Proceedings of the Annual Meeting of the Cognitive Science Society

Modern convolutional neural networks (CNNs) are able toachieve human-level object classification accuracy on specifictasks, and currently outperform competing models in explain-ing complex human visual representations. However, the cate-gorization problem is posed differently for these networks thanfor humans: the accuracy of these networks is evaluated bytheir ability to identify single labels assigned to each image.These labels often cut arbitrarily across natural psychologi-cal taxonomies (e.g., dogs are separated into breeds, but neverjointly categorized as “dogs”), and bias the resulting represen-tations. By contrast, it is common for children to hear bothdog and Dalmatian to describe the same stimulus, helping togroup perceptually disparate objects (e.g., breeds) into a com-mon mental class. In this work, we train CNN classifiers withmultiple labels for each image that correspond to different lev-els of abstraction, and use this framework to reproduce classicpatterns that appear in human generalization behavior.