HINT: Hypernetwork approach to training weight interval regions in continual learning

Patryk Krukowski, Anna Bielawska, Kamil Ksiazek, Pawel Wawrzynski, Pawel Batorski, Przemyslaw Spurek

Abstract: Highlights•HINT propagates intervals from the hypernetwork to target weights.•Our method maps interval intersections to a universally effective region.•HINT outputs hyperrectangles and is trained with the worst-case interval loss.•Non-empty interval intersections and regularization enable non-forgetting.•Our approach scales to large datasets in Task-, Class-, and Domain-IL.