- Machines learn by a simple loop: guess, check the answer, adjust — repeated at huge scale.
- Three styles: supervised (labelled), unsupervised (find structure), reinforcement (trial & reward).
- Clean, representative data matters as much as quantity — garbage in, garbage out.
01Learning from examples
Show the machine an example. Let it guess. Tell it the right answer. It nudges its settings to be a little less wrong — then does it again, millions of times.
Try it: tell this learner what each fruit is, and watch its accuracy climb.
What just happened
No single guess matters — the magic is the repetition. After enough rounds, those tiny nudges add up to settings that get the answer right most of the time, without anyone writing the rule. That's training.
The first time the learner sees a fruit it's just guessing. Once you've labelled it, it remembers — and accuracy climbs. That's supervised learning in miniature: learning from examples with the right answer attached.
02The three ways machines learn
Supervised
Learns from examples that already have the right answer labelled.
Unsupervised
Gets data with no labels and finds the hidden structure or groups by itself.
Reinforcement
Learns by trial and error, earning rewards for good moves and penalties for bad ones.
How real AI blends them
Modern AI often mixes these. A chatbot is first trained on mountains of text (a self-supervised twist on the first type), then polished with human feedback (a cousin of the third). You'll meet that again in Step 04.
03Why more — and cleaner — data means smarter AI
Varied, accurate, and representative of the real world the model will face. Leads to patterns that hold up on new examples.
Narrow, outdated, or skewed toward one group. The model faithfully learns those flaws — and repeats them confidently.
Quantity vs quality — and "garbage in, garbage out"
Quantity helps because more examples show more variety and average out flukes. See ten dogs and you might think all dogs are brown; see ten million and you learn what "dog" really covers.
Quality matters just as much. Feed a model skewed or messy examples and it learns skewed, messy patterns.
Garbage in, garbage out. A model is only as fair and accurate as the examples it learned from. Clean, representative data beats simply more data.
04Where learning goes wrong
The two failure modes to know
Memorising instead of understanding
If a model sees the same examples too many times, it can overfit — memorising the answers instead of the general idea. It aces the practice questions and flunks the real exam.
Overfitting is the student who memorises last year's exam word-for-word, then panics when this year's questions are phrased differently. Real learning generalises.
Learning our blind spots
Because models learn from human-made data, they also absorb human bias. If the examples under-represent some group, the model will too — and won't know it. We dig into fairness in Step 06.
- Machines learn by a loop: guess, check, adjust — repeated at huge scale. That's training.
- Three styles: supervised (labelled), unsupervised (find structure), reinforcement (trial and reward).
- More data helps, but clean, representative data matters just as much — garbage in, garbage out.
- Learning can misfire: overfitting (memorising) and bias (inheriting flaws in the data).