Model

A model is a calculation whose behavior additionally depends on stored numerical values — its parameters. A training algorithm, meaning a fixed sequence of steps for learning from examples, sets those values; no human decides them one by one.

Mental image: a mixing desk with a huge number of knobs. The number and arrangement of the knobs (the model’s architecture, or basic construction plan) are fixed. Their exact positions (the parameters) only emerge from training.

Once trained, a model behaves like a program again in operation: it takes an input and produces an output.

An example: A simple spam filter stores a weight for every word, say +3 for “prize” and −2 for “invoice” (made-up values). For each email, it adds up the weights of its words and compares the sum with a threshold. That calculation says nothing about spam. Which words are suspicious lives in the numbers alone. Change one and the filter decides differently.

Not to be confused with a rulebook: In a classic program, you can find the line that triggered a decision and change it. A trained model has no such line, only numbers. In large models there are so many that no single parameter means anything you could put into words.

Where you’ll come across it: In chat apps, where you can often choose between different models, in news about new model versions, and on product pages that name the model behind a feature. A language model is a model trained specifically on text.

Explained in more depth in Program, Algorithm, Model Compared.