Exploiting Data-driven Inference Towards Low-energy Implementations in Intelligent Sensors
For designers of sensor systems, faced with increasingly severe resource constraints (energy, area, bandwidth reliability), the focus on inferences from sensor data, rather than the sensor data itself, is a VERY liberating thing. While sensor data may express inferences of interest through extremely complex correlations, we now know quite broadly that these can be effectively modeled and analyzed through data-driven algorithms. What is liberating is that research in low-power systems is showing that not only can such algorithms be effectively mapped to resource-constrained implementations, but in fact such algorithms can actually relax the implementations themselves. As an example, I describe how data-driven learning enables us to select inference functions and/or parameters that are preferred from the perspective of low-energy implementation and further enables the implementations to exhibit substantially imperfect behaviors. Then, I look at how this can be exploited within systems architectures to alleviate traditional pain points (sensor acquisition, data conversion, memory operations). Measured results from several custom integrated-circuit prototypes are presented.