Learning Debiased Representations via Conditional Attribute Interpolation

From the main paper, we design a χ-shape pattern to match the training dynamics of a DNN and find Intermediate Attribute Samples (IASs) -samples near the attribute decision boundaries. Then we rectify the representation with a χ-structured metric learning objective. In this supplementary material, we present more related work of learning a debiased model from what bias information is provided in advance in Section 1. Further, we describe more implementation details in Section 2, such as the visualization of Figure 2 , matching factors A 1 and A 2 of Equation 3, and the Biased NICO dataset construction of subsection 4.1 in the main paper. We also show additional experimental observations and results in Section 3, including error intervals, robustness analysis, etc.