Communication Optimization in Large Scale Federated Learning using Autoencoder Compressed Weight Updates