Revealing task-relevant dynamics via decomposition of intrinsic and input-driven dynamics

Summarise with VIRAT
Authors: M. Schwabe, A. Behrad, R. Zeraati, S. Safavi
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Mechanistic understanding of complex behavior requires establishing a direct link between the state of behavior and the underlying neural dynamics. While various previously developed methods attempted to provide such a link, they often rely on explicit behavioral labels. Here, we introduce a label-free method that mechanistically links network input, neural dynamics, and behavior in input-driven nonlinear systems. By combining the Hankel alternative view of Koopman (HAVOK) and dynamic mode decomposition with control (DMDc), we approximate the nonlinear dynamics by a linear system given by autonomous component A and input-driven component B. We show the application of our method to an agent modeled as a nonlinear recurrent neural network (RNN) performing a naturalistic behavioral task of tracking a realistic odor plume. Projections of neural time series onto eigenvectors of matrix A are clearly correlated with both observations and behavior of the agent. Moreover, different eigenvector directions show correlation structures, which are obscured in the direct observation-behavior correlations. For instance, despite negligible correlations between transversal wind with agent turning angle, we found dynamic modes that are up-regulated with transversal wind and are correlated with turning, suggesting that the agent performs up-wind flying by aligning with the wind direction. Thus, our method identifies functionally relevant modes in neural dynamics in an unsupervised manner. The linear correlations found between dynamic mode activations, observations, and behavior can be used to develop new testable hypotheses of how nonlinear neural dynamics give rise to behavior.

Fields of Study & Subjects

Model Reduction and Neural Networks Neural dynamics and brain function Neural Networks and Applications

Publication details

Journal / source
MPG.PuRe (Max Planck Society)
Publication date
1 March 2026
Language
EN