Category learning in a transitive inference paradigm

Abstract

The implied order of a ranked set of visual images can be learned by transitive inference, without reliance on stimulus features that explicitly signal their order. Such learning is difficult to explain by associative mechanisms but can be accounted for by cognitive representations and processes such as transitive inference. Our study seeks to determine if those representations may be applied to categories of images without explicit verbal instruction. Specifically, we asked whether participants can (a) infer that images being presented belonged to familiar categories, even when every image presented during every trial is unique, and (b) perform transitive inferences about the ordering of those categories. To address these questions, we compared the performance of humans during a standard TI task, which used the same set of images throughout the session, to performance in a category TI tasks, which drew images from a set of categories. Each of the images used in the category TI task was only presented once, limiting the extent to which stimulus-outcome associations could be learned. Participants were able to learn the order of the categories based on transitive inference. However, participants in the category TI condition did not produce a symbolic distance effect. These findings collectively suggest that differing cognitive processes may underpin serial learning when learning about specific stimuli versus stimulus categories.

Emergent neural dynamics and geometry for generalization in a transitive inference task

“…Across the testing phase, these subjects performed consistently above chance (65-85% mean performance, Fig 8C ), while also showing considerable variability in performance (Fig 8D ), as is generally observed in TI studies (e.g. [130]).…”

Effects of post-transfer feedback informativeness in a transitive inference task

“…All of these design choices are intended to combat confounding reward effects, so that TI can't be "explained away" by differential reward rates. Finally, some studies continue to provide corrective feedback throughout the testing phase (Kao et al, 2020;Jensen et al, 2021), but rely on regression models to estimate response accuracy at the start of the testing phase. In effect, by modeling the evolution of the reward confound over time, this approach not only controls for it, but allows testing phases to be extended to larger numbers of trials, which in turn increases statistical power when characterizing performance of individual participants.…”

Memory reactivation in slow wave sleep enhances relational learning

“…Some of these focus on the preference of B over D in order to examine inference pairs. This occurs in both associative (Jensen et al, 2021) and conceptual models (Lazareva et al, 2020) and the key question is how participants are able to identify the linear order and select B over non-adjacent D when only adjacent pairs (A-B, B-C and C-D) of a hierarchy, i.e. A>B>C>D have been presented.…”