Fault tolerant capability of multi-layer perceptron neural network

Mean squared error is the only criteria of backpropagation training to be optimized. Some other good properties of neural networks such as generalization and fault tolerance are only be taken as side effects of neural networks. In this paper we define another energy term called constraint energy to be optimized. The role of constraint energy in the training phase is to control the fault tolerant property of neural networks. Incorporated with normal energy and constraint energy we can find out a set of weights that guarantee some degree of fault tolerance in any hidden node failure. Experimental results indicate that networks trained by fault tolerant constraints possess better generalization properties than those trained by normal backpropagation. The mechanism of training with constraint energy proposed here has the tendency to force the nodes to leave out the linear region, thus reducing the sensitivity of links and nodes. Constraint energy can be extended to tolerant multiple nodes failure