Perturbation analysis for EP

• Question

• 

  

  0

  Sign in to vote

  Hi,
  
  Consider the clutter problem of Tom Minka where the goal is to get posterior of the mean of one Gaussian component. The likelihood is a mixture of two Gaussians, one of which is just noise. The other component is the true data generator. Say I use EP to approximate this likelihood term with one Gaussian by moment matching.
  
  - What happens to the inferred mean if I perturb the mean/variance of this approximate Gaussian ?
  - Which one affects the inferred mean more between having an inaccurate mean (in approximate factors) or an inaccurate variance ?
  - In general, is there some work that analyzes how EP behaves as we vary approximate factors ?
  
  Thanks,
  Wittawat
  

  Friday, November 21, 2014 12:57 PM