Jeffrey Johnston, Ph.D.
June 15(Thu) - June 15(Thu), 2023
10am
Online zoom (ID: 728-142-6028)
Neuro@noon Seminar
Date: 10am, Thur, June 15th
Speaker: Jeffrey Johnston, Ph.D. (Columbia University)
Title: The emergence of abstract and modular representations in simple neural networks
Abstract: Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability to generalize arises from a specific representational geometry, that we call abstract and that is referred to as disentangled in machine learning. These abstract representations have been observed in recent neurophysiological studies. However, it is unknown how they emerge. Using feedforward neural networks, we demonstrate that the learning of multiple tasks causes abstract representations to emerge. We show that these abstract representations enable few-sample learning and reliable generalization on novel tasks. However, despite the benefits of abstract representations, there are many situations in which over-generalization can lead to behavioral errors. In preliminary work, we study neural networks trained to perform context-dependent tasks. We show that these networks produce modular activity patterns, where a particular subset of neurons is active in each of the trained contexts. Then, we show that the representations within each module are abstract -- and, further, that they facilitate rapid learning of new tasks and rapid acquisition of related contexts. Together, this work provides insight into how neural representations may support complex behavior.