Title: A new direction for continual learning: ask not just where to go, also how to get there
Abstract: Incrementally learning new information from a non-stationary stream of data, referred to as 'continual learning', is a key feature of natural intelligence, but an open challenge for AI and deep learning. For example, standard deep neural networks tend to catastrophically forget previous tasks or data distributions when trained on a new one. Enabling these networks to incrementally learn, and retain, new skills over time has become a topic of intense research.
I start the talk with an extended introduction, in which I discuss what continual learning is, why it is important, and why it is challenging. In the second part, I explain what the current approach to continual learning is, and I point out a fundamental limitation of that approach. I end the talk with a proposal for a new direction for continual learning, which I hope will pave the way towards practical applications of continual learning techniques.
