Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair
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 an intrinsic feature enhancement method for debiasing image classifiers that uses a bias-contrastive pair (BA and BN samples) and a bias-negative score, with no bias labels, to spatially guide the model toward intrinsic features.
Problem
Image classifiers trained on biased data rely on spuriously correlated bias attributes rather than intrinsic class features; existing debiasing methods that emphasize bias-conflicting samples do not directly tell the model where to look for intrinsic features, and many require bias labels.
Approach
Builds a bias-contrastive pair from a bias-aligned input and an auxiliary bias-negative sample selected by a loss-tracking bias-negative score from a biased model; computes an intrinsic feature enhancement weight from common-feature similarity and relative Grad-CAM exploitation, and uses it to reweight the debiased model's intermediate features as spatial guidance.
Key results
- On Waterbirds at 5% bias severity, the method reaches 66.33% vs 64.32% for HEX and 63.48% for LfF+BE.
- On BFFHQ at 0.5% severity, accuracy is 71.68% vs 67.36% for LfF+BE.
- On BAR at 1.0% severity, accuracy is 75.14% vs 73.36% for LfF+BE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BC-focused xBN selection from DBN instead of sampling xBN from D (ablation) | Intrinsic feature enhancement method · Waterbirds 1% bias severity | accuracy: 62.79 ±1.21 -> 65.10 ±0.87 | Helped |
| Add BC-focused xBN selection from DBN instead of sampling xBN from D (ablation) | Intrinsic feature enhancement method · BFFHQ 1% bias severity | accuracy: 71.04 ±2.55 -> 77.08 ±2.05 | Helped |
| Add BC-focused xBN selection from DBN instead of sampling xBN from D (ablation) | Intrinsic feature enhancement method · BAR 1% bias severity | accuracy: 73.36 ±1.40 -> 74.62 ±1.07 | Helped |
| Use the BN score s(xBN) to reweight losses (ablation) | Intrinsic feature enhancement method · Waterbirds 1% bias severity | accuracy: 65.10 ±0.87 -> 65.22 ±0.95 | Helped |
| Use the BN score s(xBN) to reweight losses (ablation) | Intrinsic feature enhancement method · BFFHQ 1% bias severity | accuracy: 77.08 ±2.05 -> 77.56 ±1.24 | Helped |
| Use the BN score s(xBN) to reweight losses (ablation) | Intrinsic feature enhancement method · BAR 1% bias severity | accuracy: 74.62 ±1.07 -> 75.14 ±0.82 | Helped |
| Add Lguide (ablation) | Intrinsic feature enhancement method · Waterbirds 1% bias severity | accuracy: 63.81 ±1.24 -> 65.22 ±0.95 | Helped |
| Add Lguide (ablation) | Intrinsic feature enhancement method · BFFHQ 1% bias severity | accuracy: 76.92 ±1.03 -> 77.56 ±1.24 | Helped |
| Add Lguide (ablation) | Intrinsic feature enhancement method · BAR 1% bias severity | accuracy: 74.03 ±1.13 -> 75.14 ±0.82 | Helped |
| Add LBN (ablation) | Intrinsic feature enhancement method · Waterbirds 1% bias severity | accuracy: 62.10 ±3.35 -> 65.22 ±0.95 | Helped |
| Add LBN (ablation) | Intrinsic feature enhancement method · BFFHQ 1% bias severity | accuracy: 74.84 ±2.00 -> 77.56 ±1.24 | Helped |
| Add LBN (ablation) | Intrinsic feature enhancement method · BAR 1% bias severity | accuracy: 74.87 ±1.51 -> 75.14 ±0.82 | Helped |
Limitations
- Relies on hyperparameters and a biased-model warm-up (T1, T2 schedule) whose sensitivity is not analyzed in the main text.
- Evaluated only on three datasets with synthetic-style bias severity levels.
Experimental setup
Datasets: Waterbirds BFFHQ BAR
Benchmarks: Waterbirds BFFHQ BAR
Metrics: classification accuracy (class-averaged on Waterbirds) best accuracy on BC test samples (BFFHQ) accuracy on BAR test set
Baselines: Vanilla HEX LNL EnD ReBias LfF DisEnt LfF+BE DisEnt+BE
Builds on: LfF BiasEnsemble JTT Grad-CAM
Closest papers by meaning
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.86)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.86)
- Mitigating Biases in Blackbox Feature Extractors for Image Classification Tasks (NeurIPS 2024 · similarity 0.85)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.85)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.85)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.85)
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.