• Journal of Internet Computing and Services
• /
• v.21 no.3
• /
• pp.123-131
• /
• 2020

Along with the rapid development of artificial intelligence technology, natural language processing, which deals with human language, is also actively studied. In particular, BERT, a language model recently proposed by Google, has been performing well in many areas of natural language processing by providing pre-trained model using a large number of corpus. Although BERT supports multilingual model, we should use the pre-trained model using large amounts of Korean corpus because there are limitations when we apply the original pre-trained BERT model directly to Korean. Also, text contains not only vocabulary, grammar, but contextual meanings such as the relation between the front and the rear, and situation. In the existing natural language processing field, research has been conducted mainly on vocabulary or grammatical meaning. Accurate identification of contextual information embedded in text plays an important role in understanding context. Knowledge graphs, which are linked using the relationship of words, have the advantage of being able to learn context easily from computer. In this paper, we propose a system to extract Korean contextual information using pre-trained BERT model with Korean language corpus and knowledge graph. We build models that can extract person, relationship, emotion, space, and time information that is important in the text and validate the proposed system through experiments.

• The KIPS Transactions:PartD
• /
• v.10D no.3
• /
• pp.447-458
• /
• 2003

The basic assumption of conventional data mining methodology is that the data set of a knowledge discovery process should be fixed and available before the process can proceed. Consequently, this assumption is valid only when the static knowledge embedded in a specific data set is the target of data mining. In addition, a conventional data mining method requires considerable computing time to produce the result of mining from a large data set. Due to these reasons, it is almost impossible to apply the mining method to a realtime analysis task in a data stream where a new transaction is continuously generated and the up-to-dated result of data mining including the newly generated transaction is needed as quickly as possible. In this paper, a new mining concept, open data mining in a data stream, is proposed for this purpose. In open data mining, whenever each transaction is newly generated, the updated mining result of whole transactions including the newly generated transactions is obtained instantly. In order to implement this mechanism efficiently, it is necessary to incorporate the delayed-insertion of newly identified information in recent transactions as well as the pruning of insignificant information in the mining result of past transactions. The proposed algorithm is analyzed through a series of experiments in order to identify the various characteristics of the proposed algorithm.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• v.40 no.9
• /
• pp.597-604
• /
• 2016

A parallel algorithm of bi-conjugate gradient method was developed based on CUDA for parallel computation of the incompressible Navier-Stokes equations. The governing equations were discretized using splitting P2P1 finite element method. Asymmetric stenotic flow problem was solved to validate the proposed algorithm, and then the parallel performance of the GPU was examined by measuring the elapsed times. Further, the GPU performance for sparse matrix-vector multiplication was also investigated with a matrix of fluid-structure interaction problem. A kernel was generated to simultaneously compute the inner product of each row of sparse matrix and a vector. In addition, the kernel was optimized to improve the performance by using both parallel reduction and memory coalescing. In the kernel construction, the effect of warp on the parallel performance of the present CUDA was also examined. The present GPU computation was more than 7 times faster than the single CPU by double precision.

• Journal of KIISE:Computing Practices and Letters
• /
• v.9 no.3
• /
• pp.341-351
• /
• 2003

Digital image watermarking is the method that can protect the copyright of the image by embedding copyright information, which is called watermark. Watermarking must have robustness to intentional or unintentional data changing, called attack. The conventional watermarking schemes are robust to waveform attacks such as image compression, filtering etc. However, they are vulnerable to geometrical attacks such as rotation, scaling, translation, and cropping. Accordingly, this paper proposes new watermarking scheme that is robust to geometrical attacks by using invariant centroid. Invariant centroid is the gravity center of a central area in a gray scale image that remains unchanged even when the image is attacked by RST including cropping and proposed scheme uses invariant centroids of original and inverted image as the template. To make geometrically invariant domain, template and angle compensated Log -Polar Map(LPM) is used. Then Discrete Cosine Transform(DCT) is performed and the watermark is embedded into the DCT coefficients. Futhermore, to prevent a watermarked image from degrading due to interpolation during coordinate system conversion, only the image of the watermark signal is extracted and added to the original image. Experimental results show that the proposed scheme is especially robust to RST attacks including cropping.

• The Journal of the Korea Contents Association
• /
• v.7 no.8
• /
• pp.31-40
• /
• 2007

Recently, Embedded systems are different with the past as loading of a simple application program that executed specific functions according to the use and are evolved in the digital convergence integrated multimedia functions and then the complication of the application program is remarkably increased. This application program is combined and evolved with many application program in accordance with the demand of the age. For develope and manage this developing application is necessary standardized interface between developer and manager. POSIX was developed as the standard of the operating system in the standard interface which has the open system structure in computing system, and there is a posix.4 to standard for the system demands the loading of real-time operating system like a digital convergence devices. In this paper, we present the contents of designing and implementing the real-time operating system UbiFOSTM to wrapper for supporting the POSIX.4. Also, Experimental results show that implemented wrapper to application program standardizing POSIX.4 in linux and UbiFOSTM is slight only $3{\sim}9{\mu}s$.

• Journal of the Korea Convergence Society
• /
• v.9 no.5
• /
• pp.33-41
• /
• 2018

The purpose of this study is to develop more accurate and robust emotion prediction neural network (EPNN) model by combining heart rate variability (HRV) and neural network. For the sake of improving the prediction performance more reliably, the proposed EPNN model is based on various types of activation functions like hyperbolic tangent, linear, and Gaussian functions, all of which are embedded in hidden nodes to improve its performance. In order to verify the validity of the proposed EPNN model, a number of HRV metrics were calculated from 20 valid and qualified participants whose emotions were induced by using money game. To add more rigor to the experiment, the participants' valence and arousal were checked and used as output node of the EPNN. The experiment results reveal that the F-Measure for Valence and Arousal is 80% and 95%, respectively, proving that the EPNN yields very robust and well-balanced performance. The EPNN performance was compared with competing models like neural network, logistic regression, support vector machine, and random forest. The EPNN was more accurate and reliable than those of the competing models. The results of this study can be effectively applied to many types of wearable computing devices when ubiquitous digital health environment becomes feasible and permeating into our everyday lives.

• The KIPS Transactions:PartA
• /
• v.12A no.5 s.95
• /
• pp.451-460
• /
• 2005

We present a methodology to design energy-efficient data paths using FPGAs. Our methodology integrates domain specific modeling, coarse-grained performance evaluation, design space exploration, and low-level simulation to understand the tradeoffs between energy, latency, and area. The domain specific modeling technique defines a high-level model by identifying various components and parameters specific to a domain that affect the system-wide energy dissipation. A domain is a family of architectures and corresponding algorithms for a given application kernel. The high-level model also consists of functions for estimating energy, latency, and area that facilitate tradeoff analysis. Design space exploration(DSE) analyzes the design space defined by the domain and selects a set of designs. Low-level simulations are used for accurate performance estimation for the designs selected by the DSE and also for final design selection We illustrate our methodology using a family of architectures and algorithms for matrix multiplication. The designs identified by our methodology demonstrate tradeoffs among energy, latency, and area. We compare our designs with a vendor specified matrix multiplication kernel to demonstrate the effectiveness of our methodology. To illustrate the effectiveness of our methodology, we used average power density(E/AT), energy/(area x latency), as themetric for comparison. For various problem sizes, designs obtained using our methodology are on average $25\%$ superior with respect to the E/AT performance metric, compared with the state-of-the-art designs by Xilinx. We also discuss the implementation of our methodology using the MILAN framework.

• Journal of KIISE:Computer Systems and Theory
• /
• v.35 no.6
• /
• pp.293-303
• /
• 2008

As the computing environment moves to the wireless and handheld system, the power efficiency is getting more important. That is the case especially in the embedded hand-held system and the power consumed by the memory system takes the second largest portion in overall. To save energy consumed in the memory system we can utilize low power mode of SDRAM. In the case of RDRAM, nap mode consumes less than 5% of the power consumed in active or standby mode. However hardware controller itself can't use this facility efficiently unless the operating system cooperates. In this paper we focus on how to minimize the number of active units of SDRAM. The operating system allocates its physical pages so that only a few units of SDRAM need to be activated and the unnecessary SDRAM can be put into nap mode. This work can be considered as a generalized and system-wide version of PAVM(Power-Aware Virtual Memory) research. We take all the physical memory into account, especially buffer cache, which takes an half of total memory usage on average. Because of the portion of buffer cache and its importance, PAVM approach cannot be robust without taking the buffer cache into account. In this paper, we analyze the RAM usage and propose power-aware page allocation policy. Especially the pages mapped into the process' address space and the buffer cache pages are considered. The relationship and interactions of these two kinds of pages are analyzed and exploited for energy saving.

• Journal of Internet Computing and Services
• /
• v.8 no.1
• /
• pp.33-45
• /
• 2007

Using switches and sensors informing the current on or off state, this paper suggests a sleeping control and remote monitoring system that not only can recognize the sleeping situations but also can control for keeping an appropriate sleeping situation remotely, And we show an example that this system is applied to the healthcare sleeping mat, Our system comprises the following 3 parts: a part for detecting the sleeping situations, a part for extracting sensing data and sending/receiving the relating situated data, and a part controlling and monitoring the all of sleeping situations. In details, in order to develop our system, we used the touch and pressure-sensitive sensors with On/Off functions for a purpose of the first part, The second part consists of the self-developed embedded board with the socket based communication as well as extracting real-time sensing data. And the third part is implemented by service modules for providing controlling and monitoring functions previously described. Finally, these service modules are implemented by the TMO scheme, one of real-time object-oriented programming models and the communications among them is supported using the TMOSM of distributed real-time middleware.

• Journal of Internet Computing and Services
• /
• v.10 no.5
• /
• pp.65-78
• /
• 2009

The most commonly adopted approach to find valuable information from tree data is to extract frequently occurring subtree patterns from them. Because mining frequent tree patterns has a wide range of applications such as xml mining, web usage mining, bioinformatics, and network multicast routing, many algorithms have been recently proposed to find the patterns. However, existing tree mining algorithms suffer from several serious pitfalls in finding frequent tree patterns from massive tree datasets. Some of the major problems are due to (1) modeling data as hierarchical tree structure, (2) the computationally high cost of the candidate maintenance, (3) the repetitious input dataset scans, and (4) the high memory dependency. These problems stem from that most of these algorithms are based on the well-known apriori algorithm and have used anti-monotone property for candidate generation and frequency counting in their algorithms. To solve the problems, we base a pattern-growth approach rather than the apriori approach, and choose to extract maximal frequent subtree patterns instead of frequent subtree patterns. The proposed method not only gets rid of the process for infrequent subtrees pruning, but also totally eliminates the problem of generating candidate subtrees. Hence, it significantly improves the whole mining process.