GAN (Generative Adversarial Network)

A GAN (Generative Adversarial Network) is a type of deep learning model used to artificially generate data such as images, audio, and text — also known as a "generative adversarial network." Proposed by Ian Goodfellow and colleagues in 2014, it is one of the landmark technologies that accelerated the field of generative AI. A GAN consists of two neural networks: 1. Generator: Creates data that looks indistinguishable from real samples, starting from random noise 2. Discriminator: Tries to tell generated data apart from real data These two networks are in an adversarial relationship — the generator tries to produce increasingly realistic data while the discriminator tries to detect fakes. Through this adversarial training, the generator learns to produce high-quality, realistic outputs. Key characteristics and advantages: • Capable of generating highly realistic images, audio, and video • Operates as an unsupervised learning method requiring no labeled data • Tends to produce more realistic outputs than other generative models such as VAEs Representative applications: • Generating face images and artwork (e.g., This Person Does Not Exist) • Style transfer (Style Transfer, Pix2Pix) • Image inpainting and denoising • Data augmentation to support model training • Synthetic data for medical, manufacturing, and advertising use On the downside, GAN training can be unstable, and issues such as **mode collapse** are known challenges. To address these, many variants have been developed — including WGAN, StyleGAN, and CycleGAN — each suited to specific use cases. GANs are a core technology in generative AI, driving innovation across image processing, content creation, simulation, and digital twin applications.