Training set coherence and set size effects on concept generalization and recognition

Abstract

Building conceptual knowledge that generalizes to novel situations is a key function of human memory. Category-learning paradigms have long been used to understand the mechanisms of knowledge generalization. In the present study, we tested the conditions that promote formation of new concepts. Participants underwent 1 of 6 training conditions that differed in the number of examples per category (set size) and their relative similarity to the category average (set coherence). Performance metrics included rates of category learning, ability to generalize categories to new items of varying similarity to prototypes, and recognition memory for individual examples. In categorization, high set coherence led to faster learning and better generalization, while set size had little effect. Recognition did not differ reliably among conditions. We also tested the nature of memory representations used for categorization and recognition decisions using quantitative prototype and exemplar models fit to behavioral responses. Prototype models posit abstract category representations based on the category's central tendency, whereas exemplar models posit that categories are represented by individual category members. Prototype strategy use during categorization increased with increasing set coherence, suggesting that coherent training sets facilitate extraction of commonalities within a category. We conclude that learning from a coherent set of examples is an efficient means of forming abstract knowledge that generalizes broadly. (PsycInfo Database Record (c) 2020 APA, all rights reserved).

Figure 1.

Category-learning task. A. Category representations and generalization to new items under the assumptions of the exemplar and prototype models. Exemplar: categories are represented as individual exemplars. New items are classified into the category with the most similar exemplars. Prototype: categories are represented by their central tendencies (prototypes). New items are classified into the category with the most similar prototype. B. Example stimuli. The leftmost stimulus is the prototype of category A and the rightmost stimulus is the prototype of category B, which shares no features with prototype A. Members of category A share more features with prototype A than prototype B and vice versa for members of category B. Stimulus coherence is computed by dividing the number of prototypical features by the total features (10) to compute the percentage of typical features. C. Participants underwent feedback-based training with one of six possible training sets that varied the size of the training set and the coherence of the examples. D. In recognition, participants were shown training (old) items and never seen category members and made old/new judgments. E. In categorization, participants were shown training (old) items and never seen category members and made categorization judgments without feedback.

Figure 2.

Training accuracy. Mean accuracy from each block of the training separated by training group. In the legend, the number of items corresponds to the set size manipulation and the percentages indicate the average percentage of typical features in the training set, corresponding to the set coherence manipulation. Error bars depict the standard error of the mean.

Figure 3.

Categorization accuracy for each item type separated by training set. Training sets are organized with increasing coherence from left to right, with specific coherence levels indicated by the percentages on the x-axis. Training set size is indicated on the x-axis by the number of items per category. Accuracies on training (old) items re-presented during the categorization test are depicted with striped bars. Accuracies for new items (common across all groups) varying in their similarity to category prototypes are depicted with solid bars. Accuracies for prototypes are depicted in the darkest bars with increasingly lighter bars for new items sharing fewer features with the prototypes. Error bars depict standard error of the mean.

Figure 4.

Recognition task. A. Overall recognition performance measured by corrected hit rates (hits – false alarms) B. Proportion of ‘old’ responses during recognition for items varying in their presentation history (old/training items v. all others/new items) and similarity to prototypes (60%–100% typical). Hit rates for training (old) items re-presented during the recognition test are depicted with striped bars. False alarm rates for new items varying in their similarity to category prototypes are depicted with solid bars. False alarm rates for prototypes are depicted in the darkest bars with increasingly lighter bars for new items sharing fewer features with the prototypes. Both graphs separate results by training group: coherence levels are indicated by the percentages on the x-axis, and training set size is indicated on the x-axis by the number of items per category. Error bars depict standard error of the mean.

Figure 5.

Raw model fit error for categorization data. Relative exemplar (x-axis) and prototype (y-axis) model fits for each subject in terms of negative log likelihood. Trend line represents equal exemplar and prototype model fit. Dots below the trend line represent subjects with smaller model error for the prototype model compared to the exemplar model. Dots above the trend line represent subjects with smaller model error for the exemplar model relative to the prototype model.

Figure 6.

Model fits. A. Models fit to categorization data and B. recognition data. The percent of subjects better fit by the prototype than exemplar model is depicted in blue, the percent better fit by the exemplar than prototype model in red, the percent who were comparably fit by both models (‘similar’) in purple, and those whose fits did not differ from chance in grey. Both graphs separate results by training group: coherence levels are indicated by the percentages on the x-axis, and training set size is indicated on the x-axis by the number of items per category. The dashed lines depict the percent of subjects best fit by the prototype model across the entire group for each respective test.