Research map / Learning theory and algorithms / Kernels and regression theory
Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains
Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, Ren Ng
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 Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 2D image Natural | PSNR: 19.32 -> 25.57 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 2D image Text | PSNR: 18.40 -> 30.47 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 3D shape | IoU: 0.864 -> 0.973 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 2D CT Shepp-Logan | PSNR: 16.75 -> 28.33 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 2D CT ATLAS | PSNR: 15.44 -> 19.88 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 3D MRI ATLAS | PSNR: 26.14 -> 34.51 | Improved |
| Add Gaussian Fourier features instead of no input mapping (ablation) | Coordinate-based MLP · 3D NeRF | PSNR: 22.41 -> 25.48 | Improved |
| Use Gaussian random features rather than positional encoding (ablation) | Coordinate-based MLP · Across reported regression tasks | PSNR or IoU: 24.95 -> 25.57; 27.57 -> 30.47; 0.960 -> 0.973; 26.89 -> 28.33; 19.55 -> 19.88; 32.23 -> 34.51; 25.28 -> 25.48 | Improved |
| Tune Fourier-feature frequency scale to match the target signal (ablation) | Fourier-feature MLP · 1D function regression | test error | Mixed |
| Use more gradual Fourier-feature frequency falloff (lower p) (ablation) | Fourier-feature MLP · 1D 1/f^1 noise | training loss | Improved |
| Use Gaussian isotropic frequencies rather than axis-aligned positional encoding (ablation) | Coordinate-based MLP · 2D sinusoid images at varying angles | fit performance | Improved |
| Jointly optimize Fourier-feature parameters with network weights rather than keeping them fixed (ablation) | Coordinate-based MLP · 2D image regression | test performance | No clear effect |
Closest papers by meaning
- How Learnable Grids Recover Fine Detail in Low Dimensions: A Neural Tangent Kernel Analysis of Multigrid Parametric Encodings (ICLR 2025 · similarity 0.88)
- Learning High-Frequency Functions Made Easy with Sinusoidal Positional Encoding (ICML 2024 · similarity 0.87)
- Learning Spatially Collaged Fourier Bases for Implicit Neural Representation (AAAI 2024 · similarity 0.86)
- FINER: Flexible Spectral-bias Tuning in Implicit NEural Representation by Variable-periodic Activation Functions (CVPR 2024 · similarity 0.86)
- Visual Fourier Prompt Tuning (NeurIPS 2024 · similarity 0.85)
- Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning (NeurIPS 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.
