Imagine a journalist looking for an illustration to his article about patriotism in a database of unannotated images. The idea of a suitable image is very vague and the best way to navigate through the database is to provide feedback to the images proposed by an Image Retrieval system in order to enable the system to learn what the ideal target image of the user is. Thus, at each search iteration a set of n images is displayed and the user must indicate how relevant they are to his/her target. When considering real-life problems we must also take into account the system's time-complexity and scalability to work with Big Data. To tackle this issue we utilize hierarchical Gaussian Process Bandits with visual Self-Organizing Map as a preprocessing technique. A prototype system called ImSe was developed and tested in experiments with real users in different types of tasks. The experiments show favorable results and indicate the benefits of proposed algorithms in different types of tasks.
