An observation concerning a classification problem and back-propagation for the feedforward neural network

A simple classification problem using a single-layer feedforward neural network in conjunction with the backpropagation training algorithm (BPTA) is examined. It has been observed that, for such a problem, the values of the input weights are closely related to the input training set. An implication of this observation is that, rather than choosing initially random weights for the BPTA, one may choose initial weights that are actually quite close to a global minimum in the BP error function. An advantage of such a choice would be faster convergence times based on knowledge of the incoming training data