Recently, I am reading the paper "Neighbourhood Components Analysis". At the last paragraph on the second page of the paper,

"Notice that by learning the overall scale of A as well as the relative directions of its rows we are also effectively learning a real-valued estimate of the optimal number of neighbours (K). This estimate appears as the effective perplexity of the distributions $p_{ij}$"

1.Is effective perplexity the perplexity?

2.I really can't see the reason. How may I get the formulation of K in your paper and connect it to effective perplexity please?

3.The equation of KL-divergence on the Euq.(6) seems different as in other books. How should I understand it?

I am so sorry if this is not the right place to ask questions about a paper. I asked this question on StackOverflow as well. I contacted authors of the paper but get no answers so far as usual.

Thank you in advance for your help.

  • $\begingroup$ Are you aware of stats.stackexchange ? It may be more suitable there, but wait and see what happens here for a little bit (I'm not optimistic though) $\endgroup$ – David Roberts Mar 15 '18 at 6:17
  • $\begingroup$ It is a machine learning paper and I am not sure stats.stackexchange a good place either. Thank you for your kind help though @DavidRoberts. $\endgroup$ – Tengerye Mar 15 '18 at 10:00
  • $\begingroup$ In that case, datascience.stackexchange.com $\endgroup$ – David Roberts Mar 15 '18 at 13:55

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.