Research map / Learning theory and algorithms / Prediction and decision losses
Batch and Sequential Unlearning for Neural Networks
Nhung Bui, Xinyang Lu, Rachael Hwee Ling Sim, See-Kiong Ng, Bryan Kian Hsiang Low
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 pseudo-inverse of the Hessian instead of exact inverse (PINV-Newton) | Newton unlearning · CNN FMNIST class-level batch | ToW: 0.13 -> 0.05 | Hurt |
| add cubic regularization with automatically chosen damping (CuReNU) to Newton unlearning | Damped Newton · CNN FMNIST class-level batch | ToW: 0.05 -> 0.93 | Improved |
| use CuReNUS rather than CuReNU | CuReNU · CNN FMNIST class-level batch | ToW: 0.93 -> 0.99 | Improved |
| use CuReNU/CuReNUS update instead of damped Newton | Damped Newton · CNN FMNIST class-level batch | Newton update norm: 838.68 ± 742.96 -> 0.38 ± 0.05 | Improved |
| use second-order CuReNUS instead of first-order GD | GD · CNN FMNIST class-level batch | ToW: 0.14 -> 0.99 | Improved |
| use CuReNUS instead of GD in sequential unlearning | GD · Llama-2 TOFU sequential (last round) | ToW: 0.60 -> 0.80 | Improved |
| use CuReNUS instead of SCRUB in sequential unlearning | SCRUB · Llama-2 TOFU sequential (last round) | ToW: 0.72 -> 0.80 | Improved |
| use CuReNUS instead of SCRUB in sequential unlearning | SCRUB · ResNet-18 CIFAR-10 sequential (last round) | ToW: 0.944 -> 0.909 | No clear effect |
| use CuReNUS instead of GA/Rand. Lbls. in sequential unlearning | GA · ResNet-18 CIFAR-10 sequential (last round) | ToW: 0.143 -> 0.909 | Improved |
| use CuReNUS (stochastic Hessian-free) instead of SCRUB for efficiency | SCRUB · Llama-2 AG-News unlearning time | unlearning time (s): 6796.16 -> 85.26 | Improved |
| use CuReNUS (Hessian-free) instead of CuReNU for efficiency | CuReNU · CNN FMNIST unlearning time | unlearning time (s): 6355.31 -> 35.54 | Improved |
| use CuReNUS instead of CuReNU, peak memory | CuReNU · CNN FMNIST peak memory | peak memory (MB): 6226 -> 1588 | Improved |
Closest papers by meaning
- Hessian-Free Online Certified Unlearning (ICLR 2025 · similarity 0.88)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.88)
- Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.86)
- Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement (NeurIPS 2024 · similarity 0.86)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.86)
- Towards Source-Free Machine Unlearning (CVPR 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.
