ConceptDSM / Denoising Score Matching
Denoising Score Matching
A core objective function for training diffusion models. Gaussian noise is added to real data, and the network is trained to predict the direction of the noise gradient, converting the score matching loss into a regression problem.
Timeline (last 90 days)
Oct 14
Denoising score matching (DSM) requires that the training data must come from the target distribution, otherwise the network learns the score of the wrong distribution, which is a fundamental obstacle to applying diffusion models to lattice field theory.
Unverified50%