| 摘要 |
PURPOSE: A node expansion learning neural net model is provided to control the number of necessary internal layer nodes in accordance with a sample vector assemblage when learning is performed and to control a learning rate of each node dynamically. CONSTITUTION: A proper scaled neural net is constructed by creating or deleting a node at need. Each internal node forms a representative vector of throngs in a pattern space. If many sample vectors, which are similar to a created node, exist, a learning intensity of the node is increased. If many sample vectors being possessed in other class exist in an adjacent position, a learning intensity of the node is decreased. If the learning intensity of the node is decreased under a proper level, the node is removed. If a learning intensity is strong, the representative vector being stored in the node is decided to express the corresponding throng properly. Thus, an amount modifying the representative vector by a new sample vector, that is, a learning rate is decreased.
|
