@Inproceedings{Antoniuk-Poster2013, IS = { zkontrolovano 15 Jan 2014 }, UPDATE = { 2013-08-02 }, title = {Statistical formulation of structured output learning from partially annotated examples}, author = {Antoniuk, Kostiantyn}, booktitle = {17th International Student Conference on Electrical Engineering}, pages = {1--5}, book_pages= {705}, address = {Technick\'a 2, Prague, Czech Republic}, isbn = {978-80-01-05242-6}, day = {16}, month = {May}, year = {2013}, publisher = {Czech Technical University in Prague}, venue = {Praha, Czech Republic}, annote = { Empirical risk minimization based methods for structured output learning have proved successful in real-life applications. A considerable deficiency of existing algorithms, like e.g. the Structured Output SVMs (SO-SVM), is the demand for fully annotated training examples. Despite several recently published works trying to extend SO-SVM for learning from partially annotated examples, two crucial problems remain open: 1) an exact statistical formulation of risk minimization based learning from partially annotated examples and 2) an efficient learning algorithm. While the existing works attempted the algorithmic issues (i.e. the second problem), in this paper we tackle the first problem. In particular, we formulate learning of the structured output classifiers from partially annotated examples as an instance of the expected risk minimization problem. We show that the minimization of the expected risk is equivalent to the minimization of a partial loss which can be evaluated on partially annotated examples only. Thus, the empirical risk minimization based methods can be applied. }, keywords = {Partially annotated examples, structured output learning, risk minimization.}, project = {SGS12/187/OHK3/3T/13, GACR P202/12/2071, Visegrad Scholarship contract No. 51200430}, psurl = {Antoniuk-Poster-2013, 204 KB}, url = {ftp://cmp.felk.cvut.cz/pub/cmp/articles/antoniuk/Antoniuk-Poster2013.pdf}, note = {CD-ROM, IC04}, }