Computation and Generalization in Neural Networks
Abstract
During this contract period, our research into backward propagation has led to a number of new theoretical and empirical results. We have developed a generalized version of backward propagation. In our generalized network, both gains and synapses are modified by a backward propagation procedure. Synapses are modified in proportion to the negative gradient of the energy with respect to the synaptic weight as in ordinary backward propagation, and gains are modified in proportion to the negative partial derivative with respect to gain. Since the resulting error signals for the gain and synaptic weights are proportional to one another, the computational complexity of our generalized network is comparable to that of the original backward propagation model.... Back propagation, Gain modification, Multilayer perceptrons.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 05, 1993
- Accession Number
- ADA263752
Entities
People
- Leon Cooper
Organizations
- Brown University