virtual adversarial training

By Miyato et al.:

A perturbation $\Delta x = \mathrm{arg max}_{\Delta x} \mathrm{KL}(p(y \vert x) \vert \vert p(y \vert x + \Delta x))

#needs-expanding