Fully Binarized Convolutional Neural Network for Accelerating Edge Vision Computing

We propose BitFlow-Net, a method to simplify binary CNN model to almost no floating-point multiplication at inference time. Recently, a lot of variations of binarized networks were proposed trying to achieve high accuracy while replacing resource-consuming floating-point multiplication with bit operation. These methods usually require scaling factor and BatchNorm to achieve comparable accuracy as their full-precision counterparts. However, data flow have to be frequently converted between floating-point data and bit data due to the multiplication with scaling factor and in BatchNorm. Such conversion will cost extra resources and time when implemented on edge hardware. Motivated by that, we further explore and reveal some basic attributes of BNN based on previous works and propose a new method to simplify binary network. As a result, our model could inference with most of its data flow remains bit flow. Such a network architecture will greatly reduce the design complexity when implemented on ASIC or FPGA. Our method performs no accuracy degradation on ImageNet compared to state-of-the-art BNN models but without extra floating-point multiplications.