Resource-efficient, self-adaptive neurosymbolic artificial intelligence for the Internet of Battlefield Things
John D. Beggs, Sean M. Coffey, Jared M. Murphy, Nathaniel D. Bastian · 2025
Artificial intelligence (AI) integration into military systems must be able to self-adapt to a variety of operating domains at test-time without requiring extensive re-training, especially in resource-constrained and communicationlimited environments. This paper proffers a neuro-symbolic AI-enabled system incorporated within a hierarchical federated learning (HFL) framework. We conduct unsupervised object detection on simulated, aerial satellite imagery to eliminate reliance on any ground-truth labels in addition to incorporating multi-modal language models to proportionally merge multiple domain-specific models at test-time. In conducting multi-domain testing, our system’s proportionally merged model outperforms single-domain models, highlighting its capability of self-adapting to domain shifts at test-time. When used within a simulated HFL setting, the central server’s global model performance improves with minor re-training. As such, we contribute a feasible means of maintaining system performance in resource-constrained, hierarchical, distributed, multi-domain deployment settings.