generative adversarial network
In plain wordsTwo neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people.
A method for training a generative modelgenerative model, n.: A model that learns what a kind of data looks like — faces, sentences, songs — well enough to produce new examples that resemble the originals without copying them. by setting two networks against each other. A generator G turns random noise vectors z into synthetic samples; a discriminator D, shown both real training examples and G's output, scores how likely each input is to be real. D learns to sort the two correctly and G learns to get its fakes accepted, so one network's gain is the other's loss (a minimax game). In the idealized analysis the contest ends when G's samples are distributed exactly like the data and D can only guess, outputting ½ everywhere. GANs generate in a single forward passforward pass, n.: The step where a neural network takes an input and runs it through its layers, one after another, to produce an output — its answer or prediction. and produced the sharpest images of the late 2010s (DCGAN, conditional GANs, StyleGANStyleGAN, prop. n.: An image-generating network from NVIDIA, famous for photorealistic faces of people who do not exist, that sets coarse traits such as head angle and face shape separately from fine details such as coloring and skin texture., BigGAN), but they are notoriously hard to train — oscillation, vanishing generator gradients, mode collapsemode collapse, n.: When a generator gets stuck producing only a few kinds of output — the same face, the same joke — instead of the full variety it was meant to learn. — and give no likelihood. From about 2021 diffusion modelsdiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static. displaced them for most media generation, but adversarial losses survive inside audio codecsneural audio codec, n.: A neural network that squeezes sound into a very compact stream of numbers and rebuilds it, used both to send audio cheaply and to let AI models 'write' speech and music piece by piece., image tokenizers such as VQGANVQGAN, n.: An image compressor that turns a picture into a small grid of codebook 'tiles' while keeping it sharp, used as the vocabulary of many image generators and of an early wave of AI art., super-resolutionsuper-resolution, n.: Making a small or blurry image larger and sharper by having AI fill in plausible fine detail, rather than just stretching the pixels. models and diffusion distillationdiffusion distillation, n.: Training a fast image generator to imitate a slow diffusion model, so that a picture that once took fifty steps can be made in one to four..
We switched from a GAN to a diffusion model and stopped babysitting the discriminator.
Origin
Introduced by Ian Goodfellow and seven co-authors at the Université de Montréal in a paper posted in June 2014 and presented at NeurIPS that year. Jürgen Schmidhuber has argued that his 1990–1991 'artificial curiosity', in which one network learns to predict another's outputs, anticipated the idea; Goodfellow disputes the equivalence. Olli Niemitalo sketched a similar, never-implemented idea in a 2010 blog post.
Usage note
Pronounced as a word ('gan'). The adversarial game here is distinct from adversarial examplesadversarial example, n.: An input altered on purpose, often too slightly for people to notice, so that an AI model gets it badly wrong — like a panda photo that a classifier confidently calls a gibbon. and adversarial trainingadversarial training, n.: Toughening an AI model by deliberately showing it attacks during training — the tricky inputs that fool it — along with the right responses, so it learns to resist them. in machine-learning security, which concern inputs crafted to fool a classifier.
Related
- generative modelgenerative model, n.: A model that learns what a kind of data looks like — faces, sentences, songs — well enough to produce new examples that resemble the originals without copying them.
- mode collapsemode collapse, n.: When a generator gets stuck producing only a few kinds of output — the same face, the same joke — instead of the full variety it was meant to learn.
- Wasserstein GANWasserstein GAN, n.: A version of the forger-versus-detective image generator that measures how far its fakes are from real data in a smoother way, making training steadier and easier to monitor.
- StyleGANStyleGAN, prop. n.: An image-generating network from NVIDIA, famous for photorealistic faces of people who do not exist, that sets coarse traits such as head angle and face shape separately from fine details such as coloring and skin texture.
- diffusion modeldiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static.
- image-to-image translationimage-to-image translation, n.: Turning one picture into another version of itself — a sketch into a photo, a day scene into night, a horse into a zebra — while keeping its layout.
- discriminative modeldiscriminative model, n.: A model that learns only to tell things apart — spam or not spam, cat or dog — without learning how to produce new examples of those things.
- deepfakedeepfake, n.: A fake but realistic video, photo or voice recording made with AI that puts words in a real person's mouth or shows them in a scene that never happened.
- variational autoencodervariational autoencoder, n.: A neural network that learns to squeeze data into a compact, slightly fuzzy code and rebuild it, arranged so that decoding random codes produces new, plausible examples.
- diffusion modeldiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static.
Sources
- Goodfellow, I. J. et al. (2014). Generative Adversarial Networks. arXiv:1406.2661 (NeurIPS 2014). arxiv.org/abs/1406.2661
- Goodfellow, I. (2016). NIPS 2016 Tutorial: Generative Adversarial Networks. arXiv:1701.00160. arxiv.org/abs/1701.00160
Mentioned in
adversarial exampleadversarial example, n.: An input altered on purpose, often too slightly for people to notice, so that an AI model gets it badly wrong — like a panda photo that a classifier confidently calls a gibbon., deepfakedeepfake, n.: A fake but realistic video, photo or voice recording made with AI that puts words in a real person's mouth or shows them in a scene that never happened., diffusion modeldiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static., generative AIgenerative AI, n.: AI that produces something new on request, such as an essay, a picture, a song or a piece of code, having absorbed from enormous numbers of existing examples what such things typically look like., AdamAdam, prop. n.: The most widely used method for training neural networks. It smooths the direction of updates over time and automatically sets a separate step size for each setting based on how large its recent changes have been., adversarial machine learningadversarial machine learning, n.: The study of how machine-learning systems can be fooled, corrupted or spied on by someone working deliberately against them, and of how to build systems that hold up., adversarial trainingadversarial training, n.: Toughening an AI model by deliberately showing it attacks during training — the tricky inputs that fool it — along with the right responses, so it learns to resist them., Alec RadfordAlec Radford, prop. n.: An American AI researcher who led OpenAI's first GPT language models and its CLIP and Whisper systems, key steps toward today's chatbots and multimodal AI., discriminative modeldiscriminative model, n.: A model that learns only to tell things apart — spam or not spam, cat or dog — without learning how to produce new examples of those things., Fréchet inception distanceFréchet inception distance, n.: A score for image generators that compares the overall statistics of their pictures with those of real photos, as seen by an image-recognition network; lower is better., game theorygame theory, n.: The mathematics of strategic decisions, where each player's best move depends on what the others do. It is used to study competition and cooperation among people, firms, nations and AI agents., generative artgenerative art, n.: Art in which the artist invents a process — dice rolls, software, written instructions — and lets that process make the work; the phrase now also loosely covers pictures from AI image generators., generative modelgenerative model, n.: A model that learns what a kind of data looks like — faces, sentences, songs — well enough to produce new examples that resemble the originals without copying them., Ian GoodfellowIan Goodfellow, prop. n.: An American computer scientist who invented generative adversarial networks, in which two neural networks compete to produce realistic fake images, and who pioneered research on fooling AI systems., image-to-image translationimage-to-image translation, n.: Turning one picture into another version of itself — a sketch into a photo, a day scene into night, a horse into a zebra — while keeping its layout., Jürgen SchmidhuberJürgen Schmidhuber, prop. n.: A German computer scientist who co-invented the LSTM, a neural network with memory that powered much of speech recognition and translation in the 2010s, and who argues that early pioneers go under-credited., latent spacelatent space, n.: The hidden internal map a model builds of its data, where each item becomes a point and similar items sit near each other. Moving smoothly through it gradually changes what the model represents or generates., MilaMila, prop. n.: Montreal's deep-learning hub, where professors and students from several Quebec universities work side by side. It grew from the lab Yoshua Bengio started in 1993 and is among the world's largest academic AI centers., mode collapsemode collapse, n.: When a generator gets stuck producing only a few kinds of output — the same face, the same joke — instead of the full variety it was meant to learn., NeurIPSNeurIPS, prop. n.: The largest yearly scientific conference on artificial intelligence and machine learning, held each December, where many of the field's landmark papers were first presented., StyleGANStyleGAN, prop. n.: An image-generating network from NVIDIA, famous for photorealistic faces of people who do not exist, that sets coarse traits such as head angle and face shape separately from fine details such as coloring and skin texture., super-resolutionsuper-resolution, n.: Making a small or blurry image larger and sharper by having AI fill in plausible fine detail, rather than just stretching the pixels., tanhtanh, n.: An S-shaped curve, like the sigmoid but centered on zero, that squeezes any number into a value between −1 and 1. It was the standard neuron rule before ReLU took over., variational autoencodervariational autoencoder, n.: A neural network that learns to squeeze data into a compact, slightly fuzzy code and rebuild it, arranged so that decoding random codes produces new, plausible examples.