Enhancements for NN

[ryangooch]

1. Normalize data to range -1 to 1 instead of standardization

2. Natural Log transform of Measurements, to match linear output of NN

3. Use an SVR as output layer

4. Weight regularization, L2 norm with C = ?

[ryangooch]

Also, I noticed that even the previous submission contains no values above 6 mm/h, the one that achieved a test set error of 23.9 on Kaggle. So obviously most of the values are low, but I wonder if I can capture those higher values any better. The log transform seems to do a better job of that, so perhaps that one is worth a submission?