Vendredi 15 mars 2024
Theoretical guarantees for Deep Generative Models: A PAC-Bayesian Approach
Sokhna Diarra Mbacke
Étudiante au doctorat au Graal
Heure: 13h30
Local: PLT-2551
Résumé: In this presentation, we study the statistical properties of deep generative models.
Generative models have become a central research area in machine learning, with applications in many different areas. The goal of a generative model is to replicate an unknown data-generating distribution, from which only a finite number of samples is available. Intuitively, a good performance metric should measure the discrepancy between the true data-generating distribution and the distribution learned by the model. This is a very challenging task, not only because the true distribution is unknown in general, but also because different discrepancy measures between probability distributions have different behaviours and interpretations.
We use PAC-Bayesian theory to study the theoretical properties of deep generative models. PAC-Bayes provides generalization bounds for machine learning models. The first part of this presentation focuses on adversarial generative models, such as the Wasserstein Generative Adversarial Network, and the second part of the presentation focuses on Variational Autoencoders.
Note: La présentation sera donnée en français.
