Guillaume Couairon

Protein Design with Variational Auto Encoders

Tue, 28 Aug 2018 10:18:00 +0000

Abstract

TrackMaster - Reinforcement Learning for Race Cars

Sun, 25 Mar 2018 00:00:00 +0000

The work presented here is a school project with Thomas Gilles, for a course on Reinforcement Learning ^[1]. We present a new Reinforcement Learning environment, TrackEnv, to teach cars how to drive. We then use this environment to compare the performances of several Reinforcement Learning Agents : SARSA, Q-learning and deep Q-Learning.

Introduction

Robust Frank-Wolfe Algorithm for Minimum Enclosing Bregman Balls

Thu, 01 Mar 2018 00:00:00 +0000

This post gathers theoretical convergence & complexity results on the Frank-Wolfe algorithm extended to Bregman divergences (instead of euclidian distance) and made robust to outliers. It is the result of a collaboration with Amaury Sabran, supervised by Pr Frank Nielsen (LIX).

We propose a robust algorithm to approximate the slack minimum enclosing ball problem for a broad class of distance functions called Bregman Divergences.
The slack minimum enclosing ball problem is a convex relaxation of the minimum enclosing ball problem that allows outliers. Bregman divergences include not only the traditional (squared) Euclidean distance but also various divergence measures based on entropic functions. We apply this algorithm to the case of Euclidean MEB, MEB in a Reproducing Kernel Hilbert Space and MEB on the probability simplex with KL-divergences.

Introduction

Kernelized Minimum Enclosing Balls for Machine Learning

Thu, 01 Feb 2018 00:00:00 +0000

In this post, we are going to talk about a very famous problem in computational geometry: the Minimum Enclosing Ball (MEB) problem, which consists in finding the ball that contains a given set of points with minimum radius. We first describe how to compute arbitrarily fine approximations of Minimum
Enclosing Balls in a feature Hilbert space induced by a kernel using the Frank-Wolfe optimization algorithm, and show how to make these balls robust by allowing outliers. We then explain several applications of these robust kernelized MEBs in machine learning by presenting applications to supervised multiclass classification, unsupervised clustering, and deep generative models.

Project in collaboration with Amaury Sabran, supervised by Pr Frank Nielsen (LIX). Video available here.

Minimum Enclosing ball

Deep Segmentation of Road Images

Sun, 01 Jan 2017 00:00:00 +0000

The market of autonomous cars is soaring and most car companies are devoting an increasing part of their research & development budget to autonomous driving. It might only be a matter of time before humans themselves will be forbidden the right to drive, but there are major ethic questions along the way. These have a chance to be solved only if autonomous vehicles are significantly more secure than humans. How can we make sure that these cars won't pose a threat to pedestrians ? That is why a car should be able to recognize its environment with a high precision level. We will try and model this problem, focusing on images taken with a camera from inside the car. Our aim is to detect pedestrians on an image, and more broadly, to classify pixels of a road image in different categories (sky, building, pedestrian...).

I would like to warmly thank Mr Le Saux, Mr Guerry and Mr Vanel for guiding me in this fascinating project.

Dataset

Awesome Youtube Channels

Fri, 01 Jan 2016 00:00:00 +0000

In this post, I present some awesome youtube channels. This list is very incomplete but there are already more high-quality videos than you can watch. If you know some good educational channels, please let me know and I will be happy to add them to the list !