RECL - Responsive Resource-Efficient Continuous Learning for Video Analytics
Offered By: USENIX via YouTube
Course Description
Overview
Explore a groundbreaking video analytics framework presented at NSDI '23 that revolutionizes continuous learning for real-time data processing. Dive into RECL (Responsive Resource-Efficient Continuous Learning), a system that seamlessly integrates model reuse and online retraining to swiftly adapt expert DNN models to specific video scenes. Discover how RECL overcomes the limitations of periodic retraining and historical model selection by implementing a shared "model zoo" across edge devices, employing rapid expert model selection, and dynamically optimizing GPU allocation. Learn about the framework's performance gains demonstrated through extensive evaluation on real-world videos for object detection and classification tasks. Gain insights into how RECL addresses the challenges of data drift and resource efficiency in video analytics, potentially transforming the field of continuous learning for large-scale applications.
Syllabus
RECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
Taught by
USENIX
Related Courses
Computer Vision: The FundamentalsUniversity of California, Berkeley via Coursera Detección de objetos
Universitat Autònoma de Barcelona (Autonomous University of Barcelona) via Coursera Deep Learning Summer School
Independent Deep Learning in Computer Vision
Higher School of Economics via Coursera Computer Vision and Image Analysis
Microsoft via edX