Multiple-Instance Active Learning
Burr Settles, Mark Craven, and Soumya Ray

#M24

We introduce a framework for active learning in the MI setting. For some domains, labels are cheaply acquired at a coarse (bag) level of granularity, and it is possible, though expensive, to obtain labels at a finer (instance) level. We develop algorithms for learning with labels at mixed granularities, by first training on inexpensive bag labels and then actively querying which instances to label.

Content-Based Image Retrieval

bag: image = { instances: segments }

Text Classification

bag: document = { instances: passages }