Noisy label correction for classification of wafer bin maps with mixed-type defect patterns

Abstract

A wafer bin map (WBM) is a spatial map of binary values for individual chips on a wafer, representing the wafer testing results for each chip. Different patterns of defective chips in WBMs are related to different root causes of process failures, and thus the classification of defect patterns in WBMs helps to detect the process failures and identify their causes. In recent studies, convolutional neural networks (CNNs) have shown effective performance on the classification of the defect patterns in WBMs owing to their high expressive power. However, previous studies implicitly assumed that the labels of WBMs used for training the CNNs are correct, although the labels can be often incorrect due to annotation errors. The possibility of such mislabeling increases when WBMs have mixed-type defect patterns. However, when trained on mislabeled data, CNNs with the standard cross entropy loss can easily overfit to mislabeled samples, which leads to poor generalization on test data. To overcome this issue, we propose a novel training algorithm called \textit{sample bootstrapping}. Sample bootstrapping identifies which samples have clean or noisy labels using a two-component beta mixture model (BMM) and measures the uncertainty of each identified label (clean or noisy) using the posterior probability of the component assignment. Then, only the samples with low uncertainty of the estimated labels are selected to compose mini-batches via weighted random sampling, where the sampling weights are determined based on the posterior probability calculated from the BNN. Finally, the CNNs are trained on these mini-batches with the dynamic bootstrapping loss, a recent modification of the cross entropy loss to account for label noise. In this way, we can correct only the samples that are highly likely to have noisy labels, and prevent the risk of false correction of actually true-labeled samples. Our experiments on simulated WBM datasets under various noise levels and noise types demonstrate better test accuracy of sample bootstrapping than other competing methods.