Prédoc III - Gabriel Huang : Learning Complex Objects
Titre : Learning Complex Objects
Lieu : Local 3195, Pavillon André-Aisenstadt, Université de Montréal
Date : Mercredi 12 décembre à 10h
Président : Aaron Courville
Membre : Pascal Vincent
Directeur : Simon Lacoste-Julien
The evolution of machine learning science and computation power has enabled us to learn more and more complex objects. We went from doing linear regression and Gaussian mixture models to predicting highly structured objects (e.g. structured prediction), generating high-dimensional images (e.g GANs), and even learning optimization and machine learning algorithms (e.g. meta-learning).
Despite the steady increase in the complexity of the tasks we solve, we need to keep in mind which problem we are actually solving (e.g. generate realistic images) and not confuse it with our usual training/validation objectives (e.g. maximize validation likelihood) which are usually only approximations/surrogates of the true objective. Indeed there is no point in using an algorithm if that algorithm is solving the wrong problem. The focus of our thesis is twofold: exploring new techniques for learning complex objects, and better understanding how those techniques actually relate to the underlying task we are trying to solve.
Our first contribution is to introduce parametric divergences and take the view that GANs minimize such divergences, which goes against the commonly assumed optimal discriminator assumption that implies GANs are minimizing traditioal, nonparametric divergences. We study properties of parametric and nonparametric divergences, and investigate ways to better align parametric divergences with the underlying task.
Our second contribution consists in indirectly modeling image distributions in scattering coefficient space. Scattering coefficients are features computed by a convolutional neural network with designed fixed weights. The motivation is that scattering coefficients are potentially a better representation for manipulating images than raw pixels. Indeed they are provably robust to small deformations of the image.
Vous êtes cordialement invité.