IMAGE ASSOCIATIVE MEMORY BY RECURRENT NEURAL SUBNETWORKS

Gray scale images are represented by recurrent neural subnetworks whic h together with a competition layer create an associative memory. The single recurrent subnetwork N-i implements a stochastic nonlinear frac tal operator F-i, constructed For the given image f(i). We show that u nder realistic assumptions 7 has a unique attractor which is located i n the vicinity of the original image, Therefore one subnetwork represe nts one original image. The associative recall is implemented in two s tages. Firstly, the competition layer Ends the most invariant subnetwo rk For the given input noisy image g. Next, the selected recurrent sub network in few (5-10) global iterations produces high quality approxim ation of the original image. The degree of invariance for the subnetwo rk N-i on the input g is measured by a norm \\g - F-i(g)\\. We have ex perimentally verified that associative recall for images of natural sc enes with pixel values in [0, 255] is successful even when Gaussian no ise has the standard deviation sigma as large as 500. Moreover, the no rm, computed only on 10% of pixels chosen randomly From images still s uccessfuly recalls a close approximation of original image. Comparing to Amari-Hopfield associative memory, our solution has no spurious sta tes, is less sensitive to noise, and its network complexity is signifi cantly lower. However, for each new stored image a new subnetwork must be added.