Approximation with neural networks activated by ramp sigmoids (Q1958429)

One proves that functions defined on a \(d\)-dimensional domain \(S\), with bounded variation with respect to a class of ellipsoids contained in \(S\) , can be accurately approximated by two-hidden-layer feedforward neural networks. The approximation error depends on the number of neurons in both hidden layers, space dimension and function variation. As activation function the rump sigmoid is considered. The author extensively refers to his earlier paper: \textit{G. H. L. Cheang} and \textit{A. R. Barron} in [J. Approximation Theory 104, No.~2, 183--203 (2000; Zbl 0991.41012)].