Handling Inconsistencies in Neural Networks

Abstract

As a nonmonotonic learning system, C-IL²P has interesting connections with Belief Revision [Gar92] and, more specifically, Truth Maintenance Systems (TMS) [Doy79] (see Chapter 2 for an overview of TMSs). When background knowledge is encoded as defeasible into a neural network, the set of training examples may specify a revision of that knowledge. In addition, the set of examples can be inconsistent with the background knowledge, and the resulting network can itself contain inconsistencies. In this chapter, we investigate how to detect and treat inconsistencies in the system. We also show the equivalence between C-IL²P networks and TMSs. As a result, the learning process of C-IL²P can be seen as a technique for handling changes of belief in a TMS.