Research map / Learning theory and algorithms / Fairness
Are Two Heads the Same as One? Identifying Disparate Treatment in Fair Neural Networks
Michael Lohaus, Matthäus Kleindessner, Krishnaram Kenthapadi, Francesco Locatello, Chris Russell
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use two-headed network with merged heads instead of Lipton per-group thresholds | two-headed approach · CelebA Attractive, 80% disparity reduction | accuracy: 0.7719 -> 0.7698 | No clear effect |
| use two-headed approach instead of fairness regularizer | regularizer R_DP · CelebA, 50% and 80% disparity reduction | achieving required fairness | Improved |
| use two-headed approach instead of Massaging preprocessing | Massaging · CelebA, 80% disparity reduction | achieving required fairness | Improved |
| increase fairness parameter lambda | fairness regularizer R_DP (ResNet50) · CelebA, 72 target-protected pairs | protected attribute recovery accuracy from last layer: 71 out of 72 | Improved |
| increase fairness parameter lambda | fairness regularizer R_DP · CelebA, one pair | protected attribute recovery (Kendall tau): p=0.64 | No clear effect |
| replace the group output g(x) with the median output of the other group | fairest regularized model · CelebA | proportion of changed predictions: around $30\%$ | Mixed |
| replace the group output g(x) in Massaging | Massaging · CelebA | proportion of changed predictions: up to $35\%$ | Mixed |
| use regularizer instead of two-headed approach to reach near-perfect fairness | regularizer · CelebA Attractive vs Male | DDP reached: saturates around $-0.1$ | Hurt |
| use abs-DP regularizer instead of R_DP (ablation) | regularizer R_DP^abs · CelebA Smiling, MobileNetV3-Small, 50% reduction | accuracy: 0.4997 | Hurt |
| swap ResNet50 for MobileNetV3-Small | two-headed approach · CelebA Attractive, 50% reduction | accuracy: 0.8021 -> 0.8023 | No clear effect |
Closest papers by meaning
- Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition (NeurIPS 2023 · similarity 0.88)
- Inherent Tradeoffs in Learning Fair Representations (NeurIPS 2019 · similarity 0.88)
- Fair Bilevel Neural Network (FairBiNN): On Balancing fairness and accuracy via Stackelberg Equilibrium (NeurIPS 2024 · similarity 0.86)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.86)
- Adversarial Fairness Network (AAAI 2024 · similarity 0.86)
- The Disparate Benefits of Deep Ensembles (ICML 2025 · similarity 0.86)
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.
