The Role of Learning in Cognitive Control: A Comparison of the Bayesian and Neural Network Learned Value of Control Models

Abstract

Recent computational models of cognitive control have focused on capturing the functions of the dorsal anterior cingulate cortex by framing control allocation as an optimization problem: the optimal control signal is the one that maximizes the expected payoff, or the difference between rewards and costs. The Learned Value of Control model introduces a method of approximating the expected value of control by associating individual stimulus features with control values, and includes a learning mechanism that uses Bayesian linear regression to update weights and improve the approximation. We reimplement the Bayesian Learned Value of Control model in the block modeling environment PsyNeuLink, and propose a novel version of the Learned Value of Control model that uses a neural network instead of Bayesian linear regression to learn. The models are tested in an environment that simulates a recent Stroop task experiment, which specifically targets the feature-related predictions of the Learned Value of Control model. We find that both the Bayesian and Neural Network Learned Value of Control models are able to learn the controlled response, and discuss potential reasons for discrepancies between our models and the original Learned Value of Control model.