Research map / Label-efficient and robust learning / Recognition and long-tail
A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective
Read the paperCheck your idea against 108,967 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 Bias-Conditioned Self-Influence (BCSI), a self-influence score computed from a heavily biased early-stage model, to build a pivotal set of bias-conflicting samples used to fine-tune and rectify biased models without bias labels or unbiased sets.
Problem
Detecting bias-conflicting samples without bias labels or an unbiased validation set is imprecise, so debiasing methods that reweight or resample them can amplify bias when detection is wrong.
Approach
Reframe bias-conflicting detection as mislabeled-sample detection and use self-influence, computed on an early-epoch, GCE-trained heavily biased model (BCSI), to rank samples; intersect top-k sets from three random seeds to form a high-purity pivotal set, then fine-tune all parameters with CE on the pivotal set plus a counterweight CE on the remaining data.
Key results
- On CIFAR10C (5%), Ours+SelecMix reaches 59.66% accuracy vs 40.24% for ERM and 54.86% for SelecMix.
- Pivotal-set detection precision for CIFAR10C (5%) reaches 96.60% with three-run intersection, versus 76.27% with a single run.
- On Waterbirds, Ours+SelecMix achieves 89.67% worst-group accuracy vs 74.72% for SelecMix; on CivilComments worst-group accuracy rises 14.8 points over JTT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SI with BCSI for detecting bias-conflicting samples | self-influence detection · CMNIST 0.5% | detection precision: 31.94 -> 92.08 | Helped |
| replace SI with BCSI for detecting bias-conflicting samples | self-influence detection · CIFAR10C 0.5% | detection precision: 63.33 -> 38.33 | Hurt |
| replace SI with BCSI for detecting bias-conflicting samples | self-influence detection · CIFAR10C 2% | detection precision: 20.19 -> 50.45 | Helped |
| apply the method to a JTT-debiased model | JTT · MultiNLI | worst-group accuracy: 70.2 -> 73.6 | Helped |
| apply the method to a JTT-debiased model | JTT · CivilComments | worst-group accuracy: 63.7 -> 78.5 | Helped |
| use more intersections of the pivotal sets (ablation) | Ours+SelecMix · CIFAR10C 0.5% | accuracy: 36.44 -> 38.85 -> 38.74 | Mixed |
| use more intersections of the pivotal sets (ablation) | Ours+SelecMix · CIFAR10C 30% | accuracy: 67.04 -> 65.90 -> 64.51 | Hurt |
| increase the pivotal-set size k (ablation) | Ours · CIFAR10C 0.5% | accuracy | Hurt |
| increase the pivotal-set size k (ablation) | Ours · CIFAR10C 5% | accuracy | Helped |
| increase the remaining-data loss weight λ (ablation) | Ours · CIFAR10C 0.5% | accuracy | Hurt |
| increase the remaining-data loss weight λ (ablation) | Ours · CIFAR10C 90% | accuracy | Helped |
| use detection models trained for 3 to 11 epochs instead of 1 epoch (ablation) | Ours with SelecMix · CIFAR10C | accuracy | No clear effect |
Surprising findings
- Directly applying self-influence or training-set influence for bias-conflicting detection performs poorly (mostly below 25% precision), unlike in the mislabeled setting.
- Last-layer retraining was insufficient in highly biased scenarios because the feature extractor also becomes biased, so the method fine-tunes all parameters.
Limitations
- Remedy is a basic fine-tuning approach; more sophisticated sample weighting or curriculum learning not explored.
- Pivotal set purity is imperfect and drops at low bias-conflicting ratios, with fewer intersections needed when bias-conflicting samples are abundant.
- Performance on low-bias settings (e.g., CIFAR10C 90%) for ERM-based variant shows slight decrease unless top-k is enlarged.
Experimental setup
Datasets: Colored MNIST Corrupted CIFAR10 Biased FFHQ Waterbirds NICO CivilComments MultiNLI
Benchmarks: Colored MNIST Corrupted CIFAR10 Biased FFHQ Waterbirds NICO CivilComments MultiNLI
Metrics: unbiased test accuracy minority-group accuracy worst-group accuracy detection precision average precision demographic parity equalized odds
Baselines: ERM GroupDRO LfF BPA DCWP DFA SelecMix JTT
Builds on: influence functions Generalized Cross Entropy LfF SelecMix JTT
Closest papers by meaning
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.88)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.87)
- An Effective Theory of Bias Amplification (ICLR 2025 · similarity 0.87)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.87)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.87)
- Ameliorate Spurious Correlations in Dataset Condensation (ICML 2024 · similarity 0.87)
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.