Research map / Continual learning / Online and replay-based
Rehearsal Revealed: The Limits and Merits of Revisiting Samples in Continual Learning
Eli Verwimp, Matthias De Lange, Tinne Tuytelaars
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER · MNIST (|M|=100) | average accuracy after 5 tasks: 72.9 -> 82.3 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER · MNIST (|M|=500) | average accuracy after 5 tasks: 87.1 -> 90.0 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER · MNIST (|M|=1000) | average accuracy after 5 tasks: 90.2 -> 91.8 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER · MNIST (|M|=200) | average accuracy after 5 tasks: 81.8 -> 87.6 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER · MNIST (|M|=2000) | average accuracy after 5 tasks: 91.8 -> 92.6 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · CIFAR10 (|M|=100) | average accuracy after 5 tasks: 46.7 -> 47.4 | No clear effect |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · CIFAR10 (|M|=500) | average accuracy after 5 tasks: 65.1 -> 66.2 | No clear effect |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · CIFAR10 (|M|=1000) | average accuracy after 5 tasks: 70.7 -> 70.8 | No clear effect |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · Mini-Imagenet (|M|=100) | average accuracy after 5 tasks: 25.7 -> 29.8 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · Mini-Imagenet (|M|=500) | average accuracy after 5 tasks: 39.7 -> 42.3 | Improved |
| add n memory-only update steps after each task minimum (ER-step) (ablation) | ER (stable) · Mini-Imagenet (|M|=1000) | average accuracy after 5 tasks: 46.6 -> 47.4 | Improved |
| add rehearsal of stored samples versus plain finetuning | finetuning · MNIST, CIFAR10, Mini-Imagenet (5 tasks, shared head) | average accuracy over all tasks: 60%, 9% and 7% gain | Improved |
Closest papers by meaning
- Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · similarity 0.95)
- A simple but strong baseline for online continual learning: Repeated Augmented Rehearsal (NeurIPS 2022 · similarity 0.92)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.90)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.90)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.89)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.88)
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.
