1
Supervised — the Teacher with an answer key
Every training example comes with a correct label attached. The model learns to match inputs to known outputs, the same way a teacher grades against an answer key.
Example: a spam filter trained on 50,000 emails each labeled "spam" or "not spam" — every example has the right answer attached.
2
Unsupervised — the Explorer with no map
No labels exist at all. The model is dropped into raw data and must discover hidden structure entirely on its own, like an explorer in an unmapped jungle.
Example: clustering millions of customers into groups based on purchase behavior, with no pre-existing "correct" grouping given to the model.
3
Reinforcement — the Dog trained with treats
An agent takes an action, receives a reward or penalty, and gradually learns which behaviors work best through trial and error, like a dog learning tricks for treats.
Example: an AI learning to play a video game by getting points for good moves and losing points for bad ones, adjusting its strategy over thousands of rounds.
Applied Walkthrough
1
A question describes a model given 100,000 photos, each already tagged "cat" or "dog" by a human.
2
Ask: are correct answers already attached to every example? Yes — every photo has a known label.
3
That's the Teacher scene: supervised learning, because the model learns to map input (photo) to a known correct output (label).
4
Contrast: if those same 100,000 photos had NO tags at all, and the model just grouped similar-looking photos together, that would instead be the Explorer scene — unsupervised learning.
Exam Application
The number one thing exams test here is whether you can identify which paradigm a described scenario belongs to. Look for the presence or absence of labels first — that instantly separates supervised (labels present) from unsupervised (no labels). Then look for reward signals and sequential decision-making — that signals reinforcement learning, which is different from both.
⚠ Common Trap
Students often confuse unsupervised learning with reinforcement learning because neither has traditional "labels." The distinguishing detail: reinforcement learning has a reward/penalty signal and a sequence of actions over time (a game, a robot navigating a room). Unsupervised learning has neither — it's a single static dataset with hidden structure to discover, no actions or rewards involved.
✓ Quick Self-Check
1. Which learning paradigm uses labeled training data?
Supervised learning — every example has a known correct answer attached, like a teacher's answer key.
Tap to reveal / hide
2. Which paradigm has no labels at all and searches for hidden structure?
Unsupervised learning — the Explorer scene, discovering patterns like clustering with no correct answer given.
Tap to reveal / hide
3. Which paradigm involves an agent taking actions and receiving rewards or penalties over time?
Reinforcement learning — the Reward Dog scene, learning through trial and error.
Tap to reveal / hide
4. A recommendation system groups users into 5 unlabeled behavior segments. What paradigm is this?
Unsupervised learning, since there are no pre-existing correct labels — the model is discovering the groupings itself.
Tap to reveal / hide
5. An AI plays thousands of rounds of a game, gaining points for wins and losing points for losses, gradually improving its strategy. What paradigm is this?
Reinforcement learning — the reward/penalty signal over a sequence of actions is the defining feature.
Tap to reveal / hide