Federated Reconstruction: Partially Local Federated Learning

Authors: Karan Singhal, Hakim Sidahmed, Zachary Garrett, Shanshan Wu, John Rush, Sushant Prakash

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We next describe experiments validating FEDRECON on matrix factorization and next word prediction. We aim to determine whether reconstruction can enable practical partially local federated learning with fast personalization for new clients, including in settings without user-specific embeddings.
Researcher Affiliation Industry Karan Singhal Google Research karansinghal@google.com Hakim Sidahmed Google Research hsidahmed@google.com Zachary Garrett Google Research zachgarrett@google.com Shanshan Wu Google Research shanshanw@google.com Keith Rush Google Research krush@google.com Sushant Prakash Google Research sush@google.com
Pseudocode Yes Algorithm 1 Federated Reconstruction Training
Open Source Code Yes Release an open-source library for evaluating algorithms across tasks in this setting.1 1https://git.io/federated_reconstruction
Open Datasets Yes We evaluate on federated matrix factorization using the popular Movie Lens 1M collaborative filtering dataset [29]. We perform next word prediction with the federated Stack Overflow dataset introduced in Tensor Flow [51].
Dataset Splits Yes We split each user s ratings into 80% train, 10% validation, and 10% test by timestamp.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'Tensor Flow Federated library' and 'Keras model' but does not specify version numbers for these software components.
Experiment Setup Yes We train for 1500 rounds using a total of 100 clients per round, 10 epochs of local training, and batch size 32 for each client. For FEDRECON, we use 10 local reconstruction steps and 1 local global update step. Additional details on hyperparameter choices for each task are provided in Appendix C.