Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Debiased Representation via Disentangled Feature Augmentation
Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, Jaegul Choo
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 disentangled feature swapping augmentation (ablation) | Debiased classifier · Colored MNIST, 1.0% bias-conflicting | accuracy (%): 74.03±2.40 -> 72.29±3.82 | Improved |
| Add disentangled feature swapping augmentation (ablation) | Debiased classifier · Corrupted CIFAR-10, 1.0% bias-conflicting | accuracy (%): 27.73±1.02 -> 32.81±2.47 | Improved |
| Add disentangled feature swapping augmentation (ablation) | Debiased classifier · BFFHQ, 0.5% bias-conflicting | accuracy (%): 59.4±2.46 -> 61.27±3.26 | Improved |
| Schedule feature augmentation after 10K iterations instead of augmenting from the beginning (ablation) | Disentangled feature augmentation · Colored MNIST, 1.0% bias-conflicting | accuracy (%): 72.29±3.82 -> 81.73±2.34 | Improved |
| Schedule feature augmentation after 10K iterations instead of augmenting from the beginning (ablation) | Disentangled feature augmentation · Corrupted CIFAR-10, 1.0% bias-conflicting | accuracy (%): 32.81±2.47 -> 52.31±1.00 | Improved |
| Schedule feature augmentation after 10K iterations instead of augmenting from the beginning (ablation) | Disentangled feature augmentation · BFFHQ, 0.5% bias-conflicting | accuracy (%): 61.27±3.26 -> 63.87±0.31 | Improved |
| Increase bias-conflicting sample diversity from 1% to 5%, sampling 50% per minibatch | Training-data sampling · Colored MNIST unbiased test | accuracy (%): 67.19±1.99 -> 83.77±2.03 | Improved |
| Use diverse samples from the 5% pool at 1% sampling instead of limited-diversity samples from the 1% pool at 50% sampling | Training-data sampling · Colored MNIST unbiased test | accuracy (%): 67.19±1.99 -> 77.97±6.00 | Improved |
| Increase bias-conflicting sample diversity from 1% to 5%, sampling 50% per minibatch | Training-data sampling · Corrupted CIFAR-10 unbiased test | accuracy (%): 33.08±0.80 -> 46.99±0.82 | Improved |
| Use diverse samples from the 5% pool at 1% sampling instead of limited-diversity samples from the 1% pool at 50% sampling | Training-data sampling · Corrupted CIFAR-10 unbiased test | accuracy (%): 33.08±0.80 -> 36.66±0.55 | Improved |
| Use feature augmentation from the beginning rather than scheduling it (ablation) | Disentangled feature augmentation · All three evaluation datasets | accuracy (%) | Hurt |
| Use the proposed method rather than the strongest baseline | Colored MNIST, 0.5%–5.0% bias-conflicting · Colored MNIST unbiased test | accuracy (%): 70.47±1.84 vs 65.22±4.41; 87.4±0.78 vs 81.73±2.34; 92.91±0.15 vs 84.79±0.95; 96.96±0.04 vs 89.66±1.09 | Mixed |
Closest papers by meaning
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.90)
- BiasAdv: Bias-Adversarial Augmentation for Model Debiasing (CVPR 2023 · similarity 0.90)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.89)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.88)
- Disentangled Representation Learning via Modular Compositional Bias (NeurIPS 2025 · similarity 0.87)
- Mitigating Biases in Blackbox Feature Extractors for Image Classification Tasks (NeurIPS 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.
