Accelerators for Video Analytics****

点击翻译按钮获取中文摘要

This project will explore the architecture of the machine intelligence accelerators (MIAs), with emphasis on the deep learning from video streams, leading to accelerated video analytics. Among the solutions for the core inference steps from video streams, we will explore the use of sparsity, arithmetic precision, as well as the nonstandard arithmetic encodings. Detailed performance evaluation with respect to throughput, circuit complexity and power, as well as the learning tasks, will be performed over multiple candidate solutions, and the architecture tuned for video analytics presentation will be implemented in Field-Programmable Gate Arrays. The performance of the selected candidate solutions will be evaluated against the realistic, large-scale real-life video streams that represent the typical benchmarks of the industrial partner, TwentyBN. The typical workloads of the software implementations by TwentyBN will be used for evaluating the performance of the overall solution. ****