Hierarchical timescales of free behavior

Animal behavior is a multiscale process. At the smallest scale, it can be deconstructed as a series of stereotypic postural movements performed by the animal. However, behavior is driven by intricate internal and external processes occurring at multiple long-timescales. In order to understand these processes it then becomes imperative for us to represent behavior at multiple timescales. Here, we use a Recurrent Neural Network (RNN) based modeling approach to create representations of animal behavior at multiple timescales. Using markerless tracking, we measured the postural dynamics of freely behaving rats as they move throughout an enriched arena over 3-hour sessions. We then train a novel multi-layer and time-dilating RNN to predict these dynamics and use latent dynamics from each layer to create multiple two-dimensional density maps of behavioral states. We find that density maps from different layers exhibit dynamics at different timescales, allowing us to predict interactions between the internal neural and neuromodulatory states that regulate the rat's behavior across these scales.