Research map / Vision-language and multimodal / Visual question answering
Bilinear Attention Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use bilinear attention over pairs of vision and language channels instead of separate modality attentions | VQA 2.0 · VQA 2.0 validation | VQA score: 64.59 $\pm$0.04 -> 65.36 $\pm$0.14 | Helped |
| Integrate multiple attention glimpses with residual learning rather than summation (ablation) | BAN-4 · VQA 2.0 validation | VQA score: 64.78 $\pm$0.08 -> 65.81 $\pm$0.09 | Helped |
| Integrate multiple attention glimpses with residual learning rather than concatenation (ablation) | BAN-4 · VQA 2.0 validation | VQA score: 64.71 $\pm$0.21 -> 65.81 $\pm$0.09 | Helped |
| Increase the number of bilinear attention glimpses from one to eight (ablation) | BAN · VQA 2.0 validation | VQA score: 65.36 $\pm$0.14 -> 66.00 $\pm$0.11 | Helped |
| Use twelve rather than eight bilinear attention glimpses (ablation) | BAN · VQA 2.0 validation | VQA score: 66.00 $\pm$0.11 -> 66.04 $\pm$0.08 | No clear effect |
| Augment Glove embeddings with semantically mixed embeddings instead of a parameter-matched duplicate (ablation) | Four-glimpse BAN · VQA 2.0 validation | VQA score: 65.72 ($\pm$ 0.11) -> 66.03 ($\pm$0.12) | Helped |
| Add a counting module (ablation) | Vanilla four-glimpse BAN · VQA 2.0 validation counting questions | Accuracy: 54.92 ($\pm$0.30) -> 58.21 ($\pm$0.49) | Helped |
| Add a counting module (ablation) | Vanilla four-glimpse BAN · VQA 2.0 validation | Overall accuracy: 65.81 ($\pm$0.09) -> 66.01 ($\pm$0.14) | Helped |
| Replace low-rank pooling in BAN-1 with MFB pooling (ablation) | BAN-1 · VQA 2.0 | VQA score | No clear effect |
| Batch phrase entities as multi-channel inputs instead of processing one entity per inference | BAN · Flickr30k Entities inference | ms/entity: 0.84 ms/entity -> 0.67 ms/entity; 25.37% improvement | Helped |
Closest papers by meaning
- Hadamard Product for Low-rank Bilinear Pooling (ICLR 2017 · similarity 0.89)
- Hierarchical Question-Image Co-Attention for Visual Question Answering (NeurIPS 2016 · similarity 0.87)
- Multi-Level Attention Networks for Visual Question Answering (CVPR 2017 · similarity 0.87)
- Cross-modal Information Flow in Multimodal Large Language Models (CVPR 2025 · similarity 0.86)
- See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning (CVPR 2026 · similarity 0.85)
- Where Does Vision Meet Language? Understanding and Refining Visual Fusion in MLLMs via Contrastive Attention (CVPR 2026 · similarity 0.85)
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.