Acquisition of internal representation by multilayered perceptrons
Bunpei Irie, Mitsuo Kawato · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 1991
Abstract One of the characteristics of the PDP model (multilayer perceptron with backward error propagation learning) is thought to be its ability to automatically find a mapping between inputs and outputs based on input‐output vector samples describing that mapping. However, since in the mapping the values of representative points are determined by table lookup, and the values of other points are determined by interpolation between them, it can be viewed as one type of Memory Based Reasoning. On the other hand, it is thought that the PDP model has the ability to extract features from the input vector. In this paper, the forementioned two points are investigated completely and it is demonstrated that entirely different results are obtained when the internal representation differs, even if the input data are exactly the same. It is stipulated that information processing in the PDP model is a process that carries out a transformation which represents the input vector distribution in an optimal metric space (internal representation) and interpolates that metric space. This internal representation acquisition capability is demonstrated by means of a simple simulation.