Modern Deep Learning & Transformers

Ian Goodfellow

Invented generative adversarial networks (GANs) in 2014, the architecture that first made realistic image generation possible by pitting two networks against each other.

Ian Goodfellow (b. 1985) came up with generative adversarial networks during a bar conversation in 2014, then reportedly went home and implemented the first working version that same night. The idea: train two networks against each other — a generator that produces fake data and a discriminator that tries to tell fake from real — and let the resulting competition push the generator toward increasingly realistic output. No one had framed generative modeling as a two-player game before, and the GAN became the dominant approach to realistic image generation for most of the following decade, before diffusion models overtook it.

Goodfellow also co-authored Deep Learning (2016) with Yoshua Bengio and Aaron Courville, for years the standard graduate-level textbook on the field, and did influential early work on adversarial examples — small, often imperceptible perturbations to an input that reliably fool a trained model, a finding that helped launch AI security as a distinct research area.

He has worked on safety and adversarial robustness at several of the major AI labs, a natural extension of research that started by showing how easily these systems can be fooled.

See also: Yoshua Bengio

Learn more: Generative Models · Wikipedia: Ian Goodfellow

Mentioned in

Lessons where this comes up in context.