• 한국정보전자통신기술학회논문지
• /
• 제14권3호
• /
• pp.201-207
• /
• 2021

머신러닝의 한 종류인 딥러닝은 각 학습 과정을 진행할 때, 가중치를 변경하면서 학습을 수행한다. 딥러닝을 수행할때 대표적으로 사용되는 Tensor Flow나 Keras의 경우 학습이 종료된 결과를 그래프 형태로 제공한다. 이에 과다학습으로 인한 퇴화 현상 또는 가중치의 잘못된 설정으로 인해 학습 결과에 오류가 발생하는 경우, 해당 학습 결과를 폐기해야한다. 이에 기존 기술은 학습 결과를 롤백하는 기능을 제공하고 있지만, 롤백 기능은 최대 5회 이내의 결과로 제한된다. 또한, 딥러닝의 모든 과정을 기록하고 있는 것이 아니기 때문에 값을 추적하기 어렵다. 이를 해결하기 위해 MLOps의 개념을 적용한 기술이 존재하지만. 해당 기술에서는 이전 시점으로 롤백하는 기능을 제공하지 않는다. 본 논문에서는 기존 기술의 문제점을 해결하기 위해 학습 과정의 중간 값을 블록체인으로 관리하여 학습 중간 과정을 기록하고, 오류가 발생할 경우 롤백할 수 있는 시스템을 구성한다. 블록체인의 기능 수행을 위해서 딥러닝 과정 및 학습 결과 롤백은 Smart Contract를 작성하여 동작하도록 설계하였다. 성능평가는 기존의 딥러닝 방식의 롤백 기능을 평가하였을 때, 제안방식은 100%의 복구율을 가지는 것에 비교하여 기존 기법에서는 6회 이후에 복구율이 감소되어 50회일 때 10%까지 감소하는 것을 확인하였다. 또한, 이더리움 블록체인의 Smart Contract를 사용할 때, 블록 1회 생성 시 157만원의 금액이 지속적으로 소모되는 것을 확인하였다.

• ETRI Journal
• /
• 제28권1호
• /
• pp.59-66
• /
• 2006

Learning behaviors of a radial basis function network (RBFN) using a singular value decomposition (SVD) and stochastic gradient (SG) algorithm, together named RBF-SVD-SG, for odor sensing systems are analyzed, and a fast training method is proposed. RBF input data is from a conducting polymer sensor array. It is revealed in this paper that the SG algorithm for the fine-tuning of centers and widths still shows ill-behaving learning results when a sufficiently small convergence coefficient is not used. Since the tuning of centers in RBFN plays a dominant role in the performance of RBFN odor sensing systems, our analysis is focused on the center-gradient variance of the RBFN-SVD-SG algorithm. We found analytically that the steadystate weight fluctuation and large values of a convergence coefficient can lead to an increase in variance of the center-gradient estimate. Based on this analysis, we propose to use the least mean square algorithm instead of SVD in adjusting the weight for stable steady-state weight behavior. Experimental results of the proposed algorithm have shown faster learning speed and better classification performance.

• 전자공학회논문지C
• /
• 제35C권6호
• /
• pp.67-75
• /
• 1998

다층퍼셉트론의 학습에서 나타나는 출력노드의 부적절한 포화를 해결하기 위해서 n차 크로스-엔트로피 오차함수가 제안되었으나, 이 오차함수를 이용한 학습성능은 오차함수의 차수에 민감하여 적절한 차수를 결정해야 하는 문제점이 있다. 이 논문에서는, 학습의 진행에 따라 학습률을 가변시키는 새로운 방법을 제시하여 다층퍼셉트론의 학습성능이 n차 크로스-엔트로피 오차함수의 차수에 덜 민감하도록 한다. 또한, 가변학습률이 매우 커지는 경우에 학습이 불안정해지는 것을 방지하기 위해서 오차신호의 크기를 제한하는 방법을 제시한다. 마지막으로, 필기체 숫자 인식 문제와 갑상선 진단 문제의 시뮬레이션으로 제안한 방법의 효용성을 검증한다.

• Journal of Electrical Engineering and Technology
• /
• 제13권5호
• /
• pp.2099-2106
• /
• 2018

This study presents a novel approach of discriminative feature vectors based on manifold learning using nonlinear dimension reduction (DR) technique to improve loss function, and combine with the Adversarial examples to regularize the object function for image classification. The traditional convolutional neural networks (CNN) with many new regularization approach has been successfully used for image classification tasks, and it achieved good results, hence it costs a lot of Calculated spacing and timing. Significantly, distrinct from traditional CNN, we discriminate the feature vectors for objects without empirically-tuned parameter, these Discriminative features intend to remain the lower-dimensional relationship corresponding high-dimension manifold after projecting the image feature vectors from high-dimension to lower-dimension, and we optimize the constrains of the preserving local features based on manifold, which narrow the mapped feature information from the same class and push different class away. Using Adversarial examples, improved loss function with additional regularization term intends to boost the Robustness and generalization of neural network. experimental results indicate that the approach based on discriminative feature of manifold learning is not only valid, but also more efficient in image classification tasks. Furthermore, the proposed approach achieves competitive classification performances for three benchmark datasets : MNIST, CIFAR-10, SVHN.

• 인간식물환경학회지
• /
• 제24권6호
• /
• pp.651-662
• /
• 2021

Background and objective: This study identifies whether children's planning-organizing executive function can be significantly classified and predicted by home environment quality and wealth factors. Methods: For empirical analysis, we used the data collected from the 10th Panel Study on Korean Children in 2017. Using machine learning tools such as support vector machine (SVM) and random forest (RF), we evaluated the accuracy of the model in which home environment factors classify and predict children's planning-organizing executive functions, and extract the relative importance of variables that determine these executive functions by income group. Results: First, SVM analysis shows that home environment quality and wealth factors show high accuracy in classification and prediction in all three groups. Second, RF analysis shows that estate had the highest predictive power in the high-income group, followed by income, asset, learning, reinforcement, and emotional environment. In the middle-income group, emotional environment showed the highest score, followed by estate, asset, reinforcement, and income. In the low-income group, estate showed the highest score, followed by income, asset, learning, reinforcement, and emotional environment. Conclusion: This study confirmed that home environment quality and wealth factors are significant factors in predicting children's planning-organizing executive functions.

• Journal of the Korean Society for Industrial and Applied Mathematics
• /
• 제27권1호
• /
• pp.56-74
• /
• 2023

In order to obtain practical and high-quality satellite images containing high-frequency components, a large aperture optical system is required, which has a limitation in that it greatly increases the payload weight. As an attempt to overcome the problem, many multi-aperture optical systems have been proposed, but in many cases, these optical systems do not include high-frequency components in all directions, and making such an high-quality image is an ill-posed problem. In this paper, we use deep learning to overcome the limitation. A deep learning model receives low-quality images as input, estimates the Point Spread Function, PSF, and combines them to output a single high-quality image. We model images obtained from three rectangular apertures arranged in a regular polygon shape. We also propose the Modulation Transfer Function Loss, MTF Loss, which can capture the high-frequency components of the images. We present qualitative and quantitative results obtained through experiments.

• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 2004년도 ICCAS
• /
• pp.694-698
• /
• 2004

In this paper, as one of these efforts for making the intelligent data mining system we propose the Associative frame of the memory according to the following three steps. First,the structured frame for performing the main brain function should be made. In this frame, the concepts of learning memory and episode memory are considered. Second,the learning mechanism for data acquisition and storing mechanism in the memory frame are provided. The obtained data are arranged and stored in the memory following the rules of the structured memory frame. Third, it is the last step of processing the inference and knowledge retrieval function using the stored knowledge in the associative memory frame. This system is applied to the area for estimating the purchasing degree from the type of customer's tastes, the pattern of commodities and the evaluation of a company.

• 조명전기설비학회논문지
• /
• 제19권6호
• /
• pp.82-86
• /
• 2005

PID 제어기는 산업자동화 설비에 널리 쓰이고 있다. 하지만 시스템 특성이 간헐 또는 연속적으로 변화할 때에 정밀제어를 위한 새로운 매개변수 결정이 쉽지 않다. 이를 해결하기 위한 방법으로 본 논문에서는 PID 제어기와 같은 기능을 갖는 학습이득조절기를 제안하였다. 시스템의 적절한 학습이득은 델타 학습규칙에 의해서 결정된다. 제안된 학습이득 조절기의 기능은 직류 전동기의 모의실험에 의해 확인하였다.

• 대한기계학회논문집A
• /
• 제20권2호
• /
• pp.399-407
• /
• 1996

A fuzzy learning algorithm to get the optimal fuzzy rules is presented in this paper. The algorithm introduces a reference model to generate a desired output and a performance index funtion instead of the performance index table. The performance index funtion is a cost function based on the error and error-rate between the reference and plant output. The cost function is minimized by a gradient method and the control input is also updated. In this case, the control rules which generate the desired response can be obtained by changing the portion of the error-rate in the cost funtion. In SISO(Single-Input Single- Output)plant, only by the learning delay, it is possible to experss the plant model and to get the desired control rules. In the long run, this algorithm gives us the good control rules with a minimal amount of prior informaiton about the environment.

• 대한전기학회:학술대회논문집
• /
• 대한전기학회 1992년도 하계학술대회 논문집 A
• /
• pp.480-483
• /
• 1992

Although the back-propagation(BP) algorithm is widely used for its simple structure and easy learning method, it has a drawback of slow convergence rate. In this paper, we propose an algorithm to improve this problem by manipulating the slope parameter of the activation function. The steepest descent method is used in learning the slope parameter, as in the case of weight. The simulation shows that the learning rates of the proposed algorithm is faster than the conventional BP algorithm.