Implementing Binary Neural Networks.

Abstract

The recent renaissance of deep neural networks has lead to impressive advancements in many domains of machine learning. However, the computational cost of these neural models in- creases in line with their performance, with many state-of-the-art models only being able to run on expensive high-end hardware. The need to efficiently deploy neural networks to commodity platforms has made network optimization a popular field of research. One particularly promising technique is network binarization, which quantizes the weights and activations of a model to only one or two bits. Although binarization offers theoretical oper- ation count reductions of up to 32X, no actual measurements have been reported. This is a symptom of the gap between theory and implementation of binary networks that exists to- day. In this work, we bridge the gap between abstract simulations and real usable high speed networks. To do so, we identify errors in the existing literature, develop novel algorithms, and introduce Riptide, an open source system that can train and deploy state-of-the-art binary neural networks to multiple hardware backends.