A dynamic model of reasoning and memory.

Abstract

Previous models of category-based induction have neglected how the process of induction unfolds over time. We conceive of induction as a dynamic process and provide the first fine-grained examination of the distribution of response times observed in inductive reasoning. We used these data to develop and empirically test the first major quantitative modeling scheme that simultaneously accounts for inductive decisions and their time course. The model assumes that knowledge of similarity relations among novel test probes and items stored in memory drive an accumulation-to-bound sequential sampling process: Test probes with high similarity to studied exemplars are more likely to trigger a generalization response, and more rapidly, than items with low exemplar similarity. We contrast data and model predictions for inductive decisions with a recognition memory task using a common stimulus set. Hierarchical Bayesian analyses across 2 experiments demonstrated that inductive reasoning and recognition memory primarily differ in the threshold to trigger a decision: Observers required less evidence to make a property generalization judgment (induction) than an identity statement about a previously studied item (recognition). Experiment 1 and a condition emphasizing decision speed in Experiment 2 also found evidence that inductive decisions use lower quality similarity-based information than recognition. The findings suggest that induction might represent a less cautious form of recognition. We conclude that sequential sampling models grounded in exemplar-based similarity, combined with hierarchical Bayesian analysis, provide a more fine-grained and informative analysis of the processes involved in inductive reasoning than is possible solely through examination of choice data.

Sample size, number of categories and sampling assumptions: Exploring some differences between categorization and generalization

“…Recognition memory experiments have much the same structure as a generalization task: people are shown items that belong to a single list (the target category) and asked whether a test item was found on the study list. The two problems are not perfectly equivalent in that the recognition memory task asks for an identification decision rather than an inductive generalization, but there is mounting evidence (Nosofsky, 1991;Hawkins, Hayes, & Heit, 2016;Nosofsky, 2016;Nosofsky, Cox, Cao, & Shiffrin, 2014) that the underlying processes between recognition memory and induction may be much the same. With that in mind, we suggest that the Nosofsky (1991) model represents the natural way to adapt the original GCM to a generalization problem.…”

Sample size, number of categories and sampling assumptions: Exploring some differences between categorization and generalization

“…Recognition memory experiments have much the same structure as a generalization task: people are shown items that belong to a single list (the target category) and asked whether a test item was found on the study list. The two problems are not perfectly equivalent in that the recognition memory task asks for an identification decision rather than an inductive generalization, but there is mounting evidence (Nosofsky, 1991;Hawkins, Hayes, & Heit, 2016;Nosofsky, 2016;Nosofsky, Cox, Cao, & Shiffrin, 2014) that the underlying processes between recognition memory and induction may be much the same. With that in mind, we suggest that the Nosofsky (1991) model represents the natural way to adapt the original GCM to a generalization problem.…”

On the efficiency of neurally-informed cognitive models to identify latent cognitive states

“…A variable derived at the level of single trials -which can be incorporated within a discrete or continuous approach -can be extracted from any property of the task environment that is relevant to performance. For example, Hawkins et al (2016) studied the similarity between study and test items in an inductive reasoning task. The similarity relations are specified at the level of individual items, and thus can be regressed against parameters of the cognitive model in the same manner as neural data.…”