= Ian Goodfellow
{c}
<Ian Goodfellow> led the coauthored 2014 paper introducing <generative adversarial networks>. A generator transforms a <noise> input into a sample, while a discriminator attempts to distinguish generated samples from <training data>[training examples]. Training sets up a <minimax>[two-player minimax] objective rather than directly requiring a tractable <probability density function>[probability density] for every generated observation.<Generative Adversarial Nets by Ian Goodfellow and others>[abstract and §2]{ref}
The original objective is
$$
\min_G\max_D\;\mathbb E_{x\sim p_{\rm data}}\log D(x)+\mathbb E_{z\sim p_z}\log(1-D(G(z))).
$$
Alternating updates seek to improve each player against the other, but the idealized <Nash equilibrium>[equilibrium] analysis does not guarantee that a practical finite <neural network> training run converges. Goodfellow also coauthored <Deep Learning by Ian Goodfellow and others> with <Yoshua Bengio> and <Aaron Courville>.
Back to article page
