<ul><li>A statistical learner's goal in noisy data is to resolve epistemic uncertainty about the test data distribution.</li><li>Epistemic uncertainty arises from various sources like multitask learning, distribution shift, and imperfect learning.</li><li>A new definition of epistemic error is introduced with a generic error bound that considers various sources of uncertainty.</li><li>Corollaries include specialized error bounds for Bayesian transfer learning, distribution shift, and generalization bounds.</li></ul>

Epistemic Errors of Imperfect Multitask Learners When Distributions Shift

Discover more