How to Teach ML

Category	Definition	Anchor example	Coding rule
Reinforcement Learning	An agent learns by acting in the environment and adjusting its policy from reward or punishment.	In 69 (Maze) activity learners have to execute a reinforcement learning algorithm.	Q‑Learning or other reinforcement‑learning algorithms
Unsupervised Learning	Algorithms discover patterns or structure in un‑labelled data.	In 63 (SmileyCluster) learners group smiley faces with k‑means and explore the effect of changing k.	Clustering, frequent itemset mining or other unsupervised algorithm
Supervised Learning	The model is trained on labelled data and learns to predict the output for new inputs.	In 78 (Scratch‑NB) learners label hand‑sign images and train a Naïve Bayes classifier to recognise them.	Classification or regression

Category	Definition	Anchor example	Coding rule
Machine	Learners embody the algorithm. They perform each step exactly as a computer would.	In 69 (Hexapawn) one of the students has to run a program as a computer with clear rules.	Whenever learners execute the algorithm step‑by‑step, simulating the computer’s role.
User	Learners interact with the model (e.g. giving labels) but do not modify its internal logic.	In 79 (ScratchML4K) students input text, add labels, click «Train», then experiment with predictions.	Whenever learners interact with the algorithm without executing or programming it themselves.
Creator	Learners design or implement substantial parts of the algorithm.	In 22 (CONVO) you can program a procedure to use a voice assistant.	When the learner can change or program the algorithm.

Category	Definition	Anchor example	Coding rule
Invisible Model Deployment	The learner integrates and applies the trained model directly into code without needing access to the inner workings of the algorithm.	In 67 (Capture It!) you can use a block to recognise a cup of coffee with your smartphone camera.	A pre‑trained model is embedded in the activity and can neither be opened, inspected, nor configured.
Invisible Return Value	The learner interacts with the model and receives a return value.	In 78 (Scratch‑NB) you can use Naïve Bayes blocks and get a numeric prediction.	Only in conjunction with model deployment. The learner receives a return value (prediction, score, probability).
Invisible Use	The learner uses the algorithm with no transparency to the internal processing or mechanisms.	In 27a (Sign Language) you interact with a model via webcam and receive a feedback score for the classification.	The learner has to rely on the algorithm’s output without being told how it is generated.
Invisible Button	The learner activates the algorithm through a simple interface action (e.g. clicking a button), while the internal operations remain hidden.	In 79 (ScratchML4K) you click a button after the labelling step to train a model.	A single click (e.g. «Train», «Classify») triggers the training of the model.
View	The learner can transparently observe each step of the algorithm's operation.	In 51 (DoodleIT) you can see the different steps of the neural network using a visualisation.	At least one internal process is displayed live or after execution.
Execute	The learner explicitly executes the algorithm step‑by‑step.	In 69 (Maze) you have to execute the given algorithm to solve a maze.	Manually execute the algorithm in steps. No change of the algorithm’s structure or parameters while stepping.
Parameter	The learner can manipulate algorithm‑specific parameters and then re‑run the algorithm.	In 63 (SmileyCluster) you can change the number of clusters in k‑means and compare the results.	Change one or more parameters, re‑run the algorithm and observe the changed result.
Edit	The learner completes or structures parts of the algorithm with substantial guidance.	In 90 (Gold Rush) you have to reorder and edit block-oriented code.	Edit, replace or reorder parts of the algorithm with scaffolding.
Create	The learner independently develops the entire algorithm or substantial parts of it without detailed guidance.	In 22 (CONVO) you can program a procedure to use a voice assistant.	Designing parts of the algorithm from an empty workspace.

Category	Definition	Anchor example	Coding rule
Data preparation	Collecting, cleaning, or transforming raw data prior to training or testing a model.	In 147 (Google Teachable Machine + Scratch) you prepare the datasets yourself.	The activity explicitly includes data collection or preprocessing steps.
Test	Check model performance on data.	In 27 (Sign Language) you test a model with different hand gestures.	Only in conjunction with data. A distinct set of unseen examples is used for evaluation.
Block‑oriented	Drag‑and‑drop, visual‑block environments (Scratch, Snap!, Blockly) are used to build, train or use the model.	In 123 (Minecraft Education AI 4 Oceans) you program an agent with visual blocks.	Programming environment uses drag‑and‑drop blocks.
Unplugged	The algorithm is enacted without computers.	In 37 (Hexapawn) you play Hexapawn offline on a board.	The learning activity requires no digital device.
Plugged	Learners work with a text‑based coding environment or GUI on a computer.	In 56 (VotestratesML) you use an online tool to analyse democratic elections.	Primary medium is typed code or a GUI on an electronic device.
Neural Network	Layered artificial neurons whose weights are repeatedly adjusted.	In 51 (DoodleIT) you draw sketches while a multilayer network highlights activations in each layer.	Whenever the learning algorithm is a neural network model.