Generative AI
Generative Adversarial Networks (GANs)
Watch a generator and discriminator compete, learning to produce realistic data through an adversarial game.
how two networks improve each other by competing, with the discriminator grading realism and the generator learning to fool it.
Stage 1 of 5: The Objective
Iteration: 0. D Loss: 0.693. G Loss: 0.693.
- Real Data
- Generated Data
- D Confident (Real)
- D Confident (Fake)
The goal of a GAN is to generate data that perfectly matches the real distribution. Here, the real data consists of two distinct clusters.
Check your understanding
1 questions in the bank. Each attempt draws a fresh set in a fresh order, so a second go is a real second go.
The Adversarial Game
A GAN consists of two networks locked in a game:
- The Generator tries to produce fake data that looks real.
- The Discriminator tries to tell the real data apart from the fakes.
In this simplified simulation, the real data consists of points clustered in two regions. The Generator starts by mapping random noise to a single spot, guessing blindly.
As they train together, the Discriminator learns the boundary separating the clusters, and the Generator uses that feedback to push its generated samples toward the regions the Discriminator thinks are real.
Mode Collapse
A common failure mode in GANs is Mode Collapse. If the Generator finds one type of sample (one mode) that consistently fools the Discriminator, it might abandon learning any other types. It produces highly realistic samples, but with zero diversity.
Reference
- Objective
- min_G max_D V(D,G)
- D Loss
- Binary Cross-Entropy
Break it on purpose
Mode Collapse