A SIGNAL SPACE INTERPRETATION OF NEURAL NETS J.A.RitceyA, L.E.Atlas~.A.SomaniA,D.Nguyen·,F.Holt·

Hopf ield neural net processors (NNP): n a v e been shown to be an interesting class '.of faul t tolerant, parallel computers for pattern recognition. In this paper, we give some limited simulation results that contrast the performance of the Hopfield NNP, whose T-matrix is in sum-of-outer-products form, and the PrOjection NNP, which uses an orthogonal prOjection onto the linear space spanned by the library elements. A Compact NNP is introduced which promises good recall ability with a low density of neuron interconnections. I. I NTRODUCT I ON S ince the introduction of neural networks (NN) to the engineering community by Hopfield (IJ, a number of applications and variaties of the basic net have been proposed. In th i s paper, we present both the Hopfield and the PrOjection neural net processors ( NNP> and compare performance based on some limited simulation results. The Pr o j ec t t o n NN (PNN). in which Hopfields T-matrix, T~, is replaced by the prOjection matrix, Tp, onto the linear subspace spanned by the library elements, is suggested by signal space concepts. In addition to these baseline performance comparisons, we present a reduced complexity neural net, the Compact NN (CNN), whose T-matrix is obtained from T H b y quantizing far off-diagonal terms to zero. Reorganization of the library elements is a key point in the development of eff icient nets of this type. Th e research at the University of Wa s h i n g t o n is supported by the Boein