Revisiting sample weights based method for noisy-label detection and classification

Tuan Hoang, Hung Tran, Santu Rana, Sunil Gupta, Svetha Venkatesh; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 4189-4204

The remarkable success of Convolutional Neural Networks (CNNs) in image classification can be attributed large clean training datasets. However, real-world data is often far from noise-free, impacting the performance of resulting deep neural network (DNN) models. Existing literature focuses on noisy label detection, often drawing a clear line between noisy and clean label samples. Nevertheless, each sample contributes differently to the final model performance; some noisy-label samples may still be valuable to a certain level, while certain clean-label samples might not significantly enhance the model. In this work, assuming that a small clean-label dataset may be available, we aim to learn a sample weight for each training sample. This weight is gradually updated as the model is training to indicate the usefulness of a particular sample in minimizing loss with respect to the clean-label dataset. Consequently, our method prioritizes high-quality data samples, minimizing the impact of harmful or unhelpful ones by assigning close-to-zero weights in a weighted loss function. We empirically demonstrate that our method is not dependent on noise type and can work well for both real-world and synthetic noise. Our method can achieve state-of-the-art performance in terms of the classification accuracy on clean test sets.