• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.609-611
• /
• 2021

기계 독해는 주어진 지문 내에서 질문에 대한 답을 기계가 찾아 답하는 문제이다. 딥러닝에서는 여러 데이터셋을 학습시킬 때에 이전에 학습했던 데이터의 weight값이 점차 사라지고 사라진 데이터에 대해 테스트 하였을때 성능이 떨어진 결과를 보인다. 이를 과거에 학습시킨 데이터의 정보를 계속 가진 채로 새로운 데이터를 학습할 수 있는 Continual learning을 통해 해결할 수 있고, 본 논문에서는 이 방법을 MRC에 적용시켜 학습시킨 후 한국어 자연어처리 Task인 Korquad 1.0의 MRC dev set을 통해 성능을 측정하였다. 세 개의 데이터셋중에서 랜덤하게 5만개를 추출하여 10stage를 학습시킨 50K 모델에서 추가로 Continual Learning의 Learning without Forgetting를 사용하여 학습시킨 50K-LWF 모델이 F1 92.57, EM 80.14의 성능을 보였고, BERT 베이스라인 모델의 성능 F1 91.68, EM 79.92에 비교하였을 때 F1, EM 각 0.89, 0.22의 향상이 있었다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.10a
• /
• pp.289-290
• /
• 2022

Arteriosclerosis is a disease in which the carotid vessel wall becomes thick, and it is important to monitor the thickness of the vessel wall for diagnosis. In this study, we propose a model for extracting 324 radiomics features from carotid MRI images and diagnosing arteriosclerosis using machine learning techniques. We learned a total of four classification models: logistic regression, support vector machine, random forest, and XGBoost through radiomics features. XGBoost model, which showed the highest performance in 5-fold cross-validation, shows the results of accuracy 0.9023, sensitivity 0.9517, specificity 0.8035, AUC 0.8776.

• The KIPS Transactions:PartB
• /
• v.15B no.4
• /
• pp.307-314
• /
• 2008

This paper presents a new robust watermarking method for curves that uses informed-detection. To embed watermarks, the presented algorithm parameterizes a curve using the B-spline model and acquires the control points of the B-spline model. For these control points, 2D mesh are created by applying Delaunay triangulation and then the mesh spectral analysis is performed to calculate the mesh spectral coefficients where watermark messages are embedded in a spread spectrum way. The watermarked coefficients are inversely transformed to the coordinates of the control points and the watermarked curve is reconstructed by calculating B-spline model with the control points. To detect the embedded watermark, we apply curve matching algorithm using inflection points of curve. After curve registration, we calculate the difference between the original and watermarked mesh spectral coefficients with the same process for embedding. By calculating correlation coefficients between the detected and candidate watermark, we decide which watermark was embedded. The experimental results prove the proposed scheme is more robust than previous watermarking schemes against print-scan process as well as geometrical distortions.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2009.10a
• /
• pp.599-601
• /
• 2009

In currently being adopted as a e-leaning standard, SCORM it is difficult to provide the customized contents to a learner by changing the learner's level at runtime, and to control selective studying. So, we designed and implemented the customized contents organization engine(CCOE) in order to complement SCORM's faults in this paper. The CCOE consists of a level evaluation module, a contents re-organization module and a question item selection module. A level evaluation module evaluates the learner's level based on a question item reaction theory. And a question item selection module selects some random items by each level or by considering the learner's level which is then provided to a studying before evaluation, a section evaluation, and a quiz. And then this module transmits the selected items to the contents reorganization module for providing the quiz. A contents re-organization module selects the customized contents based on the learner's level by searching the tagged difficulty to the content, and creates the sequence with the selected items and the transmitted items from the question item selection module. If proposed in this paper CCOE is applied, the higher effectiveness of learning is expected by providing the customized learning contents based on the re-evaluated learner's level by each section.

• Journal of Digital Convergence
• /
• v.16 no.4
• /
• pp.95-105
• /
• 2018

This study empirically test the roles of network closure and structural hole on radical innovation in technological knowledge exchange network in Gumi cluster. In doing so, we build 2,550 firm network, transforming association*firm(2-mode) to firm*firm(1-mode) network data. In addition, in order to investigate firms' attributes, we conduct survey for 101 firms in Gumi cluster using random sampling, and finally collect 86 firm samples. For analysis, we use ridge regression since network density and efficiency, indices of network closure and structural hole respectively, has a high level of multicollinearity. The findings show that structural hole has a significant and positive impact on radical innovation, but network closure has a significant and negative impact on radical innovation. This study contributes to present an empirical evidence of debate on network closure and structural hole based on past conceptual discussions and literature review and further goes a long way towards strategy formulation to establish social capital in accomplishing radical innovation. Further research is required that pays closer attention to features of technological knowledge, innovation types and interaction between network closure and structural hole, directing efforts to structural characteristics of various networks.

• Journal of KIISE:Information Networking
• /
• v.30 no.1
• /
• pp.24-29
• /
• 2003

In this paper, a new method for the mixed image separation is presented using the independent component analysis, the innovation process, and the expectation-maximization. In general, the independent component analysis (ICA) is one of the widely used statistical signal processing schemes, which represents the information from observations as a set of random variables in the from of linear combinations of another statistically independent component variables. In various useful applications, ICA provides a more meaningful representation of the data than the principal component analysis through the transformation of the data to be quasi-orthogonal to each other. which can be utilized in linear projection.. However, it has been known that ICA does not establish good performance in source separation by itself. Thus, in order to overcome this limitation, there have been many techniques that are designed to reinforce the good properties of ICA, which improves the mixed image separation. Unfortunately, the innovation process still needs to be studied since it yields inconsistent innovation process that is attached to the ICA, the expectation and maximization process is added. The results presented in this paper show that the proposed improves the image separation as presented in experiments.

• Journal of Digital Convergence
• /
• v.19 no.5
• /
• pp.263-268
• /
• 2021

The purpose of this study was to propose a model which is suitable for the actual delivery system by designing a fetal delivery hospital operation management and fetal health classification model. The number of deaths during childbirth is similar to the number of maternal mortality rate of 295,000 as of 2017. Among those numbers, 94% of deaths are preventable in most cases. Therefore, in this paper, we proposed a model that predicts the health condition of the fetus using data like heart rate of fetuses, fetal movements, uterine contractions, etc. that are extracted from the Cardiotocograms(CTG) test using a random forest. If the redundancy of the data is unbalanced, This proposed model guarantees a stable management of the fetal delivery health management system. To secure the accuracy of the fetal delivery health management system, we remove the outlier which embedded in the system, by setting thresholds for the upper and lower standard deviations. In addition, as the proportion of the sequence class uses the health status of fetus, a small number of classes were replicated by data-resampling to balance the classes. We had the 4~5% improvement and as the result we reached the accuracy of 97.75%. It is expected that the developed model will contribute to prevent death and effective fetal health management, also disease prevention by predicting and managing the fetus'deaths and diseases accurately in advance.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.18 no.5
• /
• pp.93-101
• /
• 2008

Side channel attacks are well known as one of the most powerful physical attacks against low-power cryptographic devices and do not take into account of the target's theoretical security. As an important succeeding factor in side channel attacks (specifically in DPAs), exact time-axis alignment methods are used to overcome misalignments caused by trigger jittering, noise and even some countermeasures intentionally applied to defend against side channel attacks such as random clock generation. However, the currently existing alignment methods consider only on the position of signals on time-axis, which is ineffective for certain countermeasures based on time-axis misalignments. This paper proposes a new signal alignment method based on interpolation and decimation techniques. Our proposal can align the size as well as the signals' position on time-axis. The validity of our proposed method is then evaluated experimentally with a smart card chip, and the results demonstrated that the proposed method is more efficient than the existing alignment methods.

• The Korean Journal of Applied Statistics
• /
• v.32 no.3
• /
• pp.349-374
• /
• 2019

A class imbalance problem arises when one class outnumbers the other class by a large proportion in binary data. Studies such as transforming the learning data have been conducted to solve this imbalance problem. In this study, we compared resampling methods among methods to deal with an imbalance in the classification problem. We sought to find a way to more effectively detect the minority class in the data. Through simulation, a total of 20 methods of over-sampling, under-sampling, and combined method of over- and under-sampling were compared. The logistic regression, support vector machine, and random forest models, which are commonly used in classification problems, were used as classifiers. The simulation results showed that the random under sampling (RUS) method had the highest sensitivity with an accuracy over 0.5. The next most sensitive method was an over-sampling adaptive synthetic sampling approach. This revealed that the RUS method was suitable for finding minority class values. The results of applying to some real data sets were similar to those of the simulation.

• Journal of Digital Contents Society
• /
• v.16 no.4
• /
• pp.605-613
• /
• 2015

This paper addresses the problem to make sparse the projection matrix in pattern recognition method. Recently, the size of computer program is often restricted in embedded systems. It is very often that developed programs include some constant data. For example, many pattern recognition programs use the projection matrix for dimension reduction. To improve the recognition performance, very high dimensional feature vectors are often extracted. In this case, the projection matrix can be very big. Recently, RSR(roated sparse regression) method[1] was proposed. This method has been proved one of the best algorithm that obtains the sparse matrix. We propose three methods to improve the RSR; outlier removal, sampling and elastic net RSR(E-RSR) in which the penalty term in RSR optimization function is replaced by that of the elastic net regression. The experimental results show that the proposed methods are very effective and improve the sparsity rate dramatically without sacrificing the recognition rate compared to the original RSR method.