Machine learning algorithm design for hardware performance optimization

Abstract

Machine learning has enabled us to extract and exploit information from collected data. In this thesis, we are particularly interested in how we can apply this powerful tool to enhance the performance of various hardware. The objective of our work is to combine techniques in machine learning, signal processing, and system control for hardware performance optimization. By leveraging collected data to construct a better model for both the hardware and the operating environment, machine learning enables the hardware to operate more power-efficiently, to obtain improved results, and to maintain robust performance against environmental changes. The proposed work targets three aims: (i) design data-driven signal processing algorithms which require fewer measurements taken from the sensor front-end; (ii) develop algorithm-hardware co-design techniques for hardware that performs specific machine learning tasks; (iii) design adaptive hardware control algorithms. For the first aim,we develop a compressive sensing recovery algorithm which achieves fast recovery speed and high recovery quality based on fewer compressed measurements. For the second aim, we propose a motion gesture recognition algorithm which works directly with video frames captured using compressive sensing techniques. The motion parameters are estimated in the compressed domain, and the estimation algorithm is implemented in the mixed-signal circuits. We also improve the computational and memory efficiency of existing gesture classifiers. For the third aim, we develop multiple Doherty PA control algorithms based on the bandit frameworks. By incorporating the prior information about the Doherty PA’s characteristics into our algorithm design, we improve learning efficiency and enable the PA to achieve robust and adaptive operations in time-variant environments.