Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Adversarial Autoencoder
Neural network architecture combining an autoencoder and a generative adversarial network to force the latent space to follow a specific distribution, thereby improving the quality of reconstructions and generations.
Adversarial Latent Space
Compressed representation of data in an adversarial autoencoder, whose distribution is regularized by a discriminator to approximate a target distribution (e.g., Gaussian), promoting better generalization.
Distribution Discriminator
Neural network in an adversarial autoencoder responsible for distinguishing encodings of real data from samples drawn from a prior distribution, thus forcing the encoder to produce realistic latent representations.
Adversarial Variational Autoencoder (AVAE)
Hybrid model integrating the variational regularization of a VAE with the distributional constraint of a GAN, where the discriminator acts on latent samples to refine the representation space.
Adversarial Noise
Intentionally designed perturbations to deceive the discriminator of an adversarial autoencoder, used during training to strengthen the encoder's robustness.
f-Jensen-Shannon Divergence
Generalized divergence metric used to measure the gap between the learned latent distribution and the target distribution in adversarial autoencoders, offering more flexibility than the classical KL divergence.
Mode Collapse in Autoencoders
Phenomenon where the encoder, while attempting to deceive the discriminator, maps distinct inputs to a limited number of latent representations, reducing the diversity of the latent space.
Cycle-Consistent Adversarial Autoencoder
Variant of adversarial autoencoder where a cycle consistency constraint is added, ensuring that reconstruction after passing through the latent space and back is identical to the original input.
Joint Encoder-Discriminator
Training strategy where encoder and discriminator parameters are partially shared or jointly optimized to stabilize learning and improve latent space regularization.
Structured Embedding Space
Objective of adversarial autoencoders aiming to create a latent space that is not only low-dimensional but also endowed with an exploitable semantic structure, thanks to the distributional constraint.
Adversarially Regularized Reconstruction
Data reconstruction process in an autoencoder where the reconstruction loss is complemented by an adversarial penalty, preventing overfitting and promoting more general representations.
Conditional Adversarial Autoencoder (CAAE)
Extension of the adversarial autoencoder where encoding and generation are conditioned by auxiliary information (e.g., class labels), allowing explicit control over the generated latent representations.
Wasserstein Adversarial Autoencoder
Implementation of an adversarial autoencoder using the Wasserstein loss for the discriminator, which improves training stability and provides a more meaningful measure of latent distribution convergence.
Adversarial Autoencoding Denoising
Application where an adversarial autoencoder is trained to reconstruct clean data from noisy inputs, with the discriminator ensuring that the latent representations of denoised data follow the distribution of clean data.
Adversarial Class Separability
Emergent property in some adversarial autoencoders where the latent space organizes itself to separate different data classes, facilitating downstream classification tasks.