Research map / Learning theory and algorithms / Fairness
Differentiable Optimization Layers for Guaranteed Fairness in Deep Learning
David Troxell, Noah Roemer, Guido Montufar
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 |
|---|---|---|---|
| train with fairness layer end-to-end instead of post-hoc projection | loan default classifier · Loan default prediction (25 splits) | AUC: +1.46% mean | Improved |
| use strict penalty (large lambda) instead of fairness layer | loan default classifier · Loan default prediction | AUC: about -45% AUC, -55% average precision | Hurt |
| fairness layer instead of post-hoc projection | wage prediction model · Employee raise recommendation | MSE: 0.0838 -> 0.0834 | Improved |
| fairness layer instead of Lagrangian penalty (lambda=100) | wage prediction model · Employee raise recommendation | MSE: 0.0933 -> 0.0834 | Improved |
| lower penalty weight lambda below 100 (ablation) | Penalty method · Employee raise recommendation | constraints satisfied: 10/10 -> 9/10 (lambda=10), 1/10 (lambda=1) | Hurt |
| fairness layer instead of post-hoc projection | image classifiers (5 backbones) · CelebA and FairFace | accuracy: Swin-T CelebA 0.888 -> 0.918 | Improved |
| fairness layer instead of post-hoc projection on small CNN | CustomCNN · CelebA | accuracy: 0.809 -> 0.814 | No clear effect |
| fairness layer instead of Lagrangian penalty on images | image classifiers · CelebA and FairFace | accuracy: 0.501 -> 0.918 (Swin-T CelebA) | Improved |
| fairness layer instead of baselines on synthetic regression | synthetic regression, 32 datasets · Synthetic data | test loss: ranked first in 22-27 of 32 datasets | Improved |
| fairness layer with tight bounds in large-batch regime | synthetic regression · Synthetic data, tight constraints (l=0, u=3.5) | test loss: underperformed on 6-8 datasets | Hurt |
| use Lagrangian penalty with validation-selected lambda | Penalty baseline · Synthetic data (btrain=binfer=2000) | fairness constraint satisfaction: 11 violations; 7 had gaps 7 times larger than target | Hurt |
Closest papers by meaning
- BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning (NeurIPS 2024 · similarity 0.85)
- Differentiable Distributionally Robust Optimization Layers (ICML 2024 · similarity 0.84)
- Enhancing Robust Fairness via Confusional Spectral Regularization (ICLR 2025 · similarity 0.83)
- On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning (ICML 2025 · similarity 0.83)
- Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference (NeurIPS 2025 · similarity 0.83)
- Enforcing Hard Linear Constraints in Deep Learning Models with Decision Rules (NeurIPS 2025 · similarity 0.83)
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.
