Research map / Label-efficient and robust learning / Noisy labels
A law of adversarial risk, interpolation, and label noise
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 |
|---|---|---|---|
| Interpolate uniformly noisy labels | Any classifier on a compact input subset C · Compact input subset C | adversarial risk: lower bound: \frac14 \mu(\mathcal C) | Helped |
| Interpolate uniformly noisy labels | Any classifier on a compactly supported distribution · Support of μ | adversarial risk: lower bound: \frac14 | Helped |
| Interpolate up to ⌊1.01^d⌋ noisy samples with a classifier that otherwise uses the ground-truth threshold | Constructed classifier on the sphere · Uniform distribution on sphere S^(d−1) | adversarial risk: o_d(1) | No clear effect |
| Use uniform label noise instead of noise confined to the long tail | Interpolators on the long-tailed interval mixture · Long-tailed mixture of A+B intervals | adversarial risk: O_{\rho} \left( \frac{ A}{B} \right) -> \Omega_{\rho} \left( 1 \right) | Hurt |
| Use the T-shaped hypothesis class rather than the threshold class | Interpolating classifiers on the constructed distribution · Distribution on [0,W] × {0} | adversarial risk: \mathcal{O}\br{\rho} -> \Omega\br{1} | Hurt |
| Train on uniform label noise rather than human-generated label noise at the same reported noise rate | ResNet34 · CIFAR-10/100-n | adversarial error | Hurt |
| Train on human-generated labels rather than clean labels | ResNet34 · CIFAR-10/100-n | adversarial error | No clear effect |
| Increase the uniform-noise sample size and adversarial radius relative to poisoning | Uniform-noise adversary versus Poisoner · General distribution with Ω(1) poisoning risk | adversarial risk: \Runif_{2\rho, T, \eta} \geq \frac12 \Rpoison_{\rho , N} | Helped |
Closest papers by meaning
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.88)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.88)
- Information-theoretic Limits of Online Classification with Noisy Labels (NeurIPS 2024 · similarity 0.87)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.87)
- Regretful Decisions under Label Noise (ICLR 2025 · similarity 0.86)
- Understanding Domain Generalization: A Noise Robustness Perspective (ICLR 2024 · 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.