作成者 |
|
本文言語 |
|
出版者 |
|
|
発行日 |
|
収録物名 |
|
巻 |
|
出版タイプ |
|
アクセス権 |
|
関連DOI |
|
|
関連URI |
|
|
関連情報 |
|
|
概要 |
A heteroassociative memory network for image recognition is constructed with the aid of the method in the paper [5]. This network is a three layered neural network which consists of an input layer, a ...hidden layer and an output layer. A feature of the network is to contain a sigmoid function only in the hidden units. Images to be stored in the network are real valued vectors. Weights and threshold values connecting the input layer with the hidden layer are determined such that for input reference images, a sufficiently small positive number $ \varepsilon $ or 1 - $ \varepsilon $ is output at each unit in the hidden layer. Interconnection weights between the hidden and output layers are determined so as to reconstitute the input reference images. This approach makes possible the contraction mapping analysis for the network. As done in [5], domains of attraction in the network are sought. Regions of attraction larger than these domains are also found using the smallness of $ \varepsilon $. Furthermore, a certain heteroassociative memory model is designed based on the shape of the fundamental domains of attraction, and successfully applied to recognition of facial images.続きを見る
|