Half-distributed coding makes adaptation of sigmoid-threshold useless in back-propagation networks

Abstract

The effects of the adjustment of the threshold of the hidden cells during learning in a one-hidden-layer backpropagation network with half-distributed coding of inputs are analyzed. The fundamentals of this coding method are reviewed. Although it can be applied to both inputs and outputs of the network, only the case of the inputs is considered. The effects of the modification of the thresholds during learning are analyzed. It is shown that these effects become more favorable as the task to be achieved becomes less complex. The correctness of the theoretical model was tested with a real-world application. The network has to approximate a function to realize a numerical model of a physical phenomenon.<>