'Bayesian Model Selection in Deep Learning' by Mark van der Wilk (QUVA Lab & Qualcomm meetup series)

Mark van der Wilk (Department of Computing, Imperial College London) joined us for a lecture on Deep Learning & AI on the 24th of February, 2021.

The Bayesian Deep Learning community is widely known for its efforts in bringing uncertainty estimates to deep neural networks. However, Bayesian methods have another key advantage: the ability to adjust inductive biases through model selection. Interestingly, model selection and uncertainty estimation are dual problems in the Bayesian framework. In this talk, we will discuss the current state of model selection in Bayesian deep learning, together with some of Mark van der Wilk's recent work towards this. He will discuss some theoretically grounded successes in Deep Gaussian Processes and in connecting ensembling to Bayesian inference, as well as recent empirical work on Neural Architecture Search. To finish, he would like to speculate on possible other benefits that the Bayesian framework can provide, in particular relating to asynchronous computation.

Bio:

Mark van der Wilk is a Lecturer (assistant professor) at Imperial College London in the Department of Computing. He is currently interested in automating inductive bias selection in neural networks, and making them less reliant on human design. He thinks the Occam's razor effect of Bayesian inference is fascinating, and thinks this could be part of the solution. In addition to working on trying to make it work in deep learning, he also applies Bayesian techniques in areas where they work very well (such as low-data Bayesian optimisation), particularly in collaboration with industry partners. Long-term, he is interested in developing flexible neural networks that adapt their structure to minimise computational cost as well as improving their generalisation performance.