A Framework for the Analysis of fault-tolerance in FFANN's

摘要

Feedforward Artificial Neural Networks (FFANN's) have been the architecture of choice for solving tasks as diverse as classification, regression, function approximation, density estimation, vehicle control, etc. With the availability of hardware implementation of these networks it has become desirable to measure their fault tolerant properties to a wide variety of fault types. In this paper a framework for the study of fault - tolerance properties is described. Moreover, a hierarchy of fault - models are described and error measures are suggested. Experimental results indicate that FFANN's are not necessarily fault tolerant and the sensitivity of weights nearer to the output is higher. A summary of the literature is given within the proposed framework.