Research map / Learning theory and algorithms / Fairness
KINDER: Kernel-based Independence for Fair Representation Learning via Prototype-space Erasure
Abtin Mogharabin, Jiaee Cheong, Alp Toykan Kaplan, Sinan Kalkan
Read the paperCheck your idea against 137,217 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.
Proposes KINDER, a concept-erasure method that defines debiasing directly via kernel-based statistical independence in prototype space rather than through a specific adversary or probe.
Problem
Concept erasure for fairness typically defines debiasing indirectly through the failure of a chosen adversary, probe, or fairness penalty, tying the erasure target to a particular decoder family or optimization setup.
Approach
Defines erasure of sensitive attributes directly as statistical independence between representations and sensitive attributes, measured with kernels in a prototype space, instead of depending on an adversary's failure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply KINDER post-hoc to BERT embeddings | BERT · BIOS | worst-group macro-F1: 70.7 -> 73.9 | Improved |
| apply KINDER post-hoc to BERT embeddings | BERT · MOJI | accuracy: 80.1 -> 79.2 | Hurt |
| apply KINDER post-hoc to BERT embeddings | BERT · MOJI | worst-class TPR gap: 25.6 -> 16.7 | Improved |
| keep native end-to-end task head instead of replacing it with a linear probe (ablation) | TabM + KINDER · MIMIC-IV classification | accuracy: 80.6 -> 92.0 | Improved |
| apply KINDER end-to-end with native head | TabM · MIMIC-IV classification | F1: 39.9 -> 45.5 | Improved |
| apply INLP, RLACE or LEACE post-hoc to TabM | TabM · MIMIC-IV classification | accuracy: 91.9 -> 79.5 (INLP) | Hurt |
| apply KINDER to ResNet regressor | ResNet · MIMIC-IV regression | intersectional worst-group MAE: 5.55 -> 7.01 | Hurt |
| apply KINDER to ResNet regressor | ResNet · MIMIC-IV regression | MAE: 2.59 -> 2.52 | Improved |
| apply KINDER to UNet segmentation features | UNet · FairSeg intersectional | HD95 gap: 30.7 -> 5.4 | Improved |
| apply KINDER to FactorCL | FactorCL · FairVLMed | worst-group intersectional accuracy: 0.0 -> 41.5 | Improved |
| apply KINDER to CLIP | CLIP · FairVLMed | accuracy: 73.6 -> 71.6 | Hurt |
| apply KINDER to I-JEPA | I-JEPA · CelebA | accuracy: 91.1 -> 92.2 | Improved |
Closest papers by meaning
- Obliviator Reveals the Cost of Nonlinear Guardedness in Concept Erasure (NeurIPS 2025 · similarity 0.86)
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.85)
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.83)
- Constructing Fair Latent Space for Intersection of Fairness and Explainability (AAAI 2025 · similarity 0.82)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.82)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.82)
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.
