• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2000.05a
• /
• pp.126-129
• /
• 2000

In order to approximate a nonlinear function, wavelet neural networks combining wavelet theory and neural networks have been proposed as an alterantive to coventional multi-layered neural networks. Wavelet neural networks provide better approximating performance than conventional neural networks. In this paper, an effective method to construct an optimal wavelet neural network is proposed using genetic algorithm. This is verified through experimental results.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2003.09a
• /
• pp.158-161
• /
• 2003

A cascade structured neural network called Sigma-Pi$_{t}$ Cascaded Hybrid Neural Network ($\sigma$$\pi$$_{t}$-CHNN) is Proposed. It is an extended version of the Sigma-Pi Cascaded extended Hybrid Neural Network ($\sigma$$\pi$-CHNN), where the classical multiplicative neuron ($\pi$-neuron) is replaced by the translated multiplicative ($\pi$$_{t}$-neuron) model. The learning algorithm of $\sigma$$\pi$$_{t}$-CHNN is composed of an evolutionary programming method, responsible for determining the network architecture, and of a Levenberg-Marquadt algorithm, responsible for tuning the weights of the network. The $\sigma$$\pi$$_{t}$-CHNN is evaluated in 2 pattern classification problems: the 2 spirals and the sonar problems. In the 2 spirals problem, $\sigma$$\pi$$_{t}$-CHNN can generate neural networks with 10％ less hidden neurons than that in previous neural models. In the sonar problem, $\sigma$$\pi$$_{t}$-CHNN can find the optimal solution for the problem i.e., a network with no hidden neurons. These results confirm the expanded information processing capabilities of $\sigma$$\pi$$_{t}$-CHNN, when compared to previous neural network models. network models.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.13 no.1
• /
• pp.6-11
• /
• 2003

Recently, the modular network was proposed in a way to keep the size of the neural network small. The modular network solves the problem by splitting it into sub-problems. In this aspect, fuzzy systems act in a similar way. However, in a fuzzy system, there must be an expert rule which separates the input space. To overcome this, fuzzy-neural network has been used. However, the number of fuzzy rules grows exponentially as the number of input variables grow. In this paper, we would like to solve the size problem of neural networks using modular network with the hierarchic structure. In the hierarchic structure, the output of precedent module affects only the THEN part of the rule. Finally, the rules become shorter being compared to the rule of fuzzy-neural system. Also, the relations between input and output could be understood more easily in the Proposed modular network and that makes design easier.

• Journal of Internet Computing and Services
• /
• v.19 no.4
• /
• pp.45-52
• /
• 2018

A single-image super-resolution is a process of restoring a high-resolution image from a low-resolution image. Recently, the super-resolution using the deep neural network has shown good results. In this paper, we propose a neural network structure that improves speed and performance over conventional neural network based super-resolution methods. To do this, we analyze the conventional neural network based super-resolution methods and propose solutions. The proposed method reduce the 5 stages of the conventional method to 3 stages. Then we have studied the optimal width and depth by experimenting on the width and depth of the network. Experimental results have shown that the proposed method improves the disadvantages of the conventional methods. The proposed neural network structure showed superior performance and speed than the conventional method.

• Proceedings of the Korea Committee for Ocean Resources and Engineering Conference
• /
• 2001.10a
• /
• pp.91-96
• /
• 2001

This paper presents a neural net based nonlinear adaptive controller for an autonomous underwater vehicle (AUV). AUV's dynamics are highly nonlinear and their hydrodynamic coefficients vary with different operational conditions, so it is necessary for the high performance control system of an AUV to have the capacities of learning and adapting to the change of the AUV's dynamics. In this paper a linearly parameterized neural network is used to approximate the uncertainties of the AUV's dynamics, and a sliding mode control is introduced to attenuate the effects of the neural network's reconstruction errors and the disturbances of AUV's dynamics. The presented controller is consist of three parallel schemes; linear feedback control, sliding mode control and neural network. Lyapunov theory is used to guarantee the asymptotic convergence of trajectory tracking errors and the neural network's weights errors. Numerical simulations for motion control of an AUV are performed to illustrate to effectiveness of the proposed techniques.

• Proceedings of the Korea Inteligent Information System Society Conference
• /
• 2001.01a
• /
• pp.116-121
• /
• 2001

Prediction of the top(service) speeds of high-speed trains and configuration design to trainset of them has been studied using the neural network system The traction system. The traction system of high-speed trains is composed of transformers, motor blocks, and traction motors of which locations and number in the trainset formation should be determine in the early stage of train conceptural design. Components of the traction system are the heaviest parts in a train so that it gives strong influence to the top speeds of high-speed trains. Prediction of the top speeds has been performed mainly with data associated with the traction system based on the frequently used neural network system-backpropagation. The neural network has been trained with the data of the high-speed trains such as TGV, ICE, and Shinkanse. Configuration design of the trainset determines the number of trains motor cars, traction motors, weights and power of trains. Configuration results from the neural network are more accurate if neural networks is trained with data of the same type of trains will be designed.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.39 no.4
• /
• pp.6-15
• /
• 2002

A neural network disturbance observer for a robotic manipulator is derived in this paper. The neural network used as the disturbance observer is a feedforward MLP(multiple-layered perceptron) network. The uniform ultimate boundness(UUB) of the proposed neural disturbance observer and the control error within a sufficiently small compact set is guaranteed. This neural disturbance observer method overcomes the disadvantages of the existing adaptive control methods which require the tedious analysis of the regressor matrix of the given manipulator. The effectiveness of the proposed neural disturbance observer is demonstrated by the application to the three-link robotic manipulator.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.17 no.4
• /
• pp.472-476
• /
• 2007

Learning rule affects importantly the performance of neural network. This paper proposes a new fuzzy learning rule that uses the learning rate considering the distance between the input vector and the prototypes of classes. When the learning rule updates the prototypes of classes, this consideration reduces the effect of outlier on the prototypes of classes. This comes from making the effect of the input vector, which locates near the decision boundary, larger than an outlier. Therefore, it can prevents an outlier from deteriorating the decision boundary. This new fuzzy learning rule is integrated into IAFC(Integrated Adaptive Fuzzy Clustering) fuzzy neural network. Iris data set is used to compare the performance of the proposed fuzzy neural network with those of other supervised neural networks. The results show that the proposed fuzzy neural network is better than other supervised neural networks.

• Journal of Information Processing Systems
• /
• v.17 no.2
• /
• pp.337-351
• /
• 2021

Automatically recognizing facial expressions in video sequences is a challenging task because there is little direct correlation between facial features and subjective emotions in video. To overcome the problem, a video facial expression recognition method using spatiotemporal recurrent neural network and feature fusion is proposed. Firstly, the video is preprocessed. Then, the double-layer cascade structure is used to detect a face in a video image. In addition, two deep convolutional neural networks are used to extract the time-domain and airspace facial features in the video. The spatial convolutional neural network is used to extract the spatial information features from each frame of the static expression images in the video. The temporal convolutional neural network is used to extract the dynamic information features from the optical flow information from multiple frames of expression images in the video. A multiplication fusion is performed with the spatiotemporal features learned by the two deep convolutional neural networks. Finally, the fused features are input to the support vector machine to realize the facial expression classification task. The experimental results on cNTERFACE, RML, and AFEW6.0 datasets show that the recognition rates obtained by the proposed method are as high as 88.67%, 70.32%, and 63.84%, respectively. Comparative experiments show that the proposed method obtains higher recognition accuracy than other recently reported methods.

• The Journal of Korean Association of Computer Education
• /
• v.6 no.3
• /
• pp.47-56
• /
• 2003

We describe a new method for removing non-linear phased array antenna aberration called "squint" problem. To develop a compensation scheme. theoretical antenna and artificial neural networks were used. The purpose of using the artificial neural networks is to develop an antenna system model that represents the steering function of an actual array. The artificial neural networks are also used to implement an inverse model which when concatenated with the antenna or antenna model will correct the "squint" problem. Combining the actual steering function and the inverse model contained in the artificial neural network, alters the steering command to the antenna so that the antenna will point to the desired position instead of squinting. The use of an artificial neural network provides a method of producing a non-linear system that can correct antenna performance. This paper demonstrates the feasibility of generating an inverse steering algorithm with artificial neural networks.