Does anyone know what the state-of-the-art is for composite ML problems that involve both classification and labeling sub-tasks? For instance in computer vision, classifying images into one of several categories of scenes and then detecting relevant objects in the image depending on its class. Or classifying emails into predefined topic labels and extracting relevant info (calendar events, contacts, ...) depending on the topic. The obvious approach is to solve each task in sequence but then cascading errors become a problem. Ensemble methods can combine and optimize the local results of each subtask, but at the cost of introducing another layer on top of everything. I wonder if there is some other clever approach out there for performing different ML tasks like classification and extraction on the same object in a more principled way.
[link][4 comments]