• The KIPS Transactions:PartB
• /
• v.11B no.7 s.96
• /
• pp.849-854
• /
• 2004

This paper proposes an algorithm to enhance the execution time of kNN in the document classification. The proposed algorithm is to enhance the execution time by minimizing the computing cost of the similarity between two documents by using the list of pairs, while the conventional kNN uses the iist of pairs. The 1ist of pairs can be obtained by applying the matrix transposition to the list of pairs at the training phase of the document classification. This paper analyzed the proposed algorithm in the time complexity and compared it with the conventional kNN. And it compared the proposed algorithm with the conventional kNN by using routers-21578 data experimentally. The experimental results show that the proposed algorithm outperforms kNN about $90{\%}$ in terms of the ex-ecution time.

• Journal of the Institute of Convergence Signal Processing
• /
• v.5 no.3
• /
• pp.205-209
• /
• 2004

The K-means algorithm is the most widely used method among the codebook generation algorithms in vector quantization. In this paper, we propose a codebook generation algorithm using a new updating condition to enhance the codebook performance. The conventional K-means algorithm uses a fixed weight of the distance for all training iterations, but the proposed method uses different weights according to the updating condition from the new codevectors for training iterations. Then, different weights can be applied to generate codevectors at each iteration according to this condition, and it can have a similar effect to variable weights. Experimental results show that the proposed algorithm has the better codebook performance than that of K-means algorithm.

• 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.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.1
• /
• pp.308-320
• /
• 2016

Although the face almost always has an axisymmetric structure, it is generally not symmetrical image for the face image. However, the mirror image of the face image can reflect possible variation of the poses and illumination opposite to that of the original face image. A robust minimum squared error classification (RMSEC) algorithm is proposed in this paper. Concretely speaking, the original training samples and the mirror images of the original samples are taken to form a new training set, and the generated training set is used to perform the modified minimum sqreared error classification(MMSEC) algorithm. The extensive experiments show that the accuracy rate of the proposed RMSEC is greatly increased, and the the proposed RMSEC is not sensitive to the variations of the parameters.

• The Journal of the Korea Contents Association
• /
• v.16 no.5
• /
• pp.104-112
• /
• 2016

The Adaboost chooses a good set of features in rounds. On each round, it chooses the optimal feature and its threshold value by minimizing the weighted error of classification. The involved process of classification performs a hard decision. In this paper, we expand the process of classification to a soft fuzzy decision. We believe this expansion could allow some flexibility to the Adaboost algorithm as well as a good performance especially when the size of a training data set is not large enough. The typical Adaboost algorithm assigns a same weight to each training datum on the first round of a training process. We propose a new algorithm to assign different initial weights based on some statistical properties of involved features. In experimental results, we assess that the proposed method shows higher performance than the traditional one.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.32B no.5
• /
• pp.747-755
• /
• 1995

When codebooks designed by a clustering algorithm using training sets, a time-averaged distortion, which is called the inside-training-set- distortion (ITSD), is usually calculated in each iteration of the algorithm, since the input probability function is unknown in general. The algorithm stops if the ITSD no more significantly decreases. Then, in order to test the trained codebook, the outside-training-set-distortion (OTSD) is to be calculated by a time-averaged approximation using the test set. Hence codebooks that yield small values of the OTSD are regarded as good codebooks. In other words, the calculation of the OTSD is a criterion to testing a trained codebook. But, such an argument is not always true if some conditions are not satisfied. Moreover, in order to obtain an approximation of the OTSD using the test set, it is known that a large test set is requared in general. But, large test set causes heavy calculation com0plexity. In this paper, from the analyses in [16], it has been revealed that the enough size of the test set is only the same as that of the codebook when codebook size is large. Then a simple method to testing trained codebooks is addressed. Experimental results on synthetic data and real images supporting the analysis are also provided and discussed.

• Journal of KIISE:Software and Applications
• /
• v.32 no.9
• /
• pp.912-919
• /
• 2005

When we apply binary classification to multi-class classification for text categorization, we use the One-Against-All method generally, However, this One-Against-All method has a problem. That is, documents of a negative set are not labeled by human. Thus, they can include many noisy documents in the training data. In this paper, we propose that the Sliding Window technique and the EM algorithm are applied to binary text classification for solving this problem. We here improve binary text classification through extracting noise documents from the training data by the Sliding Window technique and re-assigning categories of these documents using the EM algorithm.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.10 no.7
• /
• pp.1287-1294
• /
• 2006

In many current wireless packet data system, the short-burst transmissions are used, and training overhead is very significant for such short burst formats. So, the availability of the short training sequence and the fast converging algorithm is essential in the adaptive equalizer. In this paper, the new equalizer algorithm is proposed to improve the performance of a MTLMS (multiple-training least mean square) based DFE (decision feedback equalizer)using the short training sequence. In the proposed method, the output of the DFE is fed back to the LMS (least mean square) based adaptive DEF loop iteratively and used as an extended training sequence. Instead of the block operation using ML (maximum likelihood) estimator, the low-complexity adaptive LMS operation is used for overall processing. Simulation results show that the perfonnance of the proposed equalizer is improved with a linear computational increase as the iterations parameter in creases and can give the more robustness to the time-varying fading.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.46 no.2
• /
• pp.19-26
• /
• 2009

This paper studies on the recognition rate change with respect to illumination variance and the confidence level of PCA(Principal Component Analysis) based face recognition by measuring the cumulative match score of CMC(Cumulative Match Characteristic). We studied on the confidence level of the algorithm under illumination changes and selection of training images not only by testing multiple training images per person with illumination variance and single training image and but also by changing the illumination conditions of testing images. The experiment shows that the recognition rate drops for multiple training image case compared to single training image case. We, however, confirmed the confidence level of the algorithm under illumination variance by the fact that the training image which corresponds to the identity of testing image belongs to upper similarity lists regardless of illumination changes and the number of training images.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.49 no.2
• /
• pp.53-60
• /
• 2012

Nearest-neighbor classification predicts the class of an input data with the most frequent class among the near training data of the input data. Even though nearest-neighbor classification doesn't have a training stage, all of the training data are necessary in a predictive stage and the generalization performance depends on the quality of training data. Therefore, as the training data size increase, a nearest-neighbor classification requires the large amount of memory and the large computation time in prediction. In this paper, we propose a prototype selection algorithm that predicts the class of test data with the new set of prototypes which are near-boundary training data. Based on Tomek links and distance metric, the proposed algorithm selects boundary data and decides whether the selected data is added to the set of prototypes by considering classes and distance relationships. In the experiments, the number of prototypes is much smaller than the size of original training data and we takes advantages of storage reduction and fast prediction in a nearest-neighbor classification.