StateEnsemble: Neural State Chain-based Dynamic Model for Complex Motor Sequence Decoding

ICLR 2026 Conference Submission16675 Authors

Keywords: state space model, bayesian filter, neural decoding, brain computer interface

Abstract: Brain-computer interfaces (BCIs) have paved the way for motor function rehabilitation and reconstruction. However, accurate movement decoding is still a challenging problem, especially for complex movements. Recent studies discovered that motor sequences, particularly complex ones, are encoded through a chain of neural states, each corresponding to a movement fragment. While this neural basis could facilitate more accurate neural decoding for complex movements, existing neural decoders fall short in modeling state-level sequential information. Here, we propose a neural state chain-based dynamic model (StateEnsemble), which explicitly models the neural state transition process to perform state-dependent neural decoding. We evaluated the proposed approach with intracortical neural signals recorded from the human motor cortex during handwriting. Experimental results demonstrated that our approach can effectively capture the underlying neural state transition patterns during handwriting, and achieve significant improvements in decoding performance. The proposed StateEnsemble approach can be beneficial for diverse neural decoding tasks and facilitate high-performance BCIs.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 16675