# Ian Goodfellow > 1985– · Computer Scientist, Creator of GANs > > **Recorded contribution:** Generative Adversarial Networks (GANs, 2014); adversarial examples ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation American computer scientist Ian Goodfellow introduced generative adversarial networks (GANs) in a 2014 paper with seven co-authors and also contributed influential work on adversarial examples, deep learning, and model security. GANs reframed generative modeling as a two-player game between a generator and discriminator, producing sharp samples and a new family of research problems. The GAN reframed generation as a two-player learning problem. A generator did not need an explicitly normalized probability density; instead, it improved by producing samples that a simultaneously trained discriminator found increasingly difficult to distinguish from the data. ## 2. The problem inherited Generative models often required explicit likelihoods or difficult approximate inference and struggled to produce realistic high-dimensional samples. ## 3. The central contribution Goodfellow originated the GAN framework, training a generator indirectly through a discriminator that learns to distinguish data from generated samples. ## 4. Reconstruct the mechanism 1. Sample noise and transform it with a generator into a candidate observation. 2. Train a discriminator to separate real training examples from generated examples. 3. Update the generator using gradients that reward outputs the discriminator classifies as real. 4. Alternate the two updates and evaluate diversity and fidelity rather than relying on loss alone. ## 5. What changed downstream - GANs enabled major advances in image synthesis, translation, super-resolution, and representation learning. - Adversarial training dynamics exposed new questions about equilibrium, evaluation, dataset imitation, and synthetic-media misuse. - GANs produced major advances in image synthesis, translation, super-resolution, and representation learning and sharpened research on implicit models and adversarial optimization. ## 6. Attribution, limits, and uncertainty - The original paper has eight authors, and GANs build on game theory, neural networks, and earlier implicit generative models. - Training can be unstable and collapse modes; visually convincing samples can encode bias, violate consent or rights, and support deception. - The minimax game can cycle or collapse to limited modes, discriminator success may starve gradients, visual fidelity can conceal memorization, and generated media creates consent and provenance risks. ## 7. Reconstruction lab Train a one-dimensional GAN on a mixture of two Gaussians. Plot generator coverage over time and identify mode collapse even when discriminator loss appears stable. Track sample diversity as well as discriminator loss, deliberately induce mode collapse, and test whether the model reproduces unusually close training examples. ## 8. Evidence trail - [Generative Adversarial Nets](https://arxiv.org/abs/1406.2661) — NeurIPS - [Explaining and Harnessing Adversarial Examples](https://arxiv.org/abs/1412.6572) — ICLR - [Ian Goodfellow](https://en.wikipedia.org/wiki/Ian_Goodfellow) — Wikipedia contributors --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*