Manifold DivideMix: A Semi-Supervised Contrastive Learning Framework for Severe Label Noise

Deep neural networks have proven to be highly effective when large amounts of data with clean labels are available. However, their performance degrades when training data contains noisy labels, leading to poor generalization on the test set. Real-world datasets contain noisy label samples that either have similar visual semantics to other classes (in-distribution) or have no semantic relevance to any class (out-of-distribution) in the dataset. Most state-of-the-art methods leverage ID labeled noisy samples as unlabeled data for semi-supervised learning, but OOD labeled noisy samples cannot be used in this way because they do not belong to any class within the dataset. Hence, in this paper, we propose incorporating the information from all the training data by leveraging the benefits of self-supervised training. Our method aims to extract a meaningful and generalizable embedding space for each sample regardless of its label. Then, we employ a simple yet effective K-nearest neighbor method to remove portions of out-of-distribution samples. By discarding these samples, we propose an iterative "Manifold DivideMix" algorithm to find clean and noisy samples, and train our model in a semi-supervised way. In addition, we propose "MixEMatch", a new algorithm for the semi-supervised step that involves mixup augmentation at the input and final hidden representations of the model. This will extract better representations by interpolating both in the input and manifold spaces. Extensive experiments on multiple synthetic-noise image benchmarks and real-world web-crawled datasets demonstrate the effectiveness of our proposed framework. Code is available at https://github.com/Fahim-F/ManifoldDivideMix.