Local minima and back propagation

Abstract

It is shown that, in a feedforward net of logistic units, if there are as many hidden nodes as patterns to learn then almost certainly a solution exists, and the error function has no local minima. A large enough feedforward net can reproduce almost any finite set of targets for almost any set of input patterns, and will almost certainly not be trapped in a local minimum while learning to do so.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>