From: mailbox@dmitry-kazakov.de   
      
   On Wed, 15 Oct 2003 15:50:32 +0100, "William Silvert"   
    wrote:   
      
   >I came late to this discussion, having been away for a while, but I find the   
   >topic surprising. I did some work with a colleague from the Netherlands on   
   >using neural networks to generate fuzzy memberships, using as a training set   
   >the assignments of a team of experts, so I find the two approaches, fuzzy   
   >logic and neural networks, complementary rather than competitive.   
      
   Exactly   
      
   >About Dmitry's comment that neural nets are not tunable, it surprises me   
   >that no one has addressed this limitation.   
      
   I just reformulated "NN have no parameters". Whether it is an   
   advantage is abother question. But if we have no parameters there is   
   nothing to tune (directly at least).   
      
   >I recall once attending a talk   
   >where a NN was used to identify people from photographs, and with a training   
   >set of 10,000 photos it could make correct identification whether or not the   
   >subjects wore glasses. I asked whether the NN could be told to ignore   
   >glasses, which would reduce the training set to 5,000 pictures, but was told   
   >that couldn't be done. Why not?   
      
   That would not be easy for any system.   
      
   >As for whether expert systems require training sets, I would ask where the   
   >experts get their expertise. It seems to me that the expertise can be viewed   
   >as the result of processing the data in what could be considered a training   
   >set. If we could teach a NN to ignore eyeglasses it would become a bit more   
   >like an expert.   
      
   Right, it will have a knowledge about how "eyeglasses" and "persons"   
   are related. A model of the world.   
      
   ---   
   Regards,   
   Dmitry Kazakov   
   www.dmitry-kazakov.de   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)