Skip to content

Flow

Generative Flow

For a probability density p(x) and a transformation of coordinate x=g(z) or z=f(x), the density can be expressed using the coordinate transformations, i.e.,

(1)p(x)=p~(f(x))|detDg(f(x))|1(2)=p~(f(x))|detDf(x)|

where the Jacobian is

Dg(z)zg.

The operation gp~(z) is the push forward of p~(z). The operation g will pushforward simple distribution p~(z) to a more complex distribution p(x).

  • The generative direction: sample z from distribution p~(z), apply transformation g(z);
  • The normalizing direction: "simplify" p(x) to some simple distribution p~(z).

The key to the flow model is the chaining of the transformations

detDf(x)=Πi=1NdetDfi(xi)

where

(3)xi=gig1(z)(4)=fi+1fN(x).

  1. Liu X, Zhang F, Hou Z, Wang Z, Mian L, Zhang J, et al. Self-supervised Learning: Generative or Contrastive. arXiv [cs.LG]. 2020. Available: http://arxiv.org/abs/2006.08218 


Contributors: LM