Investigate zero padding

[knkski]

Zero padding can sometimes be useful when dealing with CNNs. We should investigate how/where to use them

[knkski]

This graph shows the effects of zero padding on a simple CNN (2x64 conv, 2x128 conv, 2x4096 dense):

The dark green line shows no zero padding, the yellow line shows zero padding before the initial conv layer, the line green line shows zero padding before each set of conv layers, and the purple line shows zero padding before each conv layer.

Based off of this graph, it looks like zero padding before the initial conv layer is useful and improves performance by about 0.2%. The other zero padding layers either had no effect or worsened performance, however.