• Markov random fields provide a convenient method for representing dependencies as well as a search methodology in a fixed site space
  
  
• modeling as a Gibbs distribution yields a dramatic decrease in computations - for an 8x8 block image and 5 states we needed 5⁶⁴ computations, now we need 5*64*iterations
  
  
• it is unclear how/if this could be used for a real-time system in speech; pre-segmenting the data would be necessary
  
  
• a natural fit for the object recognition portion of our USFS project