In score-based generative modelling, the Score Function is the gradient of the log probability density of the data with respect to the input, indicating the direction of increasing data likelihood. Diffusion models learn to estimate this score across noise levels, then use it to iteratively denoise samples drawn from a simple prior. The score function connects diffusion models to Langevin-style sampling and energy-based formulations.
Overview
- In score-based generative modelling, the Score Function is the gradient of the log probability density of the data with respect to the input, indicating the direction of increasing data likelihood.
- Diffusion models learn to estimate this score across noise levels, then use it to iteratively denoise samples drawn from a simple prior.
- The score function connects diffusion models to Langevin-style sampling and energy-based formulations.
- It is modelled as a subclass of Diffusion Model within the artificial-intelligence domain.
Key aspects
- Noise Schedule is a constituent or mechanism relevant to Score Function.
- Probability Distribution is a constituent or mechanism relevant to Score Function.
- Maximum Likelihood Estimation is a constituent or mechanism relevant to Score Function.
Mechanisms
- Score Function enables Sampling.
- Score Function enables Generative Model.
- Score Function supports Diffusion Model.
Applications
- Applied in contexts involving Sampling.
- Applied in contexts involving Generative Model.
- Applied in contexts involving Diffusion Model.
- Applied in contexts involving Artificial Intelligence.