Multi-timescale representation of rat behavior

Unconstrained animal behavior is composed of rich dynamics that span multiple timescales. How these long timescale dynamics emerge from collections of short postural movements, however, remains an open question. Using markerless tracking, we measured the postural dynamics of freely behaving rats as they move throughout an enriched arena over 3-hour sessions and train a multi-layer recurrent neural network to predict these dynamics. We use activation values from each layer to create multiple two-dimensional density maps of behavioral states. We find that in these density maps, local maxima correspond to different stereotyped behaviors that the animal performs, including locomotion at different speeds, head movements, grooming, rearing and rest. Density maps from different layers exhibit dynamics at different timescales. Lastly, we extracted the hierarchical structure of rat behavior from these data, creating a generative model that recapitulates the observed movements. The results from this model allow us to make predictions about the internal neural and neuromodulatory states that regulate the rat’s behavior at multiple timescales.