Convolution - make connections not always dense

[jminuse]

Our network probably doesn't need to be dense from top to bottom. For example, each pairwise type of feature (H-H, H-C, H-N, etc) could interact with itself for a few layers before we merge them all together. This would be analogous to the self-learned filters in image recognition.

[proteneer]

Neat idea - we should explore this more at the next meeting.