Deep Learning under Privileged Information Using Heteroscedastic Dropout