Sebastian Scholz

The cognitive science literature suggests that human reasoning is both a similarity machine and a probability engine, but a clear picture of the relationship between these mechanisms is still missing. Recent work combines Conceptual Spaces (CS) – a similarity-based theory about the structure of conceptual content – with probabilistic, or Bayesian theories of reasoning. This synthesis not only reframes the problem but also prompts a broader question – namely, whether the CS account is compatible with Predictive Processing (PP), which is a thriving research paradigm that aims to derive a grand unified theory of cognition from Bayesian principles. If so, then CS can serve as components of predictive models. This paper articulates compatibility constraints, engages critically with ideas from the aforementioned literature, and evaluates their significance for the compatibility question. The literature provides several mechanisms through which similarity and probability may interact in the human mind, including deriving priors from relative region size, distance to prototype, and updating probability distributions over CS using Gaussian mixtures. These mechanisms support what I call partial Marr-horizontal methodological compatibility. This approach is contrasted with the possibility of reducing psychological similarity to probability.

The conceptual spaces framework posits that conceptual content is structured geometrically, and is equipped with cognitive criteria of naturalness (namely, convexity and principles of cognitive economy). Its proponents suggest that cognitive naturalness is naturalness simpliciter, a novel move in a debate that is traditionally focused on how the world, and not the mind, is structured. We argue that “cognitive naturalness” is a misnomer and that the framework describes cognitive sparseness instead. To demonstrate this, we explore the approach’s shortcomings across various branches of the naturalness debate, most notably its failure to distinguish natural kinds from fictional kinds. Our diagnosis is that the evolutionary pragmatism employed by its proponents fails to establish a connection to the real world, thus failing to secure the ontological and epistemic objectivity required for a theory of naturalness. We propose an alternative view, ecological empiricism, which posits that natural concepts or properties are those revealed through interaction with the real world.

Nelson Goodman observed that we use only certain ‘good’ (viz. projectible) predicates during reasoning, with no obvious demarcation criterion in sight to distinguish them from the bad and gruesome ones. This apparent arbitrariness undermines the justifiability of our reasoning practices. Inspired by Quine’s 1969 paper on Natural Kinds, Peter Gärdenfors proposes a cognitive criterion based on his theory of Conceptual Spaces (CS). He argues the good predicates are those referring to natural concepts, and that we can capture naturalness in terms of similarity. In contrast to Quine, he does not primarily rely on logic, but geometry. He frames his account as a descriptive project, however, and it is not obvious how it addresses the normative dimension of Goodman’s Riddle. This paper develops a charitable reconstruction of his argument, based on the idea that the instrumental success of our projectible concepts is grounded in their cognitive-pragmatic naturalness. It then explores three lines of reasoning against the argument: Evolutionarily motivated skepticism, the miracles argument, and the relation between instrumental and pragmatic success. I conclude that in its current form, the CS account fails to provide any justification of reasoning beyond appealing to its instrumental success, and that a metaphysically robust notion of naturalness helps to achieve the desired goal.