A Noise-Resilient Auto-Labeling Framework With Transition Matrix
W Lee, Youngbum Hur
IF 3.6
IEEE Access
Recently, auto-labeling framework has been applied in a lot of applications across various industries. Pseudo-labeling is the most common auto-labeling method and this method is to convert unlabeled data into labeled data by assigning pseudo-labels. Unless we have a perfect model for pseudo-labeling, the additional labeled data we get from unlabeled data always include noisy labels. However, this problem has not been studied by many researchers yet. Addressing this problem, we propose a noise-resilient auto-labeling framework using a transition matrix to mitigate the impact of label noise. The framework consists of three main stages: generating pseudo-labels for unlabeled data, identifying noisy samples based on KL-divergence between estimated transition vectors and model outputs, and using noisy samples as unlabeled data and clean samples as labeled data in semi-supervised learning for training the final model. We also show how much noise is added through pseudo-labeling depending on the initial model’s accuracy. Our experiments demonstrate the proposed method outperforms the state-of-the-art methods for handling noisy labels on both standard classification benchmarks (e.g., CIFAR-10 and CIFAR-100) and real-world datasets (e.g., Clothing100K, Food-101).
TripletMatch: Wafer Map Defect Detection Using Semi-Supervised Learning and Triplet Loss With Mixup
Cheol Il Lim, Youngbum Hur
IF 3.6
IEEE Access
In the semiconductor manufacturing process, Electrical Die Sorting (EDS) is a post-production process used to assess the quality of each chip on the wafer. The results from EDS testing are visualized as a wafer bin map (WBM), which is used for quality control purposes, such as the identification of defective wafers. Recently, deep learning has emerged as a prominent approach for identifying defects in wafers. However, data on defects in the semiconductor industry remain scarce. In this paper, we propose a semi-supervised learning method, TripletMatch, which utilizes triplet loss for unlabeled data. The proposed method extends the FixMatch framework and considers Mixup to smooth decision boundaries. Our experimental results demonstrate the superiority of TripletMatch over various recent deep-learning-based methods and loss functions.
Self-Supervised Anomaly Detection Using Outliers for Multivariate Time Series
James Won‐Ki Hong, Youngbum Hur
IF 3.6
IEEE Access
Due to the difficulty of having sufficient labeled data, self-supervised learning (SSL) has recently got much attention by many researchers in time series anomaly detection. The generative adversarial network (GAN) based autoencoder model, one of the SSL models, has good performance on anomaly detection but it tends to be too sensitive (i.e., predict normal data with a small anomalous value as abnormal). In this paper, we find that mispredicted normal data have values far from the average on some sensors. We call these data as outliers. Since these data are a few in the training set, the model struggles to reconstruct these data and incorrectly predicts them as abnormal. Based on these findings, we propose a robust self-supervised anomaly detection framework that finds outliers using a clustering based on correlation features and uses them for efficient training. To evaluate our method, we compare with various deep learning-based anomaly detection methods on the real-world pump dataset. The results demonstrate the superiority of our proposed method. Through our method, we maintain sensitivity to abnormal data while reducing sensitivity to normal data with a small anomalous value.
A Noise-Resilient Auto-Labeling Framework With Transition Matrix
W Lee, Youngbum Hur
IF 3.6
IEEE Access
Recently, auto-labeling framework has been applied in a lot of applications across various industries. Pseudo-labeling is the most common auto-labeling method and this method is to convert unlabeled data into labeled data by assigning pseudo-labels. Unless we have a perfect model for pseudo-labeling, the additional labeled data we get from unlabeled data always include noisy labels. However, this problem has not been studied by many researchers yet. Addressing this problem, we propose a noise-resilient auto-labeling framework using a transition matrix to mitigate the impact of label noise. The framework consists of three main stages: generating pseudo-labels for unlabeled data, identifying noisy samples based on KL-divergence between estimated transition vectors and model outputs, and using noisy samples as unlabeled data and clean samples as labeled data in semi-supervised learning for training the final model. We also show how much noise is added through pseudo-labeling depending on the initial model’s accuracy. Our experiments demonstrate the proposed method outperforms the state-of-the-art methods for handling noisy labels on both standard classification benchmarks (e.g., CIFAR-10 and CIFAR-100) and real-world datasets (e.g., Clothing100K, Food-101).
TripletMatch: Wafer Map Defect Detection Using Semi-Supervised Learning and Triplet Loss With Mixup
Cheol Il Lim, Youngbum Hur
IF 3.6
IEEE Access
In the semiconductor manufacturing process, Electrical Die Sorting (EDS) is a post-production process used to assess the quality of each chip on the wafer. The results from EDS testing are visualized as a wafer bin map (WBM), which is used for quality control purposes, such as the identification of defective wafers. Recently, deep learning has emerged as a prominent approach for identifying defects in wafers. However, data on defects in the semiconductor industry remain scarce. In this paper, we propose a semi-supervised learning method, TripletMatch, which utilizes triplet loss for unlabeled data. The proposed method extends the FixMatch framework and considers Mixup to smooth decision boundaries. Our experimental results demonstrate the superiority of TripletMatch over various recent deep-learning-based methods and loss functions.
Self-Supervised Anomaly Detection Using Outliers for Multivariate Time Series
James Won‐Ki Hong, Youngbum Hur
IF 3.6
IEEE Access
Due to the difficulty of having sufficient labeled data, self-supervised learning (SSL) has recently got much attention by many researchers in time series anomaly detection. The generative adversarial network (GAN) based autoencoder model, one of the SSL models, has good performance on anomaly detection but it tends to be too sensitive (i.e., predict normal data with a small anomalous value as abnormal). In this paper, we find that mispredicted normal data have values far from the average on some sensors. We call these data as outliers. Since these data are a few in the training set, the model struggles to reconstruct these data and incorrectly predicts them as abnormal. Based on these findings, we propose a robust self-supervised anomaly detection framework that finds outliers using a clustering based on correlation features and uses them for efficient training. To evaluate our method, we compare with various deep learning-based anomaly detection methods on the real-world pump dataset. The results demonstrate the superiority of our proposed method. Through our method, we maintain sensitivity to abnormal data while reducing sensitivity to normal data with a small anomalous value.
Deep Learning Based Steel Plate Surface Defect Detection with Precise RoI Pooling
Hyojin Jung, Wonhee Lee, Youngbum Hur
Journal of the Korean society for quality management
Purpose: The purpose of this study is to attempt to improve the steel industry by suggesting ways to improve the quality of steel products.Methods: An attempt was made to improve the detection of defect data on the surface of steel plates through data augmentation techniques and region-of-interest pooling techniques. The tools used in this study are categorized into three dimensions: object detection model, region of interest pooling, and Mixup data augmentation.Results: The results of this study are as follows We used Mixup and Precise RoI Pooling to improve the detection performance of steel surface. We studied the effect of Mixup on the detection of steel surface defects through experiments, and found that the detection performance of certain classes is degraded when Mixup is applied. To solve this problem, we found that it is effective to apply Precise RoI pooling to improve the detection performance of the model without applying Mixup, and then we integrated Precise RoI pooling and Mixup to improve the detection performance of steel surface defects for all classes. The proposed method was found to take 0.01 seconds per sheet, which is faster than visual inspection, which takes 6 seconds per sheet.Conclusion: Improved detection of steel plate surface defects
Out-of-Distribution Detection for Semiconductor Wafer Map Defect Using GAN and Outlier-Exposure
Minju Kim, Jaehyeop Hong, Youngbum Hur
Journal of the Korean society for quality management
Purpose: Out-of-Distribution(OOD) detection plays a crucial role in semiconductor manufacturing for identifying defects and ensuring quality control. However, collecting diverse real-world defect samples is expensive and challenging, limiting model performance. This study improves OOD detection by using Deep Convolutional Generative Adversarial Networks(DCGAN) to generate synthetic defect images, which are incorporated into the Outlier Exposure(OE) framework as an auxiliary dataset to enhance model generalization.Methods: Using the WM 811K wafer dataset, we generate synthetic defect images via DCGAN and integrate them into the OE framework. The proposed approach is compared with Maximum Softmax Probability(MSP) and standard OE model to evaluate its effectiveness. Performance is measured using FPR95, AUROC, and AUPR. Additionally, we investigate the impact of DCGAN training epochs on image quality and detection performance.Results: Experimental results indicate that OE with DCGAN significantly outperforms baseline method. Incorporating DCGAN-generated data reduces FPR95 while increasing AUROC and AUPR, demonstrating improved OOD detection capabilities. The study also highlights how optimizing DCGAN training enhances synthetic data quality and overall model performance.Conclusion: This study confirms that DCGAN-generated defect images effectively mitigate data scarcity issues in semiconductor OOD detection. Future research should explore alternative generative models (e.g., StyleGAN, VQ-VAE) and address class imbalance challenges to further enhance robustness and reliability.
Pixel-level Anomaly Detection System for Casting Process Using an Unsupervised Deep Learning Model
H. S. Lim, Jung W. Suh, Youngbum Hur
Journal of Korean Institute of Industrial Engineers
Since collecting defect data in industrial environments requires considerable time and cost, unsupervised learning-based anomaly detection algorithms are being developed to solve this problem. However, research on anomaly detection in the casting process of producing impellers has focused on supervised learning approaches.<br/>Therefore, in this paper, we propose an unsupervised deep learning model for anomaly detection in the casting process. The autoencoder used in Efficient AD is limited in detecting fine-grained defect patterns in impeller data due to its upsampling reconstruction method. So, we change to an autoencoder that utilizes transposed convolution layers, a learnable upsampling method, to improve the detection of fine defects. In addition, we provide pixel-precise ground truth regions of impeller anomalies to evaluate pixel-level localization performance of various unsupervised anomaly detection algorithms in future research. Experimental results, through the impeller dataset, demonstrate the superiority in detection accuracy, inference efficiency, and particularly pixel-level localization.
A Simple Framework for Robust Out-of-Distribution Detection
Youngbum Hur, Eunho Yang, Sung Ju Hwang
IF 3.6
IEEE Access
Out-of-distribution (OOD) detection, i.e., identifying whether a given test sample is drawn from outside the training distribution, is essential for a deep classifier to be deployed in a real-world application. The existing state-of-the-art methods of OOD detection tackle this issue by utilizing the internal feature of the classification network. However, we found that such detection methods inherently struggle to detect hard OOD images, i.e., drawn near from the training distribution: a naive softmax-based baseline even outperforms them. Motivated by this, we propose a simple yet effective training scheme for further calibrating the softmax probability of a classifier to achieve high OOD detection performance under both hard and easy scenarios. In particular, we suggest to optimize consistency regularization and self-supervised loss during training. Our experiments demonstrate the superiority of our simple method under various OOD detection scenarios.
Malaysian Name-based Ethnicity Classification using LSTM
Youngbum Hur
KSII Transactions on Internet and Information Systems
Name separation (splitting full names into surnames and given names) is not a tedious task in a multiethnic country because the procedure for splitting surnames and given names is ethnicity-specific.Malaysia has multiple main ethnic groups; therefore, separating Malaysian full names into surnames and given names proves a challenge.In this study, we develop a twophase framework for Malaysian name separation using deep learning.In the initial phase, we predict the ethnicity of full names.We propose a recurrent neural network with long shortterm memory network-based model with character embeddings for prediction.Based on the predicted ethnicity, we use a rule-based algorithm for splitting full names into surnames and given names in the second phase.We evaluate the performance of the proposed model against various machine learning models and demonstrate that it outperforms them by an average of 9%.Moreover, transfer learning and fine-tuning of the proposed model with an additional dataset results in an improvement of up to 7% on average.
Appointment scheduling at a multidisciplinary outpatient clinic using stochastic programming
Youngbum Hur, Jonathan F. Bard, Douglas J. Morrice
IF 2.1
Naval Research Logistics (NRL)
Abstract The purpose of this paper is to investigate the problem of constructing an appointment template for scheduling patients at a specific type of multidisciplinary outpatient clinic called an integrated practice unit (IPU). The focus is on developing and solving a stochastic optimization model for a back pain IPU in the face of random arrivals, an uncertain patient mix, and variable service times. The deterministic version of the problem is modeled as a mixed integer program with the objective of minimizing a weighted combination of clinic closing time (duration) and total patient waiting time (length of stay). A two‐stage stochastic program is then derived to account for the randomness and the sequential nature of the decisions. Although it was not possible to solve the two‐stage problem for even a limited number of scenarios, the wait‐and‐see (WS) problem was sufficiently tractable to provide a lower bound on the stochastic solution. The introduction of valid inequalities, limiting indices, and the use of special ordered sets helped to speed up the computations. A greedy heuristic was also developed to obtain solutions much more quickly. Out of practical considerations, it was necessary to develop appointment templates with time slots at fixed intervals, which are not available from the WS solution. The first to be derived was the expected value (EV) template that is used to find the expected value of the EV solution (EEV). This solution provides an upper bound on the objective function value of the two‐stage stochastic program. The average gap between the EEV and WS solutions was 18%. Results from extensive computational testing are presented for the EV template and for our adaptation of three other templates found in the literature. Depending on the relative importance of the two objective function metrics, the results demonstrate the trade‐off that exists between them. For the templates investigated, the “closing time” ranged from an average of 235 to 275 minutes for a 300‐minute session, while the corresponding “total patient time in clinic” ranged from 80 to 71 minutes.
