New AI transformer model automates rodent behavior recognition from pose data, replacing time-consuming manual scoring methods

MSGL-Transformer uses multi-scale attention mechanisms to capture rodent social behaviors across different temporal scales through short-range, medium-range, and global attention branches

The model introduces a Behavior-Aware Modulation (BAM) block inspired by SE-Networks to emphasize behavior-relevant features before attention processing

Tested on RatSI dataset (5 behavior classes, 12D pose inputs) and CalMS21 dataset (4 behavior classes, 28D pose inputs) for evaluation

Addresses limitations of traditional manual behavior scoring by reducing human error and significantly decreasing analysis time for neuroscience research