Best Paper Award 2021
Two papers of the Graduate School received during the 3^rd Annual Meeting the Best Paper Award:
P6: Yiwen Liao, Raphaël Latty, Bin Yang, Feature Selection Using Batch-Wise Attenuation and Feature Mask Normalization, International Joint Conference on Neural Networks (IJCNN), 2021
Feature selection (also known as variables selection) aims to identify the most important features from a large amount of candidate features according to a given metric. This is a challenging task due to large search space and the lack of ground truth. This work proposes a novel Feature Mask (FM) module that can be jointly trained with a neural network in an efficient way. Specifically, this method enables better selection results than many existing approaches by introducing a novel batch-wise attenuation mechanism that targets well the feature selection requirement and considering the relations between features based on a novel feature normalization technique. As a result, this method outperforms other contemporary methods with a notable margin on benchmarking datasets. Moreover, the FM-method has been applied to post-silicon validation, where the most important variables should be identified with respect to the given figure-of-merit values. In addition to vanilla feature/variable selection use cases, this method is generic and thus has been applied to many other use cases including but not limited to enhancing autoencoder-based anomaly detection performance and measurements reduction for open circuit fault detection.
P10: Paul R. Genssler and Hussam Amrouch. “Brain-Inspired Computing for Wafer Map Defect Pattern Classification”. In: IEEE International Test Conference (ITC’21). 2021.
In semiconductor fabrication, a wafer hosts many chips. Wafer-level tests can identify defective chips fast and at a low cost. Some defects are expected, but larger patterns hint at a systematic problem in the fabrication process. Hence, detecting such patterns immediately is essential. An automatic detection enables the engineers to identify the root cause and fix the underlying problem to increase the yield. In our work, we detect such defect patterns by applying brain-inspired hyperdimensional computing, a rapidly emerging machine-learning approach. Compared to other methods like SVM or CNN, less training data is required. For binary classification, a single expert-provided sample achieves a high accuracy. Further, the pattern detection is 46 times faster than a state-of-the-art CNN while achieving competitive classification accuracies on a large data set. Our work is the first to introduce brain-inspired hyperdimensional computing to the area of semiconductor test and reliability.