RECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
summary
System integrates a shared model zoo of previously adapted lightweight experts, a learned model selector that predicts per-frame accuracy and aggregates estimates over samples, and a dynamic GPU allocation policy that trades off retraining vs selection. Built components: accuracy-estimator DNN, top-K selection with fast comparator, admission/eviction for zoo size. Evaluation on 70 hours (150 driving sessions) for detection/classification using one Nvidia A100 shows accuracy near an exhaustive-oracle, cuts end-to-end adaptation delay from tens of seconds to the first model update (~2.6s), and avoids the 70–90% retraining resource sink.
AI-written summary · about
chapters (11)
- 0:00Intro
- 1:40Existing solution: continuous learning
- 2:35Predominant adaptation method: retraining
- 4:02Challenges of continuous retraining systems
- 5:51Lightweight model selector
- 8:16Integrating reuse with retraining
- 10:53RECL significantly improves the response time
- 11:17RECL provides effective reuse suggestions
- 12:08Managing the model zoo
- 14:27RECL improves the adaptation efficiency
- 15:40System components of RECL
related talks
17mEkya: Continuous Learning of Video Analytics Models on Edge Compute Servers
35mLarge-Scale Video Intelligence
33mArchitecting Scalable Multi-Modal Video Search
41mAI applications in autonomous driving
44m10 AIs or 10B AIs
47mBest Practices for Real-time Intelligent Video Analytics
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
15mVideo Analytics with Zero-streaming Cameras