A Cognitive Category-Learning Model of Rule Abstraction, Attention Learning, and Contextual Modulation

Abstract

We introduce the CAL model (Category Abstraction Learning), a cognitive framework formally describing category learning built on similarity-based generalization, dissimilarity-based abstraction, two attention learning mechanisms, error-driven knowledge structuring and stimulus memorization. Our hypotheses draw on an array of empirical and theoretical insights connecting reinforcement and category learning, and working memory. The key novelty of the model is its explanation of how rules are learned from scratch based on three central assumptions. (1) Category rules emerge from two processes of stimulus generalization (similarity) and its direct inverse (category contrast) on independent dimensions. (2) Two attention mechanisms guide learning by focusing on rules, or on the contexts in which they produce errors. (3) Knowing about these contexts inhibits executing the rule, without correcting it, and consequently leads to applying partial rules in different situations. We show that the model decisively outperforms the established category-learning models ALCOVE (Kruschke, 1992), SUSTAIN (Love, Medin, & Gureckis, 2004) and ATRIUM (Erickson & Kruschke, 1998) on data sets from benchmark studies, including cross-validations based on trial-wise eye-movements. Additionally, CAL's three free parameters, which measure abstraction, memorization and attention control, are related to abilities measured in working memory tasks in a theoretically meaningful way. We illustrate the model's explanatory scope by simulating several phenomena (peak shift, sample size, instruction effects, extrapolation), which were so far unexplained (or unexplained within a single model). We discuss CAL's relation to existing accounts, and its promise in understanding the role of attention control and working memory in category learning and related domains.

The influence of reward magnitude on stimulus memory and stimulus generalization in categorization decisions.

“…The absence of an effect in the TR condition, however, seems more puzzling and could be related to the presence of atypical items in general or less effective reward learning, as it appeared that reward generalization was concurrently absent in the GCM group of the TR condition. However, given the exploratory nature of Study 3, and the ongoing debate regarding the special role of atypical items (or exceptions) in memory and their implications for the representation of categories (e.g., Erickson & Kruschke, 1998; Nosofsky et al, 1994; Poldrack et al, 2001; Poldrack & Foerde, 2008; Savic & Sloutsky, 2019; Schlegelmilch, Wills, & von Helversen, 2018), these explanations should be considered speculative, and further research is needed to disentangle the effects of typicality and item rewards.…”

“…Consequently, our findings highlight the need to investigate existing category-learning accounts, such as the configural cue model (e.g., Gluck & Bower, 1988), the attention learning covering map (ALCOVE; Kruschke, 1992), the supervised and unsupervised stratified adaptive incremental network (SUSTAIN; Love et al, 2004), the rational model (Sanborn et al, 2010), the divergent autoencoder (Kurtz, 2007), the S.O.S. network (Goldstone et al, 1996), the category abstraction learning model (Schlegelmilch et al, 2018), and other accounts (see Pothos & Wills, 2011) and assumptions about how reward magnitude affects learning and generalization, and to test their ability to consistently explain empirical data in different tasks.…”

A cognitive category-learning model of rule abstraction, attention learning, and contextual modulation.

“…To simulate the instruction effect in Figures 15 and 16 (D1 and D2), we sampled values from the two different γ distributions with either low values of γ from a homogeneous distribution (D1; i.e., assuming a reduction of diversity “with rule instructions”) or higher values of γ from a heterogeneous distribution (D2; “without rule instructions”). This change of γ was the main driver of the observed differences (see also Schlegelmilch et al, 2018).…”

Rethinking Norm Psychology

“…Compliance. In infancy and early childhood, compliance depends on domain-general processes of categorisation and reinforcement learning (Ayub & Wagner 2020;Foster-Hanson et al 2021;Morris & Cushman 2018;Schlegelmilch et al 2021;Wellman, Kushnir & Brink 2016). For example, in adult commentary 'giving' is more normatively loaded than 'reaching', but young children learn to categorise a variety of different body movements as (what an adult would call) giving, and that giving has positive outcomes in many contexts (e.g., hugs), in the same way that they learn to treat a variety of different body movements as reaching, and that reaching has positive outcomes in many contexts (e.g., toys; Pulverman et al 2006).…”

“…They are not necessarily or exclusively based on stored exemplars of behaviour (Sripada and Stich's criterion), and the behavioural regularities they produce can be described by rules, such as 'Children should stand when an adult enters a room', but these are very low bars for rule-hood. Even the learning of simple categories, such as 'bird', does not depend necessarily or exclusively on stored exemplars (Schlegelmilch, Wills & Helversen 2021), and any regularity in nature, anything that is not random, can be described from the outside by a rule. Implicit 42 normativity does not have rules on the inside; it is not produced by rule-like mental structures, by sentences in the head.…”