Figure 3.

A simplified geometric sketch of possible mappings between two domains, for example meanings and signals. These mappings can be unstructured, random and incompressible (A), or highly structured and compressible (B). An individual attempting to learn the latter could use similarity structure in one domain to predict what the appropriate generalization should be for unseen points. A further possibility is a degenerate mapping, which is the simplest and most compressible of all (C).