1 h
Seminar, Hosted by Kempner Institute at Harvard University
Large-Scale Brain Decoding – Taking Advantage of Diversity
McGill University; Mila
Recording
Abstract
Neural decoding supports research on sensorimotor processing, clinical diagnosis and brain–computer interfaces, yet the diversity of neural recordings makes large-scale model training difficult. Blake Richards presents a tokenization approach for training a single transformer decoder across subjects, recording modalities, stimuli, behaviors, brain regions, cell types and species. The work finds that greater data diversity improves decoding when the model is designed to share information across those differences.