Self-supervised pretraining on neural data

Learning representations from unlabeled neural recordings by solving a pretext task that needs no behavioral labels: masking and reconstructing hidden activity, predicting the next time step, or a contrastive objective that shapes the embedding geometry. The motivation is economic: aligned, labeled neural data (activity paired with well-instrumented behavior) is scarce and expensive, whereas raw recordings accumulate continuously, so the bulk of learning is pushed onto the unlabeled data.

Pretext families mirror the rest of machine learning. Generative and reconstructive objectives (the Neural Data Transformer's masked firing-rate modeling) fit a likelihood over held-out activity; contrastive objectives (BENDR's wav2vec-style scheme, CEBRA) pull related samples together without reconstructing; predictive objectives forecast future activity. Gains over training-from-scratch are real but usually modest and task-dependent, and shrink when labeled data for the target task is already plentiful.