This preprint was posted to bioRxiv on July 9, 2026, and has not yet undergone peer review. Researchers introduced face-rhythm, an unsupervised pipeline combining markerless point-tracking, spectral analysis, and non-negative tensor component analysis to decompose facial video into interpretable components. Applied to videos of mice during Pavlovian odor-reward tasks, brain-machine interface (BMI) tasks, and free behavior, the face-rhythm method identified human-interpretable behaviors including whisking, sniffing, and licking, along with more subtle behavioral motifs.
The researchers found that the resulting components were highly consistent across animals, sufficient to decode task variables and internal belief states, and effective at explaining cortical activity using low-rank representations. Notably, neural activity in face-associated primary motor cortex (M1) was well predicted by phase-invariant spectral transformations of facial movements above 0.5 Hz, while slower movements retained phase-variant representations better predicted by instantaneous facial position. These findings suggest that brain-machine interface systems could leverage the spectral features of facial movements to optimize neural prosthetic control, providing important insights for achieving more natural and effective neural prosthetic operation.