In many machine learning applications, the datasets consist of a large amount of normal instances and a small amount of abnormal instances that form multiple clusters. One is interested in identifying models for these abnormal instances
One example of such application is the registration of high-resolution mouse brain images, where identification of regions with significant textures provides important hints for good alignment.
In this paper, we propose a framework for learning scoring functions that capture significant instances in unlabelled datasets. We apply this framework to detecting significant regions in mouse brain images and learns discriminative models of their textures.