-
Notifications
You must be signed in to change notification settings - Fork 1
/
wacv_2024.html
executable file
·214 lines (182 loc) · 11.3 KB
/
wacv_2024.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
<!DOCTYPE html>
<html lang="en-US">
<head>
<meta charset="UTF-8">
<!-- Begin Jekyll SEO tag v2.5.0 -->
<title>The nuts and bolts of Uncertainty Quantification</title>
<meta name="generator" content="Jekyll v3.8.5" />
<meta property="og:title" content="Uncertainty Quantification Tutorial" />
<meta property="og:locale" content="en_US" />
<meta name="description" content="WACV 2024 Tutorial" />
<meta property="og:description" content="WACV 2024 Tutorial" />
<link rel="canonical" href="https://ensta-u2is-ai.github.io/uqt/" />
<meta property="og:url" content="https://ensta-u2is-ai.github.io/uqt/" />
<meta property="og:site_name" content="Uncertainty Quantification Tutorial" />
<script type="application/ld+json">
{"description":"WACV 2024 Tutorial","@type":"WebSite","url":"https://ensta-u2is-ai.github.io/uqt/","name":"Uncertainty Quantification Tutorial","headline":"The nuts and bolts of Uncertainty Quantification"}</script>
<!-- End Jekyll SEO tag -->
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="theme-color" content="#157878">
<link rel="stylesheet" href="https://ensta-u2is-ai.github.io/uqt/assets/css/style.css">
<link rel="stylesheet" href="https://ensta-u2is-ai.github.io/uqt/assets/mystyle.css">
</head>
<body>
<section class="page-header">
<h1 class="project-name">The nuts and bolts of Uncertainty Quantification</h1>
<h2 class="project-tagline">WACV 2024<br>Half-day event - Monday the 8th - AM</h2>
</section>
<section class="main-content">
<div class="container">
<h2>Organizers</h2>
<div>
<div class="instructor">
<a href="http://u2is.ensta-paris.fr/members/franchi/index.php?lang=fr" target="_blank">
<div class="instructorphoto"><img src="assets/franchi.jpg" width="20%" hspace="2%"> </div>
<div>Gianni Franchi<br><small>ENSTA Paris</small></div>
</a>
</div>
<div class="instructor">
<a href="https://github.com/o-laurent" target="_blank">
<div class="instructorphoto"><img src="https://avatars.githubusercontent.com/u/62881275?v=4" width="20%"
hspace="2%"> </div>
<div>Olivier Laurent<br><small>ENSTA Paris / Université Paris Saclay</small></div>
</a>
</div>
<div class="instructor">
<a href="https://qbouniot.github.io/" target="_blank">
<div class="instructorphoto"><img src="https://qbouniot.github.io/images/profil2.jpg" width="5%"
hspace="2%"> </div>
<div>Quentin Bouniot<br><small>Télécom Paris</small></div>
</a>
</div>
<div class="instructor">
<a href="https://abursuc.github.io/" target="_blank">
<div class="instructorphoto"><img src="https://abursuc.github.io/img/abursuc.jpg" width="20%" hspace="2%">
</div>
<div>Andrei Bursuc<br><small>valeo.ai</small></div>
</a>
</div>
<div style="text-align:center">
<a href="https://drive.google.com/file/d/1GpeHCq5bQDEusUtYHroGNIXDNW4fKMf1/view?usp=sharing">
<p style="text-align:center;font-size:46px;color:#088F8F"> 💻 Link to the practical session 💻 </p>
</a>
</div>
<br>
<div class="containertext">
<h2 style="text-align: center">Overview</h2>
<p> Deep Learning (DL) technique are more and more used due to their exceptional performance across various
domains such as image classification, natural language processing, and autonomous driving. However, DL
models often exhibit overconfidence and vulnerability to unreliable predictions, which can have critical
consequences, especially in safety-critical systems. To address this issue and enhance the trustworthiness
of DL models, quantifying their uncertainty is imperative.</p>
<p>In this tutorial, we delve into the theory of uncertainty quantification for Deep Neural Networks (DNNs).
We explore methods and techniques to effectively measure and interpret uncertainty, equipping practitioners
with the tools to foster more reliable and robust DL solutions.
</p>
</div>
<br>
<div class="containertext">
<h2 style="text-align: center">Outline</h2>
<h3 style="text-align: left">Why Uncertainty Quantification ?</h3>
<p> To start, we will delineate the various forms of uncertainty that exist within DNNs. This will provide an
opportunity to elucidate the primary facets of reliability and shed light on the specific issues that
different approaches aim to address. We will also introduce various real-world applications where the study
of uncertainty plays a pivotal role.</p>
<h3 style="text-align: left">Introduction to probabilistic deep models</h3>
<p> In this part, our focus will shift towards the transformation of conventional DNNs into probabilistic
models. We will expound upon the connection between cross entropy and maximum likelihood and delve into the
classical Bayesian framework and the maximum a posteriori approach. Furthermore, we will explore the
correlation with
regression tasks.</p>
<h3 style="text-align: left">Bayesian Neural Networks</h3>
<p> This part will center on the principles of Bayesian Neural Networks (BNNs), detailing how they are
constructed,
trained, and how the posterior distribution of these BNNs
is estimated. We will take a gradual approach to elucidate
the key concepts of BNNs while also highlighting the inherent limitations of these techniques.</p>
<h3 style="text-align: left">Uncertainty from Deep Ensembles</h3>
<p> Here, we will shift our attention to ensemble strategies,
which frequently yield superior performances. We will
delve into the workings of these ensemble techniques, the
reasons behind their effectiveness, and methods for optimizing their efficiency and computational resources
for computer vision applications.</p>
<h3 style="text-align: left">Deterministic Uncertainty Methods</h3>
<p> In this section, we will present specific solutions that
have been explored in the context of regression tasks and
semantic segmentation. The objective of this part is to provide concrete examples of techniques that
illustrate how
uncertainty can be quantified in the realm of computer vision tasks, aiding the audience in understanding
these approaches more comprehensively.</p>
<h3 style="text-align: left">Uncertainty quantification: Do It Yourself</h3>
<p> In this section, we will introduce the <a
href="https://github.com/ensta-u2is-ai/torch-uncertainty">TorchUncertainty
library</a> and provide guidance on how to utilize it effectively. We will demonstrate how to measure
uncertainty in the context of image classification. Using a Google Colab notebook, we will enable attendees
to actively engage
and understand the importance of uncertainty quantification, along with practical insights on how to perform
it.</p>
</div>
<a href="https://torch-uncertainty.github.io/" target="_blank">
<div><img src="assets/logoTU_full.png" width="50%" hspace="2%"> </div>
</a>
<br>
<div class="containertext">
<h2 style="text-align: center">Relation to prior tutorials and short courses</h2>
<p> This tutorial is affiliated with the <a href="https://uncv2023.github.io/">UNCV workshop</a>,
which had its inaugural edition at ECCV and the subsequent one at ICCV, although our primary
emphasis in this tutorial will be on the theoretical facets. </p>
<p> Uncertainty Quantification has received some attention
in recent times, as evidenced by its inclusion as sections in
the tutorial <a href="https://abursuc.github.io/many-faces-reliability/">'Many Faces of Reliability of Deep
Learning for Real-World Deployment'</a>. While this excellent
tutorial explored various applications associated with uncertainty, it did not place a specific emphasis on
probabilistic
models and Bayesian Neural Networks. Our tutorial aims
to provide a more in-depth exploration of uncertainty theory, accompanied by the introduction of practical
applications, including the presentation of our library, <a
href="https://github.com/ensta-u2is-ai/torch-uncertainty">TorchUncertainty</a>.</p>
</div>
<div class="containertext">
<h2 style="text-align: center">Selected References</h2>
<ol>
<li><b>Franchi, G., Bursuc, A.,</b> Aldea, E., Dubuisson, S.,
& Bloch, I. (2020). Encoding the latent posterior of
Bayesian Neural Networks for uncertainty quantification. In IEEE TPAMI.</li>
<li><b>Franchi, G., Bursuc, A.</b>, Aldea, E., Dubuisson, S., &
Bloch, I. (2020). One versus all for deep neural network
incertitude (OVNNI) quantification. In IEEE Access</li>
<li><b>Franchi, G., Bursuc, A.</b>, Aldea, E., Dubuisson, S., &
Bloch, I. (2020, August). TRADI: Tracking deep neural
network weight distributions. In ECCV 2020</li>
<li><b>Franchi, G.</b>, Yu, X., <b>Bursuc, A.</b>, Aldea, E., Dubuisson,
S., & Filliat, D. (2022, October). Latent Discriminant
deterministic Uncertainty. ECCV 2022</li>
<li><b>Laurent, O.</b>, Lafage, A., Tartaglione, E., Daniel, G.,
Martinez, J. M., <b>Bursuc, A., & Franchi, G.</b>
Packed-Ensembles for Efficient Uncertainty Estimation. In ICLR 2023</li>
<li>Yu, X., <b>Franchi, G.</b>, & Aldea, E. (2022, October). On
Monocular Depth Estimation and Uncertainty Quantification using Classification Approaches for Regression.
In ICIP.</li>
<li>Yu, X., <b>Franchi, G.</b>, & Aldea, E. (2021, Novem-
ber). SLURP: Side Learning Uncertainty for Regression Problems. In BMVC.</li>
<li>Hendrycks, D., Dietterich, T. Benchmarking Neural Network Robustness to Common Corruptions and
Perturbations. In ICLR 2019.</li>
</ol>
You will find more references in the <a
href="https://github.com/ensta-u2is-ai/awesome-uncertainty-deeplearning">Awesome Uncertainty in deep
learning.</a>
</div>
</div>
</div>
</section>
<script src="https://code.jquery.com/jquery-3.3.1.slim.min.js"
integrity="sha384-q8i/X+965DzO0rT7abK41JStQIAqVgRVzpbzo5smXKp4YfRvH+8abtTE1Pi6jizo"
crossorigin="anonymous"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.14.7/umd/popper.min.js"
integrity="sha384-UO2eT0CpHqdSJQ6hJty5KVphtPhzWj9WO1clHTMGa3JDZwrnQq4sF86dIHNDz0W1"
crossorigin="anonymous"></script>
<script src="https://stackpath.bootstrapcdn.com/bootstrap/4.3.1/js/bootstrap.min.js"
integrity="sha384-JjSmVgyd0p3pXB1rRibZUAYoIIy6OrQ6VrjIEaFf/nJGzIxFDsf4x0xIM+B07jRM"
crossorigin="anonymous"></script>
</body>
</html>