Semantically Deficient Models of Cognition – A Bayesian Case Study, Conceptual Spaces, and AI Opacity

Abstract Nina Poth criticizes Tenenbaum & Griffiths’ Bayesian model of perceptual generalization for its “semantic opaqueness” – roughly, it is unclear what the model’s probabilistic states are about – which suggests a connection with AI opacity. The first half of this paper analyzes the connection and suggests “semantically deficient models of cognition” (SEDMOC) as improved terminology. Crucially, for some purposes and in some contexts, models of cognition should reveal to cognitive scientists how semantic content produces behavior. While this is not at issue in the standard opacity debate, semantic opacity is one potential source of deficiency. The second half of this paper investigates hybrid models that combine Bayesian reasoning with geometric similarity (more specifically with Conceptual Spaces), thereby avoiding deficiency, and proposes to let the probability function shape spaces in a hierarchical model, outlining a PCA-based abstraction mechanism. It is argued that the proposed mechanism may introduce opaque dimensions which, however, are not semantically deficient. Both the terminological shift and the PCA-based abstraction mechanism contribute to better model-based explanations of cognitive behavior.