Purchase this article with an account.
Robert Geirhos, Patricia Rubisch, Jonas Rauber, Carlos R Medina Temme, Claudio Michaelis, Wieland Brendel, Matthias Bethge, Felix A Wichmann; Inducing a human-like shape bias leads to emergent human-level distortion robustness in CNNs. Journal of Vision 2019;19(10):209c. doi: https://doi.org/10.1167/19.10.209c.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Convolutional neural networks (CNNs) have been proposed as computational models for (rapid) human object recognition and the (feedforward-component) of the primate ventral stream. The usefulness of CNNs as such models obviously depends on the degree of similarity they share with human visual processing. Here we investigate two major differences between human vision and CNNs, first distortion robustness---CNNs fail to cope with novel, previously unseen distortions---and second texture bias---unlike humans, standard CNNs seem to primarily recognise objects by texture rather than shape. During our investigations we discovered an intriguing connection between the two: inducing a human-like shape bias in CNNs makes them inherently robust against many distortions. First we show that CNNs cope with novel distortions worse than humans even if many distortion-types are included in the training data. We hypothesised that the lack of generalisation in CNNs may lie in fundamentally different classification strategies: Humans primarily use object shape, whereas CNNs may rely more on (easily distorted) object texture. Thus in a second set of experiments we investigated the importance of texture vs. shape cues for human and CNN object recognition using a novel method to create texture-shape cue conflict stimuli. Our results, based on 49K human psychophysical trials and eight widely used CNNs, reveal that CNNs trained with typical “natural” images indeed depend much more on texture than on shape, a result in contrast to the recent literature claiming human-like object recognition in CNNs. However, both differences between humans and CNNs can be overcome: training CNNs on a suitable dataset induces a human-like shape bias. This resulted in an emerging human-level distortion robustness in CNNs. Taken together, our experiments highlight how key differences between human and machine vision can be harnessed to improve CNN robustness---and thus make them more similar to the human visual system---by inducing a human-like bias.
This PDF is available to Subscribers Only