Conceptual Spaces as Predictive Models

Research shows that humans reason based on similarity to past experience. But we also routinely assign probabilities to our beliefs and update them dynamically in interaction with the world. This talk explores how similarity and probability work together in reasoning. Treating Conceptual Spaces (CS) as predictive models offers a new perspective. This means that our geometrically organized conceptual knowledge serves as a model to generate expectations or priors for probability calculations, with the ultimate goal of reducing uncertainty. Thus, the discussion is contextualized with considerations about the compatibility of CS and Predictive Processing. Concrete mechanisms to be discussed are: Generating priors by relative region size, generating priors by distance to prototype, and probability density functions on CS.