Talk by Yarin Gal (Oxford) and Siddharth Swaroop (Cambridge)

We will have the following two talks from 4-5pm.

Talk by Yarin Gal from Oxford University on "Bayesian Deep Learning in Self-Driving Cars (and more)"

Talk by Siddharth Swaroop from Cambridge University on "Variational Continual Learning"

Abstract for Siddharth's talk:
In the continual learning setting, tasks are encountered sequentially, with old tasks' data inaccessible. The goal is to learn whilst i) avoiding catastrophic forgetting, ii) efficiently using model capacity, and iii) employing forward and backward transfer. In this talk, I shall introduce the Variational Continual Learning (VCL) framework, where we variationally train Bayesian Neural Networks. I shall then look at how well VCL achieves continual learning’s desiderata on two image classification benchmarks: permuted MNIST and split MNIST. By looking in detail at VCL’s solutions, we can obtain an understanding of why VCL performs as it does, and can compare its solution to what an ‘ideal’ continual learning solution might be. We can then also critically examine how well the two benchmarks test continual learning’s desiderata.