• Phonetics and Speech Sciences
• /
• v.7 no.4
• /
• pp.27-33
• /
• 2015

This paper proposes an audio event classification method using Deep Neural Networks (DNN). The proposed method applies Feed Forward Neural Network (FFNN) to generate event probabilities of ten audio events (dog barks, engine idling, and so on) for each frame. For each frame, mel scale filter bank features of its consecutive frames are used as the input vector of the FFNN. These event probabilities are accumulated for the events and the classification result is determined as the event with the highest accumulated probability. For the same dataset, the best accuracy of previous studies was reported as about 70% when the Support Vector Machine (SVM) was applied. The best accuracy of the proposed method achieves as 79.23% for the UrbanSound8K dataset when 80 mel scale filter bank features each from 7 consecutive frames (in total 560) were implemented as the input vector for the FFNN with two hidden layers and 2,000 neurons per hidden layer. In this configuration, the rectified linear unit was suggested as its activation function.

• Proceedings of the KIEE Conference
• /
• 1989.07a
• /
• pp.679-684
• /
• 1989

In this paper, we describe the PTP notion of robot manipulators by neural networks. The PTP motion requires the inverse kinematic redline and the joint trajectory generation algorithm. We use the multi-layered Perceptron neural networks and the Error Back Propagation(EBP) learning rule for inverse kinematic problems. Varying the number of hidden layers and the neurons of each hidden layer, we investigate the performance of the neural networks. Increasing the number of learning sweeps, we also discuss the performance of the neural networks. We propose a method for solving the inverse kinematic problems by adding the error compensation neural networks(ECNN). And, we implement the neural networks proposed by Grossberg et al. for automatic trajectory generation and discuss the problems in detail. Applying the neural networks to the current trajectory generation problems, we can refute the computation time for trajectory generation.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2001.05a
• /
• pp.65-68
• /
• 2001

RBFN has some problem that because the basis function isnt orthogonal to each others the number of used basis function goes to big. In this reason, the Wavelet Neural Network which uses the orthogonal basis function in the hidden node appears. In this paper, we propose the composition method of the actual function in hidden layer with the scaling function which can represent the region by which the several wavelet can be represented. In this method, we can decrease the size of the network with the pure several wavelet function. In addition to, when we determine the parameters of the scaling function we can process rough approximation and then the network becomes more stable. The other wavelets can be determined by the global solutions which is suitable for the suggested problem using the genetic algorithm and also, we use the back-propagation algorithm in the learning of the weights. In this step, we approximate the target function with fine tuning level. The complex neural network suggested in this paper is a new structure and important simultaneously in the point of handling the determination problem in the wavelet initialization.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.22 no.5
• /
• pp.1026-1035
• /
• 1997

In this paper, we studied for phoneme recognition using GPFN and PNN as a kind of RBFN. The structure of RBFN is similar to a feedforward networks but different from choosing of activation function, reference vector and learnign algorithm in a hidden layer. Expecially sigmoid function in PNN is replaced by one category included exponential function. And total calculation performance is high, because PNN performs pattern classification with out learning. In phonemerecognition experiment with 5 vowel and 12 consant, recognition rates of GPFN and PNN as a kind of RBFN reflected statistic characteristic of speech are higher than ones of MLP in case of using test data and quantizied data by VQ and LVQ.

• Proceedings of the Korean Institute of Building Construction Conference
• /
• 2022.11a
• /
• pp.223-224
• /
• 2022

The occurrence of defect, one of the major risk elements, gives rise to construction delays and additional costs. Although construction companies generally prefer to use a method of identifying and classifying the causes of defects, a system for predicting the rise of defects becomes important matter to reduce this harmful issue. However, the currently used methods are kinds of reactive systems that are focused on the defects which occurred already, and there are few studies on the occurrence of defects with prediction systems. This paper is about preliminary study on the development of judgemental algorithm that informs us whether additional works related to defect issue are needed or not. Among machine learning techniques, deep neural network was utilized as prediction model which is a major component of algorithm. It is the most suitable model to be applied to the algorithm when there are 8 hidden layers and the average number of nodes in each hidden layer is 70. Ultimately, the algorithm can identify and defects that may arise in later and contribute to minimize defect frequency.

• Communications for Statistical Applications and Methods
• /
• v.30 no.5
• /
• pp.501-516
• /
• 2023

Knowledge distillation (KD) refers to extracting knowledge from a large and complex model (teacher) and transferring it to a relatively small model (student). This can be done by training the teacher model to obtain the activation function values of the hidden or the output layers and then retraining the student model using the same training data with the obtained values. Recently, relational KD (RKD) has been proposed to extract knowledge about relative differences in training data. This method improved the performance of the student model compared to conventional KDs. In this paper, we propose a new method for RKD by introducing a new loss function for RKD. The proposed loss function is defined using the area difference between the teacher model and the student model in a specific hidden layer, and it is shown that the model can be successfully compressed, and the generalization performance of the model can be improved. We demonstrate that the accuracy of the model applying the method proposed in the study of model compression of audio data is up to 1.8% higher than that of the existing method. For the study of model generalization, we demonstrate that the model has up to 0.5% better performance in accuracy when introducing the RKD method to self-KD using image data.

• 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.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.2
• /
• pp.988-1001
• /
• 2018

Arrhythmia has recently emerged as one of the major causes of death in Koreans. Premature Ventricular Contraction (PVC) is the most common arrhythmia that can be found in clinical practice, and it may be a precursor to dangerous arrhythmias, such as paroxysmal insomnia, ventricular fibrillation, and coronary artery disease. Therefore, we need for a method that can detect an abnormal heart beat and diagnose arrhythmia early. We extracted the features corresponding to the QRS pattern from the subject's ECG signal and classify the premature ventricular contraction waveform using the features. We modified the weighting and bias values based on the error back-propagation algorithm through learning data. We classify the normal signal and the premature ventricular contraction signal through the modified weights and deflection values. MIT-BIH arrhythmia data sets were used for performance tests. We used RR interval, QS interval, QR amplitude and RS amplitude features. And the hidden layer with two nodes is composed of two layers to form a total three layers (input layer 0, output layer 3).

• Asian Pacific Journal of Cancer Prevention
• /
• v.15 no.18
• /
• pp.7775-7780
• /
• 2014

This paper addresses cancer prediction based on radial basis function neural network optimized by particle swarm optimization. Today, cancer hazard to people is increasing, and it is often difficult to cure cancer. The occurrence of cancer can be predicted by the method of the computer so that people can take timely and effective measures to prevent the occurrence of cancer. In this paper, the occurrence of cancer is predicted by the means of Radial Basis Function Neural Network Optimized by Particle Swarm Optimization. The neural network parameters to be optimized include the weight vector between network hidden layer and output layer, and the threshold of output layer neurons. The experimental data were obtained from the Wisconsin breast cancer database. A total of 12 experiments were done by setting 12 different sets of experimental result reliability. The findings show that the method can improve the accuracy, reliability and stability of cancer prediction greatly and effectively.

• Journal of Welding and Joining
• /
• v.13 no.3
• /
• pp.77-88
• /
• 1995

A prediction method for determining the welding residual stress by artificial neural network is proposed. A three-dimensional transient thermomechanical analysis has been performed for the CO$_{2}$ arc welding using the finite element method. The first part of numerical analysis performs a three-dimensional transient heat transfer analysis, and the second part then uses the results of the first part and performs a three-dimensional transient thermo-elastic-plastic analysis to compute transient and residual stresses in the weld. Data from the finite element method are used to train a backpropagation neural network to predict the residual stress. Architecturally, the fully interconnected network consists of an input layer for the voltage and current, a hidden layer to accommodate the ailure mechanism mapping, and an output layer for the residual stress. The trained network is then applied to the prediction of residual stress in the four specimens. It is concluded that the accuracy of the neural network predicting method is fully comparable with the accuracy achieved by the traditional predicting method.