Research map / Continual learning / Online and replay-based
Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models
Alvin Heng, Harold Soh
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 generative replay term over remembered classes (ablation) | Selective Amnesia (CIFAR10 DDPM) · CIFAR10 remembered classes | FID: 126 -> 9.08 | Improved |
| increase Fisher (EWC) weight lambda from 1 to 10 (ablation) | Selective Amnesia (CIFAR10 DDPM) · CIFAR10 remembered classes | FID: 19.3 -> 9.08 | Improved |
| increase lambda from 10 to 100 (ablation) | Selective Amnesia (CIFAR10 DDPM) · CIFAR10 forgotten class | classifier entropy: 1.47 -> 1.07 | Hurt |
| increase lambda from 10 to 100 (ablation) | Selective Amnesia (CIFAR10 DDPM) · CIFAR10 remembered classes | FID: 9.08 -> 8.33 | No clear effect |
| replace ELBO surrogate: maximize likelihood of uniform surrogate instead of naive ELBO minimization (ablation) | Selective Amnesia (VAE MNIST) · MNIST forget digit 0 | forgetting quality | Improved |
| apply SA with lambda=10 on STL10 | DDPM STL10 · STL10 remembered classes | FID: 14.5 -> 18.0 | Hurt |
| apply SA with lambda=10 on CIFAR10 | DDPM CIFAR10 · CIFAR10 remembered classes | FID: 9.67 -> 9.08 | Improved |
| use SA instead of ESD-x for forgetting celebrities | Stable Diffusion v1.4 · Forget Brad Pitt | proportion of images without faces: 0.347 -> 0.058 | Improved |
| use SA instead of SLD/ESD for GCD score | Stable Diffusion v1.4 · Forget Brad Pitt | GCD Score: 0.00474 / 0.0201 vs 0.0752 | Hurt |
| apply SA to forget nudity vs ESD/SLD | Stable Diffusion v1.4 · I2P nudity | nudity content | Hurt |
Closest papers by meaning
- Sculpting Memory: Multi-Concept Forgetting in Diffusion Models via Dynamic Mask and Concept-Aware Optimization (ICCV 2025 · similarity 0.89)
- Generative vs. Discriminative: Rethinking The Meta-Continual Learning (NeurIPS 2021 · similarity 0.89)
- A Comprehensive Survey of Forgetting in Deep Learning Beyond Continual Learning (TPAMI 2024 · similarity 0.89)
- Leveraging Catastrophic Forgetting to Develop Safe Diffusion Models against Malicious Finetuning (NeurIPS 2024 · similarity 0.88)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.87)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (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.
