Research map / Continual learning / Class-incremental
Versatile Incremental Learning: Towards Class and Domain-Agnostic Incremental Learning
Min-Yeong Park, Jae-Ho Lee, Gyeong-Moon Park
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 |
|---|---|---|---|
| replace constant threshold 0.5 with a dynamic threshold in the incremental classifier (ablation) | ICON · iDigits VIL | average accuracy: 74.79 -> 75.11 | Improved |
| place adapters in the first 5 layers instead of the last 5 (ablation) | ICON · iDigits VIL | average accuracy: 52.94 -> 75.11 | Improved |
| place adapters in all layers instead of the first 5 (ablation) | ICON · DomainNet VIL | average accuracy: 53.37 -> 55.14 on DomainNet; 75.11 -> 62.57 on iDigits | Mixed |
| save more than one shift per task in the shift pool (ablation) | ICON CAST · iDigits VIL | average accuracy: 75.11 -> 68.45 (10 shifts) | Hurt |
| use more shifts per task on DomainNet (ablation) | ICON CAST · DomainNet VIL | average accuracy: 53.37 -> 53.61 (10 shifts) | No clear effect |
| increase number of tasks in the VIL sequence (iDigits 8 to 20 tasks) | LAE and ICON · iDigits VIL | average accuracy: ICON 81.59 -> 75.11; LAE 79.67 -> 59.34 | Hurt |
| use ICON instead of LAE across numbers of tasks | ICON · CORe50 VIL 40 tasks | average accuracy: 77.11 -> 83.18 | Improved |
| apply ICON adapters to a ResNet-152 CNN instead of LAE | ICON · iDigits VIL (ResNet-152) | average accuracy: 30.43 -> 37.34 | Improved |
| use ICON instead of LAE on larger backbone ViT-L/16 | ICON · DomainNet VIL (ViT-L/16) | average accuracy: 51.80 -> 54.09 | Improved |
| use too small gamma in incremental classifier threshold (ablation) | ICON · iDigits VIL | average accuracy: gamma 0.5: 72.00 vs gamma 2.0: 75.11 | Hurt |
| increase beta (CAST loss weight) on CORe50 (ablation) | ICON · CORe50 VIL | average accuracy: 79.85 -> 83.18 (beta 0.01 -> 0.05) | Improved |
Closest papers by meaning
- Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need (CVPR 2025 · similarity 0.90)
- General Incremental Learning With Domain-Aware Categorical Representations (CVPR 2022 · similarity 0.89)
- Attraction Diminishing and Distributing for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.85)
- Make Domain Shift a Catastrophic Forgetting Alleviator in Class-Incremental Learning (AAAI 2025 · similarity 0.85)
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · 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.
