Research map / Architectures and efficiency / Pruning and efficient tuning
Bias in Pruned Vision Models: In-Depth Analysis and Countermeasures
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 |
|---|---|---|---|
| Prune jointly trained ResNet18 to 90% sparsity with gradual global magnitude pruning | ResNet18 trained on all 40 CelebA attributes · CelebA | accuracy: 90.4 -> 91.0 | No clear effect |
| Prune jointly trained ResNet18 to 99.5% sparsity with gradual global magnitude pruning | ResNet18 trained on all 40 CelebA attributes · CelebA | Male Bias Amplification: 0.0170 -> 0.0402 | Hurt |
| Prune jointly trained ResNet18 to 99.5% sparsity with gradual global magnitude pruning | ResNet18 trained on all 40 CelebA attributes · CelebA | uncertainty: 0.139 -> 0.276 | Hurt |
| Prune jointly trained ResNet18 to 99.5% sparsity with gradual global magnitude pruning | ResNet18 trained on all 40 CelebA attributes · CelebA | Expected Calibration Error: 0.0538 -> 0.0127 | Helped |
| Train attributes jointly rather than train each attribute singly (ablation) | ResNet18 CelebA attribute prediction · CelebA | Bias Amplification: 0.15 vs 0.04; 0.11 vs 0.03 | Helped |
| Train only the predicted attribute and its protected category together instead of jointly training all 40 attributes (ablation) | ResNet18 CelebA attribute prediction · CelebA | Bias Amplification | Mixed |
| Calibrate one prediction threshold per attribute on the validation set (ablation) | Jointly trained sparse ResNet18 · CelebA | Bias Amplification | Helped |
| Override the 5% most uncertain predictions with correct labels (ablation) | 99.5%-sparse jointly trained ResNet18 · CelebA | mean Bias Amplification: 23% | Helped |
| Override the 10% most uncertain predictions with correct labels (ablation) | 99.5%-sparse jointly trained ResNet18 · CelebA | mean Bias Amplification: 35% | Helped |
| Use uncropped rather than cropped CelebA images (ablation) | ResNet18 GMP-RI · CelebA / Wearing Necktie | Bias Amplification: 4.6 to 0.9 | Helped |
Closest papers by meaning
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.90)
- Balancing Act: Constraining Disparate Impact in Sparse Models (ICLR 2024 · similarity 0.90)
- Prospect Pruning: Finding Trainable Weights at Initialization using Meta-Gradients (ICLR 2022 · similarity 0.90)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.88)
- Entropy Induced Pruning Framework for Convolutional Neural Networks (AAAI 2024 · similarity 0.88)
- Adaptive Sharpness-Aware Pruning for Robust Sparse Networks (ICLR 2024 · similarity 0.88)
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.