Abstract: Problems of alignment or correspondence of sequences, trees and other structured objects of different types are predominant in many fields. Examples include word alignment for machine translation, functional pathway correspondence across species, matching of images/video and text. In these tasks, the similarity between the corresponding elements is difficult to specify and tune by hand, since it involves elements of different types: words of different languages, proteins that are not sequence-homologous, image regions and words. I will describe novel approaches to supervised and weakly supervised learning of alignment models from data using tools from convex optimization and probabilistic modeling.
- This event has passed.