Keynote Speakers ew conference

Prof. Hans Dermot Doran, Zürich University of Applied Sciences

Track Keynote: Dependable neural networks through redundancy – comparing architectures (Wednesday, 03 March 2021, 11:00 - 11:30)

In this presentation we make a contribution to the debate on redundant architectures in machine learning. Backing up our discussion with substantial experimental evidence on both standard GPU and FPGA platforms, we examine the computational cost of a common ML classification task, examine the real-time characteristics of the task, and, by explaining how this task is implemented on the platform, the actual real-time characteristics on these platforms. We examine the coordination/synchronisation issues between the redundant components and clearly enumerate the dependability considerations that need to be taken into account in this area as well. We also discuss the maintainability and extensionability of both the configurations and platforms. The presentation, aimed at decision makers as well as R&D interested, completes with a review and suggestions for further work/consideration.