Structured prediction refers to machine learning models that predict multiple interrelated and dependent quantities. Many applications in a wide range of domains can naturally be understood in this way. A wide variety of expressive and powerful models have been proposed, mostly tailored to specific applications. A shared common problem amongst many structured prediction models is that of intractable inference. This results in inefficient computation and in turn reduces the practical significance. We propose to study structured prediction models in the context of sequential decision making. Each decision is allowed to depend on a rich context that includes previous decisions as well as context-dependent observations. This approach puts emphasis on tractable inference and we believe that progress in this field will result to practical impact in multiple applications.
This PhD position is sponsored by Microsoft Research and is co-supervised by Peter Gehler, Max Planck Institute for Intelligent Systems and Sebastian Nowozin, Microsoft Research Cambridge.