Survey on Energy-Efficient Deep Neural Networks for Computer Vision

Abstract

Deep Neural Networks (DNNs) are greatly successful in performing many different computer vision tasks. However, the state-of-the-art DNNs are too energy, computation, and memory-intensive to be deployed on most computing devices and embedded systems. DNNs usually require server-grade CPUs and GPUs. To make computer vision more ubiquitous, recent research has focused on making DNNs more efficient. These techniques make DNNs smaller and faster through various refinements and thus are enabling computer vision on battery-powered mobile devices. Through this article, we survey the recent progress in low-power deep learning to discuss and analyze the advantages, limitations, and potential improvements to the different techniques. We particularly focus on the software-based techniques for low-power DNN inference. This survey classifies the energy-efficient DNN techniques into six broad categories: (1)Quantization, (2)Pruning, (3)Layer and Filter Compression, (4)Matrix Decomposition, (5)Neural Architecture Search, and (6)Knowledge Distillation. The techniques in each category are discussed in greater detail in this chapter. Take-aways Surveys the recent progress in low-power deep learning to analyze the advantages, limitations, and potential improvements to the different techniques. Focus on the software-based techniques for low-power DNN inference