When I lived in SF area, I often passed an AA gentleman fishing by the bay. He   
   had his fishing pole jammed in a rock pile. He made a point of saying, "Let   
   your pole do the fishing." After a while, it became a joke, as I would beat   
   him to the saying.   
      
   Maybe in another time frame and society, he would have been an AI engineer.   
   Maybe   
   not.   
      
   I never did fully satisfy myself what he meant, but in the case of knowledge   
   discovery by computer-sensor arrangements, all that is needed are heuristics   
   rather than certain certainties of which we are certain, certainly.   
      
   IOW, a way to find things out that works 'sometimes' is OK. Extra points if the   
   'sometimes' heuristic can report to itself it's success or failure -   
      
   In the case of neural nets, 95% of the work is done with something called a   
   Widrow-Hoff three layer back prop network. That is the work horse of pattern   
   discovery with ANN.   
      
   Here is the simple way in which it's success is self-reported. (though there   
   are   
    many others.) 20% of the training set is "held out" during training, eg, it is   
   not used for the training. What happens during training (by the 80% set) is   
   that   
   the hold out set accuracy peaks and then declines. The decline is of course   
   'over fitting.'   
      
   So the question is, does anyone have their favorite heuristic for looking at   
   data?   
      
   AND - It could be how you tell a cup of coffee is hot by feeling the cup   
   or looking at the steam.   
      
   But could also be imaging, spectroscopy, etc.   
      
   I don't expect a response here, but random thoughts OK.   
      
   jb   
      
   PS - And of course, ANN are not a general panacea. If ANN were really good,   
   they would be used for the pattern recognition part of speech recognition.   
   But they aren't - They play a supporting role in phoneme detection only.   
   The heart of the system is HMM - another AI paradigm entirely.   
   But ANN are easy to use and think about. When I interviewed at Google, the   
   question was "Design a complete email spam detector. (now)." White   
   boarding such a thing becomes easy with ANN as a component.   
   j   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)