Research map / Continual learning / Class-incremental
Exemplar-Free Continual Learning of Vision Transformers via Gated Class-Attention and Cascaded Feature Drift Compensation
Marco Cotogni, Fei Yang, Claudio Cusano, Andrew D. Bagdanov, Joost van de Weijer
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 gated class-attention (ablation) | fine-tuned ViT · CIFAR-100 (10 tasks) | Class-IL accuracy: 11.90 -> 31.35 | Improved |
| add gated class-attention (ablation) | fine-tuned ViT · Tiny-ImageNet (10 tasks) | Class-IL accuracy: 7.98 -> 22.79 | Improved |
| add cascaded feature drift compensation to gated class-attention + PFR (ablation) | GCAB · CIFAR-100 (10 tasks) | Class-IL accuracy: 7.96 -> 35.90 | Improved |
| add cascaded feature drift compensation to gated class-attention + PFR (ablation) | GCAB · Tiny-ImageNet (10 tasks) | Class-IL accuracy: 5.52 -> 26.82 | Improved |
| add backbone regularization (PFR) without drift compensation (ablation) | GCAB · CIFAR-100 (10 tasks) | Class-IL accuracy: 31.35 -> 7.96 | Hurt |
| add feature drift compensation on top of gated attention and PFR versus gated attention alone (ablation) | GCAB · CIFAR-100 (10 tasks) | Class-IL accuracy: 31.35 -> 35.90 | Improved |
| use feature distillation instead of PFR for backbone regularization (ablation) | GCAB · CIFAR-100 (10 tasks) | Class-IL accuracy: 35.90 -> 30.50 | Hurt |
| distill GCAB into single student block (GCAB-Fast, 80% capacity) | GCAB · CIFAR-100 (5 tasks) | Class-IL accuracy: 49.86 -> 48.85 | No clear effect |
| distill GCAB into single student block (GCAB-Fast) | GCAB · ImageNet100 (10 tasks) | Class-IL accuracy: 40.10 -> 36.22 | Hurt |
| use GCAB without exemplars versus LVT with 200 exemplars | GCAB · Tiny-ImageNet (10 tasks) | Class-IL accuracy: 17.34 -> 26.82 | Improved |
Closest papers by meaning
- Meta-Attention for ViT-Backed Continual Learning (CVPR 2022 · similarity 0.91)
- Task-Free Dynamic Sparse Vision Transformer for Continual Learning (AAAI 2024 · similarity 0.89)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · similarity 0.88)
- Visual Prompt Tuning in Null Space for Continual Learning (NeurIPS 2024 · similarity 0.87)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · 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.
