E-Book Content
Neural Networks,Vol. 10, No. 6, pp. 1143-1151, 1997 01997 Elsevier Science Ltd. All rightsreserved Printedin CheatBritain 0893–6080/97 $17.00+.03
Pergamon @
PII: S0893-6080(96)00128-1
CONTRIBUTEDARTICLE
Genetically Trained Cellular Neural Networks MICHELE ZAMPARELLI HLRZ-KFA 52425, and Dipartimento di Matematica, Universit6 degli Studi di Roma, “La Sapienza”
(Received6 June1995;accepted10 November1996)
Abstract—Real-codedgenetic algorithms on a parallel architecture are applied to optimize the synaptic couplings of a Cellular Neural Network for specific greyscale image processing tasks. Using supervised learning information in the jitnessfinction, we propose the Genetic Algorithm as a general training methodfor Cellular Neural Networks. 01997 Elsevier Science Ltd.
Keywords-Cellular neuralnetworks,Geneticalgorithms,Supervisedlearning,Imageprocessing. 1. INTRODUCTION Cellular Neural Networks (CNNS), a new type of locally connected neural network with continuous activation values, have recently demonstrated their efficacy for bipolar signal processing. Several models of cortical neurons have been proposed so far, but the time evolution in the neuron’s activation has always been rather underestimated. CNNS on the other hand, have shown how the transient regime may play a decisive role in obtaining the correct stimulus-response association. An evolutionary approach, inspired from natural laws like the survival of the fittest, has already been proposed in several works, the main argument being that the actual brain structure has itself evolved through a competitive trial and error process. Genetic Algorithms (GAs), an attempt to emulate this trial and error process, have exhibited good performance in the design of feed-forward neural network architectures (Bornholdt & Graudenz, 1992) and it is therefore quite natural to investigate their per