Learning Debiased Representations via Conditional Attribute Interpolation

Yi-Kai Zhang, Qi-Wei Wang, De-Chuan Zhan, Han-Jia Ye

Keywords: debiased representation, conditional attribute interpolation, image classification

TL;DR: This paper proposes a novel method to learn debiased representation via conditional attribute interpolation.

Abstract: An image is usually associated with more than one attribute, e.g., annotated based on both "shape" and "color". If most samples have attributes spuriously correlated with the target label, a Deep Neural Network (DNN) is prone to neglect those samples with attributes intrinsically consistent with the targets and leads to representations with large intra-class covariance. To improve the generalization ability of such a biased model, we propose a $\chi^2$-model to fill in the intra-class blanks and learn debiased representations. First, we use a $\chi$-shape pattern to match the training dynamics of a DNN and find Intermediate Attribute Samples (IASs) --- samples near decision boundaries when discerning various attributes, which indicate how attribute values change from one extreme to another. Then we rectify the decision boundary with a $\chi$-branch metric learning objective. Conditional interpolation among IASs eliminates the negative effect of peripheral attributes and facilitates making intra-class samples compact. Experiments show that $\chi^2$-model learns debiased representation effectively and achieves remarkable improvements on various datasets.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning