Crate #2: How Machines Learn

Traditional programming works like a recipe: you tell the computer exactly what to do, step by step. "If the email contains 'you won a prize,' move it to spam."

Machine Learning (ML) flips this upside down. Instead of writing rules, you show the computer thousands of examples and let it figure out the rules itself.

Imagine teaching a friend to identify dogs vs. cats. You wouldn't write a 500-page manual of rules ("if it has pointy ears AND whiskers AND..."). You'd just show them pictures: "Dog. Dog. Cat. Dog. Cat. Cat." After enough examples, they get it.

That's exactly how ML works. Show the computer 10,000 pictures of dogs and 10,000 pictures of cats, tell it which is which, and it learns to spot the difference. The "learning" is actually the computer adjusting millions of tiny numbers until it gets good at telling them apart.

SUPERVISED LEARNING — The teacher approach You give the computer both the questions AND the answers. "Here's a photo (question). It's a cat (answer)." The computer learns to match questions to answers. This is the most common type.

Real examples: spam filters, voice assistants understanding speech, medical diagnosis from X-rays.

UNSUPERVISED LEARNING — The explorer approach You give the computer data but NO answers. "Here are 50,000 shopping receipts. Find interesting patterns." The computer groups similar things together and discovers structure on its own.

Real examples: Netflix suggesting shows you might like, detecting unusual bank transactions.

REINFORCEMENT LEARNING — The gamer approach The computer learns by trial and error, like playing a video game. It tries something, gets a score (reward or punishment), and adjusts. Over thousands of attempts, it gets really good.

Real examples: AlphaGo learning to play Go, robots learning to walk, self-driving cars.

When people say they're "training" an AI, they mean they're feeding it data and letting it adjust its internal numbers (called parameters or weights) until it performs well.

Think of it like tuning a guitar. Each string (parameter) needs to be exactly right. If it's off even a tiny bit, the music sounds wrong. ML is like having a guitar with a million strings, and the computer tunes each one, a tiny bit at a time, until the whole thing sounds right.

The data you train on matters ENORMOUSLY. If you train a facial recognition system only on photos of light-skinned people, it'll be terrible at recognizing dark-skinned people. This isn't the AI being "racist" — it just never learned because no one showed it the right examples. This is why diverse, representative data is critical.

Training big AI models is expensive. Training GPT-4 reportedly cost over $100 million and used more electricity than a small city uses in a year. This is why companies don't retrain from scratch very often.