Joe Pemberton

Postdoctoral research associate

visitor
University of Washington

When an organism or machine makes a certain decision, the ensuing teaching signal (e.g. reward) is often not instanteous but comes at a later point in time. The challenge of attributing this teaching signal, and appropriately updating any parameters which led to the original decision being made, is known as the temporal credit assigment problem and is the focus of my research. In particular, I hope to bridge the ominous gap between the existing temporal learning mechanisms developed by those in pure machine learning (the most prevalent example of which at the moment is backpropogation's temporal cousin, backpropogation through time) and computational principles considered to be realistically implementable in a biological system.

I did a PhD in the group, and I am now a short-term postdoc. Before coming to Bristol, I received a BSc in mathematics at the University of Warwick, before doing an MSc in Artificial Intelligence at the University of Edinburgh.