• Journal of the Korea Society of Computer and Information
• /
• v.27 no.10
• /
• pp.29-41
• /
• 2022

Recently, with the development of deep learning technology, a variety of huge models with excellent performance have been devised by pre-training massive amounts of text data. However, in order for such a model to be applied to real-life services, the inference speed must be fast and the amount of computation must be low, so the technology for model compression is attracting attention. Knowledge distillation, a representative model compression, is attracting attention as it can be used in a variety of ways as a method of transferring the knowledge already learned by the teacher model to a relatively small-sized student model. However, knowledge distillation has a limitation in that it is difficult to solve problems with low similarity to previously learned data because only knowledge necessary for solving a given problem is learned in a teacher model and knowledge distillation to a student model is performed from the same point of view. Therefore, we propose a heterogeneous knowledge distillation method in which the teacher model learns a higher-level concept rather than the knowledge required for the task that the student model needs to solve, and the teacher model distills this knowledge to the student model. In addition, through classification experiments on about 18,000 documents, we confirmed that the heterogeneous knowledge distillation method showed superior performance in all aspects of learning efficiency and accuracy compared to the traditional knowledge distillation.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.8
• /
• pp.371-380
• /
• 2023

Relation extraction is to extract relationships between named entities from text. Traditionally, relation extraction methods only extract relations between predetermined subject and object entities. However, in end-to-end relation extraction, all possible relations must be extracted by considering the positions of the subject and object for each pair of entities, and so this method uses time and resources inefficiently. To alleviate this problem, this paper proposes a method that sets directions based on the positions of the subject and object, and extracts relations according to the directions. The proposed method utilizes existing relation extraction data to generate direction labels indicating the direction in which the subject points to the object in the sentence, adds entity position tokens and entity type to sentences to predict the directions using a pre-trained language model (KLUE-RoBERTa-base, RoBERTa-base), and generates representations of subject and object entities through probabilistic crossover operation. Then, we make use of these representations to extract relations. Experimental results show that the proposed model performs about 3 ~ 4%p better than a method for predicting integrated labels. In addition, when learning Korean and English data using the proposed model, the performance was 1.7%p higher in English than in Korean due to the number of data and language disorder and the values of the parameters that produce the best performance were different. By excluding the number of directional cases, the proposed model can reduce the waste of resources in end-to-end relation extraction.

• Journal of the Institute of Convergence Signal Processing
• /
• v.23 no.4
• /
• pp.241-248
• /
• 2022

Shear Reinforcement of Dual Anchorage(SRD) is used to reinforce the safety of reinforced concrete structures at construction sites. Welding is used to make shear reinforcement, and welding plays an important role in determining productivity and competitiveness of products. Therefore, a weld bead detection inspection is required. In this paper, we suggest an algorithm for inspecting welding beads using image data of welding beads. First, the proposed algorithm calculates a brightness value in a vertical direction in an image, and then divides a welding bead in a vertical direction by finding a position corresponding to a 50% height point of the brightness value distribution in the image. The welding bead area is also divided in the same way for the horizontal direction, and then the segmentation image is analyzed if there is a welding bead. The proposed algorithm reduced the amount of computation by performing analysis after specifying the region of interest. In addition, accuracy could be improved by using all brightness values in the vertical and horizontal directions using the difference of brightness between the base metal and the welding bead region in the SRD image. The experiment compared the analysis results using five algorithms, such as K-mean and K-neighborhood, as a method to detect if there is a welding bead, and the experimental result proved that the proposed algorithm was the most accurate.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.6
• /
• pp.536-543
• /
• 2023

Speaker diarization, which labels for "who spoken when?" in speech with multiple speakers, has been studied on a deep neural network-based end-to-end method for labeling on speech overlap and optimization of speaker diarization models. Most deep neural network-based end-to-end speaker diarization systems perform multi-label classification problem that predicts the labels of all speakers spoken in each frame of speech. However, the performance of the multi-label-based model varies greatly depending on what the threshold is set to. In this paper, it is studied a speaker diarization system using single-label classification so that speaker diarization can be performed without thresholds. The proposed model estimate labels from the output of the model by converting speaker labels into a single label. To consider speaker label permutations in the training, the proposed model is used a combination of Permutation Invariant Training (PIT) loss and cross-entropy loss. In addition, how to add the residual connection structures to model is studied for effective learning of speaker diarization models with deep structures. The experiment used the Librispech database to generate and use simulated noise data for two speakers. When compared with the proposed method and baseline model using the Diarization Error Rate (DER) performance the proposed method can be labeling without threshold, and it has improved performance by about 20.7 %.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.3
• /
• pp.137-146
• /
• 2024

The purpose of this study was to determine the effects of 12 weeks of WB-EMS training on body composition and heart rate variability based on BMI Level in Women. The subjects of the study were premenopausal women, and they were classified into the BMI-N(n=15) group for BMI<25, the BMI-1(n=16) group for BMI=25~29.9, and the BMI-2(n=9) group for BMI>30. And then, WB-EMS training was performed of 3 times a week for 12 weeks. Body composition and HRV were measured before and after the participation in exercise, which were subjected to a repeated-measures two-way ANOVA. In the case of a significant interaction between time and group, paired sample t-tests were conducted for a post-hoc analysis within each subject group. Tukey's method was used for post-hoc testing of differences between groups, and the significance level was set at 0.5. The results were as follows; First, The effect of WB-EMS training was found in all variables of body composition. In particular, Weight, BMI, FFM, and FM decreased the most in the BMI-2 group, followed by the BMI-1 and BMI-N groups. %BF and VF decreased the most in the BMI-2 group. Second, There was a difference in BPM in all groups, and the BMI-2 group showed the greatest decrease. There were differences in SDNN and RMSSD for each group, and there was no difference according to obesity level. There was no difference in LF, HF, and LF/HF ratio. In conclusion, it was confirmed that WB-EMS training can be an exercise therapy that has a positive effect on the body composition change and cardiac circulatory system in women with a high level of obesity.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.8
• /
• pp.11-21
• /
• 2024

This study explores the use of virtual reality (VR) technology to identify and label key segments of the golf swing. To address the limitations of existing VR devices, we developed a platform to collect kinematic data from various VR devices using the OpenVR SDK (Software Development Kit) and SteamVR, and developed a semi-automated labeling technique to identify and label temporal changes in kinematic behavior through LSTM (Long Short-Term Memory)-based time series data analysis. The experiment consisted of 80 participants, 20 from each of the following age groups: teenage, young-adult, middle-aged, and elderly, collecting data from five swings each to build a total of 400 kinematic datasets. The proposed technique achieved consistently high accuracy (≥0.94) and F1 Score (≥0.95) across all age groups for the seven main phases of the golf swing. This work aims to lay the groundwork for segmenting exercise data and precisely assessing athletic performance on a segment-by-segment basis, thereby providing personalized feedback to individual users during future education and training.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.53 no.6
• /
• pp.167-173
• /
• 2016

Compressed Sensing(CS) is the theory that can recover signals which are sampled below the Nyquist sampling rate to original analog signals. In this paper, we propose the estimation algorithm of ultrasonic attenuation coefficients in the frequency domain using CS. While most estimation algorithms transform the time-domain signals into the frequency-domain using the Fourier transform, the proposed method directly utilize the spectral information in the recovery process by the basis matrix without the completely recovered signals in the time domain. We apply three transform bases for sparsifying and estimate the attenuation coefficients using the Centroid Downshift method with Dual-reference diffraction compensation technique. The estimation accuracy and execution time are compared for each basis matrix. Computer simulation results show that the DCT basis matrix exhibits less than 0.35% estimation error for the compressive ratio of 50% and about 6% average error for the compressive ratio of 70%. The proposed method which directly extracts frequency information from the CS signals can be extended to estimating for other ultrasonic parameters in the Quantitative Ultrasound (QUS) Analysis.

• Journal of the Institute of Electronics Engineers of Korea TE
• /
• v.39 no.3
• /
• pp.51-58
• /
• 2002

The conventional code division multiple access(CDMA) detector shows severe degradation in communication quality as the number of users increases due to multiple access interferences(MAI). This problem thus restricts the user capacity. Various multiuser detection algorithms have been proposed to overcome the MAI problem. The existing detectors can be generally classified into one of the two categories : linear multiuser detection and subtractive interference cancellation detectors. In the linear multiuser detection, a linear transform is applied to the soft outputs of the conventional detector. In the subtractive interference cancellation detection, estimates of the interference are generated and subtracted out from the received signal. There has been great interest in the family of the subtractive interference cancellation detection because the linear multiuser detection exhibits the disadvantage of taking matrix inversion operations. The successive interference cancellation (SIC) and the parallel interference cancellation (PIC) are the two most popular structures in the subtractive interference cancellation detector. The SIC structure is very simple in hardware complexity, but has the disadvantage of increased processing delay time, while the PIC structure is good in performance, but shows the disadvantage of increased hardware complexity. In this paper we propose a mixed structure form of SIC and PIC in order to achieve good performance as well as simple hardware complexity. A performance analysis of the proposed scheme has been made, and the superior characteristics of the mixed structure are demonstrated by extensive computer simulations. 

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.4
• /
• pp.38-45
• /
• 1995

In this paper a way to improve the performance of the long term prediction is proposed, which adopts the Multi-band Excitation (MBE) method in addition to the Code-Excited Linear Prediction (CELP) method at low bit rates below 4.8 kbps. In the proposed method, the multiband long term prediction is performed on the periodic components which still remain after the long term prediction of the conventional CELP method. At this point, the whole frequency region is divided into subbands whose size is equal to the spacing between the harmonics of the fundamental frequency, and the periodic multiband excitation signals. are represented as the sum of sine waves approximately as large as the spectrum of the excitation signals, so that the actual characteristics of the excitation signals can be better taken into account. To evaluate the performance of the proposed method, computer simulation is performed at 4.8 kbps. The 4.8 kbps DoD CELP and the 4.4 kbps IMBE were chosen as the reference vocoders for the speech quality measure. The result of the perceptual speech quality measure showed that the performance of the proposed method is better than that of the 4.8 kbps DoD CELP vocoder, and similar to that of the 4.4 kbps IMBE vocoder.

• Proceedings of the Korean Institute of Electrical and Electronic Material Engineers Conference
• /
• 2008.11a
• /
• pp.174-174
• /
• 2008

초음파를 이용하여 액체 연료를 분사하면 균일한 입경과 미립화가 우수하며 에너지 절약과 공해방지등을 할 수 있다. 또한 유속과 유량에 관계없이 이용할 수 있어 반도체 분야의 웨이퍼와 평판 표시기상에 사진 석판용 화학물질의 균일도포 컴퓨터 하드 디스크의 광택제 도포등에 사용할 수 있다. 이처럼 초저의 유출 용량을 요구하는 모든 공정 및 액체연료의 분사가 요구되는 모든 산업에 적용할 수 있는 장점을 가지고 있다. 하지만 현제까지 주로 사용되고 있는 초음파노즐의 액츄에이터는 단판액츄에이터형로 높은 교류전압을 인가해주어야 하는 단점을 가지고 있다. 이 단점을 해결하기 위해 적층액츄에이터형을 사용하여 초음파 노즐 구동하면 낮은 교류 입력전압에서도 단판액츄에이터형 초음파 노즐과 같은 특성을 가질 수 있다. 또한 초음파 노즐의 구동시 기계적인 진동을 이용하므로 많은 열을 발생시켜 노즐의 온도가 상승하여 세라믹 액츄에이터에도 그 영향을 미치게 되어 열적 열화 현상이 일어날 수 있기에 높은 큐리온도를 가지는 액츄에이터가 필요하다. 본 실험에서는 $Pb(Mn_{1/3}Nb_{2/3})_{0.02}(Ni_{1/3}Nb_{2/3})_{0.12}(Zr_{0.50}Ti_{0.50})_{0.86}O_3$ 조성을 사용하여 $900^{\circ}C$의 저온에서 액상 소결하여 적층혈액츄에이터를 제작하였으며 압전 및 유전 특성을 조사하였다. 제작된 초음파노즐을 구동하기 위해서는 약 36kHz의 30V이상의 교류입력전압 할 수 있는 구동회로가 필요로 한다. 압전액츄에이터의 구동을 위해서는 정확한 정현파 입력이 필요 없다. 압전액츄에이터의 특성상 유사 정현파 입력 만으로도 임피던스 매칭이 이루어지기 때문에 설계가 쉽고 간편한 Push-Pull 방식을 이용한 PWM인버터를 사용하였고 인버터의 출력 주파수를 34~38kHz까지 가변 할 수 있게 설계하였다. 제작된 적층액츄에이터형 초음파 노즐을 PWM인버터로 실제 액체 연료인 경유를 분사하였을 때의 액츄에이터의 온도 변화에 따른 공진주파수와 온도 의존성, 전기적 특성을 조사하고 미립화 분사되는 경유의 미립자 크기 및 최대 분사량을 조사 하였다.