Custom Hardware Architecture for FPGA-Based Gated Recurrent Unit

The recurrent neural network is a two-way artificial neural network (ANN). It has its own sets of applications and variants, namely, Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). GRUs are simpler to implement compared to LSTM, being much simpler but having a capacity to model the sequential data. This reduces computational complexity and resources, making it less complicated to implement on hardware compared to the LSTM. In this paper, we propose a new GRU architecture implemented on OEM Zynq-7000 Soc XC7Z020 FPGA (field-programmable gate array) by making use of hardware re-use strategy and parallelizing computations by reducing the hardware requirements that are presented. The network achieves better speed and performance compared to the software model. We built the architecture for different network sizes and has been validated on breast cancer data set from the University of California, Irvine, Machine Learning Repository. Since validation is done on the aforementioned dataset, our neural network is application-specific for detecting breast cancer.