Research map / Vision-language and multimodal / Visual question answering
Learning Answer Embeddings for Visual Question Answering
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 answer embeddings and factorized compatibility rather than independent answer classes | Visual QA models · Visual7W, V7W, VQA2, qaVG | accuracy: 53.7 -> 65.6; 43.6 -> 55.4; 62.4 -> 63.2; 53.0 -> 62.6 | Helped |
| add the answer bi-LSTM embedding (ablation) | fPMC(SAN) · Visual7W MC, V7W MC, VQA2 Top-3k, qaVG MC | accuracy: 65.6 -> 66.0; 55.4 -> 55.5; 63.2 -> 63.9; 62.6 -> 63.4 | Helped |
| add sampled negatives beyond the minibatch answer universe (ablation) | fPMC(SAN) · VQA2 val, all questions | accuracy: 56.7 -> 60.0 | Helped |
| add sampled negatives beyond the minibatch answer universe (ablation) | fPMC(MLP) · VQA2 val, all questions | accuracy: 49.8 -> 55.2 | Helped |
| increase negative samples per minibatch from 0 to 3,000 (ablation) | fPMC · VQA2 val | accuracy | Helped |
| use multiple human annotations as multi-hot positives instead of a single dominant answer (ablation) | fPMC(SAN) · VQA2 val | accuracy: 58.0 -> 60.0 | Helped |
| use WUPS-based weighting rather than multi-hot annotations (ablation) | fPMC(SAN) · VQA2 val | accuracy: 60.0 -> 57.8 | Hurt |
| learn answer embeddings instead of fixing them to GloVe (ablation) | SAN-based PMC · VQA2, source VQA2 | accuracy: 57.5 -> 60.0 | Helped |
| learn answer embeddings instead of fixing them to GloVe (ablation) | SAN-based PMC · Visual7W, source VQA2 | accuracy: 47.5 -> 60.2 | Helped |
| learn answer embeddings instead of fixing them to GloVe (ablation) | SAN-based PMC · qaVG, source VQA2 | accuracy: 37.6 -> 54.8 | Helped |
| factorize compatibility into separate image-question and answer embeddings rather than using uPMC (ablation) | PMC · Visual7W, V7W, VQA2, qaVG | accuracy: 62.4 -> 63.1; 51.6 -> 52.4; 51.4 -> 59.3; 54.5 -> 57.7 (MLP); 65.3 -> 65.6; 55.2 -> 55.4; 56.0 -> 63.2; 61.3 -> 62.6 (SA | Mixed |
| use SAN rather than MLP for the image-question embedding (ablation) | fPMC · Visual7W, V7W, VQA2, qaVG | accuracy: 63.1 -> 65.6; 52.4 -> 55.4; 59.3 -> 63.2; 57.7 -> 62.6 | Helped |
Closest papers by meaning
- Answer-Type Prediction for Visual Question Answering (CVPR 2016 · similarity 0.92)
- Multi-Level Attention Networks for Visual Question Answering (CVPR 2017 · similarity 0.92)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.92)
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.91)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.91)
- Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models (ICLR 2025 · similarity 0.90)
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.