Towards using neural networks to perform object-oriented function approximation

Many computational methods are based on the manipulation of entities with internal structure, such as objects, records, or data structures. Most conventional approaches based on neural networks have problems dealing with such structured entities. The algorithms presented in this paper represent a novel approach to neural-symbolic integration that allows for symbolic data in the form of objects to be translated to a scalar representation that can then be used by connectionist systems. We present the implementation of two translation algorithms that aid in performing object-oriented function approximation. We argue that objects provide an abstract representation of data that is well suited for the input and output of neural networks, as well as other statistical learning techniques. By examining the results of a simple sorting example, we illustrate the efficacy of these techniques.