Collaborative Research: CPS: Medium: Real-time Criticality-Aware Neural Networks for Mission-critical Cyber-Physical Systems

人工智能(AI)的进步表明，智能系统将是科学进步的下一次飞跃，使无数未来的应用能够提高生活质量，为经济做出贡献，并增强社会对广泛破坏的适应能力。然而，人工智能的进步带来了相当大的资源成本。为了降低人工智能的成本，这个项目从生物系统中获得灵感。众所周知，人工智能中的一个关键瓶颈是感知子系统。它是让人工智能感知和理解周围环境的部分。人类非常善于理解他们的环境中什么是关键的，人类感知系统自动将有限的认知资源集中在场景中最重要的元素上，从而节省了大量的“大脑处理能力”。目前的人工智能管道没有类似的机制，导致资源成本大幅上升。该项目重新调整数据分析和机器智能管道，以便更好地确定外部刺激的优先顺序，利用并显著扩展先前在实时系统研究界开发的调度方面的进展。重构的人工智能管道将提高人工智能系统的效率和效能，使它们更安全、更灵敏，同时显著降低成本。如果成功，该项目将有助于将机器智能解决方案带给全社会。这是通过研究、教育和推广之间的互动以及多个科学社区的整合来实现的，包括(I)提供平台和调度器的嵌入式计算研究人员，(Ii)物联网和网络研究人员，以及(Iii)智能应用研究人员和应用领域专家。这项工作是网络物理计算研究的一个例子，在这个研究中，新一代数字算法学习利用对物理系统的更好理解，以改善社会成果。该项目消除了现代基于神经网络的网络物理应用中机器智能管道的系统性优先级反转。通常，在实时系统中，当不太关键(或期限较长)的计算先于较关键(或期限较短)的计算执行时，就会发生优先级反转。机器智能软件的当前状态在从感知到决策的过程中存在严重的优先级反转，导致系统对关键事件的响应能力极差，从而危及安全并增加满足应用需求的硬件成本。通过解决这一问题，该项目将提高系统对关键输入的反应能力，同时显著降低平台成本。该项目的学术价值在于研究网络物理计算的两个核心领域的交集：(I)数据分析和机器学习，以及(Ii)实时系统。具体地说，该项目对数据分析和机器智能管道进行了重构，以消除优先级反转。缓解不同系统中的优先级反转问题一直是实时社区的关键贡献之一。从机器智能管道中移除优先级反转做出了其他几项科学贡献。也就是说，(1)重构的人工智能管道提高了启用人工智能的关键任务系统的效率和效力，(2)它使自主系统能够更快地响应，同时降低了它们的成本，(3)它通过确保关键输入首先得到处理，促进了智能系统的安全。该项目希望展示现代基于机器学习的推理协议在性能方面的显著改进，同时提供显著提高对关键情况的反应的可预测性和及时性的服务差异化。如果成功，该项目将显著降低在未来的网络物理系统中部署机器智能解决方案的成本，同时提高可预测性和时间保障。除了提供这个项目的技术贡献外，这项工作的一个明确目的是促进关于智能CPS主题的教育和劳动力发展。这是通过研究、教育和扩大参与的活动之间的互动以及多个社区的整合来实现的，包括(I)提供平台和调度器的嵌入式计算研究人员，(Ii)物联网和网络研究人员，以及(Iii)智能应用和应用领域专家的研究人员。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Advances in artificial intelligence (AI) make it clear that intelligent systems will account for the next leap in scientific progress to enable a myriad of future applications that improve the quality of life, contribute to the economy, and enhance societal resilience to a broad spectrum of disruptions. Yet, advances in AI come at a considerable resource costs. To reduce the cost of AI, this project takes inspiration from biological systems. It is well-known that a key bottleneck in AI is the perception subsystem. It is the part that allows AI to perceive and understand its surroundings. Humans are very good at understanding what’s critical in their environment and the human perceptual system automatically focuses limited cognitive resources on those elements of the scene that matter most, saving a significant amount of “brain processing power”. Current AI pipelines do not have a similar mechanism, resulting in significantly higher resource costs. The project refactors data analytics and machine intelligence pipelines to allow for better prioritization of external stimuli leveraging and significantly extending advances in scheduling previously developed in the real-time systems research community. The refactored AI pipeline will improve the efficiency and efficacy of AI-enabled systems, allowing them to be safer and more responsive, while at the same time significantly lowering their cost. If successful, the project will help bring machine intelligence solutions to the benefit of all society. This is achieved through interactions between research, education, and outreach, as well as integration of multiple scientific communities, including (i) researchers on embedded computing who offer platforms and schedulers, (ii) researchers on IoT and networking, and (iii) researchers on intelligent applications and application domain experts. The work is an example of cyber-physical computing research, where a new generation of digital algorithms learn to exploit a better understanding of physical systems in order to improve societal outcomes. The project removes systemic priority inversion from machine intelligence pipelines in modern neural-network-based cyber-physical applications. In general, priority inversion occurs in real-time systems when computations that are less critical (or with longer deadlines) are performed ahead of those that are more critical (or with shorter deadlines). The current state of machine intelligence software suffers from significant priority inversion on the path from perception to decision-making, resulting in vastly inferior system responsiveness to critical events, thereby jeopardizing safety and increasing the cost of hardware to meet application needs. By resolving this problem, this project shall improve system ability to react to critical inputs, while at the same time significantly reducing platform cost. The intellectual merit of the project lies in investigating the intersection of two core areas in cyber-physical computing: (i) data analytics and machine learning and (ii) real-time systems. Specifically, the project refactors data analytics and machine intelligence pipelines to remove priority inversion. Mitigation of priority inversion problems in different systems has been one of the key contributions of the real-time community. Removal of priority inversion from machine intelligence pipelines makes several other scientific contributions. Namely, (i) the refactored AI pipeline improves the efficiency and efficacy of AI-enabled mission-critical systems, (ii) it enables autonomous systems to be more responsive, while lowering their cost, and (iii) it contributes to safety of intelligent systems by ensuring that critical inputs are processed first. The project expects to demonstrate significant improvements in performance of modern machine-learning-based inference protocols, while offering service differentiation that dramatically improves predictability and timeliness of reactions to critical situations. If successful, the project will significantly reduce the cost of deploying machine intelligence solutions in future cyber-physical systems, while improving predictability and temporal guarantees. In addition to delivering the technical contributions of this project, an explicit purpose of the work is to advance education and workforce development on Intelligent CPS topics. This is achieved through interactions between activities for research, education, and broadening participation, as well as integration of multiple communities, including (i) researchers on embedded computing who offer platforms and schedulers, (ii) researchers on IoT and networking, and (iii) researchers on intelligent applications and application domain experts.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.