Bias oriented unbiased data augmentation for cross-bias representation learning

The biased cues in the training data may build strong connections between specific targets and unexpected concepts, leading the learned representations could not be applied to real-world data that does not contain the same biased cues. To learn cross-bias representations which can generalize on unbiased datasets by only using biased data, researchers focus on reducing the influence of biased cues through unbiased sampling or augmentation on the basis of artificial experience. However, the distributions of biased cues in the dataset are neglected, which limits the performance of these methods. In this paper, we propose a bias oriented data augmentation to enhance the cross-bias generalization by enlarging "safety" and "unbiasedness" constraints in the training data without manual prior intervention. The safety constraint is proposed to maintain the class-specific information for augmentation while the unbiasedness constraint reduces the statistical correlation of bias information and class labels. Experiments under different biased proportions on four synthetic/real-world datasets show that the proposed approach could improve the performance of other SOTA debiasing approaches (colored MNIST: 0.35-26.14%, corrupted CIFAR10: 3.14-8.44%, BFFHQ: 1.50% and BAR: 1.72%).