PROBABILISTIC PERCEPTRONS

A probabilistic perceptron is a device that computes a representation of a predicate ICI as a linear threshold function. In contrast to the classical perceptron, this representation will only be correct with pr obability close to 1. For the construction of this device the method o f superimposed coding is used. The resulting advantages in comparison to the original perceptron may be summarized as follows: there is a tr ivial learning theorem; the size of the coefficients remains small; th e condition of linear separability is not required. It is argued that the advantages mentioned outweigh the disadvantages. To support this c laim, experimental work concerning full text retrieval is presented. E n passant a qualitative comparison between several information retriev al methods is obtained.