Mostrar nube de etiquetas

We propose a new framework for estimating generative models via adversarial nets, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitatively evaluation of the generated samples.

Notas/Comentarios de Juan Ignacio Godino:
El trabajo de Goodfellow y sus colegas presenta el concepto de Red Generativa Antagónica (Generative Adversarial Network -GAN). Dado un conjunto de entrenamiento, esta técnica aprende a generar nuevos datos con las mismas estadísticas que el conjunto de entrenamiento. Aunque originalmente se propuso como una forma de modelo generativo para el aprendizaje no supervisado, las GAN también han demostrado ser útiles para el aprendizaje semisupervisado, totalmente supervisado y en el aprendizaje por refuerzo.



Foro Histórico

de las Telecomunicaciones


logo COIT
C/ Almagro 2. 1º Izq. 28010. Madrid
Teléfono 91 391 10 66
logo AEIT Horizontal
C/ General Arrando, 38. 28010. Madrid
Teléfono 91 308 16 66

Copyright 2024 Foro Histórico de las Telecomunicaciones