CAREER: Coresets for Robust and Efficient Machine Learning

大数据集使现代机器学习模型在各种应用中取得了前所未有的成功，从医疗诊断到城市规划和自动驾驶，不一而足。然而，从海量数据中学习取决于异常庞大和昂贵的计算资源。这样的基础设施会消耗大量能源，产生大量的碳足迹，而且往往很快就会过时，变成电子垃圾。虽然人们一直在努力提高机器学习模型的性能和可靠性，但它们的可持续性往往被忽视。该项目旨在通过选择最相关的数据进行训练，解决机器学习的可持续性、可靠性和效率问题。由此产生的算法将广泛适用于从广泛的应用程序中的海量数据集进行学习，例如医疗诊断和环境传感。这项研究的成果将被纳入课程开发，以培养新一代机器学习和数据挖掘实践者。该项目的主要目标是开发新一代理论上严格的方法，使从海量数据集进行高效和稳健的学习成为可能。为了实现这一目标，该项目将开发可扩展的组合优化算法来提取数据的加权子集(核集)，以保证类似于对全部数据进行训练的训练动态。这使得能够从海量数据中持续、高效和准确地学习。随着数据集变得越来越大，保持它们的质量变得非常昂贵。因此，错误标记和恶意示例在大型数据集中变得无处不在。为了确保除可持续性和效率之外的可靠性，将利用开发的技术提取数据点的核心集，从而能够针对嘈杂的标签和对抗性攻击进行可证明的稳健学习。该项目还将寻求更好的学习目标，以自动提取最有价值的数据，以便从海量数据中进行高效和可靠的学习。最后，这项研究将实现高效和强大的学习框架，可以通过跨学科合作应用于许多现实世界。这一奖项反映了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。

Large datasets have enabled modern machine learning models to achieve unprecedented success in various applications, ranging from medical diagnostics to urban planning and autonomous driving, to name a few. However, learning from massive data is contingent on exceptionally large and expensive computational resources. Such infrastructures consume substantial energy, produce a massive amount of carbon footprint, and often soon become obsolete and turn into e-waste. While there has been a persistent effort to improve the performance and reliability of machine learning models, their sustainability is often neglected. This project aims to address sustainability, reliability, and efficiency of machine learning, by selecting the most relevant data for training. The resulting algorithms will be broadly applicable for learning from massive datasets across a wide range of applications, such as medical diagnosis and environment sensing. The outcomes of this research will be incorporated into curriculum development, to train a new generation of machine learning and data mining practitioners.The main objective of this project is to develop a new generation of theoretically rigorous methods that enable {efficient and robust learning from massive datasets. To achieve this goal, this project will develop scalable combinatorial optimization algorithms to extract weighted subsets (coresets) of data that guarantee similar training dynamics to that of training on the full data. This enables sustainable, efficient, and accurate learning from massive data. As datasets grow larger, maintaining their quality becomes very expensive. Hence, mislabeled and malicious examples become ubiquitous in large datasets. To ensure reliability in addition to sustainability and efficiency, the developed techniques will be leveraged to extract coresets of data points that enable provably robust learning against noisy labels and adversarial attacks. This project will also seek to learn better objectives to automatically extract the most valuable data for efficient and robust learning from massive data. Finally, this research will enable efficient and robust learning frameworks that can be applied to many real-world applications through interdisciplinary collaborations.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.