Proceedings of the Annual Meeting of the Cognitive Science Society

Deep learning networks have shown impressive performance in object recognition. We used the classification image method to probe whether a deep learning model employs the same features as humans in perceiving real and illusory contours. We adopted a deep learning network, pre-trained with natural images, and retrained the decision layer with laboratory stimuli to perform shape discrimination in the “fat/thin” task. We tested the network with real and illusory contour stimuli contaminated with luminance noise. We found that deep networks trained on natural images can be readily adapted to discriminate between psychophysical stimuli with an extremely high degree of accuracy. However, deep learning networks do not appear to represent illusory contours where they may aid performance in the fat/thin task, a process automatically performed in human vision. This divergence indicates an important difference between the kinds of visual representations formed by deep networks and by humans.