Linear Convergence of Decentralized FedAvg for PL Objectives: The Interpolation Regime

Authors: Shruti P Maralappanavar, Prashant Khanduri, Bharath B N

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on multiple real datasets corroborate our theoretical findings. In this section, we experimentally validate our theoretical findings for the decentralized versions of Fed Avg. First, we present the experimental setup for various settings.
Researcher Affiliation Academia Shruti Maralappanavar EMAIL Department of Electrical, Electronics and Communication IIT Dharwad Dharwad, India; Prashant Khanduri EMAIL Department of Computer Science Wayne State University Detroit, MI, USA; B. N. Bharath EMAIL Department of Electrical, Electronics and Communication IIT Dharwad Dharwad, India
Pseudocode Yes Algorithm 1 Decentralized Fed Avg
Open Source Code No The paper does not explicitly state that open-source code is provided, nor does it include links to a code repository. Statements like 'code will be made available' or 'available upon request' are also absent.
Open Datasets Yes Experiments on multiple real datasets corroborate our theoretical findings. We consider the image classification tasks on CIFAR-10, MNIST, and FMNIST datasets using an overparameterized simple regression and Deep Neural Network (DNN) models.
Dataset Splits Yes We consider that each device has 490 training samples and 90 test samples for the CIFAR-10 dataset. On the other hand, for MNIST and FMNIST datasets, 540 samples are used for training and 80 samples are used for testing.
Hardware Specification Yes We have implemented all our experiments on NVIDIA DGX A100.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes We set the number of local updates T = 10 and pick the tunable learning rate in the range η [0.001 : 0.01] for CIFAR-10, MNIST, and FMNIST datasets. We use 60 edge devices to run the Decentralized Fed Avg algorithm with multiple local SGD steps.