SELF-SUPERVISED LEARNING FOR RAPID FORECASTING, GENERALIZATION AND JUDICIOUS DECISION-MAKING IN DYNAMIC AND STOCHASTIC ENVIRONMENTS
Abstract
The machine learning revolution has brought a vast suite of techniques to analyze data and turn it into actionable knowledge, with growing impact on almost every scientific discipline. However, all of its noticeable successes are demonstrated in very well, and rather narrowly defined problems, where sufficient amounts data exists in order to adequately cover the input domain and effectively interpolate between the observed outputs, while difficulties arise when one seeks to understand causal effects and extrapolate in new, previously unseen, scenarios. In this work, we aspire to develop the fundamental building blocks for a new type of neural computational mathematics by introducing scalabe and robust statistical inference techniques for accelerating the modeling and simulation of complex physics-based systems. These techniques include the development of novel neural architectures, self-supervised inference techniques with strong generalization performance, as well as scalable algorithms for aleatoric and epistemic uncertainty quantification. Taken all together, these cross-cutting tools aim to provide effective rules for exploration and judicious decision making under uncertainty that will benefit a wide range of science domains, including applications in the design, planning and control of autonomous systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2021
- Source ID
- FA95502010060
Entities
People
- Paris Perdikaris
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Pennsylvania