Comparing learning and generation

Your goal

Separate how a model learns from what it produces.

To be able to compare supervised and unsupervised learning and explain how generation can use learned patterns.

Before you start

A record has pages = 12 and category = Poetry. Which is the feature?

Choices: Pages = 12 / Poetry

Which is the known answer?

Choices: Poetry / 12 pages

Hide the category. Is the page count still available?

Choices: Yes / No

Words for this lesson

Supervised learning

Learning from examples with supplied target answers.

Unsupervised learning

Finding patterns without supplied target answers.

Generation

Creating an output such as new text.

The idea

Ask two different questions. How did the model learn? It may use known target answers, or find groups without those answers. What does it produce? It may predict a label, form a group or create text. A system can receive supervised training and still generate text. A name such as AI recommender does not tell us its training method.

A completed example

Practice: a model groups anonymous items by length without category names. That is unlabelled grouping. If people name the groups afterwards, the original grouping still did not use target labels.

Check your understanding

Unlabelled data can still contain ...

1. Features
2. Guaranteed correct categories
3. No information

A system trained on labelled answers then creating text can be ...

1. Both supervised in further training and generative in output
2. Only one of those
3. A fixed date rule

AI recommender alone tells us ...

1. Too little about training
2. Definitely unsupervised
3. Definitely a language model

Task 1 · Make a prediction

Could the model group without target labels

Options: Yes / No / Not sure

Write your prediction, or state that you skipped it.

Try it using supplied records

1. Choose With answers. Inspect the four labelled booklet examples and train the small classifier. Test it on the two held-out booklet lengths.
2. Choose Without answers. Inspect the same six booklet lengths with no target labels. Predict which lengths will group together, then press Find groups.
3. Choose Creates text. Read a short scripted continuation about a booklet and its training description.
4. For each view, answer How did it learn? and What does it output? These are separate controls.

Real small algorithms and a separate scripted illustration

Complete data and source records

Expected results · supplied

These are supplied reference results. They are separate from your own explanation. No code needs to run to read them.

1. With answers computes the mean page count of each supplied target label: Short=12, Long=42. Predict the nearer target centre for held-out U2=12 and U5=42, giving Short and Long, 2/2 correct. Equal distances give Needs review. Test items do not change centres. This is a real tiny nearest-centre classifier, using one feature.

2. Without answers implements one-dimensional k-means with k=2, fixed initial centres 10 and 44. Assign to nearest centre, recompute means, repeat until assignments stop changing. Final groups U1-U3 with mean 12, U4-U6 with mean 42. Equal distance goes to the lower numbered group. Display groups as 1 and 2, not meaningful reading categories.

3. Creates text is an authored continuation from The booklet tells: The booklet tells a story about a library. Caption: Scripted illustration, not a live generator. Training description states token prediction followed by labelled question-answer training. Card answers: 1 supervised / label; 2 unsupervised / groups; 3 includes supervised further training / generated text; 4 unknown / unknown; 5 programmed rule / overdue decision.

4. Provide two independent answer fields, never a single exclusive supervised-unsupervised-generative menu. A changed learning choice must not silently change the selected output choice.

Help with this step

Task 2 · Use the evidence

1. Read the five system cards one at a time. Select the supported learning description and output description for each.
2. Attach the sentence that supports your choice. Choose Not enough information when training is not described.
3. Recommend a method for either assigning existing library categories or finding new groupings. Explain why it fits that aim.

My explanation

Write in your school assignment. Keep these answers in your own words.

Card 1

• Learning choice: Supervised / Unsupervised / Includes supervised further training / Programmed rule / Not enough information
• Output choice: Label / Groups / Created text / Rule decision / Not enough information
• Evidence
• Missing information

Card 2

• Learning choice: Supervised / Unsupervised / Includes supervised further training / Programmed rule / Not enough information
• Output choice: Label / Groups / Created text / Rule decision / Not enough information
• Evidence
• Missing information

Card 3

• Learning choice: Supervised / Unsupervised / Includes supervised further training / Programmed rule / Not enough information
• Output choice: Label / Groups / Created text / Rule decision / Not enough information
• Evidence
• Missing information

Card 4

• Learning choice: Supervised / Unsupervised / Includes supervised further training / Programmed rule / Not enough information
• Output choice: Label / Groups / Created text / Rule decision / Not enough information
• Evidence
• Missing information

Card 5

• Learning choice: Supervised / Unsupervised / Includes supervised further training / Programmed rule / Not enough information
• Output choice: Label / Groups / Created text / Rule decision / Not enough information
• Evidence
• Missing information

Library decision

• Library aim: Existing categories / New groupings
• My recommendation and reason

Explain a fresh example

A model groups books without category labels. Another learns known categories and assigns them. Compare the learning methods and say whether creating text would reveal the full training method.

My confidence

I need another explanation / I can explain with help / I can explain with my own example

My next step

Write what you will do next.

Optional depth

Why does giving a cluster a human name afterwards not change how it was formed? Could a grouping be unhelpful or unfair?

What this activity can show

Page-count groups are real computed clusters in a tiny toy dataset, not a claim about meaningful book categories. Supervised and unsupervised are learning approaches; architecture means model organisation.

This is a student learning record. Submission is completed through the school assignment system.