Research map / Vision-language and multimodal / Human and embodied modeling
Dialog without Dialog Data: Learning Visual Dialog Agents from VQA Data
Michael Cogswell, Jiasen Lu, Rishabh Jain, Stefan Lee, Devi Parikh, Dhruv Batra
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 |
|---|---|---|---|
| Fine-tune for the image-guessing task after VQA pre-training | Ours vs Zero-shot Transfer · VQA 2 Contrast 1 Round | accuracy: 0.73 -> 0.82 | Improved |
| Fine-tune for multi-round and random-pool image guessing | Ours vs Zero-shot Transfer · VQA 2 Contrast 5 Rounds; VQA 2 Random 5 Rounds; VQA 9 Random 9 Rounds | accuracy: 0.67 -> 0.87; 0.64 -> 0.95; 0.18 -> 0.53 | Improved |
| Adapt the pretrained agent using the proposed method | Ours vs Typical Transfer · AWA 9 Random 9 Rounds; CUB 9 Random 9 Rounds | accuracy: 0.48 -> 0.74; 0.38 -> 0.74 | Improved |
| Use a discrete rather than continuous latent intention (ablation) | Ours, 10-setting average · 10-setting average | accuracy: 0.75 -> 0.81 | Improved |
| Use a discrete rather than continuous latent intention (ablation) | Ours, 10-setting average · 10-setting average | relevance: 0.66 -> 0.89 | Improved |
| Use ELBO pre-training rather than MLE (ablation) | Ours, 10-setting average · 10-setting average | perplexity: 4.27 -> 2.45 | Improved |
| Use ELBO pre-training rather than MLE (ablation) | Ours, 10-setting average · 10-setting average | diversity: 4.33 -> 0.23 | Improved |
| Fine-tune the speaker instead of keeping it fixed (ablation) | Ours, 10-setting average · 10-setting average | perplexity: 2.57 -> 2.54 | No clear effect |
| Fine-tune the speaker instead of keeping it fixed (ablation) | Ours, 10-setting average · 10-setting average | accuracy: 0.81 -> 0.82 | No clear effect |
| Add Stage 2.A dialog-tracking training after Stage 1 (ablation) | Ours, 10-setting average · 10-setting average | accuracy: 0.72 -> 0.80 | Improved |
| Add Stage 2.B planner fine-tuning after Stage 2.A (ablation) | Ours, 10-setting average · 10-setting average | accuracy: 0.80 -> 0.81 | Improved |
| Allow questions to adapt to task feedback rather than use a parallel frozen question source (ablation) | Ours, 10-setting average · 10-setting average | accuracy: 0.78 -> 0.81 | Improved |
Closest papers by meaning
- Visual Dialog (CVPR 2017 · similarity 0.89)
- Large-scale Pretraining for Visual Dialog: A Simple State-of-the-Art Baseline (ECCV 2020 · similarity 0.88)
- Structure-Aware Multimodal Sequential Learning for Visual Dialog (AAAI 2024 · similarity 0.87)
- TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling (AAAI 2024 · similarity 0.86)
- MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and Instruction-Tuning Dataset for LVLMs (NeurIPS 2024 · similarity 0.86)
- DRESS: Instructing Large Vision-Language Models to Align and Interact with Humans via Natural Language Feedback (CVPR 2024 · similarity 0.86)
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.
