Using brain-machine interfaces to identify and manipulate computational principles of learning

Learning a complex motor skill like playing tennis or a music instrument requires changes in connections between many neurons in the brain. We understand a great deal about how pairs of neurons alter their connections, but how multiple brain areas collaborate to guide learning across networks of neurons is far less clear. While we’re learning a new motor skill, our sensory system observes how we are doing, and these signals must interact with learning mechanisms in the motor systems that control our actions. What target is the brain using to drive learning computations across multiple brain areas? How does sensory feedback propagate through the brain and guide changes in a subset of neural connections? Our goal is to uncover computational principles of learning motor behaviors. We take an approach that pairs computational tools from artificial neural networks (ANNs) with brain-machine interface (BMI) experiments where the activity from motor areas directly controls the movements of a computer cursor. ANNs provide powerful ways to explore how large networks perform computations, but drawing parallels between how ANNs and brains learn is challenging because we cannot match neurons in the artificial network to neurons in the brain. BMIs simplify this challenge because they let us experimentally define the ‘behavioral output layer’ of the brain. We will analyze the activity of neurons as animals learn to control a BMI and compare the observed dynamics to those of ANNs to better understand the objectives (targets) that drive learning. This will generate new ways to identify the computations underlying changes within a large group of neurons during learning. We will also perform experiments where we alter the BMI to introduce errors while we record the activity of multiple cortical areas controlling the cursor and the eyes. We will study how visual feedback of errors propagates through the network to change activity in the neurons controlling the BMI cursor and explore whether eye movements that occur as a person gathers visual information contribute to learning. This will allow us to test hypotheses about how feedback drives learning in a network of neurons.

Our long-term objective is to uncover computational principles of sensorimotor learning dynamics. We propose two parallel and complementary objectives for this pilot study: (1) to identify learning dynamics in the brain that link task objectives to neural dynamics, and (2) to use perturbations of the behavioral sensorimotor mapping in BMIs, large-scale recordings and behavioral tracking to test hypotheses of how neural populations learn. Aim 1 will leverage existing datasets from macaque monkeys learning during BMI tasks and will further benefit from new experimental data generated in aim 2. We will use latent-space embeddings of neural data and an ANN-based computational model to test predictions of how neural dynamics evolve for different task objectives. This project will provide predictions and constraints for learning machinery that will inform new experiments exploring aim 2. Aim 2 seeks to understand how brains use global signals like visual feedback to drive local plasticity in behaviorally relevant neural populations (credit assignment). We will explore how visual feedback of errors and the eye movements driving the collection of visual information contribute to targeted plasticity in neurons directly driving BMI control. Monkeys will perform motor BMI learning tasks as we make large-scale recordings from multiple motor cortical areas (M1, PMd, FEF) and track eye movements. We will perform perturbations to the BMI sensorimotor mapping inspired by ANNs and computational models to manipulate visual feedback, testing how it alters the credit assignment learning process. Analysis of eye movement behavior and FEF-M1 interactions will probe how eye movements influence learning. These experiments will test hypotheses about how global feedback drives learning across neural populations and will further inform modeling efforts to constrain hypotheses about how learning occurs across distributed networks.

This collaborative proposal leverages the co-PIs’ expertise in ANNs and large-scale electrophysiology for motor BMIs. This work will produce deeper insights into sensorimotor learning dynamics and will generate new tools and theories for studying learning in distributed large-scale brain networks.