Research map / Continual learning / Class-incremental
Mimicking the Oracle: An Initial Phase Decorrelation Approach for Class Incremental Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 CwD to LUCIR at the initial phase | LUCIR · CIFAR100 (B=50), S=2 | average incremental accuracy (%): 52.96±0.25 -> 56.81±0.21 | Helped |
| Add CwD to PODNet at the initial phase | PODNet · ImageNet100 (B=50), S=10 | average incremental accuracy (%): 75.71±0.37 -> 76.91±0.10 | Helped |
| Add CwD to AANet at the initial phase | AANet · CIFAR100 (B=50), S=10 | average incremental accuracy (%): 69.79±0.21 -> 70.30±0.37 | Helped |
| Add CwD to LUCIR at the initial phase | LUCIR · ImageNet (B=100), S=50 | average incremental accuracy (%): 52.75±0.18 -> 53.37±0.22 | Helped |
| Mimic oracle representations during initial-phase training | LUCIR · ImageNet100; initial 10 or 50 classes, then 10 classes/phase | accuracy | Helped |
| Apply oracle-representation mimicry at the initial phase | LUCIR · ImageNet100, initial phase | accuracy | No clear effect |
| Increase initial classes from 10 to 40 while using CwD (ablation) | LUCIR+CwD · ImageNet100, S=10 and S=5 | average incremental accuracy (%): +0.89; +2.02; +1.45; +2.52 | Mixed |
| Add CwD with only 10 initial classes (ablation) | LUCIR · ImageNet100, B=10, S=10 | average incremental accuracy (%): 57.01±0.14 -> 57.90±0.07; +0.89 | Helped |
| Add CwD with 5 exemplars per class (ablation) | LUCIR · ImageNet100, B=50, S=5 | average incremental accuracy (%): 60.41±0.77 -> 62.58±0.53; +2.17 | Helped |
| Increase the CwD coefficient η (ablation) | LUCIR+CwD · ImageNet100, B=50, S=10 or S=5 | average incremental accuracy (%) | Mixed |
| Use an excessively large CwD coefficient η (ablation) | LUCIR+CwD · ImageNet100, B=50, S=10 or S=5 | average incremental accuracy (%) | Hurt |
| Increase η in CwD (ablation) | Initial-phase representations · ImageNet100, 10/25/50 initial classes | top-k variance proportion α(c)k | Helped |
Closest papers by meaning
- Reducing Class-wise Confusion for Incremental Learning with Disentangled Manifolds (CVPR 2025 · similarity 0.88)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.88)
- Prediction Error-based Classification for Class-Incremental Learning (ICLR 2024 · similarity 0.88)
- Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping (CVPR 2025 · similarity 0.88)
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · similarity 0.88)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · 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.