SCenE - Self-Assessment and Continual Learning on Edge Devices

人工智能(AI)系统和机器学习算法在理论上处于现代自主的核心，但在扩展到现实系统方面遇到了瓶颈。通常，在现实环境中，人工智能系统被认为是不可信任的，缺乏适应不断变化的环境的能力，需要持续维护和调整才能保持相关性。目前的大多数人工智能系统都受到培训和开发过程中收集的知识的限制。为了使系统真正智能化，它们必须包含学习框架，这些框架意识到自己的局限性，在部署后出现故障时具有可扩展的知识库，并具有在连续和动态的真实环境中在可用能源预算内运行的能力。该项目的目标是开发一个严格和可扩展的学习框架，使数据驱动的算法能够自我评估其性能并在其先前知识的基础上不断扩展，同时在有限的能源预算下实时运行。这项工作将通过与工业合作伙伴以及地方、地区和联邦政府机构的合作，同样影响学术研究和经济发展。案例研究的例子包括医疗保健、智能交通系统、监控、恶劣天气和洪水监测、航空和旋翼飞机安全、农业、植被和濒危物种监测，以及智能和互联的校园和社区。与大西洋角社区学院的合作将作为向下一代STEM学生传播研究贡献的基础。开发的算法、源代码和硬件配置将通过开源数据共享平台向公众开放。我们的目标是解决目前人工智能系统和学习算法的局限性，这些系统和学习算法是基于确定性和过度自信的深度神经网络。这些模型的学习参数在训练后被冻结，并部署在可能能量受限的边缘平台上。这些模型不能适应非平稳环境，导致在不断变化的环境中出现故障。该项目的目标是开发一个严格的、可扩展的和开源的学习框架，以促进数据驱动算法的开发和部署，该算法可以自我评估性能并持续适应流数据集，同时以有限的能源预算实时运行。我们为机器学习系统提出了一种新的基本方法，它将：(1)通过在未知网络参数上传播分布矩来量化网络决策中的置信度，为现代学习算法的自我评估提供理论基础，(2)通过监控估计的预测分布的方差-协方差参数来促进自我评估方法的发展，(3)通过利用基于方差-协方差信息的核重要性度量来获得允许算法在给定功率预算内运行的新的训练方法，同时通过利用基于方差-协方差信息的核重要性度量来实现从流数据集的连续自适应，以及(4)使用基准公共数据集和与我们的政府、行业和学术合作的真实世界应用程序来评估数学推导和随后开发的算法的有效性。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。

Artificial intelligence (AI) systems and machine learning algorithms lay at the heart of modern autonomy in theory but have experienced a bottleneck in expansion into real-world systems. Typically, in the case of real-world environments, AI systems are considered untrustworthy and are lacking the ability to adapt to an ever-changing environment, requiring continuous maintenance and tuning to stay relevant. Most current AI systems are constrained by their knowledge gathered during training and development. In order for a system to be truly intelligent, they must incorporate learning frameworks that are aware of their own limitations, have an expandable knowledge base in case of failure after deployment, and have the capabilities to operate within available energy budgets in a continuous and dynamic real-world environment. The goal for this project is to develop a rigorous and scalable learning framework that will enable the development of data-driven algorithms that can self-assess their performance and continually expand upon their prior knowledge while operating in real-time on a limited energy budget. This work will equally impact academic research and economic development through collaborations with industrial partners as well as local, regional and federal government agencies. The case study examples include healthcare, intelligent transportation systems, surveillance, severe weather and flood monitoring, aviation and rotorcraft safety, agriculture, vegetation, and endangered species monitoring, and smart and connected campus and communities. Collaboration with the Atlantic Cape Community College will serve as a basis to disseminate the research contributions to the next generation of STEM students. The developed algorithms, source code, and hardware configurations will be made available to the public through open-source data-sharing platforms. We aim to tackle the limitations of the current AI systems and learning algorithms, which are based on deterministic and over-confident deep neural networks. The learned parameters of these models are frozen after training and deployed on possibly energy-constrained edge platforms. These models cannot adapt to non-stationary environments resulting in failures in continuously changing environments. The objective of this project is to develop a rigorous, scalable, and open-source learning framework that would facilitate the development and deployment of data-driven algorithms, which can self-assess performance and continually adapt to streaming datasets while operating in real-time on a limited energy budget. We propose a new fundamental approach to machine learning systems that will: (1) provide a theoretical foundation for self-assessment of modern learning algorithms via quantifying confidence in network decisions through the propagation of distribution moments over unknown network parameters, (2) spur the development of self-assessment methods through the monitoring of variance-covariance parameters of the estimated predictive distribution, (3) derive new training methods that allow for algorithms to operate within a given power budget while achieving continual adaptation from streaming datasets through leveraging metrics of kernel importance based on variance-covariance information, and (4) assess the validity of the mathematical derivations and subsequently developed algorithms using benchmark public datasets and real-world applications with our government, industry and academic collaborators.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.