weakly supervised multi-label classification (WSML) task, which is to learn a
multi-label classification using partially observed labels per image, is
becoming increasingly important due to its huge annotation cost. In this work,
we first regard unobserved labels as negative labels, ca