III: Medium: Collaborative Research: MUDL: Multidimensional Uncertainty-Aware Deep Learning Framework

由于缺乏信息、信息冲突和/或不确定的观察结果带来的不确定性，人们在为现实世界的问题寻找有效的决策解决方案时遇到了严重的障碍。关键的安全问题一直被强调，因为如何解释这种不确定性还没有仔细研究。如果不确定性被误解，这可能会导致不必要的风险。例如，一辆自动驾驶汽车可能会误判道路上的人。基于人工智能的医疗助理可能会将癌症误诊为良性肿瘤。此外，可以将网络钓鱼电子邮件检测为正常电子邮件。由不同类型的不确定性引起的所有这些错误检测或错误分类的后果增加了风险和潜在的不良事件。人工智能（AI）研究人员积极探索如何解决不确定性下的各种决策问题。然而，之前没有研究调查过研究人工智能中不确定性的不同方法如何相互利用。本项目研究如何衡量不确定性的不同原因，并利用它们更有效地解决各种决策问题。这个项目可以帮助开发可用于许多现实世界决策问题的可靠人工智能算法。此外，这个项目是高度跨学科的，因此它可以鼓励更广泛、更新和更多样化的方法。为了扩大该项目在研究和教育方面的影响，该项目为不同背景和代表性不足的人群的学生提供多元文化，多样性和STEM课程。该项目还包括研讨会、讲习班、短期课程和/或针对高中和社区大学生的研究项目。该项目旨在通过考虑由不同根本原因引起的多种类型的不确定性来开发一套深度学习（DL）技术，并利用它们在存在高度智能的对抗性攻击时最大限度地提高决策的有效性。本项目以协同性和变革性的研究努力来研究：(1)如何基于信念理论对不同类型的不确定性进行量化；(2)如何在基于dl的方法中考虑不同类型不确定性的估计；(3)多类型不确定性如何影响高维复杂问题决策的有效性和效率。本项目通过以下工作推进了最先进的研究：(1)提出了一个可扩展的、健壮的统一的基于dl的框架，以有效地推断敌对环境中由异构根本原因引起的预测性多维不确定性。(2)基于神经网络的多维不确定性处理。(3)考虑多维不确定性感知设计，提高决策有效性和效率。(4)采用基于仿真模型和可视化工具的先进欺骗策略，提出了在智能对抗性攻击者存在时确保其鲁棒性的方法。该奖项反映了美国国家科学基金会的法定使命，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。

People encounter serious hurdles in finding effective decision-making solutions to real world problems because of uncertainty from a lack of information, conflicting information, and/or unsure observations. Critical safety concerns have been consistently highlighted because how to interpret this uncertainty has not been carefully investigated. If the uncertainty is misinterpreted, this can result in unnecessary risk. For example, a self-driving autonomous car can misdetect a human in the road. An artificial intelligence-based medical assistant may misdiagnose cancer as a benign tumor. Further, a phishing email can be detected as a normal email. The consequences of all these misdetections or misclassifications caused by different types of uncertainty adds risk and potential adverse events. Artificial intelligence (AI) researchers have actively explored how to solve various decision-making problems under uncertainty. However, no prior research has looked into how different approaches of studying uncertainty in AI can leverage each other. This project studies how to measure different causes of uncertainty and use them to solve diverse decision-making problems more effectively. This project can help develop trustworthy AI algorithms that can be used in many real world decision-making problems. In addition, this project is highly transdisciplinary so that it can encourage broader, newer, and more diverse approaches. To magnify the impact of this project in research and education, this project leverages multicultural, diversity, and STEM programs for students with diverse backgrounds and under-represented populations. This project also includes seminar talks, workshops, short courses, and/or research projects for high school and community college students. This project aims to develop a suite of deep learning (DL) techniques by considering multiple types of uncertainties caused by different root causes and employ them to maximize the effectiveness of decision-making in the presence of highly intelligent, adversarial attacks. This project makes a synergistic but transformative research effort to study: (1) how different types of uncertainties can be quantified based on belief theory; (2) how the estimates of different types of uncertainties can be considered in DL-based approaches; and (3) how multiple types of uncertainties influence the effectiveness and efficiency of decision-making in high-dimensional, complex problems. This project advances the state-of-the-art research by performing the following: (1) Proposing a scalable, robust unified DL-based framework to effectively infer predictive multidimensional uncertainty caused by heterogeneous root causes in adversarial environments. (2) Dealing with multidimensional uncertainty based on neural networks. (3) Enhancing both decision effectiveness and efficiency by considering multidimensional uncertainty-aware designs. (4) Testing proposed approaches to ensure their robustness in the presence of intelligent adversarial attackers with advanced deception tactics based on both simulation models and visualization tools.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.