III: Small: Uncertainty Quantification and Propagation Analysis in The Visualization Pipeline

点击翻译按钮获取中文摘要

Decision making in presence of uncertainty is a notorious problem where making rational choices often requires analysis and prediction in complex environments. With the prevalence of computational tools for data visualization and analysis that are employed in diverse decision making scenarios, new challenges arise in quantifying, propagating and understanding uncertainty. This project investigates the interaction of data uncertainty with the data analysis process as different features are extracted and visualized. It addresses fundamental challenges that arise in analysis and quantification of uncertainty as the data propagates throughout various stages of the visualization pipeline. This uncertainty-aware framework allows for an end-to-end sensitivity analysis in imaging pipeline where the uncertainties in data acquisition are propagated throughout the synthesis process. The research is driven by problems that arise in uncertainty quantification in a number of application domains including: ensemble simulation and visualization, inverse problems in imaging, and large-scale data visualization. The project broadens the computer science education by providing research training for graduate students and mentoring other trainees. To model the propagation of uncertainty through the visualization pipeline, this project develops a statistical rendering framework in which each data point is a random variable whose distribution characterizes the uncertainty present at that point. The project addresses challenges that arise in imaging and visualization where data transformations (e.g., data filtering, reconstruction, classification, and isosurface extraction) lead to non-linear transformation on the uncertainty. These non-linear transformations on the distribution of random variables accumulate as the uncertainty propagates through the visualization process. This project develops novel approaches for characterizing these transformations that are common to various visualization algorithms. The developments in this project enable the integration of uncertainty, as a first class object, within the key data transformations present in direct and indirect visualization paradigms. Moreover, due to the ubiquity of these data transformations in a wide range of data processing operators, this project opens possibilities for uncertainty quantification in a diverse set of data analysis and visualization systems. The resulting algorithms, software and publications are made accessible via the project web site (http://www.cise.ufl.edu/~entezari/research/uqp/).