Publication Details
Title: Fault Tolerance in Feed-Foward Artificial Neural Networks
Author: C. H. Sequin and R. D. Clay
Group: ICSI Technical Reports
Date: July 1990
PDF: http://www.icsi.berkeley.edu/pubs/techreports/tr-90-031.pdf
Overview:
The errors resulting from defective units and faulty weights in layered feed-forward ANN's are analyzed, and techniques to make these networks more robust against such failures are discussed. First, using some simple examples of pattern classification tasks and of analog function approximation, it is demonstrated that standard architectures subjected to normal backpropagation training techniques do not lead to any noteworthy fault tolerance. Additional, redundant hardware coupled with suitable new training techniques are necessary to achieve that goal. A simple and general procedure is then introduced that develops fault tolerance in neural networks: The type of failures that one might expect to occur during operation are introduced at random during the training of the network, and the resulting output errors are used in a standard way for backpropagation and weight adjustment. The result of this training method is a modified internal representation that is not only more robust to the type of failures encountered in training, but which is also more tolerant of faults for which the network has not been explicitly trained.
Bibliographic Information:
ICSI Technical Report TR-90-031
Bibliographic Reference:
C. H. Sequin and R. D. Clay. Fault Tolerance in Feed-Foward Artificial Neural Networks. ICSI Technical Report TR-90-031, July 1990
Author: C. H. Sequin and R. D. Clay
Group: ICSI Technical Reports
Date: July 1990
PDF: http://www.icsi.berkeley.edu/pubs/techreports/tr-90-031.pdf
Overview:
The errors resulting from defective units and faulty weights in layered feed-forward ANN's are analyzed, and techniques to make these networks more robust against such failures are discussed. First, using some simple examples of pattern classification tasks and of analog function approximation, it is demonstrated that standard architectures subjected to normal backpropagation training techniques do not lead to any noteworthy fault tolerance. Additional, redundant hardware coupled with suitable new training techniques are necessary to achieve that goal. A simple and general procedure is then introduced that develops fault tolerance in neural networks: The type of failures that one might expect to occur during operation are introduced at random during the training of the network, and the resulting output errors are used in a standard way for backpropagation and weight adjustment. The result of this training method is a modified internal representation that is not only more robust to the type of failures encountered in training, but which is also more tolerant of faults for which the network has not been explicitly trained.
Bibliographic Information:
ICSI Technical Report TR-90-031
Bibliographic Reference:
C. H. Sequin and R. D. Clay. Fault Tolerance in Feed-Foward Artificial Neural Networks. ICSI Technical Report TR-90-031, July 1990