Enabling hardware relaxations through statistical learning

Machine-learning algorithms are playing an increasingly important role in embedded sensing applications, by enabling the analysis of signals derived from physically complex processes. Given the severe resource constraints faced in such applications (energy, functional capacity, reliability, etc.), there is the need to think about how the algorithms can be implemented with very high efficiency. This paper examines the opportunities on three levels: (1) inherent resilience against computational errors, enabling some degree of fault tolerance; (2) top-down training of statistical models using data explicitly affected by errors, enabling substantial fault tolerance; and (3) bottom-up specification of inference kernels based on preferred hardware implementation, enabling reduced hardware complexity. Implementations employing the last two approaches are proposed and evaluated through hardware measurements and simulation.