Tae Joon Jun

Premature Ventricular Contraction Beat Detection with Deep Neural Networks

A deep neural networks is proposed for the classification of premature ventricular contraction (PVC) beat, which is an irregular heartbeat initiated by Purkinje fibers rather than by sinoatrial node. Several machine learning approaches were proposed for the detection of PVC beats although they resulted in either achieving low accuracy of classification or using limited portion of data from existing electrocardiography (ECG) databases. In this paper, we propose an optimized deep neural networks for PVC beat classification. Our method is evaluated on TensorFlow, which is an open source machine learning platform initially developed by Google. Our method achieved overall 99.41% accuracy and a sensitivity of 96.08% with total 80,836 ECG beats including normal and PVC from the MIT-BIH Arrhythmia Database.

GPU based cloud system for high-performance arrhythmia detection with parallel k-NN algorithm

In this paper, we propose an GPU based Cloud system for high-performance arrhythmia detection. Pan-Tompkins algorithm is used for QRS detection and we optimized beat classification algorithm with K-Nearest Neighbor (K-NN). To support high performance beat classification on the system, we parallelized beat classification algorithm with CUDA to execute the algorithm on virtualized GPU devices on the Cloud system. MIT-BIH Arrhythmia database is used for validation of the algorithm. The system achieved about 93.5% of detection rate which is comparable to previous researches while our algorithm shows 2.5 times faster execution time compared to CPU only detection algorithm.In this paper, we propose an GPU based Cloud system for high-performance arrhythmia detection. Pan-Tompkins algorithm is used for QRS detection and we optimized beat classification algorithm with K-Nearest Neighbor (K-NN). To support high performance beat classification on the system, we parallelized beat classification algorithm with CUDA to execute the algorithm on virtualized GPU devices on the Cloud system. MIT-BIH Arrhythmia database is used for validation of the algorithm. The system achieved about 93.5% of detection rate which is comparable to previous researches while our algorithm shows 2.5 times faster execution time compared to CPU only detection algorithm.

HPC Supported Mission-Critical Cloud Architecture

Tactical Operations Center (TOC) system in military field is an advanced computer system composed of multiple servers and desktops to interlock internal/external weapon systems processing mission-critical applications in combat situation. However, the current TOC system has several limitations such as difficulty of integrating tactical weapon systems including missile launch system and radar system into the single TOC system due to the heterogeneity of HW and SW between systems, and an inefficient computing resource management for the weapon systems. In this paper, we proposed a novel HPC supported mission-critical Cloud architecture as TOC for Surface-to-Air-Missile (SAM) system with OpenStack Cloud OS, Data Distribution Service (DDS), and GPU virtualization techniques. With this approach, our system provides elastic resource management over the weapon systems with virtual machines, integration of heterogeneous systems with different kinds of guest OS, real-time, reliable, and high-speed communication between the virtual machines and virtualized GPU resource over the virtual machines. Evaluation of our TOC system includes DDS performance measurement over 10Gbps Ethernet and QDR InfiniBand networks on the virtualized environment with OpenStack Cloud OS, and GPU virtualization performance evaluation with two different methods, PCI pass-through and remote-API. With the evaluation results, we conclude that our system provides reasonable performance in the combat situation compared to the previous TOC system while additionally supports scalable and elastic use of computing resource through the virtual machines.

Thin-Cap Fibroatheroma Detection with Deep Neural Networks

Acute coronary syndromes (ACS) frequently results in unstable angina, acute myocardial infarction, and sudden coronary death. The most of ACS are related to coronary thrombosis that mainly caused by plaque rupture followed by plaque erosion. Thin-cap fibroatheroma (TCFA) is a well-known type of vulnerable plaque which is prone to serious plaque rupture. Intravascular ultrasound (IVUS) is the most common methods for imaging coronary arteries to determine the amount of plaque built up at the epicardial coronary artery. However, since IVUS has relatively lower resolution than that of optical coherence tomography (OCT), TCFA detection with IVUS is considerably difficult. In this paper, we propose a novel method of TCFA detection with IVUS images using machine learning technique. 12,325 IVUS images from 100 different patients were labeled with equivalent frames from OCT images. Deep feed-forward neural network (FFNN) was applied to a different number of selected features based on the Fishers exact test. As a result, IVUS derived TCFA detection achieved 0.87 area under the curve (AUC) with 78.31% specificity and 79.02% sensitivity. Our experimental result indicates a new possibility for detection of TCFA with IVUS images using machine learning technique.

ConVGPU: GPU Management Middleware in Container Based Virtualized Environment

Nowadays, Graphics Processing Unit (GPU) is essential for general-purpose high-performance computing, because of its dominant performance in parallel computing compare to that of CPU. There have been many successful trials on the use of GPU in virtualized environment. Especially, NVIDIA Docker obtained a most practical way to bring GPU into the container-based virtualized environment. However, most of these trials did not consider sharing GPU among multiple containers. Without the above consideration, a system will experience a program failure or a deadlock situation in the worst case. In this paper, we propose ConVGPU, a solution to share the GPU in multiple containers. With ConVGPU, the system can guarantee the required GPU memory which the container needs to execute. To achieve it, we introduce four scheduling algorithms that manage the GPU memory to be taken by the containers. These algorithms can prevent the system from falling into deadlock situations between containers during execution.

A Workload Prediction Approach Using Models Stacking Based on Recurrent Neural Network and Autoencoder

Workload prediction in computing systems like Cloud and Grid is an essential prerequisite for successful load balancing and achieving service-level agreements. However, since workloads in different systems and architectures have varied characteristics, providing an accurate single prediction model can be very challenging. Therefore, in this paper we have designed and implemented a model of stacking prediction algorithms to predict workload time series in Cloud and Grid systems using Recurrent Neural Network and Autoencoder. We have also performed experiments with several datasets containing different workload types and conducted comparisons with each component algorithm as well as the fixed weighted optimal combination value. Experimental results show that our model achieves lower average NRMSE in 3 datasets than the fixed weighted optimal combination value, and outperforms the component algorithms with improvements in NRMSE from 7.43% to 12.45%.

Distributed self-organized cluster-based fusion tree generation algorithm

A fusion system, which collects air-tracks from distributed radars and eliminates duplicated tracks, is required to get a Single Integrated Air Picture (SIAP) for the wide surveillance area. We are developing a distributed radar system which consists of numerous mobile radars to cover wide surveillance area. Two-tier fusion system, which is the well-known solution for wide surveillance area, was adopted. Two-tier fusion system allows to put local fusion nodes between local radars and the central fusion node. We argue that the number of processed tracks of the two-tier fusion system is highly correlated with the fusion tree which decides the local fusion nodes and their child radars. We improved the number of processed tracks of our radar system by applying the fusion tree control which has been neglected in the most of research in this field. However it is hard to generate a proper fusion tree due to the possible number of fusion tree increases exponentially as the number of radars increase. To solve this problem, we propose a novel self-organized fusion tree generation algorithm. Especially, the proposed solution generates the fusion tree without any prior information, such as network-topology, and position of radars. We evaluate the performance of the proposed solution using the OPNET network simulator and show that the proposed solution performs better than the naive methods.