Date:
Location:
(Lunch at 12:45 pm)
Daniel McNamee, Wellcome Fellow, Post-doctorate, University of Cambridge
Empirical observations and heuristic arguments have underpinned the hypothesis that the nervous system organizes behavior in hierarchies. However, the determination of behavioral and neural signatures of such hierarchical control structures are hampered by the lack of a principled computational formalism describing the allocation of cognitive or neural resources in a maximally efficient hierarchy. I will present such a theoretical model in the general context of stochastic optimal control, and then apply the model to behavioral data drawn from the literature in order to replicate a range of effects from planning compression factors to bottleneck sensitivity. Furthermore, I will describe how this general theory can be integrated with a neural policy model based on optimal population coding. This predicts spatial firing patterns which are consistent with “task-bracketing” (present in rodent striatum and infralimbic cortex), trajectory “chunking” (as recorded in rodent hippocampus), and neural indicators of environment segmentation which were observed during planning in human hippocampus and prefrontal cortex. This theory consolidates a range of unexplained data, drawn from multiple species in a variety of behavioral domains, into a common normative framework of efficient hierarchical control.