Diffusion Score Semantic Priors for Test-Time Adaptation
A technique showing that the score function a conditional diffusion model learns (its noise prediction, equivalent to the gradient of the log data density) implicitly embeds the class posterior p(y|x), turning a generative diffusion model into a structured discriminative prior. This prior lets a pre-trained image classifier or segmentation network be adapted on unlabeled out-of-distribution data at test time from a single denoising timestep, avoiding Monte Carlo likelihood estimation across many timesteps, with a candidate-selection step that shifts compute from timesteps to a few classes.
Exploring Structured Semantic Priors Underlying Diffusion Score for Test-time Adaptation Mingjia Li
This NeurIPS 2024 paper (DUSA) shows that the score function of a conditional diffusion model, the vector field the model learns to predict the noise added to an image, hides a discriminative prior: …