Neural net configuration design using theory of sensitivity and tolerances

The problem of learning neural networks to get the most convenient configuration with respect to the complexity of its technical realization is considered. By the configuration is meant the vector of synaptic weights and thresholds of formal neurons creating the network. The task of learning is considered as an optimization problem. The tools of the tolerances and sensitivity theory are used to solve this optimization problem, taking into account technological demands. The advantages of such a process of configuration design are demonstrated by an example.