Research map / Continual learning / Online and replay-based
Gradient-based Editing of Memory Examples for Online Task-free Continual Learning
Xisen Jin, Arka Sadhu, Junyi Du, Xiang Ren
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 GMED editing to memory replay | MIR · Split MNIST | final accuracy (%): 85.72 ± 1.2 -> 86.52$^{**}$ ± 1.4 | Improved |
| Add GMED editing to memory replay | MIR · Permuted MNIST | final accuracy (%): 79.13 ± 0.7 -> 79.25 ± 0.8 | No clear effect |
| Add GMED editing to augmented memory replay | ER_aug · Split CIFAR-10 | final accuracy (%): 46.29 ± 2.7 -> 47.47$^*$ ± 3.2 | Improved |
| Replace the GMED objective with random fixed-stride perturbations (ablation) | ER replay · Split mini-ImageNet | final accuracy (%): 25.92 ± 1.2 -> 26.50 ± 1.3 | Improved |
| Use GMED's interference-focused edit rather than random perturbation (ablation) | ER replay · Split CIFAR-10 | final accuracy (%): 32.26 ± 1.9 -> 34.84$^*$ ± 2.2 | Improved |
| Use GMED's interference-focused edit rather than adversarial loss maximization (ablation) | ER replay · Split CIFAR-10 | final accuracy (%): 31.69 ± 0.8 -> 34.84$^*$ ± 2.2 | Improved |
| Increase editing from one to three gradient steps (ablation) | ER_aug + GMED · Split CIFAR-10 | final accuracy (%): 34.84 ± 2.2 -> 34.93 ± 1.4 | No clear effect |
| Increase editing from one to five gradient steps (ablation) | ER_aug + GMED · Split mini-ImageNet | final accuracy (%): 27.27 ± 1.8 -> 26.36 ± 2.0 | Hurt |
| Use a nonzero regularization strength to constrain editing (ablation) | ER + GMED · Across tested datasets | final accuracy | Improved |
| Set editing decay rate γ to 1.0 instead of less than 1 (ablation) | ER + GMED and MIR + GMED · Across six datasets | final accuracy (%) | Mixed |
| Replay additional memory batches instead of editing examples (ablation) | MIR · Rotated MNIST | final accuracy (%): 77.50 ± 1.6 -> 78.07 ± 1.1 | Mixed |
| Edit memory examples using GMED | ER + GMED · Fuzzy-boundary Split CIFAR-10 | final accuracy (%): 37.15 ± 1.6 -> 40.57$^*$ ± 1.7 | Improved |
Closest papers by meaning
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.91)
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.90)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.89)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.89)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.89)
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.
