Commit a34b7bf1 authored by Christoph Lampert's avatar Christoph Lampert

more pictures

parent 015f4892
......@@ -95,10 +95,12 @@
%\vskip12ex
\vskip9ex
\centering%\raggedleft
\usebeamercolor{title in headline}{\color{fg}\textbf{\huge {\hfill \inserttitle}}\\[1ex]}
\begin{center}
%\centering%\raggedleft
\usebeamercolor{title in headline}{\color{fg}\textbf{\huge {\inserttitle}}\\[1ex]} % CENTER
\usebeamercolor{author in headline}{\color{fg}\LARGE {\insertauthor}\\[1ex]}
\usebeamercolor{institute in headline}{\color{fg}\large{\insertinstitute}\\[1ex]}
\end{center}
\end{column}
\begin{column}{.01\paperwidth}
\end{column}
......
......@@ -6,6 +6,8 @@
%\usepackage[latin1]{inputenc}
%\usepackage[T1]{fontenc}
\usepackage[scaled]{helvet}
\usepackage{pbox}
%\usepackage{algorithm}
%\usepackage{algorithmic}
......@@ -70,10 +72,43 @@
\graphicspath{{images/}}
\setbeamertemplate{itemize item}[default]
\newlength{\peopleheight}
\setlength{\peopleheight}{4cm}
\newlength{\peoplewidth}
\setlength{\peoplewidth}{4.5cm}
%\title{\LARGE iCaRL: incremental Classifier and Representation Learning}
\title{Computer Vision and Machine Learning}
\author{~}
\institute{\vskip-.5\baselineskip\large Institute of Science and Technology (IST) Austria, 3400 Klosterneuburg, Austria}
\title{\veryHuge Computer Vision and Machine Learning Group}
\author{\parbox{3.4\peoplewidth}{Members and Alumni:}\small\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/clampert.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/akolesnikov-new.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/nkonstan.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/apentina-new.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/bphuong.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/srebuffi.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/aroyer.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/gsperl.jpg}}
\parbox{\peoplewidth}{\includegraphics[height=\peopleheight{}]{people/azimin.jpg}}
\\[.5ex]
\parbox{3.4\peoplewidth}{~}\parbox{\peoplewidth}{Christoph}
\parbox{\peoplewidth}{Alex}
\parbox{\peoplewidth}{Nikola}
\parbox{\peoplewidth}{Asya}
\parbox{\peoplewidth}{Mary}
\parbox{\peoplewidth}{Sylvestre}
\parbox{\peoplewidth}{Am\'elie}
\parbox{\peoplewidth}{Georg}
\parbox{\peoplewidth}{Alex}
\\
\parbox{3.4\peoplewidth}{~}\parbox{\peoplewidth}{Lampert}
\parbox{\peoplewidth}{Kolesnikov}
\parbox{\peoplewidth}{Konstantinov}
\parbox{\peoplewidth}{Pentina}
\parbox{\peoplewidth}{Phuong}
\parbox{\peoplewidth}{Rebuffi}
\parbox{\peoplewidth}{Royer}
\parbox{\peoplewidth}{Sperl}
\parbox{\peoplewidth}{Zimin}}
\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}
%\date[]{}
......@@ -146,33 +181,31 @@
\ \ \ \begin{column}{.49\textwidth}
\begin{block}{\Large Our Research}
\newcommand{\peopleheight}{4cm}
\begin{center}
\includegraphics[height=\peopleheight{}]{people/clampert.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/akolesnikov-new.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/nkonstan.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/apentina-new.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/bphuong.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/srebuffi.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/aroyer.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/gsperl.jpg} ~~
\includegraphics[height=\peopleheight{}]{people/azimin.jpg} ~~
\end{center}
\bigskip
Object categorization methods are trained to recognize \textbf{1000s of classes}:
\begin{itemize}
\item \LARGE Lots
\item of
\item Cool
\item Stuff
\end{itemize}
\end{block}
\begin{block}{\Large Example: Flexible Fine-tuning}
\begin{block}{\Large Example: Flex-tuning}
\input{finetuning.tex}
\end{block}
\begin{block}{\Large Example: Distillation}
\begin{block}{\Large Conditional Risk Minimization}
\begin{minipage}{.45\textwidth}
\textbf{Situation:}
\begin{itemize}
\item data for more and more classes appears sequentially % $c_1,c_2,\dots,c_T$
\end{itemize}
\begin{center}
yes%\includegraphics{dummy} ~~
\end{center}
\end{minipage}
%
\begin{minipage}{.5\textwidth}
\includegraphics[width=\textwidth]{lifelong}\qquad
\end{minipage}
\end{block}
......@@ -183,32 +216,44 @@ Object categorization methods are trained to recognize \textbf{1000s of classes}
\ \ \begin{column}{.495\textwidth}
\begin{block}{\Large Example: Class-incremental learning}
\begin{block}{\Large iCaRL (Incremental Classifier and Representation Learning) {\tiny [Rebuffi et al, CVPR 2017]}}
\begin{minipage}{.38\textwidth}
\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 $X^y=\{x^y_1,\dots,x^y_{n_y}\}$
\item data for more and more classes appears sequentially % $c_1,c_2,\dots,c_T$
\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 learn a multi-class classifier for all classes so far % $c_1,c_2,\dots,c_T$% for $c_1,\dots,c_t$
\item avoid \textbf{catastrophic forgetting}
\item store a certain number, $K$, of images (a few hundreds or thousands)
\end{itemize}
\begin{center}
\includegraphics[width=.5\columnwidth]{incremental}
\end{center}
\bigskip
\textbf{Suggestion: }
\end{minipage}
%
\begin{minipage}{.58\textwidth}
\includegraphics[width=\textwidth]{incremental}
\end{minipage}
\end{block}
\begin{block}{\Large Example: Multi-task Learning}
\begin{block}{\Large Multi-task Learning with Labeled and Unlabeled Tasks {\tiny [Pentina et al, ICML 2017]}}
\begin{minipage}{.45\textwidth}
\textbf{Situation:}
\begin{itemize}
\item data for more and more classes appears sequentially % $c_1,c_2,\dots,c_T$
\end{itemize}
\end{minipage}
%
\begin{minipage}{.5\textwidth}
\includegraphics[width=\textwidth]{asya-multitask}\qquad
\end{minipage}
\begin{center}
\includegraphics[width=.5\columnwidth]{asya-multitask-with-theorem}
\end{center}
\end{block}
\begin{block}{\Large Example: Multi-output Distillation}
......@@ -217,18 +262,15 @@ Object categorization methods are trained to recognize \textbf{1000s of classes}
\includegraphics{multi-output/architecture.pdf}
\end{column}
\begin{column}{0.28\textwidth}
\includegraphics[width=0.9\textwidth,height=.6\textwidth]{fine-tuning/selection_criterion_pacs.png}
\textbf{Multi-exit architectures}
\begin{itemize}
\item a crude initial prediction that is gradually refined
\item can be stopped anytime to provide a valid prediction
\end{itemize}
\textbf{Standard training}
\begin{itemize}
\item sum of exit-wise losses
\end{itemize}
\textbf{Proposed training}
\begin{itemize}
\item distillation from later (more accurate) to earlier exits
\item Distill from later (more accurate) to earlier exits
\end{itemize}
\end{column}
\end{columns}
......
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