Ian Goodfellow

wiki
, , , ,
1

American deep learning researcher; inventor of Generative Adversarial Networks, co-pioneer of adversarial machine learning, and lead author of the canonical Deep Learning textbook.

Basic Information / Profile

Field Details
Full Name Ian J. Goodfellow
Born 1987, United States
Nationality American
Current Status Stealth startup (2025–present)
Research Fields Deep learning, generative models, adversarial machine learning, differential privacy, AI safety, fusion plasma control
PhD Advisors Yoshua Bengio, Aaron Courville
PhD Thesis Deep Learning of Representations and its Application to Computer Vision (Université de Montréal, 2014)
Personal Website iangoodfellow.com
X / Twitter @goodfellow_ian
GitHub github.com/goodfeli
Google Scholar scholar.google.com/citations?user=iYN86KEAAAAJ

Overview

Ian Goodfellow is an American computer scientist best known for inventing Generative Adversarial Networks (GANs) in 2014 — a training framework in which two neural networks compete to produce synthetic data indistinguishable from real samples, and which became the dominant paradigm for generative image modelling for nearly a decade. Trained under Yoshua Bengio at the Université de Montréal, he subsequently held research positions at Google Brain, OpenAI, Google Research, Apple, and Google DeepMind, along with a director role in Apple’s Special Projects Group. He is lead author of the MIT Press textbook Deep Learning (2016), co-authored with Bengio and Aaron Courville, which is used in over 1,500 universities across 135 countries and remains the field’s most widely assigned reference. He also co-pioneered the study of adversarial examples and adversarial training for robustness, and contributed to early work on differentially private neural network training. In 2025 he left Google DeepMind to found a stealth startup whose focus had not been made public as of mid-2026.

Early Life & Education

High School and Early Interest in AI

Goodfellow grew up in the United States and attended San Dieguito Academy High School in California. His interest in AI was kindled as a teenager through popular science coverage — particularly Scientific American — predicting that AI and nanotechnology would be the defining technologies of the 21st century. A 2006 internship at the National Institutes of Health, where he first encountered a machine learning model applied to EEG classification, sharpened that interest into a specific research direction. He changed his major to computer science and began studying under Andrew Ng.

Stanford University — BSc and MSc (2004–2010)

Goodfellow completed a Bachelor of Science and a Master of Science in Computer Science at Stanford University under Andrew Ng, working primarily on the Stanford AI Robot project in computer vision with traditional machine learning methods. During his master’s work he began engaging with deep learning through a colleague, and this catalysed his decision to pursue a doctorate in Bengio’s group. He has described learning about deep learning as the moment he became fully committed to AI research.

Université de Montréal — PhD (2010–2014)

Goodfellow was accepted into Yoshua Bengio’s lab at the Université de Montréal (home of the LISA group, the precursor to Mila), completing his doctorate in early 2014 (formally February 2015) under the joint supervision of Bengio and Aaron Courville. His thesis, Deep Learning of Representations and its Application to Computer Vision, explored representation learning, deep Boltzmann machines, and multi-prediction models. He contributed substantially to Theano and Pylearn2, the lab’s open-source deep learning software stack, which gave practitioners the infrastructure to run novel ideas quickly — a culture he has credited as central to the environment in which GANs were eventually conceived.

The GAN idea came to him approximately two weeks before the NeurIPS 2014 deadline. He recruited a group of labmates who dropped other work to run experiments across multiple datasets within that window, proving the framework learned the correct distribution and enabling submission. He has described this sprint as one of the formative collaborative experiences of his career.

Career

Google Brain — Research Scientist (2014–2016)

After his doctorate, Goodfellow joined Google Brain as a research scientist. He led the Google Brain Red Team, a group focused on adversarial machine learning and the security vulnerabilities of neural networks. In this role he was co-author of one of the first two papers to independently identify adversarial examples — inputs deliberately perturbed to cause misclassification by neural networks — and with Christian Szegedy co-invented adversarial training as a robustness defence. He also developed a system enabling Google Maps to automatically transcribe addresses from Street View photographs, and collaborated with Nicolas Papernot, Kunal Talwar, Ulfar Erlingsson, and Martin Abadi on some of the first publications on differentially private training mechanisms for neural networks.

OpenAI — Research Scientist (2016–2017)

In March 2016 Goodfellow left Google to join OpenAI as one of the organisation’s first employees, attracted by its stated mission of developing AI that benefits humanity broadly. He stayed for approximately eleven months before returning to Google Research in March 2017.

Google Research / Google Brain — Research Scientist (2017–2019)

During his second tenure at Google, Goodfellow continued work on adversarial robustness and generative modelling, contributing to the growing literature on GAN training stability, conditional generation, and the theoretical understanding of the minimax objective.

Apple — Director of Machine Learning, Special Projects Group (2019–2022)

In April 2019 Goodfellow joined Apple as Director of Machine Learning in its Special Projects Group, a confidential hardware and software skunkworks division. He has described the environment as unusually collaborative: because the project was confidential, team members had strong incentives to cooperate rather than compete for individual credit. In April 2022 he resigned, publicly citing Apple’s requirement for employees to return to in-person work — a departure that attracted wide coverage as a statement about post-pandemic workplace policy in the technology sector.

Google DeepMind — Research Scientist (2022–2025)

In July 2022 Goodfellow joined Google DeepMind as a research scientist in Oriol Vinyals’s Deep Learning team. His primary focus shifted significantly from generative modelling toward AI for fusion power generation, a direction he had sought out after reading DeepMind’s 2022 Nature paper on using reinforcement learning to control tokamak plasma configurations. At DeepMind his contribution to this effort was in plasma simulation: he wrote numerical differential equation solver code underpinning the physical models needed to train the reinforcement learning agent, rather than the RL system itself. He co-developed TORAX, an open-source differentiable plasma physics simulator written in JAX and released on GitHub, described in an arXiv paper in 2024. He also worked on LLM factuality research. In an April 2025 interview with Mila, he described himself as still at Google DeepMind; Wikipedia notes he left in 2025 to found a startup.

Stealth Startup (2025–present)

As of mid-2026, Goodfellow’s LinkedIn profile listed him as working at a stealth startup. No further details about the company’s focus or funding have been made public.

Key Contributions

  • Generative Adversarial Networks (GANs) (NeurIPS 2014, arXiv:1406.2661, with Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville, and Bengio) — Introduced a training framework in which a generator network and a discriminator network are trained simultaneously in a minimax game, enabling synthesis of highly realistic images and other data; remained the dominant paradigm for generative image modelling for nearly a decade and gave rise to StyleGAN, CycleGAN, conditional GAN, and many other variants. One of the most-cited papers in the history of computer science.
  • Adversarial Examples (2013–2014, co-discovered with Christian Szegedy and collaborators) — Co-authored two of the foundational papers establishing that neural networks are susceptible to imperceptibly small input perturbations that reliably cause misclassification; the finding opened the field of adversarial machine learning.
  • Fast Gradient Sign Method (FGSM) — Introduced a computationally efficient one-step method for generating adversarial examples by perturbing inputs in the direction of the gradient of the loss; became the standard baseline attack in adversarial robustness research.
  • Adversarial Training (with Christian Szegedy) — Co-invented adversarial training — augmenting training data with adversarial examples — as the primary empirical defence against adversarial attacks; remains one of the few defences that has not been subsequently broken.
  • Differentially Private Deep Learning (2016, with Abadi, Chu, Erlingsson, McMahan, Mironov, Papernot, Raghunathan, Ramage, Song, and Talwar) — Co-authored a foundational paper on training neural networks with differential privacy guarantees, introducing the moments accountant for tighter privacy analysis; a core reference for privacy-preserving machine learning.
  • Maxout Networks (ICML 2013, with Goodfellow, Warde-Farley, Mirza, Courville, Bengio) — Proposed a novel activation function designed to work well with dropout, enabling better generative model training and achieving state-of-the-art results on several benchmarks.
  • Pylearn2 — Led the development and popularisation of Pylearn2, the Université de Montréal’s research-grade deep learning library built on Theano; provided the software infrastructure for many early deep learning breakthroughs in the Bengio group.
  • Deep Learning textbook (MIT Press, 2016, with Yoshua Bengio and Aaron Courville) — Lead-authored the first comprehensive graduate-level textbook on deep learning; freely available at deeplearningbook.org and used as the primary text in more than 1,500 universities across 135 countries.
  • Deep Learning chapter, Artificial Intelligence: A Modern Approach (4th edition, 2020, Russell and Norvig) — Wrote the deep learning chapter for the canonical AI textbook, bringing coverage of neural networks into the most widely used undergraduate AI text.
  • TORAX — Tokamak Transport Simulator (2024, with Google DeepMind) — Co-developed an open-source differentiable plasma physics simulator in JAX designed to support reinforcement learning research for fusion power generation; released publicly as part of DeepMind’s fusion effort.

Awards & Recognition

  • MIT Technology Review Innovators Under 35 (2017) — Recognised for the invention of GANs and contributions to adversarial machine learning.
  • Foreign Policy 100 Global Thinkers (2019) — Included for work on deep learning and AI safety.

Key Relationships

  • Yoshua Bengio — Primary PhD supervisor; the intellectual environment of Bengio’s LISA lab — with its open-source culture, freedom to experiment, and emphasis on deep learning theory — directly enabled the conditions in which GANs were conceived.
  • Aaron Courville — PhD co-supervisor and co-author of the Deep Learning textbook; Courville has remained an ongoing collaborator in generative modelling research.
  • Andrew Ng — Undergraduate and master’s research supervisor at Stanford; introduced Goodfellow to machine learning through the Stanford AI Robot project and a sequence of courses.
  • Christian Szegedy — Google Brain colleague with whom Goodfellow co-discovered adversarial examples and co-invented adversarial training; the two worked independently on the phenomenon before joining forces.
  • Martin Abadi — Google Brain collaborator on differentially private neural network training; also co-designed TensorFlow.
  • Nicolas Papernot — Google Brain collaborator and co-author on privacy and adversarial robustness papers; became a leading independent researcher in trustworthy ML.
  • Oriol Vinyals — Team lead for Goodfellow’s Deep Learning team at Google DeepMind; whose research group oversaw the fusion plasma control work.
  • Jean Pouget-Abadie — Montréal PhD colleague who contributed game-theoretic analysis to the GAN framework; Goodfellow has cited him as the right collaborator for GAN’s game-theoretic formulation.

Personal Style

Goodfellow is notable in the AI research landscape for combining prolific technical output with an unusually candid public voice on the culture and consequences of the field. In a 2025 interview he described the pre-deep-learning research era at Montréal as more open-ended and curiosity-driven, and said he misses “the sense of freedom to try different things, and less of a sense that our work has consequences” — a double-edged reflection from someone whose most famous paper enabled deepfakes. His career moves have often been motivated by explicit values rather than prestige: he joined DeepMind because a paper on fusion power caught his attention, and he resigned from Apple publicly over a remote-work principle rather than quietly. He describes his research approach through Bruce Lee’s “10,000 kicks” aphorism, preferring deep mastery of a small core toolkit — Python, numerical programming, linear algebra, probability — over breadth, and advocating explicitly for knowing when to seek collaborators with complementary expertise rather than trying to acquire every skill oneself.

References