Research map / Continual learning / Online and replay-based
Smart Replay: Adaptive Scheduling of Memory Rehearsal for Computational Resource-Aware Incremental Learning
Jianting Chen, Dianzhi Yu, Irwin King
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 |
|---|---|---|---|
| schedule replay ratio adaptively with Smart Replay instead of fixed ratio | ER · CIFAR-100 limited memory, ϑ=0 | average accuracy (%): 63.84 -> 64.58 vs best linear heuristic | Improved |
| schedule replay ratio adaptively with Smart Replay | iCaRL · Tiny-ImageNet limited memory, ϑ=0.1 | average accuracy (%): 44.80 -> 49.08 | Improved |
| schedule replay ratio adaptively with Smart Replay | iCaRL · Rotated-MNIST limited memory, ϑ=0.2 | average accuracy (%): 90.93 -> 90.69 vs fixed λ=0.4 | No clear effect |
| schedule replay ratio adaptively with Smart Replay on unlimited memory | ER · CIFAR-100 unlimited memory, ϑ=0 | average accuracy (%): 74.16 -> 74.70 vs Union | Improved |
| decreasing replay ratio schedule (0.8 to 0.2) instead of increasing | ER · CIFAR-100 limited memory, ϑ=0 | average accuracy (%): 63.84 -> 57.91 | Hurt |
| add smoothness regularization on replay ratio changes (ablation) | Smart Replay · CIFAR-100 limited memory | average accuracy (%): 36.89 -> 64.58 | Improved |
| add short-term gradient norm ratio term (ablation) | Smart Replay · CIFAR-100 limited memory | average accuracy (%): 40.20 -> 64.58 | Improved |
| add long-term weight ratio term (ablation) | Smart Replay · CIFAR-100 limited memory | average accuracy (%): 61.43 -> 64.58 | Improved |
| add gradient-direction cosine term (ablation) | Smart Replay · CIFAR-100 limited memory | average accuracy (%): 64.03 -> 64.58 | Improved |
| increase decay-rate difference Δρ from 1.0 to 3.0 (ablation) | Smart Replay · CIFAR-100 limited memory | average accuracy | Improved |
| vary update step size γ, interval z and initial ratio τ (ablation) | Smart Replay · CIFAR-100 | average accuracy | No clear effect |
| use larger fixed replay ratio with limited buffer | ER · Tiny-ImageNet limited memory, ϑ=0.2 | average accuracy (%): 43.96 -> 41.20 (λ=0.2 -> 0.4) | Hurt |
Closest papers by meaning
- Autoencoder-Based Hybrid Replay for Class-Incremental Learning (ICML 2025 · similarity 0.88)
- Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning (ICML 2025 · similarity 0.87)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.87)
- MIND: Multi-Task Incremental Network Distillation (AAAI 2024 · similarity 0.86)
- Towards Efficient Replay in Federated Incremental Learning (CVPR 2024 · similarity 0.86)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 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.
