brightlab.Made for curious mindsBrightlab
Year 4 · Training data
Who disappears inside an average?
Go to the investigation ↓A revised sampling allocation and a two-habitat audit.
Whose examples shaped the reef classifier?
A training collection can describe one habitat well and another poorly. This lab learns one-dimensional class centroids separately for two synthetic reef habitats from the available samples. As collection effort changes, sparse estimates can shift and one habitat’s errors can rise. The test set stays fixed. Overall accuracy combines all test cases and can hide a smaller group’s poor result; equal sample counts alone do not guarantee fair outcomes.
Before this lesson: Compare two groups and describe a fair test. Useful earlier investigations: Beyond the collecting window; The sealed test envelope.
A training collection can describe one habitat well and another poorly. This lab learns one-dimensional class centroids separately for two synthetic reef habitats from the available samples. As collection effort changes, sparse estimates can shift and one habitat’s errors can rise. The test set stays fixed. Overall accuracy combines all test cases and can hide a smaller group’s poor result; equal sample counts alone do not guarantee fair outcomes.
Compare two groups and describe a fair test. Useful earlier investigations: Beyond the collecting window; The sealed test envelope.
Print reef and lagoon training samples and a fixed test batch. Explain that habitats are invented and the labels concern synthetic objects, never people. Rehearse how available samples change the learned centroids.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 60 minutes
Present an overall score and ask what it conceals about two habitats.
Ask: “Could most correct answers come from one place?”
Listen for: “Yes, especially if it has more test cases.”
Predict the underserved habitat’s result when more samples are allocated there.
Ask: “What must we keep fixed to compare fairly?”
Listen for: “The test objects and scoring rule.”
Move collection effort between habitats. Inspect actual training points, learned reference positions and group error counts.
Ask: “Which sample changed the reference position?”
Listen for: “The new point pulled the average toward it.”
Use equal training allocation and inspect remaining overlapping cases.
Ask: “Did equal numbers guarantee equal errors?”
Listen for: “No, the groups can still have different difficulty.”
Choose an allocation and write a report containing both group fractions and overall fraction. Include a remaining misclassified point.
Ask: “How would you explain the result without hiding the smaller group?”
Listen for: “Give each group’s count before the overall result.”
A partner reproduces the chosen allocation.
Ask: “Is a higher average enough to declare the repair successful?”
Listen for: “No, we must check who improved or got worse.”
High overall accuracy means everyone is served.
Equal training counts can still produce unequal errors because the fixed feature distributions overlap differently.
Allocate a fixed collection budget and publish group-level error counts, denominators and one unresolved case.
Check the student did not change test composition to improve the headline. Ask them to calculate one group fraction from the displayed cases.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Group evaluation | Reports only the average | Reports both group counts | Explains the aggregate’s weighting |
| Dataset repair | Adds samples indiscriminately | Targets an evidenced gap | Tests repair and acknowledges unequal residual errors |
Use two groups of five test cards and colour plus shape labels so results do not depend on colour perception.
Change collection order and explain how a small sample can make the apparent benefit unstable.
No demographic inference about students. Habitat groups are synthetic; fairness in human systems requires broader community judgement than these metrics.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
Reallocate examples to the less represented habitat. Predict its error direction before recomputing centroids.
A budget of 12 training examples starts with 11 reef examples and one lagoon example. The fixed test set has 14 reef cases and six lagoon cases. Each habitat predicts by distance to its two learned class means.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Repeat stratified collection across many deterministic seeds and plot group-error distributions, demonstrating that one favourable split is weak evidence.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
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 y4-balance-mac-pathway.zip -d y4-balance-mac cd y4-balance-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
Repeat stratified collection across many deterministic seeds and plot group-error distributions, demonstrating that one favourable split is weak evidence.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
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 y4-balance-dgx-pathway.zip -d y4-balance-dgx cd y4-balance-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 revised sampling allocation and a two-habitat audit.
Download the editable handout →