|
1. |
An asynchronous parallel branch‐and‐bound algorithm |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 1-13
Tsuyoshi Kawaguchi,
Hiroshi Masuyama,
Tamotsu Maeda,
Preview
|
PDF (975KB)
|
|
摘要:
AbstractThis paper presents a parallel branch‐and‐bound algorithm which is applicable to a loosely coupled multiprocessor with nonhierarchical interconnection network such as torus or hypercube. This algorithm is asynchronous and processing elements (PEs) start evaluation of nodes without being synchronized. Thus, when each PE finishes evaluation of one node, it can immediately start the evaluation of the next node.First, we compared the performance for parallel branch‐and‐bound algorithms which have almost the same communication overhead. From the simulation results it was confirmed that when the proposed algorithm is applied to a torus machine, an efficiency (speed‐up rate divided by the number of PEs) is obtained which is about twice better than that of the existing synchronous algorithm for the improved tree machine and about 1.4 times better than that of the synchronous algorithm for a torus machine proposed by the authors. Interconnection networks among PEs suitable for parallelization of branch‐and‐bound algorithms were investigated by applying the algorithm in this paper to several kinds of interconnection networks among PEs.As the results, it was found that even in the case in which communication time between adjacent PEs has very little effect on the execution time of an algorithm, a hypercube machine realizes almost the same efficiency as that of an interconnection network with a larger
ISSN:0882-1666
DOI:10.1002/scj.4690230401
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
2. |
A neural network model of the dynamics of a short‐term memory system in the temporal cortex |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 14-25
Masahiko Morita,
Preview
|
PDF (881KB)
|
|
摘要:
AbstractThe temporal area TE of the monkey contains a group of neurons which seems to act as the short‐term memory by a sustained firing, where the stored information seems to be retained by a certain kind of equilibrium of a dynamical system. The behavior of this group of neurons cannot be accounted for by the traditional dynamical system realized by the existing neural network models.This paper discusses the short‐term memory system in the temporal area, as well as the dynamics of associative memory, and proposes the neural network model which realizes the same dynamical behavior as that of area TE. This model differs essentially from the traditional models in that the composing unit has non‐monotonic input/output characteristics and a significant property in analyzing the behavior and the mechanics of the memory circuit in the
ISSN:0882-1666
DOI:10.1002/scj.4690230402
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
3. |
Contour extraction by local parallel and stochastic algorithm which has energy learning faculty |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 26-35
Sadayuki Hongo,
Mitsuo Kawato,
Toshio Inui,
Sei Miyake,
Preview
|
PDF (1179KB)
|
|
摘要:
AbstractWhen a combined Markov random field (MRF) image model is applied to the contour extraction problem, the values of parameters specifying the model cannot be known. However, the parameter values can be learned from image examples containing a contour line. This paper first describes details of the learning algorithm. States of line process representing contour line obeys a probability distribution function determined by three aspects: input image, energy function and energy parameters called potential.In the learning algorithm presented here, a value of local potential which specifies an MRF image model is learned using the maximum likelihood method from a contour line image given as a teacher. By applying the proposed learning algorithm to real images, energy parameters are obtained that enable contour extraction that is superior in connectivity and has little noise. Moreover, as a consideration of the generalization faculty of the energy learning algorithm, energy learning for images and extracted contours of images (which are not good for learning) are performed using learned parameters. It was verified that contour lines extracted nonlearning images that are of the same quality as that of contour lines of learning images.
ISSN:0882-1666
DOI:10.1002/scj.4690230403
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
4. |
Self‐organizing architecture for learning novel patterns |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 36-46
Kouichirou Yamauchi,
Takashi Jimbo,
Masayoshi Umeno,
Preview
|
PDF (810KB)
|
|
摘要:
AbstractAlthough the neuron model is very useful for many applications, it has a serious problem of learning. For example, a neuron model for pattern recognition works well after learning, but it must be retrained on all sample patterns if a new category of patterns is added. On the other hand, humans can add the new category immediately, without relearning all the old patterns. To solve this problem, we propose a new self‐organizing architecture for learning novel patterns. This model can learn novel patterns as soon as they are recognized to be new, just like humans d
ISSN:0882-1666
DOI:10.1002/scj.4690230404
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
5. |
Acquiring omnidirectional range information |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 47-56
Hiroshi Ishiguro,
Masashi Yamamoto,
Saburo Tsuji,
Preview
|
PDF (1287KB)
|
|
摘要:
AbstractThis paper describes a method of obtaining a precise omnidirectional view and coarse ranges to objects by rotating a single camera. The omnidirectional view is obtained by arranging image data taken through a vertical slit on the image plane while the camera rotates around the vertical axis. The omnidirectional view contains precise azimuth information determined by the resolution of the camera rotation. The range is obtained from two omnidirectional images taken through two slits when the camera moves along a circular path. This range estimate contains errors due to a finite resolution of the camera rotation system, while a conventional binocular stereo method contains errors due to quantization of images. The representation of an environment, “panoramic representation,” by the omnidirectional view and range information is useful for a vision sensor of a mobile ro
ISSN:0882-1666
DOI:10.1002/scj.4690230405
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
6. |
Edge feature determination by using neural networks based on a blurred edge model |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 57-68
Hiroshi Naruse,
Atsushi Ide,
Mitsuhiro Tateda,
Yoshihiko Nomura,
Preview
|
PDF (888KB)
|
|
摘要:
AbstractThis paper proposes a new model fit‐type edge feature measurement method. The characteristic of the proposed method is that it introduces a blurred edge model which matches well with a gray‐level pattern of an edge in an image actually observed. The blurred edge model is constructed by using not only edge features, which are the edge position and orientation within the pixel, but also the point spread function which expresses the image degradation during the image recording process as parameters. By using this model, the gray level of the multiple pixels near the edge for the various edge feature values is calculated, and a map is obtained from the edge features to the gray‐level pattern.Next, the inverse map which obtains edge features from a gray‐level pattern is obtained in advance through learning by using error backpropagation‐type neural networks consisting of three layers. By using the obtained inverse map, the edge features are determined from the gray‐level pattern of the actually observed image.Conventionally, since it was necessary to obtain this inverse map analytically, the edge model that could be used was restricted to the step‐edge type. On the other hand, with the method being proposed which utilizes the neural networks, an arbitrary optimal edge model for an individual image recording device can be used. For this reason, edge features can be determined precisely with this method even from local information. Many measurement experiments which changed the edge position and orientation were performed and the effectiveness of this method
ISSN:0882-1666
DOI:10.1002/scj.4690230406
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
7. |
Development of software system organized by minimal automata realization theorem for gait measuring |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 69-79
Yoshiyuki Saito,
Shigeru Niinomi,
Preview
|
PDF (696KB)
|
|
摘要:
AbstractWe describe the software system for improving the efficiency of the gait measuring experiment for the rehabilitation of amputees. This software has the following features: (1) The system consists of independent programs that are connected by a CHAIN statement. So it can easily adjust to the changes or progress in equipment and the measuring method; (2) the flow structure is organized in terms of the algorithm of minimal automata so that redundancy appearing in repeated measurements is eliminated and the load on the amputee is reduced.
ISSN:0882-1666
DOI:10.1002/scj.4690230407
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
8. |
A method of digital figure decomposition based on distance feature |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 80-91
Minoru Okada,
Shigeki Yokoi,
Jun‐Ichiro Toriwaki,
Preview
|
PDF (1170KB)
|
|
摘要:
AbstractThis paper proposes a method of decomposing a D‐figure based on the distance features of the figure. This method, called “breaking where it can be broken,” can be applied to digital figures of arbitrary dimension and is designed to produce the familiar decomposition which corresponds to the physical image when the rigid body has a weakness in its figure.When the information available about the digital picture is black‐and‐white, the decomposition of the figure must be based only on topological properties such as 1‐pixel connectivity, construction of a good figure decomposition is difficult because of the limited information.First, the definition of the decomposition of a D‐figure is given. Then the algorithm which realizes it in the digital space is presented. Finally, the results of experiments with the two‐dimensional (2‐D) D‐figure decomposition of artificial figures are given and the usefulness of this method is illustrated with a practical example, the analysis of a 3‐D D‐figu
ISSN:0882-1666
DOI:10.1002/scj.4690230408
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
9. |
A new information criterion combined with cross‐validation method to estimate generalization capability |
|
Systems and Computers in Japan,
Volume 23,
Issue 4,
1992,
Page 92-104
Yasuhiro Wada,
Mitsuo Kawato,
Preview
|
PDF (833KB)
|
|
摘要:
AbstractNeural network learning processes use only a limited number of examples of a given problem. Thus, generally speaking, it is not necessarily theoretically guaranteed that the trained network can give correct answers for unknown examples. A new method of selecting the optimal neural network structure with maximum generalization capability is proposed. In statistical mathematics, several information criteria, such as AIC (Akaike's information criterion), BIC (Bayesian information criterion), and MDL (minimum description length), are used widely to select a suitable model. Applications of these criteria were quite successful, especially for linear models. These criteria assume that the model parameters are estimated correctly by using the maximum likelihood method. Unfortunately, however, this assumption does not hold for conventional iterative learning processes such as backpropagation in multilayer perceptrons or Boltzmann machine learning. Thus, we should not apply AIC directly to the selection of the optimal neural network structure.In this paper, by expanding AIC, a new information criterion is proposed that can estimate generalization capability without the maximum likelihood estimator of synaptic weights. The cross‐validation method is used to calculate the new information criterion. By computer simulation, we show that the proposed information criterion can accurately predict the generalization capability of multilayer perceptrons, and thus the optimal number of hidden units can be determine
ISSN:0882-1666
DOI:10.1002/scj.4690230409
出版商:Wiley Subscription Services, Inc., A Wiley Company
年代:1992
数据来源: WILEY
|
|