• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.21 no.6
• /
• pp.374-382
• /
• 2020

The purpose of this study is to improve the learning speed of an ammunition stockpile reliability classification artificial neural network model by proposing a normalization method that reduces the number of input variables based on the characteristic of Ammunition Stockpile Reliability Program (ASRP) data without loss of classification performance. Ammunition's performance requirements are specified in the Korea Defense Specification (KDS) and Ammunition Stockpile reliability Test Procedure (ASTP). Based on the characteristic of the ASRP data, input variables can be normalized to estimate the lot percent nonconforming or failure rate. To maintain the unitary hypercube condition of the input variables, min-max normalization method is also used. Area Under the ROC Curve (AUC) of general min-max normalization and proposed 2-step normalization is over 0.95 and speed-up for marching learning based on ASRP field data is improved 1.74 ~ 1.99 times depending on the numbers of training data and of hidden layer's node.

• Science of Emotion and Sensibility
• /
• v.20 no.2
• /
• pp.161-170
• /
• 2017

Emotion affects many parts of human life such as learning ability, behavior and judgment. It is important to understand human nature. Emotion can only be inferred from facial expressions or gestures, what it actually is. In particular, emotion is difficult to classify not only because individuals feel differently about emotion but also because visually induced emotion does not sustain during whole testing period. To solve the problem, we acquired bio-signals and extracted features from those signals, which offer objective information about emotion stimulus. The emotion pattern classifier was composed of unsupervised learning algorithm with hidden nodes and feature vectors. Restricted Boltzmann machine (RBM) based on probability estimation was used in the unsupervised learning and maps emotion features to transformed dimensions. The emotion was characterized by non-linear classifiers with hidden nodes of a multi layer neural network, named deep belief network (DBN). The accuracy of DBN (about 94 %) was better than that of back-propagation neural network (about 40 %). The DBN showed good performance as the emotion pattern classifier.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.14 no.1
• /
• pp.37-53
• /
• 2012

There is considerable uncertainty in ground properties used in tunnel designs. In this study, a back analysis was performed to find optimal ground properties based on the artificial neural network facility of MATLAB program of using tunnel monitoring data. Total 81 data were constructed by changing elastic modulus and coefficient of lateral pressure which have great influence on tunnel convergence. A sensitivity analysis was conducted to establish an optimal training model by varying the number of hidden layers, the number of nodes, learning rate, and momentum. Meanwhile, the optimal training model was selected by comparing MSE (Mean Squared Error) and coefficient of determination ($R^2$) and was used to find the correct elastic moduli of layers and the coefficient of lateral pressure. In future, it is expected that the suggested method of this study can be applied to determine the optimum tunnel support pattern under given ground conditions.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.4
• /
• pp.48-53
• /
• 2001

In this paper, selectively attentive learning method has been proposed to improve the learning speed of multilayer Perceptron based on the error backpropagation algorithm. Three attention criterions are introduced to effectively determine which set of input patterns is or which portion of network is attended to for effective learning. Such criterions are based on the mean square error function of the output layer and class-selective relevance of the hidden nodes. The acceleration of learning time is achieved by lowering the computational cost per iteration. Effectiveness of the proposed method is demonstrated in a speaker adaptation task of isolated word recognition system. The experimental results show that the proposed selective attention technique can reduce the learning time more than 60% in an average sense.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.11 no.1
• /
• pp.73-80
• /
• 2016

The Car License Plate Recognition(: CLPR) is required in searching the hit-and-run car, measuring the traffic density, investigating the traffic accidents as well as in pursuing vehicle crimes according to the increasing in number of vehicles. The captured images on the real environment of the CLPR is contaminated not only by snow and rain, illumination changes, but also by the geometrical distortion due to the pose changes between camera and car at the moment of image capturing. We propose homographic transformation and intensity histogram of vertical image projection so as to transform the distorted input to the original image and cluster the character and number, respectively. Especially, in this paper, the Multilayer Perceptron Algorithm(: MLP) in the CLPR is used to not only recognize the charcters and car license plate, but also determine the optimized ratio among the number of input, hidden and output layers by the real experimental result.

• The Transactions of the Korea Information Processing Society
• /
• v.3 no.5
• /
• pp.1281-1293
• /
• 1996

In this paper, we propose a novel recognition model of on-line cursive Korean characters using HMM(Hidden Markov Model) and level building algorithm. The model is constructed as a form of recognition network with HMM for graphemes and Korean combination rules. Though the network is so flexible as to accomodate variability of input patterns, it has a problem of recognition speed caused by 11, 172 search paths. To settle the problem, we modify the level building algorithm to be adapted directly to the Korean combination rules and apply it to the model. The modified algorithm is efficient network search procedure time complexity of which depends on the number of HMMs for each grapheme, not the number of paths in the extensive recognition network. A test with 15, 000 hand written characters shows recognition rat 90% and speed of 0.72 second per character.

• The Journal of the Acoustical Society of Korea
• /
• v.11 no.3
• /
• pp.5-13
• /
• 1992

Taking devantage of the property that Korean digit is a mono-syllable word, we proposed a spoken Korean-digit recognition scheme using the multi-layer perceptron. The spoken Korean-digit is divided into three segments (initial sound, medial vowel, and final consonant) based on the voice starting / ending points and a peak point in the middle of vowel sound. The feature vectors such as cepstrum, reflection coefficients, ${\Delta}$cepstrum and ${\Delta}$energy are extracted from each segment. It has been shown that cepstrum, as an input vector to the neural network, gives higher recognition rate than reflection coefficients. Regression coefficients of cepstrum did not affect as much as we expected on the recognition rate. That is because, it is believed, we extracted features from the selected stationary segments of the input speech signal. With 150 ceptral coefficients obtained from each spoken digit, we achieved correct recognition rate of 97.8%.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.18 no.2
• /
• pp.157-168
• /
• 1993

In this paper, it is shown that neural networks trained with back propagation algorithm and dictionary can be applied to recognize handwritten English characters. To eliminate the useless data part and to minimize the variety of characters from the scanned image file, various preprocessings : that is, segmentation, centering, noise filtering, sealing and thinning are performed. After these, characteristic features are derived from thinned character pattern. The neural network is trained by using the extracted features for sample data, and all test data are classified into English alphabets according to their features through the neural network. Finally, the ways of reducing learning time and improving recognition rate, and the relationship between learning time and hidden layer nodes are considered. As a result of this study, after successful training, a high recognition rate has been obtained with this system for the trained patterns and about 93% for test patterns. Using dictionary, the recognition rate was about 97% for test pattern.

• Journal of the Korean Data and Information Science Society
• /
• v.12 no.2
• /
• pp.1-10
• /
• 2001

Neural networks we increasingly being seen as an addition to the statistics toolkit which should be considered alongside both classical and modern statistical methods. Neural networks are usually useful for classification and function estimation. In this paper we concentrate on function estimation using neural networks with weight decay factor The use of weight decay seems both to help the optimization process and to avoid overfitting. In this type of neural networks, the problem to decide the number of hidden nodes, weight decay parameter and iteration number of learning is very important. It is called the optimization of weight decay neural networks. In this paper we propose a automatic optimization based on genetic algorithms. Moreover, we compare the weight decay neural network automatically learned according to automatic optimization with ordinary neural network, projection pursuit regression and support vector machines.

• Journal of the Korea Society of Computer and Information
• /
• v.16 no.8
• /
• pp.39-47
• /
• 2011

Ergezinger's method, one of the layer-by-layer algorithms used for multilyer perceptrons, consists of an output node and can make premature saturations in the output's weight because of using linear least squared method in the output layer. These saturations are obstacles to learning time and covergence. Therefore, this paper expands Ergezinger's method to be able to use an output vector instead of an output node and introduces a learning rate to improve learning time and convergence. The learning rate is a variable rate that reflects the correlation coefficient between new weight and previous weight while updating hidden's weight. To compare the proposed method with Ergezinger's method, we tested iris recognition and nonlinear approximation. It was found that the proposed method showed better results than Ergezinger's method in learning convergence. In the CPU time considering correlation coefficient computation, the proposed method saved about 35% time than the previous method.