Page 160 - Introduction to Information Optics

2.9. Processing with Neural Networks 145

Tony Geoge
Fig. 2.53. Concept of interpattern association (I PA).

If the interconnection weights are assigned equal to 1, — 1, and zero, for
excitory, inhibitory, and null interconnections, then a tristate IPA neural net
can be constructed. For instance, in Fig. 2.55a, pixel one is the common pixel
among patterns A, B, and C, pixel two is the common pixel between A and B,
pixel three is the common pixel between A and C, and pixel four is a special
pixel, which is also an exclusive pixel with respect to pixel two. Thus, by
applying the preceding logic operations, a tristate neural network can be
constructed as shown in Fig. 2.55b, and the corresponding IPA IWM is
shown in Fig. 2.55c.
For comparison of the IPA and Hopfield models, we have used an 8 x 8
neuron NN for the tests. The training set is the 26 English letters lined up in
sequence based on their similarities. Figure 2.56 shows the error rate as a
function of stored letters. We see that the Hopfield model becomes unstable to

Fig. 2.54. Common and special subspaces.