Duncan Lecture Series- AMS Seminar: Stuart Geman (Brown University) @ Shaffer 100
Title: Real and Artificial Neural Networks
Abstract: Lately, just about everybody has been thinking about deep neural networks (DNNs). Do they work? If so, how? Do they overfit? If not, why? I will discuss these questions and suggest some uncomplicated answers, at least as a first approximation. Turning to biological learning, I will argue that the stubborn gap between human and machine performance, when it comes to interpretation (as opposed to classification), can not be substantially closed without architectures that support stronger representations. In particular, how are we to accommodate the rich collection of spatial and abstract relationships (‘on’ or ‘inside’, ‘talking’ or ‘holding hands’, ‘same’ or ‘different’) that bind parts and objects and define context? I will propose that the nonlinearities of dendritic integration in real neurons is the missing ingredient in artificial neurons. I will suggest a mechanism for embedding relationships in a generative network.