Statistical design using variable parameter variances and application to cellular neural networks

Many cellular neural network design methods result in a set of linear inequalities, which they attempt to solve by various methods. In the paper we first point out the importance of the problem for the CNN design, and then expand the statistical design method proposed by R.K. Brayton, G.D. Hachtel, and S.W. Director (1978), applying it to cellular neural networks. Instead of original assumption of constant variances of the statistical parameter distributions, we take variances to be linearly dependent on parameter nominal values, which leads us to construct an iterative process with very fast convergence. A design example of winner-take-all cellular neural network is given, showing that with our improvement one can reliably implement the network of up to 50 cells as opposed to 10 cell CNN obtained by the original method.