Pascal Koiran DIMACS, Rutgers University

Thispapertakesideasdeveloped inatheoretical frameworkbyMaass 8] andadapts them forapractical learning algorithm forfeedforward sigmoid neural networks. A number of dierent techniques are presented which are based loosely around the common theme of taking advantage of the linearity of the net input to a neuron, or in other words the fact that there is only a single non-linearity per neuron. Some experimental results are included, though many of the ideas are as yet untested. The paper can therefore be viewed as a tool box oering a selection of possible techniques for incorporation in practical, heuristic learning algorithms for multi-layer perceptrons.