Présentation Sorbonne Université
Author
Robert
Last Updated
4年前
License
Creative Commons CC BY 4.0
Abstract
Template Sorbonne Universite
Template Sorbonne Universite
\documentclass[compress,xcolor=table]{beamer}
% Packages
\usepackage[english]{babel}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{datetime}
% Possible options of the package (add/remove below in \usetheme call):
% - nosectionpages: no pages between sections
% - flama: use flama font, requires xelatex/lualatex + the font to compile
% - compressminiframes: put the heading list bullets indications pages on 1 line
\usetheme[compressminiframes]{sorbonne}
% Title page
\title{Lorem Ipsum \newline Dolor Sit Amet}
\foottitle{Lorem Ipsum Dolor Sit Amet} % optional, printed at the bottom of the slides, by default same as title, can be useful to rewrite when title has a newline for example
\subtitle{Subtitle} % optional subtitle
\date{\formatdate{22}{03}{2018}}
\author{Prénom Nom}
% \institute{LIP6 - Sorbonne Université} % Optional
% Biblatex
\usepackage[backend=bibtex, style=authoryear, citestyle=authoryear]{biblatex}
\bibliography{library.bib}
\renewcommand*{\bibfont}{\footnotesize}
%%%%
%% BEGIN OF SLIDES
%%%%
\begin{document}
\begin{frame}[plain]
\titlepage
\setcounter{framenumber}{0}
\end{frame}
\section{Noise filtering in UPMC Food-101} \subsection{}
\begin{frame}{Principle}
\begin{alertblock}{The problem}
\begin{itemize}
\item
UPMC Food-101 has been crawled from Google Images
\item
It contains a certain amount of noise
\end{itemize}
\end{alertblock}
\begin{block}{The idea}
\begin{itemize}
\item
Creating bags from images of 1 class (e.g.~pizza)
\item
Creating bags for ``rest'' class
\item
Defining the expected level of noise in the pizza bags
\item
Using \emph{Learning with Label Proportions} models to detect noise
\end{itemize}
\end{block}
\end{frame}
\begin{frame}{Experimentations}
\begin{block}{Protocol}
\begin{itemize}
\item
\textbf{Dataset:}
\((x_i, y_i, y_i^*) \in \mathbb{R}^p \times \{-1,1\} \times \{-1,1\},\quad i=1..n\)
\begin{itemize}
\item
\(x_i\) features, \(y_i\) noisy label, \(y_i^*\) true label (not
available for training)
\end{itemize}
\item
\textbf{Create bags:} Create bags \(b_j\) of 30 points having the same
\(y_i\) and give them a proportion of positive points \(p_j\)
\item
\textbf{Training:} Train the SyMIL model on bags and train an SVM on
\((b_j, p_j)\) / \((x_i, y_i)\)
\item
\textbf{Evaluation:} Use the decision frontiers of SyMIL / SVM models
to reclassify each \(x_i\)s by predicting \(\hat{y}_i^*\), and compare
with \(y_i^*\)
\end{itemize}
\end{block}
\end{frame}
\begin{frame}{Some papers for noisy images datasets}
\begin{exampleblock}{\citetitle{Azadi2015} \cite{Azadi2015}}
\begin{itemize}
\item
Define a regularized loss for training the CNN
\item
Can be seen as looking for the label of similar images for
regularization
\item
Results slightly better than Sukhbaatar model
\end{itemize}
\end{exampleblock}
\end{frame}
\section{References} \subsection{}
\begin{frame}[allowframebreaks]{References}
\printbibliography[heading=none]
\end{frame}
\end{document}