Neural Network Implementation of FPGA and Accuracy Comparison

Abstract

In this thesis, research work has been done to implement a specific trained neural network (NN) into Zybo-Z7-20 FPGA by using Xilinx Vivado software tools. The purpose of this thesis was to develop a C algorithm that can be synthesized as a register-transfer level (RTL) intellectual property (IP) by using Vivado HLS. The HLS IP was used in designing the block diagram with Zynq Processor. The result of this thesis shows that the Field Programmable Gate Array (FPGA) can compute and predict the input as accurately as software. This thesis shows that hardware can predict the result for unknown inputs same as software.

Implementing an artificial neural network (ANN) into hardware is a procedure of mapping the neural network into specific hardware in an executable form. Research in this area can be divided into two domains: general-purpose neurocomputers for a wide range of neural network models and special purpose very large-scale integration (VLSI) neural systems for specific neural network models. Nevertheless, the reconfigurability of FPGAs has shown their advantage of the flexibility and a high degree of observability into the inner working of neural algorithms for certain applications. On the other hand, the special purpose VLSI chips allow implementing a neural network with enhanced processing speed and compactness for real-time applications. The disadvantage of ANN implementation into VLSI is that the VLSI chips are fixed mapping. Hence the degree of flexibility to implement different neural models become limited. Another disadvantage of VLSI neural systems is developing such chips are time-consuming and expensive. Consequently, it is a much more feasible approach now to implement a Neural Network on FPGA than it has been in the past. Hence, this thesis focuses on how to implement a specific trained neural network into FPGA.