Generative adversarial network
An AI system that uses two competing programs: one acts like a forger trying to create fake items, while the other acts like a detective trying to spot the fakes. By constantly challenging each other, the forger eventually gets so good that its work is nearly impossible to tell apart from the real thing.
A machine learning framework consisting of two neural networks, a generator and a discriminator, locked in a zero-sum game. The generator learns to map random noise to a target data distribution, while the discriminator learns to classify inputs as real or synthetic, forcing the generator to produce increasingly realistic outputs.
A generative modeling architecture defined by a minimax game between two differentiable functions, G and D, where G learns to map a latent space to a data distribution p_g that minimizes the Jensen-Shannon divergence from the true data distribution p_data, while D is simultaneously optimized to maximize the probability of correctly assigning labels to both training examples and samples from G.
evolution
- 2014-06 · historyOriginal GAN Paper
Ian Goodfellow and colleagues introduced the GAN framework, establishing the adversarial training process between a generator and a discriminator.
- 2015-11 · historyDCGAN
Deep Convolutional GANs were introduced, stabilizing training and enabling the generation of high-quality images.
- 2017-10 · historyProGAN
NVIDIA researchers introduced progressive growing of GANs, allowing for the generation of high-resolution, photorealistic images.
- 2018-12 · historyStyleGAN
NVIDIA introduced a style-based generator architecture that allowed for unprecedented control over image features and synthesis quality.