SQUIDS Seminar - Trace-class Gaussian priors for Bayesian reinforcement learning
| Dates: | 18 October 2023 |
|---|
| Times: | 14:30 - 16:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Department of Mathematics |
|---|
| Who is it for: | University staff, External researchers, Current University students |
|---|
| Speaker: | Torben Sell |
|---|
|
Schedule:
- 2:30pm: pretalk
- 3:05pm: research talk 'Trace-class Gaussian priors for Bayesian reinforcement learning'
I will motivate the talk by looking at some challenges in reinforcement learning, where classic function priors such as Karhunen-Loève priors struggle to scale up to the domain sizes we are interested in. I will then give some background and highlight connections to existing function priors, before defining the trace-class neural network (tcNN) prior. The prior is a Gaussian neural network prior, where each weight and bias has an independent Gaussian prior, but with the key difference to standard Bayesian neural network priors that the variances decrease in the width of the network in such a way that the resulting function is almost surely well defined in the limit of an infinite width network. Benefits and disadvantages of the different priors will be discussed, and the advantages of tcNN priors over other function space priors will be highlighted in numerical examples.
Reference: Trace-class Gaussian priors for Bayesian learning of neural networks with MCMC, Torben Sell and Sumeetpal S Singh, Journal of the Royal Statistical Society, Series B, 2023, open access article (https://academic.oup.com/jrsssb/article/85/1/46/7017762).
Speaker
Torben Sell
Organisation: The University of Edinburgh
Travel and Contact Information
Find event
G.114
Alan Turing Building
Manchester
