No subject

[Guo-Zheng Sun]

   >I am interested in constructive  algorithms  for  neural  networks 
   >(capable of generating new units and layers in the NN).  Is  there 
   >any constructive algorithm for  NN  whose  units  have  continuous 
   >(e.g.: sigmoidal) activation function?
 
  The  works of this kind back to 1987 may also include:
 (1)G.Z. Sun, H.H. Chen and Y.C. Lee, "A Novel Net that Learns Sequential
Decision Processes", NIP Proceedings (1987),Denver.
 (2) G.Z. Sun, H.H.Chen and Y.C.Lee, "Parallel Sequential Induction Network:
A New Paradiagm of Neural Network Architecture", ICNN Proceedings,(1988),
San Diego.
 These two papers generalized the idea of conventional regression tree to
generate the new  NN units dynamically. The continuous  activation 
fuction is used. This  PSIN (Parallel Sequential Induction Net) algorithm 
integrates the parallel processing of neural net and sequential processing 
of decision  tree together to form a "most efficient" multilayered neural net 
classifier. The efficiency measure is the information entropy. 
  
 -G. Z. Sun
 University of Maryland