The contrasting shape representations that support object recognition in humans and CNNs
The success of Convolutional Neural Networks (CNNs) in classifying objects has led to a surge of interest in using these systems to understand human vision. Recent studies have argued that when CNNs are trained in the correct learning environment, they can emulate a key property of human vision -- learning to classify objects based on their shape. While showing a shape-bias is indeed a desirable property for any model of human object recognition, it is unclear whether the resulting shape representations learned by these networks are human-like. We explored this question in the context of a well-known observation from psychology showing that humans encode the shape of objects in terms of relations between object features. To check whether this is also true for the representations of CNNs, we ran a series of simulations where we trained CNNs on datasets of novel shapes and tested them on a set of controlled deformations of these shapes. We found that CNNs do not show any enhanced sensitivity to deformations which alter relations between features, even when explicitly trained on such deformations. This behaviour contrasted with human participants in previous studies as well as in a new experiment. We argue that these results are a consequence of a fundamental difference between how humans and CNNs learn to recognise objects: while CNNs select features that allow them to optimally classify the proximal stimulus, humans select features that they infer to be properties of the distal stimulus. This makes human representations more generalisable to novel contexts and tasks.