Adding biological constraints to CNNs makes image classification more human-like and robust

In this study, we show that when standard convolutional neural networks (CNNs) are trained end-to-end on datasets containing low-level and spatially high-frequency features, they are susceptible to learning these potentially idiosyncratic features if they are predictive of the output class. Such features are extremely unlikely to play a major role in human object recognition, where instead a strong preference for shape is observed. Through a series of empirical studies, we show that standard CNNs cannot overcome this reliance on non-shape features merely by making training more ecologically plausible or using standard regularisation methods. However, we show that these problems can be ameliorated by forgoing end-to-end learning and processing images initially with Gabor filters, in a manner that more closely resembles biological vision.