What is machine learning?
By Yassine Seghrouchni, founderUpdated
In short
Machine learning is a form of AI in which software learns patterns from examples, instead of following rules a programmer has written down. You give the system, say, a large set of resolved support tickets with their category, and it works out which characteristics go with which category. It then sorts new tickets. Language models such as ChatGPT are built with machine learning too.
Also known as: ML
How it works
A model learns in two phases. During training, it gets examples where the outcome is known and adjusts itself until it predicts those outcomes as well as possible. After that, you use it on new cases: the model gives a prediction, often with a level of confidence.
Most business applications do one of these four things:
- Classifying: which category does this email, ticket or invoice belong to?
- Predicting: how many orders will come in next week, which customer is likely to cancel?
- Spotting anomalies: which transaction or reading doesn’t fit the pattern?
- Grouping: which customers resemble each other, without defining the groups in advance?
Machine learning, AI and generative AI
AI is the umbrella term for software that does tasks that would normally need human judgement. Machine learning is how most of today’s AI comes about. Deep learning is a form of machine learning that uses large neural networks. That’s also how large language models are built.
The difference in use: classic machine learning predicts a number or a category. Generative AI creates new text, images or code. For many text tasks, such as sorting emails or pulling data out of an invoice, you no longer need to train a model of your own. An existing language model can often do it with good instructions.
Does a model keep learning by itself?
Usually not, and that’s a common misconception. A model in use doesn’t change with every new case it sees. It only learns more when someone retrains it on new data. That’s a good thing: it means you know which version produced which outcome.
It does mean a model goes out of date. If your products, your customer base or your process change, its predictions get worse. So keep measuring quality after go-live, and plan when to retrain.
Why it matters in practice
Training your own model pays off when you have a recurring prediction question and enough historical data in which the outcome is recorded. If you want to predict which quotes will be signed, you need past quotes where you know whether they were signed, with their characteristics. If that outcome isn’t recorded anywhere in your systems, there’s nothing to learn from.
A model also learns the biases in your past. If you mostly hired candidates of one profile in the past, a model learns to prefer that profile. With AI that helps make decisions about people, that weighs extra heavily: that’s where the requirements for high-risk AI apply. Which processes do lend themselves well is covered in which business processes are suitable for AI.
Frequently asked questions
What is the difference between AI and machine learning?
AI is the umbrella term for software that does tasks that would normally need human judgement. Machine learning is the method behind most of today’s AI: the system learns from examples instead of fixed rules. So all machine learning is AI, but not all AI is machine learning. Systems built on hand-written rules count as AI too.
How much data do you need for machine learning?
That depends on the question and on how much the cases differ from one another. More important than the amount is that the outcome has been recorded: without examples where you know how things turned out, there’s nothing to learn. For text, such as emails or documents, an existing language model is often a quicker start than training your own.
