annotate writeup/aistats2011_cameraready.tex @ 642:8b1a0b9fecff

made publication list shorter.
author Frederic Bastien <nouiz@nouiz.org>
date Mon, 21 Mar 2011 11:29:09 -0400
parents 507cb92d8e15
children
rev   line source
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
1 %\documentclass[twoside,11pt]{article} % For LaTeX2e
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
2 \documentclass{article} % For LaTeX2e
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
3 \usepackage[accepted]{aistats2e_2011}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
4 %\usepackage{times}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
5 \usepackage{wrapfig}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
6 \usepackage{amsthm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
7 \usepackage{amsmath}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
8 \usepackage{bbm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
9 \usepackage[utf8]{inputenc}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
10 \usepackage[psamsfonts]{amssymb}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
11 %\usepackage{algorithm,algorithmic} % not used after all
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
12 \usepackage{graphicx,subfigure}
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
13 \usepackage{natbib}
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
14 %\usepackage{afterpage}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
15
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
16 \addtolength{\textwidth}{10mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
17 \addtolength{\evensidemargin}{-5mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
18 \addtolength{\oddsidemargin}{-5mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
19
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
20 %\setlength\parindent{0mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
21
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
22 \begin{document}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
23
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
24 \twocolumn[
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
25 \aistatstitle{Deep Learners Benefit More from Out-of-Distribution Examples}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
26 \runningtitle{Deep Learners for Out-of-Distribution Examples}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
27 \runningauthor{Bengio et. al.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
28 \aistatsauthor{
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
29 Yoshua Bengio \and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
30 Frédéric Bastien \and
630
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
31 \bf Arnaud Bergeron \and
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
32 Nicolas Boulanger-Lewandowski \and \\
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
33 \bf Thomas Breuel \and
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
34 Youssouf Chherawala \and
630
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
35 \bf Moustapha Cisse \and
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
36 Myriam Côté \and
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
37 \bf Dumitru Erhan \\
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
38 \and \bf Jeremy Eustache \and
630
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
39 \bf Xavier Glorot \and
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
40 Xavier Muller \and
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
41 \bf Sylvain Pannetier Lebeuf \\
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
42 \and \bf Razvan Pascanu \and
630
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
43 \bf Salah Rifai \and
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
44 Francois Savard \and
630
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
45 \bf Guillaume Sicard \\
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
46 \vspace*{1mm}}
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
47
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
48 %I can't use aistatsaddress in a single side paragraphe.
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
49 %The document is 2 colums, but this section span the 2 colums, sot there is only 1 left
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
50 \center{Dept. IRO, U. Montreal, P.O. Box 6128, Centre-Ville branch, H3C 3J7, Montreal (Qc), Canada}
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
51 \vspace*{5mm}
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
52 ]
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
53 %\aistatsaddress{Dept. IRO, U. Montreal, P.O. Box 6128, Centre-Ville branch, H3C 3J7, Montreal (Qc), Canada}
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
54
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
55
f55f1b1499c4 better printing of all other name and the adresse.
Frederic Bastien <nouiz@nouiz.org>
parents: 627
diff changeset
56 %\vspace*{5mm}}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
57 %\date{{\tt bengioy@iro.umontreal.ca}, Dept. IRO, U. Montreal, P.O. Box 6128, Centre-Ville branch, H3C 3J7, Montreal (Qc), Canada}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
58 %\jmlrheading{}{2010}{}{10/2010}{XX/2011}{Yoshua Bengio et al}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
59 %\editor{}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
60
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
61 %\makeanontitle
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
62 %\maketitle
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
63
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
64 %{\bf Running title: Deep Self-Taught Learning}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
65
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
66 \vspace*{5mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
67 \begin{abstract}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
68 Recent theoretical and empirical work in statistical machine learning has demonstrated the potential of learning algorithms for deep architectures, i.e., function classes obtained by composing multiple levels of representation. The hypothesis evaluated here is that intermediate levels of representation, because they can be shared across tasks and examples from different but related distributions, can yield even more benefits. Comparative experiments were performed on a large-scale handwritten character recognition setting with 62 classes (upper case, lower case, digits), using both a multi-task setting and perturbed examples in order to obtain out-of-distribution examples. The results agree with the hypothesis, and show that a deep learner did {\em beat previously published results and reached human-level performance}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
69 \end{abstract}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
70 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
71
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
72 %\begin{keywords}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
73 %Deep learning, self-taught learning, out-of-distribution examples, handwritten character recognition, multi-task learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
74 %\end{keywords}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
75 %\keywords{self-taught learning \and multi-task learning \and out-of-distribution examples \and deep learning \and handwriting recognition}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
76
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
77
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
78
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
79 \section{Introduction}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
80 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
81
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
82 {\bf Deep Learning} has emerged as a promising new area of research in
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
83 statistical machine learning~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006,VincentPLarochelleH2008-very-small,ranzato-08,TaylorHintonICML2009,Larochelle-jmlr-2009,Salakhutdinov+Hinton-2009,HonglakL2009,HonglakLNIPS2009,Jarrett-ICCV2009,Taylor-cvpr-2010}. See \citet{Bengio-2009} for a review.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
84 Learning algorithms for deep architectures are centered on the learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
85 of useful representations of data, which are better suited to the task at hand,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
86 and are organized in a hierarchy with multiple levels.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
87 This is in part inspired by observations of the mammalian visual cortex,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
88 which consists of a chain of processing elements, each of which is associated with a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
89 different representation of the raw visual input. In fact,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
90 it was found recently that the features learnt in deep architectures resemble
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
91 those observed in the first two of these stages (in areas V1 and V2
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
92 of visual cortex) \citep{HonglakL2008}, and that they become more and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
93 more invariant to factors of variation (such as camera movement) in
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
94 higher layers~\citep{Goodfellow2009}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
95 It has been hypothesized that learning a hierarchy of features increases the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
96 ease and practicality of developing representations that are at once
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
97 tailored to specific tasks, yet are able to borrow statistical strength
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
98 from other related tasks (e.g., modeling different kinds of objects). Finally, learning the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
99 feature representation can lead to higher-level (more abstract, more
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
100 general) features that are more robust to unanticipated sources of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
101 variance extant in real data.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
102
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
103 Whereas a deep architecture can in principle be more powerful than a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
104 shallow one in terms of representation, depth appears to render the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
105 training problem more difficult in terms of optimization and local minima.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
106 It is also only recently that successful algorithms were proposed to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
107 overcome some of these difficulties. All are based on unsupervised
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
108 learning, often in an greedy layer-wise ``unsupervised pre-training''
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
109 stage~\citep{Bengio-2009}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
110 The principle is that each layer starting from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
111 the bottom is trained to represent its input (the output of the previous
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
112 layer). After this
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
113 unsupervised initialization, the stack of layers can be
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
114 converted into a deep supervised feedforward neural network and fine-tuned by
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
115 stochastic gradient descent.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
116 One of these layer initialization techniques,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
117 applied here, is the Denoising
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
118 Auto-encoder~(DA)~\citep{VincentPLarochelleH2008-very-small} (see
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
119 Figure~\ref{fig:da}), which performed similarly or
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
120 better~\citep{VincentPLarochelleH2008-very-small} than previously
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
121 proposed Restricted Boltzmann Machines (RBM)~\citep{Hinton06}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
122 in terms of unsupervised extraction
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
123 of a hierarchy of features useful for classification. Each layer is trained
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
124 to denoise its input, creating a layer of features that can be used as
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
125 input for the next layer, forming a Stacked Denoising Auto-encoder (SDA).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
126 Note that training a Denoising Auto-encoder
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
127 can actually been seen as training a particular RBM by an inductive
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
128 principle different from maximum likelihood~\citep{Vincent-SM-2010},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
129 namely by Score Matching~\citep{Hyvarinen-2005,HyvarinenA2008}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
130
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
131 Previous comparative experimental results with stacking of RBMs and DAs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
132 to build deep supervised predictors had shown that they could outperform
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
133 shallow architectures in a variety of settings, especially
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
134 when the data involves complex interactions between many factors of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
135 variation~\citep{LarochelleH2007,Bengio-2009}. Other experiments have suggested
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
136 that the unsupervised layer-wise pre-training acted as a useful
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
137 prior~\citep{Erhan+al-2010} that allows one to initialize a deep
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
138 neural network in a relatively much smaller region of parameter space,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
139 corresponding to better generalization.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
140
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
141 To further the understanding of the reasons for the good performance
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
142 observed with deep learners, we focus here on the following {\em hypothesis}:
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
143 intermediate levels of representation, especially when there are
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
144 more such levels, can be exploited to {\bf share
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
145 statistical strength across different but related types of examples},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
146 such as examples coming from other tasks than the task of interest
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
147 (the multi-task setting~\citep{caruana97a}), or examples coming from an overlapping
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
148 but different distribution (images with different kinds of perturbations
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
149 and noises, here). This is consistent with the hypotheses discussed
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
150 in~\citet{Bengio-2009} regarding the potential advantage
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
151 of deep learning and the idea that more levels of representation can
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
152 give rise to more abstract, more general features of the raw input.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
153
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
154 This hypothesis is related to a learning setting called
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
155 {\bf self-taught learning}~\citep{RainaR2007}, which combines principles
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
156 of semi-supervised and multi-task learning: in addition to the labeled
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
157 examples from the target distribution, the learner can exploit examples
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
158 that are unlabeled and possibly come from a distribution different from the target
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
159 distribution, e.g., from other classes than those of interest.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
160 It has already been shown that deep learners can clearly take advantage of
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
161 unsupervised learning and unlabeled examples~\citep{Bengio-2009,WestonJ2008-small}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
162 in order to improve performance on a supervised task,
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
163 but more needed to be done to explore the impact
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
164 of {\em out-of-distribution} examples and of the {\em multi-task} setting
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
165 (two exceptions are~\citet{CollobertR2008}, which shares and uses unsupervised
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
166 pre-training only with the first layer, and~\citet{icml2009_093} in the case
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
167 of video data). In particular the {\em relative
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
168 advantage of deep learning} for these settings has not been evaluated.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
169
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
170
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
171 %
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
172 The {\bf main claim} of this paper is that deep learners (with several levels of representation) can
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
173 {\bf benefit more from out-of-distribution examples than shallow learners} (with a single
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
174 level), both in the context of the multi-task setting and from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
175 perturbed examples. Because we are able to improve on state-of-the-art
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
176 performance and reach human-level performance
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
177 on a large-scale task, we consider that this paper is also a contribution
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
178 to advance the application of machine learning to handwritten character recognition.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
179 More precisely, we ask and answer the following questions:
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
180
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
181 %\begin{enumerate}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
182 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
183 Do the good results previously obtained with deep architectures on the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
184 MNIST digit images generalize to the setting of a similar but much larger and richer
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
185 dataset, the NIST special database 19, with 62 classes and around 800k examples?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
186
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
187 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
188 To what extent does the perturbation of input images (e.g. adding
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
189 noise, affine transformations, background images) make the resulting
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
190 classifiers better not only on similarly perturbed images but also on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
191 the {\em original clean examples}? We study this question in the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
192 context of the 62-class and 10-class tasks of the NIST special database 19.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
193
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
194 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
195 Do deep architectures {\em benefit {\bf more} from such out-of-distribution}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
196 examples, in particular do they benefit more from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
197 examples that are perturbed versions of the examples from the task of interest?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
198
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
199 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
200 Similarly, does the feature learning step in deep learning algorithms benefit {\bf more}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
201 from training with moderately {\em different classes} (i.e. a multi-task learning scenario) than
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
202 a corresponding shallow and purely supervised architecture?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
203 We train on 62 classes and test on 10 (digits) or 26 (upper case or lower case)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
204 to answer this question.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
205 %\end{enumerate}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
206
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
207 Our experimental results provide positive evidence towards all of these questions,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
208 as well as {\bf classifiers that reach human-level performance on 62-class isolated character
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
209 recognition and beat previously published results on the NIST dataset (special database 19)}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
210 To achieve these results, we introduce in the next section a sophisticated system
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
211 for stochastically transforming character images and then explain the methodology,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
212 which is based on training with or without these transformed images and testing on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
213 clean ones.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
214 Code for generating these transformations as well as for the deep learning
634
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 631
diff changeset
215 algorithms are made available at
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 631
diff changeset
216 {\tt http://hg.assembla.com/ift6266}.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
217
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
218 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
219 %\newpage
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
220 \section{Perturbed and Transformed Character Images}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
221 \label{s:perturbations}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
222 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
223
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
224 Figure~\ref{fig:transform} shows the different transformations we used to stochastically
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
225 transform $32 \times 32$ source images (such as the one in Fig.\ref{fig:torig})
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
226 in order to obtain data from a larger distribution which
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
227 covers a domain substantially larger than the clean characters distribution from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
228 which we start.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
229 Although character transformations have been used before to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
230 improve character recognizers, this effort is on a large scale both
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
231 in number of classes and in the complexity of the transformations, hence
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
232 in the complexity of the learning task.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
233 The code for these transformations (mostly Python) is available at
634
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 631
diff changeset
234 {\tt http://hg.assembla.com/ift6266}. All the modules in the pipeline (Figure~\ref{fig:transform}) share
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
235 a global control parameter ($0 \le complexity \le 1$) that allows one to modulate the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
236 amount of deformation or noise introduced.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
237 There are two main parts in the pipeline. The first one,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
238 from thickness to pinch, performs transformations. The second
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
239 part, from blur to contrast, adds different kinds of noise.
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
240 More details can be found in~\citet{ARXIV-2010}.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
241
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
242 \begin{figure*}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
243 \centering
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
244 \subfigure[Original]{\includegraphics[scale=0.6]{images/Original.png}\label{fig:torig}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
245 \subfigure[Thickness]{\includegraphics[scale=0.6]{images/Thick_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
246 \subfigure[Slant]{\includegraphics[scale=0.6]{images/Slant_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
247 \subfigure[Affine Transformation]{\includegraphics[scale=0.6]{images/Affine_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
248 \subfigure[Local Elastic Deformation]{\includegraphics[scale=0.6]{images/Localelasticdistorsions_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
249 \subfigure[Pinch]{\includegraphics[scale=0.6]{images/Pinch_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
250 %Noise
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
251 \subfigure[Motion Blur]{\includegraphics[scale=0.6]{images/Motionblur_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
252 \subfigure[Occlusion]{\includegraphics[scale=0.6]{images/occlusion_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
253 \subfigure[Gaussian Smoothing]{\includegraphics[scale=0.6]{images/Bruitgauss_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
254 \subfigure[Pixels Permutation]{\includegraphics[scale=0.6]{images/Permutpixel_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
255 \subfigure[Gaussian Noise]{\includegraphics[scale=0.6]{images/Distorsiongauss_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
256 \subfigure[Background Image Addition]{\includegraphics[scale=0.6]{images/background_other_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
257 \subfigure[Salt \& Pepper]{\includegraphics[scale=0.6]{images/Poivresel_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
258 \subfigure[Scratches]{\includegraphics[scale=0.6]{images/Rature_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
259 \subfigure[Grey Level \& Contrast]{\includegraphics[scale=0.6]{images/Contrast_only.png}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
260 \caption{Top left (a): example original image. Others (b-o): examples of the effect
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
261 of each transformation module taken separately. Actual perturbed examples are obtained by
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
262 a pipeline of these, with random choices about which module to apply and how much perturbation
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
263 to apply.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
264 \label{fig:transform}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
265 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
266 \end{figure*}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
267
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
268 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
269 \section{Experimental Setup}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
270 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
271
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
272 Much previous work on deep learning had been performed on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
273 the MNIST digits task~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006,Salakhutdinov+Hinton-2009},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
274 with 60,000 examples, and variants involving 10,000
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
275 examples~\citep{Larochelle-jmlr-2009,VincentPLarochelleH2008-very-small}\footnote{Fortunately, there
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
276 are more and more exceptions of course, such as~\citet{RainaICML09-small} using a million examples.}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
277 The focus here is on much larger training sets, from 10 times to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
278 to 1000 times larger, and 62 classes.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
279
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
280 The first step in constructing the larger datasets (called NISTP and P07) is to sample from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
281 a {\em data source}: {\bf NIST} (NIST database 19), {\bf Fonts}, {\bf Captchas},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
282 and {\bf OCR data} (scanned machine printed characters). See more in
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
283 Section~\ref{sec:sources} below. Once a character
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
284 is sampled from one of these sources (chosen randomly), the second step is to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
285 apply a pipeline of transformations and/or noise processes outlined in section \ref{s:perturbations}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
286
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
287 To provide a baseline of error rate comparison we also estimate human performance
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
288 on both the 62-class task and the 10-class digits task.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
289 We compare the best Multi-Layer Perceptrons (MLP) against
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
290 the best Stacked Denoising Auto-encoders (SDA), when
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
291 both models' hyper-parameters are selected to minimize the validation set error.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
292 We also provide a comparison against a precise estimate
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
293 of human performance obtained via Amazon's Mechanical Turk (AMT)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
294 service ({\tt http://mturk.com}).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
295 AMT users are paid small amounts
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
296 of money to perform tasks for which human intelligence is required.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
297 Mechanical Turk has been used extensively in natural language processing and vision.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
298 %processing \citep{SnowEtAl2008} and vision
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
299 %\citep{SorokinAndForsyth2008,whitehill09}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
300 AMT users were presented
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
301 with 10 character images (from a test set) on a screen
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
302 and asked to label them.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
303 They were forced to choose a single character class (either among the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
304 62 or 10 character classes) for each image.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
305 80 subjects classified 2500 images per (dataset,task) pair.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
306 Different humans labelers sometimes provided a different label for the same
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
307 example, and we were able to estimate the error variance due to this effect
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
308 because each image was classified by 3 different persons.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
309 The average error of humans on the 62-class task NIST test set
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
310 is 18.2\%, with a standard error of 0.1\%.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
311 We controlled noise in the labelling process by (1)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
312 requiring AMT workers with a higher than normal average of accepted
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
313 responses ($>$95\%) on other tasks (2) discarding responses that were not
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
314 complete (10 predictions) (3) discarding responses for which for which the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
315 time to predict was smaller than 3 seconds for NIST (the mean response time
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
316 was 20 seconds) and 6 seconds seconds for NISTP (average response time of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
317 45 seconds) (4) discarding responses which were obviously wrong (10
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
318 identical ones, or "12345..."). Overall, after such filtering, we kept
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
319 approximately 95\% of the AMT workers' responses.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
320
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
321 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
322 \subsection{Data Sources}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
323 \label{sec:sources}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
324 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
325
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
326 %\begin{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
327 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
328 {\bf NIST.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
329 Our main source of characters is the NIST Special Database 19~\citep{Grother-1995},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
330 widely used for training and testing character
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
331 recognition systems~\citep{Granger+al-2007,Cortes+al-2000-small,Oliveira+al-2002-short,Milgram+al-2005}.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
332 The dataset is composed of 814255 digits and characters (upper and lower cases), with hand checked classifications,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
333 extracted from handwritten sample forms of 3600 writers. The characters are labelled by one of the 62 classes
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
334 corresponding to ``0''-``9'',``A''-``Z'' and ``a''-``z''. The dataset contains 8 parts (partitions) of varying complexity.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
335 The fourth partition (called $hsf_4$, 82,587 examples),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
336 experimentally recognized to be the most difficult one, is the one recommended
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
337 by NIST as a testing set and is used in our work as well as some previous work~\citep{Granger+al-2007,Cortes+al-2000-small,Oliveira+al-2002-short,Milgram+al-2005}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
338 for that purpose. We randomly split the remainder (731,668 examples) into a training set and a validation set for
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
339 model selection.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
340 The performances reported by previous work on that dataset mostly use only the digits.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
341 Here we use all the classes both in the training and testing phase. This is especially
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
342 useful to estimate the effect of a multi-task setting.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
343 The distribution of the classes in the NIST training and test sets differs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
344 substantially, with relatively many more digits in the test set, and a more uniform distribution
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
345 of letters in the test set (whereas in the training set they are distributed
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
346 more like in natural text).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
347 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
348
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
349 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
350 {\bf Fonts.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
351 In order to have a good variety of sources we downloaded an important number of free fonts from:
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
352 {\tt http://cg.scs.carleton.ca/\textasciitilde luc/freefonts.html}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
353 % TODO: pointless to anonymize, it's not pointing to our work
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
354 Including an operating system's (Windows 7) fonts, there we uniformly chose from $9817$ different fonts.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
355 The chosen {\tt ttf} file is either used as input of the Captcha generator (see next item) or, by producing a corresponding image,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
356 directly as input to our models.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
357 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
358
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
359 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
360 {\bf Captchas.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
361 The Captcha data source is an adaptation of the \emph{pycaptcha} library (a Python-based captcha generator library) for
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
362 generating characters of the same format as the NIST dataset. This software is based on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
363 a random character class generator and various kinds of transformations similar to those described in the previous sections.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
364 In order to increase the variability of the data generated, many different fonts are used for generating the characters.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
365 Transformations (slant, distortions, rotation, translation) are applied to each randomly generated character with a complexity
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
366 depending on the value of the complexity parameter provided by the user of the data source.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
367 %Two levels of complexity are allowed and can be controlled via an easy to use facade class. %TODO: what's a facade class?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
368 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
369
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
370 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
371 {\bf OCR data.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
372 A large set (2 million) of scanned, OCRed and manually verified machine-printed
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
373 characters where included as an
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
374 additional source. This set is part of a larger corpus being collected by the Image Understanding
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
375 Pattern Recognition Research group led by Thomas Breuel at University of Kaiserslautern
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
376 ({\tt http://www.iupr.com}).%, and which will be publicly released.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
377 %TODO: let's hope that Thomas is not a reviewer! :) Seriously though, maybe we should anonymize this
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
378 %\end{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
379
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
380 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
381 \subsection{Data Sets}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
382 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
383
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
384 All data sets contain 32$\times$32 grey-level images (values in $[0,1]$) associated with one of 62 character labels.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
385 %\begin{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
386 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
387
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
388 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
389 {\bf NIST.} This is the raw NIST special database 19~\citep{Grother-1995}. It has
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
390 \{651,668 / 80,000 / 82,587\} \{training / validation / test\} examples.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
391 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
392
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
393 %\item
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
394 {\bf P07.} This dataset is obtained by taking raw characters from the above 4 sources
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
395 and sending them through the transformation pipeline described in section \ref{s:perturbations}.
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
396 For each generated example, a data source is selected with probability $10\%$ from the fonts,
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
397 $25\%$ from the captchas, $25\%$ from the OCR data and $40\%$ from NIST. The transformations are
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
398 applied in the
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
399 order given above, and for each of them we sample uniformly a \emph{complexity} in the range $[0,0.7]$.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
400 It has \{81,920,000 / 80,000 / 20,000\} \{training / validation / test\} examples
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
401 obtained from the corresponding NIST sets plus other sources.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
402 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
403
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
404 %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
405 {\bf NISTP.} This one is equivalent to P07 (complexity parameter of $0.7$ with the same proportions of data sources)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
406 except that we only apply
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
407 transformations from slant to pinch (see Fig.\ref{fig:transform}(b-f)).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
408 Therefore, the character is
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
409 transformed but without added noise, yielding images
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
410 closer to the NIST dataset.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
411 It has \{81,920,000 / 80,000 / 20,000\} \{training / validation / test\} examples
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
412 obtained from the corresponding NIST sets plus other sources.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
413 %\end{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
414
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
415 \vspace*{-3mm}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
416 \subsection{Models and their Hyper-parameters}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
417 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
418
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
419 The experiments are performed using MLPs (with a single
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
420 hidden layer) and deep SDAs.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
421 \emph{Hyper-parameters are selected based on the {\bf NISTP} validation set error.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
422
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
423 {\bf Multi-Layer Perceptrons (MLP).} The MLP output estimates the
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
424 class-conditional probabilities
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
425 \[
636
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 634
diff changeset
426 P({\rm class}|{\rm input}=x)={\rm softmax}(b_2+W_2\tanh(b_1+W_1 x)),
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
427 \]
638
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 637
diff changeset
428 i.e., two layers, where $p={\rm softmax}(a)$ means that
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 637
diff changeset
429 $p_i(x)=\exp(a_i)/\sum_j \exp(a_j)$
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
430 representing the probability
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
431 for class $i$, $\tanh$ is the element-wise
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
432 hyperbolic tangent, $b_i$ are parameter vectors, and $W_i$ are
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
433 parameter matrices (one per layer). The
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
434 number of rows of $W_1$ is called the number of hidden units (of the
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
435 single hidden layer, here), and
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
436 is one way to control capacity (the main other ways to control capacity are
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
437 the number of training iterations and optionally a regularization penalty
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
438 on the parameters, not used here because it did not help).
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
439 Whereas previous work had compared
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
440 deep architectures to both shallow MLPs and SVMs, we only compared to MLPs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
441 here because of the very large datasets used (making the use of SVMs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
442 computationally challenging because of their quadratic scaling
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
443 behavior). Preliminary experiments on training SVMs (libSVM) with subsets
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
444 of the training set allowing the program to fit in memory yielded
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
445 substantially worse results than those obtained with MLPs\footnote{RBF SVMs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
446 trained with a subset of NISTP or NIST, 100k examples, to fit in memory,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
447 yielded 64\% test error or worse; online linear SVMs trained on the whole
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
448 of NIST or 800k from NISTP yielded no better than 42\% error; slightly
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
449 better results were obtained by sparsifying the pixel intensities and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
450 projecting to a second-order polynomial (a very sparse vector), still
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
451 41\% error. We expect that better results could be obtained with a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
452 better implementation allowing for training with more examples and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
453 a higher-order non-linear projection.} For training on nearly a hundred million examples (with the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
454 perturbed data), the MLPs and SDA are much more convenient than classifiers
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
455 based on kernel methods. The MLP has a single hidden layer with $\tanh$
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
456 activation functions, and softmax (normalized exponentials) on the output
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
457 layer for estimating $P({\rm class} | {\rm input})$. The number of hidden units is
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
458 taken in $\{300,500,800,1000,1500\}$. Training examples are presented in
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
459 minibatches of size 20, i.e., the parameters are iteratively updated in the direction
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
460 of the mean gradient of the next 20 examples. A constant learning rate was chosen among $\{0.001,
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
461 0.01, 0.025, 0.075, 0.1, 0.5\}$.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
462 %through preliminary experiments (measuring performance on a validation set),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
463 %and $0.1$ (which was found to work best) was then selected for optimizing on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
464 %the whole training sets.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
465 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
466
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
467 \begin{figure*}[htb]
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
468 %\vspace*{-2mm}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
469 \centerline{\resizebox{0.8\textwidth}{!}{\includegraphics{images/denoising_autoencoder_small.pdf}}}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
470 %\vspace*{-2mm}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
471 \caption{Illustration of the computations and training criterion for the denoising
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
472 auto-encoder used to pre-train each layer of the deep architecture. Input $x$ of
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
473 the layer (i.e. raw input or output of previous layer)
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
474 s corrupted into $\tilde{x}$ and encoded into code $y$ by the encoder $f_\theta(\cdot)$.
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
475 The decoder $g_{\theta'}(\cdot)$ maps $y$ to reconstruction $z$, which
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
476 is compared to the uncorrupted input $x$ through the loss function
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
477 $L_H(x,z)$, whose expected value is approximately minimized during training
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
478 by tuning $\theta$ and $\theta'$.}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
479 \label{fig:da}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
480 %\vspace*{-2mm}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
481 \end{figure*}
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
482
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
483 %\afterpage{\clearpage}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
484
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
485 {\bf Stacked Denoising Auto-encoders (SDA).}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
486 Various auto-encoder variants and Restricted Boltzmann Machines (RBMs)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
487 can be used to initialize the weights of each layer of a deep MLP (with many hidden
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
488 layers)~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
489 apparently setting parameters in the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
490 basin of attraction of supervised gradient descent yielding better
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
491 generalization~\citep{Erhan+al-2010}.
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
492 This initial {\em unsupervised
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
493 pre-training phase} does not use the training labels.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
494 Each layer is trained in turn to produce a new representation of its input
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
495 (starting from the raw pixels).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
496 It is hypothesized that the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
497 advantage brought by this procedure stems from a better prior,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
498 on the one hand taking advantage of the link between the input
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
499 distribution $P(x)$ and the conditional distribution of interest
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
500 $P(y|x)$ (like in semi-supervised learning), and on the other hand
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
501 taking advantage of the expressive power and bias implicit in the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
502 deep architecture (whereby complex concepts are expressed as
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
503 compositions of simpler ones through a deep hierarchy).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
504
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
505 Here we chose to use the Denoising
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
506 Auto-encoder~\citep{VincentPLarochelleH2008-very-small} as the building block for
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
507 these deep hierarchies of features, as it is simple to train and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
508 explain (see Figure~\ref{fig:da}, as well as
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
509 tutorial and code there: {\tt http://deeplearning.net/tutorial}),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
510 provides efficient inference, and yielded results
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
511 comparable or better than RBMs in series of experiments
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
512 \citep{VincentPLarochelleH2008-very-small}.
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
513 Some denoising auto-encoders correspond
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
514 to a Gaussian
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
515 RBM trained by a Score Matching criterion~\cite{Vincent-SM-2010}.
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
516 During its unsupervised training, a Denoising
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
517 Auto-encoder is presented with a stochastically corrupted version $\tilde{x}$
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
518 of the input $x$ and trained to reconstruct to produce a reconstruction $z$
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
519 of the uncorrupted input $x$. Because the network has to denoise, it is
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
520 forcing the hidden units $y$ to represent the leading regularities in
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
521 the data. In a slight departure from \citet{VincentPLarochelleH2008-very-small},
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
522 the hidden units output $y$ is obtained through the tanh-affine
636
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 634
diff changeset
523 encoder
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
524 $y=\tanh(c+V x)$
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
525 and the reconstruction is obtained through the transposed transformation
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
526 $z=\tanh(d+V' y)$.
637
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
527 The training
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
528 set average of the cross-entropy
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
529 reconstruction loss (after mapping back numbers in (-1,1) into (0,1))
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
530 \[
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
531 L_H(x,z)=-\sum_i \frac{(z_i+1)}{2} \log \frac{(x_i+1)}{2} + \frac{z_i}{2} \log\frac{x_i}{2}
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
532 \]
637
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
533 is minimized.
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
534 Here we use the random binary masking corruption
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
535 (which in $\tilde{x}$ sets to 0 a random subset of the elements of $x$, and
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
536 copies the rest).
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
537 Once the first denoising auto-encoder is trained, its parameters can be used
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
538 to set the first layer of the deep MLP. The original data are then processed
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
539 through that first layer, and the output of the hidden units form a new
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
540 representation that can be used as input data for training a second denoising
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
541 auto-encoder, still in a purely unsupervised way.
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
542 This is repeated for the desired number of hidden layers.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
543 After this unsupervised pre-training stage, the parameters
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
544 are used to initialize a deep MLP (similar to the above, but
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
545 with more layers), which is fine-tuned by
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
546 the same standard procedure (stochastic gradient descent)
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
547 used to train MLPs in general (see above).
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
548 The top layer parameters of the deep MLP (the one which outputs the
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
549 class probabilities and takes the top hidden layer as input) can
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
550 be initialized at 0.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
551 The SDA hyper-parameters are the same as for the MLP, with the addition of the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
552 amount of corruption noise (we used the masking noise process, whereby a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
553 fixed proportion of the input values, randomly selected, are zeroed), and a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
554 separate learning rate for the unsupervised pre-training stage (selected
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
555 from the same above set). The fraction of inputs corrupted was selected
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
556 among $\{10\%, 20\%, 50\%\}$. Another hyper-parameter is the number
631
510220effb14 corrections demandees par reviewer
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 630
diff changeset
557 of hidden layers but it was fixed to 3 for our experiments,
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
558 based on previous work with
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
559 SDAs on MNIST~\citep{VincentPLarochelleH2008-very-small}.
637
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
560 We also compared against 1 and against 2 hidden layers,
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
561 to disantangle the effect of depth from that of unsupervised
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
562 pre-training.
637
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
563 The size of each hidden
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
564 layer was kept constant across hidden layers, and the best results
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
565 were obtained with the largest values that we tried
fe98896745a5 fitting
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 636
diff changeset
566 (1000 hidden units).
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
567
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
568 %\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
569
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
570 \begin{figure*}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
571 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
572 \centerline{\resizebox{.99\textwidth}{!}{\includegraphics{images/error_rates_charts.pdf}}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
573 %\vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
574 \caption{SDAx are the {\bf deep} models. Error bars indicate a 95\% confidence interval. 0 indicates that the model was trained
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
575 on NIST, 1 on NISTP, and 2 on P07. Left: overall results
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
576 of all models, on NIST and NISTP test sets.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
577 Right: error rates on NIST test digits only, along with the previous results from
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
578 literature~\citep{Granger+al-2007,Cortes+al-2000-small,Oliveira+al-2002-short,Milgram+al-2005}
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
579 respectively based on ART, nearest neighbors, MLPs, and SVMs.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
580 \label{fig:error-rates-charts}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
581 %\vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
582 \end{figure*}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
583
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
584
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
585 \begin{figure*}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
586 \vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
587 \centerline{\resizebox{.99\textwidth}{!}{\includegraphics{images/improvements_charts.pdf}}}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
588 \vspace*{-3mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
589 \caption{Relative improvement in error rate due to out-of-distribution examples.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
590 Left: Improvement (or loss, when negative)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
591 induced by out-of-distribution examples (perturbed data).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
592 Right: Improvement (or loss, when negative) induced by multi-task
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
593 learning (training on all classes and testing only on either digits,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
594 upper case, or lower-case). The deep learner (SDA) benefits more from
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
595 out-of-distribution examples, compared to the shallow MLP.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
596 \label{fig:improvements-charts}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
597 \vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
598 \end{figure*}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
599
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
600 \vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
601 \section{Experimental Results}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
602 \vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
603
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
604 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
605 %\subsection{SDA vs MLP vs Humans}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
606 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
607 The models are either trained on NIST (MLP0 and SDA0),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
608 NISTP (MLP1 and SDA1), or P07 (MLP2 and SDA2), and tested
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
609 on either NIST, NISTP or P07 (regardless of the data set used for training),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
610 either on the 62-class task
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
611 or on the 10-digits task. Training time (including about half
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
612 for unsupervised pre-training, for DAs) on the larger
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
613 datasets is around one day on a GPU (GTX 285).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
614 Figure~\ref{fig:error-rates-charts} summarizes the results obtained,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
615 comparing humans, the three MLPs (MLP0, MLP1, MLP2) and the three SDAs (SDA0, SDA1,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
616 SDA2), along with the previous results on the digits NIST special database
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
617 19 test set from the literature, respectively based on ARTMAP neural
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
618 networks ~\citep{Granger+al-2007}, fast nearest-neighbor search
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
619 ~\citep{Cortes+al-2000-small}, MLPs ~\citep{Oliveira+al-2002-short}, and SVMs
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
620 ~\citep{Milgram+al-2005}.% More detailed and complete numerical results
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
621 %(figures and tables, including standard errors on the error rates) can be
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
622 %found in Appendix.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
623 The deep learner not only outperformed the shallow ones and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
624 previously published performance (in a statistically and qualitatively
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
625 significant way) but when trained with perturbed data
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
626 reaches human performance on both the 62-class task
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
627 and the 10-class (digits) task.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
628 17\% error (SDA1) or 18\% error (humans) may seem large but a large
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
629 majority of the errors from humans and from SDA1 are from out-of-context
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
630 confusions (e.g. a vertical bar can be a ``1'', an ``l'' or an ``L'', and a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
631 ``c'' and a ``C'' are often indistinguishible).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
632 Regarding shallower networks pre-trained with unsupervised denoising
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
633 auto-encders, we find that the NIST test error is 21\% with one hidden
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
634 layer and 20\% with two hidden layers (vs 17\% in the same conditions
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
635 with 3 hidden layers). Compare this with the 23\% error achieved
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
636 by the MLP, i.e. a single hidden layer and no unsupervised pre-training.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
637 As found in previous work~\cite{Erhan+al-2010,Larochelle-jmlr-2009},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
638 these results show that both depth and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
639 unsupervised pre-training need to be combined in order to achieve
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
640 the best results.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
641
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
642
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
643 In addition, as shown in the left of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
644 Figure~\ref{fig:improvements-charts}, the relative improvement in error
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
645 rate brought by out-of-distribution examples is greater for the deep
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
646 SDA, and these
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
647 differences with the shallow MLP are statistically and qualitatively
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
648 significant.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
649 The left side of the figure shows the improvement to the clean
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
650 NIST test set error brought by the use of out-of-distribution examples
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
651 (i.e. the perturbed examples examples from NISTP or P07),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
652 over the models trained exclusively on NIST (respectively SDA0 and MLP0).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
653 Relative percent change is measured by taking
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
654 $100 \% \times$ (original model's error / perturbed-data model's error - 1).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
655 The right side of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
656 Figure~\ref{fig:improvements-charts} shows the relative improvement
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
657 brought by the use of a multi-task setting, in which the same model is
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
658 trained for more classes than the target classes of interest (i.e. training
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
659 with all 62 classes when the target classes are respectively the digits,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
660 lower-case, or upper-case characters). Again, whereas the gain from the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
661 multi-task setting is marginal or negative for the MLP, it is substantial
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
662 for the SDA. Note that to simplify these multi-task experiments, only the original
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
663 NIST dataset is used. For example, the MLP-digits bar shows the relative
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
664 percent improvement in MLP error rate on the NIST digits test set
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
665 as $100\% \times$ (single-task
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
666 model's error / multi-task model's error - 1). The single-task model is
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
667 trained with only 10 outputs (one per digit), seeing only digit examples,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
668 whereas the multi-task model is trained with 62 outputs, with all 62
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
669 character classes as examples. Hence the hidden units are shared across
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
670 all tasks. For the multi-task model, the digit error rate is measured by
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
671 comparing the correct digit class with the output class associated with the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
672 maximum conditional probability among only the digit classes outputs. The
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
673 setting is similar for the other two target classes (lower case characters
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
674 and upper case characters). Note however that some types of perturbations
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
675 (NISTP) help more than others (P07) when testing on the clean images.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
676 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
677 %\subsection{Perturbed Training Data More Helpful for SDA}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
678 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
679
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
680 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
681 %\subsection{Multi-Task Learning Effects}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
682 %%\vspace*{-1mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
683
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
684 \iffalse
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
685 As previously seen, the SDA is better able to benefit from the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
686 transformations applied to the data than the MLP. In this experiment we
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
687 define three tasks: recognizing digits (knowing that the input is a digit),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
688 recognizing upper case characters (knowing that the input is one), and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
689 recognizing lower case characters (knowing that the input is one). We
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
690 consider the digit classification task as the target task and we want to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
691 evaluate whether training with the other tasks can help or hurt, and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
692 whether the effect is different for MLPs versus SDAs. The goal is to find
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
693 out if deep learning can benefit more (or less) from multiple related tasks
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
694 (i.e. the multi-task setting) compared to a corresponding purely supervised
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
695 shallow learner.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
696
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
697 We use a single hidden layer MLP with 1000 hidden units, and a SDA
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
698 with 3 hidden layers (1000 hidden units per layer), pre-trained and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
699 fine-tuned on NIST.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
700
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
701 Our results show that the MLP benefits marginally from the multi-task setting
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
702 in the case of digits (5\% relative improvement) but is actually hurt in the case
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
703 of characters (respectively 3\% and 4\% worse for lower and upper class characters).
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
704 On the other hand the SDA benefited from the multi-task setting, with relative
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
705 error rate improvements of 27\%, 15\% and 13\% respectively for digits,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
706 lower and upper case characters, as shown in Table~\ref{tab:multi-task}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
707 \fi
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
708
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
709
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
710 \vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
711 \section{Conclusions and Discussion}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
712 \vspace*{-2mm}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
713
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
714 We have found that out-of-distribution examples (multi-task learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
715 and perturbed examples) are more beneficial
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
716 to a deep learner than to a traditional shallow and purely
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
717 supervised learner. More precisely,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
718 the answers are positive for all the questions asked in the introduction.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
719 %\begin{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
720
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
721 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
722 {\bf Do the good results previously obtained with deep architectures on the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
723 MNIST digits generalize to a much larger and richer (but similar)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
724 dataset, the NIST special database 19, with 62 classes and around 800k examples}?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
725 Yes, the SDA {\em systematically outperformed the MLP and all the previously
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
726 published results on this dataset} (the ones that we are aware of), {\em in fact reaching human-level
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
727 performance} at around 17\% error on the 62-class task and 1.4\% on the digits,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
728 and beating previously published results on the same data.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
729
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
730 $\bullet$ %\item
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
731 {\bf To what extent do out-of-distribution examples help deep learners,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
732 and do they help them more than shallow supervised ones}?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
733 We found that distorted training examples not only made the resulting
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
734 classifier better on similarly perturbed images but also on
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
735 the {\em original clean examples}, and more importantly and more novel,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
736 that deep architectures benefit more from such {\em out-of-distribution}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
737 examples. Shallow MLPs were helped by perturbed training examples when tested on perturbed input
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
738 images (65\% relative improvement on NISTP)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
739 but only marginally helped (5\% relative improvement on all classes)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
740 or even hurt (10\% relative loss on digits)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
741 with respect to clean examples. On the other hand, the deep SDAs
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
742 were significantly boosted by these out-of-distribution examples.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
743 Similarly, whereas the improvement due to the multi-task setting was marginal or
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
744 negative for the MLP (from +5.6\% to -3.6\% relative change),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
745 it was quite significant for the SDA (from +13\% to +27\% relative change),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
746 which may be explained by the arguments below.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
747 Since out-of-distribution data
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
748 (perturbed or from other related classes) is very common, this conclusion
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
749 is of practical importance.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
750 %\end{itemize}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
751
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
752 In the original self-taught learning framework~\citep{RainaR2007}, the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
753 out-of-sample examples were used as a source of unsupervised data, and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
754 experiments showed its positive effects in a \emph{limited labeled data}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
755 scenario. However, many of the results by \citet{RainaR2007} (who used a
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
756 shallow, sparse coding approach) suggest that the {\em relative gain of self-taught
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
757 learning vs ordinary supervised learning} diminishes as the number of labeled examples increases.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
758 We note instead that, for deep
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
759 architectures, our experiments show that such a positive effect is accomplished
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
760 even in a scenario with a \emph{large number of labeled examples},
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
761 i.e., here, the relative gain of self-taught learning and
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
762 out-of-distribution examples is probably preserved
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
763 in the asymptotic regime. However, note that in our perturbation experiments
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
764 (but not in our multi-task experiments),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
765 even the out-of-distribution examples are labeled, unlike in the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
766 earlier self-taught learning experiments~\citep{RainaR2007}.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
767
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
768 {\bf Why would deep learners benefit more from the self-taught learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
769 framework and out-of-distribution examples}?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
770 The key idea is that the lower layers of the predictor compute a hierarchy
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
771 of features that can be shared across tasks or across variants of the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
772 input distribution. A theoretical analysis of generalization improvements
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
773 due to sharing of intermediate features across tasks already points
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
774 towards that explanation~\citep{baxter95a}.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
775 Intermediate features that can be used in different
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
776 contexts can be estimated in a way that allows to share statistical
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
777 strength. Features extracted through many levels are more likely to
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
778 be more abstract and more invariant to some of the factors of variation
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
779 in the underlying distribution (as the experiments in~\citet{Goodfellow2009} suggest),
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
780 increasing the likelihood that they would be useful for a larger array
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
781 of tasks and input conditions.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
782 Therefore, we hypothesize that both depth and unsupervised
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
783 pre-training play a part in explaining the advantages observed here, and future
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
784 experiments could attempt at teasing apart these factors.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
785 And why would deep learners benefit from the self-taught learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
786 scenarios even when the number of labeled examples is very large?
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
787 We hypothesize that this is related to the hypotheses studied
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
788 in~\citet{Erhan+al-2010}. In~\citet{Erhan+al-2010}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
789 it was found that online learning on a huge dataset did not make the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
790 advantage of the deep learning bias vanish, and a similar phenomenon
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
791 may be happening here. We hypothesize that unsupervised pre-training
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
792 of a deep hierarchy with out-of-distribution examples initializes the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
793 model in the basin of attraction of supervised gradient descent
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
794 that corresponds to better generalization. Furthermore, such good
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
795 basins of attraction are not discovered by pure supervised learning
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
796 (with or without out-of-distribution examples) from random initialization, and more labeled examples
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
797 does not allow the shallow or purely supervised models to discover
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
798 the kind of better basins associated
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
799 with deep learning and out-of-distribution examples.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
800
639
507cb92d8e15 modifs mineures
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 638
diff changeset
801 A Java demo of the recognizer (where both the MLP and the SDA can be compared)
634
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 631
diff changeset
802 can be executed on-line at {\tt http://deep.host22.com}.
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
803
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
804 \iffalse
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
805 \section*{Appendix I: Detailed Numerical Results}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
806
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
807 These tables correspond to Figures 2 and 3 and contain the raw error rates for each model and dataset considered.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
808 They also contain additional data such as test errors on P07 and standard errors.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
809
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
810 \begin{table}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
811 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits +
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
812 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
813 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
814 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07)
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
815 and using a validation set to select hyper-parameters and other training choices.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
816 \{SDA,MLP\}0 are trained on NIST,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
817 \{SDA,MLP\}1 are trained on NISTP, and \{SDA,MLP\}2 are trained on P07.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
818 The human error rate on digits is a lower bound because it does not count digits that were
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
819 recognized as letters. For comparison, the results found in the literature
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
820 on NIST digits classification using the same test set are included.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
821 \label{tab:sda-vs-mlp-vs-humans}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
822 \begin{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
823 \begin{tabular}{|l|r|r|r|r|} \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
824 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
825 Humans& 18.2\% $\pm$.1\% & 39.4\%$\pm$.1\% & 46.9\%$\pm$.1\% & $1.4\%$ \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
826 SDA0 & 23.7\% $\pm$.14\% & 65.2\%$\pm$.34\% & 97.45\%$\pm$.06\% & 2.7\% $\pm$.14\%\\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
827 SDA1 & 17.1\% $\pm$.13\% & 29.7\%$\pm$.3\% & 29.7\%$\pm$.3\% & 1.4\% $\pm$.1\%\\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
828 SDA2 & 18.7\% $\pm$.13\% & 33.6\%$\pm$.3\% & 39.9\%$\pm$.17\% & 1.7\% $\pm$.1\%\\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
829 MLP0 & 24.2\% $\pm$.15\% & 68.8\%$\pm$.33\% & 78.70\%$\pm$.14\% & 3.45\% $\pm$.15\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
830 MLP1 & 23.0\% $\pm$.15\% & 41.8\%$\pm$.35\% & 90.4\%$\pm$.1\% & 3.85\% $\pm$.16\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
831 MLP2 & 24.3\% $\pm$.15\% & 46.0\%$\pm$.35\% & 54.7\%$\pm$.17\% & 4.85\% $\pm$.18\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
832 \citep{Granger+al-2007} & & & & 4.95\% $\pm$.18\% \\ \hline
642
8b1a0b9fecff made publication list shorter.
Frederic Bastien <nouiz@nouiz.org>
parents: 639
diff changeset
833 \citep{Cortes+al-2000-small} & & & & 3.71\% $\pm$.16\% \\ \hline
627
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
834 \citep{Oliveira+al-2002} & & & & 2.4\% $\pm$.13\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
835 \citep{Milgram+al-2005} & & & & 2.1\% $\pm$.12\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
836 \end{tabular}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
837 \end{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
838 \end{table}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
839
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
840 \begin{table}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
841 \caption{Relative change in error rates due to the use of perturbed training data,
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
842 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
843 A positive value indicates that training on the perturbed data helped for the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
844 given test set (the first 3 columns on the 62-class tasks and the last one is
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
845 on the clean 10-class digits). Clearly, the deep learning models did benefit more
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
846 from perturbed training data, even when testing on clean data, whereas the MLP
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
847 trained on perturbed data performed worse on the clean digits and about the same
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
848 on the clean characters. }
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
849 \label{tab:perturbation-effect}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
850 \begin{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
851 \begin{tabular}{|l|r|r|r|r|} \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
852 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
853 SDA0/SDA1-1 & 38\% & 84\% & 228\% & 93\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
854 SDA0/SDA2-1 & 27\% & 94\% & 144\% & 59\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
855 MLP0/MLP1-1 & 5.2\% & 65\% & -13\% & -10\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
856 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
857 \end{tabular}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
858 \end{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
859 \end{table}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
860
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
861 \begin{table}[ht]
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
862 \caption{Test error rates and relative change in error rates due to the use of
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
863 a multi-task setting, i.e., training on each task in isolation vs training
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
864 for all three tasks together, for MLPs vs SDAs. The SDA benefits much
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
865 more from the multi-task setting. All experiments on only on the
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
866 unperturbed NIST data, using validation error for model selection.
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
867 Relative improvement is 1 - single-task error / multi-task error.}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
868 \label{tab:multi-task}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
869 \begin{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
870 \begin{tabular}{|l|r|r|r|} \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
871 & single-task & multi-task & relative \\
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
872 & setting & setting & improvement \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
873 MLP-digits & 3.77\% & 3.99\% & 5.6\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
874 MLP-lower & 17.4\% & 16.8\% & -4.1\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
875 MLP-upper & 7.84\% & 7.54\% & -3.6\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
876 SDA-digits & 2.6\% & 3.56\% & 27\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
877 SDA-lower & 12.3\% & 14.4\% & 15\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
878 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
879 \end{tabular}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
880 \end{center}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
881 \end{table}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
882
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
883 \fi
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
884
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
885 %\afterpage{\clearpage}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
886 %\clearpage
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
887 {
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
888 %\bibliographystyle{spbasic} % basic style, author-year citations
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
889 \bibliographystyle{plainnat}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
890 \bibliography{strings,strings-short,strings-shorter,ift6266_ml,specials,aigaion-shorter}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
891 %\bibliographystyle{unsrtnat}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
892 %\bibliographystyle{apalike}
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
893 }
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
894
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
895
249a180795e3 camera ready version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
896 \end{document}