[PS-2.7] Modeling cognitive control in a hierarchical prefrontal cortex with echo state networks

Tukker, R. 1, 2 , van Rossum, A. 2, 4 , Haselager, P. 1 & Frank, S. 3

1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, The Netherlands
2 Almende B.V. , Rotterdam, The Netherlands
3 Department of Cognitive, Perceptual and Brain Sciences, University College London
4 Tilburg Centre for Cognition and Communication, Tilburg University, The Netherlands

Over the past years, considerable effort has been made to measure and model the topological mapping of abstractions over the prefrontal cortex (PFC). One central hypothesis is that the PFC is organized hierarchically, the anterior PFC corresponding to the top of the hierarchy and processing the higher abstraction levels, while the lower abstraction levels are handled by the posterior PFC at the bottom of the hierarchy. Much of this research has focused on topological maps of two types of abstractions: temporal abstraction and policy (aka task or contextual) abstraction. However, the question why these topological and hierarchical mappings would exist is not fully answered yet. One possible answer that is appealing from the cognitive perspective is that these mappings increase the performance on tasks with multiple abstraction levels by supporting the exertion of cognitive control. We investigated this hypothesis by means of computational simulation using two echo state networks as simple models of the prefrontal cortex: one with a topological mapping and one without a topological mapping of abstraction levels. The performance of these models was compared on a task similar to the n-back test for assessing temporal abstraction, and on a Wisconsin card sorting (WCS) task to investigate policy abstraction. We identified two critical factors that favor topological processing, namely a long memory length in the n-back test and a large number of low-level cues that have to be integrated and interpreted using the high-level cue in the WCS test. In contrast, a task with a short memory length or a small number of low-level cues was better handled without topological mapping. This finding can be important for theories about prefrontal cortex development, may provide a background for interpreting the many neuroimaging results, and can lead to PFC models that are embodied in a robot platform.