Vol. 50, No. 9, Sep. 2023
Digital Library


Optimizing Computation of Tensor-Train Decomposed Embedding Layer

Seungmin Yu, Hayun Lee, Dongkun Shin

http://doi.org/10.5626/JOK.2023.50.9.729

Personalized recommendation system is ubiquitous in daily life. However, the huge amount of memory requirement to store the embedding tables used by deep learning-based recommendation system models is taking up most of the resources of industrial AI data centers. To overcome this problem, one of the solutions is to use Tensor-Train (TT) decomposition, is promising compression technique in deep neural network. In this study, we analyze unnecessary computations in Tensor-Train Gather and Reduce (TT-GnR) which is the operation of embedding layer applied with TT decomposition. To solve this problem, we define a computational unit called group to bind the item vectors into a group and propose Group Reduced TT-Gather and Reduce operation to reduce unnecessary operations by calculating with groups. Since the GRT-GnR operation is calculated in groups, computational cost varies depending on how item vectors are grouped. Experimental results showed that the GRT-GnR operation had a 41% decrease in latency compared to conventional TT-GnR operation.

An Improved Algorithm of Finding a Maximal Common Subsequence

Hyeonjun Shin, Joong Chae Na, Jeong Seop Sim

http://doi.org/10.5626/JOK.2023.50.9.737

A maximal common subsequence (MCS) of two strings is a common subsequence that cannot be extended by inserting any character. Unlike the longest common subsequence (LCS), the length of MCS can vary as the longest MCS is an LCS. Although LCS is commonly used to compare similarities of two sequences, computing can take a significant amount of time. Hence, finding a longer MCS is important, as it can be computed faster than the LCS. An algorithm was proposed to compute one of the MCSs of two strings X and Y of total length n using O(kn) space and O(n√(logn/loglogn)) time. Improved algorithms were also proposed. In this paper, we present an algorithm that can check for more characters to compute an MCS. The algorithm proposed in this paper runs in O(kn) space and O(n√(logn/loglogn)) time for a given constant k. Experimental results using both real and randomly generated data showed that the length of the MCS computed by the algorithm proposed in this paper could be up to 6.31 times longer than those computed by previous algorithms.

Applying Deep Neural Networks and Random Forests to Predict the Pathogenicity of Single Nucleotide Variants in Hereditary Cancer-associated Genes

Da-Bin Lee, Seonhwa Kim, Moonjong Kang, Changbum Hong, Kyu-Baek Hwang

http://doi.org/10.5626/JOK.2023.50.9.746

The recent proliferation of genetic testing has made it possible to explore an individual"s genetic variants and use pathogenicity information to diagnose and prevent genetic diseases. However, the number of identified variants with pathogenicity information is quite small. A method for predicting the pathogenicity of variants by machine learning was proposed to address this problem. In this study, we apply and compare deep neural networks with random forests and logistic regression, which have been widely used in previous studies, to predict variant pathogenicity. The experimental data consisted of 1,068 single-nucleotide variants in genes associated with hereditary cancers. Experiments on 100 random data-sets generated for hyperparameter selection showed that random forests performed best in terms of area under the precision-recall curve. On 15 holdout gene data-sets, deep neural networks performed best on average, but the difference in performance from the second-best random forest was not significant. Logistic regression was also statistically significantly worse than that of either model. In conclusion, we found that deep neural networks and random forests were generally better than logistic regression at predicting the pathogenicity of single-nucleotide variants associated with hereditary cancer.

Activity Recognition Algorithms based on Sensor Data for Wearable Walking-Assistive Robots

Junhyeok Son, Jinho Sohn, Seungjin Choi

http://doi.org/10.5626/JOK.2023.50.9.751

The present study aimed to comprehensively perform deep learning-based activity recognition for enabling wearable walking-assistive robots to accurately identify corresponding activities using sensor data. Dataset used for this study was obtained from rehabilitation training of walking-assistive robot wearers. To optimize the training process of deep neural networks, we conducted careful processing and refinement of the data, eliminating unnecessary segments. We systematically compared various deep neural network architectures, including Convolutional Neural Network(CNN), Long Short-Term Memory(LSTM) and CNN-Gated Recurrent Unit(GRU), to identify the most effective model for activity recognition in wearable walking-assistive robots. Our experiments conducted using real-world data demonstrated that all deep neural networks used in this study performed comparably well, with LSTMs showing a slight advantage for our specific problem. This research contributes valuable insights and advancements towards enhancing capabilities of wearable walking-assistive robots, highlighting the potential for further improvements in this domain.

C3DSG: A 3D Scene Graph Generation Model Using Point Clouds of Indoor Environment

Hojun Baek, Incheol Kim

http://doi.org/10.5626/JOK.2023.50.9.758

To design an effective deep neural network model to generate 3D scene graphs from point clouds, the following three challenging issues need to be resolved: 1) to decide how to extract effective geometric features from point clouds, 2) to determine what non-geometric features are used complementarily for recognizing 3D spatial relationships between two objects, and 3) to decide which spatial reasoning mechanism is used. To address these challenging issues, we proposed a novel deep neural network model for generating 3D scene graphs from point clouds of indoor environments. The proposed model uses both geometric features of 3D point cloud extracted using Point Transformer and various non-geometric features such as linguistic features and relative comparison features that can help predict the 3D spatial relationship between objects. In addition, the proposed model uses a new NE-GAT graph neural network module that can apply attention to both object nodes and edges connecting them to effectively derive spatial context between objects. Conducting a variety of experiments using 3DSSG benchmark dataset, effectiveness and superiority of the proposed mode were proven.

An AI Web-based Interface for Mitigating Political Polarization

Jaehoon Kim, Sohyun Park, Kyungsik Han

http://doi.org/10.5626/JOK.2023.50.9.771

Political polarization promotes not only social divisions but also social costs in various fields. Although the negative impacts of news media and online communities on society have been reported for a long time, they remain challenges that society needs to address. This study proposes an AI-based web interface that allows users to consume both information consistent with their political ideology and information contrary to theirs from the perspective of promoting diversity. Pre and post-surveys scored on a 5-point Likert scale were used to measure the degree of awareness of the importance of consuming information from diverse perspectives. The interface helped to increase the awareness of users at a statistically significant level.

Machine Learning-Based Approach for Predicting Drug-Induced Liver Injury of Chemical Compounds

Soyeon Lee, Sunyong Yoo

http://doi.org/10.5626/JOK.2023.50.9.777

Drug-induced liver injury (DILI) is one of the factors constraining the distribution of investigational products on the market. Therefore, DILI risk of compounds should be assessed in advance. Although in vivo and in vitro methods can be used to test drug safety, both methods are labor-intensive, time consuming and expensive. In this study, we suggested random forest, light gradient boosting machine, logistic regression models to overcome the above problems. These models used molecular structure and physicochemical features as input to predict the DILI as output. The optimal model was random forest, which performed well for evaluation metrics overall. The proposed model is expected to help drug development process by identifying potential DILI of drug candidates in advance.

TwinAMFNet : Twin Attention-based Multi-modal Fusion Network for 3D Semantic Segmentation

Jaegeun Yoon, Jiyeon Jeon, Kwangho Song

http://doi.org/10.5626/JOK.2023.50.9.784

Recently, with the increase in the number of accidents due to misrecognition in autonomous driving, interest in 3D semantic segmentation based on sensor fusion using multi-modal sensors has increased. Accordingly, this study introduces TwinAMFNet, a novel 3D semantic segmentation neural network through sensor fusion of RGB cameras and LiDAR. The proposed neural network includes a twin neural network that processes RGB images and point cloud projection images projected on a 2D coordinate plane and through an attention-based fusion module for feature step fusion in the encoder and decoder. The proposed method shows improvement of further extended object and boundary classification. As a result, the proposed neural network recorded approximately 68% performance in 3D semantic segmentation based on mIoU, and showed approximately 4.5% improved performance compared to the ones reported in the existing studies.

Vision-based Position Deviation Fault Injection Method for Building a Collaborative Robot Motion Fault Dataset

Donghee Yun, Dongyeon Yoo, Jungwon Lee

http://doi.org/10.5626/JOK.2023.50.9.795

The data-based fault detection method, which collects data from internal and external sensors in real-time and predicts fault, is being applied to collaborative robots, which are key facilities in smart factories. The data-based fault detection method requires a large amount of data for learning, and in particular, a large amount of data labeled as a fault state is essential. However, it is difficult to obtain large amounts of actual fault data in industrial settings. Therefore, in this study, the output of the collaborative robot fault state based on a vision sensor was analyzed and compared with the output of the normal state, and a fault injection method was proposed based on the deviation between the analyzed output signals. Collaborative robot data collected in the actual fault state could be replaced with data collected in the proposed fault injection state. The comparison of the performance of the model trained with fault injection data and trained with actual fault data confirmed that there was almost no difference, with an average of 0.97 and 0.98 accuracy, thus verifying the effectiveness of the proposed fault injection method.

A GCN-based Time-Series Data Anomaly Detection Method using Sensor-specific Time Lagged Cross Correlation

Kangwoo Lee, Yunyeong Kim, Sungwon Jung

http://doi.org/10.5626/JOK.2023.50.9.805

Anomaly detection of equipment through time series data is a very important because it can prevent further damage and contribute to productivity improvement. Although research studies on time series data anomaly detection are being actively conducted, but they have the following restrictions. First, unnecessary false alarms occur because correlations with other sensors are not considered. Second, although complete graph modeling and GAT have been applied to analyze the correlation of each sensor, this method requires a lot of time due to the increase in unnecessary operations. In this paper, we propose SC-GCNAD(Sensor-specific Correlation GCN Anomaly Detection) to address these problems. SC-GCNAD can analyze the exact correlation of each sensor by applying TLCC that reflects characteristics of time series data. It utilize GCN with excellent model expressiveness. As a result, SC-GCNAD can improve F1-Score by up to 6.37% and reduce analysis time by up to 95.31% compared to the baseline model.

Design and Countermeasures of PowerShell-based Attack Techniques to Bypass Portable Executable Image Detection Methods

Jiwon Jang, Daehee Jang

http://doi.org/10.5626/JOK.2023.50.9.813

A fileless malware attack is an advanced ever-developing attack that usually exploits script functions in PowerShell, a default Windows tool. The reason is that PowerShell allows people to access computer resources easily because its tools provide infrastructure management broadly in Windows systems. This PowerShell-based fileless malware does involve using Portable Executable(PE) images in the process of attacking. One way to identify it is by tracking and detecting the flow of PE images (i.e., DLL Injection monitoring). In this paper, we demonstrate that it is possible to execute an attack without any PE images despite using fileless malware based on PowerShell and made Proof-of-Concept malware attack codes to demonstrate this method and compare the impact of such attack. Additionally, we discussed techniques to cope with this kind of advanced malicious code.

Derivation of Kotlin Flow API from Observer Pattern

Jungsun Kim

http://doi.org/10.5626/JOK.2023.50.9.821

Kotlin Flow is an asynchronous reactive stream library developed on top of Kotlin coroutine’s suspending functions. Recently, many reactive applications have increasingly adopted the Kotlin Flow instead of the well-known, but nevertheless heavy-weight, RxJava because it provides back-pressure support naturally by using coroutine’s suspending functions. However, the workings of the Flow API are difficult for most programmers to understand. One of the main reasons is due to the ignorance of the fact that its API is originated from the Observer Pattern, as well as its inherent reactive nature. In this paper, we show that Kotlin Flow API is nothing but a reactive version of the adapted Observer Pattern, together with Kotlin coroutines, by step-by-step illumination of how its API is derived from both the Observer Pattern and Kotlin’s coroutine.


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