Postprint version. Published in Proceedings of the International School on Neural Networks. Fifth Course: From Synapses to Rules: Discovering Symbolic Rules from Neural Processed Data: Erice, Sicily, Italy, February 25, 2002, pages 275-292.
Knowledge representation and reasoning methods in artificial intelligence almost exclusively rely on symbol-oriented methods: Statements describing aspects and objects of the system to be modeled are represented through symbols (mostly text strings), and these symbols are stored in a computer, and manipulated according to the inference rules prescribed by the reasoning method. This works reasonably well in situations where knowledge is available in explicit form, typically through experts or written documents. In situations where knowledge is only available implicitly, e.g. in large data sets, other methods, often based on statistical approaches, have been used more successfully. Many of these methods are based on neural network techniques, which typically represent and process knowledge at a level below symbols; this is often referred to as sub-symbolic representation. This contribution discusses approaches to integrate symbol-oriented reasoning methods with sub-symbolic ones into hybrid systems.