Dear all / Bonjour à tous,

We are happy to invite you to Vitoria Barin-Pacela's Predoc III defense on Wednesday, August 30th, at 1pm. (Hybrid event.)

Vous êtes cordialement invité.e.s à la présentation du sujet de recherche de Vitoria Barin-Pacela, le mercredi 30 août à 13h00. (Présentation hybride).

Jury

Abstract

In representation learning, it is desirable to learn factors of variation that are structured, causally interpretable and semantically meaningful, such that they can be robust to changes in environments, offer out-of-distribution generalization, and be reused for downstream tasks for higher sample efficiency. This problem has been studied through the lenses of disentanglement, with the goal of isolating the factors of variation, and nonlinear independent component analysis, where the factors of variation are independent. Statistical identifiability is a more general concept where the representation and parameters learned should be unique on the limit of infinite data, thus being an important condition for learning useful representations.

The identifiability of unsupervised models under a general nonlinear smooth map was proved impossible in literature. It is typically approached by imposing assumptions on the factors of variation and on the nonlinear map. In the main contribution, we propose quantized coordinated identifiability, a relaxed form of identifiability that requires very minimal assumptions, proving that the quantized factors of variation can be recovered. This quantization relies on axis-aligned discontinuities that are assumed to be present in the probability density function of the true factors of variation.

Finally, we present a few directions of future work, focused on demonstrating the practical benefits of identifiability in realistic scenarios. For example, we plan to extend the identifiable models to large-scale and realistic settings, evaluating them under distribution shifts and analyzing the sample-efficiency for downstream tasks.