Diverse Client Selection for Federated Learning via Submodular Maximization
Authors: Ravikumar Balakrishnan, Tian Li, Tianyi Zhou, Nageen Himayat, Virginia Smith, Jeff Bilmes
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide a thorough analysis of its convergence in the heterogeneous setting and apply it both to synthetic and to real datasets. Empirical results show several benefits of our approach, including improved learning efficiency, faster convergence, and more uniform (i.e., fair) performance across clients. |
| Researcher Affiliation | Collaboration | Ravikumar Balakrishnan* Intel Labs ravikumar.balakrishnan@intel.com Tian Li* CMU tianli@cmu.edu Tianyi Zhou* University of Washington tianyizh@uw.edu Nageen Himayat Intel Labs nageen.himayat@intel.com Virginia Smith CMU smithv@cmu.edu Jeffrey Bilmes University of Washington bilmes@uw.edu |
| Pseudocode | Yes | Algorithm 1 Div FL |
| Open Source Code | Yes | Our code is publicly available at github.com/melodi-lab/divfl. |
| Open Datasets | Yes | We evaluate the Div FL approach utilizing both synthetic and real federated datasets from the LEAF federated learning benchmark (Caldas et al., 2019). This includes image datasets (FEMNIST, Celeb A) and a language dataset (Shakespeare). |
| Dataset Splits | No | The paper mentions training loss and test accuracies but does not specify details about validation dataset splits (e.g., percentages, counts, or methodology for creation). |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not explicitly mention specific software dependencies with version numbers used for the experiments. |
| Experiment Setup | Yes | We set the mini batch-size to 10 and learning rate η = 0.01. We consider a total of 30 clients... For all the methods, we fix the number of clients per round K = 10... Each selected client performs τ = 1 round of local model update... Clients use a CNN-based 10-class classifier model with two 5x5-convolutional and 2x2-maxpooling (with a stride of 2) layers followed by a dense layer with 128 activations. For our experiments, Celeb A... a CNN-based binary classifier is utilized with 4 3x3-convolutional and 2x2-maxpooling layers followed by a dense layer. For Shakespeare, a two-layer LSTM classifier containing 100 hidden units with an 8D embedding layer is utilized. |