Parametrizable Neural Networks on Embedded FPGA Systems (Forschungspraxis)


Andreas Högger

08.07.2014, 14:00, room 4981


We investigated how a neural network can be implemented on an FPGA, and
whether there is an advantage in computational speed over a CPU solution. Our
motivation is to realize a self-learning control system that needs no
model of the plant at all. Towards this, we leveraged the fast dynamic
reconfiguration capabilities of the FPGA, which permits the learning process
in the first place, e.g., enables to change the weights of the
neural network during run-time.