AI-ordlista
Den kompletta ordlistan över AI
Variational Autoencoder
Generative neural network architecture that learns a probabilistic latent representation of input data by maximizing a lower bound on the log-likelihood.
ELBO (Evidence Lower Bound)
Objective function maximized in variational learning, combining reconstruction loss and KL regularization as a lower bound on the marginal log-likelihood.
Inference Posterior
Approximate distribution q(z|x) learned by the encoder to estimate the true posterior p(z|x), typically modeled as a diagonal Gaussian.
Convolutional VAE
VAE variant using convolutional layers to efficiently process image data, preserving spatial structure and improving reconstruction quality.
Hierarchical VAE
Multi-level architecture where latent variables are organized hierarchically, enabling more expressive modeling and progressive generation.
Continuous Latent Space
Representation space where each point corresponds to a valid sample and transitions between points are smooth, enabling interpolation and semantic manipulation.
Disentanglement Factor
Property of the latent space where each dimension captures an independent factor of variation in the data, facilitating interpretability and generative control.
Reparameterized Sampling
Sampling process enabling differentiability by separating sources of randomness from learnable parameters, essential for VAE training.
Conditional Generation
Extension of VAEs where generation is conditioned by additional information such as labels or attributes, allowing targeted control over outputs.
Variational Learning
Optimization paradigm approximating Bayesian inference through ELBO maximization, underlying theoretical foundation for training VAEs.
Distribution Regularization
Mechanism forcing the learned latent distribution to follow a specific shape (typically Gaussian), avoiding posterior collapse and ensuring a structured latent space.