I think that probability pattern matching is an interpretation of what AI is doing, rather than its actual functioning. There are no, AFAIK, explicit steps of probability pattern matching neither during its training nor during its run.
During the training, parameters (aka weights) of a function of specific form are calculated based on a set of input-output pairs. During the run, an output of this function is calculated according to accumulated history of inputs and responses. This is a basic mechanism, although there are variations, additions, modifications, randomizations, human interventions, etc.