CRII: CSR: Adaptive Federated Continuous Learning on Heterogeneous Edge Devices with Unlabeled Data

随着人工智能（AI）技术，特别是深度神经网络（DNN）的日益成功，人们一直在推动在不同领域引入AI服务，包括医疗保健，自动驾驶，图像处理等。由于数据隐私问题和未标记的客户端数据集，人工智能服务提供商必须经常收集和标记自己的数据集，然后在部署之前离线训练他们的模型。然而，这些预先训练的DNN可能无法捕获在线数据的新模式;它们通常必须在用户提供的数据上重新训练。这带来了几个挑战：（1）数据隐私：用户越来越担心未经授权访问其私人数据。(2)未标记的异构数据：边缘设备通常部署在不同的环境中，并由一系列用户拥有，导致本地数据分布的巨大变化。用户也可能缺乏充分标记其数据的动机和/或专业知识。(3)器械异质性：边缘设备具有广泛的计算和存储能力，在这些异构边缘设备上重新训练DNN模型可能非常耗时。为了克服这些挑战，该项目提出了一种自适应、联邦、连续学习系统，该系统使用一种新型的联邦、半监督学习框架，在分布式、无标签、异构数据上重新训练DNN模型，同时利用可解释的人工智能技术来加快本地训练。该项目有望通过增强隐私、适应性和效率来改善人工智能服务在现实场景中的适应性。它支持跨不同领域的AI服务的无缝集成，同时确保数据隐私并优化边缘/客户端设备上的模型性能。该项目还包含重要的教育内容。它将提供机会，让来自计算领域代表性不足的群体的学生参与进来，促进多样性和包容性。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

With the growing success of artificial intelligence (AI) techniques, especially deep neural networks (DNNs), there has been an ongoing push to introduce AI services across different domains, including healthcare, autonomous driving, image processing, and more. Due to data privacy issues and unlabeled client datasets, AI service providers must often collect and label their own datasets, and then train their models offline prior to deployment. However, these pre-trained DNNs may not capture new patterns of online data; they must typically be retrained on user-supplied data. This introduces several challenges: (1) Data Privacy: Users are increasingly concerned about unauthorized access to their private data. (2) Unlabeled Heterogeneous Data: Edge devices are typically deployed in diverse environments and owned by a range of users, leading to substantial variation in the distribution of local data. Users may also lack the motivation and/or expertise to adequately label their data. (3) Device Heterogeneity: Edge devices exhibit a wide spectrum of computing and memory capabilities, and retraining DNN models on such heterogeneous edge devices can be time-consuming. To overcome these challenges, this project proposes an adaptive, federated, continuous learning system, which uses a novel federated, semi-supervised learning framework to retrain DNN models on distributed, unlabeled, heterogeneous data, while leveraging explainable AI techniques to expedite local training. This project holds promise for improving AI service adaptation in real-world scenarios by bolstering privacy, adaptability, and efficiency. It supports seamless integration of AI services across diverse domains, while ensuring data privacy and optimizing model performance on edge/client devices. This project also contains a significant educational component. It will provide opportunities to involve students from groups underrepresented in computing, fostering diversity and inclusion. This will have a positive impact on these students’ education and careers.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.