Hand-annotating data can be expensive. Also, often times, the people who annotate it are not specialists. (Cancer image annotation). Hence, you may often have labels, which do not exactly tell you which exact class the image belongs to, but rather a less specific information like a subset of possible classes. Many firms will disgard such lower quality annotated data. This is wrong, since from the Information Theory point of view there is still enormous amounts of information present in the less exact annotation. Hence, we present methods on how to use such a "partially annotated data":