Fault tolerance of neural networks

The robustness and learning speeds of a neural network using the backpropagation algorithm are explored. An XOR experiment was performed on neural networks with one and two hidden layers. Robustness of the net was studied through removal of nodes and/or branches in hidden layers. It is observed that simulations with final output weights constrained to lie below a specified value provided superior performance, even when they were structurally damaged. Hence, for a two-hidden layer net, the weight constraint on interconnecting links yields a robust and faster-learning network.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>