14:00 – 15:00 – Rob Cornish (University of Oxford)

Title: Equivariant stochastic neural networks in Markov categories

Abstract: Traditional probability theory is often cumbersome, especially when applied to complex problems arising in modern machine learning and statistics. To address this, there has been recent interest in reorganising probability theory using techniques from category theory, which provides a variety of tools for abstracting away low-level details in order to focus on higher-level structure of interest. In this talk, I will provide an introduction to this topic that assumes no previous familiarity with category theory. I will then present a novel application to the task of parameterising a stochastic neural network that is equivariant with respect to the action of a group. The resulting procedure is flexible and compositional, and relies on minimal assumptions about the structure of the group action. Moreover, much of the underlying theory can be expressed visually in terms of string diagrams, and in a way that closely matches a computer implementation.

Refreshments available between 15:00 – 15:30, Huxley Common Room (HXLY 549)