about the reconstruction error in each level of RBMs, is there any rule for that ? how much the reconstruction error must be ? is it correct to say that this error must be low ? i implement DBN on the MINST dataset but the recostruction error is so high , with this error cay we say the DBN is working good?

asked Aug 10 '13 at 15:51

905914401's gravatar image

905914401
1112


2 Answers:

Check out this question, it is quite similar for yours. I suppose you do learning with contrastive divergence. Reconstruction error is not what you actually optimize with contrastive divergence.

answered Aug 16 '13 at 09:47

Midas's gravatar image

Midas
42151017

The reconstruction error need not be low or always go down, but usually with CD1 it does go down a bit at least through the first part of training. See this guide on training RBMs.

answered Aug 23 '13 at 18:03

gdahl's gravatar image

gdahl ♦
341453559

thank you so much

the reconstruction error in dbn is based on MSE , so how it can be more than 1 when our input is in the range [0,1] ?

(Aug 26 '13 at 06:46) 905914401
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.