AI workflow for large scale deployment of far-edge ML devices
summary
System describes a concrete pipeline: peel-and-stick sensor with accelerometer and a cellular modem sends raw vibration to cloud, containerized training spins up a separate VM per device so millions of ML models can be trained in isolation, then the cloud pushes a compiled model back for on-device inference so the modem stays asleep until an anomaly triggers upload. Uses tensorflow light for microcontrollers, Ambach Apollo 3–class low‑power MCUs, Bluetooth config and autoscaled VMs (sage maker Google tf‑x).
AI-written summary · about
chapters (10)
- 0:00Intro
- 1:41Asset Management Systems Today
- 2:02What's Shaping these Systems? Constraint
- 4:29How we're designing our next systems
- 5:13Anomaly Detection Demo TF Lite Micro on Cortex M4
- 8:08Adoptability: The two major user groups
- 9:26Adoptability: Design Considerations
- 11:18Flexible Model Development Workflows
- 12:04Auto Scaling Backends
- 13:21Decoupled Device Agents
related talks
54mSmarter Everything with ML & the IoT
59mML and the IoT: Living on the Edge
42mSo, You Want to "Do Big Data?"
49mEnable IoT with Edge Computing and Machine Learning
01h 01mML and the IoT: Living on the Edge
25mTinyML as-a-Service
18mManagement of IoT TinyML Devices
45mAIoT Ops - Using Kubernetes and ML Ops to Build Edge ML Applications