We show that for any class of functions H which has a reasonable combinatorial dimension, the vast majority of small subsets of the combinatorial cube can not be represented as a Lipschitz image of a subset of H, unless the Lipschitz constant is very large. We apply this result to the case when H consists of linear functionals of norm at most one on a Hilbert space, and thus show that "most" classification problems can not be represented as a reasonable Lipschitz loss of a kernel class. © Springer-Verlag Berlin Heidelberg 2005.
CITATION STYLE
Mendelson, S. (2005). On the limitations of embedding methods. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 3559 LNAI, pp. 353–365). Springer Verlag. https://doi.org/10.1007/11503415_24
Mendeley helps you to discover research relevant for your work.