Скрытые единицы

We investigate how neural, encoder-decoder translation systems output target strings of ap- propriate lengths, finding that a collection of hidden units learns to explicitly implement this functionality. 1...