XPost: comp.ai.neural-nets   
   From: QXUXBTVOTSAO@spammotel.com   
      
   I am interested in how nerve nets handle relationships, and am hoping   
   for some references.   
      
   Suppose the task is to look at scenes of a few objects, and produce   
   grammatical descriptions of them, such as: "There is a red square   
   above a blue circle, and a yellow diamond inside the blue circle."   
   I understand how you would go about training a nerve net so it would   
   light a a particular output for any scene with a circle, and another   
   output for any scene with a yellow object, etc.   But what sort of   
   output should the nerve net produce, that would indicate that it   
   recognized that it was the diamond that was yellow, and the circle   
   that was blue?   Or that it recognized that the square was above the   
   circle and not the other way around?   
      
   It is not vision in particular that I am interested in, but in any   
   task where the output needs to describe the relationships in the   
   input.   
      
   If you had an output for every combination: yellow square, red circle,   
   etc., you would need n-squared of them, which may not be practical.   
      
   Here's the sort of output scheme I have in mind:   
   There are outputs, called primary, for red, yellow, blue, black,   
   circle, diamond, square, etc.   
   There are outputs that represent classes of objects, such as a square   
   or diamond that is either yellow or red.   
      
   Thus, if the scene was a red square and a blue circle, the output   
   could be: red, square, blue, circle, square or diamond that is either   
   red or yellow, circle or oval that is either blue or yellow.   Since   
   the yellow, diamond, and oval outputs are not lit, we can tell from   
   this output that it is the square that is red, and the circle that is   
   blue.   
      
   I'm not sure this works.   And it doesn't feel as if my brain does it   
   this way.   But that should give an idea of the kind of thing I'm   
   looking for.    It would be helpful if I even knew the name used to   
   describe the problem, so I could search on it.   
      
   The question of how networks of neurons represent relationships, seems   
   to me a quite fundamental one for understanding how it is that real   
   brains can hold ideas in short-term memory; thus I have taken the   
   liberty of cross-posting to comp.ai as well as comp.ai.neural-nets.   
      
   thanks,   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)