FedAvg with Fine Tuning: Local Updates Lead to Representation Learning

Authors: Liam Collins, Hamed Hassani, Aryan Mokhtari, Sanjay Shakkottai

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct experiments to (I) verify our theoretical results in the linear setting and (II) determine whether our established insights generalize to deep neural networks. We use the image classification datasets CIFAR-10 and CIFAR-100 [57], which consist of 10 and 100 classes of RGB images, respectively.
Researcher Affiliation Academia Liam Collins ECE Department The University of Texas at Austin liamc@utexas.edu Hamed Hassani ESE Department University of Pennsylvania hassani@seas.upenn.edu Aryan Mokhtari ECE Department The University of Texas at Austin mokhtari@austin.utexas.edu Sanjay Shakkottai ECE Department The University of Texas at Austin sanjay.shakkottai@utexas.edu
Pseudocode No The paper describes the Fed Avg algorithm with equations (2) and (3) but does not present it or any other procedure in a structured pseudocode block or algorithm environment.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Please see Appendix C and the supplementary material.
Open Datasets Yes We use the image classification datasets CIFAR-10 and CIFAR-100 [57], which consist of 10 and 100 classes of RGB images, respectively.
Dataset Splits Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] Please see Appendix C. Image classes are heterogeneously allocated to M = 100 clients according to the Dirichlet distribution with parameter 0.6 as in [59].
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Please see Appendix C.
Software Dependencies No The provided text does not list specific software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8').
Experiment Setup Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] Please see Appendix C. Then we run Fed Avg with τ = 2 local updates and D-GD, both sampling m = M clients per round. We fine-tune using GD for τ = 200 iterations with batch size b = n.