• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.483-489
• /
• 2020

This paper proposes acoustic model training using self-attention for low-resource speech recognition. In low-resource speech recognition, it is difficult for acoustic model to distinguish certain phones. For example, plosive /d/ and /t/, plosive /g/ and /k/ and affricate /z/ and /ch/. In acoustic model training, the self-attention generates attention weights from the deep neural network model. In this study, these weights handle the similar pronunciation error for low-resource speech recognition. When the proposed method was applied to Time Delay Neural Network-Output gate Projected Gated Recurrent Unit (TNDD-OPGRU)-based acoustic model, the proposed model showed a 5.98 % word error rate. It shows absolute improvement of 0.74 % compared with TDNN-OPGRU model.

• Journal of Broadcast Engineering
• /
• v.26 no.1
• /
• pp.14-25
• /
• 2021

Change detection, which finds changes in remote sensing images of the same location captured at different times, is very important because it is used in various applications. However, registration errors, building displacement errors, and shadow errors cause false positives. To solve these problems, we propose a novle deep convolutional network called CADNet (Change Attention Dense Siamese Network). CADNet uses FPN (Feature Pyramid Network) to detect multi-scale changes, applies a Change Attention Module that attends to the changes, and uses DenseNet as a feature extractor to use feature maps that contain both low-level and high-level features for change detection. CADNet performance measured from the Precision, Recall, F1 side is 98.44%, 98.47%, 98.46% for WHU datasets and 90.72%, 91.89%, 91.30% for LEVIR-CD datasets. The results of this experiment show that CADNet can offer better performance than any other traditional change detection method.

• KIPS Transactions on Software and Data Engineering
• /
• v.11 no.1
• /
• pp.11-18
• /
• 2022

De novo drug design is the process of developing new drugs that can interact with biological targets such as protein receptors. Traditional process of de novo drug design consists of drug candidate discovery and drug development, but it requires a long time of more than 10 years to develop a new drug. Deep learning-based methods are being studied to shorten this period and efficiently find chemical compounds for new drug candidates. Many existing deep learning-based drug design models utilize recurrent neural networks to generate a chemical entity represented by SMILES strings, but due to the disadvantages of the recurrent networks, such as slow training speed and poor understanding of complex molecular formula rules, there is room for improvement. To overcome these shortcomings, we propose a deep learning model for SMILES string generation using variational autoencoders with self-attention mechanism. Our proposed model decreased the training time by 1/26 compared to the latest drug design model, as well as generated valid SMILES more effectively.

• Journal of the Korean Academy of Child and Adolescent Psychiatry
• /
• v.32 no.3
• /
• pp.85-92
• /
• 2021

Attention-deficit/hyperactivity disorder (ADHD) leads to functional decline in academic performance, interpersonal relationships, and development in school-aged children. Early diagnosis and appropriate intervention can significantly reduce the functional decline caused by ADHD. Currently, there is no established biological marker for ADHD. Some studies have suggested that various indicators from the quantitative electroencephalogram (QEEG) may be useful biological markers for the diagnosis of ADHD. Until the 2010s, theta/beta ratio (TBR) was a biomarker candidate for ADHD that consistently showed high diagnostic value. However, limitations of TBR have recently been reported. Studies have demonstrated that phase-amplitude coupling, especially theta phase-gamma amplitude coupling, are related to cognitive dysfunction and may assist in the diagnosis of ADHD. As yet, the underlying mechanism is not clearly established, and the clinical efficacy of these biomarkers needs to be proven through well-controlled studies. Based on the heterogeneous characteristics of ADHD, subgrouping through QEEG plays a key role in diagnosis and treatment planning. Sophisticated, well-designed studies and meta-analyses are necessary to confirm these findings.

• Journal of Korea Multimedia Society
• /
• v.24 no.8
• /
• pp.1026-1034
• /
• 2021

Tire wear and defect are important factors for safe driving condition. These defects are generally inspected by some specialized experts or very expensive equipments such as stereo depth camera and depth gauge. In this paper, we propose tire safety vision inspector based on deep neural network (DNN). The status of tire wear is categorized into three: 'safety', 'warning', and 'danger' based on depth of tire tread. We propose an attention mechanism for emphasizing the feature of tread area. The attention-based feature is concatenated to output feature maps of the last convolution layer of ResNet-101 to extract more robust feature. Through experiments, the proposed tire wear classification model improves 1.8% of accuracy compared to the existing ResNet-101 model. For detecting the tire defections, the developed tire defect detection model shows up-to 91% of accuracy using the Mask R-CNN model. From these results, we can see that the suggested models are useful for checking on the safety condition of working tire in real environment.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.135-140
• /
• 2022

최근 Document Retrieval을 비롯한 대부분의 자연어처리 분야에서는 BERT와 같이 self-attention을 기반으로 한 사전훈련 모델을 활용하여 SOTA(state-of-the-art)를 이루고 있다. 그러나 self-attention 메커니즘은 입력 텍스트 길이의 제곱에 비례하여 계산 복잡도가 증가하기 때문에, 해당 모델들은 선천적으로 입력 텍스트의 길이가 제한되는 한계점을 지닌다. Document Retrieval 분야에서는, 문서를 특정 토큰 길이 단위의 문단으로 나누어 각 문단의 유사 점수 또는 표현 벡터를 추출한 후 집계함으로서 길이 제한 문제를 해결하는 방법론이 하나의 주류를 이루고 있다. 그러나 논문, 특허와 같이 섹션 형식(초록, 결론 등)을 갖는 문서의 경우, 섹션 유형에 따라 고유한 정보 특성을 지닌다. 따라서 문서를 단순히 특정 길이의 문단으로 나누어 학습하는 PARADE와 같은 기존 방법론은 각 섹션이 지닌 특성을 반영하지 못한다는 한계점을 지닌다. 본 논문에서는 섹션 유형에 대한 정보를 포함하는 문단 표현을 학습한 후, 트랜스포머 인코더를 사용하여 집계함으로서, 결과적으로 섹션의 특징과 상호 정보를 학습할 수 있도록 하는 SERADE 모델을 제안하고자 한다. 실험 결과, PARADE-Transformer 모델과 비교하여 평균 3.8%의 성능 향상을 기록하였다.

• Management & Information Systems Review
• /
• v.9
• /
• pp.117-137
• /
• 2002

As attention to the Internet from both companies and individuals is rapidly on the increase, hundreds of new websites are opening in a single day. Along with such a high attention to the Internet, to set up an effective website needs efficient evaluation and reliable evaluation criterions for the website. The existing homepage contests and evaluation models are limited to certain websites in a special field or to the systemic side and to the contents, which in fact weakened the development of detailed evaluation sections and items possibly measured. This study is designed to integrate and seek out methods and success factors that should be considered when a website is built up, discovering evaluation criterions and making evaluation models objectively possible to be measured. The study focused on investigation into a new measurement standard and model by considering the previous studies, in order to suggest the followings: Centering the 7 top evaluation sections by type of each website such as (1) Service, (2) Mechanism, (3) Structure & Navigation, (4) Usability, (5) Contents (6) Community, (7) Communication, the study suggests an objective and reasonable website evaluation model on a basis of common factors considered in an integral and optimum view.

• Special Issue of the Society of Naval Architects of Korea
• /
• 2006.09a
• /
• pp.7-14
• /
• 2006

In recent years, shipboard noise control is attracting increasing attention to human environmental conditions and crew's opportunity for rest and recreation with work on board. In order to minimize the noise levels, careful attention have to be paid by the experts from initial design stage to the delivery. This paper describes the outlines of shipboard noise control including general characteristics of shipboard noise, measurement, evaluation, prediction, and control measures considering the noise transmission mechanism from source to receiver space.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.10a
• /
• pp.844-847
• /
• 2019

주가 예측은 상업적인 매력 때문에 많은 이목이 끌리는 분야이지만, 주가의 불확실성과 변동성 때문에 주가 예측은 어려운 작업이다. 최근에는 주가 예측 모델에 어텐션 메커니즘을 사용하여 주가 예측에 많은 인자들이 사용되어 생기는 성능 하락 문제를 해결하여 좋은 성능을 보여주는 연구가 존재한다. 본 연구에서는 그 모델 중 하나인 Dual-Stage Attention-Based Recurrent Neural Network(DARNN)의 어텐션 메커니즘을 변경해가며 어떤 어텐션 메커니즘이 주가 예측에 적합한지를 알아본다. KOSPI100 지수의 예측실험을 통해 location 스코어함수를 사용한 어텐션 메커니즘이 가장 뛰어난 성능을 보여주는 것을 확인하였고, 이는 기존의 스코어함수를 사용한 DARNN에 비해 약 10% 향상된 성능으로 스코어 함수가 모델의 중요한 영향을 끼치는 것을 확인하였다.

• Journal of Broadcast Engineering
• /
• v.27 no.7
• /
• pp.1011-1020
• /
• 2022

Recent high-performance panoptic segmentation models are based on transformer architectures. However, transformer-based panoptic segmentation methods are basically slower than convolution-based methods, since the attention mechanism in the transformer requires quadratic complexity w.r.t. image resolution. Also, sine and cosine computation for positional embedding in the transformer also yields a bottleneck for computation time. To address these problems, we adopt three modules to speed up the inference runtime of the transformer-based panoptic segmentation. First, we perform channel-level reduction using depth-wise separable convolution for inputs of the transformer decoder. Second, we replace sine and cosine-based positional encoding with convolution operations, called conv-embedding. We also apply a separable self-attention to the transformer encoder to lower quadratic complexity to linear one for numbers of image pixels. As result, the proposed model achieves 44% faster frame per second than baseline on ADE20K panoptic validation dataset, when we use all three modules.