Research map / Continual learning / Online and replay-based
Online Structured Laplace Approximations for Overcoming Catastrophic Forgetting
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 |
|---|---|---|---|
| use Kronecker factored curvature instead of diagonal curvature (ablation) | Online Laplace · Vision datasets, Avg. | test accuracy (%): 82.71 -> 85.40 | Helped |
| use Kronecker factored curvature instead of diagonal curvature (ablation) | Online Laplace · Permuted MNIST, 50 tasks | mean test accuracy: over 90% | Helped |
| recalculate curvature for each task rather than retain previous estimates (ablation) | Online Laplace · Permuted MNIST | mean test accuracy | No clear effect |
| increase the regularization multiplier from λ=1 to λ=3 (ablation) | Online Laplace, Kronecker factored · Permuted MNIST | average test accuracy: λ=1; λ=3 | Helped |
| increase the regularization multiplier to λ=100 (ablation) | Online Laplace · Permuted MNIST | test accuracy on the most recent task: λ=100 | Hurt |
| use Kronecker factored curvature instead of diagonal curvature (ablation) | Laplace methods · Disjoint MNIST | test accuracy | Helped |
| use Kronecker factored curvature instead of diagonal curvature (ablation) | Online Laplace · Vision datasets | test accuracy across tasks | Helped |
| use Online Laplace rather than separate per-task Laplace penalties | Laplace continual-learning methods · Permuted MNIST | mean test accuracy | Helped |
| use Online Laplace rather than Synaptic Intelligence | Continual-learning methods · Vision datasets | test accuracy across tasks | Helped |
Closest papers by meaning
- A Scalable Laplace Approximation for Neural Networks (ICLR 2018 · similarity 0.89)
- Addressing Catastrophic Forgetting in Few-Shot Problems (ICML 2021 · similarity 0.88)
- Kolmogorov-Arnold Networks Still Catastrophically Forget but Differently from MLP (AAAI 2025 · similarity 0.86)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.85)
- Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning (ICLR 2025 · similarity 0.85)
- Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online 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.