How best to restore operations of a damaged ANN?

Failure to deliver intelligent algorithms as was predicted, both by logicist approach of McCarthy and symbolic models of Newell and Simon, as well as successes of connectionist models during 80´s, prompted widespread acceptance of artificial neural network algorithms. In particular, multilayer feed forward neural network (aka MLP, an abbreviation from MultiLayer Perceptron) with back-propagation (BP) training was widely accepted as a universal approximator and used for pattern recognition problems, of different complexities. Though MLP takes long training time, once trained it could perform classification/recognition task fast. In this work, we studied how the performance of a trained MLP deteriorates, as its nodes and connection weights are damaged partially. Could the performance be restored, if further training is done on the damaged MLP, the way it was left after damage? How does the restoration depends on the complexity of the task, the network had learned before damage? Is it better to reinitialize the connection weights of the damaged MLP, before starting training again? What are the relative training times and restoration quality? Does it depend on the complexity of the problem? We have investigated these aspects in this work. In addition to explaining our results, we also conjectured to extend possible application of our findings to restore operation of a human brain, damaged by diseases like Alzheimer, or by accident.