Research map / Continual learning / Online and replay-based
On Plasticity, Invariance, and Mutually Frozen Weights in Sequential Task Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| reset mutually frozen weights before retraining | ResNet18 · CIFAR-10 -> CIFAR-100 | test accuracy: 74.2 ± 0.3 vs 73.1 ± 0.5 | Helped |
| reset mutually frozen weights before retraining | ResNet50 · ImageNet -> FashionMNIST | test accuracy: 97.8 ± 1.6 vs 97.2 ± 0.2 | Helped |
| reset mutually frozen weights before retraining | ResNet50 · ImageNet -> KMNIST | test accuracy: 99.2 ± 0.1 vs 99.2 ± 0.1 | No clear effect |
| reset mutually frozen weights before retraining | ResNet50 · ImageNet -> Patch Camelyon | test accuracy: 89.3 ± 0.6 vs 87.8 ± 1.4 | Helped |
| reset mutually frozen weights before retraining | ResNet18 · SVHN -> CIFAR-10 | test accuracy: 88.9 ± 0.2 vs 87.6 ± 0.1 | Helped |
| increase double/triple-sparse initial sparsity from 0.5 to 0.98 (ablation) | All-CNN · CIFAR-10 | test accuracy: 87.8 ± 4.9 -> 14.0 ± 7.0 | Hurt |
| increase mono-sparse initial sparsity from 0.5 to 0.98 (ablation) | All-CNN · CIFAR-10 | test accuracy: 92.1 ± 0.2 -> 92.1 ± 0.0 | No clear effect |
| increase double-sparse initial sparsity from 0.5 to 0.98 (ablation) | All-CNN · CIFAR-10 | test accuracy: 92.1 ± 0.1 -> 67.0 ± 9.0 | Hurt |
| remove trainable affine batch normalization weights (ablation) | ResNet18 · CIFAR-10 blurred -> sharp | test accuracy: 87.4 ± 0.5 vs 89.0 ± 0.5 | Mixed |
| reset frozen weights in AlexNet (no multiplicative norm weights) (ablation) | AlexNet · CIFAR-10 blurred -> sharp | test accuracy: 87.3 ± 0.2 vs 87.5 ± 0.1 | No clear effect |
| reset frozen weights | All-CNN · CIFAR-10 blurred -> sharp | test accuracy: 88.7 ± 0.2 vs 83.1 ± 4.9 | Helped |
| reset frozen weights | ResNet18 · CIFAR-10 regular -> random shift | test accuracy: 94.0 ± 0.1 vs 93.0 ± 0.3 | Helped |
Closest papers by meaning
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.88)
- Self-Normalized Resets for Plasticity in Continual Learning (ICLR 2025 · similarity 0.88)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.86)
- Feature Learning Dynamics in Infinite-Depth Neural Networks (NeurIPS 2026 · similarity 0.86)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.86)
- The Importance of Being Lazy: Scaling Limits of Continual Learning (ICML 2025 · similarity 0.85)
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.