Research map / Vision-language and multimodal / Visual question answering
Hadamard Product for Low-rank Bilinear Pooling
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 |
|---|---|---|---|
| replace low-rank bilinear pooling with MCB | MLB · VQA test-dev | overall accuracy: 65.08 -> 61.48% | Hurt |
| add ReLU for attention distributions to the MCB replacement (ablation) | MLB architecture with MCB · VQA test-dev | overall accuracy: 61.48% -> 62.11% | Helped |
| use two learning blocks instead of three (ablation) | MARN · VQA test-dev | overall accuracy: 61.68 -> 63.92 | Helped |
| increase glimpses from one to two (ablation) | MARN-L1 · VQA test-dev | overall accuracy: 63.79 -> 64.53 | Helped |
| increase glimpses from two to four (ablation) | MARN-L1 · VQA test-dev | overall accuracy: 64.53 -> 64.61 | Helped |
| add tanh activation instead of using no tanh (ablation) | MARN-L1-G2 · VQA test-dev | overall accuracy: 63.58 -> 64.53 | Helped |
| move tanh from before to after the Hadamard product (ablation) | MARN-L1-G2 · VQA test-dev | overall accuracy: 64.53 -> 64.53 | No clear effect |
| sample answers from the annotator answer distribution rather than using the mode answer (ablation) | MARN-L1-G2 · VQA test-dev | overall accuracy: 64.53 -> 64.80 | Helped |
| remove shortcut connections (ablation) | one-block MARN · VQA test-dev | overall accuracy: 64.80 -> 65.08 | Helped |
| add Visual Genome question-answer training data (ablation) | MLB · VQA test-dev | overall accuracy: 65.08 -> 65.84 | Helped |
| vary joint embedding size from d=1200 to d=800 (ablation) | MLB · VQA test-dev | overall accuracy: 65.08 -> 64.89 | Hurt |
| vary joint embedding size from d=1200 to d=1000 (ablation) | MLB · VQA test-dev | overall accuracy: 65.08 -> 65.06 | No clear effect |
Closest papers by meaning
- Visual Haystacks: A Vision-Centric Needle-In-A-Haystack Benchmark (ICLR 2025 · similarity 0.83)
- Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents (CVPR 2025 · similarity 0.83)
- MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs (ICLR 2025 · similarity 0.83)
- Star with Bilinear Mapping (CVPR 2025 · similarity 0.83)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · similarity 0.82)
- Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger (ICML 2025 · similarity 0.82)
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.