Research map / Continual learning / Online and replay-based
Gradient based sample selection for online continual learning
Rahaf Aljundi, Min Lin, Baptiste Goujaud, Yoshua Bengio
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 |
|---|---|---|---|
| Use gradient-diversity greedy selection instead of random buffer selection | online continual learning · Disjoint MNIST, buffer 300 | average test accuracy: 37.5±1.3 -> 82.6±2.9 | Improved |
| Use gradient-diversity greedy selection instead of random buffer selection | online continual learning · Disjoint CIFAR-10 | average test accuracy: 28.6±1.2 -> 33.56± 1.7 | Improved |
| Select buffer samples with gradient-based methods rather than reservoir sampling | online continual learning · Imbalanced disjoint MNIST sequences | average accuracy: 69.12±4.3 -> 77.96± 3.5 | Improved |
| Use gradient-space clustering rather than feature-space clustering | online continual learning · Permuted MNIST | average test accuracy: 77.8 ± 0.3 -> 79.74±0.2 | Improved |
| Use greedy gradient selection rather than task-aware GEM | online continual learning · Disjoint CIFAR-10 | average test performance | Improved |
| Use constrained optimization updates instead of rehearsal updates (ablation) | GSS-IQP · Disjoint MNIST, buffer 100 | average test accuracy: 71.3 -> 76.26 | Improved |
| Use constrained optimization updates instead of rehearsal updates (ablation) | GSS-IQP · Disjoint MNIST, buffer 200 | average test accuracy: 78.9 -> 83.4 | Improved |
| Increase the number of sampled buffer gradients used to estimate similarity from n=1 to n=10 (ablation) | GSS-Greedy · Disjoint MNIST | average test accuracy: Avg. Acc.=67.3 for n=1 | Improved |
| Increase n to large values in greedy gradient similarity scoring (ablation) | GSS-Greedy · Disjoint MNIST | average test accuracy | Hurt |
| Increase batch size (ablation) | GSS-Greedy · Disjoint MNIST | average test accuracy | Hurt |
| Use a very small batch size (ablation) | GSS-Greedy · Disjoint MNIST | average test accuracy | Mixed |
| Increase rehearsal buffer size from 100 to 200 (ablation) | GSS-IQP (Rehearsal) · Disjoint MNIST | average test accuracy: 71.3 -> 78.9 | Improved |
Closest papers by meaning
- GCR: Gradient Coreset Based Replay Buffer Selection for Continual Learning (CVPR 2022 · similarity 0.91)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.88)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.88)
- Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning (ICML 2025 · similarity 0.87)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.87)
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · similarity 0.87)
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.
