Edge AI Systems
summary
Covers how to get real-time, reliable video analytics by moving work from cloud or device into an edge layer and gives concrete systems: Maui methods and CloneCloud use JVM-based migration, COMET adds distributed shared memory, adaptation system AWStream profiles model latency vs accuracy, and hybrid Clownfish fuses a small local DNN with cloud results using optical flow, exponential smoothing and a tiny NN for similarity. Experiments show 10x bandwidth swings, on-device action accuracy ~51% vs cloud ~59%, and Clownfish keeps real-time throughput while recovering much of the cloud accuracy.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 1:20Rapid Development
- 3:26Market Growth
- 4:19Requirements
- 5:55Current computing landscape
- 8:49Modern distributed computing
- 10:11Defining edge
- 11:54Agenda
- 13:42Mobile Devices
- 15:40Mobile Network
- 16:31Two Approaches
- 18:05ClientServerbased
- 22:10Offloading
- 23:52MAUI
- 29:04MAUI Workflow
- 30:49CloudCloud
- 32:37CloneCloud
- 33:40Application Profiling
- 35:00Distributed Shared Memory
- 37:25Lazy Release Consistency
- 40:10Issues
- 44:29Applications
- 47:38Computing Paradigms
- 49:36High network variability
- 50:53System adaptation
- 53:50Break
related talks
36mApplied AI on the edge
50mML and AI at the edge with Google Coral, EdgeTPU and Cloud Platform
22mEdge AI Use Cases in IoT — Unlocking Intelligence at the Network Edge
31mBringing AI on-Device: From Cloud to Edge
10mThe Rise and Impact of Edge AI
47mProduction ready EdgeAI with Eclipse ioFog and Google Coral
47mBest Practices for Real-time Intelligent Video Analytics
22mExtending European Strength with Embedded AI