Adder-Only Convolutional Neural Network with Binary Input Image

Mayank Palaria, Sai Sanjeet, Bibhu Datta Sahoo, Masahiro Fujita

Abstract: Convolutional neural networks (CNNs) have performed exceptionally well on a variety of image classification tasks but need significant amount of memory and computational resources. In this paper, we propose an adder-only CNN (AO-CNN) inference engine, that has its weights and/or outputs at each layer reduced to either powers of two or zero, thus replacing the multiplication operations with a simple right or left shift and hence reducing the computation significantly. The proposed AO-CNN architecture, demonstrated using three sets of data-sets, viz., MNIST, EMNIST, and SVHN, performs with ≈ 80% accuracy or more, while using minimal hardware and taking only binary input images, i.e., 1 and 0. This could pave the way for realization of image-sensors with 1-bit resolution images and the corresponding CNN based classification engine being integrated into low-power internet-of-things (IoT) devices.