We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo

Computer Science > Machine Learning

Title: Bilevel Continual Learning

Abstract: Continual learning aims to learn continuously from a stream of tasks and data in an online-learning fashion, being capable of exploiting what was learned previously to improve current and future tasks while still being able to perform well on the previous tasks. One common limitation of many existing continual learning methods is that they often train a model directly on all available training data without validation due to the nature of continual learning, thus suffering poor generalization at test time. In this work, we present a novel framework of continual learning named "Bilevel Continual Learning" (BCL) by unifying a {\it bilevel optimization} objective and a {\it dual memory management} strategy comprising both episodic memory and generalization memory to achieve effective knowledge transfer to future tasks and alleviate catastrophic forgetting on old tasks simultaneously. Our extensive experiments on continual learning benchmarks demonstrate the efficacy of the proposed BCL compared to many state-of-the-art methods. Our implementation is available at this https URL
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2007.15553 [cs.LG]
  (or arXiv:2007.15553v1 [cs.LG] for this version)

Submission history

From: Quang Pham [view email]
[v1] Thu, 30 Jul 2020 16:00:23 GMT (402kb,D)

Link back to: arXiv, form interface, contact.