FPGA Based Real-Time Processing Architecture for Recurrent Neural Network

A field programmable gate array (FPGA)-based real-time processing architecture for recurrent neural network (RNN) is proposed and presented; the proposed FPGA processing architecture is based on echo state network (ESN) and can get the output weights of RNN in real-time. The proposed architecture and the performance have been verified on an Altera FPGA chip. Experimental results show that the real-time hardware RNN can be trained to recognize different duty cycles of the input signal. We also performed experiments to investigate the ESN demand for resources and systems convergence in FPGA.