Research map / Vision-language and multimodal / Visual question answering
Interpretable Visual Question Answering by Reasoning on Dependency Trees
Qingxing Cao, Xiaodan Liang, Bailin Li, Liang Lin
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 edge-dependent propagation (ablation) | baseline PTGRN · CLEVR val | overall accuracy: 87.62 -> 89.28 | Improved |
| add GRU gated residual composition to +Edge (ablation) | PTGRN +Edge · CLEVR val | overall accuracy: 89.28 -> 90.44 | Improved |
| add GRU gating when attention map is concatenated into hidden (ablation) | PTGRN +Edge+ConcatAtt · CLEVR val | overall accuracy: 93.36 -> 93.53 | No clear effect |
| process attention map with a separate GRU instead of concatenating with hidden (ablation) | PTGRN +Edge+GRU · CLEVR val | overall accuracy: 93.53 -> 94.42 | Improved |
| use convolutional GRU for attention map encoding to preserve spatial information (ablation) | PTGRN +Edge+GRU+GRUAtt · CLEVR val | overall accuracy: 94.42 -> 95.42 | Improved |
| use Bi-LSTM context word encoding instead of lookup table (ablation) | PTGRN · CLEVR val | overall accuracy: 87.75 -> 95.42 | Improved |
| use ground-truth program layout instead of parsed dependency tree (ablation) | PTGRN · CLEVR val | overall accuracy: 95.42 -> 96.69 | Improved |
| randomly perturb dependency relations (10% to 70%) (ablation) | PTGRN · CLEVR val | overall accuracy: 95.42 -> 86.77 (10%), 40.11 (70%) | Hurt |
| redeveloped modules (shared weights, edge-dependent, gated, ConvGRU attention) versus preliminary ACMN | PTGRN · CLEVR test | overall accuracy: 89.3 -> 95.5 | Improved |
| PTGRN versus MAC on attention accuracy without ground-truth layouts | PTGRN · CLEVR val | attention accuracy: 32.72 -> 64.71 | Improved |
| PTGRN versus FiLM on unseen color combinations | PTGRN · CLEVR-CoGenT val (train A, test B) | accuracy: 75.6 -> 83.50 | Improved |
| PTGRN versus ACMN on natural images | PTGRN · VQAv2 test-dev (All) | accuracy: 63.81 -> 67.07 | Improved |
Closest papers by meaning
- Visual Question Reasoning on General Dependency Tree (CVPR 2018 · similarity 0.92)
- It's Not About the Journey; It's About the Destination: Following Soft Paths Under Question-Guidance for Visual Reasoning (CVPR 2019 · similarity 0.90)
- MUREL: Multimodal Relational Reasoning for Visual Question Answering (CVPR 2019 · similarity 0.90)
- Interpretable Visual Reasoning via Probabilistic Formulation under Natural Supervision (ECCV 2020 · similarity 0.90)
- Explicit Reasoning over End-to-End Neural Architectures for Visual Question Answering (AAAI 2018 · similarity 0.90)
- Core-to-Global Reasoning for Compositional Visual Question Answering (AAAI 2025 · 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.
