Dynamic construction of fault tolerant multi-layer neural networks

We propose a new training algorithm for enhanced tolerance to physical defects (faults) of multi-layer neural networks (MLNs). We aim to construct such MLNs with the minimal number of hidden units. The proposed method has two characteristics, constructing MLNs dynamically and getting high fault tole...

Full description

Saved in:
Bibliographic Details
Main Authors: Haruhiko, T., Ayumi, N., Hidehiko, K., Terumine, H.
Format: Conference Proceeding
Language:English
Subjects:
Acceleration
Artificial neural networks
Fault tolerance
Heuristic algorithms
Large scale integration
Minimization methods
Multi-layer neural network
Neural networks
Output feedback
Proposals
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:We propose a new training algorithm for enhanced tolerance to physical defects (faults) of multi-layer neural networks (MLNs). We aim to construct such MLNs with the minimal number of hidden units. The proposed method has two characteristics, constructing MLNs dynamically and getting high fault tolerance easily. We proposed dynamic constructive algorithm with weight minimization approach (DCWMA) based on a DCA and WMA. DCA (dynamic constructive algorithm) is a basic dynamic constructive algorithm for MLNs. WMA (weight minimization algorithm) is a training algorithm to enhance the fault tolerance of fixed structure MLNs. The effectiveness of DCWMA is shown by some experiments.
ISSN:2161-4393
2161-4407
DOI:10.1109/IJCNN.2005.1555988