• Journal of Intelligence and Information Systems
• /
• v.5 no.1
• /
• pp.103-123
• /
• 1999

Input filtering as a preprocessing method is so much crucial to get good performance in time series forecasting. There are a few preprocessing methods (i.e. ARMA outputs as time domain filters, and Fourier transform or wavelet transform as time-frequency domain filters) for handling time series. Specially, the time-frequency domain filters describe the fractal structure of financial markets better than the time domain filters due to theoretically additional frequency information. Therefore, we, first of all, try to describe and analyze specially some issues on the effectiveness of different filtering methods from viewpoint of the performance of a neural network based forecasting. And then we discuss about neural network model architecture issues, for example, what type of neural network learning architecture is selected for our time series forecasting, and what input size should be applied to a model. In this study an input selection problem is limited to a size selection of the lagged input variables. To solve this problem, we simulate on analyzing and comparing a few neural networks having different model architecture and also use an embedding dimension measure as chaotic time series analysis or nonlinear dynamic analysis to reduce the dimensionality (i.e. the size of time delayed input variables) of the models. Throughout our study, experiments for integration methods of joint time-frequency analysis and neural network techniques are applied to a case study of daily Korean won / U. S dollar exchange returns and finally we suggest an integration framework for future research from our experimental results.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.22 no.1
• /
• pp.1-8
• /
• 2018

Recently, the convolutional neural network shows high performance in many computer vision tasks. However, convolutional neural networks require enormous amount of operation, so it is difficult to adopt them in the embedded environments. To solve this problem, many studies are performed on the ASIC or FPGA implementation, where an efficient representation method is required. The fixed-point representation is adequate for the ASIC or FPGA implementation but causes a performance degradation. This paper proposes a separate optimization of representations for the convolutional layers and the batch normalization layers. With the proposed method, the required bit width for the convolutional layers is reduced from 16 bits to 10 bits for the ResNet-50 neural network. Since the computation amount of the convolutional layers occupies the most of the entire computation, the bit width reduction in the convolutional layers enables the efficient implementation of the convolutional neural networks.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.5
• /
• pp.568-573
• /
• 2019

In the diagnosis of pulmonary diseases, auscultation technique is simpler than the other methods, and lung sounds can be used for predicting the types of pulmonary diseases as well as identifying patients with pulmonary diseases. Therefore, in this paper, we identify patients with pulmonary diseases and classify lung sounds according to their sound characteristics using various convolutional neural networks, and compare the classification performance of each neural network method. First, lung sounds over affected areas of the chest with pulmonary diseases are collected by using a single-channel lung sound recording device, and spectral features are extracted from the collected sounds in time domain and applied to each neural network. As classification methods, we use general, parallel, and residual convolutional neural network, and compare lung sound classification performance of each neural network through experiments.

• Wind and Structures
• /
• v.2 no.1
• /
• pp.25-40
• /
• 1999

In this paper, artificial neural networks, a new kind of intelligent method, are employed to model and predict amplitude dependent damping in buildings based on our full-scale measurements of buildings. The modelling method and procedure using neural networks to model the damping are studied. Comparative analysis of different neural network models of damping, which includes multi-layer perception network (MLP), recurrent neural network, and general regression neural network (GRNN), is performed and discussed in detail. The performances of the models are evaluated and discussed by tests and predictions including self-test, "one-lag" prediction and "multi-lag" prediction of the damping values at high amplitude levels. The established models of damping are used to predict the damping in the following three ways : (1) the model is established by part of the data measured from one building and is used to predict the another part of damping values which are always difficult to obtain from field measurements : the values at the high amplitude level. (2) The model is established by the damping data measured from one building and is used to predict the variation curve of damping for another building. And (3) the model is established by the data measured from more than one buildings and is used to predict the variation curve of damping for another building. The prediction results are discussed.

• Imaging Science in Dentistry
• /
• v.51 no.3
• /
• pp.237-242
• /
• 2021

Purpose: The aim of this study was to analyse and review deep learning convolutional neural networks for detecting and diagnosing early-stage dental caries on periapical radiographs. Materials and Methods: In order to conduct this review, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses(PRISMA) guidelines were followed. Studies published from 2015 to 2021 under the keywords(deep convolutional neural network) AND (caries), (deep learning caries) AND (convolutional neural network) AND (caries) were systematically reviewed. Results: When dental caries is improperly diagnosed, the lesion may eventually invade the enamel, dentin, and pulp tissue, leading to loss of tooth function. Rapid and precise detection and diagnosis are vital for implementing appropriate prevention and treatment of dental caries. Radiography and intraoral images are considered to play a vital role in detecting dental caries; nevertheless, studies have shown that 20% of suspicious areas are mistakenly diagnosed as dental caries using this technique; hence, diagnosis via radiography alone without an objective assessment is inaccurate. Identifying caries with a deep convolutional neural network-based detector enables the operator to distinguish changes in the location and morphological features of dental caries lesions. Deep learning algorithms have broader and more profound layers and are continually being developed, remarkably enhancing their precision in detecting and segmenting objects. Conclusion: Clinical applications of deep learning convolutional neural networks in the dental field have shown significant accuracy in detecting and diagnosing dental caries, and these models hold promise in supporting dental practitioners to improve patient outcomes.

• Journal of the Institute of Convergence Signal Processing
• /
• v.22 no.3
• /
• pp.99-103
• /
• 2021

Recently, convolutional neural networks (CNNs) have been widely utilized due to their outstanding performance in various computer vision fields. However, due to their computational-intensive and high memory requirements, it is difficult to deploy CNNs on hardware platforms that have limited resources, such as mobile devices and IoT devices. To address these limitations, a neural network compression research is underway to reduce the size of neural networks while maintaining their performance. This paper proposes a CNN compression technique that dynamically adjusts the thresholds of pruning, one of the neural network compression techniques. Unlike the conventional pruning that experimentally or heuristically sets the thresholds that determine the weights to be pruned, the proposed technique can dynamically find the optimal thresholds that prevent accuracy degradation and output the light-weight neural network in less time. To validate the performance of the proposed technique, the LeNet was trained using the MNIST dataset and the light-weight LeNet could be automatically obtained 1.3 to 3 times faster without loss of accuracy.

• Design & Manufacturing
• /
• v.17 no.4
• /
• pp.72-78
• /
• 2023

Injection molding is a process widely used in various industries because of its high production speed and ease of mass production during the plastic manufacturing process, and the product is molded by injecting molten plastic into the mold at high speed and pressure. Since process conditions such as resin and mold temperature mutually affect the process and the quality of the molded product, it is difficult to accurately predict quality through mathematical or statistical methods. Recently, studies to predict the quality of injection molded products by applying artificial neural networks, which are known to be very useful for analyzing nonlinear types of problems, are actively underway. In this study, structural optimization of neural networks was conducted by applying multi-task learning techniques according to the characteristics of the input and output parameters of the artificial neural network. A structure reflecting the characteristics of each process step was applied to the input parameters, and a structure reflecting the quality characteristics of the injection molded part was applied to the output parameters using multi-tasking learning. Building an artificial neural network to predict the three qualities (mass, diameter, height) of injection-molded product under six process conditions (melt temperature, mold temperature, injection speed, packing pressure, pacing time, cooling time) and comparing its performance with the existing neural network, we observed enhancements in prediction accuracy for mass, diameter, and height by approximately 69.38%, 24.87%, and 39.87%, respectively.

• KSCE Journal of Civil and Environmental Engineering Research
• /
• v.44 no.3
• /
• pp.385-396
• /
• 2024

To enhance the accuracy of sinkhole detection using GPR, this study derived a convolutional neural network that can optimally extract sinkhole characteristics from GPR B-scan grayscale images. The pre-trained convolutional neural network is evaluated to be more than twice as effective as the vanilla convolutional neural network. In pre-trained convolutional neural networks, fast feature extraction is found to cause less overfitting than feature extraction. It is analyzed that the top-1 verification accuracy and computation time are different depending on the type of architecture and simulation conditions. Among the pre-trained convolutional neural networks, InceptionV3 are evaluated as most robust for sinkhole detection in GPR B-scan grayscale images. When considering both top-1 verification accuracy and architecture efficiency index, VGG19 and VGG16 are analyzed to have high efficiency as the backbone for extracting sinkhole feature from GPR B-scan grayscale images. MobileNetV3-Large backbone is found to be suitable when mounted on GPR equipment to extract sinkhole feature in real time.

• Journal of Intelligence and Information Systems
• /
• v.5 no.1
• /
• pp.95-101
• /
• 1999

In this paper, a comparison between different methods to combine predictions from neural networks will be given. These methods are bagging, bumping, and balancing. Those are based on the analysis of the ensemble generalization error into an ambiguity term and a term incorporating generalization performances of individual networks. Neural Networks and AI machine learning models are prone to overfitting. A strategy to prevent a neural network from overfitting, is to stop training in early stage of the learning process. The complete data set is spilt up into a training set and a validation set. Training is stopped when the error on the validation set starts increasing. The stability of the networks is highly dependent on the division in training and validation set, and also on the random initial weights and the chosen minimization procedure. This causes early stopped networks to be rather unstable: a small change in the data or different initial conditions can produce large changes in the prediction. Therefore, it is advisable to apply the same procedure several times starting from different initial weights. This technique is often referred to as training ensembles of neural networks. In this paper, we presented a comparison of three statistical methods to prevent overfitting of neural network.

• Proceedings of the Korean Society of Precision Engineering Conference
• /
• 1996.11a
• /
• pp.759-763
• /
• 1996

This paper presents a new approach to the design of neural control system using digital signal processors in order to improve the precision and robustness. Robotic manipulators have become increasingly important in the field of flexible automation. High speed and high-precision trajectory tracking are indispensable capabilities for their versatile application. The need to meet demanding control requirement in increasingly complex dynamical control systems under significant uncertainties, leads toward design of intelligent manipulation robots. The TMS320C31 is used in implementing real time neural control to provide an enhanced motion control for robotic manipulators. In this control scheme, the networks introduced are neural nets with dynamic neurons, whose dynamics are distributed over all the network nodes. The nets are trained by the distributed dynamic back propagation algorithm. The proposed neural network control scheme is simple in structure, fast in computation, and suitable for implementation of real-time control. Performance of the neural controller is illustrated by simulation and experimental results for a SCARA robot.