RTML: Large: Real-Time Autonomic Decision Making on Sparsity-Aware Accelerated Hardware via Online Machine Learning and Approximation

点击翻译按钮获取中文摘要

Real-time smart and autonomic decision making involves two major stages, sensing (of sensor data and then transformation into actionable knowledge) and planning (taking decisions using this knowledge). These two stages happen in both internal and external operations of an Intelligent Physical System (IPS). In case of internal operations, sensing refers to reading data from on-board sensors and planning refers to smart execution of the firmware running on the IPS. In case of external operations, sensing refers to sensing data from externally-mounted sensors and planning refers to executing the software that constitutes an application. In the sensing stage, an IPS should be able to cope with different forms of uncertainty, especially data and model uncertainties. The goal of this research project is to achieve the objectives of online autonomic decision making on sparsity-aware accelerated hardware via Real-Time Machine Learning (RTML) and approximation for a group of IPSs such as drones performing data collection and/or multi-object tracking/classification and operating in a highly dynamic environment that is difficult to model. Remarkably, the techniques adopted in this project generalize well as they can be applied to a variety of IPS domains including natural calamities, man-made disasters, and terrorist attacks. The drone-based distributed multi-object tracking/classification will enable stakeholders such as citizens, government bodies, rescue agencies, and industries to comprehend the extent of damage, and to develop more effective mitigation policies. The research will also train students including minority and underrepresented students in the field.There are three specific tasks in this project. In Task 1, a real-time decision-making approach will be proposed via online deep reinforcement learning with inherent distributed training capability; temporal and spatial correlation in streaming video will then be exploited towards real-time multi-object tracking/detection. In Task 2, novel hardware architectures will be designed to support sparse Convolution Neural Networks (CNN). Considering the dual benefits of sparsity on both lower computational and space complexity for Deep Neural Network (DNN) models, a sparsity-aware CNN accelerator can achieve significant hardware performance improvements in term of latency, throughput, and energy efficiency over non-sparsity-aware techniques. Finally, in Task 3, hardware-aware software engineering solutions will be studied for accelerated execution. The idea of leveraging compiler optimization and the underlying hardware features in combination will be investigated in order to optimize execution performance; then, data-driven modeling techniques will be presented to replace the time-consuming segments of the ML software packages with their equivalent data-driven models, namely micro-neural networks. Once these three research tasks are validated individually via principled experimentation in terms of their stated goals, they will be integrated into a unified framework, which will be thoroughly studied via multiple trials on complementary field scenarios. The project will also collaborate with a synergistic DARPA program for related hardware development.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.