Commit 09af905b authored by Amelie Royer's avatar Amelie Royer

modify Poster skeleton

parent 3d5fb1b6
...@@ -71,9 +71,9 @@ ...@@ -71,9 +71,9 @@
\setbeamertemplate{itemize item}[default] \setbeamertemplate{itemize item}[default]
%\title{\LARGE iCaRL: incremental Classifier and Representation Learning} %\title{\LARGE iCaRL: incremental Classifier and Representation Learning}
\title{Incremental Classifier and Representation Learning} \title{Computer Vision and Machine Learning}
\author{\large Sylvestre-Alvise Rebuffi$^{\dag,*}$, Alexander Kolesnikov$^{*}$, Christoph H. Lampert$^{*}$} \author{}
\institute{\vskip-.5\baselineskip\large $^{\dag}$ CentraleSup\'elec\qquad $^{*}$ IST Austria} \institute{\vskip-.5\baselineskip\large Institute of Science and Technology (IST) Austria, 3400 Klosterneuburg, 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} %\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[]{} %\date[]{}
...@@ -139,17 +139,19 @@ ...@@ -139,17 +139,19 @@
\vspace*{-1.5cm} \vspace*{-1.5cm}
\ \ \begin{block}{\Large Abstract} \begin{block}{\Large People}
%\large \newcommand{\peopleheight}{6cm}
We introduce \bblue{iCaRL}, a method for simultaneously learning classifiers \begin{center}
and a feature representation from training data in which classes \includegraphics[height=\peopleheight{}]{people/clampert.jpg} ~~
occur incrementally. \includegraphics[height=\peopleheight{}]{people/akolesnikov-new.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/nkonstan.jpg} ~~
iCaRL uses a \blue{nearest-mean-of-exemplars} classifier, \blue{herding for \includegraphics[height=\peopleheight{}]{people/apentina-new.jpg} ~~
adaptive exemplar selection} and \blue{distillation for representation learning \includegraphics[height=\peopleheight{}]{people/bphuong.jpg} ~~
without catastrophic forgetting}. \includegraphics[height=\peopleheight{}]{people/srebuffi.jpg} ~~
% \includegraphics[height=\peopleheight{}]{people/aroyer.jpg} ~~
Experiments on CIFAR and ILSVRC\,2012 show that iCaRL can learn incrementally over a long period of time where other methods quickly fail. \includegraphics[height=\peopleheight{}]{people/gsperl.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/azimin.jpg} ~~
\end{center}
\end{block} \end{block}
\vskip-1cm \vskip-1cm
...@@ -157,7 +159,7 @@ Experiments on CIFAR and ILSVRC\,2012 show that iCaRL can learn incrementally o ...@@ -157,7 +159,7 @@ Experiments on CIFAR and ILSVRC\,2012 show that iCaRL can learn incrementally o
% %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%% First COlumn %%%%%%%%%%%%%%%%%%%%%%%%%%%%% First COlumn
\ \ \ \begin{column}{.49\textwidth} \ \ \ \begin{column}{.49\textwidth}
\begin{block}{\Large 1) Motivation} \begin{block}{\Large Multi-Task Learning}
\bigskip \bigskip
Object categorization methods are trained to recognize \textbf{1000s of classes}: Object categorization methods are trained to recognize \textbf{1000s of classes}:
...@@ -188,8 +190,9 @@ $\rightarrow$ huge computational cost, all training data must be kept around {\c ...@@ -188,8 +190,9 @@ $\rightarrow$ huge computational cost, all training data must be kept around {\c
Potential solution: \bblue{class-incremental learning} Potential solution: \bblue{class-incremental learning}
\end{block} \end{block}
\vskip4\blockskip \vskip4\blockskip
\begin{block}{\Large 3) Existing Approaches} \begin{block}{\Large Conditional Risk}
\textbf{Fixed data representation:} \textbf{Fixed data representation:}
\begin{itemize} \begin{itemize}
\item retrain classifiers on data subset with biased regularization {\scriptsize [Kuzborskij \etal, 2013]} \item retrain classifiers on data subset with biased regularization {\scriptsize [Kuzborskij \etal, 2013]}
...@@ -207,7 +210,7 @@ Potential solution: \bblue{class-incremental learning} ...@@ -207,7 +210,7 @@ Potential solution: \bblue{class-incremental learning}
\end{block} \end{block}
\vskip4\blockskip \vskip4\blockskip
\begin{block}{\Large 4) iCaRL} % {\scriptsize [arXiv \dots]}} \begin{block}{\Large iCaRL} % {\scriptsize [arXiv \dots]}}
We incrementally learn \blue{classifiers and features} with a fixed-size network. We incrementally learn \blue{classifiers and features} with a fixed-size network.
%Notation: %Notation:
...@@ -275,8 +278,10 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y} ...@@ -275,8 +278,10 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y}
\end{column} \end{column}
% %
\ \ \begin{column}{.495\textwidth} \ \ \begin{column}{.495\textwidth}
\begin{block}{\Large 2) Class-Incremental Learning} \begin{block}{\Large Multi-output Distillation}
\textbf{Situation:} \textbf{Situation:}
\begin{itemize} \begin{itemize}
\item classes appear sequentially (or in batches) % $c_1,c_2,\dots,c_T$ \item classes appear sequentially (or in batches) % $c_1,c_2,\dots,c_T$
...@@ -307,7 +312,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y} ...@@ -307,7 +312,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y}
\end{block} \end{block}
\vskip4\blockskip \vskip4\blockskip
\begin{block}{\Large 5) Experiments (excerpt)} \begin{block}{\Large Flexible Fine-tuning}
\vskip4\blockskip \vskip4\blockskip
\mbox{ \mbox{
...@@ -341,20 +346,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y} ...@@ -341,20 +346,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y}
\end{itemize} \end{itemize}
\end{block} \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 \vskip4\blockskip
\begin{block}{\Large 7) Summary} \begin{block}{\Large 7) Summary}
...@@ -364,6 +356,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y} ...@@ -364,6 +356,7 @@ p^y_k \leftarrow\!\argmin\limits_{x\in X^y} \Big\| \frac{1}{n_y}\sum_{i=1}^{n_y}
\end{itemize} \end{itemize}
\end{block} \end{block}
\bigskip\hrule\medskip\tiny \bigskip\hrule\medskip\tiny
%[Thrun \etal, "Learning one more thing", \dots] %[Thrun \etal, "Learning one more thing", \dots]
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment