NSF-BSF: RI: Small: Structured Distributions in Deep Nets

这项研究的目标是开发改善人工智能系统与人类之间互动的方法。人工智能系统在决策过程中越来越普遍，例如，在避免交通堵塞的导航系统中。然而，虽然人们可以向同行询问建议及其背后的原因，但今天的人工智能系统通常无法为其输出提供理由。因此，用户必须1）盲目地信任系统推荐，2）单独地验证可验证性，或者3）忽略推荐。为了解决这三种选择都不可取的局限性，研究开发了可以解释其输出的模型，研究开发了可以控制的算法，研究开发了允许与模型交互的方法。受人类在提出建议时专注于数据子集的启发，该研究试图通过提取数据中提供最多证据的部分来获得解释能力，控制能力和交互能力。为此，我们使用AI系统内部的概率分布。此外，这项研究将支持伊利诺伊大学厄巴纳-香槟分校的一批博士生和本科生的发展，在当地社区的推广活动，以及两个课程的开发：一个关于入门级机器学习的新型本科生课程和一个关于AI系统中分布的新型格拉德课程。从技术上讲，AI系统中的分布通常被称为注意力。注意力提供了一个令人信服的框架：1）解释歧视网络中形成的决策; 2）控制生成模型中的采样过程; 3）在强化学习系统中进行交互。本研究的技术目标分为三个方面。第一个推力将注意力机制扩展到包括多种模态的数据，并开发出更好地捕捉这些高维环境中的概率分布的算法。第二个推力将这些算法推广到更复杂的数据结构，并将这些结果用于生成高维输出的AI系统，例如，对图像的描述。该奖项反映了NSF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。

The goal of this research is to develop methods that improve interaction between AI systems and humans. AI systems are increasingly more prevalent in decision making processes, for instance, in navigation systems for avoiding traffic jams. However, while one could ask a peer for a recommendation and the reason behind it, today’s AI systems often cannot provide a justification for their output. Hence, users have to 1) trust the system recommendation blindly, 2) verify plausibility individually, or 3) ignore the recommendation. To address the limitation that none of those three options is desirable, the research develops models which can explain their output, the research develops algorithms which can be controlled, and the research develops methods which permit interaction with the model. Inspired by a human focusing on subsets of the data when making a recommendation, the research seeks to obtain explain-ability, control-ability and interact-ability by extracting which parts of the data provided most evidence. For this we use probability distributions inside AI systems. Furthermore, this research will support development of a cohort of PhD and undergraduate students at the University of Illinois at Urbana-Champaign, outreach activities in the local neighborhood and development of two classes: a novel undergrad class on entry-level machine learning and a novel grad class on distributions in AI systems.Technically, distributions inside AI systems are often referred to as attention. Attention provides a compelling framework 1) to explain the decisions formed in discriminative networks; 2) to control the sampling process in generative models; and 3) to interact in reinforcement learning systems. The technical aims of this research are divided into three thrusts. The first thrust scales attention mechanisms to data that comprises multiple modalities and develops algorithms which better capture probability distributions in those high-dimensional settings. The second thrust generalizes those algorithms to more complex data structures and leverages those results for AI systems which generate high-dimensional output, e.g., a description of an image. The third thrust studies interaction between humans and AI systems that leverage distributions.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.