Opportunity: Deep Learning at the Tactical Edge

This opportunity seeks methods to efficiently deploy CNNs, vision and multimodal transformers, and language models on resource-constrained tactical-edge platforms. The goal is to overcome the power, memory, thermal, and compute limits of deployed hardware, enabling low-latency, on-device inference and reasoning in austere or communications constrained environments. Topics include edge-computing methods, hardware–software co-design, neural architecture search, neuromorphic systems, and optimized embedded accelerators such as GPUs, NPUs, and FPGAs. Proposed approaches should improve model efficiency, accuracy, robustness, scalability, and inference management through advances in optimization, compression, abstraction layers, and deployment frameworks. Priority areas also include multimodal sensor fusion using unified transformer encoders and model-specific improvements, such as CNN feature extraction and transfer learning, as well as LLM and small-transformer prompt optimization, domain adaptation, continual learning, contextual reasoning, and efficient inference. Solutions should support U.S. Army missions including target detection, anomaly and threat assessment, autonomous navigation, and human–machine teaming.