Ian Goodfellow

Overview

Ian Goodfellow is a machine learning researcher who is the first author of "Generative Adversarial Networks", the paper that introduced a way of training generative models by setting two networks against each other^[1]. He has worked at OpenAI, Google Brain, Apple and Google DeepMind, and his name is attached both to the MaxOut network and to GAN work^[2].

Education and early research

Goodfellow earned his PhD in machine learning at Université de Montréal, where his advisors were Yoshua Bengio and Aaron Courville^[4]. The thesis is dated 2014 or 2015^[4]. Both advisors also appear as co-authors on the GAN paper, together with Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley and Sherjil Ozair^[1].

Generative adversarial networks

The paper was submitted to arXiv on 10 June 2014^[1]. Its central idea is to train two models at the same time. A generator produces candidate samples, and a discriminator examines samples and tries to decide whether they came from the training data or from the generator. The two are trained together as a two-player minimax game: the generator is pushed to produce outputs the discriminator cannot separate from real data, while the discriminator is pushed to get better at catching the fakes^[1]. In a minimax game the first player tries to make a quantity small and the second tries to make it large, so neither side has a fixed target — each one's best move depends on what the other is currently doing.

One practical consequence is stated plainly in the paper: training this way needs no Markov chains and no unrolled approximate inference networks^[1]. Those are the pieces that earlier approaches to training generative models tended to rely on, and removing them makes the training procedure comparatively simple to describe and to implement.

Positions

On 31 March 2016, OpenAI's "Team++" page listed him among its full-time members and described him as the author of the MaxOut network and of GAN work^[2]. In April 2017 a Google Research blog post was posted by him in the role of Staff Research Scientist, Google Brain Team^[3].

Wikipedia records that he joined Apple in 2019 as director of machine learning in the Special Projects Group, and that he resigned in April 2022 over Apple's in-person work requirement^[4]. He then joined Google DeepMind as a research scientist^[4].

Influence

Goodfellow's name is associated with two separate pieces of neural network work — the MaxOut network and generative adversarial networks — and OpenAI's team page named both when introducing him^[2]. The GAN paper's lasting contribution is the framing itself: instead of hand-designing a single objective for a generative model, it defines learning as a contest between two networks, one that makes samples and one that judges them^[1]. Because the setup requires no Markov chains and no unrolled approximate inference networks, it gave researchers a comparatively compact recipe for training models that produce data rather than only label it^[1].