CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data
Researchers have developed CalM, a self-supervised foundation model designed for analyzing calcium imaging data in neuroscience. This model uses a novel pretraining framework that includes a tokenizer for discrete vocabulary mapping and a transformer architecture to capture dependencies in neural and temporal data. CalM demonstrates strong performance in forecasting neural population dynamics and superior results in behavior decoding tasks compared to supervised models, while also revealing interpretable functional structures within its representations. AI
IMPACT Introduces a novel self-supervised pretraining paradigm for neural foundation models, potentially enabling broader applications in functional neural analysis.