Yongxu Zhang

This paper introduces SSLD, a framework that learns shared neural-behavioral latent dynamics and private modality-specific variables, showing it accurately reconstructs signals and isolates behaviorally relevant information across multiple datasets.

This paper introduces Time-Varying RNNs (TV-RNNs) for early classification of neural activity sequences, showing improved accuracy and earlier detection compared to standard RNNs in mouse and human motor task data.

The paper shows that pre-movement neural activity in mice contains structured dynamics that predict goal-directed actions, but not random ones, several seconds before movement.