Research map / Continual learning / Online and replay-based
Progress & Compress: A scalable framework for continual learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use P&C with online EWC consolidation rather than online EWC alone | continual learning on Omniglot · Omniglot, 50 alphabets, 5 passes | test accuracy: 79.15 { $\pm$ 1.9} -> 82.84 { $\pm$ 1.4} | Helped |
| Re-initialize the active column after each task (ablation) | P&C · Atari, positive transfer across 1–5 previous tasks | % single task performance: P&C (Active col, re-init): 127, 129, 125, 129, 128; P&C (Active col): 131, 127, 114, 106, 101 | Helped |
| Keep active-column parameters unchanged between tasks instead of re-initializing them (ablation) | P&C · Atari, positive transfer as previous tasks increase | % single task performance: P&C (Active col): 131, 127, 114, 106, 101; P&C (Active col, re-init): 127, 129, 125, 129, 128 | Hurt |
| Use P&C rather than Finetuning | continual learning on Atari · Atari, positive transfer after 5 previous tasks | % single task performance: 98 -> 128 | Helped |
| Train P&C after pre-training on a small number of other alphabets | P&C · Omniglot, learning curves on 10 unique alphabets | data efficiency | Helped |
| Use online EWC rather than EWC | sequential learning with limited training per task · Permuted MNIST, task retention and final task accuracy | accuracy: 500 and 1000 training steps: EWC accuracy is higher; 10,000 training steps: EWC makes retaining the most recent tasks mo | Mixed |
Closest papers by meaning
- Forget Forgetting: Continual Learning in a World of Abundant Memory (ICLR 2026 · similarity 0.89)
- Compacting, Picking and Growing for Unforgetting Continual Learning (NeurIPS 2019 · similarity 0.88)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Optimal Task Order for Continual Learning of Multiple Tasks (ICML 2025 · similarity 0.87)
- Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective (ICML 2025 · similarity 0.87)
- Mixture of Experts Meets Prompt-Based Continual Learning (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.