No abstract

Imperfect labels are ubiquitous in real-world datasets. Several recent successful methods for training deep neural networks (DNNs) robust to label noise have used two primary techniques: filtering samples based on loss during a warm-up phase to curate an initial set of cleanly labeled samples, and using the output of a network as a pseudo-label for subsequent loss calculations. In this paper, we evaluate different augmentation strategies for algorithms tackling the "learning with noisy labels" problem. We propose and examine multiple augmentation strategies and evaluate them using synthetic datasets based on CIFAR-10 and CIFAR-100, as well as on the real-world dataset Clothing1M. Due to several commonalities in these algorithms, we find that using one set of augmentations for loss modeling tasks and another set for learning is the most effective, improving results on the state-of-the-art and other previous methods. Furthermore, we find that applying augmentation during the warm-up period can negatively impact the loss convergence behavior of correctly versus incorrectly labeled samples. We introduce this augmentation strategy to the state-of-the-art technique and demonstrate that we can improve performance across all evaluated noise levels. In particular, we improve accuracy on the CIFAR-10 benchmark at 90% symmetric noise by more than 15% in absolute accuracy, and we also improve performance on the Clothing1M dataset.

Modern machine learning algorithms typically require large amounts of labeled training data to fit a reliable model. To minimize the cost of data collection, researchers often employ techniques such as crowdsourcing and web scraping. However, web data and human annotations are known to exhibit high margins of error, resulting in sizable amounts of incorrect labels. Poorly labeled training data can cause models to overfit to the noise distribution, crippling performance in real-world applications. In this work, we investigate the viability of using data augmentation in conjunction with semi-supervised learning to improve the label noise robustness of image classification models. We conduct several experiments using noisy variants of the CIFAR-10 image classification dataset to benchmark our method against existing algorithms. Experimental results show that our augmentative SSL approach improves upon the state-of-the-art.