• The Journal of the Acoustical Society of Korea
• /
• v.36 no.4
• /
• pp.267-272
• /
• 2017

In this paper, we propose an effective method of applying multichannel-audio feature values to GRNNs (Gated Recurrent Neural Networks) in polyphonic sound event detection. Real life sounds are often overlapped with each other, so that it is difficult to distinguish them by using a mono-channel audio features. In the proposed method, we tried to improve the performance of polyphonic sound event detection by using multi-channel audio features. In addition, we also tried to improve the performance of polyphonic sound event detection by applying a gated recurrent neural network which is simpler than LSTM (Long Short Term Memory), which shows the highest performance among the current recurrent neural networks. The experimental results show that the proposed method achieves better sound event detection performance than other existing methods.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.12 no.4
• /
• pp.685-690
• /
• 2008

In numerically evaluating the thermal performance of the heat exchanger, numerical values of thermodynamic properties such as temperature, pressure, specific volume, enthalpy and entropy are required. But the steam table or diagram itself cannot be directly used without modelling. In this study the applicability of neural networks in modelling superheated water vapor was examined. The multi-layer neural networks consist of an input layer with 2 nodes, two hidden layers with 15 and 25 nodes respectively and an output layer with 3 nodes. Quadratic spline interpolation was also applied for comparison. Neural networks model revealed smaller percentage error compared with spline interpolation. From this result, it is confirmed that the neural networks could be a powerful method in modelling the superheated water vapor.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.535-538
• /
• 2022

For Plug & Play of IoT devices, we develop a module that recognizes the type of USB, which is a typical wired interface of IoT devices, through image recognition. In order to drive an IoT device, a driver for communication and device hardware is required. The wired interface for connecting to the IoT device is recognized by using the image obtained through the camera of smartphone shooting to recognize the corresponding communication interface. For USB, which is a most popular wired interface, types of USB are classified through artificial neural network-based machine learning. In order to secure sufficient data set of artificial neural networks, USB images are collected through the Internet, and additional image data sets are secured through image processing. In addition to the convolution neural networks, recognizers are implemented with various deep artificial neural networks, and their performance is compared and evaluated.

• Smart Media Journal
• /
• v.11 no.3
• /
• pp.26-30
• /
• 2022

Spiking neural network is a neural network that applies the working principle of real brain neurons. Due to the biological mechanism of neurons, it consumes less power for training and reasoning than conventional neural networks. Recently, as deep learning models become huge and operating costs increase exponentially, the spiking neural network is attracting attention as a third-generation neural network that connects convolution neural networks and recurrent neural networks, and related research is being actively conducted. However, in order to apply the spiking neural network model to the industry, a lot of research still needs to be done, and the problem of model retraining to apply a new model must also be solved. In this paper, we propose a method to minimize the cost of model retraining by extracting the weights of the existing trained deep learning model and converting them into the weights of the spiking neural network model. In addition, it was found that weight conversion worked correctly by comparing the results of inference using the converted weights with the results of the existing model.

• Journal of the Institute of Convergence Signal Processing
• /
• v.2 no.1
• /
• pp.44-52
• /
• 2001

This paper describes an off-line method for recognizing totally unconstrained handwritten digits using modified chaotic neural networks(MCNN). The chaotic neural networks(CNN) is modified to be a useful network for solving complex pattern problems by enforcing dynamic characteristics and learning process. Since the MCNN has the characteristics of highly nonlinear dynamics in structure and neuron itself, it can be an appropriate network for the robust classification of complex handwritten digits. Digit identification starts with extraction of features from the raw digit images and then recognizes digits using the MCNN based classifier. The performance of the MCNN classifier is evaluated on the numeral database of Concordia University, Montreal, Canada. For the relative comparison of recognition performance, the MCNN classifier is compared with the recurrent neural networks(RNN) classifier. Experimental results show that the classification rate is 98.0%. It indicates that the MCNN classifier outperforms the RNN classifier as well as other classifiers that have been reported on the same database.

• KSCE Journal of Civil and Environmental Engineering Research
• /
• v.30 no.4B
• /
• pp.399-412
• /
• 2010

The goal of this research is to apply the neural networks models for the temporal disaggregation of the yearly pan evaporation (PE) data, Republic of Korea. The neural networks models consist of multilayer perceptron neural networks model (MLP-NNM) and generalized regression neural networks model (GRNNM), respectively. And, for the performances evaluation of the neural networks models, they are composed of training and test performances, respectively. The three types of data such as the historic, the generated, and the mixed data are used for the training performance. The only historic data, however, is used for the testing performance. From this research, we evaluate the application of MLP-NNM and GRNNM for the temporal disaggregation of nonlinear time series data. We should, furthermore, construct the credible monthly PE data from the temporal disaggregation of the yearly PE data, and can suggest the available data for the evaluation of irrigation and drainage networks system.

• Structural Engineering and Mechanics
• /
• v.18 no.3
• /
• pp.303-314
• /
• 2004

An artificial neural network (ANN) application is presented for flexural and axial vibration analysis of elastic beams with various support conditions. The first three natural frequencies of beams are obtained using multi layer neural network based back-propagation error learning algorithm. The natural frequencies of beams are calculated for six different boundary conditions via direct solution of governing differential equations of beams and Rayleigh's approximate method. The training of the network has been made using these data only flexural vibration case. The trained neural network, however, had been tested for cantilever beam (C-F), and both end free (F-F) in case the axial vibration, and clamped-clamped (C-C), and Guided-Pinned (G-P) support condition in case the flexural vibrations which were not included in the training set. The results found by using artificial neural network are sufficiently close to the theoretical results. It has been demonstrated that the artificial neural network approach applied in this study is highly successful for the purposes of free vibration analysis of elastic beams.

• Journal of Institute of Control, Robotics and Systems
• /
• v.9 no.11
• /
• pp.917-924
• /
• 2003

Recurrent neural networks have been successfully applied to communications channel equalization. Major disadvantages of gradient-based learning algorithms commonly employed to train recurrent neural networks are slow convergence rates and long training sequences required for satisfactory performance. In a high-speed communications system, fast convergence speed and short training symbols are essential. We propose decision feedback equalizers using a recurrent neural network trained with Kalman filtering algorithms. The main features of the proposed recurrent neural equalizers, utilizing extended Kalman filter (EKF) and unscented Kalman filter (UKF), are fast convergence rates and good performance using relatively short training symbols. Experimental results for two time-varying channels are presented to evaluate the performance of the proposed approaches over a conventional recurrent neural equalizer.

• Proceedings of the Korean Society for Technology of Plasticity Conference
• /
• 1994.10a
• /
• pp.133-140
• /
• 1994

In this paper, we have proposed a new method to determine the initial billet for the forged products using a function approximation in neural networks. the architecture of neural network is a three-layer neural network and the back propagation algorithm is employed to train the network. By utilizing the ability of function approximation of neural network, an optimal billet is determined by applying nonlinear mathematical relationship between shape ratio in the initial billet and the final products. A volume of incomplete filling in the die is measured by the rigid-plastic finite element method. The neural network is trained with the initial billet shape ratio and that of the un-filled volume. After learning, the system is able to predict the filling region which are exactly the same or slightly different to results of finite element method. It is found that the prediction of the filling shape ratio region can be made successfully and the finite element method results are represented better by the neural network.

• Transactions on Control, Automation and Systems Engineering
• /
• v.2 no.3
• /
• pp.189-195
• /
• 2000

The mathematical solutions of the stability convergence are important problems in system control. In this paper such problems are analyzed and resolved for system control using multilayer neural networks. We describe an algorithm to control an unknown nonlinear system with a disturbance, using a multilayer neural network. We include a disturbance among the modeling error, and the weight update rules of multilayer neural network are derived to satisfy Lyapunov stability. The overall control system is based upon the feedback linearization method. The weights of the neural network used to approximate a nonlinear function are updated by rules derived in this paper . The proposed control algorithm is verified through computer simulation. That is as the weights of neural network are updated at every sampling time, we show that the output error become finite within a relatively short time.