Distributed Edge Intelligence for AI-empowered Applications

System Architecture of E-tree

Distributed Edge Intelligence for Pedestrian Re-Identification

Overview

We focus on distributed edge intelligence for AI-empowered applications. We develop a new training solution, E-tree learning, that leverages the tree structure imposed on edge devices. The tree structure, locations, and aggregation orders on the tree are optimally designed to reduce synchronization time and improve training efficiency. Evaluation results show that E-tree learning outperforms existing federated learning and Gossip learning approaches in convergence and model accuracy. We also work on distributed edge inference by developing a platform for deploying AI models for mission-critical applications like collaborative real-time video surveillance. Our designed platform also includes a scheduler that offloads the AI models used for pedestrian re-identification to reduce latency and enable computation sharing among devices.

Demo

Achievements

• Mingjin Zhang, Jiannong Cao, Qianyi Chen, Yuvraj Sahni, and Lei Yang, Distributed Edge Intelligence for Collaborative Real-time Video Surveillance. HotEdgeVideo'2021: ACM MobiCom Workshop Proceedings, 3rd Workshop on Hot Topics in Video Analytics and Intelligent Edges, 2021. (under submission)
• Lei Yang, Yanyan Lu, Jiannong Cao, Jiaming Huang, and Mingjin Zhang. E-Tree Learning: A Novel Decentralized Model Learning Framework for Edge AI. IEEE Internet of Things Journal, 2021.

Members

Dr. Yuvraj Sahni, Dr. Lei Yang, Mingjin Zhang, Qianyi Chen, Sami Hormi

Funding

HK RGC General Research Fund (GRF), HK$715,500, 01/2019-12/2021