CAREER: Adversarial Machine Learning for Structured Prediction

许多重要的归纳推理问题，从理解文本和图像到实现理想的机器人行为，都是结构化预测任务。这些任务需要对许多相关变量进行联合预测，而不是对单个变量进行独立预测。例如，自动驾驶汽车的变道决策可能取决于其对附近车辆的位置和速度估计、其对道路状况的评估、其对道路上其他潜在障碍物的定位和识别等。这个NSF CAREER奖项目的目标是开发更安全和更有益的结构化预测方法。预期的改进有可能在具有关键性能指标的应用领域产生更广泛的影响，例如医疗保健和自动驾驶汽车安全。该项目通过创建数据科学的多学科课程并发布通用对抗性结构化预测工具来培养这些潜力，这些工具将向更广泛的受众展示机器学习技术。此外，在面向多样化学生群体的伊利诺伊大学芝加哥分校，本项目以本科生为对象，通过对不确定性进行推理时的最坏假设，进行结构化预测。在所提出的对抗性结构化预测公式中，该项目的主要技术目标是：（1）提供更强的理论保证（例如，Fisher一致性，更严格的泛化边界）比现有的性能度量近似方法;（2）开发可扩展的算法，用于解决一系列结构和性能度量的大型对抗性结构化预测问题;（3）当从与测试数据分布不同的分布生成的训练数据中学习时，能够实现更安全的结构化预测;以及（4）在自然语言处理、逆最优控制和计算机视觉等各种任务上演示所开发的方法。

Many important inductive reasoning problems, ranging from understanding text and images to enabling desirable robotic behavior, are structured prediction tasks. These tasks require the joint prediction of many related variables rather than independent predictions for individual variables. For example, an autonomous vehicle's lane change decisions may depend on its position and velocity estimates for nearby vehicles, its assessment of road conditions, its localization and identification of other potential obstacles on the roadway, and so on. The goal of this NSF CAREER award project is to develop safer and more beneficial structured prediction methods. Anticipated improvements have the potential for broader impact in application areas with critical performance measures, such as healthcare and autonomous vehicle safety. This project fosters these potentials by creating multidisciplinary curriculum in data science and releasing general purpose adversarial structured prediction tools that will expose machine learning techniques to a wider audience. Additionally, the project seeks to involve undergraduates in research activities at the University of Illinois at Chicago, which is an urban institution serving a diverse student population.The approach pursued in this project is to perform structured prediction by making worst-case assumptions when reasoning about uncertainty. The main technical objectives of this project within the proposed adversarial structured prediction formulation are to:(1) Provide stronger theoretical guarantees (e.g., Fisher consistency, tighter generalization bounds) than existing performance measure approximation methods;(2) Develop scalable algorithms for solving large adversarial structured prediction problems for a range of structures and performance measures;(3) Enable safer structured prediction when learning from training data that is generated from a different distribution than the testing data distribution; and (4) Demonstrate the developed methods on a diverse range of tasks from natural language processing, inverse optimal control, and computer vision.