By leveraging idle computing time, researchers can double the speed of model training while preserving accuracy.

Read the full article at news.mit.edu