Score matching is a method for fitting probability models by matching the gradient of the log-density, the score, of the model to that of the data, avoiding the intractable normalising constant. It underpins score-based generative models and diffusion models.
Semantic Classification
Content
- Score matching estimates an unnormalised probability model by minimising the difference between the model’s score function, the gradient of the log-density, and that of the data distribution. Because the score does not depend on the partition function, the method sidesteps computing an intractable normalising constant.
- Denoising score matching and related estimators connect the idea to learning the score of noise-perturbed data, which is the basis of score-based generative models and diffusion models. These approaches generate samples by following estimated scores through a reverse noising process.