Go to ICLR 2024 Conference homepageLearning with Instance-Dependent Noisy Labels by Hard Sample Selection with Anchor Hallucination
Learning from noisily-labeled data is common in real-world visual learning tasks. Mainstream Noisy-Label Learning (NLL) methods mainly focus on sample-selection approaches, which typically divide the training dataset into clean and noisy subsets according to the loss distribution of samples. However, they overlook the fact that clean samples with complex visual patterns may also yield large losses, especially for datasets with Instance-Dependent Noise (IDN), in which the probability of an image being mislabeled depends on its visual appearance. This paper extends this idea and distinguishes complex samples from noisy ones. Specifically, we first select training samples with small initial losses to form an easy subset, where these easy samples are assumed to contain simple patterns with correct labels. The remaining samples either have complex patterns or incorrect labels, forming a hard subset. Subsequently, we utilize the easy subset to hallucinate multiple anchors, which are used to select hard samples to form a clean hard subset. We further exploit samples from these subsets following a semi-supervised training scheme to better characterize the decision boundary. Extensive experiments on synthetic and real-world instance-dependent noisy datasets show that our method outperforms the State-of-The-Art NLL methods.