Performance Comparison of FPGA-based Convolutional Neural Networks by Internal Representations

Abstract

Reconfigurable Field-Programmable Gate Arrays (FPGAs) have prospects for applying in mobile and wearable electronics. FPGA-based neural networks have strong advantage in energy consumption comparing to another solutions. For further improving of their energy efficiency it is appropriate to study the individual network parameters effect on the entire system performance. By the reason, different internal representations variants of convolutional neural network (CNN) were compared and investigated. The study involves an accuracy parameters analysis with restricted memory for weights and increasing the network depth. Binary parameters were chosen for FPGA implementation as more efficient. Binarized CNN was compared with equal CNN by memory comsuption of weights. In addition, the mathematical problem statement of realizing binarized neural network is considered.