Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients

Authors: Mengmeng Ma, Tang Li, Xi Peng

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluation on a variety of real-world datasets verifies TFL s superior OOF robustness and scalability. ... 4. Experiments ... 4.2. Evaluation on OOF-resiliency ... 4.3. Evaluation on Scalability ... 4.4. Evaluation on In-federation Performance ... 4.5. Evaluation on Effectiveness of Client Clustering ... 4.6. Ablation Study
Researcher Affiliation Academia Mengmeng Ma 1 Tang Li 1 Xi Peng 1 Deep REAL Lab, Department of Computer & Information Sciences, University of Delaware.
Pseudocode Yes Algorithm 1 Topology-aware Federated Learning
Open Source Code No The paper does not contain an explicit statement about releasing source code for the described methodology or a direct link to a code repository.
Open Datasets Yes TFL is evaluated on our curated real-world datasets (①e ICU, ②Fe TS, ③TPT-48) and standard benchmarks (④CIFAR-10/-100, ⑤PACS), spanning a wide range of tasks including classification, regression, and segmentation. ... ①e ICU (Pollard et al., 2018) ... ②Fe TS (Pati et al., 2022b) ... ③TPT-48 (Vose et al., 2014) ... ④CIFAR-10/-100 (Krizhevsky & Hinton, 2009) ... ⑤PACS (Li et al., 2017)
Dataset Splits No The paper describes training and testing splits (e.g., '3 domains are used (15 clients) for training and 1 domain (5 clients) for testing' for PACS), but does not explicitly provide details about a separate validation set or its split percentage/counts.
Hardware Specification Yes All experiments are conducted using a server with 8 NVIDIA A6000 GPUs.
Software Dependencies No The paper mentions optimizers (SGD) and model architectures (ResNet18, U-Net) but does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x).
Experiment Setup Yes For the PACS dataset... learning rate of 0.01, momentum of 0.9, weight decay of 5e 4, and a batch size of 8. ... For e ICU... 30 communication rounds, using a batch size of 64 and a learning rate of 0.01, and report the performance on unseen hospitals. Within each communication round, clients perform 5 epochs (E = 5) of local optimization using SGD.