[in Japanese]

View full abstract

[in Japanese]

View full abstract

Bp learning has been used very often as methods for neural network's learning. If the network has enough neurons to classify the training data, BP learning is well known to perform good classification. However, the network is so called "Black Box", because it is difficult to give clear explanation on the relation between inputs and outputs. Recently, some methods for extracting some rules from the regularity, which appears on hidden neurons, have been proposed. These methods cannot make clear the relation of extracted rules. In this paper, we propose a learning method of neural network with lattice architecture. This network structure is consisted of hidden neurons in the lattice. To obtain the optimal network structure, we apply a neuron generation/annihilation algorithm without a dependency on an initial network structure. To verify the validity and effectiveness of proposed method, we have experimented to the function identification and classification of any points in a cube to compare with BP learning and RBF network.

View full abstract

We propose simple topographic mapping formation models from cell layer to cell layer. Our model is a discrete one, that is the state value of input and output cells takes 0 or 1 and input and output layers are represented by undirected graphs. Thus, a topographic mapping described in this model is a map which preserves the adjacency relation. We define several learning rules and a few weight normalization methods. By computer simulations we investigate topographic mapping formation conditions. We show that when an output unit normalization is considered, we have more learning rules which yield topographic mappings than the cases when an input normalization is adopted or without normalization. As to the input unit normalization, we have shown theoretically that topographic mappings are the only stable ones under the correlational type learning rule.

View full abstract

This paper considers an application of the Self-Organizing Map (SOM), an effective technique for clustering of multi-dimensional data, to the short-term prediction of the oil temperature change of the transformer. For its application, three types of processes for the prediction are investigated, i.e., (1) A SOM is obtained by using atmospheric temperature, load rate and oil temperature in the past year, then the oil temperature except those for leaning the SOM is predicted. (2) Due to the heavy load during the summer, the SOM is obtained with every three months for the duration June through October. The oil temperature for the season is predicted using similar data structure as that of process (1). (3) The prediction of the oil temperature of transformers can be realized by the SOM based on the maximum and minimum values of the forecast atmospheric temperature announced by the meteorological observatory. Using this technique, the change of the oil temperature of an outdoor transformer is well predicted, and the prediction accuracy is higher than that obtained using the conventional method.

View full abstract

The weight of qualitative information is large in the management of business strategy. It is relatively difficult to reconcile rationality and creativity in a qualitative information analysis. Even in the strategic theory made rational in management study, the multi-dimensionality is not correctly taken into consideration, though the qualitative information is essentially of multi dimensional. This paper shows that planning of business strategy is possible, by carrying out the self-organization of the qualitative information using SOM.

View full abstract

Image database of the butterfly specimen is constructed using Self-Organizing Maps (SOM). Input vector for SOM consists of color, shape, and texture characteristics. Color component is a 10-dimensional vector of a HSL color histogram. Shape component is a combination of a 24-dimensional vector of local average radii and a 2-dimensional vector of momenta. Texture component is a 52-dimensional vector of several statistical quantities such as angular second moment, contrast, correlation and entropy. As a result, the images each expressed by 88-dimensional feature vector are effectively classified to form several categories. At the retrieval, not a character-based keyword but an image is used as a query to search the most similar image stored in the database. For the user without an appropriate input image, color and texture sample palettes, and shape templates are given to a user, where the templates are obtained by the data clustering using 1-dimensional SOM.

View full abstract

Gesture recognition is an appealing tool for natural interface with computers especially for physically impaired persons. In this paper, it is proposed to use Self-organizing maps (SOM) as an image recognition system for gesture recognition, since the SOM allows alleviating many difficulties associated with gesture recognition. It is, however, required for on-line systems to reduce the recognition time to the range of normal video camera rates. To achieve this, the Randomized SOM (RSOM) is introduced. With RSOM algorithm, the recognition time is drastically reduced without accuracy deterioration. The experimental results to recognize hand gestures using RSOM are presented.

View full abstract

Vector quantization have been used for both storage and transmission of speech and image data, and often requirs the algorithm that minimizes the distortion error. To obtain the minimum distortion error in the neural networks for vector quantization, reformatory competitive learnings and so on., have been introduced. Among the number of algorithms, neural gas networks are well known for showing better performance. In this paper, we propose some self-organizing neural gas networks, self-deleting neural gas networks and ones which are combinations of them. The conventional and proposed methods are compared by the tasks to compress image data. It is shown that the method which is a combination of deleting and creating is more effective than the other algorithms.

View full abstract

Purpose: Mechanisms of budgetary resource allocations from the Japanese central government to local governments are analyzed in this paper. Method: The paper utilizes statistical methods and a Self Organizing Map (SOM). Results: All budgetary transfers to local governments are said to be redistributed on an equitable basis. As a result of the budgetary transfers to local governments which have been carried out for a long time (since the World War II) on an equitable basis, we expect inefficient investment in public goods being carried out. We also suspect political influences. This paper tries to analyze the budgetary resource allocations from three points of view - equity, efficiency and political influences. Using a cross- sectional analysis of fiscal year 1991 and a panel data analysis of 18 years from 1977 through 1995, the following results were obtained. With the cross- sectional analysis of fiscal year 1991, we found that our results were as expected- namely, the budgetary resource allocations are made on an equitable basis, but inefficient public goods investment has taken place, and there is political influence. Based on the cross sectional results, we performed the similar analysis using panel data. The results obtained utilizing SOM and the statistical method generally support our cross sectional findings. New or Breakthrough aspect of work: However, when data from 46 prefectures are clustered using SOM, and the two largest clusters are analyzed further statistically, the results differ between national data and two cluster data and also within two clusters. Furthermore, findings from an analysis of panel data by SOM provide evidence to support our assumptions. Conclusions: The result shows that SOM adds new dimension to economic analysis and that SOM is useful as a visual data mining method.

View full abstract