An Approximate Computing Approach for the Optimization of Adaptive Algorithms

Abstract

This dissertation presents a detailed investigation on the use of approximate arithmetic circuits in di erent adaptive algorithms. The analysis is rst carried out by using the traditional work ow, which is to directly implement the functional model, i.e. the truth table of these approximate arithmetic blocks. A faster yet accurate framework is proposed, which analytically model the estimation error of di erent adaptive algorithms when approximate arithmetic circuits are employed. The novel framework is veri ed by comparing it with the functional model, i.e. the truth table based implementation of approximate LMS. The performance of both implementations are assessed for various approximation scenarios and our results show that the novel framework can model the approximate arithmetic based adaptive algorithms with an accuracy of up to 92%. Furthermore, the computation time is signi - cantly reduced, i.e. the proposed model is 4000 times faster as compared to the functional model-based implementation of approximate LMS. Since approximation introduces error, further analysis is carried out to mitigate the approximation error by proposing two estimators based on the maximum likelihood estimation (MLE) and maximum a posteriori estimation (MAP). These estimators perform signi cantly better than the traditional modeling of adaptive algorithms using approximate hardware. We also analyzed the ii