Training neural networks using reduced numerical precision (fewer bits) to decrease memory usage and speed up computation.