Research map / Vision-language and multimodal / Visual question answering
Alignment Attention by Matching Key and Query Distributions
Shujian Zhang, XINJIE FAN, Huangjie Zheng, Korawat Tanwisuth, Mingyuan Zhou
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 |
|---|---|---|---|
| Replace deterministic attention with Bayesian attention (BAM-WC) | GAT · Cora, Citeseer, PubMed | accuracy: 83.00 -> 83.81; 72.50 -> 73.52; 77.26 -> 78.82 | Improved |
| Use BAM-WC attention instead of soft attention | MCAN · VQA-v2 validation, noisy images | accuracy: 61.25 -> 62.89 | Improved |
| Use BAM-WC attention instead of soft attention | MCAN · VQA-v2 validation, original images | accuracy: 66.95 -> 67.02 | No clear effect |
| Use BAM-WC attention instead of soft attention | MCAN · VQA-v2 validation, original and noisy images | PAvPU: 70.04 -> 71.21; 65.34 -> 66.75 | Improved |
| Use Bayesian attention for image captioning | Att2in · MS-COCO | BLEU-4: 32.2 -> 32.8 | Improved |
| Use BAM-WC instead of soft attention | LSTM machine translation model · IWSLT | BLEU: 32.77 -> 33.81 | Improved |
| Fine-tune with BAM-WC attention | ALBERT-base · GLUE and SQuAD | benchmark score | Mixed |
| Add a contextual prior rather than use a fixed prior (ablation) | BAM on graph classification · Cora, Citeseer, PubMed | accuracy: 83.24 -> 83.34; 72.86 -> 73.04; 78.30 -> 78.76 | Improved |
| Use Weibull rather than Lognormal attention distributions (ablation) | BAM on graph classification · Cora, Citeseer, PubMed | accuracy: 83.24 -> 83.48; 72.86 -> 73.18; 78.30 -> 78.50 | Improved |
| Add a contextual prior rather than use a fixed prior (ablation) | BAM on VQA · VQA-v2 validation, original and noisy | accuracy: 66.89 -> 66.93; 61.43 -> 61.58 | Mixed |
| Use Weibull rather than Lognormal attention distributions (ablation) | BAM on VQA · VQA-v2 validation, original and noisy | accuracy: 66.89 -> 66.93; 61.43 -> 61.60 | Mixed |
| Use semi-analytic KL regularization rather than fully sampled KL (ablation) | BAM · MS-COCO image captioning | training stability | Improved |
Closest papers by meaning
- Latent Alignment and Variational Attention (NeurIPS 2018 · similarity 0.89)
- Online and Linear-Time Attention by Enforcing Monotonic Alignments (ICML 2017 · similarity 0.86)
- Context-Aware Self-Attention Networks (AAAI 2019 · similarity 0.86)
- Compositional Attention: Disentangling Search and Retrieval (ICLR 2022 · similarity 0.86)
- Attention Alignment Between Humans and Vision-Language Models (NeurIPS 2026 · similarity 0.85)
- ResMatch: Residual Attention Learning for Feature Matching (AAAI 2024 · 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.
