Training a neural network to estimate the gradient of the log-probability (score) needed to reverse a noising process.