poster.tex 14.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
\documentclass[final]{beamer}
\usepackage{grffile}
\mode<presentation>{\usetheme{IST}}
\usepackage[english]{babel}
\usepackage{setspace}
%\usepackage[latin1]{inputenc}
%\usepackage[T1]{fontenc}
\usepackage[scaled]{helvet}

%\usepackage{algorithm}
%\usepackage{algorithmic}
%\usepackage{framed}
%\usepackage{mathrsfs}
%\usepackage{stmaryrd} % for /llbracket, /rrbracket
%\usepackage{rotating}
%\usepackage[nointegrals]{wasysym} % for \smiley, \frownie
\usepackage{marvosym}  % \Smiley{} \Frowny{}

\usepackage{pifont}
\usepackage{tikz}
\usetikzlibrary{chains,fit,shapes}
\newcommand{\mytoprule}{\specialrule{1em}{0em}{0em}}
\newcommand{\mybottomrule}{\specialrule{1em}{0em}{0em}} 

\newcommand{\bphi}{\boldmath{\phi}}

\usepackage{array}
\newcolumntype{L}[1]{>{\raggedright\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}
\newcolumntype{C}[1]{>{\centering\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}
\newcolumntype{R}[1]{>{\raggedleft\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}

\usepackage{amsmath,amsthm, amssymb, latexsym}
%\usepackage{times}\usefonttheme{professionalfonts}  % obsolete
%\usefonttheme[onlymath]{serif}
\boldmath
\usepackage[orientation=portrait,size=a0,scale=1.4, debug]{beamerposter}
% change list indention level
% \setdefaultleftmargin{3em}{}{}{}{}{}

%\setlength{\oddsidemargin}{.5cm} % adjust as necessary

%\usepackage{snapshot} % will write a .dep file with all dependencies, allows for easy bundling

%\usepackage{array,booktabs,tabularx}
%\newcolumntype{Z}{>{\centering\arraybackslash}X} % centered tabularx columns
%\newcommand{\pphantom}{\textcolor{lightgray}} % phantom introduces a vertical space in p formatted table columns??!!

\usepackage{graphicx}
\usepackage{subfigure}
\usepackage{amsmath,amssymb,stmaryrd}
%\usepackage{hyperref}
%
\usepackage{color, colortbl}

%\listfiles
\definecolor{redbg}{rgb}{0.8,0.08,0.1}
\newenvironment<>{redblock}[1]{%
  \begin{actionenv}#2%
      \def\insertblocktitle{#1}%
      \par%
      \mode<presentation>{%
        \setbeamercolor{block title}{fg=white,bg=redbg}
       \setbeamercolor{block body}{fg=black,bg=white}
       \setbeamertemplate{itemize item}[default]
     }%
      \usebeamertemplate{block begin}}
    {\par\usebeamertemplate{block end}\end{actionenv}}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\graphicspath{{images/}}
\setbeamertemplate{itemize item}[default]

%\title{\LARGE iCaRL: incremental Classifier and Representation Learning}
\title{Incremental Classifier and Representation Learning}
\author{\large Sylvestre-Alvise Rebuffi$^{\dag,*}$, Alexander Kolesnikov$^{*}$, Christoph H. Lampert$^{*}$}
\institute{\vskip-.5\baselineskip\large $^{\dag}$ CentraleSup\'elec\qquad $^{*}$ IST Austria}

%\institute{~}%Christoph Lampert} %\textsuperscript{1} ENS Rennes (Ecole Normale Sup\'{e}rieure de Rennes), Rennes, France \textsuperscript{2} IST Austria (Institute of Science and Technology Austria), Klosterneuburg, Austria}
%\date[]{}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\newlength{\columnheight}
\setlength{\columnheight}{95cm}
\setlength{\columnsep}{1cm}

\renewcommand{\P}{\pmb{\mathbb{P}}}
\newcommand{\R}{\pmb{\mathbb{R}}}
\newcommand{\Rcal}{\pmb{\mathcal{R}}}
\renewcommand{\H}{\pmb{\mathbb{H}}}
\newcommand{\Hcal}{\pmb{\mathcal{Hcal}}}

\newcommand{\E}{\mathcal{E}}
\newcommand{\F}{\mathcal{F}}
\newcommand{\G}{\mathcal{G}}
\renewcommand{\L}{\mathcal{L}}
\renewcommand{\P}{\mathbb{P}}
\newcommand{\V}{V} % {\mathcal{V}} 
\renewcommand{\S}{\mathcal{S}}
\newcommand{\X}{\mathcal{X}}
\newcommand{\x}{\mathbf{x}}
\newcommand{\Y}{\mathcal{Y}}
\newcommand{\Z}{\mathcal{Z}}
\usepackage{xspace}

\newcommand{\argmax}{\operatorname*{argmax}}
\newcommand{\argmin}{\operatorname*{argmin}}

\newcommand{\res}[2] {{\normalsize #1~$\pm$~#2}}
\newcommand{\resbold}[2] {{\textbf{\normalsize #1~$\pm$~#2}}}
\newcommand\nounfmt[1]{\textit{#1}}
\newcommand\imagenetfmt[1]{\underline{\textit{#1}}}
% abbr. with correct spacing
\makeatletter
\DeclareRobustCommand\onedot{\futurelet\@let@token\@onedot}
\def\@onedot{\ifx\@let@token.\else.\null\fi\xspace}
\def\iid{{i.i.d}\onedot}\def\Iid{{I.i.d}\onedot}
\def\eg{{e.g}\onedot} \def\Eg{{E.g}\onedot}
\def\ie{{i.e}\onedot} \def\Ie{{I.e}\onedot}
\def\cf{{c.f}\onedot} \def\Cf{{C.f}\onedot}
\def\etc{{etc}\onedot} \def\vs{{vs}\onedot}
\def\wrt{w.r.t\onedot} \def\dof{d.o.f\onedot}
\def\etal{\emph{et al}\onedot}
\makeatother

\setlength{\fboxrule}{5pt}

\definecolor{darkcyan}{RGB}{0,192,192}
\definecolor{darkyellow}{RGB}{192,192,0}
\definecolor{darkred}{RGB}{128,0,64}

\newlength{\blockskip}
\setlength{\blockskip}{.5\baselineskip}

\newcommand<>{\blue}[1]{\textcolor#2{blue}{#1}}
\newcommand<>{\bblue}[1]{\textcolor#2{blue}{\bf #1}}

\begin{document}
\begin{frame}

\vspace*{-1.5cm}

\ \ \begin{block}{\Large Abstract}
%\large 
We introduce \bblue{iCaRL}, a method for simultaneously learning classifiers 
and a feature representation from training data in which classes 
occur incrementally. 

iCaRL uses a \blue{nearest-mean-of-exemplars} classifier, \blue{herding for 
adaptive exemplar selection} and \blue{distillation for representation learning
without catastrophic forgetting}.
%
Experiments on CIFAR and ILSVRC\,2012 show that iCaRL can learn  incrementally over a long period of time where other methods quickly fail.
\end{block}

\vskip-1cm
\begin{columns}[t]
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%% First COlumn
\ \ \ \begin{column}{.49\textwidth}
\begin{block}{\Large 1) Motivation}

\bigskip
Object categorization methods are trained to recognize \textbf{1000s of classes}:

\medskip
\centerline{\includegraphics[width=.9\textwidth]{2010-imagenet}}%
\medskip
\blue{Standard training requires:}
\begin{itemize}
\item random order access to 100s of GB of training data,
\item many days to weeks of training time.
\end{itemize}

\bigskip

\medskip
\centerline{\includegraphics[width=.9\textwidth]{2010-imagenet-new}}%
\medskip

\blue{What, if a few new classes are meant to be included?}
\begin{itemize}
\item training must be re-run for all classes 

$\rightarrow$ huge computational cost, all training data must be kept around {\color{orange}\large \Frowny{}}
\end{itemize}

\bigskip
Potential solution: \bblue{class-incremental learning}
\end{block}

\vskip4\blockskip
\begin{block}{\Large 3) Existing Approaches}
\textbf{Fixed data representation:}
\begin{itemize}
\item retrain classifiers on data subset with biased regularization {\scriptsize [Kuzborskij \etal, 2013]}
\item represent classes by \blue{mean feature vectors} {\scriptsize [Mensink \etal, 2012], [Ristin \etal, 2014]}
\end{itemize}
\textbf{Learning the data representation:}
\begin{itemize}
\item grow neural network incrementally, fixing parts that are responsible for earlier class decisions
{\scriptsize [Mandziuk, Shastri. 1998], \dots, [Rusu \etal, 2016]}

\item continuously generate patterns to prevent forgetting {\scriptsize [Ans, Rousset. 2000]}

\item multi-task setting: preserve network activations by \blue{distillation} {\scriptsize [Li, Hoiem. 2016]}
\end{itemize}
\end{block}

\vskip4\blockskip
\begin{block}{\Large 4) iCaRL} % {\scriptsize [arXiv \dots]}}
We incrementally learn \blue{classifiers and features} with a fixed-size network.

%Notation:
\begin{itemize}
\item $f^1(x),\dots,f^T(x)$: probabilistic network outputs for (up to) $T$ classes
\item $\bphi$: (current) feature representation defined by the network
%\item $K$: size of extra memory that we can use to store images/exemplars  
\item $t$: number of classes observed so far
\item $P^y=(p^y_1,\dots,p^y_{m})$ set of exemplar images for class $y=1,\dots,t$
\end{itemize}

\bigskip
\bblue{iCaRL component 1: exemplar-based classification.}
\begin{itemize}
\item a new image, $x$, is classified by the \blue{nearest-mean-of-exemplars} rule
$$ y^\ast = \operatorname*{argmin}_{y=1,\dots,t} \Big\|\bphi(x) - \mu^y\Big\|
\qquad\text{for }\ 
\mu^y=\frac{1}{m}\sum_{j=1}^{m}\bphi(p^y_j).$$
\end{itemize}

\bigskip
\bblue{iCaRL component 2: representation learning.}

For new data $X^y=\{x^y_1,\dots,x^y_{n_y}\}$ of classes $y=t\!+\!1,\dots,t'$
\begin{itemize}
\item create training set of training examples and exemplars
$$\mathcal{D} \leftarrow \bigcup_{y=t+1,\dots,t'}\!\!\!\{(x,y) : x\in X^y\} 
\ \cup\!\!\bigcup_{y=1,\dots,t} \!\!\!\{(x,y) : x\in P^y\} $$
\item for all $x_i\in\mathcal{D}$, store network outputs $a^y_i = f^y(x_i)$ of classes $y=1,\dots,t$ 
\item update the network parameters, $\theta$, using BackProp on loss function
$$\ell(\theta) = -\!\!\!\!\!\!\sum_{(x_i,y_i)\in\mathcal{D}}\!\!\!\![\ \log( f^{y_i}(x_i;\theta)+\sum_{y=1}^{t}\, a^y_i\log( f^y(x_i;\theta) )\ ].$$
\end{itemize}

\bigskip
\bblue{iCaRL component 3: exemplar selection.}

When the number of observed classes increases from $t$ to $t'$: set $m'=\frac{K}{t'}$.
\begin{itemize}
\item for classes $y=1,\dots,t$, keep exemplars $p^y_{1},\dots,p^y_{m'}$, discard others % $p^j_{m'_j+1},\dots,p^j_{m_j}$
\item for classes $y=t\!+\!1,\dots,t'$, find new exemplars $p^y_{1},\dots,p^y_{m'}$ as 
\begin{align*}
p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y}\!\!\bphi(x^y_i) - \frac{1}{k}\Big[\bphi(x)+\sum_{j=1}^{k-1}\!\!\bphi(p^y_j)\Big] \Big\|
\ \text{for $k=1,\dots,m'$}.
\end{align*}
%\begin{align*}
%\operatorname*{argmin}_{|I|=m'_j} \Big\|  \frac{1}{n_j}\sum_{i=1}^{n_j}\bphi(x^j_i) - \frac{1}{m'_j}\sum_{i\in I}\bphi(x^j_i) \Big\|
%\end{align*}
\end{itemize}
\end{block}

%
%
    %\STATE // form combined training set:
    %%
    %\STATE // store network outputs with pre-update parameters:
    %\FOR{$j=1,\dots,s-1$}
    %\STATE $a^j_i\leftarrow f^j(x_i)$ \quad for all $(x_i,\cdot)\in\D$
    %\ENDFOR
    %%\STATE run network training (\eg BackProp) with loss function
    %\begin{align*}
    %\ell(\Theta) = -\!\!\!\!\!\!\sum_{(x_i,y_i)\in\D}&\!\!\!\!\!\!\big[\log( f^{y_i}(x_i))+\sum_{j=1}^{s-1}\, a^j_i\log( f^j(x_i) )\big]
    %\end{align*}
    %\textcolor{red}{\textbf{is this the right way to handle EXEMPLARS???}}


\end{column}
%
\ \ \begin{column}{.495\textwidth}
\begin{block}{\Large 2) Class-Incremental Learning}
\textbf{Situation:}
\begin{itemize}
\item classes appear sequentially (or in batches) % $c_1,c_2,\dots,c_T$
\item each class $y$ has training images $X^y=\{x^y_1,\dots,x^y_{n_y}\}$
\end{itemize}

\bigskip
\textbf{We want to/we can:}
\begin{itemize}
\item for any number of observed classes, $t$, learn a multi-class classifier% for $c_1,\dots,c_t$
\item store a certain number, $K$, of images (a few hundreds or thousands)
\end{itemize}

\bigskip
\textbf{We do not want to/we cannot:}
\begin{itemize}
\item retrain from scratch whenever new classes become available
\item store all training examples (could be millions)
\end{itemize}
\bigskip
\textbf{The dilemma:}
\begin{itemize}
  \item \textcolor{blue}{fixing the data representation}: suboptimal results on new classes. % {\tiny [Tommasi \etal, 2015]}.
  \item \textcolor{blue}{continuously improving the representation}: classifiers for earlier classes deteriorate over time 
  ("catastrophic forgetting/interference").{\scriptsize [McCloskey, Cohen. 1989]} % (catastrophic forgetting) {\tiny []}.
  % Ratcliff, R. (1990) Connectionist models of recognition memory: Constraints imposed by learning and forgetting functions. Psychological Review,97, 285-308
\end{itemize}
\end{block}

\vskip4\blockskip
\begin{block}{\Large 5) Experiments (excerpt)}

\vskip4\blockskip
\mbox{
\parbox{.48\textwidth}{ 
\!\!\!\textbf{CIFAR-100:}
\begin{itemize}
\item 100 classes, in batches of 10
\item 32-layer ResNet {\scriptsize [He \etal, 2015]}
\item evaluated by top-1 accuracy 
\item number of exemplars: 2000
\end{itemize}
}
\parbox{.48\textwidth}{ 
\textbf{ImageNet ILSVRC 2012 (subset):}
\begin{itemize}
\item 100 classes, in batches of 10
\item 18-layer ResNet {\scriptsize [He \etal, 2015]}
\item evaluated by top-5 accuracy 
\item number of exemplars: 2000
\end{itemize}
}}

\vskip4\blockskip
\textbf{Baselines:}
\begin{itemize}
\item fixed representation: freeze representation after first batch of classes
\item finetuning: ordinary NN learning, finetune whenever new classes come in
\item LwF: \emph{"Learning without Forgetting"} {\scriptsize [Li, Hoiem. 2016]}, use network itself to classify
%\item LwF+proto: like LwF, but with prototypes used for representation learning
\item iNCM: like iCaRL, but store all images and classify with true class means
\end{itemize}
\end{block}

\vskip4\blockskip
\begin{block}{\Large 6) Results (excerpt)}
  \centerline{\qquad\qquad\qquad\textbf{CIFAR-100}\qquad\hfill\textbf{ImageNet ILSVRC~2012}\quad\ \ }
  \centerline{\includegraphics[height=.3\textwidth]{cifar-cumul10-legend}\includegraphics[height=.3\textwidth]{imagenet-cumul10_top5}}

\medskip
\textbf{Discussion:}
\begin{itemize}
\item as expected: fixed representation and finetuning do not work well 
\item iCaRL is able to keep good classification accuracy for many iterations  
\item \emph{"Learning without Forgetting"} starts to forget earlier %(even with prototypes)  
\item mean-of-exemplars on par with (intractable) iNCM
\end{itemize}
    \end{block}

\vskip4\blockskip
\begin{block}{\Large 7) Summary}
\begin{itemize}
\item iCaRL learns incrementally with a fixed memory footprint
\item much better results than baselines, on par with (intractable) iNCM 
\end{itemize}
\end{block}

\bigskip\hrule\medskip\tiny

%[Thrun \etal, "Learning one more thing", \dots]

[Ans, Rousset. \emph{"Neural networks with a self-refreshing memory: Knowledge transfer in sequential learning tasks without catastrophic forgetting"}, Connection Science 12(1), 2000]

[He, Zhang, Ren, Sun. \emph{"Deep residual learning for image recognition"}. arXiv:1512.03385, 2015]

[Hinton, Vinyals, Dean. \emph{"Distilling the Knowledge in a Neural Network"}, NIPS Workshop on Deep Learning, 2014]

[Kuzborskij, Orabona, Caputo. \emph{"From N to N+1: Multiclass transfer incremental learning"}, CVPR 2013]

[Mandziuk, Shastri. \emph{"Incremental class learning approach and its application to handwritten digit recognition"}, Information Sciences, 2002]

[McCloskey, Cohen. \emph{"Catastrophic interference in connectionist networks: The sequential learning problem"}, The Psychology of Learning and Motivation, 1989]

[Mensink, Verbeek, Perronnin, Csurka. \emph{"Distance-based image classification: Generalizing to new Classes at near-zero cost"}, 2013]

[Li, Hoiem. \emph{"Learning without forgetting"}, ECCV 2016]

[Ristin, Huillaumin, Gall, van Gool. \emph{"Incremental learning of NCM forests for large-scale image classification"}, CVPR 2014]

[Rusu, Rabinowitz, Desjardins, Soyer, Kirkpatrick, Kavukcuoglu, Pascanu, Hadsell. \emph{"Progressive neural networks"}, arXiv:1606.04671 [cs.LG] 2016]

%[Ans, "Sequential Learning in Distributed Neural Networks without Catastrophic Forgetting: A Single and Realistic Self-Refreshing Memory Can Do It", Neural Information Processing--Letters and Reviews. 2004]
%[He, Zhang, Ren, Sun. Deep residual learning for image recognition. CVPR 2016]
%[Rocco De Rosa, Thomas Mensink, and Barbara Caputo, "Online Open World Recognition"] arXiv
\end{column}
\end{columns}
 
\end{frame}


\end{document}