Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals
Authors: Namhoon Kim, Sara Fridovich-Keil
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We investigate the performance of diverse INRs across a suite of 2D and 3D real and synthetic signals with varying effective bandwidth, as well as both overfitting and generalization tasks including tomography, super-resolution, and denoising. By stratifying performance according to model size as well as signal type and bandwidth, our results shed light on how different INR and grid representations allocate their capacity. |
| Researcher Affiliation | Academia | Namhoon Kim Sara Fridovich-Keil Department of Electrical and Computer Engineering Georgia Institute of Technology EMAIL |
| Pseudocode | No | The paper describes various signal representation methods and their experimental evaluation but does not include any explicitly labeled pseudocode or algorithm blocks. The methods are described in prose. |
| Open Source Code | Yes | Further implementation details are provided in Section 5.5, Table 2, and Table 3 ; the code for both signal generation and model evaluation is provided at https://github.com/voilalab/INR-benchmark. |
| Open Datasets | Yes | DIV2K Images [27] 2D High-resolution image dataset for (1) image overfitting, (2) 4 super-resolution along each axis, and (3) denoising with Gaussian noise (standard deviation ̈ {0.05, 0.1}). Computed Tomography (CT) [28] 2D X-ray CT scan of a human chest, used to evaluate signal recovery in a classic underdetermined inverse problem. 3D Dragon Shape [29] 3D A solid 3D object with approximately 1 ̇ 106 voxels (before super-resolution). |
| Dataset Splits | Yes | For CT experiments, we train models on a real chest CT slice from the dataset in Clark et al. [28], which was also used in WIRE [16]. The training data was 100 projection measurements of the original 326 ̇ 435 chest CT slice, forming a 100 ̇ 435 sinogram equivalent to approximately 30% of the total pixel count in the original image. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA RTX A6000 GPU with 48GB VRAM, with memory usage posing no limitations. |
| Software Dependencies | No | The paper mentions using the Adam optimizer with specific beta values and lists several neural network architectures (FFN, SIREN, WIRE, etc.), but it does not provide specific version numbers for software libraries or frameworks like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | All models are trained using the Adam optimizer with ̢ = (0.9, 0.999) and learning rates tuned via grid search on the Star Target image at model size 1 ̇ 104 (see Table 2). Initialization schemes follow prior work: Fourier Features use Gaussian-initialized embeddings [15] and SIREN uses small uniform weights to prevent sinusoidal explosion [14]. Mean-squared error loss is used for all experiments, with optional total-variation regularization applied to grid-based models to encourage smoothness on generalization tasks. |