Preprint2014
Generative Adversarial Networks
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu et al. · arXiv · 2014
Abstract
Frames generation as a contest between two networks trained together: one produces samples, the other tries to tell real data from generated data, and each improves by defeating the other. No explicit likelihood is ever computed, which sidesteps the intractable normalisation that made earlier generative models hard to train.
Why it matters
The adversarial framing dominated generative modelling for most of a decade before diffusion displaced it. Worth reading for the idea that a learned critic can replace a hand-specified loss — a pattern that reappears throughout modern alignment work.
generative modelsarchitecturetraining
Read the source
https://arxiv.org/abs/1406.2661