What is a generative adversarial network?

It is a pair of networks trained together: a generator that creates samples and a discriminator that tries to tell real data from generated data. The generator improves by learning to fool the discriminator, and at the ideal balance it produces realistic samples.

How do diffusion models generate data?

Diffusion models learn to reverse a process that gradually adds noise to data. Starting from pure noise, the model repeatedly removes a little noise, step by step, until a clean sample emerges, which has proven very effective for high-quality image generation.

Deep Generative Models

Deep generative models learn to produce new data resembling a training set, capturing the distribution of complex data such as images, audio, and text.

Encontrar tema com PaperMindEm breveFind papers & topics

Tools & resources

Baixar slides

Learn & explore

VídeoEm breve

Definition

A deep generative model is a neural network that represents a probability distribution over data and can generate new samples from it, learned either by maximizing likelihood, by adversarial training, or by learning to invert a corruption process, depending on the model family.

Scope

This topic covers neural models that model or sample from data distributions: variational autoencoders that pair an encoder and decoder with a latent prior, generative adversarial networks that pit a generator against a discriminator, autoregressive models that predict data one element at a time, and diffusion models that learn to reverse a noising process. It addresses how these models are trained and evaluated.

Core questions

How can a neural network learn to generate realistic new data?
What distinguishes likelihood-based, adversarial, and diffusion approaches?
How are latent variables used to control and structure generation?
How is the quality of generated samples assessed?

Key theories

Latent-variable generation: Variational autoencoders learn an encoder that maps data to a latent distribution and a decoder that reconstructs data, enabling sampling of new data by drawing latent codes from a prior.
Adversarial training: Generative adversarial networks train a generator to fool a discriminator that tries to distinguish real from generated data, a game whose equilibrium yields realistic samples.
Autoregressive and diffusion modeling: Autoregressive models factorize a distribution into a product of conditionals predicted in sequence, while diffusion models learn to reverse a gradual noising process, both giving high-quality generation.

Clinical relevance

Deep generative models power image and audio synthesis, data augmentation, drug and material design, and the generative artificial intelligence systems that produce text and images; their ability to create convincing synthetic content also raises significant concerns about misuse, authenticity, and evaluation.

History

Generative modeling with neural networks progressed from restricted Boltzmann machines and deep belief nets to variational autoencoders and, in 2014, generative adversarial networks. Autoregressive models advanced text and audio generation, and diffusion models later achieved leading image-synthesis quality, making generative modeling a central theme of modern artificial intelligence.

Key figures

Ian Goodfellow
Diederik Kingma
Yoshua Bengio

Seminal works

goodfellow2016
lecun2015
murphy2012

Frequently asked questions

What is a generative adversarial network?: It is a pair of networks trained together: a generator that creates samples and a discriminator that tries to tell real data from generated data. The generator improves by learning to fool the discriminator, and at the ideal balance it produces realistic samples.
How do diffusion models generate data?: Diffusion models learn to reverse a process that gradually adds noise to data. Starting from pure noise, the model repeatedly removes a little noise, step by step, until a clean sample emerges, which has proven very effective for high-quality image generation.