Abstract
Information present in any training set of vectors for machine learning can be interpreted in two different ways, either as whole states or as individual atomic units. In this paper, we show that these alternative information distributions are often inherently incongruent within the training set. When learning with a Boltzmann machine, modifications in the network architecture can select one type of distributional information over the other; favouring the activation of either state exemplar or atomic characteristics. This choice of distributional information is of relevance when considering the representation of knowledge in logic. Traditional logic only utilises preference that is the correlate of whole state exemplar frequency. We propose that knowledge representation derived from atomic characteristic activation frequencies is the correlate of compositional typicality, which currently has limited formal definition or application in logic. Further, we argue by counter-example, that any representation of typicality by 'most preferred model semantics' is inadequate. We provide a definition of typicality derived from the probability of characteristic features; based on neural network modelling.