• Journal of Korea Multimedia Society
• /
• v.6 no.1
• /
• pp.136-147
• /
• 2003

To achieve an optimal performance for a given problem, we need an integrative process of the parameter optimization via learning and the structure optimization via model selection. In this paper, we propose an efficient optimization method for improving generalization performance by considering the property of each sub-method and by combining them with common theoretical properties. First, weight parameters are optimized by natural gradient teaming with adaptive regularization, which uses a diverse error function. Second, the network structure is optimized by eliminating unnecessary parameters with natural pruning. Through iterating these processes, candidate models are constructed and evaluated based on the Bayesian Information Criterion so that an optimal one is finally selected. Through computational experiments on benchmark problems, we confirm the weight parameter and structure optimization performance of the proposed method.

• Journal of KIISE:Software and Applications
• /
• v.30 no.3_4
• /
• pp.251-258
• /
• 2003

A support vector machine (SVM) is supposed to provide a good generalization performance, but the actual performance of a actually implemented SVM is often far from the theoretically expected level. This is largely because the implementation is based on an approximated algorithm, due to the high complexity of time and space. To improve this limitation, we propose ensemble of SVMs by using Bagging (bootstrap aggregating) and Boosting. By a Bagging stage each individual SVM is trained independently using randomly chosen training samples via a bootstrap technique. By a Boosting stage an individual SVM is trained by choosing training samples according to their probability distribution. The probability distribution is updated by the error of independent classifiers, and the process is iterated. After the training stage, they are aggregated to make a collective decision in several ways, such ai majority voting, the LSE(least squares estimation) -based weighting, and double layer hierarchical combining. The simulation results for IRIS data classification, the hand-written digit recognition and Face detection show that the proposed SVM ensembles greatly outperforms a single SVM in terms of classification accuracy.

• Journal of Broadcast Engineering
• /
• v.18 no.4
• /
• pp.643-646
• /
• 2013

We present an improved AdaBoost algorithm to avoid overfitting phenomenon. AdaBoost is widely known as one of the best solutions for object detection. However, AdaBoost tends to be overfitting when a training dataset has noisy samples. To avoid the overfitting phenomenon of AdaBoost, the proposed method divides positive samples into K clusters using k-means algorithm, and then uses only one cluster to minimize the training error at each iteration of weak learning. Through this, excessive partitions of samples are prevented. Also, noisy samples are excluded for the training of weak learners so that the overfitting phenomenon is effectively reduced. In our experiment, the proposed method shows better classification and generalization ability than conventional boosting algorithms with various real world datasets.

• Journal of Convergence for Information Technology
• /
• v.7 no.3
• /
• pp.97-103
• /
• 2017

In this paper, we aim to investigate the generalization of ZRC codes by performing performance analysis on various coding rates and channel conditions of Zero Reduction Code (ZRC) for dimming control in VLC system. Simulation results show that the ZRC codes with relatively high coding rate achieve the enhanced BER performance as the RMS (Root Mean Square) delay spread value increases in the NLOS channelse. Therefore, it is expected that the designing of the dimming control technology using ZRC in the future visible light communication system will help to determine the coding rate according to the channel state through the analysis result in this study.

• The Korean Journal of Applied Statistics
• /
• v.17 no.1
• /
• pp.61-73
• /
• 2004

Rotation sampling designs may be classified into two categories. The first type uses the same sample unit for the entire life of the survey. The second type uses the sample unit only for a fixed number of times. In both type of designs, the entire sample is partitioned into a finite number(=G) of rotation groups. This paper is generalization of the first type designs. Since the generalized design can be identified by only G rotation groups and recall level 1, we denote this rotation system as l/G rotation design. Under l/G rotation design, variance and mean squared error (MSE) of generalized composite estimator are derived, incorporating two type of biases and exponentially decaying correlation pattern. Compromising MSE's of some selected l/G designs, we investigate design efficiency, design gap effect, ans the effects of correlation and bias.

• Journal of KIISE:Software and Applications
• /
• v.32 no.12
• /
• pp.1260-1270
• /
• 2005

Support Vector Machines (SVM) is well known for a representative learner as one of the kernel methods. SVM which is based on the statistical learning theory shows good generalization performance and has been applied to various pattern recognition problems. However, SVM is basically to deal with a two-class classification problem, so we cannot solve directly a multi-class problem with a binary SVM. One-Per-Class (OPC) and All-Pairs have been applied to solve the face recognition problem, which is one of the multi-class problems, with SVM. The two methods above are ones of the output coding methods, a general approach for solving multi-class problem with multiple binary classifiers, which decomposes a complex multi-class problem into a set of binary problems and then reconstructs the outputs of binary classifiers for each binary problem. In this paper, we introduce the output coding methods as an approach for extending binary SVM to multi-class SVM and propose new output coding schemes based on the Error-Correcting Output Codes (ECOC) which is a dominant theoretical foundation of the output coding methods. From the experiment on the face recognition, we give empirical results on the properties of output coding methods including our proposed ones.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.21 no.1
• /
• pp.87-92
• /
• 2021

Bias, variance, error and learning are important factors for performance in modeling a big data based recommendation system. The recommendation model in this system must reduce complexity while maintaining the explanatory diagram. In addition, the sparsity of the dataset and the prediction of the system are more likely to be inversely proportional to each other. Therefore, a product recommendation model has been proposed through learning the similarity between products by using a factorization method of the sparsity of the dataset. In this paper, the generalization ability of the model is improved by applying the max-norm regularization as an optimization method for the loss function of this model. The solution is to apply a stochastic projection gradient descent method that projects a gradient. The sparser data became, it was confirmed that the propsed regularization method was relatively effective compared to the existing method through lots of experiment.

• Investigative Magnetic Resonance Imaging
• /
• v.25 no.1
• /
• pp.10-22
• /
• 2021

Purpose: To overcome the difficulty in building a large data set with a high-quality in medical imaging, a concept of 'blended-transfer learning' (BTL) using a combination of both source data and target data is proposed for the target task. Materials and Methods: Source and target tasks were defined as training of the source and target networks to reconstruct cardiac CINE images from undersampled data, respectively. In transfer learning (TL), the entire neural network (NN) or some parts of the NN after conducting a source task using an open data set was adopted in the target network as the initial network to improve the learning speed and the performance of the target task. Using BTL, an NN effectively learned the target data while preserving knowledge from the source data to the maximum extent possible. The ratio of the source data to the target data was reduced stepwise from 1 in the initial stage to 0 in the final stage. Results: NN that performed BTL showed an improved performance compared to those that performed TL or standalone learning (SL). Generalization of NN was also better achieved. The learning curve was evaluated using normalized mean square error (NMSE) of reconstructed images for both target data and source data. BTL reduced the learning time by 1.25 to 100 times and provided better image quality. Its NMSE was 3% to 8% lower than with SL. Conclusion: The NN that performed the proposed BTL showed the best performance in terms of learning speed and learning curve. It also showed the highest reconstructed-image quality with the lowest NMSE for the test data set. Thus, BTL is an effective way of learning for NNs in the medical-imaging domain where both quality and quantity of data are always limited.

• Wind and Structures
• /
• v.36 no.6
• /
• pp.393-404
• /
• 2023

Skewness and kurtosis are important higher-order statistics for simulating non-Gaussian wind pressure series on low-rise buildings, but their predictions are less studied in comparison with those of the low order statistics as mean and rms. The distribution gradients of skewness and kurtosis on roofs are evidently higher than those of mean and rms, which increases their prediction difficulty. The conventional artificial neural networks (ANNs) used for predicting mean and rms show unsatisfactory accuracy in predicting skewness and kurtosis owing to the limited capacity of shallow learning of ANNs. In this work, the deep neural networks (DNNs) model with the ability of deep learning is introduced to predict the skewness and kurtosis on a low-rise building. For obtaining the optimal generalization of the DNNs model, the hyper parameters are automatically determined by Bayesian Optimization (BO). Moreover, for providing a benchmark for future studies on predicting higher order statistics, the data sets for training and testing the DNNs model are extracted from the internationally open NIST-UWO database, and the prediction errors of all taps are comprehensively quantified by various error metrices. The results show that the prediction accuracy in this study is apparently better than that in the literature, since the correlation coefficient between the predicted and experimental results is 0.99 and 0.75 in this paper and the literature respectively. In the untrained cornering wind direction, the distributions of skewness and kurtosis are well captured by DNNs on the whole building including the roof corner with strong non-normality, and the correlation coefficients between the predicted and experimental results are 0.99 and 0.95 for skewness and kurtosis respectively.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.1
• /
• pp.61-68
• /
• 2024

Despite the continuous efforts to mitigate pedestrian accidents at crosswalks, the problem persist. Vulnerable groups, including the elderly and disabled individuals are at a risk of being involved in traffic incidents. This paper proposes the implementation of object detection algorithm using the YOLO v5 model specifically for pedestrians using assistive devices like wheelchairs and crutches. For this research, data was collected and utilized through image crawling, Roboflow, and Mobility Aids datasets, which comprise of wheelchair users, crutch users, and pedestrians. Data augmentation techniques were applied to improve the model's generalization performance. Additionally, ensemble techniques were utilized to mitigate type 2 errors, resulting in 96% recall rate. This demonstrates that employing ensemble methods with a single YOLO model to target transportation-disadvantaged individuals can yield accurate detection performance without overlooking crucial objects.