Call for PapersAbstractProbabilistic graphical models are a powerful framework for multivariate prediction problems with temporal, spatial and relational interactions. They provide compact representation of complex correlations between variables and inference methods for exploiting these correlation for prediction. Kernel methods, on the other hand, are emerging as popular tools for efficient non-parametric estimation for univariate prediction problems. Their popularity is primarily due to their strong theoretical foundations and ability to exploit high dimensional feature mappings (also popularly called the kernel trick). Recently, there is considerable excitement about the convergence of results from these two seemingly different areas. For instance, the Maximum Margin Markov nets and kernelized Conditional Random Fields (CRF's) are some of the first steps in this direction. Some work has also been done on using fundamental properties of the exponential family of probability distributions to establish links. We invite high quality submissions for presentation as talks during the workshop. Revised and expanded versions of the papers will also be considered for inclusion in the proceedings of the workshop to be published tentatively in early 2005. Important Dates:
|