Learning in Gibbsian fields: how accurate and how fast can it be?

摘要: Gibbsian fields or Markov random are widely used in Bayesian image analysis, but learning Gibbs models is computationally expensive. The computational complexity pronounced by the recent minimax entropy (FRAME) which use large neighborhoods and hundreds of parameters. In this paper, we present a common framework for models. We identify two key factors that determine accuracy speed models: efficiency likelihood functions variance approximating partition using Monte Carlo integration. propose three new algorithms. particular, interested maximum satellite estimator, makes set precomputed called "satellites" to approximate functions. This algorithm can approximately estimate model textures seconds HP workstation. performances various algorithms compared our experiments.