Title: Behavioral Strategies and Neural Mechanisms of Dynamic Foraging
Advisor: Mriganka Sur
Abstract: In dynamic environments, humans and animals constantly use reward and error signals to guide action selection. As diverse strategies may be used in the task, it is unclear how the cortical encoding of reward, value and action information might change across different behavioral states. In this thesis, we develop techniques for behavioral analysis and neural dissection of the cortical contributions to dynamic foraging behavior. In part I, we find that in a dynamic foraging task, mice use a mixture of behavioral modes that are characterized by different switch offsets, sharpness and exploration rates. We develop a new computational approach, block Hidden Markov Model, to characterize and identify these discrete states of behavior, and show that they can be accurately decoded as sub-regimes of model-free or inference-based strategies. In part II, we perform widefield imaging and unsupervised analysis of the cortical activity during the behavior to uncover distinct cortical activation modes corresponding to the frontal, motor, visual and retrosplenial regions that have different dynamic representations of rewards and errors. Dissecting single-neuron responses in these candidate regions with three-photon imaging across cortical depths reveals specialized processing of reward-related variables. There is widespread representation of outcome, value and switching in all four regions, with an enriched representation of outcome in the retrosplenial cortex (RSC), and of action values in the anterior cingulate cortex (ACC). In part III, we combine the behavioral analysis and neural imaging, showing that outcome encoding is enhanced in high-efficiency behavioral states but is only weakly represented at the low-efficiency states. Optogenetic perturbation of outcome information in the RSC neural cluster decreases the frequency of these high-efficiency states, demonstrating its causal role in the expression of efficient switching behavior. Together, these computational methods and experiments provide new tools for quantification of dynamic foraging behavior, and reveal specialized and state-dependent distribution of outcome, value and switch representations in key cortical regions. These insights lead to important hypotheses about cortical-subcortical interactions during reward-guided behavior.