Neural Networks for Structured Knowledge

Franz Kurfess and Alessandro Sperduti

Course Description

This tutorial presents an overview of recent developments in using neural networks for representation and processing of structured knowledge. To date, mostly symbol-oriented methods have been used for this purpose; these methods, however, have problems with inexact and noisy data, inconsistent knowledge, brittleness, knowledge acquisition, and real-time constraints. Neural networks, on the other hand, are universal approximators, can perform automatic inference (learning), possess very good classification capabilities, and can deal with noise and incomplete data; they can frequently also be used as anytime-methods, where a (possibly non-optimal) result is available anytime during the evaluation, not only at the end.

In recent years, substantial progress has been made towards the use of neural networks for structured knowledge. Structured domains are characterized by complex patterns usually represented as lists, trees, and graphs of variable sizes and complexity. The basic problem is to overcome the limitation imposed by the fixed input size of a network when faced with the task of representing a graph of variable size and with an internal structure. Standard neural networks are well suited for dealing with unstructured patterns, and recurrent neural networks can be used to process sequences; a generalization of a recurrent neuron, the generalized recursive neuron, is capable of representing, classifying, and storing structured information very naturally.

This tutorial will present various approaches for the use of neural networks to deal with structured knowledge.

Prerequisite Knowledge

Attendees should be familiar with basic concepts of neural networks, or have background knowledge in a related area such as machine learning or statistics.

About the Lecturers

Franz Kurfess is co-director of the Software Engineering Lab and associate director of the Electronic Enterprise Engineering program at the Computer and Information Systems Department, New Jersey Institute of Technology. His main area of research is the integration of systems based on various computation methods, e.g. symbolic and connectionist systems, and in particular logic and reasoning with neural networks.

Alessandro Sperduti is an Assistant Professor at the Computer Science Department, University of Pisa, Italy. His research area is neural networks, especially for representing structured information. He has written around 40 refereed papers mainly in the areas of neural networks, fuzzy systems, and image processing.

Last modified: Thu Feb 20 13:36:33 JST 1997