Research map / Continual learning / Online and replay-based
Towards Better Plasticity-Stability Trade-Off in Incremental Learning: A Simple Linear Connector
Check 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 |
|---|---|---|---|
| Combine a null-space-projected stability model and a current-task plasticity model with the linear connector | Incremental learning with ResNet-18 · 10-split-CIFAR-100 | ACC (%): 73.77 -> 79.79 | Helped |
| Combine a null-space-projected stability model and a current-task plasticity model with the linear connector | Incremental learning with ResNet-18 · 20-split-CIFAR-100 | ACC (%): 77.16 -> 80.80 | Helped |
| Combine a null-space-projected stability model and a current-task plasticity model with the linear connector | Incremental learning with ResNet-18 · 25-split-TinyImageNet | ACC (%): 58.28 -> 64.61 | Helped |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 10-split-CIFAR-100 | BWT (%): -1.6 -> -0.92 | Helped |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 20-split-CIFAR-100 | BWT (%): -3.66 -> -5.00 | Hurt |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 25-split-TinyImageNet | BWT (%): -6.05 -> -6.00 | Helped |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 10-split-CIFAR-100 | Intransigence measure I₁₀ (%): 14.50 -> 8.10 | Helped |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 20-split-CIFAR-100 | Intransigence measure I₂₀ (%): 12.60 -> 7.00 | Helped |
| Use the linear connector instead of Adam-NSCL | Incremental learning · 25-split-TinyImageNet | Intransigence measure I₂₅ (%): 10.50 -> 5.75 | Helped |
| Set β to 0, selecting only the stability network (ablation) | Linear Connector · Incremental learning tasks shown in the β ablation | Accuracy on previous tasks | Helped |
| Set β to 1, selecting only the plasticity network (ablation) | Linear Connector · Incremental learning tasks shown in the β ablation | Accuracy on the current task | Helped |
| Set β close to 1/t to balance the two networks (ablation) | Linear Connector · 10-split-CIFAR-100 and 20-split-CIFAR-100 | Accuracy across tasks | Helped |
Closest papers by meaning
- Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective (ICML 2025 · similarity 0.89)
- Achieving a Better Stability-Plasticity Trade-Off via Auxiliary Networks in Continual Learning (CVPR 2023 · similarity 0.88)
- Activation by Interval-wise Dropout: A Simple Way to Prevent Neural Networks from Plasticity Loss (ICML 2025 · similarity 0.86)
- Mitigating Plasticity Loss in Continual Reinforcement Learning by Reducing Churn (ICML 2025 · similarity 0.86)
- A Study of Plasticity Loss in On-Policy Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Revisiting Plasticity in Visual Reinforcement Learning: Data, Modules and Training Stages (ICLR 2024 · 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.