Using Edge Machine Learning to save power and development time
summary
Demo uses a Thingy:53 with an nRF5340 to run TinyML on raw 3‑axis accelerometer data sampled at 100 Hz in 5‑second windows, classifying fan speed and flagging anomalies when imbalance occurs (demo trips a relay). Workflow uses Edge Impulse with Keras/TensorFlow export compiled to C++, then TensorFlow Lite 8‑bit quantization to drop inference from ~33 ms to ~12 ms and reduce RAM/flash. Run-time current is ~8 mA on the MCU, so inference every second can still last years on a coin cell.
AI-written summary · about
chapters (10)
- 0:00Introduction and demo
- 2:32Hardware and use cases
- 5:07Fundamentals of AI and ML
- 6:26ML development workflow
- 11:04Tooling and classification
- 17:03Anomaly detection and testing
- 18:59Model optimization
- 20:08Edge application scenarios
- 22:53Why edge AI: The BLERP framework
- 27:39Energy, future outlook, and Q&A
related talks
54mSmarter Everything with ML & the IoT
18mManagement of IoT TinyML Devices
55mEdgeAI - AI in the Palm of Your Hand
01h 01mML and the IoT: Living on the Edge
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
05mBuilding Machine Learning models with Edge Impulse
27mWhat’s TinyML good for
31mKeynote: The next big thing in machine learning is tiny