[ Home  |  FAQ-Related Q&As  |  General Q&As  |  Answered Questions ]

    Search the Q&A Archives

I would like to use cross-entropy calculation instead of the...

<< Back to: comp.ai.neural-nets FAQ, Part 7 of 7: Hardware

Question by jonmach
Submitted on 12/8/2003
Related FAQ: comp.ai.neural-nets FAQ, Part 7 of 7: Hardware
Rating: Not yet rated Rate this question: Vote
I would like to use cross-entropy calculation instead of the standard 'Generalized delta rule'.  How does this change the calculation of the error at each level of propagation through the network.

In other words, if the normal delta at the output layer is (desired-actual)*actual*(1-actual), then what needs to change in determining the error at the hidden nodes, and associated weights if the cross-entropy delta calculation is only (desired-actual)


Your answer will be published for anyone to see and rate.  Your answer will not be displayed immediately.  If you'd like to get expert points and benefit from positive ratings, please create a new account or login into an existing account below.


Your name or nickname:
If you'd like to create a new account or access your existing account, put in your password here:
Your answer:

FAQS.ORG reserves the right to edit your answer as to improve its clarity.  By submitting your answer you authorize FAQS.ORG to publish your answer on the WWW without any restrictions. You agree to hold harmless and indemnify FAQS.ORG against any claims, costs, or damages resulting from publishing your answer.

 

FAQS.ORG makes no guarantees as to the accuracy of the posts. Each post is the personal opinion of the poster. These posts are not intended to substitute for medical, tax, legal, investment, accounting, or other professional advice. FAQS.ORG does not endorse any opinion or any product or service mentioned mentioned in these posts.

 

<< Back to: comp.ai.neural-nets FAQ, Part 7 of 7: Hardware

[ Home  |  FAQ-Related Q&As  |  General Q&As  |  Answered Questions ]
© 2008 FAQS.ORG. All rights reserved.