This is to announce that there will be a panel discussion at IJCNN'2000(International Joint
Conference on Neural Networks) in Como, Italy this July on the topic:
"DOES CONNECTIONISM PERMIT READING OF RULES FROM A NETWORK?"
The abstract below summarizes the issues/questions to be adrresed by the panel. The following persons will be on the panel:
1) DAN LEVINE
2) LEE GILES
3) NOEL SHARKEY
4) ALESSANDRO SPERDUTI
5) RON SUN
6) JOHN TAYLOR
7) STEFAN WERMTER
8) PAUL WERBOS
9) ASIM ROY
This is the fourth panel discussion at these conferences on the fundamental ideas of connectionism. Further information about IJCNN'2000 is available at the
conference web site:
http://www.ims.unico.it/2000ijcnn.html
Asim Roy
Arizona State University
---------------------------------------------------------------------------
DOES CONNECTIONISM PERMIT READING OF RULES FROM A NETWORK?
Many scientists believe that the symbolic (crisp) and fuzzy (imprecise and vague) rules learned, used and expressed by humans are embedded in the networks of neurons in the brain - that these rules exist in the connection weights, the node functions and in the structure of the network. It is also believed that when humans verbalize these rules, they simply "read" the rules from the corresponding neural networks in their brains. Thus there is a growing body of work that shows that both fuzzy and symbolic rule systems can be implemented using neural networks. This body of work also shows that these fuzzy and symbolic rules can be retrieved from these networks, once they have been learned, by procedures that generally fall under the category of rule extraction. But the idea of rule extraction from a neural network involves certain procedures - specifically the reading of parameters from a network - that are not allowed by the connectionist framework that these neural networks are based on. Such rule extraction procedures imply a greater freedom and latitude about the internal mechanisms of the brain than is permitted by connectionism, as explained below.
In general, the idea of reading (extracting) rules from a neural network has a fundamental conflict with the ideas of connectionism. This is because the connectionist networks by "themselves" are inherently incapable of producing the "rules," that are embedded in the network, as output, since the "rules" are not supposed to be the outputs of connectionist networks. And in connectionism, there is no provision for an external source (a neuron or a network of neurons), in a sense a third party, to read the rules embedded in a particular connectionist network. Some more clarification perhaps is needed on this point. The connectionist framework, in the use mode, has provision only for providing certain inputs (real, binary) to a network through its input nodes and obtaining certain outputs (real, binary) from the network through its output nodes. That is, in fact, the only "mode of operation" of a connectionist network. In other words, that is all one can get from a connectionist network in terms of output - nothing else is allowed in the connectionist framework. So no symbolic or fuzzy rules can be "output" or "read" by a connectionist network. The connectionist network, in a sense, is a "closed entity" in the use mode; no other type of operation, other than the regular input-output operation, can be performed by or with the network. There is no provision for any "extra or outside procedures" in the connectionist framework to examine and interpret a network, to look into the rules it's using or the internal representation it has learned or created. So, for example, the connectionist framework has no provision for "reading" a weight from a network or for finding out the kind of rule/constraint learned by a node. The existence of any "outside procedure" for such a task, in existence outside of the network where the rules are, would go against the basic connectionist philosophy. Connectionism has never stated that the networks can be "examined and accessed in ways" other than the input-output mode.
So there is nothing in the connectionist framework that lets one develop procedures to read and extract rules from a network. So a rule extraction procedure violates in a major way the principles of connectionism by invoking a means of extracting the weights and rules and other information from a network. There is no provision/mechanism in the connectionist framework for doing that.
So the whole notion of rules existing in a network, that can be accessed and verbalized as necessary, is contradictory to the connectionist philosophy. There is absolutely no provision for "accessing networks/rules" in the connectionist framework. Connectionism forgot about the need to extract rules.
DOES ALL THIS RAISE IMPORTANT QUESTIONS ABOUT REPRESENTATION?
