Proceedings of the Annual Meeting of the Cognitive Science Society

Convolutional neural networks (CNNs) were inspired by hu-man vision and, in some settings, achieve a performance com-parable to human object recognition. This has lead to the spec-ulation that both systems use similar mechanisms to performrecognition. In this study, we conducted a series of simulationsthat indicate that there is a fundamental difference between hu-man vision and vanilla CNNs: while object recognition in hu-mans relies on analysing shape, these CNNs do not have sucha shape-bias. We teased apart the type of features selectedby the model by modifying the CIFAR-10 dataset so that, inaddition to containing objects with shape, the images concur-rently contained non-shape features, such as a noise-like mask.When trained on these modified set of images, the model didnot show any bias towards selecting shapes as features. In-stead it relied on whichever feature allowed it to perform thebest prediction – even when this feature was a noise-like maskor a single predictive pixel amongst 50176 pixels.