Demo: Spiderweb - Reliability of AI on the Edge, Effects of Hardware Disturbances on Machine Learning Software

As Artificial Intelligence is being deployed more and more on edge devices, i.e. embedded systems, this begs the question: can these algorithms deal with the (transient) hardware errors known to happen on these devices? This demo answers this question by performing fault injection on an auto-encoder algorithm executing on a typical edge device. We demonstrate that these types of algorithms are still susceptible to the effects of transient hardware errors but that traditional software-implemented fault tolerance techniques remain effective in helping them mitigate their impact.