Research map / Vision-language and multimodal / Visual question answering
Hierarchical Graph Attention Network for Few-Shot Visual-Semantic Learning
Chengxiang Yin, Kun Wu, Zhengping Che, Bo Jiang, Zhiyuan Xu, Jian Tang
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 visual-specific relationship GNN (ablation) | HGAT · Visual Genome-QA | accuracy: 76.10 -> 77.55; 82.14 -> 84.01; 63.99 -> 63.77; 66.30 -> 69.26 | Mixed |
| add semantic-specific relationship GNN (ablation) | HGAT · Visual Genome-QA | accuracy: 76.10 -> 78.14; 82.14 -> 83.88; 63.99 -> 64.23; 66.30 -> 68.61 | Improved |
| add both visual- and semantic-specific GNNs (ablation) | HGAT · Visual Genome-QA | accuracy: 76.10 -> 78.86; 82.14 -> 84.55; 63.99 -> 65.21; 66.30 -> 70.06 | Improved |
| add attention-based co-learning between modalities (ablation) | HGAT · Visual Genome-QA | accuracy: 78.86 -> 79.56; 84.55 -> 86.10; 65.21 -> 66.62; 70.06 -> 72.13 | Improved |
| replace fully-connected label prediction with Stage-2 relation-aware GNNs (ablation) | HGAT · Visual Genome-QA | accuracy: 77.47 -> 79.56; 83.26 -> 86.10; 64.90 -> 66.62; 70.03 -> 72.13 | Improved |
| semi-supervised use of unlabeled support samples (40% labeled) | HGAT · Toronto COCO-QA | 5-way 5-shot accuracy: 66.09 -> 67.16 | Improved |
| semi-supervised use of unlabeled support samples (60% labeled) | HGAT · Toronto COCO-QA | 5-way 5-shot accuracy: 69.83 -> 70.78 | Improved |
| semi-supervised use of unlabeled support samples (80% labeled) | HGAT · Toronto COCO-QA | 5-way 5-shot accuracy: 73.12 -> 73.95 | Improved |
| unlabeled support samples drawn from other classes (distractors) | HGAT · Toronto COCO-QA | 5-way 5-shot accuracy: 66.09 -> 64.25; 69.83 -> 68.94; 73.12 -> 73.01 | Hurt |
| HGAT instead of EGNN in few-shot VQA | EGNN · Visual Genome-QA | accuracy: 77.67 -> 79.56; 83.26 -> 86.10; 64.07 -> 66.62; 70.87 -> 72.13 | Improved |
Closest papers by meaning
- Multi-Level Attention Networks for Visual Question Answering (CVPR 2017 · similarity 0.90)
- Multi-Modal Graph Neural Network for Joint Reasoning on Vision and Scene Text (CVPR 2020 · similarity 0.89)
- Hypergraph Attention Networks for Multimodal Learning (CVPR 2020 · similarity 0.89)
- Deep Modular Co-Attention Networks for Visual Question Answering (CVPR 2019 · similarity 0.89)
- A Dual Attention Network with Semantic Embedding for Few-Shot Learning (AAAI 2019 · similarity 0.89)
- Hierarchical Question-Image Co-Attention for Visual Question Answering (NeurIPS 2016 · 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.
