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..
Tight High-Probability Bounds for Nonconvex Heavy-Tailed Scenario under Weaker Assumptions
Authors: Weixin An, Yuanyuan Liu, Fanhua Shang, Han Yu, Junkang Liu, Hongying Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our Fed CBG algorithm against only FL algorithms FAT-clipping-PR (PR) and FAT-clipping-PI (PI) [47] that can theoretically handle heavy-tailed noise. We also compare the well-known Fed Avg algorithm [31]. We test these methods on the CIFAR-10, CIFAR-100 [19] and Shakespeare [41] datasets. |
| Researcher Affiliation | Academia | 1Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education, School of Artificial Intelligence, Xidian University, China 2School of Computer Science and Technology, Tianjin University, China 3College of Computing and Data Science, Nanyang Technological University, Singapore 4Medical School, Tianjin University, China 5Peng Cheng Lab, Shenzhen, China |
| Pseudocode | Yes | Algorithm 1 Clipped SGD Algorithm 2 Fed CBG Algorithm |
| Open Source Code | No | The section "Experiments" does not contain any explicit statement about open-source code release or a link to a repository. The NeurIPS checklist says "Yes" but the paper content does not support it. |
| Open Datasets | Yes | We test these methods on the CIFAR-10, CIFAR-100 [19] and Shakespeare [41] datasets. |
| Dataset Splits | Yes | We use 100 randomly selected clients to train the model and the remaining 39 clients to test the model performance, which can quantify the performance gap on unseen client distributions. |
| Hardware Specification | Yes | All the experiments were performed on the Ge Force RTX 2080Ti platform with the Py Torch framework. |
| Software Dependencies | No | All the experiments were performed on the Ge Force RTX 2080Ti platform with the Py Torch framework. (No version number for PyTorch) |
| Experiment Setup | Yes | Hyperparameter selection. We conducted ablation experiments on hyperparameters γ, λ, b and K as shown in Fig. 3 and Fig. 6 in the Appendix. When global learning rate γ = 0.2 or 0.3 and λ = 3.0, our Fed CBG algorithm performs better than other choices. The performance of b = 100 and K = 10 ni/b exceeds that of other values. Experimental details. Firstly, we choose η = 1, λ = 3.0, γ = 0.3, K = ni/b and b = 100 to train all the methods. |