본문 바로가기

ML-Paper Summary

Paper Summary - Auto-Encoding Variational Bayes

Original Paper: Kingma, Diederik P, and Max Welling. 2014. “Auto-Encoding Variational Bayes.” In ICLR 2014 : International  Conference on Learning Representations (ICLR) 2014.

 

Objective

  • Probabilistic model with countinuous latent variable
  • Continuous variables leading to interactable distrubution. Effecient posterior inference. 
  • Works in inractability, when we cant track marginal likelihood.
  • Available to train on mini-batch

AEVB Algorithm

 Method of updating parameters of AE using gradient descent. 

 

Variational lower bound

 

SGVB Estimator

 For a chosen posterior q, using the SGVB estimator reparameterizes the random variable z=g(e, x) with e a noise variable. We make the latent variable z through random noise and apply it to the variational lower bound technique. 

Variational Auto Encoder

VAE is a method using the AEVB algorithm to optimize its parameters and the location-scale zipped distrubution that chooses the standard deviation as e and let g(.)=location + e * scale. Where the location(mu) and scale(sigma) are the model outputs. The KL divergence is calculated as the formula below. 

 The first sigma term is the KL divergence.