Increasing Shape Bias in ImageNet-Trained Networks Using Transfer Learning and Domain-Adversarial Methods
It addresses robustness issues in image classification for AI applications, but is incremental as it builds on existing style-transfer approaches.
This work tackled the problem of texture and color bias in CNNs by extending style-transfer methods with domain-adversarial training to increase shape bias, resulting in improved robustness and shape bias but no accuracy gain.
Convolutional Neural Networks (CNNs) have become the state-of-the-art method to learn from image data. However, recent research shows that they may include a texture and colour bias in their representation, contrary to the intuition that they learn the shapes of the image content and to human biological learning. Thus, recent works have attempted to increase the shape bias in CNNs in order to train more robust and accurate networks on tasks. One such approach uses style-transfer in order to remove texture clues from the data. This work reproduces this methodology on four image classification datasets, as well as extends the method to use domain-adversarial training in order to further increase the shape bias in the learned representation. The results show the proposed method increases the robustness and shape bias of the CNNs, while it does not provide a gain in accuracy.