Pattern classification by an exponental response neural net

The authors describe a three layer feedback neural network, with the capability for mapping real valued input to any class structure as its output. They present a Liapunov function for the network, and show that it has minima corresponding only to stored vectors. The nature of the energy surface is discussed. For a known underlying distribution statistics a one-shot local learning rule is used. For classification based on samples a fast self-organizing training algorithm, mapping class boundaries, is used. They present a Liapunov function, show that the function has no false minima, and that the network moves along a gradient descent trajectory towards its stable attractor states. The network is suitable for associative recall and pattern recognition tasks