Backlog (To Do)

• center loss (prototype)
• separate activation from layer
• speedups with dropout or drop connect (skip some convolutions)
• documentation / doxy
• function to lock/remove/mod/clean layer
• move logging code into network class
• drop connect
• batch norm
• reproduce resnet results
• activation layers - don't seem to need right now. Can use resize layer.
• finish fractional max pool (the correct way)
• finish support for non-square elements (off the table. Do want to handle 1D convolutions though)
• conv stride > 1 - in but broke fix it
• sliding window code with full image network processing
• lecun-like multi-dimensional embedding
• C wrapper with python example