We consider the problem of learning soft assignments of $N$ items to $K$
categories given two sources of information: an item-category similarity
matrix, which encourages items to be assigned to categories they are similar to
(and to not be assigned to categories they are dissimilar to