Choosing the Query
Want to be close to SVM h* trained on all labels
h* always in version space
It separates all of the labels so it separates any subset
Receiving label of a pool instance reduces version
Idea: choose pool instance that will reduce version
space as fast as possible