Generation as a Goal
Generative vs Discriminative
9 min+85 XPStep 1/10
Two ways to learn from data
Every model you've met so far has been a critic: given input x, predict label y. Cat or dog, spam or not — the model learns where the boundary between classes lies. These are discriminative models.
A generative model asks a bigger question: what does the data itself look like? Instead of finding the boundary between faces and non-faces, it learns the full distribution of faces — every pose, lighting, and expression — well enough to produce a new one.
analogy
A discriminator is an art critic — it can tell a Monet from a forgery. A generator is the forger — it has to be able to paint one. The forger's job is strictly harder.