The scope of this research project is to develop novel engineering solutions to enable real-time autonomic decision making on sparsity-aware accelerated hardware via online Machine Learning (ML) and approximation. Our research focuses on ultra-low power computing and real-time processing across multiple domains including aerial robotics, healthcare monitoring, and IoT systems. The project addresses key challenges in implementing sophisticated real-time and data-driven computing capabilities within strict power and resource limitations. This includes developing battery-less solutions for continuous monitoring, context-aware computational frameworks for drones, and efficient Neural Network (NN) architectures for resource-constrained devices. The designs are intentionally kept fundamental and theoretically solid to ensure their applicability across various cyber-physical platforms, such as autonomous driving and critical infrastructure operations including industrial control and smart grid systems.

The primary objectives include: (1) Developing real-time decision-making capabilities through online Deep Reinforcement Learning (DRL) for multi-drone coordination, focusing on extending Deep Q-Networks (DQN) and Deep Deterministic Policy Gradient (DDPG) algorithms for continuous domains; (2) Creating data-aware methods that leverage temporal and spatial correlation in streaming video for efficient object tracking and detection; (3) Designing sparsity-aware dataflow architectures for CNN inference, particularly for complex CNN architectures with shortcuts and interconnections; (4) Developing sparsity-aware architectures for real-time CNN training, including efficient DRAM access schemes for backward propagation; (5) Optimizing ML software through performance-aware runtime reconfiguration and hardware-specific optimizations; (6) Implementing approximate computing solutions to accelerate execution through dynamic library loading and neural network approximation models; and (7) Creating reliable multi-agent coordination mechanisms to address challenges including local optima, absence of global environmental information, and inconsistent communication in complex environments. These objectives collectively address the challenge of enabling real-time autonomic decision-making on resource-constrained hardware while maintaining computational efficiency.

Our research has yielded several breakthrough results with significant practical implications: (1) Successfully demonstrated battery-less operation for implantable EEG monitoring devices; (2) Achieved significant improvements in drone resource utilization and real-time performance, validated through hardware-in-the-loop emulations on NVIDIA Jetson TX2 GPU using the Microsoft AirSim simulator; (3) Developed an analog Folded Neural Network achieving optimal performance for cardiovascular monitoring with 6 layers and a hidden size of 30; (4) Created a crowd monitoring system processing patterns in just 2 milliseconds, representing a 45x speed improvement; (5) Achieved 55% reduction in computation for face authentication while maintaining security, with encryption time savings of 20 to 55 times; (6) Demonstrated superior performance in multi-agent coordination for complex Search And Rescue (SAR) operations. These advances have broad applications in healthcare, surveillance, disaster response, and IoT systems, making sophisticated AI capabilities more accessible and energy efficient. Additionally, the research has contributed to underwater joint source-channel coding, helping democratize acoustic underwater communications.

The project has made significant contributions to education and community development through: (1) Training graduate students in deep learning, hardware architecture design, and NN optimization; (2) Mentoring undergraduate students in robotics, underwater acoustics, AI, and ML through programs like the Aresty Program and departmental summer internships; (3) Organizing educational events like the HackRU Capture The Flag Hackathon focusing on cyber-physical drone security; (4) Creating new courses, such as Co-PI Yuan's course on advanced digital systems for ML and signal processing, and incorporating research findings into existing curricula; (5) Supporting minority student outreach programs and hosting diverse student internships; (6) Developing STEM literacy and workforce skills related to smart and autonomous drones for disaster response. Additionally, our research has resulted in several provisional patents filed through Rutgers University. The technology developed also shows promise in precise agriculture applications for early pest detection and disease intervention.

Last Modified: 02/10/2025
Modified by: Dario Pompili