Research map / Architectures and efficiency / Pruning and efficient tuning
Debiasing surgeon: fantastic weights and how to find them
Remi Nahon, Ivan Luiz De Moura Matos, Van-Tam Nguyen, Enzo Tartaglione
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 |
|---|---|---|---|
| prune the vanilla-trained model with FFW without fine-tuning | SimpleConvNet · Biased MNIST rho=0.99 | task accuracy (%): 88.46 -> 97.79 | Improved |
| prune the vanilla-trained model with FFW without fine-tuning | SimpleConvNet · Biased MNIST rho=0.999 | task accuracy (%): 10.02 -> 22.29 | No clear effect |
| prune with FFW on Corrupted CIFAR10 | ResNet-18 · Corrupted CIFAR10 rho=0.995 | task accuracy (%): 20.50 -> 51.57 | Improved |
| prune with FFW on CelebA BlondHair | ResNet-18 · CelebA BlondHair | task accuracy (%): 80.42 -> 87.08 | Improved |
| prune with FFW on Multi-Color MNIST | 3-layer MLP · Multi-Color MNIST average | average task accuracy (%): 58.9 -> 36.37 | Hurt |
| prune with FFW on Multi-Color MNIST worst subgroup | 3-layer MLP · Multi-Color MNIST conflicting both sides | task accuracy (%): 7.68 -> 35.85 | Improved |
| increase regularization coefficient gamma to 50 or above (ablation) | FFW · Biased MNIST rho=0.99 | task accuracy (%): 97.79 -> 93.68 (gamma 50), 23.31 (gamma 200) | Hurt |
| use FFW instead of magnitude pruning at 90% sparsity (ablation) | SimpleConvNet · Biased MNIST rho=0.99 | task accuracy (%): 31.14 -> 97.79 | Improved |
| extract sub-network from vanilla model trained for more epochs (ablation) | FFW · Biased MNIST rho=0.99 | task accuracy (%): 12.42 -> 97.96 (1 -> 40 epochs) | Improved |
Closest papers by meaning
- Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models (CVPR 2026 · similarity 0.91)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.88)
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.85)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.85)
- Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.84)
- Simplicity Bias in Overparameterized Machine Learning (AAAI 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.
