Experience has shown that both types of systems have complementary strengths and weaknesses. In particular, connectionist models (e.g., neural networks) are generally massively parallel (and thus efficient) but opaque, whilst symbolic models (e.g., decision trees) are often slower but more comprehensible. As a result of this synergy, many researchers have turned their attention to the design of hybrid connectionist-symbolic systems, as well as other forms of hybridisation (e.g., genetic algorithms and neural netowrks).
In this talk, I will review a small sample of hybrid systems we have developed in the past few years. These are chosen to show some of the benefits of hybridisation. In particular, I will present two types of connectionist/symbolic models, one motivated by efficiency and the other by anatomy. As an exemplar of the first class, I will use AA1*, a constructive, incremental learning algorithm for binary classification. As an exemplar of the second class, I will use BRAINN, a symbolic reasoning system implemented on a Hopfield network architecture based on the human cortex. In addition, I will briefly mention other forms of hybridisation, including GA-RBF, a system using a genetic algorithm to design the hidden layer of an RBF network prior to its training and FLARE, a learning algorithm that integrates lazy and eager learning.
Computer Science Home Page