FedMBridge: Bridgeable Multimodal Federated Learning

Authors: Jiayi Chen, Aidong Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on four AMFL simulations demonstrate the efficiency and effectiveness of our proposed approach. and 5. Experiments, Table 2 reports the results on all simulations, comparing Fed MBridge with baseline approaches.
Researcher Affiliation Academia 1Department of Computer Science, University of Virginia, Charlottesville, VA 22903, USA.
Pseudocode Yes Algorithm 1 Fed MBridge in Appendix B.2.
Open Source Code No We will release our AMFL simulations in the future. This statement refers to the datasets, not the source code for the proposed methodology. There is no explicit statement or link indicating that the source code for Fed MBridge is released or available.
Open Datasets Yes Scene AMF is constructed from the bimodal NYU-v2 dataset (Nathan Silberman & Fergus, 2012) Object AMF is constructed from the bimodal Model Net40 dataset (Wu et al., 2015) Emotion AMF is created from the CMU-MOSEI dataset (Liang et al., 2021) Mnist AMF is made from AVMnist (Liang et al., 2021) and Multi Mnist (Sabour et al., 2017) datasets
Dataset Splits No The paper mentions training samples and testing samples for the datasets, but does not explicitly specify the use of a validation set or provide details on how such a split would be generated or used for model selection/hyperparameter tuning. For example: Each client i have ni Normal(500, 100) training samples and 1,000 testing samples from the selected 5 classes.
Hardware Specification Yes All the approaches are implemented using Py Torch 3.7 and we ran all experiments on a single A800.
Software Dependencies Yes All the approaches are implemented using Py Torch 3.7
Experiment Setup Yes B.3. Hyperparameters The hyperparameters are listed in Table 5. Table 5. List of Hyperparameters. Where Notations/Descriptions Values local optimizers Adam local learning rate (α) 0.1 0.02 # training epoch per round 15 training batch size 128 client selection ratio (|Nr|/N) 0.25 # GNN layers (L) 4 node feature size per GNN layer [32, 64, 64, 32] layer-role embedding size (S) 32 task embedding size (F) 16 learning rate of TAHN (η) 0.07 0.005 optimizer of TAHN SGD weight decay 0.06