Deep Neural Networks Are Like The Brain, Except When They Aren’t
Francis SongDeepMind, London
Senior Common Room, Level 2 (2D17), Priory Road Complex
Date |
Thursday 3rd October 2019
Time |

The last several years have witnessed the interesting rise of the AI-researcher->neuroscientist and neuroscientist->AI-researcher. In one direction, the ease with which deep learning techniques can generate working models of traditional animal and human behavioral tasks has led to many comparisons of the behavior and activations of deep neural networks to recordings of the brain; I will describe some of my own work in this area, including what I have learned about the limitations and challenges of this approach. In the other direction, many AI researchers take inspiration from years of psychological and neuroscientific research to guide – and critique – the use of modern deep learning techniques and architecture in designing a more human-like intelligence. I will describe some of my work in the specific case of Theory of Mind, as well as work on core reinforcement learning algorithms and architectures that we hope will lead to the kind of efficient learning and generalization that many see as necessary for achieving “intelligence.”

All Welcome