Complement Sparsification: Low-Overhead Model Pruning for Federated Learning

Authors: Xiaopeng Jiang, Cristian Borcea

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate CS experimentally with two popular FL benchmark datasets.
Researcher Affiliation Academia Department of Computer Science, New Jersey Institute of Technology, Newark, NJ, USA xj8@njit.edu, borcea@njit.edu
Pseudocode Yes Algorithm 1: Complement Sparsification Pseudo-code
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes CS is evaluated with two benchmark datasets in LEAF (Caldas et al. 2018): Twitter and FEMNIST.
Dataset Splits No The training dataset is constructed with 80% of data from each user, and the rest of the data are for testing.
Hardware Specification Yes The experiments are conducted on a Ubuntu Linux cluster (Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz with 512GB memory, 4 NVIDIA P100-SXM2 GPUs with 64GB total memory).
Software Dependencies No We implement CS with Flower (Beutel et al. 2020) and Tensorflow. Specific version numbers for these software dependencies are not provided.
Experiment Setup Yes Table 1 shows the training hyper-parameters for the two models. We set the aggregation ratio (η in equation 8) to 1.5 to avoid clients training outcomes being pruned away if they are too small. We set the server model sparsity to 50%, unless otherwise specified.