Introduce network regularization techniques I

[anthonio9]

Do the parameters to below techniques manually (important):

• weight regularization, weight penalty, comes out of the box with Adam optimizer: L2

• dropout

  Test both with multiple manually set values, no automatic regularization yet

[anthonio9]

It seems that enabling dropout really helps with regularizing the validation results, the drop of RCA with increasing epochs is much lower than it is in the case of a regular model, for example ppn-split-voiced-62 drops more than ppn-split-drop-weight-68 or ppn-split-drop-voiced-64.

Let's see what happens if dropout, weight decay and augmentation with adding noise are combined together.

[anthonio9]

In the end combining all the regularization techniques didn't provide good enough results, let's close this issue and move forward