Tsuyoshi Shimada

A new self-organizing method and its application to handwritten digit recognition

A self-organizing method for neural networks is proposed. This method reduces the calculation for learning considerably, and can be applied to real application problems, where many samples must be treated. The method has been applied to handwritten digit recognition. Samples incorrectly recognized have been reduced to 1/4 (learning data) or 2/3 (unknown data), compared with the multiple similarity method, which is a conventional statistical pattern classification method.<<ETX>>

A locally active unit for neural networks and its learning method

In this paper, a type of neural network using locally active units and a learning method is proposed. A locally active unit is a processing element which is used as an artificial neuron, and which is only activated by input vectors in a bounded domain of the vector space. The unit is suitable for constructing neural networks for pattern recognition. The learning method adjusts the resolution of the units so that the neural net can handle complex boundaries between pattern categories.

Artificial neural network of locally active units with cause-oriented parameter modification

In this paper, a 3-layer neural network of locally active units is proposed. In the neural network, each constituent unit of a hidden layer is only activated by input vectors in a bounded domain of the vector space. This feature leads to additional learning, and also leads to knowing the architecture of the neural network and obtaining information suggestive of ways in which forecasting accuracy could be improved. We think that forecasting one-dimensional social quantity, for example, electric load or stock prices, makes the best use of the advantages of the proposed neural network, and we propose a method for detecting the causes of forecasting errors and improving the forecasting ability of the neural network. We examined the performance of the proposed neural network by applying it to daily peak electric load forecasting in summer. Comparing the forecasting result of the network with the conventional error back-propagation algorithm, the maximum error rate is clearly reduced. Carrying out the proposed method for detecting the causes of forecasting errors, forecasting errors are further reduced.<<ETX>>