Discriminative
x
│
├── learn p(y | x)
│
└── choose y directly

Generative
y
│
├── learn p(x | y)
├── combine with p(y)
│
└── infer the best y for x