Menu ▾ ▴

#6 New Stopping criteria: change in error per n epochs

open
nobody
None
5
2012-12-24
2010-01-21
Anonymous
No

New stopping criteria for supervised learning: Stop when the absolute rate of change in the average squared error per epoch is sufficient small for n iterations, where n can be customized.

Implementation in the SupevisedLearning.doLearningEpoch() in the same condition where the totalNetworkError is checked.

Discussion

  • Bruce Wootton

    Bruce Wootton - 2010-01-23

    Doing this will require an extra variable inside SupervisedLearning : something like previousEpochError. Would it be possible to bubble that in the same way getTotalNetworkError() ? That would be a much more useful variable for tracking learning progress of individual networks.

    If so, could the variable either transient or access synchronized. The fact that getTotalNetworkError() is not thread safe almost seems like bug.

    P.S. Willing to pitch in if useful.

     

  • Zoran Sevarac

    Zoran Sevarac - 2010-01-23

    Thanks for this comments! I agree with everything you said, and it should be done. Lets wait and see what will Riksi do and then we'll do a review and discuss. And sure you're welcome to pitch in.