INTERMITTENT-COMPUTING-AWARE DEEP LEARNING FOR BATTERYLESS IOT NODES: JOINT OPTIMIZATION OF INFERENCE SCHEDULING AND ENERGY HARVESTING DYNAMICS

Abstract

A batteryless sensor node lives on whatever energy it can collect from its surroundings, and that supply rarely behaves. Running a neural network on such a node means coping with power failures in the middle of an inference, the cost of saving progress to non-volatile memory, and results that arrive too late to be of use. We describe a scheduler that treats these problems as a single decision: how deep an exit of a multi-exit network to run, and when to run it. It looks at the energy stored in the capacitor and a running estimate of the harvested power, keeps its spending in line with what the harvester can sustain, waits briefly when a deeper exit is within reach before the deadline, and declines events it cannot afford. Every inference it starts runs to completion without a power failure. In a 24-hour simulation of a solar-powered node, with mean harvested power between 57 and 908 microwatts, we compared it with three baselines. The classifier accuracies come from a public digit recognition data set, while the energy costs are assumed values. Counting correct and timely results per sensing event, the scheduler scored between 0.505 and 0.970, against 0.069 to 0.571 for a checkpointed full network, and it beat a greedy energy aware policy by 0.9 to 4.9 points. Hardware measurements are still needed to confirm these results.