A Metric Entropy Bound is Not Sufficient for Learnability
We prove by means of a counterexample that it is not sufficient, for probably approximately correct (PAC) learning under a class of distributions, to have a uniform bound on the metric entropy of the class of concepts to be learned. This settles a conjecture of Benedek and Itai.
Index TermsLearning estimation PAC metric entropy class of distributions
Unable to display preview. Download preview PDF.
- 6.R. M. Dudley, “A course on empirical processes,” Lecture Notes in Math. Vol. 1097. New York: Springer, 1984, pp. 1–142.Google Scholar
- 9.S. R. Kulkarni, “Problems of computational and information complexity in machine vision and learning,” Ph.D. thesis, Dep. Elec. Eng. Comput. Sci., M.I.T., June 1991.Google Scholar