Label Noise: Ignorance is Bliss

on YouTube ↗42m

" summary

Defines relative signal strength as a pointwise measure of transferability between noisy and clean posteriors and proves matching upper and lower excess-risk bounds under that RSS model, showing an unavoidable epsilon-sized error term where transfer fails. Shows the simple Noise Ignorant Empirical Risk Minimization (NI-ERM) — run ERM on noisy labels as if they were clean — is minimax optimal under these conditions. Practically, fitting a linear classifier on DINOv2 features yields state‑of‑the‑art results on CIFAR-N (despite human label error rates ≈6% on ImageNet).

AI-written summary · about

" chapters (11)

1. 0:00Introduction
2. 2:58Overview
3. 9:45Label Noise
4. 14:56Signal Strength
5. 16:14Relative Signal Strength
6. 18:49A Kappa
7. 23:26Multiclass
8. 24:05Example
9. 30:54Theory vs Practice
10. 32:50Experimental Results
11. 37:56Discussion

" related talks

• 35mTesting Noise Assumptions of Learning AlgorithmsArsen Vasilyan · Joint IFML/MPG Symposium 2024
• 23mEmbracing Noise- How Data Corruption Can Make Models SmarterAayush Gauba · PyData Boston 2025
• 59mStudying Failure Modes of Deep Learning and Ways To Fix ThemSoheil Feizi · Adversarial Approaches in Machine Learning 2022
• 14mGuarding the Privacy of Label-Only Access to Neural Network Classifiers via Formal(…)Anan Kabaha, Dana Drachsler Cohen · OOPSLA 2025
• 45mOrthogonal Statistical LearningVasilis Syrgkanis · Algorithmic Aspects of Causal Inference 2022
• 01h 07mLarge-Scale Private Learning, Part IIKunal Talwar · Data Privacy: Foundations and Applications Boot Camp
• 43mBeyond Worst Case: When Complex Feedback Can Improve the Label Complexity of Active LearningKamalika Chaudhuri · Learning, Algorithm Design and Beyond Worst-Case Analysis
• 59mTutorial: Statistical Learning Theory and Neural Networks ISpencer Frei · Deep Learning Theory Workshop and Summer School 2022