arXiv AI

When majority rules, minority loses: bias amplification of gradient descent

The paper presents a formal framework for majority‑minority learning tasks, demonstrating that standard gradient‑descent training can amplify bias by favoring majority groups and producing stereotypical predictors that ignore minority‑specific features. It identifies three key theoretical findings: the close proximity between full‑data and stereotypical predictors, the dominance of a region where training the entire model mainly captures majority traits, and a lower bound on the extra training needed to mitigate this effect. Experiments on deep learning for tabular and image classification illustrate these results.

arXiv Machine Learning
4d ago

Bias-Corrected Data Synthesis for Imbalanced Learning

The paper introduces a bias‑correction method for synthetic oversampling in imbalanced learning. It estimates the loss discrepancy caused by the data generator using a held‑out majority subset and transfers this correction to the minority class under a uniform bias‑transfer assumption. The authors provide finite‑sample bounds for bias transfer and excess balanced risk, identify when SMOTE introduces significant bias, and demonstrate the method’s applicability to multi‑task learning and propensity‑score estimation, with empirical results showing greatest benefit when synthetic distortion is large.

By Pengfei Lyu, Zhengchi Ma, Linjun Zhang, Anru R. Zhang
arXiv Machine Learning
Sep 2

SAGE: Subpopulation-Aware Generative Enhancement for Mitigating Spurious Correlations

SAGE (Subpopulation-Aware Generative Enhancement) is a two-stage generative augmentation framework designed to mitigate spurious correlations in machine learning when group labels are unavailable. It uses cluster-derived sub-labels and class labels to fine‑tune a conditional generative model and text encoder, producing synthetic data that fills underrepresented regions and creates a balanced validation set for last‑layer reweighting. Experiments show SAGE improves worst‑group accuracy to 89.5%, 85.7%, and 79.1% on Waterbirds, CelebA, and MetaShift, outperforming existing group‑label‑free baselines by up to 7.7 percentage points.

By Yiming Luo, Rongqiang Zhao, Jie Liu
arXiv AI
Aug 25

Fairness-Aware Mixture-of-Experts via Subgroup Reweighting and Gate Regularization

The paper proposes a fairness-aware Mixture-of-Experts (MoE) framework that tackles routing-induced bias by applying subgroup reweighting to correct data imbalance and gate entropy regularization to prevent the gating network from collapsing onto subgroup attributes. This end-to-end approach keeps expert utilization balanced and interpretable, offering a clear view of how subgroups are allocated across experts. Experiments show that the method improves fairness while maintaining competitive predictive performance.

By Sunhee Hwang
arXiv Machine Learning
Aug 27

Controlling for Omitted Variable Bias in Deep Neural Networks

The paper introduces a control‑variable framework for deep neural networks to mitigate omitted variable bias, particularly shortcut learning where covariates like demographics influence predictions. It refits the final layer of a pre‑trained network using cross‑fitting with ridge penalisation, orthogonalises covariate effects, and marginalises predictions over covariate distributions to achieve unbiased, interpretable results. Experiments on simulated images and neuroimaging data show consistent estimation of true effects and performance close to models trained on unconfounded data.

By Manuel Pfeuffer, Roshan Prakash Rane, Kerstin Ritter, Sonja Greven
arXiv AI
Jul 17

Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers

arXiv:2607. 14984v1 Announce Type: new Abstract: Per-subgroup fairness audits of medical image classifiers face a sample-size problem: minority subgroups in held-out test sets have so few samples that the resulting confidence intervals on per-subgroup performance are wider than the bias the audit is meant to detect.

By Mahmoud Ibrahim, Bart Elen, Chang Sun, Gokhan Ertaylan, Michel Dumontier
arXiv Machine Learning
Jul 16

How the Hessian-Spectrum of Neural Networks Depends on Data

arXiv:2607. 13631v1 Announce Type: new Abstract: The Hessian matrix is an important quantity of interest when it comes to studying the loss landscape and optimization dynamics in deep learning, as well as designing measures of generalization, second-order learning algorithms, etc.

By Jasraj Singh, Enea Monzio Compagnoni, Antonio Orvieto