Title: A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity

Abstract: Low-rank recurrent neural networks (lrRNNs) serve as a model class designed to reveal the low-dimensional latent dynamics that underlie neural population activity. However, while their functional connectivity adheres to a low-rank structure, it often fails to provide independent interpretations. This limitation hinders the ability to assign specific computational roles to distinct latent dimensions. To overcome this challenge, we introduce the Factored Recurrent Neural Network (FacRNN), a generative lrRNN framework predicated on the assumption of group-wise independence among latent dynamics, while still permitting flexible entanglement within these groups. This architecture enables latent dynamics to evolve independently across groups, yet maintains internal richness to support complex computations. We recast the lrRNN within a variational autoencoder (VAE) framework, which facilitates the incorporation of a partial correlation penalty designed to foster independence between sets of latent dimensions. Evaluations using synthetic data, as well as recordings from monkey primary motor cortex (M1) and mouse voltage imaging, demonstrate that FacRNN consistently enhances the disentanglement and interpretability of learned neural latent trajectories in low-dimensional spaces and low-rank connectivity, outperforming baseline lrRNNs that do not enforce group-wise independence.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC