AI Paper Atlas

Research map / Continual learning / Class-incremental

Overcoming Catastrophic Forgetting in Incremental Few-Shot Learning by Finding Flat Minima

Guangyuan SHI, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, Xiao-Ming Wu

NeurIPS 2021 · Continual learning › Class-incremental · approach: Few-shot class-incremental prototypes · 14 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 flat-minima training and clamped fine-tuning (F2M) over frozen-base Baselineincremental few-shot learner · CIFAR-100, session 9accuracy (%): 43.83 -> 44.67Improved
F2M over frozen-base Baselineincremental few-shot learner · miniImageNet, session 9accuracy (%): 43.88 -> 44.65Improved
F2M over frozen-base Baselineincremental few-shot learner · CUB-200-2011, session 11accuracy (%): 58.12 -> 60.26Improved
freeze embedding and never adapt (Baseline) versus complex IFL/IL methodsincremental few-shot learner · CIFAR-100, session 9accuracy (%): 40.08 (joint) / 42.22 (FSLL*) -> 43.83Improved
add parameter clamping to the flat region during incremental sessions (ablation)F2M · CIFAR-100, session 9accuracy (%): 36.01 -> 44.67Improved
add prototype fixing term to base objective (ablation)F2M · CIFAR-100, session 9accuracy (%): 44.08 -> 44.67Improved
add prototype normalization (ablation)F2M · CIFAR-100, session 9accuracy (%): 44.43 -> 44.67Improved
add flat-minima noise injection in base training (with clamping only vs full) (ablation)F2M · CIFAR-100, session 9accuracy (%): 42.55 -> 44.67Improved
add parameter clamping to a model without flat-minima search (ablation)Baseline-style training · CIFAR-100, session 9accuracy (%): 06.45 -> 42.55Improved
increase flat region bound b from 0.02 to 0.08 (ablation)F2M · CIFAR-100, session 9accuracy (%): 44.67 -> 43.04Hurt
decrease flat region bound b from 0.01 to 0.0025 (ablation)F2M · CIFAR-100, session 9accuracy (%): 44.58 -> 44.16Hurt
noise-based flat minima search versus standard SGD training (ablation)base training · CIFAR-100 flatness on training setflatness indicator I (lower is flatter): 0.2993 -> 0.0506Improved

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