A method for finding the maximally robust weights of a feed forward neural network with step function neurons

The authors consider a method by which backpropagation is used to find the appropriate interpolating parameters in a network with hard limiter transfer functions. This method has two features: first, it defines a family of sigmoidal functions which converge to the hard limiter so that BP may be used as the appropriate training method; and second, it finds the place to start so that gradient descent yields the absolute minimum and does not become trapped in a relative minimum. The interpolating parameters yield a maximally robust solution