Feed-forward neural networks

被引：348

作者：

Bebis, George ^{[1
]}

Georgiopoulos, Michael ^{[1
]}

机构：

[1] Electrical and Computer Engineering Department, University of Central Florida, United States

来源：

IEEE Potentials | 1994年 / 13卷 / 04期

关键词：

Algorithms - Approximation theory - Computational complexity - Computer architecture - Correlation methods - Curve fitting - Errors - Learning systems - Optimization - Polynomials - Sensitivity analysis;

D O I：

10.1109/45.329294

中图分类号：

学科分类号：

摘要：

The paper emphasizes the importance of network size for a given application. Network size affects network complexity, learning time and generalization capabilities of the network. Included is an illustrative analogy between neural network learning and curve fitting. In the determination of hidden nodes and hidden layers it was found out that feed-forward networks can approximate virtually any function of interest to any desired degree of accuracy, provided enough hidden units are available. Small networks capable of learning the task is better for practical and theoretical reasons as compared to bigger networks. The generalization capabilities of a network can be improved by modifying the connection weights and architecture. These are specifically the pruning and constructure approaches.

引用

页码：27 / 31