ICML2008/Nonextensive Entropic Kernels

From GusWiki

Jump to: navigation, search

André Martins

  • Kernels are functions from pairs of data points to real values (Kernel : Point*Point -> Real). They are useful in density estimation and can serve to define similarity measures.
  • Instead of points, sometimes we define kernel between distributions (e.g. bag-of-words).
  • One entropic kernel is defined based on the symmetrized KL divergence.
  • a more general family of entropies is Tsallis entropies
  • improves performance in some tasks
Personal tools