CRII: CSR: Enabling On-Device Continual Learning through Enhancing Efficiency of Computing, Memory, and Data

由于边缘设备的数量（例如，随着智能手机、物联网设备（Internet of Things devices）的不断扩展，对由设备上人工智能（AI）算法驱动的智能、个性化体验的需求日益增长。这给设备上的AI带来了挑战，需要不断学习新知识，而不会忘记以前学习过的知识。为了解决这个问题，持续学习（CL）旨在构建在一系列任务中增量更新而不会忘记的AI模型。然而，当CL应用程序部署在资源有限的边缘设备上时，学习效率可能会受到影响，这在以前的工作中很少探索。该项目旨在构建一个设备上的CL框架，该框架将允许基于本地数据的持续设备学习，而不会忘记先前学习的知识，同时提高计算，内存和数据使用的效率。具体来说，该项目将追求（1）一个有效的设备上CL框架，避免灾难性遗忘;（2）满足实际部署要求的扩展，包括有限的训练数据可用性和大型数据集中有限的标签。探索这两个案例增强了所提出方法的适用性。该项目促进了由AI驱动的个性化智能用户体验，同时保护数据隐私并实时适应新信息。它将允许人工智能模型增量地从新数据中学习，减少对大型集中式数据集的需求，从而促进隐私。项目结果将在医疗保健自动移动服务的背景下进行探索，并进一步激发设备上CL的各种新兴应用的新思路，包括具有成本效益的自动驾驶和智慧城市应用。该项目还将产生一个教育平台，提供一个用户友好的学习工具，以支持教育目标的K-12学生，本科生，研究生，和学生的群体在计算中代表性不足。这个奖项反映了NSF的法定使命，并已被认为是值得通过评估使用基金会的智力价值和更广泛的影响审查标准的支持。

As the number of edge devices (e.g., smartphones, Internet of Things devices) expands, there is a growing demand for intelligent, personalized experiences driven by on-device artificial intelligence (AI) algorithms. This presents challenges for on-device AI to continually learn new knowledge without forgetting prior learnt knowledge. To address this, continual learning (CL) aims to build AI models that are incrementally updated over a sequence of tasks without forgetting. However, when CL applications are deployed on resource-limited edge devices, learning efficiency can suffer, and this is rarely explored in prior work. This project seeks to build an on-device CL framework that will allow continual on-device learning based on local data without forgetting prior learnt knowledge, while enhancing the efficiency of compute, memory, and data use. Specifically, the project will pursue (1) an efficient on-device CL framework that avoids catastrophic forgetting; and (2) extensions that address practical deployment requirements, including limited training data availability and limited labeling in large datasets. Exploring these two cases enhances the applicability of the proposed approach.This project promotes personalized, intelligent user experiences powered by AI, while preserving data privacy and adapting to new information in real-time. It will allow AI models to learn from new data incrementally, reducing the need for large, centralized datasets, thereby promoting privacy. Project results will be explored in the context of automated mobility services for healthcare, and further spur a new line of thinking for a variety of emerging applications of on-device CL, including cost-effective autonomous driving and smart city applications. The project will also yield an educational platform, offering a user-friendly learning tool to support educational goals for K-12 students, undergraduates, graduates, and students from groups underrepresented in computing.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.