Neural networks for function approximation

We describe certain results of Mhaskar concerning the approximation capabilities of neural networks with one hidden layer. In particular, these results demonstrate the construction of neural networks evaluating a squashing function or a radial basis function for optimal approximation of the Sobolev spaces. We also report on the application of some of these ideas in the construction of general-purpose networks for the prediction of time series, when the number of independent variables is known in advance, such as the Mackey-Glass series or the flour data.