Why Several Floors?

Audience

ages 10–15

Duration

30 to 40 min

Adventure

1/8

Badge to earn

Representation Builder

Today's mission

Explain how several layers can build simple patterns into more composed ones.

Proof of success: explain "Why Several Floors?", complete "Cascading Transparencies" and name the limit of the analogy.

The story

To prepare a poster, Samira receives mosaics of houses. She starts by spotting edges, groups some edges into corners, then combines corners into windows. Her classmate Émile wants to search directly for "house" in each square and gets lost. A deep network can learn a similar progression: each layer transforms the representation it receives. The useful patterns, however, are not chosen by hand and must be checked on new examples.

What is technically true

To reason like a scientist, separate what you give the model, what it computes and what you measure. Here, your anchor is "Why Several Floors?".

• One layer computes a new representation; stacking layers lets you compose several transformations.
• The first layers of a vision model can respond to local patterns, but this hierarchy varies with the data and the training.
• Deeper does not automatically mean better: you need enough data, stable optimization and a comparison against a simpler baseline.

The vertical map

The flow shows a step-by-step composition, then reminds you that the last step is a check on a fresh example. Read the diagram from top to bottom: each arrow announces a transformation or a check, never a thought inside the machine.

A useful analogy — and where it breaks

The analogy: It's like building a figure from bricks: strokes, angles, shapes, then the whole object.

Where it breaks: In a network, the "bricks" are learned numbers, not concepts that necessarily carry the same names as ours. An analogy helps you get started; it never replaces the data, the calculations or a test.

Screen-free activity — Cascading Transparencies

Materials: four transparent sheets or tracing paper, a marker and a 6 × 6 mosaic
Suggested time: 10 to 20 minutes.

1. On the first sheet, circle only the light-dark changes.
2. On the second, connect the changes that form corners.
3. On the third, group corners into small windows.
4. Stack the sheets in order, then remove one to see what becomes impossible to build.

What you should notice: each floor reuses the previous result, and information lost early cannot be magically recreated.

Optional challenge

Invent a mosaic where the simple rule "four corners equal window" fails, and explain why more depth is not enough.

Safety, fairness and human choice

A complex model can look impressive while reproducing accidental patterns. Always compare it to a simple rule on a separate test.

Responsible question: If a baseline already reaches the goal, what specific benefit would justify a deeper network?

Mini-quiz with answers

1. What does a hidden layer produce?

   Answer: A new numerical representation computed from the previous layer.

2. Does more layers guarantee a better result?

   Answer: No. Depth adds capacity but also training and control difficulties.

3. Why test on an image never seen before?

   Answer: To check that the learned hierarchy generalizes instead of memorizing the mosaics.

Competency badge

Representation Builder — You earn this badge if you can explain the diagram without reading, show the result of the activity and kindly correct a wrong answer.