Collaborative Research: RI: AF: Medium: Exchanging Knowledge Beyond Data Between Human and Machine Learner

深度学习的最新进展在解决语音识别和目标检测等基本感知任务方面取得了显著进展。为了为这些进步可能实现的许多以人为中心的应用程序铺平道路，例如在医疗保健领域，重要的是超越分类问题：认为机器学习系统不仅能产生类别预测，还能产生预测的原因。此外，这些推理模式需要为人类所理解。为了实现这一点，该项目将侧重于人类和机器学习系统之间的知识交流，以及这种数据以外的知识交流如何能够产生更好的预测，也是人类可以解释的。该项目将带来技术进步，有可能对机器学习在面向人类的应用中的可用性产生重大影响。该项目的技术目标沿着两个广泛的主题发展。第一个问题是，“我们如何在机器学习过程中引入人类反馈，以创建可解释的简明模型，并生成可解释的预测？”通过使人类能够以经验规则作为关系知识的形式提供丰富的反馈，该项目旨在推导出简洁、可解释的机器学习模型，这些模型服从简单的解释，更符合人类的因果世界观。为了增强机器学习的可解释性，该项目将进一步探讨如何利用基于关系知识的人类反馈来减少训练准确模型所需的数据集的大小。第二个问题是，“我们如何编码和利用关系信息来推导可解释和可解释的推理模型？”该项目将探索关系知识在向量空间和逻辑模型中的编码，并进一步研究如何将关系知识用于类比推理、语义理解和关系查询。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Recent advances in deep learning have made dramatic progress in solving basic perceptual tasks such as speech recognition and object detection. To pave the way for the many human-centered applications that these advances might enable, in healthcare for instance, it is important to move beyond classification problems: to think of machine learning systems as producing not just category predictions, but also the reasons for them. Moreover, these patterns of reasoning need to be comprehensible to humans. To enable this, this project will focus on the exchange of knowledge between humans and machine learning systems and how such exchange of knowledge beyond data can lead to better predictions that are also human-interpretable. The project will result in technological advances that will have the potential to significantly impact the usability of machine learning in human-facing applications.The technical aims of this project are developed along two broad themes. The first addresses the question, "How can we involve human feedback in the machine learning process to create succinct models that are interpretable and generate predictions that are explainable?" By enabling humans to provide rich feedback in the form of rules-of-thumb as relational knowledge, the project aims to derive succinct interpretable machine learning models that are amenable to simple explanations that are more compatible with the causal world-view of humans. To enhance the interpretability of machine learning, the project will further explore how human feedback based on relational knowledge can be leveraged to reduce the size of data sets required to train accurate models. The second addresses the question, "How can we encode and exploit relational information in deriving interpretable and explainable models for reasoning?" The project will explore the encoding of relational knowledge in both vector spaces and logical models and further investigate how relational knowledge can be used for analogical reasoning, semantic understanding, and relational queries.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.