Given that a learning algorithm achieves a training error \ensuremath{\hat{\epsilon}_{\rm M}} on its training set, what do we expect its test error to be? This is an inference problem (``Given A, predict B") so it must have a Bayesian answer. This note discusses the forward model and prior required to get sensible answers.
All postscript files are compressed with gzip - see this page for advice about gzip, if needed.