Improved GNN Models for Constant Matrix Inversion

It was shown that the multiplication of the left hand side of the classical Zhang neural network design rule by an appropriate positive definite matrix generates a new neural design with improved convergence rate. Our intention is to apply similar principle on the standard gradient neural network (G...

Full description

Saved in:
Bibliographic Details
Published in:Neural processing letters 2019-08, Vol.50 (1), p.321-339
Main Authors: Stanimirović, Predrag S., Petković, Marko D.
Format: Article
Language:English
Subjects:
Artificial Intelligence
Complex Systems
Computational Intelligence
Computer Science
Computer simulation
Convergence
Design standards
Eigenvalues
Investigations
Mathematical analysis
Matrices (mathematics)
Network design
Neural networks
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:It was shown that the multiplication of the left hand side of the classical Zhang neural network design rule by an appropriate positive definite matrix generates a new neural design with improved convergence rate. Our intention is to apply similar principle on the standard gradient neural network (GNN) model. To that goal, we discover that some of proposed models can be considered as the multiplication of the right hand side of the GNN model by a symmetric positive-semidefinite matrix. As a final result, we propose appropriate general pattern to define various improvements of the standard GNN design for online real-time matrix inversion in time invariant case. The leading idea in generating improved models is initiated after a combination of two GNN patterns. Improved GNN (IGNN) design shows global exponential convergence with an improved convergence rate with respect to the convergence rate of the original GNN pattern. The acceleration in the convergence rate is defined by the smallest eigenvalue of appropriate positive semidefinite matrices. IGNN models are not only generalizations of original GNN models but they comprise so far defined improvements of the standard GNN design.
ISSN:1370-4621
1573-773X
DOI:10.1007/s11063-019-10025-9