Training a deep belief network on a GPU with 100 million free parametersai.stanford.edu35 pointshenning16 years ago