Parametric fault identification and dynamic compensation techniques for cellular neural network hardware

Abstract

Listen

Translate article

Testing strategies to quantify parametric faults in a fully programmable, two-dimensional cellular neural network (CNN) are presented. The approach is intended to quantify system offsets, time constant mismatches, nonlinearities in the multipliers and state nodes, and the magnitude of the dynamic range of operation which can lead to misconvergence in the CNN array. For some cases, the authors present dynamic solutions by compensating the templates, the input data, and/or the initial condition values to minimise or cancel the undesired effects. The proposed dynamic compensation techniques can be applied to any CNN independent of the array size or topology. To demonstrate the feasibility of the proposed techniques, the authors examine their application to an actual complex VLSI CNN implementation.