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

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neural Dynamics

Background:

• Low-rank recurrent neural networks (lrRNNs) model neural population activity but lack interpretability.
• Difficulty in assigning distinct computational roles to latent dimensions hinders understanding.

Purpose of the Study:

• To develop a generative lrRNN framework, Factored Recurrent Neural Network (FacRNN), for improved disentanglement and interpretability.
• To enable group-wise independence among latent dynamics while allowing within-group entanglement.

Main Methods:

• Reformulated lrRNN within a variational autoencoder (VAE) framework.
• Introduced a partial correlation penalty to encourage independence between latent dimension groups.
• Evaluated on synthetic, monkey M1, and mouse voltage imaging data.

Main Results:

• FacRNN demonstrated improved disentanglement and interpretability of neural latent trajectories.
• The model successfully learned low-dimensional, low-rank connectivity with independent latent groups.
• Consistent performance across diverse datasets validated the framework's efficacy.

Conclusions:

• FacRNN offers a powerful approach to interpret latent neural dynamics by enforcing group-wise independence.
• This framework advances the analysis of neural population activity, facilitating the assignment of computational roles to latent dimensions.