• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.33B no.10
• /
• pp.90-97
• /
• 1996

This paper presents an efficient parallel implementation and its performance evaluations of a SOFM neural netowrk on the multicomputer. We investigate the parallel performance as the size of a neural network N, the number of the patterns L, and the number of the processors p increase. We propose an analytica performance evaluation model for eac of the parallel implementations and verified the validity of the model through experiments. Analytical result show that the number of processors for a maximum speedup of the network decomposition nd the training-set decomposition increases in proportion to .root.N and .root.L, respectively. The performances of the both decompositions depend on the number of training patterns L and the size of the neural network N and, if L.geq.0.423N, the performance of trhe training-set decomposition is proved to be better than that of the network decomposition.

• Journal of Institute of Control, Robotics and Systems
• /
• v.8 no.4
• /
• pp.338-344
• /
• 2002

In this paper, neural network force tracking control is proposed. The conventional impedance function is reformulated to have direct farce tracking capability. Neural network is used to compensate for all the uncertainties such as unknown robot dynamics, unknown environment stiffness, and unknown environment position. On line training signal of farce error for neural network is formulated. A large x-y table is built as a test-bed and neural network loaming algorithm is implemented on a DSP board mounted in a PC. Experimental studies of farce tracking on unknown environment for x-y table robot are presented to confirm the performance of the proposed technique.

• IE interfaces
• /
• v.10 no.3
• /
• pp.145-153
• /
• 1997

This paper deals with constructing a scheduling system for a mold manufacturing factory. The scheduling system is composed of 4 submodules such as pre-processor, neural network training, neural networks and simulation. Pre-processor analyzes the condition of workshop and generates input data to neural networks. Network training module is performed by using the condition of workshop, performance measures, and dispatching rules. Neural networks module presents the most optimized dispatching rule, based on previous training data according to the current condition of workshop. Simulation module predicts the earliest completion date of a mold by forward scheduling with the presented dispatching rules, and suggests a possible issue date of a material by backward tracking. The system developed shows a great potential when applied in real mold factory for automotive parts.

• Journal of the Institute of Convergence Signal Processing
• /
• v.24 no.2
• /
• pp.90-96
• /
• 2023

Securing large amounts of training data in deep learning neural networks, including convolutional neural networks, is of importance for avoiding overfitting phenomenon or for the excellent performance. However, securing labeled training data in deep learning neural networks is very limited in reality. To overcome this, several augmentation methods have been proposed in the literature to generate an additional large amount of training data through transformation or manipulation of the already acquired traing data. However, unlike training data such as images and texts, it is barely to find an augmentation method in the literature that additionally generates bio-signal training data for convolutional neural network based human activity recognition. Thus, this study proposes a simple but effective augmentation method of bio-signal training data for convolutional neural network based human activity recognition. The usefulness of the proposed augmentation method is validated by showing that human activity is recognized with high accuracy by convolutional neural network trained with its augmented bio-signal training data.

• Transactions of the Korean Society of Mechanical Engineers
• /
• v.16 no.7
• /
• pp.1322-1331
• /
• 1992

The conventional neural networks are built without considering the underlying structure of the problems. Hence, they usually contain redundant weights and require excessive training time. A novel neural network structure is proposed for symmetric problems, which alleviate some of the aforementioned drawback of the conventional neural networks. This concept is expanded to that of the constrained neural network which may be applied to general structured problems. Because these neural networks can not be trained by the conventional training algorithm, which destroys the weight structure of the neural networks, a proper training algorithm is suggested. The illustrative examples are shown to demonstrate the applicability of the proposed idea.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.44 no.2 s.314
• /
• pp.11-18
• /
• 2007

Back propagation algorithm based on a gradient-decent method has been widely used to the training of a neural network. However, this algorithm have some problems such as dropping the minimum value in a local area according to an initial value and setting the number of units in a hidden layer when training the neural network. Accordingly, to solve the above-mentioned problems, this paper proposes a genetic algorithm using the training method of the neural network. Thus, the improved genetic algorithm using a new crossover and mutation method is proposed to discriminate 3 bit parity. Experiments confirm that the proposed system is effective for training speed after demonstrating for generation gap, the number of units in the hidden layer, and the number of individuals.

• Proceedings of the IEEK Conference
• /
• 2003.07d
• /
• pp.1251-1254
• /
• 2003

In this paper, Experience Sensitive Cumulative Neural Network (ESCNN) is introduced, which can cumulate the same or similar experiences. As the same or similar training patterns are cumulated in the network, the system recognize more important information in the training patterns. The functions of forgetting less important informations and attending more important informations resided in the training patterns are surveyed and implemented by simulations. The system behaves well under the noisy circumstances due to its forgetting and/or attending properties, even in 50 percents noisy environments. This paper also describes the creation of the generalized patterns for the input training patterns.

• Journal of Advanced Information Technology and Convergence
• /
• v.8 no.2
• /
• pp.119-127
• /
• 2018

Training a very large deep neural network can be painfully slow and prone to overfitting. Many researches have done for overcoming the problem. In this paper, a combination of early stopping and ADAM based deep neural network was presented. This form of deep network is useful for handling the big data because it automatically stop the training before overfitting occurs. Also generalization ability is better than pure deep neural network model.

• International Journal of Computer Science & Network Security
• /
• v.21 no.2
• /
• pp.77-87
• /
• 2021

In this study, a novel improved second order Radial Basis Function Neural Network based method with excellent scheduling capabilities is used for the dynamic prediction of short and long-term energy required applications. The effectiveness and the reliability of the algorithm are evaluated using training operations with New England-ISO database. The dynamic prediction algorithm is implemented in Matlab and the computation of mean absolute error and mean absolute percent error, and training time for the forecasted load, are determined. The results show the impact of temperature and other input parameters on the accuracy of solar Photovoltaic load forecasting. The mean absolute percent error is found to be between 1% to 3% and the training time is evaluated from 3s to 10s. The results are also compared with the previous studies, which show that this new method predicts short and long-term load better than sigmoidal neural network and bagged regression trees. The forecasted energy is found to be the nearest to the correct values as given by England ISO database, which shows that the method can be used reliably for short and long-term load forecasting of any electrical system.

• Journal of the Korean Society of Manufacturing Technology Engineers
• /
• v.6 no.1
• /
• pp.27-33
• /
• 1997

A neural network approach has been developed to determine the depth of a surface breaking crack in a steel plate from ultrasonic backscattering data. The network is trained by the use of feedforward three-layered network together with a back-scattering algorithm for error correction. The signal used for crack insonification is a mode converted 70$^{\circ}$transverse wave. A numerical analysis of back scattered field is carried out based on elastic wave theory, by the use of the boundary element method. The numerical data are calibrated by comparison with experimental data. The numerical analysis provides synthetic data for the training of the network. The training data have been calculated for cracks with specified increments of the crack depth. The performance of the network has been tested on other synthetic data and experimental data which are different from the training data.