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 Fundamental Tradeoff between Computational Flexibility and Spiking Variability in Recurrent Neural Networks

Zachary Loschinskey¹, Michael Economo¹, Brian DePasquale¹; ¹Boston University

Presenter: Zachary Loschinskey

Neural responses are notoriously variable: when the brain is presented with identical stimuli or generates stereotyped movements, the underlying neural activity can differ substantially across repeated trials. This trial-to-trial “spiking noise” shapes how we design experiments, how we analyze neural data, and how we interpret neural codes. Despite spiking noise being a critical aspect of neural data, we lack a circuit-level understanding of how it arises and what constraints it imposes on neural computation. Building on recent theoretical work, we hypothesize that spiking noise emerges as a consequence of a neural circuit learning to perform many computations, reflecting a fundamental tradeoff between flexibility and noise. By training spiking neural networks (SNNs) to perform a variety of pattern generation tasks, we demonstrate such a tradeoff, providing a new perspective on an incompletely understood but ubiquitous property of neurons.

Topic Area: Methods, Tools, Theory & Neural Coding