|
|||
|
IMADA - Department of Mathematics and Computer Science |
|
Artificial neural networks have been the subject of massive investigation in the last twenty years. Theoretical studies on architectural and learning issues and experimental evidence are now clearly indicating their potential capabilities and limitations. In a different, apparently unrelated field, the problem of ranking Web pages for information retrieval has been studied giving rise to solutions based on a dynamical systems, which very much resemble typical neural network dynamics. In this talk, I introduce the notion of learning in web domains, which represent an abstraction of the Web, giving insights on the way neural networks and Web page scoring systems can be nicely bridged. Architectural and learning issues are discussed in a new general framework of structured domains, showing the effects of different assumptions on the input graphs. It is pointed out that the proposed framework is better suited for capturing topological features than traditional neural networks processing input vectors. I provide theoretical and preliminary experimental evidence to support the birth of a new wave of thought in neural networks, with intriguing consequences in novel fields like mining the Web, but also in classic fields like pattern recognition, where the importance of providing inputs represented by graphs has already been clearly demonstrated. Host: Klaus Meer
SDU HOME | IMADA HOME | Previous Page Last modified: April 8, 2003. Joan Boyar (joan@imada.sdu.dk) |