Rawski on Learning with Partially Ordered Sets
Jon Rawski is presenting his work on learning with partially ordered representation in the IACS seminar room this Wednesday, November 14, 1-2pm. Check it out (for the undecided: there’ll be free lunch)!
Title: Learning with Partially Ordered Representations
Abstract: When generalizing, and making compact hypotheses from sparse and impoverished data, learning algorithms often rely on domain-specific representations or features. Often these features pose a learning problem, by exponentially increasing the number of hypotheses which correctly describe data. However, features also give the learner an advantage, by structuring the space of hypotheses in a particular way. This structure gives rise to certain entailments between grammars, which I show using several linguistic examples - text recognition, speech analysis, and syntactic adjunction. I discuss how learners can exploit this structure to make inferences, and introduce a non-statistical learning method that provably identifies the responsible constraints. Integration and comparison of these insights into statistical learning is ongoing research.