Research map / Architectures and efficiency / Attention and transformers
Perceiver: General Perception with Iterative Attention
Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, Joao Carreira
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 Fourier features with learned position encodings (ablation) | Perceiver · ImageNet validation | top-1 accuracy: 78.0 -> 70.9 | Hurt |
| interleave 8 cross-attends through the network rather than placing them at the start (ablation) | Perceiver · ImageNet validation | top-1 accuracy: 73.7 -> 78.0 | Improved |
| increase cross-attends from 4 to 8 while interleaving them (ablation) | Perceiver · ImageNet validation | top-1 accuracy: 76.5 -> 78.0 | Improved |
| remove weight sharing (ablation) | Perceiver · ImageNet validation | top-1 accuracy: 78.0 -> 72.9 | Hurt |
| use weight sharing (ablation) | Perceiver · ImageNet train | top-1 accuracy: 87.7 -> 79.5 | Hurt |
| use weight sharing (ablation) | Perceiver · ImageNet validation | top-1 accuracy: 72.9 -> 78.0 | Improved |
| use crop-relative rather than raw-image position coordinates (ablation) | Perceiver · ImageNet | top-1 accuracy | Improved |
| apply video dropout during audiovisual training (ablation) | Perceiver · AudioSet audio + video | mAP: 39.9% vs. 43.2%; 39.7% vs. 43.5% | Improved |
| turn off SpecAugment and drop the spectrogram modality with 10% probability (ablation) | Perceiver mel-spectrogram model · AudioSet | mAP: 44.2 | Improved |
| use temporal unrolling, processing one frame per cross-attend (ablation) | Perceiver · AudioSet video | performance | Improved |
| use temporal unrolling, processing one frame per cross-attend (ablation) | Perceiver · AudioSet audio | performance | Hurt |
| add random per-point translation and rotation augmentation (ablation) | Perceiver · ModelNet40 | accuracy | No clear effect |
Later papers that compared with this method
1 later accepted paper reports Perceiver as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- LookupViT: Compressing visual information to a limited number of tokens (ECCV 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.86)
- TransNeXt: Robust Foveal Visual Perception for Vision Transformers (CVPR 2024 · similarity 0.85)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.85)
- Perception in Reflection (ICML 2025 · similarity 0.84)
- I-Perceive: A Foundation Model for Vision-Language Active Perception (NeurIPS 2026 · similarity 0.84)
- DICEPTION: A Generalist Diffusion Model for Visual Perceptual Tasks (NeurIPS 2025 · similarity 0.84)
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.
