Poster Presentation

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Uncovering Multi-Regional Inputs Using Input-Driven Switching Recurrent Neural Networks

Yongxu Zhang¹, Jeffrey Walker¹, Nicholas G. Hatsopoulos², Jason N MacLean², Shreya Saxena¹; ¹Yale University, ²University of Chicago

Presenter: Yongxu Zhang

A central goal in neuroscience is to understand how neural dynamics give rise to flexible, context-dependent behavior. The brain functions as a nonlinear dynamical system, where neural population activity evolves along structured trajectories that vary across states. To understand these state-dependent transitions, it is crucial to identify not only the latent neural dynamics, but also the factors that modulate them. Previous models, such as recurrent neural networks and switching dynamical systems, have revealed latent dynamics linked to cognitive or behavioral states, but often neglect the external or internal inputs driving transitions. Input-driven models like LFADS and iLQR-VAE infer external inputs from neural recordings, yet they still treat neural dynamics as autonomous processes and do not explicitly model behaviorally relevant switching. To address this gap, we introduce Input-driven Switching Recurrent Neural Networks (iSRNNs), a framework that disentangles intrinsic dynamics from input-driven modulations. iSRNNs model baseline or resting-state dynamics with a base RNN and infer unobserved, time-varying inputs that trigger transitions into distinct, behaviorally relevant dynamical regimes. Applied to both simulated and real neural data from different species, i.e., macaques and marmosets, iSRNNs reconstruct observed activity, uncover behaviorally-relevant dynamic states, and identify interpretable inputs. Surprisingly, iSRNNs fit to primary motor cortex recordings successfully recover activity from premotor cortex. Together, we develop iSRNNs as a principled and interpretable framework for uncovering how input-driven computations enable flexible neural dynamics and cross-regional coordination.

Topic Area: Methods, Tools, Theory & Neural Coding