Variational Continual Learning for Gaussian Processes and Bayesian Neural Networks

Abstract: The ability to learn continually from experience and adapt quickly to new tasks/data/environments is crucial to the development of intelligent systems. A small number of existing approaches to continual learning either use suboptimal handcrafted heuristics or suffer from catastrophic forgetting or slow updating when new data arrive. In this talk, I will discuss a principled learning algorithm based on the variational free-energy method and episodic memory. The proposed approach can sidestep the aforementioned limitations of existing methods, allowing models to learn using new data whilst maintaining expertise on tasks which they have not experienced for a long time. Some results on sequential learning for sparse Gaussian process regression and classification, Bayesian neural networks and variational autoencoder will be discussed.

This is a joint work with Cuong V. Nguyen, Yingzhen Li, Brian Trippe and Richard E. Turner.