What is machine learning?

How machine learning works

Machine learning works by showing a model many examples and adjusting it until its outputs match the examples as closely as possible.

The model starts with settings that mean nothing. During training it makes a guess for each example, compares the guess with the known answer and nudges its settings to be less wrong next time. After many rounds the settings capture a pattern, and the trained model can be used on new cases it never saw. Using a trained model in this way is called inference.

Supervised vs unsupervised learning

Supervised learning learns from examples labelled with the right answer, unsupervised learning finds structure in examples with no labels, and reinforcement learning, the third main type, learns from rewards.

• Supervised learning: each example comes with the answer, such as a deal marked won or lost, and the model learns to predict it.
• Unsupervised learning: the examples have no answers, and the model groups or summarises them, for instance clustering accounts that look alike.
• Reinforcement learning: the model acts, receives a reward or a penalty and learns which actions pay off over time, as game-playing systems do.

Machine learning vs AI

Artificial intelligence is the broad goal of making machines do tasks that need intelligence, and machine learning is one way to reach it. Symbolic AI is the other classic route: rules and logic written by people.

Deep learning sits inside machine learning as the family of methods that use many-layered neural networks, and today's large language models sit inside deep learning. So every deep learning system is machine learning, but not every AI system learns from data.

Machine learning examples in sales

The classic machine learning example in sales is a lead-scoring model that learns from past deals which traits go with a win. Give it an illustrative history of won and lost deals with firmographics such as industry, headcount and region, and it learns a weight for each trait.

A manager's scoring sheet does the same job by hand: twenty points for the right industry, ten for the right size. The model can spot combinations no one thought to write down; the sheet can be read, argued over and changed in a meeting. Other examples are finding accounts that look like your best customers, sorting inbound replies by intent and flagging contact records that are likely out of date.

What machine learning cannot tell you

Machine learning cannot tell you why a pattern holds, and it cannot see beyond the data it learned from.

If past deals came mostly from one region, the model will favour that region whether or not it matters. If the market changes, the old pattern keeps being applied until the model is retrained. And because the pattern is spread across many numbers, a model can say a lead looks promising without a reason a seller can act on. Treat its output as a signal to check, not a verdict.