Method to Approximate Initial Values for Training Lineal Neural Networks

Abstract

The present paper proposes a method to calculate a set of proposed initial values for the weight matrix and the bias vector of a neural network prior to training. The method described here applies for linear neural networks with one hidden layer, and a known proportional relationship between inputs and outputs. The algorithm and the calculations are intended to be simple, to facilitate automation in small processors The method normalizes values in a tri-level form, finds the relationships on the maximum and minimum values for all combinations of inputs and outputs, averages these results and builds the weight matrix and bias vector from these results. The end result is a set of initial values prior to training, intended to have a start point for training closer to the end result. Overall result is less training time.