RectMamba: Exploring state space models with entropy-divergence framework for noisy label rectification

The efficacy of Deep Neural Networks (DNNs) is significantly influenced by the quality of labels. Noisy labels in real-world datasets can mislead DNNs into learning incorrect patterns. Existing “learning with noisy labels” (LNL) methods often struggle in high-noise environments, where the network tends to overfit noisy labels during the initial training period, leading to suboptimal performance. Additionally, they tend to discard noisy samples too aggressively, resulting in the loss of valuable information. Moreover, many of these methods exhibit a bias toward selecting simpler, easier-to-classify instances, which causes them to overlook more complex but informative samples, further limiting their effectiveness. To address these limitations, we introduce a robust label correction mechanism within CVMamba. This mechanism refines noisy labels using hierarchical feature representations from Visual State Space (VSS) blocks and the 2D-Selective-Scan (SS2D) module, thereby enhancing model performance and resilience against label noise.Additionally, we propose two innovative components: Entropy-Divergence K-Nearest Neighbors Relabel (EDKR) and Adaptivemix data augmentation. The Entropy-Divergence method, a parameter-free label correction strategy, leverages feature similarity and entropy-divergence uncertainty assessment to accurately correct mislabeled samples.Furthermore, the KNN method evaluates the proximity of samples in the feature space, assigns normalized weights to the nearest neighbors, and utilizes a score matrix with a set threshold to select clean samples. Adaptivemix, enhanced by a refined mixup strategy, adjusts the mixing coefficient based on the training stage to optimize the combination of sample pairs and their labels.This integrated approach has demonstrated substantial improvements in label accuracy, increasing from 20 % to an average of 96 % on the CIFAR-10 dataset with 80 % symmetric label noise.Moreover, our method achieves test accuracies of 82.1 % on CIFAR-100 with 40 % asymmetric label noise, surpassing the previous SOTA accuracy of 78.0 %. On ANIMAL-10N, we achieve an accuracy of 92.5 %, exceeding the previous best of 90.5 %. On Clothing1M, our method reaches 74.93 % accuracy, nearing the SOTA performance of 75.31 %. These results illustrate RectMamba's effectiveness in managing noisy datasets and highlight its potential in improving DNN robustness against label noise. Code is available at https://github.com/ningwei-wang/RectMamba