Distributed Learning with Compressed Gradient Differences