FAI: Towards Adaptive and Interactive Post Hoc Explanations

FAI：迈向自适应和交互式事后解释

• 批准号: 2040989
• 负责人: Chenhao Tan
• 金额: $ 37.5万
• 依托单位: University of Chicago
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-02-01 至 2025-01-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2040989&HistoricalAwards=false
• 关键词: FAI; Towards; Adaptive; Interactive; Post

Explaining machine learning (ML) models have received increasing interest because of their adoption in societally-critical tasks, ranging from health care, to hiring, to criminal justice. It is crucial for the relevant parties, such as decision makers and decision subjects, to understand why a model makes a particular prediction. This proposal argues that explanations represent a communication process. In order to improve the effectiveness of explanations, explanations should be adaptive and interactive based on the subject being explained (subgroups of interest) as well as the target audience (user profiles), whose knowledge and preferences may be evolving. Therefore, this proposal aims to develop adaptive and interactive explanations of machine learning models, which will allow people to better understand the decisions being made for and about them. This proposal has three key areas of focus. First, this proposal will develop a novel formal framework for generating adaptive explanations which can be customized to account for subgroups of interest and user profiles. Second, this proposal will facilitate the explanations as an interactive communication process by dynamically incorporating user inputs. Finally, this proposal will improve existing automatic evaluation metrics such as sufficiency and comprehensiveness, and develop novel ones, especially for the understudied global explanations. The team will embed these computational approaches in real-world systems.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.

解释机器学习 (ML) 模型因其在医疗保健、招聘和刑事司法等社会关键任务中的采用而受到越来越多的关注。对于决策者和决策主体等相关方来说，理解模型为何做出特定预测至关重要。该提案认为，解释代表了一个沟通过程。为了提高解释的有效性，解释应该根据所解释的主题（感兴趣的子群体）以及目标受众（用户概况）进行适应性和交互性，目标受众的知识和偏好可能会不断变化。因此，该提案旨在开发机器学习模型的自适应和交互式解释，这将使人们更好地理解为它们做出的决策以及关于它们的决策。该提案有三个重点关注领域。首先，该提案将开发一个新颖的正式框架，用于生成自适应解释，可以对其进行定制以考虑感兴趣的子群体和用户配置文件。 其次，该提案将通过动态合并用户输入来促进作为交互式通信过程的解释。最后，该提案将改进现有的自动评估指标，例如充分性和全面性，并开发新的指标，特别是针对尚未充分研究的全局解释。该团队将把这些计算方法嵌入到现实世界的系统中。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

TalkToModel: Explaining Machine Learning Models with Interactive Natural Language Conversations.

TalkToModel：通过交互式自然语言对话解释机器学习模型。

• 发表时间: 2023-07
• 期刊: Nature machine intelligence
• 影响因子: 23.8
• 作者: Slack, Dylan;Krishna, Satyapriya;Lakkaraju, Himabindu;Singh, Sameer.
• 通讯作者: Singh, Sameer.

Towards Bridging the Gaps between the Right to Explanation and the Right to be Forgotten.

弥合解释权和被遗忘权之间的差距。

• 发表时间: 2023-06
• 期刊: International Conference on Machine Learning
• 作者: Krishna, Satyapriya;Ma, Jiaqi;Lakkaraju, Himabindu
• 通讯作者: Lakkaraju, Himabindu

Machine Explanations and Human Understanding

机器解释和人类理解

• DOI: 10.1145/3593013.3593970
• 发表时间: 2022-02-08
• 期刊: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency
• 作者: Chacha Chen;Shi Feng;Amit Sharma;Chenhao Tan
• 通讯作者: Chenhao Tan

Reliable Post hoc Explanations: Modeling Uncertainty in Explainability

可靠的事后解释：对可解释性的不确定性进行建模

• 发表时间: 2021-01
• 期刊: Advances in Neural Information Processing Systems (NeurIPS
• 作者: Slack, Dylan;Hilgard, Anna;Singh, Sameer;Lakkaraju, Himabindu
• 通讯作者: Lakkaraju, Himabindu

Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post hoc Explanations.

我应该选择哪种解释？

• 发表时间: 2022-10
• 期刊: Advances in neural information processing systems
• 作者: Han, Tessa;Srinivas, Suraj;Lakkaraju, Himabindu
• 通讯作者: Lakkaraju, Himabindu