Research map / Continual learning / Class-incremental
Knowledge Swapping via Learning and Unlearning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Knowledge Swapping, a task that simultaneously forgets user-specified classes, retains core knowledge, and learns new classes, and benchmarks it with a LoRA-based Learning-Before-Forgetting schedule with group Lasso regularization.
Problem
Pretrained models need to forget user-specified knowledge, keep essential knowledge, and learn new tasks at the same time; existing continual learning and machine unlearning methods address these only separately.
Approach
Analyzing weight changes shows learning proceeds low-to-high level while forgetting proceeds high-to-low level, so learning new classes first and then forgetting (L→F) yields stable forgetting and retention, whereas forgetting first is undone by later low-level updates.
Key results
- On ImageNet-100 with ViT-B/16, Learning-Before-Forgetting drives forgetting-set accuracy to 0% while learning-set accuracy rises above 90%.
- On object detection (DINO, COCO), L→F reduces forgetting-set mAP from 44.5 to 0.5 (CUB) and 0.6 (Oxford-dog) while retention stays near 55%.
- On semantic segmentation (Mask2Former, ADE20K), L→F drives forgetting-set mIoU to near 0 (e.g., 0.12 on VOC) whereas F→L lets forgotten classes reappear.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Learn before forgetting rather than forget before learning (ablation) | Knowledge Swapping with ViT-B/16 · ImageNet-100; four learning sets, 5- and 10-class settings | Acc_f: 80.8 -> 0; 70.8 -> 0; 77.2 -> 0; 74.8 -> 0; 82 -> 0; 82 -> 0.6; 83 -> 0.2; 82.2 -> 0 | Helped |
| Learn before forgetting rather than forget before learning (ablation) | Knowledge Swapping with Mask2Former · VOC, Oxford-IIIT Pet, COCO, and DeepGlobe Land semantic segmentation | mIoU_f: 49.42 -> 0.12; 53.67 -> 0.08; 54.47 -> 0.71; 60.25 -> 0.4 | Helped |
| Learn before forgetting rather than forget before learning (ablation) | Knowledge Swapping with DINO · COCO object detection; CUB and Oxford Dogs learning sets | mAP_f: 7.7 -> 0.5; 11.1 -> 0.6 | Helped |
| Add a second forgetting phase after the Forget→Learn sequence (ablation) | Knowledge Swapping with Mask2Former · VOC, Oxford-IIIT Pet, COCO, and DeepGlobe Land semantic segmentation | mIoU_f: 49.42 -> 0.15; 53.67 -> 0.33; 54.47 -> 0.39; 60.25 -> 0.27 | Helped |
| Add a learning phase after the Learn→Forget sequence (ablation) | Knowledge Swapping with Mask2Former · VOC, Oxford-IIIT Pet, COCO, and DeepGlobe Land semantic segmentation | mIoU_f: 0.12 -> 45.51; 0.08 -> 53.87; 0.71 -> 50.03; 0.4 -> 57.25 | Hurt |
Surprising findings
- Learning-after-forgetting causes previously forgotten classes to re-emerge (forgetting-set accuracy/mIoU rises back), because low-level updates break the forgetting in high-level layers.
Limitations
- Evaluated on a limited set of class-level forgetting/learning splits (e.g., 5 or 10 classes) and only on a few datasets.
- Sparse regularization and forgetting boundary hyperparameters (alpha, beta, BND) are set per task by hand.
- Efficiency benefits are argued but not quantified.
Experimental setup
Datasets: ImageNet-100 CUB-200-2011 Oxford-IIIT Pet RESISC45 PlantVillage Stanford Dogs COCO Pascal VOC DeepGlobe Land ADE20K
Benchmarks: ImageNet-100 classification splits (CUB, RESISC45, Oxford-pet, PlantVillage) COCO object detection (CUB, Stanford Dogs) ADE20K semantic segmentation (VOC, Oxford-pet, COCO, DeepGlobe Land)
Metrics: accuracy mAP mIoU
Baselines: Learning After Forgetting (F→L) Forgetting-first ordering
Builds on: LoRA Lasso regularization machine unlearning continual learning
Closest papers by meaning
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.85)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.85)
- Fine-Grained Knowledge Selection and Restoration for Non-exemplar Class Incremental Learning (AAAI 2024 · similarity 0.84)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.84)
- Large Scale Knowledge Washing (ICLR 2025 · similarity 0.84)
- Knowledge Graph Enhanced Generative Multi-modal Models for Class-Incremental Learning (NeurIPS 2025 · similarity 0.84)
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.