Label Noise and Machine Learning: Why Ignorance is Bliss

Watch a 42-minute lecture from the Simons Institute where Clayton Scott from the University of Michigan presents groundbreaking research on multi-class, instance-dependent label noise in machine learning. Explore the innovative concept of relative signal strength (RSS) as a point-wise measure of noisiness, and discover how it establishes matching upper and lower bounds for excess risk. Learn about the surprising effectiveness of Noise Ignorant Empirical Risk Minimization (NI-ERM), which achieves optimal results by treating data as if no label noise exists. See how these theoretical findings translate into practical applications, demonstrated through state-of-the-art performance on the CIFAR-N data challenge using a linear classifier with self-supervised feature extraction.