Visualization has become an increasingly important tool in all areas of scientific research and applications. It is crucial to diverse tasks such as evaluation of engineering designs, the understanding of large-scale simulations, the viewing of multi-modal 3D medical data, the mining of enormous web databases, and the analysis of homeland security information. However, visualization has not gained widespread use for routine data analysis in many fields, because while continuing advances are being made in visualization system software and hardware, the development of appropriate user interfaces for these tasks has lagged behind. In real-world applications, it is common to encounter data sets, which in addition to being of a spatial-temporal nature possess a large number of attributes (high dimensionality). The complexity of traditional user interfaces that were developed for low-dimensional data analysis grows rapidly as the dimensionality of the data increases, so that many users are unable or unwilling to perform the time-consuming and tedious steps required to create an informative visualization in a high dimensional space. In this project, the PI will address this challenge by incorporating machine learning into the interfaces used by scientists to explore and interact with their data, to develop intelligent interfaces targeting high-dimensional data visualization tasks with significantly improved usability. In a preliminary study the PI demonstrated the potential power of his approach in performing challenging volume classification tasks that conventional methods failed to do well, by coupling machine learning with a painting metaphor to enable more intuitive specification of user intent in the classification process. Users interactively paint directly on the volume rendered images or selected cross sections of the volume, and are given full control of what materials to classify by applying paint of one color to parts of the volume representing materials of interest and paint of another color to regions that are not desired. The system uses the painted regions (a very small subset of the volume) as training data to learn how to classify the whole volume, mapping each voxel into a value indicating the likelihood that the voxel is part of the material of interest; this uncertainty is then mapped to opacity for direct volume rendering. While the prototype system employed a deterministic artificial neural network, in this project the PI will use data sets with varying characteristics to explore whether support vector machines or Bayesian approaches can provide superior performance and better scalability in terms of being less computationally demanding while affording better interactivity.
Broader Impacts: The new visual interface technology to be developed by the PI in this project will dramatically increase the usability and efficiency of visualization systems and broaden the base of users who can successfully tackle the increasing challenges presented by complex visualization and analysis tasks. By incorporating effective machine learning into the process of data visualization and interaction, the PI will free users from repetitive tasks and the complex interfaces of today's systems. Concepts such as "learning to classify" and "learning to track" which are central to the PI's approach are very powerful, and will suggest to scientists that they rethink about how data analysis and visualization can be done, as well as enabling them to "reuse" and "share" valuable visualization experiences.
