Personalized Federated Image classification using Weight Erosion

Our goal in this project is to adapt Weight Erosion, an update aggregation scheme for personalized collaborative machine learning, to a neural network. Weight Erosion aggregates the data among different agents while ensuring privacy concerns this scheme was developed for a medical concern, to create a performing machine learning model for infectious diseases such as Ebola. We then benchmark the WE scheme against two other common baselines for image classification on the MNIST dataset. We observe that WE outperforms both federated average and local training when the number of agents is quite high, thus when the number of samples per agents is quite low.