Behavioral, Neural, and Computational Investigations of Task Representations

Abstract

Context-sensitive stimulus-response mappings, which I call task representations, are a crucial part of flexible, goal-oriented behavior. Guided Activation Theory (GAT) and the related Gating models integrate many ideas about the behavioral and neural properties of task representations within computational frameworks. In this dissertation, I present several projects that test and extend these ideas. In the first project, I follow up on recent Gating model work, introducing novel connections with reinforcement learning (RL) theory that lead to a greatly simplified version of the previous model as well as to new insights about the limited capacity of task representations. In the second project, I investigate a previously proposed normative account of task representation updating (in the domain of task switching experiments) in terms of an effort-performance tradeoff. Converging theoretical and experimental work from this project reveals that a simple speed-accuracy tradeoff best explains the data, suggesting that the proposed effort-performance tradeoff is not a valid account of the findings. The final project attempts to measure context representations, a component of task representations, in human fMRI data. This project presents novel methodological developments (model-based multi-voxel pattern analaysis), and describes the application of the novel and existing methods to two new fMRI experiments. Surprisingly, neither experimental dataset yields measurable context representations using methods claimed by others to have produced success, although control analyses are successful.