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..
Soft-consensual Federated Learning for Data Heterogeneity via Multiple Paths
Authors: Sheng Huang, Lele Fu, Fanghua Ye, Tianchi Liao, Bowen Deng, zhangchuanfu, Chuan Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments are conducted to demonstrate the advantages of the proposed method. |
| Researcher Affiliation | Collaboration | 1Sun Yat-sen University, Guangzhou, China 2Tencent Inc., Shenzhen, China EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Federated Learning Framework with Multiple Paths (Fed MP) Input: The number of communication round R, The number of initial training round Rit, The number of balance round B, the number of local epochs E, hyperparameter γ, the learning rate η and the number of selected clients K per round. Output: The global model Θ = {Ψ, Ω, Φ}. Algorithm 2 The Client Updates for Fed MP Input: The index of client p, the current round number r, the downloaded models S1, S2, S3 and the hyperparameters of Algorithm 1. Output: The local updated model Θp. |
| Open Source Code | No | Justification: We plan to release the code, datasets, and pre-trained models with sufficient instructions to faithfully reproduce our results. |
| Open Datasets | Yes | Datasets: We conduct main experiments with the proposed method using classification tasks on three datasets, which are CIFAR-10 [30], CIFAR-100 [30] and Flowers102 [31]. |
| Dataset Splits | Yes | For CIFAR-10 and CIFAR-100 datasets, we use a train set consisting of 50,000 samples and a test set consisting of 10,000 samples. For Flowers102 dataset, we use a train set consisting of 6,149 samples and a test set consisting of 1,020 samples. We partition the train sets using the Dirichlet distribution with hyperparameter α {0.3, 0.5, 1.0} to simulate the scenarios with the heterogeneous data distribution, where the smaller α is, the more unbalanced the data distribution among clients. The distributions of the train set and other settings can be found in Supplementary Material. |
| Hardware Specification | Yes | We implemented the proposed method using Py Torch 1.12, and deployed it on a machine configured with AMD R9 5900X, 64GB memory, and Nvidia RTX3090. |
| Software Dependencies | Yes | We implemented the proposed method using Py Torch 1.12, and deployed it on a machine configured with AMD R9 5900X, 64GB memory, and Nvidia RTX3090. |
| Experiment Setup | Yes | For the CIFAR-10 and CIFAR-100 datasets, we train in a system with 20 clients and the batch sizes of the data are set to 128 and 64, respectively. For the Flowers102 dataset, we use a system with 10 clients and the batch size of the data is set to 64. 50% of the clients are chosen in each round for training and aggregation, and the number of local epochs E is set to 5. We set the balance round B to 5, for the CIFAR-10 dataset we set the initial training round Rit to 5, and for the CIFAR-100 and Flowers102 datasets we set the initial training round Rit to 10. To speed up the initialization progress, we constrain the Lipschitz smoothness for the initial model. The hyperparameter γ is set to 0.1, and SGD optimizer with learning rate η = 0.05 is used. |