Deep meta-learning

Illustration of deep learning

The SiG on AI & Behavioral Change, cordially invites you for a presentation by Prof. Matthew Botvinick (Director of Neuroscience Research at DeepMind and Honorary Professor, Gatsby Computational Neuroscience Unit, UCL).

Prof. Botvinick is a leading researcher on the intersection between AI and cognitive neurosciences.

The presentation will take place on the 18th of March at 3pm (the talk will be held in Teams – see link below).

Abstract

Over the past decade there has been explosive progress in deep learning. However, many see a remaining challenge in 'sample efficiency': Deep learning is data hungry, and can be limited in its ability to adapt flexibly and rapidly -- as humans do -- to new task challenges. One approach to this problem is meta-learning or learning-to-learn, where a slow learning process, unfolding over many tasks, equips the learner to learn with increasingly nimbleness. I'll discuss my group's explorations of deep meta-leaning, focusing in particular on the questions of how well meta-learning may scale and of how it may be relevant to human-AI interaction.

> Teamslink