EFFICIENT AND ADAPTABLE MULTIMODAL NEUROSYMBOLIC ARCHITECTURES FOR ROBUST COMPLEX EVENT DETECTION.
Abstract
Deep neural networks (DNNs) produce accurate predictions from unstructured high-dimensional data such as sensory information while leveraging high-performance tensor operations in specialized hardware AI accelerators. In many tasks, they have replaced symbolic and algorithmic approaches based on first principles, human knowledge, and mechanistic models. However, the benefits of DNNs come at the cost of reduced abilities to generalize to novel situations, to provide adequate explanations of the outputs, and to reason about complex events. The project will bridge the neural vs. symbolic dichotomy by investigating efficient and adaptable multimodal neurosymbolic (NS) architectures for reasoning about complex spatiotemporal events involving multiple actors doing coordinated maneuvers in dynamic and adversarial settings. Intellectual merit of the project- (1) It re-imagines the critical bottleneck interface between neural and symbolic layers in NS architectures as a dynamic set of concepts that evolves with new data and knowledge. (2) To address efficiency barriers of symbolic layers, it exploits tensorized logic, neural proxies, and NS architecture search. (3) It proposes mechanisms to make NS architectures robust to adversarial perturbations and domain shifts. The research will be validated through datasets from physics simulators with virtual sensors, and lab-scale emulations.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2023
- Source ID
- FA95502210193
Entities
People
- Mani Srivastava
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California, Los Angeles