A Connectionist Variation of Inheritance

A connectionist architecture is outlined which makes use of RAAM to generate representations for objects in inheritance networks and extended learning to make such representations context-sensitive. The architecture embodies inheritance quite diierently by relying on associative similarities and regions in representational space. The model avoids many of the problems identiied for traditional inheritance. 1 Introduction Inheritance reasoning is concerned with transfer of properties or behaviours, sometimes referred to as defaults, between objects organised according to their class in networks. The obvious advantages are concerned with (i) avoidance of representing all knowledge explicitly by reasoning with defaults, and that (ii) inheritance generates reasonable assumptions when knowledge is incomplete. The mechanisms by which such transfer is handled are traditionally based on transi-tivity and measures derived from the network representation. Judging from the number of attempts to resolve situations that may occur in inheritance networks, reasoning with defaults is problematic 5]. A so-called exception contradicts some of the properties attached to one of its classes. Hence, an inference mechanism based solely on transitivity cannot be sound since it may return contradicting results for a member of the exception. Intuitively, an exception cancels inheritance of a particular property from its classes and thus an object should prefer to inherit the property of the nearest class. If a network contains redundant connections , a more general inheritance principle is required. A particular object should prefer to inherit the most speciic properties. That is, whenever there is a connict the property of the most specialized class should be preferred 5]. In Fig. 1 two inheritance networks are shown. The rst contains two exceptions and is problematic if a rst-order reasoner is used. The second network contains both exceptions, redundant links and multiple inheritance. To resolve such situations many have resorted to use representational and inferential measures to preempt some paths in the network 5]. Many of the attempts to formalise the notion of inheritance are indirectly constrained by accounting for the particular form of representation and the computational consequences of a particular formal model of inheritance. The model-freedom of connectionist systems relieves us from making such decisions. Con-nectionist systems acquire their own internal representation of objects through learning. Hence, representational inconsistencies, in a logical sense, can be accounted for and incorporated continuously. Appropriately conngured connection-ist systems generalise to account for data outside a training set. Hence, reasoning about objects, of which no knowledge is explicitly available, …