Transfer learning for BCI

Transfer learning reuses knowledge from source data or a pretrained model to reduce the labeled data a new BCI user or task requires. In practice this means pretraining an encoder on many subjects (or via self-supervision) and then fine-tuning some or all of its weights on a short calibration from the target user, or freezing the backbone and training only a small subject-specific head. The goal is a usable decoder from minutes rather than hours of calibration — the difference between a clinical demonstration and a practical device.

Key design choices are what to transfer (low-level temporal and spatial filters transfer well; high-level, subject-specific decision layers do not) and how much to fine-tune (too much on little data re-overfits; too little leaves the source-target gap unclosed). Few-shot and meta-learning formulations push toward near-instant calibration, but honest evaluations show a persistent accuracy gap versus fully subject-specific training that shrinks, rather than vanishes, with better pretraining.