title: "Session 2 — How does AI learn?"
author: "Applied AI · Yann Isola"
theme: "Applied AI (palette: #1A2230 / #0F7A6C / #B4612A / #E9F6F3 / #F4F7F6)"
level: " Beginner — from 12 years old, no prerequisites"

Slide 1 — Title

How does AI learn?

Applied AI · Session 2 · Beginner Level

Yann Isola

Today, we open the hood.

Slide 2 — Last time…

• Session 1: what is AI?
• One-sentence answer: a program that learns from examples, instead of just following hand-written rules.
• Question of the day: okay, but HOW does she learn?

Moderator note: have the room respond before displaying the definition.

Slide 3 — The game of teacher and student

• A volunteer becomes the “AI”
• We show him objects with the correct category: “Ball → A. Book → B…”
• Then a new object, without the answer: “Plate →?”

If he finds: he has generalized. Nobody gave him the rule!

Note: REALLY play the game (10 min). This is the best hook of the session.

Slide 4 — There are 3 main ways to learn

1. Supervised — the teacher who corrects papers
2. Unsupervised — sort candy without deposit
3. Reinforcement — learn a video game

Three ways, three analogies. At the end of the session, you will know them by heart... well no: you will have understood them.

Slide 5 — Supervised learning

We give: examples + the correct answers (the "labels")

• The AI suggests an answer → we compare it to the correct one → it corrects itself
• Like a student: exercise → teacher's correction → adjustment
• Repeated million times

Examples: "cat"/"dog" photos · "spam"/"not spam" emails · apartment prices

Slide 6 — The supervisee’s vocabulary (3 words, no more)

Learn = reduce the error, example after example.

Slide 7 — Unsupervised learning

We give: examples… and that’s it. Zero response.

• You are given a bag of sweets: “Sort.”
• You group by color, shape, size… you invent the categories
• AI does the same: it discovers hidden groups and similarities

Examples: customer families · automatic playlists · detection of “bizarre” behavior (fraud)

Slide 8 — Reinforcement learning

We give: a playing field + points

• Good deed → reward · Bad deed → punishment
• Test → result → adjustment → we start again (millions of times)
• Like your first game of a video game: you fall, you start again, you progress

Examples: AlphaGo (go world champion beaten in 2016) · robots that learn to walk · autonomous cars in simulator

Slide 9 — Recap: the 3 modes in one table

Fashion	We give him…	She learns…	Analogy
Supervised	Examples + answers	To predict	The teacher who corrects
Unsupervised	Examples only	To group	Sort candy
Strengthening	Game + points	A strategy	The video game

And ChatGPT? A mix of fashions — exact recipes evolve quickly.

Slide 10 — Data: the fuel of AI

• No data → no AI. All his “knowledge” comes from his examples.
• More data = often better learning (1000 exercises > 10 exercises)
• BUT… it’s not just a question of quantity →

Slide 11 — …not just any data!

3 traps:

1. False data → false AI (teacher who corrects with a false answer key)
2. Data not varied → blinkered AI (only labradors → unrecognized chihuahua)
3. Biased data → Unfair AI (recruitment AI disadvantages women, trained on unbalanced histories — real case)

“Garbage in, garbage out”: waste in, waste out.

Slide 12 — Training vs. use

	Workout	Usage (inference)
What ?	AI learns	The AI ​​responds
Duration	Weeks / months	< 1 second
Cost	Very expensive (millions)	Minimal per query
Frequency	Once	Billions of times/day

Slide 13 — The doctor analogy

• Train a doctor: 10 years, very expensive → training
• Consult a doctor: 20 minutes → inference
• We do not re-train the doctor for each patient!

When you talk to ChatGPT, it doesn't learn: it uses what it already knows. His “brain” has been frozen since training.

Slide 14 — The artificial neuron: a small calculator

Inspired by the brain — not a copy (a real neuron is infinitely more complex)

An artificial neuron does 3 things:

1. Receives signals (numbers)
2. The weights: some count a lot, others a little → the weights
3. Decides: if the total exceeds a threshold → it “lights up” and transmits

Slide 15 — The bouncer analogy

The nightclub bouncer:

• receives several information (age, outfit, attitude, etc.)
• gives more or less importance to each → the weights
• decides: enters / does not enter → the exit signal

An artificial neuron = a mathematical bouncer. A network = thousands of bouncers spreading the word.

Slide 16 — The network: layers of computers

• Input layer: raw data (e.g.: pixels of an image)
• Hidden layers: detect increasingly complex things
  (features → shapes → ears, whiskers → “cat!”)
• Output layer: the answer

Large AI: the equivalent of hundreds of billions of weights — figures expire each year.

Slide 17 — DEMO: the network visualizer

Interactive page, “Neural network” tab

• Watch signals cross layers
• Move the weight sliders → the output changes!
• “Auto training” button: the weights are adjusted by themselves

Learn = adjust these sliders automatically. Except there are billions of them.

Slide 18 — The secret of learning, in one sentence

Learn = adjust weights to reduce error. - The AI is wrong → it slightly shifts each weight in the right direction

• Millions of micro-adjustments later: she got strong
• That's EVERYTHING. No magic. Settings.

Slide 19 — Overfitting: the student who learns by heart

Two students, same book of 50 corrected exercises:

• Student A: understands the methods → passes the test (new exercises)
• Student B: memorizes the 50 answers → completely fails

Student B is overfitting (overlearning): brilliant on the known, poor on the new. But an AI is ONLY used for new things!

Slide 20 — How to detect rote learning?

• We keep exercises secret: the test data
• The AI NEVER sees them during training
• Training score 99.9% but test score 60%? → by heart detected!

DEMO — “Overfitting” tab: Watch the two curves diverge — the exact moment the AI starts reciting.

Slide 21 — Your turn: train your own AI!

Option A — Teachable Machine ( free Google tool, scalable interface)

1. Two classes: open hand / fist
2. ~30 photos per class varying the angles
3. “Train” (30 s) → test live!
4. The twist: test with the other hand... does it fail? For what ?

Option B (offline) — “Supervised simulator” tab of the interactive page.

Slide 22 — Workshop debriefing

• What made your model good? bad ?
• Expected answers: number of examples · diversity · quality
• You have just experienced in 10 minutes what AI engineers experience for months

The entire theory of the session has just been verified before your eyes.

Slide 23 — The 6 ideas to remember

1. AI learns from examples
2. 3 modes: teacher · candy · video game
3. Data = fuel: quantity, quality, diversity
4. Train = long and expensive · Use = fast
5. Network = layers of calculators · learn = set weights
6. Overfitting = by heart ≠ understand · the test on something never seen before is the judge

Slide 24 — Quiz!

10 questions, in teams, by show of hands

The one who understood everything without learning by heart... generalized perfectly.

Moderator note: quiz/quiz.md — commented answer key included.

Slide 25 — Next time…

Generative AI

• How can a machine write, draw, compose?
• ChatGPT, image generators… what’s really going on under the hood?

Session 3 — not to be missed.

Applied AI · Yann Isola · Beginner