Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting

Authors: Enyi Jiang, Yibo Jacky Zhang, Sanmi Koyejo

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments verify the theoretical insights and illustrate the effectiveness of the proposed methods in practice.
Researcher Affiliation Academia Enyi Jiang enyij2@illinois.edu UIUC Yibo Jacky Zhang yiboz@stanford.edu Stanford Sanmi Koyejo sanmi@cs.stanford.edu Stanford
Pseudocode Yes Algorithm 1 FDA: Gradient Projection and the Auto-Weighting Scheme
Open Source Code Yes Our code is at https://github.com/jackyzyb/Auto Fed GP.
Open Datasets Yes Colored MNIST (Arjovsky et al., 2019), VLCS (Fang et al., 2013), Terra Incognita (Beery et al., 2018) datasets... Fashion-MNIST (Xiao et al., 2017) and CIFAR-10 (Krizhevsky et al., 2009) datasets
Dataset Splits Yes For source domains, we split the training/testing data with a 20% and 80% split.
Hardware Specification Yes In a real implementation, the whole process of projection is fast, with around 0.023 seconds per call needed for N = 10 clients of Fashion-MNIST experiments on the NVIDIA TITAN Xp hardware with GPU available.
Software Dependencies No The paper mentions using Adam (Kingma & Ba, 2014) optimizer, CNN models, and ResNet-18 (He et al., 2016), but it does not specify version numbers for any software libraries or frameworks (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes We set the communication round R = 50 and the local update epoch to 1, with 10 clients (1 target, 9 source clients) in the system... For the experiments on the Fashion-MNIST dataset, we set the source learning rate to be 0.01 and the target learning rate to 0.05. For CIFAR-10, we use a 0.005 source learning rate and a 0.0025 learning rate. The source batch size is set to 64 and the target batch size is 16.