Research map / Continual learning / Online and replay-based
Learning without Isolation: Pathway Protection 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 Learning without Isolation (LwI), a data-free continual learning method that fuses old and new models via layer-wise graph matching, aligning channels with high similarity in shallow layers and low similarity in deep layers to protect task pathways without parameter isolation.
Problem
Catastrophic forgetting in sequential task learning; existing parameter-isolation methods need task IDs and reduce learnable capacity as tasks grow, and privacy limits rehearsal.
Approach
Treat model fusion between old and new task networks as a graph matching problem over channels (solved per layer with Sinkhorn/Hungarian), maximizing channel similarity in shallow layers and minimizing it in deep layers, so different tasks occupy different channel pathways without masking weights.
Key results
- On CIFAR-100 with ResNet18 (5 splits), LwI reaches 51.95% task-agnostic accuracy vs 44.66% for LwF without exemplars.
- On CIFAR-100 with ResNet32 (20 splits, task-aware), LwI achieves 83.19% vs 82.51% for WSN.
- On Tiny-ImageNet with ResNet18 (5 splits, task-agnostic), LwI reaches 34.33% vs 26.76% for LwF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LwI instead of LUCIR | ResNet32 continual learning · CIFAR-100, task-agnostic, 5 splits | accuracy: 48.48 $\pm$ 1.16 vs 43.42 $\pm$ 0.58 | Hurt |
| Use LwI instead of WSN | ResNet32 continual learning · CIFAR-100, task-aware, 5 splits | accuracy: 75.47 $\pm$ 0.48 -> 76.10 $\pm$ 0.33 | Helped |
| Use LwI instead of LUCIR | ResNet18 continual learning · CIFAR-100, task-agnostic, 5 splits | accuracy: 55.67 $\pm$ 1.04 vs 51.95 $\pm$ 0.56 | Hurt |
| Use LwI instead of LUCIR | ResNet18 continual learning · CIFAR-100, task-aware, 5 splits | accuracy: 81.22 $\pm$ 0.25 vs 81.10 $\pm$ 0.80 | Hurt |
| Use LwI instead of LUCIR | ResNet18 continual learning · Tiny-Imagenet, task-agnostic, 5 splits | accuracy: 30.17 $\pm$ 0.37 -> 34.33 $\pm$ 0.51 | Helped |
| Use LwI instead of WSN | ResNet18 continual learning · Tiny-Imagenet, task-aware, 5 splits | accuracy: 57.38 $\pm$ 0.51 -> 62.97 $\pm$ 0.14 | Helped |
| Add minimum-similarity matching to the final layer (ablation) | LwI · CIFAR-100, ResNet32, task-agnostic, 5 splits | accuracy: 26.84 $\pm$ 0.86 -> 43.42 $\pm$ 0.58 | Helped |
| Add minimum-similarity matching to the final layer (ablation) | LwI · CIFAR-100, ResNet32, task-agnostic, 10 splits | accuracy: 23.50 $\pm$ 0.30 -> 30.62 $\pm$ 1.08 | Helped |
| Add minimum-similarity matching to the final layer (ablation) | LwI · CIFAR-100, ResNet32, task-agnostic, 20 splits | accuracy: 16.05 $\pm$ 0.28 -> 20.31 $\pm$ 0.77 | Helped |
| Add task diversion (match low similarity in deep layers) (ablation) | LwI · Tiny-Imagenet, ResNet18, task-agnostic, 5 splits | accuracy: 30.75 $\pm$ 0.43 -> 34.33 $\pm$ 0.51 | Helped |
| Add task diversion (match low similarity in deep layers) (ablation) | LwI · Tiny-Imagenet, ResNet18, task-aware, 5 splits | accuracy: 62.12 $\pm$ 0.26 -> 62.97 $\pm$ 0.14 | Helped |
| Replace Euclidean distance with cosine similarity (ablation) | LwI · Tiny-Imagenet, ResNet18, task-agnostic, 5 splits | accuracy: 34.33 $\pm$ 0.51 vs 34.17 $\pm$ 0.54 | No clear effect |
Surprising findings
- Under task-agnostic evaluation, the data-free LwI underperforms exemplar-based iCaRL and LUCIR on CIFAR-100 with ResNet32, which the authors attribute to inconsistent head output sizes.
Limitations
- Validated only on small ResNet32 and ResNet18 models, not large models.
- Task-agnostic accuracy lags exemplar-based methods on CIFAR-100 ResNet32.
- Graph matching cost remains significant and could benefit from sparse matrix acceleration.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-100 Tiny-ImageNet
Metrics: task-agnostic accuracy task-aware accuracy
Baselines: EWC RWalk LwF SPG SPU GPM WSN iCaRL LUCIR
Builds on: Learning without Forgetting Winning Subnetworks (WSN) Sinkhorn algorithm Git Re-Basin-style weight matching
Closest papers by meaning
- Does Continual Learning Equally Forget All Parameters? (ICML 2023 · similarity 0.88)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.88)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.88)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · 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.