Re:New kind of recurrent neural network using attention
by jostmey on Thursday March 09, 2017 @11:12AM
(#13830220)
Oh no! I think you may be right. It is possible to correct the model for numerical stability in a manner that doesn't impact the output, but my approach fails at that. The correction from the previous step needs to be saved and carried over to the next step, just like the numerator and denominator are saved and carried over.
I think I will have to re-run all the results once I fix the code. This is a bummer
I think I will have to re-run all the results once I fix the code. This is a bummer
-- Thanks for finding this! Jared