ATD: Multimode Machine Learning and Deep GeoNetworks for Anomaly Detection

这项研究为分析时空数据集创造了数学概念和数值方法，特别强调异常检测。早期和准确地发现异常事件并预测未来的威胁对于设计有效的应对措施至关重要。当前的威胁检测算法往往无法满足众多的需求、不断变化的环境以及需要处理和分析大量的时空数据来完成这些任务。不确定性、规模、非平稳性、噪声和异质性是阻碍从数据中获取知识的各个阶段进展的基本问题。这项研究工作的目标是开发新的数学概念和计算方法，可以检测异质、大规模、时空数据集中的异常。除了该项目广泛的技术影响之外，它还为数学/工程前沿的研究和教育提供了一种跨学科活动的模型。PI将为无监督和半监督学习设计高效、鲁棒和可扩展的算法。特别地，PI将侧重于发展两种方法：(i)用于从时空数据中进行无监督异常预测的多模态扩散框架。这里，多模态是指数据可能具有不同的模态，如文本、图像、地理位置等。（ii）基于上述多模态扩散框架和专门设计用于操作时空数据的深度卷积网络的新变体，用于图结构数据的半监督学习的可扩展框架。该项目的预期成功是基于研究者在发展先进数学概念并将其转化为现实世界应用方面的现有成就。

This research effort creates mathematical concepts and numerical methods for the analysis of spatiotemporal datasets with a special emphasis on anomaly detection. Early and accurate detection of unusual events and forecasts of future threats are critical in designing an effective response to them. Current algorithms for threat detection are often unable to keep up with the numerous demands, changing environments, and the huge amounts of spatiotemporal data that need to be processed and analyzed to accomplish these tasks. Uncertainty, scale, non-stationarity, noise, and heterogeneity are fundamental issues impeding progress at all phases of the pipeline that creates knowledge from data. The goal of this research effort is to develop novel mathematical concepts and computational methods that can detect anomalies in heterogenous, large-scale, spatiotemporal datasets. Beyond the project's broad technological impact, it serves as a model for the kind of cross-disciplinary activity critical for research and education at the mathematics/engineering frontier.The PI will devise efficient, robust, and scalable algorithms for unsupervised and semi-supervised learning. In particular, the PI will focus on the development of two approaches: (i) A multimodal diffusion framework for unsupervised prediction of anomalies from spatiotemporal data. Here, multimode refers to the fact that data may have different modalities, such as text, images, geolocations, etc. (ii) A scalable framework for semi-supervised learning on graph-structured data, based on the aforementioned multimodal diffusion framework and on a novel variant of deep convolutional networks specifically designed to operate on spatiotemporal data. The expected success of this project is based on existing achievements by the investigator in developing advanced mathematical concepts and turning them into real-world applications.