# 🟢 Session 2 — Quiz: How does AI learn?

**Applied AI — Yann Isola · Beginner Level**
10 questions · Only 1 correct answer per question · Answer key commented at the end of the document.
*Facilitation tip: in groups, “show of hands” mode or teams — it’s more fun.*

---

**1. In supervised learning, what do we give the AI to learn?**

- A) Rules written by hand by programmers
- B) Examples accompanied by the correct answers (labels)
- C) Examples without any indication
- D) Rewards and punishments

**2. “Sorting a bag of candy by color without anyone telling you the categories” is the analogy of…**

- A) Supervised learning
- B) Reinforcement learning
- C) Unsupervised learning
- D) Overfitting

**3. An AI plays millions of rounds of a game, gaining points when it succeeds, losing points when it fails, and improving. This is…**

- A) Supervised learning
- B) Unsupervised learning
- C) Reinforcement learning
- D) Cheating

**4. Why do we say that data is the “fuel” of AI?**

- A) Because the servers consume gasoline
- B) Because without data, an AI cannot learn anything: all its knowledge comes from its examples
- C) Because data costs the same as oil
- D) Because the AI burns the data after reading it

**5. Your AI should recognize dogs, but you ONLY trained it with photos of Labradors. What is likely to happen when faced with a Chihuahua?**

- A) Nothing, a dog is a dog
- B) She will recognize him even better, because he is smaller
- C) She risks not recognizing it: her data lacked diversity
- D) She will refuse to answer out of caution

**6. What is the difference between training and using (inference) an AI?**

- A) None, it's the same process
- B) Training is long and expensive (the AI learns); use is quick (the AI applies what it knows)
- C) Use is longer than training
- D) Training is done after use

**7. An artificial neuron is above all…**

- A) An exact copy of a neuron in the human brain
- B) A living electronic chip
- C) A small calculator which receives signals, gives them more or less importance (weight), and sends a signal in turn
- D) A program that stores memories

**8. In a neural network, “learning” essentially means…**

- A) Add new neurons to each correct answer
- B) Gradually adjust the weights (the importance of the connections) to reduce errors
- C) Copy the data into a large memory
- D) Increase processor speed

**9. An AI obtains 99.9% success on its training examples, but only 60% on new examples. Diagnosis?**

- A) It's an excellent AI: 99.9%, it's almost perfect!
- B) It’s overfitting: she “learned by heart” instead of understanding
- C) It's normal, an AI is always bad on new things
- D) New examples are necessarily false

**10. Why do we keep test data that the AI NEVER sees during training?**

- A) To save memory
- B) Because it is a legal obligation
- C) To check that she knows how to generalize on something never seen before, like a test with new exercises
- D) Because these data are of lower quality---
---

# 🔑 Commented answer key

**1. B** — Supervised = examples + correct answers. It is the teacher who corrects the copies: the AI ​​compares its answer to the correct one and corrects itself. (A describes classical programming, C unsupervised, D reinforcement.)

**2. C** — Unsupervised: no labels, the AI ​​discovers the groups on its own — like you sorting candies by color without instructions.

**3. C** — Reinforcement: trial → reward/punishment → improvement. Just like learning a video game. This is how AlphaGo became unbeatable at Go.

**4. B** — All the “knowledge” of an AI comes from its training data. No data → no AI. And poor quality data → poor quality AI (“garbage in, garbage out”).

**5. C** — Lack of **diversity** of data prevents generalization. The AI ​​only learned “dog = labrador silhouette”. Lesson: more data, yes, but above all **varied** data.

**6. B** — Training: weeks/months, very expensive, once. Use: less than a second, billions of times a day. Analogy: 10 years of medical studies vs 20 minutes of consultation.

**7. C** — An artificial neuron is a mini-computer: input signals × weights → decision → output signal. The analogy with the brain is an **inspiration**, not a copy (wrong answer A!).

**8. B** — The whole secret of learning lies there: millions of micro-adjustments of the weights to reduce the error. Our viewer cursors, but adjusted automatically — and by the billions.

**9. B** — The huge gap between training score (99.9%) and test score (60%) is the signature of **overfitting**: the student who has memorized the 50 exercises in the book and fails the test. The training score alone proves NOTHING.

**10. C** — Test data is the final exam with never-before-seen exercises. If the AI ​​had seen them during training, it could “recite” them and the exam would be rigged. He is the only justice of the peace of true jurisdiction.

---

## Suggested scale

| Score | Verdict |
|---|---|
| 9-10 | 🏆 You could almost train your own AI! |
| 7-8 | 💪 Very solid — key concepts are in place. |
| 5-6 | 🙂 The main thing is there; review overfitting and training vs. use. |
| <5 | 🌱 Take another look at the interactive page — and the quiz will seem easy. |

---
*Applied AI — Yann Isola · Beginner Level 🟢 · Session 2 — Quiz*