Research map / Architectures and efficiency / Attention and transformers
Saccader: Improving Accuracy of Hard Attention Models for Vision
Gamaleldin Elsayed, Simon Kornblith, Quoc V Le
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 |
|---|---|---|---|
| Pretrain the location network using class-label-derived sorted-logit locations before policy-gradient training (ablation) | Saccader · ImageNet | classification accuracy | Improved |
| Add the attention network to the location-selection system (ablation) | Saccader · ImageNet | classification accuracy | Improved |
| Add the Saccader cell rather than selecting locations with the ordered-logits policy (ablation) | BagNet-77-lowD · ImageNet | classification accuracy | Improved |
| Use the learned Saccader policy instead of the ordered-logits policy | BagNet-77-lowD · ImageNet | classification accuracy | Improved |
| Use the learned Saccader policy instead of DRAM | ImageNet hard-attention classification · ImageNet | classification accuracy | Improved |
| Allow DRAM to use more glimpses than it was trained on | DRAM · ImageNet | classification accuracy | Hurt |
| Add a separate NASNet classification network for extracted patches | Saccader · ImageNet 224×224 | classification accuracy | Improved |
| Use 331×331 images for patch classification instead of 224×224 images | Saccader-NASNet · ImageNet | classification accuracy | Improved |
| Occlude the patches selected by Saccader | ResNet-v2-50 classifier · ImageNet validation | classification accuracy | Hurt |
| Use hard attention without adversarial training | Saccader and other tested attention models · ImageNet validation subset, ε = 2/255 | adversarial accuracy | No clear effect |
Closest papers by meaning
- SaccadeNet: A Fast and Accurate Object Detector (CVPR 2020 · similarity 0.87)
- Stand-Alone Self-Attention in Vision Models (NeurIPS 2019 · similarity 0.87)
- Learn to Pay Attention (ICLR 2018 · similarity 0.87)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.87)
- SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks (ICML 2021 · similarity 0.87)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.87)
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.
