AI Paper Atlas

Research map / Learning theory and algorithms / Fairness

Differentiable Optimization Layers for Guaranteed Fairness in Deep Learning

David Troxell, Noah Roemer, Guido Montufar

ICML 2026 · Learning theory and algorithms › Fairness · approach: Fairness auditing and regularization · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
train with fairness layer end-to-end instead of post-hoc projectionloan default classifier · Loan default prediction (25 splits)AUC: +1.46% meanImproved
use strict penalty (large lambda) instead of fairness layerloan default classifier · Loan default predictionAUC: about -45% AUC, -55% average precisionHurt
fairness layer instead of post-hoc projectionwage prediction model · Employee raise recommendationMSE: 0.0838 -> 0.0834Improved
fairness layer instead of Lagrangian penalty (lambda=100)wage prediction model · Employee raise recommendationMSE: 0.0933 -> 0.0834Improved
lower penalty weight lambda below 100 (ablation)Penalty method · Employee raise recommendationconstraints satisfied: 10/10 -> 9/10 (lambda=10), 1/10 (lambda=1)Hurt
fairness layer instead of post-hoc projectionimage classifiers (5 backbones) · CelebA and FairFaceaccuracy: Swin-T CelebA 0.888 -> 0.918Improved
fairness layer instead of post-hoc projection on small CNNCustomCNN · CelebAaccuracy: 0.809 -> 0.814No clear effect
fairness layer instead of Lagrangian penalty on imagesimage classifiers · CelebA and FairFaceaccuracy: 0.501 -> 0.918 (Swin-T CelebA)Improved
fairness layer instead of baselines on synthetic regressionsynthetic regression, 32 datasets · Synthetic datatest loss: ranked first in 22-27 of 32 datasetsImproved
fairness layer with tight bounds in large-batch regimesynthetic regression · Synthetic data, tight constraints (l=0, u=3.5)test loss: underperformed on 6-8 datasetsHurt
use Lagrangian penalty with validation-selected lambdaPenalty baseline · Synthetic data (btrain=binfer=2000)fairness constraint satisfaction: 11 violations; 7 had gaps 7 times larger than targetHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy