Federated Learning with Partial Model Personalization
Authors: Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, Lin Xiao
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on realworld image, text, and speech datasets demonstrate that (a) partial personalization can obtain most of the benefits of full model personalization with a small fraction of personal parameters, and, (b) the alternating update algorithm outperforms the simultaneous update algorithm by a small but consistent margin. |
| Researcher Affiliation | Collaboration | 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Meta AI. |
| Pseudocode | Yes | Algorithm 1 Fed Alt / Fed Sim |
| Open Source Code | Yes | The code to reproduce the experimental results is publicly available.1 https://github.com/krishnap25/FL_ partial_personalization |
| Open Datasets | Yes | We use the Stack Overflow dataset, where each device corresponds to the questions and answers of one user on stackoverflow.com. ... We use GLDv2 (Weyand et al., 2020), a large-scale image dataset... We use the EMNIST dataset (Cohen et al., 2017)... We construct a federated version of the Libri Speech dataset (Panayotov et al., 2015)... |
| Dataset Splits | Yes | All the tuning of hyperparameters was performed on validation data, formed by holding out 20% of the training data on each device. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'Tensor Flow Federated' and 'Py Torch' but does not provide specific version numbers for these or any other key software components. |
| Experiment Setup | Yes | Table 9. Hyperparameters for each dataset/task. ... All the tuning of hyperparameters was performed on validation data, formed by holding out 20% of the training data on each device. |