• Journal of Korea Multimedia Society
• /
• v.6 no.5
• /
• pp.824-830
• /
• 2003

RPA (Recursive Partition Averaging) method was proposed in order to improve the storage requirement and classification rate of the Memory Based Reasoning. That algorithm worked well in many areas, however, the major drawbacks of RPA are it's pattern averaging mechanism. We propose an adaptive OHC algorithm which uses the FPD(Feature-based Population Densimeter) to increase the classification rate of RPA. The proposed algorithm required only approximately 40% of memory space that is needed in k-NN classifier, and showed a superior classification performance to the RPA. Also, by reducing the number of stored patterns, it showed a excellent results in terms of classification when we compare it to the k-NN.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.3
• /
• pp.87-94
• /
• 2024

We investigated the feasibility of utilizing Large Language Models (LLMs) to perform text-based games without training on game data in advance. We adopted ChatGPT-3.5 and its state-of-the-art, ChatGPT-4, as the systems that implemented LLM. In addition, we added the persistent memory feature proposed in this paper to ChatGPT-4 to create three game player agents. We used Zork, one of the most famous text-based games, to see if the agents could navigate through complex locations, gather information, and solve puzzles. The results showed that the agent with persistent memory had the widest range of exploration and the best score among the three agents. However, all three agents were limited in solving puzzles, indicating that LLM is vulnerable to problems that require multi-level reasoning. Nevertheless, the proposed agent was still able to visit 37.3% of the total locations and collect all the items in the locations it visited, demonstrating the potential of LLM.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.48 no.4
• /
• pp.116-128
• /
• 2011

In this paper, we propose the structure of a high-performance face-detection engine that responds well to facial rotating changes using rotation transformation which minimize the required memory usage compared to the previous face-detection engine. The validity of the proposed structure has been verified through the implementation of FPGA. For high performance face detection, the MCT (Modified Census Transform) method, which is robust against lighting change, was used. The Adaboost learning algorithm was used for creating optimized learning data. And the rotation transformation method was added to maintain effectiveness against face rotating changes. The proposed hardware structure was composed of Color Space Converter, Noise Filter, Memory Controller Interface, Image Rotator, Image Scaler, MCT(Modified Census Transform), Candidate Detector / Confidence Mapper, Position Resizer, Data Grouper, Overlay Processor / Color Overlay Processor. The face detection engine was tested using a Virtex5 LX330 FPGA board, a QVGA grade CMOS camera, and an LCD Display. It was verified that the engine demonstrated excellent performance in diverse real life environments and in a face detection standard database. As a result, a high performance real time face detection engine that can conduct real time processing at speeds of at least 60 frames per second, which is effective against lighting changes and face rotating changes and can detect 32 faces in diverse sizes simultaneously, was developed.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2023.05a
• /
• pp.92-92
• /
• 2023

본 연구에서는 대한민국 도시 유역에 대하여 딥러닝 네트워크 기반의 분산형 수문 모형을 개발하였다. 개발된 모형은 완전연결계층(Fully Connected Layer)으로 연결된 여러 개의 장단기 메모리(LSTM-Long Short-Term Memory) 은닉 유닛(Hidden Unit)으로 구성되었다. 개발된 모형을 사용하여 연구 지역인 중랑천 유역을 분석하기 위해 1km² 해상도의 239개 모델 격자 셀에서 10분 단위 레이더-지상 합성 강수량과 10분 단위 기온의 시계열을 입력으로 사용하여 10분 단위 하도 유량을 모의하였다. 모형은 보정과(2013~2016년)과 검증 기간(2017~2019년)에 대한 NSE 계수는각각 0.99와 0.67로 높은 정확도를 보였다. 본 연구는 모형을 추가적으로 심층 분석하여 다음과 같은 결론을 도출하였다: (1) 모형을 기반으로 생성된 유출-강수 비율 지도는 토지 피복 데이터에서 얻은 연구 지역의 불투수율 지도와 유사하며, 이는 모형이 수문학에 대한 선험적 정보에 의존하지 않고 입력 및 출력 데이터만으로 강우-유출 분할과정을 성공적으로 학습하였음을 의미한다. (2) 모형은 연속 수문 모형의 필수 전제 조건인 토양 수분 의존 유출 프로세스를 성공적으로 재현하였다; (3) 각 LSTM 은닉 유닛은 강수 자극에 대한 시간적 민감도가 다르며, 응답이 빠른 LSTM 은닉 유닛은 유역 출구 근처에서 더 큰 출력 가중치 계수를 가졌는데, 이는 모형이 강수 입력에 대한 직접 유출과 지하수가 주도하는 기저 흐름과 같이 응답 시간의 차이가 뚜렷한 수문순환의 구성 요소를 별도로 고려하는 메커니즘을 가지고 있음을 의미한다.

• Journal of Korea Multimedia Society
• /
• v.13 no.11
• /
• pp.1643-1656
• /
• 2010

It is necessary to develop a system to judge inferiority of goods to minimize the loss at small factories in which produces various kinds of goods with small amounts. That system can be developed based on HTM theory. HTM is a model to apply the operation principles of the neocortex in human brain to the machine learning. We have to build the trained HTM network to use the HTM-based machine learning system. It requires the knowledge for the HTM theory. This paper presents the design and implementation of the training system to support the development of HTM networks which recognize the molding parts to judge its badness. This training system allows field technicians to train the HTM network with high accuracy without the knowledge of the HTM theory. It also can be applied to any kind of the HTM-based judging systems for molding parts.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.17 no.4
• /
• pp.213-220
• /
• 2024

This paper proposes an optimization strategy for performing Federated Learning between devices and edge clouds using WebAssembly (WASM). The proposed strategy aims to maximize efficiency by conducting partial training on devices and the remaining training on edge clouds. Specifically, it mathematically describes and evaluates methods to optimize data transfer between GPU memory segments and the overlapping of computational tasks to reduce overall training time and improve GPU utilization. Through various experimental scenarios, we confirmed that asynchronous data transfer and task overlap significantly reduce training time, enhance GPU utilization, and improve model accuracy. In scenarios where all optimization techniques were applied, training time was reduced by 47%, GPU utilization improved to 91.2%, and model accuracy increased to 89.5%. These results demonstrate that asynchronous data transfer and task overlap effectively reduce GPU idle time and alleviate bottlenecks. This study is expected to contribute to the performance optimization of Federated Learning systems in the future.

• Journal of IKEEE
• /
• v.24 no.3
• /
• pp.783-790
• /
• 2020

In this paper, the human detecting radar system for indoor security applications is proposed, and its FPGA-based implementation results are presented. In order to minimize the complexity and memory requirements of the computation, the top half of the spectrogram was used to extract features, excluding the feature extraction techniques that require complex computation, feature extraction techniques were proposed considering classification performance and complexity. In addition, memory requirements were minimized by designing a pipeline structure without storing the entire spectrogram. Experiments on human, dog and robot cleaners were conducted for classification, and 96.2% accuracy performance was confirmed. The proposed system was implemented using Verilog-HDL, and we confirmed that a low-area design using 1140 logics and 6.5 Kb of memory was possible.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.20 no.1
• /
• pp.70-85
• /
• 2021

Deep learning methodology, which has been actively studied in recent years, has improved the performance of artificial intelligence. Accordingly, systems utilizing deep learning have been proposed in various industries. In traffic systems, spatio-temporal graph modeling using GNN was found to be effective in predicting traffic speed. Still, it has a disadvantage that the model is trained inefficiently due to the memory bottleneck. Therefore, in this study, the road network is clustered through the graph clustering algorithm to reduce memory bottlenecks and simultaneously achieve superior performance. In order to verify the proposed method, the similarity of road speed distribution was measured using Jensen-Shannon divergence based on the analysis result of Incheon UTIC data. Then, the road network was clustered by spectrum clustering based on the measured similarity. As a result of the experiments, it was found that when the road network was divided into seven networks, the memory bottleneck was alleviated while recording the best performance compared to the baselines with MAE of 5.52km/h.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.91-96
• /
• 2021

조기 위험 검출은 실시간으로 들어오는 텍스트를 순차적으로 처리하면서 해당 대화에 위험이 있는지 조기에 분류하는 작업으로, 정확도 저하를 최소화하는 동시에 가능한 한 빨리 대화를 분류하는 것을 목적으로 한다. 이러한, 조기 위험 검출은 온라인 그루밍 검출, 보이스 피싱 검출과 같은 다양한 영역에 활용될 수 있다. 이에, 본 논문에서는 조기 위험 검출 문제를 정의하고, 이를 평가할 수 있는 데이터 셋과 Latency F1 평가 지표를 소개한다. 또한, 점진적 문장 분류 모듈과 위험 검출 결정 모듈로 구성된 점진적 조기 텍스트 분류 시스템을 제안한다. 점진적 문장 분류 모듈은 이전 문장들에 대한 메모리 벡터와 현재 문장 벡터를 통해 현재까지의 대화를 분류한다. 위험 검출 결정 모듈은 softmax 분류 점수와 강화학습을 기반으로 하여 Read 또는 Stop 판단을 내린다. 결정 모듈이 Stop 판단을 내리면, 현재까지의 대화에 대한 분류 결과를 전체 대화의 분류 결과로 간주하고 작업을 종료한다. 해당 시스템은 micro F1과 Latency F1 지표 각각에서 0.9684와 0.8918로 높은 검출 정확성 및 검출 신속성을 달성하였다.

• Journal of the Korea Society of Computer and Information
• /
• v.17 no.10
• /
• pp.25-33
• /
• 2012

In this paper, we propose a rule-based model to correct errors in a speech recognition result in the mobile device environment. The proposed model considers the mobile device environment with limited resources such as processing time and memory, as follows. In order to minimize the error correction processing time, the proposed model removes some processing steps such as morphological analysis and the composition and decomposition of syllable. Also, the proposed model utilizes the longest match rule selection method to generate one error correction candidate per point, assumed that an error occurs. For the purpose of deploying memory resource, the proposed model uses neither the Eojeol dictionary nor the morphological analyzer, and stores a combined rule list without any classification. Considering the modification and maintenance of the proposed model, the error correction rules are automatically extracted from a training corpus. Experimental results show that the proposed model improves 5.27% on the precision and 5.60% on the recall based on Eojoel unit for the speech recognition result.