AI Paper Atlas

Research map / Continual learning / Class-incremental

Forgetting Whenever You Want: A Decentralized Continual Learning Framework with On-Demand Unlearning

Xiao Zhang, Zengzhe Chen, Mingyi Li, Jing Qiao, Fuzhen Zhuang, YUAN YUAN, Dongxiao Yu

ICML 2026 · Continual learning › Class-incremental · approach: Class-balanced knowledge retention

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 changedWhereMeasuredEffect
add synthetic prototype-guided data D' to local training (ablation)DCU · RGB Dir(0.5)accuracy r% after cl0:9: 69.60 -> 75.14Improved
add replayed synthetic data D-hat for distillation of old classes (ablation)DCU · RGB Dir(0.5)accuracy r% after cl10:19: 42.51 -> 59.44Improved
add synthetic unlearning data D-hat_f guided by prototypes (ablation)DCU · RGB Dir(0.5)unlearned-class accuracy u% after ul0: 83.80 -> 0.00Improved
add prototype-based synthetic data for unlearning historical classes versus FedAF/AU requiring target dataEWC/AU, LwF/AU, FedAF · RGB Dir(0.5)remaining accuracy r% after ul0: 41.10 -> 58.77 (cl10:19 stage vs DCU ul0)Improved
use gradient ascent on real stored data (QDsga) as sequential tasks arriveEWC/QDsga, LWF/QDsga · RGB Dir(0.5)remaining accuracy r% after ul20: LWF/QDsga 5.29Hurt
use a reservoir-sampling buffer (CLMUL) in decentralized non-IID settingCLMUL · RGB Dir(0.5)remaining accuracy r% after ul20: 33.11Hurt
use diffusion prototypes instead of server-trained generator under stronger non-IIDDCU vs Lander/AU · RGB Dir(0.1)remaining accuracy r% after ul20: 34.11 -> 40.04Improved
compare DCU with Lander/AU at mild non-IIDDCU vs Lander/AU · RGB Dir(0.5)remaining accuracy r% after ul20: 50.67 vs 50.96No clear effect
standard continual learning methods without unlearning module (EWC, LwF, Target, Lander)continual learning baselines · RGB and Grayscaleforgetting gap F: e.g. -0.06 to 0.13Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy