In many modern problems across areas such as genomics, computer vision, and natural language process, one is interested in learning a Sparse Structured Input-Output Regression Model (SIORM), in which the input variables of the model such as variations on a human genome bear rich structure due to the genetic and functional dependences between entities in the genome; and the output variables such as the disease traits are also structured because of their interrelatedness. A SIORM can nicely capture rich structural properties in the data, but raises severe computational and theoretical challenge on consistent model identification.

In this talk, I will present models, algorithms, and theories that learn Sparse SIORMs of various kinds in very high dimensional input/output space, with fast and highly scalable optimization procedures, and strong statistical guarantees. I will demonstrate application of our approach to problems in large-scale genome association analysis and web image understanding.

This is joint work with Seyoung Kim, Xi Chen, Seunghak Lee, and Bin Zhao.