🎭

Ian Goodfellow

Original nameIan Goodfellow

American Computer Scientist, Inventor of the Generative Adversarial Network (GAN)

Those Who Conceived That "a Machine Could Think" · Pioneers

Generative Adversarial Networks (GANs)An Early Foundation of Generative AIThe Deep Learning Textbook

Who they are

Ian Goodfellow (1985– ) is an American computer scientist who in 2014, while pursuing his doctorate at the University of Montreal, proposed the pioneering architecture "generative adversarial network" (GAN), letting two neural networks contend against each other — one responsible for generating as realistic as possible fake samples, the other for telling real from fake as accurately as possible, the two improving together in continual contest, so that the generator network can at last produce image content that passes for real. This conception is thereafter acknowledged as one of the earliest core breakthroughs of generative AI, and though, as compute and data scale further grew, new architectures such as diffusion models gradually surpassed GANs in image-generation quality, the generative adversarial network’s historical standing as the originator of the idea of "driving generative ability by adversarial contest" is still widely acknowledged. The Deep Learning textbook Goodfellow co-authored with Bengio is one of the most widely cited introductory and advanced texts in the field. He thereafter was in charge of machine-learning safety work at Apple, resigning in 2022 over a conflict between the company’s return-to-office policy and his personal childcare arrangement, a resignation reason that provoked much public discussion of the tech industry’s remote-work policies at the time.

Primary sourcesGoodfellow et al., "Generative Adversarial Networks" (2014)

Key stories

An Inspiration in a Bar That Gave Rise to an Early Breakthrough of Generative AI

By Goodfellow’s own recollection, the core conception of the generative adversarial network was born in a late night after a gathering with friends, when, discussing why earlier generative-model methods worked poorly, he suddenly thought to let two neural networks contend against each other — one specially generating fake samples trying to "fool" the other, the other specially learning to tell real from fake, the two improving together in continual contest. He went home that night and wrote the first working prototype code, and this conception was thereafter proven one of the most influential early breakthroughs of generative AI.

Relationships

Mentor

🧭Yoshua Bengio

Disciple

🎨Emad Mostaque　The idea of "generating realistic images with a neural network" the generative adversarial network began was an important theoretical precursor of later generative image models including Stable Diffusion

Echoes today

Below are how modern works borrow or reinterpret this name or story — not the original material. The two differ, so keep them apart.

The double edge of "deepfake" technologyThe generative-image technology the generative adversarial network pioneered thereafter gave rise both to a great many creative applications and to misuse in making convincing "deepfake" content, and this double-edged effect of the technology thereafter became an important topic of public-policy discussion.

Appears in

The Deep-Learning Revival and the Shock of ImageNetThe Late-Night Inspiration of the GAN

← Back to Figures