Ian Goodfellow
Last reviewed
Jun 5, 2026
Sources
22 citations
Review status
Source-backed
Revision
v2 · 2,413 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
Jun 5, 2026
Sources
22 citations
Review status
Source-backed
Revision
v2 · 2,413 words
Add missing citations, update stale details, or suggest a clearer explanation.
Ian Goodfellow is an American computer scientist and machine learning researcher who is best known for inventing the generative adversarial network (GAN) in 2014 and for being the lead author of the textbook Deep Learning (2016), written with Yoshua Bengio and Aaron Courville. His research spans generative modeling, the security of machine learning systems, and the study of adversarial examples, inputs crafted to fool neural networks. Over his career he has held senior research roles at Google Brain, OpenAI, Apple, and Google DeepMind. [1][2]
Goodfellow was born in 1987. The GAN framework he introduced became one of the most widely used approaches to image synthesis in the second half of the 2010s, and it underlies much of the technology later associated with synthetic media and so-called deepfakes. [1][3]
| Field | Detail |
|---|---|
| Born | 1987 |
| Nationality | American |
| Fields | Machine learning, deep learning, generative modeling, adversarial machine learning |
| Education | Stanford University (BS, MS, 2009); Université de Montréal (PhD, 2015) |
| Doctoral advisors | Yoshua Bengio and Aaron Courville |
| Known for | Generative adversarial networks; Deep Learning textbook; adversarial examples; maxout; CleverHans library |
| Institutions | Google Brain; OpenAI; Apple; Google DeepMind |
Goodfellow grew up in California and graduated from San Dieguito High School Academy in Encinitas in 2004. In high school he spent three years on the debate team, an experience he has credited with sharpening his ability to build and defend an argument. [16][17]
He enrolled at Stanford University intending to prepare for a career in neuroscience, but he found the biology and chemistry coursework difficult and shifted his focus to computer science. A 2006 summer internship at the National Institutes of Health, where he worked with a machine learning model for classifying electroencephalography (EEG) signals, gave him an early exposure to the field. He has said that he did not commit fully to artificial intelligence research until a friend, Ethan Dreyfuss, introduced him to deep learning, after which the question of how to make deep networks work kept his attention. [16][17]
At Stanford, Goodfellow worked with Andrew Ng and with the computer vision researcher Gary Bradski. He carried out independent-study research on computer vision for the Stanford AI Robot using traditional machine learning, and he spent summers interning on robotics-related vision work through Stanford's CURIS undergraduate research program and at Willow Garage, a robotics company near the campus. He earned both a bachelor's degree and a master's degree in computer science from Stanford in 2009. He has recounted that, like many people who later went into the field, he was rejected from a Google internship as a student. [1][2][16][17]
He then moved to Canada for doctoral study at the Université de Montréal, completing a PhD in machine learning in February 2015 under the supervision of Yoshua Bengio and Aaron Courville. He has said he chose Montreal because few advisors at the California schools he considered, including Stanford and Berkeley, supported deep learning research at the time, whereas Bengio offered him a place in his lab. In 2013 he received what he has described as the first Google PhD Fellowship in deep learning. His dissertation was titled "Deep learning of representations and its application to computer vision," and he defended it in 2014. Montreal, through Bengio's lab (later part of the Mila research institute), was one of the centers of the deep learning research that expanded rapidly during those years, and Goodfellow's doctoral work placed him at the heart of it. [1][4][5][16]
The idea for generative adversarial networks came to Goodfellow in 2014 while he was still a doctoral student in Montreal. According to accounts he has given since, the moment arrived during an evening at Les 3 Brasseurs, a Montreal bar, where he was out with fellow students celebrating a friend's graduation. Some of them were discussing a project on getting a computer to generate realistic photographs on its own, an approach that, as they described it, would have required enormous amounts of statistical bookkeeping. Goodfellow argued that the real obstacle was not a programming problem but a question of algorithm design. [3][6]
His insight was to pit two neural networks against each other. One network, the generator, produces candidate samples such as images. A second network, the discriminator, tries to tell the generator's output apart from real data drawn from a training set. The two are trained together in a competition: the generator improves by learning to fool the discriminator, while the discriminator improves at catching fakes. As training proceeds, the generator's samples grow more realistic. Goodfellow went home that night, wrote the code, and got the system working quickly; he has recounted that he produced recognizable handwritten digits from the MNIST dataset in roughly an hour of work. He has also acknowledged a degree of luck in the episode, saying that if the code had not worked that night he might have abandoned the idea. [3][6][18]
The result was published as the paper "Generative Adversarial Nets," posted to arXiv on 10 June 2014 and presented at the Neural Information Processing Systems (NIPS, later NeurIPS) conference later that year. Its authors were Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. The work was completed under time pressure, roughly two weeks before a conference deadline, with several of Goodfellow's labmates helping to run the experiments. [5][7]
GANs were quickly taken up across the field and became a standard tool for generating images, video, and other high-dimensional data. The technique drew prominent praise: Yann LeCun, a leading figure in deep learning, called adversarial training "the coolest idea in deep learning in the last 20 years." A 2018 profile in MIT Technology Review nicknamed Goodfellow "the GANfather" in recognition of the method's influence. The same approach also became the technical basis for synthetic and manipulated media, and the term deepfake came into common use to describe such content. By the mid-2020s the original GAN paper had accumulated tens of thousands of academic citations, making it one of the most cited papers in machine learning. [3][19]
The following are among Goodfellow's most influential works. Citation counts are approximate and grow over time.
| Year | Work | Contribution |
|---|---|---|
| 2013 | "Maxout Networks" (ICML) | Introduced the maxout activation, designed to work well with dropout; set classification records on MNIST, CIFAR-10, CIFAR-100, and SVHN. [20] |
| 2013 | "Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks" | A unified convolutional model for reading multi-digit street numbers directly from images, later used to transcribe addresses for Google Maps. [21] |
| 2014 | "Generative Adversarial Nets" (NIPS) | Introduced the GAN framework; one of the most cited works in modern machine learning. [7] |
| 2015 | "Explaining and Harnessing Adversarial Examples" (ICLR) | Argued that the linear behavior of neural networks explains their vulnerability to adversarial examples and introduced the fast gradient sign method (FGSM) for generating them, along with adversarial training as a defense. [22] |
| 2016 | Deep Learning (MIT Press) | A comprehensive textbook of the field, written with Bengio and Courville. [8] |
Goodfellow is the lead author of Deep Learning, published by MIT Press in November 2016 and written together with Yoshua Bengio and Aaron Courville. The book offers a broad treatment of the field, covering the mathematical background, established network architectures, and research directions, and it became one of the most cited reference works in the area. The authors made the full text available to read for free online in addition to the print edition. [8][9]
He also wrote the chapter on deep learning (Chapter 21) for the fourth edition of Artificial Intelligence: A Modern Approach, the textbook by Stuart Russell and Peter Norvig that is widely used in university courses worldwide. [1][2]
A second major strand of Goodfellow's work concerns the robustness and security of machine learning models. He was among the early researchers to study adversarial examples, inputs altered in ways that are often imperceptible to people but that cause a trained classifier to make confident mistakes. This line of work showed that high-performing neural networks could be brittle in ways that matter for safety and security, and it helped establish adversarial machine learning as a research subfield. [1][2]
In the 2015 paper "Explaining and Harnessing Adversarial Examples," written with Jonathon Shlens and Christian Szegedy, Goodfellow proposed that the main reason neural networks are vulnerable to small adversarial perturbations is their largely linear behavior in high-dimensional spaces. The paper introduced the fast gradient sign method (FGSM), a simple and fast way to construct adversarial inputs, and it advanced adversarial training, in which a model is trained on adversarial examples in order to become more robust. FGSM became one of the most widely used baselines in the field. [22]
At Google, Goodfellow led a "red team" that probed machine learning systems for adversarial weaknesses, and he was among the first to study the security and privacy of neural networks. To support reproducible study of these vulnerabilities, he co-created CleverHans, an open-source software library, with Nicolas Papernot and other collaborators. The library provides standard reference implementations of attacks and defenses so that researchers can benchmark how robust a model is in a consistent way, since results obtained with ad hoc attack code are difficult to compare across papers. [2][10]
During his time at Google he also contributed to applied systems. The multi-digit recognition work he led allowed Google Maps to transcribe street addresses automatically from imagery captured by Street View vehicles, a system that reached high accuracy on complete street numbers, and he highlighted security weaknesses in deployed machine learning systems. [1][21]
After finishing his PhD in 2015, Goodfellow joined the Google Brain research team. In March 2016 he left Google to become one of the early research staff at OpenAI, the research laboratory founded in late 2015. He stayed there for roughly a year before returning to Google in March 2017 as a senior research scientist on Google Brain, where he remained until 2019. He has explained the return by noting that his focus on adversarial examples and related areas such as differential privacy led him to collaborate mostly with colleagues at Google. [1][2]
In 2019 he moved to Apple as director of machine learning in the company's Special Projects Group, the unit associated with Apple's special projects including its automotive work. He left Apple in 2022. His departure drew attention because it was reported to be connected to the company's return-to-office policy, which required employees to begin working on site for part of each week on a phased schedule. In a note to staff quoted in press coverage, Goodfellow wrote, "I believe strongly that more flexibility would have been the best policy for my team." [11][12]
Later in 2022 he joined Google DeepMind as a research scientist, reported to be working within the deep learning group led by Oriol Vinyals. At DeepMind his work was reported to include applying reinforcement learning to the control of nuclear fusion plasmas and research on the factual reliability of large language models. According to his own LinkedIn profile, he subsequently left Google around 2025 to work on a startup; as of 2026 details of that venture have not been publicly described. [4][13][14]
In 2017 Goodfellow was named to MIT Technology Review's list of 35 Innovators Under 35, in the artificial intelligence and robotics category, cited for having "invented a way for neural networks to get better by working together." In 2019 he was included in Foreign Policy's list of 100 Global Thinkers. His publications, led by the GAN paper and the Deep Learning textbook, are among the most heavily cited in modern machine learning, and his work is frequently referenced in discussions of the history of generative AI. [1][15]