Research map / Optimization / Implicit bias and generalization
When majority rules, minority loses: bias amplification of gradient descent
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Develops a perturbation-theoretic framework for majority-minority gradient descent, showing the full-data loss's critical points lie close to majority-only ("stereotypical") critical points, and that debiasing requires a lower-bounded extra training time, illustrated on CIFAR-10, EuroSAT and Adult.
Problem
Why standard gradient-based training under strong population and variance imbalance first learns majority-only (stereotypical) predictors and how long debiasing takes; theory of bias amplification is poorly understood.
Approach
Treats the full loss as L = L1 + L0 with L0 small in C2 norm, uses Kantorovich/Newton-type perturbation arguments to pair critical points of L with those of L1 (the stereotype gap), defines majority-training and majority-adverse zones, and derives lower bounds on the time needed to escape a stereotypical critical point (debiasing duration and catch-up overcost).
Key results
- On imbalanced CIFAR-10 with ResNet18, catch-up overcost ranges from roughly 25% to 450% depending on imbalance and threshold.
- On EuroSAT, minority-group catch-up overcost is about 45% at a 90% accuracy threshold.
- On Adult with TabNet, catch-up overcost is about 416% at a 90% accuracy threshold.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase subgroup imbalance | CIFAR-10 models · CIFAR-10, κ=90% | Catch-up overcost (%) | Helped |
| Use a larger model | CIFAR-10 models · CIFAR-10, ζ=1%, κ=90% | Catch-up overcost (%) | Helped |
| Train until minority accuracy reaches the target, rather than stopping at global accuracy | Gradient-trained models · Imbalanced CIFAR-10, EuroSAT, and Adult | Training duration | Helped |
| Use AdamW or XGBoost instead of the tested gradient-training approach | Imbalanced subgroup learning · Numerical experiments | Relative training increase for minority awareness | No clear effect |
Surprising findings
- Under strong imbalance, the minority-group accuracy stays near zero for many epochs even though global accuracy is high, so global training metrics hide the bias.
- Larger models reduce the catch-up overcost but do not eliminate the need for longer training.
Limitations
- Theoretical bounds are conservative and the worst-case biasing complexity is not established.
- Experiments are preliminary and use a small set of architectures and datasets with binary majority/minority splits.
- Assumptions (C2-smallness of L0, Morse non-degeneracy) are not verified for deep networks.
Experimental setup
Datasets: CIFAR-10 EuroSAT Adult
Benchmarks: CIFAR-10 EuroSAT Adult
Metrics: training accuracy catch-up overcost average loss per group
Baselines: AdamW XGBoost
Builds on: Newton-Kantorovich theorem ridge regularization
Closest papers by meaning
- An Effective Theory of Bias Amplification (ICLR 2025 · similarity 0.89)
- Subgroups Matter for Robust Bias Mitigation (ICML 2025 · similarity 0.86)
- fairret: a Framework for Differentiable Fairness Regularization Terms (ICLR 2024 · similarity 0.84)
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.84)
- Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (CVPR 2024 · similarity 0.84)
- Learning under Localized Minority Imbalance (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.