DualGraph: A graph-based method for reasoning about label noise

By Haiyang Zhang, Ximing Xing, Liang Liu. 2021

45 citations recorded by the source corpus.

Unreliable labels derived from large-scale dataset prevent neural networks from fully exploring the data. Existing methods of learning with noisy labels primarily take noise-cleaning-based and sample-selection-based methods. However, for numerous studies on account of the above two views, selected samples cannot take full advantage of all data points and cannot represent actual distribution of categories, in particular if label annotation is corrupted. In this paper, we start from a different perspective and propose a robust learning algorithm called DualGraph, which aims to capture structural relations among labels at two different levels with graph neural networks including instance-level and distribution-level relations. Specifically, the instance-level relation utilizes instance similarity characterize sample category, while the distribution-level relation describes instance similarity distribution from each sample to all other samples. Since the distribution-level relation is robust to label noise, our network propagates it as supervised signals to refine instance-level similarity. Combining two level relations, we design an end-to-end training paradigm to counteract noisy labels while generating reliable predictions. We conduct extensive experiments on the noisy CIFAR-10 dataset, CIFAR-100 dataset, and the Clothing1M dataset. The results demonstrate the advantageous performance of the proposed method in comparison to state-of-the-art baselines.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Continual Learning on Noisy Data Streams via Self-Purified Replay (2021, background): “Some works use graphs for offline learning from a large-scale noisy dataset [95, 93].”
Learning From Noisy Labels With Deep Neural Networks: A Survey (2020, background): “Dual-Graph [118] employs graph neural networks and reweights the examples according to the structural relations among labels, eliminating the abnormal noise examples.”