Research map / Continual learning / Online and replay-based
STAR: Stability-Inducing Weight Perturbation for Continual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STAR, a plug-and-play regularizer for rehearsal-based continual learning that perturbs weights adversarially to maximize KL divergence on correctly classified buffer samples and penalizes the resulting output change.
Problem
Catastrophic forgetting in class-incremental continual learning with rehearsal buffers; rehearsal alone still forgets, and the stability of model outputs in the local parameter neighborhood is under-explored.
Approach
Uses a surrogate for forgetting, the KL divergence between predictions on correctly classified buffer samples before and after a worst-case (gradient-ascent, layer-normalized) weight perturbation, and optimizes it with the gradient taken at the perturbed point, added to any rehearsal loss.
Key results
- Adding STAR to ER raises Split-CIFAR10 accuracy at buffer 100 from 36.39 to 51.5 (about +15 points).
- STAR improves ER-ACE, DER++ and X-DER (RPC) across all nine dataset/buffer settings in the baseline table.
- Against other regularizers on ER-ACE and DER++, STAR is best or second-best in most settings, with particularly strong gains at small buffer sizes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add STAR to DER++ | DER++ · Split-CIFAR10 | average accuracy: 57.65 -> 61.76; 64.88 -> 68.60; 72.70 -> 76.52 | Helped |
| Add STAR to ER-ACE | ER-ACE · Split-CIFAR10 | average accuracy: 52.95 -> 60.69; 61.25 -> 67.58; 71.16 -> 75.44 | Helped |
| Add STAR to X-DER (RPC) | X-DER (RPC) · Split-CIFAR10 | average accuracy: 52.75 -> 58.85; 58.48 -> 65.94; 64.77 -> 69.19 | Helped |
| Add STAR to DER++ | DER++ · Split-CIFAR100 | average accuracy: 25.11 -> 27.64; 37.13 -> 39.77; 52.08 -> 53.24 | Helped |
| Add STAR to X-DER (RPC) | X-DER (RPC) · Split-CIFAR100 | average accuracy: 37.23 -> 38.15; 48.53 -> 47.56; 57.00 -> 57.55 | Mixed |
| Add STAR to ER-ACE | ER-ACE · Split-miniImageNet | average accuracy: 17.95 -> 21.06; 22.60 -> 24.9; 27.92 -> 31.01 | Helped |
| Add STAR to DER++ | DER++ · Split-miniImageNet | average accuracy: 18.02 -> 22.4; 23.44 -> 28.19; 30.43 -> 33.36 | Helped |
| Use the gradient-based perturbation instead of random Gaussian perturbation of the same magnitude (ablation) | STAR · Split-CIFAR10 | average accuracy | Helped |
| Use only correctly classified buffer samples rather than all buffer samples in the STAR loss (ablation) | STAR · Split-CIFAR10 | average accuracy | Helped |
| Include both buffer and current-task data in the STAR loss rather than buffer data alone (ablation) | STAR · Split-CIFAR10 | average accuracy | No clear effect |
| Use only current-task data rather than buffer data in the STAR loss (ablation) | STAR · Split-CIFAR10 | forgetting | Hurt |
| Use three gradient ascent steps to calculate the perturbation instead of one (ablation) | STAR · Split-CIFAR10 | average accuracy | Mixed |
Surprising findings
- STAR increases final forgetting for ER on Split-CIFAR10 at buffer 100 (55.90 to 42.03 is a decrease, but ER+STAR standard deviation rises to 6.28), and forgetting is not uniformly reduced across settings.
- Using random Gaussian perturbations of the same magnitude instead of gradient-based worst-case perturbations yields much smaller gains.
Limitations
- Gradient approximation at the perturbed point is heuristic; accuracy of the approximation degrades with more ascent steps.
- Adds about two extra optimization steps per batch, roughly doubling or more training time (e.g., ER on miniImageNet 142 to 305 minutes).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 miniImageNet
Benchmarks: Split-CIFAR10 Split-CIFAR100 Split-miniImageNet
Metrics: average accuracy final forgetting standard deviation
Baselines: ER ER-ACE DER++ X-DER (RPC) sSGD oEWC OCM LiDER DualHSIC
Builds on: ER DER++ ER-ACE X-DER sharpness-aware minimization
Closest papers by meaning
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.92)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.90)
- Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · 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)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · 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.