Jensen Shannon Divergence Consistency Loss

• @linDivergenceMeasuresBased

• enforces smother neural network responses

• stable, consistent and insensitive across range of inputs

\mathcal{L}(p_{orig}, y)+ \lambda JS(p_{orig};p_{augmix1}; p_{augmix2})

- $$ M = \frac{p_{orig}+p_{augmix1}+ p_{augmix2}}{3}

JS(p_{orig}; p_{augmix1};p_{augmix2}) = \frac{1}{3}(KL[p_{orig}||M||]+KL[p_{augmix1}||M||]+KL[p_{augmix2}||M||])