Updated to see how well it did after world cup 2014 ended

[franz-see]

Kindly review my changes

1. Removed 2014 data from the training set
2. Modified the pyb to predict all 2014 matches and compare against actual winners

I'm not sure if I modified this correctly, and it currently has 81% accuracy (13/16).

[fisadev]

Wow, lots of changes in the learn.ipnb file. Maybe you re-run all the cells? To be able to distinguish the changes, they should be commited separatedly (first just the code changes, then the results of re-running the cells). Also, never include the checkpoints folder, for versioning we are using git, so it's redundant to also store ipython versioning folder here, and it will also differ from machine to machine.

And another problem: you removed all the 2014 data from the training set, but I did use some 2014 data to re-train the neural network between the world cup phases (specifically, between the end of the groups phase and before the tree-like phase). I think a better option would be to have a setting for the "last match to use", and filter by that (matches have an id).

Finally, thanks! and sorry about the delay, this PR went unnoticed, I discovered it today while doing some repos cleaning :)

[tonylow]

Any plan to try your system for the 2018 World Cup?