22. September, 10:30 - 11:00 Uhr mit Prof. Hans Dermot Doran, Zürich University of Applied Sciences
Making AI Architectures and Models Applicable in Functional Safety
In this presentation we review the latest state-of-the-art in safety and neural network related technologies with a specific focus on those that are applicable to industry.
We begin by considering the dependability of operations on the hardware level in AI architectures and show what techniques are used both in industry and suggested in research and appear viable for actual implementation. We then look at the monitoring approach where some bounding function is applied on the output of either the entire model or the results of individual calculations. We then describe more complex approaches where the model itself is analysed for potentially vulnerable calculations and conclusions drawn as the best way to ensure the dependability of operation.
In all these cases we focus on the practicality of the potential solutions and, where possible, estimate the resources required. We conclude with a suggestion of trends with application potential for industry in general.