Recurrent Neural Networks

A small recurrent network reads a handwritten digit one row at a time and then has to wait a number of blank steps before naming it. With 10 blank steps it names 85.8% of held-out digits correctly. With 20 it names 15.6%, barely above the 10% of guessing. The digit and the network are the same; only the distance between the evidence and the answer changed. You will run that experiment on a laptop in this chapter.