Objective Prediction of Performance in Simulated Flight Tasks Using Multimodal Physiology
Hrishikesh Rao, MIT LL
To support the acceleration of pilot training, this program is focused on multimodal physiological sensing to model cognitive state and objective predict flight performance. First, we developed a data collection testbed in which human subjects can be instruments and perform simulated flight tasks in virtual reality. Second, we collected a large dataset as subjects performed repeated runs of an instrument landing system procedure. Finally, we detail our team's results linking physiological signals to flight performance.
Kesh Rao is a technical staff member in the Human Health and Performance Systems group. His research focus is on multimodal physiological sensing and modeling for operational and clinical decision making. His specialization is in measuring eye movements and modeling cognitive states.
Topology-Driven Self-Supervised Learning for Axon Segmentation and Centerline Detection
Nina Shamsi, MIT LL
Large amounts of unannotated brain imaging data present a challenge for classification tasks due to the time and labor-intensive process of data labeling. Specifically, the dual goal of understanding how axons connect throughout the brain and intersect different brain regions would necessitate laborious annotations by subject matter experts. In this work, a Residual 3D U-Net was used to train a classifier to predict the order of permutated, sliced voxels samples of mouse brain data acquired via 3D light sheet microscopy and confocal imaging. The subsequent pretrained weights from auxiliary task training were finetuned for target task classification of axon segments and centerline detection. Preliminary results show an improvement in evaluation metrics using self-supervised learning while also tuning parameters of the soft centerline Dice loss function.
Nina Shamsi is a PhD student in electrical engineering at Northeastern University. Prior to Northeastern, Nina worked on developing physiological models of human cognition and intelligent tutoring software. Nina has a master's in electrical engineering from the University of Massachusetts, Amherst, and a BA in biochemistry from Mount Holyoke College.
AI Accelerator: Air Guardian
Costas Frost, MIT LL
Cognitive overload is a serious issue plaguing current-day pilots to which there is currently no solution. Through human–AI teaming, can we augment an inexperienced pilot to outperform an expert? Air Guardian aims to develop an autonomous system to know when to assist a pilot during periods of increased stress and overwhelming input. This can be accomplished using state-of-the-art neural circuit policies that enable R&D capabilities for utilizing focused attention and learned risk metric maps to rapidly identify dangerous situations.
Costas Frost is an associate staff member of the AI Software Architecture and Algorithms Group at MIT LL and works on simulation development. His expertise is in 3D modeling, procedural generation, and programming for simulations. He is currently working on projects in machine learning and XR spaces. He has previously worked on Emmy-nominated visual effects and has been featured by Unity.
Blade Runner: Rapid Countermeasure to Detect StyleGAN Images
Adam Wong, MIT LL
The Fifth Domain (cyber) remains a contested strategic environment for friendly forces. DeepFakes and AI-generated photographs are becoming increasingly common in use by threat actors. Defenders find themselves unable to keep up with advancing technologies due to lack of manpower, lack of supercomputing hardware, or lack of expertise in AI/ML. BLADERUNNER is a rapid countermeasure to detect AI-generated (specifically StyleGAN 1 and 2) images by exploiting weaknesses in the image post generation. With BLADERUNNER, speed, portability, and decisive action are prioritized.
Adam Dorian Wong is PhD student in cyber defense at Dakota State University. He serves as a cybersecurity and cyber threat intelligence analyst assigned to Group 52 (AI) and is a former MIT LL SOC analyst. As a NH Army national guardsman, he is a sergeant and warrant officer candidate currently assigned to 136th Cyber Security Company (Det. 1 - NH), 91st Cyber Brigade (ARNG). He mobilized as a CND analyst under Task Force Echo 3, in support of USCYBERCOM. His current research focus area is in rapid countermeasures.
CIOTER: Counter-Influence Operations Test and Evaluation (T&E) Range
Kevin Nam, MIT LL
MIT LL’s Counter Influence Operations Testbed and Evaluation Range (CIOTER) will support testing and evaluation of a variety of counter-influence operations (CIO) technologies, allow humans-in-the-loop for both testing and red-blue exercises, and enable a rapid deployment of CIO technologies. CIOTER’s open system architecture aims at providing well-defined interfaces for integration of various internal and external tools and data feeds, supporting evaluation, training, and operations sidecars. Rapid and flexible integration enables early and iterative experimentation as well as robust evaluation and training environments. In the first year of development of the testbed, we designed and implemented the open architecture and core components for evaluating two CIO technologies being developed internally at MIT LL, along with a deep dive into the evaluation approaches and metrics.
Kevin Nam is a technical staff member in the AI Software Architecture and Algorithms group. His formal training is in information science and computer science. His research areas include human-centered AI, system and data architecture, and T&E for IO technology.
LWIR Hyperspectral Imaging and Machine-Learning Algorithm for Trace Explosive Detection
Bill Barney, Jason Jong, Max Kenngott, Patrick Wen, Rod Kunz, Keegan Quigley, Mitesh Amin, Bill Herzog, MIT LL
Standoff detection of trace contaminants, such as explosives, is desirable for many applications, including checkpoint screening at secure facilities. Although many approaches have been tested in laboratory settings, no technology to date has demonstrated the ability to meet the demanding sensitivity, specificity, and area coverage rate (ACR) requirements for applications at aviation security checkpoints. As such, the preferred technology is still contact-based, e.g., surface wipes followed by ion mobility spectrometry. In the work described here, we showed that a standoff technique, longwave infrared hyperspectral imaging, has potential for explosive screening applications when augmented by machine-learning detection algorithms.
Bill Barney is a member of Group 23 (Biological and Chemical Technologies) and works mainly in the area of trace chemical detection. His formal training is in laser spectroscopy, but he has also worked on everything from trace DNA analysis to medical device development.
EmbAI: Accelerating AI Technology Impact on DoD-Relevant Applications
Vitaliy Gleyzer, Michael Yee, Michael Tierny, Paul Monticciolo, MIT LL
Many national security applications require advanced AI capabilities in challenging, SWaP-constrained environments, and therefore require high-performance embedded solutions. Since commercial and academic progress in embedded AI is moving rapidly, it is challenging for the DoD to evaluate and leverage new technology effectively.
The EmbAI program is looking to create a DoD-specific AI research environment for embedded applications. This environment would allow co-design, research, development, and direct evaluation of promising embedded AI technologies, such as new embedded hardware and machine-learning algorithms.
Vitaliy Gleyzer has been a staff member at MIT LL for more than 10 years. Prior to joining the Laboratory, he received his master’s degree in electrical and computer engineering from Carnegie Mellon University, with a research concentration on wide-area networks. His current work and research interests are primarily focused on hardware and algorithm co-design for development of novel high-performance hardware architectures, large-scale graph analytics, and embedded AI applications.
Multi-Agent Autonomous Space Technology (MAST)
Yaron Rachlin, MIT LL
Dr. Yaron Rachlin is a senior research scientist in the Artificial Intelligence Software Architectures and Algorithms Group at MIT Lincoln Laboratory. He received B.S. degrees in Electrical Engineering and Mathematics from Virginia Tech and a Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University. He has led the development, prototyping, and testing of multiple novel computational imaging systems with applications in wide field of view infrared imaging and hyperspectral imaging. He has also developed advanced algorithms for image reconstruction and for computationally efficient search for faint signals. His current research focuses on causal inference in social network data and on decision-making in non-cooperative dynamic games. He has over 40 publications and patents in diverse areas of research.
Space has increasingly become a congested, competitive, and contested domain. As a result, the challenges of non-cooperative space operations, e.g., servicing a malfunctioning satellite which has on-board collision avoidance, comprise an increasingly urgent field of research. Solutions to problems in non-cooperative space operations will require more than approaches based in classical control theory and orbital mechanics. Game theory, artificial intelligence, and machine learning techniques will be needed to solve the perception, decision-making, and control problems that arise in non-cooperative scenarios. The Multi-Agent Autonomous Space Technology Line program (MAST) is advancing research in each of these fields in order to enable a “full stack” autonomy pipeline for non-cooperative space operations.
Author Verification, Author Obfuscation, and Style Change Detection Using Transformer Models
Trang Nguyen, Kenneth Alperin, Courtland VanDam, Rohan Leehka, Charlie Dagli, MIT LL
We demonstrate how state-of-the-art natural language processing models based on transformers can be used to enable capabilities for author recognition between and within documents (author verification vs. style change detection) and the complimentary obfuscation to evade author recognition. We demonstrate the results on diverse datasets and introduce ways to use these techniques toward end-user needs.
Trang Nguyen is an associate staff member in the AI Technology and Systems Group at MIT LL. Her research focuses on the application of natural language processing and speech to problems in cyber security and authorship identification.
Defense Against Shortest Path Attacks
Ben Miller, MIT LL
Benjamin A. Miller is a Lincoln Scholar in the AI Technology and Systems Group at MIT Lincoln Laboratory and a PhD candidate in Network Science at Northeastern University. His research is focused on vulnerability and robustness in artificial intelligence applications using network data. Ben received the BS and MS degrees in Computer Science from the University of Illinois at Urbana-Champaign.
Recent research has demonstrated the vulnerability of network analytics to adversarial manipulation. We consider the scenario where an adversary can remove connections to route traffic along a specific path. Taking the defender's perspective, we show that adding random perturbations to the advertised distances between points increases the adversary's uncertainty, resulting in much larger attack costs and fewer successful attacks. In both real and simulated networks, we also demonstrate the tradeoff between robustness to attack and the impact on legitimate users of the network.