Design methodology for configurable analog to digital conversion using support vector machines

We apply a support vector machine (SVM) classifier to the design of analog to digital converters. Each output bit of the converter is the output of a binary classifier, which is a nonlinear SVM. The classifier effectively learns a folding characteristic for each bit, which is realized as the weighted sum of a set of kernel functions. In our proposal, the kernel does not need to be symmetric or positive definite, unlike in the case of a conventional SVM. This makes the approach more amenable to VLSI design, where such constraints are hard to satisfy. The SVM uses a set of binary weights, which allows the folding characteristics to be digitally corrected after fabrication. This facility is of considerable value in analog design in a deep sub micron era, where simulation models do not adequately capture the behavior of devices, or their variations. The proposed methodology reduces design time, can be automated and calibrated for process variations and ageing, by changing a set of digital scaling coefficients.