Analysis of the development of modifying loss functions in label-noise environments

In recent years, the rapid advance of artificial intelligence has intensified interest in deep learning. However, during the training process of deep learning, the paper found that it is often susceptible to label noise. It biases models and severely degrades performance. And due to the complexity of the sources and classifications of label noise, it increases the difficulty of solving the problem of label noise. Current countermeasures cluster into three streams—supervised sample selection, robust architectures, and loss-function redesign. This review focuses exclusively on the last stream, surveying loss-function innovations proposed between 2015 and 2025. The paper identifies unresolved challenges in current research—such as over-reliance on idealized noise assumptions, dependence on clean validation sets, and so on. The paper also provides promising future directions, including self-calibrating losses without clean data, differentiable meta-loss layers for hyper-parameters, etc. By consolidating this evolving landscape, the survey offers practical guidance for mitigating label noise and charts a roadmap for future research that could sustain the long-term progress of deep learning.