GAINESVILLE, Fla. –
“So what is a machine learning algorithm really doing?” That is the question a group of university students and faculty are tackling at the University of Florida (UF) under a Naval Engineering Education Consortium (NEEC) grant sponsored by NSWC Panama City Division.
The NEEC program includes university faculty and students working with NAVSEA Warfare Center mentors. Dr. Matthew Bays, NSWC Panama City Division’s senior scientist for robotics and optimization and the command’s NEEC director, said this project is using information-theoretic techniques to develop a systematic way of understanding machine learning systems.
“Machine learning processes are quickly taking over not only facets of everyday life with systems such as ChatGPT, but also have the potential to transform the maritime battlespace,” he said. “NSWC PCD has been actively pursuing machine learning techniques in automated target recognition for mine warfare. However, a growing concern both within the Department of Defense and our broader society is how we can ensure machine learning algorithms are safe and effective.”
Ben Colburn, a Ph.D. student and research assistant in UF’s Computational NeuroEngineering Lab (CNEL), explained how the researchers are using information theory ideas to understand this project.
“There is a growing use of deep learning architectures in military applications and understanding these models will be key to ensure reliability and proper application of deep learning models,” he said. “If we are deploying these models in potentially life or death situations, we need to trust them. If we are to trust them, we must understand them.
“The most current work done under this grant is submitted for review under the title, ‘The Functional Wiener Filter (FWF),’” Colburn added. “This work extends the Wiener solution for an optimal nonlinear filter to a Reproducing Kernel Hilbert space (RKHS) that is nonlinearly related to the input signal. This yields a closed-form solution to an optimal nonlinear filter in a data-dependent universal RKHS.”
Tackling important research questions while training students, like Colburn, with the goal of employment after graduation are two primary goals of the NEEC program, Bays said.
“The NEEC program has been a tremendous benefit to the Naval Research and Development Enterprise and NSWC PCD in particular, with projects recommended by all three NSWC PCD (technical) departments in recent years,” Bays said. “It’s created a sustainable pipeline of scientists and engineers with difficult-to-hire skillsets such as autonomy and machine learning.”
Dr. Jose C. Principe, Eckis Distinguished Professor of electrical and computer engineering at UF, is the NEEC faculty member and principal investigator for this project. He is responsible for the research direction effort and execution of the project.
“Machine learning is achieving amazing results, but our lack of understanding about how specifically this happens is quite frustrating,” he said.
“On the other hand, information theory is a well-developed theory of bounds, centered around the concept of information," he continued. "The big advantage of information is that it is metric and actionable. Hence, integrating machine learning with the theory of information quantifies how information is transformed within a machine learning algorithm during training. This will pinpoint bottlenecks and ways to improve training times and the choice of hyper parameters for robust performance.
“The overall project goal is to design machine learning algorithms in the same way engineers build high technology,” Principe said. “In my laboratory, students not only learn algorithms and applications, but also collaborative methodologies on how to solve problems and construct new paradigms that push the state-of-the art. I am therefore very honored to be part of the NEEC initiative.”
Isaac Sledge, NSWC Panama City Division’s senior machine learning research scientist and NEEC mentor, and his team have successfully competed for and received grants from the Office of Naval Research in the areas of information-theoretic reinforcement learning, information-theoretic deep learning and information-theoretic uncertainty quantification.
“This project is incredibly important for understanding the flow of information in deep networks and why they learn to perform complicated tasks, like automated target detection and recognition, semantic scene segmentation, and more,” Sledge said.
“This research will be instrumental in designing better automated target recognition systems and understanding why they work and where they provide spurious responses,” he said. “All of these topics are of immense interest to the U.S. Navy, as they will lead to the creation of deployable systems that remove the need for the warfighter to act in dangerous locations. They also enable us to maintain a lopsided technical edge against our adversaries.”