A survey on learning from data with label noise via deep neural networks
Baoye Song, Shihao Zhao, Linjing Dang, Haoguang Wang, Lin Xu
Abstract
Deep neural networks (DNNs) have recently achieved remarkable breakthroughs across various domains, yet their performance is heavily reliant on the quality of labelled data. In practical applications, data labelling processes are frequently susceptible to noise, leading to inaccurate or inconsistent labels that impede model training and generalization. To tackle these challenges, the field of learning with label noise (LLN) has attracted significant attention. This survey offers a thorough review of LLN methods, focussing on the types, sources, and impacts of label noise. It systematically categorizes the state-of-the-art approaches for managing label noise, encompassing supervised learning, robust regression, loss adjustment, regularization techniques, and sample selection strategies. The survey also delves into representative noisy datasets used for benchmarking, such as CIFAR-10N, Clothing1M, and WebVision, underscoring their unique challenges and applications. Moreover, the review broadens its scope to examine the implications of noisy labels in specialized domains, including fault diagnosis, where label noise poses significant risks. In conclusion, the paper sheds light on emerging research directions, such as semi-supervised learning and multimodal approaches, with the goal of enhancing robust learning techniques in complex environments.