Modeling the hidden dynamics of Drosophila behavior with recurrent neural networks

Animal behavior occurs across a variety of time scales, from single actions like grooming that lasts seconds to other behaviors like circadian rhythms that last hours. As a result, we see a hierarchy of time scales emerge in behavioral data. We can describe behaviors by observable states, or motor outputs, and the latent states which modulate them. While our ability to measure the motor outputs has greatly increased in recent years, we lack a conceptual framework for measuring the hidden dynamics. In this talk, we characterize the latent structures modulating behavior by establishing a model that reproduces the dynamics that we see in the movements of the fruit fly Drosophila melanogaster. Previous work has shown that Markov models, although commonly used, are unable to generate the full spectrum of the observed time scales. We fit recurrent neural networks (RNNs) to observed sequences of fly behavior. From the structure of the network weights, we infer the hidden state dynamics underlying the long time scales. Using this approach, we build an interpretive model in order to understand the structure of the fly’s internal states and how they drive behavior.