Memory and learning in a meso level reasoning system

In earlier work, we built on the notion of feature spaces to develop a formal paradigm for representation and reasoning that is intermediate between symbolic and neural paradigms. In our model, states of the system, that is, objects or examples of concepts, can be represented as multi-spectral images. A metric is defined on these states in terms of the energy needed to transform image intensity patterns. The metric in turn is used to define dynamics which implement the two fundamental reasoning activities of categorisation and composition. In this paper, we show how all the parameters used in the dynamics can be derived from a memory for exemplars. Such a memory allows for learning through experience.