Directly learn the score function , which is the gradient of the log density of the data distribution