Home▸AI & Machine Learning▸Generative Adversarial Networks Simulation

🧪 Generative Adversarial Networks Simulation

Watch a real generator and discriminator MLP train against each other from scratch: the discriminator's confidence field renders as 3D terrain, generated points climb toward the real data ring, and you can trigger mode collapse by cranking the learning rate.

AI & Machine Learning2DModerate60 FPS📱 Mobile-adapted

This simulation illustrates the principles of Generative Adversarial Networks by actually running one: a generator and discriminator, each a small multilayer perceptron trained with hand-written backpropagation, compete in real time in your browser. The discriminator's confidence score across the 2D data plane is rendered as living 3D terrain — height and color track D(x) exactly — while cyan points mark the fixed real data distribution and magenta points show where the generator's noise-driven output currently lands. Raise the learning rate to watch the two networks destabilize each other and trigger mode collapse; lower it and watch the generator patiently climb toward every mode of the real distribution as the adversarial game approaches a Nash equilibrium.

⚙ Under the hood

This simulation illustrates the principles of Generative Adversarial Networks (GANs), a type of machine learning model used for generating new data that resembles training data. It showcases how two networks compete to improve each other.

GANsGenerative Adversarial

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install