Inductive Simplicity and the Matrix

Gilbert Harman, Sanjeev R. KulkarnieScholarship 2003

Abstract

In certain statistical learning problems, a policy of choosing simpler rules that account fairly well for data is likely to have less error on new cases than a policy of choosing complex rules that have less error on the data.The relevant kind of simplicity is not to be measured in terms of the number of parameters needed to specify a given member of a class of rules but might be measured in terms of the VC dimension of such a class.The rationale for using simplicity so measured can be extended to allow simplicity to decide among empirically equivalent hypotheses.The extended rationale provides reasons of simplicity to reject certain sorts of philosophical skepticism.
BibTeX
@inproceedings{harman2003inductive,
  author = {Gilbert Harman and Sanjeev R. Kulkarni},
  title = {{Inductive Simplicity and the Matrix}},
  booktitle = {eScholarship},
  year = {2003},
}

Loading related content...

Figure