Machine Learning in Plain English

The doctor who saved thousands of lives — without examining a single patient

In 2017, a team at Stanford built a system that performed comparably to a panel of board-certified dermatologists on a held-out test set (Esteva et al., Nature, 2017) — results reflect a specific benchmark, not all dermatological conditions. Nobody programmed rules like "if the mole is brown and asymmetric, it's melanoma." Instead, they showed the system 130,000 photos of skin conditions — each labelled by a doctor as "cancer" or "not cancer" — and the system figured out the rules on its own.

That's machine learning. Not programmed rules. Not magic. A system that improves by looking at examples.

You already do this every day. When you were a toddler, nobody gave you a rulebook for identifying dogs. You saw hundreds of dogs — big ones, small ones, fluffy ones, scraggly ones — and your brain built its own internal "dog detector." That's exactly what machine learning does, just with math instead of neurons.

What "learning" actually means for a machine

When humans say a machine "learns," it sounds like the computer is sitting in a classroom taking notes. It's not. Here's what's really happening:

Learning = adjusting based on feedback.

Imagine you're throwing darts blindfolded. After each throw, someone tells you "too far left" or "a bit high." You adjust your next throw. After 100 throws, you're landing near the bullseye — not because someone gave you a formula for throwing, but because you adjusted based on feedback, again and again.

That's machine learning in four words: predict, check, adjust, repeat.

The machine starts with random guesses. Each time it gets feedback ("wrong!" or "close!"), it tweaks its internal settings — tiny numerical dials called parameters — to get a little closer to the right answer. After thousands or millions of rounds, those dials are tuned well enough that the system gives good answers on examples it's never seen before.

Supervised learning: the flashcard method

Remember studying with flashcards? One side has the question, the other has the answer. You look at the question, guess the answer, flip the card to check, and adjust your thinking for next time.

Supervised learning works exactly like flashcards. You give the machine a set of examples where you already know the right answers (the "labels"), and the machine uses that feedback to learn patterns.

Real examples of supervised learning:

• Email spam filter: Thousands of emails, each labelled "spam" or "not spam." The model learns which words, patterns, and senders predict spam.
• House price prediction: Thousands of houses with features (square footage, neighbourhood, bedrooms) and known sale prices. The model learns which features drive the price up or down.
• Medical diagnosis: Thousands of X-rays labelled "fracture" or "no fracture" by radiologists. The model learns what fracture patterns look like.

The key requirement: you need labelled data. Someone (usually a human) has to provide the correct answer for every training example. That's often the hardest and most expensive part of the whole process.

Unsupervised learning: sorting a messy drawer

You open your junk drawer. It's chaos — batteries, pens, receipts, rubber bands, paper clips, keys. Nobody told you the categories. Nobody labelled anything. But within minutes, you've sorted everything into natural groups: "writing stuff," "fasteners," "paper things," "electronics."

That's unsupervised learning. The machine gets data with no labels — no right answers — and finds patterns, groups, and structure on its own.

Real examples of unsupervised learning:

• Customer segmentation: A retailer feeds purchase data into a model. Nobody says "this is a budget shopper" or "this is a luxury buyer." The model discovers natural clusters: "people who buy organic food and yoga mats" vs. "people who buy bulk items and diapers."
• Anomaly detection: A bank feeds millions of normal transactions into a model. The model learns what "normal" looks like. Anything that doesn't fit the pattern gets flagged as potentially fraudulent — without anyone ever labelling individual transactions as "fraud."
• Topic discovery: You dump 100,000 news articles into a model. Nobody labels the topics. The model discovers that certain words cluster together: "election," "candidate," "polls" form one cluster; "touchdown," "quarterback," "playoff" form another.

Reinforcement learning: learning through trial and error

You've played a video game you'd never seen before. Nobody explained the rules. You pressed buttons, watched what happened, and gradually learned: "jumping on that thing gave me points, falling in the lava ended the level." After an hour, you were good at the game — not because someone taught you, but because you tried things, got feedback, and adjusted.

That's reinforcement learning (RL). The model — called an agent — takes actions in an environment, receives a reward signal (positive or negative), and learns to maximise reward over time.

Real examples of reinforcement learning:

• AlphaGo: DeepMind's system played millions of Go games against itself, rewarded for wins and penalised for losses. It discovered strategies no human player had considered in 2,500 years of the game (Silver et al., Nature, 2016).
• Robot arms: Warehouse robots learn to pick objects by trying different grip angles — rewarded when the object is picked up cleanly, penalised when it drops.
• Ad bidding: Ad platforms learn which bid to place for each impression by observing which bids led to conversions, updating in real time.

The key requirement is a reward signal — a clear definition of what "good" and "bad" outcomes look like. RL is powerful but harder to set up than the other two, which is why it's reserved for specific problems: games, robotics, and sequential decision-making.

The training loop: how a machine actually learns

Let's zoom in on that predict-check-adjust-repeat cycle with a concrete example. Say you're building a model to predict house prices.

After seeing thousands of houses, the model's internal dials are tuned so well that it can predict reasonable prices for houses it's never seen. That's the whole trick.

Key terms in the training loop:

Overfitting: the A+ student who fails the real exam

Here's a scenario you've definitely seen: a student memorises every answer in the practice test. Word for word. They score 100% on practice tests. Then they take the real exam — different questions, same concepts — and they bomb it. They memorised the answers instead of understanding the concepts.

That's overfitting. The model learned the training data too well. It memorised the specific examples instead of learning general patterns.

How do you prevent overfitting? The same way you'd advise that student:

1. Use a test set. Hold back 20% of your data. Train on 80%, test on 20%. If the model does great on training data but poorly on test data, it's overfitting.
2. Get more data. It's harder to memorise 1,000,000 examples than 100.
3. Simplify the model. A model with too many parameters relative to the data can memorise everything. Reduce complexity.
4. Regularisation. Technical term for adding a penalty when the model's parameters get too extreme — like telling the student "you can study the practice tests, but you can't bring notes into the exam."

Putting it all together: the machine learning pipeline

Here's what the full process looks like from start to finish:

Most people think machine learning is step 5 — the training. In reality, steps 1-3 (getting good data) take 80% of the time. Data scientists have a saying: "garbage in, garbage out." The fanciest algorithm in the world can't learn from bad data.

Back to the Stanford skin cancer system

The Stanford team never wrote a single rule that said "asymmetric border means melanoma." They showed the system 130,000 labelled photos and let it find the patterns itself — the same predict-check-adjust-repeat loop you worked through in the training exercise. What emerged were subtle correlations across millions of pixels that no dermatologist could consciously articulate, let alone program by hand. The system wasn't smarter than a doctor; it had just seen more examples, and it tuned billions of parameters until the patterns clicked. That's the core insight of this module: machine learning doesn't replace human expertise — it scales the signal buried inside labelled human judgments. The 130,000 photos were the dermatologists, distilled into weights.

Key takeaways

• Machine learning = adjusting based on feedback. A model makes predictions, checks them against known answers, adjusts its internal settings, and repeats — thousands of times.
• Supervised learning = flashcards. You need labelled data (question + answer pairs). Great for prediction tasks with clear right/wrong answers.
• Unsupervised learning = sorting a drawer. No labels needed. The model finds patterns and groups on its own. Great for exploration and anomaly detection.
• The training loop is predict, check, adjust, repeat. Everything else — epochs, batches, loss, learning rate — is just details of that loop.
• Overfitting = memorising instead of understanding. Always check the gap between training and test accuracy. If it's big, the model memorised the data.
• Data quality matters more than algorithm choice. 80% of the work is getting clean, representative data. The model is only as good as what you feed it.