• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.8 no.10
• /
• pp.3378-3393
• /
• 2014

We have witnessed the rapid development of information technology in recent years. One of the key phenomena is the fast, near-exponential increase of data. Consequently, most of the traditional data classification methods fail to meet the dynamic and real-time demands of today's data processing and analyzing needs--especially for continuous data streams. This paper proposes an improved incremental learning algorithm for a large-scale data stream, which is based on SVM (Support Vector Machine) and is named DS-IILS. The DS-IILS takes the load condition of the entire system and the node performance into consideration to improve efficiency. The threshold of the distance to the optimal separating hyperplane is given in the DS-IILS algorithm. The samples of the history sample set and the incremental sample set that are within the scope of the threshold are all reserved. These reserved samples are treated as the training sample set. To design a more accurate classifier, the effects of the data volumes of the history sample set and the incremental sample set are handled by weighted processing. Finally, the algorithm is implemented in a cloud computing system and is applied to study user behaviors. The results of the experiment are provided and compared with other incremental learning algorithms. The results show that the DS-IILS can improve training efficiency and guarantee relatively high classification accuracy at the same time, which is consistent with the theoretical analysis.

• The Journal of Korea Robotics Society
• /
• v.1 no.1
• /
• pp.66-72
• /
• 2006

In face recognition, learning speed of face is very important since the system should be trained again whenever the size of dataset increases. In existing methods, training time increases rapidly with the increase of data, which leads to the difficulty of training with a large dataset. To overcome this problem, we propose SVDD (Support Vector Domain Description)-based learning method that can learn a dataset of face rapidly and incrementally. In experimental results, we show that the training speed of the proposed method is much faster than those of other methods. Moreover, it is shown that our face recognition system can improve the accuracy gradually by learning faces incrementally at real environments with illumination changes.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.16 no.6
• /
• pp.736-741
• /
• 2006

The design of a classification system generally involves data acquisition module, learning module and decision module, considering their functions and it is often an important component of intelligent systems. The learning module provides a priori information and it has been playing a key role for the classification. The conventional learning techniques for classification are based on a winner take all fashion which does not reflect the description of real data where boundarues might be fuzzy Moreover they need all data for the learning of its problem domain. Generally, in many practical applications, it is not possible to prepare them at a time. In this paper, we design an adaptive classification model using incremental training fuzzy neural networks, FNN-I. To have a more useful information, it introduces the representation and membership degree by fuzzy theory. And it provides an incremental learning algorithm for continuously gathered data. We present tie experimental results on computer virus data. They show that the proposed system can learn incrementally and classify new viruses effectively.

• Journal of IKEEE
• /
• v.12 no.1
• /
• pp.65-74
• /
• 2008

One of the popular methods used for pattern classification is the MBR (Memory-Based Reasoning) algorithm. Since it simply computes distances between a test pattern and training patterns or hyperplanes stored in memory, and then assigns the class of the nearest training pattern, it is notorious for memory usage and can't learn additional information from new data. In order to overcome this problem, we propose an incremental learning algorithm (iMPA). iMPA divides the entire pattern space into fixed number partitions, and generates representatives from each partition. Also, due to the fact that it can not learn additional information from new data, we present iMPA which can learn additional information from new data and not require access to the original data, used to train. Proposed methods have been successfully shown to exhibit comparable performance to k-NN with a lot less number of patterns and better result than EACH system which implements the NGE theory using benchmark data sets from UCI Machine Learning Repository.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.1
• /
• pp.451-465
• /
• 2017

This paper presents a novel discriminative visual tracking algorithm with an adaptive incremental extreme learning machine. The parameters for an adaptive incremental extreme learning machine are initialized at the first frame with a target that is manually assigned. At each frame, the training samples are collected and random Haar-like features are extracted. The proposed tracker updates the overall output weights for each frame, and the updated tracker is used to estimate the new location of the target in the next frame. The adaptive learning rate for the update of the overall output weights is estimated by using the confidence of the predicted target location at the current frame. Our experimental results indicate that the proposed tracker can manage various difficulties and can achieve better performance than other state-of-the-art trackers.

• Proceedings of the KSRS Conference
• /
• v.2
• /
• pp.635-638
• /
• 2006

Regression of conventional prediction techniques in data mining uses the model which is generated from the training step. This model is applied to new input data without any change. If this model is applied directly to time series, the rate of prediction accuracy will be decreased. This paper proposes an incremental regression for time series prediction like typhoon track prediction. This technique considers the characteristic of time series which may be changed over time. It is composed of two steps. The first step executes a fractional process for applying input data to the regression model. The second step updates the model by using its information as new data. Additionally, the model is maintained by only recent data in a queue. This approach has the following two advantages. It maintains the minimum information of the model by using a matrix, so space complexity is reduced. Moreover, it prevents the increment of error rate by updating the model over time. Accuracy rate of the proposed method is measured by RME(Relative Mean Error) and RMSE(Root Mean Square Error). The results of typhoon track prediction experiment are performed by the proposed technique IMLR(Incremental Multiple Linear Regression) is more efficient than those of MLR(Multiple Linear Regression) and SVR(Support Vector Regression).

• The KIPS Transactions:PartB
• /
• v.13B no.2 s.105
• /
• pp.127-132
• /
• 2006

K-NN (k-Nearest Neighbors), which is a well-known instance-based learning algorithm, simply stores entire training patterns in memory, and uses a distance function to classify a test pattern. K-NN is proven to show satisfactory performance, but it is notorious formemory usage and lengthy computation. Various studies have been found in the literature in order to minimize memory usage and computation time, and NGE (Nested Generalized Exemplar) theory is one of them. In this paper, we propose RPA (Recursive Partition Averaging) and IRPA (Incremental RPA) which is an incremental version of RPA. RPA partitions the entire pattern space recursively, and generates representatives from each partition. Also, due to the fact that RPA is prone to produce excessive number of partitions as the number of features in a pattern increases, we present IRPA which reduces the number of representative patterns by processing the training set in an incremental manner. Our proposed methods have been successfully shown to exhibit comparable performance to k-NN with a lot less number of patterns and better result than EACH system which implements the NGE theory.

• The Journal of Information Technology and Database
• /
• v.8 no.1
• /
• pp.95-104
• /
• 2001

In the text classification domain, labeling the training documents is an expensive process because it requires human expertise and is a tedious, time-consuming task. Therefore, it is important to reduce the manual labeling of training documents while improving the text classifier. Selective sampling, a form of active learning, reduces the number of training documents that needs to be labeled by examining the unlabeled documents and selecting the most informative ones for manual labeling. We apply this methodology to Naive Bayes, a text classifier renowned as a successful method in text classification. One of the most important issues in selective sampling is to determine the criterion when selecting the training documents from the large pool of unlabeled documents. In this paper, we propose two measures that would determine this criterion : the Mean Absolute Deviation (MAD) and the entropy measure. The experimental results, using Renters 21578 corpus, show that this proposed learning method improves Naive Bayes text classifier more than the existing ones.

• ETRI Journal
• /
• v.32 no.3
• /
• pp.395-405
• /
• 2010

A novel multiclassifier system (MCS) strategy is proposed and applied to a text-dependent speaker verification task. The presented scheme optimizes the linear combination of classifiers on an on-line basis. In contrast to ordinary MCS approaches, neither a priori distributions nor pre-tuned parameters are required. The idea is to improve the most accurate classifier by making use of the incremental information provided by the second classifier. The on-line multiclassifier optimization approach is applicable to any pattern recognition problem. The proposed method needs neither a priori distributions nor pre-estimated weights, and does not make use of any consideration about training/testing matching conditions. Results with Yoho database show that the presented approach can lead to reductions in equal error rate as high as 28%, when compared with the most accurate classifier, and 11% against a standard method for the optimization of linear combination of classifiers.

• Proceedings of the Korean Information Science Society Conference
• /
• 2007.10c
• /
• pp.409-414
• /
• 2007

Active Appearance Model is a well-known model that can represent a non-rigid object effectively. However, since it uses the fixed appearance model, the fitting results are often unsatisfactory when the imaging condition of the target image is different from that of training images. To alleviate this problem, incremental AAM was proposed which updates its appearance bases in an on-line manner. However, it cannot deal with the sudden changes of illumination. To overcome this, we propose a novel scheme to update the appearance bases. When a new person appears in the input image, we synthesize illuminated images of that person and update the appearance bases of AAM using it. Since we update the appearance bases using synthesized illuminated images in advance, the AAM can fit their model to a target image well when the illumination changes drastically. The experimental results show that our proposed algorithm improves the fitting performance over both the incremental AAM and the original AAM.