Our paper on federated learning with small sample sizes has been accepted to the International Conference on Learning Representations (ICLR), one of the prime conferences in Machine Learning!

In this joint work with Michael Kamp and Jilles Vreeken, we show how can learn in a federated setting even when only small amounts of samples are available per client and privacy is a huge concern, as typically is the case in hospital settings. We introduce a new communication technique (visualized in the image) which we term federated daisy chaining (FedDC), which mixes traditional communication of model averages through a central server with directly passing on models in a randomized and privacy-preserving way.

In convex settings, we provably show that using appropriate aggregation techniques, our approach can properly learn a good model even on small amounts of data, where current methods fail. On synthetic benchmark and real-world biomedical imaging data, we show that for non-convex methods, such as CNNs, our method greatly outperforms state-of-the-art methods when only small amounts of data are available.

As excited as we are? Here is a preprint!