Machine Learning on Microcontrollers With Zephyr and Emlearn
summary
Activity-recognition pipeline that reads IMU (accelerometer+gyro) windows, extracts simple stats, trains scikit-learn/Keras models, and deploys them as C headers works end-to-end with Zephyr and emlearn. emlearn provides a Python one-liner to convert scikit‑learn/Keras models to C99, Zephyr supplies sensor APIs and recommends using IMU FIFO to reduce sampling jitter and allow CPU sleep, and real deployments (toothbrush timer, cattle monitoring) report roughly 50× power savings versus sending raw data.
AI-written summary · about
chapters (13)
- 0:00Introduction to TinyML
- 1:32TinyML development challenges
- 5:22Activity recognition case study
- 9:46Sensor data acquisition
- 15:05Batch processing with RTIO
- 16:29Creating datasets
- 18:14Labeling data
- 21:14Training machine learning models
- 22:25Deploying models with emlearn
- 25:50MicroPython and C integration
- 28:42MicroPython ecosystem for ML
- 29:25Using emlearn in MicroPython
- 31:19Summary and Q&A
related talks
44mMachine Learning on Microcontrollers With Zephyr and Emlearn
27mMicrocontrollers and Machine Learning with MicroPython and emlearn
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
36mTinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS
06mIntermittent Learning: On-Device Machine Learning on Intermittently Powered System
41mBeefy ML: Ultra-Low Power Algorithms on Cattle
29mMicrocontrollers + Machine Learning with MicroPython in 1-2-3
14mKeynote: Leveraging Zephyr and ML to Bring Smart Devices to Market, Faster