• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.711-715
• /
• 2023

본 연구는 대부분의 언어 모델이 사용하고 있는 서브워드 토큰화 과정을 거치지 않고, 바이트 단위의 인코딩을 그대로 다룰 수 있는 토큰-프리 사전학습 언어모델에 대한 것이다. 토큰-프리 언어모델은 명시적인 미등록어 토큰이 존재하지 않고, 전 처리 과정이 단순하며 다양한 언어 및 표현 체계에 대응할 수 있는 장점이 있다. 하지만 관련 연구가 미흡, 서브워드 모델에 대비해 학습이 어렵고 낮은 성능이 보고되어 왔다. 본 연구에서는 한국어를 중심으로 토큰-프리 언어 이해-생성 모델을 사전 학습 후, 서브워드 기반 모델과 비교하여 가능성을 살펴본다. 또한, 토큰 프리 언어모델에서 지적되는 과도한 연산량을 감소시킬 수 있는 그래디언트 기반 서브워드 토크나이저를 적용, 처리 속도를 학습 2.7배, 추론 1.46배 개선하였다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2008.05a
• /
• pp.1104-1107
• /
• 2008

These people where generally the image or the document nik technique silver document image, against the digital data of audio back all type the research is advanced being used with objective and the use which are various, is a d. Needs a low-end leveling instrument security text from the research which it sees and with substitution quantity the silver nik being simple it will be able to deliver the technique which is simple it embodied. It combined the text image first and the nose which is in the collar image of 24 bit depth which will reach ting it did and it rehabilitatedded and a higher officer technique and the result it used that the loss ratio of the text image to analyze is slight it was ascertained.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2007.05a
• /
• pp.667-670
• /
• 2007

볼륨 데이터에 대한 압축 기법의 필요성은 데이터 용량의 증가와 네트워크 사용량의 증가와 함께 더불어 증가해 왔다. 현재에는 다양한 압축 기법이 개발돼 있으며, 사용자는 데이터 유형이나 응용 분야에 맞춰 압축 기법을 선택, 적용할 수 있다. 그러나 최근에는 응용 과학자들로부터 생성되는 데이터의 용량이 기하급수적으로 증가했는데, 이렇게 응용과학 분야에서 생성되는 데이터는 대부분 3차원 볼륨 데이터다. 2차원 이미지나 3차원 동영상 데이터의 경우에는 다양한 표준 압축 방식을 사용할 수 있지만 3차원 볼륨 데이터에 적용할 수 있는 방법은 한정돼 있으며, 특히 시간가변(time-varying) 볼륨 데이터에 대한 압축 표준은 거의 존재하지 않는다고 볼 수 있다. 본 논문에서는 시간가변 볼륨 데이터에 대한 압축 방식을 제안한다. 이 방식은 가시화를 목적으로 하는 시간가변 볼륨 데이터의 인코딩을 목적으로 하며, MPEG의 I-프레임과 P-프레임 개념을 사용해서 압축률을 높인다. 본 방식은 시간가변 부동 소수점 데이터(single precision floating-point data)로 구성된 시간가변 볼륨 데이터를 대상으로 하는데, 한 블록 단위의 무작위 복원을 지원하며 Daubechies 웨이블릿 필터와 프레임간의 상관 관계를 사용, 대형 시간가변 볼륨 데이터를 이미지 화질을 보존한다.

• Journal of IKEEE
• /
• v.28 no.3
• /
• pp.375-381
• /
• 2024

In this paper, we propose a lightweight object classification system using a LiDAR sensor for delivery service robots. The 3D point cloud data is encoded into a 2D pseudo image using a Pillar Feature Network (PFN), and then passed through a lightweight classification network designed based on Depthwise Separable Convolutional Neural Networks (DS-CNN). The implementation results show that the designed classification network has 9.08K parameters and 3.49M Multiply-Accumulate (MAC) operations, while supporting a classification accuracy of 94.94%.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.28 no.2
• /
• pp.311-326
• /
• 2018

This paper has experimented using News sample video with QCIF ($176{\times}144$) resolution in JM v10.2 code of H.264/AVC-MPEG. The region of interest (ROI) to be encrypted occurred the drift by unnecessarily referring to each frame continuously in accordance with the characteristics of the motion prediction and compensation of the H.264 standard. In order to mitigate the drift, the latest related research method of re-inserting encrypted I-picture into a certain period leads to an increase in the amount of additional computation that becomes the factor increasing the bit-rate overhead of the entire video. Therefore, the reference search range of the block and the frame in the ROI to be encrypted is restricted in the motion prediction and compensation for each frame, and the reference search range in the non-ROI not to be encrypted is not restricted to maintain the normal encoding efficiency. In this way, after encoding the video with restricted reference search range, this article proposes a method of RC4 bit-stream encryption for the ROI such as the face to be able to identify in order to protect personal information in the video. Also, it is compared and analyzed the experimental results after implementing the unencrypted original video, the latest related research method, and the proposed method in the condition of the same environment. In contrast to the latest related research method, the bit-rate overhead of the proposed method is 2.35% higher than that of the original video and 14.93% lower than that of the latest related method, while mitigating temporal drift through the proposed method. These improved results have verified by experiments of this study.

• The Journal of Korean Institute of Electromagnetic Engineering and Science
• /
• v.8 no.3
• /
• pp.221-231
• /
• 1997

This paper presents an analysis of a hybrid direct-sequence/slow frequency hopped code division multiple access(DS/SFH-CDMA) system employing noncoherent M-ary frequency shift keying(MFSK) modulation in a multiple m-distribution fading environment. Multipath interfer- ence(MPI) and multiuser interference(MUI) is taken into accout and the spectral efficiency is calculated for uncoded as well as simple channel coding systems. The predetection multipath CCI canceller in conjunction with convolution coding is employed for improving the bit error rate(BER) performance. The BER of noncoherent hybrid system is obtained using a Gaussian interference approximation. From the results, we know that the error performance more deteriorates as the depth of fading becomes deeper. The DS part of the modulation combats the multipath interference, whereas the FH part is a predetection against large multiuser interference. It is shown that, for the con- sidered types of a channel coding, the use of a predetection coding is still essential for obtained a satisfactory bit error performance. The results show that the capacity of the DS/SFG-CDMA MFSK communication system increases in proportion to the length of PN code sequence in the presence of AWGN and MUI. In m-distribution fading environment the capacity increases in proportion to the fading index. The capacity is increased and error performance is improved when the CCI Canceller and Convolution code technique are adopted, respectively. From the results, it is known that the error performance of $4\times10^{-2}$ by adopting Canceller technique. Also convolutional coding technique is the improvement of error performance attains about $10^{-5}$ in code rate 1/2.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.14 no.3
• /
• pp.21-31
• /
• 2014

In this paper, we propose VLC(visible light communication) based multi-hop multimedia data transmission system. The main contributions and features of the proposed system are as follows. First, the contribution of this research is to develope the LED communication based multi-hop transmission network system which can transmit multimedia data(audio data, video data) with long distance. Second, the developed system has the following features: In transmitter, audio data and video data are transmitted via multi-hops using two channels. The relay in audio channel receives digital audio signal by using photo diode and then transmits the signal to receiver after error checking and amplifying. The receiver receives the encoded audio data via photo diode and then converts to analog audio signal by using decoding and amplifying. The relay in video channel receives video signal by using photo diode and then amplify the video signal using OP-AMP and then transmits the signal to receiver. The receiver amplifies the received signal from photo diode and then sends it to the monitor. The performance evaluation of the proposed system is conducted in the laboratory with fluorescent light source. The results of the performance evaluation confirm that the system can provide high quality multimedia data transmission from transmiter to receiver via multi-hop relays in a long distance while we can see there are differences in the transmitted multimedia(audio and video) quality according to the used LED colors.

• Journal of the Korea Society of Computer and Information
• /
• v.15 no.3
• /
• pp.65-72
• /
• 2010

We propose the asymmetric 2.5Gbps/622Mbps PON(Passive Optical Network) in order to reduce the bandwith of filter at receiver with inverse RZ(Return to Zero) code coded downstream and NRZ(Non Return to Zero) upstream re-modulation. I theoretically analyze BER(Bit Error Rate) performance and the power sensitivity with the optimal threshold level by performing simulation with MATLAB according to the types of downstream data. The results have shown that the optimal threshold level at the optical receiver could be saturated at 0.33 as the optical received power increase more than -26dBm to keep $10^{-12}$ of BER to a minimum. Also the power sensitivity is more improved by about 3dB by fixing the threshold level at 0.33 than the conventional receiver. The proposed system can be a useful technology for optical access networks with asymmetric upstream and downstream data rates because the optical receiver can be used without controlling threshold levels and that does not require a light source in optical network unit (ONU) and its control circuits in the optical line termination (OLT).

• Journal of KIISE:Information Networking
• /
• v.37 no.5
• /
• pp.363-373
• /
• 2010

The existing QoS mechanisms for video streaming are short of the consideration for various user environments and the characteristic of streaming applying programs. In order to overwhelm this problem, studies on the video streaming protocols exploiting scalable video coding (SVC), which provide spatial, temporal, and qualitative scalability in video coding, are progressing actively. However, these protocols also have the problem to deepen network congestion situation, and to lower fairness between other traffics, as they are not equipped with congestion control mechanisms. SVC based streaming protocols also have the problem to overlook the property of videos encoded in SVC, as the protocols transmit the streaming simply by extracting the bitstream which has the maximum bit rate within available bandwidth of a network. To solve these problems, this study suggests TCP-friendly network adaptive SVC streaming(T-NASS) protocol which considers both network status and SVC bitstream property. T-NASS protocol extracts the optimal SVC bitstream by calculating TCP-friendly transmission rate, and by perceiving the network status on the basis of packet loss rate and explicit congestion notification(ECN). Through the performance estimation using an ns-2 network simulator, this study identified T-NASS protocol extracts the optimal bitstream as it uses TCP-friendly transmission property and perceives the network status, and also identified the video image quality transmitted through T-NASS protocol is improved.

• Journal of Christian Education in Korea
• /
• v.72
• /
• pp.183-203
• /
• 2022

The purpose of this study is to construct factors of AI education utilization competency. AI education utilization competency is used as basic data for education to enhance the AI education competency of pre-service early childhood teachers. To this end, 7 studies related to competency factors and models were selected by searching for previous studies. Seven preceding studies were analyzed. As a result, 18 competency factors were extracted, including understanding of artificial intelligence. The extracted competency elements were divided into six areas, which are divided into understanding subject knowledge through coding, class preparation, class management, class result feedback, class guidance, and self-development. And 15 factors were constructed. The draft formed through coding was improved through review by three early childhood education experts. Factors improved through expert review were structured by classifying them into knowledge, skills, and attitudes to organize the curriculum. The validity of the structured competency factor was verified through expert Delphi. As a result of the Delphi verification, all factors were converged in the first survey. Through this, 6 competency areas, 11 competency factors, and 19 competency factors were composed of knowledge, 10 skills, and 5 attitudes. The implication is that the competency factors presented as a result of this study can be used as basic data for organizing a curriculum to improve the ability of pre-service early childhood teachers to use artificial intelligence education.