-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
483 lines (417 loc) · 18.4 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css"
integrity="sha384-B0vP5xmATw1+K9KRQjQERJvTumQW0nPEzvF6L/Z6nronJ3oUOFUFpCjEUQouq2+l" crossorigin="anonymous">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/4.7.0/css/font-awesome.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<title>Is BERT Blind?</title>
<style>
p {
line-height: 1.2;
}
.justify {
text-align: justify;
}
/* .abstract {
line-height: 1;
} */
.teaser {
position: relative;
display: inline;
}
.teaser .img-top {
position: absolute;
top: 0;
left: 0;
opacity: 0;
}
.teaser:hover .img-top {
opacity: 1;
}
.tab {
margin-left: 40px;
}
.card {
border: 0px;
margin: 5px;
}
.img-teaser {
height: 150px;
}
.emph {
font-weight: bold;
}
.emph2 {
font-style: italic;
}
.question {
font-weight: bold;
}
.carousel {
width: 640px;
height: 370px;
}
.carousel-img {
display: block;
margin-left: auto;
margin-right: auto;
}
.carousel-caption {
color: black;
position: relative;
left: auto;
right: auto;
margin: 10px;
}
.carousel-item {
margin: auto;
}
.carousel-control-next,
.carousel-control-prev,
.carousel-indicators {
filter: invert(100%);
}
.carousel-indicators {
/* margin: 10px; */
}
.paper-snapshot {
box-shadow: 10px 10px 5px grey;
}
.stroop-img {
/* box-shadow: 10px 10px 5px grey; */
border: 1px solid grey;
margin: 20px;
width: 95%
}
@media screen and (min-width: 800px) {
.paper-snapshot {
position: relative;
top: 40px;
left: 10px;
}
.stroop-img {
position: relative;
top: 15px;
}
}
.paper-snapshot-link {
max-width: 200px;
margin-left: auto;
margin-right: auto;
}
.card-img-top {
/* width: 100%; */
/* height: 15vw; */
object-fit: contain;
}
.ncd-img {
height: 100px;
/* width: auto; */
}
.ncd-card-body {
/* text-align: center; */
font-size: 18px;
font-family: 'Lucida Console', monospace;
position: relative;
left: 20px;
}
.colorbox {
width: 15%;
height: 15%;
border: 2px solid black;
margin: 5px;
}
.ncd-card-first {
position: relative;
left: 40px;
top: 5px;
}
</style>
</head>
<body class="container" style="max-width:840px">
<script src="https://code.jquery.com/jquery-3.5.1.slim.min.js"
integrity="sha384-DfXdz2htPH0lsSSs5nCTpuj/zy4C+OGpamoFVy38MVBnE+IbbVYUew+OrCXaRkfj"
crossorigin="anonymous"></script>
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js"
integrity="sha384-Piv4xVNRyMGpqkS2by6br4gNJ7DXjqk09RmUpJ8jgGtD7zP9yug3goQfGII0yAns"
crossorigin="anonymous"></script>
<!-- heading -->
<div>
<!-- title -->
<div class='row mt-5 mb-3'>
<div class='col text-center'>
<p class="h2 font-weight-normal">Is BERT Blind?</p>
<p class="h2 font-weight-normal">Exploring the Effect of Vision-and-Language Pretraining on Visual
Language Understanding</p>
</div>
</div>
<!-- authors -->
<div class="col text-center h6 font-weight-bold mb-2 ">
<span><a class="col-md-4 col-xs-6 pb-2" href="https://morrisalp.github.io/">Morris Alper *</a></span>
<span><a class="col-md-4 col-xs-6 pb-2" href="https://www.linkedin.com/in/michael-fiman/">Michael Fiman
*</a></span>
<span><a class="col-md-4 col-xs-6 pb-2" href="https://www.elor.sites.tau.ac.il/">Hadar
Averbuch-Elor</a></span>
</div>
<!-- affiliations -->
<div class='row mb-1'>
<div class='col text-center'>
<p class="h6">
<a href="https://english.tau.ac.il/"><span>Tel Aviv University</span></a>
</p>
</div>
</div>
<!-- venue -->
<div class='row mb-2'>
<div class='col text-center'>
<p class="h6">
CVPR 2023
<br>
* Equal Contribution
</p>
</div>
</div>
<!-- links -->
<div class='row mb-4'>
<div class='col text-center'>
<a href="https://arxiv.org/abs/2303.12513" target="_blank" class="btn btn-outline-primary" role="button">
<i class="ai ai-arxiv"></i>
arXiv
</a>
<a href="https://github.com/TAU-VAILab/isbertblind" target="_blank" class="btn btn-outline-primary"
role="button">
<i class="fa fa-github"></i>
Code and data
</a>
<!-- <a href="#" target="_blank" class="btn btn-outline-primary" role="button">
<i class="fa fa-database"></i>
Dataset (TBD)
</a> -->
</div>
</div>
</div>
<!-- teaser -->
<div class='row justify-content-center'>
<div id="carouselExampleIndicators" class="carousel" data-ride="carousel" data-interval="false">
<ol class="carousel-indicators">
<li data-target="#carouselExampleIndicators" data-slide-to="0" class="active"></li>
<li data-target="#carouselExampleIndicators" data-slide-to="1"></li>
<li data-target="#carouselExampleIndicators" data-slide-to="2"></li>
<li data-target="#carouselExampleIndicators" data-slide-to="3"></li>
</ol>
<div class="carousel-inner">
<div class="carousel-item active">
<img class="carousel-img" src="assets/teaser/present_lowres.jpeg" alt="picture of presents">
<div class="carousel-caption">
<h5>Alex gave Riley a <span class="emph">present</span>.</h5>
<p>How concrete is the word <span class="emph">present</span>?</p>
</div>
</div>
<div class="carousel-item">
<img class="carousel-img" src="assets/teaser/ultimatum_lowres.jpeg"
alt="picture of concepts related to an ultimatum">
<div class="carousel-caption">
<h5>Alex gave Riley an <span class="emph">ultimatum</span>.</h5>
<p>How concrete is the word <span class="emph">ultimatum</span>?</p>
</div>
</div>
<div class="carousel-item">
<img class="carousel-img" src="assets/teaser/ocean_lowres.jpeg" alt="picture of the ocean">
<div class="carousel-caption">
<h5>The ocean is ___‑colored.</h5>
<p>What word should be inserted in the blank?</p>
</div>
</div>
<div class="carousel-item">
<img class="carousel-img" src="assets/teaser/chip_lowres.jpeg"
alt="picture of a hand dipping a corn chip">
<div class="carousel-caption">
<h5>A corn chip is ___‑shaped.</h5>
<p>What word should be inserted in the blank?</p>
</div>
</div>
</div>
<a class="carousel-control-prev" href="#carouselExampleIndicators" role="button" data-slide="prev">
<span class="carousel-control-prev-icon" aria-hidden="true"></span>
<span class="sr-only">Previous</span>
</a>
<a class="carousel-control-next" href="#carouselExampleIndicators" role="button" data-slide="next">
<span class="carousel-control-next-icon" aria-hidden="true"></span>
<span class="sr-only">Next</span>
</a>
</div>
</div>
<div class='col-md-12 col-sm-12 col-xs-12 align-middle mt-1'>
<p class="text-break justify">
Although the questions above are asking about the text alone, solving these tasks (concreteness
prediction, color and shape association prediction) requires visual imagination.
</p>
</div>
<div class='col-md-12 col-sm-12 col-xs-12 align-middle mt-1'>
<p class='h6 font-weight-bold justify'>
In this work, we investigate whether vision-and-language pretraining can improve performance on text-only
tasks involving visual reasoning. We propose a suite of <emph>visual language understanding</emph> tasks for
probing the visual reasoning capabilities of text encoder models, and show that multimodally trained text
encoders
outperform unimodally trained encoders such as BERT on these tasks.
</p>
</div>
</div>
</div>
<!-- Paper section -->
<div>
<hr>
<div class="row">
<div class='col-md-12 col-sm-12 col-xs-12 align-middle mt-1'>
<p class='h4 font-weight-bold'>Abstract</p>
<p class='abstract justify'>
Most humans use visual imagination to understand and reason about language, but models such as BERT
reason about language using knowledge acquired during text-only pretraining. In this work, we
investigate whether vision-and-language pretraining can improve performance on text-only tasks that
involve implicit visual reasoning, focusing primarily on zero-shot probing methods. We propose a
suite of visual language understanding (VLU) tasks for probing the visual reasoning abilities of
text encoder models, as well as various non-visual natural language understanding (NLU) tasks for
comparison. We also contribute a novel zero-shot knowledge probing method, Stroop probing, for
applying models such as CLIP to text-only tasks without needing a prediction head such as the masked
language modelling head of models like BERT. We show that SOTA multimodally trained text encoders
outperform unimodally trained text encoders on the VLU tasks while being underperformed by them on
the NLU tasks, lending new context to previously mixed results regarding the NLU capabilities of
multimodal models. We conclude that exposure to images during pretraining affords inherent visual
reasoning knowledge that is reflected in language-only tasks that require implicit visual reasoning.
Our findings bear importance in the broader context of multimodal learning, providing principled
guidelines for the choice of text encoders used in such contexts.
</p>
</div>
<!-- <div class='col-md-3 col-sm-3 col-xs-12 text-center'>
<div class="row mt-2">
<a href="#" class="paper-snapshot-link">
<img src="assets/paper-snapshot.jpeg" alt="paper-snapshot" class="img-thumbnail paper-snapshot"
width="80%">
</a>
</div>
</div> -->
</div>
</div>
<div>
<hr>
<div class="row">
<div class='col-md-9 col-sm-9 col-xs-12 justify'>
<p class='h4 font-weight-bold'>Stroop Probing</p>
<p>
We propose a new zero-shot probing method for multimodal text encoders such as that of CLIP, taking
inspiration from the psychological phenomenon known as the <a
href="https://en.wikipedia.org/wiki/Stroop_effect">Stroop effect</a>.
</p>
<p>
<b>Try it yourself!</b> Look at the colors listed in the boxes, and try to read out loud <b>the
color in which the text is printed</b> (not the color that is written out). Which set of colors
is harder to read out loud?
</p>
<p>
Stroop probing uses this idea — that incongruent stimuli have an interference effect on the
representation of their context — in order to probe models such as CLIP for knowledge without
requiring a language modelling head.
</p>
</div>
<div class='col-md-3 col-sm-3 col-xs-12 text-center'>
<div class="row mt-2">
<img src="assets/stroop/stroop1.png" alt="stroop effect example 1" class="img-thumbnail stroop-img"
width="100%">
</div>
<div class="row mt-2">
<img src="assets/stroop/stroop3.png" alt="stroop effect example 2" class="img-thumbnail stroop-img"
width="100%">
</div>
</div>
</div>
</div>
<div>
<hr>
<div class="row">
<div class='col-md-12 col-sm-12 col-xs-12 justify'>
<p class='h4 font-weight-bold '>Sample Results</p>
<p>
Compare the results of BERT-base with masked language modelling and CLIP with Stroop probing on the
task of color association prediction. Note that the models receive only the name of the fruit or
vegetable without the given image, shown for illustration purposes only.
</p>
</div>
<div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/white.jpg" alt="white space">
<div class="card-body ncd-card-body ncd-card-first">
<p class="card-text"><b>BERT:<br>CLIP:</b></p>
</div>
</div>
<div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/Broccoli.jpg" alt="broccoli">
<div class="card-body ncd-card-body">
<p class="card-text"><img class="colorbox" src="assets/colors/green.jpeg">green<br><img
class="colorbox" src="assets/colors/green.jpeg">green</p>
</div>
</div>
<div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/carrot.jpg" alt="carrot">
<div class="card-body ncd-card-body">
<p class="card-text"><img class="colorbox" src="assets/colors/green.jpeg">green<br><img
class="colorbox" src="assets/colors/orange.jpeg">orange</p>
</div>
</div>
<div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/Corn.jpg" alt="corn">
<div class="card-body ncd-card-body">
<p class="card-text"><img class="colorbox" src="assets/colors/red.jpeg">red<br><img class="colorbox"
src="assets/colors/yellow.jpeg">yellow</p>
</div>
</div>
<!-- <div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/PG.jpg" alt="pomegranate">
<div class="card-body ncd-card-body">
<p class="card-text"><img class="colorbox" src="assets/colors/red.jpeg">red<br><img class="colorbox" src="assets/colors/red.jpeg">red</p>
</div>
</div> -->
<div class="card">
<img class="card-img-top ncd-img" src="assets/ncd/potato.jpg" alt="potato">
<div class="card-body ncd-card-body">
<p class="card-text"><img class="colorbox" src="assets/colors/white.jpeg">white<br><img
class="colorbox" src="assets/colors/brown.jpeg">brown</p>
</div>
</div>
<div class='col-md-12 col-sm-12 col-xs-12 justify'>
<p class="emph2">
In our work we examine additional visual understanding tasks including concreteness
prediction and shape association prediction. We also evaluate NLU tasks that do not directly involve
visual reasoning as baselines.
</p>
</div>
</div>
</div>
<div>
<hr>
<div class="row">
<div class='col-md-12 col-sm-12 col-xs-12'>
<p class='h4 font-weight-bold '>Acknowledgements</p>
<p>
We thank Noriyuki Kojima, Gabriel Stanovsky, and Adi Haviv for their helpful feedback.
</p>
</div>
</div>
</div>
<div>
<hr>
<div class="row">
<div class='col-md-12 col-sm-12 col-xs-12'>
<p class='h4 font-weight-bold '>Citation</p>
<pre><code>@InProceedings{alper2023:is-bert-blind,
author = {Morris Alper and Michael Fiman and Hadar Averbuch-Elor},
title = {Is BERT Blind? Exploring the Effect of Vision-and-Language Pretraining on Visual Language Understanding},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2023}
}</code></pre>
</div>
</div>
</div>
</body>
</html>