What is machine learning?

In short: Machine learning (ML) means a computer learns a function from examples instead of a human programming it by hand. You supply inputs and desired outputs, the method finds the mapping.

In more detail: A model is a function with adjustable numbers (parameters, weights). During training you compare the prediction with the correct answer, measure the error (loss) and change the parameters so that the error gets smaller. You repeat this millions of times.

In detail

• Example: spam detection. Instead of writing rules (“contains the word prize”) you show thousands of mails labelled spam/not spam. The model learns weights for words and patterns.
• The learning cycle: predict → measure error → adjust parameters → repeat until good enough.
• Kinds of learning: supervised (with answers), unsupervised (without), reinforcement (with reward), see Kinds of learning: supervised, unsupervised, reinforcement.
• Generalisation: the goal is to be good on new, unseen data, not only on the training data, see Overfitting and the bias–variance trade-off.
• Limits: the model is only as good as its data and the goal you give it.
• Everyday: recommendations, traffic jam detection, face recognition, translation, voice assistants, credit scoring.

See also: Data, features and labels, Gradient descent and learning from data