Research map / Label-efficient and robust learning / Noisy labels
Understanding Instance-Level Label Noise: Disparate Impacts and Treatments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Memorize noisy labels for an instance | Over-parameterized model · Instance-level generalization | individual excessive generalization error: Ω( \frac{l^2}{n^2} \cdot \textsf{weight}(...) \cdot \sum_{k \neq y} \tilde{\mathbb{P}}[\tilde{y}=k|x]) | Hurt |
| Apply loss correction instead of memorizing noisy labels | Loss correction · High-frequency instances | individual excessive generalization error: w.p. at least 1-e^{-2l (1/2-e_{sgn(y)}(x))^2}; Ω( \frac{l^2}{n^2} \cdot \textsf{weight}(...) ) | Helped |
| Apply loss correction instead of memorizing noisy labels | Loss correction · Low-frequency instances | prediction error: w.p. at least \frac{1}{\sqrt{2l}} \cdot e^{-l \cdot D_{\text{KL}}\left(\frac{1}{2}\|e_{sgn(y)}(x)\right)} | Hurt |
| Use label smoothing instead of loss correction | Label smoothing · Instance-level binary classification | prediction error | Mixed |
| Apply peer loss instead of memorizing noisy labels | Peer loss · High-frequency instances | individual excessive generalization error: w.p. at least 1-e^{ \frac{-2l}{p^2_{sgn(-y)}\cdot \left(1-e_+-e_-\right)^2}}; Ω( \frac{l^2}{n^2} \textsf{weight}(...) \c | Helped |
| Apply peer loss instead of memorizing noisy labels | Peer loss · Low-frequency instances | prediction error: w.p. at least \frac{1}{\sqrt{2l}}e^{-l \cdot D_{\text{KL}}\left(\frac{1}{2}\| e_{sgn(y)}\right)} | Hurt |
Closest papers by meaning
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.90)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.89)
- Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources (ICML 2024 · similarity 0.89)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.89)
- Regretful Decisions under Label Noise (ICLR 2025 · similarity 0.88)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.88)
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.