Context- and cost-aware feature selection in ultra-low-power sensor interfaces

This paper introduces the use of machine learning to improve efficiency of ultra-low-power sensor interfaces. Adaptive feature extrac- tion circuits are assisted by hardware embedded learning to dynamically activate only most relevant features. This selection is done in a context and power cost-aware way, through modification of the C4.5 algorithm. Furthermore, context dependence of different feature sets is explained. As proof-of-principle, a Voice Activity Detector is expanded with the pro- posed context- and cost-dependent voice/noise classifier, resulting in an average circuit power savings of 75%, with negligible accuracy loss.