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..
Dual-Comb Ghost Imaging with Transformer-Based Reconstruction for Optical Fiber Endomicroscopy
Authors: David Dang, Myoung-Gyun Suh, Maodong Gao, ByoungJun Park, Beyonce Hu, Yucheng Jin, Wilton Kort-Kamp, Ho Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our dual-comb ghost imaging approach, combined with the novel algorithm, outperforms classical ghost imaging techniques in both speed and accuracy, enabling real-time, high-resolution endoscopic imaging with a significantly reduced device footprint. Our contributions include: (1) First demonstration of optical fiber-based ghost imaging using a hardware-software co-design that combines dual-comb interferometry and deep learning, (2) superior reconstruction speed and resolution through the hardware-software co-design approach, (3) a robust transformer-based framework that maintains performance in noisy environments. |
| Researcher Affiliation | Collaboration | 1University of California, Irvine 2NTT Physics and Informatics Laboratories 3Los Alamos National Laboratory |
| Pseudocode | No | The paper describes the model architecture and training process using textual descriptions and mathematical formulas, for example, in the 'Transformer Modeling' section. However, it does not include any explicitly labeled pseudocode blocks or algorithmic figures. |
| Open Source Code | Yes | The scripts, specific training/validation dataset, experimental dual-comb speckle patterns, and target measurements are available at: Code Repository Link. Instructions for running the scripts can be found in the README.txt file, and the required libraries and dependencies are specified in the environment.yml file. |
| Open Datasets | Yes | Our dataset includes 19,280 from OMNIglot and 19,280 images from MNIST with a train/val split of 33,000/5,560 respectively. MNIST: Originally published by [41], downloaded via torchvision.datasets.MNIST, and licensed under the Creative Commons Attribution-Share Alike 3.0 license (see http://yann.lecun. com/exdb/mnist/). Omniglot: Originally published in [42], downloaded via torchvision.datasets.Omniglot, and licensed under the MIT License (see https://github.com/brendenlake/omniglot). |
| Dataset Splits | Yes | Our dataset includes 19,280 from OMNIglot and 19,280 images from MNIST with a train/val split of 33,000/5,560 respectively. |
| Hardware Specification | Yes | Over 30 epochs, the model took approximately 1 hour to train on a Linux workstation, using an AMD Ryzen Threadripper 3990X 64-Core Processor and four NVIDIA A6000 GPUs. |
| Software Dependencies | No | The paper states that 'the required libraries and dependencies are specified in the environment.yml file' in Appendix E. However, it does not explicitly list the software components with their specific version numbers within the text of the paper itself. |
| Experiment Setup | Yes | We use mean squared error as our loss function and an Adam W optimizer with a learning rate of 0.0003 and a weight decay of 0.001. For the following reconstruction results, we used a model with the number of attention heads set to 8 and the embedding_dim set to 32 based on a hyperparameter sweep. |