Fault tolerance design of feedforward networks

Fault-tolerant feedforward networks are designed by incorporating fault tolerance at the learning stage. This approach is particularly attractive in those instances where the network components are not accessible during normal operation. Three new methods of fault tolerance learning are investigated: min-max fault tolerance learning, fault tolerance through weight control and fault tolerance through strict learning/less strict operation. Simulation results are presented which show that considerable improvement in classification performance can be achieved over backpropagation, particularly in the case of the last two methods.