Quantifying Natural Movement Variation in the Brain and Behavior

Abstract: The ultimate goal of neural processing is to drive reliable behaviors in an animal’s natural environment, maximizing fitness in a complex environment. Our hypothesis posits a causal, evolved relationship between the complexity and structure of the brain and behavior and requires new mathematical approaches to both quantifying and recapitulating this matching between natural and neural state space. To deepen our understanding of motor encoding and control, we will integrate behavioral recording in freely moving mice executing a seed reach-to-grasp task with extensive, longitudinal tracking of neuronal activity across various layers in the motor cortex. This involves pairing detailed behavioral observations with comprehensive neuronal population recordings to characterize the mapping between the brain’s control space and the resulting movement space exhibited by the arm and paw during the challenging task. To establish a direct connection between behavioral and neural data, we will utilize machine learning tools such as VAEs and U-nets to quantify the latent space of both datasets. Our primary objective with these advanced machine learning approaches is to identify interpretable features within the representations, and to develop new mathematics to define trajectories in this feature space. The goodness of fit will be assessed by training models on behavioral data and evaluating their ability to generate realistic limb and paw trajectories, with the constraints that these are differentiable and low-dimensional. Throughout our investigation, our specific focus will be on uncovering the features of the neural response that drive variable yet successful reach movements. By examining how the brain’s code aligns or deviates from behavioral complexity, our goal is to reveal new principles of motor encoding and control that operate over both evolutionary and organismal timescales