A committee machine with a set of networks composed of two single-threshold elements as committee members

摘要

We propose a committee machine whose each committee member is a network of two threshold elements. This improves upon the ability of the usual committee machine. The subsidiary discriminant function in this case is a kind of piecewise linear discriminant function instead of a linear function. We show a sophisticated representation of the discriminant function realized by the proposed committee machine, and give a rational learning algorithm based on the function. We also show that the proposed committee machine, which is a kind of three-layer network of threshold elements, cannot be equivalently transformed to a two-layer network. This fact asserts that the proposed committee machine provides a family of discriminant functions which is intrinsically wider than that of the usual committee machine.