RTML: Small: Collaborative: A Programming Model and Platform Architecture for Real-time Machine Learning for Sub-second Systems

点击翻译按钮获取中文摘要

This project develops and evaluates novel frameworks for achieving real-time machine learning; that is, for a given target application that is producing a lot of data, how to process that data to concurrently prediction what comes next while learning from the past data at the same pace of the target application. The developed framework will produce adaptive models suitable to predict the behavior of the complex dynamics found in sub-second systems. Such systems include adaptive airbag deployment mechanisms, hypersonic vehicles, and active impact mitigation systems. Solutions will be developed to learn the dynamics at the data rates required to enable real-time decision-making systems such as those used for active control and adaptive operations. These solutions are designed for direct integration into sub-second systems to increase their resilience, robustness, safety, and viability. It follows that this research will impact society by enabling sub-second systems and empowering decision-making capabilities at speeds never reached before. Several undergraduate students will be included in the project with an emphasis on providing research experiences to underrepresented, first-generation, and low-income students by leveraging existing and valuable resources at both the University of South Carolina and Iowa State University. This project will also produce two multidisciplinary Ph.D. students with expertise in machine learning, high-rate dynamics, and control. The novelty of the approach taken in this project is to tune hyper-parameters to facilitate the use of an array of concurrent models to hide training latency. More specifically, field programmable gate arrays (FPGAs) are used to store and update the parameters of multiple concurrent long short-term memory networks as well as embed physical knowledge at the neurons' input level. This will require that the machine learning algorithm learn the temporal dependencies across operating regimes and adapt to varying dynamics. The resulting algorithm is a novel type of long short-term memory recurrent neural network that enables the prediction of nonlinear and non-stationary time series. Multiple iterations of this algorithm will be run in parallel on a single FPGA where the training time of one algorithm can be effectively hidden by another algorithm performing inference in parallel. The formulated algorithm will advance the field of real-time machine learning by furthering knowledge on: 1) how parallel models interact to hide training latency; 2) the effect of automated tuning of model parameters; 3) the role of physical knowledge in designing input spaces; 4) the benefits of subdividing non-stationary time series into local stationary systems; and 5) sustaining sufficient accuracy while meeting real-time constraints in the micro-second realm.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.