Research map / Continual learning / Online and replay-based
PAC-Bayes bounds for cumulative loss in Continual Learning
Lior Friedman, Ron Meir
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 |
|---|---|---|---|
| use VI (optimizes the bound) instead of EWC | PAC-Bayes bound · Split-MNIST | bound gap (CuL vs bound): EWC 0.9 ± 0.0 vs 4.2 ± 0.1; VI 17.6 ± 0.4 vs 19.4 ± 0.3 | Mixed |
| compare bound at final task versus averaged over tasks | PAC-Bayes bound · Split-MNIST VI | bound: 19.4 ± 0.3 -> 7.5 ± 0.5 | Improved |
| compare bound at final task versus averaged over tasks | PAC-Bayes bound · Permuted-MNIST EWC | bound: 10.6 ± 0.2 -> 5.1 ± 0.2 | Improved |
| compare bound at final task versus averaged over tasks | PAC-Bayes bound · Split-CIFAR10 VI | bound: 52.2 ± 0.2 -> 52.5 ± 0.2 | Improved |
| add experience replay to SGD | SGD · Split-ImageNet | forgetting: 7.8 ± 0.1 -> 2.8 ± 0.1 | Improved |
| add experience replay to SGD | SGD · Split-ImageNet | cumulative error: 34.8 ± 0.3 -> 55.7 ± 0.7 | Hurt |
| use EWC instead of SGD | SGD · Split-ImageNet | forgetting: 7.8 ± 0.1 -> 5.7 ± 0.4 | Improved |
| use over-parametrized deep model instead of under-parametrized | oracle bound · Swap tasks regression | cumulative error: 0.11 ± 0.01 -> 0.01 ± 0.00 | Improved |
| alternate tasks instead of one swap in linear over-parametrized regression | oracle bound · linear regression | cumulative error: 10.3 ± 0.2 -> 280.1 ± 5.4 | Hurt |
| gradual task change instead of swap in linear regression | oracle bound · linear regression | cumulative error: 10.3 ± 0.2 -> 4.7 ± 0.2 | Improved |
Closest papers by meaning
- Online PAC-Bayes Learning (NeurIPS 2022 · similarity 0.89)
- A Statistical Theory of Regularization-Based Continual Learning (ICML 2024 · similarity 0.87)
- Recursive PAC-Bayes: A Frequentist Approach to Sequential Prior Updates with No Information Loss (NeurIPS 2024 · similarity 0.87)
- Model Uncertainty Quantification by Conformal Prediction in Continual Learning (ICML 2025 · similarity 0.87)
- Temporal-Difference Variational Continual Learning (NeurIPS 2025 · similarity 0.87)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · 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.
