Research map / Vision-language and multimodal / Visual question answering
AttentionRNN: A Structured Spatial Attention Mechanism
Siddhesh Khandelwal, Leonid Sigal
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 local context attention (CTX) with AttentionRNN | CNN attention · MBG | color accuracy: 79.00 -> 85.84 | Improved |
| model combination function with a neural network instead of assuming independence (ablation) | ARNN · MBG | color accuracy: 84.27 -> 85.84 | Improved |
| sample attention values instead of using the mean (ablation) | ARNN · MBG | color accuracy: 85.84 -> 83.74 | No clear effect |
| replace CTX with ARNN for inverse attribute prediction | CNN attention · MBG-inv | accuracy: 66.37 -> 72.77 | Improved |
| use Block AttentionRNN with block size 2 instead of plain ARNN on larger features (ablation) | ARNN · MBG-b (224x224) | accuracy: 91.09 -> 92.68 | Improved |
| use block size 3 instead of 2 in BRNN (ablation) | BRNN · MBG-b (224x224) | accuracy: 92.68 -> 91.67 | Hurt |
| apply plain ARNN to large 112x112 feature map instead of no attention (ablation) | CNN · MBG-b (224x224) | accuracy: 91.43 -> 91.09 | Hurt |
| replace CBAM spatial attention with ARNN | ResNet18+CBAM · CIFAR-100 | top-1 error: 25.07 -> 24.18 | Improved |
| replace local attention with ARNN | MCB+ATT · VQA 2.0 val | accuracy: 56.89 -> 57.58 | Improved |
| replace local attention with ARNN | ModularGAN · CelebA | hair classification error: 2.5 -> 3.0 | Hurt |
| replace local attention with ARNN | ModularGAN · CelebA | gender classification error: 3.2 -> 1.4 | Improved |
| replace local attention with ARNN | ModularGAN · CelebA | smile classification error: 12.6 -> 11.4 | Improved |
Closest papers by meaning
- An Empirical Study of Spatial Attention Mechanisms in Deep Networks (ICCV 2019 · similarity 0.92)
- Structured Attention Networks (ICLR 2017 · similarity 0.90)
- Learn to Pay Attention (ICLR 2018 · similarity 0.89)
- Attention Branch Network: Learning of Attention Mechanism for Visual Explanation (CVPR 2019 · similarity 0.89)
- Attention as Implicit Structural Inference (NeurIPS 2023 · similarity 0.89)
- SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks (ICML 2021 · 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.
