Research map / Continual learning / Online and replay-based
Mitigate Modality-Asymmetric Forgetting via Stabilizing Visual Representations in CLIP-Based Class-Incremental Learning
Yuanhong Zhang, Yanan Chen, Xin Zhang, Zhaoyang Wang, Weizhan Zhang, Muyao Yuan, Hongjin Niu, LAN MA, Yuan Gao, Joey Tianyi Zhou
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 VRRM to LoRA baseline (ablation) | CLIP + LoRA · CIFAR100 B10_Inc10 | Avg accuracy (%): 84.81 -> 86.90 | Improved |
| add VRRM to LoRA baseline (ablation) | CLIP + LoRA · ImageNet-R B20_Inc20 | Avg accuracy (%): 84.01 -> 87.24 | Improved |
| add structural distillation loss L_distill (ablation) | CLIP + LoRA · CIFAR100 B10_Inc10 | Last accuracy (%): 77.33 -> 78.93 | Improved |
| add L_distill on top of VRRM (ablation) | VRRM · CIFAR100 B10_Inc10 | Last accuracy (%): 79.35 -> 80.40 | Improved |
| add VRRM on top of L_distill (ablation) | L_distill · ImageNet-R B20_Inc20 | Last accuracy (%): 82.32 -> 82.83 | Improved |
| use second-order self-similarity attention instead of QK or KK attention (ablation) | VRRM · CIFAR100 | Avg accuracy (%) | Improved |
| text-aware layer weighting instead of last-layer or uniform average (ablation) | VRRM · CIFAR100 | Avg accuracy (%) | Improved |
| replace KL/MSE/CKC distillation with matrix-entropy structural distillation (ablation) | L_distill · CIFAR100 | Avg accuracy (%) | Improved |
| integrate VRRM into other distillation losses (ablation) | KL/MSE/CKC distillation · CIFAR100 | Avg accuracy (%) | Improved |
| vary distillation weight lambda and fusion coefficient beta away from 0.5 and 0.1 (ablation) | VAST · CIFAR100 B10_Inc10 | Avg accuracy (%): 87.40 at peak, 86.76-87.33 elsewhere (grid in Fig. 7a) | Mixed |
| choose self-similarity layer M (ablation) | VRRM · CIFAR100 B10_Inc10 | Avg accuracy (%) | No clear effect |
| apply VAST vs LoRA baseline for visual feature separability | CLIP + LoRA · CIFAR100 | FDR after Task 1: 2.08 -> 2.97 | Improved |
Closest papers by meaning
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.90)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.88)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.88)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.87)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 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.
