Machine Learning methods to enable Naturalistic Neuroscience and Neuroengineering

Abstract

Recent advances in brain recording technology and in algorithms for analyzing behavioral data are enabling the study of neural activity underlying spontaneous behaviors.This new paradigm called "naturalistic neuroscience” goes beyond the confines of traditional neuroscience experiments that rely on cued, repeated trials, and a priori experimental design. In this dissertation, we describe how we use machine-learning to study increasingly naturalistic behaviors and associated neural data in two distinctly different settings. In the first "data-rich" setting, we study large (~250 GB/subject) opportunistically-collected datasets of simultaneously recorded long-term (7--10 day) electrocorticography and naturalistic behavior video data for 12 human subjects. Our approach uncovers and annotates thousands of instances of human upper-limb movement events from the video recordings, using a pipeline comprising computer-vision, discrete latent-variable modeling, string pattern-matching and event metadata extraction. We curate these events into a database that can be used for many downstream applications in neuroscience and neuroengineering, two of which we prototype -- (1) time-frequency analysis and (2) movement initiation decoding. We have published our curated dataset, making available a resource that captures naturalistic neural and behavioral variability at a scale not previously available. In the second "simulation-based" setting, we study plume-tracking, a complex control problem requiring multimodal sensory integration and robustness to odor intermittency, wind non-stationarity and spatiotemporal plume variability. Flying insects routinely track plumes, often over long distances, in pursuit of odors originating from food or mates. Isolated aspects of this remarkable behavior have been studied in detail in many experimental studies. We take a complementary in silico approach, using artificial agents trained with reinforcement learning, to develop an integrated understanding of the behaviors and neural computations that support plume tracking. Specifically, we use deep reinforcement learning to train recurrent neural-network (RNN) based agents to locate the source of simulated turbulent plumes. Interestingly, the agents' emergent behaviors resemble those of flying insects, and the RNNs learn to represent task-relevant variables such as head-direction and time between odor encounters. Exploiting the simulator's flexibility and the full observability of the RNNs' neural activity, we also generate insights into behavior modularity, memory capacity, neural computations and network connectivity that support plume tracking in a variety of easily simulated but hard-to-realize plume configurations. Our in silico approach provides key intuitions for an integrated understanding of turbulent plume tracking and motivates future targeted experimental and theoretical developments.