The cost of data annotation is a substantial impediment for multi-label image
classification: in every image, every category must be labeled as present or
absent. Single positive multi-label (SPML) learning is a cost-effective
solution, where models are trained on a single positive lab