| "In information technology, big data is a collection of data sets so large and complex that it becomes difficult to process using traditional data processing applications" [151]. In a world of increasing sensor modalities, cheaper storage, and more data oriented questions, we are quickly passing the limits of tractable computations using traditional statistical analysis methods. Methods which often show great results on simple data have difficulties processing complicated multidimensional data. Accuracy alone can no longer justify unwarranted memory use and computational complexity. Improving the scaling properties of these methods for multidimensional data is the only way to make these methods relevant. In this work we explore methods for improving the scaling properties of parametric and nonparametric models. Namely, we focus on the structure of the data to lower the complexity of a specific family of problems. The two types of structures considered in this work are distributive optimization with separable constraints (Chapters 2-3), and scaling Gaussian processes for multidimensional lattice input (Chapters 4-5). By improving the scaling of these methods, we can expand their use to a wide range of applications which were previously intractable open the door to new research questions. |
