Machine Learning-assisted Modeling and Design of Approximate Computing with Generalizability and Interpretability

点击翻译按钮获取中文摘要

By 2040, the projected energy consumed by computers will exceed the electricity the world can generate, unless radical changes are made in the way we design computers. This project aims to develop approximate computing techniques to drastically reduce the energy consumption in modern computation-intensive computing, for example, video/image processing and machine learning applications. Approximate computing is a promising technique that can trade off accuracy for energy saving and performance improvement. This project addresses one of the fundamental obstacles that has been impeding the practical usage of approximate computing: how to accurately and quickly design approximate computing systems to maximize the benefits of approximate computing without introducing too much accuracy loss or errors. The success of this project can greatly improve the practicality of approximate computing and enable its wide usage in real-world applications, such as video/image processing and machine learning. By paving the way for future approximate computing, this project can lead to considerable energy saving for future computing paradigm and carbon footprint reduction. It also advances the applications of machine learning algorithms in circuit design area, as well as identifies several fundamental machine learning questions motived by special features of the circuit design problem, such that machine learning algorithms can better benefit hardware design. This can lead to new design and testing approaches for a broad range of computing systems, from low-power embedded systems to high-performance data centers. The educational plan will integrate research activities into curriculum development and will provide students with early research training. The team is committed to broadening the participation of undergraduates and underrepresented groups in engineering research and in STEM outreach activities.Given the huge amount of energy consumed by modern computation-intensive computing such as machine learning and video/image processing applications, energy efficient computing is an urgent need. Approximate computing, by slightly trading off accuracy for better performance and/or efficiency, e.g., computation latency, area, and energy and power, has been a promising new computing paradigm. Many approximate computing approaches, such as low-precision computing, voltage scaling, inexact approximate circuits and memory, have achieved orders of speedup or energy saving. However, to safely deploy approximate computing in practice, two major challenges need to be addressed: (1) how to accurately and quickly estimate the impact of approximation on application output quality; and (2) how to accurately and quickly find the best approximation configuration to maximize the benefits of approximate computing. This proposal presents three closely-interacted research tasks to address these two challenges and to seek the wide-reaching benefits of approximate computing: (1) to develop input-aware error models of approximate circuits and input-aware simulation platform for approximate computing; (2) to develop a graph neural networks (GNNs)-based framework to quickly estimate application output quality; (3) to develop a resource-aware approximation configuration framework to optimize performance/energy while satisfying user-defined quality constraints. The goal of this project is to unveil the underlining knowledge of the intrinsic relations between output quality and input data, approximate circuits, and approximate program structures. The project will provide a practical, generalizable, and interpretable toolset that can learn to configure approximate computing once and for all.The intellectual merits of this project include both AxC design innovations as well as machine learning innovations. (1) This project will develop input-aware error models for approximate circuits considering the impact of input data, which are usually overlooked; then, it exposes the circuit-level errors to behavioral-level approximate programs by developing an input-aware simulation platform. This will form the foundation for a much-needed holistic evaluation of approximate computing. (2) This project will develop inductive GNN models and machine learning models to predict the output quality of any unseen approximate programs and approximation configurations. This will provide key generalizability and interpretability of approximate computing. In addition, the investigation of GNN in this project will uncover two new fundamental studies to GNN community, increasing GNN expressiveness power by amending graph connectivity, and utilizing graph regularity. (3) This project will design a resource-aware reinforcement learning (RL) based approach to automatically configure approximation settings for optimal performance/energy-quality tradeoff. In addition, the investigation of GNN and RL in this project will uncover a new research question, the joint optimization of the RL agent and the surrogate model. This project is a pioneering approach to the joint areas of reinforcement learning, graph neural networks, and approximate computing. It aims to establish the technological foundation for practical approximate computing.This project will bring an unprecedented transformation in our ability to understanding and designing approximate computing for practical use, by enabling a more disciplined, generalizable, and interpretable approximation. The research team will release models, tools, and infrastructures to the research and industry community. This can lead to new design and testing approaches for a broad range of computing systems, from low-power embedded systems to high-performance data centers. The educational plan will integrate research activities into curriculum development and will provide students with early exposure to research. The PIs are committed to broadening the participation of undergraduates and underrepresented groups in engineering research and in STEM outreach activities.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.