Vol. 50, No. 2,
Feb. 2023
Digital Library
Analyses of the Memory Usage Efficiency of the Scudo Memory Allocator
KyungSeok Lee, Taehyung Lee, Young Ik Eom
http://doi.org/10.5626/JOK.2023.50.2.103
Optimizing memory usage is important in mobile systems because diverse applications share limited system memory in those systems. Although the default memory allocator of Android, Scudo, is designed to mitigate vulnerabilities related to heap memory and optimize the memory allocation performance, it has inefficiency in terms of memory usage aspects. This paper addresses two problems based on the structure and operation mechanism of Scudo. First, the problems that occur due to the static number of local caches are analyzed. Second, the internal fragmentation problems induced by the inefficient free-chunk management policy are analyzed. The evaluation results show that, according to the characteristics of applications, there are some performance overheads and memory waste due to the static number of local caches in Android systems. Furthermore, we confirm that repeated allocation and free operations fragment almost all allocated physical pages in Scudo; hence pages are not actively reclaimed to the OS even after 90% of allocated chunks are freed.
HLS-based Efficient CNN Accelerator Design Supporting Data Reuse
Sumin Kim, Taeyun An, Youngmin Yi
http://doi.org/10.5626/JOK.2023.50.2.111
This paper presents an efficient and scalable convolution neural network (CNN) accelerator design and implementation using high-level synthesis (HLS). The proposed CNN accelerator consists of tile buffers to reduce DRAM accesses and a 2-dimensional processing element (PE) array with weight stationary architecture. Using HLS directives, we could easily implement an efficient accelerator that exploited PE parallelism and the pipelined parallelism among tasks. However, unlike HDL, it is not easy to implement dynamic data reuse using HLS, which is required to skip costly DRAM access when the PE re-reads the same data in the tile buffer. We implemented data reuse using Stream, a FIFO buffer provided by HLS library. With data reuse, the performance was increased by 13.7%, which corresponding to an inference time of 49 ms with ResNet50 on a Xilinx Alveo U200 FPGA board. We could easily scale PE numbers and confirmed performance scalability, achieving up to 835 GOPS and 35.3 GOPS/W at the server system level with 64×64 PEs.
Graph Neural Networks with Prototype Nodes for Few-shot Image Classification
http://doi.org/10.5626/JOK.2023.50.2.127
The remarkable performance of deep learning models is based on a large amount of training data. However, there are a number of domains where it is difficult to obtain such a large amount of data, and in these domains a large amount of resources must be invested for data collection and refining. To overcome these limitations, research on few-shot learning, which enables learning with only a small number of data, is being actively conducted. In particular, among meta learning methodologies, metric-based learning which utilizes similarity between data has the advantage that it does not require fine-tuning of the model for a new task, and recent studies using graph neural networks have shown good results. A few-shot classification model based on a graph neural network can explicitly process data characteristics and the relationship between data by constructing a task graph using data of a given support set and query set as nodes. The EGNN(Edge-labeling Graph Neural Net) model expresses the similarity between data in the form of edge labels and models the intra-class and inter-class similarity more clearly. In this paper, we propose a method of applying a prototype node representing each class to few-shot task graph to model the similarity between data and class-data at the same time. The proposed model provides a generalized prototype node that is created based on task data and class configuration, and it can perform two different few-shot image classification predictions based on the prototype-query edge label or the Euclidean distance between prototype-query nodes. Comparing the 5-way 5-shot classification performance on the mini-ImageNet dataset with the EGNN model and other meta-learning-based few-shot classification models, the proposed model showed significant performance improvement.
A Pre-processing Method for Learning Data Using eXplainable Artificial Intelligence
Changhong Lee, Jaemin Lee, Donghyun Kim, Jongdeok Kim
http://doi.org/10.5626/JOK.2023.50.2.133
Artificial intelligence model generation proceeds to the stages of learning data processing, model learning, and model evaluation. Data pre-processing techniques for creating quality learning data contribute many of the methods for improving model accuracy. Existing pre-processing techniques tend to rely heavily on the experience of model generators. If pre-processing is performed based on experience, it is difficult to explain the basis for selecting the corresponding pre-processing technique. However, the reason why generators are forced to rely on experience is that the learning model becomes huge and complicated to a level that is difficult for humans to interpret. Therefore, research is being conducted to explain the operation method of the model by introducing eXplainable AI. In this paper, we propose a learning data pre-processing system using eXplainable AI. The system operation process is trained with data that has not been pre-processed, the learned model is analyzed using eXplainable AI, and the data pre-processing is repeated based on that information. Finally, we will improve the model performance, explain pre-processing reliability, and show the practicality of the system.
OS-in-the-Loop Concolic Testing for Multitask Embedded Software
http://doi.org/10.5626/JOK.2023.50.2.140
Verification of multitask embedded software still depends on manual effort because of its hardware-dependent structure and complicated multitasking works. We present the OSin-the-loop(OiL) concolic test, the new approach for automatic and efficient verification of multitask embedded software. Given the OS model, hardware stubs that replace platform-dependent code, and annotated application code with explicit context switch control logic, it provides a hardwareindependent environment and automatically checks properties with the concolic test. In the application of our approach on a representative multitask embedded software, Object-Follower and concolic testing tool CROWN 2.0, when there is an OS model, our method achieves fewer false alarms from 91.67% to 5.13% than without OS.
Change Description Difference Analysis between Human and Code Differencing Techniques
Moojun Kim, Beomcheol Kim, Jindae Kim
http://doi.org/10.5626/JOK.2023.50.2.150
This study investigated the difference between descriptions of code changes made by source code differencing tools and humans. We applied two popular source code differencing techniques to collected code changes. We found that these tools often generated different descriptions for the same changes, and only 3% of the changes have the same descriptions from both tools. On the other hand, human participants agree on change descriptions for 50% of the given changes. Furthermore, many of the different descriptions were caused by simple mistakes. If we ignore differences caused by these mistakes, human participants described 71% of the changes similarly. We also compared change type and entity type of edit scripts generated by human and the source code differencing techniques for the same changes. We found that the techniques generated the same description as humans for only 8.20～35.65% of the changes, which indicates that these techniques require significant improvement to provide descriptions similar to human’s.
Developing a Testability Prediction Model for High Complexity Software using Regression Analysis
http://doi.org/10.5626/JOK.2023.50.2.162
Testability is the degree to which the software supports testing in a given test context. Early prediction of testability can help developers identify software components that require a lot of effort to ensure software quality, plan testing activities, and recognize the need for refactoring to reduce testing effort. Existing studies have been conducted to predict testability by performing regression analysis using software metrics and code coverage. These studies used training data with a large proportion of simple software structures. However, prediction models trained with imbalanced data with a large proportion of simple structures may have low testability prediction accuracy of high complexity software. We used the training data generated based on the metric acceptance criteria of industry domain standards to build a prediction model considering high complexity software. As a result of building a testability prediction model using three regression analyses, we construct a predictive model with a branch coverage error of about 4.4% and a coefficient of determination of 0.86.
Risk Scheduling-based Optimistic Exploration for Distributional Reinforcement Learning
Jihwan Oh, Joonkee Kim, Se-Young Yun
http://doi.org/10.5626/JOK.2023.50.2.172
Distributional reinforcement learning demonstrates state-of-the-art performance in continuous and discrete control systems with the features of variance and risk, which can be used to explore action space. However, the exploration method employing the risk property is hard to find, although numerous exploration methods in Distributional RL employ the variance of the return distribution for an action. This paper presents risk scheduling approaches that explore risk levels and optimistic behaviors from a risk perspective in Distributional reinforcement learning. We demonstrate the performance (win-rate) enhancement of the DMIX, DDN, and DIQL algorithms, which integrate Distributional reinforcement learning into a multi-agent system using risk scheduling in a multi-agent setting with comprehensive experiments.
A Study of Metric and Framework Improving Fairness-utility Trade-off in Link Prediction
Heeyoon Yang, YongHoon Kang, Gahyung Kim, Jiyoung Lim, SuHyun Yoon, Ho Seung Kim, Jee-Hyong Lee
http://doi.org/10.5626/JOK.2023.50.2.179
The advance in artificial intelligence (AI) technology has shown remarkable improvements over the last decade. However, sometimes, AI makes biased predictions based on real-world big data that intrinsically contain discriminative social factors. This problem often arises in friend recommendations in Social Network Services (SNS). In the case of social network datasets, Graph Neural Network (GNN) is utilized for training these datasets, but it has a high tendency to connect similar nodes (Homophily effect). Furthermore, it is more likely to make a biased prediction based on socially sensitive attributes, such as, gender or religion, making it ethically more problematic. To overcome these problems, various fairness-aware AI models and fairness metrics have been proposed. However, most of the studies used different metrics to evaluate fairness and did not consider the trade-off relationship that existed between accuracy and fairness. Thus, we propose a novel fairness metric called Fairβ-metri which takes both accuracy and prediction into consideration, and a framework called FairU that shows outstanding performance in the proposed metric.
Attack Success Rate Analysis of Adversarial Patch in Physical Environment
Hyeon-Jae Jeong, Jubin Lee, Yu Seung Ma, Seung-Ik Lee
http://doi.org/10.5626/JOK.2023.50.2.185
Adversarial patches are widely known as representative adversarial example attacks in physical environment. However, most studies on adversarial patches have demonstrated robust attack success rates based on digital environment rather than physical environment. This study investigated the robustness of adversarial patches in physical environment. To this end, 5 types of generation conditions and 3 types of attachment conditions were derived. The attack success rates of digital patches in physical environment were reviewed according to the changes in conditions. As a basic condition, location, angle, and size variables were targeted as presented in the original adversarial patch paper. Additionally, learning epoch, intent class, and neural network under simulated attack were newly considered and tested as digital patch generation conditions. As a result, the condition which greatly influenced the attack success rates of digital patches was the size. As a learning condition for digital patch generation, digital patches showed sufficient attack success rates with only one to two small learning epochs and simple intent class images. In conclusion, the attack success rate of digital patches in physical environment was not robust unlike in the digital environment.
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