@online{Kamp2110.03469,
TITLE = {Federated Learning from Small Datasets},
AUTHOR = {Kamp, Michael and Fischer, Jonas and Vreeken, Jilles},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2110.03469},
EPRINT = {2110.03469},
EPRINTTYPE = {arXiv},
YEAR = {2021},
ABSTRACT = {Federated learning allows multiple parties to collaboratively train a joint<br>model without sharing local data. This enables applications of machine learning<br>in settings of inherently distributed, undisclosable data such as in the<br>medical domain. In practice, joint training is usually achieved by aggregating<br>local models, for which local training objectives have to be in expectation<br>similar to the joint (global) objective. Often, however, local datasets are so<br>small that local objectives differ greatly from the global objective, resulting<br>in federated learning to fail. We propose a novel approach that intertwines<br>model aggregations with permutations of local models. The permutations expose<br>each local model to a daisy chain of local datasets resulting in more efficient<br>training in data-sparse domains. This enables training on extremely small local<br>datasets, such as patient data across hospitals, while retaining the training<br>efficiency and privacy benefits of federated learning.<br>},
}