Universal approximations of invariant maps by neural networks (Q2117338): Difference between revisions

Neural networks are currently one of the most-investigated methods in applied mathematics. The author of this paper puts the approach in a new, more general context in mathematics, namely the universal approximations of invariant or equivariant or permutation invariant maps (there are several new versions of the universal approximation theorem). This is carried out in one instance with the help of constructing intermediate layers of the shallow networks by polynomials. The latter allows the approximations to be invariant or equivariant and complete. The construction with polynomials can then be extended by using polarisation. The groups on which the maps are formed are for example general compact groups or special translation-invariant groups. The inputs that are admitted to the network include continuous multivariable signals from infinite dimensional spaces. Both shallow and deep neural networks are addressed.