generative vs discriminative

• generative
  • learn \(p(x,y)\)
  • You have to learn \(p(x)\), which is the distribution for the data, which is very hard
  • Imagine you're learning a face generator. Then, you need to learn the distribution of plausible faces
• discriminative
  • learn \(p(y\mid x)\)
  • just have to learn a few key features