Learning from examples, not instructions
Ordinary software does exactly what it is told, step by step. Machine learning flips that: instead of coding the rule, you show the system many examples and let it infer the rule itself. Feed it thousands of labelled photos and it learns what a cat looks like; feed it years of sales figures and it learns what tends to precede a good quarter. The whole craft is getting a model to generalise — to be right about examples it has never seen, not just to memorise the ones it has.
Three broad styles
- Supervised — learn from labelled examples (this email is spam, this one isn't) to predict labels for new cases.
- Unsupervised — find structure in unlabelled data, such as grouping customers who behave alike.
- Reinforcement — learn by trial and error against a reward, the way a system learns to play a game.
Most of the modern excitement — neural networks and deep learning — is a powerful sub-area of machine learning, not something separate. And good machine learning still rests on the unglamorous craft of data analysis: a model is only as honest as the data behind it.