model

In machine learning, a model is the thing that does the predicting — a recipe that turns an input into an answer. Feed it the features of a house and it returns a price; feed it an email and it returns "spam" or "not spam." Before training, the model is a blank recipe with adjustable dials; training is the process of turning those dials until the recipe gives good answers on the examples it has seen. Afterwards, the trained model is just a frozen set of those dial settings, packaged so it can answer new questions.

It helps to separate three things people lazily all call "the model." First there is the form — a straight line, a decision tree, a neural network — which decides the shape of patterns it can express. Second there are the parameters, the numbers learned from data that fill in that form. Third is the trained artifact, form plus learned parameters, the file you actually deploy. A useful slogan: the form is the empty bookshelf, the parameters are the books, and training is the act of stocking the shelf.

A model is a deliberate simplification of reality, never reality itself — as the statistician George Box put it, all models are wrong, but some are useful. A weather model does not contain the sky; it captures just enough pattern to be helpful. So a model's worth is not whether it is "true" but whether it makes good-enough predictions on data it has never seen. A model that dazzles on its training examples but stumbles in the real world has failed at the only job that matters.