Research map / Continual learning / Online and replay-based
REMIND Your Neural Network to Prevent Catastrophic Forgetting
Tyler L. Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, Christopher Kanan
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add product quantization instead of k-means (32 codebooks vs 1) (ablation) | REMIND · ImageNet streaming | Omega_all: 0.778 -> 0.855 | Improved |
| add augmentation (random crop and manifold mixup) during replay (ablation) | REMIND · ImageNet streaming | Omega_all: 0.818 -> 0.855 | Improved |
| store entire dataset (100% buffer) instead of 959665 examples (ablation) | REMIND · ImageNet streaming | Omega_all: 0.855 -> 0.856 | No clear effect |
| use real features instead of quantized features (ablation) | REMIND · ImageNet streaming | Omega_all: 0.855 -> 0.868 | No clear effect |
| use real features instead of quantized features (memory) (ablation) | REMIND · ImageNet streaming | memory GB: 1.51 -> 24.08 | Hurt |
| use REMIND instead of SLDA | ResNet-18 · ImageNet streaming | Omega_all: 0.752 -> 0.855 | Improved |
| use REMIND instead of regularization method A-GEM (with task labels) | CORe50 cls iid · CORe50 | Omega_all: 0.925 -> 0.995 | Improved |
| use REMIND instead of A-GEM (no task labels) | CORe50 cls iid · CORe50 | Omega_all: 0.417 -> 0.978 | Improved |
| use REMIND instead of SLDA for streaming VQA | SAN · TDIUC q-type | Omega_all: 0.644 -> 0.919 | Improved |
| use REMIND instead of SLDA for streaming VQA | MAC · CLEVR q-type | Omega_all: 0.496 -> 0.985 | Improved |
| reduce VQA buffer from 100% to 25% (ablation) | REMIND · TDIUC q-type | Omega_all: 0.931 -> 0.936 | No clear effect |
| reduce VQA buffer from 100% to 25% (ablation) | REMIND · CLEVR q-type | Omega_all: 0.985 -> 0.960 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.90)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.89)
- Understanding the Forgetting of (Replay-based) Continual Learning via Feature Learning: Angle Matters (ICML 2025 · similarity 0.89)
- Memory Replay GANs: Learning to Generate New Categories without Forgetting (NeurIPS 2018 · similarity 0.89)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.88)
- Towards Understanding Catastrophic Forgetting in Two-layer Convolutional Neural Networks (ICML 2025 · similarity 0.87)
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.
