A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity
Researchers have developed a new framework called the Factored Recurrent Neural Network (FacRNN) to improve the interpretability of low-rank RNNs used in analyzing neural activity. This model introduces group-wise independence among latent dynamics, allowing for more distinct computational roles to be assigned to different neural dimensions. Experiments on synthetic and real neural data demonstrated that FacRNN enhances the disentanglement and interpretability of learned neural trajectories and connectivity compared to standard low-rank RNNs. AI
IMPACT Introduces a novel method for disentangling neural dynamics, potentially improving the understanding and application of RNNs in neuroscience.