Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning has a disparate impact on model accuracy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| apply fairness mitigation after pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 30% pruning | fairness violation: 0.179 -> 0.172 | Helped |
| apply fairness mitigation after pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 50% pruning | fairness violation: 0.186 -> 0.161 | Helped |
| apply fairness mitigation after pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 70% pruning | fairness violation: 0.152 -> 0.163 | Hurt |
| apply fairness mitigation after pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 90% pruning | fairness violation: 0.134 -> 0.05 | Helped |
| apply fairness mitigation before pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 90% pruning | fairness violation: 0.134 -> 0.238 | Hurt |
| apply fairness mitigation both before and after pruning instead of standard pruning | CNN · UTKFaces, age protected attribute, 50% pruning | fairness violation: 0.186 -> 0.144 | Helped |
| replace the full gradient-norm and Hessian constraints with group-loss proxy constraints (ablation) | Fair pruning mitigation · SVHN | overall accuracy: 0.857 -> 0.795 | Hurt |
| replace the full gradient-norm and Hessian constraints with group-loss proxy constraints (ablation) | Fair pruning mitigation · UTK-age bins | overall accuracy: 0.395 -> 0.390 | No clear effect |
| replace the full gradient-norm and Hessian constraints with group-loss proxy constraints (ablation) | Fair pruning mitigation · MNIST | overall accuracy: 0.993 -> 0.986 | Hurt |
| prune more parameters from the trained model | Standard pruning · All protected groups | excessive loss upper bound | Hurt |
| balance group gradient norms and Hessian eigenvalues during training | Pruned classifier · Multiple datasets and architectures | fairness violation | Helped |
Closest papers by meaning
- Balancing Act: Constraining Disparate Impact in Sparse Models (ICLR 2024 · similarity 0.91)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.86)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.86)
- DARE the Extreme: Revisiting Delta-Parameter Pruning For Fine-Tuned Models (ICLR 2025 · similarity 0.84)
- In defense of parameter sharing for model-compression (ICLR 2024 · similarity 0.84)
- Evaluating Model Bias Requires Characterizing its Mistakes (ICML 2024 · 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.