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..
Neural Tangent Knowledge Distillation for Optical Convolutional Networks
Authors: Jinlin Xiang, Minho Choi, Yubo Zhang, Zhihao Zhou, Arka Majumdar, Eli Shlizerman
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple datasets (e.g., MNIST, CIFAR, Carvana Image Masking Dataset) and hardware configurations show that our pipeline consistently improves ONN performance and enables practical deployment in both pre-fabrication simulations and physical implementations. |
| Researcher Affiliation | Academia | Department of Applied Mathematics, University of Washington, Seattle, WA 98195, USA. Department of Electrical and Computer Engineering, University of Washington, Seattle, WA 98195, USA. Department of Physics, University of Washington, Seattle, WA 98195, USA. Department of Electrical Engineering, Ulsan National Institute of Science and Technology, Ulsan 44919, South Korea ||Corresponding authors: EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology in prose sections like "3.1 Optical Frontend Design", "3.2 Knowledge Transfer Training", and "3.3 Error Compensation", but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper uses open-source benchmark datasets but does not explicitly state that the code for their proposed methodology (NTKD pipeline) is open-source or provide a link to a code repository. The 'Justification' for Question 5 in the context refers to datasets, not code. |
| Open Datasets | Yes | The MNIST and CIFAR-10 datasets each consist of 50, 000 training images and 10, 000 testing images. The Carvana dataset, originally introduced in Kaggle s Carvana Image Masking Challenge, contains 5,088 high-resolution 1920 1280 car images. with additional experiments conducted using Res Net variants and more complex datasets such as Image Net-100 and COCO-Stuff 10k. |
| Dataset Splits | Yes | The MNIST and CIFAR-10 datasets each consist of 50, 000 training images and 10, 000 testing images. |
| Hardware Specification | No | The paper primarily describes the optical hardware used in their ONNs (e.g., metasurfaces, meta-optics, color camera, PSF-engineered meta-optics) and discusses system-level energy consumption and MACs, but it does not specify the computational hardware (e.g., specific GPU or CPU models) used for training or evaluating the digital backend components of their hybrid optical-electronic systems. |
| Software Dependencies | No | The paper mentions general techniques and models such as 'gradient descent algorithm', 'Le Net', 'Alex Net', 'U-Net', and 'Res Net' but does not specify any software libraries, frameworks, or programming languages with their version numbers that were used for implementation. |
| Experiment Setup | Yes | We use batch size of 128 for MNIST and CIFAR-10/100, batch size of 64 for Image Net-100, and batch size of 8 for segmentation tasks. Then, we minimize a weighted sum of two losses, controlled by hyperparameters α and β. λ is a regularization parameter, which is selected via grid search on a validation set. |