Almost optimal estimates for approximation and learning by radial basis function networks (Q2251472)

The paper is devoted to approximation of differentiable multivariate functions by the radial basis function networks (RBFN) defined by a formula of the following type: \[ R(x)=\sum_{k=0}^N c_k \sigma (w_k|x - \theta_k|). \] Here \(\sigma\) is the activation function, \(c_k,w_k \in\mathbb R\), \(\theta_k \in\mathbb R^d\). Let \(B^d\) be the unit cube in \(\mathbb R^d\). The authors prove that for any given polynomial \(P\) and sufficiently smooth function \(\sigma\) there exists an RBFN approximating \(P\) arbitrarily closely in \(C(B^d)\). The authors also study machine learning. They prove that using the standard empirical risk minimization, the RBFN can realize an almost optimal learning rate.