Research map / Vision-language and multimodal / CLIP and zero-shot
Generative Negative Text Replay for Continual Vision-Language Pretraining
Shipeng Yan, Lanqing Hong, Hang Xu, Jianhua Han, Tinne Tuytelaars, Zhenguo Li, Xuming He
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 knowledge distillation loss (ablation) | IncCLIP · CC2M class incremental (4 steps), zero-shot ImageNet | top-1 accuracy: 16.01 -> 17.67 | Improved |
| add hard negative text generation (ablation) | IncCLIP · CC2M class incremental (4 steps), zero-shot ImageNet | top-1 accuracy: 17.67 -> 18.85 | Improved |
| add hard negative text generation (ablation) | IncCLIP · Flickr30K image-to-text (class incremental, 4 steps) | R@1: 22.8 -> 24.1 | Improved |
| add knowledge distillation loss (ablation) | IncCLIP · Flickr30K image-to-text (class incremental, 4 steps) | R@1: 21.3 -> 22.8 | Improved |
| add hard negative text generation (ablation) | IncCLIP · Flickr30K text-to-image (class incremental, 4 steps) | R@1: 15.10 -> 17.14 | Improved |
| add hard negative text generation (ablation) | IncCLIP · MSCOCO image-to-text (class incremental, 4 steps) | R@1: 11.55 -> 12.38 | Improved |
| add weight norm clipping (ablation) | IncCLIP · CC2M class incremental (4 steps), zero-shot ImageNet | top-1 accuracy | Improved |
| increase memory size from 1% to 50% of dataset (ablation) | IncCLIP · CC2M class incremental (4 steps), zero-shot ImageNet | top-1 accuracy: 12.82 -> 28.29 | Improved |
| increase memory size from 1% to 50% of dataset (ablation) | IncCLIP · Flickr30K image-to-text (class incremental, 4 steps) | R@1: 18.2 -> 31.6 | Improved |
| use IncCLIP instead of Co2L | class incremental 4 steps · CC2M class incremental (4 steps), average of 8 datasets | top-1 accuracy: 16.84 -> 21.57 | Improved |
| use IncCLIP instead of ER for forgetting | continual VLP · CC2M class incremental, forgetting analysis | BWT: -31.93 -> -18.48 | Improved |
| increase number of steps from 4 to 8 | IncCLIP · CC2M class incremental, average of 8 datasets | top-1 accuracy: 21.57 -> 17.24 | Hurt |
Closest papers by meaning
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.91)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.90)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.89)
- Continual Vision-Language Retrieval via Dynamic Knowledge Rectification (AAAI 2024 · similarity 0.89)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.89)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.88)
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.
