Gain Modification Enhances High Momentum Backward Propagation
Abstract
We present a backward propagation network which simultaneously modifies the gain parameters and the synaptic weights. Gain modification is shown to enhance the improvement in convergence rate obtained by high momentum in standard synaptic backward propagation. These improvements occur without degrading the generalization capabilities of the final solutions obtained by the network. Keywords: Neural nets; Backward propagation; Gain modification; Momentum; Effective time-dependent; Step constant.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 08, 1989
- Accession Number
- ADA216032
Entities
People
- Charles M. Bachmann
Organizations
- Brown University