An upper estimate of the error of approximation of continuous multivariable functions by KBF networks

摘要

We present an estimate of approximation error of multivariable continuous funtions by networks with kernel basis function (KBF) units. The estimate is a function of the nunber of hidden units and of the total variation of the convolution of the function to be approximated with a kernel basis function. We also present here known estimates of the error of approximation of continuous multivariable functions by networks with sigmoidal and spline activation functions. All the presented estimates of errors depend indirectly on the number of units in the hidden layer.