Brady@UDEL.EDU (11/15/87)
I am interested in inferring concepts from data, and have been reading about back propagation in neural nets as a way to make such inferences. I am confused about the little red riding hood article in BYTE. The article seems to suggest that the nodes in the middle layer (representing the concepts wolf, granny, woodcutter) are INFERRED during training. Other literature on back propagation that I have seen also suggest that concepts can be inferred that way. But a look at the BPSIM code that implements the little red riding hood network seems to suggest the existance of these three nodes before training begins. So my question is: if one wants to infer concepts from data, can one do that by using back propagation? Or do you still have to a priori anticipate the existance of the concepts? [I haven't seen the example in question, but the usual neural network learning procedure does use predefined nodes. The nodes of the center layer are identical except for random variations in the initial weights. After training, these nodes take on very different roles characterized by their weight vectors. Determining what these roles are can be quite difficult, so it is not clear how much of the inference is done by the network and how much by the human -- but clearly the network has done part of the work. This strategy permits nodes to be deleted (via zeroed weights), but not created. For creation of nodes you may have to investigate genetic learning algorithms. -- KIL]
tsai@POLLUX.USC.EDU.UUCP (11/16/87)
In article <8711151153.aa02040@Dewey.UDEL.EDU> Brady@UDEL.EDU writes: >I am confused about the little red riding hood article in BYTE. >The article seems to suggest that the nodes in the middle layer > ..... and KIL's comment follows: > This strategy permits nodes to be > deleted (via zeroed weights), but not created. For creation of nodes > you may have to investigate genetic learning algorithms. -- KIL] ^^^^^^^^^^^^^^^^^^^^^^^^^^^ I am interested in these genetic learning algorithms used in a neural network implementaion. Can somebody in the Netland gives me some references? Please response by e-mail to me. Thanks in advance! Y. C. Tsai :-) tsai@pollux.usc.edu fot Internet, {sdcrdc,cit-cav}!uscvax!tsai for UUCP EE-Systems, University of Southern California, Ca. 90089-0781