Mercurial > ift6266

annotate writeup/nipswp_submission.tex @ 599:587674030c5a

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added authors and emails list
author boulanni <nicolas_boulanger@hotmail.com>
date Fri, 15 Oct 2010 14:14:06 -0400
parents a0fdc1f134da
children
rev   line source
597
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
1 %\documentclass[twoside,11pt]{article} % For LaTeX2e
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
2 \documentclass{article} % For LaTeX2e
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
3 \usepackage{nips10submit_e}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
4 \usepackage{times}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
5 \usepackage{wrapfig}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
6 \usepackage{amsthm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
7 \usepackage{amsmath}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
8 \usepackage{bbm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
9 \usepackage[utf8]{inputenc}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
10 \usepackage[psamsfonts]{amssymb}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
11 %\usepackage{algorithm,algorithmic} % not used after all
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
12 \usepackage{graphicx,subfigure}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
13 \usepackage[numbers]{natbib}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
14
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
15 \addtolength{\textwidth}{10mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
16 \addtolength{\evensidemargin}{-5mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
17 \addtolength{\oddsidemargin}{-5mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
18
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
19 %\setlength\parindent{0mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
20
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
21 \begin{document}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
22
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
23 \title{Deep Self-Taught Learning for Handwritten Character Recognition}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
24 \author{
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
25 Yoshua Bengio \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
26 Frédéric Bastien \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
27 Arnaud Bergeron \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
28 Nicolas Boulanger-Lewandowski \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
29 Thomas Breuel \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
30 Youssouf Chherawala \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
31 Moustapha Cisse \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
32 Myriam Côté \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
33 Dumitru Erhan \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
34 Jeremy Eustache \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
35 Xavier Glorot \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
36 Xavier Muller \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
37 Sylvain Pannetier Lebeuf \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
38 Razvan Pascanu \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
39 Salah Rifai \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
40 Francois Savard \and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
41 Guillaume Sicard
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
42 }
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
43 \date{{\tt bengioy@iro.umontreal.ca}, Dept. IRO, U. Montreal, P.O. Box 6128, Centre-Ville branch, H3C 3J7, Montreal (Qc), Canada}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
44 %\jmlrheading{}{2010}{}{10/2010}{XX/2011}{Yoshua Bengio et al}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
45 %\editor{}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
46
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
47 %\makeanontitle
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
48 \maketitle
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
49
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
50 %{\bf Running title: Deep Self-Taught Learning}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
51
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
52 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
53 \begin{abstract}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
54 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. Self-taught learning (exploiting unlabeled examples or examples from other distributions) has already been applied to deep learners, but mostly to show the advantage of unlabeled examples. Here we explore the advantage brought by {\em out-of-distribution examples}. For this purpose we developed a powerful generator of stochastic variations and noise processes for character images, including not only affine transformations but also slant, local elastic deformations, changes in thickness, background images, grey level changes, contrast, occlusion, and various types of noise. The out-of-distribution examples are obtained from these highly distorted images or by including examples of object classes different from those in the target test set. We show that {\em deep learners benefit more from out-of-distribution examples than a corresponding shallow learner}, at least in a large-scale handwritten character recognition setting. In fact, we show that they {\em beat previously published results and reach human-level performance}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
55 \end{abstract}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
56 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
57
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
58 %\begin{keywords}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
59 %Deep learning, self-taught learning, out-of-distribution examples, handwritten character recognition, multi-task learning
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
60 %\end{keywords}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
61 %\keywords{self-taught learning \and multi-task learning \and out-of-distribution examples \and deep learning \and handwriting recognition}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
62
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
63
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
64
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
65 \section{Introduction}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
66 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
67
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
68 {\bf Deep Learning} has emerged as a promising new area of research in
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
69 statistical machine learning~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006,VincentPLarochelleH2008,ranzato-08,TaylorHintonICML2009,Larochelle-jmlr-2009,Salakhutdinov+Hinton-2009,HonglakL2009,HonglakLNIPS2009,Jarrett-ICCV2009,Taylor-cvpr-2010}. See \citet{Bengio-2009} for a review.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
70 Learning algorithms for deep architectures are centered on the learning
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
71 of useful representations of data, which are better suited to the task at hand,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
72 and are organized in a hierarchy with multiple levels.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
73 This is in part inspired by observations of the mammalian visual cortex,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
74 which consists of a chain of processing elements, each of which is associated with a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
75 different representation of the raw visual input. In fact,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
76 it was found recently that the features learnt in deep architectures resemble
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
77 those observed in the first two of these stages (in areas V1 and V2
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
78 of visual cortex) \citep{HonglakL2008}, and that they become more and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
79 more invariant to factors of variation (such as camera movement) in
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
80 higher layers~\citep{Goodfellow2009}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
81 Learning a hierarchy of features increases the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
82 ease and practicality of developing representations that are at once
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
83 tailored to specific tasks, yet are able to borrow statistical strength
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
84 from other related tasks (e.g., modeling different kinds of objects). Finally, learning the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
85 feature representation can lead to higher-level (more abstract, more
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
86 general) features that are more robust to unanticipated sources of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
87 variance extant in real data.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
88
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
89 {\bf Self-taught learning}~\citep{RainaR2007} is a paradigm that combines principles
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
90 of semi-supervised and multi-task learning: the learner can exploit examples
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
91 that are unlabeled and possibly come from a distribution different from the target
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
92 distribution, e.g., from other classes than those of interest.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
93 It has already been shown that deep learners can clearly take advantage of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
94 unsupervised learning and unlabeled examples~\citep{Bengio-2009,WestonJ2008-small},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
95 but more needs to be done to explore the impact
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
96 of {\em out-of-distribution} examples and of the {\em multi-task} setting
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
97 (one exception is~\citep{CollobertR2008}, which uses a different kind
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
98 of learning algorithm). In particular the {\em relative
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
99 advantage of deep learning} for these settings has not been evaluated.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
100 The hypothesis discussed in the conclusion is that in the context of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
101 multi-task learning and the availability of out-of-distribution training examples,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
102 a deep hierarchy of features
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
103 may be better able to provide {\em sharing of statistical strength}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
104 between different regions in input space or different tasks, compared to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
105 a shallow learner.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
106
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
107 \iffalse
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
108 Whereas a deep architecture can in principle be more powerful than a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
109 shallow one in terms of representation, depth appears to render the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
110 training problem more difficult in terms of optimization and local minima.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
111 It is also only recently that successful algorithms were proposed to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
112 overcome some of these difficulties. All are based on unsupervised
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
113 learning, often in an greedy layer-wise ``unsupervised pre-training''
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
114 stage~\citep{Bengio-2009}. One of these layer initialization techniques,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
115 applied here, is the Denoising
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
116 Auto-encoder~(DA)~\citep{VincentPLarochelleH2008-very-small} (see Figure~\ref{fig:da}),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
117 which
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
118 performed similarly or better than previously proposed Restricted Boltzmann
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
119 Machines in terms of unsupervised extraction of a hierarchy of features
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
120 useful for classification. Each layer is trained to denoise its
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
121 input, creating a layer of features that can be used as input for the next layer.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
122 \fi
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
123
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
124 %The principle is that each layer starting from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
125 %the bottom is trained to encode its input (the output of the previous
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
126 %layer) and to reconstruct it from a corrupted version. After this
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
127 %unsupervised initialization, the stack of DAs can be
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
128 %converted into a deep supervised feedforward neural network and fine-tuned by
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
129 %stochastic gradient descent.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
130
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
131 %
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
132 The {\bf main claim} of this paper is that deep learners (with several levels of representation) can
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
133 {\bf benefit more from self-taught learning than shallow learners} (with a single
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
134 level), both in the context of the multi-task setting and from {\em
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
135 out-of-distribution examples} in general. Because we are able to improve on state-of-the-art
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
136 performance and reach human-level performance
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
137 on a large-scale task, we consider that this paper is also a contribution
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
138 to advance the application of machine learning to handwritten character recognition.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
139 More precisely, we ask and answer the following questions:
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
140
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
141 %\begin{enumerate}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
142 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
143 Do the good results previously obtained with deep architectures on the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
144 MNIST digit images generalize to the setting of a similar but much larger and richer
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
145 dataset, the NIST special database 19, with 62 classes and around 800k examples?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
146
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
147 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
148 To what extent does the perturbation of input images (e.g. adding
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
149 noise, affine transformations, background images) make the resulting
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
150 classifiers better not only on similarly perturbed images but also on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
151 the {\em original clean examples}? We study this question in the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
152 context of the 62-class and 10-class tasks of the NIST special database 19.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
153
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
154 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
155 Do deep architectures {\em benefit {\bf more} from such out-of-distribution}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
156 examples, i.e. do they benefit more from the self-taught learning~\citep{RainaR2007} framework?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
157 We use highly perturbed examples to generate out-of-distribution examples.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
158
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
159 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
160 Similarly, does the feature learning step in deep learning algorithms benefit {\bf more}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
161 from training with moderately {\em different classes} (i.e. a multi-task learning scenario) than
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
162 a corresponding shallow and purely supervised architecture?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
163 We train on 62 classes and test on 10 (digits) or 26 (upper case or lower case)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
164 to answer this question.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
165 %\end{enumerate}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
166
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
167 Our experimental results provide positive evidence towards all of these questions,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
168 as well as {\em classifiers that reach human-level performance on 62-class isolated character
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
169 recognition and beat previously published results on the NIST dataset (special database 19)}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
170 To achieve these results, we introduce in the next section a sophisticated system
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
171 for stochastically transforming character images and then explain the methodology,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
172 which is based on training with or without these transformed images and testing on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
173 clean ones. We measure the relative advantage of out-of-distribution examples
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
174 (perturbed or out-of-class)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
175 for a deep learner vs a supervised shallow one.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
176 Code for generating these transformations as well as for the deep learning
598
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
177 algorithms are made available at {\tt http://anonymous.url.net}.%{\tt http://hg.assembla.com/ift6266}.
597
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
178 We also estimate the relative advantage for deep learners of training with
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
179 other classes than those of interest, by comparing learners trained with
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
180 62 classes with learners trained with only a subset (on which they
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
181 are then tested).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
182 The conclusion discusses
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
183 the more general question of why deep learners may benefit so much from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
184 the self-taught learning framework. Since out-of-distribution data
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
185 (perturbed or from other related classes) is very common, this conclusion
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
186 is of practical importance.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
187
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
188 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
189 %\newpage
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
190 \section{Perturbed and Transformed Character Images}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
191 \label{s:perturbations}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
192 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
193
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
194 Figure~\ref{fig:transform} shows the different transformations we used to stochastically
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
195 transform $32 \times 32$ source images (such as the one in Fig.\ref{fig:torig})
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
196 in order to obtain data from a larger distribution which
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
197 covers a domain substantially larger than the clean characters distribution from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
198 which we start.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
199 Although character transformations have been used before to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
200 improve character recognizers, this effort is on a large scale both
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
201 in number of classes and in the complexity of the transformations, hence
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
202 in the complexity of the learning task.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
203 The code for these transformations (mostly python) is available at
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
204 {\tt http://anonymous.url.net}. All the modules in the pipeline share
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
205 a global control parameter ($0 \le complexity \le 1$) that allows one to modulate the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
206 amount of deformation or noise introduced.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
207 There are two main parts in the pipeline. The first one,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
208 from slant to pinch below, performs transformations. The second
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
209 part, from blur to contrast, adds different kinds of noise.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
210 More details can be found in~\citep{ift6266-tr-anonymous}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
211
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
212 \begin{figure}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
213 \centering
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
214 \subfigure[Original]{\includegraphics[scale=0.6]{images/Original.png}\label{fig:torig}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
215 \subfigure[Thickness]{\includegraphics[scale=0.6]{images/Thick_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
216 \subfigure[Slant]{\includegraphics[scale=0.6]{images/Slant_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
217 \subfigure[Affine Transformation]{\includegraphics[scale=0.6]{images/Affine_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
218 \subfigure[Local Elastic Deformation]{\includegraphics[scale=0.6]{images/Localelasticdistorsions_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
219 \subfigure[Pinch]{\includegraphics[scale=0.6]{images/Pinch_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
220 %Noise
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
221 \subfigure[Motion Blur]{\includegraphics[scale=0.6]{images/Motionblur_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
222 \subfigure[Occlusion]{\includegraphics[scale=0.6]{images/occlusion_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
223 \subfigure[Gaussian Smoothing]{\includegraphics[scale=0.6]{images/Bruitgauss_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
224 \subfigure[Pixels Permutation]{\includegraphics[scale=0.6]{images/Permutpixel_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
225 \subfigure[Gaussian Noise]{\includegraphics[scale=0.6]{images/Distorsiongauss_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
226 \subfigure[Background Image Addition]{\includegraphics[scale=0.6]{images/background_other_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
227 \subfigure[Salt \& Pepper]{\includegraphics[scale=0.6]{images/Poivresel_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
228 \subfigure[Scratches]{\includegraphics[scale=0.6]{images/Rature_only.png}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
229 \subfigure[Grey Level \& Contrast]{\includegraphics[scale=0.6]{images/Contrast_only.png}}
598
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
230 \caption{Top left (a): example original image. Others (b-o): examples of the effect
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
231 of each transformation module taken separately. Actual perturbed examples are obtained by
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
232 a pipeline of these, with random choices about which module to apply and how much perturbation
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
233 to apply.}
597
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
234 \label{fig:transform}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
235 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
236 \end{figure}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
237
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
238 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
239 \section{Experimental Setup}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
240 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
241
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
242 Much previous work on deep learning had been performed on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
243 the MNIST digits task~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006,Salakhutdinov+Hinton-2009},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
244 with 60~000 examples, and variants involving 10~000
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
245 examples~\citep{Larochelle-jmlr-toappear-2008,VincentPLarochelleH2008}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
246 The focus here is on much larger training sets, from 10 times to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
247 to 1000 times larger, and 62 classes.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
248
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
249 The first step in constructing the larger datasets (called NISTP and P07) is to sample from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
250 a {\em data source}: {\bf NIST} (NIST database 19), {\bf Fonts}, {\bf Captchas},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
251 and {\bf OCR data} (scanned machine printed characters). Once a character
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
252 is sampled from one of these sources (chosen randomly), the second step is to
598
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
253 apply a pipeline of transformations and/or noise processes outlined in section \ref{s:perturbations}.
597
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
254
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
255 To provide a baseline of error rate comparison we also estimate human performance
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
256 on both the 62-class task and the 10-class digits task.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
257 We compare the best Multi-Layer Perceptrons (MLP) against
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
258 the best Stacked Denoising Auto-encoders (SDA), when
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
259 both models' hyper-parameters are selected to minimize the validation set error.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
260 We also provide a comparison against a precise estimate
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
261 of human performance obtained via Amazon's Mechanical Turk (AMT)
598
a0fdc1f134da minor changes to nips workshop submission
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 597
diff changeset
262 service ({\tt http://mturk.com}).
597
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
263 AMT users are paid small amounts
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
264 of money to perform tasks for which human intelligence is required.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
265 Mechanical Turk has been used extensively in natural language processing and vision.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
266 %processing \citep{SnowEtAl2008} and vision
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
267 %\citep{SorokinAndForsyth2008,whitehill09}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
268 AMT users were presented
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
269 with 10 character images (from a test set) and asked to choose 10 corresponding ASCII
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
270 characters. They were forced to choose a single character class (either among the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
271 62 or 10 character classes) for each image.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
272 80 subjects classified 2500 images per (dataset,task) pair.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
273 Different humans labelers sometimes provided a different label for the same
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
274 example, and we were able to estimate the error variance due to this effect
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
275 because each image was classified by 3 different persons.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
276 The average error of humans on the 62-class task NIST test set
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
277 is 18.2\%, with a standard error of 0.1\%.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
278
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
279 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
280 \subsection{Data Sources}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
281 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
282
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
283 %\begin{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
284 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
285 {\bf NIST.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
286 Our main source of characters is the NIST Special Database 19~\citep{Grother-1995},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
287 widely used for training and testing character
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
288 recognition systems~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
289 The dataset is composed of 814255 digits and characters (upper and lower cases), with hand checked classifications,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
290 extracted from handwritten sample forms of 3600 writers. The characters are labelled by one of the 62 classes
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
291 corresponding to ``0''-``9'',``A''-``Z'' and ``a''-``z''. The dataset contains 8 parts (partitions) of varying complexity.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
292 The fourth partition (called $hsf_4$, 82587 examples),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
293 experimentally recognized to be the most difficult one, is the one recommended
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
294 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,Oliveira+al-2002-short,Milgram+al-2005}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
295 for that purpose. We randomly split the remainder (731668 examples) into a training set and a validation set for
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
296 model selection.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
297 The performances reported by previous work on that dataset mostly use only the digits.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
298 Here we use all the classes both in the training and testing phase. This is especially
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
299 useful to estimate the effect of a multi-task setting.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
300 The distribution of the classes in the NIST training and test sets differs
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
301 substantially, with relatively many more digits in the test set, and a more uniform distribution
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
302 of letters in the test set (whereas in the training set they are distributed
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
303 more like in natural text).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
304 %\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
305
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
306 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
307 {\bf Fonts.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
308 In order to have a good variety of sources we downloaded an important number of free fonts from:
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
309 {\tt http://cg.scs.carleton.ca/\textasciitilde luc/freefonts.html}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
310 % TODO: pointless to anonymize, it's not pointing to our work
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
311 Including the operating system's (Windows 7) fonts, there is a total of $9817$ different fonts that we can choose uniformly from.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
312 The chosen {\tt ttf} file is either used as input of the Captcha generator (see next item) or, by producing a corresponding image,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
313 directly as input to our models.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
314 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
315
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
316 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
317 {\bf Captchas.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
318 The Captcha data source is an adaptation of the \emph{pycaptcha} library (a python based captcha generator library) for
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
319 generating characters of the same format as the NIST dataset. This software is based on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
320 a random character class generator and various kinds of transformations similar to those described in the previous sections.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
321 In order to increase the variability of the data generated, many different fonts are used for generating the characters.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
322 Transformations (slant, distortions, rotation, translation) are applied to each randomly generated character with a complexity
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
323 depending on the value of the complexity parameter provided by the user of the data source.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
324 %Two levels of complexity are allowed and can be controlled via an easy to use facade class. %TODO: what's a facade class?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
325 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
326
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
327 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
328 {\bf OCR data.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
329 A large set (2 million) of scanned, OCRed and manually verified machine-printed
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
330 characters where included as an
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
331 additional source. This set is part of a larger corpus being collected by the Image Understanding
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
332 Pattern Recognition Research group led by Thomas Breuel at University of Kaiserslautern
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
333 ({\tt http://www.iupr.com}), and which will be publicly released.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
334 %TODO: let's hope that Thomas is not a reviewer! :) Seriously though, maybe we should anonymize this
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
335 %\end{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
336
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
337 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
338 \subsection{Data Sets}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
339 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
340
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
341 All data sets contain 32$\times$32 grey-level images (values in $[0,1]$) associated with a label
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
342 from one of the 62 character classes.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
343 %\begin{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
344 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
345
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
346 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
347 {\bf NIST.} This is the raw NIST special database 19~\citep{Grother-1995}. It has
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
348 \{651668 / 80000 / 82587\} \{training / validation / test\} examples.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
349 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
350
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
351 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
352 {\bf P07.} This dataset is obtained by taking raw characters from all four of the above sources
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
353 and sending them through the transformation pipeline described in section \ref{s:perturbations}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
354 For each new example to generate, a data source is selected with probability $10\%$ from the fonts,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
355 $25\%$ from the captchas, $25\%$ from the OCR data and $40\%$ from NIST. We apply all the transformations in the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
356 order given above, and for each of them we sample uniformly a \emph{complexity} in the range $[0,0.7]$.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
357 It has \{81920000 / 80000 / 20000\} \{training / validation / test\} examples.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
358 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
359
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
360 %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
361 {\bf NISTP.} This one is equivalent to P07 (complexity parameter of $0.7$ with the same proportions of data sources)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
362 except that we only apply
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
363 transformations from slant to pinch. Therefore, the character is
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
364 transformed but no additional noise is added to the image, giving images
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
365 closer to the NIST dataset.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
366 It has \{81920000 / 80000 / 20000\} \{training / validation / test\} examples.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
367 %\end{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
368
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
369 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
370 \subsection{Models and their Hyperparameters}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
371 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
372
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
373 The experiments are performed using MLPs (with a single
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
374 hidden layer) and SDAs.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
375 \emph{Hyper-parameters are selected based on the {\bf NISTP} validation set error.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
376
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
377 {\bf Multi-Layer Perceptrons (MLP).}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
378 Whereas previous work had compared deep architectures to both shallow MLPs and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
379 SVMs, we only compared to MLPs here because of the very large datasets used
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
380 (making the use of SVMs computationally challenging because of their quadratic
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
381 scaling behavior). Preliminary experiments on training SVMs (libSVM) with subsets of the training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
382 set allowing the program to fit in memory yielded substantially worse results
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
383 than those obtained with MLPs. For training on nearly a billion examples
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
384 (with the perturbed data), the MLPs and SDA are much more convenient than
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
385 classifiers based on kernel methods.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
386 The MLP has a single hidden layer with $\tanh$ activation functions, and softmax (normalized
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
387 exponentials) on the output layer for estimating $P(class | image)$.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
388 The number of hidden units is taken in $\{300,500,800,1000,1500\}$.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
389 Training examples are presented in minibatches of size 20. A constant learning
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
390 rate was chosen among $\{0.001, 0.01, 0.025, 0.075, 0.1, 0.5\}$.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
391 %through preliminary experiments (measuring performance on a validation set),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
392 %and $0.1$ (which was found to work best) was then selected for optimizing on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
393 %the whole training sets.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
394 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
395
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
396
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
397 {\bf Stacked Denoising Auto-Encoders (SDA).}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
398 Various auto-encoder variants and Restricted Boltzmann Machines (RBMs)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
399 can be used to initialize the weights of each layer of a deep MLP (with many hidden
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
400 layers)~\citep{Hinton06,ranzato-07-small,Bengio-nips-2006},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
401 apparently setting parameters in the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
402 basin of attraction of supervised gradient descent yielding better
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
403 generalization~\citep{Erhan+al-2010}. This initial {\em unsupervised
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
404 pre-training phase} uses all of the training images but not the training labels.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
405 Each layer is trained in turn to produce a new representation of its input
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
406 (starting from the raw pixels).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
407 It is hypothesized that the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
408 advantage brought by this procedure stems from a better prior,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
409 on the one hand taking advantage of the link between the input
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
410 distribution $P(x)$ and the conditional distribution of interest
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
411 $P(y|x)$ (like in semi-supervised learning), and on the other hand
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
412 taking advantage of the expressive power and bias implicit in the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
413 deep architecture (whereby complex concepts are expressed as
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
414 compositions of simpler ones through a deep hierarchy).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
415
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
416 \begin{figure}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
417 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
418 \centerline{\resizebox{0.8\textwidth}{!}{\includegraphics{images/denoising_autoencoder_small.pdf}}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
419 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
420 \caption{Illustration of the computations and training criterion for the denoising
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
421 auto-encoder used to pre-train each layer of the deep architecture. Input $x$ of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
422 the layer (i.e. raw input or output of previous layer)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
423 s corrupted into $\tilde{x}$ and encoded into code $y$ by the encoder $f_\theta(\cdot)$.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
424 The decoder $g_{\theta'}(\cdot)$ maps $y$ to reconstruction $z$, which
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
425 is compared to the uncorrupted input $x$ through the loss function
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
426 $L_H(x,z)$, whose expected value is approximately minimized during training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
427 by tuning $\theta$ and $\theta'$.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
428 \label{fig:da}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
429 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
430 \end{figure}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
431
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
432 Here we chose to use the Denoising
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
433 Auto-encoder~\citep{VincentPLarochelleH2008} as the building block for
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
434 these deep hierarchies of features, as it is simple to train and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
435 explain (see Figure~\ref{fig:da}, as well as
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
436 tutorial and code there: {\tt http://deeplearning.net/tutorial}),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
437 provides efficient inference, and yielded results
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
438 comparable or better than RBMs in series of experiments
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
439 \citep{VincentPLarochelleH2008}. During training, a Denoising
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
440 Auto-encoder is presented with a stochastically corrupted version
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
441 of the input and trained to reconstruct the uncorrupted input,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
442 forcing the hidden units to represent the leading regularities in
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
443 the data. Here we use the random binary masking corruption
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
444 (which sets to 0 a random subset of the inputs).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
445 Once it is trained, in a purely unsupervised way,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
446 its hidden units' activations can
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
447 be used as inputs for training a second one, etc.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
448 After this unsupervised pre-training stage, the parameters
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
449 are used to initialize a deep MLP, which is fine-tuned by
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
450 the same standard procedure used to train them (see previous section).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
451 The SDA hyper-parameters are the same as for the MLP, with the addition of the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
452 amount of corruption noise (we used the masking noise process, whereby a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
453 fixed proportion of the input values, randomly selected, are zeroed), and a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
454 separate learning rate for the unsupervised pre-training stage (selected
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
455 from the same above set). The fraction of inputs corrupted was selected
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
456 among $\{10\%, 20\%, 50\%\}$. Another hyper-parameter is the number
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
457 of hidden layers but it was fixed to 3 based on previous work with
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
458 SDAs on MNIST~\citep{VincentPLarochelleH2008}. The size of the hidden
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
459 layers was kept constant across hidden layers, and the best results
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
460 were obtained with the largest values that we could experiment
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
461 with given our patience, with 1000 hidden units.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
462
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
463 \vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
464
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
465 \begin{figure}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
466 %\vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
467 \centerline{\resizebox{.99\textwidth}{!}{\includegraphics{images/error_rates_charts.pdf}}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
468 %\vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
469 \caption{SDAx are the {\bf deep} models. Error bars indicate a 95\% confidence interval. 0 indicates that the model was trained
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
470 on NIST, 1 on NISTP, and 2 on P07. Left: overall results
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
471 of all models, on NIST and NISTP test sets.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
472 Right: error rates on NIST test digits only, along with the previous results from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
473 literature~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
474 respectively based on ART, nearest neighbors, MLPs, and SVMs.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
475 \label{fig:error-rates-charts}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
476 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
477 \end{figure}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
478
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
479
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
480 \begin{figure}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
481 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
482 \centerline{\resizebox{.99\textwidth}{!}{\includegraphics{images/improvements_charts.pdf}}}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
483 \vspace*{-3mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
484 \caption{Relative improvement in error rate due to self-taught learning.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
485 Left: Improvement (or loss, when negative)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
486 induced by out-of-distribution examples (perturbed data).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
487 Right: Improvement (or loss, when negative) induced by multi-task
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
488 learning (training on all classes and testing only on either digits,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
489 upper case, or lower-case). The deep learner (SDA) benefits more from
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
490 both self-taught learning scenarios, compared to the shallow MLP.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
491 \label{fig:improvements-charts}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
492 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
493 \end{figure}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
494
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
495 \section{Experimental Results}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
496 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
497
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
498 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
499 %\subsection{SDA vs MLP vs Humans}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
500 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
501 The models are either trained on NIST (MLP0 and SDA0),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
502 NISTP (MLP1 and SDA1), or P07 (MLP2 and SDA2), and tested
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
503 on either NIST, NISTP or P07, either on the 62-class task
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
504 or on the 10-digits task. Training (including about half
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
505 for unsupervised pre-training, for DAs) on the larger
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
506 datasets takes around one day on a GPU-285.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
507 Figure~\ref{fig:error-rates-charts} summarizes the results obtained,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
508 comparing humans, the three MLPs (MLP0, MLP1, MLP2) and the three SDAs (SDA0, SDA1,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
509 SDA2), along with the previous results on the digits NIST special database
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
510 19 test set from the literature, respectively based on ARTMAP neural
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
511 networks ~\citep{Granger+al-2007}, fast nearest-neighbor search
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
512 ~\citep{Cortes+al-2000}, MLPs ~\citep{Oliveira+al-2002-short}, and SVMs
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
513 ~\citep{Milgram+al-2005}.% More detailed and complete numerical results
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
514 %(figures and tables, including standard errors on the error rates) can be
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
515 %found in Appendix.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
516 The deep learner not only outperformed the shallow ones and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
517 previously published performance (in a statistically and qualitatively
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
518 significant way) but when trained with perturbed data
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
519 reaches human performance on both the 62-class task
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
520 and the 10-class (digits) task.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
521 17\% error (SDA1) or 18\% error (humans) may seem large but a large
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
522 majority of the errors from humans and from SDA1 are from out-of-context
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
523 confusions (e.g. a vertical bar can be a ``1'', an ``l'' or an ``L'', and a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
524 ``c'' and a ``C'' are often indistinguishible).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
525
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
526 In addition, as shown in the left of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
527 Figure~\ref{fig:improvements-charts}, the relative improvement in error
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
528 rate brought by self-taught learning is greater for the SDA, and these
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
529 differences with the MLP are statistically and qualitatively
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
530 significant.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
531 The left side of the figure shows the improvement to the clean
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
532 NIST test set error brought by the use of out-of-distribution examples
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
533 (i.e. the perturbed examples examples from NISTP or P07).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
534 Relative percent change is measured by taking
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
535 $100 \% \times$ (original model's error / perturbed-data model's error - 1).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
536 The right side of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
537 Figure~\ref{fig:improvements-charts} shows the relative improvement
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
538 brought by the use of a multi-task setting, in which the same model is
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
539 trained for more classes than the target classes of interest (i.e. training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
540 with all 62 classes when the target classes are respectively the digits,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
541 lower-case, or upper-case characters). Again, whereas the gain from the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
542 multi-task setting is marginal or negative for the MLP, it is substantial
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
543 for the SDA. Note that to simplify these multi-task experiments, only the original
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
544 NIST dataset is used. For example, the MLP-digits bar shows the relative
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
545 percent improvement in MLP error rate on the NIST digits test set
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
546 is $100\% \times$ (single-task
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
547 model's error / multi-task model's error - 1). The single-task model is
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
548 trained with only 10 outputs (one per digit), seeing only digit examples,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
549 whereas the multi-task model is trained with 62 outputs, with all 62
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
550 character classes as examples. Hence the hidden units are shared across
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
551 all tasks. For the multi-task model, the digit error rate is measured by
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
552 comparing the correct digit class with the output class associated with the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
553 maximum conditional probability among only the digit classes outputs. The
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
554 setting is similar for the other two target classes (lower case characters
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
555 and upper case characters).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
556 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
557 %\subsection{Perturbed Training Data More Helpful for SDA}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
558 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
559
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
560 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
561 %\subsection{Multi-Task Learning Effects}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
562 %%\vspace*{-1mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
563
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
564 \iffalse
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
565 As previously seen, the SDA is better able to benefit from the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
566 transformations applied to the data than the MLP. In this experiment we
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
567 define three tasks: recognizing digits (knowing that the input is a digit),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
568 recognizing upper case characters (knowing that the input is one), and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
569 recognizing lower case characters (knowing that the input is one). We
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
570 consider the digit classification task as the target task and we want to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
571 evaluate whether training with the other tasks can help or hurt, and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
572 whether the effect is different for MLPs versus SDAs. The goal is to find
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
573 out if deep learning can benefit more (or less) from multiple related tasks
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
574 (i.e. the multi-task setting) compared to a corresponding purely supervised
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
575 shallow learner.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
576
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
577 We use a single hidden layer MLP with 1000 hidden units, and a SDA
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
578 with 3 hidden layers (1000 hidden units per layer), pre-trained and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
579 fine-tuned on NIST.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
580
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
581 Our results show that the MLP benefits marginally from the multi-task setting
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
582 in the case of digits (5\% relative improvement) but is actually hurt in the case
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
583 of characters (respectively 3\% and 4\% worse for lower and upper class characters).
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
584 On the other hand the SDA benefited from the multi-task setting, with relative
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
585 error rate improvements of 27\%, 15\% and 13\% respectively for digits,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
586 lower and upper case characters, as shown in Table~\ref{tab:multi-task}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
587 \fi
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
588
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
589
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
590 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
591 \section{Conclusions and Discussion}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
592 \vspace*{-2mm}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
593
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
594 We have found that the self-taught learning framework is more beneficial
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
595 to a deep learner than to a traditional shallow and purely
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
596 supervised learner. More precisely,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
597 the answers are positive for all the questions asked in the introduction.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
598 %\begin{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
599
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
600 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
601 {\bf Do the good results previously obtained with deep architectures on the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
602 MNIST digits generalize to a much larger and richer (but similar)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
603 dataset, the NIST special database 19, with 62 classes and around 800k examples}?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
604 Yes, the SDA {\em systematically outperformed the MLP and all the previously
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
605 published results on this dataset} (the ones that we are aware of), {\em in fact reaching human-level
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
606 performance} at around 17\% error on the 62-class task and 1.4\% on the digits,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
607 and beating previously published results on the same data.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
608
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
609 $\bullet$ %\item
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
610 {\bf To what extent do self-taught learning scenarios help deep learners,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
611 and do they help them more than shallow supervised ones}?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
612 We found that distorted training examples not only made the resulting
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
613 classifier better on similarly perturbed images but also on
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
614 the {\em original clean examples}, and more importantly and more novel,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
615 that deep architectures benefit more from such {\em out-of-distribution}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
616 examples. MLPs were helped by perturbed training examples when tested on perturbed input
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
617 images (65\% relative improvement on NISTP)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
618 but only marginally helped (5\% relative improvement on all classes)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
619 or even hurt (10\% relative loss on digits)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
620 with respect to clean examples . On the other hand, the deep SDAs
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
621 were significantly boosted by these out-of-distribution examples.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
622 Similarly, whereas the improvement due to the multi-task setting was marginal or
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
623 negative for the MLP (from +5.6\% to -3.6\% relative change),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
624 it was quite significant for the SDA (from +13\% to +27\% relative change),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
625 which may be explained by the arguments below.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
626 %\end{itemize}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
627
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
628 In the original self-taught learning framework~\citep{RainaR2007}, the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
629 out-of-sample examples were used as a source of unsupervised data, and
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
630 experiments showed its positive effects in a \emph{limited labeled data}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
631 scenario. However, many of the results by \citet{RainaR2007} (who used a
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
632 shallow, sparse coding approach) suggest that the {\em relative gain of self-taught
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
633 learning vs ordinary supervised learning} diminishes as the number of labeled examples increases.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
634 We note instead that, for deep
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
635 architectures, our experiments show that such a positive effect is accomplished
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
636 even in a scenario with a \emph{large number of labeled examples},
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
637 i.e., here, the relative gain of self-taught learning is probably preserved
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
638 in the asymptotic regime.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
639
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
640 {\bf Why would deep learners benefit more from the self-taught learning framework}?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
641 The key idea is that the lower layers of the predictor compute a hierarchy
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
642 of features that can be shared across tasks or across variants of the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
643 input distribution. A theoretical analysis of generalization improvements
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
644 due to sharing of intermediate features across tasks already points
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
645 towards that explanation~\cite{baxter95a}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
646 Intermediate features that can be used in different
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
647 contexts can be estimated in a way that allows to share statistical
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
648 strength. Features extracted through many levels are more likely to
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
649 be more abstract and more invariant to some of the factors of variation
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
650 in the underlying distribution (as the experiments in~\citet{Goodfellow2009} suggest),
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
651 increasing the likelihood that they would be useful for a larger array
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
652 of tasks and input conditions.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
653 Therefore, we hypothesize that both depth and unsupervised
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
654 pre-training play a part in explaining the advantages observed here, and future
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
655 experiments could attempt at teasing apart these factors.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
656 And why would deep learners benefit from the self-taught learning
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
657 scenarios even when the number of labeled examples is very large?
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
658 We hypothesize that this is related to the hypotheses studied
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
659 in~\citet{Erhan+al-2010}. In~\citet{Erhan+al-2010}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
660 it was found that online learning on a huge dataset did not make the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
661 advantage of the deep learning bias vanish, and a similar phenomenon
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
662 may be happening here. We hypothesize that unsupervised pre-training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
663 of a deep hierarchy with self-taught learning initializes the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
664 model in the basin of attraction of supervised gradient descent
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
665 that corresponds to better generalization. Furthermore, such good
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
666 basins of attraction are not discovered by pure supervised learning
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
667 (with or without self-taught settings) from random initialization, and more labeled examples
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
668 does not allow the shallow or purely supervised models to discover
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
669 the kind of better basins associated
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
670 with deep learning and self-taught learning.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
671
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
672 A Flash demo of the recognizer (where both the MLP and the SDA can be compared)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
673 can be executed on-line at {\tt http://deep.host22.com}.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
674
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
675 \iffalse
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
676 \section*{Appendix I: Detailed Numerical Results}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
677
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
678 These tables correspond to Figures 2 and 3 and contain the raw error rates for each model and dataset considered.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
679 They also contain additional data such as test errors on P07 and standard errors.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
680
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
681 \begin{table}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
682 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits +
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
683 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
684 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
685 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07)
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
686 and using a validation set to select hyper-parameters and other training choices.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
687 \{SDA,MLP\}0 are trained on NIST,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
688 \{SDA,MLP\}1 are trained on NISTP, and \{SDA,MLP\}2 are trained on P07.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
689 The human error rate on digits is a lower bound because it does not count digits that were
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
690 recognized as letters. For comparison, the results found in the literature
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
691 on NIST digits classification using the same test set are included.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
692 \label{tab:sda-vs-mlp-vs-humans}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
693 \begin{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
694 \begin{tabular}{|l|r|r|r|r|} \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
695 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
696 Humans& 18.2\% $\pm$.1\% & 39.4\%$\pm$.1\% & 46.9\%$\pm$.1\% & $1.4\%$ \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
697 SDA0 & 23.7\% $\pm$.14\% & 65.2\%$\pm$.34\% & 97.45\%$\pm$.06\% & 2.7\% $\pm$.14\%\\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
698 SDA1 & 17.1\% $\pm$.13\% & 29.7\%$\pm$.3\% & 29.7\%$\pm$.3\% & 1.4\% $\pm$.1\%\\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
699 SDA2 & 18.7\% $\pm$.13\% & 33.6\%$\pm$.3\% & 39.9\%$\pm$.17\% & 1.7\% $\pm$.1\%\\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
700 MLP0 & 24.2\% $\pm$.15\% & 68.8\%$\pm$.33\% & 78.70\%$\pm$.14\% & 3.45\% $\pm$.15\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
701 MLP1 & 23.0\% $\pm$.15\% & 41.8\%$\pm$.35\% & 90.4\%$\pm$.1\% & 3.85\% $\pm$.16\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
702 MLP2 & 24.3\% $\pm$.15\% & 46.0\%$\pm$.35\% & 54.7\%$\pm$.17\% & 4.85\% $\pm$.18\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
703 \citep{Granger+al-2007} & & & & 4.95\% $\pm$.18\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
704 \citep{Cortes+al-2000} & & & & 3.71\% $\pm$.16\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
705 \citep{Oliveira+al-2002} & & & & 2.4\% $\pm$.13\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
706 \citep{Milgram+al-2005} & & & & 2.1\% $\pm$.12\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
707 \end{tabular}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
708 \end{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
709 \end{table}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
710
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
711 \begin{table}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
712 \caption{Relative change in error rates due to the use of perturbed training data,
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
713 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
714 A positive value indicates that training on the perturbed data helped for the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
715 given test set (the first 3 columns on the 62-class tasks and the last one is
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
716 on the clean 10-class digits). Clearly, the deep learning models did benefit more
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
717 from perturbed training data, even when testing on clean data, whereas the MLP
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
718 trained on perturbed data performed worse on the clean digits and about the same
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
719 on the clean characters. }
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
720 \label{tab:perturbation-effect}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
721 \begin{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
722 \begin{tabular}{|l|r|r|r|r|} \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
723 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
724 SDA0/SDA1-1 & 38\% & 84\% & 228\% & 93\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
725 SDA0/SDA2-1 & 27\% & 94\% & 144\% & 59\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
726 MLP0/MLP1-1 & 5.2\% & 65\% & -13\% & -10\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
727 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
728 \end{tabular}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
729 \end{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
730 \end{table}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
731
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
732 \begin{table}[ht]
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
733 \caption{Test error rates and relative change in error rates due to the use of
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
734 a multi-task setting, i.e., training on each task in isolation vs training
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
735 for all three tasks together, for MLPs vs SDAs. The SDA benefits much
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
736 more from the multi-task setting. All experiments on only on the
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
737 unperturbed NIST data, using validation error for model selection.
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
738 Relative improvement is 1 - single-task error / multi-task error.}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
739 \label{tab:multi-task}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
740 \begin{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
741 \begin{tabular}{|l|r|r|r|} \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
742 & single-task & multi-task & relative \\
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
743 & setting & setting & improvement \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
744 MLP-digits & 3.77\% & 3.99\% & 5.6\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
745 MLP-lower & 17.4\% & 16.8\% & -4.1\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
746 MLP-upper & 7.84\% & 7.54\% & -3.6\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
747 SDA-digits & 2.6\% & 3.56\% & 27\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
748 SDA-lower & 12.3\% & 14.4\% & 15\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
749 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
750 \end{tabular}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
751 \end{center}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
752 \end{table}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
753
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
754 \fi
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
755
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
756 %\afterpage{\clearpage}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
757 %\clearpage
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
758 {
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
759 %\bibliographystyle{spbasic} % basic style, author-year citations
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
760 \bibliographystyle{plainnat}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
761 \bibliography{strings,strings-short,strings-shorter,ift6266_ml,specials,aigaion-shorter}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
762 %\bibliographystyle{unsrtnat}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
763 %\bibliographystyle{apalike}
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
764 }
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
765
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
766
5ab605c9a7d9 NIPS deep learning workshop submission new .tex file compressed to 8 pages
boulanni <nicolas_boulanger@hotmail.com>
parents:
diff changeset
767 \end{document}