Research map / Label-efficient and robust learning / Recognition and long-tail
AracNet: Revealing Debiasing Signals across Layers with Shallow Monitors
Vito Paolo Pastore, Massimiliano Ciranni, Enzo Tartaglione, Vittorio Murino
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 |
|---|---|---|---|
| add the surprisal weight w_surprisal to the monitor-loss weight (ablation) | AracNet · Waterbirds (ResNet50) | worst-group accuracy: 88.78 -> 91.27 | Improved |
| add the monitor-loss weight w_l* to the surprisal weight (ablation) | AracNet · Waterbirds (ResNet50) | worst-group accuracy: 84.94 -> 91.27 | Improved |
| set surprisal exponent lambda to 2 rather than 1, 3, 4 or 5 (ablation) | AracNet · Waterbirds (ResNet50) | worst-group accuracy: 84.47 (lambda 5) -> 91.27 (lambda 2) | Improved |
| pretrain biased backbone for 200 epochs instead of 50 (ablation) | AracNet · Waterbirds (ResNet50) | worst-group accuracy: 91.27 -> 90.85 | No clear effect |
| use monitor from a ResNet18 backbone to debias ResNet50 or ViT-S/16 (ablation) | AracNet · Waterbirds | worst-group accuracy: 91.27 -> 91.06 (ResNet50); 90.68 (ViT-S/16) | No clear effect |
| apply AracNet to ViT-S/16 with a monitor per block | ViT-S/16 · Waterbirds | worst-group accuracy: 67.69 -> 90.66 | Improved |
| AracNet versus DeNetDM on Corrupted CIFAR-10 at highest bias | Corrupted CIFAR-10 (rho=0.995) · Corrupted CIFAR-10 (rho=0.995) | accuracy: 38.93 (DeNetDM) vs 33.23 | Hurt |
| AracNet versus best baselines at low bias on Corrupted CIFAR-10 | Corrupted CIFAR-10 (rho=0.950) · Corrupted CIFAR-10 (rho=0.950) | accuracy: 56.30 -> 56.80 | Improved |
| AracNet versus DDB on UrbanCars | UrbanCars · UrbanCars | I.D. accuracy: I.D. 86.39 (DDB) vs 85.95; BG-CoObj gap 0.12 vs 0.35 | Mixed |
| AracNet versus LoYS on BAR (rho=0.990) | BAR · BAR (rho=0.990) | accuracy: 75.92 -> 76.22 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeNetDM — DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024) · better on 3 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.88)
- NeuronTune: Towards Self-Guided Spurious Bias Mitigation (ICML 2025 · similarity 0.87)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.87)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.86)
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment (NeurIPS 2025 · similarity 0.86)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.86)
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.
