TinyML
Syllabus · Historical materials
| Courseware | TinyML slides and readings from the HarvardX program; each slide chapter has matching readings |
| Textbook | Volume I, as a companion reference |
| Prerequisites | Python and basic linear algebra |
| Format | 10–12 weeks, two 75-minute lectures per week |
| Hardware | One microcontroller board per student or team, from week 7; see Hardware |
Description
Students train machine learning models and deploy them on microcontrollers, where memory, energy, and latency are tightly limited. The course moves from fundamentals to applications and then to deployment on real hardware. Part IV, on MLOps, is optional.
Learning outcomes
By the end of the course, students can:
- Train and quantize a model small enough for a microcontroller.
- Deploy keyword spotting, visual wake words, and gesture recognition on a board.
- Collect a custom dataset and build an end-to-end TinyML application.
Reading schedule
Each week pairs a chapter of the TinyML courseware with a Volume I chapter. The courseware also includes Colab notebooks and, from week 7, hardware labs.
| Week | Topic | Reading | Slides |
|---|---|---|---|
| Part I: Fundamentals | |||
| 1 | Welcome to TinyML | Introduction | Ch 1.1–1.3 |
| 2 | The Machine Learning Paradigm | ML Lifecycle | Ch 2.1–2.2 |
| 3 | CNNs and Computer Vision | Neural Computation | Ch 2.3–2.5 |
| Part II: Applications | |||
| 4 | TensorFlow Lite and Quantization | Model Compression | Ch 3.1–3.4 |
| 5 | Keyword Spotting | Benchmarking | Ch 3.5–3.6 |
| 6 | Visual Wake Words and Anomaly Detection | Hardware Acceleration | Ch 3.7–3.9 |
| Part III: Deployment | |||
| 7 | Embedded Systems and TFLite Micro | Hardware Acceleration | Ch 4.1–4.4 |
| 8 | Deploying KWS and Custom Datasets | Model Serving | Ch 4.5–4.6 |
| 9 | Deploying VWW and Gesture Recognition | Model Serving | Ch 4.7–4.9 |
| Part IV: MLOps and Scaling (optional) | |||
| 10 | MLOps Fundamentals and ML Development | ML Operations | Ch 5.1–5.3 |
| 11 | Training, Conversion, and Deployment at Scale | Responsible Engineering | Ch 5.4–5.7 |
| 12 | Monitoring, Management, and Responsible AI | Responsible Engineering | Ch 5.8–5.12 |
Hardware
The HarvardX courseware uses the Arduino Nano 33 BLE Sense from the TinyML kit. The Hardware Kits labs also support these boards, with approximate prices from that site:
| Board | Approximate price |
|---|---|
| Grove Vision AI V2 | ~$25 |
| XIAOML Kit | ~$40 |
| Arduino Nicla Vision | ~$95 |
Assessment
A suggested split; each offering sets its own.
| Component | Weight |
|---|---|
| Colab assignments (weeks 2–6) | 30% |
| Hardware labs (weeks 7–9) | 30% |
| Quizzes from the edX materials | 15% |
| Final project: an end-to-end TinyML application | 25% |
Adapting this syllabus
- Each offering sets its own dates, deadlines, and policies on late work, collaboration, and AI tools.
- 6-week intensive: Parts I–III only.
- 16 weeks: Part IV spread over four weeks, with a midterm after week 6.
- Module in the introductory course: weeks 4–9, taught within Introduction to Machine Learning Systems.