Closed-Form Implicit Neural Representations
Organizations: TelePIX, Seoul, Republic of Korea.
Abstract
Training implicit neural representations (INRs) to capture fine-scale details typically relies on iterative backpropagation and is often hindered by spectral bias when the target exhibits highly non-uniform frequency content. We propose ELM-INR, a backpropagation-free INR that decomposes the domain into overlapping subdomains and fits each local problem using an Extreme Learning Machine (ELM) in closed form, replacing iterative optimization with stable linear least-squares solutions. This design yields fast and numerically robust reconstruction by combining local predictors through a partition of unity. To understand where approximation becomes difficult under fixed local capacity, we analyze the method from a spectral Barron norm perspective, which reveals that global reconstruction error is dominated by regions with high spectral complexity. Building on this insight, we introduce BEAM, an adaptive mesh refinement strategy that balances spectral complexity across subdomains to improve reconstruction quality in capacity-constrained regimes.
Figures & tables
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Description |
| Global function and domain notation | |
| Continuous input coordinate in -dimensional space | |
| Ground-truth target function | |
| Global ELM-INR approximation of | |
| Input dimensionality | |
| Global spatial domain | |
| Dataset | SIREN | FFN | GaussNet | WIRE | ELM-INR |
| kodim05 | 31.12 0.22 | 25.51 0.25 | 24.32 0.12 | 30.68 0.27 | 79.03 1.59 |
| kodim08 | 27.63 0.35 | 22.88 0.15 | 21.97 0.08 | 30.43 0.16 | 78.82 1.73 |
| kodim20 | 40.11 0.64 | 36.02 0.12 | 32.45 0.05 | 39.85 0.94 | 78.42 1.20 |
| kodim24 | 28.03 0.40 | 25.97 0.26 | 24.50 0.06 | 31.10 0.05 | 79.11 1.49 |
| WorldView-3, CDG | 33.06 0.42 | 38.60 0.55 | 30.61 0.38 | 30.89 0.61 | 79.19 0.42 |
| WorldView-3, LAX | 30.56 0.47 | 35.08 0.33 | 28.87 0.52 | 29.20 0.68 | 78.49 0.71 |
| Metric | SIREN | FFN | GaussNet | WIRE | ELM-INR |
| MAE | |||||
| Time (sec) | 1,960 | 1,924 | 2,838 | 2,402 | 0.015 |
| Metric | CO | SIREN | FFN | GaussNet | WIRE | ELM-INR |
| 20.2 | 11.5 | 11.7 | 5.8 | 77.8 | ||
| PSNR (dB) | 25.1 | 13.1 | 16.0 | 9.1 | – | |
| 27.1 | 14.9 | 19.4 | 13.4 | – | ||
| 34 | 20 | 11 | 5 | – | ||
| Steps | 79 | 92 | 48 | 28 | – | |
| 135 | 178 | 93 | 54 | – |
| Band | Description | Wave. Range (nm) | Primary Usage |
| B1 | Coastal | Coastal monitoring and atmospheric correction | |
| B2 | Blue | Aerosol/haze effects and water quality cues | |
| B3 | Green | Vegetation vigor and true-color rendering | |
| B4 | Yellow | Vegetation stress and chlorosis cues | |
| B5 | Red | Chlorophyll absorption and vegetation discrimination | |
| B6 | Red Edge | Chlorophyll-related sensitivity and crop condition monitoring |
| Band | Description | Wave. Range (nm) | Primary Usage |
| B1 | Coastal aerosol | 433–453 | Atmospheric correction |
| B2 | Blue | 458–523 | Water body analysis |
| B3 | Green | 543–578 | Vegetation monitoring |
| B4 | Red | 650–680 | Vegetation and soil analysis |
| B5 | Red-edge 1 | 698–713 | Vegetation monitoring |
| B6 | Red-edge 2 | 733–748 | Vegetation structure analysis |
| Bach | Counting | |||||
| Method | MSE | Time (s) | Params | MSE | Time (s) | Params |
| SIREN | 373.7 | 0.8M | 644.8 | 0.8M | ||
| FFN | 232.9 | 1.1M | 534.4 | 1.1M | ||
| GaussNet | 534.4 | 1.1M | 896.1 | 1.1M | ||
| WIRE | 671.8 | 1.1M | 653.1 | 1.1M | ||
| ELM-INR | 0.015 | 0.1M | 0.020 | 0.1M | ||
| Dataset | Metric | Local-Global | MINER | TensoRF | K-Planes | GA-Planes | ELM-INR |
| kodim05 | PSNR | 28.34 0.42 | 45.36 0.18 | 44.44 0.16 | 46.09 0.01 | 46.36 0.81 | 79.03 1.59 |
| SSIM | 0.9810 0.0003 | 0.9996 0.0000 | 0.9996 0.0000 | 0.9997 0.0000 | 0.9997 0.0001 | 1.0000 0.0000 | |
| WorldView-3 | PSNR | 41.34 0.27 | 43.84 0.15 | 48.38 0.06 | 43.03 0.52 | 48.79 0.36 | 79.19 0.42 |
| SSIM | 0.9840 0.0004 | 0.9910 0.0002 | 0.9962 0.0000 | 0.9950 0.0006 | 0.9987 0.0001 | 1.0000 0.0000 |
| Linear model | ELM-INR | |||
| Image | PSNR (dB) | SSIM | PSNR (dB) | SSIM |
| kodim05 | 16.81 0.58 | 0.6667 0.0614 | 79.03 1.59 | 1.0000 0.0000 |
| kodim08 | 15.46 0.53 | 0.7256 0.0424 | 78.82 1.73 | 1.0000 0.0000 |
| kodim20 | 18.89 0.60 | 0.8165 0.0301 | 78.42 1.20 | 1.0000 0.0000 |
| kodim24 | 19.76 0.51 | 0.8174 0.0258 | 79.11 1.49 | 1.0000 0.0000 |
| cameraman | kodim05 | |||||
| Activation | PSNR (dB) | SSIM | MSE | PSNR (dB) | SSIM | MSE |
| ReLU | 79.98 0.01 | 1.0000 | 79.88 0.10 | 1.0000 | ||
| Sine | 79.97 0.01 | 1.0000 | 79.98 0.01 | 1.0000 | ||
| Tanh | 79.99 0.00 | 1.0000 | 79.99 0.00 | 1.0000 | ||
| GELU | 79.01 1.31 | 1.0000 | 76.91 3.90 | 1.0000 | ||
| Swish | 79.96 0.04 | 1.0000 | 79.67 0.31 | 1.0000 | ||
| Partition | Construction (ms) | Fitting (ms) | Total (ms) | PSNR (dB) | SSIM | |
| 256 | Regular mesh | – | 3.4 | 3.4 | 26.02 | 0.9686 |
| 256 | BEAM | 14.7 | 18.4 | 33.1 | 27.54 | 0.9781 |
| 512 | Regular mesh | – | 7.0 | 7.0 | 29.60 | 0.9864 |
| 512 | BEAM | 14.7 | 45.5 | 60.2 | 32.29 | 0.9928 |
Implicit Neural Representations
Implicit neural representations encode signals and fields as continuous functions that can be queried at arbitrary coordinates and resolutions.
See all 200 papers