Research map / Vision-language and multimodal / CLIP and zero-shot
ViLTA: Enhancing Vision-Language Pre-training through Textual Augmentation
Weihan Wang, Zhen Yang, Bin Xu, Juanzi Li, Yankui Sun
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 synthetic hard negatives for ITM (ablation) | ViLTA (MLM+ITM) · VQAv2 | accuracy: 77.34 -> 78.06 | Improved |
| add cross-distillation soft labels for MLM (ablation) | ViLTA (MLM+ITM) · VQAv2 | accuracy: 77.34 -> 77.85 | Improved |
| combine hard negatives and cross-distillation (ablation) | ViLTA · VQAv2 | accuracy: 78.06 -> 78.47 | Improved |
| add ITM task to MLM-only pre-training (ablation) | ViLTA · COCO Captions | BLEU@4: 38.2 -> 38.8 | Improved |
| use dynamically generated negatives instead of in-batch random (ablation) | ViLTA ITM · VQAv2 test-dev | accuracy: 74.08 -> 74.93 | Improved |
| use WordNet antonym negatives instead of random negatives (ablation) | ViLTA ITM · VQAv2 test-dev | accuracy: 74.08 -> 74.24 | No clear effect |
| add WordNet negatives to dynamically generated negatives (ablation) | ViLTA ITM · VQAv2 test-dev | accuracy: 74.93 -> 74.71 | Hurt |
| regenerate negatives dynamically each step instead of once before training (ablation) | ViLTA ITM · VQAv2 test-dev | accuracy: 74.52 -> 74.93 | Improved |
| add language modeling pre-training phase for captioning (ablation) | ViLTA · COCO Captions | BLEU@4: 39.6 -> 40.0 (w/o CA); 40.4 -> 41.0 (w/ CA) | Improved |
| keep cross-attention module on top of the image encoder (ablation) | ViLTA captioning · COCO Captions | CIDEr: 131.5 -> 135.1 (w/ LM) | Improved |
| ViLTA pre-trained with only MLM and ITM (no ITC) on COCO retrieval | ViLTA-BASE · COCO text retrieval | TR@1: 73.3 vs METER 76.2 | Hurt |
| add both techniques to pre-training (ablation) | ViLTA · pre-training cost | GPU-hours per epoch: 34.68 -> 38.43 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VLMo-LARGE — VLMo: Unified Vision-Language Pre-Training with Mixture-of-Modality-Experts (NeurIPS 2022) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unified Vision-Language Pre-Training for Image Captioning and VQA (AAAI 2020 · similarity 0.91)
- SimVLM: Simple Visual Language Model Pretraining with Weak Supervision (ICLR 2022 · similarity 0.91)
- GRIT-VLP: Grouped Mini-Batch Sampling for Efficient Vision and Language Pre-training (ECCV 2022 · similarity 0.90)
- Enhancing Fine-Grained Vision-Language Pretraining with Negative Augmented Samples (AAAI 2025 · similarity 0.90)
- Probing Visual Language Priors in VLMs (ICML 2025 · similarity 0.89)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · 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.
