Teacher pilot · 45 minutes
Examples change a prediction
Help learners predict, label examples and test what changes. Review the language and activities for your group before teaching this pilot.
What learners will show
- Make and record a prediction before seeing the test result.
- Describe one human choice in labels, features or training coverage.
- Explain that this method stores labelled examples rather than fitting neural weights.
- Recognize that performance on a tiny reused test is not general reliability.
Before the session
- Check the lab route when available; otherwise use the complete paper alternative.
- Prepare the exact colour/roundness table and a 0-to-1 grid, with the test-label column covered until predictions are recorded.
- Keep test labels covered until learners commit a prediction; keep all test IDs out of training.
- Use synthetic botanical examples only; no personal data, uploads, external model accounts or open chat.
- Provide the source readings in a form the group can read or discuss.
Local 1-nearest-neighbour using Euclidean distance in the two-feature space. It stores learner-labelled training examples; it does not fit neural weights. Colour and roundness have equal weight. The pictures are illustrations of hand-set features. With the intended labels, the four starter examples give 4/8 correct decisions; adding the intended varied pair gives 8/8 only on this small reused set. Learner label changes can change these results.
Colour runs from 0 (green) to 1 (violet); roundness runs from 0 (pointed) to 1 (rounded).
The source readings
- What training on data actually means · 7 min. Contrast neural weight adjustment with the simpler stored-example algorithm.
- Where the data came from · 3 min. Discuss who chooses the examples and labels.
The 45-minute sequence
0–4 min · Predict before testing
Show a green pointed sample and a violet rounded sample. Learners record which feature they believe matters; accept more than one plausible rule.
4–14 min · Read and compare methods
Use the two linked readings (10 authored minutes). Explain that their neural-weight mechanism differs from the local nearest-example experiment.
14–17 min · Make the rule inspectable
Demonstrate Euclidean nearest-example selection on the colour/roundness grid. Conflicting equal-distance nearest labels mean Cannot decide.
17–29 min · Label, train and test
Label the four starter training examples, train, predict before running the eight fixed test examples, then record observed results. Test IDs must never enter training.
29–35 min · Change coverage, then retest
Add the varied training pair (violet leaf and green bloom), choose their labels, retrain and compare decisions on the same test set. Do not count repeated-test improvement as independent validation.
35–40 min · Discuss and challenge
In pairs, identify one wrong or unknown decision and one input/label choice that changed it. Discuss why extra copies of the same examples need not fix missing coverage.
40–45 min · Record an artifact and exit reflection
Complete the experiment card. State what is stored, a remaining limitation and one genuinely new test you would need.
A paper alternative
Plot the same synthetic labelled training examples and fixed test coordinates on a grid. For each test, locate the nearest training point by Euclidean distance (comparing squared distances also works), copy its label, and abstain if equally nearest examples conflict. Change only training examples/labels, then repeat the reference tests.
Read the coordinate table aloud and use squared-distance arithmetic instead of colour vision, dragging or spatial pointing.
| ID | Stage | Colour | Roundness | Intended label |
|---|---|---|---|---|
| train-01 | Starter | 0.08 | 0.18 | Leaf |
| train-02 | Starter | 0.22 | 0.25 | Leaf |
| train-03 | Starter | 0.78 | 0.78 | Bloom |
| train-04 | Starter | 0.92 | 0.84 | Bloom |
| train-05 | First varied pair | 0.78 | 0.19 | Leaf |
| train-06 | First varied pair | 0.1 | 0.73 | Bloom |
| train-07 | Optional later pair | 0.92 | 0.28 | Leaf |
| train-08 | Optional later pair | 0.24 | 0.85 | Bloom |
| train-09 | Optional later pair | 0.48 | 0.17 | Leaf |
| train-10 | Optional later pair | 0.52 | 0.81 | Bloom |
| ID | Colour | Roundness | Known label |
|---|---|---|---|
| test-01 | 0.12 | 0.2 | Leaf |
| test-02 | 0.18 | 0.16 | Leaf |
| test-03 | 0.82 | 0.81 | Bloom |
| test-04 | 0.88 | 0.76 | Bloom |
| test-05 | 0.84 | 0.23 | Leaf |
| test-06 | 0.89 | 0.21 | Leaf |
| test-07 | 0.14 | 0.8 | Bloom |
| test-08 | 0.18 | 0.77 | Bloom |
A separate tie example
Use just two training points: Leaf at (0, 0) and Bloom at (1, 1). Test (0.5, 0.5). Both points are equally near and have conflicting labels, so the result is “Cannot decide”.
A separate original paper tie example, not a promised preset in the digital lab.
Misconceptions to discuss
- This experiment trains a language model.
- It classifies two numeric features using stored labelled examples. Neural language models fit many weights and predict tokens; these are different methods.
- More examples always fix bias.
- Additional copies of an unrepresentative group can leave missing coverage and label errors untouched.
- 8/8 means the model is reliable everywhere.
- It describes only these eight examples, reused after inspection. New distributions and genuinely new tests can fail.
- A forced guess is better than an unknown result.
- Conflicting equal-distance labels leave insufficient evidence; Cannot decide is the specified honest result.
- The model discovers the true meaning of Leaf and Bloom.
- Humans choose features and labels; the geometric rule copies a label and does not understand botany.
Look for understanding
Teacher observation and feedback only; separate from official Addaly source checks, exams or credit.
- Records a prediction before seeing a test result and compares it with the outcome.
- Identifies a label or coverage change and traces how it affected a decision.
- Explains stored nearest examples, the two equally weighted features, explicit retraining after edits, and unknown conflicting ties.
- Limits the score to the reused toy test and proposes a genuinely new check.
My model experiment card: prediction; training choices; before/after test decisions; one error or unknown; limitation; next fresh test.
Exit questions
- What did the model store, and what did it do when equally close examples disagreed?
- Which choice changed a prediction?
- Why does a higher score on this reused set not prove the model will work elsewhere?
Original AI-authored teaching pilot. The complete lessons remain available. No grade-band or school-standard alignment is claimed. Chapter films are being developed; this plan uses a teacher/grid demonstration.