CIF: Medium: Learning, refining, and understanding models through relational feedback

大多数机器智能系统需要学习算法和提供监督的oracle专家之间的合作。有许多丰富的学习场景，其中来自oracle的数据由关系反馈组成，这些反馈要么是部分顺序，要么是项目之间的偏好，要么是指示项目之间相似性或差异性感知的标签，要么是oracle执行一系列决策以优化自己的效用时的丰富组合。在所有这些设置中，都有一个隐含于oracle的数据空间，研究人员希望利用这个空间中的结构（即学习），明智地从各种数据源中引出和组合知识，以加速这一过程（即提炼），并将该结构的各个方面转移到其他学习模型中（即理解）。研究人员提出了一个合作研究议程，将以两种不同的方式通过关系反馈来改变学习、改进和理解模型的能力：(a)通过利用潜在空间表示来构建更好的模型和算法，(b)通过利用新的关系查询范式来更高效、有效和可解释地从神谕中提取信息。研究人员正在让各个级别和多个部门的学生参与这项高度跨学科的研究工作，预计将有广泛的应用，包括（但不限于）信息检索系统、（逆）强化学习和心理物理实验设计。拟议研究的结果旨在提高对排名系统的理解，排名系统影响新闻和社交媒体网站上的内容显示方式，甚至影响招聘和大学录取的方式。研究人员还打算开发新的工具，用于汇总人类的主观偏好，使用关系查询使人工智能系统与人类价值观保持一致。此外，警务司正从这个项目中拨出资源，扩大传统上代表性不足的群体的参与。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。

Most machine intelligence systems require cooperation between a learning algorithm and an oracle expert providing supervision. There are many rich learning scenarios in which data from the oracle consists of relational feedback, either in terms of partial orders or preferences among items, labels indicating the perception of similarity or differences between items, or rich combinations thereof when the oracle executes a sequence of decisions to optimize her own utility. In all of these settings, there is a data space that is implicit to the oracle, and the investigators wish to exploit structure in this space (i.e., learn), to judiciously elicit and combine knowledge from a variety of data sources to accelerate this process (i.e., refine), and to transfer aspects of this structure into other learned models (i.e., understand). The investigators propose a collaborative research agenda that will transform the ability to learn, refine, and understand models through relational feedback in two distinct ways: (a) by exploiting latent space representations to build better models and algorithms, and (b) by exploiting new relational query paradigms for more efficient, effective, and interpretable information extraction from oracles. The investigators are involving students at all levels and across multiple departments in this highly interdisciplinary research effort that is expected to have broad applications, including (but not limited to) information-retrieval systems, (inverse) reinforcement learning, and psychophysical experimental design. The results of the proposed research are intended to improve the understanding of ranking systems impacting how content is displayed on news and social media sites, and even how hiring and college admissions is conducted. The researchers also intend to develop new tools for aggregating subjective human preferences that use relational queries to align AI systems with human values. Additionally, the PIs are devoting resources from this project to broaden participation among traditionally under-represented groups.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.