NICE: NON-LINEAR INDEPENDENT COMPONENTS ESTIMATION
Laurent Dinh, David Krueger, Yoshua Bengio
Abstract
modeling complex high-dimensional densities
basic assumption: a good representation is one in which the data has a distribution that is easy to model
maps it to a latent space so as to make the transformed data conform to a factorized distribution
computing the determinant of the Jacobian and inverse Jacobian is trivial, yet maintain the ability to learn complex non-linear transformations
(a) maxlogpX(x)=logpH(f(x))+log|det(∂f∂x)| max log p X ( x ) = log p H ( f ( x ) ) + log | det ( ∂ f ∂ x ) |
(b) h∼pH(h),x=f−1(h) h ∼ p H ( h ) , x = f − 1 ( h )
(c) maxxHlogpX((xO,xH)) max x H log p X ( ( x O , x H ) )
Introduction
ask the learner to find a transformation h=f(x) h = f ( x ) of the data into a new space such that the resulting distribution factorizes, i.e., the components hd h d are independent
h=f(x) h = f ( x ) , f f is invertible,
and x x have the same dimension