Jason Z Kim
December 2(Thu) - December 2(Thu), 2021
12PM
ZOOM (ID: 728-142-6028)
Neuro@noon Seminar
Date: 12PM, Thursday, December 2nd
ZOOM 회의 참가 ID: 728-142-6028
Speaker: Jason Z Kim
(Department of Bioengineering, Univ. of Pennsylvania)
Title: Distributed representations and control in structural brain networks
Abstract:
Neural systems display complex and distributed patterns of interactions between components. These interactions often occur along structural connections that link neural units in a hard-wired connection topology, supporting a variety of dynamical behaviors such as cognitive control. But precisely what are the distributed patterns of connections that enable such control? In this talk, I will present on two projects addressing this question. In the first project, I use linear network control theory to derive accurate closed-form expressions that relate the structural connectivity of the brain to the energy required to control the brain to different states. I identify specific topological features that enable energetically favorable control, and discover more of these features in high-resolution brain networks of increasingly complex species. In the second project, I teach nonlinear recurrent neural networks (RNNs) to form internal representations, and to manipulate these representations and make accurate predictions far outside of their training data. To explain how neural networks extrapolate, I demonstrate that RNNs form a symbolic representation of their inputs, and derive a closed-form relationship between the connection topology and the symbolic representation. Together, these results bring us closer to making sense of the complex and distributed topology of brain networks.