-
Tomokazu Ueyama, Yasuaki Kuroe, Takehiro Mori
Session ID: 5007
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
It is known that some particular neurons in our brains produce various periodic patterns, such as locomotion, heart rhythm and so on. We have already proposed a method to synthesize spiking neural oscillators which generate periodic spike trains in spiking neural networks.In applying neural oscillators to engineering problems, their stability often become an important issue. Thus, we have proposed a method to analyze stability of periodic spike trains. In this paper, we propose a learning method to synthesize spiking neural oscillators with considering their stability.
View full abstract
-
Yoshiyuki Yagi, Keiji Tatsumi, Tetsuzo Tanino
Session ID: 5008
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Shogo Motoyashiki, Hirokazu Nishio, Yasushi Masuda, Shigehiko Kanaya, ...
Session ID: 5009
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Takaharu Onishi, Kazushi Ikeda, Noboru Murata
Session ID: 5010
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
Support Vector Machine (SVM) is a new classification technique which has a high generalization ability, however, it requires a heavy computational load since margin maximization results in a quadratic program. In this article, we show that the task results in a p-th order program if we employ the Lp norm instead of the L2 norm. This means that we can reduce the computational load by using the L1 norm. A geometrical view of SVM with the Lp norm explains that the change of p little affects the generalization performance of SVM in practice. This is confirmed by computer simulations.
View full abstract
-
Yukinori Nakamura, Yasuaki Kuroe, Takehiro Mori
Session ID: 5013
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Yu Mizoguchi, Tadahiko Murata
Session ID: 5014
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Shousuke Kimura, Seiichi Ozawa, Shigeo Abe
Session ID: 5016
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
We have proposed an intelligent face recognition system with the following two desirable features: one-pass incremental learning and automatic generation of training data. In this work, we mainly improve the computation time and robustness in the detection part of this system. Concretely, we introduce the mechanism of calculating area of the detected skin regions and deleting small regions from them. In our computer simulations, it is verified that the detection part is enhanced in speed and accuracy.
View full abstract
-
Miyoko Nakano, Fumiko Yasukata, Minoru Fukumi
Session ID: 5017
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Takuya Akashi, Minoru Fukumi, Norio Akamatsu
Session ID: 5018
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In this paper, as a new visual front end of audio-visual speech recognition systems, a simple method for lips region extraction is proposed. This method has robustness for varying shape and three-dimensional geometric changes by using only one image template. The image data is represented by x component (redness) of the Yxy colour space. The system environment is that a camera and human head run around separately for mobile devices. In our previous study, we tried to extract lips region, which has varying shape by speech, and the geometric changes of lips on two-dimension. Then, we extend this previous method to three-dimension by using projective geometry. In our simulations, we try to extract lips region from a face image with background, varying shape by speech, geometric changes, and face horizontal direction change. By this simulation, we can obtain high extraction accuracy.
View full abstract
-
Koichi Ikuta, Hiroaki Tanaka, Ken-ichi Tanaka, Kazuo Kyuma
Session ID: 5022
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
Most of the learning algorithms of the recognition system require specific training patterns for learning of the feature extraction filters. However, there are many cases that the recognition tasks don't have specific training patterns. In this paper, we propose a learning algorithm which reconstructs feature extraction filters on the basis of reinforcement signals. The recognition system constructed by the learning algorithm is robust against environmental variation.
View full abstract
-
Hiroaki Tanaka, Koichi Ikuta, Ken-ichi Tanaka, Kazuo Kyuma
Session ID: 5023
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
An algorithm which doesn't need exact information about recognized object is introduced in "Learning of the feature extraction filters by reinforcement signals". The algorithm uses reinforcement signal and acquires feature extraction filter automatically. It is effective against environmental variation and image recognition which includes unclear feature, because it is difficult to construct feature extraction method exactly. In this paper, we propose a learning recognition system with the use of the algorithm assuming two-class discrimination issue using some images.
View full abstract
-
Hiroshi Matsumoto, Tadahiko Murata
Session ID: 5024
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Kazunori Iwata, Kazushi Ikeda, Hideaki Sakai
Session ID: 5025
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In reinforcement learning, the return maximization is characterized bytwo factors, a stochastic complexity and a quantity depending on theparameters of environment.By changing the parameters of action selection strategy, we cancontrol the stochastic complexity but not the quantity.Hence, we examine the sensitivity of stochastic complexity, which isuseful in appropriately tuning the parameters.
View full abstract
-
Kazunori HANADA, daisuke KUSHIDA, Akira KITAMURA, Satoshi NISHIZONO, M ...
Session ID: 5026
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Tomomi Sugie, Hajime Murao, Hisashi Tamaki, Shinzo Kitamura
Session ID: 5027
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This paper proposes a method for AIBO, a four-legged robot, to acquire asmooth walking pattern automatically. In our approach for pattern acquisition, we prepare a variety ofAIBO's postures as templates, and apply Q-learning to find a sequenceof postures which realizes a smooth walk.Here the templates contribute to reducethe search space. The usefulness of the proposed method is evaluatedthrough experiments.
View full abstract
-
Yuki Miyaura, Yoshio Mogami, Kenji Ikeda, Takao Shimomura
Session ID: 5028
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Katsuhiro Honda, Kazumori Tomita, Hidetomo Ichihashi
Session ID: 5029
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
Local linear modeling is a useful tool for knowledge discovery in databases and handling missing values is an important issue in real world application. It is an effective strategy to ignore only the missing values and use all the remaining elements of data matrices, and linear fuzzy clustering with missing values has been proposed, in which missing values are ignored by using least squares criterion based on lower rank approximation of data matrices. This paper considers extending the fuzzy clustering technique to the case where the missing values do not arise randomly. Two types of additional weights are introduced to the least squares criterion to adjust the effects of the number of the missing values. One is the weight for samples and the other is that for attributes. Numerical experiment shows that the weights are useful for handling missing values that do not arise randomly.
View full abstract
-
Katsuhiro Honda, Hidetomo Ichihashi
Session ID: 5030
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
Fuzzy c-Lines (FCL) is a technique for partitioning a data set into several linear clusters whose prototypes are lines. This paper considers the linear fuzzy clustering of data with uncertainties based on intervals. The clustering criterion is defined by considering two different metrics, Minimum Distance and Maximum Distance, and the optimal prototypes are estimated by using a linear search algorithm. Numerical example shows that the result of the proposed method provides a tool for interpretation of local features of the data with uncertainties.
View full abstract
-
Katsuhiro Honda, Yoshihito Nakamura, Taichi Makise, Hidetomo Ichihashi
Session ID: 5031
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
Simultaneous approaches to fuzzy clustering and quantification of categorical data have been applied to the KDD (Knowledge Discovery in Databases). In the approaches, a data set is partitioned considering similarities of samples or variables. This paper shows an analysis of questionnaire data about information technology using a simultaneous application of homogeneity analysis and fuzzy clustering. Using statistical table together with the plots of the feature values derived by the local quantification method, characteristic features of large scale categorical data sets are revealed intuitively.
View full abstract
-
Masahiro Takatani, Tomoharu Nakashima, Masayo Udo, Hisao Ishibuchi
Session ID: 5032
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In this paper, we show applications of fuzzy rule-based systems to soccer agents. The task for them is to pass the ball to each other without being captured it by an opponent agent. That is, agents should reach quickly to the ball when receiving it, and consider pass timing in terms of an opponent's situation when passing.We define the former situation as the ball intercept problem, and the latter as the ball pass/dribble problem.We have proposed the learning method for the ball intercept problem. In this paper, we first introduce this method for acquiring the behavior.problem, and improve the behavior of soccer agents using another fuzzy rule-based system that turn angles. Then the ball/pass problem is described. Through computational experiments, we conclude that the learning agents can successfully acquire the behavioraccording to situations.
View full abstract
-
Hiroko Kitano, Tomoharu Nakashima, Hisao Ishibuchi
Session ID: 5033
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Kaname Narukawa, Hisao Ishibuchi
Session ID: 5034
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
The aim of this paper is to propose a simple multiobjective genetic local search (S-MOGLS) algorithm with high search ability and to evaluate its performance. First we explain our S-MOGLS algorithm, which can be easily understood, easily implemented and efficiently excuted with small memory storage and short CPU time. Then we examine the performance of our S-MOGLS algorithm in comparison with the NSGA-II through computational experiments on multiobjective 0/1 knapsack problems.
View full abstract
-
Takeru Yokoi, Hidekazu Yanagimoto, Sigeru Omatu
Session ID: 5035
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Masashi Okubo, Tadahiko Murata
Session ID: 5037
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Yuji Matsumura, Minoru Fukumi, Norio Akamatsu, Fumiaki Takeda
Session ID: 5038
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This paper tries to recognize EMG signals by neural networks. The electrodes under the dry state are attached to wrists and then EMG is measured. These EMG signals are classified into seven categories, such as neutral, up and down, right and left, twist to inside, twist to outside by neural networks. The NN learns FFT spectra to classify them. Moreover, we structuralized NN using multiple PCA for accuracy improvement of the network. It is shown that our approach is effective to classify the EMG signals by means of computer simulations.
View full abstract
-
Yuichi Sasaki
Session ID: 5040
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Kouji Urawaki, Yasushi Masuda, Yoshitsugu Manabe, Kunihiro Chihara
Session ID: 5041
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In learning of the correct movement in sport, which is necessary for the improvement of skill acquiring, both the correct positioning of the body and muscular activities are important. However, conventional VR-based learning systems lack representation for intensities and/or timing of muscular activities. This paper proposes a system for form measurement/learning which visualizes muscular activities derived from Electromyographic(EMG) data. The fundamental experiment examined analysis method for muscular intensity, and performed visualization in CG avator.
View full abstract
-
Akio Gofuku, Akira Yamasaki, Kazuhiko Mitsuoka, Hiroyuki Hashizume
Session ID: 5042
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This study investigates a method to represent the knowledge of disease diagnosis procedure of doctors as a decision tree by the RDR (Ripple Down Rule) technique and to reconstruct it to an effective procedure by the ID3 (Iterative Dichotomiser 3) technique. In the reconstruction, the clinical system is considered as restrictions. The proposed method is implemented as an interface system by the Java. The applicability of the proposed method is examined by sample problems for disease diagnosis of hands.
View full abstract
-
Yasunobu Maeda, Norio Kurita
Session ID: 5043
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
To manage new emerging environmental and technological risks, for example, food safety risks such as BSE, and health risks of electro magnetic fields, it is necessary to catch early signals of the risks, and to originate warning at the early stage. For this purpose, a clearinghouse for food safety risks was designed and contents of the clearinghouse was collected from the WWW by using natural language processing techniques. The collected contents were reviewed by experts in food safety risk analysis.
View full abstract
-
Mitsunori MIKI, Tomoyuki HIROYASU, Rintaro NAGANO
Session ID: 5044
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Toshiyuki SAKAI, Chisato KIHARA, Masaki IDA, Masuzo YANAGIDA
Session ID: 5045
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Yusuke Takaguchi, Tsutomu Inamoto, Hisashi Tamaki, Hajime Murao, Shinz ...
Session ID: 5048
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Chitomi Kadowaki, Daisuke Kushida, Akira Kitamura, Nobuyuki Tomochika, ...
Session ID: 5049
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Jung's collective unconscious approach
joe imae, nobuyuki otsuki, tomoaki kobayashi
Session ID: 5050
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
When it comes to the minimax controller design, it would be extremely difficult to obtain such controllers in the nonlinear situations. One of the reasons is that the minimax controller should be robust against any kind of disturbances. Based on the genetic programming and Jung's collective unconscious, this paper presents a very simple design technique to solve the minimax control problems, where the minimax controller may be constructed paying attention only one kind of disturbance in a sense. In addition, several simulations/experiments are given to demonstrate the usefulness of the proposed design technique with H-inf control problems, including inverted pendulum problems.
View full abstract
-
Hitoshi Iima
Session ID: 5051
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In this paper, a method based on the particle swarm optimization method is proposed for solving flow shop scheduling problems. In the proposed method, a solution is represented by a product sequence. A solution is updated by copying a subsequence of the best solution in the solutions searched. The effectiveness of the proposed method is investigated through the computational result.
View full abstract
-
Shinsuke Oh-hara, Kenji Hirta, Hiroshi Minemura, Kei-ichi Kamimura
Session ID: 5053
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Hayato Nakada, Kiyotsugu Takaba, Tohru Katayama
Session ID: 5055
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This paper is concerned with the recursive identification of piecewise affine systems.The method to be presented is composed of incremental clustering of the data,recursive least squares estimation of the parameter corresponding to each cluster,and least square support vector machineto approximately estimate boundary hyperplanesbetween two adjacent polyhedral regions on the regression space.
View full abstract
-
Hideki Nishigaito, Shigeru Yamamoto
Session ID: 5056
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
We consider state boundedness of a system with a state reset mechanism in which some of the elements of the state vector are reseted to zero whenever the output exceeds a given threshold. We explicitly derive a bound of the state for the case where the system realization is given as the observable canonical form.
View full abstract
-
Kiminao Kogiso, Kenji Hirata
Session ID: 5057
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This paper investigates experimental evaluations of controller switching strategies for systems with state and control magnitude constraints. For the goal, we implement the switching rule and a number of static feedback controllers, which are desinged in advance, to a personal computer.Performing an experimental control of a regulation for a state to zeros, we report the effectiveness of the switching control method and an issue about the management and implementation.
View full abstract
-
Izumi Masubuchi, Seiji Yabuki
Session ID: 5058
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
This paper is concerned with L
2-gain analysis of hybrid dynamical systems. By extending Sum of Squares (SOS) to a matrix-valued one, we formulate LMI problems to solve matrix inequalities with functional unknowns. Numerical examples will be provided.
View full abstract
-
Masataka Imura, Yusuke Sakan, Yoshihiro Yasumuro, Yoshitsugu Manabe, K ...
Session ID: 6001
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
K. Fukunaka, M. Imura, Y. Yasumuro, Y. Manabe, K. Chihara
Session ID: 6002
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Tadayuki TAKEDA, Hiroshi SASAKI, Masataka IMURA, Yoshihiro YASUMURO, Y ...
Session ID: 6003
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
With development of immersive projection technique, display quality is worth to perform practical use.On the other hand, we need to interact with the displayed contents freely for effective task in immersive environments.We consider that natural interaction between the user and the virtual objects can be achieved by intuitive motions of body parts including hands and fingers to operate objects directly as we do in daily life.In this paper, we propose the intuitive interface which enables users to operate virtual environments using body action recognized by motion capture system.
View full abstract
-
Arito Mochizuki, Takashi Amada, Kazuhiro Kohyama, Sayuri Sawa, Tadayuk ...
Session ID: 6004
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In VR field, olfactory information is important to improve reality. But when some scented air is emitted by using olfactory displays, it is difficult to control the scent after the emission. To display multiple scent, thus, it is necessary to display the scent with spatio-temporal locality. For this purpose, we introduce an olfactory display system which includes a mechanism that estimates the appropriate time of user's sniffing.
View full abstract
-
–use of haptic device–
Masahiro Kono
Session ID: 6005
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
For learning the Japanese character, it is desirable that we master how to write correctly through handwriting. Students must study the way of handwriting by observing the teacher's model sequence of handwriting in order to acquiring the writing skills of letter shape and stroke order. However, our Japanese learning environment is not sufficient for individual learning such as the teacher guides each student. In this report, we describe the handwriting learning system that the learner can get the feeling of brush movement in handwriting. The system shows the brush sequential movement of the model character through force feedback pen (haptic device).
View full abstract
-
Hisanao Otsuka, Masataka Imura, Yoshihiro Yasumuro, Yoshitsugu Manabe, ...
Session ID: 6006
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
In this work, we present a modeling method that utilises images of actual clothes and is applicable for virtual fitting rooms. Virtual fitting rooms can work only if well prepared clothes models are available. Thus, it is important to acquire varieties of clothes models by simple process. First, we estimate shape features of the clothes using images. Then, the features are used to fit basic geometry models. Finally, texture in the cloth image is merged to the models based on the applied feature points. This paper demonstrate feature estimation using only 2 images of front and back side of each cloth.
View full abstract
-
Yukio Tada, Yusuke Kawata
Session ID: 6007
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
It has been disputed whether "Ogawabon-Shinsousenjimon" wasan originally written work or a work made by unknown copying technique, "transparency copy".One of factors for the judgement is the strength of stroke and press in brush movement.In this paper, through the examination of the outlines of characters with the image processing,we compare cracked state of characters in "Ogawabon-Shinsousenjimon" with those of two old works; one is known as an originally written work and the other is known that it was written by "transparency copy".It is concluded from the comparizon of chain codes thatThe feature of "Sinsousenjimon" is different form that ofthe "Souranjou" which is known as the work by "transparency copy".
View full abstract
-
Daisuke MASUI, Yuta SAKOGUCHI, Tetsuo SHIMADA
Session ID: 6008
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Yuji Takabe, Motonori Doi
Session ID: 6010
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
-
Hiroyuki Fujioka, Hiroyuki Kano
Session ID: 6011
Published: 2004
Released on J-STAGE: September 01, 2004
CONFERENCE PROCEEDINGS
FREE ACCESS
A problem of designing cursive characters is considered. Based on the generation method of "dynamic font", a novel redesigning scheme is developed by employing a dynamic model of robotic arm as the generation mechanism of new characters. This scheme is used to transform the typeface of predesigned Japanese characters to the omitted running style. A design example is included.
View full abstract