Présentation de l'examen prédoctoral
Salah Rifai
Jeudi, 16 mai (demain), 13h00
Salle 3195, André-Aisenstadt
Titre:
Auto-encoders and their geometry.
Résumé:
Using a geometrically motivated approach to unsupervised
feature extraction, we propose a regularization and generation framework
for auto-encoders based on differential geometry. We show that regularizing
the variations of auto-encoders with respect to their input yields in a
space localized representation useful for a wide range of machine learning
tasks. We present an efficient way to approximate the manifold structure of
the input signal by building a topological atlas of charts with an
auto-encoder . The local chart associated to an input sample is the set of
feature vectors spanning the local tangent space to the manifold as
approximated by the auto-encoder. In a supervised setting, these charts can
be used to encourage the insensitivity of the classifier's output when
moving locally along the manifold. Another interesting use of these
local charts is the generation of samples from the input distribution in
an unsupervised way. The procedure we propose to generate samples exploits
the local structure approximated by the auto-encoder to explore
efficiently high-dimensional multimodal input distributions. Finally, we
relate
auto-encoders to a particular form of non-linear Independent
Component Analysis(ICA) and we use this link to propose a novel form
of auto-encoder that learns locally independent features.
Jury:
Aaron Courville (président)
Yoshua Bengio (directeur)
Roland Memisevic (membre)
Venez en grands nombres...
-- Aaron
--
Aaron C. Courville
Département d’Informatique et
de recherche opérationnelle
Université de Montréal
email:Aaron.Courville@gmail.com
