Neural Encoding of Prior Experience in Sensorimotor Behavior

One influential hypothesis in neuroscience holds that the nervous system learns statistical regularities in the environment to optimize behavior based on past experiences. The main challenge in evaluating this hypothesis is to reconcile conceptual views that have historically been developed at different scales. At the behavioral scale, the effect of statistical regularities is often described by the Bayesian theory in terms of prior distributions that represent knowledge previously gathered about the environment. At the neural scale, the effects of prior experience have been described by the theory of predictive processing in terms of efficient coding principles that govern the response properties of neurons. The major contribution of this thesis is to bridge these two levels of description. Using a series of time-intervalreproduction tasks in rhesus macaques, I first establish a quantitative link between temporal regularities in the environment and coding properties of neurons in the frontal cortex. Specifically, I show that patterns of activity across populations of neurons are precisely rescaled in time to match the statistical mean of a learned temporal distribution, in accordance with predictive processing. Second, I show that the structure of the underlying neural representation implements the effect of a Bayesian prior, and biases behavioral responses toward prior expectations as predicted by the theory. Third, I demonstrate that the results hold in non-stationary environments when animals adapt to new temporal statistics. Fourth, I present a computational model that recapitulates the behavioral and neural findings and provides a solution for incorporating temporal expectations in neural dynamics. Finally, I conclude with a broader perspective on sensorimotor learning.Ph.D