Research map / Continual learning / Online and replay-based
Learning curves for continual learning in neural networks: Self-knowledge transfer and forgetting
Ryo Karakida, Shotaro Akaho
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 |
|---|---|---|---|
| Train sequentially on task A before task B, with target similarity above the critical value | NTK continual learning · Subsequent task B | generalization error: E_{A\rightarrow B}/E_B \sim 2(1-\rho); \rho>1/2 | Improved |
| Train sequentially on task A before task B when target similarity is sufficiently low | NTK continual learning · Subsequent task B | generalization error: \rho < \rho^*:= \sqrt{\gamma_A}/(1+\sqrt{\gamma_A}) | Hurt |
| Train on task B after task A when targets are even subtly dissimilar | NTK continual learning · Previous task A | backward generalization error: E_{A\rightarrow B}^{back}(\rho) \sim \sum_i \eta_i^2 (1-\rho) | Hurt |
| Use equal sample sizes for the two tasks with the same target | NTK continual learning · Same target, two tasks | generalization error: E_{A\rightarrow B}(1) < E_{ave} < E_A=E_B | Improved |
| Make task B's sample size much larger than task A's for the same target | NTK continual learning · Same target, task B | generalization error relative to task-B-only training: E_{A\rightarrow B}(1)/E_B \sim 1/(1-\gamma_A) | Hurt |
| Make task B's sample size much smaller than task A's for the same target | NTK continual learning · Same target, previous task A | generalization error relative to task-A-only training: E_{A\rightarrow B}(1)/E_A \sim 1/(1-\gamma_B) | Hurt |
| Train on additional tasks with equal sample sizes and the same target | NTK continual learning · Same target, multiple tasks | generalization error: E_{n+1} \leq E_n; if all eigenvalues are positive, E_{n+1} < E_n | Improved |
| Add label noise | Sequential training on the same target · Synthetic data and deep-network experiments | generalization error | Hurt |
| Use unbalanced sample sizes across tasks | Sequential training on the same target · Synthetic data, MNIST, and CIFAR-10 | generalization error | Mixed |
Closest papers by meaning
- Understanding the Dynamics of Forgetting and Generalization in Continual Learning via the Neural Tangent Kernel (ICLR 2026 · similarity 0.93)
- Theory on Forgetting and Generalization of Continual Learning (ICML 2023 · similarity 0.92)
- Disentangling and mitigating the impact of task similarity for continual learning (NeurIPS 2024 · similarity 0.89)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.88)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.88)
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.
