Hardware Specialization in Low-power Sensing Applications to Address Energy and Resilience

This paper explores implications of introducing machine learning capabilities within a hardware-specialized platform for low power embedded sensing applications. Such a platform enables algorithms well suited for analyzing complex sensor signals under strict energy constraints. However, the benefits go further, enabling the effects of errors to be overcome in the presence of hardware faults within the platform. Although errors can result in substantial bit-level perturbations, the approach described views these an alteration on the way that information is encoded within the embedded data. The new information encoding can thus be learned in the form of an error-aware model. The energy implications of hardware-specialized machine-learning kernels are analyzed using a fabricated custom IC, and the hardware-resilience implications are analyzed using an FPGA platform, which permits controllable and randomized injection of logical hardwarefaults.