Research map / Continual learning / Class-incremental
R-DFCIL: Relation-Guided Representation Learning for Data-Free Class Incremental Learning
Qiankun Gao, Chen Zhao, Bernard Ghanem, Jian Zhang
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 relational knowledge distillation (RKD) (ablation) | R-DFCIL · CIFAR100 (equal split) N=20 | last accuracy: 21.63 -> 30.75 | Improved |
| add hard knowledge distillation (HKD) on synthetic data (ablation) | R-DFCIL · CIFAR100 (equal split) N=20 | last accuracy: 5.37 -> 30.75 | Improved |
| add classification head refinement (CHR) (ablation) | R-DFCIL · CIFAR100 (equal split) N=20 | last accuracy: 23.75 -> 30.75 | Improved |
| add adaptive linear transformation functions in RKD (ablation) | R-DFCIL · CIFAR100 (equal split) N=20 | last accuracy: about 2% | Improved |
| replace RKD with HKD on new data (ablation) | R-DFCIL · CIFAR100 N=20 (single run) | last accuracy A_20: 30.92% -> 21.35% | Hurt |
| replace HKD with RKD on synthetic old data (ablation) | R-DFCIL · CIFAR100 N=20 (single run) | last accuracy A_20: 30.92% -> 5.81% | Hurt |
| use RKD instead of newer relational KD (CRCD, CRD) (ablation) | R-DFCIL · CIFAR100 | average accuracy: 5% (CRCD) and 10% (CRD) | Improved |
| raise hard KD weight lambda_hkd to 5.0 (ablation) | R-DFCIL · CIFAR100 N=5 | accuracy | Hurt |
| vary lambda_rkd over 0.05, 0.5, 5.0 (ablation) | R-DFCIL · CIFAR100 N=5 | accuracy | No clear effect |
| use R-DFCIL instead of ABD with more tasks (N=5 to N=20) | ABD · CIFAR100 (equal split) | last accuracy: +3.11 at N=5, +8.46 at N=20 | Improved |
| use R-DFCIL instead of ABD | ABD · Tiny-ImageNet200 (equal split) N=20 | last accuracy: 15.20 -> 24.43 | Improved |
| replace real old data with synthetic old data in replay methods | UCIR/PODNet · CIFAR100 (50-class first task) N=26 | last accuracy: 50.08 -> 9.62 (UCIR); 49.14 -> 20.18 (PODNet) | Hurt |
Later papers that compared with this method
4 later accepted papers report R-DFCIL as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- EFC++: Elastic Feature Consolidation with Prototype Re-balancing for Cold Start Exemplar-free Incremental Learning (IJCV 2026 · better on 2 of 2 shared results)
- Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning (NeurIPS 2025 · better on 6 of 6 shared results)
- DiffClass: Diffusion-Based Class Incremental Learning (ECCV 2024 · better on 2 of 2 shared results)
- Data-Free Class-Incremental Hand Gesture Recognition (ICCV 2023 · better on 4 of 4 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
- Few-Shot Class-Incremental Learning via Class-Aware Bilateral Distillation (CVPR 2023 · similarity 0.88)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.86)
- Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental Learning (ICML 2025 · similarity 0.86)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.85)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.85)
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.
