Method for convolutional neural network hardware implementation based on a residue number system

Convolutional Neural Networks (CNN) show high accuracy in pattern recognition solving problem but have high computational complexity, which leads to slow data processing. To increase the speed of CNN, we propose a hardware implementation method with calculations in the residue number system with moduli of a special type $2^\alpha$ and $2^\alpha-1$. A hardware simulation of the proposed method on Field-Programmable Gate Array for LeNet-5 CNN is trained with the MNIST, FMNIST, and CIFAR-10 image databases. It has shown that the proposed approach can increase the clock frequency and performance of the device by 11%-12%, compared with the traditional approach based on the positional number system.