Fiche de TRAN Ba-hien

Titre : Advancing Bayesian Deep Learning: Sensible Priors and Accelerated Inference
Date : 13 octobre 2023
Heure: 14:00
Résumé : In this thesis, we propose solutions to address challenges in Bayesian deep learning, including the selection of sensible priors and accelerated inference. The first contribution is a study of the pathologies associated with poor prior choices in Bayesian neural networks for supervised learning. We propose a practical and effective solution by reasoning in terms of functional priors. The second contribution is a novel framework for model selection in Bayesian autoencoders for unsupervised tasks. We consider the marginal likelihood of these models using functional priors and propose a fully sample-based optimization approach. The third contribution is a fully Bayesian autoencoder model that handles both local latent variables and the global decoder. We propose an efficient amortized MCMC scheme and sparse Gaussian process priors over the latent space. The last contribution is a simple yet effective approach to improve likelihood-based generative models through data mollification.
Lieu : EURECOM