Optimizing the Collaboration Structure in Cross-Silo Federated Learning

Authors: Wenxuan Bao, Haohan Wang, Jun Wu, Jingrui He

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our framework with a variety of datasets, models, and types of non-IIDness. Our results demonstrate that FEDCOLLAB effectively mitigates negative transfer across a wide range of FL algorithms and consistently outperforms other clustered FL algorithms.
Researcher Affiliation Academia 1University of Illinois Urbana-Champaign, Champaign, IL, USA.
Pseudocode Yes Algorithm 1 Training client discriminator
Open Source Code Yes Our code is available at https://github.com/baowenxuan/Fed Collab.
Open Datasets Yes We evaluate our framework on three models and datasets: we train a 3-layer MLP for Fashion MNIST (Xiao et al., 2017), a 5-layer CNN for CIFAR-10 (Krizhevsky, 2009), and an Image Net pre-trained Res Net-18 (He et al., 2016) for CIFAR-100 (with 20 coarse labels).
Dataset Splits No Table 4 shows 'training / testing samples' but does not explicitly provide validation splits for the main model training. Algorithm 1 mentions 'Train-valid split' for the client discriminator, but not for the primary FL model training.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments are mentioned in the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We train a 3-layer MLP for Fashion MNIST (Xiao et al., 2017), a 5-layer CNN for CIFAR-10 (Krizhevsky, 2009), and an Image Net pre-trained Res Net-18 (He et al., 2016) for CIFAR-100 (with 20 coarse labels)... For all three datasets, we use a light-weight two-layer MLP as the client discriminator to estimate pairwise distribution distances... We run C = 6, 8, 10 on all three settings, and report the best result. Finally, we choose C = 10 for MNIST and CIFAR-10 experiments, and C = 8 for CIFAR-100 experiments.