EGD: Error-Entropy-Guided Distillation for Noisy Multi-View Classification

In recent years, multi-view learning has received extensive research interest. Most existing multi-view learning methods often rely on well-annotated data to improve decision accuracy. However, noisy labels are ubiquitous in multi-view data due to imperfect annotations. Although some methods have achieved promising performance using robust-loss designs and implicit regularization, they fail to explicitly model the reliability of the supervision signal and fail to dynamically correct noisy labels during training. Clearly, this largely constrains their performance ceiling. To deal with this problem, we propose an Error-Entropy-Guided Distillation network (EGD) for noisy multi-view classification. In this framework, we first design an Error-Entropy (EE) metric to explicitly evaluate the reliability of sample-wise supervision, which serves as the basis for identifying and filtering noisy labels. On this basis, we adopt the distillation paradigm based on Error-Entropy (EE). The teacher model provides the student with soft label distributions that are less affected by noisy labels in the early training stage. To further mitigate noise memorization and accumulated confirmation bias, we propose a periodic memory-clearing strategy and supervision signal update strategy to prevent the teacher from error memorization and accumulating confirmation bias. Meanwhile, the student model learns from the soft supervision of the teacher to capture structured inter-class relationships. Additionally, a consistency module is employed to enhance the consistency of the student across multiple views. Extensive experiments on five benchmark datasets demonstrate that EGD consistently outperforms state-of-the-art multi-view learning methods under various noise levels.