Data Expansion Approach with Attention Mechanism for Learning with Noisy Labels

Abstract

In recent years, the development of deep learning has contributed to various areas of machine learning. However, deep learning requires a huge amount of data to train the model, and data collection techniques such as web crawling can easily generate incorrect labels. If a training dataset has noisy labels, the generalization performance of deep learning significantly decreases. Some recent works have successfully divided the dataset into samples with clean labels and ones with noisy labels. In light of these studies, we propose a novel data expansion framework to robustly train the models on noisy labels with the attention mechanisms. First, our method trains a deep learning model with the sample selection approach and saves the samples selected as clean at the end of training. The original noisy dataset is then extended with the selected samples and the model is trained on the dataset again. To prevent over-fitting and allow the model to learn different patterns of the selected samples, we leverage the attention mechanism of deep learning to modify the representation of the selected samples. We evaluated our method with synthetic noisy labels on CIFAR-10 and CUB-200-2011 and real-world dataset Clothing1M. Our method obtained comparable results to baseline CNNs and state-of-the-art methods.