Relational neurocomputing

Knowledge may be characterized in terms of the relations between the elements of the relevant knowledge base. The development of cognitive powers in a machine requires the ability for the machine to be able to learn and use such relations. A machine controlled by a neural network processing such abilities will be termed a `relational neurocomputer´. The purpose of this paper is to describe some possible principles on which such a style of neural computation might be based. The problem faced by any machine attempting to be effective in achieving its goals in its environment is that of building useful representations of the significant objects around it, and manipulating such representations to achieve effective actions. What determines that significance, and hence what gets represented and how, will be strongly determined by the actions available to the machine. We conclude that any relations between the elements of the machines knowledge base, and even in the elements themselves, are importantly determined by the actions that are used in discovering the relations. Support for an action-based approach to the development of cognition, especially the learning and use of relations, is seen in the action-scripts based nature of infant category learning. There is also evidence for this position from the frontal lobe storage of verbs, and other linguistic elements for relations, near the premotor convex and the position of storage of action schemas