Machine learning explained: how software learns from examples

Machine learning is what happens when you show a program thousands of examples that already have the right answer attached, and let it work out the relationship between them. Nobody writes the rule. The program finds a mathematical function that maps inputs to outputs with as few mistakes as possible across the examples, and then you apply that function to cases nobody has seen. Everything else that gets filed under artificial intelligence, from a credit decision to a chatbot, is a variation on this one idea at different scales.

One example, all the way through: predicting a house price

Say you want software that estimates what a house will sell for. You obtain the records of 50,000 completed sales in your region over the last five years. Each record has the sale price and a set of details about the property.

The details you feed in are called features. The known answer is called the label.

Here is the whole procedure.

1. Split the 50,000 records. Around 80 percent become the training set. The remaining 20 percent are locked away as the test set and the model never sees them.
2. Start the model with essentially random settings. Show it a house, let it guess, compare the guess with the real price. The gap is the error.
3. Adjust the settings slightly in the direction that would have made the error smaller. This is the only thing "learning" means.
4. Repeat across the training set many times over. The average error falls, quickly at first, then slowly.
5. Now unlock the test set and run the model on those 10,000 houses it has never seen. That score, and only that score, tells you whether it learned anything transferable.
6. Deploy it. A new listing arrives, you supply the features, the model returns a price estimate.

The same skeleton applies whether the label is a price, a category (spam or not spam) or the next word in a sentence. Change the data and the model size and you get how a language model is trained.

The three kinds of learning you will hear named

Supervised learning is the house price case: every example comes with the correct answer attached. It covers the large majority of machine learning in commercial use.

Unsupervised learning has no labels. You hand the system 50,000 customer records and ask it to group them by similarity, and it returns clusters you then have to interpret yourself. Useful for discovering structure, harder to evaluate, because there is no right answer to check against.

Reinforcement learning has no fixed examples at all. An agent acts, receives a reward or a penalty, and adjusts to earn more reward over time. It is how game playing systems are trained, and a version of it is used to tune chatbots toward answers people rate as helpful.

Overfitting: memorizing instead of learning

A model with enough capacity can drive its training error to almost zero by memorizing the individual houses rather than learning what makes a house expensive. It notices that the property on Oak Street sold for an odd amount and encodes that quirk. On the training set it looks superb. On the test set it falls apart, because none of the memorized quirks apply.

This is overfitting, and it is the central failure of the field. The signal is a large gap between training accuracy and test accuracy. The fixes are all forms of making memorization harder: more data, a smaller model, deliberately holding back part of the network during training, or simply stopping the training early while the test score is still improving.

The opposite problem, underfitting, is a model too simple to capture the pattern at all. It is easier to spot, because it performs poorly everywhere.

Distribution shift: why models quietly rot

Even a well built model assumes tomorrow resembles the data it learned from. That assumption expires.

A house price model trained on 2019 to 2024 sales has interest rates, construction costs and buyer behavior from that window baked into its weights. When rates move sharply, its estimates drift high or low across the board, and it will never mention this. A fraud model learns the patterns of the scams that existed while it was trained, and fraudsters adapt within weeks. A demand forecast built on pre pandemic shopping was useless by mid 2020.

This is distribution shift, and it is why deployed models need monitoring rather than just launching. The practical defense is to keep comparing predictions against outcomes that arrive later, and to retrain on a schedule rather than waiting for someone to complain.

The shortcut problem

Models optimize for the score you give them, not for the thing you meant. If there is an easier signal in the data that happens to correlate with the label, they will take it.

The best documented version of this comes from medical imaging. Skin lesion classifiers trained on clinical photographs turned out to be partly detecting rulers, because clinicians place a measuring ruler beside lesions they already suspect are malignant. The model scored well in testing, because the ruler correlation held in the test set too. It had learned something real about the photographs and nothing about skin.

The same pattern shows up everywhere: a model that identifies animals by the background rather than the animal, a hiring model that keys on the name of a university, a quality inspection model that reads the timestamp burned into the image. Each looks fine until the shortcut disappears. This is also one route by which bias enters AI systems, since a shortcut is often a proxy for something you were not allowed to use directly.

What to check first

When someone tells you a system is accurate, ask what it was trained on, when that data was collected, and whether the test set came from a genuinely different pool than the training set. Those three answers predict real world performance better than any headline number. If the answer to the first question is vague, where AI training data comes from explains why that vagueness is so common, and what it usually hides.