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..
Adversarial generalization of unfolding (model-based) networks
Authors: Vicky Kouni
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We further present a series of experiments on real-world data, with results corroborating our derived theory, consistently for all data. We evaluate our mathematical results with numerical experiments on real-world data. |
| Researcher Affiliation | Academia | Vicky Kouni LAMSADE, Paris Dauphine PSL Research University EMAIL |
| Pseudocode | No | The paper describes iterative schemes and mathematical formulations (e.g., in Appendix A.2 and A.3) but does not present a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | Upon acceptance, we will provide a link to a public github repository with pytorch code, and sufficient documentation for reproducibility of all the experimental results that accompany the paper. |
| Open Datasets | Yes | We train and test ADMM-DAD on two real-world image datasets: CIFAR10 (50000 training and 10000 test 32x32 coloured image examples) and SVHN (73257 training and 26032 test 32x32 colored image examples). |
| Dataset Splits | Yes | CIFAR10 (50000 training and 10000 test 32x32 coloured image examples) and SVHN (73257 training and 26032 test 32x32 colored image examples). |
| Hardware Specification | Yes | For the course of our experiments, we have utilized a node of 4 H100 GPUs. |
| Software Dependencies | No | We implement all models in Py Torch [19] and train them using the Adam algorithm [21]. While these tools are mentioned, specific version numbers are not provided, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | For CIFAR10, we set ρ = 1 and λ = 10^-4, while for SVHN we alternate ρ and λ depending on the value of N. Particularly, for N = [10, 20, 30, 40, 50], we set λ = [10^-5, 10^-4, 10^-4, 10^-3, 10^-5] and ρ = [100, 1, 1, 1, 10], respectively. All of Adam's parameters are set to their default values, except for the learning rate ϵlr. Specifically, for the CIFAR10 dataset, we train the 5- and 10-layer ADMM-DAD with ϵlr = 10^-5 and ϵlr = 10^-4, respectively. For the SVHN dataset, we train the 10- and 15-layer ADMM-DAD with ϵlr = 10^-4 and ϵlr = 10^-5, respectively. We train all models on all datasets using early-stopping with respect to the adversarial empirical generalization error (adversarial EGE) (21). We repeat all the experiments at least 10 times and average the results over the runs. We initialize the learnable overcomplete sparsifier W RN n using a Xavier normal distribution [14]. We implement all models in Py Torch [19] and train them using the Adam algorithm [21], with batch size 128. |