Latest Issue
Vol. 53,
No. 3,
Mar.
2026
All Issues
Bandwidth-Aware Memory Migration for Tiered Memory Systems
Sunyoung Hwang, Wonjung Lee, Eojin Lee
http://doi.org/10.5626/JOK.2026.53.3.181
The growing demand for large-capacity memory has underscored the significance of CXL-based tiered memory architectures. Traditional tiered memory management typically allocates low-utilization cold data to slower memory tiers. However, this approach can result in performance degradation when bandwidth requirements are high. To address this issue, we present Bandwidth- Aware Memory Migration (BAMM), a mechanism that continuously monitors both memory and bandwidth utilization. When bandwidth demand increases, BAMM selectively migrates data with moderate access frequencies—not just cold data—to optimize bandwidth utilization in slower memory. Experimental results demonstrate that BAMM achieves an average performance improvement of 11.22% and up to 27.11% under bandwidth-intensive workloads compared to conventional methods, while also enhancing overall system bandwidth utilization by an average of 18.29 GB/s, reaching high as 33.48 GB/s.
Forward-Forward Algorithm with Arc Memory in Few-Shot Learning
Taewook Hwang, Hyein Seo, Sangkeun Jung
http://doi.org/10.5626/JOK.2026.53.3.190
The Forward-Forward algorithm introduces a new approach to replace backpropagation, which learns independently for each layer, akin to the functioning of the human brain. However, this independent learning raises concerns about the transmission of information between layers. In this study, we propose a method that assigns a dedicated memory space to each layer, emulating human memory, to enhance information transfer between layers. We store the average output values for each label in memory and employ an angular margin loss function to measure the difference between these stored values and the current layer's output, ensuring that each label is represented at distinct angles. Additionally, we compare the memory values from the previous layer with the current layer's output using the angular margin loss function to facilitate alignment with the previous layer's angle. Experimental results indicate that, despite initial limitations with very limited data, our proposed method achieved 90.71% accuracy on the MNIST dataset with a sufficient data volume (100-shot), outperforming the existing Forward-Forward algorithm, which achieved 89.63% accuracy.
A Study on the Effective Operation and Practice of Artificial Intelligence Online Liberal Arts Education for Non-Majors
Jaeho Hwang, Minseok Hur, Sooyon Seo, Moohong Min
http://doi.org/10.5626/JOK.2026.53.3.196
This paper presents a case study analyzing a questionnaire administered before and after a course on basic artificial intelligence tailored for non-majors. The curriculum was designed to enhance students' understanding of artificial intelligence through practical assignments. Survey results indicated a significant improvement in students' grasp of artificial intelligence concepts after completing the course, with practical assignments playing a crucial role in this learning enhancement. Additionally, student satisfaction throughout the course was notably high. These findings have important implications for the future development and refinement of artificial intelligence education curricula. This study highlights the importance of a well-structured curriculum that incorporates practical assignments to bolster both students' understanding and interest in the subject.
Korean Paper Based Retrieval Augmented Generation Dataset
Junho Han, Minjun Choi, Keunha Kim, Youngjoong Ko
http://doi.org/10.5626/JOK.2026.53.3.205
Large language models (LLMs) trained on general domain data have limitations in specialized fields that are rich in information and technical terminology. Retrieval-augmented generation (RAG) improves answer accuracy and reliability by referencing external knowledge, making it particularly effective in specialized domains where pre-training data is scarce. However, there is a lack of public datasets for Korean specialized domains, highlighting the need for a dedicated retrieval-augmented generation dataset. This paper introduces a new Korean RAG dataset based on scientific and technical papers to support research in this area. We preprocessed existing document-query data to create a searchable corpus and extracted key phrases and key sentences suited for specialized applications. Additionally, we conducted a comprehensive quantitative evaluation of the dataset‘s quality. By reflecting the unique characteristics of scientific and technical papers, this dataset serves as a robust foundation for Korean RAG systems.
Automatic Bug Report Generation for Open Source Projects via QLoRA Fine-Tuning, CTQRS-Structured Prompting, and the Integration of CoT and Few-Shot Strategies
http://doi.org/10.5626/JOK.2026.53.3.217
Bug reports are crucial for tracking defects and maintaining software. However, in open-source environments, they are often created by non-experts, which can result in incomplete, inconsistent and less reproducible reports. Previous studies have primarily focused on template-based methods or simple fine-tuning, without fully utilizing multidimensional quality metrics like CTQRS or systematically assessing the effectiveness of few-shot prompting. This paper proposes a novel approach that integrates QLoRA-4bit fine-tuning of large language models with CTQRS-based structured prompting, Chain-of-Thought reasoning, and one or two-shot examples. Experiments conducted on a Bugzilla dataset of 3,966 pairs demonstrated significant improvements: CTQRS increased from 77% to 94%, ROUGE-1 Recall rose from 0.61 to 0.87, and SBERT similarity improved from 85 to 90. Additionally, QLoRA alone outperformed the baseline, with the supplementary strategies contributing complementary gains. These findings empirically validate that structured prompting, reasoning guidance, and minimal example provision are critical factors in enhancing performance, highlighting the practical potential of resource-efficient fine-tuning for open-source software maintenance.
Enhanced Object Detection in Fog Using a Multi-Stage Preprocessing Ensemble with Hyperparameter Optimization in YOLOv8
Jabeen Koo, Yujin Kim, Bokyung Amy Kwon, Kyungtae Kang
http://doi.org/10.5626/JOK.2026.53.3.230
Conventional deep learning-based object detection models often struggle with detection performance in foggy environments. To improve this performance, in specific environments, research has focused on image correction and blending techniques. However, most single-processing methods have drawbacks, such as over-brightening certain areas or darkening edges. In this study, we propose a multi-processing model designed to enhance object detection performance in foggy conditions using continuous hyperparameter optimization based on the YOLOv8 model. We aim to address limitations of individual techniques by combining multiple preprocessing methods, specifically optimizing hyperparameters for gamma correction and histogram matching. Our performance evaluation results indicate that the ensemble model we developed outperformed individual single processing models on various metrics, suggesting that it can significantly enhance object detection performance in challenging environments like fog.
A Structured Three-Level Classification Model for Hand-Finger Gesture Recognition in Virtual Reality
Sehyeok Yoo, Kyungmin Kim, Youngho Chai
http://doi.org/10.5626/JOK.2026.53.3.239
Hand-based interaction in virtual reality (VR) is intuitive, but vision-based hand tracking often struggles to accurately capture user intent due to issues like occlusion, lighting variations, and tracking noise. To improve the stability of existing binary (straight/bent) classification and reduce the cognitive load of multi-level schemes, this study introduces a rule-based gesture recognition framework that categorizes finger flexion into three states: straight, intermediate, and bent. A multi-view webcam setup, combined with exponential moving-average filtering, was implemented to enhance robustness against occlusion and jitter. User evaluations across three VR scenarios showed high recognition accuracy and controllability, regardless of variations in hand size or morphology. However, gestures involving unfamiliar hand shapes highlighted areas for usability improvement. These results suggest that a practical, extensible, and reliable VR gesture input system can be developed without relying on complex machine-learning models, indicating potential for broader application and future enhancements.
A Method for Predicting Tappability of Interface Elements in Mobile Application Using Layout Features
http://doi.org/10.5626/JOK.2026.53.3.248
In mobile applications, a disconnect between how users perceive tappability and the designer's intent can lead to users tapping unintended elements or encountering no response (e.g., dead clicks), which negatively impacts the user experience. Previous research has primarily focused on visual features to predict tappability, but these methods can only be applied once the design is finalized. To address this limitation, this study introduces an AI-based approach that predicts tappability using only the layout features of UI elements. We defined layout features-such as element position, size, and spacing between neighboring elements-that can be directly extracted from layout files, and utilized them in a Support Vector Machine (SVM) model. The proposed model achieved an F1 score of 0.772, indicating strong performance. These findings suggest that tappability prediction can be effectively carried out using only layout information in the early design stages, offering a practical solution for proactively detecting usability issues throughout the development process.
Recommendation Systems based on Feature-Adaptive Graph Attention Network
Jin-Soo Ahn, Min-Cheol Park, Min-Jeong Oh, Do-Jin Choi
http://doi.org/10.5626/JOK.2026.53.3.256
With the rapid advancement of information and communication technology, the volume and variety of online content have grown explosively, causing users to suffer from information overload and increasing the need for effective recommender systems. In this paper, we propose a graph learning method that models user–content interactions as a graph and effectively captures domain-specific structural characteristics and noise. To achieve data-adaptive graph learning, we build on an enhanced Graph Attention Network (GAT) that applies different operations to each head and design specialized heads such as RatingConv and PopConv to reflect data characteristics during training. The proposed method achieves approximately a 10% performance improvement in the NDCG@20 metric compared to existing baseline models on standard recommendation benchmarks, including the MovieLens-1M, MovieLens-25M, and FilmTrust datasets.
Throughput-aware and DQN-based Heterogeneous Interface Selection Mechanism in 5G NR-V2X Environments
http://doi.org/10.5626/JOK.2026.53.3.265
As autonomous driving technology nears commercialization, maintaining stable communication performance in diverse Vehicle-to-Everything (V2X) environments becomes increasingly critical. However, delays, disconnections, and packet loss that occur during transitions between different interfaces can degrade throughput performance. While previous studies have mainly concentrated on seamless handover, they fall short in optimizing the data transmission and reception required in real-world autonomous driving scenarios. This paper proposes a Deep Q-Network (DQN)-based interface selection algorithm to address these limitations and enhance throughput. Performance analysis using ns-3 demonstrates that the proposed mechanism achieves an approximate 14% improvement in throughput compared to existing handover 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