Research map / Continual learning / Online and replay-based
Continual evaluation for lifelong learning: Identifying the stability gap
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 |
|---|---|---|---|
| Increase evaluation interval from every iteration to every 1000 iterations | ER · Split-MiniImagenet | min-ACC: 0.5 ± 0.2 -> 3.6 ± 1.1 | Mixed |
| Evaluate every iteration instead of every 100 iterations | ER · Split-MNIST | min-ACC: 91.3 ± 0.2 -> 73.0 ± 4.4 | Hurt |
| Evaluate every iteration instead of every 100 iterations | ER · Split-CIFAR10 | min-ACC: 2.0 ± 2.5 -> 0.0 ± 0.0 | Hurt |
| Increase consecutive-task rotation from φ=10° to φ=80° | ER · Rotated-MNIST | min-ACC: 94.3 ± 0.3 -> 69.2 ± 2.2 | Hurt |
| Increase consecutive-task rotation from φ=10° to φ=80° | ER · Rotated-MNIST | WF10: 3.5 ± 0.3 -> 17.8 ± 2.0 | Hurt |
| Use continual evaluation rather than task-transition-only evaluation | ER and continual learning methods studied · Class-, task-, and domain-incremental benchmarks | worst-case accuracy | Helped |
| Begin learning a new task, when replay-buffer stability gradients are initially near zero | Experience Replay · Five class- and domain-incremental benchmarks | accuracy on previous tasks | Mixed |
| Apply gradient-constrained replay | GEM · Split-MNIST | min-ACC | Hurt |
| Apply knowledge distillation across tasks | LwF · Split-MNIST and Rotated-MNIST | accuracy on previous tasks | Mixed |
| Apply parameter regularization | EWC and SI · Split-MNIST and Rotated-MNIST | accuracy on previous tasks | Mixed |
Closest papers by meaning
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.91)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.90)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.90)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.90)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.90)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.90)
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.