CIF: Small: Resource-Efficient Statistical Inference in Networked Environments

近年来，人们对分布式信息处理的需求有了巨大的增长，在分布式信息处理中，数据是在一个大型的分布式代理网络中收集、存储和处理的。例如，使用移动的设备训练更好的机器智能以改善用户体验，使用传感器网络监测环境以减少和减轻野火，改善无人驾驶飞行器在监视中的协调。该项目开始了一个新的框架，以解决以资源有效的方式处理不断增长的大规模数据集的新挑战，这些挑战迄今为止尚未解决。研究人员将通过长期的指导和推广活动，积极招募和培训具有不同背景的学生，包括代表性不足的少数民族和STEM领域的女性。该项目将大大推进从高维分布式数据中进行统计学习和推理的算法实践，通过在网络环境中开发资源有效的分布式统计推断算法，可证明实现最佳交易，在统计误差、计算和通信成本方面，这使得分散和异构数据的可扩展处理成为可能。该项目呼吁将高维统计推断和大规模分散优化紧密结合起来，包括三个主要目标：（i）开发具有可证明收敛保证的资源有效的分散统计估计算法;（ii）发展系统的处理方法，以确保算法收敛，即使在存在高度不平衡和异质数据的情况下，以及促进多样性的异构代理使用图形正则化;（iii）促进分散非参数估计的最佳和自适应早期停止标准。本文开发的工具和技术将进一步促进包括高维统计、大规模优化、统计信号处理和机器学习在内的广泛领域之间的相互作用。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Recent years have witnessed tremendous growth in the need of distributed information processing, where the data are collected, stored, and processed in and through a large network of distributed agents. Examples include training better machine intelligence using mobile devices to improve user experiences, monitoring environments using a network of sensors to reduce and mitigate wildfire, improving the coordination of unmanned aerial vehicles in surveillance. This project embarks on a new framework to address emerging challenges in processing the ever-growing large-scale datasets in a resource-efficient manner that have been unaddressed until now. The investigators will actively recruit and train students with diverse backgrounds including underrepresented minorities and women in STEM through long-term mentoring and outreach activities.This project will substantially advance the algorithmic practice of statistical learning and inference from high-dimensional distributed data, by developing resource-efficient distributed statistical inference algorithms in network environments that provably achieve the optimal trade-offs in statistical error, computation and communication costs, thereby enabling scalable processing of decentralized and heterogenous data. Calling for a tight integration of high-dimensional statistical inference and large-scale decentralized optimization, this project consists of three major thrusts: (i) develop resource-efficient decentralized statistical estimation algorithms with provable convergence guarantees; (ii) develop systematic treatments to ensure algorithmic convergence even in the presence of highly unbalanced and heterogeneous data, as well as promote diversity for heterogeneous agents using graph regularization; (iii) promote optimal and adaptive early stopping criteria for decentralized nonparametric estimation. The tools and techniques developed herein will further foster the interplay between a broad range of fields including high-dimensional statistics, large-scale optimization, statistical signal processing, and machine learning.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.