AI Paper Atlas

Research map / Continual learning / Online and replay-based

Flattening Sharpness for Dynamic Gradient Projection Memory Benefits Continual Learning

Danruo DENG, Guangyong Chen, Jianye HAO, Qiong Wang, Pheng-Ann Heng

NeurIPS 2021 · Continual learning › Online and replay-based · approach: Gradient projection and plasticity preservation · 23 citations (OpenAlex)

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 flattening sharpness (adversarial weight perturbation) to GPM with dynamic bases (ablation)DGPM · CIFAR-100 SplitACC (%): 73.78 -> 74.33Improved
add flattening sharpness to DGPM (ablation)DGPM · CIFAR-100 SuperclassACC (%): 56.78 -> 58.81Improved
add dynamic soft weights on bases to FS-GPM (ablation)FS-GPM · CIFAR-100 SplitACC (%): 73.96 -> 74.33Improved
add dynamic soft weights on bases to FS-GPM (ablation)FS-GPM · CIFAR-100 SuperclassACC (%): 58.61 -> 58.81Improved
use adversarial weight perturbation instead of look-ahead descent perturbation (ablation)FS-DGPM · CIFAR-100 SplitACC (%): 73.74 -> 74.33Improved
add flattening sharpness to DGPM to reduce forgetting (ablation)DGPM · CIFAR-100 SplitBWT (%): -3.67 -> -2.71Improved
add flattening sharpness to vanilla ERER · PMNISTACC (%): 86.16 -> 90.44Improved
use FS-DGPM instead of GPMGPM · CIFAR-100 SplitBWT (%): -1.17 -> -2.71Hurt
use FS-DGPM instead of GPMGPM · TinyImageNetACC (%): 67.39 -> 70.41Improved
use FS-DGPM instead of GPMGPM · PMNISTACC (%): 74.54 -> 76.96Improved

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