Functional Priors for Bayesian Deep Learning

Explore the innovative framework for imposing functional priors on modern neural networks in this 58-minute lecture by Maurizio Filippone at the Finnish Center for Artificial Intelligence FCAI. Delve into the challenges of specifying prior distributions over weight and bias parameters in Bayesian neural networks, and discover how Gaussian processes offer a rigorous nonparametric approach to define priors over function spaces. Learn about a novel method that minimizes the Wasserstein distance between samples of stochastic processes to implement functional priors. Examine experimental results demonstrating significant performance improvements when combining these priors with scalable Markov chain Monte Carlo sampling, compared to alternative prior choices and state-of-the-art approximate Bayesian deep learning techniques. Gain insights from Maurizio Filippone, an associate professor and AXA Chair of Computational Statistics at EURECOM, France, with extensive expertise in Bayesian statistics and inference of Gaussian processes and neural networks.

Maurizio Filippone: Functional Priors for Bayesian Deep Learning