One way to improve the energy efficiency of executing neural networks (NN) is to deploy and execute them on Field Programmable Gate Arrays (FPGA). This allows to execute NNs efficiently "in hardware" while being able to adapt their structure at any time. Due to the limited amount of available resources on embedded FPGAs however NNs have to be optimised so they can be executable on them.
Instead of try to implement generic artificial neurons for a preexisting NN strtucture, we design tailormade neurons, optimised for the strengths and weaknesses of FPGAs. These neurons can then be deployed with very few hardware resources (i.e. by using a limited amount of Lookup Tables or LUTs) and can therefore be realised much more energy efficiently. We showcase the potential for this solution by detecting artifacts in ECG heart data.
Project start: 01.10.2019
Project end: 31.12.2020
Using pretrained and preoptimised neural networks we are developing AI solutions that can be executed highly efficiently on FPGAs. We show the potential by detecting artifacts in ECG heart data.