Future experiments

v8/v9 + no dropout: to check if the dropout network isn't expressive enough. If no dropout network learned training perfectly, I should increase the size of the dropout network. Otherwise, I need to inspect why I can't overfit on the training set. Might want to continue E-93. (https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf)

E99/E100 + higher learning rate, bigger momentum, and max-norm reluarization. (https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf)

Check if I can overfit on the [-2, 2] range.

Add more data