Bookworm continual learning: beyond zero-shot learning and continual learning

Authors: Kai Wang, Luis Herranz, Anjan Dutta, Joost van de Weijer

Publication: TASK-CV workshop in ECCV 2020

Publication year: 2020

PAPER| VIRTUAL MEETING

Abstract:

We propose bookworm continual learning (BCL), a setting combining continual updates and semantic descriptions to infer unseen classes. We address BCL with the bidirectional imagination (BImag), a feature generation framework that generates synthetic features for both past and future classes. We observe that conditioning the feature generator on attributes can actually harm the continual learning ability, and propose two variants to alleviate this problem.

Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>