A Cyclic Adaptation-Generalization Framework with Uncertainty-Guided Self-Paced Learning for Long-Term Brain-Machine Interfaces

The paper proposes Uncertainty-guided Self-paced Cycling (UnSPC) for long-term invasive brain-machine interfaces, where neural drift gradually degrades decoding performance. Its Uncertainty-guided Self-paced Pseudo-labeling (UnSPL) ranks and mines reliable pseudo-labeled samples, while Cycling Adaptation and Generalization (CycAG) alternates domain adaptation and domain generalization to address both global and subdomain shifts. The abstract reports experiments on multiple neural decoding datasets and claims improved robustness, but provides no dataset names, baseline comparisons, or quantitative results in the supplied material.