Machine Learning with Noisy Labels

Chapter 3: Theoretical aspects of noisy-label learning

Bias-variance decomposition, label transition distribution, and PAC learning

Abstract

This chapter introduces three approaches developed to improve our basic understanding of the noisy-label learning problem. We first explain the bias-variance decomposition that divides the training error into three components (i.e., bias, variance, and irreducible noise), which behave differently when symmetric, asymmetric, and instance-dependent noise affect the training set. Then, we explain the necessary and sufficient conditions for the identifiability of the transition matrix for noisy-label learning problems. We conclude the chapter with a brief overview of the use of Valiant's probably approximately ...