A workflow proposal for Safe-AI

The implication of the term safe and dependable AI is reliable execution of reliable AI models. Due to the massive computing-parallelism required by many AI-based classifiers to operate under deadline constraints, achieving reliable execution is considered a challenge – but not an insurmountable one. The state of the art in the manufacture of reliable AI models is less encouraging. Regardless of the state of the art it is anticipatable that a workflow compatible with functional safety and safety of the intended function is necessary. We approach this workflow based on industry use cases offered by the REBECCA project. In this presentation we introduce the state of the art in reliable execution, reliable AI modelling and the architectures to achieve reliable execution. We discuss why we choose CNNs as an AI architecture, how we architect the CNN into a safety/reliability framework, how the reliable execution is achieved. We then discuss a proposal to reduce the additional computational complexity introduced by reliable (redundant) execution. We then discuss the general requirements of a safe-AI workflow and present a proposal. The proposal centres around the ONNX modelling concept. We present the first trials including, not always successful, efforts with state-of-the-art tools. Relevant performance characteristics resulting from application of the workflow are presented.