Glossary

Model

A model is a calculation whose behavior additionally depends on stored numerical values — its parameters. An algorithm (the training algorithm) sets those values from examples; 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 works only with its fixed parameters: it takes an input and calculates an output from it.

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, without a single line of code changing.

Not to be confused with a rulebook: In a classic program, you can find the line that triggered a decision and change it. A purely trained model has no such line, only numbers. In large models there are so many that a single parameter usually can’t be described in 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 service or app. A language model is a model trained specifically on text.

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

Last changed on · commit 602ed93