Ptolemy had (…) a universal approximator. He just didn’t know it. With enough epicycles and cycles, one could approximate any orbital system as closely as one wanted. In essence, the mistake (but he couldn’t have known it) was to try to construct a theory to fit the details of the approximation machinery.
Fuzzy logic rules and neural nets are other universal approximators. I wouldn’t go around trying to construct significance into their details, either.
Tha was a commenter at Andrew Gelman´s blog. I find the insight really powerfull. Think, for instance, of IQ and the theory of general intelligence. Or, for that matter, linear regression with interactions and polinomal, or splines or alike.