Research map / Continual learning / Class-incremental
Selective Amnesia using Contrastive Subnet Erasure for Class Level Unlearning in Vision Models
Vishal Pramanik, Maisha Maliha, Susmit Jha, Alvaro Velasquez, Olivera Kotevska, Sumit Kumar Jha
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 a non-target set with semantic overlap (bird, ship) instead of none (ablation) | CSE · CIFAR-10 forget airplane | retain-test accuracy: 0.76 -> 0.93 | Improved |
| use semantically similar non-target classes instead of dissimilar ones (ablation) | CSE · CIFAR-10 forget airplane | retain-test accuracy: 0.81 -> 0.93 | Improved |
| raise non-target samples from 5% to 10% per category (ablation) | CSE · CIFAR-10 forget airplane | retain-test accuracy: 0.85 -> 0.93 | Improved |
| raise non-target samples from 10% to 15% per category (ablation) | CSE · CIFAR-10 forget airplane | retain-test accuracy: 0.93 -> 0.94 | No clear effect |
| training-free contrastive channel attenuation instead of ESC global projection | CSE · CIFAR-10 single-class, ResNet-18 | forget-test accuracy: .10 -> .01 | Improved |
| contrastive channel attenuation instead of DELETE fine-tuning | CSE · ImageNet single-class, ResNet-18 | H-Mean: .65 -> .73 | Improved |
| increase forgotten classes from 2 to 5 | CSE · CIFAR-100 multi-class | forget-test accuracy: .02 -> .04 | No clear effect |
| increase forgotten classes from 2 to 5 | ESC · CIFAR-100 multi-class | H-Mean: .80 -> .75 | Hurt |
| apply CSE to a face model to forget one LFW identity | CSE · LFW, ResNet-18 | target identity verification accuracy: 0.91 -> 0.04 | Improved |
| apply CSE to a face model to forget one LFW identity | CSE · LFW, ResNet-18 | non-target identity accuracy: 0.88 -> 0.86 | No clear effect |
Closest papers by meaning
- Can Vision Models Truly Forget? Mirage: Representation-Level Certification of Visual Unlearning (ECCV 2026 · similarity 0.90)
- POUR: A Provably Optimal Method for Unlearning Representation via Neural Collapse (CVPR 2026 · similarity 0.89)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.88)
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.88)
- Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models (ICLR 2024 · similarity 0.88)
- Black-Box Forgetting (NeurIPS 2024 · 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.
