Source: https://mokxi.com/learn/path/tinyml-from-scratch
Updated: 2026-11-20

All paths

# Machine learning on a microcontroller

Gesture recognition from scratch in plain Python on a Raspberry Pi: features, k-nearest neighbors, decision trees, a neuron, gradient descent and a tiny neural network, and weights exported as C arrays for a microcontroller.

Raspberry Pi 4 Model B Free 7 lessons 5 h 4 chapters 3,070 XP Coding · Expert +400 XP

0 of 7 done

Start the course free Data and features Lesson plan for teachers

What you will build

## 3 projects, built step by step

7 lessons in 4 chapters, about 5 hours on a simulated Raspberry Pi 4 Model B. Free, in your browser, every step checked by the simulator.

Chapter 1 project

### k-nearest neighbors

The simplest model there is: label a gesture like its nearest training gestures. Distance, votes, accuracy, choosing k, a confusion matrix, why the z-scores matter, and what k-NN costs a microcontroller.
Open it in the simulator

Chapter 2 project

### A neuron

One artificial neuron: a weighted sum, a bias and a step. Set its weights by hand, then let the perceptron rule learn them, find where a single neuron gives up (a problem that is not linearly separable with one feature), and fire a lamp from it.
Open it in the simulator

Chapter 3 project

### Weights in C

Ship the model: print the trained weights and the z-score constants as C arrays a microcontroller sketch can paste in, quantize them to int8, prove the small model still agrees, and print the C classify function that uses them.
Open it in the simulator

The map

## Chapter by chapter

Each lesson opens when the one before it is done, and each chapter ends in a project.

Chapter 1

## Data first

Turn raw sensor traces into features and split them honestly.

Open 330 XP

### Data and features

Gestures from a wrist accelerometer: raw traces, six features that summarize half a second in six numbers, class averages that show why it can work, an honest train/test split, and z-scores computed from the training set only.

0 of 5 steps

Start What is in it

Project Locked 660 XP

### k-nearest neighbors

The simplest model there is: label a gesture like its nearest training gestures. Distance, votes, accuracy, choosing k, a confusion matrix, why the z-scores matter, and what k-NN costs a microcontroller.

0 of 5 steps

5 questions, graded as you go.

Chapter 2

## Models you can read

Decision trees and a single neuron, and where each one breaks.

Locked 280 XP

### Decision trees

A model you can read: the best first question by Gini impurity, a depth-2 tree printed as rules, overfitting made visible with three mislabeled gestures, and the tree turned into C if statements.

0 of 4 steps

Project Locked 560 XP

### A neuron

One artificial neuron: a weighted sum, a bias and a step. Set its weights by hand, then let the perceptron rule learn them, find where a single neuron gives up (a problem that is not linearly separable with one feature), and fire a lamp from it.

0 of 4 steps

4 questions, graded as you go.

Chapter 3

## Train and ship

Gradient descent on the Pi, then weights a microcontroller can run.

Locked 330 XP

### Training on the Pi

Gradient descent from scratch: softmax regression for all three gestures, the cross-entropy loss falling epoch by epoch, test accuracy and confidence, learning rates too small and too big, and a 6-4-3 neural network trained by backpropagation.

0 of 5 steps

Project Locked 560 XP

### Weights in C

Ship the model: print the trained weights and the z-score constants as C arrays a microcontroller sketch can paste in, quantize them to int8, prove the small model still agrees, and print the C classify function that uses them.

0 of 4 steps

5 questions, graded as you go.

Chapter 4

## Boss

Live gesture recognition on the traffic lights.

Locked 350 XP

### Boss: gesture recognition

The boss: the whole pipeline live. Press the button, a fresh gesture is recorded, featurized, z-scored and classified, and the traffic lights show wave, punch or circle; a confidence threshold says "not sure" to a gesture it was never taught; and 30 never-seen gestures measure it honestly.

0 of 6 steps

3 questions, graded as you go.

Syllabus

## Every lesson in the course

All 7 lessons, in order, each with what it covers. Any of them opens in the editor right away.

Chapter 1

### Data first

- Data and features Gestures from a wrist accelerometer: raw traces, six features that summarize half a second in six numbers, class averages that show why it can work, an honest train/test split, and z-scores computed from the training set only.
- k-nearest neighbors The simplest model there is: label a gesture like its nearest training gestures. Distance, votes, accuracy, choosing k, a confusion matrix, why the z-scores matter, and what k-NN costs a microcontroller.

Chapter 2

### Models you can read

- Decision trees A model you can read: the best first question by Gini impurity, a depth-2 tree printed as rules, overfitting made visible with three mislabeled gestures, and the tree turned into C if statements.
- A neuron One artificial neuron: a weighted sum, a bias and a step. Set its weights by hand, then let the perceptron rule learn them, find where a single neuron gives up (a problem that is not linearly separable with one feature), and fire a lamp from it.

Chapter 3

### Train and ship

- Training on the Pi Gradient descent from scratch: softmax regression for all three gestures, the cross-entropy loss falling epoch by epoch, test accuracy and confidence, learning rates too small and too big, and a 6-4-3 neural network trained by backpropagation.
- Weights in C Ship the model: print the trained weights and the z-score constants as C arrays a microcontroller sketch can paste in, quantize them to int8, prove the small model still agrees, and print the C classify function that uses them.

Chapter 4

### Boss

- Boss: gesture recognition The boss: the whole pipeline live. Press the button, a fresh gesture is recorded, featurized, z-scored and classified, and the traffic lights show wave, punch or circle; a confidence threshold says "not sure" to a gesture it was never taught; and 30 never-seen gestures measure it honestly.

Kits

## Build it on your bench too

Every lesson runs in the simulator with nothing to buy. When you want the real thing, these kits have the parts.

Have a kit already?

### Find your kit

Pick the starter kit you own and run every lesson in its manual with the same parts on the same pins.
See every kit

Questions

## About Machine learning on a microcontroller

Is Machine learning on a microcontroller free?

Yes. All 7 lessons of Machine learning on a microcontroller are free and run in your browser, with nothing to install and no parts to buy before you start.

Do I need a real Raspberry Pi 4 Model B?

No. Every lesson runs on a simulated Raspberry Pi 4 Model B with its parts on the canvas, and the simulator checks each step as you go. When you want to build it for real, the same wiring and code work on a real board.

How long does Machine learning on a microcontroller take?

About 5 hours: 7 lessons in 4 chapters, each one a few short steps. Your progress is kept, so you can stop after any lesson and pick up where you left off.

What will I build?

Each chapter ends in a project that puts it together: k-nearest neighbors, A neuron and Weights in C.

Where do I start?

With "Data and features", and each lesson builds on the one before it. New to circuits? Electronics basics is the course that starts from zero.
