Reproduce “Online Continual Learning with Maximally Interfered Retrieval” 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
Rahaf Aljundi, Lucas Caccia, Eugene Belilovsky, Massimo Caccia, Min Lin, Laurent Charlin, Tinne Tuytelaars
A new method for online continual learning, Maximally Interfered Retrieval (MIR), improves knowledge retention and significantly reduces catastrophic forgetting by selecting past samples for replay that are most negatively impacted by learning new information. This targeted approach consistently enhances performance across various replay paradigms and challenging datasets.
Key insights
- Actively selecting samples that are predicted to be maximally interfered by new learning is consistently more effective for preserving past knowledge than random sampling in online continual learning.
- The core principle of estimating interference can be successfully applied across different replay paradigms, including storing actual past data, generating synthetic data, and utilizing compressed latent representations.
- For generative and hybrid replay, crucial factors include ensuring the quality of synthetic samples (e.g., via an entropy penalty) and maintaining consistency in data processing (e.g., autoencoding all inputs) to prevent the classifier from exploiting distribution shifts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.