Inference

In machine learning, inference means applying a prepared model to an input to produce an output. It can return a category, a score, a numerical prediction, or generated content. This operational meaning differs from statistical inference, which concerns drawing conclusions from data about unknown quantities or processes.

Apply the model that was prepared

Suppose a delivery model uses distance and parcel weight. For a new request, the system builds those inputs in the expected units, applies the model, and interprets its output. The input need not be unseen: repeating a request can also be inference.

Ordinary inference does not fit model parameters to each request. A system can separately learn online, but that update should be distinguished from producing the current prediction. Google's machine-learning glossary describes inference as using a trained model.

The calculation is part of a larger path

Authorization, feature retrieval, preprocessing, queuing, and response formatting can surround the model calculation. Serving provides that usable path; deployment makes the selected model and runtime available. Inference may happen on a device, on a server, or in a batch job.

A score is not automatically a calibrated probability, and a generated response is not automatically factual. Record the model version and input definitions, and evaluate the outputs for their actual use. Successful execution means that a computation finished, not that its prediction was correct.

Reference: Google: Machine learning glossary.