Vol. 48, No. 2,
Feb. 2021
Digital Library
Implementation and Evaluation of a DMA Controller for PCIe-based FPGA Boards
http://doi.org/10.5626/JOK.2021.48.2.141
An FPGA is an integrated circuit designed to be reconfigurable multiple times at runtime, which shows great performance and energy efficiency in modern applications such as deep learning and big data processing. Major FPGA vendors produce PCIe-based FPGA boards to use FPGAs as accelerators. To transfer large data between a host system and an FPGA, a DMA controller should be implemented inside the FPGA. In previous work, however, controllers did not fully utilize the PCIe bandwidth or were unable to send and receive simultaneously. This paper presents a new DMA controller architecture that can utilize the full-duplex bandwidth of a PCIe link. The DMA controller is implemented and evaluated on a board with Intel Stratix 10 FPGA. The results show that our controller is up to 2.3 times faster than the controller shipped with Intel FPGA Acceleration Stack.
Learning Disentangled Representation of Web Addresses via Convolutional-Recurrent Triplet Network for Phishing URL Classification
http://doi.org/10.5626/JOK.2021.48.2.147
Automated classification of phishing URLs propagated through hyperlinks is critical in environments reinforcing personal connections due to the explosive growth of social media services. Deep learning models for the classification of phishing URLs based on convolutional-recurrent neural networks yielded the best performance in terms of accuracy by modeling the character-level and word-level features. However, the deep learning-based classifier focused on the fitting of a given task via accumulated URLs is limited due to the class imbalance of the phishing attacks that are generated and discarded immediately. We address the class imbalance issue in terms of deep learning-based URL feature space generation task. We propose a modified triplet network structure that explicitly learns the similarity between URLs based on Euclidean distance to alleviate the limitations of the existing deep phishing classifiers. Experiments investigating the real-world dataset of 60,000 URLs collected from web addresses showed the highest performance among the latest deep learning methods, despite the hostile class imbalance. We also demonstrate that the generated URL feature space from the proposed method improved recall by 45.85% compared to the existing methods.
Cross-Validated Ensemble Methods in Natural Language Inference
Kisu Yang, Taesun Whang, Dongsuk Oh, Chanjun Park, Heuiseok Lim
http://doi.org/10.5626/JOK.2021.48.2.154
An ensemble method is a machine learning technique that combines several models to make the final prediction, which guarantees improved performance for deep learning models. However, most techniques require additional models or operations only for an ensemble. To address this problem, we propose a cross-validated ensemble method for reducing the costs of ensemble operations with cross-validation and for improving the generalization effects with the ensemble. To demonstrate the effectiveness of the proposed method, we show the improved performances of the proposed ensemble over the previous ensemble methods using the BiLSTM, CNN, ELMo and BERT models on the MRPC and RTE datasets. We also discuss the generalization mechanism involved in cross-validation along with the performance changes caused by the hyper-parameter of cross-validation.
Method of Reflecting Various Personas in a Chatbot
Shinhyeok Oh, Seok-won Jung, Harksoo Kim
http://doi.org/10.5626/JOK.2021.48.2.160
A chatbot is a computer program that simulates human conversation. Research on generative chatbots that provide various responses based on personal characteristics has been increasing. Representatively, there are persona chatbots that reflect personal characteristics in chatbots. Persona chatbots refers to a chatbot that reflects persona, which means personal characteristics, and are gaining popularity due to the movement to reflect a brand personality in various services. In response to this trend, this paper proposes a chatbot model that can generate different responses for each persona by suggesting sentence persona encoder and table persona encoder that reflects personas based on dual WGAN generative chatbot. The performance of the proposed model is verified through comparative experiments and experimental examples for each module using quantitative and qualitative evaluation.
Joint Model of Morphological Analysis and Named Entity Recognition Using Shared Layer
Hongjin Kim, Seongsik Park, Harksoo Kim
http://doi.org/10.5626/JOK.2021.48.2.167
Named entity recognition is a natural language processing technology that finds words with unique meanings such as human names, place names, organization names, dates, and time in sentences and attaches them. Morphological analysis in Korean is generally divided into morphological analysis and part-of-speech tagging. In general, named entity recognition and morphological analysis studies conducted in independently. However, in this architecture, the error of morphological analysis propagates to named entity recognition. To alleviate the error propagation problem, we propose an integrated model using Label Attention Network (LAN). As a result of the experiment, our model shows better performance than the single model of named entity recognition and morphological analysis. Our model also demonstrates better performance than previous integration models.
Comparison of False Alarm Detection using KLEE and CBMC for Effective Multitask Program Verification
http://doi.org/10.5626/JOK.2021.48.2.174
OiL-CEGAR[1] verifies the composition of a formal OS model and an abstracted application program for accurate verification. Due to the use of the abstract program, however false-alarms can be reported and executability checking for identifying false-alarms requires a high cost. Therefore, efficient executability checking is essential to improve verification performance. To find an effective executability checking method, this study introduces and compares two different techniques that perform executability checking. The first one collects the Boolean formula for the entire program and checks the reachability of all the program blocks in the counterexample by using CBMC. While the second one uses KLEE and identifies non-executable blocks in the counterexample through the binary search-based executability checking. The suggested executability checking methods are applied to a window controller program from the automotive domain. Results show that executability checking using KLEE takes only 1/2000 time compared to that of CBMC and reduces 11.78% of OiL-CEGAR verification costs.
Building a Korean Sentence-Compression Corpus by Analyzing Sentences and Deleting Words
GyoungHo Lee, Yo-Han Park, Kong Joo Lee
http://doi.org/10.5626/JOK.2021.48.2.183
Developing a sentence-compression system based on deep learning models requires a parallel corpus consisting of both original sentences and compressed sentences. In this paper, we propose a sentence-compression algorithm that can compress an original sentence into a short sentence. Our basic approach is to delete nodes from a syntactic-dependency tree of the original sentence while maintaining the grammaticality of the compressed sentence. The algorithm chooses nodes to be deleted using the structural constraints and semantically obligatory information of the sentence. By applying the algorithm to the first sentences and headlines of news articles, we built a Korean sentence-compression corpus consisting of approximately 140,000 pairs. We manually assessed the quality of the compression in terms of readability and informativeness, which yielded results of 4.75 and 4.53 out of 5, respectively.
Autoencoder-based Learning Contribution Measurement Method for Training Data Selection
Yuna Jeong, Myunggwon Hwang, Wonkyung Sung
http://doi.org/10.5626/JOK.2021.48.2.195
Despite recent significant performance improvements, the iterative process of machine-learning algorithms makes development and utilization difficult and time-consuming. In this paper, we present a data-selection method that reduces the time required by providing an approximate solution . First, data are mapped to a feature vector in latent space based on an Autoencoder, with high weight given to data with high learning contribution that are relatively difficult to learn. Finally, data are ranked and selected based on weight and used for training. Experimental results showed that the proposed method selected data that achieve higher performance than random sampling.
Bounded Search Strategies of Concolic Testing for Effective and Efficient Structural Coverage Achievement
http://doi.org/10.5626/JOK.2021.48.2.201
This paper proposes a loop-bounded search strategy for effective and efficient coverage achievement in concolic testing. In selecting a new path to explore, a loop-bounded search strategy limits the number of iterations in a loop to a certain loop-bound, so that the concolic testing is guided to explore various program behaviors within a limited range. In addition, to extend the range of path exploration gradually, this search strategy increments the loop-bound over test executions based on their coverage achievement rates. We implemented three versions of loop-bounded search strategies based on three existing concolic search strategies of CREST. The experiments with 4 real-world target programs (Vim, Grep, Busybox Awk, and Busybox Sed) showed that CREST achieves a higher branch coverage more quickly when the loop-bounded search strategies are applied.
Ensemble Modeling with Convolutional Neural Networks for Application in Visual Object Tracking
Minji Kim, Ilchae Jung, Bohyung Han
http://doi.org/10.5626/JOK.2021.48.2.211
In the area of computer vision, visual object tracking aims to estimate the status of a target object from an input video stream, which can be broadly applicable to industries such as surveillance and the military. Recently, deep learning-based tracking algorithms have gone through significant improvements by using tracking-by-detection or template-based approach. However, these approaches are still suffering from inherent limitations caused by each strategy. In this paper, we propose a novel method to model ensemble trackers by fusing the two strategies, tracking-by-detection and template-based approach. We report significantly enhanced performance on widely adopted visual object tracking benchmarks, OTB100, UAV123, and LaSOT.
Hyperbolic Graph Transformer Networks for non-Euclidean Data Analysis on Heterogeneous Graphs
Seunghun Lee, Hyeonjin Park, Hyunwoo J Kim
http://doi.org/10.5626/JOK.2021.48.2.217
Convolution Neural Networks (CNNs), which are based on convolution operations, are used for various tasks in image classification, image generation, time series analysis, etc. Since the convolution operations are not directly applicable to non-Euclidean spaces such as graphs and manifolds, a variety of Graph Neural Networks (GNNs) have extended convolutional neural networks to homogeneous graphs, which has a single type of edges and nodes. However, in real-world applications, heterogeneous and hierarchical graph data often occur. To expand the operating range of GNNs to the graphs that have multiple types of nodes and edges with the hierarchy, herein, we propose a new model that integrates Hyperbolic Graph Convolution Networks (HGCNs) and Graph Transformer Networks (GTNs).
SuperPoint-High Resolution Network (HRN); Interest Point Detection using HRNet
http://doi.org/10.5626/JOK.2021.48.2.226
Interest point detector is a fundamental method in computer vision for image matching and image recognition. SIFT, ORB, etc. have been used in many computer vision applications. As acquiring ground truth of interest points is difficult, SuperPoint has been devised to make pseudo ground truth using synthetic corners in training the model. The SuperPoint produced competitive performance to leading conventional computer vision interest point detectors. This paper proposes SuperPoint-High Resolution Network (HRN) to improve repeatability that is one of the most important features in an interest point detector by substituting HRN exchanging information among different resolution activation maps for the shared encoder implemented with a convolutional neural network and by modifying the detector head to accommodate high-resolution activation maps. The proposed method demonstrated immense improvements in HPatches data over the SuperPoint in terms of repeatability and localization errors of interest points.
Predicting Chemical Structure of Drugs Using Deep Learning
Soohyun Ko, Chihyun Park, Jaegyoon Ahn
http://doi.org/10.5626/JOK.2021.48.2.234
Numerous computer-based methods have been investigated in attempts to reduce the time and cost of drug development. In particular, with the recent development of deep learning techniques, various generation models for generating the chemical formulas of candidate compounds and reinforcement learning models to generate chemical formulas that satisfy specific conditions have been presented. In this paper, we propose a reinforcement learning model that exploits predicted binding affinity information between specific proteins and generated compounds. More specifically, the generative model used in this paper is Stack-RNN, and reinforcement learning is implemented by using Stack-RNN as a policy to ensure that the generated formula has specific chemical properties and high binding affinity with specific proteins. The proposed model generates paper, we generated the chemical formulas of compounds that are similar to three anti-cancer drugs (Sorafenib, Sunitinib, and Dasatinib) by using the target protein information of these three anti-cancer drugs.
Service Migration Based on Reinforcement Learning in Vehicular Edge Computing
http://doi.org/10.5626/JOK.2021.48.2.243
As edge computing can provide low latency and real-time services, it is emerging as a promising technology that can lead the Internet of things(IoT). However, user mobility and limited coverage of edge computing result in service interruption and reduce Quality of Service(QoS). Thus, service migration is considered an important issue to guarantee seamless service. In this paper, a migration decision algorithm is proposed using Q-learning, as a reinforcement learning method in the vehicular edge computing environment. The proposed algorithm decides whether or not to migrate and where to migrate in order to meet delay constraint and minimize system cost. In the performance evaluation, we compared propsed algorithm with other algorithms in terms of deciding whether or not to migrate and where to migrate, and our proposed algorithm shows better performance compared to the other algorithms.
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