Vol. 48, No. 1,
Jan. 2021
Digital Library
LFA-SkipList: Optimizing SkipList by Reducing Access in a NUMA-Aware System
Sunghwan Ahn, Yujin Jang, Seungjun Ha, Beomseok Nam
http://doi.org/10.5626/JOK.2021.48.1.1
Intel"s Optane DC Persistent Memory is a non-volatile memory that works faster than storage devices and stores data persistently. However, in the NUMA system, it takes a longer latency to access the remote memory of another CPU socket than for local NUMA access. Therefore, performance is degraded when configuring the SkipList in multiple non-volatile memories. In this paper, an LFA-SkipList was proposed to solve this problem. The LFA-SkipList has a newly added local pointer and uses it to access the local node first and then the remote node, thereby reducing unnecessary remote node access and improving performance. The study found the LFA-SkipList demonstrated a much shorter search time than that of the legacy SkipList.
Improvement on Parallel Matrix Multiplication Routines in ScaLAPACK using Blocked Matrix Multiplication Algorithm on Intel KNL Clusters with AVX-512
Thi My Tuyen Nguyen, Yoosang Park, Jaeyoung Choi
http://doi.org/10.5626/JOK.2021.48.1.7
General matrix multiplication (GEMM) is a core computation algorithm in linear algebra, machine learning, statistics, and many other domains. Optimizations of such routines, including GEMM, have been conducted by vendors and researches with auto-tuning techniques. To achieve high performance for parallel matrix multiplication, a matrix multiplication processing scheme based on the optimization of local matrix multiplication at each node should be necessarily applied. In this paper, the application of parallel double-precision general matrix multiplication (PDGEMM) on Intel KNL was examined. The application of DGEMM calculated sub-matrices multiplication at each node. Details of the proposed DGEMM were introduced, including a blocked matrix multiplication algorithm with AVX-512 instruction sets and several optimization techniques, such as the data prefetching, loop unrolling, and cache blocking. This study found that the proposed PDGEMM performance was better than that in the ordinary cases of PDGEMM from the Intel Math Kernel Library (MKL) on both 4 and 16-node KNL clusters, with the flop rate improvements of 6% and 68%, respectively.
An Embedding Technique for Weighted Graphs using LSTM Autoencoders
http://doi.org/10.5626/JOK.2021.48.1.13
Graph embedding is the representation of graphs as vectors in a low-dimensional space. Recently, research on graph embedding using deep learning technology have been conducted. However, most research to date has focused mainly on the topology of nodes, and there are few studies on graph embedding for weighted graphs, which has an arbitrary weight on the edges between the nodes. Therefore, in this paper, we proposed a new graph embedding technique for weighted graphs. Given weighted graphs to be embedded, the proposed technique first extracts node-weight sequences that exist inside the graphs, and then encodes each node-weight sequence into a fixed-length vector using an LSTM (Long Short-Term Memory) autoencoder. Finally, for each graph, the proposed technique combines the encoding vectors of node-weight sequences extracted from the graph to generate one final embedding vector. The embedding vectors of the weighted graphs obtained by the proposed technique can be used for measuring the similarity between weighted graphs or classifying weighted graphs. Experiments on synthetic and real datasets consisting of groups of similar weighted graphs showed that the proposed technique provided more than 94% accuracy in finding similar weighted graphs.
Korean Dependency Parsing using Token-Level Contextual Representation in Pre-trained Language Model
http://doi.org/10.5626/JOK.2021.48.1.27
Dependency parsing is a problem of disambiguating sentence structure by recognizing dependencies and labels between words in sentences. In contrast to previous studies that have applied additional RNNs to the pre-trained language model, this paper proposes a dependency parsing method that uses fine-tuning alone to maximize the self-attention mechanism of the pre-trained language model, and also proposes a technique for using relative distance parameters and SEP tokens. In the results of evaluating the Sejong parsing corpus of TTA standard guidelines, the KorBERT_base model showed 95.73% UAS and 93.39% LAS while the KorBERT_large model showed 96.31% UAS and 94.17% LAS. This represents an improvement of about 3% compared to the results of previous studies that did not use the pre-trained language model. Next, the results of the word-morpheme mixed transformation corpus of the previous study showed that the KorBERT_base model was 94.19% UAS and that the KorBERT_large model was 94.76% UAS.
Improved Prediction for Configuration Bug Report Using Text Mining and Dimensionality Reduction
Jeongwhan Choi, Jiwon Choi, Duksan Ryu, Suntae Kim
http://doi.org/10.5626/JOK.2021.48.1.35
Configuration bugs are one of the main causes of software failure. Software organizations collect and manage bug reports using an issue tracking system. The bug assignor can spend excessive amounts of time identifying whether a bug is a configuration bug or not. Configuration bug prediction can help the bug assignor reduce classification efforts and aid decision making. In this paper, we propose an improved classification model using text mining and dimensionality reduction. This paper extracts 4,457 bug reports from six open-source software projects, trains a model to classify configuration bug reports, and evaluates prediction performance. The best performance method is obtained using the k-Nearest Neighbors model with the SMOTEENN sampling technique after extracting the feature with Bag of Words and then reducing the dimension of the feature using Linear Discriminant Analysis. The results show that ROC-AUC is 0.9812 and MCC is 0.942. This indicates better performance than Xia et al."s method and solves the class imbalance problem of our previous study. By predicting these enhanced configuration bug reports, our proposed approach can provide the bug assignors with information they need to make informed decisions.
A Fusion of CNN-based Frame Vector for Segment-level Video Partial Copy Detection
http://doi.org/10.5626/JOK.2021.48.1.43
Recently, the demand for media has grown rapidly, led by multimedia content platforms such as YouTube and Instagram. As a result, problems such as copyright protection and the spread of illegal content have arisen. To solve these problems, studies have been proposed to extract unique identifiers based on the content. However, existing studies were designed for simulated transformation and failed to detect whether the copied videos were actually shared. In this paper, we proposed a deep learning-based segment fingerprint that fused frame information for partial copy detection that was robust for various variations in the actually shared video. We used TIRI for data-level fusion and Pooling for feature-level fusion. We also designed a detection system with a segment fingerprint that was trained with Triplet loss. We evaluated the performance with VCDB, a dataset collected based on YouTube, and obtained 66% performance by fusing frame features sampled for 5 seconds with Max pooling for detecting video partial-copy problems.
Dual-Use Encoder-Decoder Model for License Plate Recognition
http://doi.org/10.5626/JOK.2021.48.1.51
Due to the rapid, continuous development of machine learning, the neural network model shows high performance in the field of license plate recognition. The most important factors in the performance of machine learning are the data and the model. Most license plate datasets are only given character sequence labels. In such cases, an encoder-decoder model is typically used to recognize character sequences. Detection-based models are better than encoder-decoder models, but they can only be used when character-bounding box labels are included in the dataset, which requires a high labeling cost. In this paper, we suggest an encoder-decoder model that can be used regardless of the presence or absence of character bounding box labels. It includes a combination of the Resnet [1] encoder and the Transformer [2] decoder. The proposed model not only achieves high recognition performance in the absence of character-bounding box labels, but also improves the performance by exploiting bounding box labels when they are available. In case of the Taiwan AOLP [3] dataset containing character sequence labels only, the proposed model shows 99.55% accuracy, which is higher than those of the conventional methods. Further, for the Korean KarPlate [4] dataset which includes additional character bounding box labels, the accuracy of the proposed model is 98.82%. This is still slightly higher than those of the conventional methods, but it is worth noting that the accuracy is improved to 99.25% when the data without character bounding box labels are added.
Design and Implementation of Framework Based Emulator Considering Expansion of MIDS LVT Platform
Sangtae Lee, Jongseo Kim, Sounghyouk Wi, Taegwon Lee, Seungbae Jee, Seungchan Lee
http://doi.org/10.5626/JOK.2021.48.1.61
MIDS LVT is communication equipment mounted with Link-16-based weapon systems to provide the Link-16 operating environment between weapon systems. Currently, the military operates the MIDS LVT BU1, but it will be changed to the BU2 and JTRS according to the performance improvement. The communication interface, message data format, and message composition of the MIDS LVT are different depending on the platform family (BU1/BU2/JTRS) and type (A,D,J). In this paper, we propose a framework based emulator design and implementation method that considers the MIDS LVT platform extension to improve these problems. In consideration of the quality attributes of the sw architecture, we designed a common based framework for modifiability, reusability, and extensibility. The MIDS LVT emulator comprises the MIDS LVT emulator processing, link interface processing, and monitoring tool. The MIDS LVT emulator was implemented by deriving and improving functions through the analysis of the functions of the previously developed overseas tools. Through the development of the MIDS LVT emulator, it can be used to develop and verify the developed Link-16 host system.
Development and Application of Guidelines for Compliance with IEC 62304 International Standards for AI Medical Device Software
DongYeop Kim, Ye-Seul Park, Byungjeong Lee, Jung-Won Lee
http://doi.org/10.5626/JOK.2021.48.1.71
Medical device software developers must implement the processes required by IEC 62304, the international standard for medical device software life-cycle processes, and they must also have a large amount of artifacts to obtain a license. Recently, AI medical device software based on medical images has been actively developed, and since it is treated as standalone software, it must be approved in accordance with IEC 62304 for medical device software. The international standard for AI technology is currently in the discussion stage, and the developer should arbitrarily establish the life-cycle process of AI medical device software, and by matching the specifications of IEC 62304, the performance and safety of AI products will be evaluated. It is unclear which quality management technique should be used to produce the best artifact. This paper provides a quality control technique for fulfilling the scope and requirements of IEC 62304 compliance for AI medical device software in the form of guidelines. These guidelines are also applied to actual AI products to check their potential use in real applications.
Deep Neural Network Structure Selection Using Coverage Methods
http://doi.org/10.5626/JOK.2021.48.1.82
Recently, there has been an immense increase in the use of neural networks in various fields. Consequently, diverse, studies have been conducted to verify and test deep neural networks. One of the most popular studies aimed to test the coverage method about deep neural networks. This study proposes a novel idea that the test coverage can be used in finding the optimal deep neural network structure for given training data using various test coverage methods. To this end, it calculates test coverages for multiple neural network structures that are trained for predicting temperature with the same dataset and select the most suitable neural network. Specifically, the proposed method successfully finds the optimal deep neural network structure out of a total of thirteen neural network structures consisting of one to three Long-Short Term Memory (LSTM) layers and a fully connected layer with 2 to 20 neurons.
Anomaly Detection by a Surveillance System through the Combination of C3D and Object-centric Motion Information
Seulgi Park, Myungduk Hong, Geunsik Jo
http://doi.org/10.5626/JOK.2021.48.1.91
In the existing closed-circuit television (CCTV) videos, the deep learning-based anomaly detection reported in the literature detected anomalies using only the object"s action value. For this reason, it is difficult to extract the action value of an object depending upon the situation, and there is a problem that information is reduced over time. Since the cause of abnormalities in CCTV videos involves several factors such as frame complexity and information according to time series analysis, there is a limit to detecting an abnormality using only the action value of the object. To solve this problem, in this paper, we designed a new deep learning-based anomaly detection model that combined optical flow with C3D to use various feature values centered on the objects. The proposed anomaly detection model used the UCF-Crime dataset, and the experimental results achieved an area under the curve (AUC) of 76.44. Compared to previous studies, this study worked more effectively in fast-moving videos such as explosions. Finally, we concluded that it was appropriate to use the information according to different feature values and time series analysis considering various aspects of the behavior of an object when designing an anomaly detection model.
Gradient Flow Analysis and Performance Comparison of CNN Models
http://doi.org/10.5626/JOK.2021.48.1.100
Among the various deep learning techniques available, convolutional neural networks(CNNs) are widely used due to their superior performance in the fields of computer vision and natural language processing. CNNs can effectively extract the locality and correlation of input data using structures wherein convolutional layers are successively applied to the input data. The performance of neural networks has generally been improved as the depth of CNNs has increased. However, an increase in the depth of a CNN is not always accompanied by a corresponding increase in the accuracy of the neural network. This is because the gradient vanishing problem may occur, thereby causing the weights of the weighted layers to fail to converge. Accordingly, in the present study, the gradient flows of the VGGNet model, ResNet model, and DenseNet model were analyzed and compared, and reasons for the differences in the error rate performances of the models was derived.
A Traceability Analysis for Integrated Relationship Analysis of Development/Safety Artifacts of Cyber Physical Systems
Sejin Jung, Eui-Sub Kim, Junbeom Yoo
http://doi.org/10.5626/JOK.2021.48.1.107
A cyber-physical system (CPS), that is to be used a safety important system, needs to analyze the traceability of development artifacts. The traceability analysis of the CPS should be performed integrating development artifacts and safety/hazard analysis elements because CPS has several features such as heterogeneity, dynamic reconfiguration, and interoperability. However, there is a limitation in terms of expressing all traceability relationships by identically connecting and analyzing the traceability between development artifacts and safety analysis elements. This paper proposes an analysis method and relationships of traceability for CPS. The proposed method uses an abstract model for development artifacts and safety analysis elements that are defined in this paper. The traceability relationships define the relations between elements of the model. The proposed method makes it possible to analyze integrated relationships from development artifacts and safety/hazard analysis elements. The case study shows integrated relationships according to each element of several artifacts.
Using Vertical and Horizonal Hidden Vector of BERT, Attention-based Separated Transfer Learning Model for Dialog Response Selection
http://doi.org/10.5626/JOK.2021.48.1.119
The purpose of this paper is to create a dialog response selection system that accurately identifies the next utterance (one correct answer out of 100 candidates) of a given dialog based on data provided by DSTC. To this end, BERT was used; BERT can be used for multiple purposes and achieves high performance, but it is not easy to customize the model, and it is also difficult to transform the input data format for performance optimization. To address these issues, we propose an effective data augmentation method, and we also propose an independent transfer learning model that involves extracting contextual attention information (self-attention vector) from the BERT model. This made it possible to achieve a performance improvement of 22.85% over the previous value.
Research on WGAN models with Rényi Differential Privacy
Sujin Lee, Cheolhee Park, Dowon Hong, Jae-kum Kim
http://doi.org/10.5626/JOK.2021.48.1.128
Personal data is collected through various services and managers extract values from the collected data and provide individually customized services by analyzing the results. However, data that contains sensitive information, such as medical data, must be protected from privacy breaches. Accordingly, to mitigate privacy invasion, Generative Adversarial Network(GAN) is widely used as a model for generating synthetic data. Still, privacy vulnerabilities exist because GAN models can learn not only the characteristics of the original data but also the sensitive information contained in the original data. Hence, many studies have been conducted to protect the privacy of GAN models. In particular, research has been actively conducted in the field of differential privacy, which is a strict privacy notion. But it is insufficient to apply it to real environments in terms of the usefulness of the data. In this paper, we studied GAN models with Rényi differential privacy, which preserve the utility of the original data while ensuring privacy protection. Specifically, we focused on WGAN and WGAN-GP models, compared synthetic data generated from non-private and differentially private models, and analyzed data utility in each scenario.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr