Dynamic Sensor-Frontend Tuning for Resource Efficient Embedded Classification

Emerging portable applications, including human activity recognition and robot navigation, require always-on sensing technologies to continuously monitor the environment. Continuous sensing, however, strongly dominates the hardware devices' power consumption and consequently hampers the systems' always-on functionality. In this paper, we propose a circuit-aware machine learning scheme that exploits the devices' ability to dynamically tune the quality of its sensors to trade-off system level accuracy versus total system-level power consumption. To this end, we: 1) analytically derive the power-quality trade-off space in sensory front-ends; 2) use these equations to make the probabilistic relations between sensory features and their degraded versions explicit in a Bayesian network classifier; and 3) propose a methodology building upon this model to control the required sensory information quality at run-time. We show how this enables to tune the circuit's power consumption versus inference accuracy trade-off space with fine granularity, and achieve significant power savings at almost no accuracy loss. In addition, our methodology is able to cope with sensor failure and with a dynamically changing environment by means of an efficient on-line tuning strategy. This dynamic power-vs-quality scalability is empirically shown on various machine learning benchmarking data sets.