Dependable Neural Networks Through Redundancy, A Comparison of Redundant Architectures

With edge-AI finding an increasing number of real-world applications,\nespecially in industry, the question of functionally safe applications using AI\nhas begun to be asked. In this body of work, we explore the issue of achieving\ndependable operation of neural networks. We discuss the issue of dependability\nin general implementation terms before examining lockstep solutions. We intuit\nthat it is not necessarily a given that two similar neural networks generate\nresults at precisely the same time and that synchronization between the\nplatforms will be required. We perform some preliminary measurements that may\nsupport this intuition and introduce some work in implementing lockstep neural\nnetwork engines.\n