Add automatic relevance determination

[Jack-H-Buckner]

Add functions to regularize the neural network that automatically reduces the effect of inputs with little or no impact on the outputs to zero. This can be done by adding a layer to the beginning of the neural network that multiplies each input by a constant. We can then apply regularization to these scaling factors to modify the effect of each input.

[Jack-H-Buckner]

Note I think this should be for version 1.1, but I had a breakthrough on how to make this happen so I wanted to flag it.