Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

A thalamocortical RNN architecture for flexible and robust motor sequencing

Abhijnana Das¹, G Sean Escola², Laureline Logiaco¹; ¹University of Colorado Anschutz, ²Columbia University

Presenter: Abhijnana Das

The ability to learn an extendable library of motor motifs that can be flexibly strung together without rehearsing transitions between motifs is a hallmark of biological motor control, as exemplified by phoneme production during speech. Motor cortical regions are thought to rely on preparation and trigger mechanisms for this process, but how such mechanisms might be used in machine learning models remains unclear. Here, we show that standard RNNs cannot robustly improvise transitions between motifs drawn from an extendable library. Specifically, while many parameter-segregated architectures support continual learning of individual motifs without interference, we find that these simple architectures do not generalize well during transitions. We then show that a transition module inspired by neural preparatory and trigger mechanisms enables networks to transition robustly between motifs.

Topic Area: Methods, Tools, Theory & Neural Coding