From: ted.dunning@gmail.com   
      
   On Apr 1, 2:46 am, castiro...@gmail.com wrote:   
   > Question: In Neural and Bayesian nets, is every node updated on every   
   > clock pulse?  Or are selective ones?  I am thinking that you could   
   > (not a pro.) update weights O( f ) less and grow the number of   
   > connections O( f ).  What kind of advanced AI is going on?   
      
   You use the word "clock pulse" which tends to imply that you think   
   that nodes are updated in real time.   
      
   That is a very unusual way to operate either kind of network.   
      
   It is much more common to use the training data to derive the weights   
   and structure of the network and then freeze the network at that   
   point.  One major reason for this is simply that if you have an   
   important application, then you probably don't want to take the chance   
   of your system going nuts without warning.  This can happen for any   
   number of humdrum and silly reasons that have little to do with   
   mathematical truths and proofs.   
      
   As such, training algorithms are really just mathematical constructs   
   and the question of when and how often node weights are updated has   
   very little to do with anything except possibly the overall cost of   
   training the network.   
      
   There are classes of machine learning algorithms known as online   
   algorithms that are designed to be practical to do updates to the   
   system as new data arrive, but with neural networks at least, on-line   
   updates really devolve to batch updates performed periodically.  This   
   is entirely practical for most applications and it allows extensive   
   quality analysis to be done on the resulting model before it is   
   deployed.   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)