Small-Vote Sample Selection for Label-Noise Learning

The small-loss criterion is widely used in recent label-noise learning methods. However, such a only considers the loss of each training sample mini-batch but ignores distribution whole set. Moreover, selection clean samples depends on heuristic data rate. As result, some noisy-labeled are easily identified as ones, and vice versa. In this paper, we propose novel yet simple method, which mainly consists Hierarchical Voting Scheme (HVS) an Adaptive Clean rate Estimation Strategy (ACES), to accurately identify for robust learning. Specifically, HVS effectively combine global vote local vote, so that both epoch-level batch-level information exploited assign hierarchical sample. Based HVS, further develop ACES adaptively estimate by leveraging 1D Gaussian Mixture Model (GMM). Experimental results show our proposed method consistently outperforms several state-of-the-art methods synthetic real-world noisy benchmark datasets.