GANs: the art of the duel
A 4-lesson mini course: why creating is harder than judging, the adversarial game, the math of the equilibrium, and the failure modes that made GANs famous.
Why creating is hard
Judging vs creating
By 2014, networks classified images superbly — judging is a well-posed problem with an obvious loss (were you right?). Creating is different: what's the loss function for 'this generated face looks real'? Pixel-distance to some training image? That punishes perfectly valid NEW faces and rewards blurry averages. Realism refuses to be written down as a formula.
The 2014 generative toolbox
Existing generative models paid heavy prices: Boltzmann machines needed Markov-chain sampling with intractable normalizing constants; variational autoencoders (contemporaneous) optimized a likelihood bound and tended toward blurry samples. Everyone was approximating an explicit probability — and the approximations showed.
The reframe
Goodfellow's move: stop trying to define realism — LEARN it. Make the loss function itself a neural network (the discriminator) that continuously studies the generator's current fakes. The definition of 'looks real' sharpens automatically, exactly as fast as the forger improves. The duel below runs this idea live.
Key takeaway
There's no formula for 'looks real' — so GANs make the loss a learned opponent. The discriminator is an adaptive, self-sharpening definition of realism.
Hands-on — experiment before the quiz
GAN Lab: forger vs detective
Two real networks in a live adversarial duel
100% real: both networks train live with true backpropagation — generator gradients genuinely flow through the discriminator. The instabilities you can trigger (oscillation, mode collapse) are the same ones that made GANs famously hard to train.
Two networks, one duel: the forger (teal) turns random noise into samples; the detective (white curve) rates every point — 1 means 'looks real'. Press play to start the arms race.
Analogy: A counterfeiter and an art detective training each other: every caught fake teaches the counterfeiter, every convincing fake sharpens the detective — until the fakes are indistinguishable.
Checkpoint
Why not train a generator by minimizing pixel distance to training images?