Spectral envelopes of facial movements predict intention, cortical representations, and neural prosthetic control

1Animals, including humans, use coordinated facial movements to sample the environment, ingest nutrients, and communicate. Rodents, in particular, produce rhythmic facial movements during spontaneous behavior and cognitive tasks. Measuring these movements precisely and linking them to neural activity remains challenging. We introduce face-rhythm, an unsupervised pipeline that combines markerless pointtracking, spectral analysis, and non-negative tensor component analysis to decompose facial video into a small set of interpretable components. Applied to videos of mice during a Pavlovian odor-reward task, a brain-machine interface (BMI) task, and free behavior, face-rhythm recovers human-interpretable behaviors such as whisking, sniffing, licking, and subtler motifs. The resulting components are consistent across animals, are sufficient to decode task variables or internal belief states, and explain cortical activity using a low-rank representation. We also find that the activity of neurons in face-associated primary motor cortex (M1) is predicted well by a phase-invariant spectral transformation of facial movements above [~]0.5 Hz, while slower movements retain a phase-variant representation better predicted by the instantaneous position of the face; individual neurons can show either or both forms of tuning. A systematic comparison against deeplearning point-tracking models, contrastive-learning embeddings, and vision-transformer features places face-rhythm competitively across tasks while also achieving the goal of producing a low-dimensional, interpretable description of rodent facial behavior that is closely linked to cortical activity.