Mine new information from a huge unlabeled image corpus

You have a very large corpus of unlabeled images. Design a system that surfaces examples unlike anything already in the training data, and that finds the images containing a given object of interest. Cover representations, indexing at scale, and how you check that what you found is useful.