Bias Formulation, Characterization, and Learning (BIFOCAL)

A key challenge to higher level information fusion is fusing data originally collected for disparate tasks. Although much data may be available, data collected for a particular task may not be well suited to subsequent tasks because it embodies biases inherent in the original data collection process. A major challenge to applying data to a new purpose is understanding and correcting for biases in the data. To address this challenge, we propose a tool for Bias Formulation, Characterization and Learning (BIFOCAL). BIFOCAL uses deep learning to detect and characterize the biases in a dataset, thereby better enabling the data to be fused with other data to accomplish different purposes. Our approach is based on the idea that even biased data contains useful relationships, but we must have a deep understanding of the data. To accomplish this understanding, we will perform a systematic study of biases in complex and machine-generated data and use deep machine learning to build models of data sets and data generating processes as well as their target applications.