Masked modeling of neural activity

A BERT-style pretext task: hide a subset of the neural array and train the model to reconstruct it from the surrounding context. For spikes the reconstruction is usually a count likelihood (Poisson or negative-binomial) on the masked bins, so the model learns to infer missing activity from co-active neurons and recent history. The axis you mask along matters a great deal — masking time bins teaches temporal structure, while masking whole neurons teaches cross-neuron structure and directly rehearses the ability to fill in units that drop out or appear anew.

Multi-task masking (MtM) generalizes this by randomly masking along several axes at once (time, neuron, and region), which pushes a model toward inferring any part of the population from any other part — a step the authors frame as a 'universal translator' for neural dynamics. This neuron-masking capability is what links masked pretraining to the practical problems of variable electrode yield and cross-session unit turnover.