Self-Aware Personalized Federated Learning

Authors: Huili Chen, Jie Ding, Eric W. Tramel, Shuang Wu, Anit Kumar Sahu, Salman Avestimehr, Tao Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental With experimental studies on synthetic data, Amazon Alexa audio data, and public datasets such as MNIST, FEMNIST and Sent140, we show that our proposed method can achieve significantly improved personalization performance compared with the existing counterparts.
Researcher Affiliation Industry Alexa AI, Amazon {chehuili,jiedi,eritrame,wushuan,anitsah,avestime,taozhng}@amazon.com
Pseudocode Yes Algorithm 1 Self-aware Personal FL (Self-FL)
Open Source Code Yes We have included the codes in the supplementary material.
Open Datasets Yes MNIST Data. This image dataset has 10 output classes and input dimension of 28 28. We use a multinomial logistic regression model for this task. The FL system has a total of 1, 000 clients. We use the same non-IID MNIST dataset as Fed Prox (Li, 2020), where each client has samples of two-digit classes, and the local sample size of clients follows a power law (Li et al., 2020b).
Dataset Splits No The paper mentions 'training data' and 'test data' but does not explicitly describe a separate 'validation' dataset split or its size/proportion for reproducibility.
Hardware Specification Yes We use AWS p316 instances for all experiments.
Software Dependencies No The paper mentions 'Py Torch example Py Torch (2022)' but does not specify a version number for PyTorch itself or other libraries. It also refers to Github repositories for Fed Prox, p Fed Me, and Per Fed Avg but does not list their specific software dependencies with version numbers.
Experiment Setup Yes For MNIST, we use learning rate = 0.03 and batch size 10 as suggested in Fed Prox (Li, 2020). For FEMNIST, we use = 0.01 and batch size 10. The FL training starts from scratch and runs for 200 rounds. For comparison, we implement Fed Avg and three other personalized FL techniques, DITTO (Li et al., 2021), p Fed Me (Dinh et al., 2020), and Per Fed Avg (Fallah et al., 2020b) with the same hyper-parameter configurations. ... We use a fixed local training step of 20 for other FL algorithms and set lmax = 40 for Self-FL.