Network Reduction Using Error Prediction
Abstract
This thesis investigates gradient descent learning algorithms for multi-layer feed forward neural networks. A technique is developed which uses error prediction to reduce the number of weights/nodes in a network. The research begins by studying the first and second order back-prop training algorithms along with their convergence properties. A network is reduced by making an estimate of the amount of error which would occur when a weight(s) is removed. This error estimate is then used to determine if a particular weight is essential to the operation of the network. If not, it is removed and the network retrained. The process is repeated until the network is reduced to the desired size, or the error becomes unacceptable.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1990
- Accession Number
- ADA230755
Entities
People
- Michael V. Gilsdorf
Organizations
- Air Force Institute of Technology