Supervisor: Dr Tim Hospedales
In traditional supervised machine learning, the human defines, collects, and annotates the data the machine should learn from. Thus the machine is a passive learner, with no ability to influence its training program. The result in many cases is dramatically sub-optimal learning, with much more data (and more importantly and expensively, more annotation) required than necessary. The alternative is active learning or optimal experimental design, where the machine performs more humanlike introspection, and decides the most informative data, explicitly asking the human teacher for the explanation (annotation) of that data. In this project, we will investigate cost-sensitive active learning models for structured domains such as multimedia. Technical challenges will be developing introspection models to estimate and trade-off the expected benefit of learning from different kinds of data and annotation, as well as developing efficient incremental learning algorithms to make this feasible. Possible application areas include computer vision/multimedia, medical diagnosis, risk management, etc.