Research map / Continual learning / Online and replay-based
Exemplar-Free Continual Learning for State Space Models
Isaac Ning Lee, Leila Mahmoodi, Trung Le, Mehrtash Harandi
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 |
|---|---|---|---|
| integrate Inf-SSM into existing CIL baselines (ER, LUCIR, X-DER, L2P-R, CLFD) | CIL baselines on Vim-small · ImageNet-R, CIFAR-100, Caltech-256 (5 and 10 tasks) | AA and FM: AA +8.31% and FM -9.36% on average | Improved |
| add Inf-SSM to X-DER | X-DER · ImageNet-R, CIFAR-100, Caltech-256 | AA and FM: AA +19.61%, FM -23.16% on average | Improved |
| add Inf-SSM to LUCIR | LUCIR · ImageNet-R (5 tasks) | AIA: 60.97 -> 59.53 | Hurt |
| add Inf-SSM to ER | ER · ImageNet-R (5 tasks) | AA: 30.91 -> 31.26 | No clear effect |
| regularize observability subspace (A,C) with Inf-SSM instead of EWC/SI/MAS/LwF on (A,C) | Vim-small EFCIL · ImageNet-R, CIFAR-100, Caltech-256 | FM and AA: FM -40.18%, AA +21.73% on average | Improved |
| use Inf-SSM (A,C only) instead of baselines regularizing A,B,C | Vim-small EFCIL · ImageNet-R, CIFAR-100, Caltech-256 | FM and AA: FM -14.56%, AA +6.79% on average | Improved |
| use diagonal-A closed-form Sylvester solution instead of Bartels-Stewart (ablation) | Inf-SSM · Vim-small, n=16 | FLOPS: 25n^3 -> 4n^2, about 100x fewer FLOPS | Improved |
| regularize more Vim blocks with Inf-SSM (ablation) | Inf-SSM · Vim-small | performance | Improved |
| add Frobenius regularization of B (Inf-SSM+) (ablation) | Inf-SSM · CL benchmarks | AA, AIA, FM | No clear effect |
| use Inf-SSM loss instead of X-DER loss | CL loss cost · Vim-small, A40 GPU, per batch | latency: 1.2534 s -> 0.0960 s | Improved |
| replace EWC with Inf-SSM loss | regularizer cost · Vim-small, A40 GPU | per-batch latency: 0.0095 -> 0.0960 s per batch, but no 181.5038 s per-task Fisher pass | Hurt |
| Inf-SSM benefit grows with number of tasks (10 vs 5 tasks) | Inf-SSM integrated with baselines · Vim-small | AA and FM | Improved |
Closest papers by meaning
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- StableSSM: Alleviating the Curse of Memory in State-space Models through Stable Reparameterization (ICML 2024 · similarity 0.87)
- A Statistical Theory of Regularization-Based Continual Learning (ICML 2024 · similarity 0.87)
- MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models (NeurIPS 2026 · similarity 0.87)
- LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning (CVPR 2025 · similarity 0.86)
- Continuous Subspace Optimization for Continual Learning (NeurIPS 2025 · 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.
