Federated Learning on Filecoin

This project is about doing federated learning on Filecoin. Federated learning is a way how to train machine learning models on distributed data without gathering the data to one central place. Filecoin is a perfect place where to do federated learning, because it stores huge amounts of data distributedly and there are already tools how to do compute over those data with bacalhau.org

We create an easy to use tool for anyone to do federated learning. Users can just specify csv datasets on which they would like to train some model. There is a predefined list of supported models. Unfortunately right now, we were able to provide just a simple use case for computing average on last column of data. But the list of supported models could be extended.

This project uses bacalhau for off-chain compute over data. We provide docker image which can run local training (training on individual datasets) and aggregation (aggregating results from local trainings)

Apart from algorithms which make federated learning possible, we provide python script which runs this process easily for you.