Reproduce “Generalizable, real-time neural decoding with hybrid state-space models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Avery Hee-Woon Ryoo, Nanda H. Krishna, Ximeng Mao, Mehdi Azabou, Eva L. Dyer, Matthew G. Perich, Guillaume Lajoie
POSSM decodes neural activity by embedding individual spikes, compressing short chunks, and carrying recurrent state forward to predict behavior. It adapts to new recordings through unit and session embeddings or full finetuning. On imagined handwriting, pretrained and finetuned POSSM-GRU reached 97.37 ± 2.32% accuracy versus 95.82 ± 3.41% from scratch. Offline inference took 2.44 ms per chunk; evaluations were not online.
Key insights
- Integrating flexible input processing, such as individual spike tokenization, with efficient recurrent state-space models effectively bridges the gap between strong generalization capabilities and real-time inference speed in neural decoding.
- Multi-dataset pretraining allows a single model to learn transferable neural dynamics across diverse contexts, significantly improving data and training compute efficiency for adaptation to new subjects and tasks.
- Cross-species transfer learning is feasible and beneficial, demonstrating that neural dynamics learned from abundant non-human primate data can improve decoding performance on human clinical tasks, addressing data scarcity challenges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.