-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
446 lines (391 loc) · 23.6 KB
/
index.html
File metadata and controls
446 lines (391 loc) · 23.6 KB
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
<!doctype html>
<html lang="en">
<head>
<!-- Required meta tags -->
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<!-- Primary Meta Tags -->
<title>If your data distribution shifts, use self-learning</title>
<meta name="title" content="If your data distribution shifts, use self-learning">
<meta name="description" content="Test-time adaptation with self-learning improves robustness of large-scale computer vision models on ImageNet-C, -R, and -A.">
<!-- Open Graph / Facebook -->
<meta property="og:type" content="website">
<meta property="og:url" content="https://metatags.io/">
<meta property="og:title" content="If your data distribution shifts, use self-learning">
<meta property="og:description" content="Test-time adaptation with self-learning improves robustness of large-scale computer vision models on ImageNet-C, -R, and -A.">
<!-- Twitter -->
<meta property="twitter:card" content="summary_large_image">
<meta property="twitter:url" content="https://metatags.io/">
<meta property="twitter:title" content="If your data distribution shifts, use self-learning">
<meta property="twitter:description" content="Test-time adaptation with self-learning improves robustness of large-scale computer vision models on ImageNet-C, -R, and -A.">
<!-- Bootstrap CSS -->
<link rel="stylesheet" href="https://stackpath.bootstrapcdn.com/bootstrap/4.5.0/css/bootstrap.min.css"
integrity="sha384-9aIt2nRpC12Uk9gS9baDl411NQApFmC26EwAOH8WgZl5MYYxFfc+NcPb1dKGj7Sk" crossorigin="anonymous">
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<link href="https://fonts.googleapis.com/css2?family=IBM+Plex+Sans+Condensed&display=swap" rel="stylesheet">
<link href="https://fonts.googleapis.com/css2?family=IBM+Plex+Mono&display=swap" rel="stylesheet">
<link href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.13.1/css/all.min.css" rel="stylesheet">
<script src="https://cdnjs.cloudflare.com/ajax/libs/Chart.js/2.9.3/Chart.min.js"></script>
<style>
.main {
font-family: 'IBM Plex Sans Condensed', sans-serif;
}
.code {
font-family: 'IBM Plex Mono', monospace;
}
.row {
padding-bottom: 20px;
}
.a {
color: gainsboro;
font-family: 'IBM Plex Sans Condensed', sans-serif;
}
td {
padding: 0 15px;
}
p {
text-align: justify;
}
</style>
</head>
<body>
<div style="display: none;">
\[
\renewcommand{\d}{{\bf{d}}}
\renewcommand{\b}{{\bf{b}}}
\newcommand{\J}{{\bf{J}}}
\newcommand{\A}{\bf{A}}
\newcommand{\B}{\bf{B}}
\newcommand{\RR}{\mathbf{R}}
\newcommand{\h}{{\bf{h}}}
\newcommand{\x}{{\bf{x}}}
\newcommand{\bfa}{{\bf{a}}}
\newcommand{\bfb}{{\bf{b}}}
\newcommand{\bfc}{{\bf{c}}}
\newcommand{\y}{{\bf{y}}}
\newcommand{\z}{{\bf{z}}}
\newcommand{\w}{{\bf{w}}}
\newcommand{\f}{{\bf{f}}}
\newcommand{\tf}{{\bf{\tilde f}}}
\newcommand{\tx}{{\bf{\tilde x}}}
\renewcommand{\d}{{\rm{d}}}
\newcommand{\s}{{\bf{s}}}
\newcommand{\g}{{\bf{g}}}
\newcommand{\W}{{\bf{W}}}
\newcommand{\vol}{{\operatorname{vol}}}
\newcommand{\zz}{\mathbf{z}}
\newcommand{\xx}{\mathbf{x}}
\newcommand{\bdelta}{\bm{\delta}}
\renewcommand{\H}{\mathbf{H}}
\newcommand{\txx}{{\tilde{\mathbf{x}}}}
\newcommand{\tzz}{{\tilde{\mathbf{z}}}}
\newcommand{\tyy}{{\tilde{\mathbf{y}}}}
\newcommand{\invf}{f^{-1}}
\newcommand{\Sp}{\mathbb{S}}
\]
</div>
<div class="container main">
<div class="row">
<div class="col-sm-2">
</div>
<div class="col-sm-8" id="main-content">
<div class="row text-center my-5" id="#">
<h1>If your data distribution shifts, use self-learning</h1>
</div>
<!-- Begin author list-->
<div class="row text-center mb-4">
<div class="col-sm-4 mb-4">
Evgenia Rusak*
<a href="mailto:evgenia.rusak@bethgelab.org"><i class="far fa-envelope"></i></a></br>
University of Tübingen & <nobr>IMPRS-IS</nobr>
</div>
<div class="col-sm-4 mb-4">
Steffen Schneider*
<a href="mailto:steffen@bethgelab.org"><i class="far fa-envelope"></i></a>
<a href="https://stes.io" target="_blank"><i class="fas fa-link"></i></a></br>
University of Tübingen & <nobr>IMPRS-IS</nobr> & Amazon <small>(internship)</small>
</div>
<div class="col-sm-4 mb-4">
George Pachitariu<br/>
University of Tübingen
</div>
<div class="col-sm-4 mb-4">
Luisa Eck
<a href="https://luisaeck.de/" target="_blank"><i class="fas fa-link"></i></a></br>
University of Oxford
</div>
<div class="col-sm-4 mb-4">
Peter Gehler
<a href="http://gehler.io" target="_blank"><i class="fas fa-link"></i></a></br>
Amazon Tübingen
</div>
<div class="col-sm-4 mb-4">
Oliver Bringmann
<a href="https://www.embedded.uni-tuebingen.de/en/team/oliver-bringmann/" target="_blank"><i class="fas fa-link"></i></a></br>
University of Tübingen
</div>
<div class="col-sm-2 mb-2">
</div>
<div class="col-sm-4 mb-4">
Wieland Brendel
<a href="https://is.mpg.de/en/person/wbrendel" target="_blank"><i class="fas fa-link"></i></a></br>
University of Tübingen
</div>
<div class="col-sm-4 mb-4">
Matthias Bethge
<a href="http://bethgelab.org/people/matthias" target="_blank"><i class="fas fa-link"></i></a><br>
University of Tübingen
</div>
</div>
<!-- End author list-->
<div class="row text-center">
<div class="col-sm-3 mb-3">
<h4>
<a href="https://openreview.net/pdf?id=vqRzLv6POg" target="_blank">
<i class="fas fa-file-alt"></i>
Paper
</a>
</h4>
</div>
<div class="col-sm-3 mb-3">
<h4>
<a href="https://arxiv.org/abs/2104.12928" target="_blank">
<i class="fas fa-file-alt"></i>
Pre-Print
</a>
</h4>
</div>
<div class="col-sm-3 mb-3">
<h4>
<a class="btn-link" href="https://github.com/shift-happens-benchmark/icml-2022/tree/main/shifthappens/tasks/imagenet_d" target="_blank">
<i class="far fa-chart-bar"></i>
Dataset
</a>
</h4>
</div>
<div class="col-sm-3 mb-3">
<h4>
<a href="https://github.com/bethgelab/robustness" target="_blank"> <i
class="fab fa-github"></i>
Code
</a>
</h4>
</div>
</div>
<div class="row text-center">
<p>
<b>tl;dr:</b>
<span class="text-muted">
Test-time adaptation with self-learning improves robustness of large-scale computer vision models on ImageNet-C, -R, and -A.
</span>
</p>
</div>
<div class="row mt-2">
<h3>News</h3>
</div>
<div class="row">
<table>
<tr>
<td class="mr-10">
<span class="badge badge-pill badge-primary">November '22</span>
</td>
<td>
<b>The paper was accepted for publication in the Transactions of Machine Learning Research (TMLR).</b>
The reviews and our comments are published <a href="https://openreview.net/forum?id=vqRzLv6POg">on OpenReview</a>.
</td>
</tr>
<tr>
<td class="mr-10">
<span class="badge badge-pill badge-primary">July '22</span>
</td>
<td>
We presented the <a href="https://openreview.net/forum?id=LiC2vmzbpMO">ImageNet-D dataset</a> at the ICML 2023 Shift Happens Workshop!
The dataset is now part of the <a href="https://github.com/shift-happens-benchmark/icml-2022/tree/main/shifthappens/tasks/imagenet_d">Shift Happens Benchmark</a>.
</td>
</tr>
<tr>
<td class="mr-10">
<span class="badge badge-pill badge-primary">May '21</span>
</td>
<td>
We released <a href="#implementation">a first reference implementation</a> of robust pseudo-labeling. Stay tuned for the full code release.
</td>
</tr>
<tr>
<td class="mr-10">
<span class="badge badge-pill badge-primary">April '21</span>
</td>
<td>
The first pre-print, titled "Adapting ImageNet-scale models to complex distribution shifts with self-learning" is now
available on arXiv: <a
href="https://arxiv.org/abs/2104.12928v2" target="_blank">arxiv.org/abs/2104.12928</a>.
</td>
</tr>
<tr>
<td>
<span class="badge badge-pill badge-primary">April '21</span>
</td>
<td>
A preliminary version of the paper with the title "Better adaptation to distribution shifts with Robust Pseudo-Labeling" was selected for a contributed talk at the ICLR <a href="https://weasul.github.io/accpapers/">Workshop on Weakly Supervised Learning</a>.
</td>
</tr>
</table>
</div>
<div class="row mt-2">
<h3>Abstract</h3>
</div>
<div class="row">
<p>
We demonstrate that self-learning techniques like entropy minimization and pseudo-labeling are simple and effective at improving performance of a deployed computer vision model under systematic domain shifts. We conduct a wide range of large-scale experiments and show consistent improvements irrespective of the model architecture, the pre-training technique or the type of distribution shift. At the same time, self-learning is simple to use in practice because it does not require knowledge or access to the original training data or scheme, is robust to hyperparameter choices, is straight-forward to implement and requires only a few adaptation epochs. This makes self-learning techniques highly attractive for any practitioner who applies machine learning algorithms in the real world. We present state-of-the-art adaptation results on CIFAR10-C (8.5% error), ImageNet-C (22.0% mCE), ImageNet-R (17.4% error) and ImageNet-A (14.8% error), theoretically study the dynamics of self-supervised adaptation methods and propose a new classification dataset (ImageNet-D) which is challenging even with adaptation.
</p>
</div>
<div class="row mt-2">
<img src="img/overview.svg" style="width: 100%;" />
<small class="text-muted">
Robust pseudo labeling (RPL) achieves a new state of the art on ImageNet-C, ImageNet-A and ImageNet-P across various model architectures.
</small>
</div>
<div class="row mt-2">
<h3>Contributions</h3>
</div>
<ol>
<li>
We obtain state-of-the-art adaptation performance on all common robustness datasets (IN-C: 22.0% mCE, IN-A: 14.8% top-1 error, IN-R: 17.4% top-1 error) and improve upon existing strategies for increasing model robustness for all tested model types.
</li>
<li>
We find that self-learning with short update intervals and a limited number of both adaptable and distributed parameters is crucial for success. We leverage label noise robustness methods to enable adaptation with hard labels and a limited number of images, a problem not typically present in smaller scale domain adaptation.
</li>
<li>
Given the huge performance boost on all robustness datasets, we re-purpose the closest candidate to an ImageNet-scale domain adaptation dataset---the dataset used in the Visual Domain Adaptation Challenge 2019---and propose a subset of it as an additional robustness benchmark for the robustness community. We refer to it as ImageNet-D.
</li>
</ol>
<div class="row mt-2">
<h3>Key Experiments</h3>
</div>
<div class="row mt-2">
<div class="col-12">
We build on the paradigm for robustness evaluation considered <a href="https://domainadaptation.org/batchnorm">in our previous work</a>:
Assuming access to unlabeled test samples for adapting ImageNet-trained computer vision models.
In this work, we use various forms of self-learning for adapting the models. We test pseudo labeling approaches with "hard" and "soft" labels, along with entropy minimization and a variant of hard-pseudolabling which uses the <a href="https://arxiv.org/abs/1805.07836">generalized cross entropy loss</a>.
Model selection is entirely done on the development (or holdout) corruptions in ImageNet-C:
</div>
<div class="col-10 offset-lg-1 my-5">
<img src="img/model_selection.svg" style="width: 100%;" />
<small class="text-muted">
Model selection is done on the four dev corruptions in ImageNet-C (left). We use the resulting hyperparameters to evaluate models on the ImageNet-C test set, ImageNet-A and ImageNet-R.
</small>
</div>
<div class="col-12">
Self-learning enables further improvements across various model architectures.
Notably, in contrast to test-time adaptation of batch norm parameters, the technique works for
large scale models pre-trained on datasets like IG-3.5 and JFT300M.
</div>
<div class="col-12 my-5">
<img src="img/full_results.svg" style="width: 100%;" />
<small class="text-muted">
mCE (lower is better) ImageNet-C in %.
Entropy minimization (ENT) and pseudo-labeling paired with a robust loss function (RPL) reduce the mean Corruption Error (mCE) on IN-C for different models. We report the dev score on the holdout corruptions that were used for hyper-parameter tuning and the "test" score on the 15 test corruptions, evaluated with the best hyper-parameters found on the dev set. We compare vanilla trained (Baseline) and the best known robust variants of different architectures.
*) For the EfficientNet-L2 model, we evaluate the mCE on dev on the severities [1,3,5] to save computational resources. For the ResNet50 model, we show results averaged over three seeds as "mean (unbiased std)".
</small>
</div>
Refer to the table in the intro as well as the full papers for results on ImageNet-A and ImageNet-R.
</div>
<div class="row mt-2">
<h3>ImageNet-D: A new challenging robustness benchmark</h3>
</div>
<div class="row mt-2">
<p>
We propose a subset of the dataset from the <a href="https://ai.bu.edu/visda-2019/" target="blank_">Visual Domain Adaptation Challenge 2019</a> as an additional robustness benchmark.
We only consider the subset of the original dataset whose classes can be mapped to ImageNet classes to enable an off-the-shelf evaluation of ImageNet trained models.
</p>
</div>
<div class="col-10 offset-lg-1 my-5">
<img class="w-100" src="img/imagenet-d.png"/>
<small class="text-muted">
Overview of six domains in ImageNet-D. The dataset is a filtered version of the VisDa dataset common in domain adaptation research.
To make the dataset easy to use in a context similar to other robustness datasets, we filtered and remapped the original VisDa dataset onto ImageNet labels.
</small>
</div>
<div class="row mt-2">
<p>
Our best model <span>—</span> the Noisy Student EfficientNet-L2 model <span>—</span> performs considerably worse on this dataset compared to the other robustness benchmarks, making this dataset an interesting future benchmark for the robustness community!
</p>
</div>
<div class="row mt-2" id="implementation">
<h3>Implementation</h3>
</div>
<div class="row">
<p>
Robust pseudo-labeling is conceptually easy to implement.
Notably, the best performing variant does not require to cache any values and computes the pseudo-labels on the fly.
Have a look at <a href="https://gist.github.com/stes/d623b52a1ed4256c71384a596474fb31" target="blank_">our reference implementation</a>.
</p>
<img class="w-100" src="img/gce.gif" style="border-radius: 10px;" />
</div>
<div class="row">
<h3>BibTeX</h3>
</div>
<div class="row">
<p>If you find our analysis helpful, please cite our pre-print:</p>
</div>
<div class="row justify-content-md-center">
<div class="col-sm-8 rounded p-3 m-2" style="background-color:lightgray;">
<small class="code">
@article{rusak2021selflearning,<br>
author = { <br>
Rusak, Evgenia. and<br>
Schneider, Steffen and<br>
Pachitariu, George and<br>
Eck, Luisa and<br>
Gehler, Peter and<br>
Bringmann, Oliver and<br>
Brendel, Wieland and<br>
Bethge, Matthias<br>
},<br>
title = {<br>
If your data distribution shifts,<br>
use self-learning<br>
},<br>
journal={<br>
Transactions of Machine Learning Research<br>
},<br>
year={2022},<br>
url={https://openreview.net/forum?id=vqRzLv6POg},<br>
}
</small>
</div>
</div>
<div class="row">
<h3>Notes</h3>
</div>
<div class="row text-muted">
<ul>
<li>
Concurrent to this work, the CLIP model from Radford et al. [<a href="https://arxiv.org/abs/2103.00020">2103.00020</a>] has shown to be effective at various robustness datasets.
In particular, zero shot transfer to ImageNet-R is at 11.1% (vs. 17.4% for our best <i>adapted</i> model). On ImageNet-A, we still slightly outperform CLIP: 22.9% vs.
16.5% for the non-adapted Efficient Net model and 14.8% for an adapted EfficientNet model.
Given the impressive performance on the ImageNet sketch dataset, it is conceivable that CLIP will also get good performance on ImageNet-D.
</li>
</ul>
</div>
<div class="row">
<small class="text-muted">Webpage designed using Bootstrap 4.5.</small>
<a href="#" class="ml-auto"><i class="fas fa-sort-up"></i></a>
</div>
</div>
</div>
</div>
<!-- Optional JavaScript -->
<!-- jQuery first, then Popper.js, then Bootstrap JS -->
<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/popper.js@1.16.0/dist/umd/popper.min.js"
integrity="sha384-Q6E9RHvbIyZFJoft+2mJbHaEWldlvI9IOYy5n3zV9zzTtmI3UksdQRVvoxMfooAo"
crossorigin="anonymous"></script>
<script src="https://stackpath.bootstrapcdn.com/bootstrap/4.5.0/js/bootstrap.min.js"
integrity="sha384-OgVRvuATP1z7JjHLkuOU7Xw704+h835Lr+6QL9UvYjZE3Ipu6Tp75j7Bh/kR0JKI"
crossorigin="anonymous"></script>
</body>
</html>