States that a standard feed forward Neural Network with a single hidden layer can approximate any continuous function to any desired degree of accuracy, provided the network has a finite but arbitrarily large number of neurons and uses a non-linear activation function.