WhatAreNormalizingFlows
Title: What are Normalizing Flows? Sources: https://www.youtube.com/watch?v=i7LjDvsLWCg
- Bi-jecttion from known sample (Usually Gaussian) to complicated thing your trying to model.
- Bi-jection is a composition of many other bijections (each of which is invertable)
- Trick is setting up a bijections that is by construction invertable and has many pramaters that can be learned.
Done with Coupling Layers. split data at that layer into two contiguous regions. (A and B) yA = xA yB = h(xB; Theta(xA))
Results in bijections that with a Jacobian that is triangular => |Det| is production of diagonals (These preserve area)
- Separate scaling matrix to change area
- Trick to reduce the diminionsality of latent space (above has the same as the input data) Can enforce smaller latent dim. with something called "Real NVP"