Cognitive Bayesian Algorithmic Modeling with Controlled Coherence Variables: Perceptual Attention, Novelty Detection and Top-Down Influence of Predicted Patterns

Abstract The study of natural cognitive systems suggests complex architectures of knowledge representation and reasoning. In the probabilistic framework, modeling such architectures yields hierarchical models, and Bayesian inference yields elaborate information pathways, which can be dynamically evolving. In this paper, we present mathematical tools to help lay out hierarchical probabilistic architectures, and carefully tease apart, modulate, measure and control the resulting information flows. These mathematical tools are called “controlled coherence variables” and are an extension of coherence variables. We present three use cases, first to modulate, second to measure and third to control asymmetrically the probabilistic information flow in the model’s architecture. Then, we describe the BRAID ( Bayesian word Recognition with Attention, Interference and Dynamics ) model, a model of visual word recognition, as a case study to show how controlled coherence variables have been used to model three cognitive mechanisms: first, modulating the information flow defines a model of visual attention; second, measuring the information flow defines a model of familiarity evaluation and novelty detection; third and finally, asymmetrical control of information flows defines a model of the top-down influence of predicted patterns during perception. In each case, we provide simulation results to illustrate the properties of controlled coherence variables.