Machine learning means writing programs that get better at a task by learning from data — instead of us hand-coding every rule.

Where it's used: spam filters, recommendations, search, face recognition, translation, fraud detection.

Four words to know:

• Data — examples, written as vectors of numbers.
• Model — a function with adjustable parameters that turns input into output.
• Learning — picking parameters that fit the data, usually by minimizing some error.
• Generalization — the real goal. Fitting data you already have is easy; doing well on new data is the point.

Exam trap: a model that fits training data perfectly isn't automatically good — that's overfitting (Session 6).