Network-Based Visualization Algorithms for Improved Feature Engineering, Analysis, and Modeling

BlueLightAI is developing methods for addressing crucial issues confronting Machine Learning and Artificial Intelligence development and implementation.  These issues include massive data requirements as well as massive space and time resources for building accurate models.  In addition, Deep Learning is often seen as a black box, with little transparency and explainability, which, presents problems for adoption in heavily regulated industries such as health care and finance as well as important DoD applications.  Finally, while development of the resulting models requires substantial mathematical, statistical, and engineering expertise on the part of the developer, there is the general understanding that Deep Learning cannot be engineered, in the sense that it is not the result of an iterative procedure in which each step provides improvement, but instead is the result of extensive experimentation. To address all these issues, BlueLightAI has built a platform for interactive, visual data analysis and exploration for complex data based on Topological Data Analysis.  In the platform, users can explore graph-based representations of their data and features, and investigate the relationships found in the dataset through interactive colorings of these visualizations. Among other things, this allows them to discover clusters of features that readily distinguish two groups of interest, or to stratify groups based on similarity or proximity in the graphs. The goal of this proposal is to take the next step; namely, to extend the capabilities of our platform to include building and deploying predictive models. With these additional capabilities, users will be able to take an insight obtained by exploratory analysis and transform it into a system that delivers real-world value. These new modeling tools will be based on novel techniques that are tightly integrated with the graph-based visualization and analysis tools already present in the BlueLightAI platform. These techniques are currently in the initial stages of research, and the commercialization process will begin with validating and refining them on relevant datasets, as well as developing prototypes of the user-facing interface. In this base period in our proposal, we will concentrate on two main themes: (1) Developing novel functionality that supports building and delivering explainable and performant predictive models, validated by application to a real-world dataset. (2) Building a prototype application for real time predictions using an appropriately chosen datasets to transition to the commercialization of our work. Finally, a substantial body of research initiated and conducted by the PI, his academic group, and the company itself has demonstrated proofs of concept on a variety of examples and establishes feasibility for the pioneering BlueLightAI approach detailed here.