Optimal resource usage in ultra-low-power sensor interfaces through context- and resource-cost-aware machine learning

Abstract

This paper introduces an approach that combines machine learning and adaptive hardware to improve the e ciency of ultra-low-power sensor interfaces. Adaptive feature extraction circuits are assisted by hardware embedded training to dynamically activate only the most relevant features. This selection is done in a context-and power cost-aware manner, through modification of the C4.5 algorithm. As proof-of-principle, a Voice Activity Detector illustrates the contextdependent relevance of features, demonstrating average circuit power savings of 70%, without accuracy loss. The RECAS database developed for experimenting with this context-and dynamic resource-cost-aware training is presented and made open-source for the research community.

A 90 nm CMOS, Power-Proportional Acoustic Sensing Frontend for Voice Activity Detection

“…The extracted feature subset, af 5 -af 12 , is passed on to the on-chip classifier (Fig. 5) while the complete feature-set af 1 -af 16 can be passed on to an off-chip ADC for more complex information extraction, such as context-change detection and retraining the classifier as in [22].…”

“…-af 12 . The selected feature (sf i ) is then compared with a reference voltage (Vref i ) determined by a modified C4.5 machine learning algorithm [22], generating the output decision b i of each node: ,…”

“…Our modification to C4.5 maximizes the information-gain/watt and therefore outputs a resourceefficient model that maximizes the information capture while minimizing the power [22]. This is enabled as each feature extracts information from a higher frequency band so that the power cost increases from af 1 to af 16 .…”

“…The training runs on the microcontroller to generate a discriminating feature subset and reference levels for the comparison in the DT. The training results of the past context are not stored but dynamically learnt, as context-change is detected [22].…”