Implement bachnorm layer and add it to some tests

[Mikolaj]

https://en.wikipedia.org/wiki/Batch_normalization https://penkovsky.com/neural-networks/day4/

I'm sure there are good blog posts about that. Looking this up in pytorch and/or tensorflow API may help, too.

Perhaps adding this to the existing tests with convolutional and recurrent neural networks would be a good idea. We could measure how the training improves. With batchnorm perhaps we could switch activation function to relu, in examples where it's not already relu, and not lose convergence. Probably normalizing data from -1 to 1 (zero mean, unit variance) would also be useful for relu, so a convenient tools for normalizing is a related task (probably batchnorm is just much more general, e.g., it needs to be differentiable).

Two other keywords that may be related or may be even the same thing as one of the above, so worth googling:

• regularization functions, to keep params not too small or big

• activity_regularizer: Regularizer function applied to the output of the layer (its "activation")