Machine Learning-Based Self-Compensating Approximate Computing

Abstract

Dedicated hardware accelerators are suitable for parallel computational tasks. Moreover, they have the tendency to accept inexact results. These hardware accelerators are extensively used in image processing and computer vision applications, e.g., to process the dense 3-D maps required for self-driving cars. Such error-tolerant hardware accelerators can be designed approximately for reduced power consumption and/or processing time. However, since for some inputs the output errors may reach unacceptable levels, the main challenge is to enhance the accuracy of the results of approximate accelerators and keep the error magnitude within an allowed range. Towards this goal, in this paper, we propose a novel machine learning-based self-compensating approximate accelerators for energy efficient systems. The proposed error compensation module, which is integrated within the architecture of approximate hardware accelerators, efficiently reduces the accumulated error at its output. It utilizes lightweight supervised machine learning techniques, i.e., decision tree, to capture input dependency of the error. We consider image blending application in multiplication mode to demonstrate a practical application of self-compensating approximate computing. Simulation results show that the proposed design of self-compensating approximate accelerator can achieve about 9% accuracy enhancement, with negligible overhead in other performance measures, i.e., power, area, delay and energy.