What do adversarial images tell us about human vision?

Abstract

Deep convolutional neural networks (DCNNs) are frequently described as promising models of human and primate vision. An obvious challenge to this claim is the existence of adversarial images that fool DCNNs but are uninterpretable to humans. However, recent research has suggested that there may be similarities in how humans and DCNNs interpret these seemingly nonsense images. In this study, we reanalysed data from a high-profile paper and conducted four experiments controlling for different ways in which these images can be generated and selected. We show that agreement between humans and DCNNs is much weaker and more variable than previously reported, and that the weak agreement is contingent on the choice of adversarial images and the design of the experiment. Indeed, it is easy to generate images with no agreement. We conclude that adversarial images still challenge the claim that DCNNs constitute promising models of human and primate vision.