The project group will follow two approaches: First, we will study the approximation of ML techniques to be able to cope with resource limitations. Approximation can be done at the algorithm/software level, for example by (properly) deleting edges in a neural network to reduce the number of computations. Approximation can also be supported by the hardware level, for example by reduced precision data types or inaccurate arithmetic components. Both types of approximations trade off reductions in energy and hardware for decreased ML accuracy. Second, we will employ modern reconfigurable systems-on-chip platforms for implementation. Such platforms, for example the Xilinx UltraScale+MPSoC, provide a combination of different ARM CPU cores (64 and 32 bit), an embedded GPU, reconfigurable hardware, and memory and peripherals. We will implement ML techniques such as neural networks, kth-nearest neighbour, and random markov fields, on such a heterogeneous platform and leverage optimized hardware accelerators.
The project group will be done in cooperation with Weidmüller. We will discuss our goals and developments with industry and demonstrate the performance of our EML techniques on real industrial data provided by Weidmüller. In the course of the project, we will also schedule an excursion to a Weidmüller production site.
For more information, please check out the slide set from the project group presentation or contact the advisors: