
0.47119140625,
0.91552734375,
0.47119140625,
0.91552734375,
0.48583984375,
0.0809326171875,
0.48583984375
],
[
0.0821533203125,
0.491455078125,
0.91552734375,
0.491455078125,
0.91552734375,
0.505859375,
0.0821533203125,
0.505859375
],
[
0.0809326171875,
0.50927734375,
0.91796875,
0.50927734375,
0.91796875,
0.5263671875,
0.0809326171875,
0.5263671875
],
[
0.0784912109375,
0.5263671875,
0.91552734375,
0.52783203125,
0.91552734375,
0.544921875,
0.0784912109375,
0.54443359375
],
[
0.0809326171875,
0.5498046875,
0.91162109375,
0.54833984375,
0.91162109375,
0.56298828125,
0.0809326171875,
0.564453125
],
[
0.11834716796875,
0.57275390625,
0.91650390625,
0.57275390625,
0.91650390625,
0.58740234375,
0.11834716796875,
0.58740234375
],
[
0.0821533203125,
0.59228515625,
0.91552734375,
0.59228515625,
0.91552734375,
0.60693359375,
0.0821533203125,
0.60693359375
],
[
0.0821533203125,
0.61181640625,
0.91650390625,
0.61181640625,
0.91650390625,
0.62646484375,
0.0821533203125,
0.62646484375
],
[
0.0809326171875,
0.63134765625,
0.91552734375,
0.63134765625,
0.91552734375,
0.64599609375,
0.0809326171875,
0.64599609375
],
[
0.0821533203125,
0.64990234375,
0.51953125,
0.64990234375,
0.51953125,
0.66455078125,
0.0821533203125,
0.66455078125
],
[
0.11834716796875,
0.67431640625,
0.91552734375,
0.67431640625,
0.91552734375,
0.68896484375,
0.11834716796875,
0.68896484375
],
[
0.0772705078125,
0.69384765625,
0.91650390625,
0.69140625,
0.91650390625,
0.708984375,
0.0772705078125,
0.7109375
],
[
0.0784912109375,
0.71240234375,
0.91796875,
0.71240234375,
0.91796875,
0.7294921875,
0.0784912109375,
0.7294921875
],
[
0.0809326171875,
0.73291015625,
0.91650390625,
0.73291015625,
0.91650390625,
0.74755859375,
0.0809326171875,
0.74755859375
],
[
0.0809326171875,
0.75048828125,
0.91552734375,
0.75146484375,
0.91552734375,
0.76611328125,
0.0809326171875,
0.76513671875
],
[
0.0809326171875,
0.77197265625,
0.75830078125,
0.77197265625,
0.75830078125,
0.7861328125,
0.0809326171875,
0.7861328125
],
[
0.11474609375,
0.79443359375,
0.91552734375,
0.79443359375,
0.91552734375,
0.8125,
0.11474609375,
0.8125
],
[
0.0809326171875,
0.8154296875,
0.91650390625,
0.8154296875,
0.91650390625,
0.82958984375,
0.0809326171875,
0.82958984375
],
[
0.0821533203125,
0.8349609375,
0.91552734375,
0.8349609375,
0.91552734375,
0.84912109375,
0.0821533203125,
0.84912109375
],
[
0.0821533203125,
0.8544921875,
0.9130859375,
0.8544921875,
0.9130859375,
0.869140625,
0.0821533203125,
0.869140625
],
[
0.0809326171875,
0.87255859375,
0.91650390625,
0.87255859375,
0.91650390625,
0.88720703125,
0.0809326171875,
0.88720703125
],
[
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
],
[
0.0772705078125,
0.9287109375,
0.87451171875,
0.93017578125,
0.87451171875,
0.947265625,
0.0772705078125,
0.9462890625
]
] |
[
"2",
"Huang et al.",
"",
"O",
"福",
"aLe",
"ionandrdien",
"ALO",
"TG",
"CLC",
"Hnderadien",
"",
"O",
"",
"EUE",
"aLes",
"aL/e",
"indonnderadien",
"Fig.1.The essence of our Representation Self-Challenging (RSC)training method:",
"top two panels: the algorithm mutes the feature representations associated with the",
"highest gradient, such that the network is forced to predict the labels through other",
"features; bottom panel:aftertraining, the model is expected to leveragemorefeatures",
"forprediction in comparisontomodels trained convetionally.",
"out-of-domain data, samples collected from the distributions that are similar",
"to, but different from the distributions of thetraining samples). To account for",
"this discrepancy,technologies have been invented under the domain adaptation",
"regime [2,3], where the goal is to train a model invariant to the distributional",
"differences between the source domain (i.e.,the distribution of the training sam-",
"ples)and the target domain (i.e.,thedistributionof thetesting samples)[5.32]",
"As the influenceof machinelearning increases,theindustrystarts todemand",
"the models that can beapplied tothe domains that are not seen during the train-",
"ing phase.Domain generalization [i8l,as an extension of domain adaptation,has",
"been studied as a response. The central goal is to train a model that can align",
"the signals from multiple source domains.",
"Further, Wang et al. extend the problem to ask how to train a model that",
"generalizes to an arbitrary domain with only the training samples, but not the",
"corresponding domain information, as these domain information may not be",
"available in the real world [31l.Our paper builds upon this set-up and aims to",
"offer a solution that allows the model to be robustly trained without domain",
"information and to empirically perform well on unseen domains.",
"In this paper, we introduce a simple training heuristic that improves cross-",
"domain generalization. This approach discards the representations associated",
"with the higher gradients at each epoch, and forces the model to predict with",
"remaining information.Intuitively,in a image classification problem, our heuris",
"tic works like a“self-challenging\"mechanism as it prevents the fully-connected",
"layers topredict withthe most predictive subsets of features,such as the most",
"frequent color,edges, or shapes inthe training data.We name our method Rep",
"resentation Self Challenging (RSC) and illustrate its main idea in Figure 1."
] |
{
"arxiv_id": "2007.02454",
"language": "en",
"page_idx": 1
}
|
||
2007.02454-en-2
|
2007.02454-en-2.jpg
|
3
Self-Challenging Improves Cross-Domain Generalization
We present mathematical analysis that RSC induces a smaller generalization bound. We further demonstrate the empirical strength of our method with domain-agnostic cross-domain evaluations, following previous setup [31]. We also conduct ablation study to examine the alignment between its empirical performance and our intuitive understanding. The inspections also shed light upon the choices of its extra hyperparameter.
## 2 Related Work
We summarize the related DG works from two perspectives: learning domain invariant features and augmenting source domain data. Further, as RSC can be broadly viewed as a generic training heuristic for CNN, we also briefly discuss the general-purpose regularizations that appear similar to our method.
DG through Learning Domain Invariant Features: These methods typically minimize the discrepancy between source domains assuming that the resulting features will be domain-invariant and generalize well for unseen target distributions. Along this track, Muandet $et\textit{al. employed Maximum Mean}$ Discrepancy (MMD) [18]. Ghifary $et\textit{al. proposed a multi- domain reconstruc- }$ tion auto-encoder [10]. Li $et\textit{al. applied MMD constraints to an autoencoder via}$ adversarial training [15].
Recently, meta-learning based techniques start to be used to solve DG problems. Li $et\textit{al. alternates domain- specific feature extractors and classifiers across}$ domains via episodic training, but without using inner gradient descent update [14]. Balaji $et\textit{al. proposed MetaReg that learns a regularization function}$ (e.g., weighted $\ell_1$ loss) particularly for the networks classification layer, while excluding the feature extractor [1].
Further, recent DG works forgo the requirement of source domains partitions and directly learn the cross-domain generalizable representations through a mixed collection of training data. Wang $et\textit{al. extracted robust feature rep- }$ resentation by projecting out superficial patterns like color and texture [31]. Wang $et\textit{al. penalized models tendency in predicting with local features in order}$ to extract robust globe representation [30]. RSC follows this more recent path and directly activates more features in all source domain data for DG without knowledge of the partition of source domains.
DG through Augmenting Source Domain: These methods augment the source domain to a wider span of the training data space, enlarging the possibility of covering the span of the data in the target domain. For example, An auxiliary domain classifer has been introduced to augment the data by perturbing input data based on the domain classification signal [23]. Volpi $et\textit{al. developed an}$ adversarial approach, in which samples are perturbed according to fictitious target distributions within a certain Wasserstein distance from the source [29]. A recent method with state-of-the art performance is JiGen [4], which leverages self-supervised signals by solving jigsaw puzzles.
$Key\textit{ difference: These approaches usually introduce a model- specific DG}$ model and rely on prior knowledge of the target domain, for instance, the tar-
|
[
[
0.287353515625,
0.0138092041015625,
0.8408203125,
0.01543426513671875,
0.8408203125,
0.033294677734375,
0.287353515625,
0.03167724609375
],
[
0.90087890625,
0.0162506103515625,
0.92138671875,
0.0162506103515625,
0.92138671875,
0.03167724609375,
0.90087890625,
0.03167724609375
],
[
0.11712646484375,
0.056060791015625,
0.9130859375,
0.056854248046875,
0.9130859375,
0.07470703125,
0.11712646484375,
0.07391357421875
],
[
0.0821533203125,
0.07879638671875,
0.91552734375,
0.07879638671875,
0.91552734375,
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,
0.9130859375,
0.11700439453125,
0.9130859375,
0.131591796875,
0.0809326171875,
0.131591796875
],
[
0.0809326171875,
0.1373291015625,
0.91552734375,
0.1373291015625,
0.91552734375,
0.15185546875,
0.0809326171875,
0.15185546875
],
[
0.0809326171875,
0.15673828125,
0.460205078125,
0.15673828125,
0.460205078125,
0.17138671875,
0.0809326171875,
0.17138671875
],
[
0.0809326171875,
0.2080078125,
0.10870361328125,
0.2080078125,
0.10870361328125,
0.2218017578125,
0.0809326171875,
0.2218017578125
],
[
0.10748291015625,
0.2120361328125,
0.11834716796875,
0.2120361328125,
0.11834716796875,
0.2169189453125,
0.10748291015625,
0.2169189453125
],
[
0.1256103515625,
0.2071533203125,
0.326171875,
0.2071533203125,
0.326171875,
0.2218017578125,
0.1256103515625,
0.2218017578125
],
[
0.0784912109375,
0.24609375,
0.91650390625,
0.247802734375,
0.91650390625,
0.264892578125,
0.0784912109375,
0.263916015625
],
[
0.0809326171875,
0.269775390625,
0.91650390625,
0.269775390625,
0.91650390625,
0.284423828125,
0.0809326171875,
0.284423828125
],
[
0.0809326171875,
0.289306640625,
0.91650390625,
0.289306640625,
0.91650390625,
0.3037109375,
0.0809326171875,
0.3037109375
],
[
0.0772705078125,
0.30712890625,
0.8310546875,
0.305419921875,
0.8310546875,
0.3232421875,
0.0772705078125,
0.324951171875
],
[
0.11834716796875,
0.327392578125,
0.91552734375,
0.327392578125,
0.91552734375,
0.342041015625,
0.11834716796875,
0.342041015625
],
[
0.0809326171875,
0.34765625,
0.91552734375,
0.34765625,
0.91552734375,
0.3623046875,
0.0809326171875,
0.3623046875
],
[
0.0809326171875,
0.3671875,
0.9130859375,
0.3671875,
0.9130859375,
0.3818359375,
0.0809326171875,
0.3818359375
],
[
0.0821533203125,
0.385986328125,
0.91650390625,
0.385986328125,
0.91650390625,
0.400390625,
0.0821533203125,
0.400390625
],
[
0.0821533203125,
0.40625,
0.9130859375,
0.40625,
0.9130859375,
0.4208984375,
0.0821533203125,
0.4208984375
],
[
0.0809326171875,
0.42578125,
0.91552734375,
0.42578125,
0.91552734375,
0.440185546875,
0.0809326171875,
0.440185546875
],
[
0.0772705078125,
0.44189453125,
0.338134765625,
0.443603515625,
0.338134765625,
0.46142578125,
0.0772705078125,
0.459716796875
],
[
0.11712646484375,
0.4638671875,
0.91162109375,
0.4638671875,
0.91162109375,
0.478515625,
0.11712646484375,
0.478515625
],
[
0.0809326171875,
0.484130859375,
0.91650390625,
0.484130859375,
0.91650390625,
0.498779296875,
0.0809326171875,
0.498779296875
],
[
0.0821533203125,
0.50341796875,
0.9130859375,
0.50341796875,
0.9130859375,
0.51806640625,
0.0821533203125,
0.51806640625
],
[
0.0821533203125,
0.52294921875,
0.91650390625,
0.52294921875,
0.91650390625,
0.53759765625,
0.0821533203125,
0.53759765625
],
[
0.0821533203125,
0.54248046875,
0.91796875,
0.54248046875,
0.91796875,
0.55712890625,
0.0821533203125,
0.55712890625
],
[
0.0809326171875,
0.56201171875,
0.44677734375,
0.56201171875,
0.44677734375,
0.57666015625,
0.0809326171875,
0.57666015625
],
[
0.11712646484375,
0.580078125,
0.91552734375,
0.580078125,
0.91552734375,
0.59765625,
0.11712646484375,
0.59765625
],
[
0.0809326171875,
0.60009765625,
0.91552734375,
0.60009765625,
0.91552734375,
0.61474609375,
0.0809326171875,
0.61474609375
],
[
0.0772705078125,
0.6181640625,
0.91552734375,
0.619140625,
0.91552734375,
0.63671875,
0.0772705078125,
0.63623046875
],
[
0.0809326171875,
0.63916015625,
0.91162109375,
0.63916015625,
0.91162109375,
0.65380859375,
0.0809326171875,
0.65380859375
],
[
0.0821533203125,
0.65869140625,
0.91552734375,
0.65869140625,
0.91552734375,
0.67333984375,
0.0821533203125,
0.67333984375
],
[
0.0809326171875,
0.67919921875,
0.91552734375,
0.67919921875,
0.91552734375,
0.69384765625,
0.0809326171875,
0.69384765625
],
[
0.0821533203125,
0.69775390625,
0.91796875,
0.69775390625,
0.91796875,
0.71240234375,
0.0821533203125,
0.71240234375
],
[
0.0821533203125,
0.71630859375,
0.560546875,
0.71630859375,
0.560546875,
0.73095703125,
0.0821533203125,
0.73095703125
],
[
0.11834716796875,
0.73681640625,
0.91552734375,
0.73681640625,
0.91552734375,
0.75146484375,
0.11834716796875,
0.75146484375
],
[
0.0809326171875,
0.75732421875,
0.9130859375,
0.75732421875,
0.9130859375,
0.77197265625,
0.0809326171875,
0.77197265625
],
[
0.0772705078125,
0.7734375,
0.91650390625,
0.77490234375,
0.91650390625,
0.7919921875,
0.0772705078125,
0.791015625
],
[
0.0809326171875,
0.7919921875,
0.91796875,
0.7919921875,
0.91796875,
0.81005859375,
0.0809326171875,
0.81005859375
],
[
0.0809326171875,
0.81298828125,
0.91796875,
0.81298828125,
0.91796875,
0.8310546875,
0.0809326171875,
0.8310546875
],
[
0.0809326171875,
0.83447265625,
0.91552734375,
0.83447265625,
0.91552734375,
0.84912109375,
0.0809326171875,
0.84912109375
],
[
0.0821533203125,
0.85400390625,
0.91162109375,
0.85400390625,
0.91162109375,
0.8681640625,
0.0821533203125,
0.8681640625
],
[
0.0821533203125,
0.873046875,
0.91552734375,
0.873046875,
0.91552734375,
0.8876953125,
0.0821533203125,
0.8876953125
],
[
0.0772705078125,
0.89013671875,
0.591796875,
0.89111328125,
0.591796875,
0.9091796875,
0.0772705078125,
0.908203125
],
[
0.11712646484375,
0.90966796875,
0.91552734375,
0.9091796875,
0.91552734375,
0.927734375,
0.11712646484375,
0.9287109375
],
[
0.0809326171875,
0.931640625,
0.91552734375,
0.931640625,
0.91552734375,
0.9462890625,
0.0809326171875,
0.9462890625
]
] |
[
"Self-Challenging Improves Cross-Domain Generalization",
"3",
"We present mathematical analysis that RsC induces a smaller generaliza",
"tion bound.Wefurther demonstrate the empirical strength of our method with",
"domain-agnostic cross-domain evaluations,following previous setup3i.Wealso",
"conduct ablation study to examine the alignment between its empirical perfor-",
"mance and our intuitive understanding.Theinspections also shed light upon the",
"choices ofits extrahyperparameter.",
"2",
"",
"Related Work",
"We summarizethe related DG works from two perspectives:learning domain",
"invariant features and augmenting source domain data. Further, as RsC can be",
"broadly viewed as a generic training heuristic for CNN,we also briefly discuss",
"the general-purpose regularizations that appear similar to our method.",
"DG through Learning Domain Invariant Features: These methods",
"typically minimize the discrepancy between source domains assuming that the",
"resulting features will be domain-invariant and generalize well for unseen tar-",
"get distributions. Along this track, Muandet et al.employed Maximum Mean",
"Discrepancy (MMD)[18].Ghifary etal.proposed amulti-domainreconstruc",
"tion auto-encoder[1ol.Li et al.applied MMD constraints to an autoencoder via",
"adversarial training [15].",
"Recently,meta-learning based techniques start to beused to solve DG prob",
"lems.Li et al. alternates domain-specific feature extractors and classifiers across",
"domains via episodic training,but without using inner gradient descent up",
"date[14l.Balaji et al.proposed MetaReg that learns aregularization function",
"(e.g.,weighted i loss)particularly for the networks classification layer,while",
"excluding the feature extractor [1].",
"Further, recent DG works forgo the requirement of source domains parti-",
"tions and directly learn the cross-domain generalizable representations through",
"a mixed collection of training data. Wang et al. extracted robust feature rep",
"resentation byprojecting out superficial patterns like color and texture [31]",
"Wang et al.penalized modelstendency in predicting with local features in order",
"to extract robust globe representation3ol.RSCfollowsthismorerecent path",
"and directly activates more features in all source domain data for DG without",
"knowledgeofthepartitionof sourcedomains.",
"DG through Augmenting Source Domain:Thesemethodsaugment the",
"source domainto awider span ofthetrainingdata space,enlargingthepossibility",
"of covering the span of thedatain thetargetdomain.For example, An auxiliary",
"domain classifier has been introduced to augment the data by perturbing input",
"data based on the domain classification signal [23]. Volpi et al. developed an",
"adversarial approach,in which samples are perturbed according tofictitious",
"target distributions within a certain Wasserstein distance from the source29",
"Arecentmethod withstate-of-theart performanceisJiGen4l,whichleverages",
"self-supervised signals by solving jigsaw puzzles.",
"Key difference: These approaches usually introduce a model-specific DG",
"model and rely on prior knowledge of the target domain, for instance,thetar-"
] |
{
"arxiv_id": "2007.02454",
"language": "en",
"page_idx": 2
}
|
||
2007.02454-en-3
|
2007.02454-en-3.jpg
|
4 Huang et al.
get spatial permutation is assumed by JiGen [4]. In contrast, RSC is a modelagnostic training algorithm that aims to improve the cross-domain robustness of any given model. More importantly, RSC does not utilize any knowledge of partitions of domains, either source domain or target domain, which is the general scenario in real world application.
Generic Model Regularization: CNNs are powerful models and tend to overfit on source domain datasets. From this perspective, model regularization, $e.g.$, weight decay [19], early stopping, and shake-shake regularization [8],could also improve the DG performance. Dropout [25] mutes features by randomly zeroing each hidden unit of the neural network during the training phase. In this way, the network benefit from the assembling effect of small subnetworks to achieve a good regularization effect. Cutout [6] and HaS [24] randomly drop patches of input images. SpatialDropout [26] randomly drops channels of a feature map. DropBlock [9] drops contiguous regions from feature maps instead of random units. DropPath [11] zeroes out an entire layer in training, not just a particular unit. MaxDrop [20] selectively drops features of high activations across the feature map or across the channels. Adversarial Dropout [21] dropouts for maximizing the divergence between the training supervision and the outputs from the network. [12] leverages Adversarial Dropout [21] to learn discriminative features by enforcing the cluster assumption.
$Key\textit{ difference: RSC differs from above methods in that RSC locates and}$ mutes most predictive parts of feature maps by gradients instead of randomness, activation or prediction divergence maximization. This selective process plays an important role in improving the convergence, as we will briefly argue later.
## 3 Method
$\textbf{Notations: ( x, y) denotes a sample- label pair from the data collection }( \mathbf{X} , \mathbf{Y} )$ with $n$ samples, and z (or $\mathbf{Z})$ denotes the feature representation of $(\mathbf{x},\mathbf{y})$ learned by a neural network. $f(\cdot;\theta)$ denotes the CNN model, whose parameters are denoted as $\theta.h(\cdot;\theta^\text{top})$ denotes the task component of $f(\cdot;\theta);h(\cdot;\theta^\text{top})$ takes z as input and outputs the logits prior to a softmax function; $\theta^\mathrm{top}$ denotes the parameters of $h(\cdot;\theta^{\mathrm{top}}).l(\cdot,\cdot)$ denotes a generic loss function. RSC requires one extra scalar hyperparameter: the percentage of the representations to be discarded, denoted as $p.$ Further, we use$^{~}\cdot$to denote the estimated quantities, use $\tilde{\cdot}$ to denote the quantities after the representations are discarded, and use $t$ in the subscript to index the iteration. For example, $\hat{\theta}_t$ means the estimated parameter at iteration $t.$
### 3.1 Self-Challenging Algorithm
As a generic deep learning training method, RSC solves the same standard loss
function as the ones used by many other neural networks, $i.e.$,
$$\widehat{\theta}=\arg\operatorname*{min}_{\theta}\sum_{\langle\mathbf{x},\mathbf{y}\rangle\sim\langle\mathbf{X},\mathbf{Y}\rangle}l(f(\mathbf{x};\theta),\mathbf{y}),$$
|
[
[
0.0784912109375,
0.01300048828125,
0.1026611328125,
0.017059326171875,
0.09661865234375,
0.033294677734375,
0.07244873046875,
0.0284271240234375
],
[
0.15576171875,
0.0138092041015625,
0.283935546875,
0.0162506103515625,
0.283935546875,
0.03411865234375,
0.15576171875,
0.03167724609375
],
[
0.0809326171875,
0.05767822265625,
0.91552734375,
0.056060791015625,
0.91552734375,
0.07391357421875,
0.0809326171875,
0.0755615234375
],
[
0.0784912109375,
0.0780029296875,
0.92041015625,
0.0755615234375,
0.92041015625,
0.09344482421875,
0.0784912109375,
0.09503173828125
],
[
0.0809326171875,
0.0966796875,
0.9130859375,
0.0966796875,
0.9130859375,
0.11456298828125,
0.0809326171875,
0.11456298828125
],
[
0.0809326171875,
0.11614990234375,
0.91650390625,
0.11614990234375,
0.91650390625,
0.134033203125,
0.0809326171875,
0.134033203125
],
[
0.0784912109375,
0.1356201171875,
0.434814453125,
0.1373291015625,
0.434814453125,
0.15185546875,
0.0784912109375,
0.1502685546875
],
[
0.11712646484375,
0.1527099609375,
0.91552734375,
0.1527099609375,
0.91552734375,
0.1705322265625,
0.11712646484375,
0.1705322265625
],
[
0.0772705078125,
0.1722412109375,
0.91552734375,
0.1746826171875,
0.91552734375,
0.19091796875,
0.0772705078125,
0.1893310546875
],
[
0.0809326171875,
0.193359375,
0.91650390625,
0.193359375,
0.91650390625,
0.211181640625,
0.0809326171875,
0.211181640625
],
[
0.0809326171875,
0.2144775390625,
0.91552734375,
0.2144775390625,
0.91552734375,
0.2291259765625,
0.0809326171875,
0.2291259765625
],
[
0.0809326171875,
0.2340087890625,
0.91552734375,
0.2340087890625,
0.91552734375,
0.24853515625,
0.0809326171875,
0.24853515625
],
[
0.0821533203125,
0.25341796875,
0.91650390625,
0.25341796875,
0.91650390625,
0.267333984375,
0.0821533203125,
0.267333984375
],
[
0.0772705078125,
0.2705078125,
0.91650390625,
0.271240234375,
0.91650390625,
0.289306640625,
0.0772705078125,
0.288330078125
],
[
0.0821533203125,
0.291748046875,
0.91162109375,
0.291748046875,
0.91162109375,
0.305419921875,
0.0821533203125,
0.305419921875
],
[
0.0809326171875,
0.31201171875,
0.92041015625,
0.31201171875,
0.92041015625,
0.32666015625,
0.0809326171875,
0.32666015625
],
[
0.0772705078125,
0.3291015625,
0.92041015625,
0.329833984375,
0.92041015625,
0.34765625,
0.0772705078125,
0.346923828125
],
[
0.0821533203125,
0.350830078125,
0.91796875,
0.350830078125,
0.91796875,
0.365478515625,
0.0821533203125,
0.365478515625
],
[
0.0821533203125,
0.370361328125,
0.91650390625,
0.370361328125,
0.91650390625,
0.385009765625,
0.0821533203125,
0.385009765625
],
[
0.0772705078125,
0.385986328125,
0.91650390625,
0.38818359375,
0.91650390625,
0.404541015625,
0.0772705078125,
0.40283203125
],
[
0.0784912109375,
0.406982421875,
0.91552734375,
0.406982421875,
0.91552734375,
0.4248046875,
0.0784912109375,
0.4248046875
],
[
0.0821533203125,
0.42822265625,
0.6015625,
0.42822265625,
0.6015625,
0.44189453125,
0.0821533203125,
0.44189453125
],
[
0.11712646484375,
0.445068359375,
0.91796875,
0.4443359375,
0.91796875,
0.462158203125,
0.11712646484375,
0.463134765625
],
[
0.0809326171875,
0.467041015625,
0.9130859375,
0.467041015625,
0.9130859375,
0.481689453125,
0.0809326171875,
0.481689453125
],
[
0.0821533203125,
0.4873046875,
0.91552734375,
0.4873046875,
0.91552734375,
0.501953125,
0.0821533203125,
0.501953125
],
[
0.0809326171875,
0.5068359375,
0.86962890625,
0.5068359375,
0.86962890625,
0.521484375,
0.0809326171875,
0.521484375
],
[
0.0784912109375,
0.55078125,
0.10992431640625,
0.55078125,
0.10992431640625,
0.56689453125,
0.0784912109375,
0.56689453125
],
[
0.1014404296875,
0.55078125,
0.2427978515625,
0.55078125,
0.2427978515625,
0.56884765625,
0.1014404296875,
0.56884765625
],
[
0.0809326171875,
0.5859375,
0.91552734375,
0.5859375,
0.91552734375,
0.603515625,
0.0809326171875,
0.603515625
],
[
0.0821533203125,
0.607421875,
0.91552734375,
0.607421875,
0.91552734375,
0.6220703125,
0.0821533203125,
0.6220703125
],
[
0.0809326171875,
0.62646484375,
0.91650390625,
0.62646484375,
0.91650390625,
0.64111328125,
0.0809326171875,
0.64111328125
],
[
0.0772705078125,
0.642578125,
0.92138671875,
0.642578125,
0.92138671875,
0.6630859375,
0.0772705078125,
0.6630859375
],
[
0.0772705078125,
0.66455078125,
0.91796875,
0.662109375,
0.91796875,
0.68017578125,
0.0772705078125,
0.681640625
],
[
0.0809326171875,
0.68408203125,
0.91650390625,
0.68408203125,
0.91650390625,
0.70166015625,
0.0809326171875,
0.70166015625
],
[
0.0809326171875,
0.705078125,
0.91650390625,
0.705078125,
0.91650390625,
0.7197265625,
0.0809326171875,
0.7197265625
],
[
0.0821533203125,
0.724609375,
0.9130859375,
0.724609375,
0.9130859375,
0.7392578125,
0.0821533203125,
0.7392578125
],
[
0.0809326171875,
0.7431640625,
0.91552734375,
0.7431640625,
0.91552734375,
0.7578125,
0.0809326171875,
0.7578125
],
[
0.0772705078125,
0.76025390625,
0.92041015625,
0.76123046875,
0.92041015625,
0.77880859375,
0.0772705078125,
0.77734375
],
[
0.0772705078125,
0.783203125,
0.340576171875,
0.7822265625,
0.340576171875,
0.7958984375,
0.0772705078125,
0.7978515625
],
[
0.0809326171875,
0.8251953125,
0.1473388671875,
0.8251953125,
0.1473388671875,
0.83984375,
0.0809326171875,
0.83984375
],
[
0.13525390625,
0.82373046875,
0.475830078125,
0.82470703125,
0.475830078125,
0.84228515625,
0.13525390625,
0.841796875
],
[
0.0821533203125,
0.8544921875,
0.91552734375,
0.8544921875,
0.91552734375,
0.869140625,
0.0821533203125,
0.869140625
],
[
0.0772705078125,
0.87255859375,
0.74169921875,
0.873046875,
0.74169921875,
0.89111328125,
0.0772705078125,
0.89013671875
],
[
0.30908203125,
0.8994140625,
0.439697265625,
0.904296875,
0.4384765625,
0.92529296875,
0.306884765625,
0.92138671875
],
[
0.475830078125,
0.90234375,
0.51806640625,
0.90234375,
0.51806640625,
0.92431640625,
0.475830078125,
0.92431640625
],
[
0.54833984375,
0.90478515625,
0.6845703125,
0.90478515625,
0.6845703125,
0.923828125,
0.54833984375,
0.923828125
],
[
0.39013671875,
0.91796875,
0.408203125,
0.92529296875,
0.3974609375,
0.9375,
0.379150390625,
0.9287109375
],
[
0.4384765625,
0.93017578125,
0.55419921875,
0.93017578125,
0.55419921875,
0.94482421875,
0.4384765625,
0.94482421875
]
] |
[
"4",
"Huang et al.",
"get spatial permutation is assumed by JiGen [4]. In contrast, RSC is a model-",
"agnostic training algorithm that aims to improve the cross-domain robustness of",
"any given model. More importantly, RsC does not utilize any knowledge of par-",
"titions of domains, either source domain or target domain, which is the general",
"scenario inrealworldapplication",
"Generic Model Regularization: CNNs are powerful models and tend to",
"overfit on source domain datasets.From this perspective, model regularization,",
"e.g., weight decay [19], early stopping, and shake-shake regularization [8], could",
"also improve the DG performance.Dropout25mutes features by randomly",
"zeroing each hidden unit of the neural network during the training phase. In",
"thiswaythe network benefit from the assembling effect of small subnetworks",
"to achieve a good regularization effect. Cutout [6] and Has [24] randomly drop",
"patches of input images.SpatialDropout[26l randomly drops channels of afea",
"ture map.DropBlock9] drops contiguous regions from feature maps instead of",
"random units. DropPath [1i] zeroes out an entire layer in training, not just a",
"particular unit.MaxDropJ20lselectively drops features of high activations across",
"the feature map or across the channels. Adversarial Dropout21dropouts for",
"maximizing the divergence between the training supervision and the outputs",
"from the network. [12] leverages Adversarial Dropout [21] to learn discrimina-",
"tive features by enforcing the cluster assumption.",
"Key difference: RSC differs from above methods in that RSC locates and",
"mutes most predictiveparts of feature maps by gradients instead of randomness",
"activation or prediction divergence maximization. This selective process plays an",
"important role in improving the convergence, as we will briefly arguelater.",
"3 ",
"Method",
"Notations: (x, y) denotes a sample-label pair from the data collection (X, Y)",
"with n samples,and z (or Z)denotesthefeature representation of (x,y)learned",
"by a neural network.f(;) denotes the CNN model,whose parameters are",
"denoted as 0. h(; otop) denotes the task component of f(-; 0); h(-; otop) takes",
"z as input and outputs the logits prior to a softmax function: gtop denotes",
"the parameters of h(:; otop). l(, ) denotes a generic loss function. RSC requires",
"one extra scalar hyperparameter:the percentage of therepresentationsto be",
"discarded, denoted as p. Further, we use :to denote the estimated quantities",
"use : to denote the quantities after the representations are discarded, and use",
"t in the subscriptto index the iteration.For example, ot means the estimated",
"parameter at iteration t.",
"3.1",
"Self-Challenging Algorithm",
"As a generic deep learning training method,RsC solves the same standard loss",
"function as the ones used by many other neural networks, i.e.,",
"=argmin",
"7",
"l(f(x; 0),y),",
"0",
"(x,y)~(X,Y)"
] |
{
"arxiv_id": "2007.02454",
"language": "en",
"page_idx": 3
}
|
||
2007.02454-en-4
|
2007.02454-en-4.jpg
|
Self-Challenging Improves Cross-Domain Generalization 5
but RSC solves it in a different manner.
At each iteration, RSC inspects the gradient, identifies and then mutes the most predictive subset of the representation z (by setting the corresponding values to zero), and fnally updates the entire model.
This simple heuristic has three steps (for simplicity, we drop the indices of
samples and assume the batch size is 1 in the following equations):
1. Locate: RSC first calculates the gradient of upper layers with respect to the
representation as follows:
$$\mathbf{g_z}=\partial(h(\mathbf{z};\widehat{\theta}_t^{\mathrm{top}})\odot\mathbf{y})/\partial\mathbf{z},$$
(1)
where $\odot$denotes an element-wise product. Then RSC computes the (100- $p)^{\mathrm{th~percentile,~denoted~as~}q_p.~\mathrm{Then~it~constructs~a~masking~vector~m~in~the}}$ same dimension of g as follows. For the $i^\mathrm{th}$ element:
$$\mathbf{m}(i)=\begin{cases}0,&\text{if}\quad\mathbf{g_z}(i)\ge q_p\\1,&\text{otherwise}\end{cases}$$
(2)
In other words, RSC creates a masking vector $\mathbf{m}$, whose element is set to 0 if the corresponding element in g is one of the top percentage elements in $\mathbf{g}$,and set to 1 otherwise.
2. Mute: For every representation z, RSC masks out the bits associated with
${\mathrm{larger~gradients~by:}}$
$$\tilde{\mathbf{z}}=\mathbf{z}\odot\mathbf{m}$$
(3)
3. Update: RSC computes the softmax with perturbed representation with
$$\tilde{\mathbf{s}}=\mathrm{softmax}(h(\tilde{\mathbf{z}};\widehat{\theta}_{t}^{\mathrm{top}})),$$
(4)
and then use the gradient
$$\tilde{\mathbf{g}}_{\theta}=\partial l(\tilde{\mathbf{s}},\mathbf{y})/\partial\widehat{\theta}_{t}$$
(5)
to update the entire model for $\widehat{\theta}_{t+1}$ with optimizers such as SGD or ADAM.
We summarize the procedure of RSC in Algorithm 1.No that operations of RSC comprise of only few simple operations such as pooling, threshold and element-wise product. Besides the weights of the original network, no extra parameter needs to be learned.
### 3.2 Theoretical Evidence
To expand the theoretical discussion smoothly, we will refer to the“dog”vs. “cat”classification example repeatedly as we progress. The basic set-up, as we introduced in the beginning of this paper, is the scenario of a child trying to learn the concepts of “dog”vs.“cat" from illustrations in her book: while the
|
[
[
0.287353515625,
0.0138092041015625,
0.8408203125,
0.01543426513671875,
0.8408203125,
0.033294677734375,
0.287353515625,
0.03167724609375
],
[
0.90087890625,
0.0162506103515625,
0.92138671875,
0.0162506103515625,
0.92138671875,
0.03167724609375,
0.90087890625,
0.03167724609375
],
[
0.0821533203125,
0.05767822265625,
0.50244140625,
0.05767822265625,
0.50244140625,
0.07232666015625,
0.0821533203125,
0.07232666015625
],
[
0.11474609375,
0.07635498046875,
0.91552734375,
0.0755615234375,
0.91552734375,
0.09344482421875,
0.11474609375,
0.09423828125
],
[
0.0784912109375,
0.0966796875,
0.91650390625,
0.0966796875,
0.91650390625,
0.11456298828125,
0.0784912109375,
0.11456298828125
],
[
0.0821533203125,
0.11700439453125,
0.63623046875,
0.11700439453125,
0.63623046875,
0.131591796875,
0.0821533203125,
0.131591796875
],
[
0.11834716796875,
0.1373291015625,
0.91796875,
0.1373291015625,
0.91796875,
0.15185546875,
0.11834716796875,
0.15185546875
],
[
0.0809326171875,
0.156005859375,
0.78515625,
0.156005859375,
0.78515625,
0.1705322265625,
0.0809326171875,
0.1705322265625
],
[
0.088134765625,
0.18359375,
0.91650390625,
0.1851806640625,
0.91650390625,
0.203857421875,
0.088134765625,
0.203125
],
[
0.11712646484375,
0.20556640625,
0.391357421875,
0.2047119140625,
0.391357421875,
0.2225341796875,
0.11712646484375,
0.223388671875
],
[
0.3779296875,
0.238037109375,
0.658203125,
0.2347412109375,
0.658203125,
0.256591796875,
0.3779296875,
0.260009765625
],
[
0.88427734375,
0.238037109375,
0.92138671875,
0.238037109375,
0.92138671875,
0.260009765625,
0.88427734375,
0.260009765625
],
[
0.1231689453125,
0.274658203125,
0.9130859375,
0.274658203125,
0.9130859375,
0.29248046875,
0.1231689453125,
0.29248046875
],
[
0.1231689453125,
0.29638671875,
0.91552734375,
0.29638671875,
0.91552734375,
0.31103515625,
0.1231689453125,
0.31103515625
],
[
0.11834716796875,
0.314453125,
0.6728515625,
0.314453125,
0.6728515625,
0.332275390625,
0.11834716796875,
0.332275390625
],
[
0.450439453125,
0.34765625,
0.65478515625,
0.351806640625,
0.65478515625,
0.373779296875,
0.450439453125,
0.36962890625
],
[
0.35986328125,
0.364013671875,
0.41796875,
0.364013671875,
0.41796875,
0.380126953125,
0.35986328125,
0.380126953125
],
[
0.88427734375,
0.360595703125,
0.92041015625,
0.360595703125,
0.92041015625,
0.38427734375,
0.88427734375,
0.38427734375
],
[
0.46142578125,
0.37451171875,
0.60986328125,
0.37451171875,
0.60986328125,
0.39306640625,
0.46142578125,
0.39306640625
],
[
0.11956787109375,
0.41015625,
0.92041015625,
0.4111328125,
0.92041015625,
0.428955078125,
0.11956787109375,
0.42822265625
],
[
0.11956787109375,
0.43212890625,
0.91650390625,
0.43212890625,
0.91650390625,
0.44921875,
0.11956787109375,
0.44921875
],
[
0.11834716796875,
0.450927734375,
0.391357421875,
0.448486328125,
0.391357421875,
0.46630859375,
0.11834716796875,
0.46875
],
[
0.090576171875,
0.469482421875,
0.91650390625,
0.469482421875,
0.91650390625,
0.4873046875,
0.090576171875,
0.4873046875
],
[
0.11834716796875,
0.489013671875,
0.32861328125,
0.489013671875,
0.32861328125,
0.5068359375,
0.11834716796875,
0.5068359375
],
[
0.460205078125,
0.52294921875,
0.5771484375,
0.52294921875,
0.5771484375,
0.541015625,
0.460205078125,
0.541015625
],
[
0.88427734375,
0.5224609375,
0.92138671875,
0.5224609375,
0.92138671875,
0.54443359375,
0.88427734375,
0.54443359375
],
[
0.091796875,
0.5595703125,
0.8818359375,
0.5595703125,
0.8818359375,
0.57421875,
0.091796875,
0.57421875
],
[
0.391357421875,
0.5908203125,
0.64111328125,
0.5908203125,
0.64111328125,
0.61181640625,
0.391357421875,
0.61181640625
],
[
0.88427734375,
0.59130859375,
0.92138671875,
0.59130859375,
0.92138671875,
0.61328125,
0.88427734375,
0.61328125
],
[
0.12078857421875,
0.62939453125,
0.393798828125,
0.62939453125,
0.393798828125,
0.64404296875,
0.12078857421875,
0.64404296875
],
[
0.42626953125,
0.662109375,
0.60888671875,
0.65869140625,
0.60888671875,
0.6806640625,
0.42626953125,
0.68408203125
],
[
0.88525390625,
0.6630859375,
0.92041015625,
0.6630859375,
0.92041015625,
0.681640625,
0.88525390625,
0.681640625
],
[
0.11956787109375,
0.7001953125,
0.91650390625,
0.7001953125,
0.91650390625,
0.71826171875,
0.11956787109375,
0.71826171875
],
[
0.11712646484375,
0.73193359375,
0.91650390625,
0.734375,
0.91650390625,
0.75146484375,
0.11712646484375,
0.75
],
[
0.0784912109375,
0.75146484375,
0.91796875,
0.7529296875,
0.91796875,
0.77001953125,
0.0784912109375,
0.76953125
],
[
0.0772705078125,
0.77197265625,
0.91552734375,
0.7724609375,
0.91552734375,
0.79052734375,
0.0772705078125,
0.78857421875
],
[
0.0809326171875,
0.79345703125,
0.38525390625,
0.79345703125,
0.38525390625,
0.80810546875,
0.0809326171875,
0.80810546875
],
[
0.0772705078125,
0.83837890625,
0.402099609375,
0.83935546875,
0.402099609375,
0.85693359375,
0.0772705078125,
0.8564453125
],
[
0.0821533203125,
0.87255859375,
0.9130859375,
0.87255859375,
0.9130859375,
0.88720703125,
0.0821533203125,
0.88720703125
],
[
0.0821533203125,
0.89111328125,
0.91796875,
0.89111328125,
0.91796875,
0.9091796875,
0.0821533203125,
0.9091796875
],
[
0.0784912109375,
0.91162109375,
0.91796875,
0.91162109375,
0.91796875,
0.9287109375,
0.0784912109375,
0.9287109375
],
[
0.0772705078125,
0.9287109375,
0.91796875,
0.93017578125,
0.91796875,
0.947265625,
0.0772705078125,
0.9462890625
]
] |
[
"Self-Challenging Improves Cross-Domain Generalization",
"5",
"but RSC solves it in a different manner.",
"At each iteration, RSC inspects the gradient, identifies and then mutes the",
"most predictive subset of the representation z (by setting the corresponding",
"values to zero), and finally updates the entire model.",
"This simple heuristic has three steps (for simplicity,we drop the indices of",
"samples and assume the batch size is 1 in the following equations):",
"1. Locate: RSC first calculates the gradient of upper layers with respect to the",
"representation as follows:",
"gz = 0(h(z; otop) o y)/8z,",
"(1)",
"where denotes an element-wise product. Then RSC computes the (100-",
"p)th percentile, denoted as qp.Then it constructs a masking vectorm inthe",
"same dimension of g as follows. For the ith element:",
"[o, if gz(i) ≥qp",
"m(i)",
"(2)",
"1,otherwise",
"In other words, RSC creates a masking vector m, whose element is set to 0",
"if the corresponding element in g is one of the top p percentage elements in",
"g, and set to 1 otherwise.",
"2. Mute: For every representation z, RsC masks out the bits associated with",
"larger gradients by:",
"z=zom",
"(3)",
"3.Update:RsC computesthe softmaxwithperturbedrepresentationwith",
"s = softmax(h(z; otop))",
"(4)",
"and then use the gradient",
"go = 0l(s,y)/00t",
"(5)",
"to update the entire model for t+1 with optimizers such as SGD or ADAM.",
"We summarize the procedure of RSC in Algorithm1.No that operations",
"of RsC comprise of only few simple operations such as pooling, threshold and",
"element-wise product.Besides the weights of the original network, no extra pa-",
"rameterneedsto be learned.",
"3.2Theoretical Evidence",
"To expand the theoretical discussion smoothly,we will refer to the “dog\"vs",
"\"cat\" classification example repeatedly as we progress. The basic set-up, as we",
"introduced in the beginning of this paper, is the scenario of a child trying to",
"learn the concepts of\"dog\"vs.“cat\"from illustrations in her book:while the"
] |
{
"arxiv_id": "2007.02454",
"language": "en",
"page_idx": 4
}
|
||
2004.00280-en-0
|
2004.00280-en-0.jpg
|
# $\textbf{Creating Something from Nothing: }$ $\textbf{Unsupervised Knowledge Distillation for Cross- Modal Hashing}$
Hengtong Hu$^1, 2* $, Lingxi Xie$^3$, Richang Hong$^1, 2\dagger $, Qi Tian$^3$
$^{1}$School of Computer Science and Information Engineering, Hefei University of Technology, $^{2}\hat{\text{Key Laboratory of Knowledge Engineering with Big Data, Hefei University of Technology,}}$
$^{3}$Huawei Inc.
huhengtong.hfut@gmail.com, 198808xc@gmail.com, hongrc@hfut.edu.cn, tian.qil@huawei.com
## $\mathbf{Abstract}$
<table>
<tbody>
<tr>
<th> </th>
<th>$WL?$</th>
<th>$ER?$</th>
<th>$KD?$</th>
</tr>
<tr>
<td>DCMH [17]</td>
<td> </td>
<td>$\overline{{\sqrt{7}}}$</td>
<td> </td>
</tr>
<tr>
<td>$SSAH$ [20]</td>
<td> </td>
<td>$\overline{V}$</td>
<td> </td>
</tr>
<tr>
<td>$UCH$ [21]</td>
<td>1</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{UGACH}$ [45]</td>
<td>$√$</td>
<td>$\overline{V}$</td>
<td> </td>
</tr>
<tr>
<td>$UKD$</td>
<td>$\sqrt{1}$</td>
<td>$\sqrt{v}$</td>
<td>$\overline{J}$</td>
</tr>
</tbody>
</table>
Table 1. The difference between our approach and some recent cross-modal hashing methods. Here, 'WL' indicates that training without using labels, “ER’indicates that the method utilizes extensive relevance information rather than only the pairwise information, and ‘KD’ indicates utilizing knowledge distillation in the training process.
In recent years, cross-modal hashing (CMH) has attracted increasing attentions, mainly because its potential ability of mapping contents from different modalities, especially in vision and language, into the same space, so that it becomes efficient in cross-modal data retrieval. There are $two mainframeworks for CMH, differing from each other in$ whether semantic supervision is required. Compared to the unsupervised methods, the supervised methods often enjoy more accurate results, but require much heavier labors in data annotation. In this paper, we propose a novel approach that enables guiding a supervised method using outputs produced by an unsupervised method. Specifically, we make
$\begin{array}{lll}{\mathrm{N}}&{useofteacher-student~optimization~}\end{array}$
$benchmarks,i.e.,theMIRFlickrandNUS-WIDEdatasets.$ Our approach outperforms all existing unsupervised methods by a large margin.
## is built with hashing codes.
## $1. \textbf{Introduction}$
State-of-the-art CMH methods can be roughly categorized into two parts, namely, supervised and unsupervised methods. Both of them learn to shrink the gap between the distributions of two sets of training data (e.g., using adversarial-learning-based approaches [20,21,13]), but they differ from each other in whether an instance-level annotation is provided during the training stage. From this perspective, the supervised CMH methods[1,25,22,34, 44], receiving additional supervision, often produce more accurate results, and the unsupervised counterparts, while achieving lower performance, are relatively easier to deploy to real-world scenarios.
This paper combines the benefits of both methods by a simple yet effective idea, known as creating something from nothing. The core idea is straightforward: the supervised methods do not really require each instance to be labeled, but they use the labels to estimate the similarity between each pair of cross-modal data. Such information, in case of no supervision, can also be obtained from calculating the distance between their feature vectors, with the features provided by a trained unsupervised CMH method. Our approach, unsupervised knowledge distillation (UKD),
Recently, with the rapid increase of multimedia data, cross-modal retrieval [37,46,18,47,7,1,10,25,22] has attracted more and more attentions in both academia and industry. The goal is to retrieve instances from one modality using a query instance from another modality, e.g., finding an image with a few textual tags. One of the most popular pipeline for this purpose, named cross-modal hashing $(\mathbf{CMH})[1,19,23,47,7]$,involves mapping contents in different modalities into a common hamming space. By compressing each instance into a fxed-length binary code, the storage cost can be dramatically reduced and the time complexity for retrieval is constant since the indexing structure
*This work was done when the first author was an intern at Huawei
Noah's Ark Lab
$^{\dagger}$Corresponding author.
1
|
[
[
0.310546875,
0.1319580078125,
0.66015625,
0.1331787109375,
0.66015625,
0.15087890625,
0.310546875,
0.149658203125
],
[
0.1690673828125,
0.1546630859375,
0.80224609375,
0.1558837890625,
0.80224609375,
0.173583984375,
0.1690673828125,
0.17236328125
],
[
0.2418212890625,
0.20263671875,
0.7294921875,
0.20263671875,
0.7294921875,
0.2203369140625,
0.2418212890625,
0.2203369140625
],
[
0.125,
0.2203369140625,
0.8486328125,
0.2215576171875,
0.8486328125,
0.2392578125,
0.125,
0.238037109375
],
[
0.125,
0.2392578125,
0.84619140625,
0.2412109375,
0.84619140625,
0.255615234375,
0.125,
0.25390625
],
[
0.431396484375,
0.2568359375,
0.53759765625,
0.2587890625,
0.53759765625,
0.2734375,
0.431396484375,
0.271484375
],
[
0.0196075439453125,
0.26513671875,
0.062103271484375,
0.26513671875,
0.06451416015625,
0.37890625,
0.0228729248046875,
0.37890625
],
[
0.096435546875,
0.27783203125,
0.6953125,
0.27783203125,
0.6953125,
0.291748046875,
0.096435546875,
0.291748046875
],
[
0.70166015625,
0.279052734375,
0.87353515625,
0.279052734375,
0.87353515625,
0.2890625,
0.70166015625,
0.2890625
],
[
0.236083984375,
0.32568359375,
0.314453125,
0.32568359375,
0.314453125,
0.343505859375,
0.236083984375,
0.343505859375
],
[
0.68115234375,
0.328857421875,
0.72216796875,
0.328857421875,
0.72216796875,
0.34033203125,
0.68115234375,
0.34033203125
],
[
0.73388671875,
0.328857421875,
0.76953125,
0.328857421875,
0.76953125,
0.34033203125,
0.73388671875,
0.34033203125
],
[
0.77783203125,
0.328857421875,
0.82275390625,
0.328857421875,
0.82275390625,
0.34033203125,
0.77783203125,
0.34033203125
],
[
0.56884765625,
0.343505859375,
0.66015625,
0.34521484375,
0.66015625,
0.35986328125,
0.56884765625,
0.35791015625
],
[
0.744140625,
0.348388671875,
0.75732421875,
0.348388671875,
0.75732421875,
0.359130859375,
0.744140625,
0.359130859375
],
[
0.098876953125,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.373779296875,
0.098876953125,
0.373779296875
],
[
0.57177734375,
0.35986328125,
0.65185546875,
0.35986328125,
0.65185546875,
0.375,
0.57177734375,
0.375
],
[
0.744140625,
0.3642578125,
0.75732421875,
0.3642578125,
0.75732421875,
0.375,
0.744140625,
0.375
],
[
0.03350830078125,
0.375,
0.037567138671875,
0.375,
0.037567138671875,
0.386962890625,
0.03350830078125,
0.386962890625
],
[
0.037567138671875,
0.371826171875,
0.047393798828125,
0.371826171875,
0.047393798828125,
0.375,
0.037567138671875,
0.375
],
[
0.08087158203125,
0.375,
0.46728515625,
0.375,
0.46728515625,
0.38818359375,
0.08087158203125,
0.38818359375
],
[
0.57275390625,
0.37890625,
0.6435546875,
0.37890625,
0.6435546875,
0.39013671875,
0.57275390625,
0.39013671875
],
[
0.69384765625,
0.380615234375,
0.70751953125,
0.380615234375,
0.70751953125,
0.391357421875,
0.69384765625,
0.391357421875
],
[
0.080078125,
0.39013671875,
0.46728515625,
0.39013671875,
0.46728515625,
0.404052734375,
0.080078125,
0.404052734375
],
[
0.57177734375,
0.392578125,
0.66845703125,
0.392578125,
0.66845703125,
0.4072265625,
0.57177734375,
0.4072265625
],
[
0.6953125,
0.395751953125,
0.7060546875,
0.395751953125,
0.7060546875,
0.406005859375,
0.6953125,
0.406005859375
],
[
0.744140625,
0.396484375,
0.75634765625,
0.396484375,
0.75634765625,
0.40478515625,
0.744140625,
0.40478515625
],
[
0.080078125,
0.40478515625,
0.4697265625,
0.40478515625,
0.4697265625,
0.419189453125,
0.080078125,
0.419189453125
],
[
0.02532958984375,
0.397705078125,
0.0555419921875,
0.397705078125,
0.0555419921875,
0.4462890625,
0.02532958984375,
0.4462890625
],
[
0.56884765625,
0.4072265625,
0.615234375,
0.4072265625,
0.615234375,
0.42236328125,
0.56884765625,
0.42236328125
],
[
0.7509765625,
0.4091796875,
0.75732421875,
0.41357421875,
0.7509765625,
0.42041015625,
0.744140625,
0.416748046875
],
[
0.080078125,
0.42041015625,
0.4697265625,
0.42041015625,
0.4697265625,
0.434326171875,
0.080078125,
0.434326171875
],
[
0.080078125,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.50341796875,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.50341796875,
0.448974609375
],
[
0.03350830078125,
0.443115234375,
0.053924560546875,
0.443115234375,
0.053924560546875,
0.473388671875,
0.03350830078125,
0.473388671875
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.080078125,
0.464111328125
],
[
0.5009765625,
0.448974609375,
0.89208984375,
0.449462890625,
0.89208984375,
0.46337890625,
0.5009765625,
0.462158203125
],
[
0.080078125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50244140625,
0.46337890625,
0.890625,
0.46337890625,
0.890625,
0.4765625,
0.50244140625,
0.4765625
],
[
0.080078125,
0.48095703125,
0.468017578125,
0.48095703125,
0.468017578125,
0.494384765625,
0.080078125,
0.494384765625
],
[
0.50341796875,
0.479248046875,
0.8896484375,
0.479248046875,
0.8896484375,
0.4892578125,
0.50341796875,
0.4892578125
],
[
0.080078125,
0.496826171875,
0.4697265625,
0.496826171875,
0.4697265625,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.50244140625,
0.4912109375,
0.89208984375,
0.4912109375,
0.89208984375,
0.5048828125,
0.50244140625,
0.5048828125
],
[
0.02777099609375,
0.498779296875,
0.053924560546875,
0.498779296875,
0.053924560546875,
0.54541015625,
0.02777099609375,
0.54541015625
],
[
0.080078125,
0.51123046875,
0.468017578125,
0.51123046875,
0.468017578125,
0.5244140625,
0.080078125,
0.5244140625
],
[
0.50341796875,
0.5068359375,
0.60107421875,
0.5068359375,
0.60107421875,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.080078125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.080078125,
0.54052734375
],
[
0.0269622802734375,
0.5361328125,
0.056365966796875,
0.5361328125,
0.056365966796875,
0.580078125,
0.0269622802734375,
0.580078125
],
[
0.080078125,
0.54150390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.55615234375,
0.080078125,
0.55615234375
],
[
0.50244140625,
0.5498046875,
0.6806640625,
0.5498046875,
0.6806640625,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.07843017578125,
0.556640625,
0.4697265625,
0.5546875,
0.4697265625,
0.568359375,
0.07843017578125,
0.57080078125
],
[
0.080078125,
0.5712890625,
0.468017578125,
0.5712890625,
0.468017578125,
0.58544921875,
0.080078125,
0.58544921875
],
[
0.521484375,
0.56640625,
0.890625,
0.56640625,
0.890625,
0.58056640625,
0.521484375,
0.58056640625
],
[
0.0236968994140625,
0.5771484375,
0.058013916015625,
0.5771484375,
0.058013916015625,
0.705078125,
0.0236968994140625,
0.705078125
],
[
0.037567138671875,
0.57763671875,
0.047393798828125,
0.57763671875,
0.047393798828125,
0.58349609375,
0.037567138671875,
0.58349609375
],
[
0.08087158203125,
0.5859375,
0.468017578125,
0.5859375,
0.468017578125,
0.599609375,
0.08087158203125,
0.599609375
],
[
0.50244140625,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.035125732421875,
0.59765625,
0.0400390625,
0.59765625,
0.0400390625,
0.60498046875,
0.035125732421875,
0.60498046875
],
[
0.08087158203125,
0.60107421875,
0.2279052734375,
0.60107421875,
0.2279052734375,
0.6142578125,
0.08087158203125,
0.6142578125
],
[
0.50244140625,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.50244140625,
0.611328125
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.50341796875,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.50341796875,
0.642578125
],
[
0.08251953125,
0.646484375,
0.20751953125,
0.646484375,
0.20751953125,
0.6611328125,
0.08251953125,
0.6611328125
],
[
0.50341796875,
0.64453125,
0.8896484375,
0.64453125,
0.8896484375,
0.65478515625,
0.50341796875,
0.65478515625
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.09722900390625,
0.67041015625,
0.468017578125,
0.671875,
0.468017578125,
0.68603515625,
0.09722900390625,
0.68505859375
],
[
0.50244140625,
0.67236328125,
0.890625,
0.671875,
0.890625,
0.68603515625,
0.50244140625,
0.6875
],
[
0.08087158203125,
0.6875,
0.4697265625,
0.6875,
0.4697265625,
0.70068359375,
0.08087158203125,
0.70068359375
],
[
0.50341796875,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.08087158203125,
0.7041015625,
0.465576171875,
0.7041015625,
0.465576171875,
0.7138671875,
0.08087158203125,
0.7138671875
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.70263671875,
0.89208984375,
0.71630859375,
0.50244140625,
0.71630859375
],
[
0.080078125,
0.71728515625,
0.46728515625,
0.71728515625,
0.46728515625,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.50341796875,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.73291015625,
0.468017578125,
0.73291015625,
0.468017578125,
0.7470703125,
0.080078125,
0.7470703125
],
[
0.50244140625,
0.734375,
0.65869140625,
0.734375,
0.65869140625,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.080078125,
0.74853515625,
0.46728515625,
0.74853515625,
0.46728515625,
0.76220703125,
0.080078125,
0.76220703125
],
[
0.521484375,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.521484375,
0.76513671875
],
[
0.080078125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.080078125,
0.7763671875
],
[
0.50244140625,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.7763671875,
0.468017578125,
0.77783203125,
0.468017578125,
0.79248046875,
0.080078125,
0.79150390625
],
[
0.50244140625,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.08087158203125,
0.79345703125,
0.468017578125,
0.79345703125,
0.468017578125,
0.806640625,
0.08087158203125,
0.806640625
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.08087158203125,
0.8095703125,
0.4697265625,
0.8095703125,
0.4697265625,
0.82275390625,
0.08087158203125,
0.82275390625
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82470703125,
0.46728515625,
0.82470703125,
0.46728515625,
0.83544921875,
0.08087158203125,
0.83544921875
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.83837890625,
0.471435546875,
0.83837890625,
0.471435546875,
0.8515625,
0.080078125,
0.8515625
],
[
0.5009765625,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.096435546875,
0.861328125,
0.471435546875,
0.8623046875,
0.471435546875,
0.87548828125,
0.096435546875,
0.875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.87548828125,
0.1690673828125,
0.87548828125,
0.1690673828125,
0.88720703125,
0.080078125,
0.88720703125
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.09478759765625,
0.88623046875,
0.223876953125,
0.88720703125,
0.223876953125,
0.9013671875,
0.09478759765625,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88525390625,
0.890625,
0.8994140625,
0.5009765625,
0.9013671875
],
[
0.483642578125,
0.92822265625,
0.488525390625,
0.92822265625,
0.488525390625,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Creating Something from Nothing:",
"Unsupervised Knowledge Distillation for Cross-Modal Hashing",
"Hengtong Hul.2* Lingxi Xie3, Richang Hongl.2↑ Qi Tian3",
"ISchool of Computer Science and Information Engineering, Hefei University of Technology,",
"2Key Laboratory of Knowledge Engineering with Big Data, Hefei University of Technology",
"3Huawei Inc.",
"000",
"huhengtong.hfut@gmail.com,198808xc@gmail.com,hongrc@hfut.edu.cn,",
"tian.gil@huawei.com",
"Abstract",
"WL?",
"ER?",
"KD?",
"DCMH [17]",
"",
"In recent years, cross-modal hashing (CMH) has at-",
"SSAH[20]",
"",
"",
"",
"tracted increasing attentions, mainly because its potential",
"UCH[21]",
"V",
"ability of mapping contents from different modalities, espe-",
"UGACH [45]",
"V",
"V",
"cially in vision and language, into the same space, so that",
"LAD",
"UKD",
"J",
"it becomes efficient in cross-modal data retrieval. There are",
"two main frameworks for CMH, differing from each other in",
"Table 1. The difference between our approach and some recent",
"'SO",
"whether semantic supervision is required. Compared to the",
"cross-modal hashing methods.Here,WL' indicates that training",
"unsupervised methods, the supervised methods often enjoy",
"without using labels, 'ER' indicates that the method utilizes ex-",
"more accurate results, but require much heavier labors in",
"tensiverelevanceinformation ratherthanonlythepairwise infor-",
"data annotation. In this paper, we propose a novel approach",
"mation, and 'KD' indicates utilizing knowledge distillation in the",
"A082",
"that enables guiding a supervised method using outputs pro-",
"training process.",
"duced by an unsupervised method. Specifically, we make",
"3200",
"use of teacher-student optimization for propagating knowl-",
"is built with hashing codes.",
"edge.Experiments areperformed on two popular CMH",
"benchmarks, i.e., the MIRFlickr and NUS-WIDE datasets.",
"State-of-the-art CMH methods can be roughly catego-",
"",
"",
"Our approach outperforms all existing unsupervised meth-",
"rized into two parts, namely, supervised and unsupervised",
"",
"ods by a large margin.",
"methods. Both of them learn to shrink the gap between",
"the distributions of two sets of training data (e.g., using",
"adversarial-learning-based approaches [20, 21, 13]), but",
"1. Introduction",
"theydifferfromeachotherinwhetheraninstance-levelan-",
"notation is provided during the training stage.From this",
"Recently, with the rapid increase of multimedia data,",
"perspective, the supervised CMH methods [1, 25, 22, 34,",
"cross-modal retrieval [37, 46, 18, 47, 7, 1, 10, 25, 22] has",
"44, receiving additional supervision, often produce more",
"attractedmoreandmoreattentionsinbothacademiaandin-",
"accurate results, and the unsupervised counterparts, while",
"dustry. The goal is to retrieve instances from one modality",
"achieving lower performance,are relatively easier to deploy",
"using a query instance from another modality, e.g., finding",
"to real-world scenarios.",
"an image with a few textual tags.One of the most pop-",
"This paper combines the benefits of both methods by",
"ular pipeline for this purpose, named cross-modal hashing",
"a simple yet effective idea, known as creating something",
"(CMH) [1, 19, 23, 47, 7], involves mapping contents in dif-",
"from nothing. The core idea is straightforward: the su-",
"ferent modalities into a common hamming space. By com-",
"pervised methods do not really require each instance to be",
"pressing each instance into a fixed-length binary code, the",
"labeled, but they use the labels to estimate the similarity be-",
"storagecostcanbedramaticallyreduced andthetimecom-",
"tween each pair of cross-modal data. Such information, in",
"plexity for retrieval is constant since the indexing structure",
"case of no supervision, can also be obtained from calculat-",
"\"This work was done when the first author was an intern at Huawei",
"ing the distance between their feature vectors, with the fea-",
"Noah's Ark Lab.",
"tures provided by a trained unsupervised CMH method. Our",
"tCorresponding author.",
"approach, unsupervised knowledge distillation (UKD),",
"1"
] |
{
"arxiv_id": "2004.00280",
"language": "en",
"page_idx": 0
}
|
||
2004.00280-en-1
|
2004.00280-en-1.jpg
|
contains an unsupervised CMH module followed by another supervised one, both of which can be freely replaced by new and more powerful models in the future.
Our research paves the way towards an interesting direction that using an unsupervised method to guide a supervised method, for which CMH is a good scenario to test on. We perform experiments on two popular cross-modality retrieval datasets, $i.e.$, MIRFlickr and NUS-WIDE, and demonstrate state-of-the-art performance, outperforming existing unsupervised CMH methods by a significant margin. Moreover, we delve deep into the benefits of supervision, and point out a few directions for future research.
The remainder of this paper is organized as follows. Section 2 briefly reviews the preliminaries of cross-modal retrieval and hashing, and Section 3 describes the unsupervised knowledge distillation approach. Experimental results are shown in Section 4 and conclusions are drawn in Section 5.
## $2. \textbf{Related Work}$
$2. 1. \textbf{ Cross- Modal Retrieval and Hashing}$
Cross-modal retrieval aims to search semantically similar instances in one modality using a query from another modality [37, 39]. Throughout this paper, we consider the retrieval task between vision and language, $i.e.$, involving images and texts. To map them into the same space, two models need to be trained, one for each modality. The goal is to make the image-text pairs with relevant semantics to be close in the feature space. To train and evaluate the mapping functions, a dataset with image-text pairs is present. The dataset is further split into a training set and a query set, $i.e.$, the testing stage is performed on the query set.
In the past decade, many efforts were made on this topic [18,46,37]. However, most of them suffered from high computation costs in real-world, high-dimensional data. To scale up these models to real-world scenarios, researchers often compressed the output of these models into binary vectors of a fixed length $[1,19,10,24],i.e.$, Hashing codes. In this situation, this task is often referred to as cross-modality hashing.
$2. 2. \textbf{ Supervised Cross- Modal Hashing Methods}$
The fundamental challenges of cross-modal hashing lie in learning reliable mapping functions to bridge the modality gap. Supervised methods [25,47,38,20,39,7] achieved this goal by exploiting semantic labels to capture rich correlation information among data from different modalities. Traditional supervised learning methods were mostly based on handcrafted features, and aimed to understand the semantic relevance in the common space. SePH[22] proposed a semantics-preserving hashing method which aimed to approximate the distribution of semantic labels with
hash codes on the Hamming space via minimizing the KLdivergence. Wang et al.[34] proposed to leverage listwise supervision into a principled framework of learning the hashing function.
With the rapid development of deep learning, researchers started to build supervised methods upon more powerful yet discriminative features. DCMH[17] proposed a deep crossmodal hashing method by integrating feature learning and binary quantization into one framework. SSAH[20]improved this work by proposing a self-supervised approach, which incorporated adversarial learning into cross-modal hashing. Zhang et al. [47] also investigated a similar idea by proposing an adversarial hashing network with an attention mechanism to enhance the measurement of content-level similarities. These supervised methods achieved superior performance, arguably by acquiring correlation information from the semantic labels of both images and texts. However, acquiring a large amount of such labels is often expensive and thus intractable, which makes the supervised approaches infeasible in the real-world applications.
2.3. Unsupervised Cross-Modal Hashing Methods
Compared with the supervised counterparts, unsupervised cross-modal hashing methods [8,45,13,36] only relied on the correlation information from the paired data, making it easier to be deployed to other scenarios. These methods usually learned hashing codes by preserving interand intra-correlations. For example, Song et $al.[32]$ proposed inter-media hashing to establish a common Hamming space by maintaining inter-media and intra-media consistency. Recently, several works introduced deep learning to improve unsupervised cross-modal hashing. UGACH[45] utilized a generative adversarial network to exploit the underlying manifold structure of cross-modal data. As an improvement, UCH [21] coupled the generative adversarial network to build two cycled networks in an unified framework to learn common representations and hash mapping simultaneously
Despite the superiority in reducing the burden of data annotation, the accuracy of unsupervised cross-modal hashing methods is often below satisfaction, in particular, much lower than the supervised counterparts. The main reason lies in lacking the knowledge of pairwise similarity for the training data pairs. On the other hand, we notice that the output of an unsupervised model contains, though somewhat inaccurate, such semantic information. This motivates us to guide a supervised model by the output of an unsupervised model. This is yet another type of research which distills knowledge to assist model training.
# 3. Our Approach
In this work, we focus on the idea named creating something from nothing, $i.e.$, a supervised cross-modal hash-
2
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
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,
0.346435546875,
0.12249755859375,
0.346435546875,
0.1363525390625,
0.080078125,
0.1363525390625
],
[
0.50341796875,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.50341796875,
0.136962890625
],
[
0.096435546875,
0.136962890625,
0.468017578125,
0.1363525390625,
0.468017578125,
0.15087890625,
0.096435546875,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.64453125,
0.1383056640625,
0.64453125,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.154052734375,
0.4697265625,
0.154052734375,
0.4697265625,
0.1673583984375,
0.080078125,
0.1673583984375
],
[
0.52294921875,
0.1514892578125,
0.890625,
0.1514892578125,
0.890625,
0.1654052734375,
0.52294921875,
0.1654052734375
],
[
0.080078125,
0.1685791015625,
0.4697265625,
0.1685791015625,
0.4697265625,
0.1817626953125,
0.080078125,
0.1817626953125
],
[
0.5009765625,
0.1666259765625,
0.89208984375,
0.1673583984375,
0.89208984375,
0.1817626953125,
0.5009765625,
0.18115234375
],
[
0.080078125,
0.184326171875,
0.468017578125,
0.184326171875,
0.468017578125,
0.1976318359375,
0.080078125,
0.1976318359375
],
[
0.50244140625,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.50341796875,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.210205078125,
0.50341796875,
0.210205078125
],
[
0.080078125,
0.2139892578125,
0.4697265625,
0.2139892578125,
0.4697265625,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50244140625,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.4697265625,
0.2291259765625,
0.4697265625,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50244140625,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.07843017578125,
0.2442626953125,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.07843017578125,
0.25830078125
],
[
0.50341796875,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.438720703125,
0.2587890625,
0.438720703125,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.50244140625,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.271484375,
0.50244140625,
0.271484375
],
[
0.1004638671875,
0.274658203125,
0.46728515625,
0.274658203125,
0.46728515625,
0.28857421875,
0.1004638671875,
0.28857421875
],
[
0.5009765625,
0.2734375,
0.89208984375,
0.272705078125,
0.89208984375,
0.287353515625,
0.5009765625,
0.28857421875
],
[
0.08087158203125,
0.291748046875,
0.46728515625,
0.291748046875,
0.46728515625,
0.3017578125,
0.08087158203125,
0.3017578125
],
[
0.50244140625,
0.2890625,
0.8935546875,
0.2890625,
0.8935546875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.468017578125,
0.304931640625,
0.468017578125,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50244140625,
0.3037109375,
0.8935546875,
0.3037109375,
0.8935546875,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.080078125,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.333984375,
0.080078125,
0.333984375
],
[
0.50341796875,
0.32080078125,
0.890625,
0.32080078125,
0.890625,
0.330810546875,
0.50341796875,
0.330810546875
],
[
0.080078125,
0.335205078125,
0.4697265625,
0.335205078125,
0.4697265625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50244140625,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.07843017578125,
0.349853515625,
0.1502685546875,
0.349853515625,
0.1502685546875,
0.3642578125,
0.07843017578125,
0.3642578125
],
[
0.50341796875,
0.349853515625,
0.8896484375,
0.349853515625,
0.8896484375,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.50244140625,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.376953125,
0.50244140625,
0.376953125
],
[
0.08087158203125,
0.380126953125,
0.2181396484375,
0.380126953125,
0.2181396484375,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50341796875,
0.380126953125,
0.8447265625,
0.380126953125,
0.8447265625,
0.393310546875,
0.50341796875,
0.393310546875
],
[
0.080078125,
0.40283203125,
0.3896484375,
0.40478515625,
0.3896484375,
0.419189453125,
0.080078125,
0.417236328125
],
[
0.50341796875,
0.401611328125,
0.88232421875,
0.401611328125,
0.88232421875,
0.416748046875,
0.50341796875,
0.416748046875
],
[
0.098876953125,
0.4287109375,
0.46728515625,
0.4287109375,
0.46728515625,
0.442626953125,
0.098876953125,
0.442626953125
],
[
0.521484375,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.521484375,
0.439453125
],
[
0.080078125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.458251953125,
0.080078125,
0.458251953125
],
[
0.50341796875,
0.44140625,
0.890625,
0.44140625,
0.890625,
0.455078125,
0.50341796875,
0.455078125
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.473388671875,
0.080078125,
0.473388671875
],
[
0.50244140625,
0.455810546875,
0.890625,
0.455810546875,
0.890625,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.080078125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.4716796875,
0.89208984375,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.07843017578125,
0.488037109375,
0.468017578125,
0.4892578125,
0.468017578125,
0.50390625,
0.07843017578125,
0.50244140625
],
[
0.50244140625,
0.486083984375,
0.89208984375,
0.486083984375,
0.89208984375,
0.5,
0.50244140625,
0.5
],
[
0.080078125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.08087158203125,
0.521484375,
0.46728515625,
0.521484375,
0.46728515625,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.080078125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.548828125,
0.080078125,
0.548828125
],
[
0.50341796875,
0.5322265625,
0.890625,
0.5322265625,
0.890625,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.08087158203125,
0.5498046875,
0.46728515625,
0.5498046875,
0.46728515625,
0.56298828125,
0.08087158203125,
0.56298828125
],
[
0.50244140625,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.50244140625,
0.56103515625
],
[
0.080078125,
0.56298828125,
0.468017578125,
0.564453125,
0.468017578125,
0.5791015625,
0.080078125,
0.57763671875
],
[
0.50244140625,
0.5625,
0.890625,
0.5625,
0.890625,
0.57568359375,
0.50244140625,
0.57568359375
],
[
0.080078125,
0.580078125,
0.439453125,
0.580078125,
0.439453125,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50341796875,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.50341796875,
0.59033203125
],
[
0.098876953125,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.098876953125,
0.609375
],
[
0.50244140625,
0.59228515625,
0.89208984375,
0.59228515625,
0.89208984375,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.08087158203125,
0.6123046875,
0.46728515625,
0.6123046875,
0.46728515625,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.50341796875,
0.60693359375,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.07843017578125,
0.62548828125,
0.468017578125,
0.625,
0.468017578125,
0.6396484375,
0.07843017578125,
0.64013671875
],
[
0.50244140625,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.080078125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.5009765625,
0.63623046875,
0.89208984375,
0.63720703125,
0.89208984375,
0.6513671875,
0.5009765625,
0.65087890625
],
[
0.080078125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.080078125,
0.66845703125
],
[
0.50341796875,
0.65380859375,
0.6044921875,
0.65380859375,
0.6044921875,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.080078125,
0.67041015625,
0.4697265625,
0.67041015625,
0.4697265625,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.52294921875,
0.66748046875,
0.89208984375,
0.66748046875,
0.89208984375,
0.68115234375,
0.52294921875,
0.68115234375
],
[
0.080078125,
0.68603515625,
0.4697265625,
0.68603515625,
0.4697265625,
0.69970703125,
0.080078125,
0.69970703125
],
[
0.50341796875,
0.68310546875,
0.890625,
0.68310546875,
0.890625,
0.6962890625,
0.50341796875,
0.6962890625
],
[
0.080078125,
0.7021484375,
0.238525390625,
0.7021484375,
0.238525390625,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50341796875,
0.69873046875,
0.89208984375,
0.69873046875,
0.89208984375,
0.7119140625,
0.50341796875,
0.7119140625
],
[
0.50244140625,
0.7119140625,
0.890625,
0.7119140625,
0.890625,
0.72607421875,
0.50244140625,
0.72607421875
],
[
0.080078125,
0.72607421875,
0.438720703125,
0.72607421875,
0.438720703125,
0.7392578125,
0.080078125,
0.7392578125
],
[
0.5048828125,
0.72998046875,
0.890625,
0.72998046875,
0.890625,
0.74072265625,
0.5048828125,
0.74072265625
],
[
0.50244140625,
0.74365234375,
0.89208984375,
0.74365234375,
0.89208984375,
0.7568359375,
0.50244140625,
0.7568359375
],
[
0.098876953125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.098876953125,
0.7646484375
],
[
0.5009765625,
0.7568359375,
0.8896484375,
0.7568359375,
0.8896484375,
0.77099609375,
0.5009765625,
0.77099609375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.7734375,
0.89208984375,
0.7734375,
0.89208984375,
0.787109375,
0.50244140625,
0.787109375
],
[
0.080078125,
0.7822265625,
0.4697265625,
0.7822265625,
0.4697265625,
0.79541015625,
0.080078125,
0.79541015625
],
[
0.50244140625,
0.78857421875,
0.890625,
0.78857421875,
0.890625,
0.8017578125,
0.50244140625,
0.8017578125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.8037109375,
0.89208984375,
0.8037109375,
0.89208984375,
0.8173828125,
0.50341796875,
0.8173828125
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.5009765625,
0.8173828125,
0.78369140625,
0.818359375,
0.78369140625,
0.83349609375,
0.5009765625,
0.83203125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50244140625,
0.8447265625,
0.64892578125,
0.8447265625,
0.64892578125,
0.8623046875,
0.50244140625,
0.8623046875
],
[
0.080078125,
0.85693359375,
0.465576171875,
0.85693359375,
0.465576171875,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.51953125,
0.869140625,
0.89208984375,
0.87060546875,
0.89208984375,
0.88525390625,
0.51953125,
0.8837890625
],
[
0.07843017578125,
0.88720703125,
0.4697265625,
0.88623046875,
0.4697265625,
0.8994140625,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50244140625,
0.8994140625
],
[
0.48046875,
0.923828125,
0.49267578125,
0.923828125,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"contains an unsupervised CMH module followed by another",
"hash codes on the Hamming space via minimizing the KL-",
"supervised one, both of which can be freely replaced by new",
"divergence. Wang et al. [34] proposed to leverage list-",
"and more powerful models in the future.",
"wise supervision into a principled framework of learning",
"Our research paves the way towards an interesting di-",
"the hashing function.",
"rection that using an unsupervised method to guide a su-",
"With the rapid development of deep learning,researchers",
"pervised method, for which CMH is a good scenario to test",
"started to build supervised methods upon more powerful yet",
"on.We perform experiments on two popular cross-modality",
"discriminative features. DCMH [17] proposed a deep cross-",
"retrieval datasets, i.e., MIRFlickr and NUS-WIDE, and",
"modalhashingmethodbyintegratingfeaturelearningand",
"demonstrate state-of-the-art performance, outperforming",
"binary quantization into one framework. SSAH [20] im-",
"existing unsupervised CMH methods by a significant mar-",
"proved this work by proposing a self-supervised approach,",
"gin.Moreover, we delvedeep into the benefits of supervi-",
"which incorporated adversarial learning into cross-modal",
"sion, and point out a few directions for future research.",
"hashing. Zhang et al. [47] also investigated a similar idea by",
"The remainder of this paper is organized as follows. Sec-",
"proposing an adversarial hashing network with an attention",
"tion 2 brieflyreviewsthe preliminaries of cross-modal re-",
"mechanism to enhance the measurement of content-level",
"trieval and hashing, and Section 3 describes the unsuper-",
"similarities. These supervised methods achieved superior",
"vised knowledge distillation approach.Experimental re-",
"performance,arguablybyacguiringcorrelationinformation",
"sults are shown in Section 4 and conclusions are drawn in",
"from the semantic labels of both images and texts. How-",
"Section 5.",
"ever, acquiring a large amount of such labels is often ex-",
"pensive and thus intractable, which makes the supervised",
"2. Related Work",
"approaches infeasible in the real-world applications.",
"2.1. Cross-Modal Retrieval and Hashing",
"2.3. Unsupervised Cross-Modal Hashing Methods",
"Cross-modal retrieval aims to search semantically sim-",
"Compared with the supervised counterparts, unsuper-",
"ilar instances in one modality using a query from another",
"vised cross-modal hashing methods [8, 45, 13, 36] only",
"modality [37, 39]. Throughout this paper, we consider the",
"relied on the correlation information from the paired data",
"retrieval task between vision and language, i.e., involving",
"making it easier to be deployed to other scenarios. These",
"images and texts. To map them into the same space, two",
"methods usually learned hashing codes by preserving inter-",
"models need to be trained, one for each modality. The goal",
"and intra-correlations. For example, Song et al. [32] pro-",
"istomaketheimage-textpairswithrelevantsemanticsto",
"posed inter-media hashing to establish a common Hamming",
"be close in the feature space. To train and evaluate the map-",
"space by maintaining inter-media and intra-media consis-",
"ping functions, a dataset with image-text pairs is present.",
"tency. Recently, several works introduced deep learning to",
"The dataset is further split into a training set and a query",
"improve unsupervised cross-modal hashing. UGACH [45]",
"set, i.e., the testing stage is performed on the query set.",
"utilized a generative adversarial network to exploit the un-",
"In the past decade, many efforts were made on this",
"derlying manifold structure of cross-modal data. As an im-",
"topic[18.46.371.However,mostofthemsufferedfrom",
"provement, UCH [21] coupled the generative adversarial",
"high computation costs in real-world, high-dimensional",
"network to build two cycled networks in an unified frame-",
"data. To scale up these models to real-world scenarios, re-",
"work to learn common representations and hash mapping",
"searchers often compressed the output of these models into",
"simultaneously",
"binary vectors of a fixed length [1, 19, 10, 24], i.e., Hash-",
"Despite the superiority in reducing the burden of data",
"ing codes. In this situation, this task is often referred to as",
"annotation, the accuracy of unsupervised cross-modal hash-",
"cross-modality hashing.",
"ing methods is often below satisfaction, in particular, much",
"lower than the supervised counterparts. The main reason",
"2.2. Supervised Cross-Modal Hashing Methods",
"liesinlackingtheknowledgeofpairwisesimilarityforthe",
"training data pairs. On the other hand, we notice that the",
"The fundamental challenges of cross-modal hashing lie",
"output of an unsupervised model contains, though some-",
"in learning reliable mapping functions to bridge the modal-",
"what inaccurate, such semantic information. This motivates",
"ity gap. Supervised methods [25, 47, 38, 20, 39, 7] achieved",
"us to guide a supervised model by the output of an unsu-",
"this goal by exploiting semantic labels to capture rich cor-",
"pervised model. This is yet another type of research which",
"relation information among data from different modalities.",
"distills knowledge to assist model training.",
"Traditional supervised learning methods were mostly based",
"on handcrafted features, and aimed to understand the se-",
"3. Our Approach",
"mantic relevance in the common space. SePH [22] pro-",
"posed a semantics-preserving hashing method which aimed",
"In this work, we focus on the idea named creating some-",
"to approximate the distribution of semantic labels with",
"thing from nothing, i.e., a supervised cross-modal hash-",
"2"
] |
{
"arxiv_id": "2004.00280",
"language": "en",
"page_idx": 1
}
|
||
2004.00280-en-2
|
2004.00280-en-2.jpg
|
<FigureHere>
Figure 1. The proposed UKD framework which involves training a teacher model in an unsupervised manner, constructing the similarity matrix S by distilling knowledge from the teacher model, and using $\mathbf{S}$ to supervise the student model. Each dot represents an intermediate feature. Please zoom in to see the details of this figure.
Given S, the objective of training involves minimizing
the total distance with respect to $\theta^\mathrm{I}$ and $\theta^\mathrm{T},i.e.$,
(1)
$$\boldsymbol{\theta}^{\mathrm{I},\star},\boldsymbol{\theta}^{\mathrm{T},\star}=\arg\min_{\boldsymbol{\theta}^{\mathrm{I}},\boldsymbol{\theta}^{\mathrm{T}}}=\sum_{i,j}S_{i,j}\cdot\left|\mathbf{f}_{i}^{\mathrm{I}}-\mathbf{f}_{j}^{\mathrm{T}}\right|.$$
ing method can be guided by the output of an unsupervised method, which reveals the similarity between training data pairs. Figure 1 shows the framework of the proposed UKD. In what follows, we frst explain the motivation of our approach, and then introduce the proposed pipeline, unsupervised knowledge distillation, from two aspects, namely, how to distill similarity from an unsupervised model, and how to utilize it efficiently to optimize a supervised model.
3.1. Supervised and Unsupervised Baselines
Therefore, the definition of S forms the major challenge of the learning task. According to whether extra labels of images and texts, besides the paired information, are used, existing methods can be categorized into either supervised or unsupervised learning. In the supervised setting, instancelevel annotations $(e.g.$, classification tags) are used to mea-
training set contains paired data, $ie, \mathcal{D} = \left \{ \mathbf{v} _{n}^{\mathrm{I} }, \mathbf{v} _{n}^{\mathrm{T} }\right \} _{n= 1}^{N}$, sure whether two instances are relevant, while in the unsu- where $N$ is the number of image- text pairs. Here, $\mathbf{v} _{n}^{\mathrm{I} }\in \mathbb{R} ^{D_{1}}$pervised setting. n
the former provides more accurate estimation on S than the latter and, consequently, stronger models for cross-modal hashing. However, collecting additional annotations, even at the instance level, can be a large burden especially when the dataset is very large. Hence, we focus on improving the performance of unsupervised learning methods which are easier to be deployed to real-world scenarios.
## $3. 2. \textbf{ Unsupervised Knowledge Distillation}$
Throughout this paper, we consider the case that the I and T denote `images' and `texts', and $D_{\mathrm{I}}$ and $D_{\mathrm{T}}$ denote the dimensionality of the feature spaces, respectively. $D_\mathrm{I}$ and $D_{\mathrm{T}}$ can be different, $e.g.$, as in our experiments. The models that map them into the same space are denoted as $\mathbf{f} _{n}^{\mathrm{I} }$ $\doteq$ $f^{\mathrm{I} }\left ( \mathbf{v} _{n}^{\mathrm{I} }; \mathbf{\theta } ^{\mathrm{I} }\right )$ $\in$ $\mathbb{R} ^{K}$ and $\mathbf{f} _{n}^{\mathrm{T} }$ $\doteq$ $f^{\mathrm{T} }\left ( \mathbf{v} _{n}^{\mathrm{T} }; \mathbf{\theta } ^{\mathrm{T} }\right )$ $\in$ $\mathbb{R} ^{K}$, respectively, where $K$ is the dimensionality of the common feature space, and $\theta^\mathrm{I}$ and $\theta^\mathrm{T}$ are model parameters. The compressed hashing code for images and texts are denoted $\mathbf{by}$ $\mathbf{b} _{n}^{\mathrm{I} }$ $\doteq$sgn$\left ( \mathbf{f} _{n}^{\mathrm{I} }\right )$ and$\mathbf{b} _{n}^{\mathrm{T} }$ $\doteq$sgn$\bigl(\mathbf{f}_{n}^{\mathrm{T}}\bigr)$,respectively$,i.e.$, $\mathrm{both~b}_{n}^{\mathrm{I}}$and $\mathbf{b}_n^{\mathrm{T}}$fall within$\left\{-1,+1\right\}^K.$
The key to cross-modal hashing lies in recognizing which pairs of image-text data are semantically relevant while other are not, so that the model can learn to pull the features of relevant pairs closer in the common space. A straightforward idea is to define all paired image and text instances to be relevant and all others irrelevant. However, this strategy produces a very small positive set and a much larger negative set, which often causes data imbalance during the training stage. A generalized yet more effective solution is to define a similarity matrix $\mathbf{S}\in\{0,1\}^{N\times N}$, so that when $S_{i,j}=1$ defines a positive pair $(i,j)$ and vice versa. The original sampling strategy is equivalent to $\mathbf{S}\equiv\mathbf{I}.$
Our idea originates from the fact that, as shown above, the difference between supervised and unsupervised crossmodal hashing algorithms is not big, but supervised methods often report much higher accuracy than the unsupervised counterparts. Moreover, the supervised algorithms do not require real supervision, namely, the manually labeled image/text tags, but only need to know, or estimate, the similarity between any pair of data, $i.e.$, elements in S. Beyond the unsupervised baseline that estimates S using raw image/text features (extracted from a pre-trained deep network or computed using bag-of-words statistics), we seek for the possibility that a cross-modal retrieval model, after trained in an unsupervised manner, can produce a more accurate estimation of S. We illustrate an example in Figure 2. Later,
3
|
[
[
0.2091064453125,
0.0947265625,
0.294189453125,
0.0947265625,
0.294189453125,
0.1060791015625,
0.2091064453125,
0.1060791015625
],
[
0.363525390625,
0.0947265625,
0.476318359375,
0.0947265625,
0.476318359375,
0.10797119140625,
0.363525390625,
0.10797119140625
],
[
0.66259765625,
0.09344482421875,
0.744140625,
0.09344482421875,
0.744140625,
0.10479736328125,
0.66259765625,
0.10479736328125
],
[
0.36279296875,
0.10797119140625,
0.414306640625,
0.10797119140625,
0.414306640625,
0.11932373046875,
0.36279296875,
0.11932373046875
],
[
0.4306640625,
0.10919189453125,
0.483642578125,
0.10919189453125,
0.483642578125,
0.11932373046875,
0.4306640625,
0.11932373046875
],
[
0.57177734375,
0.10919189453125,
0.6611328125,
0.10919189453125,
0.6611328125,
0.12060546875,
0.57177734375,
0.12060546875
],
[
0.361083984375,
0.12371826171875,
0.414306640625,
0.12371826171875,
0.414306640625,
0.1337890625,
0.361083984375,
0.1337890625
],
[
0.428955078125,
0.121826171875,
0.484375,
0.121826171875,
0.484375,
0.1363525390625,
0.428955078125,
0.1363525390625
],
[
0.68310546875,
0.1256103515625,
0.70166015625,
0.1256103515625,
0.70166015625,
0.1337890625,
0.68310546875,
0.1337890625
],
[
0.69921875,
0.1256103515625,
0.7138671875,
0.1256103515625,
0.7138671875,
0.1337890625,
0.69921875,
0.1337890625
],
[
0.73779296875,
0.125,
0.74609375,
0.125,
0.74609375,
0.1337890625,
0.73779296875,
0.1337890625
],
[
0.373291015625,
0.1337890625,
0.39794921875,
0.1337890625,
0.39794921875,
0.154052734375,
0.373291015625,
0.154052734375
],
[
0.65673828125,
0.1275634765625,
0.66259765625,
0.1275634765625,
0.66259765625,
0.1337890625,
0.65673828125,
0.1337890625
],
[
0.44775390625,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.15283203125,
0.44775390625,
0.15283203125
],
[
0.779296875,
0.1383056640625,
0.8310546875,
0.1383056640625,
0.8310546875,
0.1483154296875,
0.779296875,
0.1483154296875
],
[
0.65185546875,
0.14453125,
0.71337890625,
0.146484375,
0.71337890625,
0.1610107421875,
0.65185546875,
0.1590576171875
],
[
0.348876953125,
0.154052734375,
0.40185546875,
0.154052734375,
0.40185546875,
0.1641845703125,
0.348876953125,
0.1641845703125
],
[
0.50244140625,
0.1546630859375,
0.56201171875,
0.1546630859375,
0.56201171875,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.73779296875,
0.15087890625,
0.74267578125,
0.15087890625,
0.74267578125,
0.1578369140625,
0.73779296875,
0.1578369140625
],
[
0.19775390625,
0.160400390625,
0.3203125,
0.160400390625,
0.3203125,
0.17041015625,
0.19775390625,
0.17041015625
],
[
0.449462890625,
0.1578369140625,
0.48779296875,
0.1578369140625,
0.48779296875,
0.169189453125,
0.449462890625,
0.169189453125
],
[
0.56787109375,
0.1634521484375,
0.7138671875,
0.1610107421875,
0.7138671875,
0.1717529296875,
0.56787109375,
0.173583984375
],
[
0.73779296875,
0.1634521484375,
0.74658203125,
0.1634521484375,
0.74658203125,
0.17041015625,
0.73779296875,
0.17041015625
],
[
0.50244140625,
0.1654052734375,
0.56201171875,
0.1654052734375,
0.56201171875,
0.1767578125,
0.50244140625,
0.1767578125
],
[
0.3505859375,
0.1641845703125,
0.370849609375,
0.1641845703125,
0.370849609375,
0.17236328125,
0.3505859375,
0.17236328125
],
[
0.373291015625,
0.179931640625,
0.378173828125,
0.179931640625,
0.378173828125,
0.1900634765625,
0.373291015625,
0.1900634765625
],
[
0.41259765625,
0.1767578125,
0.423095703125,
0.1767578125,
0.423095703125,
0.1881103515625,
0.41259765625,
0.1881103515625
],
[
0.20751953125,
0.1849365234375,
0.316162109375,
0.1849365234375,
0.316162109375,
0.1956787109375,
0.20751953125,
0.1956787109375
],
[
0.443603515625,
0.186279296875,
0.4697265625,
0.1881103515625,
0.46728515625,
0.20263671875,
0.441162109375,
0.199462890625
],
[
0.5703125,
0.1817626953125,
0.650390625,
0.1817626953125,
0.650390625,
0.1932373046875,
0.5703125,
0.1932373046875
],
[
0.1356201171875,
0.1944580078125,
0.1405029296875,
0.1944580078125,
0.1405029296875,
0.199462890625,
0.1356201171875,
0.199462890625
],
[
0.370849609375,
0.1944580078125,
0.381591796875,
0.1956787109375,
0.378173828125,
0.2108154296875,
0.36767578125,
0.208984375
],
[
0.373291015625,
0.1900634765625,
0.3798828125,
0.1900634765625,
0.3798828125,
0.1932373046875,
0.373291015625,
0.1932373046875
],
[
0.410888671875,
0.1976318359375,
0.423095703125,
0.1976318359375,
0.423095703125,
0.210205078125,
0.410888671875,
0.210205078125
],
[
0.73779296875,
0.1976318359375,
0.74658203125,
0.1976318359375,
0.74658203125,
0.212158203125,
0.73779296875,
0.212158203125
],
[
0.316162109375,
0.212158203125,
0.361083984375,
0.2139892578125,
0.359375,
0.2266845703125,
0.314453125,
0.2247314453125
],
[
0.60107421875,
0.212158203125,
0.60888671875,
0.212158203125,
0.60888671875,
0.216552734375,
0.60107421875,
0.216552734375
],
[
0.73779296875,
0.212158203125,
0.744140625,
0.212158203125,
0.744140625,
0.216552734375,
0.73779296875,
0.216552734375
],
[
0.77392578125,
0.208984375,
0.8310546875,
0.208984375,
0.8310546875,
0.2203369140625,
0.77392578125,
0.2203369140625
],
[
0.373291015625,
0.2171630859375,
0.378173828125,
0.2171630859375,
0.378173828125,
0.2266845703125,
0.373291015625,
0.2266845703125
],
[
0.410888671875,
0.21533203125,
0.4248046875,
0.21533203125,
0.4248046875,
0.2310791015625,
0.410888671875,
0.2310791015625
],
[
0.31787109375,
0.2266845703125,
0.340576171875,
0.2266845703125,
0.340576171875,
0.23486328125,
0.31787109375,
0.23486328125
],
[
0.736328125,
0.2215576171875,
0.74658203125,
0.2215576171875,
0.74658203125,
0.23486328125,
0.736328125,
0.23486328125
],
[
0.416748046875,
0.231689453125,
0.42724609375,
0.240478515625,
0.416748046875,
0.248779296875,
0.406005859375,
0.240478515625
],
[
0.441162109375,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2392578125,
0.441162109375,
0.2392578125
],
[
0.373291015625,
0.238037109375,
0.378173828125,
0.238037109375,
0.378173828125,
0.24560546875,
0.373291015625,
0.24560546875
],
[
0.57421875,
0.238037109375,
0.63818359375,
0.238037109375,
0.63818359375,
0.248779296875,
0.57421875,
0.248779296875
],
[
0.736328125,
0.2374267578125,
0.74658203125,
0.2374267578125,
0.74658203125,
0.246826171875,
0.736328125,
0.246826171875
],
[
0.359375,
0.248779296875,
0.486083984375,
0.248779296875,
0.486083984375,
0.261962890625,
0.359375,
0.261962890625
],
[
0.080078125,
0.271484375,
0.890625,
0.271484375,
0.890625,
0.285400390625,
0.080078125,
0.285400390625
],
[
0.080078125,
0.285400390625,
0.89208984375,
0.285400390625,
0.89208984375,
0.298583984375,
0.080078125,
0.298583984375
],
[
0.080078125,
0.2998046875,
0.40283203125,
0.2998046875,
0.40283203125,
0.313232421875,
0.080078125,
0.313232421875
],
[
0.07843017578125,
0.337158203125,
0.4697265625,
0.33837890625,
0.4697265625,
0.352783203125,
0.07843017578125,
0.3515625
],
[
0.521484375,
0.337158203125,
0.89208984375,
0.33837890625,
0.89208984375,
0.352783203125,
0.521484375,
0.3515625
],
[
0.080078125,
0.354248046875,
0.4697265625,
0.354248046875,
0.4697265625,
0.367431640625,
0.080078125,
0.367431640625
],
[
0.5009765625,
0.3515625,
0.81884765625,
0.3515625,
0.81884765625,
0.3662109375,
0.5009765625,
0.3662109375
],
[
0.08087158203125,
0.37060546875,
0.465576171875,
0.37060546875,
0.465576171875,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.5439453125,
0.375,
0.859375,
0.376953125,
0.859375,
0.397705078125,
0.5439453125,
0.395751953125
],
[
0.07843017578125,
0.38330078125,
0.46728515625,
0.3837890625,
0.46728515625,
0.3984375,
0.07843017578125,
0.397705078125
],
[
0.86767578125,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.393310546875,
0.86767578125,
0.393310546875
],
[
0.65869140625,
0.393310546875,
0.69140625,
0.393310546875,
0.69140625,
0.40087890625,
0.65869140625,
0.40087890625
],
[
0.080078125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.71337890625,
0.397705078125,
0.736328125,
0.397705078125,
0.736328125,
0.410400390625,
0.71337890625,
0.410400390625
],
[
0.07843017578125,
0.41357421875,
0.468017578125,
0.4140625,
0.468017578125,
0.4287109375,
0.07843017578125,
0.427978515625
],
[
0.50341796875,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.080078125,
0.429931640625,
0.4697265625,
0.429931640625,
0.4697265625,
0.443115234375,
0.080078125,
0.443115234375
],
[
0.50244140625,
0.43310546875,
0.89208984375,
0.43310546875,
0.89208984375,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.080078125,
0.44580078125,
0.465576171875,
0.44580078125,
0.465576171875,
0.458984375,
0.080078125,
0.458984375
],
[
0.50341796875,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.50341796875,
0.462158203125
],
[
0.50244140625,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50244140625,
0.4765625
],
[
0.08087158203125,
0.4716796875,
0.416748046875,
0.4716796875,
0.416748046875,
0.48486328125,
0.08087158203125,
0.48486328125
],
[
0.50341796875,
0.479248046875,
0.89208984375,
0.479248046875,
0.89208984375,
0.492431640625,
0.50341796875,
0.492431640625
],
[
0.1004638671875,
0.494384765625,
0.468017578125,
0.494384765625,
0.468017578125,
0.5087890625,
0.1004638671875,
0.5087890625
],
[
0.5009765625,
0.49365234375,
0.8896484375,
0.49365234375,
0.8896484375,
0.5068359375,
0.5009765625,
0.5068359375
],
[
0.0718994140625,
0.50244140625,
0.455810546875,
0.50244140625,
0.455810546875,
0.53515625,
0.0718994140625,
0.53515625
],
[
0.5009765625,
0.50830078125,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.5009765625,
0.521484375
],
[
0.07598876953125,
0.52587890625,
0.4697265625,
0.5234375,
0.4697265625,
0.54443359375,
0.07598876953125,
0.546875
],
[
0.50244140625,
0.5234375,
0.89208984375,
0.5234375,
0.89208984375,
0.537109375,
0.50244140625,
0.537109375
],
[
0.5009765625,
0.537109375,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.5009765625,
0.5517578125
],
[
0.080078125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.080078125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57373046875,
0.080078125,
0.57373046875
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.50244140625,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.080078125,
0.57470703125,
0.468017578125,
0.57470703125,
0.468017578125,
0.58837890625,
0.080078125,
0.58837890625
],
[
0.50244140625,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50244140625,
0.59765625
],
[
0.07843017578125,
0.59033203125,
0.4697265625,
0.59033203125,
0.4697265625,
0.60400390625,
0.07843017578125,
0.60400390625
],
[
0.07843017578125,
0.60498046875,
0.4697265625,
0.60498046875,
0.4697265625,
0.619140625,
0.07843017578125,
0.619140625
],
[
0.50341796875,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.611328125,
0.50341796875,
0.611328125
],
[
0.5009765625,
0.61279296875,
0.890625,
0.61279296875,
0.890625,
0.626953125,
0.5009765625,
0.626953125
],
[
0.08087158203125,
0.6220703125,
0.1993408203125,
0.6220703125,
0.1993408203125,
0.63623046875,
0.08087158203125,
0.63623046875
],
[
0.191162109375,
0.62060546875,
0.418212890625,
0.62060546875,
0.418212890625,
0.63818359375,
0.191162109375,
0.63818359375
],
[
0.414306640625,
0.62353515625,
0.46728515625,
0.62353515625,
0.46728515625,
0.63525390625,
0.414306640625,
0.63525390625
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.080078125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.080078125,
0.65478515625
],
[
0.5009765625,
0.64306640625,
0.7998046875,
0.64306640625,
0.7998046875,
0.6572265625,
0.5009765625,
0.6572265625
],
[
0.080078125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.07843017578125,
0.67041015625,
0.4697265625,
0.671875,
0.4697265625,
0.68603515625,
0.07843017578125,
0.68505859375
],
[
0.50244140625,
0.66552734375,
0.8203125,
0.66650390625,
0.8203125,
0.68115234375,
0.50244140625,
0.6796875
],
[
0.07598876953125,
0.68310546875,
0.473876953125,
0.681640625,
0.473876953125,
0.70263671875,
0.07598876953125,
0.7041015625
],
[
0.521484375,
0.689453125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.521484375,
0.70263671875
],
[
0.0767822265625,
0.7041015625,
0.330078125,
0.70263671875,
0.330078125,
0.71728515625,
0.0767822265625,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.09722900390625,
0.7197265625,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.09722900390625,
0.734375
],
[
0.50244140625,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.73486328125,
0.50244140625,
0.73486328125
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50244140625,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.75048828125,
0.4697265625,
0.75048828125,
0.4697265625,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50341796875,
0.75244140625,
0.890625,
0.75244140625,
0.890625,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50341796875,
0.7666015625,
0.89208984375,
0.7666015625,
0.89208984375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.78271484375,
0.4697265625,
0.78271484375,
0.4697265625,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50341796875,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
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,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.5009765625,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5009765625,
0.85498046875
],
[
0.07598876953125,
0.85400390625,
0.443603515625,
0.849609375,
0.443603515625,
0.8720703125,
0.07598876953125,
0.87548828125
],
[
0.43701171875,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.869140625,
0.43701171875,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.8720703125,
0.471435546875,
0.8720703125,
0.471435546875,
0.88623046875,
0.07843017578125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.0767822265625,
0.88623046875,
0.46728515625,
0.88525390625,
0.46728515625,
0.8994140625,
0.0767822265625,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.5009765625,
0.8994140625
],
[
0.48046875,
0.923828125,
0.491943359375,
0.923828125,
0.491943359375,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"Teacher Model",
"Generative Module",
"Student Model",
"00.00",
"00-00",
"Image Modality",
"00...00",
"",
"",
"?",
"?",
"国",
"",
"店",
"0000",
"888",
"generating",
"Knowledge",
"",
"+ImageFeatureExtraction",
"reward",
"conv1-1 convl-2 conv2-1conv5-4 fe6fe7",
"fc&",
"Distillation",
"pair",
"",
"8",
"Text Feature Extraction",
"",
"Text Modality",
"2",
"",
"-",
"8",
"C",
"selecting",
"",
".",
"000000",
"",
"：",
"pairs",
"00",
"C",
"",
"",
"multi-sceale fusion",
"fel",
"Discriminative Module",
"Figure 1. The proposed UKD framework which involves training a teacher model in an unsupervised manner, constructing the similarity",
"matrix S by distilling knowledge from the teacher model, and using S to supervise the student model. Each dot represents an intermediate",
"feature. Please zoom in to see the details of this figure.",
"ing method can be guided by the output of an unsupervised",
"Given S, the objective of training involves minimizing",
"method, which reveals the similarity between training data",
"the total distance with respect to o' and T, i.e..",
"pairs.FigureIshowstheframeworkoftheproposedUKD",
"g1*,oT*= arg min = Si,3 f! -fTI.",
"In what follows, we first explain the motivation of our ap-",
"(1)",
"O1OT",
"proach, and then introduce the proposed pipeline, unsuper-",
"i.j",
"vised knowledge distillation, from two aspects, namely,",
"Therefore, the definition of S forms the major challenge of",
"how to distill similarity from an unsupervised model, and",
"the learning task. According to whether extra labels of im-",
"how to utilize it efficiently to optimize a supervised model.",
"ages and texts, besides the paired information, are used, ex-",
"isting methods can be categorized into either supervised or",
"3.1. Supervised and Unsupervised Baselines",
"unsupervised learning. In the supervised setting, instance-",
"Throughout this paper, we consider the case that the",
"level annotations (e.g., classification tags) are used to mea-",
"trainin se onans pired data (",
"sure whether two instances are relevant, while in the unsu-",
"where N is the number of image-text pairs. Here, v, e RD",
"pervised setting, no additional labels are available and thus",
"the raw features are the only source of judgment. Obviously,",
"be an image and vT e RDr be a text, where the superscripts",
"I and T denote \"images' and 'texts', and Dr and DT denote",
"the former provides more accurate estimation on S than the",
"latter and, consequently, stronger models for cross-modal",
"the dimensionality of the feature spaces, respectively. Dr",
"hashing. However, collecting additional annotations, even",
"and Dr can be different, e.g., as in our experiments. The",
"models that map them into the same space are denoted as",
"attheinstancelevel,canbealargeburdenespeciallywhen",
"the dataset is very large. Hence, we focus on improving the",
"fI=fI(v;0)",
") RK and fT = fT(vT;oT)",
"ERK",
"performance of unsupervised learning methods which are",
"respectively,where K is the dimensionality of the common",
"easier to be deployed to real-world scenarios.",
"feature space, and e' and oT are model parameters. The",
"compressed hashing code for images and texts are denoted",
"3.2. Unsupervised Knowledge Distillation",
"by b, = sgn(fI) and bT = sgn(fT), respectively, i.e.,",
"Our idea originates from the fact that, as shown above,",
"both b, and bT fall within f-1,+1?",
"the difference between supervised and unsupervised cross-",
"The key to cross-modal hashing lies in recognizing",
"modal hashing algorithms is not big, but supervised meth-",
"which pairs of image-text data are semantically relevant",
"ods often report much higher accuracy than the unsuper-",
"while other are not, so that the model can learn to pull the",
"vised counterparts.Moreover,the supervised algorithms do",
"features of relevant pairs closer in the common space. :A",
"not require real supervision, namely, the manually labeled",
"straightforward idea is to define all paired image and text",
"image/texttags,butonlyneed toknow,or estimate,the sim-",
"instances to be relevant and all others irrelevant. However,",
"ilarity between any pair of data, i.e., elements in S. Beyond",
"this strategy produces a very small positive set and a much",
"the unsupervised baseline that estimates S using raw im-",
"larger negative set, which often causes data imbalance dur-",
"age/text features (extracted from a pre-trained deep network",
"ing the training stage. A generalized yet more effective so-",
"or computed using bag-of-words statistics), we seek for the",
"lution is to define a similarity matrix s e (0,1)NxN",
",so",
"possibility that a cross-modal retrieval model, after trained",
"that when Si,j = 1 defines a positive pair (i, j) and vice",
"in an unsupervised manner, can produce a more accurate es-",
"versa. The original sampling strategy is equivalent to S = I.",
"timation of S. We illustrate an example in Figure 2. Later,",
"3"
] |
{
"arxiv_id": "2004.00280",
"language": "en",
"page_idx": 2
}
|
||
2004.00280-en-3
|
2004.00280-en-3.jpg
|
<table>
<tbody>
<tr>
<th>Function</th>
<th>new</th>
<th>image</th>
<th>text</th>
<th>${\mathrm{P@1000}}$</th>
<th>${\mathrm{P@500C}}$</th>
</tr>
<tr>
<td>$S_{i,}$ 2 - $\left|\mathbf{v}_{i}^{\mathrm{I}}-\mathbf{v}_{i}^{\mathrm{I}}\right|_{2}$ /2</td>
<td> </td>
<td>$√$</td>
<td> </td>
<td>$74.6\%$</td>
<td>$64.4\%$</td>
</tr>
<tr>
<td>$S_{i,.}$ 2 - $\left|\mathbf{v}_{i}^{\mathrm{T}}\right|$ $\Gamma-\mathbf{v}_{i}^{\mathrm{T}}\Big|$ /2</td>
<td> </td>
<td> </td>
<td>$√$</td>
<td>$57.1\%$</td>
<td>$55.6\%$</td>
</tr>
<tr>
<td>$S_{i,.}$ $\left|2-\left|\mathbf{f}_{i}^{\mathrm{I}}-\mathbf{f}_{j}^{\mathrm{I}}\right|_{2}\right|_{2}$ /2</td>
<td>$\sqrt{1}$</td>
<td>$√$</td>
<td> </td>
<td>$84.6\%$</td>
<td>$74.6\%$</td>
</tr>
<tr>
<td>$S_{i,.}$ 2 - $\left|\mathbf{f}_{i}^{\mathrm{T}}-\mathbf{f}_{i}^{\mathrm{T}}\right|$ /2</td>
<td>$√$</td>
<td> </td>
<td>$√$</td>
<td>$75.9\%$</td>
<td>$67.9\%$</td>
</tr>
<tr>
<td>$S_{i.i}$ $4-\left|\mathbf{f}_{i}^{\mathrm{I}}-\mathbf{f}_{i}^{\mathrm{I}}\right|_{a}-\left|\mathbf{f}_{i}^{\mathrm{T}}-\mathbf{f}_{i}^{\mathrm{T}}\right.$</td>
<td>4</td>
<td>$\sqrt{\text{v}}$</td>
<td>$√$</td>
<td>$83.9\%$</td>
<td>$73.4\%$</td>
</tr>
</tbody>
</table>
Table 2. Comparison among different functions to measure
similarity between image-text pairs.
results are computed using
features extracted from a UGACH [45] model trained on the MIRFlickr dataset. Here we consider four properties: `new’means that the new feature space, learned by the teacher model, is used; `image’and `text' means the corresponding features used, and índiv’means image and text features are used individually. P@ 10000P@ 5000 indicate the accuracy rates among the top 1000075000 retrieved pairs.
<FigureHere>
Figure 2. Knowledge distilled from an unsupervised model (best viewed in color). Compared to the retrieval results in the original feature space, our approach produces more accurate information about the tag of an image and, more importantly, better estimation on the relevance of image-text pairs.
First, we point out that $S_{i,i}\equiv1$ for all $i.$ When $i\neq j$, $S_{i,j}$ takes four vectors, $\mathbf{f}_i^\mathrm{I},\mathbf{f}_i^\mathrm{T},\mathbf{f}_j^\mathrm{I}$ and $\mathbf{f}_j^\mathrm{T}$, into consideration. The design of $S_{i,j}$ can have various forms. For example, it can consider both image and text features so that is the Euclidean distance between two vectors which lies in the range of [0,2] for two normalized vectors. Also, it is possible for $S_{i,j}$ to consider only single-modal informa- $\begin{array}{l}{\mathrm{tion,~e.g.,~S_{i,j}=\left(2-\left|\mathbf{f_{i}^{I}}-\mathbf{f_{j}^{I}}\right|_{2}\right)/2~in~which~only~image}}\\{\mathrm{c.~d~}}\end{array}$ features are used for measuring similarity.
Here, we take several definitions of $S_{i,j}$ into consideration, and compare their performance in finding true-positive pairs. Results are shown in Table 2. We can observe several important properties that are useful for similarity measurement. First, the features trained for cross-modal hashing are indeed better than those without being fine-tuned; Second, measuring similarity in the image feature space is more accurate than that in the text feature space; Third, directly combining image and text similarity into one does not improve accuracy beyond using image similarity alone, though we expect that text features to provide auxiliary information. Motivated by these results, we use image features and text features to retrieve two lists of relevant pairs and then merge them into one. This strategy reports a precision of 76.1% at top-5000 instances, surpassing that using image and text features alone. We fix this setting throughout the remainder of this paper.
we will show in experiments, with the help of oracle annotations, that the updated estimation of S is indeed more accurate in terms of finding relevant pairs.
Note that the estimated S can be used to train either supervised or unsupervised models, with the formulations de- $\hat{\text{tailed above. When S is used for unsupervised learning, the}}$ only effect is to provide a better sampling strategy, so as to increase the portion of true-positive image-text pairs in the chosen training set. This alleviates the risk that the model learns to pull the features of actually irrelevant pairs. When it is used for supervised learning, we are actually creating something from nothing, i.e., guiding a supervised model with the output of an unsupervised model.
The proposed framework, unsupervised knowledge distillation (UKD), works as follows. After the teacher model has been trained, we obtain both $f^{\mathrm{I}}\left(\cdot;\boldsymbol{\theta}^{\mathrm{I}}\right)$ and $f^{\mathrm{T}}\left(\cdot;\boldsymbol{\theta}^{\mathrm{T}}\right)$ for image and text feature embedding, respectively. It remains to determine each element of S. Without loss of generality, we assume that the feature vectors extracted from either modality, $i.e.,\textbf{f}_n^{\mathrm{I}}$ or f$_i^\mathrm{T}$,have a $\ell_2$-norm. This is to ease the following calculations.
## 3.3. Models and Implementation Details
We first illustrate the supervised and unsupervised methods we have used. We take DCMH[17] as an example of supervised learning. Here we utilize the framework of DCMH but modify its architecture for higher accuracy. This model contains two deep neural networks, designed for the image modality and the text modality, respectively. The image modality network consists of 19 layers, the first eighteen layers are the same as those in VGG19 network [31], and the last layer maps features into the Hamming space. For the
4
|
[
[
0.53125,
0.09027099609375,
0.57666015625,
0.0921630859375,
0.576171875,
0.1060791015625,
0.529296875,
0.10418701171875
],
[
0.6650390625,
0.0889892578125,
0.72607421875,
0.091552734375,
0.72607421875,
0.1060791015625,
0.6650390625,
0.10418701171875
],
[
0.4697265625,
0.0921630859375,
0.50244140625,
0.0921630859375,
0.50244140625,
0.10479736328125,
0.4697265625,
0.10479736328125
],
[
0.1650390625,
0.0921630859375,
0.2296142578125,
0.0921630859375,
0.2296142578125,
0.10418701171875,
0.1650390625,
0.10418701171875
],
[
0.60693359375,
0.0921630859375,
0.6357421875,
0.0921630859375,
0.6357421875,
0.10479736328125,
0.60693359375,
0.10479736328125
],
[
0.74169921875,
0.0921630859375,
0.8056640625,
0.0921630859375,
0.8056640625,
0.10418701171875,
0.74169921875,
0.10418701171875
],
[
0.1666259765625,
0.10986328125,
0.2115478515625,
0.1142578125,
0.2091064453125,
0.1307373046875,
0.1641845703125,
0.126953125
],
[
0.6787109375,
0.11236572265625,
0.72607421875,
0.11236572265625,
0.72607421875,
0.1275634765625,
0.6787109375,
0.1275634765625
],
[
0.75634765625,
0.11236572265625,
0.8056640625,
0.11236572265625,
0.8056640625,
0.1275634765625,
0.75634765625,
0.1275634765625
],
[
0.31787109375,
0.11297607421875,
0.35546875,
0.11297607421875,
0.35546875,
0.126953125,
0.31787109375,
0.126953125
],
[
0.221435546875,
0.11297607421875,
0.324462890625,
0.11297607421875,
0.324462890625,
0.1275634765625,
0.221435546875,
0.1275634765625
],
[
0.54736328125,
0.1142578125,
0.5595703125,
0.1142578125,
0.5595703125,
0.1256103515625,
0.54736328125,
0.1256103515625
],
[
0.1666259765625,
0.1363525390625,
0.2132568359375,
0.1363525390625,
0.2132568359375,
0.1514892578125,
0.1666259765625,
0.1514892578125
],
[
0.25244140625,
0.1351318359375,
0.330810546875,
0.1351318359375,
0.330810546875,
0.15283203125,
0.25244140625,
0.15283203125
],
[
0.6787109375,
0.1363525390625,
0.72607421875,
0.1363525390625,
0.72607421875,
0.15087890625,
0.6787109375,
0.15087890625
],
[
0.75634765625,
0.1363525390625,
0.8056640625,
0.1363525390625,
0.8056640625,
0.15087890625,
0.75634765625,
0.15087890625
],
[
0.61279296875,
0.1383056640625,
0.6259765625,
0.1383056640625,
0.6259765625,
0.149658203125,
0.61279296875,
0.149658203125
],
[
0.313720703125,
0.1383056640625,
0.36279296875,
0.1351318359375,
0.363525390625,
0.149658203125,
0.314453125,
0.15283203125
],
[
0.1666259765625,
0.1590576171875,
0.2156982421875,
0.1590576171875,
0.2156982421875,
0.1749267578125,
0.1666259765625,
0.1749267578125
],
[
0.221435546875,
0.1590576171875,
0.3505859375,
0.1590576171875,
0.3505859375,
0.173583984375,
0.221435546875,
0.173583984375
],
[
0.67724609375,
0.1590576171875,
0.72802734375,
0.1590576171875,
0.72802734375,
0.173583984375,
0.67724609375,
0.173583984375
],
[
0.75634765625,
0.1590576171875,
0.8056640625,
0.1590576171875,
0.8056640625,
0.173583984375,
0.75634765625,
0.173583984375
],
[
0.54833984375,
0.1634521484375,
0.5595703125,
0.1634521484375,
0.5595703125,
0.17236328125,
0.54833984375,
0.17236328125
],
[
0.4794921875,
0.1622314453125,
0.49267578125,
0.1622314453125,
0.49267578125,
0.17236328125,
0.4794921875,
0.17236328125
],
[
0.250732421875,
0.18115234375,
0.361083984375,
0.18115234375,
0.361083984375,
0.1988525390625,
0.250732421875,
0.1988525390625
],
[
0.6787109375,
0.1817626953125,
0.72802734375,
0.1817626953125,
0.72802734375,
0.1962890625,
0.6787109375,
0.1962890625
],
[
0.75634765625,
0.1817626953125,
0.8056640625,
0.1817626953125,
0.8056640625,
0.1962890625,
0.75634765625,
0.1962890625
],
[
0.61376953125,
0.1849365234375,
0.62744140625,
0.1849365234375,
0.62744140625,
0.1956787109375,
0.61376953125,
0.1956787109375
],
[
0.1666259765625,
0.18310546875,
0.2132568359375,
0.18310546875,
0.2132568359375,
0.1976318359375,
0.1666259765625,
0.1976318359375
],
[
0.226318359375,
0.1875,
0.2548828125,
0.1875,
0.2548828125,
0.1912841796875,
0.226318359375,
0.1912841796875
],
[
0.473876953125,
0.189453125,
0.48779296875,
0.1787109375,
0.498291015625,
0.1881103515625,
0.486083984375,
0.1976318359375
],
[
0.1666259765625,
0.2052001953125,
0.2115478515625,
0.20703125,
0.2099609375,
0.2215576171875,
0.1650390625,
0.2197265625
],
[
0.6787109375,
0.2052001953125,
0.72802734375,
0.2052001953125,
0.72802734375,
0.2197265625,
0.6787109375,
0.2197265625
],
[
0.75732421875,
0.2052001953125,
0.8037109375,
0.2052001953125,
0.8037109375,
0.2197265625,
0.75732421875,
0.2197265625
],
[
0.314453125,
0.205810546875,
0.410888671875,
0.205810546875,
0.410888671875,
0.2203369140625,
0.314453125,
0.2203369140625
],
[
0.25244140625,
0.205810546875,
0.32275390625,
0.205810546875,
0.32275390625,
0.2203369140625,
0.25244140625,
0.2203369140625
],
[
0.414306640625,
0.20703125,
0.44287109375,
0.20703125,
0.44287109375,
0.2183837890625,
0.414306640625,
0.2183837890625
],
[
0.48046875,
0.2083740234375,
0.491943359375,
0.2083740234375,
0.491943359375,
0.2171630859375,
0.48046875,
0.2171630859375
],
[
0.61376953125,
0.2083740234375,
0.625,
0.2083740234375,
0.625,
0.2171630859375,
0.61376953125,
0.2171630859375
],
[
0.07843017578125,
0.222900390625,
0.89208984375,
0.2235107421875,
0.89208984375,
0.238037109375,
0.07843017578125,
0.2374267578125
],
[
0.080078125,
0.238037109375,
0.89208984375,
0.238037109375,
0.89208984375,
0.251953125,
0.080078125,
0.251953125
],
[
0.07843017578125,
0.251953125,
0.89208984375,
0.251953125,
0.89208984375,
0.26513671875,
0.07843017578125,
0.26513671875
],
[
0.0767822265625,
0.266357421875,
0.86376953125,
0.26513671875,
0.86376953125,
0.279052734375,
0.0767822265625,
0.27978515625
],
[
0.52294921875,
0.304931640625,
0.890625,
0.304931640625,
0.890625,
0.318115234375,
0.52294921875,
0.318115234375
],
[
0.294189453125,
0.3125,
0.431396484375,
0.3125,
0.431396484375,
0.322509765625,
0.294189453125,
0.322509765625
],
[
0.294189453125,
0.3212890625,
0.431396484375,
0.3212890625,
0.431396484375,
0.33203125,
0.294189453125,
0.33203125
],
[
0.50244140625,
0.318115234375,
0.890625,
0.318115234375,
0.890625,
0.333251953125,
0.50244140625,
0.333251953125
],
[
0.295654296875,
0.333251953125,
0.435546875,
0.333251953125,
0.435546875,
0.343505859375,
0.295654296875,
0.343505859375
],
[
0.50244140625,
0.335205078125,
0.8896484375,
0.335205078125,
0.8896484375,
0.348388671875,
0.50244140625,
0.348388671875
],
[
0.295654296875,
0.34521484375,
0.4306640625,
0.3466796875,
0.4306640625,
0.35791015625,
0.295654296875,
0.356689453125
],
[
0.293212890625,
0.35595703125,
0.426513671875,
0.356689453125,
0.426513671875,
0.36865234375,
0.293212890625,
0.367431640625
],
[
0.50244140625,
0.35107421875,
0.89208984375,
0.35107421875,
0.89208984375,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.294189453125,
0.36865234375,
0.443603515625,
0.36865234375,
0.443603515625,
0.37890625,
0.294189453125,
0.37890625
],
[
0.50244140625,
0.367431640625,
0.890625,
0.367431640625,
0.890625,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.293212890625,
0.3818359375,
0.426513671875,
0.380126953125,
0.426513671875,
0.391357421875,
0.293212890625,
0.393310546875
],
[
0.293212890625,
0.392578125,
0.441162109375,
0.392578125,
0.441162109375,
0.40283203125,
0.293212890625,
0.40283203125
],
[
0.50341796875,
0.386962890625,
0.89208984375,
0.386962890625,
0.89208984375,
0.40087890625,
0.50341796875,
0.40087890625
],
[
0.294189453125,
0.404052734375,
0.404296875,
0.404052734375,
0.404296875,
0.4140625,
0.294189453125,
0.4140625
],
[
0.5048828125,
0.404052734375,
0.890625,
0.404052734375,
0.890625,
0.4140625,
0.5048828125,
0.4140625
],
[
0.295654296875,
0.417236328125,
0.404296875,
0.417236328125,
0.404296875,
0.427978515625,
0.295654296875,
0.427978515625
],
[
0.50341796875,
0.417236328125,
0.89208984375,
0.417236328125,
0.89208984375,
0.431884765625,
0.50341796875,
0.431884765625
],
[
0.294189453125,
0.4267578125,
0.435546875,
0.4267578125,
0.435546875,
0.4375,
0.294189453125,
0.4375
],
[
0.291748046875,
0.43505859375,
0.426513671875,
0.436279296875,
0.426513671875,
0.449462890625,
0.291748046875,
0.448974609375
],
[
0.5009765625,
0.434326171875,
0.89453125,
0.434326171875,
0.89453125,
0.455078125,
0.5009765625,
0.455078125
],
[
0.50244140625,
0.455078125,
0.78125,
0.455078125,
0.78125,
0.4697265625,
0.50244140625,
0.4697265625
],
[
0.08087158203125,
0.47021484375,
0.4697265625,
0.47021484375,
0.4697265625,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.51953125,
0.4697265625,
0.89208984375,
0.47021484375,
0.89208984375,
0.48486328125,
0.51953125,
0.484130859375
],
[
0.080078125,
0.484130859375,
0.4697265625,
0.484130859375,
0.4697265625,
0.49755859375,
0.080078125,
0.49755859375
],
[
0.50244140625,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.080078125,
0.498779296875,
0.4697265625,
0.498779296875,
0.4697265625,
0.51220703125,
0.080078125,
0.51220703125
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.08087158203125,
0.5146484375,
0.46728515625,
0.5146484375,
0.46728515625,
0.5244140625,
0.08087158203125,
0.5244140625
],
[
0.5009765625,
0.5166015625,
0.89208984375,
0.51513671875,
0.89208984375,
0.52978515625,
0.5009765625,
0.53076171875
],
[
0.07843017578125,
0.5263671875,
0.295654296875,
0.5263671875,
0.295654296875,
0.54052734375,
0.07843017578125,
0.54052734375
],
[
0.50244140625,
0.5322265625,
0.89208984375,
0.5322265625,
0.89208984375,
0.54541015625,
0.50244140625,
0.54541015625
],
[
0.50244140625,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.50244140625,
0.56005859375
],
[
0.5009765625,
0.56103515625,
0.890625,
0.56103515625,
0.890625,
0.57470703125,
0.5009765625,
0.57470703125
],
[
0.080078125,
0.5673828125,
0.46728515625,
0.5673828125,
0.46728515625,
0.58056640625,
0.080078125,
0.58056640625
],
[
0.50244140625,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.50244140625,
0.59033203125
],
[
0.07843017578125,
0.58349609375,
0.4697265625,
0.58349609375,
0.4697265625,
0.5966796875,
0.07843017578125,
0.5966796875
],
[
0.50244140625,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.07843017578125,
0.5986328125,
0.35693359375,
0.5986328125,
0.35693359375,
0.6123046875,
0.07843017578125,
0.6123046875
],
[
0.5009765625,
0.60791015625,
0.89208984375,
0.60693359375,
0.89208984375,
0.62060546875,
0.5009765625,
0.6220703125
],
[
0.098876953125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.098876953125,
0.62744140625
],
[
0.50244140625,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.63720703125,
0.89208984375,
0.63720703125,
0.89208984375,
0.65087890625,
0.50244140625,
0.65087890625
],
[
0.080078125,
0.64453125,
0.4697265625,
0.64453125,
0.4697265625,
0.65771484375,
0.080078125,
0.65771484375
],
[
0.50244140625,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.50244140625,
0.66650390625
],
[
0.080078125,
0.66015625,
0.4697265625,
0.66015625,
0.4697265625,
0.673828125,
0.080078125,
0.673828125
],
[
0.080078125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.080078125,
0.6875
],
[
0.50244140625,
0.66845703125,
0.89208984375,
0.66845703125,
0.89208984375,
0.681640625,
0.50244140625,
0.681640625
],
[
0.50244140625,
0.68310546875,
0.89208984375,
0.68310546875,
0.89208984375,
0.6962890625,
0.50244140625,
0.6962890625
],
[
0.08087158203125,
0.69140625,
0.468017578125,
0.69140625,
0.468017578125,
0.7021484375,
0.08087158203125,
0.7021484375
],
[
0.50244140625,
0.69873046875,
0.89208984375,
0.69873046875,
0.89208984375,
0.7119140625,
0.50244140625,
0.7119140625
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.71337890625,
0.66259765625,
0.71337890625,
0.66259765625,
0.7265625,
0.50244140625,
0.7265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.07843017578125,
0.73486328125,
0.468017578125,
0.734375,
0.468017578125,
0.74853515625,
0.07843017578125,
0.75
],
[
0.50341796875,
0.7412109375,
0.810546875,
0.7412109375,
0.810546875,
0.7548828125,
0.50341796875,
0.7548828125
],
[
0.080078125,
0.75048828125,
0.35693359375,
0.75048828125,
0.35693359375,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.098876953125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.098876953125,
0.77978515625
],
[
0.521484375,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
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,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.07843017578125,
0.79736328125,
0.468017578125,
0.79736328125,
0.468017578125,
0.8125,
0.07843017578125,
0.8125
],
[
0.5009765625,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.50244140625,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08251953125,
0.8193359375,
0.4697265625,
0.8212890625,
0.4697265625,
0.8388671875,
0.08251953125,
0.8369140625
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84033203125,
0.4697265625,
0.84033203125,
0.4697265625,
0.85400390625,
0.080078125,
0.85400390625
],
[
0.50244140625,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.46728515625,
0.85791015625,
0.46728515625,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.35302734375,
0.88623046875,
0.35302734375,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.5009765625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5009765625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9345703125,
0.48046875,
0.9345703125
]
] |
[
"image",
"P@1000",
"new",
"Function",
"text",
"P@5000",
"Si.j",
"74.6%",
"64.4%",
"2)/2",
"2 - /v) -v)l2",
"V",
"Si,j =",
"IvT-VTI,",
"57.1%",
"55.6%",
"V",
"12)/2",
"Si.j =",
"(2-fl-f)l)/2",
"84.6%",
"74.6%",
"V",
"V",
"fT-fTI2) /2",
"75.9%",
"67.9%",
"V",
"Si,j =",
"",
"L",
"Sij:",
"83.9%",
"73.4%",
"-|fT-fTl2",
"fl-fil?",
"0./4",
"Y",
"y",
"Table 2. Comparison among different functions to measure the similarity between image-text pairs.All the results are computed using",
"features extracted from a UGACH [45] model trained on the MIRFlickr dataset. Here we consider four properties: 'new' means that the",
"new feature space, learned by the teacher model, is used; image' and 'text' means the corresponding features used, and indiv' means",
"image and textfeatures are used individually.P@1000/P@5000 indicate the accuracy rates among the top 1000/5000 retrieved pairs.",
"First, we point out that Si= 1 for all i. When i + i,",
"ball, bird, people,dog,water",
"dog, baby,room, light,white",
"Sii takes four vectors, fi, fT, ff and fT, into considera-",
"kite,plane,clouds,man,blue",
"tion. The design of Si.i can have various forms. For ex-",
"grassskyclouds,dog.gre",
"huskygray,leaves,tree,sun",
"ample, it can consider both image and text features so that",
"night,sky,dogs,oldman,moon",
"Si.j =(4-fl -filz -|fT-fTl2) /4, where |fi-f2l2",
"house,women,sofa,cat.",
"keyboard.phone,cats,toy,pens",
"is the Euclidean distance between two vectors which lies",
"bag.clock,papers,man",
"inthe range of[o.2for twonormalized vectors.Also,it",
"wooden,cat,light,room",
"is possible for Si., to consider only single-modal informa-",
"floor,chair,person,cigaret,ca",
"kittygrass,by,yllow,",
"tion, e.g., Si,j = (2-f} - fl2) /2 in which only image",
"features are used for measuring similarity.",
"Figure 2. Knowledge distilled from an unsupervised model (best",
"Here, we take several definitions of Si, into considera-",
"viewed in color). Compared to the retrieval results in the original",
"tion, and compare their performance in finding true-positive",
"feature space, our approach produces more accurate information",
"pairs. Results are shown in Table 2. We can observe several",
"aboutthetagofanimageand,moreimportantly,betterestimation",
"important properties that are useful for similarity measure-",
"on the relevance of image-text pairs.",
"ment. First, the features trained for cross-modal hashing",
"are indeed better than those without being fine-tuned; Sec-",
"ond, measuring similarity in the image feature space is more",
"we will show in experiments, with the help of oracle an-",
"accurate than that in the text feature space; Third, directly",
"notations, that the updated estimation of S is indeed more",
"combining image and text similarity into one does not im-",
"accurate in terms of finding relevant pairs.",
"prove accuracy beyond using image similarity alone, though",
"Note that the estimated S can be used to train either su-",
"we expect that text features to provide auxiliary informa-",
"pervised or unsupervised models, with the formulations de-",
"tion. Motivated by these results, we use image features and",
"tailed above. When S is used for unsupervised learning, the",
"text features to retrieve two lists of relevant pairs and then",
"only effect is to provide a better sampling strategy, so as to",
"increase the portion of true-positive image-text pairs in the",
"merge them into one. This strategy reports a precision of",
"76.1% at top-5000 instances, surpassing that using image",
"chosen training set. This alleviates the risk that the model",
"and text features alone. We fix this setting throughout the",
"learns to pull the features of actually irrelevant pairs. When",
"remainder of this paper.",
"it is used for supervised learning, we are actually creating",
"something from nothing, i.e., guiding a supervised model",
"3.3. Models and Implementation Details",
"with the output of an unsupervised model.",
"The proposed framework, unsupervised knowledge",
"We first illustrate the supervised and unsupervised meth-",
"distillation (UKD), works as follows. After the teacher",
"ods we have used. We take DCMH[17] as an example of su-",
"model has been trained, we obtain both fI(,α') and",
"pervised learning. Here we utilize the framework of DCMH",
"but modify its architecture for higher accuracy. This model",
"fT (-; oT) for image and text feature embedding, respec-",
"contains two deep neural networks, designed for the image",
"tively. It remains to determine each element of S. Without",
"modality and the text modality, respectively.The image",
"loss of generality, we assume that the feature vectors ex-",
"modalitynetwork consists of 19 layers, the first eighteen",
"tracted from either modality, i.e., f or fT, have a l2-norm.",
"layers are the same as those in VGG19 network [31], and the",
"This is to ease the following calculations.",
"last layer maps features into the Hamming space. For the",
"4"
] |
{
"arxiv_id": "2004.00280",
"language": "en",
"page_idx": 3
}
|
||
2004.00280-en-4
|
2004.00280-en-4.jpg
|
text modality, a multi-scale fusion model from SSAH[20] which consists of multiple average pooling layers and a 1×1 convolutional layer is used to extract the text features. Then, a hash layer follows to map the text features into the Hamming space.
On the other hand, we investigate UGACH [45], a representative unsupervised learning method as the teacher model. It consists of a generative module and a discriminative module. The discriminator receives the data selected by the generator as negative instances, and take the data sampled using S as positive instances. Then a triplet loss is used to optimize to obtain better discriminate ability for the discriminator. Both the generative and discriminative modules have a two-pathway architecture, each of which has two fully-connected layers. We set the dimension of representation layer to 4096 in our experiments. The dimension of the hash layer is same as the hash code length.
For the supervised model, we take the raw pixels as inputs. In pre-processing, we resize all images into $256\times256$ and crop a $224\times224$ patches randomly. We select relevant instances for the student model by using the teacher model with the highest precision (128 bits in all experiments). We set the number of relevant instances to be 10,000 for the supervised student model, and 20 for the unsupervised student model. We train our approach in a batch-based manner and set the batch size to 256. We train the model using an SGD optimizer with a weight decay of 0.01. For the compared methods, we apply the same implementations as provided in the original work.
## 3.4. Relationship to Previous Work
Our method is related to knowledge distillation [29, 43, 33], which was proposed to extract knowledge from a teacher model to assist training a student model. Hinton $et$ $al.[15]$ suggested that there should be some‘dark knowledge' that can be propagated during this process. Recently, many efforts were made to study what the dark knowledge is [41, 40], and/or how to efficiently take advantage of such knowledge[11,42,35,2]. In particular, DarkRank[5] distilled knowledge for deep metric learning by matching two probability distributions over ranking, while our approach utilized knowledge by selecting relevant instances. On the other hand, both [42] and [27] transferred knowledge to improve the student models by designing a distillation loss, while our approach enables guiding a supervised method by an unsupervised method, in which no extra loss is used.
We also notice the connection between our approach and the self-learning algorithms for semi-supervised learning, $e.g.$, medical image analysis[48]. The shared idea is to start with a small part of labeled data (in our case, labeled image-text pairs) and try to explore the unlabeled part (in our case, other image-text pairs with unknown relevance), but the methods to gain additional supervision are differ-
ent. Also, the idea that `training a stronger model at the second time’is related to the coarse-to-fine learning approaches [14, 49] which often adopted iteration for larger improvements.
Our approach shares the same idea with some prior work that guided a supervised model with the output of an unsupervised model. DeepCluster[3] groups the features with a standard clustering algorithm and uses the subsequent assignments as supervision to update the weights of the network. Gomez et al. [12] performed self-supervised learning of visual features by mining a large scale corpus of multimodal (text and image) documents. Differently, our approach makes use of teacher-student optimization to combine the supervised and unsupervised models. Experiment results show the effectiveness of knowledge distillation.
## $4. \textbf{Experiments}$
## 4.1. Datasets, Evaluation, and Baselines
We evaluate our approach on two benchmark datasets: MIRFlickr and NUS-WIDE. MIRFlickr-25K[16] consists of 25,000 images downloaded from Flickr. Each image is associated with text tags and annotated with at least one among 24 pre-defined categories. Following UGACH [45], we use 20,015 image-text pairs in our experiments, where 2,000 are preserved as the query set and the rest are used for retrieving. We represent each image by a 4096-dimensional feature vector, extracted from a pre-trained VGGNet[31] of 19 layers, and each text by a 1386-dimensional bag-ofwords features.
NUS-WIDE[6] is much larger than MIRFlickr, which contains 269,498 images and the associated text tags from Flickr. It defined 81 categories, but there are considerable overlaps among them. Still, following UGACH[45], 10 largest categories and the corresponding 186,577 imagetext pairs are used in the experiments. We preserve 1% of data as the query database and use the rest as the retrieval set. Each image is represented by a 4096-dimensional feature vector extracted from the same VGGNet, and each text by a 1000-dimensional bag-of-words vector.
Following the convention, we adopt the mean Average Precision (mAP) criterion to evaluate the retrieval performance of all methods. The mAP score is computed as the mean value of the average precision scores for all queries.
We compare our approach against 9 previous methods. 4 of them used additional supervision (CMSSH[1], SCM[44], DCMH[17], and SSAH[20]), and while 5 others (CVH[19], PDH[28], CMFH[9], and CCQ[26]), and UGACH [45]) did not. Following our direct baseline, UGACH, we use a 19-layer VGGNet [31] pre-trained on the ImageNet dataset [30] to extract deep features and, for a fair comparison, use them to replace the features used in other baselines, including those using handcrafted features.
5
|
[
[
0.07843017578125,
0.091552734375,
0.468017578125,
0.09027099609375,
0.468017578125,
0.10479736328125,
0.07843017578125,
0.1060791015625
],
[
0.5009765625,
0.0921630859375,
0.89208984375,
0.0921630859375,
0.89208984375,
0.1060791015625,
0.5009765625,
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
],
[
0.50341796875,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.50341796875,
0.136962890625
],
[
0.080078125,
0.136962890625,
0.46728515625,
0.136962890625,
0.46728515625,
0.15087890625,
0.080078125,
0.15087890625
],
[
0.50341796875,
0.1395263671875,
0.60107421875,
0.1395263671875,
0.60107421875,
0.15087890625,
0.50341796875,
0.15087890625
],
[
0.08087158203125,
0.1546630859375,
0.16015625,
0.1546630859375,
0.16015625,
0.1666259765625,
0.08087158203125,
0.1666259765625
],
[
0.521484375,
0.15283203125,
0.89208984375,
0.1514892578125,
0.89208984375,
0.1666259765625,
0.521484375,
0.1673583984375
],
[
0.098876953125,
0.1685791015625,
0.46728515625,
0.1685791015625,
0.46728515625,
0.1817626953125,
0.098876953125,
0.1817626953125
],
[
0.50341796875,
0.1685791015625,
0.8896484375,
0.1685791015625,
0.8896484375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.080078125,
0.184326171875,
0.4697265625,
0.184326171875,
0.4697265625,
0.1976318359375,
0.080078125,
0.1976318359375
],
[
0.50244140625,
0.184326171875,
0.89208984375,
0.184326171875,
0.89208984375,
0.1976318359375,
0.50244140625,
0.1976318359375
],
[
0.08087158203125,
0.2008056640625,
0.46728515625,
0.2008056640625,
0.46728515625,
0.2108154296875,
0.08087158203125,
0.2108154296875
],
[
0.50341796875,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.21337890625,
0.50341796875,
0.21337890625
],
[
0.080078125,
0.2139892578125,
0.468017578125,
0.2139892578125,
0.468017578125,
0.2279052734375,
0.080078125,
0.2279052734375
],
[
0.50244140625,
0.21533203125,
0.890625,
0.21533203125,
0.890625,
0.2291259765625,
0.50244140625,
0.2291259765625
],
[
0.080078125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.24365234375,
0.080078125,
0.24365234375
],
[
0.5009765625,
0.2291259765625,
0.89208984375,
0.2298583984375,
0.89208984375,
0.24365234375,
0.5009765625,
0.242431640625
],
[
0.08087158203125,
0.246826171875,
0.468017578125,
0.246826171875,
0.468017578125,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.50244140625,
0.2442626953125,
0.890625,
0.2442626953125,
0.890625,
0.25830078125,
0.50244140625,
0.25830078125
],
[
0.080078125,
0.2587890625,
0.4697265625,
0.2587890625,
0.4697265625,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.5009765625,
0.25830078125,
0.890625,
0.2587890625,
0.890625,
0.2734375,
0.5009765625,
0.272705078125
],
[
0.08087158203125,
0.274658203125,
0.468017578125,
0.274658203125,
0.468017578125,
0.28857421875,
0.08087158203125,
0.28857421875
],
[
0.50341796875,
0.274658203125,
0.89208984375,
0.274658203125,
0.89208984375,
0.28857421875,
0.50341796875,
0.28857421875
],
[
0.07843017578125,
0.2890625,
0.4697265625,
0.28857421875,
0.4697265625,
0.302978515625,
0.07843017578125,
0.3037109375
],
[
0.50244140625,
0.2890625,
0.89208984375,
0.2890625,
0.89208984375,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.304931640625,
0.4697265625,
0.304931640625,
0.4697265625,
0.318115234375,
0.080078125,
0.318115234375
],
[
0.50244140625,
0.304931640625,
0.86767578125,
0.304931640625,
0.86767578125,
0.318115234375,
0.50244140625,
0.318115234375
],
[
0.080078125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.080078125,
0.333251953125
],
[
0.080078125,
0.335205078125,
0.398681640625,
0.335205078125,
0.398681640625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.50341796875,
0.3359375,
0.63134765625,
0.3359375,
0.63134765625,
0.35107421875,
0.50341796875,
0.35107421875
],
[
0.098876953125,
0.35107421875,
0.468017578125,
0.35107421875,
0.468017578125,
0.3642578125,
0.098876953125,
0.3642578125
],
[
0.080078125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.080078125,
0.37890625
],
[
0.50341796875,
0.361083984375,
0.80810546875,
0.361083984375,
0.80810546875,
0.375732421875,
0.50341796875,
0.375732421875
],
[
0.080078125,
0.380615234375,
0.4697265625,
0.380615234375,
0.4697265625,
0.39453125,
0.080078125,
0.39453125
],
[
0.51953125,
0.3837890625,
0.890625,
0.385009765625,
0.890625,
0.399658203125,
0.51953125,
0.3984375
],
[
0.08087158203125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.50244140625,
0.40087890625,
0.890625,
0.40087890625,
0.890625,
0.4140625,
0.50244140625,
0.4140625
],
[
0.080078125,
0.410888671875,
0.4697265625,
0.410888671875,
0.4697265625,
0.4248046875,
0.080078125,
0.4248046875
],
[
0.50244140625,
0.416748046875,
0.89208984375,
0.416748046875,
0.89208984375,
0.429931640625,
0.50244140625,
0.429931640625
],
[
0.080078125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.50244140625,
0.43115234375,
0.89208984375,
0.43115234375,
0.89208984375,
0.4443359375,
0.50244140625,
0.4443359375
],
[
0.080078125,
0.44140625,
0.4697265625,
0.44140625,
0.4697265625,
0.455078125,
0.080078125,
0.455078125
],
[
0.50244140625,
0.4462890625,
0.89208984375,
0.4462890625,
0.89208984375,
0.460205078125,
0.50244140625,
0.460205078125
],
[
0.080078125,
0.45703125,
0.4697265625,
0.45703125,
0.4697265625,
0.47021484375,
0.080078125,
0.47021484375
],
[
0.50341796875,
0.4609375,
0.89208984375,
0.4609375,
0.89208984375,
0.474853515625,
0.50341796875,
0.474853515625
],
[
0.07843017578125,
0.47021484375,
0.468017578125,
0.4716796875,
0.468017578125,
0.486083984375,
0.07843017578125,
0.48486328125
],
[
0.50244140625,
0.4765625,
0.890625,
0.4765625,
0.890625,
0.490478515625,
0.50244140625,
0.490478515625
],
[
0.080078125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.080078125,
0.50048828125
],
[
0.50341796875,
0.4912109375,
0.890625,
0.4912109375,
0.890625,
0.505859375,
0.50341796875,
0.505859375
],
[
0.080078125,
0.50244140625,
0.4697265625,
0.50244140625,
0.4697265625,
0.5166015625,
0.080078125,
0.5166015625
],
[
0.50341796875,
0.5068359375,
0.890625,
0.5068359375,
0.890625,
0.52099609375,
0.50341796875,
0.52099609375
],
[
0.07843017578125,
0.517578125,
0.2156982421875,
0.517578125,
0.2156982421875,
0.53076171875,
0.07843017578125,
0.53076171875
],
[
0.50244140625,
0.521484375,
0.890625,
0.521484375,
0.890625,
0.53515625,
0.50244140625,
0.53515625
],
[
0.080078125,
0.54443359375,
0.348876953125,
0.54443359375,
0.348876953125,
0.55810546875,
0.080078125,
0.55810546875
],
[
0.5048828125,
0.53857421875,
0.60546875,
0.53857421875,
0.60546875,
0.548828125,
0.5048828125,
0.548828125
],
[
0.521484375,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.098876953125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.098876953125,
0.58203125
],
[
0.50244140625,
0.568359375,
0.89208984375,
0.568359375,
0.89208984375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.07843017578125,
0.58203125,
0.468017578125,
0.58349609375,
0.468017578125,
0.59765625,
0.07843017578125,
0.5966796875
],
[
0.50244140625,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.080078125,
0.5986328125,
0.4697265625,
0.5986328125,
0.4697265625,
0.6123046875,
0.080078125,
0.6123046875
],
[
0.50244140625,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.08087158203125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.08087158203125,
0.62744140625
],
[
0.5009765625,
0.6123046875,
0.890625,
0.61279296875,
0.890625,
0.62744140625,
0.5009765625,
0.626953125
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.6298828125,
0.468017578125,
0.64306640625,
0.080078125,
0.64306640625
],
[
0.50244140625,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.50244140625,
0.642578125
],
[
0.08087158203125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.65576171875,
0.08087158203125,
0.65576171875
],
[
0.50244140625,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.08087158203125,
0.6611328125,
0.468017578125,
0.6611328125,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
0.50244140625,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.50244140625,
0.67236328125
],
[
0.080078125,
0.6748046875,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.50244140625,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50244140625,
0.6875
],
[
0.07843017578125,
0.68798828125,
0.4697265625,
0.689453125,
0.4697265625,
0.7041015625,
0.07843017578125,
0.70263671875
],
[
0.5009765625,
0.68798828125,
0.79736328125,
0.6875,
0.79736328125,
0.7021484375,
0.5009765625,
0.70263671875
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.521484375,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.521484375,
0.71826171875
],
[
0.080078125,
0.7197265625,
0.471435546875,
0.7197265625,
0.471435546875,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.7216796875,
0.8896484375,
0.7216796875,
0.8896484375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.5048828125,
0.736328125,
0.8896484375,
0.736328125,
0.8896484375,
0.7470703125,
0.5048828125,
0.7470703125
],
[
0.08087158203125,
0.75,
0.468017578125,
0.75,
0.468017578125,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.50244140625,
0.75048828125,
0.88232421875,
0.75048828125,
0.88232421875,
0.7646484375,
0.50244140625,
0.7646484375
],
[
0.08087158203125,
0.76513671875,
0.4697265625,
0.76513671875,
0.4697265625,
0.77880859375,
0.08087158203125,
0.77880859375
],
[
0.521484375,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.521484375,
0.77978515625
],
[
0.080078125,
0.77978515625,
0.465576171875,
0.77978515625,
0.465576171875,
0.79345703125,
0.080078125,
0.79345703125
],
[
0.5009765625,
0.77978515625,
0.890625,
0.78076171875,
0.890625,
0.79541015625,
0.5009765625,
0.79443359375
],
[
0.098876953125,
0.79443359375,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.098876953125,
0.8095703125
],
[
0.50341796875,
0.79736328125,
0.890625,
0.79736328125,
0.890625,
0.80810546875,
0.50341796875,
0.80810546875
],
[
0.080078125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5009765625,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.50244140625,
0.8720703125,
0.8935546875,
0.8720703125,
0.8935546875,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"text modality, a multi-scale fusion model from SSAH[20]",
"ent. Also, the idea that 'training a stronger model at the",
"which consists of multiple average pooling layers and a 1 x 1",
"second time' is related to the coarse-to-fine learning ap-",
"convolutional layer is used to extract the text features. Then,",
"proaches [14, 49] which often adopted iteration for larger",
"a hash layer follows to map the text features into the Ham-",
"improvements.",
"ming space.",
"Ourapproachsharesthe same idea with someprior work",
"On the other hand, we investigate UGACH [45], a rep-",
"that guided a supervised model with the output of an unsu-",
"resentative unsupervised learning method as the teacher",
"pervised model. DeepCluster [3] groups the features with",
"model.It consists of agenerative moduleand a discrimina-",
"a standard clustering algorithm and uses the subsequent as-",
"tive module. The discriminator receives the data selected by",
"signments as supervision to update the weights of the net-",
"the generator as negative instances, and take the data sam-",
"work.Gomez etal.[12] performed self-supervised learning",
"pled using S as positive instances.Then a triplet loss is",
"of visual features by mining a large scale corpus of multi-",
"used to optimize to obtain better discriminate ability for the",
"modal (text and image) documents. Differently, our ap-",
"discriminator. Both the generative and discriminative mod-",
"proach makes use of teacher-student optimization to com-",
"ules have a two-pathway architecture, each of which has",
"bine the supervised and unsupervised models. Experiment",
"two fully-connected layers. We set the dimension of repre-",
"results show the effectiveness of knowledge distillation.",
"sentation layer to 4096 in our experiments.The dimension",
"of the hash layer is same as the hash code length",
"4.Experiments",
"For the supervised model, we take the raw pixels as in-",
"puts. In pre-processing, we resize all images into 256 x 256",
"4.1. Datasets, Evaluation, and Baselines",
"and crop a 224 × 224 patches randomly. We select relevant",
"We evaluate our approach on two benchmark datasets:",
"instancesforthestudentmodelbyusingtheteachermodel",
"MIRFlickr and NUS-WIDE. MIRFlickr-25K [16] consists",
"with the highest precision (128 bits in all experiments). We",
"of 25,000 images downloaded from Flickr. Each image is",
"set the number of relevant instances to be 10,000 for the su-",
"associated with text tags and annotated with at least one",
"pervised student model, and 20 for the unsupervised student",
"among 24 pre-defined categories. Following UGACH [45],",
"model. We train our approach in a batch-based manner and",
"we use 20,015 image-text pairs in our experiments, where",
"set the batch size to 256. We train the model using an SGD",
"2,000 are preserved as the query set and the rest are used for",
"optimizer with a weight decay of 0.01. For the compared",
"retrieving. We represent each image by a 4096-dimensional",
"methods, we apply the same implementations as provided",
"feature vector, extracted from a pre-trained VGGNet [31]",
"in the original work.",
"of 19 layers, and each text by a 1386-dimensional bag-of-",
"3.4. Relationship to Previous Work",
"words features.",
"NUS-WIDE [6] is much larger than MIRFlickr, which",
"Our method is related to knowledge distillation [29,",
"contains 269,498 images and the associated text tags from",
"43, 33], which was proposed to extract knowledge from a",
"Flickr. It defined 81 categories, but there are considerable",
"teacher model to assist training a student model. Hinton et",
"overlaps among them. Still, following UGACH [45], 10",
"al. [15] suggested that there should be some 'dark knowl-",
"largest categories and the corresponding 186,577 image-",
"edge' that can be propagated during this process. Recently,",
"text pairs are used in the experiments. We preserve 1% of",
"manyefforts weremadeto studywhat thedarkknowledge",
"data as the query database and use the rest as the retrieval",
"is[41,40l,and/orhowtoefficientlytakeadvantageof such",
"set. Each image is represented by a 4096-dimensional fea-",
"knowledge [11, 42, 35, 2]. In particular, DarkRank [5] dis-",
"ture vector extracted from the same VGGNet, and each text",
"tilled knowledge for deep metric learning by matching two",
"by a 1000-dimensional bag-of-words vector.",
"probability distributions over ranking, while our approach",
"Following the convention, we adopt the mean Average",
"utilized knowledge by selecting relevant instances. On the",
"Precision (mAP)criterion to evaluate the retrieval perfor-",
"other hand, both [42] and [27] transferred knowledge to im-",
"manceof allmethods.ThemAP scoreiscomputedasthe",
"prove the student models by designing a distillation loss,",
"mean value of the average precision scores for all queries.",
"while our approach enables guiding a supervised method",
"We compare our approach against 9 previous meth-",
"by an unsupervised method, in which no extra loss is used.",
"ods. 4 of them used additional supervision (CMSSH [1],",
"We also notice the connection between our approach and",
"SCM [44], DCMH [17], and SSAH [20D), and while 5",
"the self-learning algorithms for semi-supervised learning,",
"others (CVH [19], PDH [28], CMFH [9], and CCQ [26]],",
"e.g., medical image analysis [48]. The shared idea is to",
"and UGACH [45] did not. Following our direct baseline,",
"start with a small part of labeled data (in our case, labeled",
"UGACH, we use a 19-layer VGGNet [31] pre-trained on",
"image-text pairs) and try to explore the unlabeled part (in",
"the ImageNet dataset [30] to extract deep features and, for",
"our case, other image-text pairs with unknown relevance),",
"a fair comparison, use them to replace the features used in",
"but the methods to gain additional supervision are differ-",
"other baselines, including those using handcrafted features.",
"5"
] |
{
"arxiv_id": "2004.00280",
"language": "en",
"page_idx": 4
}
|
||
2307.08476-en-1
|
2307.08476-en-1.jpg
|
action classes on FineGym dataset[62]. It can be seen that the 2D poses can accurately capture human skeletons and motion details.
Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton sequence is more suitable to distinguish similar actions with subtle differences than the RGB data. To capture discriminative spatial-temporal motion patterns, most of the existing skeleton-based action recognition methods [11,88,5] are fully supervised and usually require large amounts of labeled data for training elaborate models, which is timeconsuming and labor-intensive. To mitigate the problem of limited labeled training data, self-supervised skeletonbased action recognition methods [32, 18,68] have attracted increasing attention recently. Some contrastive learning methods [32, 18] adopted data augmentation to generate pairs of positive and negative samples, but they rely heavily on the number of contrastive pairs. With the popularity of the encoder-decoder [67,51], some methods [94,68] reconstructed the masked skeleton sequence by link reconstruction to encourage the topological closeness following the paradigm of graph encoder-decoder. However, these methods are usually good at link prediction and node clustering but are unsatisfactory in node and graph classifications. For accurate action recognition, the fine-grained dependencies among different skeleton joints (i.e., graph classifications) are essential. Therefore, previous self-supervised learningbased methods tend to ignore the fine-grained dependencies among different skeleton joints, which restricts the generalization of self-supervised skeleton representation. As shown in Fig. 1 (b)-(c), the arm joints and edges are essential to discriminate between these two similar actions. Different from the randomly masking strategy of MAE[20], our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Our SkeletonMAE utilizes prior human topology knowledge to guide the reconstruction of the masked skeleton joints and edges to achieve a comprehensive perception of the joints, topology, and action.
To address the aforementioned challenges, we propose an efficient skeleton sequence representation learning framework, named Skeleton Sequence Learning (SSL). To fully discover the fine-grained dependencies among different skeleton joints, we build a novel asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton joint sequences in Graph Convolutional Network (GCN).The SkeletonMAE aims to reconstruct the masked human skeleton joints and edges based on prior human topology knowledge so that it can infer the underlying topology of the joints and obtain a comprehensive perception of human action. To learn discriminative spatial-temporal skeleton representation, the pre-trained SkeletonMAE encoder is integrated
with the Spatial-Temporal Representation Learning (STRL) module to learn spatial-temporal dependencies. Finally, the SSL is fine-tuned on action recognition datasets. Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 show that our SSL generalizes well across different datasets and outperforms the state-of-the-art methods significantly. Our contributions are summarized as follows:
· To comprehensively capture human pose and obtain discriminative skeleton sequence representation, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology.
·To learn comprehensive spatial-temporal dependencies for skeleton sequence, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL),which integrates the pre-trained SkeletonMAE encoder with the SpatialTemporal Representation Learning (STRL) module.
$\cdot\hat{\text{Extensive experimental results on FineGym, Div-}}$
ing48, NTU 60 and NTU 120 datasets show that our SSL outperforms the state-of-the-art selfsupervised skeleton-based action recognition methods and achieves comparable performance compared with the state-of-the-art fully supervised methods.
## $2. \textbf{Related Work}$
Action Recognition. One of the most challenging tasks for action recognition is to distinguish similar actions from subtle differences. Recently, some challenging action recognition datasets like FineGym [62], Diving48 [35], NTU RGB+D 60[61] and NTU RGB+D 120[38] are proposed. These datasets contain a large number of challenging actions that require discriminative and finegrained action representation learning. For example, in FineGym [62], an action is divided into action units, subactions, or phases, and the model is required to distinguish between“split leap with 1 turn”and“switch leap with 1 turn". The higher inter-class similarity and a new level of granularity in the fine-grained setting make it a challenging task, which makes coarse-grained backbones and methods [14,2,74,83] struggle to overcome. To tackle the more challenging fine-grained action recognition task, most of the existing works [53,34] are fully supervised and consider fine-grained actions as distinct categories and supervise the model to learn action semantics. However, collecting and labeling these fine-grained actions is timeconsuming and labor-intensive, which limits the generalization of a well-trained model to different datasets. To utilize unlabeled data, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE.
|
[
[
0.080078125,
0.091552734375,
0.4697265625,
0.091552734375,
0.4697265625,
0.1060791015625,
0.080078125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.080078125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.080078125,
0.121826171875
],
[
0.50341796875,
0.10919189453125,
0.890625,
0.10919189453125,
0.890625,
0.11932373046875,
0.50341796875,
0.11932373046875
],
[
0.08087158203125,
0.12371826171875,
0.1805419921875,
0.12371826171875,
0.1805419921875,
0.1351318359375,
0.08087158203125,
0.1351318359375
],
[
0.5009765625,
0.121826171875,
0.89208984375,
0.12249755859375,
0.89208984375,
0.136962890625,
0.5009765625,
0.1363525390625
],
[
0.09722900390625,
0.141357421875,
0.468017578125,
0.141357421875,
0.468017578125,
0.1558837890625,
0.09722900390625,
0.1558837890625
],
[
0.50244140625,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.1572265625,
0.4697265625,
0.1572265625,
0.4697265625,
0.17041015625,
0.080078125,
0.17041015625
],
[
0.50341796875,
0.15283203125,
0.890625,
0.15283203125,
0.890625,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.17236328125,
0.4697265625,
0.17236328125,
0.4697265625,
0.186279296875,
0.080078125,
0.186279296875
],
[
0.50341796875,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.50341796875,
0.18115234375
],
[
0.07843017578125,
0.186279296875,
0.4697265625,
0.1875,
0.4697265625,
0.2020263671875,
0.07843017578125,
0.2008056640625
],
[
0.50244140625,
0.18310546875,
0.89208984375,
0.18310546875,
0.89208984375,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.07843017578125,
0.2020263671875,
0.468017578125,
0.20263671875,
0.468017578125,
0.216552734375,
0.07843017578125,
0.21533203125
],
[
0.521484375,
0.210205078125,
0.89208984375,
0.210205078125,
0.89208984375,
0.2235107421875,
0.521484375,
0.2235107421875
],
[
0.080078125,
0.2183837890625,
0.468017578125,
0.2183837890625,
0.468017578125,
0.231689453125,
0.080078125,
0.231689453125
],
[
0.53515625,
0.2259521484375,
0.89208984375,
0.2259521484375,
0.89208984375,
0.2392578125,
0.53515625,
0.2392578125
],
[
0.080078125,
0.23291015625,
0.4697265625,
0.23291015625,
0.4697265625,
0.246826171875,
0.080078125,
0.246826171875
],
[
0.53515625,
0.2412109375,
0.89208984375,
0.2412109375,
0.89208984375,
0.255126953125,
0.53515625,
0.255126953125
],
[
0.080078125,
0.2474365234375,
0.46728515625,
0.2474365234375,
0.46728515625,
0.261474609375,
0.080078125,
0.261474609375
],
[
0.53515625,
0.255615234375,
0.89208984375,
0.255615234375,
0.89208984375,
0.26953125,
0.53515625,
0.26953125
],
[
0.080078125,
0.26318359375,
0.4697265625,
0.26318359375,
0.4697265625,
0.276611328125,
0.080078125,
0.276611328125
],
[
0.53515625,
0.271484375,
0.89208984375,
0.271484375,
0.89208984375,
0.285400390625,
0.53515625,
0.285400390625
],
[
0.080078125,
0.27783203125,
0.468017578125,
0.27783203125,
0.468017578125,
0.291748046875,
0.080078125,
0.291748046875
],
[
0.53515625,
0.285888671875,
0.89208984375,
0.285888671875,
0.89208984375,
0.2998046875,
0.53515625,
0.2998046875
],
[
0.080078125,
0.292236328125,
0.468017578125,
0.292236328125,
0.468017578125,
0.30615234375,
0.080078125,
0.30615234375
],
[
0.07843017578125,
0.306884765625,
0.468017578125,
0.30810546875,
0.468017578125,
0.322509765625,
0.07843017578125,
0.3212890625
],
[
0.53515625,
0.3017578125,
0.8349609375,
0.3017578125,
0.8349609375,
0.31494140625,
0.53515625,
0.31494140625
],
[
0.521484375,
0.31494140625,
0.89208984375,
0.31640625,
0.89208984375,
0.330810546875,
0.521484375,
0.330078125
],
[
0.080078125,
0.323974609375,
0.4697265625,
0.323974609375,
0.4697265625,
0.337158203125,
0.080078125,
0.337158203125
],
[
0.53515625,
0.33203125,
0.89208984375,
0.33203125,
0.89208984375,
0.34521484375,
0.53515625,
0.34521484375
],
[
0.080078125,
0.339111328125,
0.468017578125,
0.339111328125,
0.468017578125,
0.352783203125,
0.080078125,
0.352783203125
],
[
0.53515625,
0.3466796875,
0.89208984375,
0.3466796875,
0.89208984375,
0.35986328125,
0.53515625,
0.35986328125
],
[
0.080078125,
0.352783203125,
0.46728515625,
0.352783203125,
0.46728515625,
0.3662109375,
0.080078125,
0.3662109375
],
[
0.5361328125,
0.3623046875,
0.89208984375,
0.3623046875,
0.89208984375,
0.375732421875,
0.5361328125,
0.375732421875
],
[
0.08087158203125,
0.37060546875,
0.465576171875,
0.37060546875,
0.465576171875,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.5361328125,
0.375732421875,
0.890625,
0.375732421875,
0.890625,
0.389404296875,
0.5361328125,
0.389404296875
],
[
0.080078125,
0.3837890625,
0.468017578125,
0.3837890625,
0.468017578125,
0.397705078125,
0.080078125,
0.397705078125
],
[
0.5361328125,
0.393310546875,
0.8759765625,
0.393310546875,
0.8759765625,
0.404052734375,
0.5361328125,
0.404052734375
],
[
0.080078125,
0.399658203125,
0.468017578125,
0.399658203125,
0.468017578125,
0.41357421875,
0.080078125,
0.41357421875
],
[
0.52294921875,
0.4072265625,
0.890625,
0.4072265625,
0.890625,
0.42041015625,
0.52294921875,
0.42041015625
],
[
0.080078125,
0.4140625,
0.46728515625,
0.4140625,
0.46728515625,
0.427978515625,
0.080078125,
0.427978515625
],
[
0.53515625,
0.421630859375,
0.89208984375,
0.421630859375,
0.89208984375,
0.436279296875,
0.53515625,
0.436279296875
],
[
0.08087158203125,
0.4287109375,
0.4697265625,
0.4287109375,
0.4697265625,
0.443115234375,
0.08087158203125,
0.443115234375
],
[
0.53515625,
0.4375,
0.89208984375,
0.4375,
0.89208984375,
0.45068359375,
0.53515625,
0.45068359375
],
[
0.08087158203125,
0.44580078125,
0.468017578125,
0.44580078125,
0.468017578125,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.080078125,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.5361328125,
0.453857421875,
0.8896484375,
0.453857421875,
0.8896484375,
0.464111328125,
0.5361328125,
0.464111328125
],
[
0.53515625,
0.46728515625,
0.89208984375,
0.46728515625,
0.89208984375,
0.48095703125,
0.53515625,
0.48095703125
],
[
0.080078125,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.53515625,
0.48291015625,
0.8310546875,
0.48291015625,
0.8310546875,
0.496826171875,
0.53515625,
0.496826171875
],
[
0.080078125,
0.490478515625,
0.46728515625,
0.490478515625,
0.46728515625,
0.50390625,
0.080078125,
0.50390625
],
[
0.080078125,
0.5048828125,
0.46728515625,
0.5048828125,
0.46728515625,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.50341796875,
0.51318359375,
0.64208984375,
0.51318359375,
0.64208984375,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.08087158203125,
0.521484375,
0.46728515625,
0.521484375,
0.46728515625,
0.5322265625,
0.08087158203125,
0.5322265625
],
[
0.080078125,
0.53515625,
0.4697265625,
0.53515625,
0.4697265625,
0.548828125,
0.080078125,
0.548828125
],
[
0.50341796875,
0.5390625,
0.89208984375,
0.5390625,
0.89208984375,
0.55322265625,
0.50341796875,
0.55322265625
],
[
0.080078125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.080078125,
0.564453125
],
[
0.50244140625,
0.5546875,
0.89208984375,
0.5546875,
0.89208984375,
0.568359375,
0.50244140625,
0.568359375
],
[
0.080078125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.5791015625,
0.080078125,
0.5791015625
],
[
0.50244140625,
0.5693359375,
0.890625,
0.5693359375,
0.890625,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.080078125,
0.580078125,
0.468017578125,
0.580078125,
0.468017578125,
0.59326171875,
0.080078125,
0.59326171875
],
[
0.50244140625,
0.58544921875,
0.89208984375,
0.58544921875,
0.89208984375,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.08087158203125,
0.5966796875,
0.468017578125,
0.5966796875,
0.468017578125,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.50341796875,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.50341796875,
0.61279296875
],
[
0.080078125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.50244140625,
0.61572265625,
0.89208984375,
0.61572265625,
0.89208984375,
0.62890625,
0.50244140625,
0.62890625
],
[
0.080078125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.5009765625,
0.6298828125,
0.890625,
0.62890625,
0.890625,
0.64306640625,
0.5009765625,
0.64453125
],
[
0.080078125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.080078125,
0.65380859375
],
[
0.50341796875,
0.64599609375,
0.89208984375,
0.64599609375,
0.89208984375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.080078125,
0.65576171875,
0.4697265625,
0.65576171875,
0.4697265625,
0.66943359375,
0.080078125,
0.66943359375
],
[
0.50244140625,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.080078125,
0.671875,
0.30810546875,
0.671875,
0.30810546875,
0.68505859375,
0.080078125,
0.68505859375
],
[
0.5009765625,
0.6748046875,
0.89208984375,
0.673828125,
0.89208984375,
0.68798828125,
0.5009765625,
0.689453125
],
[
0.098876953125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.098876953125,
0.7041015625
],
[
0.50244140625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
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
],
[
0.07843017578125,
0.73486328125,
0.4697265625,
0.736328125,
0.4697265625,
0.75,
0.07843017578125,
0.74853515625
],
[
0.50244140625,
0.736328125,
0.89208984375,
0.736328125,
0.89208984375,
0.75,
0.50244140625,
0.75
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
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,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5009765625,
0.78076171875,
0.89208984375,
0.77978515625,
0.89208984375,
0.79443359375,
0.5009765625,
0.79541015625
],
[
0.08087158203125,
0.79736328125,
0.468017578125,
0.79736328125,
0.468017578125,
0.80810546875,
0.08087158203125,
0.80810546875
],
[
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,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5009765625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
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
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
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,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5009765625,
0.88720703125,
0.8095703125,
0.88720703125,
0.8095703125,
0.90087890625,
0.5009765625,
0.90087890625
]
] |
[
"action classes on FineGym dataset [62]. It can be seen that",
"with the Spatial-Temporal Representation Learning (STRL)",
"the 2D poses can accurately capture human skeletons and",
"moduletolearn spatial-temporaldependencies.Finally,the",
"motiondetails.",
"SSL is fine-tuned on action recognition datasets. Extensive",
"Due to the promising ability to model multiple granu-",
"experimental results on FineGym, Diving48, NTU 60 and",
"larities and large variations in human motion, the skeleton",
"NTU 120 show that our SSL generalizes well across dif-",
"sequence is more suitable to distinguish similar actions with",
"ferent datasets and outperforms the state-of-the-art methods",
"subtle differences than the RGB data. To capture discrim-",
"significantly. Our contributions are summarized as follows:",
"inative spatial-temporal motion patterns,mostof the exist-",
": To comprehensively capture human pose and obtain",
"ing skeleton-based action recognition methods [11, 88, 5]",
"discriminative skeleton sequence representation, we",
"are fully supervised and usually require large amounts of",
"propose a graph-based encoder-decoder pre-training",
"labeled data for training elaborate models, which is time-",
"architecture named SkeletonMAE, that embeds skele-",
"consuming and labor-intensive.To mitigate the problem",
"ton joint sequence into GCN and utilize the prior hu-",
"of limited labeled training data, self-supervised skeleton-",
"man topology knowledge to guide the reconstruction",
"based action recognition methods [32, 18, 68] have attracted",
"increasing attention recently.Some contrastive learning",
"of the underlying masked joints and topology.",
": To learn comprehensive spatial-temporal dependen-",
"methods [32, 18] adopted data augmentation to generate",
"cies for skeleton sequence, we propose an efficient",
"pairs of positive and negative samples, but they rely heavily",
"skeleton sequence learning framework, named Skele-",
"on the number of contrastive pairs.With the popularity of",
"ton Sequence Learning (SSL), which integrates the",
"theencoder-decoder[67,511.somemethods[94.681recon-",
"pre-trained SkeletonMAE encoder with the Spatial-",
"structed the masked skeleton sequence by link reconstruc-",
"TemporalRepresentationLearning(STRL)module.",
"tion to encourage the topological closeness following the",
": Extensive experimental results on FineGym, Div-",
"paradigm of graph encoder-decoder. However, these meth-",
"ing48, NTU 60 and NTU 120 datasets show",
"ods are usually good at link prediction and node clustering",
"that our SSL outperforms the state-of-the-art self-",
"butareunsatisfactoryinnodeandgraphclassifications.For",
"accurate action recognition, the fine-grained dependencies",
"supervised skeleton-basedactionrecognitionmethods",
"and achieves comparable performance compared with",
"among different skeleton joints (i.e., graph classifications)",
"the state-of-the-art fully supervised methods.",
"are essential. Therefore, previous self-supervised learning-",
"based methods tend to ignore the fine-grained dependen-",
"2.Related Work",
"ciesamongdifferent skeletonjoints,whichrestrictsthegen-",
"eralization of self-supervised skeleton representation. As",
"Action Recognition.One of the most challenging tasks",
"shown in Fig. 1 (b)-(c), the arm joints and edges are essen-",
"for action recognition is to distinguish similar actions",
"tial to discriminate between these two similar actions. Dif-",
"from subtle differences. Recently, some challenging ac-",
"ferent from the randomly masking strategy of MAE [20],",
"tion recognition datasets like FineGym [62], Diving48",
"our masking strategy is action-sensitive and reconstructs",
"[35], NTU RGB+D 60 [61] and NTU RGB+D 120 [38]",
"specific limbs or body parts that dominate the given ac-",
"are proposed. These datasets contain a large number of",
"tion class. Our SkeletonMAE utilizes prior human topology",
"challenging actions that require discriminative and fine-",
"knowledge to guide the reconstruction of the masked skele-",
"grained action representation learning. For example, in",
"ton joints and edges to achieve a comprehensive perception",
"FineGym [62], an action is divided into action units, sub-",
"of the joints, topology, and action.",
"actions, or phases, and the model is required to distinguish",
"To address the aforementioned challenges, we pro-",
"between “split leap with 1 turn\" and \"switch leap with 1",
"pose an efficient skeleton sequence representation learn-",
"turn\". The higher inter-class similarity and a new level",
"ing framework, named Skeleton Sequence Learning (SSL).",
"of granularity in the fine-grained setting make it a chal-",
"To fully discover the fine-grained dependencies among",
"lenging task, which makes coarse-grained backbones and",
"different skeleton joints, we build a novel asymmet-",
"methods [14, 2, 74, 83] struggle to overcome. To tackle",
"ric graph-based encoder-decoder pre-training architecture",
"the more challenging fine-grained action recognition task,",
"named SkeletonMAE that embeds skeleton joint sequences",
"most of the existing works [53, 34] are fully supervised",
"in Graph Convolutional Network (GCN). The Skeleton-",
"and consider fine-grained actions as distinct categories and",
"MAE aims to reconstruct the masked human skeleton joints",
"supervise the model to learn action semantics. However,",
"and edges based on prior human topology knowledge so",
"collecting and labeling these fine-grained actions is time-",
"that it can infer the underlying topology of the joints and",
"consuming and labor-intensive, which limits the generaliza-",
"obtain a comprehensive perception of human action. To",
"tion of a well-trained model to different datasets.To utilize",
"learn discriminative spatial-temporal skeleton representa-",
"unlabeled data, we propose a graph-based encoder-decoder",
"tion, the pre-trained SkeletonMAE encoder is integrated",
"pre-training architecture named SkeletonMAE"
] |
{
"arxiv_id": "2307.08476",
"language": "en",
"page_idx": 1
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.