SES: Noisy Label Correction via Semantic Embedding Similarity

In the field of noisy label learning, errors in labels often lead to uncertainty in features and distortion in the embedding space distribution, resulting in inconsistencies between feature and semantic space, which significantly limit the performance. To address this limitation, we propose a Semantically Embedded Similarity (SES)-based approach. Firstly, SES incorporates prototype-based classifiers alongside traditional linear classifiers to capture class semantics features effectively. Secondly, SES employs a consistent regularization strategy to ensure consistent output distribution of reliable data based on prototypes and linear classifiers. This process aids in learning a certain feature space and aligning it with the semantic space. Finally, SES proposes a mutual constraint strategy to enhance the stability of two classifiers in uncertain spaces by encouraging them to adjust to each other. This strategy leverages their differences to promote collaborative correction of noisy labels, thereby enhancing label accuracy post-correction. Extensive experimentation across multiple benchmark datasets demonstrates SES’s state-of-the-art performance. Notably, our methods both achieved impressive top-1 results under asymmetric noise conditions, significantly outperforming other methods. In addition, SES exhibits promise in real-world noise-label datasets.