RANDOMNESS IN GENERALIZATION ABILITY - A SOURCE TO IMPROVE IT

Among several models of neurons and their interconnections, feedforwar d artificial neural networks (FFANN's) are most popular, because of th eir simplicity and effectiveness. Some obstacles, however, are yet to be cleared to make them truly reliable-smart information processing sy stems, Difficulties such as long learning time and local minima may no t affect FFANN's as much as the question of generalization ability, be cause a network needs only one training, and then it may be used for a long time, The question of generalization ability of ANN's, however, is of great interest for both theoretical understanding and practical use, This paper reports our observations about randomness in generaliz ation ability of FFANN's, A novel method for measuring generalization ability is defined, This method can be used to identify degree of rand omness in generalization ability of learning systems, If an FFANN arch itecture shows randomness in generalization ability for a given proble m, multiple networks can be used to improve it, We have developed a mo del, called voting model, for predicting generalization ability of mul tiple networks, It has been shown that if correct classification proba bility of a single network is greater than half, then as the number of networks in a voting network is increased so does its generalization ability, Further analysis has shown that VC-dimension of the voting ne twork model may increase monotonically as the number of networks in th e voting network is increased, This result is counter intuitive, since it is generally believed that the smaller the VC-dimension, the bette r the generalization ability.