Constraining Bayesian inference with cognitive architectures: An updated associative learning mechanism in ACT-R

Abstract

Bayesian inference has been shown to be an efficient mechanism for describing models of learning; however, concerns over a lack of constraint in Bayesian models (e.g., Jones & Love, 2011) has limited their influence as being a description of the ‘real’ processes of human cognition. In this paper, we review some of these concerns and argue that cognitive architectures can address these concerns by constraining the hypothesis space of Bayesian models and providing a biologically-plausible mechanism for setting priors and performing inference. This is done in the context of the ACT-R functional cognitive architecture (Anderson & Lebiere, 1998), whose sub-symbolic information processing is essentially Bayesian. To that end, our focus in this paper is on an updated associative learning mechanism for ACT-R that implements the constraints of Hebbian-inspired learning in a Bayesian-compatible framework.