II.A.1.c.ii.(1) Computational Bayesian Learning Via Topology (CoBiLT)
Abstract
Artificial Intelligence (AI) opens up new avenues that will help future warfighters to make accurate tactical decisions via human-machine teaming capabilities. For example, the brain computer interface (BCI) technology, which does not depend on any external device or muscle intervention, is needed where response time is crucial as in a battlefield setting. BCIs harness information generated by brain waves and analyze it to communicate with computers, drones, vehicles and other devices. Extracting information generated by brain activity, electroencephalography (EEG) recordings have been widely used. However, it is expected that humans in a human-machine team will be actively engaged, which will yield nonstationary, nonlinear and noisy (3N) EEG signals. Consequently, current methodologies for analyzing these signals often fall short as they rely on assumptions, which violate at least one of the 3N nature of EEG signals. Our Computational Bayesian Learning via Topology (CoBiLT) method provides an effective, flexible and noise-resilient scheme to analyze these complex signals by relying on their shape peculiarities while respecting their 3N nature. CoBiLT views the EEG analysis via topological data analysis (TDA) lenses by tracking the evolution of homological features in signals over time intervals and producing summary representations, namely persistence diagrams (PDs). Due to the variability in these complex signals, novel data analysis methods are urgently needed, and the associated persistence diagrams give rise to great variability. CoBiLT quantifies this uncertainty through a novel Bayesian framework. CoBiLTÕs Bayesian framework allows us to incorporate prior knowledge from historical data, and/or past experience of an individual who is engaged in the human-machine team. To that end, posterior distributions for each task and/or individual, which quantify the associated uncertainty, are generated. The generation of these posterior distributions rely on an approximation method due to the high complexity of signals. CoBiLT estimates posterior distributions by developing the first computational Bayesian methods for TDA, namely (i) importance sampling Monte Carlo to address rare events; (ii) sequential Monte Carlo to tackle the problem of analyzing signals recorded during long periods of times; and (iii) Markov Chain Monte Carlo schemes for PDs as an alternative to the two aforementioned methods. The posterior distribution estimation will in turn be employed to a robust classification of individualsÕ EEG signals, which measure brain activity, using a true Bayesian formulation where the experience and knowledge through historical data is allowed to further improve soldiersÕ performance. CoBiLT will produce a powerful machinery with direct applications in AI that will help a future warfighter to make tactical decisions. Precisely, CoBiLTÕs powerful methods will greatly impact the human-machine teaming by providing with concepts and capabilities to dominate a future battlefield where the decision will be assisted by AI. Indeed, PI Maroulas, a Senior Research Fellow at the US Army Research Lab, will collaborate with scientists at Aberdeen Proving Ground (APG), to help military officers in making accurate tactical decisions in a harsh environment by harnessing information via physiological signals. Gained knowledge from CoBiLT will be delivered into a one year course in the new PhD in Data Science at UTK, of which the PI is one of the founding members. This course will uniquely expose graduate students to Army relevant research.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110094
Entities
People
- Vasileios Maroulas
Organizations
- Army Contracting Command
- United States Army
- University of Tennessee