CCA Alumna Lucia Pezzetti and Chair Stefano Favaro publish in top-tier AI Conference Proceedings

We are proud to announce that CCA Alumna Lucia Pezzetti – now PhD student in Computer Science at ETH AI Center – and Carlo Alberto Chair Stefano Favaro have recently published in top-tier Artificial Intelligence Conference Proceedings.

The article “Function-space MCMC for Bayesian wide neural networks” (also with Stefano Peluchetti) recently appeared on the Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS).

In this paper, the authors propose a novel strategy for posterior sampling in Bayesian Neural Networks. Specifically, they investigate the use of the preconditioned Crank-Nicolson (pCN) algorithm and its Langevin version to sample from a re-parameterized posterior distribution of the network’s weights in the large-width regime. They establish that these algorithms remain robust in the infinite-dimensional setting, and prove that the acceptance probability of the proposed algorithms converges to one as the network width increases, independently of step-size tuning. This result suggests that, in wide Bayesian Neural Networks, the pCN algorithm enables more efficient sampling of the re-parameterized posterior, as evidenced by a higher effective sample size and improved diagnostic metrics compared to existing approaches.

The publication offers a significant contribution in the field of Bayesian Neural Networks, which represent a fascinating confluence of deep learning and probabilistic reasoning, offering a compelling framework for understanding uncertainty in complex predictive models.

We congratulate Lucia and Stefano on this outstanding achievement and look forward to seeing the impact of this work on the academic community.