optimisationofneuralnetworks.PDF

Correlating matched-filter model for analysis and optimisation of neural networks D.R. Selviah, MA Prof. J.E. Midwinter, OBE, FEng, FIEE, FRS A.W. Rivers, MSc K.W. Lung, BSc Indexing terms: Active networks, Signal processing, Pattern recognition outputs the stored binary, bipolar (+ 1, -1) pattern si Abstract: A new formalism is described for mod- (i = 1to M) which most closely resembles the input. elling neural networks by means of which a clear The neural net architecture chosen for discussion in physical understanding of the network behaviour this paper is the synchronously operated Little-Hopfield can be gained. In essence, the neural net is rep- neural net [6, 71 (having threshold levels of zero), which resented by an equivalent network of matched has recently received much attention due to its suitability filters which is then analysed by standard correla- for electronic and optical implementation [8-121. The tion techniques. The procedure is demonstrated network is shown in Fig. 1, where an input unknown on the synchronous Little-Hopfield network. It is shown how the ability of this network to discrimi- nate between stored binary, bipolar codes is opti- nnection non-linear mised if the stored codes are chosen to be processing orthogonal. However, such a choice will not often be possible and so a new neural network architec- ture is proposed which enables the same discrimi- nation to be obtained for arbitrary stored codes.