Research map / Continual learning / Class-incremental
End-to-End Incremental Learning
Francisco M. Castro, Manuel J. Marin-Jimenez, Nicolas Guil, Cordelia Schmid, Karteek Alahari
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 data augmentation to the base model (ablation) | Our-CNN-Base · CIFAR-100 ablation, fixed memory (20 classes/step) | accuracy: 50.1 -> 57.2 | Improved |
| add balanced fine-tuning to the base model (ablation) | Our-CNN-Base · CIFAR-100 ablation, fixed memory (20 classes/step) | accuracy: 50.1 -> 57.1 | Improved |
| add both augmentation and balanced fine-tuning (ablation) | Our-CNN-Base · CIFAR-100 ablation, fixed memory (10 classes/step) | accuracy: 53.7 -> 64.0 | Improved |
| add balanced fine-tuning to augmentation-only model (ablation) | Our-CNN-DA · CIFAR-100 ablation, fixed memory (5 classes/step) | accuracy: 59.2 -> 63.8 | Improved |
| add data augmentation to model with balanced fine-tuning (ablation) | Our-CNN-BF · CIFAR-100 ablation, fixed memory (5 classes/step) | accuracy: 57.9 -> 63.8 | Improved |
| use herding selection instead of random selection for memory (ablation) | Representative memory · CIFAR-100 (10 classes/step) | accuracy: 63.1 -> 63.6 | No clear effect |
| use herding selection instead of histogram selection (ablation) | Representative memory · CIFAR-100 (10 classes/step) | accuracy: 59.1 -> 63.6 | Improved |
| use end-to-end cross-distilled model instead of iCaRL | Incremental learner · CIFAR-100 fixed memory (2 classes/step) | average incremental accuracy: 54.1 -> 63.8 | Improved |
| use end-to-end model instead of iCaRL with many classes per step | Incremental learner · CIFAR-100 fixed memory (50 classes/step) | average incremental accuracy: 61.8 -> 60.8 | Hurt |
| use end-to-end model instead of iCaRL | Incremental learner · ImageNet fixed memory (100 classes/step) | accuracy: 62.5 -> 69.4 | Improved |
| increase exemplars per old class from 50 to 100 | Our-CNN · CIFAR-100 fixed samples (5 classes/step) | accuracy: 62.4 -> 68.6 | Improved |
| use end-to-end model instead of GEM with small memory | Incremental learner · CIFAR-100 task known a priori (memory 200) | final accuracy: 48.0 -> 77.9 | Improved |
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 7 of 8 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 by Knowledge Distillation With Adaptive Feature Consolidation (CVPR 2022 · similarity 0.91)
- Large Scale Incremental Learning (CVPR 2019 · similarity 0.90)
- Class-Incremental Learning: A Survey (TPAMI 2024 · similarity 0.90)
- Incremental Learning in Online Scenario (CVPR 2020 · similarity 0.90)
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.89)
- Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning (ICCV 2021 · 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.
