Research map / Continual learning / Class-incremental
IL2M: Class Incremental Learning With Dual Memory
Eden Belouadah, Adrian Popescu
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 score rectification with class-level means (IL2M 1) (ablation) | FT · ILSVRC, Z=10, K=5k | top-1 average accuracy: 32.23 -> 42.20 | Improved |
| add score rectification with class-level means (IL2M 1) (ablation) | FT · ILSVRC, Z=10, K=20k | top-1 average accuracy: 51.13 -> 53.45 | Improved |
| add model-level mean normalization alone (IL2M 2) (ablation) | FT · ILSVRC, Z=10, K=5k | top-1 average accuracy: 32.23 -> 31.74 | Hurt |
| add model-level mean normalization alone (IL2M 2) (ablation) | FT · ILSVRC, Z=10, K=20k | top-1 average accuracy: 51.13 -> 51.94 | Improved |
| add model-level means on top of class-level means (IL2M 1+2) (ablation) | IL2M 1 · ILSVRC, Z=10, K=10k | top-1 average accuracy: 47.64 -> 49.57 | Improved |
| add restriction to rectify only images first predicted as new class (ablation) | IL2M 1+2 · ILSVRC, Z=10, K=10k | top-1 average accuracy: 49.57 -> 50.82 | Improved |
| ablate distillation loss from iCaRL (FT NEM) (ablation) | iCaRL · ILSVRC, Z=10, K=10k | top-1 average accuracy: 33.6 -> 49.1 | Improved |
| ablate distillation loss from iCaRL (FT NEM) (ablation) | iCaRL · VGGFace2, Z=10, K=10k | top-1 average accuracy: 65.3 -> 87.6 | Improved |
| add distillation loss to fine tuning with no memory (ablation) | FT · ILSVRC, Z=10, K=0 | top-1 average accuracy: 18.3 -> 20.8 | Improved |
| use NEM classifier instead of network classification layer (FT NEM) | FT · Landmarks, Z=10, K=10k | top-1 average accuracy: 90.1 -> 88.5 | Hurt |
| add balanced fine-tuning step (FT BAL) | FT · ILSVRC, Z=10, K=10k | top-1 average accuracy: 42.3 -> 47.0 | Improved |
| add IL2M score rectification | FT · ILSVRC, Z=10, K=20k and K=5k | top-1 average accuracy: +5.3 and +11.9 points | Improved |
Later papers that compared with this method
2 later accepted papers report il2m as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Hierarchical Augmentation and Distillation for Class Incremental Audio-Visual Video Recognition (TPAMI 2024 · better on 9 of 9 shared results)
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · better on 1 of 1 shared result)
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 5 of 6 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
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.92)
- A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning (ICLR 2023 · similarity 0.91)
- Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning (ICCV 2021 · similarity 0.91)
- Learning Without Memorizing (CVPR 2019 · similarity 0.91)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.90)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.89)
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.
