| Paper: | PS-2A.42 |
| Session: | Poster Session 2A |
| Location: | Symphony/Overture |
| Session Time: | Friday, September 7, 17:15 - 19:15 |
| Presentation Time: | Friday, September 7, 17:15 - 19:15 |
| Presentation: |
Poster
|
| Publication: |
2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania |
| Paper Title: |
Corticostriatal signatures of learning efficient internal models for control |
| Manuscript: |
Click here to view manuscript |
| DOI: |
https://doi.org/10.32470/CCN.2018.1125-0 |
| Authors: |
Daniel McNamee, University of Cambridge, United Kingdom; Matthew Botvinick, DeepMind, United Kingdom; Samuel Gershman, Harvard University, United States |
| Abstract: |
Control of high-dimensional, dynamical systems such as the body or the world imposes large complexity costs on the subserving neural hardware. We consider the hypothesis that, in order to make efficient use of its resources, the brain adaptively compresses its internal models via reinforcement learning. We developed a normative measure of the importance of stimulus information in determining future action-outcome trajectories which can be updated via prediction errors. In a planning task, we found that decision reaction times were modulated by these predictions errors and that behavioral efficiency was strongly correlated with the strength of this modulation on a per-participant basis. Analysis of functional magnetic resonance imaging data indicated that three essential components of the model were encoded in focal cortical and striatal regions which were known to contain the necessary stimulus and action representations a priori. We suggest these data provide preliminary evidence that the brain monitors the efficiency of its internal models and updates them accordingly in the associative corticostriatal loop and that the results of these computations are reflected in sensorimotor loop activity and behavior. |
