Chapter 4 · The deep learning era (2006–2016)2014
GANs: generator vs discriminator
Two networks compete: one fakes, one detects, and both improve until fakes look real.
Chapter 4 · The deep learning era (2006–2016)2014
Two networks compete: one fakes, one detects, and both improve until fakes look real.
Round 1. The forger draws fakes, but they sit far from the real numbers.
The forger never sees the real numbers. It only learns from the detective's verdicts, and that is enough.
The generator (the forger) makes fake examples from random noise. The discriminator (the detective) is shown a mix of real and fake examples and says which is which.
Press Train or Auto-train. Green dots are real, coral dots are the forger's fakes, the violet curve is how real the detective thinks each spot looks.
Press Train. The forger starts far away from the real numbers.
Try: set detective practice to 0. With nobody judging the fakes properly, the forger gets no useful signal. Then move the real data mid-way and watch the forger chase it. A simplified teaching version: 24 dots, one number to learn.
What does the generator learn from?
The detective scores fakes at about 50%. What does that mean?
What is a common difficulty with GANs?