This commit changes the way coffee breaks work -> each break now returns a dict in which statistics are recorded.
These dicts can then be recoreded by the AccumulateStatistics break and saved to a hdf5 file.
This is quite useful for monitoring training.

it can be used to plot data that was written by the AccumulateStatistics break
and even works during training.
This is analogous to the plot_monitor.py script of pylearn2.

…ishs dropout paper.

This is a first quick sketch that gives the desired result of ~100 errors on MNIST
using a fully connected network of 2 layers a 1200 units with dropout.
NOTE: this is still missing the constraint on the L2 norm of the weights, will add them in a later comit.

…iple files for plotting