HARDWARE ACCELERATION OF CONVOLUTIONAL NEURAL NETWORKS USING APPROXIMATE COMPUTING AND DYNAMIC PARTIAL RECONFIGURATION

Abstract

In this work, I have trained new four different convolutional neural networks (CNNs) to recognize four different datasets MNIST, Fashion MNIST, SVHN and CIFAR-10. Then, the CNNs are tested for recognition. The resulting trainable weights are approximated using precision scaling. The four networks are tested again while using this approximation. A new hardware architecture is proposed to recognize three datasets (MNIST- Fashion MNIST- SVHN) while using precision scaling approximation. This architecture is implemented on Xilinx XC7Z020 FPGA. The resulting power and energy consumed to recognize each image in each dataset is reported. The results show significant reduction in energy consumption while having minor loss in accuracy. This approximation is significant because CNN requires a lot of computation, and hence, consumes large power.