• 전자공학회논문지B
• /
• 제31B권4호
• /
• pp.145-154
• /
• 1994

This paper proposes an efficient method for improving the training performance of the neural network by using a hybrid of a stochastic approximation and a backpropagation algorithm. The proposed method improves the performance of the training by appliying a global optimization method which is a hybrid of a stochastic approximation and a backpropagation algorithm. The approximate initial point for a stochastic approximation and a backpropagation algorihtm. The approximate initial point for fast global optimization is estimated first by applying the stochastic approximation, and then the backpropagation algorithm, which is the fast gradient descent method, is applied for a high speed global optimization. And further speed-up of training is made possible by adjusting the training parameters of each of the output and the hidden layer adaptively to the standard deviation of the neuron output of each layer. The proposed method has been applied to the parity checking and the pattern classification, and the simulation results show that the performance of the proposed method is superior to that of the backpropagation, the Baba's MROM, and the Sun's method with randomized initial point settings. The results of adaptive adjusting of the training parameters show that the proposed method further improves the convergence speed about 20% in training.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제11권12호
• /
• pp.6038-6053
• /
• 2017

In this study, we propose an automatic melody extraction algorithm using deep learning. In this algorithm, feature images, generated using the energy of frequency band, are extracted from polyphonic audio files and a deep learning technique, a convolutional neural network (CNN), is applied on the feature images. In the training data, a short frame of polyphonic music is labeled as a musical note and a classifier based on CNN is learned in order to determine a pitch value of a short frame of audio signal. We want to build a novel structure of melody extraction, thus the proposed algorithm has a simple structure and instead of using various signal processing techniques for melody extraction, we use only a CNN to find a melody from a polyphonic audio. Despite of simple structure, the promising results are obtained in the experiments. Compared with state-of-the-art algorithms, the proposed algorithm did not give the best result, but comparable results were obtained and we believe they could be improved with the appropriate training data. In this paper, melody extraction and the proposed algorithm are introduced first, and the proposed algorithm is then further explained in detail. Finally, we present our experiment and the comparison of results follows.

• 산업공학
• /
• 제24권2호
• /
• pp.97-104
• /
• 2011

Research for multi-variate classification has been studied through two kinds of procedures which are feature selection and classification. Feature Selection techniques have been applied to select important features and the other one has improved classification performances through classifier applications. In general, each technique has been independently studied, however consideration of the interaction between both procedures has not been widely explored which leads to a degraded performance. In this paper, through integrating these two procedures, classification performance can be improved. The proposed model takes advantage of KBANN (Knowledge-Based Artificial Neural Network) which uses prior knowledge to learn NN (Neural Network) as training information. Each NN learns characteristics of the Feature Selection and Classification techniques as training sets. The integrated NN can be learned again to modify features appropriately and enhance classification performance. This innovative technique is called ALBNN (Algorithm Learning-Based Neural Network). The experiments' results show improved performance in various classification problems.

• 한국정밀공학회지
• /
• 제20권11호
• /
• pp.64-70
• /
• 2003

This paper suggests a systematic methodology to predict chip forms using the experimental design technique and the neural network. Significant factors determined with ANOVA analysis are used as input variables of the neural network back-propagation algorithm. It has been shown that cutting conditions and cutting tool shapes have distinct effects on the chip forms, so chip breaking. Cutting tools are represented using the Z-map method, which differs from existing methods using some chip breaker parameters. After training the neural network with selected input variables, chip forms are predicted and compared with original chip forms obtained from experiments under same input conditions, showing that chip forms are same at all conditions. To verify the suggested model, one tool not used in training the model is chosen and input to the model. Under various cutting conditions, predicted chip forms agree well with those obtained from cutting experiments. The suggested method could reduce the cost and time significantly in designing cutting tools as well as replacing the“trial-and-error”design method.

• 제어로봇시스템학회논문지
• /
• 제5권2호
• /
• pp.189-199
• /
• 1999

This paper presents an optimization algorithm for a stable Self Dynamic Neural Network(SDNN) using genetic algorithm. Optimized SDNN is applied to a problem of controlling nonlinear dynamical systems. SDNN is dynamic mapping and is better suited for dynamical systems than static forward neural network. The real-time implementation is very important, and thus the neuro controller also needs to be designed such that it converges with a relatively small number of training cycles. SDW has considerably fewer weights than DNN. Since there is no interlink among the hidden layer. The object of proposed algorithm is that the number of self dynamic neuron node and the gradient of activation functions are simultaneously optimized by genetic algorithms. To guarantee convergence, an analytic method based on the Lyapunov function is used to find a stable learning for the SDNN. The ability and effectiveness of identifying and controlling a nonlinear dynamic system using the proposed optimized SDNN considering stability is demonstrated by case studies.

• 한국산업정보학회논문지
• /
• 제2권1호
• /
• pp.73-87
• /
• 1997

Multilayer neural network using a modified beackpropagation learning algorithm was introduced to achieve automatic identification of different types of aircraft in a variety of 3-D orientations. A 3-D shape of an aircraft can be described by a library of 2-D images corresponding to the projected views of an aircraft. From each 2-D binary aircraft image we extracted 2-D invariant (L, Φ) feature vector to be used for training neural network aircraft classifier. Simulations concerning the neural network classification rate was compared using nearest-neighbor classfier (NNC) which has been widely served as a performance benchmark. And we also introduced reliability measure of the designed neural network classifier.

• 한국전자통신학회논문지
• /
• 제6권2호
• /
• pp.237-242
• /
• 2011

본 논문에서는 다층 퍼셉트론 신경회로망을 사용하여 각 프레임에서의 유성음, 무성음, 그리고 묵음 구간을 검출하는 구간검출 알고리즘을 제안한다. 다층 퍼셉트론 신경회로망의 입력으로는 고속 푸리에변환에 의한 전력스펙트럼 및 고속 푸리에변환 계수가 사용되어 네트워크가 학습된다. 본 실험에서는 원 음성에 백색잡음이 중첩된 음성을 신경회로망에 입력함으로서 각 프레임에서의 유성음, 무성음, 묵음 구간의 검출성능 결과를 나타낸다. 본 실험에서는 신경회로망의 학습 데이터 및 평가 데이터가 다를 경우에도 이러한 음성 및 백색잡음에 대하여 92% 이상의 검출율을 구할 수 있었다.

• 한국정밀공학회지
• /
• 제12권11호
• /
• pp.118-124
• /
• 1995

In the paper, we have proposed a new technique to determine the initial billet for the forged products using a function approximation in neural network. A three-layer neural network is used and a back propagation algorithm is employed to train the network. An optimal billet which satisfied the forming limitation, minimum of incomplete filling in the die cavity, load and energy as well as more uniform distribution of effective strain, is determined by applying the ability of function approximation of the neural network. The amount of incomplete filling in the die, load and forming energy as well as effective strain are measured by the rigid-plastic finite element method. This new technique is applied to find the optimal billet size for the axisymmetric rib-web product in hot forging. This would reduce the number of finite element simulation for determining the optimal billet of forging products, further it is usefully adopted to physical modeling for the forging design

• 한국공작기계학회논문집
• /
• 제17권1호
• /
• pp.77-85
• /
• 2008

The objective of the study is to construct a sensor fusion system for tool-condition monitoring (TCM) that will lead to a more efficient and economical drill usage. Drill-wear monitoring has an important attribute in the automatic machining processes as it can help preventing the damage of tools and workpieces, and optimizing the drill usage. In this study, we present the architectures of a multi-layer feed-forward neural network with Levenberg-Marquardt training algorithm based on sensor fusion for the monitoring of drill-wear condition. The input features to the neural networks were extracted from AE, vibration and current signals using the wavelet packet transform (WPT) analysis. Training and testing were performed at a moderate range of cutting conditions in the dry drilling of steel plates. The results show good performance in drill- wear monitoring by the proposed method of sensor fusion and neural network analysis.

• 대한임베디드공학회논문지
• /
• 제16권5호
• /
• pp.179-186
• /
• 2021

For stereo matching based on deep learning, the design of network structure is crucial to the calculation of matching cost, and the time-consuming problem of convolutional neural network in image processing also needs to be solved urgently. In this paper, a method of stereo matching using sparse loss volume in parallax dimension is proposed. A sparse 3D loss volume is constructed by using a wide step length translation of the right view feature map, which reduces the video memory and computing resources required by the 3D convolution module by several times. In order to improve the accuracy of the algorithm, the nonlinear up-sampling of the matching loss in the parallax dimension is carried out by using the method of multi-category output, and the training model is combined with two kinds of loss functions. Compared with the benchmark algorithm, the proposed algorithm not only improves the accuracy but also shortens the running time by about 30%.