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..
Rethinking Gradient Step Denoiser: Towards Truly Pseudo-Contractive Operator
Authors: Shuchang Zhang, Yaoyun Zeng, Kangkang Deng, Hongxia Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive numerical experiments confirm that the learned GS denoiser satisfies the truly pseudo-contractive property and, when integrated into RED-PRO, provides a favorable trade-off between interpretability and empirical performance on inverse problems. |
| Researcher Affiliation | Academia | Shuchang Zhang College of Science National University of Defense Technology Changsha, 410073 EMAIL |
| Pseudocode | Yes | Algorithm 1 RED-PRO with the learned truly SPC denoiser |
| Open Source Code | No | The paper does not explicitly state that source code for the methodology is available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | For the MNIST dataset, ICNN is implemented... For BSD400 and Celeb A, ICNN uses... We test the SPC property of the learned GS denoiser Tθ = I ψθ with ICNN ψθ, MMO [48], and SPCNet [61] on two MNIST and test12 datasets. ...trained on the public Mayo-CT dataset [?]. |
| Dataset Splits | No | All models are trained for Gaussian denoising with a noise level of σ = 5/255 and a batch size of 128. Training spans 50 epochs for MNIST and BSD400, and 30 epochs for Celeb A. Deblurring results on Celeb A over 20 samples. Numerical results for CT reconstruction on the Mayo-CT dataset, computed over 128 test images. |
| Hardware Specification | Yes | All experiments are conducted on one NVIDIA A800 GPU using the Py Torch framework. |
| Software Dependencies | No | All experiments are conducted on one NVIDIA A800 GPU using the Py Torch framework. |
| Experiment Setup | Yes | ICNN models start with an initial learning rate of 10 3, decaying to 10 4 after half the epochs. The Dn CNN models begin with 10 4, decaying to 5 10 5 mid-training. For the MNIST dataset, ICNN is implemented with four convolutional layers, each containing 64 hidden neurons and softplus activation function φ(x) = 1 β log(1 + eβx) with β = 10. For BSD400 and Celeb A, ICNN uses 256 hidden neurons with β = 100. All models are trained for Gaussian denoising with a noise level of σ = 5/255 and a batch size of 128. Training spans 50 epochs for MNIST and BSD400, and 30 epochs for Celeb A. Require: initialization x0 Rn, µk = c (1+k)α , w (0, 2 Lθ ) |