Keynote: Incremental Deep Learning For Satellite with KubeEdge...
summary
Experiment shows cloud-native incremental learning running on LEO satellites using KubeEdge Sedna with a TinyMS front end and MindSpore training on ground. A YOLOv3-tiny style model <=30 MB runs onboard (edge runtime ~17 MB), hard samples compress and upload over a ~1 Mbps link, a 3 MB compressed partial model is pushed back and applied to update weights, and iterative retraining on ground measurably increases detection accuracy for remote sensing.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:51Low Earth Orbit Satellites
- 2:04Cloud Native + Al For Space
- 2:50Cloud Native For Space
- 3:50Community Growth
- 4:17KubeEdge Architecture
- 5:41KubeEdge Sedna
- 6:47Cloud-Native Satellite
- 10:27MindSpore: Open Source All-Scenario Al Framework
- 12:46MindSpore: Top Al Open Source Community Building in China
- 13:28MindSpore: Incremental Deep Learning Made Simple
- 14:01Deep Space Learning and Cloud Native Space Computing
- 14:31Join the futuristic open source community!
related talks
32mFrom Ground to Space: Cloud-Native Edge Machine-Learning Case Studies with KubeEdge-Sedna
10mKeynote: Break Through Gravity! Cloud Native Towards Space With KubeEdge
18mDeep learning for Space imagery applications
46mDeep Learning for Satellite Image Processing, Lessons Learned
30mAI On-board of Satellites, Towards Autonomous Scientific Instruments
32mKubeEdge DeepDive: Extending Kubernetes To the Edge With Real-World Industry Use Case
07mKubeEdge Everywhere: From Graduation To Global Adoption
57mMachine Learning in the Age of IoT