FacRNN framework disentangles neural dynamics with group-wise independence
New generative RNN model reveals independent latent groups in brain activity data
Chengrui Li and colleagues have introduced FacRNN (Factored Recurrent Neural Network), a novel generative framework that extends low-rank RNNs (lrRNNs) to uncover independent latent dynamics in neural population activity. Traditional lrRNNs capture low-dimensional structure but lack independence between latent dimensions, making it hard to assign distinct computational roles. FacRNN addresses this by enforcing group-wise independence—allowing groups of latent neurons to evolve separately while retaining rich internal computation.
Built on a variational autoencoder (VAE) architecture, FacRNN incorporates a partial correlation penalty that encourages independence between latent groups. Experiments on synthetic data, primate motor cortex (monkey M1), and mouse voltage imaging show that FacRNN consistently yields better disentangled neural trajectories and interpretable low-rank connectivity compared to baseline lrRNNs. This approach could empower neuroscientists to map distinct brain functions to separable neural populations, advancing both basic science and neural interface design.
- FacRNN uses a VAE with a partial correlation penalty to enforce group-wise independence among latent dynamics.
- The model was validated on synthetic data, monkey M1 recordings, and mouse voltage imaging, outperforming standard lrRNNs in disentanglement.
- Independent latent groups enable distinct computational roles for different neural subpopulations, improving interpretability.
Why It Matters
FacRNN bridges machine learning and neuroscience, enabling clearer interpretation of neural circuits for brain-computer interfaces.