Sebra: DeBiasing through Self-Guided Bias Ranking

Adarsh Kappiyath¹, Abhra Chaudhuri ², Ajay Jaiswal³ , Ziquan Liu⁴ , Yunpeng Li¹ , Xiatian Zhu¹ , Lu Yin^1*

¹University of Surrey, ² Fujitsu Research of Europe, ^3,University of Texas, ⁴ Queen Mary University of London
ICLR 2025
^*Corresponding author

Paper Code

W&B

Overview

Sebra is a cutting-edge solution designed to enhance ML pipelines for addressing biases and improving data quality. Its standout features include:

Unsupervised Bias Ranking: Automatically detects biases and ranks datapoints based on their spuriosity.
Multi-Bias Mitigation: Simultaneously addresses multiple forms of bias to create fairer models.
In-the-Wild Dataset Compatibility: Designed to work effectively with challenging datasets like ImageNet-1K.
Bias Discovery: Facilitates discovery of unknown biases that might otherwise go unnoticed.
Outlier and Noise Detection: Facilitates identification of outliers and noisy samples for data refinement.

Methodology

Sebra involves three process at each epoch : Selection, Upweighted Training and Ranking. The ranking obtained via Sebra is utilised for contrastive DeBiasing.

Results

Bias Ranking obtained via Sebra facilitates discovery of previously unknown biases even in specifically curated bias datasets.

Identifying outliers and noisy samples

Sebra assigns high rank to outlier and mislabelled training samples thereby enabling easy identification of noisy samples and outliers.

Bias Mitigation

Contrastive Debiasing utilises bias ranking obtained from Sebra to mitigate multiple biases and obtain unbiased models even on in-the wild datasets like Imagenet-1K

BibTeX

@inproceedings{
      anonymous2025sebra,
      title={{SEBRA} : Debiasing through Self-Guided Bias Ranking},
      author={Anonymous},
      booktitle={The Thirteenth International Conference on Learning Representations},
      year={2025},
      url={https://openreview.net/forum?id=MyVC4X5B2X}
      },
  }