Probably Approximately Correct (PAC) learning is a term used in machine learning to describe whether it is theoretically possible to construct a suitable model or hypothesis to solve a problem from some training data.

Approximately correct means the error in the hypothesis is very small. Probably approximately correct means that the learning algorithm will come up with an approximately correct hypothesis most of the time, and PAC learning means that this is done using a reasonable amount of training data in a reasonable amount of time.

This has similarities to deciding whether a computation problem is class P or NP-hard.