Using hindsight to anchor past knowledge in continual learning
In continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement diff...
| Main Authors: | , , , , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
Association for the Advancement of Artificial Intelligence
2021
|
| _version_ | 1797098240734134272 |
|---|---|
| author | Chaudhry, A Gordo, A Dokania, P Torr, P Lopez-Paz, D |
| author_facet | Chaudhry, A Gordo, A Dokania, P Torr, P Lopez-Paz, D |
| author_sort | Chaudhry, A |
| collection | OXFORD |
| description | In continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement different types of experience replay, re-learning on past data stored in a small buffer known as episodic memory. In this work, we complement experience replay with a new objective that we call ``anchoring'', where the learner uses bilevel optimization to update its knowledge on the current task, while keeping intact the predictions on some anchor points of past tasks. These anchor points are learned using gradient-based optimization to maximize forgetting, which is approximated by fine-tuning the currently trained model on the episodic memory of past tasks. Experiments on several supervised learning benchmarks for continual learning demonstrate that our approach improves the standard experience replay in terms of both accuracy and forgetting metrics and for various sizes of episodic memory. |
| first_indexed | 2024-03-07T05:06:43Z |
| format | Conference item |
| id | oxford-uuid:da2b2e66-8c0d-44d2-8174-dead0bc543bd |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T05:06:43Z |
| publishDate | 2021 |
| publisher | Association for the Advancement of Artificial Intelligence |
| record_format | dspace |
| spelling | oxford-uuid:da2b2e66-8c0d-44d2-8174-dead0bc543bd2022-03-27T09:01:14ZUsing hindsight to anchor past knowledge in continual learningConference itemhttp://purl.org/coar/resource_type/c_5794uuid:da2b2e66-8c0d-44d2-8174-dead0bc543bdEnglishSymplectic Elements Association for the Advancement of Artificial Intelligence2021Chaudhry, AGordo, ADokania, PTorr, PLopez-Paz, DIn continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement different types of experience replay, re-learning on past data stored in a small buffer known as episodic memory. In this work, we complement experience replay with a new objective that we call ``anchoring'', where the learner uses bilevel optimization to update its knowledge on the current task, while keeping intact the predictions on some anchor points of past tasks. These anchor points are learned using gradient-based optimization to maximize forgetting, which is approximated by fine-tuning the currently trained model on the episodic memory of past tasks. Experiments on several supervised learning benchmarks for continual learning demonstrate that our approach improves the standard experience replay in terms of both accuracy and forgetting metrics and for various sizes of episodic memory. |
| spellingShingle | Chaudhry, A Gordo, A Dokania, P Torr, P Lopez-Paz, D Using hindsight to anchor past knowledge in continual learning |
| title | Using hindsight to anchor past knowledge in continual learning |
| title_full | Using hindsight to anchor past knowledge in continual learning |
| title_fullStr | Using hindsight to anchor past knowledge in continual learning |
| title_full_unstemmed | Using hindsight to anchor past knowledge in continual learning |
| title_short | Using hindsight to anchor past knowledge in continual learning |
| title_sort | using hindsight to anchor past knowledge in continual learning |
| work_keys_str_mv | AT chaudhrya usinghindsighttoanchorpastknowledgeincontinuallearning AT gordoa usinghindsighttoanchorpastknowledgeincontinuallearning AT dokaniap usinghindsighttoanchorpastknowledgeincontinuallearning AT torrp usinghindsighttoanchorpastknowledgeincontinuallearning AT lopezpazd usinghindsighttoanchorpastknowledgeincontinuallearning |