Unifying Perspectives on Learning Biases: A Data-Centric Intervention for Holistic Fairness, Robustness, and Generalization
Abstract
Fairness, robustness, and out-of-distribution generalization are typically studied in isolation despite sharing common roots in dataset acquisition bias, spurious shortcuts, and representational imbalance. We propose a unified data-centric framework that diagnoses and intervenes on learning biases prior to model fitting, demonstrating simultaneous gains across fairness disparities, environmental perturbations, and cross-domain transfer in resource-constrained deployments.
Key Findings & Contributions
- Formalizes procedural alignment, structuring deep learning pipelines to mirror explicit human segmentation-and-attribute grading stages.
- Evaluates 22 architectures across five model families, demonstrating that high clean-set accuracy (99.71%) coincides with negligible explanation faithfulness (2.5-5.8% EER).
- Exposes severe domain-shift vulnerability in monolithic black box classifiers (F1 dropping to 31.2-45.8%) compared with 97.2% retained F1 in decomposed architectures.
- Unifies fairness, out-of-distribution robustness, and explainability as linked consequences of background shortcut bias.
Benchmark Results
Accuracy, Explanation Evidence Ratio (EER), and Domain-Shift Resilience (ICLR 2026)
| Architecture | Clean F1 (%) | EER (%) | Shifted F1 (%) | Params (M) |
|---|---|---|---|---|
| U-Net + ResNet-18 (Decomposed) | 99.45 | 100.0 | 97.2 | 13.2 |
| Swin-Base (Monolithic) | 99.57 | 3.7 ± 3.2 | 45.8 | 88.0 |
| DeiT-Base (Monolithic) | 99.57 | 4.2 ± 3.6 | 42.1 | 86.6 |
| DenseNet-201 (Monolithic) | 99.71 | 3.1 ± 2.8 | 38.7 | 20.0 |
| EfficientNet-B2 (Monolithic) | 99.71 | 2.8 ± 3.4 | 34.5 | 9.1 |
| ResNet-50 (Monolithic) | 99.71 | 2.5 ± 3.1 | 31.2 | 25.6 |
BibTeX
@inproceedings{vincent2026unifying,
title={Unifying Perspectives on Learning Biases: A Data-Centric Intervention for Holistic Fairness, Robustness, and Generalization},
author={Vincent, Patrick Ndowo and Nyalala, Innocent},
booktitle={ICLR 2026 Workshop on Principled Design for Trustworthy AI: Interpretability, Robustness, and Safety across Modalities},
address={Rio de Janeiro, Brazil},
year={2026},
url={https://openreview.net/forum?id=RCEEHTbSDk}
} Vincent, P. N., & Nyalala, I. (2026). Unifying Perspectives on Learning Biases: A Data-Centric Intervention for Holistic Fairness, Robustness, and Generalization. In ICLR 2026 Workshop on Principled Design for Trustworthy AI, Rio de Janeiro. https://openreview.net/forum?id=RCEEHTbSDk