Federated learning (FL) has been intensively investigated in terms of communication efficiency, privacy, and fairness. However, efficient annotation, which is a pain point in real-world FL applications, is less studied. In this project, we propose to apply active learning (AL) to the FL framework to reduce the annotation workload. We expect that the AL and FL can improve the performance of each other complementarily. In our proposed federated active learning (F-AL) method, the clients collaboratively execute the AL to obtain the instances which are considered informative to FL in a distributed optimization manner. We compare the test accuracies of the global FL models using the conventional random sampling strategy, client-level separate AL (S-AL), and the proposed F-AL. We empirically demonstrate that the F-AL outperforms baseline methods in image classification tasks.

Federated learning is a distributed machine learning framework that enables a large number of devices to cooperatively train a model without data sharing. However, because federated learning trains a model using non-independent and identically distributed (non-IID) data stored at local devices, the weight divergence causes a performance loss. This paper focuses on solving the non-IID problems and proposes Kalman filter-based clustering federated learning method called K-FL to get performance gain by providing a specific model with low variance to the device. To the best of our knowledge, it is the first clustering federated learning method that can train a model requiring fewer communication rounds under the premise that non-IID environment without any prior knowledge and an initial value set by the user. From simulations, we demonstrate that the proposed K-FL can train a model much faster, requiring fewer communication rounds than FedAvg and LG-FedAvg when testing neural networks using the MNIST, FMNIST, and CIFAR-10 datasets. As a numerical result, it is shown that the accuracy is improved in all datasets while the computational time cost is reduced by 1.43×, 1.67×, and 1.63× compared to FedAvg, respectively.

In the semiconductor industry, achieving a high production yield is a very important issue. Wafer bin maps (WBMs) provide critical information for identifying anomalies in the manufacturing process. A WBM forms a certain defect pattern according to the error occurring during the process, and by accurately classifying the defect pattern existing in the WBM, the root causes of the anomalies that have occurred during the process can be inferred. Therefore, WBM defect pattern recognition and classification tasks are important for improving yield. In this paper, we propose a deep convolutional generative adversarial network (DCGAN)-based data augmentation method to improve the accuracy of a convolutional neural network (CNN)-based defect pattern classifier in the presence of extremely imbalanced data. The proposed method forms various defect patterns compared to the data augmentation method by using a convolutional autoencoder (CAE), and the formed defect patterns are classified into the same pattern as the original pattern through a CNN-based defect pattern classifier. Here, we introduce a new quantitative index called PGI to compare the effectiveness of the augmented models, and propose a masking process to refine the augmented images. The proposed method was tested using the WM-811k dataset. The proposed method helps to improve the classification performance of the pattern classifier by effectively solving the data imbalance issue compared to the CAE-based augmentation method. The experimental results showed that the proposed method improved the accuracy of each defect pattern by about 5.31% on average compared to the CAE-based augmentation method.

Due to recent technological developments such as online navigation, augmented reality (AR), virtual reality (VR), and digital twins, and the high demand from users for various location-based services (LBS), research on location estimation techniques is being actively conducted. As a result, there is an increasing demand for effective localization technologies that can be used in places where the use of Global Positioning System (GPS) is limited, especially in indoor spaces with very large areas. In this paper, a new structure for an indoor localization system in which wireless fingerprinting and visual-based positioning are hierarchically combined—the so-called Fi-Vi system—is proposed. This scheme consists of two steps: fingerprint-based localization (FBL) followed by visual-based localization (VBL). In the first positioning step (i.e., the FBL stage), the entire area of a significantly broad range for localization is divided into multiple regions, the size and the number of which depend on the target accuracy of this step. Moreover, a machine-learning (ML) or deep-learning (DL) model trained on a Wi-Fi fingerprint radio map selects suitable candidate regions among these multiple regions. In the second positioning step (i.e., the VBL stage), the final location is precisely estimated through visual-based positioning based on the received information regarding the candidate regions. The FBL stage uses a sparse radio map (SRM) for fingerprinting, which can be constructed with relatively little effort and cost compared to radio maps used in conventional fingerprinting methods. As a result, it can be easily combined with existing visual-based positioning methods with almost negligible implementation complexity. Because of the hierarchical structure and SRM, the proposed scheme shows a significant performance improvement in terms of computational load and time required for indoor localization compared to the use of the existing visual-based indoor positioning method alone. In addition, it provides high accuracy and robustness even in a dynamically changing indoor wireless environment where conventional wireless fingerprinting methods show significant performance degradation. Finally, the performance analysis of the proposed scheme was performed using the UJIIndoorLoc dataset. Experiments and theoretical analysis have shown that when the estimation accuracy of the candidate region for the test dataset was achieved at about 99% through the FBL stage, the average computational amount of the VBL stage for the final position estimation was only about 16% of that in cases where the visual-based positioning method was used alone.