Cross-subject / cross-session transfer

The central promise of the field: a model pretrained across many subjects and recording sessions transfers to a new subject or a later session with little or no new labeled data. The obstacle is that neural recordings have no fixed input correspondence — every session has different neurons, electrode placements, counts, and tuning — so a decoder trained on one subject cannot be applied verbatim to another. Foundation-model approaches attack this with channel-agnostic tokenization and shared latent representations rather than a fixed input vector.

It is worth being precise about what has been shown. Within-species, within-task transfer with a light calibration step — for example fine-tuning only the new session's unit or session embeddings while freezing the backbone — has been demonstrated for reaching and other motor tasks (POYO+, NDT2). Broad cross-species, cross-area, or cross-task transfer remains largely aspirational, and true zero-shot (no calibration at all on the new subject) is the exception, not the rule. Reporting should always separate few-shot from zero-shot and state how much target data was used.