This paper investigates the generation of neural networks through the induction of binary trees of threshold logic units (TLUs). Initially, we describe the framework for our tree construction algorithm and show how it helps to bridge the gap between pure connectionist (neural network) and symbolic (decision tree) paradigms. We also show how the trees of threshold units that we induce can be transformed into an isomorphic neural network topology. Several methods for learning the linear discriminant functions at each node of the tree structure are examined and shown to produce accuracy results that are comparable to classical information theoretic methods for constructing decision trees (which use single feature tests at each node), but produce trees that are smaller and thus easier to understand. Moreover, our results also show that it is possible to simultaneously learn both the topology and weight settings of a neural network simply using the training data set that we are initially given.
展开▼
