brightlab.Made for curious mindsBrightlab
Year 2 · Classification
Which known creature is most like your new one? Change one feature.
Go to the investigation ↓
Investigate with Pip the possumA repaired training set with three fixed-query prediction traces.
Which known creature is most like your new one? Change one feature.
This classifier remembers labelled examples. It measures how far a new creature is from each example using two invented features: antenna length and shell width. It copies the label of the closest example. A label is supplied by a person; it is not hidden inside the creature. This tiny nearest-neighbour model calculates distances but does not understand living things.
Before this lesson: Compare lengths and count spots using two features. Useful earlier investigations: The sorting stream.
This classifier remembers labelled examples. It measures how far a new creature is from each example using two invented features: antenna length and shell width. It copies the label of the closest example. A label is supplied by a person; it is not hidden inside the creature. This tiny nearest-neighbour model calculates distances but does not understand living things.
Compare lengths and count spots using two features. Useful earlier investigations: The sorting stream.
Draw four invented creatures with differing antenna lengths and shell widths. Give two the label Dune and two Reef. Place them on a large two-axis floor or desk map. Do not use real animal taxonomy as ground truth.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 45 minutes
Students describe antenna and shell before reading labels. Place paper examples on the map.
Ask: “Is Dune a measurement or a name we supplied?”
Listen for: “A name; shell width is a measurement.”
Mark the unknown creature’s nearest example before revealing distances.
Ask: “Which example do you expect the machine to copy?”
Listen for: “The one closest on the map.”
Move the unknown creature using two coordinates. Watch distance lines and the predicted label. Keep training examples fixed and record points on opposite sides of a change.
Ask: “Did the machine see a real creature, or just two numbers?”
Listen for: “Just the numbers we gave it.”
Change the label of the closest training example without moving it or the query.
Ask: “Why did the name change when the features did not?”
Listen for: “The machine copied our changed label.”
Find a wrongly labelled example, repair it and test three fixed query locations. Keep an original prediction snapshot.
Ask: “What evidence supports your repair?”
Listen for: “The label disagreed with the rule we had agreed for our invented creatures.”
Each pair gives one oral explanation supported by a distance trace.
Ask: “Does the closest example guarantee the right label?”
Listen for: “No; the example could be wrong or missing.”
The label lives inside the creature.
Relabelling the nearest example changes the prediction without changing any creature features. Equal distances also require a tie rule.
Repair one deliberately inconsistent label, then evaluate the same three unknown creatures before and after.
Ask the student to point to the training label responsible for a prediction. If they appeal to the creature’s personality, return to the actual numeric inputs.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Representation | Confuses feature and label | Names features and labels separately | Explains what the representation omits |
| Prediction evidence | Guesses a creature name | Identifies the nearest example | Uses distance and relabelling tests to explain a change |
Use a one-dimensional antenna line first and tangible rulers. Allow a partner to operate coordinates while the student predicts.
Find a tie and specify a deterministic tie rule; compare copying one neighbour with voting among three.
Use only fictional creatures. No classmates, face images or cultural material are classified. Labels describe invented habitats, not worth or ability.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
Which known creature is most like your new one? Change one feature.
Examples: 1 = (2,2), Dune; 2 = (3,7), Dune; 3 = (7,3), Reef; 4 = (8,8), Reef. The unknown starts at (4,5). Coordinates are antenna length and shell width; copy the nearest label.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Compute nearest-neighbour votes over a dense query grid on MPS/CUDA. Compare one and three neighbours, and measure how relabelling one training example changes the boundary.
Teacher-operated extension: students predict and interpret the visual report together.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download Mac Studio investigation ↓unzip y2-creature-mac-pathway.zip -d y2-creature-mac cd y2-creature-mac bash setup-mac.sh source .venv/bin/activate python experiment.py --device mps --output results marimo edit notebook.py
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
NVIDIA DGX Spark · PyTorch CUDA
Compute nearest-neighbour votes over a dense query grid on MPS/CUDA. Compare one and three neighbours, and measure how relabelling one training example changes the boundary.
Teacher-operated extension: students predict and interpret the visual report together.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download DGX Spark investigation ↓unzip y2-creature-dgx-pathway.zip -d y2-creature-dgx cd y2-creature-dgx bash run-dgx.sh # Open the localhost notebook URL printed by the container.
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
A repaired training set with three fixed-query prediction traces.
Download the editable handout →