similarities at every level of abstraction is the key to human memory organization. Accepting this notion requires one to have a means of computing similarities between large numbers of potentially relevant episodic traces, both for memory access and update. The aforementioned hierarchy suggests that goal similarities are crucial, planning-level similarities are almost as important, and similarities across other dimensions are of progressively lower importance. Hence, if memory were organized according to the computational criteria required for metaphor comprehension, it follows that memory searchers for "good metaphors" (those preserving high-invariance properties) require less work (either to generate or comprehend) and prove more rewarding for the understander as they index relevant memory more readily.
Metaphor is a linguistic realization of an inference phenomenon. As such, it should reflect underlying memory structure, as well as suggest the types of inferences people can perform most readily. If we ask why creative metaphors are used, the most logical answer appears to be that the writer is trying to induce the reader to perform the necessary inferences required to comprehend the new material. Metaphor serves as a vehicle to suggest a fruitful domain from which the relevant inferences can be mapped onto the new domain. Hence, when Senator Joe McCarthy referred to Communism as a "dreaded plague," he was inducing the inference in the minds of his listeners that Communism must be actively "eradicated" or it will spread. The metaphor is effective only because the appropriate inference structure was already in existence in the source domain, and McCarthy knew this at the time he created the metaphor.
An interesting avenue of future research is automating metaphor generation. If the model discussed here is correct, metaphor generation requires that the writer have a model of the knowledge state (including goals, strategies) of the reader, as well as an integrated episodic memory where the computed similarity metric incorporates the invariance hierarchy. (That is, two domains are considered similar if the same types of problems and inference processes are present in planning effective behavior in both domains.)
In order to clear possible misconceptions. I emphasize that no distinct, localized, "conscious" existence for the invariance hierarchy is postulated as part of a human memory model. My hypothesis is that the regularities manifested in the hierarchy are epiphenomenal reflections of human memory organization. As such, the invariance hierarchy summarizes a phenomenon that must be explained by comprehensive memory-organization models, and hence it ought to be taken into account in the model-formulation process.
This research was sponsored in part by the Office of Naval Research (ONR) under grant number N00014-79-C-0661, and in part by the Defense Advanced Research Projects