Johnson Distributions for Fitting Weighted Sums of Sigmoided Neuron Outputs

Abstract : In this paper, it is shown that a continuum of distributions best characterizes the hidden layer outputs of a multilayer perceptron when trained as a 0-1 classifier and tested with a range of signal-to-noise ratio (SNR) input distributions. A four parameter system of transformed normal distributions, known as the Johnson system of distributions, is utilized to illustrate the shape of output distributions as a function of input SNR levels. Neural networks, Active signal processing.