Why Are Eight Bits Enough for Deep Neural Networks? (2015) (petewarden.com)