Research map / Architectures and efficiency / Attention and transformers
Learnable Fourier Features for Multi-dimensional Spatial Positional Encoding
Yang Li, Si Si, Gang Li, Cho-Jui Hsieh, Samy Bengio
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 |
|---|---|---|---|
| Add MLP on top of learnable Fourier features (ablation) | Learnable-Fourier-2/2 · Widget captioning | CIDEr: 95.3 -> 98.0 | Improved |
| Make Fourier features learnable instead of fixed (ablation) | Fourier+MLP-2/2 · Widget captioning | CIDEr: 100.2 vs 98.0 | Mixed |
| Add MLP to sinusoidal encoding (ablation) | Sine-4D · Widget captioning | CIDEr: 94.9 -> 98.1 | Improved |
| Add Fourier features to MLP-only encoding (ablation) | MLP · Widget captioning | CIDEr: 70.3 -> 100.7 | Improved |
| Add MLP to learnable Fourier features with normal init (ablation) | Learnable-Fourier · COCO 2017 val | AP: 39.1 -> 40.2 | Improved |
| Add MLP to learnable Fourier features with uniform init (ablation) | Learnable-Fourier · COCO 2017 val | AP: 38.3 -> 40.0 | Improved |
| Normal init instead of uniform [0,1] init (ablation) | Learnable-Fourier · COCO 2017 val | AP: 38.3 -> 39.1 | Improved |
| Replace Embed-2D with learnable Fourier+MLP | DETR · COCO unseen sizes | AP: 36.6 -> 39.5 | Improved |
| Replace Sine-2D with learnable Fourier+MLP | DETR · COCO 2017 val | AP: 40.1 -> 40.2 | Improved |
| Replace Embed-4D with Learnable-Fourier+MLP-4/1 | widget captioning model · unseen positions | CIDEr: 78.5 -> 82.2 | Improved |
| Replace Embed-4D with Learnable-Fourier+MLP-4/1 | widget captioning model · seen positions | CIDEr: 123.4 -> 123.4 | No clear effect |
| Group coordinates (2/2 or 4/1) rather than single 1/4 group (ablation) | Learnable-Fourier+MLP · Widget captioning | CIDEr: 99.1 -> 100.7 | Improved |
Closest papers by meaning
- Learning High-Frequency Functions Made Easy with Sinusoidal Positional Encoding (ICML 2024 · similarity 0.87)
- RelFlexformer: Efficient Attention Transformers for Integrable Relative Positional Encodings (NeurIPS 2026 · similarity 0.85)
- Learning Spatially Collaged Fourier Bases for Implicit Neural Representation (AAAI 2024 · similarity 0.85)
- How Learnable Grids Recover Fine Detail in Low Dimensions: A Neural Tangent Kernel Analysis of Multigrid Parametric Encodings (ICLR 2025 · similarity 0.85)
- Cameras as Relative Positional Encoding (NeurIPS 2025 · similarity 0.85)
- GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers (ICLR 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.
