See http://www.fast.ai/2018/07/02/adam-weight-decay/ and the original
paper https://arxiv.org/abs/1711.05101.pdf for context.
I don't know what I'm doing, and this is quite possibly wrong - but on
a simple Char-RNN I have lying around on my harddisk, this seems to
improve the rate of learning consistently for different hyperparameters
vs. standard ADAM with the same decay constant.
`@. p.Δ = η * p.Δ / √acc` parses correctly while `@. p.Δ /= √acc*η` seems to parse like `@. p.Δ /= (√acc*η)`, hence the step size was de facto interpreted as `1/η`
