Research map / Continual learning / Online and replay-based
Overcoming Catastrophic Forgetting by Bayesian Generative Regularization
Pei-Hung Chen, Wei Wei, Cho-Jui Hsieh, Bo Dai
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 |
|---|---|---|---|
| add EBM generative regularization to Bayesian VCL (ablation) | VCL · Fashion-MNIST | average classification accuracy (%): 76.9 -> 97.2 | Improved |
| add EBM generative regularization to Bayesian VCL (ablation) | VCL · CUB | average classification accuracy (%): 67.4 -> 78.8 | Improved |
| remove Bayesian parameter sampling and KL divergence, retaining the generative loss (ablation) | BGR · Fashion-MNIST | average classification accuracy (%): 97.2 -> 87.9 | Hurt |
| remove Bayesian parameter sampling and KL divergence, retaining the generative loss (ablation) | BGR · CUB | average classification accuracy (%): 78.8 -> 74.0 | Hurt |
| add generative EBM loss without Bayesian updates (ablation) | VCL · Fashion-MNIST | average classification accuracy (%): 76.9 -> 87.9 | Improved |
| add generative EBM loss without Bayesian updates (ablation) | VCL · CUB | average classification accuracy (%): 67.4 -> 74.0 | Improved |
| add L2 regularization to GEN (ablation) | GEN · Fashion-MNIST | average classification accuracy (%): 87.9 -> 90.9 | Improved |
| add L2 regularization to GEN (ablation) | GEN · CUB | average classification accuracy (%): 74.0 -> 72.8 | Hurt |
| use a small GAN of similar size instead of the larger VGR generator (ablation) | VGR · Split-MNIST | average classification accuracy (%): 88 | Hurt |
| reinitialize the classifier for each task in VGR-GAN-ONLY (ablation) | VGR · Split-MNIST | average classification accuracy (%): 66 | Hurt |
| add BGR generative regularization to iTAML | iTAML · Split-MNIST | average classification accuracy (%): 97.8 -> 98.4 | Improved |
| train with generative regularization rather than SGD alone | BGR · Permuted-MNIST, digit task after another task | accuracy (%): 99.7% to 54.2% -> 99.7% to 95.0% | Improved |
Closest papers by meaning
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.88)
- Temporal-Difference Variational Continual Learning (NeurIPS 2025 · similarity 0.88)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.87)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.87)
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.
