forked from sthalles/carp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
237 lines (213 loc) · 10.6 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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>CARP</title>
<link href="./css/index.css" rel="stylesheet">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bootstrap-icons@1.11.1/font/bootstrap-icons.css">
<link href="https://cdn.jsdelivr.net/npm/bootstrap@5.3.2/dist/css/bootstrap.min.css" rel="stylesheet"
integrity="sha384-T3c6CoIi6uLrA9TneNEoa7RxnatzjcDSCmG1MXxSR1GAsXEV/Dwwykc2MPK8M2HN" crossorigin="anonymous">
</head>
<body>
<div class="container">
<div class="row">
<article>
<h1 class="display-4">
Representation Learning via Consistent Assignment of Views over Random Partitions
</h1>
<p class="fs-4">
Thalles Santos Silva, Adín Ramírez Rivera
</p>
<p class="fs-5">
37th Conference on Neural Information Processing Systems (NeurIPS 2023).
</p>
</article>
<!-- <img src="./images/carp_overview.png" class="img-fluid" alt="CARP overview architecture."> -->
<img src="./images/carp_overview.png" class="rounded float-start" alt="...">
<!-- <img src="./images/carp_rp.png" class="rounded float-end" alt="..."> -->
</div>
<div class="row">
<h2 class="display-6">Abstract</h2>
<p class="fst-italic">
We present Consistent Assignment of Views over Random Partitions (CARP), a self-supervised clustering
method for representation learning of visual features. CARP learns prototypes in an end-to-end online
fashion using gradient descent without additional non-differentiable modules to solve the cluster
assignment problem. CARP optimizes a new pretext task based on random partitions of prototypes that
regularizes the model and enforces consistency between views' assignments. Additionally, our method
improves training stability and prevents collapsed solutions in joint-embedding training. Through an
extensive evaluation, we demonstrate that CARP's representations are suitable for learning downstream
tasks. We evaluate CARP's representations capabilities in 17 datasets across many standard protocols,
including linear evaluation, few-shot classification, k-NN, k-means, image retrieval, and copy
detection. We compare CARP performance to 11 existing self-supervised methods. We extensively ablate our
method and demonstrate that our proposed random partition pretext task improves the quality of the
learned representations by devising multiple random classification tasks. In transfer learning tasks,
CARP achieves the best performance on average against many SSL methods trained for a longer time.
</p>
</div>
<div class="row">
<h2 class="display-6">Short video presentation</h2>
<p>TODO</p>
</div>
<div class="row">
<h2 class="display-6">Pre-trained models</h2>
<p>Models trained with ResNet50 encoders.</p>
<table class="table table-hover">
<thead>
<tr>
<th scope="col"></th>
<th scope="col">Epochs</th>
<th scope="col">Multicrop</th>
<th scope="col">Linear</th>
<th scope="col">K-NN</th>
<th scope="col">URL</th>
</tr>
</thead>
<tbody class="table-group-divider">
<tr>
<th scope="row">CARP</th>
<td>100</td>
<td>2x224 + 6x96</td>
<td>72.5</td>
<td>63.5</td>
<td><a class="link-opacity-75-hover"
href="https://drive.google.com/drive/folders/1Kj7pp2CcUcEoLYv2d4vs8hQK6xTK6VbR?usp=sharing">checkpoints</a>
</td>
</tr>
<tr>
<th scope="row">CARP</th>
<td>200</td>
<td>2x224 + 6x96</td>
<td>74.2</td>
<td>66.5</td>
<td><a class="link-opacity-75-hover"
href="https://drive.google.com/drive/folders/1NmEAzD4BtM33rOgjEw3o8YS9vHj7qrH9?usp=sharing">checkpoints</a>
</td>
</tr>
<tr>
<th scope="row">CARP</th>
<td>400</td>
<td>2x224</td>
<td>73.0</td>
<td>67.6</td>
<td><a class="link-opacity-75-hover"
href="https://drive.google.com/drive/folders/1xlDsn0JsD_tB11HA1qjdDJxVVy85pV2z?usp=sharing">checkpoints</a>
</td>
</tr>
<tr>
<th scope="row">CARP</th>
<td>400</td>
<td>2x224 + 6x96</td>
<td>75.3</td>
<td>67.7</td>
<td><a class="link-opacity-75-hover"
href="https://drive.google.com/drive/folders/1xlDsn0JsD_tB11HA1qjdDJxVVy85pV2z?usp=sharing">checkpoints</a>
</td>
</tr>
</tbody>
</table>
</div>
<div class="row">
<h2 class="display-6">Important links</h2>
<table class="table">
<thead>
<tr>
<th scope="col">
<a class="icon-link" href="https://github.com/sthalles/carp">
<i class="bi-code" style="font-size: 2rem; color: cornflowerblue;"></i>
Code
</a>
</th>
<th scope="col">
<a class="icon-link" href="https://arxiv.org/abs/2310.12692">
<i class="bi-files" style="font-size: 2rem; color: cornflowerblue;"></i>
arXiv
</a>
</th>
<th scope="col">
<a class="icon-link" href="#">
<i class="bi-paperclip" style="font-size: 2rem; color: cornflowerblue;"></i>
NeurIPS 2023 Proceedings
</a>
</th>
<th scope="col">
<a class="icon-link"
href="https://openreview.net/forum?id=fem6BIJkdv&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DNeurIPS.cc%2F2023%2FConference%2FAuthors%23your-submissions)">
<i class="bi-paperclip" style="font-size: 2rem; color: cornflowerblue;"></i>
OpenReview
</a>
</th>
</tr>
</thead>
</table>
</div>
<div class="row">
<h2 class="display-6">Reference</h2>
<pre class="citation"><code >
@inproceedings{
Silva2023,
title={Representation Learning via Consistent Assignment of Views over Random Partitions},
author={Silva, Thalles and Ram\'irez Rivera, Ad\'in},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems ({NeurIPS})},
year={2023},
url={https://openreview.net/forum?id=fem6BIJkdv}
}
</code></pre>
</div>
<div class="row">
<h2 class="display-6">Authors</h2>
<div class="col">
<div class="span4"></div>
<div class="span4">
<img class="author-profile-images center-block img-thumbnail rounded-circle"
src="./images/thalles.jpeg" />
<p class="text-center">Thalles Santos Silva</p>
</div>
<div class="span4"></div>
</div>
<div class="col">
<div class="span4"></div>
<div class="span4">
<img class="author-profile-images center-block img-thumbnail rounded-circle" src="./images/adin.jpg" />
<p class="text-center">Adín Ramírez Rivera</p>
</div>
<div class="span4"></div>
</div>
</div>
<div class="row">
<h2 class="display-6">Acknowledgements</h2>
<p>
The computations were performed in part on resources provided by Sigma2---the National Infrastructure
for High Performance Computing and Data Storage in Norway---through Project NN8104K.
This work was funded in part by the Research Council of Norway, through its Centre for Research-based
Innovation funding scheme (grant no. 309439), and Consortium Partners.
This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível
Superior---Brasil (CAPES)---Finance Code 001
</p>
</div>
<div class="row">
<table class="table">
<thead>
<tr>
<th scope="col">
<img class="logo center-block" src="./images/unicamp.png" />
</th>
<th scope="col">
<img class="logo center-block" src="./images/ic.png" />
</th>
<th scope="col">
<img class="logo center-block" src="./images/recod.png" />
</th>
<th scope="col">
<img class="logo center-block" src="./images/uio_logo.png" />
</th>
</tr>
</thead>
</table>
</div>
</div>
<script src="https://cdn.jsdelivr.net/npm/bootstrap@5.3.2/dist/js/bootstrap.bundle.min.js"
integrity="sha384-C6RzsynM9kWDrMNeT87bh95OGNyZPhcTNXj1NW7RuBCsyN/o0jlpcV8Qyq46cDfL"
crossorigin="anonymous"></script>
</body>
</html>