The mammalian brain is particularly well suited for managing streams of (often noisy) evidence, both
internally and externally generated, to converge to a decision. This evidence accumulation process can
adapt to changing environments and reward opportunities, mediated by cortico-basal-ganglia-thalamic
(CBGT) circuits that both contribute to action selection and use feedback signals to modify future
decisions. Dysfunction in how these pathways use feedback to guide future decisions is a primary
mechanism for many addictive behaviors (e.g., opioid addiction, obesity). Our prior work has identified
subsystems, which we call control ensembles, within the CBGT pathways that regulate dimensions of the
evidence accumulation process, leading to various neural states with differing receptivity to the evidence
streams that drive decisions, encapsulated in a particular decision policy.
We propose a series of empirical and theoretical investigations that bridge across levels of analysis to
understand the flow of information through CBGT circuits during the decision-making process. On the
theory side we will use our models to understand the low-dimensional representational space of CBGT
circuits throughout the decision-making process, using energy landscape models coupled with
dimensionality reduction. Using computational models we will model decision trajectories through CBGT
networks by applying entropy based analyses to the network behavior and building predictions of observed
dynamics in both discrete and continuous actions (Specific Aim 1). Empirically, we will test predictions
emerging from our network model and provide new observations to support model refinement using
experiments in rodents (optogenetics, electrophysiology) as they perform both tasks with dynamic reward
contingencies featuring either discrete choices or continuous motor control (Specific Aim 2). Our theoretical
and empirical work will evolve in a mutual-development cycle, with theoretical experiments being used to
derive novel behavioral and neural predictions and empirical experimental results being used to revise and
update the generative model properties that lead to subsequent predictions.