Exploiting System Configurability Towards Dynamic Accuracy-Power Trade-Offs In Sensor Front-Ends

Analog-to-information converters and Compressed Sampling (CS) sensor front-ends try to only extract the relevant, information-bearing elements of an incoming data stream. Information extraction and recognition tasks can run directly on the compressed data stream without needing full signal reconstruction. The accuracy of the extracted information or classification is strongly determined by the front-end settings and tolerated level of hardware impairments. Exploiting this, allows to dynamically tune accuracy for power consumption. This paper discusses this trade-off and introduces a theoretical framework to guide the selection of optimal hardware settings under given power or accuracy constraints. This is illustrated with two circuit realizations: 1) an analog-to-information converter for voice activity detection (VAD), and 2) a CS photopletysmographic (PPG) heart rate (HR) extraction application.