• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.622-624
• /
• 2021

현재 BERT와 같은 대용량의 코퍼스로부터 학습된 사전 학습 언어 모델을 자연어 응용 태스크에 적용하기 위해 일반적으로 널리 사용되는 방법은 Fine-tuning으로 각 응용 태스크에 적용 시 모델의 모든 파라미터를 조정하기 때문에 모든 파라미터를 조정하는데 필요한 시간적 비용과 함께 업데이트된 파라미터를 저장하기 위한 별도의 저장공간이 요구된다. 언어 모델이 커지면 커질수록 저장 공간의 비용이 증대됨에 따라 이러한 언어모델을 효율적으로 튜닝 할 수 있는 방법들이 연구되었다. 본 연구에서는 문장의 입력 임베딩에 연속적 태스크 특화 벡터인 prefix를 추가하여 해당 prefix와 관련된 파라미터만 튜닝하는 prefix-tuning을 한국어 네이버 감성 분석 데이터 셋에 적용 후 실험결과를 보인다.

• The KIPS Transactions:PartA
• /
• v.13A no.3 s.100
• /
• pp.241-252
• /
• 2006

In cluster file systems, the availability of files has been supported by replicating entire files or generating parities on parity servers. However, those methods require very large temporal and spatial cost, and cannot handle massive failures situation on the file system. So we propose HA-PVFS, a cluster file system supporting high data availability adaptive to temporal locality. HA-PVFS restricts replication or parity generation to some important files, for that it employs an efficient algorithm to estimate file access patterns from limited information. Moreover, in order to minimize the performance degradation of the file system, it uses delayed update method and relay replication.

• Journal of Korean Society for Geospatial Information Science
• /
• v.21 no.1
• /
• pp.45-52
• /
• 2013

EIASS(Environment Impact Assessment Supporting System) provided the various updated quality environment impact assessment data and the assessment reports through database system to support cost effective and stable services. Through this system users can access environment impact assessment information and have great convenience in their relevant works. This study introduces geo-spatial information usages in EIASS and how this system developed the use of convenience and expansion of service ranges by user surveys. This study also demonstrated step-wise delivery and links on one-stop database inventory for data collection, manipulation, and usages of geo-spatial information among environmental information.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2019.05a
• /
• pp.30-30
• /
• 2019

본 연구는 지점별 연평균 강우침식인자 값을 2018년 호우사상을 포함한 최신자료로 업데이트하기 위한 것이다. 이를 위하여 기상자료개방포털의 QC 테스트를 통과한 강수량 자료를 2018년 자료까지 수집하여 입력데이터로 활용하였다. 본 연구에서는 국립농업과학원에서 개발한 강우인자 계산기를 이용하여 지점별 연평균 강우침식인자 값을 재산정한 뒤, 기존 연구의 결과값과의 차이점을 비교 분석하였다. 산정된 지점별 강우침식인자 값을 바탕으로 강우침식인자의 공간분포 지도인 등강우침식도를 작성하였으며, 기존에 학계에 보고된 선행연구 결과의 등강우침식도와 비교하여, 강우침식인자의 지역별 분포의 최근 변화 양상을 분석하여 제시하였다.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2017.05a
• /
• pp.476-476
• /
• 2017

최근 갑작스런 폭우로 인한 제방 붕괴, 침수 및 지진 등과 같은 재해 발생 시 추가 피해를 방지하고 주민들의 긴급대피를 도운 건 SNS를 통한 현장 정보와 경보 메시지의 지속적인 전파이다. 최근의 SNS는 재난정보에서도 활용할 수 있을 정도로 진화하였다. 국가재난정보 중 수재해 관련 정보를 추출하여 다양한 주제도위에 중첩으로 공간정보를 제공할 수 있는 재난정보 제공을 위한 웹서비스를 개발하고자 하였다. 수재해 정보를 필터링하기 위하여 우선 관련된 키워드 선정이 필요하며, 기본적인 키워드는 하천일람표를 참고하여 6개 권역 및 하천이름을 선정하였다. 또한, 한강 홍수 통제소의 수자원 용어사전과 (사)한국물학술단체연합회에서 발간한 물용어집을 참고하여 수재해 관련 용어들 약 300여개를 추가하였다. 선정된 용어들은 1차적으로 적재된 데이터베이스에서 수재해 정보 관련 필터링을 하는데 사용되며, 비정형 데이터들을 필터링하고 주소 정보 검색 및 추출을 통하여 정형화 하게 된다. 추출된 주소정보에 대하여 개발한 지오코딩 모듈을 적용하여 수재해 항목에 대해 좌표정보를 업데이트 하게 된다. 가뭄, 집중호우, 홍수 등의 수재해 정보별, 또한 일자별 그룹화 및 구조화를 진행하고 해당되는 정보를 공간정보 오픈플랫폼 API를 활용하여 지도상에 가시화할 수 있다. 개발한 지오코딩 모듈을 이용하여 실제 테이블 정보를 구성하여 데이터베이스에 수재해 정보 지오코딩 테이블을 구성하여 테스트 모의하였다. 재난정보 중 홍수, 가뭄에 대한 선택정보와 시간정보를 매개변수로 받는 XML 웹서비스 테스트로 검증을 하였다. 본 연구를 통하여 재난정보 가시화에 있어서 사용자가 조회하고자 하는 유형별, 날짜별 선택이 가능한 공간적 정보를 검색 및 확인할 수 있게 되었다. 개발한 수재해 정보 지오코딩 모듈 프로토 타입은 수재해 정보 플랫폼 융합기술 연구단에서 개발하는 핵심 목표시스템 내 재난정보 제공시스템에 적용 가능하며, 수재해 정보에 대하여 대국민 서비스가 가능할 것으로 사료된다.

• Journal of IKEEE
• /
• v.23 no.3
• /
• pp.805-810
• /
• 2019

We present a time-frequency analysis algorithm based on the multitaper method and the state-space frameworks. In general, time-frequency representations have a trade-off between the time duration and the spectral bandwidth by the uncertainty principle. To optimize the trade-off problems, the short-time Fourier transform and wavelet based algorithms have been developed. Alternatively, the authors proposed the state-space frameworks based on the multitaper method in the previous work. In this paper, we develop a real-time algorithm to estimate variances and spectrum using the state-space framework. We test our algorithm in spectral analysis of simulated data.

• Journal of Korea Game Society
• /
• v.10 no.6
• /
• pp.157-168
• /
• 2010

A terrain is an essential element for constructing a virtual world in which game characters and objects make various interactions with one another. Creating a terrain requires a great deal of time and repetitive editing processes. This paper presents a 3D terrain reconstruction system to create 3D terrain in virtual space based on real terrain data. In this system, it converts the coordinate system of the height maps which are generated from a stereo camera and a laser scanner from global GPS into 3D world using the x and z axis vectors of the global GPS coordinate system. It calculates the movement vectors and the rotation matrices frame by frame. Terrain meshes are dynamically generated and rendered in the virtual areas which are represented in an undirected graph. The rendering meshes are exactly created and updated by correcting terrain data errors. In our experiments, the FPS of the system was regularly checked until the terrain was reconstructed by our system, and the visualization quality of the terrain was reviewed. As a result, our system shows that it has 3 times higher FPS than other terrain management systems with Quadtree for small area, improves 40% than others for large area. The visualization of terrain data maintains the same shape as the contour of real terrain. This system could be used for the terrain system of realtime 3D games to generate terrain on real time, and for the terrain design work of CG Movies.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.2
• /
• pp.1-8
• /
• 2022

In this paper, we design and implement a GPU-based parallel algorithm to effectively solve the traveling salesman problem through an ant color system. The repetition process of generating hundreds or thousands of tours simultaneously in TSP utilizes GPU's task-level parallelism, and the update process of pheromone trails data actively exploits data parallelism by 32x32 thread blocks. In particular, through simultaneous memory access of multiple threads, the coalesced accesses on continuous memory addresses and concurrent accesses on shared memory are supported. This experiment used 127 to 1002 city data provided by TSPLIB, and compared the performance of sequential and parallel algorithms by using Intel Core i9-9900K CPU and Nvidia Titan RTX system. Performance improvement by GPU parallelization shows speedup of about 10.13 to 11.37 times.

• Journal of the Korean Association of Geographic Information Studies
• /
• v.25 no.3
• /
• pp.17-28
• /
• 2022

This study is to develop the direction of the development of the next-generation mapping of marine information required to develop a base of the utilization localization of maritime production tools. The GIS data-processing products and technologies currently used in the Korea's marine sector depend on external applications which is renewal costs, technical updates, and unreflected characteristics. Meanwhile, the S-100 standard, the next generation hydrographic data model that complements S-57's problems in marine GIS data processing, was adopted as a new marine data standard. This study aims to present the current status and problems of marine GIS technology in Korea and to suggest the development direction of GIS software based on the next generation hydrogrphic data model S-100 standard of IHO(International Hydrographic Organization). S-100-based marine GIS localization technology development and industrial ecosystem development research is expected to scientific decision-making on policy issues that occur with other countries such as marine territory management and development and use of marine resources.

• Journal of KIISE:Databases
• /
• v.37 no.5
• /
• pp.238-249
• /
• 2010

Flash memory has been studied as a storage medium in order to improve the performance of the system using its high computing speed in the DBMS field where frequent data access is needed. The most difficulty using the flash memory is the performance degradation and the life span shortening of flash memory coming from inefficient in-place update. Log based approaches have been studied to solve inefficient in-place update problem in the DBMS where write operations occur in smaller size of data than page frequently. However the existing log based approaches suffer from the frequent merging operations, which are the principal cause of performance deterioration. Thus is because their fixed log area management can not guarantee a sufficient space for logs. In this paper, we propose non-fixed log area management technique that can minimize the occurrence of the merging operations by promising an enough space for logs. We also suggest the cost calculation model of the optimal log sector number minimizing the system operation cost in a block. In experiment, we show that our non-fixed log area management technique can have the improved performance compared to existing approaches.