Fault-tolerance characteristics of neural networks
A methodology for measuring and improving the fault tolerance characteristics of neural networks is presented. Sensitivity analysis and headroom analysis programs have been developed using fault models more realistic and appropriate for emulating hardware failures than those used previously. The potential mode of failure is simulated as a corruption of stored weight and threshold values. These analysis tools enable the fault tolerance characteristics of neural networks to be evaluated. It is demonstrated how functionally identical neural networks can have significantly different reliability characteristics should they be subjected to hardware platform failures. Criteria for selection of globally optimal architecture and trained state are discussed using results provided by the sensitivity and headroom analysis programs. These criteria, combined with empirical results, lead to implied design rules which can be adopted by engineers of neural networks for improving and maximizing the fault tolerance characteristics of the system. A novel modification to the backward error propagation training algorithm is discussed and evaluated on its effectiveness in improving the robustness of the trained network. The modification involves the deliberate injection of a small amount of random white noise on the network's weights and thresholds to expedite and increase the likelihood of a more optimal convergence stat occurring for the network from the aspect of fault tolerance. The methodology is demonstrated using an iterative design scenario and is shown to be effective under certain circumstance.
- Research Organization:
- California Univ., Los Angeles, CA (USA)
- OSTI ID:
- 5923848
- Resource Relation:
- Other Information: Thesis (Ph. D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Final Technical Report - Applications of Machine Learning Techniques to Geothermal Play Fairway Analysis in the Great Basin Region, Nevada
Identification of Plasma Parameters and Optimization of Magnetic Sensors in the Superconducting Steady-State Tokamak-1 Using Neural Networks