Collaborative Research: III: Medium: Towards Effective Detection and Mitigation for Shortcut Learning: A Data Modeling Framework

深度神经网络的泛化已经成为一个具有挑战性的问题。当数据分布发生变化或输入中存在小干扰时，许多dnn不能保持预测。这一挑战的一个主要原因是快速学习，它指的是基于存在的数据关系（但不是因果关系）做出的决策。当模型被转移到现实世界场景时，由于虚假的相关性，这些决策失败了。本项目旨在研究深度学习中的快捷识别和缓解方法。这项研究的成功结果将导致理论理解的进步，并开发鲁棒和可推广的DNN算法来分析具有各种捷径类型的数据集。整合了机器学习、工业工程和健康信息学的教育计划旨在培养学生掌握信息系统中基本的数据分析工具，以吸引、指导和留住来自代表性不足群体的成员。该项目的主要目标是从数据中心的角度系统地研究快捷特征的识别和缓解，以促进深度学习的推广。所开发的以数据为中心的机制可以直接应用于实际的数据分析系统。具体而言，本项目研究了实例级、特征级、任务级等不同层次的捷径识别与检测，并通过数据增强和训练正则化实现捷径缓解。该项目还演示了如何将提出的研究创新嵌入到两个基于深度神经网络的真实医疗信息系统中。本文提出的框架通过对不同分布转移类别的快捷特征进行校正，揭示了快捷学习的内在特性，并使研究人员和实践者能够理解和采用这些框架。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。

Generalization of Deep Neural Networks (DNNs) has become a challenging problem. Many DNNs do not remain predictive when the distribution of data changes or there are small disturbances in the input. A major reason for this challenge is shortcut learning, which refer to decisions based on relationships in the data that exist, but which are not causal. These decisions fail when the model is transferred to real-world scenarios because of spurious correlations. This project is to investigate shortcut identification and mitigation in deep learning. The successful outcome of this research will lead to advances in providing theoretical understandings, and developing robust and generalizable DNN algorithms to analyze datasets with various types of shortcuts. The education program that integrates machine learning, industrial engineering, and health informatics is to train students with essential data analytics tools in information systems, to attract, mentor and retain members from underrepresented groups.The primary goal of this project is to systematically investigate the identification and mitigation of shortcut features from a data-centric perspective to facilitate the generalization of deep learning. The developed data-centric mechanisms could be directly adopted in real-world data analytics systems. Specifically, this project studies shortcut identification and detection at different levels, including instance-, feature-, and task-levels, and then performs shortcut mitigation through data augmentation and training regularization. This project also demonstrates how the proposed research innovations could be embedded in two DNN based real medical informatics systems. The proposed frameworks uncover the intrinsic properties of shortcut learning by calibrating shortcut features from different categories of distribution shift, and enable their comprehension and adoption for researchers and practitioners.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.