Matthew Botvinick
Wednesday 6th July 2016
Time: 4.00pm
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Deep Learning to Learn
Recent advances in deep reinforcement learning have attracted great interest in neuroscience and psychology. However, while neural networks have attained human (and even superhuman) levels of performance in a growing range of task environments, they still compare poorly with human learners in terms of both their sample efficiency and their behavioral flexibility. I will argue that, contrary to prevailing opinion, there is a readily available strategy for overcoming these apparent limitations in deep learning. The key lies in reviving an old but neglected idea for how to endow neural networks with an ability to ‘learn how to learn.’ In addition to reviewing the approach and describing some recent applications, I will also touch on some aspects of our work on hierarchical reinforcement learning, which led us to our current interest in learning to learn.
Matthew Botvinick, M.D., Ph.D.
Director of Neuroscience Research, Google DeepMind
Honorary Professor, Gatsby Computational Neuroscience Unit, UCL