Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining

A research team from McGill University and the Vector Institute has introduced POYO-CAP, a novel pretraining strategy that significantly improves AI's ability to decode dynamic visual experiences from calcium imaging data of the brain. The core challenge in this neuroAI field is neural heterogeneity—recordings mix highly predictable neurons with stochastic, stimulus-contingent ones, which destabilizes standard self-supervised learning (SSL). POYO-CAP addresses this with a biologically grounded, two-stage curriculum: it first performs masked reconstruction with lightweight supervision on statistically regular neurons (identified by their skewness and kurtosis), then fine-tunes on the more variable populations. This turns a fundamental obstacle into a scalable learning advantage.

The technical breakthrough lies in making statistical predictability an explicit data-selection criterion. When tested on the benchmark Allen Brain Observatory dataset, POYO-CAP yielded 12–13% relative improvements over standard from-scratch training. Crucially, it enabled smooth, monotonic performance scaling with increased model size, whereas baseline models trained on mixed neural populations would plateau or become unstable. This work, detailed in arXiv preprint 2510.18516, demonstrates that a curriculum informed by neural statistics is key to building more reliable and scalable brain decoding models. The implications are significant for next-generation brain-computer interfaces (BCIs) and for fundamental neuroscience research aiming to reconstruct perception from neural activity.