A scalable genome representation for neural-symbolic networks

Abstract

Neural networks that are capable of representing symbolic information such as logic programs are said to be neural-symbolic. Because the human mind is composed of interconnected neurons and is capable of storing and processing symbolic information, neural-symbolic networks contribute towards a model of human cognition. Given that natural evolution and development are capable of producing biological networks that are able to process logic, it may be possible to produce their artificial counterparts through evolutionary algorithms that have developmental properties. The first step towards this goal is to design a genome representation of a neural-symbolic network. This paper presents a genome that directs the growth of neural-symbolic networks constructed according to a model known as SHRUTI. The genome is successful in producing SHRUTI networks that learn to represent relations between logical predicates based on observations of sequences of predicate instances. A practical advantage of the genome is that its length is independent of the size of the network it encodes, because rather than explicitly encoding a network topology, it encodes a set of developmental rules. This approach to encoding structure in a genome also has biological grounding.