Robustness of cellular neural network implementations

Image filtering is computationally intensive on digital computers. Analog VLSI implementations can perform the filtering in less time and less power. Parallel analog networks capable of sensing two-dimensional input and computing the output quickly have been attracting growing interest. One drawback of these analog filters is the loss of accuracy due to limited precision of the circuit components with which they are implemented. Analysis often assumes these components, eg. resistors and transconductance amplifiers are perfectly matched linear elements. In practical realizations of these networks, deviations from these conditions, i.e. spatial mismatch among nominally identical elements and non-linearities in the elements will lead to errors between the ideal and actual outputs. If the effect of these nonidealities on the output can be quantified, the sensitivity of the different circuit realizations to the nonidealities can be compared. In this work, we analyse the distortion caused by these nonidealities by modelling their effects using equivalent input noise generators, a method commonly exploited by circuit designers. Using this methodology, the original problem can be simplified and easily analyzed. With the robustness analysis, circuit designers are able to predict the circuit performance and choose the implementations which are most robust to nonidealities in the circuit elements