Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Factored VAE
VAE architecture modifying the objective function to decompose total covariance into independent factors, promoting disentangled representation learning in the latent space.
InfoVAE
VAE variant based on an upper bound of mutual information, replacing KL divergence with MMD distance to improve generation capacity and avoid posterior collapse.
β-TCVAE
Combination of β-VAE and TC-VAE using a unified framework with three β parameters to independently control capacity, independence encouragement, and effective dimensionality.
Annealed VAE
VAE variant implementing progressive annealing of the KL regularization term, starting with a low weight and gradually increasing to avoid posterior collapse during initial training.
Cyclical Annealing Schedule
Training strategy cyclically varying the weight of the KL term, alternating reconstruction and regularization phases to improve convergence and representation quality.
Mutual Information Neural Estimation
Technique for estimating mutual information between inputs and latent variables via adversarial neural networks, used to evaluate and optimize latent space capacity.
IWAE
Importance Weighted Autoencoder using a multi-particle importance sampling likelihood estimator, providing a tighter lower bound on the marginal log-likelihood.
β-VAE with Disentanglement
Extension of β-VAE specifically optimizing the disentanglement of variation factors, adjusting β to maximize independence between latent dimensions while preserving reconstruction quality.
DIP-VAE II
Version améliorée du DIP-VAE utilisant une régularisation plus agressive sur les covariances latentes, implémentée via des contraintes sur la matrice de covariance empirique du code.
VAMP
Variational Autoencoder with a Mixture of Posteriors utilisant un a priori appris comme mélange de distributions postérieures, réduisant le biais de l'a priori standard et améliorant la capacité générative.
Semi-supervised VAE
Extension VAE combinant apprentissage supervisé et non-supervisé, utilisant des labels partiels pour guider l'apprentissage de représentations désenchevêtrées et interprétables.
Conditional VAE
CVAE modélisant p(x|y) en conditionnant l'encodeur et le décodeur sur des variables auxiliaires y, permettant un contrôle explicite des attributs générés et une meilleure désentanglement.
Adversarial Autoencoder
AAE utilisant un réseau adversaire pour forcer la distribution latente à suivre une distribution a priori arbitraire, combinant régularisation VAE et flexibilité de modélisation.