Evidence in favour of multiple-systems of category learning needs to be treated cautiously

Recently, Dr Charlotte Edmunds, a member of the Psychological Science Research Group (PSRG) at UWE Bristol, has recently published a correspondence in Nature Reviews Psychology called “Structure-based dissociations provide agnostic evidence to the multiple-systems debate” with external collaborators Prof Andy Wills and Dr Fraser Milton. This publication responds to a review of the evidence for multiple systems of category learning (Minda, Roark, Kalra & Cruz, 2024) published in the same journal.

Understanding how humans learn to classify the world into meaningful categories is foundational to cognition and successfully navigating the world. From distinguishing edible plants to identifying emotional expressions, categorisation underpins much of out everyday decision-making. The majority of Dr Edmunds’ work in this area has focused on determining whether category learning is supported by multiple competing systems or would be better (or more simply) explained by a single system.

Recently, Minda, Roark, Kalra and Cruz (2024) published a review of the evidence for multiple systems of category learning. This review focused on one particular model of category learning called COVIS (COmpetition between Verbal and Implicit Systems). As might be suggested by the name, this model claims that humans rely on both an explicit and an implicit system that compete to control responding. The system that “wins” is determined by the structure of the categories you are trying to learn. If the category is easy to describe verbally (e.g., short vs long lines) people use simple rules to assign them to a category. This process is hypothesised to be explicit (i.e., available to consciousness) and require working memory. In contrast, if the category is hard to describe verbally, COVIS hypothesises that people implicitly associate stimuli with responses with a form of procedural learning that is not accessible to consciousness. The review article covered the variety of evidence in favour of this approach to category learning.

In response, Dr Edmunds and their collaborators published a brief comment that points out that the literature investigating this model tends to conflate two popular category structures and these two proposed systems. In these experiments so called “rule-based” structure are always assumed to be learned explicitly. In contrast, “information-integration” category structures are assumed to be learned implicitly. To check this is the case, researchers in this field use formal modelling of participants strategies to check that people in each category structure condition are using a strategy that is consistent with the “correct” system. Unfortunately, previous work by Dr Edmunds has shown (Edmunds et al., 2018) that this modelling is flawed: the results are biased toward confirming this pairing, regardless of what strategy the participant was actually using. Thus, without a clear determination of which system is being used, Dr Edmunds’ and colleague conclude that the evidence for multiple categorisation system is agnostic: it may be the case that people are using two systems, or it may be that our analyses just fail to correctly identify people’s underlying strategies to categorisation.