Modeling the neural correlates of movement

Voluntary movement is a result of a highly coordinated activity between the nervous system and the rest of the body. However, much is still unknown about the quantitative relationship between neural activity, muscle activity, and thence movement. Movement is driven by coordinated activity across a network of reciprocally connected brain areas. Many questions surround the relationship between these different brain regions and their ultimate product: muscle activity and thence movement. Often, these questions have a similar form: how do the properties of one neural population - features of the response, encoded variables, population dynamics, etc. - relate to those of another population. This question has been difficult to resolve even when comparing the neural population activity in primary motor cortex with the muscle response. An enduring hurdle is that the field often lacks appropriate computational frameworks and analytical methods for determining what is shared and what is different between populations. For example, areas that perform very different computations may display superficially similar responses. Conversely, commonalities between communicating regions can be lost amid surface-level differences. It is thus essential that comparisons be guided by concrete hypotheses, constrained both by biological reality and by theory regarding what computational mechanisms are plausible or desirable. The goal of this project is to leverage hypothesis- guided analysis methods to characterize and compare populations in data from multiple motor and non-motor areas of the primate and rodent. Modeling such neural activity in the context of movement generation using cutting-edge machine learning models is still a nascent yet growing field. Nonlinear models of neural dynamics, for example, neural networks involving dynamical models such as recurrent neural networks have been gaining popularity due to the rapid advances in computational power that have led to tractable learning of deep and complex network structures. However, the inference of these models is still challenging due to the underconstrained nature of high-dimensional models; thus the question becomes how best to constrain or regularize these in the presence of neural data.I will develop generalizable dynamical models on data collected on mice and macaque monkeys with the goal of gaining understanding of the computational principles governing movement-related neural activity. Specifically, in Aim 1, I will use a goal-driven architecture to learn a recurrent neural network that reproduces the computational structure seen in motor cortex data recorded during movements made by a macaque monkey at different speeds. In Aim 2, I will explore a data-driven approach to model neural activity from multiple regions while a mouse is making spontaneous and task-driven movements, and explore differences between the dynamics of the neural activity during these modes. In Aim 3, I will combine the two approaches, namely goal-driven and data-driven, to develop a dynamical model learnt using end-to-end training for the generation of task-driven movements. This will help elucidate the role of neural activity from different regions in the context of behavior. While we hope to make significant scientific advances while addressing the aims as discussed below, there will doubtless be computational developments that may even surpass the scientific contributions made while addressing the aims of this proposal.