• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.281-284
• /
• 2018

본 논문에서는 machine reading 분야에서 기존의 long short-term memory (LSTM) 모델이 가지는 문제점을 해결하는 새로운 네트워크를 제안하고자 한다. 기존의 LSTM 모델은 크게 두가지 제한점을 가지는데, 그 중 첫째는 forget gate로 인해 잊혀진 중요한 문맥 정보들이 복원될 수 있는 방법이 없다는 것이다. 자연어에서 과거의 문맥 정보에 따라 현재의 단어의 의미가 크게 좌지우지될 수 있으므로 올바른 문장의 이해를 위해 필요한 과거 문맥의 정보 유지는 필수적이다. 또 다른 문제는 자연어는 그 자체로 단어들 간의 복잡한 구조를 통해 문장이 이루어지는 반면 기존의 시계열 모델들은 단어들 간의 관계를 추론할 수 있는 직접적인 방법을 가지고 있지 않다는 것이다. 본 논문에서는 최근 딥 러닝 분야에서 널리 쓰이는 attention mechanism과 본 논문이 제안하는 restore gate를 결합한 네트워크를 통해 상기 문제를 해결하고자 한다. 본 논문의 실험에서는 기존의 다른 시계열 모델들과 비교를 통해 제안한 모델의 우수성을 확인하였다.

• 제어로봇시스템학회:학술대회논문집
• /
• 1987.10a
• /
• pp.770-775
• /
• 1987

In this paper, design and development experiences of a vision based robotic assembly system for electronic components are described. Specifically, the overall system consists of the following three subsystems each of which employs a 16 bit Preprocessor MC 68000 : supervisory controller, real-time vision system, and servo system. The three microprocessors are interconnected using the time shared common memory bus structure with hardwired bus arbitration scheme and operated as a master-slave type in which each slave is functionally fixed in view of software. With this system architecture, the followings are developed and implemented in this research; (i) the system programming language, called 'CLRC', for man-machine interface including the robot motion and vision primitives, (ii) real-time vision system using hardwired chain coder, (iii) the high-precision servo techniques for high speed de motors and high speed stepping motors. The proposed control system were implemented and tested in real-time successfully.

• Proceedings of the IEEK Conference
• /
• 2003.07b
• /
• pp.927-930
• /
• 2003

In this paper, a design of processor IP for TCP/IP protocol stack is described. The processor consists of input and output buffer memory with dual bank structure, 32-bit RISC microprocessor core, DMA unit with on-the-fly checksum capability. To handle the various modes of TCP/IP protocol, hardware and software co-design approach is used rather than the conventional state machine based design. To eliminate delay time due to the data transfer and checksum operation, DAM module which can execute the checksum operation on-the-fly along with data transfer operation is adopted. By programming the on-chip code ROM of RISC processor differently. the designed stack processor can support the packet format conversion operations required in the various TCP/IP protocols.

• Transactions of the Korean Society of Mechanical Engineers
• /
• v.19 no.8
• /
• pp.1832-1843
• /
• 1995

Tool interference is one of the most critical problems in sculptured surface machining. When machining cavities and concaves, the tool frequently overcuts the portions of the surface, which cause inaccuracy in machining. So tool interference-free paths must be generated for rough cutting more efficiently. In this paper a software using SSI(Surface/Surface intersection) algorithm is developed for eliminating tool interference which occurs in an offset surface in 3-dimensional free form surface modeling. this work consists of two stages : using the offset data, the intersection curves are rapidly checked by this algorithm at the first stage. CL(cutter location) data are obtained by deleting the loop section of intersected offset patches at the second stage. This algorithm can reduce the amount of memory required to store machining data and also easily check region which have the possibility of intersection. Also, This software is verified to be useful in machining a curved object on a DNC milling machine.

• Proceedings of the Korean Society of Machine Tool Engineers Conference
• /
• 1998.10a
• /
• pp.267.1-270
• /
• 1998

We measured the dielectric properties with YMnO$_3$ ceramics using solution method based procedure via by citrate. The crystalline phases were determined using XRD. Also we observed morphologies of YMnO$_3$ ceramics using SEM. We proved the structure of YMnO$_3$ ceramics which is hexagonal. But lots of pores were observed the microstructure. It would be considered as volatile organic. The maximum density of YMn03 ceramics is obtained sintering temperature 135$0^{\circ}C$ and the ratio 0.95/1.05 of Y/Mn. But even though the density we obtained is the highest, that is lower than theoretical density because of remaining organics by citric acid. Dielectric constant and dissipation factor were improved as increasing sintering temperature and the Y/Mn ratio (0.95/1.05)

• Proceedings of the Korean Information Science Society Conference
• /
• 2002.10d
• /
• pp.643-645
• /
• 2002

자바가상기계는 클래스 파일에 있는 클래스, 상수, 필드, 메소드 등의 내부 정보를 읽고 자바 응용 프로그램을 실행한다. 보조기억장치가 없는 내장형 시스템의 경우 클래스 파일은 주기억장치, 즉 메모리에 두어야 하지만 클래스 파일 자체는 크기도 클 뿐 아니라 내부 정보에 접근하는 것도 효율적이지 못하다. 따라서 대개의 경우 클래스 파일을 변형한 형태로 메모리에 배치하는데, 본 논문에서는 특히 simpleRTJ 라고 하는 상용 내장형 자바가상기계에서 적용된 방식에 대해 조사해보았다. 이 플렛폼에서의 분석을 통해 클래스 파일의 크기가 얼마까지 줄어들 수 있고, 내부 정보에 대해서는 얼마나 효율적으로 접근할 수 있는지에 대해 고찰하였으며, 그 결과를 바탕으로 향후 더 개선된 형태로 클래스 파일을 메모리에 배치할 수 있는 방안에 대해 연구하고자 한다.

• Proceedings of the Korean Information Science Society Conference
• /
• 2004.04a
• /
• pp.805-807
• /
• 2004

최근 IT 산업의 발전과 더불어, 리소스가 제한된 소형 기기들의 사용이 비약적으로 증가하고 있는 추세이다. 자바는 플랫폼 독립성(Platform Independency), 보안성(Security), 네트워크 이동성(Network Mobility) 등의 장점을 가지고 있어, 이러한 소형 기기들에 자바 환경을 적용하게 되면 여러 가지 이점을 가지게 된다. 임베디드 장치나 모바일 같은 제한된 리소스를 사용하는 기기들에는 SUN 사의 CLDC(Connected, Limited Device Configuration)에서 정의하고 있는 K 가상 머신(K Virtual Machine: KVM)을 탑재하여 사용하게 된다. 본 논문에서는 실시간 운영체제 iRTOS$^{TM}$와 KVM 을 탑재한 소형 기기에서 좀더 효율적으로 KVM 의 메모리를 관리하기 위한 Garbage Collection기법을 설계하고 구현한 내용을 설명한다.

• ETRI Journal
• /
• v.39 no.6
• /
• pp.832-840
• /
• 2017

Recently, deep recurrent neural networks have achieved great success in various machine learning tasks, and have also been applied for sound event detection. The detection of temporally overlapping sound events in realistic environments is much more challenging than in monophonic detection problems. In this paper, we present an approach to improve the accuracy of polyphonic sound event detection in multichannel audio based on gated recurrent neural networks in combination with auditory spectral features. In the proposed method, human hearing perception-based spatial and spectral-domain noise-reduced harmonic features are extracted from multichannel audio and used as high-resolution spectral inputs to train gated recurrent neural networks. This provides a fast and stable convergence rate compared to long short-term memory recurrent neural networks. Our evaluation reveals that the proposed method outperforms the conventional approaches.

• Journal of the Korean Data and Information Science Society
• /
• v.26 no.6
• /
• pp.1537-1545
• /
• 2015

Support vector quantile regression (SVQR) can be obtained by applying support vector machine with a check function instead of an e-insensitive loss function into the quantile regression, which still requires to solve a quadratic program (QP) problem which is time and memory expensive. In this paper we propose an SVQR whose objective function is composed of an asymmetric quadratic loss function. The proposed method overcomes the weak point of the SVQR with the check function. We use the iterative procedure to solve the objective problem. Furthermore, we introduce the generalized cross validation function to select the hyper-parameters which affect the performance of SVQR. Experimental results are then presented, which illustrate the performance of proposed SVQR.

• ETRI Journal
• /
• v.33 no.4
• /
• pp.652-655
• /
• 2011

In this letter, we propose a new compression method for a high dimensional support vector machine (SVM). We used singular value decomposition (SVD) to compress the norm part of a radial basis function SVM. By deleting the least significant vectors that are extracted from the decomposition, we can compress each vector with minimized energy loss. We select the compressed vector dimension according to the predefined threshold which can limit the energy loss to design criteria. We verified the proposed vector compressed SVM (VCSVM) for conventional datasets. Experimental results show that VCSVM can reduce computational complexity and memory by more than 40% without reduction in accuracy when classifying a 20,958 dimension dataset.