Unravelling structure in behavioral variability across scales

Animal behavior is highly variable, reflecting contrasting internal states or persistent inter-individual differences. Understanding the structure in this variation and its biological significance is a challenging task, as it can manifest across multiple time scales in behavior. How do we find structure in variation among such multiscale, unpredictable dynamical systems from finite observations? We construct multiscale behavioral representations, which we tune to maximize scale separation among phenotypic groups. As two finite recordings can be generated by the same dynamical process, phenotypes are uncertain, rendering bottom-up clustering approaches unstable. We thus use simulations to obtain uncertainty scales for each recording and use them explicitly in a Hierarchical Multiplicative Diffusive clustering to reveal robust phenotypic groups from the top down. We essentially map the phenotypic space onto a multiplicative diffusion process with state-dependent temperatures capturing the varying uncertainties of each behavioral recording. We then use the eigenspectrum of the diffusion operator to reveal structure in variability across multiple scales, providing a precise handle on the difference among phenotypic groups. Applied to a dataset of ~500 larval zebrafish swimming in different sensory contexts, our approach reveals how internal states can override immediate sensory cues to drive variation along an exploration-exploitation axis.