Tom McCoy, Ph.D.
Assistant Professor, Linguistics
Yale University

Website

Zoom: https://arizona.zoom.us/j/88638472463
 

 

Image

Understanding neural networks via the problem they are trained to solve

Abstract: Neural network language models, such as ChatGPT, have shown tremendous advances in recent years. However, these systems are notoriously difficult to understand and control - an issue that limits their usefulness as models of cognition. In this talk, I will discuss a perspective that mitigates these issues: approaching neural networks from the perspective of the problem that they are trained to solve (i.e., viewing them at Marr’s computational level). The first part of the talk will discuss how this perspective enables us to predict some important limitations of large language models - an analysis supported by evidence of surprising failure modes that GPT-4 displays on seemingly simple tasks. In the second part of the talk, I will show how the same perspective can also enable us to control neural networks by connecting them to structured Bayesian models. We apply this approach in a case study based on language learning, where our goal is to distill the symbolic priors of a Bayesian model into a neural network. Like a Bayesian model, the resulting system can learn linguistic patterns from a small number of examples; like a neural network, it can also learn aspects of English syntax from a naturalistic corpus. Overall, these results show how it is both possible and beneficial to combine Bayesian models and neural networks, two popular approaches within computational cognitive science that have often been viewed as antagonistic.