Training

Training is the process of fitting a model to data according to an objective. A supervised model learns from inputs and target outcomes; other learning settings obtain their learning signal differently. Training changes the model's learned state. Inference uses a prepared model to produce outputs.

What changes during training?

Imagine predicting a delivery duration from distance. A simple model starts with a coefficient that assigns minutes per kilometre. Training uses examples to adjust that coefficient so predictions better match observed durations under the chosen loss, a measure of prediction error. This is an illustration, not a claim that distance alone is sufficient.

Many models learn through repeated updates, but training is not synonymous with one algorithm such as gradient descent. Google's glossary distinguishes training, model parameters, and inference.

Fitting the examples is not enough

A model may memorize training examples or learn a relationship that fails in another period. Use separate data to guide choices and estimate performance on the intended use. Prevent information that would be unavailable at prediction time from entering the training features.

Training produces a candidate model, not automatic approval for production. Save the data scope, preprocessing, configuration, and evaluation evidence. Deployment makes an accepted model available in an execution environment; ordinary prediction requests need not repeat its training.

Reference: Google: Machine learning glossary.