Show HN: Parametric Activation Pools greatly increase performance in ConvNets

What does 'loss' mean on your graphs? Error fraction?