Applying the Residue Number System to Network Inference

Use of neural networks for computer vision, speech recognition, and other applications has exploded in recent years, in part due to their unprecedented performance on a variety of benchmarks. Nonetheless, highthroughput and energy-efficient evaluation of such neural networks, and in particular, convolutional neural networks (CNNs), remains an active field of research. Evaluation of networks is memory and compute intensive, with the bottleneck depending on the network topology and layer types (convolutional or fully-connected [FCN]).