Research map / Vision-language and multimodal / CLIP and zero-shot
Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models
Yujin Jo, Taesup Kim
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 |
|---|---|---|---|
| make null-space bases U_n and V_n trainable instead of freezing them (ablation) | NuSA-CL · MTIL full-shot | Last: 82.79 -> 77.32 | Hurt |
| adapt both text and vision encoders instead of one (ablation) | NuSA-CL · MTIL full-shot | Avg.: 72.62 / 70.49 -> 75.08 | Improved |
| update Tail (null-like) directions instead of Top or Random (ablation) | NuSA-CL · MTIL | forgetting: 4.44% / 4.57% -> 2.57% | Improved |
| increase update rank from 32 to 256 (ablation) | NuSA-CL · MTIL | forgetting and target accuracy: forgetting 1.46% -> 5.11% | Mixed |
| raise energy cutoff rho to 0.999 (ablation) | NuSA-CL · MTIL full-shot | Last: 82.79 -> 79.16 | Hurt |
| vary energy cutoff rho between 0.80 and 0.99 (ablation) | NuSA-CL · MTIL full-shot | Avg.: 74.87 to 75.08 | No clear effect |
| use data-agnostic SVD null space instead of InfLoRA gradient projection | NuSA-CL · MTIL | init time and Avg.: ~81 min -> <1 min; Avg 74.2 -> 75.1 | Improved |
| use null-space LoRA instead of standard LoRA | CLIP continual learning · MTIL full-shot | Avg.: 70.1 -> 75.1 | Improved |
| initialize LoRA in minor singular components (MiLoRA) without persistent constraint | LoRA · MTIL full-shot | Avg.: 70.1 -> 68.7 | No clear effect |
| NuSA-CL vs storage-based MoE-Adapters | NuSA-CL · MTIL full-shot | Avg. / efficiency: Avg 75.1 vs 76.7; 1.5M vs 59.8M params | Mixed |
Closest papers by meaning
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.93)
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · similarity 0.90)
- Dual Memory Networks: A Versatile Adaptation Approach for Vision-Language Models (CVPR 2024 · similarity 0.90)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.89)
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.
