Experimental exploration of the performance of binary networks

Abstract

Binary neural networks for object recognition are desirable especially for small and embedded systems because of their arithmetic and memory efficiency coming from the restriction of the bit-depth of network weights and activations. Neural networks in general have a tradeoff between the accuracy and efficiency in choosing a model architecture, and this tradeoff matters more for binary networks because of the limited bit-depth. This paper then examines the performance of binary networks by modifying architecture parameters (depth and width parameters) and reports the best-performing settings for specific datasets. These findings will be useful for designing binary networks for practical uses.