Few-shot classification of wafer bin maps using transfer learning and ensemble learning

Abstract

The emergence of new defect patterns in wafer bin maps (WBMs) has increased the need for the classification of such defect patterns with a limited number of WBMs, namely few-shot WBM classification, owing to the high cost of collection and annotation. Existing few-shot WBM classification algorithms mainly utilize meta-learning methods that leverage knowledge learned in several episodes. Meta-learning methods require a large amount of additional real WBMs, which can be unrealistic. To help train a network with a few real WBMs while avoiding this challenge, we propose to use simulated WBMs. In this process, we employ transfer learning, which is known to be simple and well-performing in few-shot learning. We further implement ensemble learning to solve the overfitting problem in transfer learning. Specifically, we apply transfer learning by pre-training a network with simulated WBMs and then fine-tuning it with real WBMs. We then apply ensemble learning by fine-tuning multiple sets of classification layers. A series of experiments on a real dataset demonstrate that our model compares favorably to widely used meta-learning methods in few-shot WBM classification. Additionally, we empirically verify that transfer learning and ensemble learning, the two most important yet simple components of our model, reduce the bias and variance of the learning algorithm under a few-shot scenario without a significant rise in training time.