[Paper Review] Redes Generativas Adversarias (GAN) Fundamentos Teóricos y Aplicaciones

J de la Torre|arXiv (Cornell University)|Feb 18, 2023

Anomaly Detection Techniques and Applications4 citations

TL;DR

This paper provides a comprehensive theoretical and practical overview of Generative Adversarial Networks (GANs), explaining how a generator and discriminator network compete to learn the underlying data distribution and generate realistic synthetic samples. It covers core architectures, training dynamics, and applications across computer vision, NLP, and time series synthesis, with a focus on making GAN theory accessible to Spanish-speaking researchers.

ABSTRACT

Generative adversarial networks (GANs) are a method based on the training of two neural networks, one called generator and the other discriminator, competing with each other to generate new instances that resemble those of the probability distribution of the training data. GANs have a wide range of applications in fields such as computer vision, semantic segmentation, time series synthesis, image editing, natural language processing, and image generation from text, among others. Generative models model the probability distribution of a data set, but instead of providing a probability value, they generate new instances that are close to the original distribution. GANs use a learning scheme that allows the defining attributes of the probability distribution to be encoded in a neural network, allowing instances to be generated that resemble the original probability distribution. This article presents the theoretical foundations of this type of network as well as the basic architecture schemes and some of its applications. This article is in Spanish to facilitate the arrival of this scientific knowledge to the Spanish-speaking community.

Motivation & Objective

• To present the theoretical foundations of Generative Adversarial Networks (GANs) in an accessible format for Spanish-speaking researchers.
• To explain the competitive training mechanism between generator and discriminator networks in GANs.
• To illustrate key architectural designs and training dynamics that enable high-fidelity data generation.
• To survey major applications of GANs across computer vision, natural language processing, and time series modeling.
• To bridge the accessibility gap in GAN research by providing a didactic, comprehensive review in Spanish.

Proposed method

• The paper introduces GANs as a generative modeling framework based on adversarial training between two neural networks: a generator and a discriminator.
• It explains that the generator learns to map random noise to data samples that mimic the true data distribution.
• The discriminator is trained to distinguish between real data samples and those generated by the generator.
• The training process follows a minimax optimization objective, where the generator aims to fool the discriminator.
• Theoretical analysis includes how GANs encode the statistical properties of data distributions within neural network weights.
• The paper presents common GAN architectures and training schemes, including loss functions and training stability considerations.

Experimental results

Research questions

• RQ1 How do generator and discriminator networks interact during adversarial training to model complex data distributions?
• RQ2 What are the core architectural components and training dynamics that enable GANs to generate realistic data?
• RQ3 How do GANs differ from traditional generative models in learning and sampling from data distributions?
• RQ4 What are the main application domains where GANs demonstrate superior performance or unique capabilities?
• RQ5 How can the theoretical principles of GANs be effectively communicated and taught to non-English-speaking researchers?

Key findings

• GANs effectively model complex data distributions by training a generator to produce samples indistinguishable from real data.
• The adversarial training process enables the generator to learn intricate patterns in data, such as those in images and text sequences.
• Applications span multiple domains, including image generation, semantic segmentation, and time series synthesis.
• The paper demonstrates that GANs can generate high-fidelity samples when properly trained, especially with stabilized training techniques.
• The theoretical framework of GANs allows for the encoding of data distribution attributes within neural network parameters.
• The review highlights that GANs are particularly effective in tasks requiring realistic data synthesis, such as data augmentation and text-to-image generation.