Limited Gradient Descent: Learning With Noisy Labels
share
Label noise may handicap the generalization of classifiers, and it is an important issue how to effectively learn main pattern from samples with noisy labels. Recent studies have witnessed that deep neural networks tend to prioritize learning simple patterns and then memorize noise patterns. This suggests a method to search the best generalization, which learns the main pattern until the noise begins to be memorized. A natural idea is to use a supervised approach to find the stop timing of learning, for example resorting clean verification set. In practice, however, a clean verification set is sometimes not easy to obtain. To solve this problem, we propose an unsupervised method called limited gradient descent to estimate the best stop timing. We modified the labels of few samples in noisy dataset to be almost false labels as reverse pattern. By monitoring the learning progresses of the noisy samples and the reverse samples, we can determine the stop timing of learning. In this paper, we also provide some sufficient conditions on learning with noisy labels. Experimental results on CIFAR-10 demonstrate that our approach has similar generalization performance to those supervised methods. For uncomplicated datasets, such as MNIST, we add relabeling strategy to further improve generalization and achieve state-of-the-art performance.
READ FULL TEXT
