Sparsity and heterogeneous dropout for continual learning in the null space of neural activations
| dc.contributor.author | Abbasi, Ali | |
| dc.contributor.author | Nooralinejad, Parsa | |
| dc.contributor.author | Braverman, Vladimir | |
| dc.contributor.author | Pirsiavash, Hamed | |
| dc.contributor.author | Kolouri, Soheil | |
| dc.date.accessioned | 2022-11-14T15:46:18Z | |
| dc.date.available | 2022-11-14T15:46:18Z | |
| dc.date.issued | 2022 | |
| dc.description | 1st Conference on Lifelong Learning Agents; Montreal, Canada; August 22 - 24, 2022 | |
| dc.description.abstract | Continual/lifelong learning from a non-stationary input data stream is a cornerstone of intelligence. Despite their phenomenal performance in a wide variety of applications, deep neural networks are prone to forgetting their previously learned information upon learning new ones. This phenomenon is called “catastrophic forgetting” and is deeply rooted in the stability-plasticity dilemma. Overcoming catastrophic forgetting in deep neural networks has become an active field of research in recent years. In particular, gradient projection-based methods have recently shown exceptional performance at overcoming catastrophic forgetting. This paper proposes two biologically-inspired mechanisms based on sparsity and heterogeneous dropout that significantly increase a continual learner’s performance over a long sequence of tasks. Our proposed approach builds on the Gradient Projection Memory (GPM) framework. We leverage k-winner activations in each layer of a neural network to enforce layer-wise sparse activations for each task, together with a between-task heterogeneous dropout that encourages the network to use non-overlapping activation patterns between different tasks. In addition, we introduce two new benchmarks for continual learning under distributional shift, namely Continual Swiss Roll and ImageNet SuperDog-40. Lastly, we provide an in-depth analysis of our proposed method and demonstrate a significant performance boost on various benchmark continual learning problems. | en_US |
| dc.description.sponsorship | SK and HP were supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR00112190135. VB was supported by DARPA under Contract No. HR00112190130. | en_US |
| dc.description.uri | https://proceedings.mlr.press/v199/abbasi22a | en_US |
| dc.format.extent | 12 pages | en_US |
| dc.genre | conference papers and proceedings | en_US |
| dc.genre | preprints | en_US |
| dc.identifier | doi:10.13016/m2wgwn-iyjz | |
| dc.identifier.citation | Abbasi, Ali, Parsa Nooralinejad, Vladimir Braverman, Hamed Pirsiavash, and Soheil Kolouri. “Sparsity and Heterogeneous Dropout for Continual Learning in the Null Space of Neural Activations.” In Proceedings of The 1st Conference on Lifelong Learning Agents, 617–28. PMLR, 2022. https://proceedings.mlr.press/v199/abbasi22a.html. | |
| dc.identifier.uri | http://hdl.handle.net/11603/26318 | |
| dc.language.iso | en_US | en_US |
| dc.publisher | MLResearchPress | |
| dc.relation.isAvailableAt | The University of Maryland, Baltimore County (UMBC) | |
| dc.relation.ispartof | UMBC Computer Science and Electrical Engineering Department Collection | |
| dc.relation.ispartof | UMBC Faculty Collection | |
| dc.rights | This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author. | en_US |
| dc.title | Sparsity and heterogeneous dropout for continual learning in the null space of neural activations | en_US |
| dc.type | Text | en_US |