Gammelsæter, Martin (Master thesis, 2015)
In many of the problem domains typically tackled by deep learning, data
is plentiful and cheap but labeling of the data is tedious and expensive.
Letting a model actively select the data instances it is uncertain about ...