KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Fisher Score
In statistics, the gradient of the log-likelihood with respect to the model parameters, a fundamental concept underlying score matching for parametric estimation.
Sliced Score Matching
A variant of score matching that reduces computational complexity by projecting the gradient onto random directions, making training more efficient for high-dimensional data.
Poisson Equation
Partial differential equation linking the vector field of the score to the Laplacian of the logarithm of the density, whose resolution is central in score-based methods.
Stochastic Langevin
Diffusion process or sampling algorithm that uses the gradient of the log-density (the score) to guide sampling towards high-probability regions.
Score Vector Field
Spatial representation of the score at each point in the data space, indicating the direction and magnitude of the strongest increase in probability density.
Stein Divergence
Dissimilarity metric between distributions based on Stein test functions, closely related to the score matching objective and used to evaluate the quality of the score model.
Denoising Paradox
Phenomenon where training a score model on noisy data (denoising) yields better results for estimating the score of clean data than direct training.
Score-Based Generative Modeling
Generative modeling paradigm where a neural network is trained to estimate the score of the data distribution at multiple noise levels, then used for generation via a reverse diffusion process.
Score Multiplicity
Concept where a single score model can be used to generate samples from different distributions by changing the noise level or initial condition of the diffusion process.
Fokker-Planck Equation
Partial differential equation describing the time evolution of the probability density of a stochastic process, fundamental for understanding the theory behind diffusion and score models.
Score Bias Correction
Technique aimed at adjusting score predictions to compensate for biases introduced by model approximation or the use of noisy data, essential for accurate estimation.
Score-Based Normalizing Flow
Hybrid approach where score information is used to design or improve transformations in a normalizing flow model, combining the advantages of both paradigms.
Score Matching Criterion
Objective function, often a form of distance between the predicted score and the true score, that is minimized during training to learn an accurate score model.
Score Interpolation
Process of estimating the score for intermediate noise levels by interpolating model predictions, used in multi-scale diffusion models.
Conditional Score
Extension of score matching where the learned score is conditioned on metadata (e.g., class labels), allowing directional control over the generation process.