Research map / Continual learning / Online and replay-based
Scalable Recollections for Continual Lifelong 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 compressed recollections instead of real-example storage with GEM at the same incremental storage | GEM · MNIST-Rotations | retention: 62.5 -> 79.0 (100); 67.4 -> 81.5 (200) | Helped |
| Use compressed recollections instead of real-example storage at equal incremental storage | Replay · MNIST-Rotations | retention: 63.4 -> 75.6 (100); 71.3 -> 81.1 (200) | Helped |
| Use compressed recollections instead of real-example storage at equal incremental storage | Replay · Incremental CIFAR-100 | retention: 29.4 -> 39.7; 33.4 -> 47.9; 43.0 -> 51.6 | Mixed |
| Use compressed recollections instead of real-example storage at equal incremental storage | GEM · Incremental CIFAR-100 | retention: 23.4 -> 52.4; 40.6 -> 56.0; 48.7 -> 59.0 | Helped |
| Initialize the recollection autoencoder using CIFAR-10 transfer (ablation) | Replay Recollections · Incremental CIFAR-100 | retention: 43.7 -> 49.7 | Helped |
| Initialize the recollection autoencoder using CIFAR-10 transfer (ablation) | GEM Recollections · Incremental CIFAR-100 | retention: 43.7 -> 54.2 | Helped |
| Increase replay training from one to three gradient descent steps per example (ablation) | Replay Real Storage · Omniglot | retention accuracy: 4.3% -> 4.8% | Helped |
| Increase training from one to three steps per example (ablation) | Scalable Recollections · Omniglot | retention accuracy: 9.3% -> 13.0% | Helped |
| Use a recollection index buffer rather than sampling latent codes directly (ablation) | Discrete VAE recollection module · MNIST training examples | nearest neighbor distortion: 0.074 -> 0.054 (38 2d variables); 0.081 -> 0.021 (168 2d variables) | Helped |
| Use a recollection buffer rather than direct latent-code sampling for distillation (ablation) | CNN teacher to CNN student · MNIST | student accuracy | Helped |
| Use CIFAR-10 pretraining for the reasoning model (ablation) | Online model without replay buffer · Incremental CIFAR-100 | L1 reconstruction distortion | Helped |
| Use active and diverse sampling instead of random sampling for generative distillation (ablation) | CNN teacher and student · MNIST | student accuracy: 10 episodes: 10.67 -> 11.51; 100: 27.05 -> 29.93; 1000: 94.81 -> 92.54 | Mixed |
Closest papers by meaning
- Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · similarity 0.87)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.87)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.86)
- Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond (ICLR 2025 · similarity 0.85)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.85)
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.