Proceedings of the Annual Meeting of the Cognitive Science Society

Deep neural networks are increasingly being used in cognitivemodeling as a means of deriving representations for complexstimuli such as images. While the predictive power of thesenetworks is high, it is often not clear whether they also offeruseful explanations of the task at hand. Convolutional neuralnetwork representations have been shown to be predictive ofhuman similarity judgments for images after appropriate adap-tation. However, these high-dimensional representations aredifficult to interpret. Here we present a method for reducingthese representations to a low-dimensional space which is stillpredictive of similarity judgments. We show that these low-dimensional representations also provide insightful explana-tions of factors underlying human similarity judgments.