Two complementary perspectives to continual learning: ask not only what to optimize, but also how

Abstract:

Continually learning from a stream of non-stationary data is difficult for deep neural networks. When these networks are trained on something new, they tend to quickly forget what was learned before. In recent years, considerable progress has been made towards overcoming such “catastrophic forgetting”, predominantly thanks to approaches that add replay or regularization terms to the loss function to approximate the joint loss over all tasks so far. However, in this talk I will show that even with a perfect approximation to the joint loss, these approaches still suffer from temporary but substantial forgetting when starting to train on a new task. Motivated by this “stability gap”, I propose that continual learning research must focus not only on the optimization objective, but also on the way this objective is optimized.

Short bio:

Gido van de Ven is a MSCA postdoctoral fellow at KU Leuven (Belgium), where he performs research at the intersection of deep learning, computational neuroscience and cognitive science. In his research, Gido van de Ven uses insights and intuitions from neuroscience to make the behavior of deep neural networks more human-like. In particular, he is interested in the problem of continual learning, with generative models a principal tool to address this problem. Previously, for doctoral research, Gido van de Ven used optogenetics and electrophysiological recordings in mice to study the role of replay in memory consolidation in the brain.