Tuesday, 5th of Mars

14h30 (room R2014, 660 building) (see location)

Gwendoline de Bie

(ENSAE/TAU)

Title: Stochastic Deep Networks

Abstract

In many settings of practical interest in machine learning,
input data are not accurately described by vectors in an Euclidean space,
but can be modelled using the notion of measure or random vector. Treating
such data as measures comes with enormous advantages in term of storage
(no need for a fixed resolution, no need to store a dense embedding
volume) and is also mathematically elegant (no need for an explicit
parametrization of surfaces for instance). This however poses challenging
issues, since most known deep architectures use Euclidean build-blocks
over a fixed topology. To address this goal, we introduce a general
framework to deal with measures in supervised and unsupervised learning
settings. We will describe the proposed framework, review its robustness
and show its efficiency on instanciations of discriminative, generative
and predictive tasks.



Contact: guillaume.charpiat at inria.fr
All TAU seminars: here