Loading...
Please wait, while we are loading the content...
Similar Documents
NNES: A Neural Network Explanation System for Transforming Trained Neural Networks into Decision Trees
| Content Provider | Semantic Scholar |
|---|---|
| Author | Moody, John E. Hanson, Steve J. |
| Copyright Year | 1995 |
| Abstract | Development of an intelligent forensic system for hair analysis and comparison. 14 neural network. NNES provides a means to make this conversion. By taking a trained neural network and examining the IM values of the hyperplanes of that network, NNES is able to create a decision tree representation of the network. 8 Acknowledgements We would like to thank Jim Hooman for his insightful comments and ideas which made this research more complete than it might have otherwise been. We would also like to thank Rich Relue for reviewing an earlier draft of this paper and providing us with a great many useful comments. We used software distributed with McClelland and Rumelhart, 1988 ] for many of our simulations. 13 6 Related Work Quite a bit of work has been previously done in describing various means of extracting information from neural networks. A few such approaches are brieey described below. One system for extracting symbolic rules from a neural network is that of Towell and Shavlik, 1993 ]. Their MofN method extracts a series of m-of-n rules based on the weights and biases of the hidden and output units. m-of-n rules are those in which at least M elements of a rule consisting of N elements are true. Craven and Shavlik, 1993 ] also incorporate the m-of-n concept into a query-based rule extraction algorithm. Another system for extracting rules from a neural network is Validity Interval Analysis (VIA), described by Thrun, 1993 ]. Here, rules are determined based on an analysis of the relations of node activation values. One of the main diierences between VIA and NNES (aside from the fact that VIA extracts rules while NNES extracts trees) is that VIA veriies the extracted rules. This allows for a greater degree of accuracy. A third algorithm is the KT algorithm of Fu, 1991 ]. KT extracts production rules from a network. It distinguishes between positive attributes (those which enable activations to approach 1) and negative attributes (those which send activations towards 0). These attributes are then analyzed with regard to the network and the rule information is extracted from this analysis. KT and NNES are similar in that they both work on any back propagation neural network. One major diierence is that as part of the rule-extraction process used by KT, it analyzes the certainty and conndence level s of the various rules. The NNES method, in contrast, makes no … |
| File Format | PDF HTM / HTML |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
