• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.37-41
• /
• 2020

본 논문에서는 Joint CTC/Attention 모델에 CTC ratio scheduling을 이용한 end-to-end 한국어 음성인식을 연구하였다. Joint CTC/Attention은 CTC와 attention의 장점을 결합한 모델로서 attention, CTC 단일 모델보다 좋은 성능을 보여주지만, 학습이 진행될수록 CTC가 attention의 학습을 저해하는 요인이 된다. 본 논문에서는 이러한 문제를 해결하기 위해, 학습 진행에 따라 CTC의 비율(ratio)를 줄여나가는 CTC ratio scheduling 방법을 제안한다. CTC ratio scheduling를 이용하여 학습한 결과물은 기존 Joint CTC/Attention, 단일 attention 모델 대비 좋은 성능을 보여주는 것을 확인하였다.

• 한국어정보학회:학술대회논문집
• /
• 2016.10a
• /
• pp.20-24
• /
• 2016

Sequence-to-sequence 모델은 입력열을 길이가 다른 출력열로 변환하는 모델로, 단일 신경망 구조만을 사용하는 End-to-end 방식의 모델이다. 본 논문에서는 Sequence-to-sequence 모델을 한국어 구구조 구문 분석에 적용한다. 이를 위해 구구조 구문 트리를 괄호와 구문 태그 및 어절로 이루어진 출력열의 형태로 만들고 어절들을 단일 기호 'XX'로 치환하여 출력 단어 사전의 수를 줄였다. 그리고 최근 기계번역의 성능을 높이기 위해 연구된 Attention mechanism과 Input-feeding을 적용하였다. 실험 결과, 세종말뭉치의 구구조 구문 분석 데이터에 대해 기존의 연구보다 높은 F1 89.03%의 성능을 보였다.

• Annual Conference on Human and Language Technology
• /
• 2016.10a
• /
• pp.20-24
• /
• 2016

Sequence-to-sequence 모델은 입력열을 길이가 다른 출력열로 변환하는 모델로, 단일 신경망 구조만을 사용하는 End-to-end 방식의 모델이다. 본 논문에서는 Sequence-to-sequence 모델을 한국어 구구조 구문 분석에 적용한다. 이를 위해 구구조 구문 트리를 괄호와 구문 태그 및 어절로 이루어진 출력열의 형태로 만들고 어절들을 단일 기호 'XX'로 치환하여 출력 단어 사전의 수를 줄였다. 그리고 최근 기계번역의 성능을 높이기 위해 연구된 Attention mechanism과 Input-feeding을 적용하였다. 실험 결과, 세종말뭉치의 구구조 구문 분석 데이터에 대해 기존의 연구보다 높은 F1 89.03%의 성능을 보였다.

• International journal of advanced smart convergence
• /
• v.7 no.4
• /
• pp.19-26
• /
• 2018

One of the distinct features of Wireless Sensor Networks (WSNs) is duty cycling mechanism, which is used to conserve energy and extend the network lifetime. Large duty cycle interval introduces lower energy consumption, meanwhile longer end-to-end (E2E) delay. In this paper, we introduce an energy consumption minimization problem for duty-cycled WSNs. We have applied Q-learning algorithm to obtain the maximum duty cycle interval which supports various delay requirements and given Delay Success ratio (DSR) i.e. the required probability of packets arriving at the sink before given delay bound. Our approach only requires sink to compute Q-leaning which makes it practical to implement. Nodes in the different group have the different duty cycle interval in our proposed method and nodes don't need to know the information of the neighboring node. Performance metrics show that our proposed scheme outperforms existing algorithms in terms of energy efficiency while assuring the required delay bound and DSR.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.5
• /
• pp.193-204
• /
• 2019

The explosion of data due to the improvement of sensor technology and computing performance has become the basis for analyzing the situation in the industrial fields, and various attempts to detect events based on such data are increasing recently. In particular, sound signals collected from sensors are used as important information to classify events in various application fields as an advantage of efficiently collecting field information at a relatively low cost. However, the performance of sound-event classification in the field cannot be guaranteed if noise can not be removed. That is, in order to implement a system that can be practically applied, robust performance should be guaranteed even in various noise conditions. In this study, we propose a system that can classify the sound event after generating the enhanced sound signal based on the deep learning algorithm. Especially, to remove noise from the sound signal itself, the enhanced sound data against the noise is generated using SEGAN applied to the GAN with a VAE technique. Then, an end-to-end based sound-event classification system is designed to classify the sound events using the enhanced sound signal as input data of CNN structure without a data conversion process. The performance of the proposed method was verified experimentally using sound data obtained from the industrial field, and the f1 score of 99.29% (railway industry) and 97.80% (livestock industry) was confirmed.

• International Journal of Control, Automation, and Systems
• /
• v.6 no.4
• /
• pp.526-533
• /
• 2008

We propose a wavelet neural network (WNN) control method for active queue management (AQM) in an end-to-end TCP network, which is trained by adaptive learning rates (ALRs). In the TCP network, AQM is important to regulate the queue length by passing or dropping the packets at the intermediate routers. RED, PI, and PID algorithms have been used for AQM. But these algorithms show weaknesses in the detection and control of congestion under dynamically changing network situations. In our method, the WNN controller using ALRs is designed to overcome these problems. It adaptively controls the dropping probability of the packets and is trained by gradient-descent algorithm. We apply Lyapunov theorem to verify the stability of the WNN controller using ALRs. Simulations are carried out to demonstrate the effectiveness of the proposed method.

• Journal of Korea Multimedia Society
• /
• v.24 no.10
• /
• pp.1319-1325
• /
• 2021

It is important to remove fog for accurate object recognition and detection during preprocessing because images taken in foggy adverse weather suffer from poor quality of images due to scattering and absorption of light, resulting in poor performance of various vision-based applications. This paper proposes an end-to-end deep learning-based single image de-fogging method using U-Net architecture. The loss function used in the algorithm is a loss function based on Mahalanobis distance with fog features, which solves the problem of domain shifts, and demonstrates superior performance by comparing qualitative and quantitative numerical evaluations with conventional methods. We also design it to generate fog through the VGG19 loss function and use it as the next training dataset.

• Journal of Advanced Navigation Technology
• /
• v.24 no.2
• /
• pp.155-162
• /
• 2020

Recently, with the development of computer vision and deep learning technology, research on human action recognition has been actively conducted for video analysis, video surveillance, interactive multimedia, and human machine interaction applications. Diverse techniques have been introduced for human action understanding and classification by many researchers using RGB image, depth image, skeleton and inertial data. However, skeleton-based action discrimination is still a challenging research topic for human machine-interaction. In this paper, we propose an end-to-end skeleton joints mapping of action for generating spatio-temporal image so-called dynamic image. Then, an efficient deep convolution neural network is devised to perform the classification among the action classes. We use publicly accessible UTD-MHAD skeleton dataset for evaluating the performance of the proposed method. As a result of the experiment, the proposed system shows better performance than the existing methods with high accuracy of 97.45%.

• Journal of Korea Water Resources Association
• /
• v.54 no.10
• /
• pp.759-768
• /
• 2021

This work introduces a new approach that classifies individual household water usage by examining the characteristics of smart meter end-user demand data. Here, one of the most well-known unsupervised machine learning, K-means algorithm, is applied to classify water consumptions by each household. The intensity and duration of end-user demands are used as main features to determine the households with similar water consumption pattern. The results showed that 21 households are classified into 13 clusters with each cluster having one, two, three, or five houses. The reasoning why multiple households are classified into the same cluster is described in this paper with respect to the collected data and end-user water consumption behavior.

• Journal of the Korean Society of Manufacturing Technology Engineers
• /
• v.22 no.5
• /
• pp.837-844
• /
• 2013

This study performs erection of a pendulum hanging at a free end of an arm by rotating the arm to the upright position. A mathematical model of a rotary inverted pendulum system (RIPS) is derived. A brain emotional learning based intelligent controller (BELBIC) is designed and used as a controller for swinging up and balancing the pendulum of the RIPS. In simulations performed in the study, a pendulum is initially inclined at $45^{\circ}$ with respect to the upright position. A simulation is also performed for evaluating the adaptiveness of the designed BELBIC in the case of system variation. In addition, a simulation is performed for evaluating the robustness of the designed BELBIC against a disturbance in the control input.