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

深度神经网络（DNN）泛化是一个具有挑战性的问题。当数据分布发生变化或输入受到小干扰时，许多深度神经网络不能保持预测。这种行为的一个常见原因是“捷径学习”，在这种学习中，DNN根据数据中观察到的关系学习做出决定，但这不是因果关系。当模型被转移到现实场景时，这些决策就会失败，因为网络已经锁定了虚假的相关性。本项目研究如何识别和缓解dnn中的快速学习。这项研究的成功结果将导致理论理解的进步，以及避免捷径的鲁棒和可推广的DNN算法。该教育计划整合了机器学习、工业工程和健康信息学，以培养学生掌握信息系统中基本的数据分析工具，并吸引、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为中心的角度系统地研究快捷特征的识别和缓解，以促进深度学习的泛化。所开发的以数据为中心的机制可以直接应用于现实世界的数据分析系统中，以减轻快捷学习的缺点。项目从实例、特征、任务等不同层次研究快捷方式识别与检测，通过数据增强和训练正则化实现快捷方式缓解。该项目还演示了拟议的研究创新如何嵌入到两个真实的基于dnn的医疗信息系统中。提出的框架通过校准不同类型分布转移的快捷特征，揭示了快捷学习的内在属性，应该为研究人员和实践者提供支持。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。

Deep Neural Network (DNN) generalization is a challenging problem. Many DNNs do not remain predictive when the distribution of data changes or there are small disturbances to their input. A common reason for this behavior is “shortcut learning”, in which the DNN learns to make decisions based on relationships observed in the data, but that are not causal. These decisions fail when the model is transferred to real-world scenarios because the network has latched onto spurious correlations. This project investigates how to identify and mitigate shortcut learning in DNNs. A successful outcome of this research will lead to advances in theoretical understanding, as well as robust and generalizable DNN algorithms that avoid shortcuts. The education program integrates machine learning, industrial engineering, and health informatics to train students with essential data analytics tools in information systems, as well as 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 generalization in deep learning. The developed data-centric mechanisms could be directly adopted in real-world data analytics systems to mitigate the drawbacks of shortcut learning. The 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. The project also demonstrates how the proposed research innovations could be embedded into two real DNN-based medical informatics systems. The proposed framework uncovers intrinsic properties of shortcut learning by calibrating shortcut features across different types of distribution shifts, and should support both 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.