ImageNet-trained CNNs are biased towards texture
This is a talk by Robert Geirhos presented at ICLR in 2019 titled ' ImageNet-trained CNNs are biased towards texture: increasing shape bias improves accuracy and robustness'.
Key sentences from their abstract below:"We show that ImageNet-trained CNNs are strongly biased towards recognising textures rather than shapes, which is in stark contrast to human behavioural evidence and reveals fundamentally different classification strategies. We then demonstrate that the same standard architecture (ResNet-50) that learns a texture-based representation on ImageNet is able to learn a shape-based representation instead when trained on 'Stylized-ImageNet', a stylized version of ImageNet. This provides a much better fit for human behavioural performance".
1: Once again, the importance of the prior information inherent in the training set.
2: Shape bias in training set leads to noise robustness. What does this tell us about how we should approach data augmentation?
You can find the paper here.
Post a Comment