Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of Biases-Specific Experts

ResNet-18 DebiAN PnD (ours) (a) (b) (c) Figure 1: Taking the age attribute in the CelebA dataset as an example for analyzing the agnostic biases problem. (a) Representative samples in CelebA containing multiple biases. By analyzing the attribute distribution of all data within the young/old category, we found that three attributes can be biased: gender (female/male), attractiveness (attractive/not attractive), and wearing lipstick (lipstick/no lipstick). (b) Proportion of samples with 0 -3 biases in a single image within the young and old groups. This indicates that the number of images with multiple biases dominated other cases in the dataset. (c) Age classification accuracy (%) of the existing methods for the worst groups of three bias attributes in CelebA degrades under this realistic bias scenario. For clarification, when discussing biases in this paper, we refer to abstract words like age as "attribute", and the italic words which describe the labels of age like young/old as "category".