• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

Article

Вычислимые комбинаторные оценки вероятности переобучения

Journal of machine learning and data analysis. 2013. Vol. 1. No. 6. P. 734-743.
K.V. Vorontsov, Sokolov E., Frey A.

Computable combinatorial data dependent on generalization bounds are studied. This approach
is based on simpli ed probabilistic assumptions: it is assumed that the instance space is nite,
the labeling function is deterministic, and the loss function is binary. A random walk across
a set of linear classi ers with low error rate is used to compute the bound eciently. The
experimental evidence to con rm that this approach leads to practical over tting bounds in
classi cation tasks is provided.