— Computer & Engineering Sciences for Nonproliferation
forthcoming.
May 5, 2022
- Dr. Craig Bakker (PNNL), “Learning how to Dial 0: Calling the Koopman Operator to Transfer You to a Linear Problem”
- Prof. Jonathan Rogers (GT), “Source Term Estimation for Multi-Source Radiological Search in Obstacle Environments”
April 25, 2022
- Prof. Al Hero (UM), “A Geometry-Driven Approach to Longitudinal Topic Modeling of Nuclear-Scientific Literature”
- Prof. John Fisher (MIT), “Meta-data Integration with mmHDPs”
March 18, 2022
- Prof. Jonathan How (MIT), “Collaborative SLAM for Facilitating Radiological Search and Mapping on a UWB Enabled Multi-Agent Aerial Platform”
- Prof. Milton Garces (UH), “Hunga Tonga eruption data fusion”
March 7, 2022
- Prof. Pavel Tsvetkov (TAMU), “Data and remote sensing in multimodal global surveillance”
February 8, 2021
- TA1 thrust coordination meeting, led by Prof. Paul Wilson (UW)
January 12, 2021
- Prof. Aurora Clark (WSU), “The Shape of Data in Chemistry” and from Dr. Karl Pazdernik on “Microstructural feature characterization of LiAlO2 pellets by deep learning and statistical methods”
December 14, 2020
- Dr. Tenzing Joshi (LBL), “Machine Learning Methods for Radiological Anomaly Detection and Identification”
November 10, 2020
- Prof. Robert Nowak (UW), “Generalized Chernoff Sampling: From Optimal Sensor Selection to Training Deep Neural Nets”
- Prof. Jonathan Rogers (GT), “Time-Optimal Resource Allocation and Planning Algorithms for Multi-UAV Radiological Search”
August 6, 2020
- Prof. Jon How (MIT), “Active Reward Learning for Co-Robotic Vision Based Exploration in Bandwidth Limited Environments”
Abstract: Efficient robotic exploration in novel environments requires a human-in-the-loop who understands the mission objectives. Communication bandwidth constraints necessitate that the human-robot team make intelligent communication decisions regarding what information to send and receive during the mission. In this online variant of the active learning problem, we find that typical information-theoretic heuristics to active learning are not optimal. Instead, a novel heuristic based on minimizing planning-regret enables the robot explorer to choose questions that better help it to identify the highest reward trajectories. When using this heuristic as part of the solution to the newly presented co-robotic exploration POMDP, the robot is able to collect up to 17% more mission reward than with the next-best heuristic.
June 4, 2020
- Prof. Larry Carin (Duke), “Continual learning with neural networks for sensor-based monitoring”
Abstract: Naively trained neural networks tend to experience catastrophic forgetting in sequential task settings, where data from previous tasks are unavailable. A principled Bayesian nonparametric approach is developed, letting the data determine how much to expand the model complexity over time. We pair this with a factorization of the neural network’s weight matrices. The effectiveness of this method is demonstrated on several continual learning benchmarks.
- Prof. Pavel Tsvetkov (TAMU), “Multi-modal global surveillance using cube satellite platforms – problem definition and application feasibility”
Abstract: The talk will update on the current status of our efforts to develop an on-demand capabilities for characterization of localized developments on the earth surface, subsurface and within atmosphere. The use of orbital survey methods offers access options for any location in 3D from subsurface up to upper atmosphere, continuously and over discrete periods of interest. The project is a synthesis of high TRL observational platforms (cube satellites) with lower TRL sensors and predictive methods including fusion and machine learning to yield a robust multi-modal surveillance and prediction capability. This is the collaboration that is progressing together with experts from Sandia and BNL.
May 7, 2020
- Prof. John Fisher (MIT), “Optimal experiment design with bounded rewards”
Abstract: Information-based sensor planning subject to resource constraints can be formulated as a Bayesian optimal experimental design (BOED) problem. However, practical implementations of BOED generally rely on costly estimates of information measures or computationally efficient proxies, but which eliminate performance guarantees. We present an approach that utilizes iterative bound refinement on mutual information (MI) while preserving performance guarantees. The resulting method explicitly incorporates a tradeoff between computational resource expenditures with the quality of the information reward.
- Prof. Al Hero (UM), “Spectral unmixing and prediction with generative probabilistic graphical models”
Abstract: Probabilistic graphical modeling is a general strategy for learning about the hidden structure of observational data that has the benefit of interpretabality in terms of a few simple latent variables that form a generative mechanism of the data. We illustrate the flexibility and power of this machine learning approach for applications including nuclear spectral unmixing, nuclear unfolding, and spatio-temporal event prediction.