Iconic Language Representation in a Recursive Neural System

In a totally reduced view of the interaction of a learning machine with a “world”, such a machine can receive world-state information, it can act on the world and it can receive data from the world that controls learning. We see the learning machine as a classical state machine whose states are its constructs of the experienced world. Learning is a process of developing state structure that is related to taking appropriate actions in the world. There are only two ways that states can acquire an assignment — unsupervised and supervised. The difficulty with unsupevised learning is that should the machine ever be required to describe some of its own constructs there will need to be internal decodings of the arbitrarily assigned states. We believe that this does not remove the need for some states having arbitrary assignments but the only way that the machine can have an “understanding” of the world is through the following alternative way of assigning values to states. It is suggested that the states of the world are best represented through a many-to-some relationship between the state variables of the world and the state variables of the state machine. As the perception of the states of the world can only occur through some senses our definition of an iconic state is ‘a state assigned to the state variables of the state machine in such a way that it represents the sensory pattern that is generated by a world state’.