AI Paper Atlas

Research map / Continual learning / Online and replay-based

Online Continual Learning from Imbalanced Data

Aristotelis Chrysakis, Marie-Francine Moens

ICML 2020 · Continual learning › Online and replay-based · approach: Plug-and-play online training enhancements · 51 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
Replace reservoir sampling / GSS memory population with class-balancing reservoir sampling (CBRS)online continual learning memory population · MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100 (imbalanced streams)test accuracy: MNIST 83.3 vs 74.2 (GSS); CIFAR-100 40.2 vs 28.1 (reservoir)Improved
Use uniform instead of weighted replay with CBRS (ablation)CBRS · MNIST m=1000 and CIFAR-10 m=1000test accuracy: MNIST 86.0 vs 86.1; CIFAR-10 74.4 vs 73.7 (m=1000)No clear effect
Use weighted instead of uniform replay with reservoir sampling (ablation)RESERVOIR · MNIST m=1000 and CIFAR-10 m=1000test accuracy: MNIST 76.3 vs 71.3; CIFAR-10 56.7 vs 54.2Improved
Use weighted instead of uniform replay with CBRS at m=5000 (ablation)CBRS · MNIST and CIFAR-10 m=5000test accuracy: MNIST 90.3 vs 85.6; CIFAR-10 78.3 vs 76.8Mixed
Increase memory size from 100 to 5000 with CBRSCBRS · Fashion-MNISTtest accuracy: 66.5 at m=100 to 77.1 at m=5000Improved
Add CBRS memory overhead (index structure per class) vs GSS gradient storagememory population · CIFAR-100 computewall-clock time per batch and memory overhead: CBRS 0.03% overhead vs GSS 3639.86%; CBRS 3671.1 ms vs GSS 364.9 ms (time per batch)Improved
Replace reservoir sampling with GSS-Greedy memory populationonline continual learning memory population · CIFAR-100 (ResNet-18, m=1000)test accuracy: 23.0 vs 28.1 for reservoirHurt
Use weighted replay (inverse-class-count sampling) instead of uniform replay for all methods (ablation)online CL replay · MNIST, CIFAR-10 at m=1000 and m=5000test accuracy: Gains for RESERVOIR/GSS; none for CBRS at m=1000 (86.0 vs 86.1)Mixed

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