Reproduce “Rectification-based Knowledge Retention for Continual Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pravendra Singh, Pratik Mazumder, Piyush Rai, Vinay P. Namboodiri
Rectification-based Knowledge Retention (RKR) introduces task-specific modifications to deep neural networks, adapting existing weights and intermediate outputs to mitigate catastrophic forgetting in continual learning. It achieves state-of-the-art accuracy in both standard and generalized zero-shot task incremental learning while maintaining exceptional parameter efficiency.
Key insights
- Both the learned task-specific weight rectifications and the affine scaling factors for intermediate outputs are crucial for performance, with the removal of either component leading to substantial degradation.
- Low-rank approximation effectively balances the need for task-specific adaptations with minimal parameter overhead, allowing for scalable continual learning.
- Leveraging 'forward knowledge transfer' by initializing new task parameters from the previous task consistently improves performance across various settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.