Optimizing Neurosymbolic AI-ML Models for Latency, Accuracy, and Robustness on Resource-constrained Edge Computing Platforms
Abstract
The research is developing a novel hardware-in-loop neurosymbolic architecture search approach for multi-domain operations that enables creation of systems combining neural and symbolic AI-ML models for real-time processing of sensory information and wireless signals on extremely low-SWAP platforms. The underlying Bayesian hyperparameter optimization is however extremely compute and I-O intensive. This Defense University Instrumentation Program (DURIP) proposal aims to establish a system to perform the hardware-in-loop hyperparameter search for neurosymbolic models with desired end-to-end latency, prediction and decision accuracy, and robustness characteristics for ultra-low-resource platforms while accommodating a range of compute targets including embedded processors, neural accelerators, field-programmable gate arrays, and embedded GPUs. The proposed system consists of a multi-GPU server allowing parallel hyperparameter search for rapid architecture space exploration, a custom-fabricated flexible heterogeneous edge platform for real-time hardware-in-loop evaluation of candidate neurosymbolic models on diverse edge platforms, a cluster of AI-ML enabled software defined radios for neurosymbolic models in networked settings, and an edge server for code-generation and edge platform orchestration. Besides allowing scaling of the AFOSR-funded research in terms of diversity of edge platform types and real-time performance, the proposed system and associated research will also benefit synergistic Department of Defense projects at UCLA that involve learning-enabled cyber-physical systems on resource constrained platforms, such as research under the Army Research Laboratory (ARL) funded Internet of Battlefield Things (IoBT) CRA and research on human-robot collaboration under the AI-ML Research for Expeditionary Maneuver and Air-Ground Reconnaissance program. The system and associated research will lead to new trustworthy AI architectures that efficiently learn and reason in dynamic and complex all-domain operations, and open-source software and synthetic datasets to enable research validation and reproducibility.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2024
- Source ID
- FA95502310559
Entities
People
- Mani Srivastava
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California, Los Angeles