Research map / Label-efficient and robust learning / Recognition and long-tail
R-ESC: Robustly Erasing Space Concepts via Stochastic Feature Remapping
Kang Eun Jeon, Yunsung Kang, Do Kang, Tae-Young Lee, Gyeong-Moon Park, Jong Hwan Ko
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 prototype-orthogonal projection to remapping (ablation) | Remap · CIFAR-10 (KR setting, lr 0.1) | HM: 0.27 -> 66.89 | Improved |
| add prototype-orthogonal projection to remapping (ablation) | Remap · CIFAR-10 (class erasure) | HM: 94.47 -> 99.94 | Improved |
| add prototype-orthogonal projection to erasure (ablation) | Erase · CIFAR-10 (class erasure) | D_f: 14.38 -> 0.00 | Improved |
| add prototype-orthogonal projection to erasure (ablation) | Erase · CIFAR-10 (KR setting, lr 0.1) | HM: 1.94 -> 1.97 | No clear effect |
| replace erasure with remapping (with PO) (ablation) | Erase · CIFAR-10 (KR setting, lr 0.1) | HM: 1.97 -> 66.89 | Improved |
| replace single-expert remapping with multi-expert stochastic remapping (ablation) | Remap · CIFAR-10 (KR setting, lr 0.1) | HM: 66.89 -> 92.37 | Improved |
| replace single-expert remapping with multi-expert stochastic remapping (ablation) | Remap · CIFAR-100 (KR setting, lr 0.1) | HM: 62.76 -> 99.89 | Improved |
| replace single-expert remapping with multi-expert stochastic remapping (ablation) | Remap · Tiny-ImageNet (KR setting, lr 0.1) | HM: 59.60 -> 97.99 | Improved |
| remap forget class 4 to class 2 instead of class 0 (ablation) | Remap · CIFAR-10 (KR setting, lr 0.1) | HM: 66.89 -> 34.06 | Hurt |
| use trained conditional router with load balancing (P-T-B) (ablation) | R-ESC · CIFAR-10 (KR setting, N=8, lr 0.1) | HM_t: 85.24 -> 91.79 | Improved |
| use trained conditional router with load balancing (P-T-B) (ablation) | R-ESC · CIFAR-100 (KR setting, N=80, lr 0.1) | HM: 99.97 -> 97.34 | Hurt |
| apply remapping instead of erasure at third-last layer (ablation) | Erase · CIFAR-10 (third-last layer) | HM: 69.40 -> 84.76 | Improved |
Closest papers by meaning
- ESC: Erasing Space Concept for Knowledge Deletion (CVPR 2025 · similarity 0.88)
- Obliviator Reveals the Cost of Nonlinear Guardedness in Concept Erasure (NeurIPS 2025 · similarity 0.87)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.87)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.86)
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.86)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (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.
