SBIR Phase I: Technology Translation of a Universal Summarization System

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to make machine-learning (ML) and artificial intelligence (AI) less costly, less biased, more accurate, more scalable, and easier to use through the process of commercializing “universal data summarization.” This summarization process works on any kind of data and significantly reduces dataset size without loss of information. Cost reductions include computational, core memory, storage, labor, and energy, all while reducing AI's environmental impact to provide "green AI." Universal summarization can also be used to measure and remove bias in data used to train AI/ML systems. Biases, caused by concepts in the data that are vastly over-represented while others are under-represented, will be reduced since for a small summary to be representative, it must be diverse and inclusive. In addition, accuracy will be increased by improving human analytics. Many data science tasks involve analysis by humans who must examine data to discover insight. These are arduous, expensive, time-consuming, and error-prone tasks made worse by human alert and decision fatigue caused by redundancy and repetitiveness. By reducing size and eliminating redundancy, human fatigue is reduced, human accuracy and efficiency are increased, and annotation costs are mitigated. This Small Business Innovation Research (SBIR) Phase I project will develop and commercialize the ability to simply and affordably perform universal summarization of massive datasets. A summarization is a process that selects from a dataset a small subset of data items – the few selected summary items accurately represent the information contained in the many remaining unselected items. The innovation is called “calibrated submodular summarization,” a technology that involves quickly, cost-effectively, and accurately measuring information in subsets of data, and then algorithmically selecting small subsets that have mathematical information content guarantees. This technology strips away redundancy, leaving behind an efficient representation of the core information in the dataset. The proposed activities will create a commercial service that can summarize massive amounts of data (of any kind) quickly, easily, and without requiring user expertise in machine learning, data science, or submodularity. This will greatly reduce costs, time-to-market, environmental impact and data bias for any data-rich industry. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.