Tests of formal models of human categorization have traditionally been restricted to artificial categories because deriving psychological representations for large numbers of natural stimuli has been an intractable task. We show that deep learning may be used to solve this problem. We train an ensemble of convolutional neural networks (CNNs) to produce the multidimensional scaling (MDS) coordinates of images of rocks. We then show that not only are the CNNs able to predict the MDS coordinates of a held-out test set of rocks, but that the CNN-derived representations can be used in combination with a formal psychological model to predict human categorization behavior on a completely new set of rocks.

A rule-plus-exception model of category learning, RULEX (Nosofsky, Palmeri, & McKinley, 1992), and an exemplar-based connectionist model of category learning, A L C O V E (Kxuschke, 1992), were evaluated on their ability to predict the types of generalization patterns exhibited by h u m a n subjects. Although both models were able to predict the average transfer data extremely well, each model had difficulty predicting certain types of generalizations shown by individual subjects. In particular, RULEX accurately predicted the prominence of rule-based generalizations, whereas A L C O V E accurately predicted the prominence of similarity-based generalizations. A hybrid model, incorporating both rules and similarity to exemplars, might best account for category learning. Furthermore, a stochastic learning rule, such as that used in RULEX , might be crucial for captiiring the different types of generalizations patterns exhibited by h u m a n s .

A classic issue in the cognitive-science of human categorylearning has involved the contrast between exemplar and prototypemodels. However, experimental tests to distinguish the models haverelied almost solely on use of artificial categories composed ofsimplified stimuli. Here we contrast the predictions from the modelsin a real-world natural-science category domain – geologic rocktypes. Previous work in this domain used a set of complementarymethods, including multidimensional scaling and direct dimensionratings, to derive a high-dimensional feature space in which the rockstimuli are embedded. The present work compares the category-learning predictions of exemplar and prototype models that makereference to this derived feature space. The experiments includeconditions that should be favorable to prototype abstraction,including use of large-size categories, delayed transfer testing, andreal-world natural category structures. Nevertheless, the results ofthe qualitative and quantitative model comparisons point toward theexemplar model as providing a better account of the observedresults. Limitations and directions of future work are discussed.

Shiffrin and Schneider (1977) demonstrated that highly efficient memory- and visual-search performance could be achieved through consistent item-to-response mapping (CM) training. It is theorized that subjects shifted from relying on working memory to learned item-response associations in long-term memory (Logan, 1988). The theory was tested and explored mostly through behavioral experiments and computational modeling. In a recent series of articles involving visual search (e.g. Woodman et al, 2013; Carlisle et al. 2011), Woodman and colleagues found that the contralateral-delay activity (CDA) of human event-related potentials is related to the maintenance of information in visual working memory and that the magnitude of the CDA decreases when target information is stored in long-term memory. We employed the CDA and other neural measures to study the nature of memory retrieval in CM memory search tasks. We observed a significant reduction in the magnitude of the CDA in CM training compared to a control condition in which item-response mappings varied from trial to trial (VM). The results provided converging evidence supporting the classic theoretical interpretation of the bases for CM and VM memory search. The results also raised interesting questions concerning the detailed interpretation of CDA.