Embracing Noise- How Data Corruption Can Make Models Smarter
summary
Three controlled CIFAR-10 experiments test specific claims: adding Gaussian feature noise during training yields a small robustness gain (clean baseline ~57% vs noisy-features ~53% under perturbation); corrupting inputs trades a few clean points (58–59% down to ~54%) for much better high-severity performance (clean model falls to ~15% vs noisy-trained ~28%); training with pixel shuffle and random erasing boosts resilience on those corruptions (clean ~26% vs unstructured-trained ~35–36%). Conclusion: robustness improves for the corruption family actually used.
AI-written summary · about
chapters (11)
- 0:00Noise in machine learning
- 1:29Structured noise and Gaussian
- 4:40Speckle noise effects
- 6:30Unstructured noise and corruption
- 8:39Random erasing simulation
- 11:08Experiment setup
- 12:33Experiment one: feature noise
- 16:07Experiment two: noisy images
- 17:56Experiment three: unstructured augmentation
- 19:50Unstructured noise robustness
- 20:14Q&A
related talks
01h 09mLearning Generative Models from Corrupted Data
12mDistribution-Free Robustness
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
59mStudying Failure Modes of Deep Learning and Ways To Fix Them
31mRealizing Robustness
42mLabel Noise: Ignorance is Bliss
49mA New Perspective on Adversarial Perturbations
50mTraining on the Test Set and Other Heresies