Research map / 3D vision / Face and shape
Closed-Form Implicit Neural Representations
Woojin Cho, Junghwan Park
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ELM-INR, a backpropagation-free implicit neural representation that decomposes the domain into overlapping subdomains and fits each with a closed-form Extreme Learning Machine.
Problem
Training implicit neural representations to capture fine-scale details relies on iterative backpropagation and suffers from spectral bias when target signals have highly non-uniform frequency content.
Approach
Replaces iterative gradient-based training with closed-form least-squares fitting of Extreme Learning Machines on overlapping spatial subdomains, so each local problem is solved directly rather than by backpropagation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace backprop-trained INRs with closed-form local ELM + partition of unity | ELM-INR vs SIREN/FFN/GaussNet/WIRE · Kodak, WorldView-3, Sentinel-2, MRI | PSNR (dB): kodim05 31.12 (SIREN) -> 79.03 | Improved |
| replace backprop-trained INRs with closed-form ELM-INR | Navier-Stokes volume · Navier-Stokes (2+1)D | MAE / time: MAE 4.22e-3 -> 5.68e-5; time 1,960 s -> 0.015 s | Improved |
| add BEAM adaptive partitioning instead of regular mesh (same N) (ablation) | ELM-INR, m=256 · kodim20 | PSNR (dB): 26.0 -> 27.5 | Improved |
| add BEAM adaptive partitioning instead of regular mesh (same N) (ablation) | ELM-INR, m=512 · kodim20 | PSNR (dB): 29.6 -> 32.3 | Improved |
| add BEAM adaptive partitioning instead of regular mesh (same N) (ablation) | ELM-INR, m=1024 · ERA5 winter | MAE: 0.398 -> 0.363 | Improved |
| add BEAM adaptive partitioning instead of regular mesh (same N) (ablation) | ELM-INR, m=2048 · ERA5 winter | MAE: 0.317 -> 0.288 | Improved |
| increase hidden nodes m (ablation) | ELM-INR · cameraman 512x512 | PSNR | Improved |
| add Random Fourier Features encoding (ablation) | ELM-INR · cameraman 512x512 | PSNR saturation vs m | Improved |
| give baselines up to 10x CO2/compute budget (ablation) | SIREN, FFN, GaussNet, WIRE · cameraman 256x256 | PSNR (dB): SIREN 20.2 -> 27.1 (still far below ELM-INR 77.8) | Improved |
Experimental setup
Builds on: Extreme Learning Machine implicit neural representations
Closest papers by meaning
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.88)
- Learning Transferable Features for Implicit Neural Representations (NeurIPS 2024 · similarity 0.87)
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.86)
- SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.86)
- Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations (ICML 2025 · similarity 0.86)
- Few-shot Implicit Function Generation via Equivariance (CVPR 2025 · similarity 0.86)
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.
