How Neural Networks for Pattern Recognition Can Be Synthesized
スポンサーリンク
概要
- 論文の詳細を見る
This paper discusses synthesis of neural networks for pattern recognition. First the saturation characteristics of the sigmoid function are shown to be essential for pattern recognition. Then neural networks for pattern recognition are proved to be synthesized by the following steps: (1) If the input data are separated by k hyperplanes, take k hidden neurons setting the weights between the input and hidden neurons as the coefficients of linear equations that define the hyperplanes. (2) If a class is separated into a single region by hyperplanes selected from the k hyperplanes, a three-layered neural network is synthesized to recognize that class; if not, a four-layered neural network is synthesized. Finally, a method to accelerate learning is proposed and is verified for a parity circuit, classification of two-dimensional patterns, and number recognition.
- 一般社団法人情報処理学会の論文
- 1991-12-31
著者
-
Takenaga Hiroshi
Hitachi Research Laboratory Hitachi Ltd.
-
ABE SHIGEO
Hitachi Research Laboratory, Hitachi, Ltd.
-
KAYAMA MASAHIRO
Hitachi Research Laboratory, Hitachi, Ltd.
-
Kayama M
Hitachi Research Laboratory Hitachi Ltd.
-
Abe Shigeo
Hitachi Research Laboratory Hitachi Ltd.
関連論文
- How Neural Networks for Pattern Recognition Can Be Synthesized
- Fuzzy Clustering Networks : Design Criteria for Approximation and Prediction
- Fuzzy Rule Extraction with Optimized Hyperboxes for Approximating Class Regions
- Optimizing Composite Neural Networks for Very Hard Classification Problems