Machine Learning ML

Machine learning is a way of teaching software a rule by showing it thousands of solved examples, so it works out the pattern instead of being told it.

Also known as ML learning from data predictive models

Definition

Machine learning is how you teach software a rule without writing the rule down. You show it thousands of examples where the answer is already known, and it finds the pattern that connects them. A small child learns the word cat the same way. Nobody reads a two-year-old a definition; somebody points at cats until it clicks, and a model needs the same pointing with far more cats.

Card fraud is the classic case, because nobody can write the rules for it. You cannot list every sign that a payment is stolen, and the thieves would change method the day you did. Show a model a million payments already marked good or bad, and it finds the shape of the bad ones. That shape includes things no person would think to check, such as how fast the card number was typed.

That power comes with a catch that goes on no brochure, because a model learns everything the examples contain, including the mistakes. If your staff only ever checked payments from one country, the model learns that habit and repeats it like a law. Answers can only be as good as the history behind them, so cleaning old records is most of the project, which is the opposite of what people expect. Linkysoft has turned work down at exactly this point, because the records were not fit to learn from.

A model gives you a score, not a verdict. It says this payment is eighty-two out of a hundred suspicious, and a person still decides where the line sits. That line is a business choice about how many real customers you will annoy to catch one thief. So when Linkysoft builds an AI system, the line sits on a screen the manager controls, not buried in code.

The last surprise is the running cost, because a model goes stale as the world it learned from moves on. Demand in a shop changes the week a competitor opens next door, which is why Storek retrains its stock forecast instead of trusting one forever. So ask who retrains yours, how often, and what it costs a year. Linkysoft writes that number into the contract rather than leaving it to be found later. The same question fits any web application with a model inside, and the wider idea of artificial intelligence it belongs to.

Questions about Machine Learning

How is machine learning different from normal software?
Normal software follows rules a person typed in. A model is shown solved examples and works the rule out itself, which is why it can spot patterns nobody thought to look for.
How many examples do I need?
Thousands, not dozens, and each one must already carry the right answer. Marking up those old records is usually the biggest hidden cost of the whole project.
Can the model be wrong?
Always, and it should tell you how sure it is. You choose the line between accepting an answer and checking it by hand, and that line is a business decision, not a technical one.
Does a model need looking after once it is built?
Yes. It slowly stops fitting, because prices, habits and competitors change. Agree from the start who retrains it, how often, and what that costs each year.
Can a model inherit our old mistakes?
Yes, and it is the most common problem. It copies whatever the old records contain, habits and bias included, so someone should compare its decisions with real cases every few months.

Still not sure how this applies to your project?

Tell us what you are building and we will answer in plain language.