• Journal of Korea Multimedia Society
• /
• v.23 no.8
• /
• pp.977-985
• /
• 2020

Aiming at the problem of complex feature extraction and low accuracy in human action recognition, this paper proposed a network structure combining batch normalization algorithm with GoogLeNet network model. Applying Batch Normalization idea in the field of image classification to action recognition field, it improved the algorithm by normalizing the network input training sample by mini-batch. For convolutional network, RGB image was the spatial input, and stacked optical flows was the temporal input. Then, it fused the spatio-temporal networks to get the final action recognition result. It trained and evaluated the architecture on the standard video actions benchmarks of UCF101 and HMDB51, which achieved the accuracy of 93.42% and 67.82%. The results show that the improved convolutional neural network has a significant improvement in improving the recognition rate and has obvious advantages in action recognition.

• Proceedings of the IEEK Conference
• /
• 2008.06a
• /
• pp.667-668
• /
• 2008

This paper demonstrates that the recurrent neural networks can be used successfully for the identification of finite automata (FAs). A new type of recurrent neural network (RNN) is proposed and the offline training algorithm, regulated Levenberg-Marquadt (LM) algorithm, for the network is developed. Simulation result shows that the identification and the extraction of FAs are practically achievable.

• Proceedings of the KAIS Fall Conference
• /
• 2006.11a
• /
• pp.186-190
• /
• 2006

"Dynamic Neural Unit"(DNU) based upon the topology of a reverberating circuit in a neuronal pool of the central nervous system. In this thesis, we present a genetic DNU-control scheme for unknown nonlinear systems. Our methodis different from those using supervised learning algorithms, such as the backpropagation (BP) algorithm, that needs training information in each step. The contributions of this thesis are the new approach to constructing neural network architecture and its trainin.

• Proceedings of the Korean Society for Noise and Vibration Engineering Conference
• /
• 2007.05a
• /
• pp.978-982
• /
• 2007

A new neuro-control scheme for active control of structures is proposed. It utilizes lattice pattern of state vector as training data of probabilistic neural network (PNN). Therefore, it is the so-called lattice probabilistic neural network (LPNN). PNN makes control forces by using all the training patterns. Therefore, it takes much time to obtain a control force in application. This inevitably may delay the control action. However, control force of LPNN is calculated by using only the adjacent information of LPNN input. So, the response of LPNN is greatly faster than PNN. The proposed control algorithm is applied for one story building under California and El Centro earthquakes. Also, control results of the LPNN are compared with those of the conventional PNN. The structural responses have been suppressed effectively by the proposed algorithm.

• JSTS:Journal of Semiconductor Technology and Science
• /
• v.6 no.3
• /
• pp.182-188
• /
• 2006

An artificial neural network model for the microwave characteristics of an InGaAs/InP hemt for 70 nm gate length has been developed. The small-signal microwave parameters have been evaluated to determine the transconductance and drain-conductance. We have further investigated the frequency characteristics of the device. The neural network training have been done using the three layer architecture using Levenberg-Marqaurdt Backpropagation algorithm. The results have been compared with the experimental data, which shows a close agreement and the validity of our proposed model.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.6
• /
• pp.505-514
• /
• 2020

Since a large amount of training data is typically needed to train Deep Neural Networks (DNNs), a parallel training approach is required to train the DNNs. The Stochastic Gradient Descent (SGD) algorithm is one of the most widely used methods to train the DNNs. However, since the SGD is an inherently sequential process, it requires some sort of approximation schemes to parallelize the SGD algorithm. In this paper, we review various efforts on parallelizing the SGD algorithm, and analyze the computational overhead, communication overhead, and the effects of the approximations.

• Proceedings of the KIEE Conference
• /
• 2003.11c
• /
• pp.425-428
• /
• 2003

In this paper, the stabilization of an inverted pendulum system is studied. Here, the PID control method is adopted to make the system stable. In order to adjust the PID gains, a three-layer neural network, which is based on the back propagation method, is used. Meanwhile, the time for training the neural network depends on the initial values of PID gains and connection weights. Hence, the genetic algorithm Is considered to shorten the time to find the desired values. Simulation results show the effectiveness of the proposed approach.

• MALSORI
• /
• no.48
• /
• pp.93-105
• /
• 2003

Since telephone channel has bandlimited frequency characteristics, speech signal over the telephone channel shows degraded speech quality. In this paper, we propose an algorithm using neural network to reconstruct wideband speech from its narrowband version. Although single neural network is a good tool for direct mapping, it has difficulty in training for vast and complicated data. To alleviate this problem, we modularize the neural networks based on appropriate clustering of the acoustic space. We also introduce fuzzy computing to compensate for probable misclassification at the cluster boundaries. According to our simulation, the proposed algorithm showed improved performance over the single neural network and conventional codebook mapping method in both objective and subjective evaluations.

• Journal of the Korean Society for Industrial and Applied Mathematics
• /
• v.22 no.2
• /
• pp.115-123
• /
• 2018

This paper discusses extended Kalman filter method for solving learning problems of multilayered neural networks. A lot of learning algorithms for deep layered network are sincerely suffered from complex computation and slow convergence because of a very large number of free parameters. We consider an efficient learning algorithm for deep neural network. Extended Kalman filter method is applied to parameter estimation of neural network to improve convergence and computation complexity. We discuss how an efficient algorithm should be developed for neural network learning by using Extended Kalman filter.

• Proceedings of the KIEE Conference
• /
• 1999.07g
• /
• pp.2932-2934
• /
• 1999

In this paper, a scheme for recognition of handwritten digits using a multilayer neural network trained with the back-propagation algorithm using generalized delta rule is proposed. The neural network is trained with hand written digit data of different writers and different styles. One of the purpose of the work with neural networks is the minimization of the mean square error(MSE) between actual output and desired one. The back-propagation algorithm is an efficient and very classical method. The back-propagation algorithm for training the weights in a multilayer net uses the steepest descent minimization procedure and the sigmoid threshold function. As an error rate is reduced, recognition rate is improved. Therefore we propose a method that is reduced an error rate.