Error bounds for approximation with neural networks (Q5959036)

The paper considers neural network approximations with shifts of so-called ridge functions \(\sigma(a_j^Tx+b_j)\). Linear combinations of such ridge functions correspond to a neural network with a single hidden layer. Of particular interest are the approximation orders that are obtainable with these approximations to functions from Sobolev spaces. In the special case when the approximand can be written as a continuous convolution with the kernel used for the approximation, error estimates are given. The error estimates depend on the Sobolev smoothness of the kernel as well as the approximand's smoothness and the number of kernel functions in the neural network. The analysis applies to general kernels from a class that contains ridge functions. Applications of the theoretic results to perceptrons are provided.