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..
Revisiting 1-peer exponential graph for enhancing decentralized learning efficiency
Authors: Kenta Niwa, Yuki Takezawa, Guoqiang Zhang, W. Kleijn
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments confirm the effectiveness of these new graphs, particularly the null-cascade graph, in most test settings. ... Numerical experiments on image classification tasks demonstrated that the null-cascade graph achieved the highest test accuracy compared to conventional graphs across most test settings. |
| Researcher Affiliation | Collaboration | Kenta Niwa NTT Communication Science Laboratories <EMAIL> Yuki Takezawa Kyoto University Okinawa Institute of Science and Technology <EMAIL> Guoqiang Zhang University of Exeter <EMAIL> W. Bastiaan Kleijn Victoria University of Wellington <EMAIL> |
| Pseudocode | Yes | Algorithm 1 Null-cascade graph Algorithm 2 Circulant shifting function Algorithm 3 SNF: a subroutine of Alg. 1 Algorithm 4 An implementation of subroutine to select roots, expansions, and communication orders Algorithm 5 New algorithm to construct base-(k +1) graph Algorithm 6 DSGD used in the experiments in Sec. 5 |
| Open Source Code | Yes | https://github.com/garden1984/Null Cascade Graph |
| Open Datasets | Yes | We investigated decentralized learning performance of each graph using image classification benchmark tests using CIFAR-10 and CIFAR-100 with Res Net18 [8]. ... The training dataset of CIFAR-10 and CIFAR-100 5 were divided into n local datasets... 5https://www.cs.toronto.edu/~kriz/cifar.html |
| Dataset Splits | Yes | The training dataset was divided into n local datasets Di (i {1, . . . , n}) to follow a Dirichlet distribution with concentration hyperparameter α [38]. We set α = 0.1, representing a scenario with strong data heterogeneity. ... The data distributions across the n nodes are illustrated in Appendix G. |
| Hardware Specification | Yes | We used computing servers employing 8 GPUs (NVIDIA RTX 6000 Ada (48 GB)) and 2 CPUs (AMD EPYC 9354, 3.25 GHz, 32-Core Processor). |
| Software Dependencies | No | The paper mentions "Py Torch" for data augmentation but does not specify a version number for it or any other key software libraries used in the implementation. |
| Experiment Setup | Yes | We set the number of inner multiple local updates to T = 100 for CIFAR-10 and T = 10 for CIFAR-100. The minibatch size to compute stochastic gradient was 64. ... a learning rate of η = 0.01, which gradually reduces to η = 0.001 through cosine annealing, was chosen for all combinations of datasets (CIFAR-10 / CIFAR-100) and the model (Res Net-18). ... Momentum 0.9 Weight decay 0.005 Batch size 64 Learning rate scheduler Cosine annealing Learning warmup 10 epochs (η = 5e-6) |